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
)
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 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
, unsigned long *flags
)
201 __acquires(p
->pi_lock
)
207 raw_spin_lock_irqsave(&p
->pi_lock
, *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 lockdep_pin_lock(&rq
->lock
);
230 raw_spin_unlock(&rq
->lock
);
231 raw_spin_unlock_irqrestore(&p
->pi_lock
, *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;
318 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
320 int cpu
= (int)(long)hcpu
;
323 case CPU_UP_CANCELED
:
324 case CPU_UP_CANCELED_FROZEN
:
325 case CPU_DOWN_PREPARE
:
326 case CPU_DOWN_PREPARE_FROZEN
:
328 case CPU_DEAD_FROZEN
:
329 hrtick_clear(cpu_rq(cpu
));
336 static __init
void init_hrtick(void)
338 hotcpu_notifier(hotplug_hrtick
, 0);
342 * Called to set the hrtick timer state.
344 * called with rq->lock held and irqs disabled
346 void hrtick_start(struct rq
*rq
, u64 delay
)
349 * Don't schedule slices shorter than 10000ns, that just
350 * doesn't make sense. Rely on vruntime for fairness.
352 delay
= max_t(u64
, delay
, 10000LL);
353 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
354 HRTIMER_MODE_REL_PINNED
);
357 static inline void init_hrtick(void)
360 #endif /* CONFIG_SMP */
362 static void init_rq_hrtick(struct rq
*rq
)
365 rq
->hrtick_csd_pending
= 0;
367 rq
->hrtick_csd
.flags
= 0;
368 rq
->hrtick_csd
.func
= __hrtick_start
;
369 rq
->hrtick_csd
.info
= rq
;
372 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
373 rq
->hrtick_timer
.function
= hrtick
;
375 #else /* CONFIG_SCHED_HRTICK */
376 static inline void hrtick_clear(struct rq
*rq
)
380 static inline void init_rq_hrtick(struct rq
*rq
)
384 static inline void init_hrtick(void)
387 #endif /* CONFIG_SCHED_HRTICK */
390 * cmpxchg based fetch_or, macro so it works for different integer types
392 #define fetch_or(ptr, mask) \
394 typeof(ptr) _ptr = (ptr); \
395 typeof(mask) _mask = (mask); \
396 typeof(*_ptr) _old, _val = *_ptr; \
399 _old = cmpxchg(_ptr, _val, _val | _mask); \
407 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
409 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
410 * this avoids any races wrt polling state changes and thereby avoids
413 static bool set_nr_and_not_polling(struct task_struct
*p
)
415 struct thread_info
*ti
= task_thread_info(p
);
416 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
420 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
422 * If this returns true, then the idle task promises to call
423 * sched_ttwu_pending() and reschedule soon.
425 static bool set_nr_if_polling(struct task_struct
*p
)
427 struct thread_info
*ti
= task_thread_info(p
);
428 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
431 if (!(val
& _TIF_POLLING_NRFLAG
))
433 if (val
& _TIF_NEED_RESCHED
)
435 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
444 static bool set_nr_and_not_polling(struct task_struct
*p
)
446 set_tsk_need_resched(p
);
451 static bool set_nr_if_polling(struct task_struct
*p
)
458 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
460 struct wake_q_node
*node
= &task
->wake_q
;
463 * Atomically grab the task, if ->wake_q is !nil already it means
464 * its already queued (either by us or someone else) and will get the
465 * wakeup due to that.
467 * This cmpxchg() implies a full barrier, which pairs with the write
468 * barrier implied by the wakeup in wake_up_list().
470 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
473 get_task_struct(task
);
476 * The head is context local, there can be no concurrency.
479 head
->lastp
= &node
->next
;
482 void wake_up_q(struct wake_q_head
*head
)
484 struct wake_q_node
*node
= head
->first
;
486 while (node
!= WAKE_Q_TAIL
) {
487 struct task_struct
*task
;
489 task
= container_of(node
, struct task_struct
, wake_q
);
491 /* task can safely be re-inserted now */
493 task
->wake_q
.next
= NULL
;
496 * wake_up_process() implies a wmb() to pair with the queueing
497 * in wake_q_add() so as not to miss wakeups.
499 wake_up_process(task
);
500 put_task_struct(task
);
505 * resched_curr - mark rq's current task 'to be rescheduled now'.
507 * On UP this means the setting of the need_resched flag, on SMP it
508 * might also involve a cross-CPU call to trigger the scheduler on
511 void resched_curr(struct rq
*rq
)
513 struct task_struct
*curr
= rq
->curr
;
516 lockdep_assert_held(&rq
->lock
);
518 if (test_tsk_need_resched(curr
))
523 if (cpu
== smp_processor_id()) {
524 set_tsk_need_resched(curr
);
525 set_preempt_need_resched();
529 if (set_nr_and_not_polling(curr
))
530 smp_send_reschedule(cpu
);
532 trace_sched_wake_idle_without_ipi(cpu
);
535 void resched_cpu(int cpu
)
537 struct rq
*rq
= cpu_rq(cpu
);
540 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
543 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int i
, cpu
= smp_processor_id();
559 struct sched_domain
*sd
;
561 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
565 for_each_domain(cpu
, sd
) {
566 for_each_cpu(i
, sched_domain_span(sd
)) {
567 if (!idle_cpu(i
) && is_housekeeping_cpu(cpu
)) {
574 if (!is_housekeeping_cpu(cpu
))
575 cpu
= housekeeping_any_cpu();
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 static void wake_up_idle_cpu(int cpu
)
592 struct rq
*rq
= cpu_rq(cpu
);
594 if (cpu
== smp_processor_id())
597 if (set_nr_and_not_polling(rq
->idle
))
598 smp_send_reschedule(cpu
);
600 trace_sched_wake_idle_without_ipi(cpu
);
603 static bool wake_up_full_nohz_cpu(int cpu
)
606 * We just need the target to call irq_exit() and re-evaluate
607 * the next tick. The nohz full kick at least implies that.
608 * If needed we can still optimize that later with an
611 if (tick_nohz_full_cpu(cpu
)) {
612 if (cpu
!= smp_processor_id() ||
613 tick_nohz_tick_stopped())
614 tick_nohz_full_kick_cpu(cpu
);
621 void wake_up_nohz_cpu(int cpu
)
623 if (!wake_up_full_nohz_cpu(cpu
))
624 wake_up_idle_cpu(cpu
);
627 static inline bool got_nohz_idle_kick(void)
629 int cpu
= smp_processor_id();
631 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
634 if (idle_cpu(cpu
) && !need_resched())
638 * We can't run Idle Load Balance on this CPU for this time so we
639 * cancel it and clear NOHZ_BALANCE_KICK
641 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
645 #else /* CONFIG_NO_HZ_COMMON */
647 static inline bool got_nohz_idle_kick(void)
652 #endif /* CONFIG_NO_HZ_COMMON */
654 #ifdef CONFIG_NO_HZ_FULL
655 bool sched_can_stop_tick(struct rq
*rq
)
659 /* Deadline tasks, even if single, need the tick */
660 if (rq
->dl
.dl_nr_running
)
664 * If there are more than one RR tasks, we need the tick to effect the
665 * actual RR behaviour.
667 if (rq
->rt
.rr_nr_running
) {
668 if (rq
->rt
.rr_nr_running
== 1)
675 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
676 * forced preemption between FIFO tasks.
678 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
683 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
684 * if there's more than one we need the tick for involuntary
687 if (rq
->nr_running
> 1)
692 #endif /* CONFIG_NO_HZ_FULL */
694 void sched_avg_update(struct rq
*rq
)
696 s64 period
= sched_avg_period();
698 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
700 * Inline assembly required to prevent the compiler
701 * optimising this loop into a divmod call.
702 * See __iter_div_u64_rem() for another example of this.
704 asm("" : "+rm" (rq
->age_stamp
));
705 rq
->age_stamp
+= period
;
710 #endif /* CONFIG_SMP */
712 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
713 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
715 * Iterate task_group tree rooted at *from, calling @down when first entering a
716 * node and @up when leaving it for the final time.
718 * Caller must hold rcu_lock or sufficient equivalent.
720 int walk_tg_tree_from(struct task_group
*from
,
721 tg_visitor down
, tg_visitor up
, void *data
)
723 struct task_group
*parent
, *child
;
729 ret
= (*down
)(parent
, data
);
732 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
739 ret
= (*up
)(parent
, data
);
740 if (ret
|| parent
== from
)
744 parent
= parent
->parent
;
751 int tg_nop(struct task_group
*tg
, void *data
)
757 static void set_load_weight(struct task_struct
*p
)
759 int prio
= p
->static_prio
- MAX_RT_PRIO
;
760 struct load_weight
*load
= &p
->se
.load
;
763 * SCHED_IDLE tasks get minimal weight:
765 if (idle_policy(p
->policy
)) {
766 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
767 load
->inv_weight
= WMULT_IDLEPRIO
;
771 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
772 load
->inv_weight
= sched_prio_to_wmult
[prio
];
775 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
778 if (!(flags
& ENQUEUE_RESTORE
))
779 sched_info_queued(rq
, p
);
780 p
->sched_class
->enqueue_task(rq
, p
, flags
);
783 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
786 if (!(flags
& DEQUEUE_SAVE
))
787 sched_info_dequeued(rq
, p
);
788 p
->sched_class
->dequeue_task(rq
, p
, flags
);
791 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
793 if (task_contributes_to_load(p
))
794 rq
->nr_uninterruptible
--;
796 enqueue_task(rq
, p
, flags
);
799 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
801 if (task_contributes_to_load(p
))
802 rq
->nr_uninterruptible
++;
804 dequeue_task(rq
, p
, flags
);
807 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
810 * In theory, the compile should just see 0 here, and optimize out the call
811 * to sched_rt_avg_update. But I don't trust it...
813 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
814 s64 steal
= 0, irq_delta
= 0;
816 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
817 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
820 * Since irq_time is only updated on {soft,}irq_exit, we might run into
821 * this case when a previous update_rq_clock() happened inside a
824 * When this happens, we stop ->clock_task and only update the
825 * prev_irq_time stamp to account for the part that fit, so that a next
826 * update will consume the rest. This ensures ->clock_task is
829 * It does however cause some slight miss-attribution of {soft,}irq
830 * time, a more accurate solution would be to update the irq_time using
831 * the current rq->clock timestamp, except that would require using
834 if (irq_delta
> delta
)
837 rq
->prev_irq_time
+= irq_delta
;
840 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
841 if (static_key_false((¶virt_steal_rq_enabled
))) {
842 steal
= paravirt_steal_clock(cpu_of(rq
));
843 steal
-= rq
->prev_steal_time_rq
;
845 if (unlikely(steal
> delta
))
848 rq
->prev_steal_time_rq
+= steal
;
853 rq
->clock_task
+= delta
;
855 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
856 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
857 sched_rt_avg_update(rq
, irq_delta
+ steal
);
861 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
863 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
864 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
868 * Make it appear like a SCHED_FIFO task, its something
869 * userspace knows about and won't get confused about.
871 * Also, it will make PI more or less work without too
872 * much confusion -- but then, stop work should not
873 * rely on PI working anyway.
875 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
877 stop
->sched_class
= &stop_sched_class
;
880 cpu_rq(cpu
)->stop
= stop
;
884 * Reset it back to a normal scheduling class so that
885 * it can die in pieces.
887 old_stop
->sched_class
= &rt_sched_class
;
892 * __normal_prio - return the priority that is based on the static prio
894 static inline int __normal_prio(struct task_struct
*p
)
896 return p
->static_prio
;
900 * Calculate the expected normal priority: i.e. priority
901 * without taking RT-inheritance into account. Might be
902 * boosted by interactivity modifiers. Changes upon fork,
903 * setprio syscalls, and whenever the interactivity
904 * estimator recalculates.
906 static inline int normal_prio(struct task_struct
*p
)
910 if (task_has_dl_policy(p
))
911 prio
= MAX_DL_PRIO
-1;
912 else if (task_has_rt_policy(p
))
913 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
915 prio
= __normal_prio(p
);
920 * Calculate the current priority, i.e. the priority
921 * taken into account by the scheduler. This value might
922 * be boosted by RT tasks, or might be boosted by
923 * interactivity modifiers. Will be RT if the task got
924 * RT-boosted. If not then it returns p->normal_prio.
926 static int effective_prio(struct task_struct
*p
)
928 p
->normal_prio
= normal_prio(p
);
930 * If we are RT tasks or we were boosted to RT priority,
931 * keep the priority unchanged. Otherwise, update priority
932 * to the normal priority:
934 if (!rt_prio(p
->prio
))
935 return p
->normal_prio
;
940 * task_curr - is this task currently executing on a CPU?
941 * @p: the task in question.
943 * Return: 1 if the task is currently executing. 0 otherwise.
945 inline int task_curr(const struct task_struct
*p
)
947 return cpu_curr(task_cpu(p
)) == p
;
951 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
952 * use the balance_callback list if you want balancing.
954 * this means any call to check_class_changed() must be followed by a call to
955 * balance_callback().
957 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
958 const struct sched_class
*prev_class
,
961 if (prev_class
!= p
->sched_class
) {
962 if (prev_class
->switched_from
)
963 prev_class
->switched_from(rq
, p
);
965 p
->sched_class
->switched_to(rq
, p
);
966 } else if (oldprio
!= p
->prio
|| dl_task(p
))
967 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
970 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
972 const struct sched_class
*class;
974 if (p
->sched_class
== rq
->curr
->sched_class
) {
975 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
977 for_each_class(class) {
978 if (class == rq
->curr
->sched_class
)
980 if (class == p
->sched_class
) {
988 * A queue event has occurred, and we're going to schedule. In
989 * this case, we can save a useless back to back clock update.
991 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
992 rq_clock_skip_update(rq
, true);
997 * This is how migration works:
999 * 1) we invoke migration_cpu_stop() on the target CPU using
1001 * 2) stopper starts to run (implicitly forcing the migrated thread
1003 * 3) it checks whether the migrated task is still in the wrong runqueue.
1004 * 4) if it's in the wrong runqueue then the migration thread removes
1005 * it and puts it into the right queue.
1006 * 5) stopper completes and stop_one_cpu() returns and the migration
1011 * move_queued_task - move a queued task to new rq.
1013 * Returns (locked) new rq. Old rq's lock is released.
1015 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
1017 lockdep_assert_held(&rq
->lock
);
1019 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1020 dequeue_task(rq
, p
, 0);
1021 set_task_cpu(p
, new_cpu
);
1022 raw_spin_unlock(&rq
->lock
);
1024 rq
= cpu_rq(new_cpu
);
1026 raw_spin_lock(&rq
->lock
);
1027 BUG_ON(task_cpu(p
) != new_cpu
);
1028 enqueue_task(rq
, p
, 0);
1029 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1030 check_preempt_curr(rq
, p
, 0);
1035 struct migration_arg
{
1036 struct task_struct
*task
;
1041 * Move (not current) task off this cpu, onto dest cpu. We're doing
1042 * this because either it can't run here any more (set_cpus_allowed()
1043 * away from this CPU, or CPU going down), or because we're
1044 * attempting to rebalance this task on exec (sched_exec).
1046 * So we race with normal scheduler movements, but that's OK, as long
1047 * as the task is no longer on this CPU.
1049 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1051 if (unlikely(!cpu_active(dest_cpu
)))
1054 /* Affinity changed (again). */
1055 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1058 rq
= move_queued_task(rq
, p
, dest_cpu
);
1064 * migration_cpu_stop - this will be executed by a highprio stopper thread
1065 * and performs thread migration by bumping thread off CPU then
1066 * 'pushing' onto another runqueue.
1068 static int migration_cpu_stop(void *data
)
1070 struct migration_arg
*arg
= data
;
1071 struct task_struct
*p
= arg
->task
;
1072 struct rq
*rq
= this_rq();
1075 * The original target cpu might have gone down and we might
1076 * be on another cpu but it doesn't matter.
1078 local_irq_disable();
1080 * We need to explicitly wake pending tasks before running
1081 * __migrate_task() such that we will not miss enforcing cpus_allowed
1082 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1084 sched_ttwu_pending();
1086 raw_spin_lock(&p
->pi_lock
);
1087 raw_spin_lock(&rq
->lock
);
1089 * If task_rq(p) != rq, it cannot be migrated here, because we're
1090 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1091 * we're holding p->pi_lock.
1093 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1094 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1095 raw_spin_unlock(&rq
->lock
);
1096 raw_spin_unlock(&p
->pi_lock
);
1103 * sched_class::set_cpus_allowed must do the below, but is not required to
1104 * actually call this function.
1106 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1108 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1109 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1112 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1114 struct rq
*rq
= task_rq(p
);
1115 bool queued
, running
;
1117 lockdep_assert_held(&p
->pi_lock
);
1119 queued
= task_on_rq_queued(p
);
1120 running
= task_current(rq
, p
);
1124 * Because __kthread_bind() calls this on blocked tasks without
1127 lockdep_assert_held(&rq
->lock
);
1128 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1131 put_prev_task(rq
, p
);
1133 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1136 p
->sched_class
->set_curr_task(rq
);
1138 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1142 * Change a given task's CPU affinity. Migrate the thread to a
1143 * proper CPU and schedule it away if the CPU it's executing on
1144 * is removed from the allowed bitmask.
1146 * NOTE: the caller must have a valid reference to the task, the
1147 * task must not exit() & deallocate itself prematurely. The
1148 * call is not atomic; no spinlocks may be held.
1150 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1151 const struct cpumask
*new_mask
, bool check
)
1153 unsigned long flags
;
1155 unsigned int dest_cpu
;
1158 rq
= task_rq_lock(p
, &flags
);
1161 * Must re-check here, to close a race against __kthread_bind(),
1162 * sched_setaffinity() is not guaranteed to observe the flag.
1164 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1169 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1172 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
1177 do_set_cpus_allowed(p
, new_mask
);
1179 /* Can the task run on the task's current CPU? If so, we're done */
1180 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1183 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
1184 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1185 struct migration_arg arg
= { p
, dest_cpu
};
1186 /* Need help from migration thread: drop lock and wait. */
1187 task_rq_unlock(rq
, p
, &flags
);
1188 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1189 tlb_migrate_finish(p
->mm
);
1191 } else if (task_on_rq_queued(p
)) {
1193 * OK, since we're going to drop the lock immediately
1194 * afterwards anyway.
1196 lockdep_unpin_lock(&rq
->lock
);
1197 rq
= move_queued_task(rq
, p
, dest_cpu
);
1198 lockdep_pin_lock(&rq
->lock
);
1201 task_rq_unlock(rq
, p
, &flags
);
1206 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1208 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1210 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1212 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1214 #ifdef CONFIG_SCHED_DEBUG
1216 * We should never call set_task_cpu() on a blocked task,
1217 * ttwu() will sort out the placement.
1219 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1223 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1224 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1225 * time relying on p->on_rq.
1227 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1228 p
->sched_class
== &fair_sched_class
&&
1229 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1231 #ifdef CONFIG_LOCKDEP
1233 * The caller should hold either p->pi_lock or rq->lock, when changing
1234 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1236 * sched_move_task() holds both and thus holding either pins the cgroup,
1239 * Furthermore, all task_rq users should acquire both locks, see
1242 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1243 lockdep_is_held(&task_rq(p
)->lock
)));
1247 trace_sched_migrate_task(p
, new_cpu
);
1249 if (task_cpu(p
) != new_cpu
) {
1250 if (p
->sched_class
->migrate_task_rq
)
1251 p
->sched_class
->migrate_task_rq(p
);
1252 p
->se
.nr_migrations
++;
1253 perf_event_task_migrate(p
);
1256 __set_task_cpu(p
, new_cpu
);
1259 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1261 if (task_on_rq_queued(p
)) {
1262 struct rq
*src_rq
, *dst_rq
;
1264 src_rq
= task_rq(p
);
1265 dst_rq
= cpu_rq(cpu
);
1267 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1268 deactivate_task(src_rq
, p
, 0);
1269 set_task_cpu(p
, cpu
);
1270 activate_task(dst_rq
, p
, 0);
1271 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1272 check_preempt_curr(dst_rq
, p
, 0);
1275 * Task isn't running anymore; make it appear like we migrated
1276 * it before it went to sleep. This means on wakeup we make the
1277 * previous cpu our targer instead of where it really is.
1283 struct migration_swap_arg
{
1284 struct task_struct
*src_task
, *dst_task
;
1285 int src_cpu
, dst_cpu
;
1288 static int migrate_swap_stop(void *data
)
1290 struct migration_swap_arg
*arg
= data
;
1291 struct rq
*src_rq
, *dst_rq
;
1294 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1297 src_rq
= cpu_rq(arg
->src_cpu
);
1298 dst_rq
= cpu_rq(arg
->dst_cpu
);
1300 double_raw_lock(&arg
->src_task
->pi_lock
,
1301 &arg
->dst_task
->pi_lock
);
1302 double_rq_lock(src_rq
, dst_rq
);
1304 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1307 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1310 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1313 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1316 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1317 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1322 double_rq_unlock(src_rq
, dst_rq
);
1323 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1324 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1330 * Cross migrate two tasks
1332 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1334 struct migration_swap_arg arg
;
1337 arg
= (struct migration_swap_arg
){
1339 .src_cpu
= task_cpu(cur
),
1341 .dst_cpu
= task_cpu(p
),
1344 if (arg
.src_cpu
== arg
.dst_cpu
)
1348 * These three tests are all lockless; this is OK since all of them
1349 * will be re-checked with proper locks held further down the line.
1351 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1354 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1357 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1360 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1361 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1368 * wait_task_inactive - wait for a thread to unschedule.
1370 * If @match_state is nonzero, it's the @p->state value just checked and
1371 * not expected to change. If it changes, i.e. @p might have woken up,
1372 * then return zero. When we succeed in waiting for @p to be off its CPU,
1373 * we return a positive number (its total switch count). If a second call
1374 * a short while later returns the same number, the caller can be sure that
1375 * @p has remained unscheduled the whole time.
1377 * The caller must ensure that the task *will* unschedule sometime soon,
1378 * else this function might spin for a *long* time. This function can't
1379 * be called with interrupts off, or it may introduce deadlock with
1380 * smp_call_function() if an IPI is sent by the same process we are
1381 * waiting to become inactive.
