4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex
);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
97 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
99 void update_rq_clock(struct rq
*rq
)
103 lockdep_assert_held(&rq
->lock
);
105 if (rq
->clock_skip_update
& RQCF_ACT_SKIP
)
108 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
112 update_rq_clock_task(rq
, delta
);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug
unsigned int sysctl_sched_features
=
123 #include "features.h"
129 * Number of tasks to iterate in a single balance run.
130 * Limited because this is done with IRQs disabled.
132 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
135 * period over which we average the RT time consumption, measured
140 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
143 * period over which we measure -rt task cpu usage in us.
146 unsigned int sysctl_sched_rt_period
= 1000000;
148 __read_mostly
int scheduler_running
;
151 * part of the period that we allow rt tasks to run in us.
154 int sysctl_sched_rt_runtime
= 950000;
156 /* cpus with isolated domains */
157 cpumask_var_t cpu_isolated_map
;
160 * this_rq_lock - lock this runqueue and disable interrupts.
162 static struct rq
*this_rq_lock(void)
169 raw_spin_lock(&rq
->lock
);
175 * __task_rq_lock - lock the rq @p resides on.
177 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
182 lockdep_assert_held(&p
->pi_lock
);
186 raw_spin_lock(&rq
->lock
);
187 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
188 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
191 raw_spin_unlock(&rq
->lock
);
193 while (unlikely(task_on_rq_migrating(p
)))
199 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
201 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
202 __acquires(p
->pi_lock
)
208 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
210 raw_spin_lock(&rq
->lock
);
212 * move_queued_task() task_rq_lock()
215 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
216 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
217 * [S] ->cpu = new_cpu [L] task_rq()
221 * If we observe the old cpu in task_rq_lock, the acquire of
222 * the old rq->lock will fully serialize against the stores.
224 * If we observe the new cpu in task_rq_lock, the acquire will
225 * pair with the WMB to ensure we must then also see migrating.
227 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
228 rf
->cookie
= lockdep_pin_lock(&rq
->lock
);
231 raw_spin_unlock(&rq
->lock
);
232 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
234 while (unlikely(task_on_rq_migrating(p
)))
239 #ifdef CONFIG_SCHED_HRTICK
241 * Use HR-timers to deliver accurate preemption points.
244 static void hrtick_clear(struct rq
*rq
)
246 if (hrtimer_active(&rq
->hrtick_timer
))
247 hrtimer_cancel(&rq
->hrtick_timer
);
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
254 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
256 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
258 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
260 raw_spin_lock(&rq
->lock
);
262 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
263 raw_spin_unlock(&rq
->lock
);
265 return HRTIMER_NORESTART
;
270 static void __hrtick_restart(struct rq
*rq
)
272 struct hrtimer
*timer
= &rq
->hrtick_timer
;
274 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
278 * called from hardirq (IPI) context
280 static void __hrtick_start(void *arg
)
284 raw_spin_lock(&rq
->lock
);
285 __hrtick_restart(rq
);
286 rq
->hrtick_csd_pending
= 0;
287 raw_spin_unlock(&rq
->lock
);
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq
*rq
, u64 delay
)
297 struct hrtimer
*timer
= &rq
->hrtick_timer
;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta
= max_t(s64
, delay
, 10000LL);
306 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
308 hrtimer_set_expires(timer
, time
);
310 if (rq
== this_rq()) {
311 __hrtick_restart(rq
);
312 } else if (!rq
->hrtick_csd_pending
) {
313 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
314 rq
->hrtick_csd_pending
= 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq
*rq
, u64 delay
)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay
= max_t(u64
, delay
, 10000LL);
331 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
332 HRTIMER_MODE_REL_PINNED
);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq
*rq
)
339 rq
->hrtick_csd_pending
= 0;
341 rq
->hrtick_csd
.flags
= 0;
342 rq
->hrtick_csd
.func
= __hrtick_start
;
343 rq
->hrtick_csd
.info
= rq
;
346 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
347 rq
->hrtick_timer
.function
= hrtick
;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq
*rq
)
354 static inline void init_rq_hrtick(struct rq
*rq
)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct
*p
)
385 struct thread_info
*ti
= task_thread_info(p
);
386 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct
*p
)
397 struct thread_info
*ti
= task_thread_info(p
);
398 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
401 if (!(val
& _TIF_POLLING_NRFLAG
))
403 if (val
& _TIF_NEED_RESCHED
)
405 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
414 static bool set_nr_and_not_polling(struct task_struct
*p
)
416 set_tsk_need_resched(p
);
421 static bool set_nr_if_polling(struct task_struct
*p
)
428 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
430 struct wake_q_node
*node
= &task
->wake_q
;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
443 get_task_struct(task
);
446 * The head is context local, there can be no concurrency.
449 head
->lastp
= &node
->next
;
452 void wake_up_q(struct wake_q_head
*head
)
454 struct wake_q_node
*node
= head
->first
;
456 while (node
!= WAKE_Q_TAIL
) {
457 struct task_struct
*task
;
459 task
= container_of(node
, struct task_struct
, wake_q
);
461 /* task can safely be re-inserted now */
463 task
->wake_q
.next
= NULL
;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task
);
470 put_task_struct(task
);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq
*rq
)
483 struct task_struct
*curr
= rq
->curr
;
486 lockdep_assert_held(&rq
->lock
);
488 if (test_tsk_need_resched(curr
))
493 if (cpu
== smp_processor_id()) {
494 set_tsk_need_resched(curr
);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr
))
500 smp_send_reschedule(cpu
);
502 trace_sched_wake_idle_without_ipi(cpu
);
505 void resched_cpu(int cpu
)
507 struct rq
*rq
= cpu_rq(cpu
);
510 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
513 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i
, cpu
= smp_processor_id();
529 struct sched_domain
*sd
;
531 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
535 for_each_domain(cpu
, sd
) {
536 for_each_cpu(i
, sched_domain_span(sd
)) {
540 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
547 if (!is_housekeeping_cpu(cpu
))
548 cpu
= housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu
)
565 struct rq
*rq
= cpu_rq(cpu
);
567 if (cpu
== smp_processor_id())
570 if (set_nr_and_not_polling(rq
->idle
))
571 smp_send_reschedule(cpu
);
573 trace_sched_wake_idle_without_ipi(cpu
);
576 static bool wake_up_full_nohz_cpu(int cpu
)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (tick_nohz_full_cpu(cpu
)) {
585 if (cpu
!= smp_processor_id() ||
586 tick_nohz_tick_stopped())
587 tick_nohz_full_kick_cpu(cpu
);
594 void wake_up_nohz_cpu(int cpu
)
596 if (!wake_up_full_nohz_cpu(cpu
))
597 wake_up_idle_cpu(cpu
);
600 static inline bool got_nohz_idle_kick(void)
602 int cpu
= smp_processor_id();
604 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
607 if (idle_cpu(cpu
) && !need_resched())
611 * We can't run Idle Load Balance on this CPU for this time so we
612 * cancel it and clear NOHZ_BALANCE_KICK
614 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
618 #else /* CONFIG_NO_HZ_COMMON */
620 static inline bool got_nohz_idle_kick(void)
625 #endif /* CONFIG_NO_HZ_COMMON */
627 #ifdef CONFIG_NO_HZ_FULL
628 bool sched_can_stop_tick(struct rq
*rq
)
632 /* Deadline tasks, even if single, need the tick */
633 if (rq
->dl
.dl_nr_running
)
637 * If there are more than one RR tasks, we need the tick to effect the
638 * actual RR behaviour.
640 if (rq
->rt
.rr_nr_running
) {
641 if (rq
->rt
.rr_nr_running
== 1)
648 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
649 * forced preemption between FIFO tasks.
651 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
656 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
657 * if there's more than one we need the tick for involuntary
660 if (rq
->nr_running
> 1)
665 #endif /* CONFIG_NO_HZ_FULL */
667 void sched_avg_update(struct rq
*rq
)
669 s64 period
= sched_avg_period();
671 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
673 * Inline assembly required to prevent the compiler
674 * optimising this loop into a divmod call.
675 * See __iter_div_u64_rem() for another example of this.
677 asm("" : "+rm" (rq
->age_stamp
));
678 rq
->age_stamp
+= period
;
683 #endif /* CONFIG_SMP */
685 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
686 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
688 * Iterate task_group tree rooted at *from, calling @down when first entering a
689 * node and @up when leaving it for the final time.
691 * Caller must hold rcu_lock or sufficient equivalent.
693 int walk_tg_tree_from(struct task_group
*from
,
694 tg_visitor down
, tg_visitor up
, void *data
)
696 struct task_group
*parent
, *child
;
702 ret
= (*down
)(parent
, data
);
705 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
712 ret
= (*up
)(parent
, data
);
713 if (ret
|| parent
== from
)
717 parent
= parent
->parent
;
724 int tg_nop(struct task_group
*tg
, void *data
)
730 static void set_load_weight(struct task_struct
*p
)
732 int prio
= p
->static_prio
- MAX_RT_PRIO
;
733 struct load_weight
*load
= &p
->se
.load
;
736 * SCHED_IDLE tasks get minimal weight:
738 if (idle_policy(p
->policy
)) {
739 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
740 load
->inv_weight
= WMULT_IDLEPRIO
;
744 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
745 load
->inv_weight
= sched_prio_to_wmult
[prio
];
748 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
751 if (!(flags
& ENQUEUE_RESTORE
))
752 sched_info_queued(rq
, p
);
753 p
->sched_class
->enqueue_task(rq
, p
, flags
);
756 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
759 if (!(flags
& DEQUEUE_SAVE
))
760 sched_info_dequeued(rq
, p
);
761 p
->sched_class
->dequeue_task(rq
, p
, flags
);
764 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 if (task_contributes_to_load(p
))
767 rq
->nr_uninterruptible
--;
769 enqueue_task(rq
, p
, flags
);
772 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
774 if (task_contributes_to_load(p
))
775 rq
->nr_uninterruptible
++;
777 dequeue_task(rq
, p
, flags
);
780 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
783 * In theory, the compile should just see 0 here, and optimize out the call
784 * to sched_rt_avg_update. But I don't trust it...
786 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
787 s64 steal
= 0, irq_delta
= 0;
789 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
790 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
793 * Since irq_time is only updated on {soft,}irq_exit, we might run into
794 * this case when a previous update_rq_clock() happened inside a
797 * When this happens, we stop ->clock_task and only update the
798 * prev_irq_time stamp to account for the part that fit, so that a next
799 * update will consume the rest. This ensures ->clock_task is
802 * It does however cause some slight miss-attribution of {soft,}irq
803 * time, a more accurate solution would be to update the irq_time using
804 * the current rq->clock timestamp, except that would require using
807 if (irq_delta
> delta
)
810 rq
->prev_irq_time
+= irq_delta
;
813 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
814 if (static_key_false((¶virt_steal_rq_enabled
))) {
815 steal
= paravirt_steal_clock(cpu_of(rq
));
816 steal
-= rq
->prev_steal_time_rq
;
818 if (unlikely(steal
> delta
))
821 rq
->prev_steal_time_rq
+= steal
;
826 rq
->clock_task
+= delta
;
828 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
829 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
830 sched_rt_avg_update(rq
, irq_delta
+ steal
);
834 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
836 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
837 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
841 * Make it appear like a SCHED_FIFO task, its something
842 * userspace knows about and won't get confused about.
844 * Also, it will make PI more or less work without too
845 * much confusion -- but then, stop work should not
846 * rely on PI working anyway.
848 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
850 stop
->sched_class
= &stop_sched_class
;
853 cpu_rq(cpu
)->stop
= stop
;
857 * Reset it back to a normal scheduling class so that
858 * it can die in pieces.
860 old_stop
->sched_class
= &rt_sched_class
;
865 * __normal_prio - return the priority that is based on the static prio
867 static inline int __normal_prio(struct task_struct
*p
)
869 return p
->static_prio
;
873 * Calculate the expected normal priority: i.e. priority
874 * without taking RT-inheritance into account. Might be
875 * boosted by interactivity modifiers. Changes upon fork,
876 * setprio syscalls, and whenever the interactivity
877 * estimator recalculates.
879 static inline int normal_prio(struct task_struct
*p
)
883 if (task_has_dl_policy(p
))
884 prio
= MAX_DL_PRIO
-1;
885 else if (task_has_rt_policy(p
))
886 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
888 prio
= __normal_prio(p
);
893 * Calculate the current priority, i.e. the priority
894 * taken into account by the scheduler. This value might
895 * be boosted by RT tasks, or might be boosted by
896 * interactivity modifiers. Will be RT if the task got
897 * RT-boosted. If not then it returns p->normal_prio.
899 static int effective_prio(struct task_struct
*p
)
901 p
->normal_prio
= normal_prio(p
);
903 * If we are RT tasks or we were boosted to RT priority,
904 * keep the priority unchanged. Otherwise, update priority
905 * to the normal priority:
907 if (!rt_prio(p
->prio
))
908 return p
->normal_prio
;
913 * task_curr - is this task currently executing on a CPU?
914 * @p: the task in question.
916 * Return: 1 if the task is currently executing. 0 otherwise.
918 inline int task_curr(const struct task_struct
*p
)
920 return cpu_curr(task_cpu(p
)) == p
;
924 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
925 * use the balance_callback list if you want balancing.
927 * this means any call to check_class_changed() must be followed by a call to
928 * balance_callback().
930 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
931 const struct sched_class
*prev_class
,
934 if (prev_class
!= p
->sched_class
) {
935 if (prev_class
->switched_from
)
936 prev_class
->switched_from(rq
, p
);
938 p
->sched_class
->switched_to(rq
, p
);
939 } else if (oldprio
!= p
->prio
|| dl_task(p
))
940 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
943 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
945 const struct sched_class
*class;
947 if (p
->sched_class
== rq
->curr
->sched_class
) {
948 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
950 for_each_class(class) {
951 if (class == rq
->curr
->sched_class
)
953 if (class == p
->sched_class
) {
961 * A queue event has occurred, and we're going to schedule. In
962 * this case, we can save a useless back to back clock update.
964 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
965 rq_clock_skip_update(rq
, true);
970 * This is how migration works:
972 * 1) we invoke migration_cpu_stop() on the target CPU using
974 * 2) stopper starts to run (implicitly forcing the migrated thread
976 * 3) it checks whether the migrated task is still in the wrong runqueue.
977 * 4) if it's in the wrong runqueue then the migration thread removes
978 * it and puts it into the right queue.
979 * 5) stopper completes and stop_one_cpu() returns and the migration
984 * move_queued_task - move a queued task to new rq.
986 * Returns (locked) new rq. Old rq's lock is released.
988 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
990 lockdep_assert_held(&rq
->lock
);
992 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
993 dequeue_task(rq
, p
, 0);
994 set_task_cpu(p
, new_cpu
);
995 raw_spin_unlock(&rq
->lock
);
997 rq
= cpu_rq(new_cpu
);
999 raw_spin_lock(&rq
->lock
);
1000 BUG_ON(task_cpu(p
) != new_cpu
);
1001 enqueue_task(rq
, p
, 0);
1002 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1003 check_preempt_curr(rq
, p
, 0);
1008 struct migration_arg
{
1009 struct task_struct
*task
;
1014 * Move (not current) task off this cpu, onto dest cpu. We're doing
1015 * this because either it can't run here any more (set_cpus_allowed()
1016 * away from this CPU, or CPU going down), or because we're
1017 * attempting to rebalance this task on exec (sched_exec).
1019 * So we race with normal scheduler movements, but that's OK, as long
1020 * as the task is no longer on this CPU.
1022 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
1024 if (unlikely(!cpu_active(dest_cpu
)))
1027 /* Affinity changed (again). */
1028 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1031 rq
= move_queued_task(rq
, p
, dest_cpu
);
1037 * migration_cpu_stop - this will be executed by a highprio stopper thread
1038 * and performs thread migration by bumping thread off CPU then
1039 * 'pushing' onto another runqueue.
1041 static int migration_cpu_stop(void *data
)
1043 struct migration_arg
*arg
= data
;
1044 struct task_struct
*p
= arg
->task
;
1045 struct rq
*rq
= this_rq();
1048 * The original target cpu might have gone down and we might
1049 * be on another cpu but it doesn't matter.
1051 local_irq_disable();
1053 * We need to explicitly wake pending tasks before running
1054 * __migrate_task() such that we will not miss enforcing cpus_allowed
1055 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1057 sched_ttwu_pending();
1059 raw_spin_lock(&p
->pi_lock
);
1060 raw_spin_lock(&rq
->lock
);
1062 * If task_rq(p) != rq, it cannot be migrated here, because we're
1063 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1064 * we're holding p->pi_lock.
1066 if (task_rq(p
) == rq
&& task_on_rq_queued(p
))
1067 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1068 raw_spin_unlock(&rq
->lock
);
1069 raw_spin_unlock(&p
->pi_lock
);
1076 * sched_class::set_cpus_allowed must do the below, but is not required to
1077 * actually call this function.
1079 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1081 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1082 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1085 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1087 struct rq
*rq
= task_rq(p
);
1088 bool queued
, running
;
1090 lockdep_assert_held(&p
->pi_lock
);
1092 queued
= task_on_rq_queued(p
);
1093 running
= task_current(rq
, p
);
1097 * Because __kthread_bind() calls this on blocked tasks without
1100 lockdep_assert_held(&rq
->lock
);
1101 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1104 put_prev_task(rq
, p
);
1106 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1109 p
->sched_class
->set_curr_task(rq
);
1111 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1115 * Change a given task's CPU affinity. Migrate the thread to a
1116 * proper CPU and schedule it away if the CPU it's executing on
1117 * is removed from the allowed bitmask.
1119 * NOTE: the caller must have a valid reference to the task, the
1120 * task must not exit() & deallocate itself prematurely. The
1121 * call is not atomic; no spinlocks may be held.
1123 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1124 const struct cpumask
*new_mask
, bool check
)
1126 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1127 unsigned int dest_cpu
;
1132 rq
= task_rq_lock(p
, &rf
);
1134 if (p
->flags
& PF_KTHREAD
) {
1136 * Kernel threads are allowed on online && !active CPUs
1138 cpu_valid_mask
= cpu_online_mask
;
1142 * Must re-check here, to close a race against __kthread_bind(),
1143 * sched_setaffinity() is not guaranteed to observe the flag.
1145 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1150 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1153 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1158 do_set_cpus_allowed(p
, new_mask
);
1160 if (p
->flags
& PF_KTHREAD
) {
1162 * For kernel threads that do indeed end up on online &&
1163 * !active we want to ensure they are strict per-cpu threads.
1165 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1166 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1167 p
->nr_cpus_allowed
!= 1);
1170 /* Can the task run on the task's current CPU? If so, we're done */
1171 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1174 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1175 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1176 struct migration_arg arg
= { p
, dest_cpu
};
1177 /* Need help from migration thread: drop lock and wait. */
1178 task_rq_unlock(rq
, p
, &rf
);
1179 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1180 tlb_migrate_finish(p
->mm
);
1182 } else if (task_on_rq_queued(p
)) {
1184 * OK, since we're going to drop the lock immediately
1185 * afterwards anyway.
1187 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
1188 rq
= move_queued_task(rq
, p
, dest_cpu
);
1189 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
1192 task_rq_unlock(rq
, p
, &rf
);
1197 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1199 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1201 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1203 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1205 #ifdef CONFIG_SCHED_DEBUG
1207 * We should never call set_task_cpu() on a blocked task,
1208 * ttwu() will sort out the placement.
1210 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1214 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1215 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1216 * time relying on p->on_rq.
1218 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1219 p
->sched_class
== &fair_sched_class
&&
1220 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1222 #ifdef CONFIG_LOCKDEP
1224 * The caller should hold either p->pi_lock or rq->lock, when changing
1225 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1227 * sched_move_task() holds both and thus holding either pins the cgroup,
1230 * Furthermore, all task_rq users should acquire both locks, see
1233 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1234 lockdep_is_held(&task_rq(p
)->lock
)));
1238 trace_sched_migrate_task(p
, new_cpu
);
1240 if (task_cpu(p
) != new_cpu
) {
1241 if (p
->sched_class
->migrate_task_rq
)
1242 p
->sched_class
->migrate_task_rq(p
);
1243 p
->se
.nr_migrations
++;
1244 perf_event_task_migrate(p
);
1247 __set_task_cpu(p
, new_cpu
);
1250 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1252 if (task_on_rq_queued(p
)) {
1253 struct rq
*src_rq
, *dst_rq
;
1255 src_rq
= task_rq(p
);
1256 dst_rq
= cpu_rq(cpu
);
1258 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1259 deactivate_task(src_rq
, p
, 0);
1260 set_task_cpu(p
, cpu
);
1261 activate_task(dst_rq
, p
, 0);
1262 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1263 check_preempt_curr(dst_rq
, p
, 0);
1266 * Task isn't running anymore; make it appear like we migrated
1267 * it before it went to sleep. This means on wakeup we make the
1268 * previous cpu our targer instead of where it really is.
1274 struct migration_swap_arg
{
1275 struct task_struct
*src_task
, *dst_task
;
1276 int src_cpu
, dst_cpu
;
1279 static int migrate_swap_stop(void *data
)
1281 struct migration_swap_arg
*arg
= data
;
1282 struct rq
*src_rq
, *dst_rq
;
1285 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1288 src_rq
= cpu_rq(arg
->src_cpu
);
1289 dst_rq
= cpu_rq(arg
->dst_cpu
);
1291 double_raw_lock(&arg
->src_task
->pi_lock
,
1292 &arg
->dst_task
->pi_lock
);
1293 double_rq_lock(src_rq
, dst_rq
);
1295 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1298 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1301 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1304 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1307 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1308 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1313 double_rq_unlock(src_rq
, dst_rq
);
1314 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1315 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1321 * Cross migrate two tasks
1323 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1325 struct migration_swap_arg arg
;
1328 arg
= (struct migration_swap_arg
){
1330 .src_cpu
= task_cpu(cur
),
1332 .dst_cpu
= task_cpu(p
),
1335 if (arg
.src_cpu
== arg
.dst_cpu
)
1339 * These three tests are all lockless; this is OK since all of them
1340 * will be re-checked with proper locks held further down the line.
1342 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1345 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1348 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1351 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1352 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1359 * wait_task_inactive - wait for a thread to unschedule.
1361 * If @match_state is nonzero, it's the @p->state value just checked and
1362 * not expected to change. If it changes, i.e. @p might have woken up,
1363 * then return zero. When we succeed in waiting for @p to be off its CPU,
1364 * we return a positive number (its total switch count). If a second call
1365 * a short while later returns the same number, the caller can be sure that
1366 * @p has remained unscheduled the whole time.
1368 * The caller must ensure that the task *will* unschedule sometime soon,
1369 * else this function might spin for a *long* time. This function can't
1370 * be called with interrupts off, or it may introduce deadlock with
1371 * smp_call_function() if an IPI is sent by the same process we are
1372 * waiting to become inactive.
