4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <uapi/linux/sched/types.h>
11 #include <linux/sched/loadavg.h>
12 #include <linux/sched/hotplug.h>
13 #include <linux/wait_bit.h>
14 #include <linux/cpuset.h>
15 #include <linux/delayacct.h>
16 #include <linux/init_task.h>
17 #include <linux/context_tracking.h>
18 #include <linux/rcupdate_wait.h>
20 #include <linux/blkdev.h>
21 #include <linux/cpufreq_times.h>
22 #include <linux/kprobes.h>
23 #include <linux/mmu_context.h>
24 #include <linux/module.h>
25 #include <linux/nmi.h>
26 #include <linux/prefetch.h>
27 #include <linux/profile.h>
28 #include <linux/security.h>
29 #include <linux/syscalls.h>
30 #include <linux/debug-snapshot.h>
31 #include <linux/ems.h>
33 #include <asm/switch_to.h>
35 #ifdef CONFIG_PARAVIRT
36 #include <asm/paravirt.h>
40 #include "../workqueue_internal.h"
41 #include "../smpboot.h"
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/sched.h>
47 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
50 * Debugging: various feature bits
53 #define SCHED_FEAT(name, enabled) \
54 (1UL << __SCHED_FEAT_##name) * enabled |
56 const_debug
unsigned int sysctl_sched_features
=
63 * Number of tasks to iterate in a single balance run.
64 * Limited because this is done with IRQs disabled.
66 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
69 * period over which we average the RT time consumption, measured
74 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
77 * period over which we measure -rt task CPU usage in us.
80 unsigned int sysctl_sched_rt_period
= 1000000;
82 __read_mostly
int scheduler_running
;
85 * part of the period that we allow rt tasks to run in us.
88 int sysctl_sched_rt_runtime
= 950000;
90 /* CPUs with isolated domains */
91 cpumask_var_t cpu_isolated_map
;
94 * __task_rq_lock - lock the rq @p resides on.
96 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
101 lockdep_assert_held(&p
->pi_lock
);
105 raw_spin_lock(&rq
->lock
);
106 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
110 raw_spin_unlock(&rq
->lock
);
112 while (unlikely(task_on_rq_migrating(p
)))
118 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
120 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
121 __acquires(p
->pi_lock
)
127 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
129 raw_spin_lock(&rq
->lock
);
131 * move_queued_task() task_rq_lock()
134 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
135 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
136 * [S] ->cpu = new_cpu [L] task_rq()
140 * If we observe the old cpu in task_rq_lock, the acquire of
141 * the old rq->lock will fully serialize against the stores.
143 * If we observe the new CPU in task_rq_lock, the acquire will
144 * pair with the WMB to ensure we must then also see migrating.
146 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
150 raw_spin_unlock(&rq
->lock
);
151 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
153 while (unlikely(task_on_rq_migrating(p
)))
159 * RQ-clock updating methods:
162 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
165 * In theory, the compile should just see 0 here, and optimize out the call
166 * to sched_rt_avg_update. But I don't trust it...
168 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
169 s64 steal
= 0, irq_delta
= 0;
171 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
172 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
175 * Since irq_time is only updated on {soft,}irq_exit, we might run into
176 * this case when a previous update_rq_clock() happened inside a
179 * When this happens, we stop ->clock_task and only update the
180 * prev_irq_time stamp to account for the part that fit, so that a next
181 * update will consume the rest. This ensures ->clock_task is
184 * It does however cause some slight miss-attribution of {soft,}irq
185 * time, a more accurate solution would be to update the irq_time using
186 * the current rq->clock timestamp, except that would require using
189 if (irq_delta
> delta
)
192 rq
->prev_irq_time
+= irq_delta
;
195 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
196 if (static_key_false((¶virt_steal_rq_enabled
))) {
197 steal
= paravirt_steal_clock(cpu_of(rq
));
198 steal
-= rq
->prev_steal_time_rq
;
200 if (unlikely(steal
> delta
))
203 rq
->prev_steal_time_rq
+= steal
;
208 rq
->clock_task
+= delta
;
210 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
211 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
212 sched_rt_avg_update(rq
, irq_delta
+ steal
);
216 void update_rq_clock(struct rq
*rq
)
220 lockdep_assert_held(&rq
->lock
);
222 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
225 #ifdef CONFIG_SCHED_DEBUG
226 if (sched_feat(WARN_DOUBLE_CLOCK
))
227 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
228 rq
->clock_update_flags
|= RQCF_UPDATED
;
231 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
235 update_rq_clock_task(rq
, delta
);
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
);
259 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
263 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
266 return HRTIMER_NORESTART
;
271 static void __hrtick_restart(struct rq
*rq
)
273 struct hrtimer
*timer
= &rq
->hrtick_timer
;
275 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
279 * called from hardirq (IPI) context
281 static void __hrtick_start(void *arg
)
287 __hrtick_restart(rq
);
288 rq
->hrtick_csd_pending
= 0;
293 * Called to set the hrtick timer state.
295 * called with rq->lock held and irqs disabled
297 void hrtick_start(struct rq
*rq
, u64 delay
)
299 struct hrtimer
*timer
= &rq
->hrtick_timer
;
304 * Don't schedule slices shorter than 10000ns, that just
305 * doesn't make sense and can cause timer DoS.
307 delta
= max_t(s64
, delay
, 10000LL);
308 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
310 hrtimer_set_expires(timer
, time
);
312 if (rq
== this_rq()) {
313 __hrtick_restart(rq
);
314 } else if (!rq
->hrtick_csd_pending
) {
315 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
316 rq
->hrtick_csd_pending
= 1;
322 * Called to set the hrtick timer state.
324 * called with rq->lock held and irqs disabled
326 void hrtick_start(struct rq
*rq
, u64 delay
)
329 * Don't schedule slices shorter than 10000ns, that just
330 * doesn't make sense. Rely on vruntime for fairness.
332 delay
= max_t(u64
, delay
, 10000LL);
333 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
334 HRTIMER_MODE_REL_PINNED
);
336 #endif /* CONFIG_SMP */
338 static void init_rq_hrtick(struct rq
*rq
)
341 rq
->hrtick_csd_pending
= 0;
343 rq
->hrtick_csd
.flags
= 0;
344 rq
->hrtick_csd
.func
= __hrtick_start
;
345 rq
->hrtick_csd
.info
= rq
;
348 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
349 rq
->hrtick_timer
.function
= hrtick
;
351 #else /* CONFIG_SCHED_HRTICK */
352 static inline void hrtick_clear(struct rq
*rq
)
356 static inline void init_rq_hrtick(struct rq
*rq
)
359 #endif /* CONFIG_SCHED_HRTICK */
362 * cmpxchg based fetch_or, macro so it works for different integer types
364 #define fetch_or(ptr, mask) \
366 typeof(ptr) _ptr = (ptr); \
367 typeof(mask) _mask = (mask); \
368 typeof(*_ptr) _old, _val = *_ptr; \
371 _old = cmpxchg(_ptr, _val, _val | _mask); \
379 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
381 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
382 * this avoids any races wrt polling state changes and thereby avoids
385 static bool set_nr_and_not_polling(struct task_struct
*p
)
387 struct thread_info
*ti
= task_thread_info(p
);
388 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
392 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
394 * If this returns true, then the idle task promises to call
395 * sched_ttwu_pending() and reschedule soon.
397 static bool set_nr_if_polling(struct task_struct
*p
)
399 struct thread_info
*ti
= task_thread_info(p
);
400 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
403 if (!(val
& _TIF_POLLING_NRFLAG
))
405 if (val
& _TIF_NEED_RESCHED
)
407 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
416 static bool set_nr_and_not_polling(struct task_struct
*p
)
418 set_tsk_need_resched(p
);
423 static bool set_nr_if_polling(struct task_struct
*p
)
430 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
432 struct wake_q_node
*node
= &task
->wake_q
;
435 * Atomically grab the task, if ->wake_q is !nil already it means
436 * its already queued (either by us or someone else) and will get the
437 * wakeup due to that.
439 * This cmpxchg() implies a full barrier, which pairs with the write
440 * barrier implied by the wakeup in wake_up_q().
442 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
447 get_task_struct(task
);
450 * The head is context local, there can be no concurrency.
453 head
->lastp
= &node
->next
;
457 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
,
458 int sibling_count_hint
);
460 void wake_up_q(struct wake_q_head
*head
)
462 struct wake_q_node
*node
= head
->first
;
464 while (node
!= WAKE_Q_TAIL
) {
465 struct task_struct
*task
;
467 task
= container_of(node
, struct task_struct
, wake_q
);
469 /* Task can safely be re-inserted now: */
471 task
->wake_q
.next
= NULL
;
474 * try_to_wake_up() implies a wmb() to pair with the queueing
475 * in wake_q_add() so as not to miss wakeups.
477 try_to_wake_up(task
, TASK_NORMAL
, 0, head
->count
);
478 put_task_struct(task
);
483 * resched_curr - mark rq's current task 'to be rescheduled now'.
485 * On UP this means the setting of the need_resched flag, on SMP it
486 * might also involve a cross-CPU call to trigger the scheduler on
489 void resched_curr(struct rq
*rq
)
491 struct task_struct
*curr
= rq
->curr
;
494 lockdep_assert_held(&rq
->lock
);
496 if (test_tsk_need_resched(curr
))
501 if (cpu
== smp_processor_id()) {
502 set_tsk_need_resched(curr
);
503 set_preempt_need_resched();
507 if (set_nr_and_not_polling(curr
))
508 smp_send_reschedule(cpu
);
510 trace_sched_wake_idle_without_ipi(cpu
);
513 void resched_cpu(int cpu
)
515 struct rq
*rq
= cpu_rq(cpu
);
518 raw_spin_lock_irqsave(&rq
->lock
, flags
);
519 if (cpu_online(cpu
) || cpu
== smp_processor_id())
521 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
525 #ifdef CONFIG_NO_HZ_COMMON
527 * In the semi idle case, use the nearest busy CPU for migrating timers
528 * from an idle CPU. This is good for power-savings.
530 * We don't do similar optimization for completely idle system, as
531 * selecting an idle CPU will add more delays to the timers than intended
532 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
534 int get_nohz_timer_target(void)
536 int i
, cpu
= smp_processor_id();
537 struct sched_domain
*sd
;
539 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
543 for_each_domain(cpu
, sd
) {
544 for_each_cpu(i
, sched_domain_span(sd
)) {
548 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
555 if (!is_housekeeping_cpu(cpu
))
556 cpu
= housekeeping_any_cpu();
563 * When add_timer_on() enqueues a timer into the timer wheel of an
564 * idle CPU then this timer might expire before the next timer event
565 * which is scheduled to wake up that CPU. In case of a completely
566 * idle system the next event might even be infinite time into the
567 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
568 * leaves the inner idle loop so the newly added timer is taken into
569 * account when the CPU goes back to idle and evaluates the timer
570 * wheel for the next timer event.
572 static void wake_up_idle_cpu(int cpu
)
574 struct rq
*rq
= cpu_rq(cpu
);
576 if (cpu
== smp_processor_id())
579 if (set_nr_and_not_polling(rq
->idle
))
580 smp_send_reschedule(cpu
);
582 trace_sched_wake_idle_without_ipi(cpu
);
585 static bool wake_up_full_nohz_cpu(int cpu
)
588 * We just need the target to call irq_exit() and re-evaluate
589 * the next tick. The nohz full kick at least implies that.
590 * If needed we can still optimize that later with an
593 if (cpu_is_offline(cpu
))
594 return true; /* Don't try to wake offline CPUs. */
595 if (tick_nohz_full_cpu(cpu
)) {
596 if (cpu
!= smp_processor_id() ||
597 tick_nohz_tick_stopped())
598 tick_nohz_full_kick_cpu(cpu
);
606 * Wake up the specified CPU. If the CPU is going offline, it is the
607 * caller's responsibility to deal with the lost wakeup, for example,
608 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
610 void wake_up_nohz_cpu(int cpu
)
612 if (!wake_up_full_nohz_cpu(cpu
))
613 wake_up_idle_cpu(cpu
);
616 static inline bool got_nohz_idle_kick(void)
618 int cpu
= smp_processor_id();
620 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
623 if (idle_cpu(cpu
) && !need_resched())
627 * We can't run Idle Load Balance on this CPU for this time so we
628 * cancel it and clear NOHZ_BALANCE_KICK
630 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
634 #else /* CONFIG_NO_HZ_COMMON */
636 static inline bool got_nohz_idle_kick(void)
641 #endif /* CONFIG_NO_HZ_COMMON */
643 #ifdef CONFIG_NO_HZ_FULL
644 bool sched_can_stop_tick(struct rq
*rq
)
648 /* Deadline tasks, even if single, need the tick */
649 if (rq
->dl
.dl_nr_running
)
653 * If there are more than one RR tasks, we need the tick to effect the
654 * actual RR behaviour.