1383 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1385 unsigned long flags
;
1386 int running
, queued
;
1392 * We do the initial early heuristics without holding
1393 * any task-queue locks at all. We'll only try to get
1394 * the runqueue lock when things look like they will
1400 * If the task is actively running on another CPU
1401 * still, just relax and busy-wait without holding
1404 * NOTE! Since we don't hold any locks, it's not
1405 * even sure that "rq" stays as the right runqueue!
1406 * But we don't care, since "task_running()" will
1407 * return false if the runqueue has changed and p
1408 * is actually now running somewhere else!
1410 while (task_running(rq
, p
)) {
1411 if (match_state
&& unlikely(p
->state
!= match_state
))
1417 * Ok, time to look more closely! We need the rq
1418 * lock now, to be *sure*. If we're wrong, we'll
1419 * just go back and repeat.
1421 rq
= task_rq_lock(p
, &flags
);
1422 trace_sched_wait_task(p
);
1423 running
= task_running(rq
, p
);
1424 queued
= task_on_rq_queued(p
);
1426 if (!match_state
|| p
->state
== match_state
)
1427 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1428 task_rq_unlock(rq
, p
, &flags
);
1431 * If it changed from the expected state, bail out now.
1433 if (unlikely(!ncsw
))
1437 * Was it really running after all now that we
1438 * checked with the proper locks actually held?
1440 * Oops. Go back and try again..
1442 if (unlikely(running
)) {
1448 * It's not enough that it's not actively running,
1449 * it must be off the runqueue _entirely_, and not
1452 * So if it was still runnable (but just not actively
1453 * running right now), it's preempted, and we should
1454 * yield - it could be a while.
1456 if (unlikely(queued
)) {
1457 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1459 set_current_state(TASK_UNINTERRUPTIBLE
);
1460 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1465 * Ahh, all good. It wasn't running, and it wasn't
1466 * runnable, which means that it will never become
1467 * running in the future either. We're all done!
1476 * kick_process - kick a running thread to enter/exit the kernel
1477 * @p: the to-be-kicked thread
1479 * Cause a process which is running on another CPU to enter
1480 * kernel-mode, without any delay. (to get signals handled.)
1482 * NOTE: this function doesn't have to take the runqueue lock,
1483 * because all it wants to ensure is that the remote task enters
1484 * the kernel. If the IPI races and the task has been migrated
1485 * to another CPU then no harm is done and the purpose has been
1488 void kick_process(struct task_struct
*p
)
1494 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1495 smp_send_reschedule(cpu
);
1498 EXPORT_SYMBOL_GPL(kick_process
);
1501 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1503 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1505 int nid
= cpu_to_node(cpu
);
1506 const struct cpumask
*nodemask
= NULL
;
1507 enum { cpuset
, possible
, fail
} state
= cpuset
;
1511 * If the node that the cpu is on has been offlined, cpu_to_node()
1512 * will return -1. There is no cpu on the node, and we should
1513 * select the cpu on the other node.
1516 nodemask
= cpumask_of_node(nid
);
1518 /* Look for allowed, online CPU in same node. */
1519 for_each_cpu(dest_cpu
, nodemask
) {
1520 if (!cpu_online(dest_cpu
))
1522 if (!cpu_active(dest_cpu
))
1524 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1530 /* Any allowed, online CPU? */
1531 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1532 if (!cpu_online(dest_cpu
))
1534 if (!cpu_active(dest_cpu
))
1539 /* No more Mr. Nice Guy. */
1542 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1543 cpuset_cpus_allowed_fallback(p
);
1549 do_set_cpus_allowed(p
, cpu_possible_mask
);
1560 if (state
!= cpuset
) {
1562 * Don't tell them about moving exiting tasks or
1563 * kernel threads (both mm NULL), since they never
1566 if (p
->mm
&& printk_ratelimit()) {
1567 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1568 task_pid_nr(p
), p
->comm
, cpu
);
1576 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1579 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1581 lockdep_assert_held(&p
->pi_lock
);
1583 if (p
->nr_cpus_allowed
> 1)
1584 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1587 * In order not to call set_task_cpu() on a blocking task we need
1588 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1591 * Since this is common to all placement strategies, this lives here.
1593 * [ this allows ->select_task() to simply return task_cpu(p) and
1594 * not worry about this generic constraint ]
1596 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1598 cpu
= select_fallback_rq(task_cpu(p
), p
);
1603 static void update_avg(u64
*avg
, u64 sample
)
1605 s64 diff
= sample
- *avg
;
1611 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1612 const struct cpumask
*new_mask
, bool check
)
1614 return set_cpus_allowed_ptr(p
, new_mask
);
1617 #endif /* CONFIG_SMP */
1620 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1622 #ifdef CONFIG_SCHEDSTATS
1623 struct rq
*rq
= this_rq();
1626 int this_cpu
= smp_processor_id();
1628 if (cpu
== this_cpu
) {
1629 schedstat_inc(rq
, ttwu_local
);
1630 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1632 struct sched_domain
*sd
;
1634 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1636 for_each_domain(this_cpu
, sd
) {
1637 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1638 schedstat_inc(sd
, ttwu_wake_remote
);
1645 if (wake_flags
& WF_MIGRATED
)
1646 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1648 #endif /* CONFIG_SMP */
1650 schedstat_inc(rq
, ttwu_count
);
1651 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1653 if (wake_flags
& WF_SYNC
)
1654 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1656 #endif /* CONFIG_SCHEDSTATS */
1659 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1661 activate_task(rq
, p
, en_flags
);
1662 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1664 /* if a worker is waking up, notify workqueue */
1665 if (p
->flags
& PF_WQ_WORKER
)
1666 wq_worker_waking_up(p
, cpu_of(rq
));
1670 * Mark the task runnable and perform wakeup-preemption.
1673 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1675 check_preempt_curr(rq
, p
, wake_flags
);
1676 p
->state
= TASK_RUNNING
;
1677 trace_sched_wakeup(p
);
1680 if (p
->sched_class
->task_woken
) {
1682 * Our task @p is fully woken up and running; so its safe to
1683 * drop the rq->lock, hereafter rq is only used for statistics.
1685 lockdep_unpin_lock(&rq
->lock
);
1686 p
->sched_class
->task_woken(rq
, p
);
1687 lockdep_pin_lock(&rq
->lock
);
1690 if (rq
->idle_stamp
) {
1691 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1692 u64 max
= 2*rq
->max_idle_balance_cost
;
1694 update_avg(&rq
->avg_idle
, delta
);
1696 if (rq
->avg_idle
> max
)
1705 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
1707 lockdep_assert_held(&rq
->lock
);
1710 if (p
->sched_contributes_to_load
)
1711 rq
->nr_uninterruptible
--;
1714 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
1715 ttwu_do_wakeup(rq
, p
, wake_flags
);
1719 * Called in case the task @p isn't fully descheduled from its runqueue,
1720 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1721 * since all we need to do is flip p->state to TASK_RUNNING, since
1722 * the task is still ->on_rq.
1724 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1729 rq
= __task_rq_lock(p
);
1730 if (task_on_rq_queued(p
)) {
1731 /* check_preempt_curr() may use rq clock */
1732 update_rq_clock(rq
);
1733 ttwu_do_wakeup(rq
, p
, wake_flags
);
1736 __task_rq_unlock(rq
);
1742 void sched_ttwu_pending(void)
1744 struct rq
*rq
= this_rq();
1745 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1746 struct task_struct
*p
;
1747 unsigned long flags
;
1752 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1753 lockdep_pin_lock(&rq
->lock
);
1756 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1757 llist
= llist_next(llist
);
1758 ttwu_do_activate(rq
, p
, 0);
1761 lockdep_unpin_lock(&rq
->lock
);
1762 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1765 void scheduler_ipi(void)
1768 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1769 * TIF_NEED_RESCHED remotely (for the first time) will also send
1772 preempt_fold_need_resched();
1774 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1778 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1779 * traditionally all their work was done from the interrupt return
1780 * path. Now that we actually do some work, we need to make sure
1783 * Some archs already do call them, luckily irq_enter/exit nest
1786 * Arguably we should visit all archs and update all handlers,
1787 * however a fair share of IPIs are still resched only so this would
1788 * somewhat pessimize the simple resched case.
1791 sched_ttwu_pending();
1794 * Check if someone kicked us for doing the nohz idle load balance.
1796 if (unlikely(got_nohz_idle_kick())) {
1797 this_rq()->idle_balance
= 1;
1798 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1803 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
1805 struct rq
*rq
= cpu_rq(cpu
);
1807 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1808 if (!set_nr_if_polling(rq
->idle
))
1809 smp_send_reschedule(cpu
);
1811 trace_sched_wake_idle_without_ipi(cpu
);
1815 void wake_up_if_idle(int cpu
)
1817 struct rq
*rq
= cpu_rq(cpu
);
1818 unsigned long flags
;
1822 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1825 if (set_nr_if_polling(rq
->idle
)) {
1826 trace_sched_wake_idle_without_ipi(cpu
);
1828 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1829 if (is_idle_task(rq
->curr
))
1830 smp_send_reschedule(cpu
);
1831 /* Else cpu is not in idle, do nothing here */
1832 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1839 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1841 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1843 #endif /* CONFIG_SMP */
1845 static void ttwu_queue(struct task_struct
*p
, int cpu
)
1847 struct rq
*rq
= cpu_rq(cpu
);
1849 #if defined(CONFIG_SMP)
1850 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1851 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1852 ttwu_queue_remote(p
, cpu
);
1857 raw_spin_lock(&rq
->lock
);
1858 lockdep_pin_lock(&rq
->lock
);
1859 ttwu_do_activate(rq
, p
, 0);
1860 lockdep_unpin_lock(&rq
->lock
);
1861 raw_spin_unlock(&rq
->lock
);
1865 * Notes on Program-Order guarantees on SMP systems.
1869 * The basic program-order guarantee on SMP systems is that when a task [t]
1870 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1871 * execution on its new cpu [c1].
1873 * For migration (of runnable tasks) this is provided by the following means:
1875 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1876 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1877 * rq(c1)->lock (if not at the same time, then in that order).
1878 * C) LOCK of the rq(c1)->lock scheduling in task
1880 * Transitivity guarantees that B happens after A and C after B.
1881 * Note: we only require RCpc transitivity.
1882 * Note: the cpu doing B need not be c0 or c1
1891 * UNLOCK rq(0)->lock
1893 * LOCK rq(0)->lock // orders against CPU0
1895 * UNLOCK rq(0)->lock
1899 * UNLOCK rq(1)->lock
1901 * LOCK rq(1)->lock // orders against CPU2
1904 * UNLOCK rq(1)->lock
1907 * BLOCKING -- aka. SLEEP + WAKEUP
1909 * For blocking we (obviously) need to provide the same guarantee as for
1910 * migration. However the means are completely different as there is no lock
1911 * chain to provide order. Instead we do:
1913 * 1) smp_store_release(X->on_cpu, 0)
1914 * 2) smp_cond_acquire(!X->on_cpu)
1918 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1920 * LOCK rq(0)->lock LOCK X->pi_lock
1923 * smp_store_release(X->on_cpu, 0);
1925 * smp_cond_acquire(!X->on_cpu);
1931 * X->state = RUNNING
1932 * UNLOCK rq(2)->lock
1934 * LOCK rq(2)->lock // orders against CPU1
1937 * UNLOCK rq(2)->lock
1940 * UNLOCK rq(0)->lock
1943 * However; for wakeups there is a second guarantee we must provide, namely we
1944 * must observe the state that lead to our wakeup. That is, not only must our
1945 * task observe its own prior state, it must also observe the stores prior to
1948 * This means that any means of doing remote wakeups must order the CPU doing
1949 * the wakeup against the CPU the task is going to end up running on. This,
1950 * however, is already required for the regular Program-Order guarantee above,
1951 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1956 * try_to_wake_up - wake up a thread
1957 * @p: the thread to be awakened
1958 * @state: the mask of task states that can be woken
1959 * @wake_flags: wake modifier flags (WF_*)
1961 * Put it on the run-queue if it's not already there. The "current"
1962 * thread is always on the run-queue (except when the actual
1963 * re-schedule is in progress), and as such you're allowed to do
1964 * the simpler "current->state = TASK_RUNNING" to mark yourself
1965 * runnable without the overhead of this.
1967 * Return: %true if @p was woken up, %false if it was already running.
1968 * or @state didn't match @p's state.
1971 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1973 unsigned long flags
;
1974 int cpu
, success
= 0;
1977 * If we are going to wake up a thread waiting for CONDITION we
1978 * need to ensure that CONDITION=1 done by the caller can not be
1979 * reordered with p->state check below. This pairs with mb() in
1980 * set_current_state() the waiting thread does.
1982 smp_mb__before_spinlock();
1983 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1984 if (!(p
->state
& state
))
1987 trace_sched_waking(p
);
1989 success
= 1; /* we're going to change ->state */
1992 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
1997 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1998 * possible to, falsely, observe p->on_cpu == 0.
2000 * One must be running (->on_cpu == 1) in order to remove oneself
2001 * from the runqueue.
2003 * [S] ->on_cpu = 1; [L] ->on_rq
2007 * [S] ->on_rq = 0; [L] ->on_cpu
2009 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2010 * from the consecutive calls to schedule(); the first switching to our
2011 * task, the second putting it to sleep.
2016 * If the owning (remote) cpu is still in the middle of schedule() with
2017 * this task as prev, wait until its done referencing the task.
2019 * Pairs with the smp_store_release() in finish_lock_switch().
2021 * This ensures that tasks getting woken will be fully ordered against
2022 * their previous state and preserve Program Order.
2024 smp_cond_acquire(!p
->on_cpu
);
2026 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2027 p
->state
= TASK_WAKING
;
2029 if (p
->sched_class
->task_waking
)
2030 p
->sched_class
->task_waking(p
);
2032 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2033 if (task_cpu(p
) != cpu
) {
2034 wake_flags
|= WF_MIGRATED
;
2035 set_task_cpu(p
, cpu
);
2037 #endif /* CONFIG_SMP */
2041 if (schedstat_enabled())
2042 ttwu_stat(p
, cpu
, wake_flags
);
2044 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2050 * try_to_wake_up_local - try to wake up a local task with rq lock held
2051 * @p: the thread to be awakened
2053 * Put @p on the run-queue if it's not already there. The caller must
2054 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2057 static void try_to_wake_up_local(struct task_struct
*p
)
2059 struct rq
*rq
= task_rq(p
);
2061 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2062 WARN_ON_ONCE(p
== current
))
2065 lockdep_assert_held(&rq
->lock
);
2067 if (!raw_spin_trylock(&p
->pi_lock
)) {
2069 * This is OK, because current is on_cpu, which avoids it being
2070 * picked for load-balance and preemption/IRQs are still
2071 * disabled avoiding further scheduler activity on it and we've
2072 * not yet picked a replacement task.
2074 lockdep_unpin_lock(&rq
->lock
);
2075 raw_spin_unlock(&rq
->lock
);
2076 raw_spin_lock(&p
->pi_lock
);
2077 raw_spin_lock(&rq
->lock
);
2078 lockdep_pin_lock(&rq
->lock
);
2081 if (!(p
->state
& TASK_NORMAL
))
2084 trace_sched_waking(p
);
2086 if (!task_on_rq_queued(p
))
2087 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2089 ttwu_do_wakeup(rq
, p
, 0);
2090 if (schedstat_enabled())
2091 ttwu_stat(p
, smp_processor_id(), 0);
2093 raw_spin_unlock(&p
->pi_lock
);
2097 * wake_up_process - Wake up a specific process
2098 * @p: The process to be woken up.
2100 * Attempt to wake up the nominated process and move it to the set of runnable
2103 * Return: 1 if the process was woken up, 0 if it was already running.
2105 * It may be assumed that this function implies a write memory barrier before
2106 * changing the task state if and only if any tasks are woken up.
2108 int wake_up_process(struct task_struct
*p
)
2110 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2112 EXPORT_SYMBOL(wake_up_process
);
2114 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2116 return try_to_wake_up(p
, state
, 0);
2120 * This function clears the sched_dl_entity static params.
2122 void __dl_clear_params(struct task_struct
*p
)
2124 struct sched_dl_entity
*dl_se
= &p
->dl
;
2126 dl_se
->dl_runtime
= 0;
2127 dl_se
->dl_deadline
= 0;
2128 dl_se
->dl_period
= 0;
2132 dl_se
->dl_throttled
= 0;
2133 dl_se
->dl_yielded
= 0;
2137 * Perform scheduler related setup for a newly forked process p.
2138 * p is forked by current.
2140 * __sched_fork() is basic setup used by init_idle() too:
2142 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2147 p
->se
.exec_start
= 0;
2148 p
->se
.sum_exec_runtime
= 0;
2149 p
->se
.prev_sum_exec_runtime
= 0;
2150 p
->se
.nr_migrations
= 0;
2152 INIT_LIST_HEAD(&p
->se
.group_node
);
2154 #ifdef CONFIG_FAIR_GROUP_SCHED
2155 p
->se
.cfs_rq
= NULL
;
2158 #ifdef CONFIG_SCHEDSTATS
2159 /* Even if schedstat is disabled, there should not be garbage */
2160 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2163 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2164 init_dl_task_timer(&p
->dl
);
2165 __dl_clear_params(p
);
2167 INIT_LIST_HEAD(&p
->rt
.run_list
);
2169 p
->rt
.time_slice
= sched_rr_timeslice
;
2173 #ifdef CONFIG_PREEMPT_NOTIFIERS
2174 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2177 #ifdef CONFIG_NUMA_BALANCING
2178 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2179 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2180 p
->mm
->numa_scan_seq
= 0;
2183 if (clone_flags
& CLONE_VM
)
2184 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2186 p
->numa_preferred_nid
= -1;
2188 p
->node_stamp
= 0ULL;
2189 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2190 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2191 p
->numa_work
.next
= &p
->numa_work
;
2192 p
->numa_faults
= NULL
;
2193 p
->last_task_numa_placement
= 0;
2194 p
->last_sum_exec_runtime
= 0;
2196 p
->numa_group
= NULL
;
2197 #endif /* CONFIG_NUMA_BALANCING */
2200 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2202 #ifdef CONFIG_NUMA_BALANCING
2204 void set_numabalancing_state(bool enabled
)
2207 static_branch_enable(&sched_numa_balancing
);
2209 static_branch_disable(&sched_numa_balancing
);
2212 #ifdef CONFIG_PROC_SYSCTL
2213 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2214 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2218 int state
= static_branch_likely(&sched_numa_balancing
);
2220 if (write
&& !capable(CAP_SYS_ADMIN
))
2225 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2229 set_numabalancing_state(state
);
2235 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2237 #ifdef CONFIG_SCHEDSTATS
2238 static void set_schedstats(bool enabled
)
2241 static_branch_enable(&sched_schedstats
);
2243 static_branch_disable(&sched_schedstats
);
2246 void force_schedstat_enabled(void)
2248 if (!schedstat_enabled()) {
2249 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2250 static_branch_enable(&sched_schedstats
);
2254 static int __init
setup_schedstats(char *str
)
2260 if (!strcmp(str
, "enable")) {
2261 set_schedstats(true);
2263 } else if (!strcmp(str
, "disable")) {
2264 set_schedstats(false);
2269 pr_warn("Unable to parse schedstats=\n");
2273 __setup("schedstats=", setup_schedstats
);
2275 #ifdef CONFIG_PROC_SYSCTL
2276 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2277 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2281 int state
= static_branch_likely(&sched_schedstats
);
2283 if (write
&& !capable(CAP_SYS_ADMIN
))
2288 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2292 set_schedstats(state
);
2299 * fork()/clone()-time setup:
2301 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2303 unsigned long flags
;
2304 int cpu
= get_cpu();
2306 __sched_fork(clone_flags
, p
);
2308 * We mark the process as running here. This guarantees that
2309 * nobody will actually run it, and a signal or other external
2310 * event cannot wake it up and insert it on the runqueue either.
2312 p
->state
= TASK_RUNNING
;
2315 * Make sure we do not leak PI boosting priority to the child.
2317 p
->prio
= current
->normal_prio
;
2320 * Revert to default priority/policy on fork if requested.
2322 if (unlikely(p
->sched_reset_on_fork
)) {
2323 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2324 p
->policy
= SCHED_NORMAL
;
2325 p
->static_prio
= NICE_TO_PRIO(0);
2327 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2328 p
->static_prio
= NICE_TO_PRIO(0);
2330 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2334 * We don't need the reset flag anymore after the fork. It has
2335 * fulfilled its duty:
2337 p
->sched_reset_on_fork
= 0;
2340 if (dl_prio(p
->prio
)) {
2343 } else if (rt_prio(p
->prio
)) {
2344 p
->sched_class
= &rt_sched_class
;
2346 p
->sched_class
= &fair_sched_class
;
2349 if (p
->sched_class
->task_fork
)
2350 p
->sched_class
->task_fork(p
);
2353 * The child is not yet in the pid-hash so no cgroup attach races,
2354 * and the cgroup is pinned to this child due to cgroup_fork()
2355 * is ran before sched_fork().
2357 * Silence PROVE_RCU.
2359 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2360 set_task_cpu(p
, cpu
);
2361 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2363 #ifdef CONFIG_SCHED_INFO
2364 if (likely(sched_info_on()))
2365 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2367 #if defined(CONFIG_SMP)
2370 init_task_preempt_count(p
);
2372 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2373 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2380 unsigned long to_ratio(u64 period
, u64 runtime
)
2382 if (runtime
== RUNTIME_INF
)
2386 * Doing this here saves a lot of checks in all
2387 * the calling paths, and returning zero seems
2388 * safe for them anyway.
2393 return div64_u64(runtime
<< 20, period
);
2397 inline struct dl_bw
*dl_bw_of(int i
)
2399 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2400 "sched RCU must be held");
2401 return &cpu_rq(i
)->rd
->dl_bw
;
2404 static inline int dl_bw_cpus(int i
)
2406 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2409 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2410 "sched RCU must be held");
2411 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2417 inline struct dl_bw
*dl_bw_of(int i
)
2419 return &cpu_rq(i
)->dl
.dl_bw
;
2422 static inline int dl_bw_cpus(int i
)
2429 * We must be sure that accepting a new task (or allowing changing the
2430 * parameters of an existing one) is consistent with the bandwidth
2431 * constraints. If yes, this function also accordingly updates the currently
2432 * allocated bandwidth to reflect the new situation.