1374 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1376 int running
, queued
;
1383 * We do the initial early heuristics without holding
1384 * any task-queue locks at all. We'll only try to get
1385 * the runqueue lock when things look like they will
1391 * If the task is actively running on another CPU
1392 * still, just relax and busy-wait without holding
1395 * NOTE! Since we don't hold any locks, it's not
1396 * even sure that "rq" stays as the right runqueue!
1397 * But we don't care, since "task_running()" will
1398 * return false if the runqueue has changed and p
1399 * is actually now running somewhere else!
1401 while (task_running(rq
, p
)) {
1402 if (match_state
&& unlikely(p
->state
!= match_state
))
1408 * Ok, time to look more closely! We need the rq
1409 * lock now, to be *sure*. If we're wrong, we'll
1410 * just go back and repeat.
1412 rq
= task_rq_lock(p
, &rf
);
1413 trace_sched_wait_task(p
);
1414 running
= task_running(rq
, p
);
1415 queued
= task_on_rq_queued(p
);
1417 if (!match_state
|| p
->state
== match_state
)
1418 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1419 task_rq_unlock(rq
, p
, &rf
);
1422 * If it changed from the expected state, bail out now.
1424 if (unlikely(!ncsw
))
1428 * Was it really running after all now that we
1429 * checked with the proper locks actually held?
1431 * Oops. Go back and try again..
1433 if (unlikely(running
)) {
1439 * It's not enough that it's not actively running,
1440 * it must be off the runqueue _entirely_, and not
1443 * So if it was still runnable (but just not actively
1444 * running right now), it's preempted, and we should
1445 * yield - it could be a while.
1447 if (unlikely(queued
)) {
1448 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
1450 set_current_state(TASK_UNINTERRUPTIBLE
);
1451 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1456 * Ahh, all good. It wasn't running, and it wasn't
1457 * runnable, which means that it will never become
1458 * running in the future either. We're all done!
1467 * kick_process - kick a running thread to enter/exit the kernel
1468 * @p: the to-be-kicked thread
1470 * Cause a process which is running on another CPU to enter
1471 * kernel-mode, without any delay. (to get signals handled.)
1473 * NOTE: this function doesn't have to take the runqueue lock,
1474 * because all it wants to ensure is that the remote task enters
1475 * the kernel. If the IPI races and the task has been migrated
1476 * to another CPU then no harm is done and the purpose has been
1479 void kick_process(struct task_struct
*p
)
1485 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1486 smp_send_reschedule(cpu
);
1489 EXPORT_SYMBOL_GPL(kick_process
);
1492 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1494 * A few notes on cpu_active vs cpu_online:
1496 * - cpu_active must be a subset of cpu_online
1498 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1499 * see __set_cpus_allowed_ptr(). At this point the newly online
1500 * cpu isn't yet part of the sched domains, and balancing will not
1503 * - on cpu-down we clear cpu_active() to mask the sched domains and
1504 * avoid the load balancer to place new tasks on the to be removed
1505 * cpu. Existing tasks will remain running there and will be taken
1508 * This means that fallback selection must not select !active CPUs.
1509 * And can assume that any active CPU must be online. Conversely
1510 * select_task_rq() below may allow selection of !active CPUs in order
1511 * to satisfy the above rules.
1513 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1515 int nid
= cpu_to_node(cpu
);
1516 const struct cpumask
*nodemask
= NULL
;
1517 enum { cpuset
, possible
, fail
} state
= cpuset
;
1521 * If the node that the cpu is on has been offlined, cpu_to_node()
1522 * will return -1. There is no cpu on the node, and we should
1523 * select the cpu on the other node.
1526 nodemask
= cpumask_of_node(nid
);
1528 /* Look for allowed, online CPU in same node. */
1529 for_each_cpu(dest_cpu
, nodemask
) {
1530 if (!cpu_active(dest_cpu
))
1532 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1538 /* Any allowed, online CPU? */
1539 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1540 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1542 if (!cpu_online(dest_cpu
))
1547 /* No more Mr. Nice Guy. */
1550 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1551 cpuset_cpus_allowed_fallback(p
);
1557 do_set_cpus_allowed(p
, cpu_possible_mask
);
1568 if (state
!= cpuset
) {
1570 * Don't tell them about moving exiting tasks or
1571 * kernel threads (both mm NULL), since they never
1574 if (p
->mm
&& printk_ratelimit()) {
1575 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1576 task_pid_nr(p
), p
->comm
, cpu
);
1584 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1587 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1589 lockdep_assert_held(&p
->pi_lock
);
1591 if (tsk_nr_cpus_allowed(p
) > 1)
1592 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1594 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1597 * In order not to call set_task_cpu() on a blocking task we need
1598 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1601 * Since this is common to all placement strategies, this lives here.
1603 * [ this allows ->select_task() to simply return task_cpu(p) and
1604 * not worry about this generic constraint ]
1606 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1608 cpu
= select_fallback_rq(task_cpu(p
), p
);
1613 static void update_avg(u64
*avg
, u64 sample
)
1615 s64 diff
= sample
- *avg
;
1621 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1622 const struct cpumask
*new_mask
, bool check
)
1624 return set_cpus_allowed_ptr(p
, new_mask
);
1627 #endif /* CONFIG_SMP */
1630 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1632 #ifdef CONFIG_SCHEDSTATS
1633 struct rq
*rq
= this_rq();
1636 int this_cpu
= smp_processor_id();
1638 if (cpu
== this_cpu
) {
1639 schedstat_inc(rq
, ttwu_local
);
1640 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
1642 struct sched_domain
*sd
;
1644 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
1646 for_each_domain(this_cpu
, sd
) {
1647 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1648 schedstat_inc(sd
, ttwu_wake_remote
);
1655 if (wake_flags
& WF_MIGRATED
)
1656 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
1658 #endif /* CONFIG_SMP */
1660 schedstat_inc(rq
, ttwu_count
);
1661 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
1663 if (wake_flags
& WF_SYNC
)
1664 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
1666 #endif /* CONFIG_SCHEDSTATS */
1669 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1671 activate_task(rq
, p
, en_flags
);
1672 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1674 /* if a worker is waking up, notify workqueue */
1675 if (p
->flags
& PF_WQ_WORKER
)
1676 wq_worker_waking_up(p
, cpu_of(rq
));
1680 * Mark the task runnable and perform wakeup-preemption.
1682 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1683 struct pin_cookie cookie
)
1685 check_preempt_curr(rq
, p
, wake_flags
);
1686 p
->state
= TASK_RUNNING
;
1687 trace_sched_wakeup(p
);
1690 if (p
->sched_class
->task_woken
) {
1692 * Our task @p is fully woken up and running; so its safe to
1693 * drop the rq->lock, hereafter rq is only used for statistics.
1695 lockdep_unpin_lock(&rq
->lock
, cookie
);
1696 p
->sched_class
->task_woken(rq
, p
);
1697 lockdep_repin_lock(&rq
->lock
, cookie
);
1700 if (rq
->idle_stamp
) {
1701 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1702 u64 max
= 2*rq
->max_idle_balance_cost
;
1704 update_avg(&rq
->avg_idle
, delta
);
1706 if (rq
->avg_idle
> max
)
1715 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1716 struct pin_cookie cookie
)
1718 int en_flags
= ENQUEUE_WAKEUP
;
1720 lockdep_assert_held(&rq
->lock
);
1723 if (p
->sched_contributes_to_load
)
1724 rq
->nr_uninterruptible
--;
1726 if (wake_flags
& WF_MIGRATED
)
1727 en_flags
|= ENQUEUE_MIGRATED
;
1730 ttwu_activate(rq
, p
, en_flags
);
1731 ttwu_do_wakeup(rq
, p
, wake_flags
, cookie
);
1735 * Called in case the task @p isn't fully descheduled from its runqueue,
1736 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1737 * since all we need to do is flip p->state to TASK_RUNNING, since
1738 * the task is still ->on_rq.
1740 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1746 rq
= __task_rq_lock(p
, &rf
);
1747 if (task_on_rq_queued(p
)) {
1748 /* check_preempt_curr() may use rq clock */
1749 update_rq_clock(rq
);
1750 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
.cookie
);
1753 __task_rq_unlock(rq
, &rf
);
1759 void sched_ttwu_pending(void)
1761 struct rq
*rq
= this_rq();
1762 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1763 struct pin_cookie cookie
;
1764 struct task_struct
*p
;
1765 unsigned long flags
;
1770 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1771 cookie
= lockdep_pin_lock(&rq
->lock
);
1776 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1777 llist
= llist_next(llist
);
1779 if (p
->sched_remote_wakeup
)
1780 wake_flags
= WF_MIGRATED
;
1782 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1785 lockdep_unpin_lock(&rq
->lock
, cookie
);
1786 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1789 void scheduler_ipi(void)
1792 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1793 * TIF_NEED_RESCHED remotely (for the first time) will also send
1796 preempt_fold_need_resched();
1798 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1802 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1803 * traditionally all their work was done from the interrupt return
1804 * path. Now that we actually do some work, we need to make sure
1807 * Some archs already do call them, luckily irq_enter/exit nest
1810 * Arguably we should visit all archs and update all handlers,
1811 * however a fair share of IPIs are still resched only so this would
1812 * somewhat pessimize the simple resched case.
1815 sched_ttwu_pending();
1818 * Check if someone kicked us for doing the nohz idle load balance.
1820 if (unlikely(got_nohz_idle_kick())) {
1821 this_rq()->idle_balance
= 1;
1822 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1827 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1829 struct rq
*rq
= cpu_rq(cpu
);
1831 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1833 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1834 if (!set_nr_if_polling(rq
->idle
))
1835 smp_send_reschedule(cpu
);
1837 trace_sched_wake_idle_without_ipi(cpu
);
1841 void wake_up_if_idle(int cpu
)
1843 struct rq
*rq
= cpu_rq(cpu
);
1844 unsigned long flags
;
1848 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1851 if (set_nr_if_polling(rq
->idle
)) {
1852 trace_sched_wake_idle_without_ipi(cpu
);
1854 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1855 if (is_idle_task(rq
->curr
))
1856 smp_send_reschedule(cpu
);
1857 /* Else cpu is not in idle, do nothing here */
1858 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1865 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1867 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1869 #endif /* CONFIG_SMP */
1871 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1873 struct rq
*rq
= cpu_rq(cpu
);
1874 struct pin_cookie cookie
;
1876 #if defined(CONFIG_SMP)
1877 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1878 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
1879 ttwu_queue_remote(p
, cpu
, wake_flags
);
1884 raw_spin_lock(&rq
->lock
);
1885 cookie
= lockdep_pin_lock(&rq
->lock
);
1886 ttwu_do_activate(rq
, p
, wake_flags
, cookie
);
1887 lockdep_unpin_lock(&rq
->lock
, cookie
);
1888 raw_spin_unlock(&rq
->lock
);
1892 * Notes on Program-Order guarantees on SMP systems.
1896 * The basic program-order guarantee on SMP systems is that when a task [t]
1897 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1898 * execution on its new cpu [c1].
1900 * For migration (of runnable tasks) this is provided by the following means:
1902 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1903 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1904 * rq(c1)->lock (if not at the same time, then in that order).
1905 * C) LOCK of the rq(c1)->lock scheduling in task
1907 * Transitivity guarantees that B happens after A and C after B.
1908 * Note: we only require RCpc transitivity.
1909 * Note: the cpu doing B need not be c0 or c1
1918 * UNLOCK rq(0)->lock
1920 * LOCK rq(0)->lock // orders against CPU0
1922 * UNLOCK rq(0)->lock
1926 * UNLOCK rq(1)->lock
1928 * LOCK rq(1)->lock // orders against CPU2
1931 * UNLOCK rq(1)->lock
1934 * BLOCKING -- aka. SLEEP + WAKEUP
1936 * For blocking we (obviously) need to provide the same guarantee as for
1937 * migration. However the means are completely different as there is no lock
1938 * chain to provide order. Instead we do:
1940 * 1) smp_store_release(X->on_cpu, 0)
1941 * 2) smp_cond_load_acquire(!X->on_cpu)
1945 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1947 * LOCK rq(0)->lock LOCK X->pi_lock
1950 * smp_store_release(X->on_cpu, 0);
1952 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1958 * X->state = RUNNING
1959 * UNLOCK rq(2)->lock
1961 * LOCK rq(2)->lock // orders against CPU1
1964 * UNLOCK rq(2)->lock
1967 * UNLOCK rq(0)->lock
1970 * However; for wakeups there is a second guarantee we must provide, namely we
1971 * must observe the state that lead to our wakeup. That is, not only must our
1972 * task observe its own prior state, it must also observe the stores prior to
1975 * This means that any means of doing remote wakeups must order the CPU doing
1976 * the wakeup against the CPU the task is going to end up running on. This,
1977 * however, is already required for the regular Program-Order guarantee above,
1978 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1983 * try_to_wake_up - wake up a thread
1984 * @p: the thread to be awakened
1985 * @state: the mask of task states that can be woken
1986 * @wake_flags: wake modifier flags (WF_*)
1988 * Put it on the run-queue if it's not already there. The "current"
1989 * thread is always on the run-queue (except when the actual
1990 * re-schedule is in progress), and as such you're allowed to do
1991 * the simpler "current->state = TASK_RUNNING" to mark yourself
1992 * runnable without the overhead of this.
1994 * Return: %true if @p was woken up, %false if it was already running.
1995 * or @state didn't match @p's state.
1998 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2000 unsigned long flags
;
2001 int cpu
, success
= 0;
2004 * If we are going to wake up a thread waiting for CONDITION we
2005 * need to ensure that CONDITION=1 done by the caller can not be
2006 * reordered with p->state check below. This pairs with mb() in
2007 * set_current_state() the waiting thread does.
2009 smp_mb__before_spinlock();
2010 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2011 if (!(p
->state
& state
))
2014 trace_sched_waking(p
);
2016 success
= 1; /* we're going to change ->state */
2020 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2021 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2022 * in smp_cond_load_acquire() below.
2024 * sched_ttwu_pending() try_to_wake_up()
2025 * [S] p->on_rq = 1; [L] P->state
2026 * UNLOCK rq->lock -----.
2030 * LOCK rq->lock -----'
2034 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2036 * Pairs with the UNLOCK+LOCK on rq->lock from the
2037 * last wakeup of our task and the schedule that got our task
2041 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2046 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2047 * possible to, falsely, observe p->on_cpu == 0.
2049 * One must be running (->on_cpu == 1) in order to remove oneself
2050 * from the runqueue.
2052 * [S] ->on_cpu = 1; [L] ->on_rq
2056 * [S] ->on_rq = 0; [L] ->on_cpu
2058 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2059 * from the consecutive calls to schedule(); the first switching to our
2060 * task, the second putting it to sleep.
2065 * If the owning (remote) cpu is still in the middle of schedule() with
2066 * this task as prev, wait until its done referencing the task.
2068 * Pairs with the smp_store_release() in finish_lock_switch().
2070 * This ensures that tasks getting woken will be fully ordered against
2071 * their previous state and preserve Program Order.
2073 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2075 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2076 p
->state
= TASK_WAKING
;
2078 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2079 if (task_cpu(p
) != cpu
) {
2080 wake_flags
|= WF_MIGRATED
;
2081 set_task_cpu(p
, cpu
);
2083 #endif /* CONFIG_SMP */
2085 ttwu_queue(p
, cpu
, wake_flags
);
2087 if (schedstat_enabled())
2088 ttwu_stat(p
, cpu
, wake_flags
);
2090 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2096 * try_to_wake_up_local - try to wake up a local task with rq lock held
2097 * @p: the thread to be awakened
2099 * Put @p on the run-queue if it's not already there. The caller must
2100 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2103 static void try_to_wake_up_local(struct task_struct
*p
, struct pin_cookie cookie
)
2105 struct rq
*rq
= task_rq(p
);
2107 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2108 WARN_ON_ONCE(p
== current
))
2111 lockdep_assert_held(&rq
->lock
);
2113 if (!raw_spin_trylock(&p
->pi_lock
)) {
2115 * This is OK, because current is on_cpu, which avoids it being
2116 * picked for load-balance and preemption/IRQs are still
2117 * disabled avoiding further scheduler activity on it and we've
2118 * not yet picked a replacement task.
2120 lockdep_unpin_lock(&rq
->lock
, cookie
);
2121 raw_spin_unlock(&rq
->lock
);
2122 raw_spin_lock(&p
->pi_lock
);
2123 raw_spin_lock(&rq
->lock
);
2124 lockdep_repin_lock(&rq
->lock
, cookie
);
2127 if (!(p
->state
& TASK_NORMAL
))
2130 trace_sched_waking(p
);
2132 if (!task_on_rq_queued(p
))
2133 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2135 ttwu_do_wakeup(rq
, p
, 0, cookie
);
2136 if (schedstat_enabled())
2137 ttwu_stat(p
, smp_processor_id(), 0);
2139 raw_spin_unlock(&p
->pi_lock
);
2143 * wake_up_process - Wake up a specific process
2144 * @p: The process to be woken up.
2146 * Attempt to wake up the nominated process and move it to the set of runnable
2149 * Return: 1 if the process was woken up, 0 if it was already running.
2151 * It may be assumed that this function implies a write memory barrier before
2152 * changing the task state if and only if any tasks are woken up.
2154 int wake_up_process(struct task_struct
*p
)
2156 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2158 EXPORT_SYMBOL(wake_up_process
);
2160 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2162 return try_to_wake_up(p
, state
, 0);
2166 * This function clears the sched_dl_entity static params.
2168 void __dl_clear_params(struct task_struct
*p
)
2170 struct sched_dl_entity
*dl_se
= &p
->dl
;
2172 dl_se
->dl_runtime
= 0;
2173 dl_se
->dl_deadline
= 0;
2174 dl_se
->dl_period
= 0;
2178 dl_se
->dl_throttled
= 0;
2179 dl_se
->dl_yielded
= 0;
2183 * Perform scheduler related setup for a newly forked process p.
2184 * p is forked by current.
2186 * __sched_fork() is basic setup used by init_idle() too:
2188 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2193 p
->se
.exec_start
= 0;
2194 p
->se
.sum_exec_runtime
= 0;
2195 p
->se
.prev_sum_exec_runtime
= 0;
2196 p
->se
.nr_migrations
= 0;
2198 INIT_LIST_HEAD(&p
->se
.group_node
);
2200 #ifdef CONFIG_FAIR_GROUP_SCHED
2201 p
->se
.cfs_rq
= NULL
;
2204 #ifdef CONFIG_SCHEDSTATS
2205 /* Even if schedstat is disabled, there should not be garbage */
2206 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2209 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2210 init_dl_task_timer(&p
->dl
);
2211 __dl_clear_params(p
);
2213 INIT_LIST_HEAD(&p
->rt
.run_list
);
2215 p
->rt
.time_slice
= sched_rr_timeslice
;
2219 #ifdef CONFIG_PREEMPT_NOTIFIERS
2220 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2223 #ifdef CONFIG_NUMA_BALANCING
2224 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2225 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2226 p
->mm
->numa_scan_seq
= 0;
2229 if (clone_flags
& CLONE_VM
)
2230 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2232 p
->numa_preferred_nid
= -1;
2234 p
->node_stamp
= 0ULL;
2235 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2236 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2237 p
->numa_work
.next
= &p
->numa_work
;
2238 p
->numa_faults
= NULL
;
2239 p
->last_task_numa_placement
= 0;
2240 p
->last_sum_exec_runtime
= 0;
2242 p
->numa_group
= NULL
;
2243 #endif /* CONFIG_NUMA_BALANCING */
2246 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2248 #ifdef CONFIG_NUMA_BALANCING
2250 void set_numabalancing_state(bool enabled
)
2253 static_branch_enable(&sched_numa_balancing
);
2255 static_branch_disable(&sched_numa_balancing
);
2258 #ifdef CONFIG_PROC_SYSCTL
2259 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2260 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2264 int state
= static_branch_likely(&sched_numa_balancing
);
2266 if (write
&& !capable(CAP_SYS_ADMIN
))
2271 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2275 set_numabalancing_state(state
);
2281 #ifdef CONFIG_SCHEDSTATS
2283 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2284 static bool __initdata __sched_schedstats
= false;
2286 static void set_schedstats(bool enabled
)
2289 static_branch_enable(&sched_schedstats
);
2291 static_branch_disable(&sched_schedstats
);
2294 void force_schedstat_enabled(void)
2296 if (!schedstat_enabled()) {
2297 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2298 static_branch_enable(&sched_schedstats
);
2302 static int __init
setup_schedstats(char *str
)
2309 * This code is called before jump labels have been set up, so we can't
2310 * change the static branch directly just yet. Instead set a temporary
2311 * variable so init_schedstats() can do it later.
2313 if (!strcmp(str
, "enable")) {
2314 __sched_schedstats
= true;
2316 } else if (!strcmp(str
, "disable")) {
2317 __sched_schedstats
= false;
2322 pr_warn("Unable to parse schedstats=\n");
2326 __setup("schedstats=", setup_schedstats
);
2328 static void __init
init_schedstats(void)
2330 set_schedstats(__sched_schedstats
);
2333 #ifdef CONFIG_PROC_SYSCTL
2334 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2335 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2339 int state
= static_branch_likely(&sched_schedstats
);
2341 if (write
&& !capable(CAP_SYS_ADMIN
))
2346 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2350 set_schedstats(state
);
2353 #endif /* CONFIG_PROC_SYSCTL */
2354 #else /* !CONFIG_SCHEDSTATS */
2355 static inline void init_schedstats(void) {}
2356 #endif /* CONFIG_SCHEDSTATS */
2359 * fork()/clone()-time setup:
2361 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2363 unsigned long flags
;
2364 int cpu
= get_cpu();
2366 __sched_fork(clone_flags
, p
);
2368 * We mark the process as NEW here. This guarantees that
2369 * nobody will actually run it, and a signal or other external
2370 * event cannot wake it up and insert it on the runqueue either.
2372 p
->state
= TASK_NEW
;
2375 * Make sure we do not leak PI boosting priority to the child.
2377 p
->prio
= current
->normal_prio
;
2380 * Revert to default priority/policy on fork if requested.
2382 if (unlikely(p
->sched_reset_on_fork
)) {
2383 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2384 p
->policy
= SCHED_NORMAL
;
2385 p
->static_prio
= NICE_TO_PRIO(0);
2387 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2388 p
->static_prio
= NICE_TO_PRIO(0);
2390 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2394 * We don't need the reset flag anymore after the fork. It has
2395 * fulfilled its duty:
2397 p
->sched_reset_on_fork
= 0;
2400 if (dl_prio(p
->prio
)) {
2403 } else if (rt_prio(p
->prio
)) {
2404 p
->sched_class
= &rt_sched_class
;
2406 p
->sched_class
= &fair_sched_class
;
2409 init_entity_runnable_average(&p
->se
);
2412 * The child is not yet in the pid-hash so no cgroup attach races,
2413 * and the cgroup is pinned to this child due to cgroup_fork()
2414 * is ran before sched_fork().
2416 * Silence PROVE_RCU.
2418 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2420 * We're setting the cpu for the first time, we don't migrate,
2421 * so use __set_task_cpu().