656 if (rq
->rt
.rr_nr_running
) {
657 if (rq
->rt
.rr_nr_running
== 1)
664 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
665 * forced preemption between FIFO tasks.
667 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
672 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
673 * if there's more than one we need the tick for involuntary
676 if (rq
->nr_running
> 1)
681 #endif /* CONFIG_NO_HZ_FULL */
683 void sched_avg_update(struct rq
*rq
)
685 s64 period
= sched_avg_period();
687 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
689 * Inline assembly required to prevent the compiler
690 * optimising this loop into a divmod call.
691 * See __iter_div_u64_rem() for another example of this.
693 asm("" : "+rm" (rq
->age_stamp
));
694 rq
->age_stamp
+= period
;
699 #endif /* CONFIG_SMP */
701 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
702 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
704 * Iterate task_group tree rooted at *from, calling @down when first entering a
705 * node and @up when leaving it for the final time.
707 * Caller must hold rcu_lock or sufficient equivalent.
709 int walk_tg_tree_from(struct task_group
*from
,
710 tg_visitor down
, tg_visitor up
, void *data
)
712 struct task_group
*parent
, *child
;
718 ret
= (*down
)(parent
, data
);
721 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
728 ret
= (*up
)(parent
, data
);
729 if (ret
|| parent
== from
)
733 parent
= parent
->parent
;
740 int tg_nop(struct task_group
*tg
, void *data
)
746 static void set_load_weight(struct task_struct
*p
)
748 int prio
= p
->static_prio
- MAX_RT_PRIO
;
749 struct load_weight
*load
= &p
->se
.load
;
752 * SCHED_IDLE tasks get minimal weight:
754 if (idle_policy(p
->policy
)) {
755 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
756 load
->inv_weight
= WMULT_IDLEPRIO
;
760 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
761 load
->inv_weight
= sched_prio_to_wmult
[prio
];
764 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 if (!(flags
& ENQUEUE_NOCLOCK
))
769 if (!(flags
& ENQUEUE_RESTORE
))
770 sched_info_queued(rq
, p
);
772 p
->sched_class
->enqueue_task(rq
, p
, flags
);
775 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
777 if (!(flags
& DEQUEUE_NOCLOCK
))
780 if (!(flags
& DEQUEUE_SAVE
))
781 sched_info_dequeued(rq
, p
);
783 p
->sched_class
->dequeue_task(rq
, p
, flags
);
786 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
788 if (task_contributes_to_load(p
))
789 rq
->nr_uninterruptible
--;
791 enqueue_task(rq
, p
, flags
);
794 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
796 if (task_contributes_to_load(p
))
797 rq
->nr_uninterruptible
++;
799 dequeue_task(rq
, p
, flags
);
803 * __normal_prio - return the priority that is based on the static prio
805 static inline int __normal_prio(struct task_struct
*p
)
807 return p
->static_prio
;
811 * Calculate the expected normal priority: i.e. priority
812 * without taking RT-inheritance into account. Might be
813 * boosted by interactivity modifiers. Changes upon fork,
814 * setprio syscalls, and whenever the interactivity
815 * estimator recalculates.
817 static inline int normal_prio(struct task_struct
*p
)
821 if (task_has_dl_policy(p
))
822 prio
= MAX_DL_PRIO
-1;
823 else if (task_has_rt_policy(p
))
824 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
826 prio
= __normal_prio(p
);
831 * Calculate the current priority, i.e. the priority
832 * taken into account by the scheduler. This value might
833 * be boosted by RT tasks, or might be boosted by
834 * interactivity modifiers. Will be RT if the task got
835 * RT-boosted. If not then it returns p->normal_prio.
837 static int effective_prio(struct task_struct
*p
)
839 p
->normal_prio
= normal_prio(p
);
841 * If we are RT tasks or we were boosted to RT priority,
842 * keep the priority unchanged. Otherwise, update priority
843 * to the normal priority:
845 if (!rt_prio(p
->prio
))
846 return p
->normal_prio
;
851 * task_curr - is this task currently executing on a CPU?
852 * @p: the task in question.
854 * Return: 1 if the task is currently executing. 0 otherwise.
856 inline int task_curr(const struct task_struct
*p
)
858 return cpu_curr(task_cpu(p
)) == p
;
862 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
863 * use the balance_callback list if you want balancing.
865 * this means any call to check_class_changed() must be followed by a call to
866 * balance_callback().
868 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
869 const struct sched_class
*prev_class
,
872 if (prev_class
!= p
->sched_class
) {
873 if (prev_class
->switched_from
)
874 prev_class
->switched_from(rq
, p
);
876 p
->sched_class
->switched_to(rq
, p
);
877 } else if (oldprio
!= p
->prio
|| dl_task(p
))
878 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
881 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
883 const struct sched_class
*class;
885 if (p
->sched_class
== rq
->curr
->sched_class
) {
886 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
888 for_each_class(class) {
889 if (class == rq
->curr
->sched_class
)
891 if (class == p
->sched_class
) {
899 * A queue event has occurred, and we're going to schedule. In
900 * this case, we can save a useless back to back clock update.
902 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
903 rq_clock_skip_update(rq
, true);
908 static inline bool is_per_cpu_kthread(struct task_struct
*p
)
910 if (!(p
->flags
& PF_KTHREAD
))
913 if (p
->nr_cpus_allowed
!= 1)
920 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
921 * __set_cpus_allowed_ptr() and select_fallback_rq().
923 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
925 if (!cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
928 if (is_per_cpu_kthread(p
))
929 return cpu_online(cpu
);
931 return cpu_active(cpu
);
935 * This is how migration works:
937 * 1) we invoke migration_cpu_stop() on the target CPU using
939 * 2) stopper starts to run (implicitly forcing the migrated thread
941 * 3) it checks whether the migrated task is still in the wrong runqueue.
942 * 4) if it's in the wrong runqueue then the migration thread removes
943 * it and puts it into the right queue.
944 * 5) stopper completes and stop_one_cpu() returns and the migration
949 * move_queued_task - move a queued task to new rq.
951 * Returns (locked) new rq. Old rq's lock is released.
953 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
954 struct task_struct
*p
, int new_cpu
)
956 lockdep_assert_held(&rq
->lock
);
958 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
959 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
960 set_task_cpu(p
, new_cpu
);
963 rq
= cpu_rq(new_cpu
);
966 BUG_ON(task_cpu(p
) != new_cpu
);
967 enqueue_task(rq
, p
, 0);
968 p
->on_rq
= TASK_ON_RQ_QUEUED
;
969 check_preempt_curr(rq
, p
, 0);
974 struct migration_arg
{
975 struct task_struct
*task
;
980 * Move (not current) task off this CPU, onto the destination CPU. We're doing
981 * this because either it can't run here any more (set_cpus_allowed()
982 * away from this CPU, or CPU going down), or because we're
983 * attempting to rebalance this task on exec (sched_exec).
985 * So we race with normal scheduler movements, but that's OK, as long
986 * as the task is no longer on this CPU.
988 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
989 struct task_struct
*p
, int dest_cpu
)
991 /* Affinity changed (again). */
992 if (!is_cpu_allowed(p
, dest_cpu
))
996 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1002 * migration_cpu_stop - this will be executed by a highprio stopper thread
1003 * and performs thread migration by bumping thread off CPU then
1004 * 'pushing' onto another runqueue.
1006 static int migration_cpu_stop(void *data
)
1008 struct migration_arg
*arg
= data
;
1009 struct task_struct
*p
= arg
->task
;
1010 struct rq
*rq
= this_rq();
1014 * The original target CPU might have gone down and we might
1015 * be on another CPU but it doesn't matter.
1017 local_irq_disable();
1019 * We need to explicitly wake pending tasks before running
1020 * __migrate_task() such that we will not miss enforcing cpus_allowed
1021 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1023 sched_ttwu_pending();
1025 raw_spin_lock(&p
->pi_lock
);
1028 * If task_rq(p) != rq, it cannot be migrated here, because we're
1029 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1030 * we're holding p->pi_lock.
1032 if (task_rq(p
) == rq
) {
1033 if (task_on_rq_queued(p
))
1034 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1036 p
->wake_cpu
= arg
->dest_cpu
;
1039 raw_spin_unlock(&p
->pi_lock
);
1046 * sched_class::set_cpus_allowed must do the below, but is not required to
1047 * actually call this function.
1049 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1051 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1052 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1055 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1057 struct rq
*rq
= task_rq(p
);
1058 bool queued
, running
;
1060 lockdep_assert_held(&p
->pi_lock
);
1062 queued
= task_on_rq_queued(p
);
1063 running
= task_current(rq
, p
);
1067 * Because __kthread_bind() calls this on blocked tasks without
1070 lockdep_assert_held(&rq
->lock
);
1071 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1074 put_prev_task(rq
, p
);
1076 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1079 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1081 set_curr_task(rq
, p
);
1085 * Change a given task's CPU affinity. Migrate the thread to a
1086 * proper CPU and schedule it away if the CPU it's executing on
1087 * is removed from the allowed bitmask.
1089 * NOTE: the caller must have a valid reference to the task, the
1090 * task must not exit() & deallocate itself prematurely. The
1091 * call is not atomic; no spinlocks may be held.
1093 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1094 const struct cpumask
*new_mask
, bool check
)
1096 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1097 unsigned int dest_cpu
;
1102 rq
= task_rq_lock(p
, &rf
);
1103 update_rq_clock(rq
);
1105 if (p
->flags
& PF_KTHREAD
) {
1107 * Kernel threads are allowed on online && !active CPUs
1109 cpu_valid_mask
= cpu_online_mask
;
1113 * Must re-check here, to close a race against __kthread_bind(),
1114 * sched_setaffinity() is not guaranteed to observe the flag.
1116 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1121 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1124 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1129 do_set_cpus_allowed(p
, new_mask
);
1131 if (p
->flags
& PF_KTHREAD
) {
1133 * For kernel threads that do indeed end up on online &&
1134 * !active we want to ensure they are strict per-CPU threads.
1136 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1137 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1138 p
->nr_cpus_allowed
!= 1);
1141 /* Can the task run on the task's current CPU? If so, we're done */
1142 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1145 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1146 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1147 struct migration_arg arg
= { p
, dest_cpu
};
1148 /* Need help from migration thread: drop lock and wait. */
1149 task_rq_unlock(rq
, p
, &rf
);
1150 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1151 tlb_migrate_finish(p
->mm
);
1153 } else if (task_on_rq_queued(p
)) {
1155 * OK, since we're going to drop the lock immediately
1156 * afterwards anyway.
1158 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1161 task_rq_unlock(rq
, p
, &rf
);
1166 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1168 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1170 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1172 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1174 #ifdef CONFIG_SCHED_DEBUG
1176 * We should never call set_task_cpu() on a blocked task,
1177 * ttwu() will sort out the placement.
1179 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1183 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1184 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1185 * time relying on p->on_rq.
1187 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1188 p
->sched_class
== &fair_sched_class
&&
1189 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1191 #ifdef CONFIG_LOCKDEP
1193 * The caller should hold either p->pi_lock or rq->lock, when changing
1194 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1196 * sched_move_task() holds both and thus holding either pins the cgroup,
1199 * Furthermore, all task_rq users should acquire both locks, see
1202 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1203 lockdep_is_held(&task_rq(p
)->lock
)));
1206 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1208 WARN_ON_ONCE(!cpu_online(new_cpu
));
1211 trace_sched_migrate_task(p
, new_cpu
);
1213 if (task_cpu(p
) != new_cpu
) {
1214 if (p
->sched_class
->migrate_task_rq
)
1215 p
->sched_class
->migrate_task_rq(p
);
1216 p
->se
.nr_migrations
++;
1217 perf_event_task_migrate(p
);
1219 walt_fixup_busy_time(p
, new_cpu
);
1222 __set_task_cpu(p
, new_cpu
);
1225 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1227 if (task_on_rq_queued(p
)) {
1228 struct rq
*src_rq
, *dst_rq
;
1229 struct rq_flags srf
, drf
;
1231 src_rq
= task_rq(p
);
1232 dst_rq
= cpu_rq(cpu
);
1234 rq_pin_lock(src_rq
, &srf
);
1235 rq_pin_lock(dst_rq
, &drf
);
1237 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1238 deactivate_task(src_rq
, p
, 0);
1239 set_task_cpu(p
, cpu
);
1240 activate_task(dst_rq
, p
, 0);
1241 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1242 check_preempt_curr(dst_rq
, p
, 0);
1244 rq_unpin_lock(dst_rq
, &drf
);
1245 rq_unpin_lock(src_rq
, &srf
);
1249 * Task isn't running anymore; make it appear like we migrated
1250 * it before it went to sleep. This means on wakeup we make the
1251 * previous CPU our target instead of where it really is.