2434 * This function is called while holding p's rq->lock.
2436 * XXX we should delay bw change until the task's 0-lag point, see
2439 static int dl_overflow(struct task_struct
*p
, int policy
,
2440 const struct sched_attr
*attr
)
2443 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2444 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2445 u64 runtime
= attr
->sched_runtime
;
2446 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2449 /* !deadline task may carry old deadline bandwidth */
2450 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2454 * Either if a task, enters, leave, or stays -deadline but changes
2455 * its parameters, we may need to update accordingly the total
2456 * allocated bandwidth of the container.
2458 raw_spin_lock(&dl_b
->lock
);
2459 cpus
= dl_bw_cpus(task_cpu(p
));
2460 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2461 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2462 __dl_add(dl_b
, new_bw
);
2464 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2465 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2466 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2467 __dl_add(dl_b
, new_bw
);
2469 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2470 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2473 raw_spin_unlock(&dl_b
->lock
);
2478 extern void init_dl_bw(struct dl_bw
*dl_b
);
2481 * wake_up_new_task - wake up a newly created task for the first time.
2483 * This function will do some initial scheduler statistics housekeeping
2484 * that must be done for every newly created context, then puts the task
2485 * on the runqueue and wakes it.
2487 void wake_up_new_task(struct task_struct
*p
)
2489 unsigned long flags
;
2492 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2493 /* Initialize new task's runnable average */
2494 init_entity_runnable_average(&p
->se
);
2497 * Fork balancing, do it here and not earlier because:
2498 * - cpus_allowed can change in the fork path
2499 * - any previously selected cpu might disappear through hotplug
2501 set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2503 /* Post initialize new task's util average when its cfs_rq is set */
2504 post_init_entity_util_avg(&p
->se
);
2506 rq
= __task_rq_lock(p
);
2507 activate_task(rq
, p
, 0);
2508 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2509 trace_sched_wakeup_new(p
);
2510 check_preempt_curr(rq
, p
, WF_FORK
);
2512 if (p
->sched_class
->task_woken
) {
2514 * Nothing relies on rq->lock after this, so its fine to
2517 lockdep_unpin_lock(&rq
->lock
);
2518 p
->sched_class
->task_woken(rq
, p
);
2519 lockdep_pin_lock(&rq
->lock
);
2522 task_rq_unlock(rq
, p
, &flags
);
2525 #ifdef CONFIG_PREEMPT_NOTIFIERS
2527 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2529 void preempt_notifier_inc(void)
2531 static_key_slow_inc(&preempt_notifier_key
);
2533 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2535 void preempt_notifier_dec(void)
2537 static_key_slow_dec(&preempt_notifier_key
);
2539 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2542 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2543 * @notifier: notifier struct to register
2545 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2547 if (!static_key_false(&preempt_notifier_key
))
2548 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2550 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2552 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2555 * preempt_notifier_unregister - no longer interested in preemption notifications
2556 * @notifier: notifier struct to unregister
2558 * This is *not* safe to call from within a preemption notifier.
2560 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2562 hlist_del(¬ifier
->link
);
2564 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2566 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2568 struct preempt_notifier
*notifier
;
2570 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2571 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2574 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2576 if (static_key_false(&preempt_notifier_key
))
2577 __fire_sched_in_preempt_notifiers(curr
);
2581 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2582 struct task_struct
*next
)
2584 struct preempt_notifier
*notifier
;
2586 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2587 notifier
->ops
->sched_out(notifier
, next
);
2590 static __always_inline
void
2591 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2592 struct task_struct
*next
)
2594 if (static_key_false(&preempt_notifier_key
))
2595 __fire_sched_out_preempt_notifiers(curr
, next
);
2598 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2600 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2605 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2606 struct task_struct
*next
)
2610 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2613 * prepare_task_switch - prepare to switch tasks
2614 * @rq: the runqueue preparing to switch
2615 * @prev: the current task that is being switched out
2616 * @next: the task we are going to switch to.
2618 * This is called with the rq lock held and interrupts off. It must
2619 * be paired with a subsequent finish_task_switch after the context
2622 * prepare_task_switch sets up locking and calls architecture specific
2626 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2627 struct task_struct
*next
)
2629 sched_info_switch(rq
, prev
, next
);
2630 perf_event_task_sched_out(prev
, next
);
2631 fire_sched_out_preempt_notifiers(prev
, next
);
2632 prepare_lock_switch(rq
, next
);
2633 prepare_arch_switch(next
);
2637 * finish_task_switch - clean up after a task-switch
2638 * @prev: the thread we just switched away from.
2640 * finish_task_switch must be called after the context switch, paired
2641 * with a prepare_task_switch call before the context switch.
2642 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2643 * and do any other architecture-specific cleanup actions.
2645 * Note that we may have delayed dropping an mm in context_switch(). If
2646 * so, we finish that here outside of the runqueue lock. (Doing it
2647 * with the lock held can cause deadlocks; see schedule() for
2650 * The context switch have flipped the stack from under us and restored the
2651 * local variables which were saved when this task called schedule() in the
2652 * past. prev == current is still correct but we need to recalculate this_rq
2653 * because prev may have moved to another CPU.
2655 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2656 __releases(rq
->lock
)
2658 struct rq
*rq
= this_rq();
2659 struct mm_struct
*mm
= rq
->prev_mm
;
2663 * The previous task will have left us with a preempt_count of 2
2664 * because it left us after:
2667 * preempt_disable(); // 1
2669 * raw_spin_lock_irq(&rq->lock) // 2
2671 * Also, see FORK_PREEMPT_COUNT.
2673 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2674 "corrupted preempt_count: %s/%d/0x%x\n",
2675 current
->comm
, current
->pid
, preempt_count()))
2676 preempt_count_set(FORK_PREEMPT_COUNT
);
2681 * A task struct has one reference for the use as "current".
2682 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2683 * schedule one last time. The schedule call will never return, and
2684 * the scheduled task must drop that reference.
2686 * We must observe prev->state before clearing prev->on_cpu (in
2687 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2688 * running on another CPU and we could rave with its RUNNING -> DEAD
2689 * transition, resulting in a double drop.
2691 prev_state
= prev
->state
;
2692 vtime_task_switch(prev
);
2693 perf_event_task_sched_in(prev
, current
);
2694 finish_lock_switch(rq
, prev
);
2695 finish_arch_post_lock_switch();
2697 fire_sched_in_preempt_notifiers(current
);
2700 if (unlikely(prev_state
== TASK_DEAD
)) {
2701 if (prev
->sched_class
->task_dead
)
2702 prev
->sched_class
->task_dead(prev
);
2705 * Remove function-return probe instances associated with this
2706 * task and put them back on the free list.
2708 kprobe_flush_task(prev
);
2709 put_task_struct(prev
);
2712 tick_nohz_task_switch();
2718 /* rq->lock is NOT held, but preemption is disabled */
2719 static void __balance_callback(struct rq
*rq
)
2721 struct callback_head
*head
, *next
;
2722 void (*func
)(struct rq
*rq
);
2723 unsigned long flags
;
2725 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2726 head
= rq
->balance_callback
;
2727 rq
->balance_callback
= NULL
;
2729 func
= (void (*)(struct rq
*))head
->func
;
2736 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2739 static inline void balance_callback(struct rq
*rq
)
2741 if (unlikely(rq
->balance_callback
))
2742 __balance_callback(rq
);
2747 static inline void balance_callback(struct rq
*rq
)
2754 * schedule_tail - first thing a freshly forked thread must call.
2755 * @prev: the thread we just switched away from.
2757 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2758 __releases(rq
->lock
)
2763 * New tasks start with FORK_PREEMPT_COUNT, see there and
2764 * finish_task_switch() for details.
2766 * finish_task_switch() will drop rq->lock() and lower preempt_count
2767 * and the preempt_enable() will end up enabling preemption (on
2768 * PREEMPT_COUNT kernels).
2771 rq
= finish_task_switch(prev
);
2772 balance_callback(rq
);
2775 if (current
->set_child_tid
)
2776 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2780 * context_switch - switch to the new MM and the new thread's register state.
2782 static __always_inline
struct rq
*
2783 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2784 struct task_struct
*next
)
2786 struct mm_struct
*mm
, *oldmm
;
2788 prepare_task_switch(rq
, prev
, next
);
2791 oldmm
= prev
->active_mm
;
2793 * For paravirt, this is coupled with an exit in switch_to to
2794 * combine the page table reload and the switch backend into
2797 arch_start_context_switch(prev
);
2800 next
->active_mm
= oldmm
;
2801 atomic_inc(&oldmm
->mm_count
);
2802 enter_lazy_tlb(oldmm
, next
);
2804 switch_mm_irqs_off(oldmm
, mm
, next
);
2807 prev
->active_mm
= NULL
;
2808 rq
->prev_mm
= oldmm
;
2811 * Since the runqueue lock will be released by the next
2812 * task (which is an invalid locking op but in the case
2813 * of the scheduler it's an obvious special-case), so we
2814 * do an early lockdep release here:
2816 lockdep_unpin_lock(&rq
->lock
);
2817 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2819 /* Here we just switch the register state and the stack. */
2820 switch_to(prev
, next
, prev
);
2823 return finish_task_switch(prev
);
2827 * nr_running and nr_context_switches:
2829 * externally visible scheduler statistics: current number of runnable
2830 * threads, total number of context switches performed since bootup.
2832 unsigned long nr_running(void)
2834 unsigned long i
, sum
= 0;
2836 for_each_online_cpu(i
)
2837 sum
+= cpu_rq(i
)->nr_running
;
2843 * Check if only the current task is running on the cpu.
2845 * Caution: this function does not check that the caller has disabled
2846 * preemption, thus the result might have a time-of-check-to-time-of-use
2847 * race. The caller is responsible to use it correctly, for example:
2849 * - from a non-preemptable section (of course)
2851 * - from a thread that is bound to a single CPU
2853 * - in a loop with very short iterations (e.g. a polling loop)
2855 bool single_task_running(void)
2857 return raw_rq()->nr_running
== 1;
2859 EXPORT_SYMBOL(single_task_running
);
2861 unsigned long long nr_context_switches(void)
2864 unsigned long long sum
= 0;
2866 for_each_possible_cpu(i
)
2867 sum
+= cpu_rq(i
)->nr_switches
;
2872 unsigned long nr_iowait(void)
2874 unsigned long i
, sum
= 0;
2876 for_each_possible_cpu(i
)
2877 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2882 unsigned long nr_iowait_cpu(int cpu
)
2884 struct rq
*this = cpu_rq(cpu
);
2885 return atomic_read(&this->nr_iowait
);
2888 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2890 struct rq
*rq
= this_rq();
2891 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2892 *load
= rq
->load
.weight
;
2898 * sched_exec - execve() is a valuable balancing opportunity, because at
2899 * this point the task has the smallest effective memory and cache footprint.
2901 void sched_exec(void)
2903 struct task_struct
*p
= current
;
2904 unsigned long flags
;
2907 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2908 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2909 if (dest_cpu
== smp_processor_id())
2912 if (likely(cpu_active(dest_cpu
))) {
2913 struct migration_arg arg
= { p
, dest_cpu
};
2915 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2916 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2920 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2925 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2926 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2928 EXPORT_PER_CPU_SYMBOL(kstat
);
2929 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2932 * Return accounted runtime for the task.
2933 * In case the task is currently running, return the runtime plus current's
2934 * pending runtime that have not been accounted yet.
2936 unsigned long long task_sched_runtime(struct task_struct
*p
)
2938 unsigned long flags
;
2942 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2944 * 64-bit doesn't need locks to atomically read a 64bit value.
2945 * So we have a optimization chance when the task's delta_exec is 0.
2946 * Reading ->on_cpu is racy, but this is ok.
2948 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2949 * If we race with it entering cpu, unaccounted time is 0. This is
2950 * indistinguishable from the read occurring a few cycles earlier.
2951 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2952 * been accounted, so we're correct here as well.
2954 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
2955 return p
->se
.sum_exec_runtime
;
2958 rq
= task_rq_lock(p
, &flags
);
2960 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2961 * project cycles that may never be accounted to this
2962 * thread, breaking clock_gettime().
2964 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
2965 update_rq_clock(rq
);
2966 p
->sched_class
->update_curr(rq
);
2968 ns
= p
->se
.sum_exec_runtime
;
2969 task_rq_unlock(rq
, p
, &flags
);
2975 * This function gets called by the timer code, with HZ frequency.
2976 * We call it with interrupts disabled.
2978 void scheduler_tick(void)
2980 int cpu
= smp_processor_id();
2981 struct rq
*rq
= cpu_rq(cpu
);
2982 struct task_struct
*curr
= rq
->curr
;
2986 raw_spin_lock(&rq
->lock
);
2987 update_rq_clock(rq
);
2988 curr
->sched_class
->task_tick(rq
, curr
, 0);
2989 cpu_load_update_active(rq
);
2990 calc_global_load_tick(rq
);
2991 raw_spin_unlock(&rq
->lock
);
2993 perf_event_task_tick();
2996 rq
->idle_balance
= idle_cpu(cpu
);
2997 trigger_load_balance(rq
);
2999 rq_last_tick_reset(rq
);
3002 #ifdef CONFIG_NO_HZ_FULL
3004 * scheduler_tick_max_deferment
3006 * Keep at least one tick per second when a single
3007 * active task is running because the scheduler doesn't
3008 * yet completely support full dynticks environment.
3010 * This makes sure that uptime, CFS vruntime, load
3011 * balancing, etc... continue to move forward, even
3012 * with a very low granularity.
3014 * Return: Maximum deferment in nanoseconds.
3016 u64
scheduler_tick_max_deferment(void)
3018 struct rq
*rq
= this_rq();
3019 unsigned long next
, now
= READ_ONCE(jiffies
);
3021 next
= rq
->last_sched_tick
+ HZ
;
3023 if (time_before_eq(next
, now
))
3026 return jiffies_to_nsecs(next
- now
);
3030 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3031 defined(CONFIG_PREEMPT_TRACER))
3033 * If the value passed in is equal to the current preempt count
3034 * then we just disabled preemption. Start timing the latency.
3036 static inline void preempt_latency_start(int val
)
3038 if (preempt_count() == val
) {
3039 unsigned long ip
= get_lock_parent_ip();
3040 #ifdef CONFIG_DEBUG_PREEMPT
3041 current
->preempt_disable_ip
= ip
;
3043 trace_preempt_off(CALLER_ADDR0
, ip
);
3047 void preempt_count_add(int val
)
3049 #ifdef CONFIG_DEBUG_PREEMPT
3053 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3056 __preempt_count_add(val
);
3057 #ifdef CONFIG_DEBUG_PREEMPT
3059 * Spinlock count overflowing soon?
3061 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3064 preempt_latency_start(val
);
3066 EXPORT_SYMBOL(preempt_count_add
);
3067 NOKPROBE_SYMBOL(preempt_count_add
);
3070 * If the value passed in equals to the current preempt count
3071 * then we just enabled preemption. Stop timing the latency.
3073 static inline void preempt_latency_stop(int val
)
3075 if (preempt_count() == val
)
3076 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3079 void preempt_count_sub(int val
)
3081 #ifdef CONFIG_DEBUG_PREEMPT
3085 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3088 * Is the spinlock portion underflowing?
3090 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3091 !(preempt_count() & PREEMPT_MASK
)))
3095 preempt_latency_stop(val
);
3096 __preempt_count_sub(val
);
3098 EXPORT_SYMBOL(preempt_count_sub
);
3099 NOKPROBE_SYMBOL(preempt_count_sub
);
3102 static inline void preempt_latency_start(int val
) { }
3103 static inline void preempt_latency_stop(int val
) { }
3107 * Print scheduling while atomic bug:
3109 static noinline
void __schedule_bug(struct task_struct
*prev
)
3111 if (oops_in_progress
)
3114 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3115 prev
->comm
, prev
->pid
, preempt_count());
3117 debug_show_held_locks(prev
);
3119 if (irqs_disabled())
3120 print_irqtrace_events(prev
);
3121 #ifdef CONFIG_DEBUG_PREEMPT
3122 if (in_atomic_preempt_off()) {
3123 pr_err("Preemption disabled at:");
3124 print_ip_sym(current
->preempt_disable_ip
);
3129 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3133 * Various schedule()-time debugging checks and statistics:
3135 static inline void schedule_debug(struct task_struct
*prev
)
3137 #ifdef CONFIG_SCHED_STACK_END_CHECK
3138 BUG_ON(task_stack_end_corrupted(prev
));
3141 if (unlikely(in_atomic_preempt_off())) {
3142 __schedule_bug(prev
);
3143 preempt_count_set(PREEMPT_DISABLED
);
3147 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3149 schedstat_inc(this_rq(), sched_count
);
3153 * Pick up the highest-prio task:
3155 static inline struct task_struct
*
3156 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
3158 const struct sched_class
*class = &fair_sched_class
;
3159 struct task_struct
*p
;
3162 * Optimization: we know that if all tasks are in
3163 * the fair class we can call that function directly:
3165 if (likely(prev
->sched_class
== class &&
3166 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3167 p
= fair_sched_class
.pick_next_task(rq
, prev
);
3168 if (unlikely(p
== RETRY_TASK
))
3171 /* assumes fair_sched_class->next == idle_sched_class */
3173 p
= idle_sched_class
.pick_next_task(rq
, prev
);
3179 for_each_class(class) {
3180 p
= class->pick_next_task(rq
, prev
);
3182 if (unlikely(p
== RETRY_TASK
))
3188 BUG(); /* the idle class will always have a runnable task */
3192 * __schedule() is the main scheduler function.
3194 * The main means of driving the scheduler and thus entering this function are:
3196 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3198 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3199 * paths. For example, see arch/x86/entry_64.S.
3201 * To drive preemption between tasks, the scheduler sets the flag in timer
3202 * interrupt handler scheduler_tick().
3204 * 3. Wakeups don't really cause entry into schedule(). They add a
3205 * task to the run-queue and that's it.
3207 * Now, if the new task added to the run-queue preempts the current
3208 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3209 * called on the nearest possible occasion:
3211 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3213 * - in syscall or exception context, at the next outmost
3214 * preempt_enable(). (this might be as soon as the wake_up()'s
3217 * - in IRQ context, return from interrupt-handler to
3218 * preemptible context
3220 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3223 * - cond_resched() call
3224 * - explicit schedule() call
3225 * - return from syscall or exception to user-space
3226 * - return from interrupt-handler to user-space
3228 * WARNING: must be called with preemption disabled!
3230 static void __sched notrace
__schedule(bool preempt
)
3232 struct task_struct
*prev
, *next
;
3233 unsigned long *switch_count
;
3237 cpu
= smp_processor_id();
3242 * do_exit() calls schedule() with preemption disabled as an exception;
3243 * however we must fix that up, otherwise the next task will see an
3244 * inconsistent (higher) preempt count.
3246 * It also avoids the below schedule_debug() test from complaining
3249 if (unlikely(prev
->state
== TASK_DEAD
))
3250 preempt_enable_no_resched_notrace();
3252 schedule_debug(prev
);
3254 if (sched_feat(HRTICK
))
3257 local_irq_disable();
3258 rcu_note_context_switch();
3261 * Make sure that signal_pending_state()->signal_pending() below
3262 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3263 * done by the caller to avoid the race with signal_wake_up().
3265 smp_mb__before_spinlock();
3266 raw_spin_lock(&rq
->lock
);
3267 lockdep_pin_lock(&rq
->lock
);
3269 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3271 switch_count
= &prev
->nivcsw
;
3272 if (!preempt
&& prev
->state
) {
3273 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3274 prev
->state
= TASK_RUNNING
;
3276 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3280 * If a worker went to sleep, notify and ask workqueue
3281 * whether it wants to wake up a task to maintain
3284 if (prev
->flags
& PF_WQ_WORKER
) {
3285 struct task_struct
*to_wakeup
;
3287 to_wakeup
= wq_worker_sleeping(prev
);
3289 try_to_wake_up_local(to_wakeup
);
3292 switch_count
= &prev
->nvcsw
;
3295 if (task_on_rq_queued(prev
))
3296 update_rq_clock(rq
);
3298 next
= pick_next_task(rq
, prev
);
3299 clear_tsk_need_resched(prev
);
3300 clear_preempt_need_resched();
3301 rq
->clock_skip_update
= 0;
3303 if (likely(prev
!= next
)) {
3308 trace_sched_switch(preempt
, prev
, next
);
3309 rq
= context_switch(rq
, prev
, next
); /* unlocks the rq */
3311 lockdep_unpin_lock(&rq
->lock
);
3312 raw_spin_unlock_irq(&rq
->lock
);
3315 balance_callback(rq
);
3317 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3319 static inline void sched_submit_work(struct task_struct
*tsk
)
3321 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3324 * If we are going to sleep and we have plugged IO queued,
3325 * make sure to submit it to avoid deadlocks.
3327 if (blk_needs_flush_plug(tsk
))
3328 blk_schedule_flush_plug(tsk
);
3331 asmlinkage __visible
void __sched
schedule(void)
3333 struct task_struct
*tsk
= current
;
3335 sched_submit_work(tsk
);
3339 sched_preempt_enable_no_resched();
3340 } while (need_resched());
3342 EXPORT_SYMBOL(schedule
);
3344 #ifdef CONFIG_CONTEXT_TRACKING
3345 asmlinkage __visible
void __sched
schedule_user(void)
3348 * If we come here after a random call to set_need_resched(),
3349 * or we have been woken up remotely but the IPI has not yet arrived,
3350 * we haven't yet exited the RCU idle mode. Do it here manually until
3351 * we find a better solution.
3353 * NB: There are buggy callers of this function. Ideally we
3354 * should warn if prev_state != CONTEXT_USER, but that will trigger
3355 * too frequently to make sense yet.