2423 __set_task_cpu(p
, cpu
);
2424 if (p
->sched_class
->task_fork
)
2425 p
->sched_class
->task_fork(p
);
2426 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2428 #ifdef CONFIG_SCHED_INFO
2429 if (likely(sched_info_on()))
2430 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2432 #if defined(CONFIG_SMP)
2435 init_task_preempt_count(p
);
2437 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2438 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2445 unsigned long to_ratio(u64 period
, u64 runtime
)
2447 if (runtime
== RUNTIME_INF
)
2451 * Doing this here saves a lot of checks in all
2452 * the calling paths, and returning zero seems
2453 * safe for them anyway.
2458 return div64_u64(runtime
<< 20, period
);
2462 inline struct dl_bw
*dl_bw_of(int i
)
2464 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2465 "sched RCU must be held");
2466 return &cpu_rq(i
)->rd
->dl_bw
;
2469 static inline int dl_bw_cpus(int i
)
2471 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2474 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2475 "sched RCU must be held");
2476 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2482 inline struct dl_bw
*dl_bw_of(int i
)
2484 return &cpu_rq(i
)->dl
.dl_bw
;
2487 static inline int dl_bw_cpus(int i
)
2494 * We must be sure that accepting a new task (or allowing changing the
2495 * parameters of an existing one) is consistent with the bandwidth
2496 * constraints. If yes, this function also accordingly updates the currently
2497 * allocated bandwidth to reflect the new situation.
2499 * This function is called while holding p's rq->lock.
2501 * XXX we should delay bw change until the task's 0-lag point, see
2504 static int dl_overflow(struct task_struct
*p
, int policy
,
2505 const struct sched_attr
*attr
)
2508 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2509 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2510 u64 runtime
= attr
->sched_runtime
;
2511 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2514 /* !deadline task may carry old deadline bandwidth */
2515 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2519 * Either if a task, enters, leave, or stays -deadline but changes
2520 * its parameters, we may need to update accordingly the total
2521 * allocated bandwidth of the container.
2523 raw_spin_lock(&dl_b
->lock
);
2524 cpus
= dl_bw_cpus(task_cpu(p
));
2525 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2526 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2527 __dl_add(dl_b
, new_bw
);
2529 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2530 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2531 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2532 __dl_add(dl_b
, new_bw
);
2534 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2535 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2538 raw_spin_unlock(&dl_b
->lock
);
2543 extern void init_dl_bw(struct dl_bw
*dl_b
);
2546 * wake_up_new_task - wake up a newly created task for the first time.
2548 * This function will do some initial scheduler statistics housekeeping
2549 * that must be done for every newly created context, then puts the task
2550 * on the runqueue and wakes it.
2552 void wake_up_new_task(struct task_struct
*p
)
2557 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2558 p
->state
= TASK_RUNNING
;
2561 * Fork balancing, do it here and not earlier because:
2562 * - cpus_allowed can change in the fork path
2563 * - any previously selected cpu might disappear through hotplug
2565 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2566 * as we're not fully set-up yet.
2568 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2570 rq
= __task_rq_lock(p
, &rf
);
2571 post_init_entity_util_avg(&p
->se
);
2573 activate_task(rq
, p
, 0);
2574 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2575 trace_sched_wakeup_new(p
);
2576 check_preempt_curr(rq
, p
, WF_FORK
);
2578 if (p
->sched_class
->task_woken
) {
2580 * Nothing relies on rq->lock after this, so its fine to
2583 lockdep_unpin_lock(&rq
->lock
, rf
.cookie
);
2584 p
->sched_class
->task_woken(rq
, p
);
2585 lockdep_repin_lock(&rq
->lock
, rf
.cookie
);
2588 task_rq_unlock(rq
, p
, &rf
);
2591 #ifdef CONFIG_PREEMPT_NOTIFIERS
2593 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2595 void preempt_notifier_inc(void)
2597 static_key_slow_inc(&preempt_notifier_key
);
2599 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2601 void preempt_notifier_dec(void)
2603 static_key_slow_dec(&preempt_notifier_key
);
2605 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2608 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2609 * @notifier: notifier struct to register
2611 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2613 if (!static_key_false(&preempt_notifier_key
))
2614 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2616 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2618 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2621 * preempt_notifier_unregister - no longer interested in preemption notifications
2622 * @notifier: notifier struct to unregister
2624 * This is *not* safe to call from within a preemption notifier.
2626 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2628 hlist_del(¬ifier
->link
);
2630 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2632 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2634 struct preempt_notifier
*notifier
;
2636 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2637 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2640 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2642 if (static_key_false(&preempt_notifier_key
))
2643 __fire_sched_in_preempt_notifiers(curr
);
2647 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2648 struct task_struct
*next
)
2650 struct preempt_notifier
*notifier
;
2652 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2653 notifier
->ops
->sched_out(notifier
, next
);
2656 static __always_inline
void
2657 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2658 struct task_struct
*next
)
2660 if (static_key_false(&preempt_notifier_key
))
2661 __fire_sched_out_preempt_notifiers(curr
, next
);
2664 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2666 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2671 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2672 struct task_struct
*next
)
2676 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2679 * prepare_task_switch - prepare to switch tasks
2680 * @rq: the runqueue preparing to switch
2681 * @prev: the current task that is being switched out
2682 * @next: the task we are going to switch to.
2684 * This is called with the rq lock held and interrupts off. It must
2685 * be paired with a subsequent finish_task_switch after the context
2688 * prepare_task_switch sets up locking and calls architecture specific
2692 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2693 struct task_struct
*next
)
2695 sched_info_switch(rq
, prev
, next
);
2696 perf_event_task_sched_out(prev
, next
);
2697 rseq_sched_out(prev
);
2698 fire_sched_out_preempt_notifiers(prev
, next
);
2699 prepare_lock_switch(rq
, next
);
2700 prepare_arch_switch(next
);
2704 * finish_task_switch - clean up after a task-switch
2705 * @prev: the thread we just switched away from.
2707 * finish_task_switch must be called after the context switch, paired
2708 * with a prepare_task_switch call before the context switch.
2709 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2710 * and do any other architecture-specific cleanup actions.
2712 * Note that we may have delayed dropping an mm in context_switch(). If
2713 * so, we finish that here outside of the runqueue lock. (Doing it
2714 * with the lock held can cause deadlocks; see schedule() for
2717 * The context switch have flipped the stack from under us and restored the
2718 * local variables which were saved when this task called schedule() in the
2719 * past. prev == current is still correct but we need to recalculate this_rq
2720 * because prev may have moved to another CPU.
2722 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2723 __releases(rq
->lock
)
2725 struct rq
*rq
= this_rq();
2726 struct mm_struct
*mm
= rq
->prev_mm
;
2730 * The previous task will have left us with a preempt_count of 2
2731 * because it left us after:
2734 * preempt_disable(); // 1
2736 * raw_spin_lock_irq(&rq->lock) // 2
2738 * Also, see FORK_PREEMPT_COUNT.
2740 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2741 "corrupted preempt_count: %s/%d/0x%x\n",
2742 current
->comm
, current
->pid
, preempt_count()))
2743 preempt_count_set(FORK_PREEMPT_COUNT
);
2748 * A task struct has one reference for the use as "current".
2749 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2750 * schedule one last time. The schedule call will never return, and
2751 * the scheduled task must drop that reference.
2753 * We must observe prev->state before clearing prev->on_cpu (in
2754 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2755 * running on another CPU and we could rave with its RUNNING -> DEAD
2756 * transition, resulting in a double drop.
2758 prev_state
= prev
->state
;
2759 vtime_task_switch(prev
);
2760 perf_event_task_sched_in(prev
, current
);
2761 finish_lock_switch(rq
, prev
);
2762 finish_arch_post_lock_switch();
2764 fire_sched_in_preempt_notifiers(current
);
2767 if (unlikely(prev_state
== TASK_DEAD
)) {
2768 if (prev
->sched_class
->task_dead
)
2769 prev
->sched_class
->task_dead(prev
);
2772 * Remove function-return probe instances associated with this
2773 * task and put them back on the free list.
2775 kprobe_flush_task(prev
);
2776 put_task_struct(prev
);
2779 tick_nohz_task_switch();
2785 /* rq->lock is NOT held, but preemption is disabled */
2786 static void __balance_callback(struct rq
*rq
)
2788 struct callback_head
*head
, *next
;
2789 void (*func
)(struct rq
*rq
);
2790 unsigned long flags
;
2792 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2793 head
= rq
->balance_callback
;
2794 rq
->balance_callback
= NULL
;
2796 func
= (void (*)(struct rq
*))head
->func
;
2803 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2806 static inline void balance_callback(struct rq
*rq
)
2808 if (unlikely(rq
->balance_callback
))
2809 __balance_callback(rq
);
2814 static inline void balance_callback(struct rq
*rq
)
2821 * schedule_tail - first thing a freshly forked thread must call.
2822 * @prev: the thread we just switched away from.
2824 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2825 __releases(rq
->lock
)
2830 * New tasks start with FORK_PREEMPT_COUNT, see there and
2831 * finish_task_switch() for details.
2833 * finish_task_switch() will drop rq->lock() and lower preempt_count
2834 * and the preempt_enable() will end up enabling preemption (on
2835 * PREEMPT_COUNT kernels).
2838 rq
= finish_task_switch(prev
);
2839 balance_callback(rq
);
2842 if (current
->set_child_tid
)
2843 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2847 * context_switch - switch to the new MM and the new thread's register state.
2849 static __always_inline
struct rq
*
2850 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2851 struct task_struct
*next
, struct pin_cookie cookie
)
2853 struct mm_struct
*mm
, *oldmm
;
2855 prepare_task_switch(rq
, prev
, next
);
2858 oldmm
= prev
->active_mm
;
2860 * For paravirt, this is coupled with an exit in switch_to to
2861 * combine the page table reload and the switch backend into
2864 arch_start_context_switch(prev
);
2867 next
->active_mm
= oldmm
;
2868 atomic_inc(&oldmm
->mm_count
);
2869 enter_lazy_tlb(oldmm
, next
);
2871 switch_mm_irqs_off(oldmm
, mm
, next
);
2874 prev
->active_mm
= NULL
;
2875 rq
->prev_mm
= oldmm
;
2878 * Since the runqueue lock will be released by the next
2879 * task (which is an invalid locking op but in the case
2880 * of the scheduler it's an obvious special-case), so we
2881 * do an early lockdep release here:
2883 lockdep_unpin_lock(&rq
->lock
, cookie
);
2884 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2886 /* Here we just switch the register state and the stack. */
2887 switch_to(prev
, next
, prev
);
2890 return finish_task_switch(prev
);
2894 * nr_running and nr_context_switches:
2896 * externally visible scheduler statistics: current number of runnable
2897 * threads, total number of context switches performed since bootup.
2899 unsigned long nr_running(void)
2901 unsigned long i
, sum
= 0;
2903 for_each_online_cpu(i
)
2904 sum
+= cpu_rq(i
)->nr_running
;
2910 * Check if only the current task is running on the cpu.
2912 * Caution: this function does not check that the caller has disabled
2913 * preemption, thus the result might have a time-of-check-to-time-of-use
2914 * race. The caller is responsible to use it correctly, for example:
2916 * - from a non-preemptable section (of course)
2918 * - from a thread that is bound to a single CPU
2920 * - in a loop with very short iterations (e.g. a polling loop)
2922 bool single_task_running(void)
2924 return raw_rq()->nr_running
== 1;
2926 EXPORT_SYMBOL(single_task_running
);
2928 unsigned long long nr_context_switches(void)
2931 unsigned long long sum
= 0;
2933 for_each_possible_cpu(i
)
2934 sum
+= cpu_rq(i
)->nr_switches
;
2939 unsigned long nr_iowait(void)
2941 unsigned long i
, sum
= 0;
2943 for_each_possible_cpu(i
)
2944 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2949 unsigned long nr_iowait_cpu(int cpu
)
2951 struct rq
*this = cpu_rq(cpu
);
2952 return atomic_read(&this->nr_iowait
);
2955 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2957 struct rq
*rq
= this_rq();
2958 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2959 *load
= rq
->load
.weight
;
2965 * sched_exec - execve() is a valuable balancing opportunity, because at
2966 * this point the task has the smallest effective memory and cache footprint.
2968 void sched_exec(void)
2970 struct task_struct
*p
= current
;
2971 unsigned long flags
;
2974 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2975 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2976 if (dest_cpu
== smp_processor_id())
2979 if (likely(cpu_active(dest_cpu
))) {
2980 struct migration_arg arg
= { p
, dest_cpu
};
2982 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2983 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
2987 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2992 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
2993 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
2995 EXPORT_PER_CPU_SYMBOL(kstat
);
2996 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
2999 * The function fair_sched_class.update_curr accesses the struct curr
3000 * and its field curr->exec_start; when called from task_sched_runtime(),
3001 * we observe a high rate of cache misses in practice.
3002 * Prefetching this data results in improved performance.
3004 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3006 #ifdef CONFIG_FAIR_GROUP_SCHED
3007 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3009 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3012 prefetch(&curr
->exec_start
);
3016 * Return accounted runtime for the task.
3017 * In case the task is currently running, return the runtime plus current's
3018 * pending runtime that have not been accounted yet.
3020 unsigned long long task_sched_runtime(struct task_struct
*p
)
3026 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3028 * 64-bit doesn't need locks to atomically read a 64bit value.
3029 * So we have a optimization chance when the task's delta_exec is 0.
3030 * Reading ->on_cpu is racy, but this is ok.
3032 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3033 * If we race with it entering cpu, unaccounted time is 0. This is
3034 * indistinguishable from the read occurring a few cycles earlier.
3035 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3036 * been accounted, so we're correct here as well.
3038 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3039 return p
->se
.sum_exec_runtime
;
3042 rq
= task_rq_lock(p
, &rf
);
3044 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3045 * project cycles that may never be accounted to this
3046 * thread, breaking clock_gettime().
3048 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3049 prefetch_curr_exec_start(p
);
3050 update_rq_clock(rq
);
3051 p
->sched_class
->update_curr(rq
);
3053 ns
= p
->se
.sum_exec_runtime
;
3054 task_rq_unlock(rq
, p
, &rf
);
3060 * This function gets called by the timer code, with HZ frequency.
3061 * We call it with interrupts disabled.
3063 void scheduler_tick(void)
3065 int cpu
= smp_processor_id();
3066 struct rq
*rq
= cpu_rq(cpu
);
3067 struct task_struct
*curr
= rq
->curr
;
3071 raw_spin_lock(&rq
->lock
);
3072 update_rq_clock(rq
);
3073 curr
->sched_class
->task_tick(rq
, curr
, 0);
3074 cpu_load_update_active(rq
);
3075 calc_global_load_tick(rq
);
3076 raw_spin_unlock(&rq
->lock
);
3078 perf_event_task_tick();
3081 rq
->idle_balance
= idle_cpu(cpu
);
3082 trigger_load_balance(rq
);
3084 rq_last_tick_reset(rq
);
3087 #ifdef CONFIG_NO_HZ_FULL
3089 * scheduler_tick_max_deferment
3091 * Keep at least one tick per second when a single
3092 * active task is running because the scheduler doesn't
3093 * yet completely support full dynticks environment.
3095 * This makes sure that uptime, CFS vruntime, load
3096 * balancing, etc... continue to move forward, even
3097 * with a very low granularity.
3099 * Return: Maximum deferment in nanoseconds.
3101 u64
scheduler_tick_max_deferment(void)
3103 struct rq
*rq
= this_rq();
3104 unsigned long next
, now
= READ_ONCE(jiffies
);
3106 next
= rq
->last_sched_tick
+ HZ
;
3108 if (time_before_eq(next
, now
))
3111 return jiffies_to_nsecs(next
- now
);
3115 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3116 defined(CONFIG_PREEMPT_TRACER))
3118 * If the value passed in is equal to the current preempt count
3119 * then we just disabled preemption. Start timing the latency.
3121 static inline void preempt_latency_start(int val
)
3123 if (preempt_count() == val
) {
3124 unsigned long ip
= get_lock_parent_ip();
3125 #ifdef CONFIG_DEBUG_PREEMPT
3126 current
->preempt_disable_ip
= ip
;
3128 trace_preempt_off(CALLER_ADDR0
, ip
);
3132 void preempt_count_add(int val
)
3134 #ifdef CONFIG_DEBUG_PREEMPT
3138 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3141 __preempt_count_add(val
);
3142 #ifdef CONFIG_DEBUG_PREEMPT
3144 * Spinlock count overflowing soon?
3146 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3149 preempt_latency_start(val
);
3151 EXPORT_SYMBOL(preempt_count_add
);
3152 NOKPROBE_SYMBOL(preempt_count_add
);
3155 * If the value passed in equals to the current preempt count
3156 * then we just enabled preemption. Stop timing the latency.
3158 static inline void preempt_latency_stop(int val
)
3160 if (preempt_count() == val
)
3161 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3164 void preempt_count_sub(int val
)
3166 #ifdef CONFIG_DEBUG_PREEMPT
3170 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3173 * Is the spinlock portion underflowing?
3175 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3176 !(preempt_count() & PREEMPT_MASK
)))
3180 preempt_latency_stop(val
);
3181 __preempt_count_sub(val
);
3183 EXPORT_SYMBOL(preempt_count_sub
);
3184 NOKPROBE_SYMBOL(preempt_count_sub
);
3187 static inline void preempt_latency_start(int val
) { }
3188 static inline void preempt_latency_stop(int val
) { }
3192 * Print scheduling while atomic bug:
3194 static noinline
void __schedule_bug(struct task_struct
*prev
)
3196 if (oops_in_progress
)
3199 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3200 prev
->comm
, prev
->pid
, preempt_count());
3202 debug_show_held_locks(prev
);
3204 if (irqs_disabled())
3205 print_irqtrace_events(prev
);
3206 #ifdef CONFIG_DEBUG_PREEMPT
3207 if (in_atomic_preempt_off()) {
3208 pr_err("Preemption disabled at:");
3209 print_ip_sym(current
->preempt_disable_ip
);
3214 panic("scheduling while atomic\n");
3217 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3221 * Various schedule()-time debugging checks and statistics:
3223 static inline void schedule_debug(struct task_struct
*prev
)
3225 #ifdef CONFIG_SCHED_STACK_END_CHECK
3226 if (task_stack_end_corrupted(prev
))
3227 panic("corrupted stack end detected inside scheduler\n");
3230 if (unlikely(in_atomic_preempt_off())) {
3231 __schedule_bug(prev
);
3232 preempt_count_set(PREEMPT_DISABLED
);
3236 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3238 schedstat_inc(this_rq(), sched_count
);
3242 * Pick up the highest-prio task:
3244 static inline struct task_struct
*
3245 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct pin_cookie cookie
)
3247 const struct sched_class
*class = &fair_sched_class
;
3248 struct task_struct
*p
;
3251 * Optimization: we know that if all tasks are in
3252 * the fair class we can call that function directly:
3254 if (likely(prev
->sched_class
== class &&
3255 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3256 p
= fair_sched_class
.pick_next_task(rq
, prev
, cookie
);
3257 if (unlikely(p
== RETRY_TASK
))
3260 /* assumes fair_sched_class->next == idle_sched_class */
3262 p
= idle_sched_class
.pick_next_task(rq
, prev
, cookie
);
3268 for_each_class(class) {
3269 p
= class->pick_next_task(rq
, prev
, cookie
);
3271 if (unlikely(p
== RETRY_TASK
))
3277 BUG(); /* the idle class will always have a runnable task */
3281 * __schedule() is the main scheduler function.
3283 * The main means of driving the scheduler and thus entering this function are:
3285 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3287 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3288 * paths. For example, see arch/x86/entry_64.S.
3290 * To drive preemption between tasks, the scheduler sets the flag in timer
3291 * interrupt handler scheduler_tick().
3293 * 3. Wakeups don't really cause entry into schedule(). They add a
3294 * task to the run-queue and that's it.
3296 * Now, if the new task added to the run-queue preempts the current
3297 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3298 * called on the nearest possible occasion:
3300 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3302 * - in syscall or exception context, at the next outmost
3303 * preempt_enable(). (this might be as soon as the wake_up()'s
3306 * - in IRQ context, return from interrupt-handler to
3307 * preemptible context
3309 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3312 * - cond_resched() call
3313 * - explicit schedule() call
3314 * - return from syscall or exception to user-space
3315 * - return from interrupt-handler to user-space
3317 * WARNING: must be called with preemption disabled!
3319 static void __sched notrace
__schedule(bool preempt
)
3321 struct task_struct
*prev
, *next
;
3322 unsigned long *switch_count
;
3323 struct pin_cookie cookie
;
3327 cpu
= smp_processor_id();
3332 * do_exit() calls schedule() with preemption disabled as an exception;
3333 * however we must fix that up, otherwise the next task will see an
3334 * inconsistent (higher) preempt count.
3336 * It also avoids the below schedule_debug() test from complaining
3339 if (unlikely(prev
->state
== TASK_DEAD
))
3340 preempt_enable_no_resched_notrace();
3342 schedule_debug(prev
);
3344 if (sched_feat(HRTICK
))
3347 local_irq_disable();
3348 rcu_note_context_switch();
3351 * Make sure that signal_pending_state()->signal_pending() below
3352 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3353 * done by the caller to avoid the race with signal_wake_up().
3355 smp_mb__before_spinlock();
3356 raw_spin_lock(&rq
->lock
);
3357 cookie
= lockdep_pin_lock(&rq
->lock
);
3359 rq
->clock_skip_update
<<= 1; /* promote REQ to ACT */
3361 switch_count
= &prev
->nivcsw
;
3362 if (!preempt
&& prev
->state
) {
3363 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3364 prev
->state
= TASK_RUNNING
;
3366 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3370 * If a worker went to sleep, notify and ask workqueue
3371 * whether it wants to wake up a task to maintain
3374 if (prev
->flags
& PF_WQ_WORKER
) {
3375 struct task_struct
*to_wakeup
;
3377 to_wakeup
= wq_worker_sleeping(prev
);
3379 try_to_wake_up_local(to_wakeup
, cookie
);
3382 switch_count
= &prev
->nvcsw
;
3385 if (task_on_rq_queued(prev
))
3386 update_rq_clock(rq
);
3388 next
= pick_next_task(rq
, prev
, cookie
);
3389 clear_tsk_need_resched(prev
);
3390 clear_preempt_need_resched();
3391 rq
->clock_skip_update
= 0;
3393 if (likely(prev
!= next
)) {
3398 trace_sched_switch(preempt
, prev
, next
);
3399 rq
= context_switch(rq
, prev
, next
, cookie
); /* unlocks the rq */
3401 lockdep_unpin_lock(&rq
->lock
, cookie
);
3402 raw_spin_unlock_irq(&rq
->lock
);
3405 balance_callback(rq
);
3407 STACK_FRAME_NON_STANDARD(__schedule
); /* switch_to() */
3409 static inline void sched_submit_work(struct task_struct
*tsk
)
3411 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3414 * If we are going to sleep and we have plugged IO queued,
3415 * make sure to submit it to avoid deadlocks.