1257 struct migration_swap_arg
{
1258 struct task_struct
*src_task
, *dst_task
;
1259 int src_cpu
, dst_cpu
;
1262 static int migrate_swap_stop(void *data
)
1264 struct migration_swap_arg
*arg
= data
;
1265 struct rq
*src_rq
, *dst_rq
;
1268 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1271 src_rq
= cpu_rq(arg
->src_cpu
);
1272 dst_rq
= cpu_rq(arg
->dst_cpu
);
1274 double_raw_lock(&arg
->src_task
->pi_lock
,
1275 &arg
->dst_task
->pi_lock
);
1276 double_rq_lock(src_rq
, dst_rq
);
1278 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1281 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1284 if (!cpumask_test_cpu(arg
->dst_cpu
, &arg
->src_task
->cpus_allowed
))
1287 if (!cpumask_test_cpu(arg
->src_cpu
, &arg
->dst_task
->cpus_allowed
))
1290 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1291 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1296 double_rq_unlock(src_rq
, dst_rq
);
1297 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1298 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1304 * Cross migrate two tasks
1306 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1308 struct migration_swap_arg arg
;
1311 arg
= (struct migration_swap_arg
){
1313 .src_cpu
= task_cpu(cur
),
1315 .dst_cpu
= task_cpu(p
),
1318 if (arg
.src_cpu
== arg
.dst_cpu
)
1322 * These three tests are all lockless; this is OK since all of them
1323 * will be re-checked with proper locks held further down the line.
1325 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1328 if (!cpumask_test_cpu(arg
.dst_cpu
, &arg
.src_task
->cpus_allowed
))
1331 if (!cpumask_test_cpu(arg
.src_cpu
, &arg
.dst_task
->cpus_allowed
))
1334 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1335 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1342 * wait_task_inactive - wait for a thread to unschedule.
1344 * If @match_state is nonzero, it's the @p->state value just checked and
1345 * not expected to change. If it changes, i.e. @p might have woken up,
1346 * then return zero. When we succeed in waiting for @p to be off its CPU,
1347 * we return a positive number (its total switch count). If a second call
1348 * a short while later returns the same number, the caller can be sure that
1349 * @p has remained unscheduled the whole time.
1351 * The caller must ensure that the task *will* unschedule sometime soon,
1352 * else this function might spin for a *long* time. This function can't
1353 * be called with interrupts off, or it may introduce deadlock with
1354 * smp_call_function() if an IPI is sent by the same process we are
1355 * waiting to become inactive.
1357 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1359 int running
, queued
;
1366 * We do the initial early heuristics without holding
1367 * any task-queue locks at all. We'll only try to get
1368 * the runqueue lock when things look like they will
1374 * If the task is actively running on another CPU
1375 * still, just relax and busy-wait without holding
1378 * NOTE! Since we don't hold any locks, it's not
1379 * even sure that "rq" stays as the right runqueue!
1380 * But we don't care, since "task_running()" will
1381 * return false if the runqueue has changed and p
1382 * is actually now running somewhere else!
1384 while (task_running(rq
, p
)) {
1385 if (match_state
&& unlikely(p
->state
!= match_state
))
1391 * Ok, time to look more closely! We need the rq
1392 * lock now, to be *sure*. If we're wrong, we'll
1393 * just go back and repeat.
1395 rq
= task_rq_lock(p
, &rf
);
1396 trace_sched_wait_task(p
);
1397 running
= task_running(rq
, p
);
1398 queued
= task_on_rq_queued(p
);
1400 if (!match_state
|| p
->state
== match_state
)
1401 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1402 task_rq_unlock(rq
, p
, &rf
);
1405 * If it changed from the expected state, bail out now.
1407 if (unlikely(!ncsw
))
1411 * Was it really running after all now that we
1412 * checked with the proper locks actually held?
1414 * Oops. Go back and try again..
1416 if (unlikely(running
)) {
1422 * It's not enough that it's not actively running,
1423 * it must be off the runqueue _entirely_, and not
1426 * So if it was still runnable (but just not actively
1427 * running right now), it's preempted, and we should
1428 * yield - it could be a while.
1430 if (unlikely(queued
)) {
1431 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1433 set_current_state(TASK_UNINTERRUPTIBLE
);
1434 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1439 * Ahh, all good. It wasn't running, and it wasn't
1440 * runnable, which means that it will never become
1441 * running in the future either. We're all done!
1450 * kick_process - kick a running thread to enter/exit the kernel
1451 * @p: the to-be-kicked thread
1453 * Cause a process which is running on another CPU to enter
1454 * kernel-mode, without any delay. (to get signals handled.)
1456 * NOTE: this function doesn't have to take the runqueue lock,
1457 * because all it wants to ensure is that the remote task enters
1458 * the kernel. If the IPI races and the task has been migrated
1459 * to another CPU then no harm is done and the purpose has been
1462 void kick_process(struct task_struct
*p
)
1468 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1469 smp_send_reschedule(cpu
);
1472 EXPORT_SYMBOL_GPL(kick_process
);
1475 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1477 * A few notes on cpu_active vs cpu_online:
1479 * - cpu_active must be a subset of cpu_online
1481 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1482 * see __set_cpus_allowed_ptr(). At this point the newly online
1483 * CPU isn't yet part of the sched domains, and balancing will not
1486 * - on CPU-down we clear cpu_active() to mask the sched domains and
1487 * avoid the load balancer to place new tasks on the to be removed
1488 * CPU. Existing tasks will remain running there and will be taken
1491 * This means that fallback selection must not select !active CPUs.
1492 * And can assume that any active CPU must be online. Conversely
1493 * select_task_rq() below may allow selection of !active CPUs in order
1494 * to satisfy the above rules.
1496 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1498 int nid
= cpu_to_node(cpu
);
1499 const struct cpumask
*nodemask
= NULL
;
1500 enum { cpuset
, possible
, fail
} state
= cpuset
;
1504 * If the node that the CPU is on has been offlined, cpu_to_node()
1505 * will return -1. There is no CPU on the node, and we should
1506 * select the CPU on the other node.
1509 nodemask
= cpumask_of_node(nid
);
1511 /* Look for allowed, online CPU in same node. */
1512 for_each_cpu(dest_cpu
, nodemask
) {
1513 if (!cpu_active(dest_cpu
))
1515 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
1521 /* Any allowed, online CPU? */
1522 for_each_cpu(dest_cpu
, &p
->cpus_allowed
) {
1523 if (!is_cpu_allowed(p
, dest_cpu
))
1529 /* No more Mr. Nice Guy. */
1532 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1533 cpuset_cpus_allowed_fallback(p
);
1539 do_set_cpus_allowed(p
, cpu_possible_mask
);
1550 if (state
!= cpuset
) {
1552 * Don't tell them about moving exiting tasks or
1553 * kernel threads (both mm NULL), since they never
1556 if (p
->mm
&& printk_ratelimit()) {
1557 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1558 task_pid_nr(p
), p
->comm
, cpu
);
1566 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1569 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
,
1570 int sibling_count_hint
)
1572 lockdep_assert_held(&p
->pi_lock
);
1574 if (p
->nr_cpus_allowed
> 1)
1575 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
,
1576 sibling_count_hint
);
1578 cpu
= cpumask_any(&p
->cpus_allowed
);
1581 * In order not to call set_task_cpu() on a blocking task we need
1582 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1585 * Since this is common to all placement strategies, this lives here.
1587 * [ this allows ->select_task() to simply return task_cpu(p) and
1588 * not worry about this generic constraint ]
1590 if (unlikely(!is_cpu_allowed(p
, cpu
)))
1591 cpu
= select_fallback_rq(task_cpu(p
), p
);
1596 static void update_avg(u64
*avg
, u64 sample
)
1598 s64 diff
= sample
- *avg
;
1602 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1604 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1605 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1609 * Make it appear like a SCHED_FIFO task, its something
1610 * userspace knows about and won't get confused about.
1612 * Also, it will make PI more or less work without too
1613 * much confusion -- but then, stop work should not
1614 * rely on PI working anyway.
1616 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1618 stop
->sched_class
= &stop_sched_class
;
1621 cpu_rq(cpu
)->stop
= stop
;
1625 * Reset it back to a normal scheduling class so that
1626 * it can die in pieces.
1628 old_stop
->sched_class
= &rt_sched_class
;
1634 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1635 const struct cpumask
*new_mask
, bool check
)
1637 return set_cpus_allowed_ptr(p
, new_mask
);
1640 #endif /* CONFIG_SMP */
1643 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1647 if (!schedstat_enabled())
1653 if (cpu
== rq
->cpu
) {
1654 schedstat_inc(rq
->ttwu_local
);
1655 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1657 struct sched_domain
*sd
;
1659 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1661 for_each_domain(rq
->cpu
, sd
) {
1662 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1663 schedstat_inc(sd
->ttwu_wake_remote
);
1670 if (wake_flags
& WF_MIGRATED
)
1671 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1672 #endif /* CONFIG_SMP */
1674 schedstat_inc(rq
->ttwu_count
);
1675 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1677 if (wake_flags
& WF_SYNC
)
1678 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1681 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1683 activate_task(rq
, p
, en_flags
);
1684 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1686 /* If a worker is waking up, notify the workqueue: */
1687 if (p
->flags
& PF_WQ_WORKER
)
1688 wq_worker_waking_up(p
, cpu_of(rq
));
1692 * Mark the task runnable and perform wakeup-preemption.
1694 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1695 struct rq_flags
*rf
)
1697 check_preempt_curr(rq
, p
, wake_flags
);
1698 p
->state
= TASK_RUNNING
;
1699 trace_sched_wakeup(p
);
1702 if (p
->sched_class
->task_woken
) {
1704 * Our task @p is fully woken up and running; so its safe to
1705 * drop the rq->lock, hereafter rq is only used for statistics.
1707 rq_unpin_lock(rq
, rf
);
1708 p
->sched_class
->task_woken(rq
, p
);
1709 rq_repin_lock(rq
, rf
);
1712 if (rq
->idle_stamp
) {
1713 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1714 u64 max
= 2*rq
->max_idle_balance_cost
;
1716 update_avg(&rq
->avg_idle
, delta
);
1718 if (rq
->avg_idle
> max
)
1727 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1728 struct rq_flags
*rf
)
1730 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
1732 lockdep_assert_held(&rq
->lock
);
1735 if (p
->sched_contributes_to_load
)
1736 rq
->nr_uninterruptible
--;
1738 if (wake_flags
& WF_MIGRATED
)
1739 en_flags
|= ENQUEUE_MIGRATED
;
1742 ttwu_activate(rq
, p
, en_flags
);
1743 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1747 * Called in case the task @p isn't fully descheduled from its runqueue,
1748 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1749 * since all we need to do is flip p->state to TASK_RUNNING, since
1750 * the task is still ->on_rq.
1752 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1758 rq
= __task_rq_lock(p
, &rf
);
1759 if (task_on_rq_queued(p
)) {
1760 /* check_preempt_curr() may use rq clock */
1761 update_rq_clock(rq
);
1762 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1765 __task_rq_unlock(rq
, &rf
);
1771 void sched_ttwu_pending(void)
1773 struct rq
*rq
= this_rq();
1774 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1775 struct task_struct
*p
, *t
;
1781 rq_lock_irqsave(rq
, &rf
);
1782 update_rq_clock(rq
);
1784 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
1785 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
1787 rq_unlock_irqrestore(rq
, &rf
);
1790 void scheduler_ipi(void)
1793 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1794 * TIF_NEED_RESCHED remotely (for the first time) will also send
1797 preempt_fold_need_resched();
1799 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1803 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1804 * traditionally all their work was done from the interrupt return
1805 * path. Now that we actually do some work, we need to make sure
1808 * Some archs already do call them, luckily irq_enter/exit nest
1811 * Arguably we should visit all archs and update all handlers,
1812 * however a fair share of IPIs are still resched only so this would
1813 * somewhat pessimize the simple resched case.
1816 sched_ttwu_pending();
1819 * Check if someone kicked us for doing the nohz idle load balance.