3357 enum ctx_state prev_state
= exception_enter();
3359 exception_exit(prev_state
);
3364 * schedule_preempt_disabled - called with preemption disabled
3366 * Returns with preemption disabled. Note: preempt_count must be 1
3368 void __sched
schedule_preempt_disabled(void)
3370 sched_preempt_enable_no_resched();
3375 static void __sched notrace
preempt_schedule_common(void)
3379 * Because the function tracer can trace preempt_count_sub()
3380 * and it also uses preempt_enable/disable_notrace(), if
3381 * NEED_RESCHED is set, the preempt_enable_notrace() called
3382 * by the function tracer will call this function again and
3383 * cause infinite recursion.
3385 * Preemption must be disabled here before the function
3386 * tracer can trace. Break up preempt_disable() into two
3387 * calls. One to disable preemption without fear of being
3388 * traced. The other to still record the preemption latency,
3389 * which can also be traced by the function tracer.
3391 preempt_disable_notrace();
3392 preempt_latency_start(1);
3394 preempt_latency_stop(1);
3395 preempt_enable_no_resched_notrace();
3398 * Check again in case we missed a preemption opportunity
3399 * between schedule and now.
3401 } while (need_resched());
3404 #ifdef CONFIG_PREEMPT
3406 * this is the entry point to schedule() from in-kernel preemption
3407 * off of preempt_enable. Kernel preemptions off return from interrupt
3408 * occur there and call schedule directly.
3410 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3413 * If there is a non-zero preempt_count or interrupts are disabled,
3414 * we do not want to preempt the current task. Just return..
3416 if (likely(!preemptible()))
3419 preempt_schedule_common();
3421 NOKPROBE_SYMBOL(preempt_schedule
);
3422 EXPORT_SYMBOL(preempt_schedule
);
3425 * preempt_schedule_notrace - preempt_schedule called by tracing
3427 * The tracing infrastructure uses preempt_enable_notrace to prevent
3428 * recursion and tracing preempt enabling caused by the tracing
3429 * infrastructure itself. But as tracing can happen in areas coming
3430 * from userspace or just about to enter userspace, a preempt enable
3431 * can occur before user_exit() is called. This will cause the scheduler
3432 * to be called when the system is still in usermode.
3434 * To prevent this, the preempt_enable_notrace will use this function
3435 * instead of preempt_schedule() to exit user context if needed before
3436 * calling the scheduler.
3438 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3440 enum ctx_state prev_ctx
;
3442 if (likely(!preemptible()))
3447 * Because the function tracer can trace preempt_count_sub()
3448 * and it also uses preempt_enable/disable_notrace(), if
3449 * NEED_RESCHED is set, the preempt_enable_notrace() called
3450 * by the function tracer will call this function again and
3451 * cause infinite recursion.
3453 * Preemption must be disabled here before the function
3454 * tracer can trace. Break up preempt_disable() into two
3455 * calls. One to disable preemption without fear of being
3456 * traced. The other to still record the preemption latency,
3457 * which can also be traced by the function tracer.
3459 preempt_disable_notrace();
3460 preempt_latency_start(1);
3462 * Needs preempt disabled in case user_exit() is traced
3463 * and the tracer calls preempt_enable_notrace() causing
3464 * an infinite recursion.
3466 prev_ctx
= exception_enter();
3468 exception_exit(prev_ctx
);
3470 preempt_latency_stop(1);
3471 preempt_enable_no_resched_notrace();
3472 } while (need_resched());
3474 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3476 #endif /* CONFIG_PREEMPT */
3479 * this is the entry point to schedule() from kernel preemption
3480 * off of irq context.
3481 * Note, that this is called and return with irqs disabled. This will
3482 * protect us against recursive calling from irq.
3484 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3486 enum ctx_state prev_state
;
3488 /* Catch callers which need to be fixed */
3489 BUG_ON(preempt_count() || !irqs_disabled());
3491 prev_state
= exception_enter();
3497 local_irq_disable();
3498 sched_preempt_enable_no_resched();
3499 } while (need_resched());
3501 exception_exit(prev_state
);
3504 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3507 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3509 EXPORT_SYMBOL(default_wake_function
);
3511 #ifdef CONFIG_RT_MUTEXES
3514 * rt_mutex_setprio - set the current priority of a task
3516 * @prio: prio value (kernel-internal form)
3518 * This function changes the 'effective' priority of a task. It does
3519 * not touch ->normal_prio like __setscheduler().
3521 * Used by the rt_mutex code to implement priority inheritance
3522 * logic. Call site only calls if the priority of the task changed.
3524 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3526 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3528 const struct sched_class
*prev_class
;
3530 BUG_ON(prio
> MAX_PRIO
);
3532 rq
= __task_rq_lock(p
);
3535 * Idle task boosting is a nono in general. There is one
3536 * exception, when PREEMPT_RT and NOHZ is active:
3538 * The idle task calls get_next_timer_interrupt() and holds
3539 * the timer wheel base->lock on the CPU and another CPU wants
3540 * to access the timer (probably to cancel it). We can safely
3541 * ignore the boosting request, as the idle CPU runs this code
3542 * with interrupts disabled and will complete the lock
3543 * protected section without being interrupted. So there is no
3544 * real need to boost.
3546 if (unlikely(p
== rq
->idle
)) {
3547 WARN_ON(p
!= rq
->curr
);
3548 WARN_ON(p
->pi_blocked_on
);
3552 trace_sched_pi_setprio(p
, prio
);
3555 if (oldprio
== prio
)
3556 queue_flag
&= ~DEQUEUE_MOVE
;
3558 prev_class
= p
->sched_class
;
3559 queued
= task_on_rq_queued(p
);
3560 running
= task_current(rq
, p
);
3562 dequeue_task(rq
, p
, queue_flag
);
3564 put_prev_task(rq
, p
);
3567 * Boosting condition are:
3568 * 1. -rt task is running and holds mutex A
3569 * --> -dl task blocks on mutex A
3571 * 2. -dl task is running and holds mutex A
3572 * --> -dl task blocks on mutex A and could preempt the
3575 if (dl_prio(prio
)) {
3576 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3577 if (!dl_prio(p
->normal_prio
) ||
3578 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3579 p
->dl
.dl_boosted
= 1;
3580 queue_flag
|= ENQUEUE_REPLENISH
;
3582 p
->dl
.dl_boosted
= 0;
3583 p
->sched_class
= &dl_sched_class
;
3584 } else if (rt_prio(prio
)) {
3585 if (dl_prio(oldprio
))
3586 p
->dl
.dl_boosted
= 0;
3588 queue_flag
|= ENQUEUE_HEAD
;
3589 p
->sched_class
= &rt_sched_class
;
3591 if (dl_prio(oldprio
))
3592 p
->dl
.dl_boosted
= 0;
3593 if (rt_prio(oldprio
))
3595 p
->sched_class
= &fair_sched_class
;
3601 p
->sched_class
->set_curr_task(rq
);
3603 enqueue_task(rq
, p
, queue_flag
);
3605 check_class_changed(rq
, p
, prev_class
, oldprio
);
3607 preempt_disable(); /* avoid rq from going away on us */
3608 __task_rq_unlock(rq
);
3610 balance_callback(rq
);
3615 void set_user_nice(struct task_struct
*p
, long nice
)
3617 int old_prio
, delta
, queued
;
3618 unsigned long flags
;
3621 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3624 * We have to be careful, if called from sys_setpriority(),
3625 * the task might be in the middle of scheduling on another CPU.
3627 rq
= task_rq_lock(p
, &flags
);
3629 * The RT priorities are set via sched_setscheduler(), but we still
3630 * allow the 'normal' nice value to be set - but as expected
3631 * it wont have any effect on scheduling until the task is
3632 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3634 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3635 p
->static_prio
= NICE_TO_PRIO(nice
);
3638 queued
= task_on_rq_queued(p
);
3640 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3642 p
->static_prio
= NICE_TO_PRIO(nice
);
3645 p
->prio
= effective_prio(p
);
3646 delta
= p
->prio
- old_prio
;
3649 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3651 * If the task increased its priority or is running and
3652 * lowered its priority, then reschedule its CPU:
3654 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3658 task_rq_unlock(rq
, p
, &flags
);
3660 EXPORT_SYMBOL(set_user_nice
);
3663 * can_nice - check if a task can reduce its nice value
3667 int can_nice(const struct task_struct
*p
, const int nice
)
3669 /* convert nice value [19,-20] to rlimit style value [1,40] */
3670 int nice_rlim
= nice_to_rlimit(nice
);
3672 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3673 capable(CAP_SYS_NICE
));
3676 #ifdef __ARCH_WANT_SYS_NICE
3679 * sys_nice - change the priority of the current process.
3680 * @increment: priority increment
3682 * sys_setpriority is a more generic, but much slower function that
3683 * does similar things.
3685 SYSCALL_DEFINE1(nice
, int, increment
)
3690 * Setpriority might change our priority at the same moment.
3691 * We don't have to worry. Conceptually one call occurs first
3692 * and we have a single winner.
3694 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3695 nice
= task_nice(current
) + increment
;
3697 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3698 if (increment
< 0 && !can_nice(current
, nice
))
3701 retval
= security_task_setnice(current
, nice
);
3705 set_user_nice(current
, nice
);
3712 * task_prio - return the priority value of a given task.
3713 * @p: the task in question.
3715 * Return: The priority value as seen by users in /proc.
3716 * RT tasks are offset by -200. Normal tasks are centered
3717 * around 0, value goes from -16 to +15.
3719 int task_prio(const struct task_struct
*p
)
3721 return p
->prio
- MAX_RT_PRIO
;
3725 * idle_cpu - is a given cpu idle currently?
3726 * @cpu: the processor in question.
3728 * Return: 1 if the CPU is currently idle. 0 otherwise.
3730 int idle_cpu(int cpu
)
3732 struct rq
*rq
= cpu_rq(cpu
);
3734 if (rq
->curr
!= rq
->idle
)
3741 if (!llist_empty(&rq
->wake_list
))
3749 * idle_task - return the idle task for a given cpu.
3750 * @cpu: the processor in question.
3752 * Return: The idle task for the cpu @cpu.
3754 struct task_struct
*idle_task(int cpu
)
3756 return cpu_rq(cpu
)->idle
;
3760 * find_process_by_pid - find a process with a matching PID value.
3761 * @pid: the pid in question.
3763 * The task of @pid, if found. %NULL otherwise.
3765 static struct task_struct
*find_process_by_pid(pid_t pid
)
3767 return pid
? find_task_by_vpid(pid
) : current
;
3771 * This function initializes the sched_dl_entity of a newly becoming
3772 * SCHED_DEADLINE task.
3774 * Only the static values are considered here, the actual runtime and the
3775 * absolute deadline will be properly calculated when the task is enqueued
3776 * for the first time with its new policy.
3779 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3781 struct sched_dl_entity
*dl_se
= &p
->dl
;
3783 dl_se
->dl_runtime
= attr
->sched_runtime
;
3784 dl_se
->dl_deadline
= attr
->sched_deadline
;
3785 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3786 dl_se
->flags
= attr
->sched_flags
;
3787 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3790 * Changing the parameters of a task is 'tricky' and we're not doing
3791 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3793 * What we SHOULD do is delay the bandwidth release until the 0-lag
3794 * point. This would include retaining the task_struct until that time
3795 * and change dl_overflow() to not immediately decrement the current
3798 * Instead we retain the current runtime/deadline and let the new
3799 * parameters take effect after the current reservation period lapses.
3800 * This is safe (albeit pessimistic) because the 0-lag point is always
3801 * before the current scheduling deadline.
3803 * We can still have temporary overloads because we do not delay the
3804 * change in bandwidth until that time; so admission control is
3805 * not on the safe side. It does however guarantee tasks will never
3806 * consume more than promised.
3811 * sched_setparam() passes in -1 for its policy, to let the functions
3812 * it calls know not to change it.
3814 #define SETPARAM_POLICY -1
3816 static void __setscheduler_params(struct task_struct
*p
,
3817 const struct sched_attr
*attr
)
3819 int policy
= attr
->sched_policy
;
3821 if (policy
== SETPARAM_POLICY
)
3826 if (dl_policy(policy
))
3827 __setparam_dl(p
, attr
);
3828 else if (fair_policy(policy
))
3829 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3832 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3833 * !rt_policy. Always setting this ensures that things like
3834 * getparam()/getattr() don't report silly values for !rt tasks.
3836 p
->rt_priority
= attr
->sched_priority
;
3837 p
->normal_prio
= normal_prio(p
);
3841 /* Actually do priority change: must hold pi & rq lock. */
3842 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3843 const struct sched_attr
*attr
, bool keep_boost
)
3845 __setscheduler_params(p
, attr
);
3848 * Keep a potential priority boosting if called from
3849 * sched_setscheduler().
3852 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3854 p
->prio
= normal_prio(p
);
3856 if (dl_prio(p
->prio
))
3857 p
->sched_class
= &dl_sched_class
;
3858 else if (rt_prio(p
->prio
))
3859 p
->sched_class
= &rt_sched_class
;
3861 p
->sched_class
= &fair_sched_class
;
3865 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3867 struct sched_dl_entity
*dl_se
= &p
->dl
;
3869 attr
->sched_priority
= p
->rt_priority
;
3870 attr
->sched_runtime
= dl_se
->dl_runtime
;
3871 attr
->sched_deadline
= dl_se
->dl_deadline
;
3872 attr
->sched_period
= dl_se
->dl_period
;
3873 attr
->sched_flags
= dl_se
->flags
;
3877 * This function validates the new parameters of a -deadline task.
3878 * We ask for the deadline not being zero, and greater or equal
3879 * than the runtime, as well as the period of being zero or
3880 * greater than deadline. Furthermore, we have to be sure that
3881 * user parameters are above the internal resolution of 1us (we
3882 * check sched_runtime only since it is always the smaller one) and
3883 * below 2^63 ns (we have to check both sched_deadline and
3884 * sched_period, as the latter can be zero).
3887 __checkparam_dl(const struct sched_attr
*attr
)
3890 if (attr
->sched_deadline
== 0)
3894 * Since we truncate DL_SCALE bits, make sure we're at least
3897 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3901 * Since we use the MSB for wrap-around and sign issues, make
3902 * sure it's not set (mind that period can be equal to zero).
3904 if (attr
->sched_deadline
& (1ULL << 63) ||
3905 attr
->sched_period
& (1ULL << 63))
3908 /* runtime <= deadline <= period (if period != 0) */
3909 if ((attr
->sched_period
!= 0 &&
3910 attr
->sched_period
< attr
->sched_deadline
) ||
3911 attr
->sched_deadline
< attr
->sched_runtime
)
3918 * check the target process has a UID that matches the current process's
3920 static bool check_same_owner(struct task_struct
*p
)
3922 const struct cred
*cred
= current_cred(), *pcred
;
3926 pcred
= __task_cred(p
);
3927 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
3928 uid_eq(cred
->euid
, pcred
->uid
));
3933 static bool dl_param_changed(struct task_struct
*p
,
3934 const struct sched_attr
*attr
)
3936 struct sched_dl_entity
*dl_se
= &p
->dl
;
3938 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
3939 dl_se
->dl_deadline
!= attr
->sched_deadline
||
3940 dl_se
->dl_period
!= attr
->sched_period
||
3941 dl_se
->flags
!= attr
->sched_flags
)
3947 static int __sched_setscheduler(struct task_struct
*p
,
3948 const struct sched_attr
*attr
,
3951 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
3952 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
3953 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
3954 int new_effective_prio
, policy
= attr
->sched_policy
;
3955 unsigned long flags
;
3956 const struct sched_class
*prev_class
;
3959 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3961 /* may grab non-irq protected spin_locks */
3962 BUG_ON(in_interrupt());
3964 /* double check policy once rq lock held */
3966 reset_on_fork
= p
->sched_reset_on_fork
;
3967 policy
= oldpolicy
= p
->policy
;
3969 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
3971 if (!valid_policy(policy
))
3975 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
3979 * Valid priorities for SCHED_FIFO and SCHED_RR are
3980 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3981 * SCHED_BATCH and SCHED_IDLE is 0.
3983 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
3984 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
3986 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
3987 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
3991 * Allow unprivileged RT tasks to decrease priority:
3993 if (user
&& !capable(CAP_SYS_NICE
)) {
3994 if (fair_policy(policy
)) {
3995 if (attr
->sched_nice
< task_nice(p
) &&
3996 !can_nice(p
, attr
->sched_nice
))
4000 if (rt_policy(policy
)) {
4001 unsigned long rlim_rtprio
=
4002 task_rlimit(p
, RLIMIT_RTPRIO
);
4004 /* can't set/change the rt policy */
4005 if (policy
!= p
->policy
&& !rlim_rtprio
)
4008 /* can't increase priority */
4009 if (attr
->sched_priority
> p
->rt_priority
&&
4010 attr
->sched_priority
> rlim_rtprio
)
4015 * Can't set/change SCHED_DEADLINE policy at all for now
4016 * (safest behavior); in the future we would like to allow
4017 * unprivileged DL tasks to increase their relative deadline
4018 * or reduce their runtime (both ways reducing utilization)
4020 if (dl_policy(policy
))
4024 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4025 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4027 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4028 if (!can_nice(p
, task_nice(p
)))
4032 /* can't change other user's priorities */
4033 if (!check_same_owner(p
))
4036 /* Normal users shall not reset the sched_reset_on_fork flag */
4037 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4042 retval
= security_task_setscheduler(p
);
4048 * make sure no PI-waiters arrive (or leave) while we are
4049 * changing the priority of the task:
4051 * To be able to change p->policy safely, the appropriate
4052 * runqueue lock must be held.
4054 rq
= task_rq_lock(p
, &flags
);
4057 * Changing the policy of the stop threads its a very bad idea
4059 if (p
== rq
->stop
) {
4060 task_rq_unlock(rq
, p
, &flags
);
4065 * If not changing anything there's no need to proceed further,
4066 * but store a possible modification of reset_on_fork.
4068 if (unlikely(policy
== p
->policy
)) {
4069 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4071 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4073 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4076 p
->sched_reset_on_fork
= reset_on_fork
;
4077 task_rq_unlock(rq
, p
, &flags
);
4083 #ifdef CONFIG_RT_GROUP_SCHED
4085 * Do not allow realtime tasks into groups that have no runtime
4088 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4089 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4090 !task_group_is_autogroup(task_group(p
))) {
4091 task_rq_unlock(rq
, p
, &flags
);
4096 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4097 cpumask_t
*span
= rq
->rd
->span
;
4100 * Don't allow tasks with an affinity mask smaller than
4101 * the entire root_domain to become SCHED_DEADLINE. We
4102 * will also fail if there's no bandwidth available.
4104 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4105 rq
->rd
->dl_bw
.bw
== 0) {
4106 task_rq_unlock(rq
, p
, &flags
);
4113 /* recheck policy now with rq lock held */
4114 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4115 policy
= oldpolicy
= -1;
4116 task_rq_unlock(rq
, p
, &flags
);
4121 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4122 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4125 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4126 task_rq_unlock(rq
, p
, &flags
);
4130 p
->sched_reset_on_fork
= reset_on_fork
;
4135 * Take priority boosted tasks into account. If the new
4136 * effective priority is unchanged, we just store the new
4137 * normal parameters and do not touch the scheduler class and
4138 * the runqueue. This will be done when the task deboost
4141 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4142 if (new_effective_prio
== oldprio
)
4143 queue_flags
&= ~DEQUEUE_MOVE
;
4146 queued
= task_on_rq_queued(p
);
4147 running
= task_current(rq
, p
);
4149 dequeue_task(rq
, p
, queue_flags
);
4151 put_prev_task(rq
, p
);
4153 prev_class
= p
->sched_class
;
4154 __setscheduler(rq
, p
, attr
, pi
);
4157 p
->sched_class
->set_curr_task(rq
);
4160 * We enqueue to tail when the priority of a task is
4161 * increased (user space view).
4163 if (oldprio
< p
->prio
)
4164 queue_flags
|= ENQUEUE_HEAD
;
4166 enqueue_task(rq
, p
, queue_flags
);
4169 check_class_changed(rq
, p
, prev_class
, oldprio
);
4170 preempt_disable(); /* avoid rq from going away on us */
4171 task_rq_unlock(rq
, p
, &flags
);
4174 rt_mutex_adjust_pi(p
);
4177 * Run balance callbacks after we've adjusted the PI chain.
4179 balance_callback(rq
);
4185 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4186 const struct sched_param
*param
, bool check
)
4188 struct sched_attr attr
= {
4189 .sched_policy
= policy
,
4190 .sched_priority
= param
->sched_priority
,
4191 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4194 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4195 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4196 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4197 policy
&= ~SCHED_RESET_ON_FORK
;
4198 attr
.sched_policy
= policy
;
4201 return __sched_setscheduler(p
, &attr
, check
, true);
4204 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4205 * @p: the task in question.
4206 * @policy: new policy.
4207 * @param: structure containing the new RT priority.
4209 * Return: 0 on success. An error code otherwise.
4211 * NOTE that the task may be already dead.
4213 int sched_setscheduler(struct task_struct
*p
, int policy
,
4214 const struct sched_param
*param
)
4216 return _sched_setscheduler(p
, policy
, param
, true);
4218 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4220 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4222 return __sched_setscheduler(p
, attr
, true, true);
4224 EXPORT_SYMBOL_GPL(sched_setattr
);
4227 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4228 * @p: the task in question.
4229 * @policy: new policy.
4230 * @param: structure containing the new RT priority.
4232 * Just like sched_setscheduler, only don't bother checking if the
4233 * current context has permission. For example, this is needed in
4234 * stop_machine(): we create temporary high priority worker threads,
4235 * but our caller might not have that capability.
4237 * Return: 0 on success. An error code otherwise.
4239 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4240 const struct sched_param
*param
)
4242 return _sched_setscheduler(p
, policy
, param
, false);
4244 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4247 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4249 struct sched_param lparam
;
4250 struct task_struct
*p
;
4253 if (!param
|| pid
< 0)
4255 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4260 p
= find_process_by_pid(pid
);
4262 retval
= sched_setscheduler(p
, policy
, &lparam
);
4269 * Mimics kernel/events/core.c perf_copy_attr().
4271 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4272 struct sched_attr
*attr
)
4277 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4281 * zero the full structure, so that a short copy will be nice.