3417 if (blk_needs_flush_plug(tsk
))
3418 blk_schedule_flush_plug(tsk
);
3421 asmlinkage __visible
void __sched
schedule(void)
3423 struct task_struct
*tsk
= current
;
3425 sched_submit_work(tsk
);
3429 sched_preempt_enable_no_resched();
3430 } while (need_resched());
3432 EXPORT_SYMBOL(schedule
);
3434 #ifdef CONFIG_CONTEXT_TRACKING
3435 asmlinkage __visible
void __sched
schedule_user(void)
3438 * If we come here after a random call to set_need_resched(),
3439 * or we have been woken up remotely but the IPI has not yet arrived,
3440 * we haven't yet exited the RCU idle mode. Do it here manually until
3441 * we find a better solution.
3443 * NB: There are buggy callers of this function. Ideally we
3444 * should warn if prev_state != CONTEXT_USER, but that will trigger
3445 * too frequently to make sense yet.
3447 enum ctx_state prev_state
= exception_enter();
3449 exception_exit(prev_state
);
3454 * schedule_preempt_disabled - called with preemption disabled
3456 * Returns with preemption disabled. Note: preempt_count must be 1
3458 void __sched
schedule_preempt_disabled(void)
3460 sched_preempt_enable_no_resched();
3465 static void __sched notrace
preempt_schedule_common(void)
3469 * Because the function tracer can trace preempt_count_sub()
3470 * and it also uses preempt_enable/disable_notrace(), if
3471 * NEED_RESCHED is set, the preempt_enable_notrace() called
3472 * by the function tracer will call this function again and
3473 * cause infinite recursion.
3475 * Preemption must be disabled here before the function
3476 * tracer can trace. Break up preempt_disable() into two
3477 * calls. One to disable preemption without fear of being
3478 * traced. The other to still record the preemption latency,
3479 * which can also be traced by the function tracer.
3481 preempt_disable_notrace();
3482 preempt_latency_start(1);
3484 preempt_latency_stop(1);
3485 preempt_enable_no_resched_notrace();
3488 * Check again in case we missed a preemption opportunity
3489 * between schedule and now.
3491 } while (need_resched());
3494 #ifdef CONFIG_PREEMPT
3496 * this is the entry point to schedule() from in-kernel preemption
3497 * off of preempt_enable. Kernel preemptions off return from interrupt
3498 * occur there and call schedule directly.
3500 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3503 * If there is a non-zero preempt_count or interrupts are disabled,
3504 * we do not want to preempt the current task. Just return..
3506 if (likely(!preemptible()))
3509 preempt_schedule_common();
3511 NOKPROBE_SYMBOL(preempt_schedule
);
3512 EXPORT_SYMBOL(preempt_schedule
);
3515 * preempt_schedule_notrace - preempt_schedule called by tracing
3517 * The tracing infrastructure uses preempt_enable_notrace to prevent
3518 * recursion and tracing preempt enabling caused by the tracing
3519 * infrastructure itself. But as tracing can happen in areas coming
3520 * from userspace or just about to enter userspace, a preempt enable
3521 * can occur before user_exit() is called. This will cause the scheduler
3522 * to be called when the system is still in usermode.
3524 * To prevent this, the preempt_enable_notrace will use this function
3525 * instead of preempt_schedule() to exit user context if needed before
3526 * calling the scheduler.
3528 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3530 enum ctx_state prev_ctx
;
3532 if (likely(!preemptible()))
3537 * Because the function tracer can trace preempt_count_sub()
3538 * and it also uses preempt_enable/disable_notrace(), if
3539 * NEED_RESCHED is set, the preempt_enable_notrace() called
3540 * by the function tracer will call this function again and
3541 * cause infinite recursion.
3543 * Preemption must be disabled here before the function
3544 * tracer can trace. Break up preempt_disable() into two
3545 * calls. One to disable preemption without fear of being
3546 * traced. The other to still record the preemption latency,
3547 * which can also be traced by the function tracer.
3549 preempt_disable_notrace();
3550 preempt_latency_start(1);
3552 * Needs preempt disabled in case user_exit() is traced
3553 * and the tracer calls preempt_enable_notrace() causing
3554 * an infinite recursion.
3556 prev_ctx
= exception_enter();
3558 exception_exit(prev_ctx
);
3560 preempt_latency_stop(1);
3561 preempt_enable_no_resched_notrace();
3562 } while (need_resched());
3564 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3566 #endif /* CONFIG_PREEMPT */
3569 * this is the entry point to schedule() from kernel preemption
3570 * off of irq context.
3571 * Note, that this is called and return with irqs disabled. This will
3572 * protect us against recursive calling from irq.
3574 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3576 enum ctx_state prev_state
;
3578 /* Catch callers which need to be fixed */
3579 BUG_ON(preempt_count() || !irqs_disabled());
3581 prev_state
= exception_enter();
3587 local_irq_disable();
3588 sched_preempt_enable_no_resched();
3589 } while (need_resched());
3591 exception_exit(prev_state
);
3594 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3597 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3599 EXPORT_SYMBOL(default_wake_function
);
3601 #ifdef CONFIG_RT_MUTEXES
3604 * rt_mutex_setprio - set the current priority of a task
3606 * @prio: prio value (kernel-internal form)
3608 * This function changes the 'effective' priority of a task. It does
3609 * not touch ->normal_prio like __setscheduler().
3611 * Used by the rt_mutex code to implement priority inheritance
3612 * logic. Call site only calls if the priority of the task changed.
3614 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3616 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3617 const struct sched_class
*prev_class
;
3621 BUG_ON(prio
> MAX_PRIO
);
3623 rq
= __task_rq_lock(p
, &rf
);
3626 * Idle task boosting is a nono in general. There is one
3627 * exception, when PREEMPT_RT and NOHZ is active:
3629 * The idle task calls get_next_timer_interrupt() and holds
3630 * the timer wheel base->lock on the CPU and another CPU wants
3631 * to access the timer (probably to cancel it). We can safely
3632 * ignore the boosting request, as the idle CPU runs this code
3633 * with interrupts disabled and will complete the lock
3634 * protected section without being interrupted. So there is no
3635 * real need to boost.
3637 if (unlikely(p
== rq
->idle
)) {
3638 WARN_ON(p
!= rq
->curr
);
3639 WARN_ON(p
->pi_blocked_on
);
3643 trace_sched_pi_setprio(p
, prio
);
3646 if (oldprio
== prio
)
3647 queue_flag
&= ~DEQUEUE_MOVE
;
3649 prev_class
= p
->sched_class
;
3650 queued
= task_on_rq_queued(p
);
3651 running
= task_current(rq
, p
);
3653 dequeue_task(rq
, p
, queue_flag
);
3655 put_prev_task(rq
, p
);
3658 * Boosting condition are:
3659 * 1. -rt task is running and holds mutex A
3660 * --> -dl task blocks on mutex A
3662 * 2. -dl task is running and holds mutex A
3663 * --> -dl task blocks on mutex A and could preempt the
3666 if (dl_prio(prio
)) {
3667 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3668 if (!dl_prio(p
->normal_prio
) ||
3669 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3670 p
->dl
.dl_boosted
= 1;
3671 queue_flag
|= ENQUEUE_REPLENISH
;
3673 p
->dl
.dl_boosted
= 0;
3674 p
->sched_class
= &dl_sched_class
;
3675 } else if (rt_prio(prio
)) {
3676 if (dl_prio(oldprio
))
3677 p
->dl
.dl_boosted
= 0;
3679 queue_flag
|= ENQUEUE_HEAD
;
3680 p
->sched_class
= &rt_sched_class
;
3682 if (dl_prio(oldprio
))
3683 p
->dl
.dl_boosted
= 0;
3684 if (rt_prio(oldprio
))
3686 p
->sched_class
= &fair_sched_class
;
3692 p
->sched_class
->set_curr_task(rq
);
3694 enqueue_task(rq
, p
, queue_flag
);
3696 check_class_changed(rq
, p
, prev_class
, oldprio
);
3698 preempt_disable(); /* avoid rq from going away on us */
3699 __task_rq_unlock(rq
, &rf
);
3701 balance_callback(rq
);
3706 void set_user_nice(struct task_struct
*p
, long nice
)
3708 int old_prio
, delta
, queued
;
3712 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3715 * We have to be careful, if called from sys_setpriority(),
3716 * the task might be in the middle of scheduling on another CPU.
3718 rq
= task_rq_lock(p
, &rf
);
3720 * The RT priorities are set via sched_setscheduler(), but we still
3721 * allow the 'normal' nice value to be set - but as expected
3722 * it wont have any effect on scheduling until the task is
3723 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3725 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3726 p
->static_prio
= NICE_TO_PRIO(nice
);
3729 queued
= task_on_rq_queued(p
);
3731 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3733 p
->static_prio
= NICE_TO_PRIO(nice
);
3736 p
->prio
= effective_prio(p
);
3737 delta
= p
->prio
- old_prio
;
3740 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3742 * If the task increased its priority or is running and
3743 * lowered its priority, then reschedule its CPU:
3745 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3749 task_rq_unlock(rq
, p
, &rf
);
3751 EXPORT_SYMBOL(set_user_nice
);
3754 * can_nice - check if a task can reduce its nice value
3758 int can_nice(const struct task_struct
*p
, const int nice
)
3760 /* convert nice value [19,-20] to rlimit style value [1,40] */
3761 int nice_rlim
= nice_to_rlimit(nice
);
3763 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3764 capable(CAP_SYS_NICE
));
3767 #ifdef __ARCH_WANT_SYS_NICE
3770 * sys_nice - change the priority of the current process.
3771 * @increment: priority increment
3773 * sys_setpriority is a more generic, but much slower function that
3774 * does similar things.
3776 SYSCALL_DEFINE1(nice
, int, increment
)
3781 * Setpriority might change our priority at the same moment.
3782 * We don't have to worry. Conceptually one call occurs first
3783 * and we have a single winner.
3785 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3786 nice
= task_nice(current
) + increment
;
3788 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3789 if (increment
< 0 && !can_nice(current
, nice
))
3792 retval
= security_task_setnice(current
, nice
);
3796 set_user_nice(current
, nice
);
3803 * task_prio - return the priority value of a given task.
3804 * @p: the task in question.
3806 * Return: The priority value as seen by users in /proc.
3807 * RT tasks are offset by -200. Normal tasks are centered
3808 * around 0, value goes from -16 to +15.
3810 int task_prio(const struct task_struct
*p
)
3812 return p
->prio
- MAX_RT_PRIO
;
3816 * idle_cpu - is a given cpu idle currently?
3817 * @cpu: the processor in question.
3819 * Return: 1 if the CPU is currently idle. 0 otherwise.
3821 int idle_cpu(int cpu
)
3823 struct rq
*rq
= cpu_rq(cpu
);
3825 if (rq
->curr
!= rq
->idle
)
3832 if (!llist_empty(&rq
->wake_list
))
3840 * idle_task - return the idle task for a given cpu.
3841 * @cpu: the processor in question.
3843 * Return: The idle task for the cpu @cpu.
3845 struct task_struct
*idle_task(int cpu
)
3847 return cpu_rq(cpu
)->idle
;
3851 * find_process_by_pid - find a process with a matching PID value.
3852 * @pid: the pid in question.
3854 * The task of @pid, if found. %NULL otherwise.
3856 static struct task_struct
*find_process_by_pid(pid_t pid
)
3858 return pid
? find_task_by_vpid(pid
) : current
;
3862 * This function initializes the sched_dl_entity of a newly becoming
3863 * SCHED_DEADLINE task.
3865 * Only the static values are considered here, the actual runtime and the
3866 * absolute deadline will be properly calculated when the task is enqueued
3867 * for the first time with its new policy.
3870 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3872 struct sched_dl_entity
*dl_se
= &p
->dl
;
3874 dl_se
->dl_runtime
= attr
->sched_runtime
;
3875 dl_se
->dl_deadline
= attr
->sched_deadline
;
3876 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3877 dl_se
->flags
= attr
->sched_flags
;
3878 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3881 * Changing the parameters of a task is 'tricky' and we're not doing
3882 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3884 * What we SHOULD do is delay the bandwidth release until the 0-lag
3885 * point. This would include retaining the task_struct until that time
3886 * and change dl_overflow() to not immediately decrement the current
3889 * Instead we retain the current runtime/deadline and let the new
3890 * parameters take effect after the current reservation period lapses.
3891 * This is safe (albeit pessimistic) because the 0-lag point is always
3892 * before the current scheduling deadline.
3894 * We can still have temporary overloads because we do not delay the
3895 * change in bandwidth until that time; so admission control is
3896 * not on the safe side. It does however guarantee tasks will never
3897 * consume more than promised.
3902 * sched_setparam() passes in -1 for its policy, to let the functions
3903 * it calls know not to change it.
3905 #define SETPARAM_POLICY -1
3907 static void __setscheduler_params(struct task_struct
*p
,
3908 const struct sched_attr
*attr
)
3910 int policy
= attr
->sched_policy
;
3912 if (policy
== SETPARAM_POLICY
)
3917 if (dl_policy(policy
))
3918 __setparam_dl(p
, attr
);
3919 else if (fair_policy(policy
))
3920 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3923 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3924 * !rt_policy. Always setting this ensures that things like
3925 * getparam()/getattr() don't report silly values for !rt tasks.
3927 p
->rt_priority
= attr
->sched_priority
;
3928 p
->normal_prio
= normal_prio(p
);
3932 /* Actually do priority change: must hold pi & rq lock. */
3933 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3934 const struct sched_attr
*attr
, bool keep_boost
)
3936 __setscheduler_params(p
, attr
);
3939 * Keep a potential priority boosting if called from
3940 * sched_setscheduler().
3943 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
3945 p
->prio
= normal_prio(p
);
3947 if (dl_prio(p
->prio
))
3948 p
->sched_class
= &dl_sched_class
;
3949 else if (rt_prio(p
->prio
))
3950 p
->sched_class
= &rt_sched_class
;
3952 p
->sched_class
= &fair_sched_class
;
3956 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
3958 struct sched_dl_entity
*dl_se
= &p
->dl
;
3960 attr
->sched_priority
= p
->rt_priority
;
3961 attr
->sched_runtime
= dl_se
->dl_runtime
;
3962 attr
->sched_deadline
= dl_se
->dl_deadline
;
3963 attr
->sched_period
= dl_se
->dl_period
;
3964 attr
->sched_flags
= dl_se
->flags
;
3968 * This function validates the new parameters of a -deadline task.
3969 * We ask for the deadline not being zero, and greater or equal
3970 * than the runtime, as well as the period of being zero or
3971 * greater than deadline. Furthermore, we have to be sure that
3972 * user parameters are above the internal resolution of 1us (we
3973 * check sched_runtime only since it is always the smaller one) and
3974 * below 2^63 ns (we have to check both sched_deadline and
3975 * sched_period, as the latter can be zero).
3978 __checkparam_dl(const struct sched_attr
*attr
)
3981 if (attr
->sched_deadline
== 0)
3985 * Since we truncate DL_SCALE bits, make sure we're at least
3988 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
3992 * Since we use the MSB for wrap-around and sign issues, make
3993 * sure it's not set (mind that period can be equal to zero).
3995 if (attr
->sched_deadline
& (1ULL << 63) ||
3996 attr
->sched_period
& (1ULL << 63))
3999 /* runtime <= deadline <= period (if period != 0) */
4000 if ((attr
->sched_period
!= 0 &&
4001 attr
->sched_period
< attr
->sched_deadline
) ||
4002 attr
->sched_deadline
< attr
->sched_runtime
)
4009 * check the target process has a UID that matches the current process's
4011 static bool check_same_owner(struct task_struct
*p
)
4013 const struct cred
*cred
= current_cred(), *pcred
;
4017 pcred
= __task_cred(p
);
4018 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4019 uid_eq(cred
->euid
, pcred
->uid
));
4024 static bool dl_param_changed(struct task_struct
*p
,
4025 const struct sched_attr
*attr
)
4027 struct sched_dl_entity
*dl_se
= &p
->dl
;
4029 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
4030 dl_se
->dl_deadline
!= attr
->sched_deadline
||
4031 dl_se
->dl_period
!= attr
->sched_period
||
4032 dl_se
->flags
!= attr
->sched_flags
)
4038 static int __sched_setscheduler(struct task_struct
*p
,
4039 const struct sched_attr
*attr
,
4042 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4043 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4044 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4045 int new_effective_prio
, policy
= attr
->sched_policy
;
4046 const struct sched_class
*prev_class
;
4049 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
4052 /* may grab non-irq protected spin_locks */
4053 BUG_ON(in_interrupt());
4055 /* double check policy once rq lock held */
4057 reset_on_fork
= p
->sched_reset_on_fork
;
4058 policy
= oldpolicy
= p
->policy
;
4060 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4062 if (!valid_policy(policy
))
4066 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4070 * Valid priorities for SCHED_FIFO and SCHED_RR are
4071 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4072 * SCHED_BATCH and SCHED_IDLE is 0.
4074 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4075 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4077 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4078 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4082 * Allow unprivileged RT tasks to decrease priority:
4084 if (user
&& !capable(CAP_SYS_NICE
)) {
4085 if (fair_policy(policy
)) {
4086 if (attr
->sched_nice
< task_nice(p
) &&
4087 !can_nice(p
, attr
->sched_nice
))
4091 if (rt_policy(policy
)) {
4092 unsigned long rlim_rtprio
=
4093 task_rlimit(p
, RLIMIT_RTPRIO
);
4095 /* can't set/change the rt policy */
4096 if (policy
!= p
->policy
&& !rlim_rtprio
)
4099 /* can't increase priority */
4100 if (attr
->sched_priority
> p
->rt_priority
&&
4101 attr
->sched_priority
> rlim_rtprio
)
4106 * Can't set/change SCHED_DEADLINE policy at all for now
4107 * (safest behavior); in the future we would like to allow
4108 * unprivileged DL tasks to increase their relative deadline
4109 * or reduce their runtime (both ways reducing utilization)
4111 if (dl_policy(policy
))
4115 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4116 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4118 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4119 if (!can_nice(p
, task_nice(p
)))
4123 /* can't change other user's priorities */
4124 if (!check_same_owner(p
))
4127 /* Normal users shall not reset the sched_reset_on_fork flag */
4128 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4133 retval
= security_task_setscheduler(p
);
4139 * make sure no PI-waiters arrive (or leave) while we are
4140 * changing the priority of the task:
4142 * To be able to change p->policy safely, the appropriate
4143 * runqueue lock must be held.
4145 rq
= task_rq_lock(p
, &rf
);
4148 * Changing the policy of the stop threads its a very bad idea
4150 if (p
== rq
->stop
) {
4151 task_rq_unlock(rq
, p
, &rf
);
4156 * If not changing anything there's no need to proceed further,
4157 * but store a possible modification of reset_on_fork.
4159 if (unlikely(policy
== p
->policy
)) {
4160 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4162 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4164 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4167 p
->sched_reset_on_fork
= reset_on_fork
;
4168 task_rq_unlock(rq
, p
, &rf
);
4174 #ifdef CONFIG_RT_GROUP_SCHED
4176 * Do not allow realtime tasks into groups that have no runtime
4179 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4180 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4181 !task_group_is_autogroup(task_group(p
))) {
4182 task_rq_unlock(rq
, p
, &rf
);
4187 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4188 cpumask_t
*span
= rq
->rd
->span
;
4191 * Don't allow tasks with an affinity mask smaller than
4192 * the entire root_domain to become SCHED_DEADLINE. We
4193 * will also fail if there's no bandwidth available.
4195 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4196 rq
->rd
->dl_bw
.bw
== 0) {
4197 task_rq_unlock(rq
, p
, &rf
);
4204 /* recheck policy now with rq lock held */
4205 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4206 policy
= oldpolicy
= -1;
4207 task_rq_unlock(rq
, p
, &rf
);
4212 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4213 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4216 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4217 task_rq_unlock(rq
, p
, &rf
);
4221 p
->sched_reset_on_fork
= reset_on_fork
;
4226 * Take priority boosted tasks into account. If the new
4227 * effective priority is unchanged, we just store the new
4228 * normal parameters and do not touch the scheduler class and
4229 * the runqueue. This will be done when the task deboost
4232 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4233 if (new_effective_prio
== oldprio
)
4234 queue_flags
&= ~DEQUEUE_MOVE
;
4237 queued
= task_on_rq_queued(p
);
4238 running
= task_current(rq
, p
);
4240 dequeue_task(rq
, p
, queue_flags
);
4242 put_prev_task(rq
, p
);
4244 prev_class
= p
->sched_class
;
4245 __setscheduler(rq
, p
, attr
, pi
);
4248 p
->sched_class
->set_curr_task(rq
);
4251 * We enqueue to tail when the priority of a task is
4252 * increased (user space view).
4254 if (oldprio
< p
->prio
)
4255 queue_flags
|= ENQUEUE_HEAD
;
4257 enqueue_task(rq
, p
, queue_flags
);
4260 check_class_changed(rq
, p
, prev_class
, oldprio
);
4261 preempt_disable(); /* avoid rq from going away on us */
4262 task_rq_unlock(rq
, p
, &rf
);
4265 rt_mutex_adjust_pi(p
);
4268 * Run balance callbacks after we've adjusted the PI chain.
4270 balance_callback(rq
);
4276 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4277 const struct sched_param
*param
, bool check
)
4279 struct sched_attr attr
= {
4280 .sched_policy
= policy
,
4281 .sched_priority
= param
->sched_priority
,
4282 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4285 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4286 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4287 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4288 policy
&= ~SCHED_RESET_ON_FORK
;
4289 attr
.sched_policy
= policy
;
4292 return __sched_setscheduler(p
, &attr
, check
, true);
4295 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4296 * @p: the task in question.
4297 * @policy: new policy.
4298 * @param: structure containing the new RT priority.
4300 * Return: 0 on success. An error code otherwise.
4302 * NOTE that the task may be already dead.
4304 int sched_setscheduler(struct task_struct
*p
, int policy
,
4305 const struct sched_param
*param
)
4307 return _sched_setscheduler(p
, policy
, param
, true);
4309 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4311 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4313 return __sched_setscheduler(p
, attr
, true, true);
4315 EXPORT_SYMBOL_GPL(sched_setattr
);
4318 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4319 * @p: the task in question.
4320 * @policy: new policy.
4321 * @param: structure containing the new RT priority.
4323 * Just like sched_setscheduler, only don't bother checking if the
4324 * current context has permission. For example, this is needed in
4325 * stop_machine(): we create temporary high priority worker threads,
4326 * but our caller might not have that capability.
4328 * Return: 0 on success. An error code otherwise.
4330 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4331 const struct sched_param
*param
)
4333 return _sched_setscheduler(p
, policy
, param
, false);
4335 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4338 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4340 struct sched_param lparam
;
4341 struct task_struct
*p
;
4344 if (!param
|| pid
< 0)
4346 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4351 p
= find_process_by_pid(pid
);
4353 retval
= sched_setscheduler(p
, policy
, &lparam
);
4360 * Mimics kernel/events/core.c perf_copy_attr().