1821 if (unlikely(got_nohz_idle_kick())) {
1822 this_rq()->idle_balance
= 1;
1823 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1828 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1830 struct rq
*rq
= cpu_rq(cpu
);
1832 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1834 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1835 if (!set_nr_if_polling(rq
->idle
))
1836 smp_send_reschedule(cpu
);
1838 trace_sched_wake_idle_without_ipi(cpu
);
1842 void wake_up_if_idle(int cpu
)
1844 struct rq
*rq
= cpu_rq(cpu
);
1849 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1852 if (set_nr_if_polling(rq
->idle
)) {
1853 trace_sched_wake_idle_without_ipi(cpu
);
1855 rq_lock_irqsave(rq
, &rf
);
1856 if (is_idle_task(rq
->curr
))
1857 smp_send_reschedule(cpu
);
1858 /* Else CPU is not idle, do nothing here: */
1859 rq_unlock_irqrestore(rq
, &rf
);
1866 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1868 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1870 #endif /* CONFIG_SMP */
1872 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1874 struct rq
*rq
= cpu_rq(cpu
);
1877 #if defined(CONFIG_SMP)
1878 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1879 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
1880 ttwu_queue_remote(p
, cpu
, wake_flags
);
1886 update_rq_clock(rq
);
1887 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
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 #ifdef CONFIG_SCHED_WALT
1984 /* utility function to update walt signals at wakeup */
1985 static inline void walt_try_to_wake_up(struct task_struct
*p
)
1987 struct rq
*rq
= cpu_rq(task_cpu(p
));
1991 rq_lock_irqsave(rq
, &rf
);
1992 wallclock
= walt_ktime_clock();
1993 walt_update_task_ravg(rq
->curr
, rq
, TASK_UPDATE
, wallclock
, 0);
1994 walt_update_task_ravg(p
, rq
, TASK_WAKE
, wallclock
, 0);
1995 rq_unlock_irqrestore(rq
, &rf
);
1998 #define walt_try_to_wake_up(a) {}
2003 * try_to_wake_up - wake up a thread
2004 * @p: the thread to be awakened
2005 * @state: the mask of task states that can be woken
2006 * @wake_flags: wake modifier flags (WF_*)
2007 * @sibling_count_hint: A hint at the number of threads that are being woken up
2010 * If (@state & @p->state) @p->state = TASK_RUNNING.
2012 * If the task was not queued/runnable, also place it back on a runqueue.
2014 * Atomic against schedule() which would dequeue a task, also see
2015 * set_current_state().
2017 * Return: %true if @p->state changes (an actual wakeup was done),
2021 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
,
2022 int sibling_count_hint
)
2024 unsigned long flags
;
2025 int cpu
, success
= 0;
2028 * If we are going to wake up a thread waiting for CONDITION we
2029 * need to ensure that CONDITION=1 done by the caller can not be
2030 * reordered with p->state check below. This pairs with mb() in
2031 * set_current_state() the waiting thread does.
2033 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2034 smp_mb__after_spinlock();
2035 if (!(p
->state
& state
))
2038 trace_sched_waking(p
);
2040 /* We're going to change ->state: */
2045 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2046 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2047 * in smp_cond_load_acquire() below.
2049 * sched_ttwu_pending() try_to_wake_up()
2050 * [S] p->on_rq = 1; [L] P->state
2051 * UNLOCK rq->lock -----.
2055 * LOCK rq->lock -----'
2059 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2061 * Pairs with the UNLOCK+LOCK on rq->lock from the
2062 * last wakeup of our task and the schedule that got our task
2066 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2071 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2072 * possible to, falsely, observe p->on_cpu == 0.
2074 * One must be running (->on_cpu == 1) in order to remove oneself
2075 * from the runqueue.
2077 * [S] ->on_cpu = 1; [L] ->on_rq
2081 * [S] ->on_rq = 0; [L] ->on_cpu
2083 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2084 * from the consecutive calls to schedule(); the first switching to our
2085 * task, the second putting it to sleep.
2090 * If the owning (remote) CPU is still in the middle of schedule() with
2091 * this task as prev, wait until its done referencing the task.
2093 * Pairs with the smp_store_release() in finish_lock_switch().
2095 * This ensures that tasks getting woken will be fully ordered against
2096 * their previous state and preserve Program Order.
2098 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2100 walt_try_to_wake_up(p
);
2102 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2103 p
->state
= TASK_WAKING
;
2106 delayacct_blkio_end(p
);
2107 atomic_dec(&task_rq(p
)->nr_iowait
);
2110 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
,
2111 sibling_count_hint
);
2112 if (task_cpu(p
) != cpu
) {
2113 wake_flags
|= WF_MIGRATED
;
2114 set_task_cpu(p
, cpu
);
2117 #else /* CONFIG_SMP */
2120 delayacct_blkio_end(p
);
2121 atomic_dec(&task_rq(p
)->nr_iowait
);
2124 #endif /* CONFIG_SMP */
2126 ttwu_queue(p
, cpu
, wake_flags
);
2128 ttwu_stat(p
, cpu
, wake_flags
);
2130 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2136 * try_to_wake_up_local - try to wake up a local task with rq lock held
2137 * @p: the thread to be awakened
2138 * @rf: request-queue flags for pinning
2140 * Put @p on the run-queue if it's not already there. The caller must
2141 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2144 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2146 struct rq
*rq
= task_rq(p
);
2148 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2149 WARN_ON_ONCE(p
== current
))
2152 lockdep_assert_held(&rq
->lock
);
2154 if (!raw_spin_trylock(&p
->pi_lock
)) {
2156 * This is OK, because current is on_cpu, which avoids it being
2157 * picked for load-balance and preemption/IRQs are still
2158 * disabled avoiding further scheduler activity on it and we've
2159 * not yet picked a replacement task.
2162 raw_spin_lock(&p
->pi_lock
);
2166 if (!(p
->state
& TASK_NORMAL
))
2169 trace_sched_waking(p
);
2171 if (!task_on_rq_queued(p
)) {
2172 u64 wallclock
= walt_ktime_clock();
2174 walt_update_task_ravg(rq
->curr
, rq
, TASK_UPDATE
, wallclock
, 0);
2175 walt_update_task_ravg(p
, rq
, TASK_WAKE
, wallclock
, 0);
2178 delayacct_blkio_end(p
);
2179 atomic_dec(&rq
->nr_iowait
);
2181 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
);
2184 ttwu_do_wakeup(rq
, p
, 0, rf
);
2185 ttwu_stat(p
, smp_processor_id(), 0);
2187 raw_spin_unlock(&p
->pi_lock
);
2191 * wake_up_process - Wake up a specific process
2192 * @p: The process to be woken up.
2194 * Attempt to wake up the nominated process and move it to the set of runnable
2197 * Return: 1 if the process was woken up, 0 if it was already running.
2199 * It may be assumed that this function implies a write memory barrier before
2200 * changing the task state if and only if any tasks are woken up.
2202 int wake_up_process(struct task_struct
*p
)
2204 return try_to_wake_up(p
, TASK_NORMAL
, 0, 1);
2206 EXPORT_SYMBOL(wake_up_process
);
2208 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2210 return try_to_wake_up(p
, state
, 0, 1);
2214 * Perform scheduler related setup for a newly forked process p.
2215 * p is forked by current.
2217 * __sched_fork() is basic setup used by init_idle() too:
2219 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2224 p
->se
.exec_start
= 0;
2225 p
->se
.sum_exec_runtime
= 0;
2226 p
->se
.prev_sum_exec_runtime
= 0;
2227 p
->se
.nr_migrations
= 0;
2229 #ifdef CONFIG_SCHED_WALT
2230 p
->last_sleep_ts
= 0;
2233 INIT_LIST_HEAD(&p
->se
.group_node
);
2234 #ifdef CONFIG_SCHED_EMS
2235 rcu_assign_pointer(p
->band
, NULL
);
2236 INIT_LIST_HEAD(&p
->band_members
);
2238 walt_init_new_task_load(p
);
2240 #ifdef CONFIG_FAIR_GROUP_SCHED
2241 p
->se
.cfs_rq
= NULL
;
2244 #ifdef CONFIG_SCHEDSTATS
2245 /* Even if schedstat is disabled, there should not be garbage */
2246 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2249 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2250 init_dl_task_timer(&p
->dl
);
2251 init_dl_inactive_task_timer(&p
->dl
);
2252 __dl_clear_params(p
);
2254 INIT_LIST_HEAD(&p
->rt
.run_list
);
2256 p
->rt
.time_slice
= sched_rr_timeslice
;
2260 #ifdef CONFIG_PREEMPT_NOTIFIERS
2261 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2264 #ifdef CONFIG_NUMA_BALANCING
2265 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2266 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2267 p
->mm
->numa_scan_seq
= 0;
2270 if (clone_flags
& CLONE_VM
)
2271 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2273 p
->numa_preferred_nid
= -1;
2275 p
->node_stamp
= 0ULL;
2276 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2277 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2278 p
->numa_work
.next
= &p
->numa_work
;
2279 p
->numa_faults
= NULL
;
2280 p
->last_task_numa_placement
= 0;
2281 p
->last_sum_exec_runtime
= 0;
2283 p
->numa_group
= NULL
;
2284 #endif /* CONFIG_NUMA_BALANCING */
2287 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2289 #ifdef CONFIG_NUMA_BALANCING
2291 void set_numabalancing_state(bool enabled
)
2294 static_branch_enable(&sched_numa_balancing
);
2296 static_branch_disable(&sched_numa_balancing
);
2299 #ifdef CONFIG_PROC_SYSCTL
2300 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2301 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2305 int state
= static_branch_likely(&sched_numa_balancing
);
2307 if (write
&& !capable(CAP_SYS_ADMIN
))
2312 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2316 set_numabalancing_state(state
);
2322 #ifdef CONFIG_SCHEDSTATS
2324 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2325 static bool __initdata __sched_schedstats
= false;
2327 static void set_schedstats(bool enabled
)
2330 static_branch_enable(&sched_schedstats
);
2332 static_branch_disable(&sched_schedstats
);
2335 void force_schedstat_enabled(void)
2337 if (!schedstat_enabled()) {
2338 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2339 static_branch_enable(&sched_schedstats
);
2343 static int __init
setup_schedstats(char *str
)
2350 * This code is called before jump labels have been set up, so we can't
2351 * change the static branch directly just yet. Instead set a temporary
2352 * variable so init_schedstats() can do it later.
2354 if (!strcmp(str
, "enable")) {
2355 __sched_schedstats
= true;
2357 } else if (!strcmp(str
, "disable")) {
2358 __sched_schedstats
= false;
2363 pr_warn("Unable to parse schedstats=\n");
2367 __setup("schedstats=", setup_schedstats
);
2369 static void __init
init_schedstats(void)
2371 set_schedstats(__sched_schedstats
);
2374 #ifdef CONFIG_PROC_SYSCTL
2375 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2376 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2380 int state
= static_branch_likely(&sched_schedstats
);
2382 if (write
&& !capable(CAP_SYS_ADMIN
))
2387 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2391 set_schedstats(state
);
2394 #endif /* CONFIG_PROC_SYSCTL */
2395 #else /* !CONFIG_SCHEDSTATS */
2396 static inline void init_schedstats(void) {}
2397 #endif /* CONFIG_SCHEDSTATS */
2400 * fork()/clone()-time setup:
2402 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2404 unsigned long flags
;
2405 int cpu
= get_cpu();
2407 __sched_fork(clone_flags
, p
);
2409 * We mark the process as NEW here. This guarantees that
2410 * nobody will actually run it, and a signal or other external
2411 * event cannot wake it up and insert it on the runqueue either.
2413 p
->state
= TASK_NEW
;
2416 * Make sure we do not leak PI boosting priority to the child.
2418 p
->prio
= current
->normal_prio
;
2421 * Revert to default priority/policy on fork if requested.
2423 if (unlikely(p
->sched_reset_on_fork
)) {
2424 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2425 p
->policy
= SCHED_NORMAL
;
2426 p
->static_prio
= NICE_TO_PRIO(0);
2428 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2429 p
->static_prio
= NICE_TO_PRIO(0);
2431 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2435 * We don't need the reset flag anymore after the fork. It has
2436 * fulfilled its duty:
2438 p
->sched_reset_on_fork
= 0;
2441 if (dl_prio(p
->prio
)) {
2444 } else if (rt_prio(p
->prio
)) {
2445 p
->sched_class
= &rt_sched_class
;
2447 p
->sched_class
= &fair_sched_class
;
2450 init_entity_runnable_average(&p
->se
);
2451 init_rt_entity_runnable_average(&p
->rt
);
2454 * The child is not yet in the pid-hash so no cgroup attach races,
2455 * and the cgroup is pinned to this child due to cgroup_fork()
2456 * is ran before sched_fork().
2458 * Silence PROVE_RCU.
2460 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2462 * We're setting the CPU for the first time, we don't migrate,
2463 * so use __set_task_cpu().
2465 __set_task_cpu(p
, cpu
);
2466 if (p
->sched_class
->task_fork
)
2467 p
->sched_class
->task_fork(p
);
2468 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2470 #ifdef CONFIG_SCHED_INFO
2471 if (likely(sched_info_on()))
2472 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2474 #if defined(CONFIG_SMP)
2477 init_task_preempt_count(p
);
2479 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2480 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2487 unsigned long to_ratio(u64 period
, u64 runtime
)
2489 if (runtime
== RUNTIME_INF
)
2493 * Doing this here saves a lot of checks in all
2494 * the calling paths, and returning zero seems
2495 * safe for them anyway.