4283 memset(attr
, 0, sizeof(*attr
));
4285 ret
= get_user(size
, &uattr
->size
);
4289 if (size
> PAGE_SIZE
) /* silly large */
4292 if (!size
) /* abi compat */
4293 size
= SCHED_ATTR_SIZE_VER0
;
4295 if (size
< SCHED_ATTR_SIZE_VER0
)
4299 * If we're handed a bigger struct than we know of,
4300 * ensure all the unknown bits are 0 - i.e. new
4301 * user-space does not rely on any kernel feature
4302 * extensions we dont know about yet.
4304 if (size
> sizeof(*attr
)) {
4305 unsigned char __user
*addr
;
4306 unsigned char __user
*end
;
4309 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4310 end
= (void __user
*)uattr
+ size
;
4312 for (; addr
< end
; addr
++) {
4313 ret
= get_user(val
, addr
);
4319 size
= sizeof(*attr
);
4322 ret
= copy_from_user(attr
, uattr
, size
);
4327 * XXX: do we want to be lenient like existing syscalls; or do we want
4328 * to be strict and return an error on out-of-bounds values?
4330 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4335 put_user(sizeof(*attr
), &uattr
->size
);
4340 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4341 * @pid: the pid in question.
4342 * @policy: new policy.
4343 * @param: structure containing the new RT priority.
4345 * Return: 0 on success. An error code otherwise.
4347 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4348 struct sched_param __user
*, param
)
4350 /* negative values for policy are not valid */
4354 return do_sched_setscheduler(pid
, policy
, param
);
4358 * sys_sched_setparam - set/change the RT priority of a thread
4359 * @pid: the pid in question.
4360 * @param: structure containing the new RT priority.
4362 * Return: 0 on success. An error code otherwise.
4364 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4366 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4370 * sys_sched_setattr - same as above, but with extended sched_attr
4371 * @pid: the pid in question.
4372 * @uattr: structure containing the extended parameters.
4373 * @flags: for future extension.
4375 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4376 unsigned int, flags
)
4378 struct sched_attr attr
;
4379 struct task_struct
*p
;
4382 if (!uattr
|| pid
< 0 || flags
)
4385 retval
= sched_copy_attr(uattr
, &attr
);
4389 if ((int)attr
.sched_policy
< 0)
4394 p
= find_process_by_pid(pid
);
4396 retval
= sched_setattr(p
, &attr
);
4403 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4404 * @pid: the pid in question.
4406 * Return: On success, the policy of the thread. Otherwise, a negative error
4409 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4411 struct task_struct
*p
;
4419 p
= find_process_by_pid(pid
);
4421 retval
= security_task_getscheduler(p
);
4424 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4431 * sys_sched_getparam - get the RT priority of a thread
4432 * @pid: the pid in question.
4433 * @param: structure containing the RT priority.
4435 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4438 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4440 struct sched_param lp
= { .sched_priority
= 0 };
4441 struct task_struct
*p
;
4444 if (!param
|| pid
< 0)
4448 p
= find_process_by_pid(pid
);
4453 retval
= security_task_getscheduler(p
);
4457 if (task_has_rt_policy(p
))
4458 lp
.sched_priority
= p
->rt_priority
;
4462 * This one might sleep, we cannot do it with a spinlock held ...
4464 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4473 static int sched_read_attr(struct sched_attr __user
*uattr
,
4474 struct sched_attr
*attr
,
4479 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4483 * If we're handed a smaller struct than we know of,
4484 * ensure all the unknown bits are 0 - i.e. old
4485 * user-space does not get uncomplete information.
4487 if (usize
< sizeof(*attr
)) {
4488 unsigned char *addr
;
4491 addr
= (void *)attr
+ usize
;
4492 end
= (void *)attr
+ sizeof(*attr
);
4494 for (; addr
< end
; addr
++) {
4502 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4510 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4511 * @pid: the pid in question.
4512 * @uattr: structure containing the extended parameters.
4513 * @size: sizeof(attr) for fwd/bwd comp.
4514 * @flags: for future extension.
4516 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4517 unsigned int, size
, unsigned int, flags
)
4519 struct sched_attr attr
= {
4520 .size
= sizeof(struct sched_attr
),
4522 struct task_struct
*p
;
4525 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4526 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4530 p
= find_process_by_pid(pid
);
4535 retval
= security_task_getscheduler(p
);
4539 attr
.sched_policy
= p
->policy
;
4540 if (p
->sched_reset_on_fork
)
4541 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4542 if (task_has_dl_policy(p
))
4543 __getparam_dl(p
, &attr
);
4544 else if (task_has_rt_policy(p
))
4545 attr
.sched_priority
= p
->rt_priority
;
4547 attr
.sched_nice
= task_nice(p
);
4551 retval
= sched_read_attr(uattr
, &attr
, size
);
4559 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4561 cpumask_var_t cpus_allowed
, new_mask
;
4562 struct task_struct
*p
;
4567 p
= find_process_by_pid(pid
);
4573 /* Prevent p going away */
4577 if (p
->flags
& PF_NO_SETAFFINITY
) {
4581 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4585 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4587 goto out_free_cpus_allowed
;
4590 if (!check_same_owner(p
)) {
4592 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4594 goto out_free_new_mask
;
4599 retval
= security_task_setscheduler(p
);
4601 goto out_free_new_mask
;
4604 cpuset_cpus_allowed(p
, cpus_allowed
);
4605 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4608 * Since bandwidth control happens on root_domain basis,
4609 * if admission test is enabled, we only admit -deadline
4610 * tasks allowed to run on all the CPUs in the task's
4614 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4616 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4619 goto out_free_new_mask
;
4625 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4628 cpuset_cpus_allowed(p
, cpus_allowed
);
4629 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4631 * We must have raced with a concurrent cpuset
4632 * update. Just reset the cpus_allowed to the
4633 * cpuset's cpus_allowed
4635 cpumask_copy(new_mask
, cpus_allowed
);
4640 free_cpumask_var(new_mask
);
4641 out_free_cpus_allowed
:
4642 free_cpumask_var(cpus_allowed
);
4648 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4649 struct cpumask
*new_mask
)
4651 if (len
< cpumask_size())
4652 cpumask_clear(new_mask
);
4653 else if (len
> cpumask_size())
4654 len
= cpumask_size();
4656 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4660 * sys_sched_setaffinity - set the cpu affinity of a process
4661 * @pid: pid of the process
4662 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4663 * @user_mask_ptr: user-space pointer to the new cpu mask
4665 * Return: 0 on success. An error code otherwise.
4667 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4668 unsigned long __user
*, user_mask_ptr
)
4670 cpumask_var_t new_mask
;
4673 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4676 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4678 retval
= sched_setaffinity(pid
, new_mask
);
4679 free_cpumask_var(new_mask
);
4683 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4685 struct task_struct
*p
;
4686 unsigned long flags
;
4692 p
= find_process_by_pid(pid
);
4696 retval
= security_task_getscheduler(p
);
4700 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4701 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4702 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4711 * sys_sched_getaffinity - get the cpu affinity of a process
4712 * @pid: pid of the process
4713 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4714 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4716 * Return: 0 on success. An error code otherwise.
4718 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4719 unsigned long __user
*, user_mask_ptr
)
4724 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4726 if (len
& (sizeof(unsigned long)-1))
4729 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4732 ret
= sched_getaffinity(pid
, mask
);
4734 size_t retlen
= min_t(size_t, len
, cpumask_size());
4736 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4741 free_cpumask_var(mask
);
4747 * sys_sched_yield - yield the current processor to other threads.
4749 * This function yields the current CPU to other tasks. If there are no
4750 * other threads running on this CPU then this function will return.
4754 SYSCALL_DEFINE0(sched_yield
)
4756 struct rq
*rq
= this_rq_lock();
4758 schedstat_inc(rq
, yld_count
);
4759 current
->sched_class
->yield_task(rq
);
4762 * Since we are going to call schedule() anyway, there's
4763 * no need to preempt or enable interrupts:
4765 __release(rq
->lock
);
4766 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4767 do_raw_spin_unlock(&rq
->lock
);
4768 sched_preempt_enable_no_resched();
4775 int __sched
_cond_resched(void)
4777 if (should_resched(0)) {
4778 preempt_schedule_common();
4783 EXPORT_SYMBOL(_cond_resched
);
4786 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4787 * call schedule, and on return reacquire the lock.
4789 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4790 * operations here to prevent schedule() from being called twice (once via
4791 * spin_unlock(), once by hand).
4793 int __cond_resched_lock(spinlock_t
*lock
)
4795 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4798 lockdep_assert_held(lock
);
4800 if (spin_needbreak(lock
) || resched
) {
4803 preempt_schedule_common();
4811 EXPORT_SYMBOL(__cond_resched_lock
);
4813 int __sched
__cond_resched_softirq(void)
4815 BUG_ON(!in_softirq());
4817 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4819 preempt_schedule_common();
4825 EXPORT_SYMBOL(__cond_resched_softirq
);
4828 * yield - yield the current processor to other threads.
4830 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4832 * The scheduler is at all times free to pick the calling task as the most
4833 * eligible task to run, if removing the yield() call from your code breaks
4834 * it, its already broken.
4836 * Typical broken usage is:
4841 * where one assumes that yield() will let 'the other' process run that will
4842 * make event true. If the current task is a SCHED_FIFO task that will never
4843 * happen. Never use yield() as a progress guarantee!!
4845 * If you want to use yield() to wait for something, use wait_event().
4846 * If you want to use yield() to be 'nice' for others, use cond_resched().
4847 * If you still want to use yield(), do not!
4849 void __sched
yield(void)
4851 set_current_state(TASK_RUNNING
);
4854 EXPORT_SYMBOL(yield
);
4857 * yield_to - yield the current processor to another thread in
4858 * your thread group, or accelerate that thread toward the
4859 * processor it's on.
4861 * @preempt: whether task preemption is allowed or not
4863 * It's the caller's job to ensure that the target task struct
4864 * can't go away on us before we can do any checks.
4867 * true (>0) if we indeed boosted the target task.
4868 * false (0) if we failed to boost the target.
4869 * -ESRCH if there's no task to yield to.
4871 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4873 struct task_struct
*curr
= current
;
4874 struct rq
*rq
, *p_rq
;
4875 unsigned long flags
;
4878 local_irq_save(flags
);
4884 * If we're the only runnable task on the rq and target rq also
4885 * has only one task, there's absolutely no point in yielding.
4887 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4892 double_rq_lock(rq
, p_rq
);
4893 if (task_rq(p
) != p_rq
) {
4894 double_rq_unlock(rq
, p_rq
);
4898 if (!curr
->sched_class
->yield_to_task
)
4901 if (curr
->sched_class
!= p
->sched_class
)
4904 if (task_running(p_rq
, p
) || p
->state
)
4907 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
4909 schedstat_inc(rq
, yld_count
);
4911 * Make p's CPU reschedule; pick_next_entity takes care of
4914 if (preempt
&& rq
!= p_rq
)
4919 double_rq_unlock(rq
, p_rq
);
4921 local_irq_restore(flags
);
4928 EXPORT_SYMBOL_GPL(yield_to
);
4931 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4932 * that process accounting knows that this is a task in IO wait state.
4934 long __sched
io_schedule_timeout(long timeout
)
4936 int old_iowait
= current
->in_iowait
;
4940 current
->in_iowait
= 1;
4941 blk_schedule_flush_plug(current
);
4943 delayacct_blkio_start();
4945 atomic_inc(&rq
->nr_iowait
);
4946 ret
= schedule_timeout(timeout
);
4947 current
->in_iowait
= old_iowait
;
4948 atomic_dec(&rq
->nr_iowait
);
4949 delayacct_blkio_end();
4953 EXPORT_SYMBOL(io_schedule_timeout
);
4956 * sys_sched_get_priority_max - return maximum RT priority.
4957 * @policy: scheduling class.
4959 * Return: On success, this syscall returns the maximum
4960 * rt_priority that can be used by a given scheduling class.
4961 * On failure, a negative error code is returned.
4963 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4970 ret
= MAX_USER_RT_PRIO
-1;
4972 case SCHED_DEADLINE
:
4983 * sys_sched_get_priority_min - return minimum RT priority.
4984 * @policy: scheduling class.
4986 * Return: On success, this syscall returns the minimum
4987 * rt_priority that can be used by a given scheduling class.
4988 * On failure, a negative error code is returned.
4990 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
4999 case SCHED_DEADLINE
:
5009 * sys_sched_rr_get_interval - return the default timeslice of a process.
5010 * @pid: pid of the process.
5011 * @interval: userspace pointer to the timeslice value.
5013 * this syscall writes the default timeslice value of a given process
5014 * into the user-space timespec buffer. A value of '0' means infinity.
5016 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5019 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5020 struct timespec __user
*, interval
)
5022 struct task_struct
*p
;
5023 unsigned int time_slice
;
5024 unsigned long flags
;
5034 p
= find_process_by_pid(pid
);
5038 retval
= security_task_getscheduler(p
);
5042 rq
= task_rq_lock(p
, &flags
);
5044 if (p
->sched_class
->get_rr_interval
)
5045 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5046 task_rq_unlock(rq
, p
, &flags
);
5049 jiffies_to_timespec(time_slice
, &t
);
5050 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5058 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5060 void sched_show_task(struct task_struct
*p
)
5062 unsigned long free
= 0;
5064 unsigned long state
= p
->state
;
5067 state
= __ffs(state
) + 1;
5068 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5069 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5070 #if BITS_PER_LONG == 32
5071 if (state
== TASK_RUNNING
)
5072 printk(KERN_CONT
" running ");
5074 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5076 if (state
== TASK_RUNNING
)
5077 printk(KERN_CONT
" running task ");
5079 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5081 #ifdef CONFIG_DEBUG_STACK_USAGE
5082 free
= stack_not_used(p
);
5087 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5089 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5090 task_pid_nr(p
), ppid
,
5091 (unsigned long)task_thread_info(p
)->flags
);
5093 print_worker_info(KERN_INFO
, p
);
5094 show_stack(p
, NULL
);
5097 void show_state_filter(unsigned long state_filter
)
5099 struct task_struct
*g
, *p
;
5101 #if BITS_PER_LONG == 32
5103 " task PC stack pid father\n");
5106 " task PC stack pid father\n");
5109 for_each_process_thread(g
, p
) {
5111 * reset the NMI-timeout, listing all files on a slow
5112 * console might take a lot of time:
5114 touch_nmi_watchdog();
5115 if (!state_filter
|| (p
->state
& state_filter
))
5119 touch_all_softlockup_watchdogs();
5121 #ifdef CONFIG_SCHED_DEBUG
5123 sysrq_sched_debug_show();
5127 * Only show locks if all tasks are dumped:
5130 debug_show_all_locks();
5133 void init_idle_bootup_task(struct task_struct
*idle
)
5135 idle
->sched_class
= &idle_sched_class
;
5139 * init_idle - set up an idle thread for a given CPU
5140 * @idle: task in question
5141 * @cpu: cpu the idle task belongs to
5143 * NOTE: this function does not set the idle thread's NEED_RESCHED
5144 * flag, to make booting more robust.
5146 void init_idle(struct task_struct
*idle
, int cpu
)
5148 struct rq
*rq
= cpu_rq(cpu
);
5149 unsigned long flags
;
5151 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5152 raw_spin_lock(&rq
->lock
);
5154 __sched_fork(0, idle
);
5155 idle
->state
= TASK_RUNNING
;
5156 idle
->se
.exec_start
= sched_clock();
5158 kasan_unpoison_task_stack(idle
);
5162 * Its possible that init_idle() gets called multiple times on a task,
5163 * in that case do_set_cpus_allowed() will not do the right thing.
5165 * And since this is boot we can forgo the serialization.
5167 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5170 * We're having a chicken and egg problem, even though we are
5171 * holding rq->lock, the cpu isn't yet set to this cpu so the
5172 * lockdep check in task_group() will fail.
5174 * Similar case to sched_fork(). / Alternatively we could
5175 * use task_rq_lock() here and obtain the other rq->lock.
5180 __set_task_cpu(idle
, cpu
);
5183 rq
->curr
= rq
->idle
= idle
;
5184 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5188 raw_spin_unlock(&rq
->lock
);
5189 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5191 /* Set the preempt count _outside_ the spinlocks! */
5192 init_idle_preempt_count(idle
, cpu
);
5195 * The idle tasks have their own, simple scheduling class:
5197 idle
->sched_class
= &idle_sched_class
;
5198 ftrace_graph_init_idle_task(idle
, cpu
);
5199 vtime_init_idle(idle
, cpu
);
5201 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5205 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5206 const struct cpumask
*trial
)
5208 int ret
= 1, trial_cpus
;
5209 struct dl_bw
*cur_dl_b
;
5210 unsigned long flags
;
5212 if (!cpumask_weight(cur
))
5215 rcu_read_lock_sched();
5216 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5217 trial_cpus
= cpumask_weight(trial
);
5219 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5220 if (cur_dl_b
->bw
!= -1 &&
5221 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5223 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5224 rcu_read_unlock_sched();
5229 int task_can_attach(struct task_struct
*p
,
5230 const struct cpumask
*cs_cpus_allowed
)
5235 * Kthreads which disallow setaffinity shouldn't be moved
5236 * to a new cpuset; we don't want to change their cpu
5237 * affinity and isolating such threads by their set of
5238 * allowed nodes is unnecessary. Thus, cpusets are not
5239 * applicable for such threads. This prevents checking for
5240 * success of set_cpus_allowed_ptr() on all attached tasks
5241 * before cpus_allowed may be changed.
5243 if (p
->flags
& PF_NO_SETAFFINITY
) {
5249 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5251 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5256 unsigned long flags
;
5258 rcu_read_lock_sched();
5259 dl_b
= dl_bw_of(dest_cpu
);
5260 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5261 cpus
= dl_bw_cpus(dest_cpu
);
5262 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5267 * We reserve space for this task in the destination
5268 * root_domain, as we can't fail after this point.
5269 * We will free resources in the source root_domain
5270 * later on (see set_cpus_allowed_dl()).
5272 __dl_add(dl_b
, p
->dl
.dl_bw
);
5274 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5275 rcu_read_unlock_sched();
5285 #ifdef CONFIG_NUMA_BALANCING
5286 /* Migrate current task p to target_cpu */
5287 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5289 struct migration_arg arg
= { p
, target_cpu
};
5290 int curr_cpu
= task_cpu(p
);
5292 if (curr_cpu
== target_cpu
)
5295 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5298 /* TODO: This is not properly updating schedstats */
5300 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5301 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5305 * Requeue a task on a given node and accurately track the number of NUMA
5306 * tasks on the runqueues
5308 void sched_setnuma(struct task_struct
*p
, int nid
)
5311 unsigned long flags
;
5312 bool queued
, running
;
5314 rq
= task_rq_lock(p
, &flags
);
5315 queued
= task_on_rq_queued(p
);
5316 running
= task_current(rq
, p
);
5319 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5321 put_prev_task(rq
, p
);
5323 p
->numa_preferred_nid
= nid
;
5326 p
->sched_class
->set_curr_task(rq
);
5328 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5329 task_rq_unlock(rq
, p
, &flags
);
5331 #endif /* CONFIG_NUMA_BALANCING */
5333 #ifdef CONFIG_HOTPLUG_CPU
5335 * Ensures that the idle task is using init_mm right before its cpu goes
5338 void idle_task_exit(void)
5340 struct mm_struct
*mm
= current
->active_mm
;
5342 BUG_ON(cpu_online(smp_processor_id()));
5344 if (mm
!= &init_mm
) {
5345 switch_mm_irqs_off(mm
, &init_mm
, current
);
5346 finish_arch_post_lock_switch();
5352 * Since this CPU is going 'away' for a while, fold any nr_active delta
5353 * we might have. Assumes we're called after migrate_tasks() so that the
5354 * nr_active count is stable.
5356 * Also see the comment "Global load-average calculations".
5358 static void calc_load_migrate(struct rq
*rq
)
5360 long delta
= calc_load_fold_active(rq
);
5362 atomic_long_add(delta
, &calc_load_tasks
);
5365 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5369 static const struct sched_class fake_sched_class
= {
5370 .put_prev_task
= put_prev_task_fake
,
5373 static struct task_struct fake_task
= {
5375 * Avoid pull_{rt,dl}_task()
5377 .prio
= MAX_PRIO
+ 1,
5378 .sched_class
= &fake_sched_class
,
5382 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5383 * try_to_wake_up()->select_task_rq().
5385 * Called with rq->lock held even though we'er in stop_machine() and
5386 * there's no concurrency possible, we hold the required locks anyway
5387 * because of lock validation efforts.
5389 static void migrate_tasks(struct rq
*dead_rq
)
5391 struct rq
*rq
= dead_rq
;
5392 struct task_struct
*next
, *stop
= rq
->stop
;
5396 * Fudge the rq selection such that the below task selection loop
5397 * doesn't get stuck on the currently eligible stop task.
5399 * We're currently inside stop_machine() and the rq is either stuck
5400 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5401 * either way we should never end up calling schedule() until we're
5407 * put_prev_task() and pick_next_task() sched
5408 * class method both need to have an up-to-date
5409 * value of rq->clock[_task]
5411 update_rq_clock(rq
);
5415 * There's this thread running, bail when that's the only
5418 if (rq
->nr_running
== 1)
5422 * pick_next_task assumes pinned rq->lock.
5424 lockdep_pin_lock(&rq
->lock
);
5425 next
= pick_next_task(rq
, &fake_task
);
5427 next
->sched_class
->put_prev_task(rq
, next
);
5430 * Rules for changing task_struct::cpus_allowed are holding
5431 * both pi_lock and rq->lock, such that holding either
5432 * stabilizes the mask.
5434 * Drop rq->lock is not quite as disastrous as it usually is
5435 * because !cpu_active at this point, which means load-balance
5436 * will not interfere. Also, stop-machine.
5438 lockdep_unpin_lock(&rq
->lock
);
5439 raw_spin_unlock(&rq
->lock
);
5440 raw_spin_lock(&next
->pi_lock
);
5441 raw_spin_lock(&rq
->lock
);
5444 * Since we're inside stop-machine, _nothing_ should have
5445 * changed the task, WARN if weird stuff happened, because in
5446 * that case the above rq->lock drop is a fail too.