4362 static int sched_copy_attr(struct sched_attr __user
*uattr
,
4363 struct sched_attr
*attr
)
4368 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4372 * zero the full structure, so that a short copy will be nice.
4374 memset(attr
, 0, sizeof(*attr
));
4376 ret
= get_user(size
, &uattr
->size
);
4380 if (size
> PAGE_SIZE
) /* silly large */
4383 if (!size
) /* abi compat */
4384 size
= SCHED_ATTR_SIZE_VER0
;
4386 if (size
< SCHED_ATTR_SIZE_VER0
)
4390 * If we're handed a bigger struct than we know of,
4391 * ensure all the unknown bits are 0 - i.e. new
4392 * user-space does not rely on any kernel feature
4393 * extensions we dont know about yet.
4395 if (size
> sizeof(*attr
)) {
4396 unsigned char __user
*addr
;
4397 unsigned char __user
*end
;
4400 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4401 end
= (void __user
*)uattr
+ size
;
4403 for (; addr
< end
; addr
++) {
4404 ret
= get_user(val
, addr
);
4410 size
= sizeof(*attr
);
4413 ret
= copy_from_user(attr
, uattr
, size
);
4418 * XXX: do we want to be lenient like existing syscalls; or do we want
4419 * to be strict and return an error on out-of-bounds values?
4421 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4426 put_user(sizeof(*attr
), &uattr
->size
);
4431 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4432 * @pid: the pid in question.
4433 * @policy: new policy.
4434 * @param: structure containing the new RT priority.
4436 * Return: 0 on success. An error code otherwise.
4438 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4439 struct sched_param __user
*, param
)
4441 /* negative values for policy are not valid */
4445 return do_sched_setscheduler(pid
, policy
, param
);
4449 * sys_sched_setparam - set/change the RT priority of a thread
4450 * @pid: the pid in question.
4451 * @param: structure containing the new RT priority.
4453 * Return: 0 on success. An error code otherwise.
4455 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4457 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4461 * sys_sched_setattr - same as above, but with extended sched_attr
4462 * @pid: the pid in question.
4463 * @uattr: structure containing the extended parameters.
4464 * @flags: for future extension.
4466 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4467 unsigned int, flags
)
4469 struct sched_attr attr
;
4470 struct task_struct
*p
;
4473 if (!uattr
|| pid
< 0 || flags
)
4476 retval
= sched_copy_attr(uattr
, &attr
);
4480 if ((int)attr
.sched_policy
< 0)
4485 p
= find_process_by_pid(pid
);
4487 retval
= sched_setattr(p
, &attr
);
4494 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4495 * @pid: the pid in question.
4497 * Return: On success, the policy of the thread. Otherwise, a negative error
4500 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4502 struct task_struct
*p
;
4510 p
= find_process_by_pid(pid
);
4512 retval
= security_task_getscheduler(p
);
4515 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4522 * sys_sched_getparam - get the RT priority of a thread
4523 * @pid: the pid in question.
4524 * @param: structure containing the RT priority.
4526 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4529 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4531 struct sched_param lp
= { .sched_priority
= 0 };
4532 struct task_struct
*p
;
4535 if (!param
|| pid
< 0)
4539 p
= find_process_by_pid(pid
);
4544 retval
= security_task_getscheduler(p
);
4548 if (task_has_rt_policy(p
))
4549 lp
.sched_priority
= p
->rt_priority
;
4553 * This one might sleep, we cannot do it with a spinlock held ...
4555 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4564 static int sched_read_attr(struct sched_attr __user
*uattr
,
4565 struct sched_attr
*attr
,
4570 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4574 * If we're handed a smaller struct than we know of,
4575 * ensure all the unknown bits are 0 - i.e. old
4576 * user-space does not get uncomplete information.
4578 if (usize
< sizeof(*attr
)) {
4579 unsigned char *addr
;
4582 addr
= (void *)attr
+ usize
;
4583 end
= (void *)attr
+ sizeof(*attr
);
4585 for (; addr
< end
; addr
++) {
4593 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4601 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4602 * @pid: the pid in question.
4603 * @uattr: structure containing the extended parameters.
4604 * @size: sizeof(attr) for fwd/bwd comp.
4605 * @flags: for future extension.
4607 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4608 unsigned int, size
, unsigned int, flags
)
4610 struct sched_attr attr
= {
4611 .size
= sizeof(struct sched_attr
),
4613 struct task_struct
*p
;
4616 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4617 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4621 p
= find_process_by_pid(pid
);
4626 retval
= security_task_getscheduler(p
);
4630 attr
.sched_policy
= p
->policy
;
4631 if (p
->sched_reset_on_fork
)
4632 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4633 if (task_has_dl_policy(p
))
4634 __getparam_dl(p
, &attr
);
4635 else if (task_has_rt_policy(p
))
4636 attr
.sched_priority
= p
->rt_priority
;
4638 attr
.sched_nice
= task_nice(p
);
4642 retval
= sched_read_attr(uattr
, &attr
, size
);
4650 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4652 cpumask_var_t cpus_allowed
, new_mask
;
4653 struct task_struct
*p
;
4658 p
= find_process_by_pid(pid
);
4664 /* Prevent p going away */
4668 if (p
->flags
& PF_NO_SETAFFINITY
) {
4672 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4676 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4678 goto out_free_cpus_allowed
;
4681 if (!check_same_owner(p
)) {
4683 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4685 goto out_free_new_mask
;
4690 retval
= security_task_setscheduler(p
);
4692 goto out_free_new_mask
;
4695 cpuset_cpus_allowed(p
, cpus_allowed
);
4696 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4699 * Since bandwidth control happens on root_domain basis,
4700 * if admission test is enabled, we only admit -deadline
4701 * tasks allowed to run on all the CPUs in the task's
4705 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4707 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4710 goto out_free_new_mask
;
4716 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4719 cpuset_cpus_allowed(p
, cpus_allowed
);
4720 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4722 * We must have raced with a concurrent cpuset
4723 * update. Just reset the cpus_allowed to the
4724 * cpuset's cpus_allowed
4726 cpumask_copy(new_mask
, cpus_allowed
);
4731 free_cpumask_var(new_mask
);
4732 out_free_cpus_allowed
:
4733 free_cpumask_var(cpus_allowed
);
4739 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4740 struct cpumask
*new_mask
)
4742 if (len
< cpumask_size())
4743 cpumask_clear(new_mask
);
4744 else if (len
> cpumask_size())
4745 len
= cpumask_size();
4747 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4751 * sys_sched_setaffinity - set the cpu affinity of a process
4752 * @pid: pid of the process
4753 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4754 * @user_mask_ptr: user-space pointer to the new cpu mask
4756 * Return: 0 on success. An error code otherwise.
4758 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4759 unsigned long __user
*, user_mask_ptr
)
4761 cpumask_var_t new_mask
;
4764 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4767 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4769 retval
= sched_setaffinity(pid
, new_mask
);
4770 free_cpumask_var(new_mask
);
4774 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4776 struct task_struct
*p
;
4777 unsigned long flags
;
4783 p
= find_process_by_pid(pid
);
4787 retval
= security_task_getscheduler(p
);
4791 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4792 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4793 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4802 * sys_sched_getaffinity - get the cpu affinity of a process
4803 * @pid: pid of the process
4804 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4805 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4807 * Return: size of CPU mask copied to user_mask_ptr on success. An
4808 * error code otherwise.
4810 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4811 unsigned long __user
*, user_mask_ptr
)
4816 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4818 if (len
& (sizeof(unsigned long)-1))
4821 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4824 ret
= sched_getaffinity(pid
, mask
);
4826 size_t retlen
= min_t(size_t, len
, cpumask_size());
4828 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4833 free_cpumask_var(mask
);
4839 * sys_sched_yield - yield the current processor to other threads.
4841 * This function yields the current CPU to other tasks. If there are no
4842 * other threads running on this CPU then this function will return.
4846 SYSCALL_DEFINE0(sched_yield
)
4848 struct rq
*rq
= this_rq_lock();
4850 schedstat_inc(rq
, yld_count
);
4851 current
->sched_class
->yield_task(rq
);
4854 * Since we are going to call schedule() anyway, there's
4855 * no need to preempt or enable interrupts:
4857 __release(rq
->lock
);
4858 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4859 do_raw_spin_unlock(&rq
->lock
);
4860 sched_preempt_enable_no_resched();
4867 int __sched
_cond_resched(void)
4869 if (should_resched(0)) {
4870 preempt_schedule_common();
4875 EXPORT_SYMBOL(_cond_resched
);
4878 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4879 * call schedule, and on return reacquire the lock.
4881 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4882 * operations here to prevent schedule() from being called twice (once via
4883 * spin_unlock(), once by hand).
4885 int __cond_resched_lock(spinlock_t
*lock
)
4887 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4890 lockdep_assert_held(lock
);
4892 if (spin_needbreak(lock
) || resched
) {
4895 preempt_schedule_common();
4903 EXPORT_SYMBOL(__cond_resched_lock
);
4905 int __sched
__cond_resched_softirq(void)
4907 BUG_ON(!in_softirq());
4909 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4911 preempt_schedule_common();
4917 EXPORT_SYMBOL(__cond_resched_softirq
);
4920 * yield - yield the current processor to other threads.
4922 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4924 * The scheduler is at all times free to pick the calling task as the most
4925 * eligible task to run, if removing the yield() call from your code breaks
4926 * it, its already broken.
4928 * Typical broken usage is:
4933 * where one assumes that yield() will let 'the other' process run that will
4934 * make event true. If the current task is a SCHED_FIFO task that will never
4935 * happen. Never use yield() as a progress guarantee!!
4937 * If you want to use yield() to wait for something, use wait_event().
4938 * If you want to use yield() to be 'nice' for others, use cond_resched().
4939 * If you still want to use yield(), do not!
4941 void __sched
yield(void)
4943 set_current_state(TASK_RUNNING
);
4946 EXPORT_SYMBOL(yield
);
4949 * yield_to - yield the current processor to another thread in
4950 * your thread group, or accelerate that thread toward the
4951 * processor it's on.
4953 * @preempt: whether task preemption is allowed or not
4955 * It's the caller's job to ensure that the target task struct
4956 * can't go away on us before we can do any checks.
4959 * true (>0) if we indeed boosted the target task.
4960 * false (0) if we failed to boost the target.
4961 * -ESRCH if there's no task to yield to.
4963 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
4965 struct task_struct
*curr
= current
;
4966 struct rq
*rq
, *p_rq
;
4967 unsigned long flags
;
4970 local_irq_save(flags
);
4976 * If we're the only runnable task on the rq and target rq also
4977 * has only one task, there's absolutely no point in yielding.
4979 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
4984 double_rq_lock(rq
, p_rq
);
4985 if (task_rq(p
) != p_rq
) {
4986 double_rq_unlock(rq
, p_rq
);
4990 if (!curr
->sched_class
->yield_to_task
)
4993 if (curr
->sched_class
!= p
->sched_class
)
4996 if (task_running(p_rq
, p
) || p
->state
)
4999 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5001 schedstat_inc(rq
, yld_count
);
5003 * Make p's CPU reschedule; pick_next_entity takes care of
5006 if (preempt
&& rq
!= p_rq
)
5011 double_rq_unlock(rq
, p_rq
);
5013 local_irq_restore(flags
);
5020 EXPORT_SYMBOL_GPL(yield_to
);
5023 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5024 * that process accounting knows that this is a task in IO wait state.
5026 long __sched
io_schedule_timeout(long timeout
)
5028 int old_iowait
= current
->in_iowait
;
5032 current
->in_iowait
= 1;
5033 blk_schedule_flush_plug(current
);
5035 delayacct_blkio_start();
5037 atomic_inc(&rq
->nr_iowait
);
5038 ret
= schedule_timeout(timeout
);
5039 current
->in_iowait
= old_iowait
;
5040 atomic_dec(&rq
->nr_iowait
);
5041 delayacct_blkio_end();
5045 EXPORT_SYMBOL(io_schedule_timeout
);
5048 * sys_sched_get_priority_max - return maximum RT priority.
5049 * @policy: scheduling class.
5051 * Return: On success, this syscall returns the maximum
5052 * rt_priority that can be used by a given scheduling class.
5053 * On failure, a negative error code is returned.
5055 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5062 ret
= MAX_USER_RT_PRIO
-1;
5064 case SCHED_DEADLINE
:
5075 * sys_sched_get_priority_min - return minimum RT priority.
5076 * @policy: scheduling class.
5078 * Return: On success, this syscall returns the minimum
5079 * rt_priority that can be used by a given scheduling class.
5080 * On failure, a negative error code is returned.
5082 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5091 case SCHED_DEADLINE
:
5101 * sys_sched_rr_get_interval - return the default timeslice of a process.
5102 * @pid: pid of the process.
5103 * @interval: userspace pointer to the timeslice value.
5105 * this syscall writes the default timeslice value of a given process
5106 * into the user-space timespec buffer. A value of '0' means infinity.
5108 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5111 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5112 struct timespec __user
*, interval
)
5114 struct task_struct
*p
;
5115 unsigned int time_slice
;
5126 p
= find_process_by_pid(pid
);
5130 retval
= security_task_getscheduler(p
);
5134 rq
= task_rq_lock(p
, &rf
);
5136 if (p
->sched_class
->get_rr_interval
)
5137 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5138 task_rq_unlock(rq
, p
, &rf
);
5141 jiffies_to_timespec(time_slice
, &t
);
5142 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5150 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5152 void sched_show_task(struct task_struct
*p
)
5154 unsigned long free
= 0;
5156 unsigned long state
= p
->state
;
5159 state
= __ffs(state
) + 1;
5160 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5161 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5162 #if BITS_PER_LONG == 32
5163 if (state
== TASK_RUNNING
)
5164 printk(KERN_CONT
" running ");
5166 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5168 if (state
== TASK_RUNNING
)
5169 printk(KERN_CONT
" running task ");
5171 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5173 #ifdef CONFIG_DEBUG_STACK_USAGE
5174 free
= stack_not_used(p
);
5179 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5181 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5182 task_pid_nr(p
), ppid
,
5183 (unsigned long)task_thread_info(p
)->flags
);
5185 print_worker_info(KERN_INFO
, p
);
5186 show_stack(p
, NULL
);
5189 void show_state_filter(unsigned long state_filter
)
5191 struct task_struct
*g
, *p
;
5193 #if BITS_PER_LONG == 32
5195 " task PC stack pid father\n");
5198 " task PC stack pid father\n");
5201 for_each_process_thread(g
, p
) {
5203 * reset the NMI-timeout, listing all files on a slow
5204 * console might take a lot of time:
5205 * Also, reset softlockup watchdogs on all CPUs, because
5206 * another CPU might be blocked waiting for us to process
5209 touch_nmi_watchdog();
5210 touch_all_softlockup_watchdogs();
5211 if (!state_filter
|| (p
->state
& state_filter
))
5215 #ifdef CONFIG_SCHED_DEBUG
5217 sysrq_sched_debug_show();
5221 * Only show locks if all tasks are dumped:
5224 debug_show_all_locks();
5227 void init_idle_bootup_task(struct task_struct
*idle
)
5229 idle
->sched_class
= &idle_sched_class
;
5233 * init_idle - set up an idle thread for a given CPU
5234 * @idle: task in question
5235 * @cpu: cpu the idle task belongs to
5237 * NOTE: this function does not set the idle thread's NEED_RESCHED
5238 * flag, to make booting more robust.
5240 void init_idle(struct task_struct
*idle
, int cpu
)
5242 struct rq
*rq
= cpu_rq(cpu
);
5243 unsigned long flags
;
5245 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5246 raw_spin_lock(&rq
->lock
);
5248 __sched_fork(0, idle
);
5249 idle
->state
= TASK_RUNNING
;
5250 idle
->se
.exec_start
= sched_clock();
5252 kasan_unpoison_task_stack(idle
);
5256 * Its possible that init_idle() gets called multiple times on a task,
5257 * in that case do_set_cpus_allowed() will not do the right thing.
5259 * And since this is boot we can forgo the serialization.
5261 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5264 * We're having a chicken and egg problem, even though we are
5265 * holding rq->lock, the cpu isn't yet set to this cpu so the
5266 * lockdep check in task_group() will fail.
5268 * Similar case to sched_fork(). / Alternatively we could
5269 * use task_rq_lock() here and obtain the other rq->lock.
5274 __set_task_cpu(idle
, cpu
);
5277 rq
->curr
= rq
->idle
= idle
;
5278 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5282 raw_spin_unlock(&rq
->lock
);
5283 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5285 /* Set the preempt count _outside_ the spinlocks! */
5286 init_idle_preempt_count(idle
, cpu
);
5289 * The idle tasks have their own, simple scheduling class:
5291 idle
->sched_class
= &idle_sched_class
;
5292 ftrace_graph_init_idle_task(idle
, cpu
);
5293 vtime_init_idle(idle
, cpu
);
5295 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5299 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5300 const struct cpumask
*trial
)
5302 int ret
= 1, trial_cpus
;
5303 struct dl_bw
*cur_dl_b
;
5304 unsigned long flags
;
5306 if (!cpumask_weight(cur
))
5309 rcu_read_lock_sched();
5310 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5311 trial_cpus
= cpumask_weight(trial
);
5313 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5314 if (cur_dl_b
->bw
!= -1 &&
5315 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5317 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5318 rcu_read_unlock_sched();
5323 int task_can_attach(struct task_struct
*p
,
5324 const struct cpumask
*cs_cpus_allowed
)
5329 * Kthreads which disallow setaffinity shouldn't be moved
5330 * to a new cpuset; we don't want to change their cpu
5331 * affinity and isolating such threads by their set of
5332 * allowed nodes is unnecessary. Thus, cpusets are not
5333 * applicable for such threads. This prevents checking for
5334 * success of set_cpus_allowed_ptr() on all attached tasks
5335 * before cpus_allowed may be changed.
5337 if (p
->flags
& PF_NO_SETAFFINITY
) {
5343 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5345 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5350 unsigned long flags
;
5352 rcu_read_lock_sched();
5353 dl_b
= dl_bw_of(dest_cpu
);
5354 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5355 cpus
= dl_bw_cpus(dest_cpu
);
5356 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5361 * We reserve space for this task in the destination
5362 * root_domain, as we can't fail after this point.
5363 * We will free resources in the source root_domain
5364 * later on (see set_cpus_allowed_dl()).
5366 __dl_add(dl_b
, p
->dl
.dl_bw
);
5368 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5369 rcu_read_unlock_sched();
5379 static bool sched_smp_initialized __read_mostly
;
5381 #ifdef CONFIG_NUMA_BALANCING
5382 /* Migrate current task p to target_cpu */
5383 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5385 struct migration_arg arg
= { p
, target_cpu
};
5386 int curr_cpu
= task_cpu(p
);
5388 if (curr_cpu
== target_cpu
)
5391 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5394 /* TODO: This is not properly updating schedstats */
5396 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5397 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5401 * Requeue a task on a given node and accurately track the number of NUMA
5402 * tasks on the runqueues
5404 void sched_setnuma(struct task_struct
*p
, int nid
)
5406 bool queued
, running
;
5410 rq
= task_rq_lock(p
, &rf
);
5411 queued
= task_on_rq_queued(p
);
5412 running
= task_current(rq
, p
);
5415 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5417 put_prev_task(rq
, p
);
5419 p
->numa_preferred_nid
= nid
;
5422 p
->sched_class
->set_curr_task(rq
);
5424 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5425 task_rq_unlock(rq
, p
, &rf
);
5427 #endif /* CONFIG_NUMA_BALANCING */
5429 #ifdef CONFIG_HOTPLUG_CPU
5431 * Ensures that the idle task is using init_mm right before its cpu goes
5434 void idle_task_exit(void)
5436 struct mm_struct
*mm
= current
->active_mm
;
5438 BUG_ON(cpu_online(smp_processor_id()));
5440 if (mm
!= &init_mm
) {
5441 switch_mm_irqs_off(mm
, &init_mm
, current
);
5442 finish_arch_post_lock_switch();
5448 * Since this CPU is going 'away' for a while, fold any nr_active delta
5449 * we might have. Assumes we're called after migrate_tasks() so that the
5450 * nr_active count is stable. We need to take the teardown thread which
5451 * is calling this into account, so we hand in adjust = 1 to the load
5454 * Also see the comment "Global load-average calculations".
5456 static void calc_load_migrate(struct rq
*rq
)
5458 long delta
= calc_load_fold_active(rq
, 1);
5460 atomic_long_add(delta
, &calc_load_tasks
);
5463 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5467 static const struct sched_class fake_sched_class
= {
5468 .put_prev_task
= put_prev_task_fake
,
5471 static struct task_struct fake_task
= {
5473 * Avoid pull_{rt,dl}_task()
5475 .prio
= MAX_PRIO
+ 1,
5476 .sched_class
= &fake_sched_class
,
5480 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5481 * try_to_wake_up()->select_task_rq().
5483 * Called with rq->lock held even though we'er in stop_machine() and
5484 * there's no concurrency possible, we hold the required locks anyway
5485 * because of lock validation efforts.
5487 static void migrate_tasks(struct rq
*dead_rq
)
5489 struct rq
*rq
= dead_rq
;
5490 struct task_struct
*next
, *stop
= rq
->stop
;
5491 struct pin_cookie cookie
;
5495 * Fudge the rq selection such that the below task selection loop
5496 * doesn't get stuck on the currently eligible stop task.
5498 * We're currently inside stop_machine() and the rq is either stuck
5499 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5500 * either way we should never end up calling schedule() until we're
5506 * put_prev_task() and pick_next_task() sched
5507 * class method both need to have an up-to-date
5508 * value of rq->clock[_task]
5510 update_rq_clock(rq
);
5514 * There's this thread running, bail when that's the only
5517 if (rq
->nr_running
== 1)
5521 * pick_next_task assumes pinned rq->lock.
5523 cookie
= lockdep_pin_lock(&rq
->lock
);
5524 next
= pick_next_task(rq
, &fake_task
, cookie
);
5526 next
->sched_class
->put_prev_task(rq
, next
);
5529 * Rules for changing task_struct::cpus_allowed are holding
5530 * both pi_lock and rq->lock, such that holding either
5531 * stabilizes the mask.
5533 * Drop rq->lock is not quite as disastrous as it usually is
5534 * because !cpu_active at this point, which means load-balance
5535 * will not interfere. Also, stop-machine.
5537 lockdep_unpin_lock(&rq
->lock
, cookie
);
5538 raw_spin_unlock(&rq
->lock
);
5539 raw_spin_lock(&next
->pi_lock
);
5540 raw_spin_lock(&rq
->lock
);
5543 * Since we're inside stop-machine, _nothing_ should have
5544 * changed the task, WARN if weird stuff happened, because in
5545 * that case the above rq->lock drop is a fail too.