2500 return div64_u64(runtime
<< BW_SHIFT
, period
);
2504 * wake_up_new_task - wake up a newly created task for the first time.
2506 * This function will do some initial scheduler statistics housekeeping
2507 * that must be done for every newly created context, then puts the task
2508 * on the runqueue and wakes it.
2510 void wake_up_new_task(struct task_struct
*p
)
2515 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2517 newbie_join_band(p
);
2519 walt_init_new_task_load(p
);
2521 p
->state
= TASK_RUNNING
;
2524 * Fork balancing, do it here and not earlier because:
2525 * - cpus_allowed can change in the fork path
2526 * - any previously selected CPU might disappear through hotplug
2528 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2529 * as we're not fully set-up yet.
2531 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0, 1));
2533 rq
= __task_rq_lock(p
, &rf
);
2534 update_rq_clock(rq
);
2535 post_init_entity_util_avg(&p
->se
);
2537 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2538 walt_mark_task_starting(p
);
2540 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2541 trace_sched_wakeup_new(p
);
2542 check_preempt_curr(rq
, p
, WF_FORK
);
2544 if (p
->sched_class
->task_woken
) {
2546 * Nothing relies on rq->lock after this, so its fine to
2549 rq_unpin_lock(rq
, &rf
);
2550 p
->sched_class
->task_woken(rq
, p
);
2551 rq_repin_lock(rq
, &rf
);
2554 task_rq_unlock(rq
, p
, &rf
);
2557 #ifdef CONFIG_PREEMPT_NOTIFIERS
2559 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2561 void preempt_notifier_inc(void)
2563 static_key_slow_inc(&preempt_notifier_key
);
2565 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2567 void preempt_notifier_dec(void)
2569 static_key_slow_dec(&preempt_notifier_key
);
2571 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2574 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2575 * @notifier: notifier struct to register
2577 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2579 if (!static_key_false(&preempt_notifier_key
))
2580 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2582 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2584 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2587 * preempt_notifier_unregister - no longer interested in preemption notifications
2588 * @notifier: notifier struct to unregister
2590 * This is *not* safe to call from within a preemption notifier.
2592 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2594 hlist_del(¬ifier
->link
);
2596 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2598 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2600 struct preempt_notifier
*notifier
;
2602 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2603 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2606 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2608 if (static_key_false(&preempt_notifier_key
))
2609 __fire_sched_in_preempt_notifiers(curr
);
2613 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2614 struct task_struct
*next
)
2616 struct preempt_notifier
*notifier
;
2618 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2619 notifier
->ops
->sched_out(notifier
, next
);
2622 static __always_inline
void
2623 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2624 struct task_struct
*next
)
2626 if (static_key_false(&preempt_notifier_key
))
2627 __fire_sched_out_preempt_notifiers(curr
, next
);
2630 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2632 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2637 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2638 struct task_struct
*next
)
2642 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2645 * prepare_task_switch - prepare to switch tasks
2646 * @rq: the runqueue preparing to switch
2647 * @prev: the current task that is being switched out
2648 * @next: the task we are going to switch to.
2650 * This is called with the rq lock held and interrupts off. It must
2651 * be paired with a subsequent finish_task_switch after the context
2654 * prepare_task_switch sets up locking and calls architecture specific
2658 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2659 struct task_struct
*next
)
2661 sched_info_switch(rq
, prev
, next
);
2662 perf_event_task_sched_out(prev
, next
);
2663 fire_sched_out_preempt_notifiers(prev
, next
);
2664 prepare_lock_switch(rq
, next
);
2665 prepare_arch_switch(next
);
2669 * finish_task_switch - clean up after a task-switch
2670 * @prev: the thread we just switched away from.
2672 * finish_task_switch must be called after the context switch, paired
2673 * with a prepare_task_switch call before the context switch.
2674 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2675 * and do any other architecture-specific cleanup actions.
2677 * Note that we may have delayed dropping an mm in context_switch(). If
2678 * so, we finish that here outside of the runqueue lock. (Doing it
2679 * with the lock held can cause deadlocks; see schedule() for
2682 * The context switch have flipped the stack from under us and restored the
2683 * local variables which were saved when this task called schedule() in the
2684 * past. prev == current is still correct but we need to recalculate this_rq
2685 * because prev may have moved to another CPU.
2687 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2688 __releases(rq
->lock
)
2690 struct rq
*rq
= this_rq();
2691 struct mm_struct
*mm
= rq
->prev_mm
;
2695 * The previous task will have left us with a preempt_count of 2
2696 * because it left us after:
2699 * preempt_disable(); // 1
2701 * raw_spin_lock_irq(&rq->lock) // 2
2703 * Also, see FORK_PREEMPT_COUNT.
2705 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2706 "corrupted preempt_count: %s/%d/0x%x\n",
2707 current
->comm
, current
->pid
, preempt_count()))
2708 preempt_count_set(FORK_PREEMPT_COUNT
);
2713 * A task struct has one reference for the use as "current".
2714 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2715 * schedule one last time. The schedule call will never return, and
2716 * the scheduled task must drop that reference.
2718 * We must observe prev->state before clearing prev->on_cpu (in
2719 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2720 * running on another CPU and we could rave with its RUNNING -> DEAD
2721 * transition, resulting in a double drop.
2723 prev_state
= prev
->state
;
2724 vtime_task_switch(prev
);
2725 perf_event_task_sched_in(prev
, current
);
2727 * The membarrier system call requires a full memory barrier
2728 * after storing to rq->curr, before going back to user-space.
2730 * TODO: This smp_mb__after_unlock_lock can go away if PPC end
2731 * up adding a full barrier to switch_mm(), or we should figure
2732 * out if a smp_mb__after_unlock_lock is really the proper API
2735 smp_mb__after_unlock_lock();
2736 finish_lock_switch(rq
, prev
);
2737 finish_arch_post_lock_switch();
2739 fire_sched_in_preempt_notifiers(current
);
2742 if (unlikely(prev_state
== TASK_DEAD
)) {
2743 if (prev
->sched_class
->task_dead
)
2744 prev
->sched_class
->task_dead(prev
);
2747 * Remove function-return probe instances associated with this
2748 * task and put them back on the free list.
2750 kprobe_flush_task(prev
);
2752 /* Task is done with its stack. */
2753 put_task_stack(prev
);
2755 put_task_struct(prev
);
2758 tick_nohz_task_switch();
2764 /* rq->lock is NOT held, but preemption is disabled */
2765 static void __balance_callback(struct rq
*rq
)
2767 struct callback_head
*head
, *next
;
2768 void (*func
)(struct rq
*rq
);
2769 unsigned long flags
;
2771 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2772 head
= rq
->balance_callback
;
2773 rq
->balance_callback
= NULL
;
2775 func
= (void (*)(struct rq
*))head
->func
;
2782 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2785 static inline void balance_callback(struct rq
*rq
)
2787 if (unlikely(rq
->balance_callback
))
2788 __balance_callback(rq
);
2793 static inline void balance_callback(struct rq
*rq
)
2800 * schedule_tail - first thing a freshly forked thread must call.
2801 * @prev: the thread we just switched away from.
2803 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2804 __releases(rq
->lock
)
2809 * New tasks start with FORK_PREEMPT_COUNT, see there and
2810 * finish_task_switch() for details.
2812 * finish_task_switch() will drop rq->lock() and lower preempt_count
2813 * and the preempt_enable() will end up enabling preemption (on
2814 * PREEMPT_COUNT kernels).
2817 rq
= finish_task_switch(prev
);
2818 balance_callback(rq
);
2821 if (current
->set_child_tid
)
2822 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2826 * context_switch - switch to the new MM and the new thread's register state.
2828 static __always_inline
struct rq
*
2829 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2830 struct task_struct
*next
, struct rq_flags
*rf
)
2832 struct mm_struct
*mm
, *oldmm
;
2834 prepare_task_switch(rq
, prev
, next
);
2837 oldmm
= prev
->active_mm
;
2839 * For paravirt, this is coupled with an exit in switch_to to
2840 * combine the page table reload and the switch backend into
2843 arch_start_context_switch(prev
);
2846 next
->active_mm
= oldmm
;
2848 enter_lazy_tlb(oldmm
, next
);
2850 switch_mm_irqs_off(oldmm
, mm
, next
);
2853 prev
->active_mm
= NULL
;
2854 rq
->prev_mm
= oldmm
;
2857 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
2860 * Since the runqueue lock will be released by the next
2861 * task (which is an invalid locking op but in the case
2862 * of the scheduler it's an obvious special-case), so we
2863 * do an early lockdep release here:
2865 rq_unpin_lock(rq
, rf
);
2866 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2868 /* Here we just switch the register state and the stack. */
2869 switch_to(prev
, next
, prev
);
2872 return finish_task_switch(prev
);
2876 * nr_running and nr_context_switches:
2878 * externally visible scheduler statistics: current number of runnable
2879 * threads, total number of context switches performed since bootup.
2881 unsigned long nr_running(void)
2883 unsigned long i
, sum
= 0;
2885 for_each_online_cpu(i
)
2886 sum
+= cpu_rq(i
)->nr_running
;
2892 * Check if only the current task is running on the CPU.
2894 * Caution: this function does not check that the caller has disabled
2895 * preemption, thus the result might have a time-of-check-to-time-of-use
2896 * race. The caller is responsible to use it correctly, for example:
2898 * - from a non-preemptable section (of course)
2900 * - from a thread that is bound to a single CPU
2902 * - in a loop with very short iterations (e.g. a polling loop)
2904 bool single_task_running(void)
2906 return raw_rq()->nr_running
== 1;
2908 EXPORT_SYMBOL(single_task_running
);
2910 unsigned long long nr_context_switches(void)
2913 unsigned long long sum
= 0;
2915 for_each_possible_cpu(i
)
2916 sum
+= cpu_rq(i
)->nr_switches
;
2922 * IO-wait accounting, and how its mostly bollocks (on SMP).
2924 * The idea behind IO-wait account is to account the idle time that we could
2925 * have spend running if it were not for IO. That is, if we were to improve the
2926 * storage performance, we'd have a proportional reduction in IO-wait time.
2928 * This all works nicely on UP, where, when a task blocks on IO, we account
2929 * idle time as IO-wait, because if the storage were faster, it could've been
2930 * running and we'd not be idle.
2932 * This has been extended to SMP, by doing the same for each CPU. This however
2935 * Imagine for instance the case where two tasks block on one CPU, only the one
2936 * CPU will have IO-wait accounted, while the other has regular idle. Even
2937 * though, if the storage were faster, both could've ran at the same time,
2938 * utilising both CPUs.
2940 * This means, that when looking globally, the current IO-wait accounting on
2941 * SMP is a lower bound, by reason of under accounting.
2943 * Worse, since the numbers are provided per CPU, they are sometimes
2944 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2945 * associated with any one particular CPU, it can wake to another CPU than it
2946 * blocked on. This means the per CPU IO-wait number is meaningless.
2948 * Task CPU affinities can make all that even more 'interesting'.
2951 unsigned long nr_iowait(void)
2953 unsigned long i
, sum
= 0;
2955 for_each_possible_cpu(i
)
2956 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2962 * Consumers of these two interfaces, like for example the cpufreq menu
2963 * governor are using nonsensical data. Boosting frequency for a CPU that has
2964 * IO-wait which might not even end up running the task when it does become
2968 unsigned long nr_iowait_cpu(int cpu
)
2970 struct rq
*this = cpu_rq(cpu
);
2971 return atomic_read(&this->nr_iowait
);
2974 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2976 struct rq
*rq
= this_rq();
2977 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2978 *load
= rq
->load
.weight
;
2984 * sched_exec - execve() is a valuable balancing opportunity, because at
2985 * this point the task has the smallest effective memory and cache footprint.
2987 void sched_exec(void)
2989 struct task_struct
*p
= current
;
2990 unsigned long flags
;
2993 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2994 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0, 1);
2995 if (dest_cpu
== smp_processor_id())
2998 if (likely(cpu_active(dest_cpu
))) {
2999 struct migration_arg arg
= { p
, dest_cpu
};
3001 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3002 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3006 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3011 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3012 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3014 EXPORT_PER_CPU_SYMBOL(kstat
);
3015 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3018 * The function fair_sched_class.update_curr accesses the struct curr
3019 * and its field curr->exec_start; when called from task_sched_runtime(),
3020 * we observe a high rate of cache misses in practice.
3021 * Prefetching this data results in improved performance.
3023 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3025 #ifdef CONFIG_FAIR_GROUP_SCHED
3026 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3028 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3031 prefetch(&curr
->exec_start
);
3035 * Return accounted runtime for the task.
3036 * In case the task is currently running, return the runtime plus current's
3037 * pending runtime that have not been accounted yet.