5448 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5449 raw_spin_unlock(&next
->pi_lock
);
5453 /* Find suitable destination for @next, with force if needed. */
5454 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5456 rq
= __migrate_task(rq
, next
, dest_cpu
);
5457 if (rq
!= dead_rq
) {
5458 raw_spin_unlock(&rq
->lock
);
5460 raw_spin_lock(&rq
->lock
);
5462 raw_spin_unlock(&next
->pi_lock
);
5467 #endif /* CONFIG_HOTPLUG_CPU */
5469 static void set_rq_online(struct rq
*rq
)
5472 const struct sched_class
*class;
5474 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5477 for_each_class(class) {
5478 if (class->rq_online
)
5479 class->rq_online(rq
);
5484 static void set_rq_offline(struct rq
*rq
)
5487 const struct sched_class
*class;
5489 for_each_class(class) {
5490 if (class->rq_offline
)
5491 class->rq_offline(rq
);
5494 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5500 * migration_call - callback that gets triggered when a CPU is added.
5501 * Here we can start up the necessary migration thread for the new CPU.
5504 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5506 int cpu
= (long)hcpu
;
5507 unsigned long flags
;
5508 struct rq
*rq
= cpu_rq(cpu
);
5510 switch (action
& ~CPU_TASKS_FROZEN
) {
5512 case CPU_UP_PREPARE
:
5513 rq
->calc_load_update
= calc_load_update
;
5514 account_reset_rq(rq
);
5518 /* Update our root-domain */
5519 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5521 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5525 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5528 #ifdef CONFIG_HOTPLUG_CPU
5530 sched_ttwu_pending();
5531 /* Update our root-domain */
5532 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5534 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5538 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
5539 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5543 calc_load_migrate(rq
);
5548 update_max_interval();
5554 * Register at high priority so that task migration (migrate_all_tasks)
5555 * happens before everything else. This has to be lower priority than
5556 * the notifier in the perf_event subsystem, though.
5558 static struct notifier_block migration_notifier
= {
5559 .notifier_call
= migration_call
,
5560 .priority
= CPU_PRI_MIGRATION
,
5563 static void set_cpu_rq_start_time(void)
5565 int cpu
= smp_processor_id();
5566 struct rq
*rq
= cpu_rq(cpu
);
5567 rq
->age_stamp
= sched_clock_cpu(cpu
);
5570 static int sched_cpu_active(struct notifier_block
*nfb
,
5571 unsigned long action
, void *hcpu
)
5573 int cpu
= (long)hcpu
;
5575 switch (action
& ~CPU_TASKS_FROZEN
) {
5577 set_cpu_rq_start_time();
5580 case CPU_DOWN_FAILED
:
5581 set_cpu_active(cpu
, true);
5589 static int sched_cpu_inactive(struct notifier_block
*nfb
,
5590 unsigned long action
, void *hcpu
)
5592 switch (action
& ~CPU_TASKS_FROZEN
) {
5593 case CPU_DOWN_PREPARE
:
5594 set_cpu_active((long)hcpu
, false);
5601 static int __init
migration_init(void)
5603 void *cpu
= (void *)(long)smp_processor_id();
5606 /* Initialize migration for the boot CPU */
5607 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5608 BUG_ON(err
== NOTIFY_BAD
);
5609 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5610 register_cpu_notifier(&migration_notifier
);
5612 /* Register cpu active notifiers */
5613 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
5614 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
5618 early_initcall(migration_init
);
5620 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5622 #ifdef CONFIG_SCHED_DEBUG
5624 static __read_mostly
int sched_debug_enabled
;
5626 static int __init
sched_debug_setup(char *str
)
5628 sched_debug_enabled
= 1;
5632 early_param("sched_debug", sched_debug_setup
);
5634 static inline bool sched_debug(void)
5636 return sched_debug_enabled
;
5639 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5640 struct cpumask
*groupmask
)
5642 struct sched_group
*group
= sd
->groups
;
5644 cpumask_clear(groupmask
);
5646 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5648 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5649 printk("does not load-balance\n");
5651 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5656 printk(KERN_CONT
"span %*pbl level %s\n",
5657 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5659 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5660 printk(KERN_ERR
"ERROR: domain->span does not contain "
5663 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5664 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5668 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5672 printk(KERN_ERR
"ERROR: group is NULL\n");
5676 if (!cpumask_weight(sched_group_cpus(group
))) {
5677 printk(KERN_CONT
"\n");
5678 printk(KERN_ERR
"ERROR: empty group\n");
5682 if (!(sd
->flags
& SD_OVERLAP
) &&
5683 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5684 printk(KERN_CONT
"\n");
5685 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5689 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5691 printk(KERN_CONT
" %*pbl",
5692 cpumask_pr_args(sched_group_cpus(group
)));
5693 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5694 printk(KERN_CONT
" (cpu_capacity = %d)",
5695 group
->sgc
->capacity
);
5698 group
= group
->next
;
5699 } while (group
!= sd
->groups
);
5700 printk(KERN_CONT
"\n");
5702 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5703 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5706 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5707 printk(KERN_ERR
"ERROR: parent span is not a superset "
5708 "of domain->span\n");
5712 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5716 if (!sched_debug_enabled
)
5720 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5724 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5727 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5735 #else /* !CONFIG_SCHED_DEBUG */
5736 # define sched_domain_debug(sd, cpu) do { } while (0)
5737 static inline bool sched_debug(void)
5741 #endif /* CONFIG_SCHED_DEBUG */
5743 static int sd_degenerate(struct sched_domain
*sd
)
5745 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5748 /* Following flags need at least 2 groups */
5749 if (sd
->flags
& (SD_LOAD_BALANCE
|
5750 SD_BALANCE_NEWIDLE
|
5753 SD_SHARE_CPUCAPACITY
|
5754 SD_SHARE_PKG_RESOURCES
|
5755 SD_SHARE_POWERDOMAIN
)) {
5756 if (sd
->groups
!= sd
->groups
->next
)
5760 /* Following flags don't use groups */
5761 if (sd
->flags
& (SD_WAKE_AFFINE
))
5768 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5770 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5772 if (sd_degenerate(parent
))
5775 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5778 /* Flags needing groups don't count if only 1 group in parent */
5779 if (parent
->groups
== parent
->groups
->next
) {
5780 pflags
&= ~(SD_LOAD_BALANCE
|
5781 SD_BALANCE_NEWIDLE
|
5784 SD_SHARE_CPUCAPACITY
|
5785 SD_SHARE_PKG_RESOURCES
|
5787 SD_SHARE_POWERDOMAIN
);
5788 if (nr_node_ids
== 1)
5789 pflags
&= ~SD_SERIALIZE
;
5791 if (~cflags
& pflags
)
5797 static void free_rootdomain(struct rcu_head
*rcu
)
5799 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5801 cpupri_cleanup(&rd
->cpupri
);
5802 cpudl_cleanup(&rd
->cpudl
);
5803 free_cpumask_var(rd
->dlo_mask
);
5804 free_cpumask_var(rd
->rto_mask
);
5805 free_cpumask_var(rd
->online
);
5806 free_cpumask_var(rd
->span
);
5810 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5812 struct root_domain
*old_rd
= NULL
;
5813 unsigned long flags
;
5815 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5820 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5823 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5826 * If we dont want to free the old_rd yet then
5827 * set old_rd to NULL to skip the freeing later
5830 if (!atomic_dec_and_test(&old_rd
->refcount
))
5834 atomic_inc(&rd
->refcount
);
5837 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5838 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5841 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5844 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5847 static int init_rootdomain(struct root_domain
*rd
)
5849 memset(rd
, 0, sizeof(*rd
));
5851 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5853 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5855 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5857 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5860 init_dl_bw(&rd
->dl_bw
);
5861 if (cpudl_init(&rd
->cpudl
) != 0)
5864 if (cpupri_init(&rd
->cpupri
) != 0)
5869 free_cpumask_var(rd
->rto_mask
);
5871 free_cpumask_var(rd
->dlo_mask
);
5873 free_cpumask_var(rd
->online
);
5875 free_cpumask_var(rd
->span
);
5881 * By default the system creates a single root-domain with all cpus as
5882 * members (mimicking the global state we have today).
5884 struct root_domain def_root_domain
;
5886 static void init_defrootdomain(void)
5888 init_rootdomain(&def_root_domain
);
5890 atomic_set(&def_root_domain
.refcount
, 1);
5893 static struct root_domain
*alloc_rootdomain(void)
5895 struct root_domain
*rd
;
5897 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5901 if (init_rootdomain(rd
) != 0) {
5909 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5911 struct sched_group
*tmp
, *first
;
5920 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5925 } while (sg
!= first
);
5928 static void free_sched_domain(struct rcu_head
*rcu
)
5930 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5933 * If its an overlapping domain it has private groups, iterate and
5936 if (sd
->flags
& SD_OVERLAP
) {
5937 free_sched_groups(sd
->groups
, 1);
5938 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5939 kfree(sd
->groups
->sgc
);
5945 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5947 call_rcu(&sd
->rcu
, free_sched_domain
);
5950 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5952 for (; sd
; sd
= sd
->parent
)
5953 destroy_sched_domain(sd
, cpu
);
5957 * Keep a special pointer to the highest sched_domain that has
5958 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5959 * allows us to avoid some pointer chasing select_idle_sibling().
5961 * Also keep a unique ID per domain (we use the first cpu number in
5962 * the cpumask of the domain), this allows us to quickly tell if
5963 * two cpus are in the same cache domain, see cpus_share_cache().
5965 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5966 DEFINE_PER_CPU(int, sd_llc_size
);
5967 DEFINE_PER_CPU(int, sd_llc_id
);
5968 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5969 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5970 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5972 static void update_top_cache_domain(int cpu
)
5974 struct sched_domain
*sd
;
5975 struct sched_domain
*busy_sd
= NULL
;
5979 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5981 id
= cpumask_first(sched_domain_span(sd
));
5982 size
= cpumask_weight(sched_domain_span(sd
));
5983 busy_sd
= sd
->parent
; /* sd_busy */
5985 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5987 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5988 per_cpu(sd_llc_size
, cpu
) = size
;
5989 per_cpu(sd_llc_id
, cpu
) = id
;
5991 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5992 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5994 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5995 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5999 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6000 * hold the hotplug lock.
6003 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6005 struct rq
*rq
= cpu_rq(cpu
);
6006 struct sched_domain
*tmp
;
6008 /* Remove the sched domains which do not contribute to scheduling. */
6009 for (tmp
= sd
; tmp
; ) {
6010 struct sched_domain
*parent
= tmp
->parent
;
6014 if (sd_parent_degenerate(tmp
, parent
)) {
6015 tmp
->parent
= parent
->parent
;
6017 parent
->parent
->child
= tmp
;
6019 * Transfer SD_PREFER_SIBLING down in case of a
6020 * degenerate parent; the spans match for this
6021 * so the property transfers.
6023 if (parent
->flags
& SD_PREFER_SIBLING
)
6024 tmp
->flags
|= SD_PREFER_SIBLING
;
6025 destroy_sched_domain(parent
, cpu
);
6030 if (sd
&& sd_degenerate(sd
)) {
6033 destroy_sched_domain(tmp
, cpu
);
6038 sched_domain_debug(sd
, cpu
);
6040 rq_attach_root(rq
, rd
);
6042 rcu_assign_pointer(rq
->sd
, sd
);
6043 destroy_sched_domains(tmp
, cpu
);
6045 update_top_cache_domain(cpu
);
6048 /* Setup the mask of cpus configured for isolated domains */
6049 static int __init
isolated_cpu_setup(char *str
)
6053 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6054 ret
= cpulist_parse(str
, cpu_isolated_map
);
6056 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
6061 __setup("isolcpus=", isolated_cpu_setup
);
6064 struct sched_domain
** __percpu sd
;
6065 struct root_domain
*rd
;
6076 * Build an iteration mask that can exclude certain CPUs from the upwards
6079 * Asymmetric node setups can result in situations where the domain tree is of
6080 * unequal depth, make sure to skip domains that already cover the entire
6083 * In that case build_sched_domains() will have terminated the iteration early
6084 * and our sibling sd spans will be empty. Domains should always include the
6085 * cpu they're built on, so check that.
6088 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6090 const struct cpumask
*span
= sched_domain_span(sd
);
6091 struct sd_data
*sdd
= sd
->private;
6092 struct sched_domain
*sibling
;
6095 for_each_cpu(i
, span
) {
6096 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6097 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6100 cpumask_set_cpu(i
, sched_group_mask(sg
));
6105 * Return the canonical balance cpu for this group, this is the first cpu
6106 * of this group that's also in the iteration mask.
6108 int group_balance_cpu(struct sched_group
*sg
)
6110 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6114 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6116 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6117 const struct cpumask
*span
= sched_domain_span(sd
);
6118 struct cpumask
*covered
= sched_domains_tmpmask
;
6119 struct sd_data
*sdd
= sd
->private;
6120 struct sched_domain
*sibling
;
6123 cpumask_clear(covered
);
6125 for_each_cpu(i
, span
) {
6126 struct cpumask
*sg_span
;
6128 if (cpumask_test_cpu(i
, covered
))
6131 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6133 /* See the comment near build_group_mask(). */
6134 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6137 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6138 GFP_KERNEL
, cpu_to_node(cpu
));
6143 sg_span
= sched_group_cpus(sg
);
6145 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6147 cpumask_set_cpu(i
, sg_span
);
6149 cpumask_or(covered
, covered
, sg_span
);
6151 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6152 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6153 build_group_mask(sd
, sg
);
6156 * Initialize sgc->capacity such that even if we mess up the
6157 * domains and no possible iteration will get us here, we won't
6160 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6163 * Make sure the first group of this domain contains the
6164 * canonical balance cpu. Otherwise the sched_domain iteration
6165 * breaks. See update_sg_lb_stats().
6167 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6168 group_balance_cpu(sg
) == cpu
)
6178 sd
->groups
= groups
;
6183 free_sched_groups(first
, 0);
6188 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6190 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6191 struct sched_domain
*child
= sd
->child
;
6194 cpu
= cpumask_first(sched_domain_span(child
));
6197 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6198 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6199 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6206 * build_sched_groups will build a circular linked list of the groups
6207 * covered by the given span, and will set each group's ->cpumask correctly,
6208 * and ->cpu_capacity to 0.
6210 * Assumes the sched_domain tree is fully constructed
6213 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6215 struct sched_group
*first
= NULL
, *last
= NULL
;
6216 struct sd_data
*sdd
= sd
->private;
6217 const struct cpumask
*span
= sched_domain_span(sd
);
6218 struct cpumask
*covered
;
6221 get_group(cpu
, sdd
, &sd
->groups
);
6222 atomic_inc(&sd
->groups
->ref
);
6224 if (cpu
!= cpumask_first(span
))
6227 lockdep_assert_held(&sched_domains_mutex
);
6228 covered
= sched_domains_tmpmask
;
6230 cpumask_clear(covered
);
6232 for_each_cpu(i
, span
) {
6233 struct sched_group
*sg
;
6236 if (cpumask_test_cpu(i
, covered
))
6239 group
= get_group(i
, sdd
, &sg
);
6240 cpumask_setall(sched_group_mask(sg
));
6242 for_each_cpu(j
, span
) {
6243 if (get_group(j
, sdd
, NULL
) != group
)
6246 cpumask_set_cpu(j
, covered
);
6247 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6262 * Initialize sched groups cpu_capacity.
6264 * cpu_capacity indicates the capacity of sched group, which is used while
6265 * distributing the load between different sched groups in a sched domain.
6266 * Typically cpu_capacity for all the groups in a sched domain will be same
6267 * unless there are asymmetries in the topology. If there are asymmetries,
6268 * group having more cpu_capacity will pickup more load compared to the
6269 * group having less cpu_capacity.
6271 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6273 struct sched_group
*sg
= sd
->groups
;
6278 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6280 } while (sg
!= sd
->groups
);
6282 if (cpu
!= group_balance_cpu(sg
))
6285 update_group_capacity(sd
, cpu
);
6286 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6290 * Initializers for schedule domains
6291 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6294 static int default_relax_domain_level
= -1;
6295 int sched_domain_level_max
;
6297 static int __init
setup_relax_domain_level(char *str
)
6299 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6300 pr_warn("Unable to set relax_domain_level\n");
6304 __setup("relax_domain_level=", setup_relax_domain_level
);
6306 static void set_domain_attribute(struct sched_domain
*sd
,
6307 struct sched_domain_attr
*attr
)
6311 if (!attr
|| attr
->relax_domain_level
< 0) {
6312 if (default_relax_domain_level
< 0)
6315 request
= default_relax_domain_level
;
6317 request
= attr
->relax_domain_level
;
6318 if (request
< sd
->level
) {
6319 /* turn off idle balance on this domain */
6320 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6322 /* turn on idle balance on this domain */
6323 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6327 static void __sdt_free(const struct cpumask
*cpu_map
);
6328 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6330 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6331 const struct cpumask
*cpu_map
)
6335 if (!atomic_read(&d
->rd
->refcount
))
6336 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6338 free_percpu(d
->sd
); /* fall through */
6340 __sdt_free(cpu_map
); /* fall through */
6346 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6347 const struct cpumask
*cpu_map
)
6349 memset(d
, 0, sizeof(*d
));
6351 if (__sdt_alloc(cpu_map
))
6352 return sa_sd_storage
;
6353 d
->sd
= alloc_percpu(struct sched_domain
*);
6355 return sa_sd_storage
;
6356 d
->rd
= alloc_rootdomain();
6359 return sa_rootdomain
;
6363 * NULL the sd_data elements we've used to build the sched_domain and
6364 * sched_group structure so that the subsequent __free_domain_allocs()
6365 * will not free the data we're using.
6367 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6369 struct sd_data
*sdd
= sd
->private;
6371 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6372 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6374 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6375 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6377 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6378 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6382 static int sched_domains_numa_levels
;
6383 enum numa_topology_type sched_numa_topology_type
;
6384 static int *sched_domains_numa_distance
;
6385 int sched_max_numa_distance
;
6386 static struct cpumask
***sched_domains_numa_masks
;
6387 static int sched_domains_curr_level
;
6391 * SD_flags allowed in topology descriptions.
6393 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6394 * SD_SHARE_PKG_RESOURCES - describes shared caches
6395 * SD_NUMA - describes NUMA topologies
6396 * SD_SHARE_POWERDOMAIN - describes shared power domain
6399 * SD_ASYM_PACKING - describes SMT quirks
6401 #define TOPOLOGY_SD_FLAGS \
6402 (SD_SHARE_CPUCAPACITY | \
6403 SD_SHARE_PKG_RESOURCES | \
6406 SD_SHARE_POWERDOMAIN)
6408 static struct sched_domain
*
6409 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6411 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6412 int sd_weight
, sd_flags
= 0;
6416 * Ugly hack to pass state to sd_numa_mask()...
6418 sched_domains_curr_level
= tl
->numa_level
;
6421 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6424 sd_flags
= (*tl
->sd_flags
)();
6425 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6426 "wrong sd_flags in topology description\n"))
6427 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6429 *sd
= (struct sched_domain
){
6430 .min_interval
= sd_weight
,
6431 .max_interval
= 2*sd_weight
,
6433 .imbalance_pct
= 125,
6435 .cache_nice_tries
= 0,
6442 .flags
= 1*SD_LOAD_BALANCE
6443 | 1*SD_BALANCE_NEWIDLE
6448 | 0*SD_SHARE_CPUCAPACITY
6449 | 0*SD_SHARE_PKG_RESOURCES
6451 | 0*SD_PREFER_SIBLING
6456 .last_balance
= jiffies
,
6457 .balance_interval
= sd_weight
,
6459 .max_newidle_lb_cost
= 0,
6460 .next_decay_max_lb_cost
= jiffies
,
6461 #ifdef CONFIG_SCHED_DEBUG
6467 * Convert topological properties into behaviour.
6470 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6471 sd
->flags
|= SD_PREFER_SIBLING
;
6472 sd
->imbalance_pct
= 110;
6473 sd
->smt_gain
= 1178; /* ~15% */
6475 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6476 sd
->imbalance_pct
= 117;
6477 sd
->cache_nice_tries
= 1;
6481 } else if (sd
->flags
& SD_NUMA
) {
6482 sd
->cache_nice_tries
= 2;
6486 sd
->flags
|= SD_SERIALIZE
;
6487 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6488 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6495 sd
->flags
|= SD_PREFER_SIBLING
;
6496 sd
->cache_nice_tries
= 1;
6501 sd
->private = &tl
->data
;
6507 * Topology list, bottom-up.
6509 static struct sched_domain_topology_level default_topology
[] = {
6510 #ifdef CONFIG_SCHED_SMT
6511 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6513 #ifdef CONFIG_SCHED_MC
6514 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6516 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6520 static struct sched_domain_topology_level
*sched_domain_topology
=
6523 #define for_each_sd_topology(tl) \
6524 for (tl = sched_domain_topology; tl->mask; tl++)
6526 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6528 sched_domain_topology
= tl
;
6533 static const struct cpumask
*sd_numa_mask(int cpu
)
6535 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6538 static void sched_numa_warn(const char *str
)
6540 static int done
= false;
6548 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6550 for (i
= 0; i
< nr_node_ids
; i
++) {
6551 printk(KERN_WARNING
" ");
6552 for (j
= 0; j
< nr_node_ids
; j
++)
6553 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6554 printk(KERN_CONT
"\n");
6556 printk(KERN_WARNING
"\n");
6559 bool find_numa_distance(int distance
)
6563 if (distance
== node_distance(0, 0))
6566 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6567 if (sched_domains_numa_distance
[i
] == distance
)
6575 * A system can have three types of NUMA topology:
6576 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6577 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6578 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6580 * The difference between a glueless mesh topology and a backplane
6581 * topology lies in whether communication between not directly
6582 * connected nodes goes through intermediary nodes (where programs
6583 * could run), or through backplane controllers. This affects
6584 * placement of programs.
6586 * The type of topology can be discerned with the following tests:
6587 * - If the maximum distance between any nodes is 1 hop, the system
6588 * is directly connected.
6589 * - If for two nodes A and B, located N > 1 hops away from each other,
6590 * there is an intermediary node C, which is < N hops away from both
6591 * nodes A and B, the system is a glueless mesh.