5547 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5548 raw_spin_unlock(&next
->pi_lock
);
5552 /* Find suitable destination for @next, with force if needed. */
5553 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5555 rq
= __migrate_task(rq
, next
, dest_cpu
);
5556 if (rq
!= dead_rq
) {
5557 raw_spin_unlock(&rq
->lock
);
5559 raw_spin_lock(&rq
->lock
);
5561 raw_spin_unlock(&next
->pi_lock
);
5566 #endif /* CONFIG_HOTPLUG_CPU */
5568 static void set_rq_online(struct rq
*rq
)
5571 const struct sched_class
*class;
5573 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5576 for_each_class(class) {
5577 if (class->rq_online
)
5578 class->rq_online(rq
);
5583 static void set_rq_offline(struct rq
*rq
)
5586 const struct sched_class
*class;
5588 for_each_class(class) {
5589 if (class->rq_offline
)
5590 class->rq_offline(rq
);
5593 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5598 static void set_cpu_rq_start_time(unsigned int cpu
)
5600 struct rq
*rq
= cpu_rq(cpu
);
5602 rq
->age_stamp
= sched_clock_cpu(cpu
);
5605 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
5607 #ifdef CONFIG_SCHED_DEBUG
5609 static __read_mostly
int sched_debug_enabled
;
5611 static int __init
sched_debug_setup(char *str
)
5613 sched_debug_enabled
= 1;
5617 early_param("sched_debug", sched_debug_setup
);
5619 static inline bool sched_debug(void)
5621 return sched_debug_enabled
;
5624 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5625 struct cpumask
*groupmask
)
5627 struct sched_group
*group
= sd
->groups
;
5629 cpumask_clear(groupmask
);
5631 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5633 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5634 printk("does not load-balance\n");
5636 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5641 printk(KERN_CONT
"span %*pbl level %s\n",
5642 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
5644 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5645 printk(KERN_ERR
"ERROR: domain->span does not contain "
5648 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5649 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5653 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5657 printk(KERN_ERR
"ERROR: group is NULL\n");
5661 if (!cpumask_weight(sched_group_cpus(group
))) {
5662 printk(KERN_CONT
"\n");
5663 printk(KERN_ERR
"ERROR: empty group\n");
5667 if (!(sd
->flags
& SD_OVERLAP
) &&
5668 cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5669 printk(KERN_CONT
"\n");
5670 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5674 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5676 printk(KERN_CONT
" %*pbl",
5677 cpumask_pr_args(sched_group_cpus(group
)));
5678 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
) {
5679 printk(KERN_CONT
" (cpu_capacity = %d)",
5680 group
->sgc
->capacity
);
5683 group
= group
->next
;
5684 } while (group
!= sd
->groups
);
5685 printk(KERN_CONT
"\n");
5687 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5688 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5691 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5692 printk(KERN_ERR
"ERROR: parent span is not a superset "
5693 "of domain->span\n");
5697 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5701 if (!sched_debug_enabled
)
5705 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5709 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5712 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
5720 #else /* !CONFIG_SCHED_DEBUG */
5721 # define sched_domain_debug(sd, cpu) do { } while (0)
5722 static inline bool sched_debug(void)
5726 #endif /* CONFIG_SCHED_DEBUG */
5728 static int sd_degenerate(struct sched_domain
*sd
)
5730 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5733 /* Following flags need at least 2 groups */
5734 if (sd
->flags
& (SD_LOAD_BALANCE
|
5735 SD_BALANCE_NEWIDLE
|
5738 SD_SHARE_CPUCAPACITY
|
5739 SD_SHARE_PKG_RESOURCES
|
5740 SD_SHARE_POWERDOMAIN
)) {
5741 if (sd
->groups
!= sd
->groups
->next
)
5745 /* Following flags don't use groups */
5746 if (sd
->flags
& (SD_WAKE_AFFINE
))
5753 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5755 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5757 if (sd_degenerate(parent
))
5760 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5763 /* Flags needing groups don't count if only 1 group in parent */
5764 if (parent
->groups
== parent
->groups
->next
) {
5765 pflags
&= ~(SD_LOAD_BALANCE
|
5766 SD_BALANCE_NEWIDLE
|
5769 SD_SHARE_CPUCAPACITY
|
5770 SD_SHARE_PKG_RESOURCES
|
5772 SD_SHARE_POWERDOMAIN
);
5773 if (nr_node_ids
== 1)
5774 pflags
&= ~SD_SERIALIZE
;
5776 if (~cflags
& pflags
)
5782 static void free_rootdomain(struct rcu_head
*rcu
)
5784 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
5786 cpupri_cleanup(&rd
->cpupri
);
5787 cpudl_cleanup(&rd
->cpudl
);
5788 free_cpumask_var(rd
->dlo_mask
);
5789 free_cpumask_var(rd
->rto_mask
);
5790 free_cpumask_var(rd
->online
);
5791 free_cpumask_var(rd
->span
);
5795 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
5797 struct root_domain
*old_rd
= NULL
;
5798 unsigned long flags
;
5800 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5805 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
5808 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
5811 * If we dont want to free the old_rd yet then
5812 * set old_rd to NULL to skip the freeing later
5815 if (!atomic_dec_and_test(&old_rd
->refcount
))
5819 atomic_inc(&rd
->refcount
);
5822 cpumask_set_cpu(rq
->cpu
, rd
->span
);
5823 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
5826 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5829 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
5832 static int init_rootdomain(struct root_domain
*rd
)
5834 memset(rd
, 0, sizeof(*rd
));
5836 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
5838 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
5840 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
5842 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
5845 init_dl_bw(&rd
->dl_bw
);
5846 if (cpudl_init(&rd
->cpudl
) != 0)
5849 if (cpupri_init(&rd
->cpupri
) != 0)
5854 free_cpumask_var(rd
->rto_mask
);
5856 free_cpumask_var(rd
->dlo_mask
);
5858 free_cpumask_var(rd
->online
);
5860 free_cpumask_var(rd
->span
);
5866 * By default the system creates a single root-domain with all cpus as
5867 * members (mimicking the global state we have today).
5869 struct root_domain def_root_domain
;
5871 static void init_defrootdomain(void)
5873 init_rootdomain(&def_root_domain
);
5875 atomic_set(&def_root_domain
.refcount
, 1);
5878 static struct root_domain
*alloc_rootdomain(void)
5880 struct root_domain
*rd
;
5882 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
5886 if (init_rootdomain(rd
) != 0) {
5894 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
5896 struct sched_group
*tmp
, *first
;
5905 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
5910 } while (sg
!= first
);
5913 static void free_sched_domain(struct rcu_head
*rcu
)
5915 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
5918 * If its an overlapping domain it has private groups, iterate and
5921 if (sd
->flags
& SD_OVERLAP
) {
5922 free_sched_groups(sd
->groups
, 1);
5923 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
5924 kfree(sd
->groups
->sgc
);
5930 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
5932 call_rcu(&sd
->rcu
, free_sched_domain
);
5935 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
5937 for (; sd
; sd
= sd
->parent
)
5938 destroy_sched_domain(sd
, cpu
);
5942 * Keep a special pointer to the highest sched_domain that has
5943 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5944 * allows us to avoid some pointer chasing select_idle_sibling().
5946 * Also keep a unique ID per domain (we use the first cpu number in
5947 * the cpumask of the domain), this allows us to quickly tell if
5948 * two cpus are in the same cache domain, see cpus_share_cache().
5950 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
5951 DEFINE_PER_CPU(int, sd_llc_size
);
5952 DEFINE_PER_CPU(int, sd_llc_id
);
5953 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
5954 DEFINE_PER_CPU(struct sched_domain
*, sd_busy
);
5955 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
5957 static void update_top_cache_domain(int cpu
)
5959 struct sched_domain
*sd
;
5960 struct sched_domain
*busy_sd
= NULL
;
5964 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
5966 id
= cpumask_first(sched_domain_span(sd
));
5967 size
= cpumask_weight(sched_domain_span(sd
));
5968 busy_sd
= sd
->parent
; /* sd_busy */
5970 rcu_assign_pointer(per_cpu(sd_busy
, cpu
), busy_sd
);
5972 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
5973 per_cpu(sd_llc_size
, cpu
) = size
;
5974 per_cpu(sd_llc_id
, cpu
) = id
;
5976 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
5977 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
5979 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
5980 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
5984 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5985 * hold the hotplug lock.
5988 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
5990 struct rq
*rq
= cpu_rq(cpu
);
5991 struct sched_domain
*tmp
;
5993 /* Remove the sched domains which do not contribute to scheduling. */
5994 for (tmp
= sd
; tmp
; ) {
5995 struct sched_domain
*parent
= tmp
->parent
;
5999 if (sd_parent_degenerate(tmp
, parent
)) {
6000 tmp
->parent
= parent
->parent
;
6002 parent
->parent
->child
= tmp
;
6004 * Transfer SD_PREFER_SIBLING down in case of a
6005 * degenerate parent; the spans match for this
6006 * so the property transfers.
6008 if (parent
->flags
& SD_PREFER_SIBLING
)
6009 tmp
->flags
|= SD_PREFER_SIBLING
;
6010 destroy_sched_domain(parent
, cpu
);
6015 if (sd
&& sd_degenerate(sd
)) {
6018 destroy_sched_domain(tmp
, cpu
);
6023 sched_domain_debug(sd
, cpu
);
6025 rq_attach_root(rq
, rd
);
6027 rcu_assign_pointer(rq
->sd
, sd
);
6028 destroy_sched_domains(tmp
, cpu
);
6030 update_top_cache_domain(cpu
);
6033 /* Setup the mask of cpus configured for isolated domains */
6034 static int __init
isolated_cpu_setup(char *str
)
6038 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6039 ret
= cpulist_parse(str
, cpu_isolated_map
);
6041 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
6046 __setup("isolcpus=", isolated_cpu_setup
);
6049 struct sched_domain
** __percpu sd
;
6050 struct root_domain
*rd
;
6061 * Build an iteration mask that can exclude certain CPUs from the upwards
6064 * Asymmetric node setups can result in situations where the domain tree is of
6065 * unequal depth, make sure to skip domains that already cover the entire
6068 * In that case build_sched_domains() will have terminated the iteration early
6069 * and our sibling sd spans will be empty. Domains should always include the
6070 * cpu they're built on, so check that.
6073 static void build_group_mask(struct sched_domain
*sd
, struct sched_group
*sg
)
6075 const struct cpumask
*span
= sched_domain_span(sd
);
6076 struct sd_data
*sdd
= sd
->private;
6077 struct sched_domain
*sibling
;
6080 for_each_cpu(i
, span
) {
6081 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6082 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6085 cpumask_set_cpu(i
, sched_group_mask(sg
));
6090 * Return the canonical balance cpu for this group, this is the first cpu
6091 * of this group that's also in the iteration mask.
6093 int group_balance_cpu(struct sched_group
*sg
)
6095 return cpumask_first_and(sched_group_cpus(sg
), sched_group_mask(sg
));
6099 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
6101 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
6102 const struct cpumask
*span
= sched_domain_span(sd
);
6103 struct cpumask
*covered
= sched_domains_tmpmask
;
6104 struct sd_data
*sdd
= sd
->private;
6105 struct sched_domain
*sibling
;
6108 cpumask_clear(covered
);
6110 for_each_cpu(i
, span
) {
6111 struct cpumask
*sg_span
;
6113 if (cpumask_test_cpu(i
, covered
))
6116 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
6118 /* See the comment near build_group_mask(). */
6119 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
6122 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6123 GFP_KERNEL
, cpu_to_node(cpu
));
6128 sg_span
= sched_group_cpus(sg
);
6130 cpumask_copy(sg_span
, sched_domain_span(sibling
->child
));
6132 cpumask_set_cpu(i
, sg_span
);
6134 cpumask_or(covered
, covered
, sg_span
);
6136 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, i
);
6137 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
6138 build_group_mask(sd
, sg
);
6141 * Initialize sgc->capacity such that even if we mess up the
6142 * domains and no possible iteration will get us here, we won't
6145 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
6148 * Make sure the first group of this domain contains the
6149 * canonical balance cpu. Otherwise the sched_domain iteration
6150 * breaks. See update_sg_lb_stats().
6152 if ((!groups
&& cpumask_test_cpu(cpu
, sg_span
)) ||
6153 group_balance_cpu(sg
) == cpu
)
6163 sd
->groups
= groups
;
6168 free_sched_groups(first
, 0);
6173 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
6175 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
6176 struct sched_domain
*child
= sd
->child
;
6179 cpu
= cpumask_first(sched_domain_span(child
));
6182 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
6183 (*sg
)->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
6184 atomic_set(&(*sg
)->sgc
->ref
, 1); /* for claim_allocations */
6191 * build_sched_groups will build a circular linked list of the groups
6192 * covered by the given span, and will set each group's ->cpumask correctly,
6193 * and ->cpu_capacity to 0.
6195 * Assumes the sched_domain tree is fully constructed
6198 build_sched_groups(struct sched_domain
*sd
, int cpu
)
6200 struct sched_group
*first
= NULL
, *last
= NULL
;
6201 struct sd_data
*sdd
= sd
->private;
6202 const struct cpumask
*span
= sched_domain_span(sd
);
6203 struct cpumask
*covered
;
6206 get_group(cpu
, sdd
, &sd
->groups
);
6207 atomic_inc(&sd
->groups
->ref
);
6209 if (cpu
!= cpumask_first(span
))
6212 lockdep_assert_held(&sched_domains_mutex
);
6213 covered
= sched_domains_tmpmask
;
6215 cpumask_clear(covered
);
6217 for_each_cpu(i
, span
) {
6218 struct sched_group
*sg
;
6221 if (cpumask_test_cpu(i
, covered
))
6224 group
= get_group(i
, sdd
, &sg
);
6225 cpumask_setall(sched_group_mask(sg
));
6227 for_each_cpu(j
, span
) {
6228 if (get_group(j
, sdd
, NULL
) != group
)
6231 cpumask_set_cpu(j
, covered
);
6232 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6247 * Initialize sched groups cpu_capacity.
6249 * cpu_capacity indicates the capacity of sched group, which is used while
6250 * distributing the load between different sched groups in a sched domain.
6251 * Typically cpu_capacity for all the groups in a sched domain will be same
6252 * unless there are asymmetries in the topology. If there are asymmetries,
6253 * group having more cpu_capacity will pickup more load compared to the
6254 * group having less cpu_capacity.
6256 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
6258 struct sched_group
*sg
= sd
->groups
;
6263 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
6265 } while (sg
!= sd
->groups
);
6267 if (cpu
!= group_balance_cpu(sg
))
6270 update_group_capacity(sd
, cpu
);
6271 atomic_set(&sg
->sgc
->nr_busy_cpus
, sg
->group_weight
);
6275 * Initializers for schedule domains
6276 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6279 static int default_relax_domain_level
= -1;
6280 int sched_domain_level_max
;
6282 static int __init
setup_relax_domain_level(char *str
)
6284 if (kstrtoint(str
, 0, &default_relax_domain_level
))
6285 pr_warn("Unable to set relax_domain_level\n");
6289 __setup("relax_domain_level=", setup_relax_domain_level
);
6291 static void set_domain_attribute(struct sched_domain
*sd
,
6292 struct sched_domain_attr
*attr
)
6296 if (!attr
|| attr
->relax_domain_level
< 0) {
6297 if (default_relax_domain_level
< 0)
6300 request
= default_relax_domain_level
;
6302 request
= attr
->relax_domain_level
;
6303 if (request
< sd
->level
) {
6304 /* turn off idle balance on this domain */
6305 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6307 /* turn on idle balance on this domain */
6308 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6312 static void __sdt_free(const struct cpumask
*cpu_map
);
6313 static int __sdt_alloc(const struct cpumask
*cpu_map
);
6315 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6316 const struct cpumask
*cpu_map
)
6320 if (!atomic_read(&d
->rd
->refcount
))
6321 free_rootdomain(&d
->rd
->rcu
); /* fall through */
6323 free_percpu(d
->sd
); /* fall through */
6325 __sdt_free(cpu_map
); /* fall through */
6331 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6332 const struct cpumask
*cpu_map
)
6334 memset(d
, 0, sizeof(*d
));
6336 if (__sdt_alloc(cpu_map
))
6337 return sa_sd_storage
;
6338 d
->sd
= alloc_percpu(struct sched_domain
*);
6340 return sa_sd_storage
;
6341 d
->rd
= alloc_rootdomain();
6344 return sa_rootdomain
;
6348 * NULL the sd_data elements we've used to build the sched_domain and
6349 * sched_group structure so that the subsequent __free_domain_allocs()
6350 * will not free the data we're using.
6352 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
6354 struct sd_data
*sdd
= sd
->private;
6356 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
6357 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
6359 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
6360 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
6362 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
6363 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
6367 static int sched_domains_numa_levels
;
6368 enum numa_topology_type sched_numa_topology_type
;
6369 static int *sched_domains_numa_distance
;
6370 int sched_max_numa_distance
;
6371 static struct cpumask
***sched_domains_numa_masks
;
6372 static int sched_domains_curr_level
;
6376 * SD_flags allowed in topology descriptions.
6378 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6379 * SD_SHARE_PKG_RESOURCES - describes shared caches
6380 * SD_NUMA - describes NUMA topologies
6381 * SD_SHARE_POWERDOMAIN - describes shared power domain
6384 * SD_ASYM_PACKING - describes SMT quirks
6386 #define TOPOLOGY_SD_FLAGS \
6387 (SD_SHARE_CPUCAPACITY | \
6388 SD_SHARE_PKG_RESOURCES | \
6391 SD_SHARE_POWERDOMAIN)
6393 static struct sched_domain
*
6394 sd_init(struct sched_domain_topology_level
*tl
, int cpu
)
6396 struct sched_domain
*sd
= *per_cpu_ptr(tl
->data
.sd
, cpu
);
6397 int sd_weight
, sd_flags
= 0;
6401 * Ugly hack to pass state to sd_numa_mask()...
6403 sched_domains_curr_level
= tl
->numa_level
;
6406 sd_weight
= cpumask_weight(tl
->mask(cpu
));
6409 sd_flags
= (*tl
->sd_flags
)();
6410 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
6411 "wrong sd_flags in topology description\n"))
6412 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
6414 *sd
= (struct sched_domain
){
6415 .min_interval
= sd_weight
,
6416 .max_interval
= 2*sd_weight
,
6418 .imbalance_pct
= 125,
6420 .cache_nice_tries
= 0,
6427 .flags
= 1*SD_LOAD_BALANCE
6428 | 1*SD_BALANCE_NEWIDLE
6433 | 0*SD_SHARE_CPUCAPACITY
6434 | 0*SD_SHARE_PKG_RESOURCES
6436 | 0*SD_PREFER_SIBLING
6441 .last_balance
= jiffies
,
6442 .balance_interval
= sd_weight
,
6444 .max_newidle_lb_cost
= 0,
6445 .next_decay_max_lb_cost
= jiffies
,
6446 #ifdef CONFIG_SCHED_DEBUG
6452 * Convert topological properties into behaviour.
6455 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
6456 sd
->flags
|= SD_PREFER_SIBLING
;
6457 sd
->imbalance_pct
= 110;
6458 sd
->smt_gain
= 1178; /* ~15% */
6460 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
6461 sd
->imbalance_pct
= 117;
6462 sd
->cache_nice_tries
= 1;
6466 } else if (sd
->flags
& SD_NUMA
) {
6467 sd
->cache_nice_tries
= 2;
6471 sd
->flags
|= SD_SERIALIZE
;
6472 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
6473 sd
->flags
&= ~(SD_BALANCE_EXEC
|
6480 sd
->flags
|= SD_PREFER_SIBLING
;
6481 sd
->cache_nice_tries
= 1;
6486 sd
->private = &tl
->data
;
6492 * Topology list, bottom-up.
6494 static struct sched_domain_topology_level default_topology
[] = {
6495 #ifdef CONFIG_SCHED_SMT
6496 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
6498 #ifdef CONFIG_SCHED_MC
6499 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
6501 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
6505 static struct sched_domain_topology_level
*sched_domain_topology
=
6508 #define for_each_sd_topology(tl) \
6509 for (tl = sched_domain_topology; tl->mask; tl++)
6511 void set_sched_topology(struct sched_domain_topology_level
*tl
)
6513 sched_domain_topology
= tl
;
6518 static const struct cpumask
*sd_numa_mask(int cpu
)
6520 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
6523 static void sched_numa_warn(const char *str
)
6525 static int done
= false;
6533 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
6535 for (i
= 0; i
< nr_node_ids
; i
++) {
6536 printk(KERN_WARNING
" ");
6537 for (j
= 0; j
< nr_node_ids
; j
++)
6538 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
6539 printk(KERN_CONT
"\n");
6541 printk(KERN_WARNING
"\n");
6544 bool find_numa_distance(int distance
)
6548 if (distance
== node_distance(0, 0))
6551 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6552 if (sched_domains_numa_distance
[i
] == distance
)
6560 * A system can have three types of NUMA topology:
6561 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6562 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6563 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6565 * The difference between a glueless mesh topology and a backplane
6566 * topology lies in whether communication between not directly
6567 * connected nodes goes through intermediary nodes (where programs
6568 * could run), or through backplane controllers. This affects
6569 * placement of programs.
6571 * The type of topology can be discerned with the following tests:
6572 * - If the maximum distance between any nodes is 1 hop, the system
6573 * is directly connected.
6574 * - If for two nodes A and B, located N > 1 hops away from each other,
6575 * there is an intermediary node C, which is < N hops away from both
6576 * nodes A and B, the system is a glueless mesh.
6578 static void init_numa_topology_type(void)
6582 n
= sched_max_numa_distance
;
6584 if (sched_domains_numa_levels
<= 1) {
6585 sched_numa_topology_type
= NUMA_DIRECT
;
6589 for_each_online_node(a
) {
6590 for_each_online_node(b
) {
6591 /* Find two nodes furthest removed from each other. */
6592 if (node_distance(a
, b
) < n
)
6595 /* Is there an intermediary node between a and b? */
6596 for_each_online_node(c
) {
6597 if (node_distance(a
, c
) < n
&&
6598 node_distance(b
, c
) < n
) {
6599 sched_numa_topology_type
=
6605 sched_numa_topology_type
= NUMA_BACKPLANE
;
6611 static void sched_init_numa(void)
6613 int next_distance
, curr_distance
= node_distance(0, 0);
6614 struct sched_domain_topology_level
*tl
;
6618 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
6619 if (!sched_domains_numa_distance
)
6623 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6624 * unique distances in the node_distance() table.
6626 * Assumes node_distance(0,j) includes all distances in
6627 * node_distance(i,j) in order to avoid cubic time.
6629 next_distance
= curr_distance
;
6630 for (i
= 0; i
< nr_node_ids
; i
++) {
6631 for (j
= 0; j
< nr_node_ids
; j
++) {
6632 for (k
= 0; k
< nr_node_ids
; k
++) {
6633 int distance
= node_distance(i
, k
);
6635 if (distance
> curr_distance
&&
6636 (distance
< next_distance
||
6637 next_distance
== curr_distance
))
6638 next_distance
= distance
;
6641 * While not a strong assumption it would be nice to know
6642 * about cases where if node A is connected to B, B is not
6643 * equally connected to A.