3039 unsigned long long task_sched_runtime(struct task_struct
*p
)
3045 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3047 * 64-bit doesn't need locks to atomically read a 64bit value.
3048 * So we have a optimization chance when the task's delta_exec is 0.
3049 * Reading ->on_cpu is racy, but this is ok.
3051 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3052 * If we race with it entering CPU, unaccounted time is 0. This is
3053 * indistinguishable from the read occurring a few cycles earlier.
3054 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3055 * been accounted, so we're correct here as well.
3057 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3058 return p
->se
.sum_exec_runtime
;
3061 rq
= task_rq_lock(p
, &rf
);
3063 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3064 * project cycles that may never be accounted to this
3065 * thread, breaking clock_gettime().
3067 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3068 prefetch_curr_exec_start(p
);
3069 update_rq_clock(rq
);
3070 p
->sched_class
->update_curr(rq
);
3072 ns
= p
->se
.sum_exec_runtime
;
3073 task_rq_unlock(rq
, p
, &rf
);
3079 * This function gets called by the timer code, with HZ frequency.
3080 * We call it with interrupts disabled.
3082 void scheduler_tick(void)
3084 int cpu
= smp_processor_id();
3085 struct rq
*rq
= cpu_rq(cpu
);
3086 struct task_struct
*curr
= rq
->curr
;
3093 walt_set_window_start(rq
, &rf
);
3094 walt_update_task_ravg(rq
->curr
, rq
, TASK_UPDATE
,
3095 walt_ktime_clock(), 0);
3096 update_rq_clock(rq
);
3097 curr
->sched_class
->task_tick(rq
, curr
, 0);
3098 cpu_load_update_active(rq
);
3099 calc_global_load_tick(rq
);
3103 perf_event_task_tick();
3106 rq
->idle_balance
= idle_cpu(cpu
);
3107 trigger_load_balance(rq
);
3109 rq_last_tick_reset(rq
);
3111 update_band(curr
, -1);
3114 #ifdef CONFIG_NO_HZ_FULL
3116 * scheduler_tick_max_deferment
3118 * Keep at least one tick per second when a single
3119 * active task is running because the scheduler doesn't
3120 * yet completely support full dynticks environment.
3122 * This makes sure that uptime, CFS vruntime, load
3123 * balancing, etc... continue to move forward, even
3124 * with a very low granularity.
3126 * Return: Maximum deferment in nanoseconds.
3128 u64
scheduler_tick_max_deferment(void)
3130 struct rq
*rq
= this_rq();
3131 unsigned long next
, now
= READ_ONCE(jiffies
);
3133 next
= rq
->last_sched_tick
+ HZ
;
3135 if (time_before_eq(next
, now
))
3138 return jiffies_to_nsecs(next
- now
);
3142 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3143 defined(CONFIG_PREEMPT_TRACER))
3145 * If the value passed in is equal to the current preempt count
3146 * then we just disabled preemption. Start timing the latency.
3148 static inline void preempt_latency_start(int val
)
3150 if (preempt_count() == val
) {
3151 unsigned long ip
= get_lock_parent_ip();
3152 #ifdef CONFIG_DEBUG_PREEMPT
3153 current
->preempt_disable_ip
= ip
;
3155 trace_preempt_off(CALLER_ADDR0
, ip
);
3159 void preempt_count_add(int val
)
3161 #ifdef CONFIG_DEBUG_PREEMPT
3165 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3168 __preempt_count_add(val
);
3169 #ifdef CONFIG_DEBUG_PREEMPT
3171 * Spinlock count overflowing soon?
3173 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3176 preempt_latency_start(val
);
3178 EXPORT_SYMBOL(preempt_count_add
);
3179 NOKPROBE_SYMBOL(preempt_count_add
);
3182 * If the value passed in equals to the current preempt count
3183 * then we just enabled preemption. Stop timing the latency.
3185 static inline void preempt_latency_stop(int val
)
3187 if (preempt_count() == val
)
3188 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3191 void preempt_count_sub(int val
)
3193 #ifdef CONFIG_DEBUG_PREEMPT
3197 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3200 * Is the spinlock portion underflowing?
3202 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3203 !(preempt_count() & PREEMPT_MASK
)))
3207 preempt_latency_stop(val
);
3208 __preempt_count_sub(val
);
3210 EXPORT_SYMBOL(preempt_count_sub
);
3211 NOKPROBE_SYMBOL(preempt_count_sub
);
3214 static inline void preempt_latency_start(int val
) { }
3215 static inline void preempt_latency_stop(int val
) { }
3218 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3220 #ifdef CONFIG_DEBUG_PREEMPT
3221 return p
->preempt_disable_ip
;
3228 * Print scheduling while atomic bug:
3230 static noinline
void __schedule_bug(struct task_struct
*prev
)
3232 /* Save this before calling printk(), since that will clobber it */
3233 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3235 if (oops_in_progress
)
3238 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3239 prev
->comm
, prev
->pid
, preempt_count());
3241 debug_show_held_locks(prev
);
3243 if (irqs_disabled())
3244 print_irqtrace_events(prev
);
3245 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3246 && in_atomic_preempt_off()) {
3247 pr_err("Preemption disabled at:");
3248 print_ip_sym(preempt_disable_ip
);
3252 panic("scheduling while atomic\n");
3255 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3259 * Various schedule()-time debugging checks and statistics:
3261 static inline void schedule_debug(struct task_struct
*prev
)
3263 #ifdef CONFIG_SCHED_STACK_END_CHECK
3264 if (task_stack_end_corrupted(prev
))
3265 panic("corrupted stack end detected inside scheduler\n");
3268 if (unlikely(in_atomic_preempt_off())) {
3269 __schedule_bug(prev
);
3270 preempt_count_set(PREEMPT_DISABLED
);
3274 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3276 schedstat_inc(this_rq()->sched_count
);
3280 * Pick up the highest-prio task:
3282 static inline struct task_struct
*
3283 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3285 const struct sched_class
*class;
3286 struct task_struct
*p
;
3289 * Optimization: we know that if all tasks are in the fair class we can
3290 * call that function directly, but only if the @prev task wasn't of a
3291 * higher scheduling class, because otherwise those loose the
3292 * opportunity to pull in more work from other CPUs.
3294 if (likely((prev
->sched_class
== &idle_sched_class
||
3295 prev
->sched_class
== &fair_sched_class
) &&
3296 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3298 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3299 if (unlikely(p
== RETRY_TASK
))
3302 /* Assumes fair_sched_class->next == idle_sched_class */
3304 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3310 for_each_class(class) {
3311 p
= class->pick_next_task(rq
, prev
, rf
);
3313 if (unlikely(p
== RETRY_TASK
))
3319 /* The idle class should always have a runnable task: */
3324 * __schedule() is the main scheduler function.
3326 * The main means of driving the scheduler and thus entering this function are:
3328 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3330 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3331 * paths. For example, see arch/x86/entry_64.S.
3333 * To drive preemption between tasks, the scheduler sets the flag in timer
3334 * interrupt handler scheduler_tick().
3336 * 3. Wakeups don't really cause entry into schedule(). They add a
3337 * task to the run-queue and that's it.
3339 * Now, if the new task added to the run-queue preempts the current
3340 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3341 * called on the nearest possible occasion:
3343 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3345 * - in syscall or exception context, at the next outmost
3346 * preempt_enable(). (this might be as soon as the wake_up()'s
3349 * - in IRQ context, return from interrupt-handler to
3350 * preemptible context
3352 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3355 * - cond_resched() call
3356 * - explicit schedule() call
3357 * - return from syscall or exception to user-space
3358 * - return from interrupt-handler to user-space
3360 * WARNING: must be called with preemption disabled!
3362 static void __sched notrace
__schedule(bool preempt
)
3364 struct task_struct
*prev
, *next
;
3365 unsigned long *switch_count
;
3371 cpu
= smp_processor_id();
3375 schedule_debug(prev
);
3377 if (sched_feat(HRTICK
))
3380 local_irq_disable();
3381 rcu_note_context_switch(preempt
);
3384 * Make sure that signal_pending_state()->signal_pending() below
3385 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3386 * done by the caller to avoid the race with signal_wake_up().
3389 smp_mb__after_spinlock();
3391 /* Promote REQ to ACT */
3392 rq
->clock_update_flags
<<= 1;
3393 update_rq_clock(rq
);
3395 switch_count
= &prev
->nivcsw
;
3396 if (!preempt
&& prev
->state
) {
3397 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3398 prev
->state
= TASK_RUNNING
;
3400 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
3403 if (prev
->in_iowait
) {
3404 atomic_inc(&rq
->nr_iowait
);
3405 delayacct_blkio_start();
3409 * If a worker went to sleep, notify and ask workqueue
3410 * whether it wants to wake up a task to maintain
3413 if (prev
->flags
& PF_WQ_WORKER
) {
3414 struct task_struct
*to_wakeup
;
3416 to_wakeup
= wq_worker_sleeping(prev
);
3418 try_to_wake_up_local(to_wakeup
, &rf
);
3421 switch_count
= &prev
->nvcsw
;
3424 next
= pick_next_task(rq
, prev
, &rf
);
3425 wallclock
= walt_ktime_clock();
3426 walt_update_task_ravg(prev
, rq
, PUT_PREV_TASK
, wallclock
, 0);
3427 walt_update_task_ravg(next
, rq
, PICK_NEXT_TASK
, wallclock
, 0);
3428 clear_tsk_need_resched(prev
);
3429 clear_preempt_need_resched();
3431 if (likely(prev
!= next
)) {
3432 #ifdef CONFIG_SCHED_WALT
3434 prev
->last_sleep_ts
= wallclock
;
3439 * The membarrier system call requires each architecture
3440 * to have a full memory barrier after updating
3441 * rq->curr, before returning to user-space. For TSO
3442 * (e.g. x86), the architecture must provide its own
3443 * barrier in switch_mm(). For weakly ordered machines
3444 * for which spin_unlock() acts as a full memory
3445 * barrier, finish_lock_switch() in common code takes
3446 * care of this barrier. For weakly ordered machines for
3447 * which spin_unlock() acts as a RELEASE barrier (only
3448 * arm64 and PowerPC), arm64 has a full barrier in
3449 * switch_to(), and PowerPC has
3450 * smp_mb__after_unlock_lock() before
3451 * finish_lock_switch().
3455 trace_sched_switch(preempt
, prev
, next
);
3457 /* Also unlocks the rq: */
3458 rq
= context_switch(rq
, prev
, next
, &rf
);
3460 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3461 rq_unlock_irq(rq
, &rf
);
3464 dbg_snapshot_task(smp_processor_id(), rq
->curr
);
3465 balance_callback(rq
);
3468 void __noreturn
do_task_dead(void)
3470 /* Causes final put_task_struct in finish_task_switch(): */
3471 set_special_state(TASK_DEAD
);
3473 /* Tell freezer to ignore us: */
3474 current
->flags
|= PF_NOFREEZE
;
3479 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3484 static inline void sched_submit_work(struct task_struct
*tsk
)
3486 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3489 * If we are going to sleep and we have plugged IO queued,
3490 * make sure to submit it to avoid deadlocks.
3492 if (blk_needs_flush_plug(tsk
))
3493 blk_schedule_flush_plug(tsk
);
3496 asmlinkage __visible
void __sched
schedule(void)
3498 struct task_struct
*tsk
= current
;
3500 sched_submit_work(tsk
);
3504 sched_preempt_enable_no_resched();
3505 } while (need_resched());
3507 EXPORT_SYMBOL(schedule
);
3510 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3511 * state (have scheduled out non-voluntarily) by making sure that all
3512 * tasks have either left the run queue or have gone into user space.
3513 * As idle tasks do not do either, they must not ever be preempted
3514 * (schedule out non-voluntarily).
3516 * schedule_idle() is similar to schedule_preempt_disable() except that it
3517 * never enables preemption because it does not call sched_submit_work().
3519 void __sched
schedule_idle(void)
3522 * As this skips calling sched_submit_work(), which the idle task does
3523 * regardless because that function is a nop when the task is in a
3524 * TASK_RUNNING state, make sure this isn't used someplace that the
3525 * current task can be in any other state. Note, idle is always in the
3526 * TASK_RUNNING state.
3528 WARN_ON_ONCE(current
->state
);
3531 } while (need_resched());
3534 #ifdef CONFIG_CONTEXT_TRACKING
3535 asmlinkage __visible
void __sched
schedule_user(void)
3538 * If we come here after a random call to set_need_resched(),
3539 * or we have been woken up remotely but the IPI has not yet arrived,
3540 * we haven't yet exited the RCU idle mode. Do it here manually until
3541 * we find a better solution.
3543 * NB: There are buggy callers of this function. Ideally we
3544 * should warn if prev_state != CONTEXT_USER, but that will trigger
3545 * too frequently to make sense yet.