6593 static void init_numa_topology_type(void)
6597 n
= sched_max_numa_distance
;
6599 if (sched_domains_numa_levels
<= 1) {
6600 sched_numa_topology_type
= NUMA_DIRECT
;
6604 for_each_online_node(a
) {
6605 for_each_online_node(b
) {
6606 /* Find two nodes furthest removed from each other. */
6607 if (node_distance(a
, b
) < n
)
6610 /* Is there an intermediary node between a and b? */
6611 for_each_online_node(c
) {
6612 if (node_distance(a
, c
) < n
&&
6613 node_distance(b
, c
) < n
) {
6614 sched_numa_topology_type
=
6620 sched_numa_topology_type
= NUMA_BACKPLANE
;
6626 static void sched_init_numa(void)
6628 int next_distance
, curr_distance
= node_distance(0, 0);
6629 struct sched_domain_topology_level
*tl
;
6633 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6634 if (!sched_domains_numa_distance
)
6638 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6639 * unique distances in the node_distance() table.
6641 * Assumes node_distance(0,j) includes all distances in
6642 * node_distance(i,j) in order to avoid cubic time.
6644 next_distance
= curr_distance
;
6645 for (i
= 0; i
< nr_node_ids
; i
++) {
6646 for (j
= 0; j
< nr_node_ids
; j
++) {
6647 for (k
= 0; k
< nr_node_ids
; k
++) {
6648 int distance
= node_distance(i
, k
);
6650 if (distance
> curr_distance
&&
6651 (distance
< next_distance
||
6652 next_distance
== curr_distance
))
6653 next_distance
= distance
;
6656 * While not a strong assumption it would be nice to know
6657 * about cases where if node A is connected to B, B is not
6658 * equally connected to A.
6660 if (sched_debug() && node_distance(k
, i
) != distance
)
6661 sched_numa_warn("Node-distance not symmetric");
6663 if (sched_debug() && i
&& !find_numa_distance(distance
))
6664 sched_numa_warn("Node-0 not representative");
6666 if (next_distance
!= curr_distance
) {
6667 sched_domains_numa_distance
[level
++] = next_distance
;
6668 sched_domains_numa_levels
= level
;
6669 curr_distance
= next_distance
;
6674 * In case of sched_debug() we verify the above assumption.
6684 * 'level' contains the number of unique distances, excluding the
6685 * identity distance node_distance(i,i).
6687 * The sched_domains_numa_distance[] array includes the actual distance
6692 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6693 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6694 * the array will contain less then 'level' members. This could be
6695 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6696 * in other functions.
6698 * We reset it to 'level' at the end of this function.
6700 sched_domains_numa_levels
= 0;
6702 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6703 if (!sched_domains_numa_masks
)
6707 * Now for each level, construct a mask per node which contains all
6708 * cpus of nodes that are that many hops away from us.
6710 for (i
= 0; i
< level
; i
++) {
6711 sched_domains_numa_masks
[i
] =
6712 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6713 if (!sched_domains_numa_masks
[i
])
6716 for (j
= 0; j
< nr_node_ids
; j
++) {
6717 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6721 sched_domains_numa_masks
[i
][j
] = mask
;
6724 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6727 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6732 /* Compute default topology size */
6733 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6735 tl
= kzalloc((i
+ level
+ 1) *
6736 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6741 * Copy the default topology bits..
6743 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6744 tl
[i
] = sched_domain_topology
[i
];
6747 * .. and append 'j' levels of NUMA goodness.
6749 for (j
= 0; j
< level
; i
++, j
++) {
6750 tl
[i
] = (struct sched_domain_topology_level
){
6751 .mask
= sd_numa_mask
,
6752 .sd_flags
= cpu_numa_flags
,
6753 .flags
= SDTL_OVERLAP
,
6759 sched_domain_topology
= tl
;
6761 sched_domains_numa_levels
= level
;
6762 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6764 init_numa_topology_type();
6767 static void sched_domains_numa_masks_set(int cpu
)
6770 int node
= cpu_to_node(cpu
);
6772 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6773 for (j
= 0; j
< nr_node_ids
; j
++) {
6774 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6775 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6780 static void sched_domains_numa_masks_clear(int cpu
)
6783 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6784 for (j
= 0; j
< nr_node_ids
; j
++)
6785 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6790 * Update sched_domains_numa_masks[level][node] array when new cpus
6793 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6794 unsigned long action
,
6797 int cpu
= (long)hcpu
;
6799 switch (action
& ~CPU_TASKS_FROZEN
) {
6801 sched_domains_numa_masks_set(cpu
);
6805 sched_domains_numa_masks_clear(cpu
);
6815 static inline void sched_init_numa(void)
6819 static int sched_domains_numa_masks_update(struct notifier_block
*nfb
,
6820 unsigned long action
,
6825 #endif /* CONFIG_NUMA */
6827 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6829 struct sched_domain_topology_level
*tl
;
6832 for_each_sd_topology(tl
) {
6833 struct sd_data
*sdd
= &tl
->data
;
6835 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6839 sdd
->sg
= alloc_percpu(struct sched_group
*);
6843 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6847 for_each_cpu(j
, cpu_map
) {
6848 struct sched_domain
*sd
;
6849 struct sched_group
*sg
;
6850 struct sched_group_capacity
*sgc
;
6852 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6853 GFP_KERNEL
, cpu_to_node(j
));
6857 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6859 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6860 GFP_KERNEL
, cpu_to_node(j
));
6866 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6868 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6869 GFP_KERNEL
, cpu_to_node(j
));
6873 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6880 static void __sdt_free(const struct cpumask
*cpu_map
)
6882 struct sched_domain_topology_level
*tl
;
6885 for_each_sd_topology(tl
) {
6886 struct sd_data
*sdd
= &tl
->data
;
6888 for_each_cpu(j
, cpu_map
) {
6889 struct sched_domain
*sd
;
6892 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6893 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6894 free_sched_groups(sd
->groups
, 0);
6895 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6899 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6901 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6903 free_percpu(sdd
->sd
);
6905 free_percpu(sdd
->sg
);
6907 free_percpu(sdd
->sgc
);
6912 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6913 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6914 struct sched_domain
*child
, int cpu
)
6916 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6920 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6922 sd
->level
= child
->level
+ 1;
6923 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6927 if (!cpumask_subset(sched_domain_span(child
),
6928 sched_domain_span(sd
))) {
6929 pr_err("BUG: arch topology borken\n");
6930 #ifdef CONFIG_SCHED_DEBUG
6931 pr_err(" the %s domain not a subset of the %s domain\n",
6932 child
->name
, sd
->name
);
6934 /* Fixup, ensure @sd has at least @child cpus. */
6935 cpumask_or(sched_domain_span(sd
),
6936 sched_domain_span(sd
),
6937 sched_domain_span(child
));
6941 set_domain_attribute(sd
, attr
);
6947 * Build sched domains for a given set of cpus and attach the sched domains
6948 * to the individual cpus
6950 static int build_sched_domains(const struct cpumask
*cpu_map
,
6951 struct sched_domain_attr
*attr
)
6953 enum s_alloc alloc_state
;
6954 struct sched_domain
*sd
;
6956 int i
, ret
= -ENOMEM
;
6958 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6959 if (alloc_state
!= sa_rootdomain
)
6962 /* Set up domains for cpus specified by the cpu_map. */
6963 for_each_cpu(i
, cpu_map
) {
6964 struct sched_domain_topology_level
*tl
;
6967 for_each_sd_topology(tl
) {
6968 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6969 if (tl
== sched_domain_topology
)
6970 *per_cpu_ptr(d
.sd
, i
) = sd
;
6971 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6972 sd
->flags
|= SD_OVERLAP
;
6973 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6978 /* Build the groups for the domains */
6979 for_each_cpu(i
, cpu_map
) {
6980 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6981 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6982 if (sd
->flags
& SD_OVERLAP
) {
6983 if (build_overlap_sched_groups(sd
, i
))
6986 if (build_sched_groups(sd
, i
))
6992 /* Calculate CPU capacity for physical packages and nodes */
6993 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6994 if (!cpumask_test_cpu(i
, cpu_map
))
6997 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6998 claim_allocations(i
, sd
);
6999 init_sched_groups_capacity(i
, sd
);
7003 /* Attach the domains */
7005 for_each_cpu(i
, cpu_map
) {
7006 sd
= *per_cpu_ptr(d
.sd
, i
);
7007 cpu_attach_domain(sd
, d
.rd
, i
);
7013 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7017 static cpumask_var_t
*doms_cur
; /* current sched domains */
7018 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7019 static struct sched_domain_attr
*dattr_cur
;
7020 /* attribues of custom domains in 'doms_cur' */
7023 * Special case: If a kmalloc of a doms_cur partition (array of
7024 * cpumask) fails, then fallback to a single sched domain,
7025 * as determined by the single cpumask fallback_doms.
7027 static cpumask_var_t fallback_doms
;
7030 * arch_update_cpu_topology lets virtualized architectures update the
7031 * cpu core maps. It is supposed to return 1 if the topology changed
7032 * or 0 if it stayed the same.
7034 int __weak
arch_update_cpu_topology(void)
7039 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7042 cpumask_var_t
*doms
;
7044 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7047 for (i
= 0; i
< ndoms
; i
++) {
7048 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7049 free_sched_domains(doms
, i
);
7056 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7059 for (i
= 0; i
< ndoms
; i
++)
7060 free_cpumask_var(doms
[i
]);
7065 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7066 * For now this just excludes isolated cpus, but could be used to
7067 * exclude other special cases in the future.
7069 static int init_sched_domains(const struct cpumask
*cpu_map
)
7073 arch_update_cpu_topology();
7075 doms_cur
= alloc_sched_domains(ndoms_cur
);
7077 doms_cur
= &fallback_doms
;
7078 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7079 err
= build_sched_domains(doms_cur
[0], NULL
);
7080 register_sched_domain_sysctl();
7086 * Detach sched domains from a group of cpus specified in cpu_map
7087 * These cpus will now be attached to the NULL domain
7089 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7094 for_each_cpu(i
, cpu_map
)
7095 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7099 /* handle null as "default" */
7100 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7101 struct sched_domain_attr
*new, int idx_new
)
7103 struct sched_domain_attr tmp
;
7110 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7111 new ? (new + idx_new
) : &tmp
,
7112 sizeof(struct sched_domain_attr
));
7116 * Partition sched domains as specified by the 'ndoms_new'
7117 * cpumasks in the array doms_new[] of cpumasks. This compares
7118 * doms_new[] to the current sched domain partitioning, doms_cur[].
7119 * It destroys each deleted domain and builds each new domain.
7121 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7122 * The masks don't intersect (don't overlap.) We should setup one
7123 * sched domain for each mask. CPUs not in any of the cpumasks will
7124 * not be load balanced. If the same cpumask appears both in the
7125 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7128 * The passed in 'doms_new' should be allocated using
7129 * alloc_sched_domains. This routine takes ownership of it and will
7130 * free_sched_domains it when done with it. If the caller failed the
7131 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7132 * and partition_sched_domains() will fallback to the single partition
7133 * 'fallback_doms', it also forces the domains to be rebuilt.
7135 * If doms_new == NULL it will be replaced with cpu_online_mask.
7136 * ndoms_new == 0 is a special case for destroying existing domains,
7137 * and it will not create the default domain.
7139 * Call with hotplug lock held
7141 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7142 struct sched_domain_attr
*dattr_new
)
7147 mutex_lock(&sched_domains_mutex
);
7149 /* always unregister in case we don't destroy any domains */
7150 unregister_sched_domain_sysctl();
7152 /* Let architecture update cpu core mappings. */
7153 new_topology
= arch_update_cpu_topology();
7155 n
= doms_new
? ndoms_new
: 0;
7157 /* Destroy deleted domains */
7158 for (i
= 0; i
< ndoms_cur
; i
++) {
7159 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7160 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7161 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7164 /* no match - a current sched domain not in new doms_new[] */
7165 detach_destroy_domains(doms_cur
[i
]);
7171 if (doms_new
== NULL
) {
7173 doms_new
= &fallback_doms
;
7174 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7175 WARN_ON_ONCE(dattr_new
);
7178 /* Build new domains */
7179 for (i
= 0; i
< ndoms_new
; i
++) {
7180 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7181 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7182 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7185 /* no match - add a new doms_new */
7186 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7191 /* Remember the new sched domains */
7192 if (doms_cur
!= &fallback_doms
)
7193 free_sched_domains(doms_cur
, ndoms_cur
);
7194 kfree(dattr_cur
); /* kfree(NULL) is safe */
7195 doms_cur
= doms_new
;
7196 dattr_cur
= dattr_new
;
7197 ndoms_cur
= ndoms_new
;
7199 register_sched_domain_sysctl();
7201 mutex_unlock(&sched_domains_mutex
);
7204 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7207 * Update cpusets according to cpu_active mask. If cpusets are
7208 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7209 * around partition_sched_domains().
7211 * If we come here as part of a suspend/resume, don't touch cpusets because we
7212 * want to restore it back to its original state upon resume anyway.
7214 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7218 case CPU_ONLINE_FROZEN
:
7219 case CPU_DOWN_FAILED_FROZEN
:
7222 * num_cpus_frozen tracks how many CPUs are involved in suspend
7223 * resume sequence. As long as this is not the last online
7224 * operation in the resume sequence, just build a single sched
7225 * domain, ignoring cpusets.
7228 if (likely(num_cpus_frozen
)) {
7229 partition_sched_domains(1, NULL
, NULL
);
7234 * This is the last CPU online operation. So fall through and
7235 * restore the original sched domains by considering the
7236 * cpuset configurations.
7240 cpuset_update_active_cpus(true);
7248 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7251 unsigned long flags
;
7252 long cpu
= (long)hcpu
;
7258 case CPU_DOWN_PREPARE
:
7259 rcu_read_lock_sched();
7260 dl_b
= dl_bw_of(cpu
);
7262 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7263 cpus
= dl_bw_cpus(cpu
);
7264 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7265 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7267 rcu_read_unlock_sched();
7270 return notifier_from_errno(-EBUSY
);
7271 cpuset_update_active_cpus(false);
7273 case CPU_DOWN_PREPARE_FROZEN
:
7275 partition_sched_domains(1, NULL
, NULL
);
7283 void __init
sched_init_smp(void)
7285 cpumask_var_t non_isolated_cpus
;
7287 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7288 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7293 * There's no userspace yet to cause hotplug operations; hence all the
7294 * cpu masks are stable and all blatant races in the below code cannot
7297 mutex_lock(&sched_domains_mutex
);
7298 init_sched_domains(cpu_active_mask
);
7299 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7300 if (cpumask_empty(non_isolated_cpus
))
7301 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7302 mutex_unlock(&sched_domains_mutex
);
7304 hotcpu_notifier(sched_domains_numa_masks_update
, CPU_PRI_SCHED_ACTIVE
);
7305 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7306 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7310 /* Move init over to a non-isolated CPU */
7311 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7313 sched_init_granularity();
7314 free_cpumask_var(non_isolated_cpus
);
7316 init_sched_rt_class();
7317 init_sched_dl_class();
7320 void __init
sched_init_smp(void)
7322 sched_init_granularity();
7324 #endif /* CONFIG_SMP */
7326 int in_sched_functions(unsigned long addr
)
7328 return in_lock_functions(addr
) ||
7329 (addr
>= (unsigned long)__sched_text_start
7330 && addr
< (unsigned long)__sched_text_end
);
7333 #ifdef CONFIG_CGROUP_SCHED
7335 * Default task group.
7336 * Every task in system belongs to this group at bootup.
7338 struct task_group root_task_group
;
7339 LIST_HEAD(task_groups
);
7341 /* Cacheline aligned slab cache for task_group */
7342 static struct kmem_cache
*task_group_cache __read_mostly
;
7345 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7347 void __init
sched_init(void)
7350 unsigned long alloc_size
= 0, ptr
;
7352 #ifdef CONFIG_FAIR_GROUP_SCHED
7353 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7355 #ifdef CONFIG_RT_GROUP_SCHED
7356 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7359 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7361 #ifdef CONFIG_FAIR_GROUP_SCHED
7362 root_task_group
.se
= (struct sched_entity
**)ptr
;
7363 ptr
+= nr_cpu_ids
* sizeof(void **);
7365 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7366 ptr
+= nr_cpu_ids
* sizeof(void **);
7368 #endif /* CONFIG_FAIR_GROUP_SCHED */
7369 #ifdef CONFIG_RT_GROUP_SCHED
7370 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7371 ptr
+= nr_cpu_ids
* sizeof(void **);
7373 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7374 ptr
+= nr_cpu_ids
* sizeof(void **);
7376 #endif /* CONFIG_RT_GROUP_SCHED */
7378 #ifdef CONFIG_CPUMASK_OFFSTACK
7379 for_each_possible_cpu(i
) {
7380 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7381 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7383 #endif /* CONFIG_CPUMASK_OFFSTACK */
7385 init_rt_bandwidth(&def_rt_bandwidth
,
7386 global_rt_period(), global_rt_runtime());
7387 init_dl_bandwidth(&def_dl_bandwidth
,
7388 global_rt_period(), global_rt_runtime());
7391 init_defrootdomain();
7394 #ifdef CONFIG_RT_GROUP_SCHED
7395 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7396 global_rt_period(), global_rt_runtime());
7397 #endif /* CONFIG_RT_GROUP_SCHED */
7399 #ifdef CONFIG_CGROUP_SCHED
7400 task_group_cache
= KMEM_CACHE(task_group
, 0);
7402 list_add(&root_task_group
.list
, &task_groups
);
7403 INIT_LIST_HEAD(&root_task_group
.children
);
7404 INIT_LIST_HEAD(&root_task_group
.siblings
);
7405 autogroup_init(&init_task
);
7406 #endif /* CONFIG_CGROUP_SCHED */
7408 for_each_possible_cpu(i
) {
7412 raw_spin_lock_init(&rq
->lock
);
7414 rq
->calc_load_active
= 0;
7415 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7416 init_cfs_rq(&rq
->cfs
);
7417 init_rt_rq(&rq
->rt
);
7418 init_dl_rq(&rq
->dl
);
7419 #ifdef CONFIG_FAIR_GROUP_SCHED
7420 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7421 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7423 * How much cpu bandwidth does root_task_group get?
7425 * In case of task-groups formed thr' the cgroup filesystem, it
7426 * gets 100% of the cpu resources in the system. This overall
7427 * system cpu resource is divided among the tasks of
7428 * root_task_group and its child task-groups in a fair manner,
7429 * based on each entity's (task or task-group's) weight
7430 * (se->load.weight).
7432 * In other words, if root_task_group has 10 tasks of weight
7433 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7434 * then A0's share of the cpu resource is:
7436 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7438 * We achieve this by letting root_task_group's tasks sit
7439 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7441 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7442 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7443 #endif /* CONFIG_FAIR_GROUP_SCHED */
7445 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7446 #ifdef CONFIG_RT_GROUP_SCHED
7447 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7450 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7451 rq
->cpu_load
[j
] = 0;
7456 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7457 rq
->balance_callback
= NULL
;
7458 rq
->active_balance
= 0;
7459 rq
->next_balance
= jiffies
;
7464 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7465 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7467 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7469 rq_attach_root(rq
, &def_root_domain
);
7470 #ifdef CONFIG_NO_HZ_COMMON
7471 rq
->last_load_update_tick
= jiffies
;
7474 #ifdef CONFIG_NO_HZ_FULL
7475 rq
->last_sched_tick
= 0;
7477 #endif /* CONFIG_SMP */
7479 atomic_set(&rq
->nr_iowait
, 0);
7482 set_load_weight(&init_task
);
7484 #ifdef CONFIG_PREEMPT_NOTIFIERS
7485 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7489 * The boot idle thread does lazy MMU switching as well:
7491 atomic_inc(&init_mm
.mm_count
);
7492 enter_lazy_tlb(&init_mm
, current
);
7495 * During early bootup we pretend to be a normal task:
7497 current
->sched_class
= &fair_sched_class
;
7500 * Make us the idle thread. Technically, schedule() should not be
7501 * called from this thread, however somewhere below it might be,
7502 * but because we are the idle thread, we just pick up running again
7503 * when this runqueue becomes "idle".
7505 init_idle(current
, smp_processor_id());
7507 calc_load_update
= jiffies
+ LOAD_FREQ
;
7510 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7511 /* May be allocated at isolcpus cmdline parse time */
7512 if (cpu_isolated_map
== NULL
)
7513 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7514 idle_thread_set_boot_cpu();
7515 set_cpu_rq_start_time();
7517 init_sched_fair_class();
7519 scheduler_running
= 1;
7522 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7523 static inline int preempt_count_equals(int preempt_offset
)
7525 int nested
= preempt_count() + rcu_preempt_depth();
7527 return (nested
== preempt_offset
);
7530 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7533 * Blocking primitives will set (and therefore destroy) current->state,
7534 * since we will exit with TASK_RUNNING make sure we enter with it,
7535 * otherwise we will destroy state.