6645 if (sched_debug() && node_distance(k
, i
) != distance
)
6646 sched_numa_warn("Node-distance not symmetric");
6648 if (sched_debug() && i
&& !find_numa_distance(distance
))
6649 sched_numa_warn("Node-0 not representative");
6651 if (next_distance
!= curr_distance
) {
6652 sched_domains_numa_distance
[level
++] = next_distance
;
6653 sched_domains_numa_levels
= level
;
6654 curr_distance
= next_distance
;
6659 * In case of sched_debug() we verify the above assumption.
6669 * 'level' contains the number of unique distances, excluding the
6670 * identity distance node_distance(i,i).
6672 * The sched_domains_numa_distance[] array includes the actual distance
6677 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6678 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6679 * the array will contain less then 'level' members. This could be
6680 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6681 * in other functions.
6683 * We reset it to 'level' at the end of this function.
6685 sched_domains_numa_levels
= 0;
6687 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
6688 if (!sched_domains_numa_masks
)
6692 * Now for each level, construct a mask per node which contains all
6693 * cpus of nodes that are that many hops away from us.
6695 for (i
= 0; i
< level
; i
++) {
6696 sched_domains_numa_masks
[i
] =
6697 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
6698 if (!sched_domains_numa_masks
[i
])
6701 for (j
= 0; j
< nr_node_ids
; j
++) {
6702 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
6706 sched_domains_numa_masks
[i
][j
] = mask
;
6709 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
6712 cpumask_or(mask
, mask
, cpumask_of_node(k
));
6717 /* Compute default topology size */
6718 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
6720 tl
= kzalloc((i
+ level
+ 1) *
6721 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
6726 * Copy the default topology bits..
6728 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
6729 tl
[i
] = sched_domain_topology
[i
];
6732 * .. and append 'j' levels of NUMA goodness.
6734 for (j
= 0; j
< level
; i
++, j
++) {
6735 tl
[i
] = (struct sched_domain_topology_level
){
6736 .mask
= sd_numa_mask
,
6737 .sd_flags
= cpu_numa_flags
,
6738 .flags
= SDTL_OVERLAP
,
6744 sched_domain_topology
= tl
;
6746 sched_domains_numa_levels
= level
;
6747 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
6749 init_numa_topology_type();
6752 static void sched_domains_numa_masks_set(unsigned int cpu
)
6754 int node
= cpu_to_node(cpu
);
6757 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6758 for (j
= 0; j
< nr_node_ids
; j
++) {
6759 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
6760 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6765 static void sched_domains_numa_masks_clear(unsigned int cpu
)
6769 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
6770 for (j
= 0; j
< nr_node_ids
; j
++)
6771 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
6776 static inline void sched_init_numa(void) { }
6777 static void sched_domains_numa_masks_set(unsigned int cpu
) { }
6778 static void sched_domains_numa_masks_clear(unsigned int cpu
) { }
6779 #endif /* CONFIG_NUMA */
6781 static int __sdt_alloc(const struct cpumask
*cpu_map
)
6783 struct sched_domain_topology_level
*tl
;
6786 for_each_sd_topology(tl
) {
6787 struct sd_data
*sdd
= &tl
->data
;
6789 sdd
->sd
= alloc_percpu(struct sched_domain
*);
6793 sdd
->sg
= alloc_percpu(struct sched_group
*);
6797 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
6801 for_each_cpu(j
, cpu_map
) {
6802 struct sched_domain
*sd
;
6803 struct sched_group
*sg
;
6804 struct sched_group_capacity
*sgc
;
6806 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
6807 GFP_KERNEL
, cpu_to_node(j
));
6811 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
6813 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6814 GFP_KERNEL
, cpu_to_node(j
));
6820 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
6822 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
6823 GFP_KERNEL
, cpu_to_node(j
));
6827 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
6834 static void __sdt_free(const struct cpumask
*cpu_map
)
6836 struct sched_domain_topology_level
*tl
;
6839 for_each_sd_topology(tl
) {
6840 struct sd_data
*sdd
= &tl
->data
;
6842 for_each_cpu(j
, cpu_map
) {
6843 struct sched_domain
*sd
;
6846 sd
= *per_cpu_ptr(sdd
->sd
, j
);
6847 if (sd
&& (sd
->flags
& SD_OVERLAP
))
6848 free_sched_groups(sd
->groups
, 0);
6849 kfree(*per_cpu_ptr(sdd
->sd
, j
));
6853 kfree(*per_cpu_ptr(sdd
->sg
, j
));
6855 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
6857 free_percpu(sdd
->sd
);
6859 free_percpu(sdd
->sg
);
6861 free_percpu(sdd
->sgc
);
6866 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
6867 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6868 struct sched_domain
*child
, int cpu
)
6870 struct sched_domain
*sd
= sd_init(tl
, cpu
);
6874 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
6876 sd
->level
= child
->level
+ 1;
6877 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
6881 if (!cpumask_subset(sched_domain_span(child
),
6882 sched_domain_span(sd
))) {
6883 pr_err("BUG: arch topology borken\n");
6884 #ifdef CONFIG_SCHED_DEBUG
6885 pr_err(" the %s domain not a subset of the %s domain\n",
6886 child
->name
, sd
->name
);
6888 /* Fixup, ensure @sd has at least @child cpus. */
6889 cpumask_or(sched_domain_span(sd
),
6890 sched_domain_span(sd
),
6891 sched_domain_span(child
));
6895 set_domain_attribute(sd
, attr
);
6901 * Build sched domains for a given set of cpus and attach the sched domains
6902 * to the individual cpus
6904 static int build_sched_domains(const struct cpumask
*cpu_map
,
6905 struct sched_domain_attr
*attr
)
6907 enum s_alloc alloc_state
;
6908 struct sched_domain
*sd
;
6910 int i
, ret
= -ENOMEM
;
6912 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6913 if (alloc_state
!= sa_rootdomain
)
6916 /* Set up domains for cpus specified by the cpu_map. */
6917 for_each_cpu(i
, cpu_map
) {
6918 struct sched_domain_topology_level
*tl
;
6921 for_each_sd_topology(tl
) {
6922 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
6923 if (tl
== sched_domain_topology
)
6924 *per_cpu_ptr(d
.sd
, i
) = sd
;
6925 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
6926 sd
->flags
|= SD_OVERLAP
;
6927 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
6932 /* Build the groups for the domains */
6933 for_each_cpu(i
, cpu_map
) {
6934 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6935 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
6936 if (sd
->flags
& SD_OVERLAP
) {
6937 if (build_overlap_sched_groups(sd
, i
))
6940 if (build_sched_groups(sd
, i
))
6946 /* Calculate CPU capacity for physical packages and nodes */
6947 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
6948 if (!cpumask_test_cpu(i
, cpu_map
))
6951 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
6952 claim_allocations(i
, sd
);
6953 init_sched_groups_capacity(i
, sd
);
6957 /* Attach the domains */
6959 for_each_cpu(i
, cpu_map
) {
6960 sd
= *per_cpu_ptr(d
.sd
, i
);
6961 cpu_attach_domain(sd
, d
.rd
, i
);
6967 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6971 static cpumask_var_t
*doms_cur
; /* current sched domains */
6972 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6973 static struct sched_domain_attr
*dattr_cur
;
6974 /* attribues of custom domains in 'doms_cur' */
6977 * Special case: If a kmalloc of a doms_cur partition (array of
6978 * cpumask) fails, then fallback to a single sched domain,
6979 * as determined by the single cpumask fallback_doms.
6981 static cpumask_var_t fallback_doms
;
6984 * arch_update_cpu_topology lets virtualized architectures update the
6985 * cpu core maps. It is supposed to return 1 if the topology changed
6986 * or 0 if it stayed the same.
6988 int __weak
arch_update_cpu_topology(void)
6993 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
6996 cpumask_var_t
*doms
;
6998 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7001 for (i
= 0; i
< ndoms
; i
++) {
7002 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7003 free_sched_domains(doms
, i
);
7010 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7013 for (i
= 0; i
< ndoms
; i
++)
7014 free_cpumask_var(doms
[i
]);
7019 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7020 * For now this just excludes isolated cpus, but could be used to
7021 * exclude other special cases in the future.
7023 static int init_sched_domains(const struct cpumask
*cpu_map
)
7027 arch_update_cpu_topology();
7029 doms_cur
= alloc_sched_domains(ndoms_cur
);
7031 doms_cur
= &fallback_doms
;
7032 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7033 err
= build_sched_domains(doms_cur
[0], NULL
);
7034 register_sched_domain_sysctl();
7040 * Detach sched domains from a group of cpus specified in cpu_map
7041 * These cpus will now be attached to the NULL domain
7043 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7048 for_each_cpu(i
, cpu_map
)
7049 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7053 /* handle null as "default" */
7054 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7055 struct sched_domain_attr
*new, int idx_new
)
7057 struct sched_domain_attr tmp
;
7064 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7065 new ? (new + idx_new
) : &tmp
,
7066 sizeof(struct sched_domain_attr
));
7070 * Partition sched domains as specified by the 'ndoms_new'
7071 * cpumasks in the array doms_new[] of cpumasks. This compares
7072 * doms_new[] to the current sched domain partitioning, doms_cur[].
7073 * It destroys each deleted domain and builds each new domain.
7075 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7076 * The masks don't intersect (don't overlap.) We should setup one
7077 * sched domain for each mask. CPUs not in any of the cpumasks will
7078 * not be load balanced. If the same cpumask appears both in the
7079 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7082 * The passed in 'doms_new' should be allocated using
7083 * alloc_sched_domains. This routine takes ownership of it and will
7084 * free_sched_domains it when done with it. If the caller failed the
7085 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7086 * and partition_sched_domains() will fallback to the single partition
7087 * 'fallback_doms', it also forces the domains to be rebuilt.
7089 * If doms_new == NULL it will be replaced with cpu_online_mask.
7090 * ndoms_new == 0 is a special case for destroying existing domains,
7091 * and it will not create the default domain.
7093 * Call with hotplug lock held
7095 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7096 struct sched_domain_attr
*dattr_new
)
7101 mutex_lock(&sched_domains_mutex
);
7103 /* always unregister in case we don't destroy any domains */
7104 unregister_sched_domain_sysctl();
7106 /* Let architecture update cpu core mappings. */
7107 new_topology
= arch_update_cpu_topology();
7109 n
= doms_new
? ndoms_new
: 0;
7111 /* Destroy deleted domains */
7112 for (i
= 0; i
< ndoms_cur
; i
++) {
7113 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7114 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7115 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7118 /* no match - a current sched domain not in new doms_new[] */
7119 detach_destroy_domains(doms_cur
[i
]);
7125 if (doms_new
== NULL
) {
7127 doms_new
= &fallback_doms
;
7128 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7129 WARN_ON_ONCE(dattr_new
);
7132 /* Build new domains */
7133 for (i
= 0; i
< ndoms_new
; i
++) {
7134 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7135 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7136 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7139 /* no match - add a new doms_new */
7140 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7145 /* Remember the new sched domains */
7146 if (doms_cur
!= &fallback_doms
)
7147 free_sched_domains(doms_cur
, ndoms_cur
);
7148 kfree(dattr_cur
); /* kfree(NULL) is safe */
7149 doms_cur
= doms_new
;
7150 dattr_cur
= dattr_new
;
7151 ndoms_cur
= ndoms_new
;
7153 register_sched_domain_sysctl();
7155 mutex_unlock(&sched_domains_mutex
);
7158 static int num_cpus_frozen
; /* used to mark begin/end of suspend/resume */
7161 * Update cpusets according to cpu_active mask. If cpusets are
7162 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7163 * around partition_sched_domains().
7165 * If we come here as part of a suspend/resume, don't touch cpusets because we
7166 * want to restore it back to its original state upon resume anyway.
7168 static void cpuset_cpu_active(void)
7170 if (cpuhp_tasks_frozen
) {
7172 * num_cpus_frozen tracks how many CPUs are involved in suspend
7173 * resume sequence. As long as this is not the last online
7174 * operation in the resume sequence, just build a single sched
7175 * domain, ignoring cpusets.
7178 if (likely(num_cpus_frozen
)) {
7179 partition_sched_domains(1, NULL
, NULL
);
7183 * This is the last CPU online operation. So fall through and
7184 * restore the original sched domains by considering the
7185 * cpuset configurations.
7188 cpuset_update_active_cpus(true);
7191 static int cpuset_cpu_inactive(unsigned int cpu
)
7193 unsigned long flags
;
7198 if (!cpuhp_tasks_frozen
) {
7199 rcu_read_lock_sched();
7200 dl_b
= dl_bw_of(cpu
);
7202 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
7203 cpus
= dl_bw_cpus(cpu
);
7204 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
7205 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
7207 rcu_read_unlock_sched();
7211 cpuset_update_active_cpus(false);
7214 partition_sched_domains(1, NULL
, NULL
);
7219 int sched_cpu_activate(unsigned int cpu
)
7221 struct rq
*rq
= cpu_rq(cpu
);
7222 unsigned long flags
;
7224 set_cpu_active(cpu
, true);
7226 if (sched_smp_initialized
) {
7227 sched_domains_numa_masks_set(cpu
);
7228 cpuset_cpu_active();
7232 * Put the rq online, if not already. This happens:
7234 * 1) In the early boot process, because we build the real domains
7235 * after all cpus have been brought up.
7237 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7240 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7242 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7245 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7247 update_max_interval();
7252 int sched_cpu_deactivate(unsigned int cpu
)
7256 set_cpu_active(cpu
, false);
7258 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7259 * users of this state to go away such that all new such users will
7262 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7263 * not imply sync_sched(), so wait for both.
7265 * Do sync before park smpboot threads to take care the rcu boost case.
7267 if (IS_ENABLED(CONFIG_PREEMPT
))
7268 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
7272 if (!sched_smp_initialized
)
7275 ret
= cpuset_cpu_inactive(cpu
);
7277 set_cpu_active(cpu
, true);
7280 sched_domains_numa_masks_clear(cpu
);
7284 static void sched_rq_cpu_starting(unsigned int cpu
)
7286 struct rq
*rq
= cpu_rq(cpu
);
7288 rq
->calc_load_update
= calc_load_update
;
7289 update_max_interval();
7292 int sched_cpu_starting(unsigned int cpu
)
7294 set_cpu_rq_start_time(cpu
);
7295 sched_rq_cpu_starting(cpu
);
7299 #ifdef CONFIG_HOTPLUG_CPU
7300 int sched_cpu_dying(unsigned int cpu
)
7302 struct rq
*rq
= cpu_rq(cpu
);
7303 unsigned long flags
;
7305 /* Handle pending wakeups and then migrate everything off */
7306 sched_ttwu_pending();
7307 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7309 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7313 BUG_ON(rq
->nr_running
!= 1);
7314 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7315 calc_load_migrate(rq
);
7316 update_max_interval();
7317 nohz_balance_exit_idle(cpu
);
7323 void __init
sched_init_smp(void)
7325 cpumask_var_t non_isolated_cpus
;
7327 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7328 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7333 * There's no userspace yet to cause hotplug operations; hence all the
7334 * cpu masks are stable and all blatant races in the below code cannot
7337 mutex_lock(&sched_domains_mutex
);
7338 init_sched_domains(cpu_active_mask
);
7339 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7340 if (cpumask_empty(non_isolated_cpus
))
7341 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7342 mutex_unlock(&sched_domains_mutex
);
7344 /* Move init over to a non-isolated CPU */
7345 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7347 sched_init_granularity();
7348 free_cpumask_var(non_isolated_cpus
);
7350 init_sched_rt_class();
7351 init_sched_dl_class();
7352 sched_smp_initialized
= true;
7355 static int __init
migration_init(void)
7357 sched_rq_cpu_starting(smp_processor_id());
7360 early_initcall(migration_init
);
7363 void __init
sched_init_smp(void)
7365 sched_init_granularity();
7367 #endif /* CONFIG_SMP */
7369 int in_sched_functions(unsigned long addr
)
7371 return in_lock_functions(addr
) ||
7372 (addr
>= (unsigned long)__sched_text_start
7373 && addr
< (unsigned long)__sched_text_end
);
7376 #ifdef CONFIG_CGROUP_SCHED
7378 * Default task group.
7379 * Every task in system belongs to this group at bootup.
7381 struct task_group root_task_group
;
7382 LIST_HEAD(task_groups
);
7384 /* Cacheline aligned slab cache for task_group */
7385 static struct kmem_cache
*task_group_cache __read_mostly
;
7388 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7390 void __init
sched_init(void)
7393 unsigned long alloc_size
= 0, ptr
;
7395 #ifdef CONFIG_FAIR_GROUP_SCHED
7396 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7398 #ifdef CONFIG_RT_GROUP_SCHED
7399 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7402 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7404 #ifdef CONFIG_FAIR_GROUP_SCHED
7405 root_task_group
.se
= (struct sched_entity
**)ptr
;
7406 ptr
+= nr_cpu_ids
* sizeof(void **);
7408 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7409 ptr
+= nr_cpu_ids
* sizeof(void **);
7411 #endif /* CONFIG_FAIR_GROUP_SCHED */
7412 #ifdef CONFIG_RT_GROUP_SCHED
7413 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7414 ptr
+= nr_cpu_ids
* sizeof(void **);
7416 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7417 ptr
+= nr_cpu_ids
* sizeof(void **);
7419 #endif /* CONFIG_RT_GROUP_SCHED */
7421 #ifdef CONFIG_CPUMASK_OFFSTACK
7422 for_each_possible_cpu(i
) {
7423 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7424 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7426 #endif /* CONFIG_CPUMASK_OFFSTACK */
7428 init_rt_bandwidth(&def_rt_bandwidth
,
7429 global_rt_period(), global_rt_runtime());
7430 init_dl_bandwidth(&def_dl_bandwidth
,
7431 global_rt_period(), global_rt_runtime());
7434 init_defrootdomain();
7437 #ifdef CONFIG_RT_GROUP_SCHED
7438 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7439 global_rt_period(), global_rt_runtime());
7440 #endif /* CONFIG_RT_GROUP_SCHED */
7442 #ifdef CONFIG_CGROUP_SCHED
7443 task_group_cache
= KMEM_CACHE(task_group
, 0);
7445 list_add(&root_task_group
.list
, &task_groups
);
7446 INIT_LIST_HEAD(&root_task_group
.children
);
7447 INIT_LIST_HEAD(&root_task_group
.siblings
);
7448 autogroup_init(&init_task
);
7449 #endif /* CONFIG_CGROUP_SCHED */
7451 for_each_possible_cpu(i
) {
7455 raw_spin_lock_init(&rq
->lock
);
7457 rq
->calc_load_active
= 0;
7458 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7459 init_cfs_rq(&rq
->cfs
);
7460 init_rt_rq(&rq
->rt
);
7461 init_dl_rq(&rq
->dl
);
7462 #ifdef CONFIG_FAIR_GROUP_SCHED
7463 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7464 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7466 * How much cpu bandwidth does root_task_group get?
7468 * In case of task-groups formed thr' the cgroup filesystem, it
7469 * gets 100% of the cpu resources in the system. This overall
7470 * system cpu resource is divided among the tasks of
7471 * root_task_group and its child task-groups in a fair manner,
7472 * based on each entity's (task or task-group's) weight
7473 * (se->load.weight).
7475 * In other words, if root_task_group has 10 tasks of weight
7476 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7477 * then A0's share of the cpu resource is:
7479 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7481 * We achieve this by letting root_task_group's tasks sit
7482 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7484 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7485 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7486 #endif /* CONFIG_FAIR_GROUP_SCHED */
7488 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7489 #ifdef CONFIG_RT_GROUP_SCHED
7490 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7493 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7494 rq
->cpu_load
[j
] = 0;
7499 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7500 rq
->balance_callback
= NULL
;
7501 rq
->active_balance
= 0;
7502 rq
->next_balance
= jiffies
;
7507 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7508 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7510 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7512 rq_attach_root(rq
, &def_root_domain
);
7513 #ifdef CONFIG_NO_HZ_COMMON
7514 rq
->last_load_update_tick
= jiffies
;
7517 #ifdef CONFIG_NO_HZ_FULL
7518 rq
->last_sched_tick
= 0;
7520 #endif /* CONFIG_SMP */
7522 atomic_set(&rq
->nr_iowait
, 0);
7525 set_load_weight(&init_task
);
7527 #ifdef CONFIG_PREEMPT_NOTIFIERS
7528 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7532 * The boot idle thread does lazy MMU switching as well:
7534 atomic_inc(&init_mm
.mm_count
);
7535 enter_lazy_tlb(&init_mm
, current
);
7538 * During early bootup we pretend to be a normal task:
7540 current
->sched_class
= &fair_sched_class
;
7543 * Make us the idle thread. Technically, schedule() should not be
7544 * called from this thread, however somewhere below it might be,
7545 * but because we are the idle thread, we just pick up running again
7546 * when this runqueue becomes "idle".
7548 init_idle(current
, smp_processor_id());
7550 calc_load_update
= jiffies
+ LOAD_FREQ
;
7553 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
7554 /* May be allocated at isolcpus cmdline parse time */
7555 if (cpu_isolated_map
== NULL
)
7556 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7557 idle_thread_set_boot_cpu();
7558 set_cpu_rq_start_time(smp_processor_id());
7560 init_sched_fair_class();
7564 scheduler_running
= 1;
7567 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7568 static inline int preempt_count_equals(int preempt_offset
)
7570 int nested
= preempt_count() + rcu_preempt_depth();
7572 return (nested
== preempt_offset
);
7575 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7578 * Blocking primitives will set (and therefore destroy) current->state,
7579 * since we will exit with TASK_RUNNING make sure we enter with it,
7580 * otherwise we will destroy state.
7582 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7583 "do not call blocking ops when !TASK_RUNNING; "
7584 "state=%lx set at [<%p>] %pS\n",
7586 (void *)current
->task_state_change
,
7587 (void *)current
->task_state_change
);
7589 ___might_sleep(file
, line
, preempt_offset
);
7591 EXPORT_SYMBOL(__might_sleep
);
7593 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7595 static unsigned long prev_jiffy
; /* ratelimiting */
7597 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7598 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7599 !is_idle_task(current
)) ||
7600 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7602 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7604 prev_jiffy
= jiffies
;
7607 "BUG: sleeping function called from invalid context at %s:%d\n",
7610 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7611 in_atomic(), irqs_disabled(),
7612 current
->pid
, current
->comm
);
7614 if (task_stack_end_corrupted(current
))
7615 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7617 debug_show_held_locks(current
);
7618 if (irqs_disabled())
7619 print_irqtrace_events(current
);
7620 #ifdef CONFIG_DEBUG_PREEMPT
7621 if (!preempt_count_equals(preempt_offset
)) {
7622 pr_err("Preemption disabled at:");
7623 print_ip_sym(current
->preempt_disable_ip
);
7629 EXPORT_SYMBOL(___might_sleep
);
7632 #ifdef CONFIG_MAGIC_SYSRQ
7633 void normalize_rt_tasks(void)
7635 struct task_struct
*g
, *p
;
7636 struct sched_attr attr
= {
7637 .sched_policy
= SCHED_NORMAL
,
7640 read_lock(&tasklist_lock
);
7641 for_each_process_thread(g
, p
) {
7643 * Only normalize user tasks:
7645 if (p
->flags
& PF_KTHREAD
)
7648 p
->se
.exec_start
= 0;
7649 #ifdef CONFIG_SCHEDSTATS
7650 p
->se
.statistics
.wait_start
= 0;
7651 p
->se
.statistics
.sleep_start
= 0;
7652 p
->se
.statistics
.block_start
= 0;
7655 if (!dl_task(p
) && !rt_task(p
)) {
7657 * Renice negative nice level userspace
7660 if (task_nice(p
) < 0)
7661 set_user_nice(p
, 0);
7665 __sched_setscheduler(p
, &attr
, false, false);
7667 read_unlock(&tasklist_lock
);
7670 #endif /* CONFIG_MAGIC_SYSRQ */
7672 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7674 * These functions are only useful for the IA64 MCA handling, or kdb.