3547 enum ctx_state prev_state
= exception_enter();
3549 exception_exit(prev_state
);
3554 * schedule_preempt_disabled - called with preemption disabled
3556 * Returns with preemption disabled. Note: preempt_count must be 1
3558 void __sched
schedule_preempt_disabled(void)
3560 sched_preempt_enable_no_resched();
3565 static void __sched notrace
preempt_schedule_common(void)
3569 * Because the function tracer can trace preempt_count_sub()
3570 * and it also uses preempt_enable/disable_notrace(), if
3571 * NEED_RESCHED is set, the preempt_enable_notrace() called
3572 * by the function tracer will call this function again and
3573 * cause infinite recursion.
3575 * Preemption must be disabled here before the function
3576 * tracer can trace. Break up preempt_disable() into two
3577 * calls. One to disable preemption without fear of being
3578 * traced. The other to still record the preemption latency,
3579 * which can also be traced by the function tracer.
3581 preempt_disable_notrace();
3582 preempt_latency_start(1);
3584 preempt_latency_stop(1);
3585 preempt_enable_no_resched_notrace();
3588 * Check again in case we missed a preemption opportunity
3589 * between schedule and now.
3591 } while (need_resched());
3594 #ifdef CONFIG_PREEMPT
3596 * this is the entry point to schedule() from in-kernel preemption
3597 * off of preempt_enable. Kernel preemptions off return from interrupt
3598 * occur there and call schedule directly.
3600 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3603 * If there is a non-zero preempt_count or interrupts are disabled,
3604 * we do not want to preempt the current task. Just return..
3606 if (likely(!preemptible()))
3609 preempt_schedule_common();
3611 NOKPROBE_SYMBOL(preempt_schedule
);
3612 EXPORT_SYMBOL(preempt_schedule
);
3615 * preempt_schedule_notrace - preempt_schedule called by tracing
3617 * The tracing infrastructure uses preempt_enable_notrace to prevent
3618 * recursion and tracing preempt enabling caused by the tracing
3619 * infrastructure itself. But as tracing can happen in areas coming
3620 * from userspace or just about to enter userspace, a preempt enable
3621 * can occur before user_exit() is called. This will cause the scheduler
3622 * to be called when the system is still in usermode.
3624 * To prevent this, the preempt_enable_notrace will use this function
3625 * instead of preempt_schedule() to exit user context if needed before
3626 * calling the scheduler.
3628 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3630 enum ctx_state prev_ctx
;
3632 if (likely(!preemptible()))
3637 * Because the function tracer can trace preempt_count_sub()
3638 * and it also uses preempt_enable/disable_notrace(), if
3639 * NEED_RESCHED is set, the preempt_enable_notrace() called
3640 * by the function tracer will call this function again and
3641 * cause infinite recursion.
3643 * Preemption must be disabled here before the function
3644 * tracer can trace. Break up preempt_disable() into two
3645 * calls. One to disable preemption without fear of being
3646 * traced. The other to still record the preemption latency,
3647 * which can also be traced by the function tracer.
3649 preempt_disable_notrace();
3650 preempt_latency_start(1);
3652 * Needs preempt disabled in case user_exit() is traced
3653 * and the tracer calls preempt_enable_notrace() causing
3654 * an infinite recursion.
3656 prev_ctx
= exception_enter();
3658 exception_exit(prev_ctx
);
3660 preempt_latency_stop(1);
3661 preempt_enable_no_resched_notrace();
3662 } while (need_resched());
3664 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3666 #endif /* CONFIG_PREEMPT */
3669 * this is the entry point to schedule() from kernel preemption
3670 * off of irq context.
3671 * Note, that this is called and return with irqs disabled. This will
3672 * protect us against recursive calling from irq.
3674 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3676 enum ctx_state prev_state
;
3678 /* Catch callers which need to be fixed */
3679 BUG_ON(preempt_count() || !irqs_disabled());
3681 prev_state
= exception_enter();
3687 local_irq_disable();
3688 sched_preempt_enable_no_resched();
3689 } while (need_resched());
3691 exception_exit(prev_state
);
3694 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
3697 return try_to_wake_up(curr
->private, mode
, wake_flags
, 1);
3699 EXPORT_SYMBOL(default_wake_function
);
3701 #ifdef CONFIG_RT_MUTEXES
3703 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
3706 prio
= min(prio
, pi_task
->prio
);
3711 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3713 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3715 return __rt_effective_prio(pi_task
, prio
);
3719 * rt_mutex_setprio - set the current priority of a task
3721 * @pi_task: donor task
3723 * This function changes the 'effective' priority of a task. It does
3724 * not touch ->normal_prio like __setscheduler().
3726 * Used by the rt_mutex code to implement priority inheritance
3727 * logic. Call site only calls if the priority of the task changed.
3729 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
3731 int prio
, oldprio
, queued
, running
, queue_flag
=
3732 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
3733 const struct sched_class
*prev_class
;
3737 /* XXX used to be waiter->prio, not waiter->task->prio */
3738 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
3741 * If nothing changed; bail early.
3743 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
3746 rq
= __task_rq_lock(p
, &rf
);
3747 update_rq_clock(rq
);
3749 * Set under pi_lock && rq->lock, such that the value can be used under
3752 * Note that there is loads of tricky to make this pointer cache work
3753 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3754 * ensure a task is de-boosted (pi_task is set to NULL) before the
3755 * task is allowed to run again (and can exit). This ensures the pointer
3756 * points to a blocked task -- which guaratees the task is present.
3758 p
->pi_top_task
= pi_task
;
3761 * For FIFO/RR we only need to set prio, if that matches we're done.
3763 if (prio
== p
->prio
&& !dl_prio(prio
))
3767 * Idle task boosting is a nono in general. There is one
3768 * exception, when PREEMPT_RT and NOHZ is active:
3770 * The idle task calls get_next_timer_interrupt() and holds
3771 * the timer wheel base->lock on the CPU and another CPU wants
3772 * to access the timer (probably to cancel it). We can safely
3773 * ignore the boosting request, as the idle CPU runs this code
3774 * with interrupts disabled and will complete the lock
3775 * protected section without being interrupted. So there is no
3776 * real need to boost.
3778 if (unlikely(p
== rq
->idle
)) {
3779 WARN_ON(p
!= rq
->curr
);
3780 WARN_ON(p
->pi_blocked_on
);
3784 trace_sched_pi_setprio(p
, pi_task
);
3787 if (oldprio
== prio
)
3788 queue_flag
&= ~DEQUEUE_MOVE
;
3790 prev_class
= p
->sched_class
;
3791 queued
= task_on_rq_queued(p
);
3792 running
= task_current(rq
, p
);
3794 dequeue_task(rq
, p
, queue_flag
);
3796 put_prev_task(rq
, p
);
3799 * Boosting condition are:
3800 * 1. -rt task is running and holds mutex A
3801 * --> -dl task blocks on mutex A
3803 * 2. -dl task is running and holds mutex A
3804 * --> -dl task blocks on mutex A and could preempt the
3807 if (dl_prio(prio
)) {
3808 if (!dl_prio(p
->normal_prio
) ||
3809 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3810 p
->dl
.dl_boosted
= 1;
3811 queue_flag
|= ENQUEUE_REPLENISH
;
3813 p
->dl
.dl_boosted
= 0;
3814 p
->sched_class
= &dl_sched_class
;
3815 } else if (rt_prio(prio
)) {
3816 if (dl_prio(oldprio
))
3817 p
->dl
.dl_boosted
= 0;
3819 queue_flag
|= ENQUEUE_HEAD
;
3820 p
->sched_class
= &rt_sched_class
;
3822 if (dl_prio(oldprio
))
3823 p
->dl
.dl_boosted
= 0;
3824 if (rt_prio(oldprio
))
3826 p
->sched_class
= &fair_sched_class
;
3832 enqueue_task(rq
, p
, queue_flag
);
3834 set_curr_task(rq
, p
);
3836 check_class_changed(rq
, p
, prev_class
, oldprio
);
3838 /* Avoid rq from going away on us: */
3840 __task_rq_unlock(rq
, &rf
);
3842 balance_callback(rq
);
3846 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
3852 void set_user_nice(struct task_struct
*p
, long nice
)
3854 bool queued
, running
;
3855 int old_prio
, delta
;
3859 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3862 * We have to be careful, if called from sys_setpriority(),
3863 * the task might be in the middle of scheduling on another CPU.
3865 rq
= task_rq_lock(p
, &rf
);
3866 update_rq_clock(rq
);
3869 * The RT priorities are set via sched_setscheduler(), but we still
3870 * allow the 'normal' nice value to be set - but as expected
3871 * it wont have any effect on scheduling until the task is
3872 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3874 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3875 p
->static_prio
= NICE_TO_PRIO(nice
);
3878 queued
= task_on_rq_queued(p
);
3879 running
= task_current(rq
, p
);
3881 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
3883 put_prev_task(rq
, p
);
3885 p
->static_prio
= NICE_TO_PRIO(nice
);
3888 p
->prio
= effective_prio(p
);
3889 delta
= p
->prio
- old_prio
;
3892 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
3894 * If the task increased its priority or is running and
3895 * lowered its priority, then reschedule its CPU:
3897 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3901 set_curr_task(rq
, p
);
3903 task_rq_unlock(rq
, p
, &rf
);
3905 EXPORT_SYMBOL(set_user_nice
);
3908 * can_nice - check if a task can reduce its nice value
3912 int can_nice(const struct task_struct
*p
, const int nice
)
3914 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3915 int nice_rlim
= nice_to_rlimit(nice
);
3917 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3918 capable(CAP_SYS_NICE
));
3921 #ifdef __ARCH_WANT_SYS_NICE
3924 * sys_nice - change the priority of the current process.
3925 * @increment: priority increment
3927 * sys_setpriority is a more generic, but much slower function that
3928 * does similar things.
3930 SYSCALL_DEFINE1(nice
, int, increment
)
3935 * Setpriority might change our priority at the same moment.
3936 * We don't have to worry. Conceptually one call occurs first
3937 * and we have a single winner.
3939 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3940 nice
= task_nice(current
) + increment
;
3942 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3943 if (increment
< 0 && !can_nice(current
, nice
))
3946 retval
= security_task_setnice(current
, nice
);
3950 set_user_nice(current
, nice
);
3957 * task_prio - return the priority value of a given task.
3958 * @p: the task in question.
3960 * Return: The priority value as seen by users in /proc.
3961 * RT tasks are offset by -200. Normal tasks are centered
3962 * around 0, value goes from -16 to +15.
3964 int task_prio(const struct task_struct
*p
)
3966 return p
->prio
- MAX_RT_PRIO
;
3970 * idle_cpu - is a given CPU idle currently?
3971 * @cpu: the processor in question.
3973 * Return: 1 if the CPU is currently idle. 0 otherwise.
3975 int idle_cpu(int cpu
)
3977 struct rq
*rq
= cpu_rq(cpu
);
3979 if (rq
->curr
!= rq
->idle
)
3982 if (rq
->nr_running
== 1)
3986 if (!llist_empty(&rq
->wake_list
))
3994 * idle_task - return the idle task for a given CPU.
3995 * @cpu: the processor in question.
3997 * Return: The idle task for the CPU @cpu.
3999 struct task_struct
*idle_task(int cpu
)
4001 return cpu_rq(cpu
)->idle
;
4005 * find_process_by_pid - find a process with a matching PID value.
4006 * @pid: the pid in question.
4008 * The task of @pid, if found. %NULL otherwise.
4010 static struct task_struct
*find_process_by_pid(pid_t pid
)
4012 return pid
? find_task_by_vpid(pid
) : current
;
4016 * sched_setparam() passes in -1 for its policy, to let the functions
4017 * it calls know not to change it.
4019 #define SETPARAM_POLICY -1
4021 static void __setscheduler_params(struct task_struct
*p
,
4022 const struct sched_attr
*attr
)
4024 int policy
= attr
->sched_policy
;
4026 if (policy
== SETPARAM_POLICY
)
4031 if (dl_policy(policy
))
4032 __setparam_dl(p
, attr
);
4033 else if (fair_policy(policy
))
4034 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4037 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4038 * !rt_policy. Always setting this ensures that things like
4039 * getparam()/getattr() don't report silly values for !rt tasks.
4041 p
->rt_priority
= attr
->sched_priority
;
4042 p
->normal_prio
= normal_prio(p
);
4046 /* Actually do priority change: must hold pi & rq lock. */
4047 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4048 const struct sched_attr
*attr
, bool keep_boost
)
4050 __setscheduler_params(p
, attr
);
4053 * Keep a potential priority boosting if called from
4054 * sched_setscheduler().