7537 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7538 "do not call blocking ops when !TASK_RUNNING; "
7539 "state=%lx set at [<%p>] %pS\n",
7541 (void *)current
->task_state_change
,
7542 (void *)current
->task_state_change
);
7544 ___might_sleep(file
, line
, preempt_offset
);
7546 EXPORT_SYMBOL(__might_sleep
);
7548 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7550 static unsigned long prev_jiffy
; /* ratelimiting */
7552 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7553 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7554 !is_idle_task(current
)) ||
7555 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7557 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7559 prev_jiffy
= jiffies
;
7562 "BUG: sleeping function called from invalid context at %s:%d\n",
7565 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7566 in_atomic(), irqs_disabled(),
7567 current
->pid
, current
->comm
);
7569 if (task_stack_end_corrupted(current
))
7570 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7572 debug_show_held_locks(current
);
7573 if (irqs_disabled())
7574 print_irqtrace_events(current
);
7575 #ifdef CONFIG_DEBUG_PREEMPT
7576 if (!preempt_count_equals(preempt_offset
)) {
7577 pr_err("Preemption disabled at:");
7578 print_ip_sym(current
->preempt_disable_ip
);
7584 EXPORT_SYMBOL(___might_sleep
);
7587 #ifdef CONFIG_MAGIC_SYSRQ
7588 void normalize_rt_tasks(void)
7590 struct task_struct
*g
, *p
;
7591 struct sched_attr attr
= {
7592 .sched_policy
= SCHED_NORMAL
,
7595 read_lock(&tasklist_lock
);
7596 for_each_process_thread(g
, p
) {
7598 * Only normalize user tasks:
7600 if (p
->flags
& PF_KTHREAD
)
7603 p
->se
.exec_start
= 0;
7604 #ifdef CONFIG_SCHEDSTATS
7605 p
->se
.statistics
.wait_start
= 0;
7606 p
->se
.statistics
.sleep_start
= 0;
7607 p
->se
.statistics
.block_start
= 0;
7610 if (!dl_task(p
) && !rt_task(p
)) {
7612 * Renice negative nice level userspace
7615 if (task_nice(p
) < 0)
7616 set_user_nice(p
, 0);
7620 __sched_setscheduler(p
, &attr
, false, false);
7622 read_unlock(&tasklist_lock
);
7625 #endif /* CONFIG_MAGIC_SYSRQ */
7627 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7629 * These functions are only useful for the IA64 MCA handling, or kdb.
7631 * They can only be called when the whole system has been
7632 * stopped - every CPU needs to be quiescent, and no scheduling
7633 * activity can take place. Using them for anything else would
7634 * be a serious bug, and as a result, they aren't even visible
7635 * under any other configuration.
7639 * curr_task - return the current task for a given cpu.
7640 * @cpu: the processor in question.
7642 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7644 * Return: The current task for @cpu.
7646 struct task_struct
*curr_task(int cpu
)
7648 return cpu_curr(cpu
);
7651 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7655 * set_curr_task - set the current task for a given cpu.
7656 * @cpu: the processor in question.
7657 * @p: the task pointer to set.
7659 * Description: This function must only be used when non-maskable interrupts
7660 * are serviced on a separate stack. It allows the architecture to switch the
7661 * notion of the current task on a cpu in a non-blocking manner. This function
7662 * must be called with all CPU's synchronized, and interrupts disabled, the
7663 * and caller must save the original value of the current task (see
7664 * curr_task() above) and restore that value before reenabling interrupts and
7665 * re-starting the system.
7667 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7669 void set_curr_task(int cpu
, struct task_struct
*p
)
7676 #ifdef CONFIG_CGROUP_SCHED
7677 /* task_group_lock serializes the addition/removal of task groups */
7678 static DEFINE_SPINLOCK(task_group_lock
);
7680 static void sched_free_group(struct task_group
*tg
)
7682 free_fair_sched_group(tg
);
7683 free_rt_sched_group(tg
);
7685 kmem_cache_free(task_group_cache
, tg
);
7688 /* allocate runqueue etc for a new task group */
7689 struct task_group
*sched_create_group(struct task_group
*parent
)
7691 struct task_group
*tg
;
7693 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7695 return ERR_PTR(-ENOMEM
);
7697 if (!alloc_fair_sched_group(tg
, parent
))
7700 if (!alloc_rt_sched_group(tg
, parent
))
7706 sched_free_group(tg
);
7707 return ERR_PTR(-ENOMEM
);
7710 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7712 unsigned long flags
;
7714 spin_lock_irqsave(&task_group_lock
, flags
);
7715 list_add_rcu(&tg
->list
, &task_groups
);
7717 WARN_ON(!parent
); /* root should already exist */
7719 tg
->parent
= parent
;
7720 INIT_LIST_HEAD(&tg
->children
);
7721 list_add_rcu(&tg
->siblings
, &parent
->children
);
7722 spin_unlock_irqrestore(&task_group_lock
, flags
);
7725 /* rcu callback to free various structures associated with a task group */
7726 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7728 /* now it should be safe to free those cfs_rqs */
7729 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7732 void sched_destroy_group(struct task_group
*tg
)
7734 /* wait for possible concurrent references to cfs_rqs complete */
7735 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7738 void sched_offline_group(struct task_group
*tg
)
7740 unsigned long flags
;
7742 /* end participation in shares distribution */
7743 unregister_fair_sched_group(tg
);
7745 spin_lock_irqsave(&task_group_lock
, flags
);
7746 list_del_rcu(&tg
->list
);
7747 list_del_rcu(&tg
->siblings
);
7748 spin_unlock_irqrestore(&task_group_lock
, flags
);
7751 /* change task's runqueue when it moves between groups.
7752 * The caller of this function should have put the task in its new group
7753 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7754 * reflect its new group.
7756 void sched_move_task(struct task_struct
*tsk
)
7758 struct task_group
*tg
;
7759 int queued
, running
;
7760 unsigned long flags
;
7763 rq
= task_rq_lock(tsk
, &flags
);
7765 running
= task_current(rq
, tsk
);
7766 queued
= task_on_rq_queued(tsk
);
7769 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7770 if (unlikely(running
))
7771 put_prev_task(rq
, tsk
);
7774 * All callers are synchronized by task_rq_lock(); we do not use RCU
7775 * which is pointless here. Thus, we pass "true" to task_css_check()
7776 * to prevent lockdep warnings.
7778 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7779 struct task_group
, css
);
7780 tg
= autogroup_task_group(tsk
, tg
);
7781 tsk
->sched_task_group
= tg
;
7783 #ifdef CONFIG_FAIR_GROUP_SCHED
7784 if (tsk
->sched_class
->task_move_group
)
7785 tsk
->sched_class
->task_move_group(tsk
);
7788 set_task_rq(tsk
, task_cpu(tsk
));
7790 if (unlikely(running
))
7791 tsk
->sched_class
->set_curr_task(rq
);
7793 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7795 task_rq_unlock(rq
, tsk
, &flags
);
7797 #endif /* CONFIG_CGROUP_SCHED */
7799 #ifdef CONFIG_RT_GROUP_SCHED
7801 * Ensure that the real time constraints are schedulable.
7803 static DEFINE_MUTEX(rt_constraints_mutex
);
7805 /* Must be called with tasklist_lock held */
7806 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7808 struct task_struct
*g
, *p
;
7811 * Autogroups do not have RT tasks; see autogroup_create().
7813 if (task_group_is_autogroup(tg
))
7816 for_each_process_thread(g
, p
) {
7817 if (rt_task(p
) && task_group(p
) == tg
)
7824 struct rt_schedulable_data
{
7825 struct task_group
*tg
;
7830 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7832 struct rt_schedulable_data
*d
= data
;
7833 struct task_group
*child
;
7834 unsigned long total
, sum
= 0;
7835 u64 period
, runtime
;
7837 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7838 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7841 period
= d
->rt_period
;
7842 runtime
= d
->rt_runtime
;
7846 * Cannot have more runtime than the period.
7848 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7852 * Ensure we don't starve existing RT tasks.
7854 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7857 total
= to_ratio(period
, runtime
);
7860 * Nobody can have more than the global setting allows.
7862 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7866 * The sum of our children's runtime should not exceed our own.
7868 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7869 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7870 runtime
= child
->rt_bandwidth
.rt_runtime
;
7872 if (child
== d
->tg
) {
7873 period
= d
->rt_period
;
7874 runtime
= d
->rt_runtime
;
7877 sum
+= to_ratio(period
, runtime
);
7886 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7890 struct rt_schedulable_data data
= {
7892 .rt_period
= period
,
7893 .rt_runtime
= runtime
,
7897 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7903 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7904 u64 rt_period
, u64 rt_runtime
)
7909 * Disallowing the root group RT runtime is BAD, it would disallow the
7910 * kernel creating (and or operating) RT threads.
7912 if (tg
== &root_task_group
&& rt_runtime
== 0)
7915 /* No period doesn't make any sense. */
7919 mutex_lock(&rt_constraints_mutex
);
7920 read_lock(&tasklist_lock
);
7921 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7925 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7926 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7927 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7929 for_each_possible_cpu(i
) {
7930 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7932 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7933 rt_rq
->rt_runtime
= rt_runtime
;
7934 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7936 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7938 read_unlock(&tasklist_lock
);
7939 mutex_unlock(&rt_constraints_mutex
);
7944 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
7946 u64 rt_runtime
, rt_period
;
7948 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7949 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
7950 if (rt_runtime_us
< 0)
7951 rt_runtime
= RUNTIME_INF
;
7953 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7956 static long sched_group_rt_runtime(struct task_group
*tg
)
7960 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
7963 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
7964 do_div(rt_runtime_us
, NSEC_PER_USEC
);
7965 return rt_runtime_us
;
7968 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
7970 u64 rt_runtime
, rt_period
;
7972 rt_period
= rt_period_us
* NSEC_PER_USEC
;
7973 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7975 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
7978 static long sched_group_rt_period(struct task_group
*tg
)
7982 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7983 do_div(rt_period_us
, NSEC_PER_USEC
);
7984 return rt_period_us
;
7986 #endif /* CONFIG_RT_GROUP_SCHED */
7988 #ifdef CONFIG_RT_GROUP_SCHED
7989 static int sched_rt_global_constraints(void)
7993 mutex_lock(&rt_constraints_mutex
);
7994 read_lock(&tasklist_lock
);
7995 ret
= __rt_schedulable(NULL
, 0, 0);
7996 read_unlock(&tasklist_lock
);
7997 mutex_unlock(&rt_constraints_mutex
);
8002 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8004 /* Don't accept realtime tasks when there is no way for them to run */
8005 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8011 #else /* !CONFIG_RT_GROUP_SCHED */
8012 static int sched_rt_global_constraints(void)
8014 unsigned long flags
;
8017 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8018 for_each_possible_cpu(i
) {
8019 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8021 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8022 rt_rq
->rt_runtime
= global_rt_runtime();
8023 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8025 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8029 #endif /* CONFIG_RT_GROUP_SCHED */
8031 static int sched_dl_global_validate(void)
8033 u64 runtime
= global_rt_runtime();
8034 u64 period
= global_rt_period();
8035 u64 new_bw
= to_ratio(period
, runtime
);
8038 unsigned long flags
;
8041 * Here we want to check the bandwidth not being set to some
8042 * value smaller than the currently allocated bandwidth in
8043 * any of the root_domains.
8045 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8046 * cycling on root_domains... Discussion on different/better
8047 * solutions is welcome!
8049 for_each_possible_cpu(cpu
) {
8050 rcu_read_lock_sched();
8051 dl_b
= dl_bw_of(cpu
);
8053 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8054 if (new_bw
< dl_b
->total_bw
)
8056 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8058 rcu_read_unlock_sched();
8067 static void sched_dl_do_global(void)
8072 unsigned long flags
;
8074 def_dl_bandwidth
.dl_period
= global_rt_period();
8075 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8077 if (global_rt_runtime() != RUNTIME_INF
)
8078 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8081 * FIXME: As above...
8083 for_each_possible_cpu(cpu
) {
8084 rcu_read_lock_sched();
8085 dl_b
= dl_bw_of(cpu
);
8087 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8089 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8091 rcu_read_unlock_sched();
8095 static int sched_rt_global_validate(void)
8097 if (sysctl_sched_rt_period
<= 0)
8100 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8101 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8107 static void sched_rt_do_global(void)
8109 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8110 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8113 int sched_rt_handler(struct ctl_table
*table
, int write
,
8114 void __user
*buffer
, size_t *lenp
,
8117 int old_period
, old_runtime
;
8118 static DEFINE_MUTEX(mutex
);
8122 old_period
= sysctl_sched_rt_period
;
8123 old_runtime
= sysctl_sched_rt_runtime
;
8125 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8127 if (!ret
&& write
) {
8128 ret
= sched_rt_global_validate();
8132 ret
= sched_dl_global_validate();
8136 ret
= sched_rt_global_constraints();
8140 sched_rt_do_global();
8141 sched_dl_do_global();
8145 sysctl_sched_rt_period
= old_period
;
8146 sysctl_sched_rt_runtime
= old_runtime
;
8148 mutex_unlock(&mutex
);
8153 int sched_rr_handler(struct ctl_table
*table
, int write
,
8154 void __user
*buffer
, size_t *lenp
,
8158 static DEFINE_MUTEX(mutex
);
8161 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8162 /* make sure that internally we keep jiffies */
8163 /* also, writing zero resets timeslice to default */
8164 if (!ret
&& write
) {
8165 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8166 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8168 mutex_unlock(&mutex
);
8172 #ifdef CONFIG_CGROUP_SCHED
8174 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8176 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8179 static struct cgroup_subsys_state
*
8180 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8182 struct task_group
*parent
= css_tg(parent_css
);
8183 struct task_group
*tg
;
8186 /* This is early initialization for the top cgroup */
8187 return &root_task_group
.css
;
8190 tg
= sched_create_group(parent
);
8192 return ERR_PTR(-ENOMEM
);
8194 sched_online_group(tg
, parent
);
8199 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8201 struct task_group
*tg
= css_tg(css
);
8203 sched_offline_group(tg
);
8206 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8208 struct task_group
*tg
= css_tg(css
);
8211 * Relies on the RCU grace period between css_released() and this.
8213 sched_free_group(tg
);
8216 static void cpu_cgroup_fork(struct task_struct
*task
)
8218 sched_move_task(task
);
8221 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8223 struct task_struct
*task
;
8224 struct cgroup_subsys_state
*css
;
8226 cgroup_taskset_for_each(task
, css
, tset
) {
8227 #ifdef CONFIG_RT_GROUP_SCHED
8228 if (!sched_rt_can_attach(css_tg(css
), task
))
8231 /* We don't support RT-tasks being in separate groups */
8232 if (task
->sched_class
!= &fair_sched_class
)
8239 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8241 struct task_struct
*task
;
8242 struct cgroup_subsys_state
*css
;
8244 cgroup_taskset_for_each(task
, css
, tset
)
8245 sched_move_task(task
);
8248 #ifdef CONFIG_FAIR_GROUP_SCHED
8249 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8250 struct cftype
*cftype
, u64 shareval
)
8252 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8255 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8258 struct task_group
*tg
= css_tg(css
);
8260 return (u64
) scale_load_down(tg
->shares
);
8263 #ifdef CONFIG_CFS_BANDWIDTH
8264 static DEFINE_MUTEX(cfs_constraints_mutex
);
8266 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8267 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8269 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8271 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8273 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8274 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8276 if (tg
== &root_task_group
)
8280 * Ensure we have at some amount of bandwidth every period. This is
8281 * to prevent reaching a state of large arrears when throttled via
8282 * entity_tick() resulting in prolonged exit starvation.
8284 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8288 * Likewise, bound things on the otherside by preventing insane quota
8289 * periods. This also allows us to normalize in computing quota
8292 if (period
> max_cfs_quota_period
)
8296 * Prevent race between setting of cfs_rq->runtime_enabled and
8297 * unthrottle_offline_cfs_rqs().
8300 mutex_lock(&cfs_constraints_mutex
);
8301 ret
= __cfs_schedulable(tg
, period
, quota
);
8305 runtime_enabled
= quota
!= RUNTIME_INF
;
8306 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8308 * If we need to toggle cfs_bandwidth_used, off->on must occur
8309 * before making related changes, and on->off must occur afterwards
8311 if (runtime_enabled
&& !runtime_was_enabled
)
8312 cfs_bandwidth_usage_inc();
8313 raw_spin_lock_irq(&cfs_b
->lock
);
8314 cfs_b
->period
= ns_to_ktime(period
);
8315 cfs_b
->quota
= quota
;
8317 __refill_cfs_bandwidth_runtime(cfs_b
);
8318 /* restart the period timer (if active) to handle new period expiry */
8319 if (runtime_enabled
)
8320 start_cfs_bandwidth(cfs_b
);
8321 raw_spin_unlock_irq(&cfs_b
->lock
);
8323 for_each_online_cpu(i
) {
8324 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8325 struct rq
*rq
= cfs_rq
->rq
;
8327 raw_spin_lock_irq(&rq
->lock
);
8328 cfs_rq
->runtime_enabled
= runtime_enabled
;
8329 cfs_rq
->runtime_remaining
= 0;
8331 if (cfs_rq
->throttled
)
8332 unthrottle_cfs_rq(cfs_rq
);
8333 raw_spin_unlock_irq(&rq
->lock
);
8335 if (runtime_was_enabled
&& !runtime_enabled
)
8336 cfs_bandwidth_usage_dec();
8338 mutex_unlock(&cfs_constraints_mutex
);
8344 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8348 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8349 if (cfs_quota_us
< 0)
8350 quota
= RUNTIME_INF
;
8352 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8354 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8357 long tg_get_cfs_quota(struct task_group
*tg
)
8361 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8364 quota_us
= tg
->cfs_bandwidth
.quota
;
8365 do_div(quota_us
, NSEC_PER_USEC
);
8370 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8374 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8375 quota
= tg
->cfs_bandwidth
.quota
;
8377 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8380 long tg_get_cfs_period(struct task_group
*tg
)
8384 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8385 do_div(cfs_period_us
, NSEC_PER_USEC
);
8387 return cfs_period_us
;
8390 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8393 return tg_get_cfs_quota(css_tg(css
));
8396 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8397 struct cftype
*cftype
, s64 cfs_quota_us
)
8399 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8402 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8405 return tg_get_cfs_period(css_tg(css
));
8408 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8409 struct cftype
*cftype
, u64 cfs_period_us
)
8411 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8414 struct cfs_schedulable_data
{
8415 struct task_group
*tg
;
8420 * normalize group quota/period to be quota/max_period
8421 * note: units are usecs
8423 static u64
normalize_cfs_quota(struct task_group
*tg
,
8424 struct cfs_schedulable_data
*d
)
8432 period
= tg_get_cfs_period(tg
);
8433 quota
= tg_get_cfs_quota(tg
);
8436 /* note: these should typically be equivalent */
8437 if (quota
== RUNTIME_INF
|| quota
== -1)
8440 return to_ratio(period
, quota
);
8443 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8445 struct cfs_schedulable_data
*d
= data
;
8446 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8447 s64 quota
= 0, parent_quota
= -1;
8450 quota
= RUNTIME_INF
;
8452 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8454 quota
= normalize_cfs_quota(tg
, d
);
8455 parent_quota
= parent_b
->hierarchical_quota
;
8458 * ensure max(child_quota) <= parent_quota, inherit when no
8461 if (quota
== RUNTIME_INF
)
8462 quota
= parent_quota
;
8463 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8466 cfs_b
->hierarchical_quota
= quota
;
8471 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8474 struct cfs_schedulable_data data
= {
8480 if (quota
!= RUNTIME_INF
) {
8481 do_div(data
.period
, NSEC_PER_USEC
);
8482 do_div(data
.quota
, NSEC_PER_USEC
);
8486 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8492 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8494 struct task_group
*tg
= css_tg(seq_css(sf
));
8495 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8497 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8498 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8499 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8503 #endif /* CONFIG_CFS_BANDWIDTH */
8504 #endif /* CONFIG_FAIR_GROUP_SCHED */
8506 #ifdef CONFIG_RT_GROUP_SCHED
8507 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8508 struct cftype
*cft
, s64 val
)
8510 return sched_group_set_rt_runtime(css_tg(css
), val
);
8513 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8516 return sched_group_rt_runtime(css_tg(css
));
8519 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8520 struct cftype
*cftype
, u64 rt_period_us
)
8522 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8525 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8528 return sched_group_rt_period(css_tg(css
));
8530 #endif /* CONFIG_RT_GROUP_SCHED */
8532 static struct cftype cpu_files
[] = {
8533 #ifdef CONFIG_FAIR_GROUP_SCHED
8536 .read_u64
= cpu_shares_read_u64
,
8537 .write_u64
= cpu_shares_write_u64
,
8540 #ifdef CONFIG_CFS_BANDWIDTH
8542 .name
= "cfs_quota_us",
8543 .read_s64
= cpu_cfs_quota_read_s64
,
8544 .write_s64
= cpu_cfs_quota_write_s64
,
8547 .name
= "cfs_period_us",
8548 .read_u64
= cpu_cfs_period_read_u64
,
8549 .write_u64
= cpu_cfs_period_write_u64
,
8553 .seq_show
= cpu_stats_show
,
8556 #ifdef CONFIG_RT_GROUP_SCHED
8558 .name
= "rt_runtime_us",
8559 .read_s64
= cpu_rt_runtime_read
,
8560 .write_s64
= cpu_rt_runtime_write
,
8563 .name
= "rt_period_us",
8564 .read_u64
= cpu_rt_period_read_uint
,
8565 .write_u64
= cpu_rt_period_write_uint
,
8571 struct cgroup_subsys cpu_cgrp_subsys
= {
8572 .css_alloc
= cpu_cgroup_css_alloc
,
8573 .css_released
= cpu_cgroup_css_released
,
8574 .css_free
= cpu_cgroup_css_free
,
8575 .fork
= cpu_cgroup_fork
,
8576 .can_attach
= cpu_cgroup_can_attach
,
8577 .attach
= cpu_cgroup_attach
,
8578 .legacy_cftypes
= cpu_files
,
8582 #endif /* CONFIG_CGROUP_SCHED */
8584 void dump_cpu_task(int cpu
)
8586 pr_info("Task dump for CPU %d:\n", cpu
);
8587 sched_show_task(cpu_curr(cpu
));
8591 * Nice levels are multiplicative, with a gentle 10% change for every
8592 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8593 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8594 * that remained on nice 0.
8596 * The "10% effect" is relative and cumulative: from _any_ nice level,
8597 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8598 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8599 * If a task goes up by ~10% and another task goes down by ~10% then
8600 * the relative distance between them is ~25%.)
8602 const int sched_prio_to_weight
[40] = {
8603 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8604 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8605 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8606 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8607 /* 0 */ 1024, 820, 655, 526, 423,
8608 /* 5 */ 335, 272, 215, 172, 137,
8609 /* 10 */ 110, 87, 70, 56, 45,
8610 /* 15 */ 36, 29, 23, 18, 15,
8614 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8616 * In cases where the weight does not change often, we can use the
8617 * precalculated inverse to speed up arithmetics by turning divisions
8618 * into multiplications:
8620 const u32 sched_prio_to_wmult
[40] = {
8621 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8622 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8623 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8624 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8625 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8626 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8627 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8628 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,