7676 * They can only be called when the whole system has been
7677 * stopped - every CPU needs to be quiescent, and no scheduling
7678 * activity can take place. Using them for anything else would
7679 * be a serious bug, and as a result, they aren't even visible
7680 * under any other configuration.
7684 * curr_task - return the current task for a given cpu.
7685 * @cpu: the processor in question.
7687 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7689 * Return: The current task for @cpu.
7691 struct task_struct
*curr_task(int cpu
)
7693 return cpu_curr(cpu
);
7696 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7700 * set_curr_task - set the current task for a given cpu.
7701 * @cpu: the processor in question.
7702 * @p: the task pointer to set.
7704 * Description: This function must only be used when non-maskable interrupts
7705 * are serviced on a separate stack. It allows the architecture to switch the
7706 * notion of the current task on a cpu in a non-blocking manner. This function
7707 * must be called with all CPU's synchronized, and interrupts disabled, the
7708 * and caller must save the original value of the current task (see
7709 * curr_task() above) and restore that value before reenabling interrupts and
7710 * re-starting the system.
7712 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7714 void set_curr_task(int cpu
, struct task_struct
*p
)
7721 #ifdef CONFIG_CGROUP_SCHED
7722 /* task_group_lock serializes the addition/removal of task groups */
7723 static DEFINE_SPINLOCK(task_group_lock
);
7725 static void sched_free_group(struct task_group
*tg
)
7727 free_fair_sched_group(tg
);
7728 free_rt_sched_group(tg
);
7730 kmem_cache_free(task_group_cache
, tg
);
7733 /* allocate runqueue etc for a new task group */
7734 struct task_group
*sched_create_group(struct task_group
*parent
)
7736 struct task_group
*tg
;
7738 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7740 return ERR_PTR(-ENOMEM
);
7742 if (!alloc_fair_sched_group(tg
, parent
))
7745 if (!alloc_rt_sched_group(tg
, parent
))
7751 sched_free_group(tg
);
7752 return ERR_PTR(-ENOMEM
);
7755 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7757 unsigned long flags
;
7759 spin_lock_irqsave(&task_group_lock
, flags
);
7760 list_add_rcu(&tg
->list
, &task_groups
);
7762 WARN_ON(!parent
); /* root should already exist */
7764 tg
->parent
= parent
;
7765 INIT_LIST_HEAD(&tg
->children
);
7766 list_add_rcu(&tg
->siblings
, &parent
->children
);
7767 spin_unlock_irqrestore(&task_group_lock
, flags
);
7769 online_fair_sched_group(tg
);
7772 /* rcu callback to free various structures associated with a task group */
7773 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7775 /* now it should be safe to free those cfs_rqs */
7776 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7779 void sched_destroy_group(struct task_group
*tg
)
7781 /* wait for possible concurrent references to cfs_rqs complete */
7782 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7785 void sched_offline_group(struct task_group
*tg
)
7787 unsigned long flags
;
7789 /* end participation in shares distribution */
7790 unregister_fair_sched_group(tg
);
7792 spin_lock_irqsave(&task_group_lock
, flags
);
7793 list_del_rcu(&tg
->list
);
7794 list_del_rcu(&tg
->siblings
);
7795 spin_unlock_irqrestore(&task_group_lock
, flags
);
7798 static void sched_change_group(struct task_struct
*tsk
, int type
)
7800 struct task_group
*tg
;
7803 * All callers are synchronized by task_rq_lock(); we do not use RCU
7804 * which is pointless here. Thus, we pass "true" to task_css_check()
7805 * to prevent lockdep warnings.
7807 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7808 struct task_group
, css
);
7809 tg
= autogroup_task_group(tsk
, tg
);
7810 tsk
->sched_task_group
= tg
;
7812 #ifdef CONFIG_FAIR_GROUP_SCHED
7813 if (tsk
->sched_class
->task_change_group
)
7814 tsk
->sched_class
->task_change_group(tsk
, type
);
7817 set_task_rq(tsk
, task_cpu(tsk
));
7821 * Change task's runqueue when it moves between groups.
7823 * The caller of this function should have put the task in its new group by
7824 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7827 void sched_move_task(struct task_struct
*tsk
)
7829 int queued
, running
;
7833 rq
= task_rq_lock(tsk
, &rf
);
7835 running
= task_current(rq
, tsk
);
7836 queued
= task_on_rq_queued(tsk
);
7839 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
7840 if (unlikely(running
))
7841 put_prev_task(rq
, tsk
);
7843 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7845 if (unlikely(running
))
7846 tsk
->sched_class
->set_curr_task(rq
);
7848 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
7850 task_rq_unlock(rq
, tsk
, &rf
);
7852 #endif /* CONFIG_CGROUP_SCHED */
7854 #ifdef CONFIG_RT_GROUP_SCHED
7856 * Ensure that the real time constraints are schedulable.
7858 static DEFINE_MUTEX(rt_constraints_mutex
);
7860 /* Must be called with tasklist_lock held */
7861 static inline int tg_has_rt_tasks(struct task_group
*tg
)
7863 struct task_struct
*g
, *p
;
7866 * Autogroups do not have RT tasks; see autogroup_create().
7868 if (task_group_is_autogroup(tg
))
7871 for_each_process_thread(g
, p
) {
7872 if (rt_task(p
) && task_group(p
) == tg
)
7879 struct rt_schedulable_data
{
7880 struct task_group
*tg
;
7885 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
7887 struct rt_schedulable_data
*d
= data
;
7888 struct task_group
*child
;
7889 unsigned long total
, sum
= 0;
7890 u64 period
, runtime
;
7892 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
7893 runtime
= tg
->rt_bandwidth
.rt_runtime
;
7896 period
= d
->rt_period
;
7897 runtime
= d
->rt_runtime
;
7901 * Cannot have more runtime than the period.
7903 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
7907 * Ensure we don't starve existing RT tasks.
7909 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
7912 total
= to_ratio(period
, runtime
);
7915 * Nobody can have more than the global setting allows.
7917 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
7921 * The sum of our children's runtime should not exceed our own.
7923 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
7924 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
7925 runtime
= child
->rt_bandwidth
.rt_runtime
;
7927 if (child
== d
->tg
) {
7928 period
= d
->rt_period
;
7929 runtime
= d
->rt_runtime
;
7932 sum
+= to_ratio(period
, runtime
);
7941 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
7945 struct rt_schedulable_data data
= {
7947 .rt_period
= period
,
7948 .rt_runtime
= runtime
,
7952 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
7958 static int tg_set_rt_bandwidth(struct task_group
*tg
,
7959 u64 rt_period
, u64 rt_runtime
)
7964 * Disallowing the root group RT runtime is BAD, it would disallow the
7965 * kernel creating (and or operating) RT threads.
7967 if (tg
== &root_task_group
&& rt_runtime
== 0)
7970 /* No period doesn't make any sense. */
7974 mutex_lock(&rt_constraints_mutex
);
7975 read_lock(&tasklist_lock
);
7976 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
7980 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7981 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
7982 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
7984 for_each_possible_cpu(i
) {
7985 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
7987 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
7988 rt_rq
->rt_runtime
= rt_runtime
;
7989 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
7991 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
7993 read_unlock(&tasklist_lock
);
7994 mutex_unlock(&rt_constraints_mutex
);
7999 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8001 u64 rt_runtime
, rt_period
;
8003 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8004 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8005 if (rt_runtime_us
< 0)
8006 rt_runtime
= RUNTIME_INF
;
8008 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8011 static long sched_group_rt_runtime(struct task_group
*tg
)
8015 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8018 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8019 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8020 return rt_runtime_us
;
8023 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
8025 u64 rt_runtime
, rt_period
;
8027 rt_period
= rt_period_us
* NSEC_PER_USEC
;
8028 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8030 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8033 static long sched_group_rt_period(struct task_group
*tg
)
8037 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8038 do_div(rt_period_us
, NSEC_PER_USEC
);
8039 return rt_period_us
;
8041 #endif /* CONFIG_RT_GROUP_SCHED */
8043 #ifdef CONFIG_RT_GROUP_SCHED
8044 static int sched_rt_global_constraints(void)
8048 mutex_lock(&rt_constraints_mutex
);
8049 read_lock(&tasklist_lock
);
8050 ret
= __rt_schedulable(NULL
, 0, 0);
8051 read_unlock(&tasklist_lock
);
8052 mutex_unlock(&rt_constraints_mutex
);
8057 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8059 /* Don't accept realtime tasks when there is no way for them to run */
8060 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8066 #else /* !CONFIG_RT_GROUP_SCHED */
8067 static int sched_rt_global_constraints(void)
8069 unsigned long flags
;
8072 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8073 for_each_possible_cpu(i
) {
8074 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8076 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8077 rt_rq
->rt_runtime
= global_rt_runtime();
8078 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8080 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8084 #endif /* CONFIG_RT_GROUP_SCHED */
8086 static int sched_dl_global_validate(void)
8088 u64 runtime
= global_rt_runtime();
8089 u64 period
= global_rt_period();
8090 u64 new_bw
= to_ratio(period
, runtime
);
8093 unsigned long flags
;
8096 * Here we want to check the bandwidth not being set to some
8097 * value smaller than the currently allocated bandwidth in
8098 * any of the root_domains.
8100 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8101 * cycling on root_domains... Discussion on different/better
8102 * solutions is welcome!
8104 for_each_possible_cpu(cpu
) {
8105 rcu_read_lock_sched();
8106 dl_b
= dl_bw_of(cpu
);
8108 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8109 if (new_bw
< dl_b
->total_bw
)
8111 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8113 rcu_read_unlock_sched();
8122 static void sched_dl_do_global(void)
8127 unsigned long flags
;
8129 def_dl_bandwidth
.dl_period
= global_rt_period();
8130 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
8132 if (global_rt_runtime() != RUNTIME_INF
)
8133 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
8136 * FIXME: As above...
8138 for_each_possible_cpu(cpu
) {
8139 rcu_read_lock_sched();
8140 dl_b
= dl_bw_of(cpu
);
8142 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
8144 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
8146 rcu_read_unlock_sched();
8150 static int sched_rt_global_validate(void)
8152 if (sysctl_sched_rt_period
<= 0)
8155 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
8156 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
8162 static void sched_rt_do_global(void)
8164 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8165 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
8168 int sched_rt_handler(struct ctl_table
*table
, int write
,
8169 void __user
*buffer
, size_t *lenp
,
8172 int old_period
, old_runtime
;
8173 static DEFINE_MUTEX(mutex
);
8177 old_period
= sysctl_sched_rt_period
;
8178 old_runtime
= sysctl_sched_rt_runtime
;
8180 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8182 if (!ret
&& write
) {
8183 ret
= sched_rt_global_validate();
8187 ret
= sched_dl_global_validate();
8191 ret
= sched_rt_global_constraints();
8195 sched_rt_do_global();
8196 sched_dl_do_global();
8200 sysctl_sched_rt_period
= old_period
;
8201 sysctl_sched_rt_runtime
= old_runtime
;
8203 mutex_unlock(&mutex
);
8208 int sched_rr_handler(struct ctl_table
*table
, int write
,
8209 void __user
*buffer
, size_t *lenp
,
8213 static DEFINE_MUTEX(mutex
);
8216 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8217 /* make sure that internally we keep jiffies */
8218 /* also, writing zero resets timeslice to default */
8219 if (!ret
&& write
) {
8220 sched_rr_timeslice
= sched_rr_timeslice
<= 0 ?
8221 RR_TIMESLICE
: msecs_to_jiffies(sched_rr_timeslice
);
8223 mutex_unlock(&mutex
);
8227 #ifdef CONFIG_CGROUP_SCHED
8229 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8231 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8234 static struct cgroup_subsys_state
*
8235 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8237 struct task_group
*parent
= css_tg(parent_css
);
8238 struct task_group
*tg
;
8241 /* This is early initialization for the top cgroup */
8242 return &root_task_group
.css
;
8245 tg
= sched_create_group(parent
);
8247 return ERR_PTR(-ENOMEM
);
8249 sched_online_group(tg
, parent
);
8254 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8256 struct task_group
*tg
= css_tg(css
);
8258 sched_offline_group(tg
);
8261 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8263 struct task_group
*tg
= css_tg(css
);
8266 * Relies on the RCU grace period between css_released() and this.
8268 sched_free_group(tg
);
8272 * This is called before wake_up_new_task(), therefore we really only
8273 * have to set its group bits, all the other stuff does not apply.
8275 static void cpu_cgroup_fork(struct task_struct
*task
)
8280 rq
= task_rq_lock(task
, &rf
);
8282 sched_change_group(task
, TASK_SET_GROUP
);
8284 task_rq_unlock(rq
, task
, &rf
);
8287 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8289 struct task_struct
*task
;
8290 struct cgroup_subsys_state
*css
;
8293 cgroup_taskset_for_each(task
, css
, tset
) {
8294 #ifdef CONFIG_RT_GROUP_SCHED
8295 if (!sched_rt_can_attach(css_tg(css
), task
))
8298 /* We don't support RT-tasks being in separate groups */
8299 if (task
->sched_class
!= &fair_sched_class
)
8303 * Serialize against wake_up_new_task() such that if its
8304 * running, we're sure to observe its full state.
8306 raw_spin_lock_irq(&task
->pi_lock
);
8308 * Avoid calling sched_move_task() before wake_up_new_task()
8309 * has happened. This would lead to problems with PELT, due to
8310 * move wanting to detach+attach while we're not attached yet.
8312 if (task
->state
== TASK_NEW
)
8314 raw_spin_unlock_irq(&task
->pi_lock
);
8322 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8324 struct task_struct
*task
;
8325 struct cgroup_subsys_state
*css
;
8327 cgroup_taskset_for_each(task
, css
, tset
)
8328 sched_move_task(task
);
8331 #ifdef CONFIG_FAIR_GROUP_SCHED
8332 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8333 struct cftype
*cftype
, u64 shareval
)
8335 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8338 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8341 struct task_group
*tg
= css_tg(css
);
8343 return (u64
) scale_load_down(tg
->shares
);
8346 #ifdef CONFIG_CFS_BANDWIDTH
8347 static DEFINE_MUTEX(cfs_constraints_mutex
);
8349 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8350 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8352 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8354 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8356 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8357 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8359 if (tg
== &root_task_group
)
8363 * Ensure we have at some amount of bandwidth every period. This is
8364 * to prevent reaching a state of large arrears when throttled via
8365 * entity_tick() resulting in prolonged exit starvation.
8367 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8371 * Likewise, bound things on the otherside by preventing insane quota
8372 * periods. This also allows us to normalize in computing quota
8375 if (period
> max_cfs_quota_period
)
8379 * Prevent race between setting of cfs_rq->runtime_enabled and
8380 * unthrottle_offline_cfs_rqs().
8383 mutex_lock(&cfs_constraints_mutex
);
8384 ret
= __cfs_schedulable(tg
, period
, quota
);
8388 runtime_enabled
= quota
!= RUNTIME_INF
;
8389 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8391 * If we need to toggle cfs_bandwidth_used, off->on must occur
8392 * before making related changes, and on->off must occur afterwards
8394 if (runtime_enabled
&& !runtime_was_enabled
)
8395 cfs_bandwidth_usage_inc();
8396 raw_spin_lock_irq(&cfs_b
->lock
);
8397 cfs_b
->period
= ns_to_ktime(period
);
8398 cfs_b
->quota
= quota
;
8400 __refill_cfs_bandwidth_runtime(cfs_b
);
8401 /* restart the period timer (if active) to handle new period expiry */
8402 if (runtime_enabled
)
8403 start_cfs_bandwidth(cfs_b
);
8404 raw_spin_unlock_irq(&cfs_b
->lock
);
8406 for_each_online_cpu(i
) {
8407 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8408 struct rq
*rq
= cfs_rq
->rq
;
8410 raw_spin_lock_irq(&rq
->lock
);
8411 cfs_rq
->runtime_enabled
= runtime_enabled
;
8412 cfs_rq
->runtime_remaining
= 0;
8414 if (cfs_rq
->throttled
)
8415 unthrottle_cfs_rq(cfs_rq
);
8416 raw_spin_unlock_irq(&rq
->lock
);
8418 if (runtime_was_enabled
&& !runtime_enabled
)
8419 cfs_bandwidth_usage_dec();
8421 mutex_unlock(&cfs_constraints_mutex
);
8427 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8431 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8432 if (cfs_quota_us
< 0)
8433 quota
= RUNTIME_INF
;
8435 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8437 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8440 long tg_get_cfs_quota(struct task_group
*tg
)
8444 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8447 quota_us
= tg
->cfs_bandwidth
.quota
;
8448 do_div(quota_us
, NSEC_PER_USEC
);
8453 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8457 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8458 quota
= tg
->cfs_bandwidth
.quota
;
8460 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8463 long tg_get_cfs_period(struct task_group
*tg
)
8467 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8468 do_div(cfs_period_us
, NSEC_PER_USEC
);
8470 return cfs_period_us
;
8473 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8476 return tg_get_cfs_quota(css_tg(css
));
8479 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8480 struct cftype
*cftype
, s64 cfs_quota_us
)
8482 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8485 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8488 return tg_get_cfs_period(css_tg(css
));
8491 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8492 struct cftype
*cftype
, u64 cfs_period_us
)
8494 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8497 struct cfs_schedulable_data
{
8498 struct task_group
*tg
;
8503 * normalize group quota/period to be quota/max_period
8504 * note: units are usecs
8506 static u64
normalize_cfs_quota(struct task_group
*tg
,
8507 struct cfs_schedulable_data
*d
)
8515 period
= tg_get_cfs_period(tg
);
8516 quota
= tg_get_cfs_quota(tg
);
8519 /* note: these should typically be equivalent */
8520 if (quota
== RUNTIME_INF
|| quota
== -1)
8523 return to_ratio(period
, quota
);
8526 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8528 struct cfs_schedulable_data
*d
= data
;
8529 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8530 s64 quota
= 0, parent_quota
= -1;
8533 quota
= RUNTIME_INF
;
8535 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8537 quota
= normalize_cfs_quota(tg
, d
);
8538 parent_quota
= parent_b
->hierarchical_quota
;
8541 * ensure max(child_quota) <= parent_quota, inherit when no
8544 if (quota
== RUNTIME_INF
)
8545 quota
= parent_quota
;
8546 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8549 cfs_b
->hierarchical_quota
= quota
;
8554 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8557 struct cfs_schedulable_data data
= {
8563 if (quota
!= RUNTIME_INF
) {
8564 do_div(data
.period
, NSEC_PER_USEC
);
8565 do_div(data
.quota
, NSEC_PER_USEC
);
8569 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8575 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
8577 struct task_group
*tg
= css_tg(seq_css(sf
));
8578 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8580 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8581 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8582 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8586 #endif /* CONFIG_CFS_BANDWIDTH */
8587 #endif /* CONFIG_FAIR_GROUP_SCHED */
8589 #ifdef CONFIG_RT_GROUP_SCHED
8590 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8591 struct cftype
*cft
, s64 val
)
8593 return sched_group_set_rt_runtime(css_tg(css
), val
);
8596 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8599 return sched_group_rt_runtime(css_tg(css
));
8602 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8603 struct cftype
*cftype
, u64 rt_period_us
)
8605 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8608 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8611 return sched_group_rt_period(css_tg(css
));
8613 #endif /* CONFIG_RT_GROUP_SCHED */
8615 static struct cftype cpu_files
[] = {
8616 #ifdef CONFIG_FAIR_GROUP_SCHED
8619 .read_u64
= cpu_shares_read_u64
,
8620 .write_u64
= cpu_shares_write_u64
,
8623 #ifdef CONFIG_CFS_BANDWIDTH
8625 .name
= "cfs_quota_us",
8626 .read_s64
= cpu_cfs_quota_read_s64
,
8627 .write_s64
= cpu_cfs_quota_write_s64
,
8630 .name
= "cfs_period_us",
8631 .read_u64
= cpu_cfs_period_read_u64
,
8632 .write_u64
= cpu_cfs_period_write_u64
,
8636 .seq_show
= cpu_stats_show
,
8639 #ifdef CONFIG_RT_GROUP_SCHED
8641 .name
= "rt_runtime_us",
8642 .read_s64
= cpu_rt_runtime_read
,
8643 .write_s64
= cpu_rt_runtime_write
,
8646 .name
= "rt_period_us",
8647 .read_u64
= cpu_rt_period_read_uint
,
8648 .write_u64
= cpu_rt_period_write_uint
,
8654 struct cgroup_subsys cpu_cgrp_subsys
= {
8655 .css_alloc
= cpu_cgroup_css_alloc
,
8656 .css_released
= cpu_cgroup_css_released
,
8657 .css_free
= cpu_cgroup_css_free
,
8658 .fork
= cpu_cgroup_fork
,
8659 .can_attach
= cpu_cgroup_can_attach
,
8660 .attach
= cpu_cgroup_attach
,
8661 .legacy_cftypes
= cpu_files
,
8665 #endif /* CONFIG_CGROUP_SCHED */
8667 void dump_cpu_task(int cpu
)
8669 pr_info("Task dump for CPU %d:\n", cpu
);
8670 sched_show_task(cpu_curr(cpu
));
8674 * Nice levels are multiplicative, with a gentle 10% change for every
8675 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8676 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8677 * that remained on nice 0.
8679 * The "10% effect" is relative and cumulative: from _any_ nice level,
8680 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8681 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8682 * If a task goes up by ~10% and another task goes down by ~10% then
8683 * the relative distance between them is ~25%.)
8685 const int sched_prio_to_weight
[40] = {
8686 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8687 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8688 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8689 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8690 /* 0 */ 1024, 820, 655, 526, 423,
8691 /* 5 */ 335, 272, 215, 172, 137,
8692 /* 10 */ 110, 87, 70, 56, 45,
8693 /* 15 */ 36, 29, 23, 18, 15,
8697 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8699 * In cases where the weight does not change often, we can use the
8700 * precalculated inverse to speed up arithmetics by turning divisions
8701 * into multiplications:
8703 const u32 sched_prio_to_wmult
[40] = {
8704 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8705 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8706 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8707 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8708 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8709 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8710 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8711 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,