4056 p
->prio
= normal_prio(p
);
4058 p
->prio
= rt_effective_prio(p
, p
->prio
);
4060 if (dl_prio(p
->prio
))
4061 p
->sched_class
= &dl_sched_class
;
4062 else if (rt_prio(p
->prio
))
4063 p
->sched_class
= &rt_sched_class
;
4065 p
->sched_class
= &fair_sched_class
;
4069 * Check the target process has a UID that matches the current process's:
4071 static bool check_same_owner(struct task_struct
*p
)
4073 const struct cred
*cred
= current_cred(), *pcred
;
4077 pcred
= __task_cred(p
);
4078 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4079 uid_eq(cred
->euid
, pcred
->uid
));
4084 static int __sched_setscheduler(struct task_struct
*p
,
4085 const struct sched_attr
*attr
,
4088 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4089 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4090 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4091 int new_effective_prio
, policy
= attr
->sched_policy
;
4092 const struct sched_class
*prev_class
;
4095 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4098 /* The pi code expects interrupts enabled */
4099 BUG_ON(pi
&& in_interrupt());
4101 /* Double check policy once rq lock held: */
4103 reset_on_fork
= p
->sched_reset_on_fork
;
4104 policy
= oldpolicy
= p
->policy
;
4106 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4108 if (!valid_policy(policy
))
4112 if (attr
->sched_flags
&
4113 ~(SCHED_FLAG_RESET_ON_FORK
| SCHED_FLAG_RECLAIM
))
4117 * Valid priorities for SCHED_FIFO and SCHED_RR are
4118 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4119 * SCHED_BATCH and SCHED_IDLE is 0.
4121 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4122 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4124 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4125 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4129 * Allow unprivileged RT tasks to decrease priority:
4131 if (user
&& !capable(CAP_SYS_NICE
)) {
4132 if (fair_policy(policy
)) {
4133 if (attr
->sched_nice
< task_nice(p
) &&
4134 !can_nice(p
, attr
->sched_nice
))
4138 if (rt_policy(policy
)) {
4139 unsigned long rlim_rtprio
=
4140 task_rlimit(p
, RLIMIT_RTPRIO
);
4142 /* Can't set/change the rt policy: */
4143 if (policy
!= p
->policy
&& !rlim_rtprio
)
4146 /* Can't increase priority: */
4147 if (attr
->sched_priority
> p
->rt_priority
&&
4148 attr
->sched_priority
> rlim_rtprio
)
4153 * Can't set/change SCHED_DEADLINE policy at all for now
4154 * (safest behavior); in the future we would like to allow
4155 * unprivileged DL tasks to increase their relative deadline
4156 * or reduce their runtime (both ways reducing utilization)
4158 if (dl_policy(policy
))
4162 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4163 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4165 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4166 if (!can_nice(p
, task_nice(p
)))
4170 /* Can't change other user's priorities: */
4171 if (!check_same_owner(p
))
4174 /* Normal users shall not reset the sched_reset_on_fork flag: */
4175 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4180 retval
= security_task_setscheduler(p
);
4186 * Make sure no PI-waiters arrive (or leave) while we are
4187 * changing the priority of the task:
4189 * To be able to change p->policy safely, the appropriate
4190 * runqueue lock must be held.
4192 rq
= task_rq_lock(p
, &rf
);
4193 update_rq_clock(rq
);
4196 * Changing the policy of the stop threads its a very bad idea:
4198 if (p
== rq
->stop
) {
4199 task_rq_unlock(rq
, p
, &rf
);
4204 * If not changing anything there's no need to proceed further,
4205 * but store a possible modification of reset_on_fork.
4207 if (unlikely(policy
== p
->policy
)) {
4208 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4210 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4212 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4215 p
->sched_reset_on_fork
= reset_on_fork
;
4216 task_rq_unlock(rq
, p
, &rf
);
4222 #ifdef CONFIG_RT_GROUP_SCHED
4224 * Do not allow realtime tasks into groups that have no runtime
4227 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4228 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4229 !task_group_is_autogroup(task_group(p
))) {
4230 task_rq_unlock(rq
, p
, &rf
);
4235 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4236 cpumask_t
*span
= rq
->rd
->span
;
4239 * Don't allow tasks with an affinity mask smaller than
4240 * the entire root_domain to become SCHED_DEADLINE. We
4241 * will also fail if there's no bandwidth available.
4243 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4244 rq
->rd
->dl_bw
.bw
== 0) {
4245 task_rq_unlock(rq
, p
, &rf
);
4252 /* Re-check policy now with rq lock held: */
4253 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4254 policy
= oldpolicy
= -1;
4255 task_rq_unlock(rq
, p
, &rf
);
4260 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4261 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4264 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
4265 task_rq_unlock(rq
, p
, &rf
);
4269 p
->sched_reset_on_fork
= reset_on_fork
;
4274 * Take priority boosted tasks into account. If the new
4275 * effective priority is unchanged, we just store the new
4276 * normal parameters and do not touch the scheduler class and
4277 * the runqueue. This will be done when the task deboost
4280 new_effective_prio
= rt_effective_prio(p
, newprio
);
4281 if (new_effective_prio
== oldprio
)
4282 queue_flags
&= ~DEQUEUE_MOVE
;
4285 queued
= task_on_rq_queued(p
);
4286 running
= task_current(rq
, p
);
4288 dequeue_task(rq
, p
, queue_flags
);
4290 put_prev_task(rq
, p
);
4292 prev_class
= p
->sched_class
;
4293 __setscheduler(rq
, p
, attr
, pi
);
4297 * We enqueue to tail when the priority of a task is
4298 * increased (user space view).
4300 if (oldprio
< p
->prio
)
4301 queue_flags
|= ENQUEUE_HEAD
;
4303 enqueue_task(rq
, p
, queue_flags
);
4306 set_curr_task(rq
, p
);
4308 check_class_changed(rq
, p
, prev_class
, oldprio
);
4310 /* Avoid rq from going away on us: */
4312 task_rq_unlock(rq
, p
, &rf
);
4315 rt_mutex_adjust_pi(p
);
4317 /* Run balance callbacks after we've adjusted the PI chain: */
4318 balance_callback(rq
);
4324 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4325 const struct sched_param
*param
, bool check
)
4327 struct sched_attr attr
= {
4328 .sched_policy
= policy
,
4329 .sched_priority
= param
->sched_priority
,
4330 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4333 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4334 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4335 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4336 policy
&= ~SCHED_RESET_ON_FORK
;
4337 attr
.sched_policy
= policy
;
4340 return __sched_setscheduler(p
, &attr
, check
, true);
4343 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4344 * @p: the task in question.
4345 * @policy: new policy.
4346 * @param: structure containing the new RT priority.
4348 * Return: 0 on success. An error code otherwise.
4350 * NOTE that the task may be already dead.
4352 int sched_setscheduler(struct task_struct
*p
, int policy
,
4353 const struct sched_param
*param
)
4355 return _sched_setscheduler(p
, policy
, param
, true);
4357 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4359 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4361 return __sched_setscheduler(p
, attr
, true, true);
4363 EXPORT_SYMBOL_GPL(sched_setattr
);
4366 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4367 * @p: the task in question.
4368 * @policy: new policy.
4369 * @param: structure containing the new RT priority.
4371 * Just like sched_setscheduler, only don't bother checking if the
4372 * current context has permission. For example, this is needed in
4373 * stop_machine(): we create temporary high priority worker threads,
4374 * but our caller might not have that capability.
4376 * Return: 0 on success. An error code otherwise.
4378 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4379 const struct sched_param
*param
)
4381 return _sched_setscheduler(p
, policy
, param
, false);
4383 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4386 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4388 struct sched_param lparam
;
4389 struct task_struct
*p
;
4392 if (!param
|| pid
< 0)
4394 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4399 p
= find_process_by_pid(pid
);
4401 retval
= sched_setscheduler(p
, policy
, &lparam
);
4408 * Mimics kernel/events/core.c perf_copy_attr().
4410 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
4415 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4418 /* Zero the full structure, so that a short copy will be nice: */
4419 memset(attr
, 0, sizeof(*attr
));
4421 ret
= get_user(size
, &uattr
->size
);
4425 /* Bail out on silly large: */
4426 if (size
> PAGE_SIZE
)
4429 /* ABI compatibility quirk: */
4431 size
= SCHED_ATTR_SIZE_VER0
;
4433 if (size
< SCHED_ATTR_SIZE_VER0
)
4437 * If we're handed a bigger struct than we know of,
4438 * ensure all the unknown bits are 0 - i.e. new
4439 * user-space does not rely on any kernel feature
4440 * extensions we dont know about yet.
4442 if (size
> sizeof(*attr
)) {
4443 unsigned char __user
*addr
;
4444 unsigned char __user
*end
;
4447 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4448 end
= (void __user
*)uattr
+ size
;
4450 for (; addr
< end
; addr
++) {
4451 ret
= get_user(val
, addr
);
4457 size
= sizeof(*attr
);
4460 ret
= copy_from_user(attr
, uattr
, size
);
4465 * XXX: Do we want to be lenient like existing syscalls; or do we want
4466 * to be strict and return an error on out-of-bounds values?
4468 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4473 put_user(sizeof(*attr
), &uattr
->size
);
4478 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4479 * @pid: the pid in question.
4480 * @policy: new policy.
4481 * @param: structure containing the new RT priority.
4483 * Return: 0 on success. An error code otherwise.
4485 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
4490 return do_sched_setscheduler(pid
, policy
, param
);
4494 * sys_sched_setparam - set/change the RT priority of a thread
4495 * @pid: the pid in question.
4496 * @param: structure containing the new RT priority.
4498 * Return: 0 on success. An error code otherwise.
4500 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4502 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4506 * sys_sched_setattr - same as above, but with extended sched_attr
4507 * @pid: the pid in question.
4508 * @uattr: structure containing the extended parameters.
4509 * @flags: for future extension.
4511 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4512 unsigned int, flags
)
4514 struct sched_attr attr
;
4515 struct task_struct
*p
;
4518 if (!uattr
|| pid
< 0 || flags
)
4521 retval
= sched_copy_attr(uattr
, &attr
);
4525 if ((int)attr
.sched_policy
< 0)
4530 p
= find_process_by_pid(pid
);
4532 retval
= sched_setattr(p
, &attr
);
4539 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4540 * @pid: the pid in question.
4542 * Return: On success, the policy of the thread. Otherwise, a negative error
4545 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4547 struct task_struct
*p
;
4555 p
= find_process_by_pid(pid
);
4557 retval
= security_task_getscheduler(p
);
4560 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4567 * sys_sched_getparam - get the RT priority of a thread
4568 * @pid: the pid in question.
4569 * @param: structure containing the RT priority.
4571 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4574 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4576 struct sched_param lp
= { .sched_priority
= 0 };
4577 struct task_struct
*p
;
4580 if (!param
|| pid
< 0)
4584 p
= find_process_by_pid(pid
);
4589 retval
= security_task_getscheduler(p
);
4593 if (task_has_rt_policy(p
))
4594 lp
.sched_priority
= p
->rt_priority
;
4598 * This one might sleep, we cannot do it with a spinlock held ...
4600 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4609 static int sched_read_attr(struct sched_attr __user
*uattr
,
4610 struct sched_attr
*attr
,
4615 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4619 * If we're handed a smaller struct than we know of,
4620 * ensure all the unknown bits are 0 - i.e. old
4621 * user-space does not get uncomplete information.
4623 if (usize
< sizeof(*attr
)) {
4624 unsigned char *addr
;
4627 addr
= (void *)attr
+ usize
;
4628 end
= (void *)attr
+ sizeof(*attr
);
4630 for (; addr
< end
; addr
++) {
4638 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4646 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4647 * @pid: the pid in question.
4648 * @uattr: structure containing the extended parameters.
4649 * @size: sizeof(attr) for fwd/bwd comp.
4650 * @flags: for future extension.
4652 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4653 unsigned int, size
, unsigned int, flags
)
4655 struct sched_attr attr
= {
4656 .size
= sizeof(struct sched_attr
),
4658 struct task_struct
*p
;
4661 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4662 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4666 p
= find_process_by_pid(pid
);
4671 retval
= security_task_getscheduler(p
);
4675 attr
.sched_policy
= p
->policy
;
4676 if (p
->sched_reset_on_fork
)
4677 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4678 if (task_has_dl_policy(p
))
4679 __getparam_dl(p
, &attr
);
4680 else if (task_has_rt_policy(p
))
4681 attr
.sched_priority
= p
->rt_priority
;
4683 attr
.sched_nice
= task_nice(p
);
4687 retval
= sched_read_attr(uattr
, &attr
, size
);
4695 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4697 cpumask_var_t cpus_allowed
, new_mask
;
4698 struct task_struct
*p
;
4703 p
= find_process_by_pid(pid
);
4709 /* Prevent p going away */
4713 if (p
->flags
& PF_NO_SETAFFINITY
) {
4717 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4721 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4723 goto out_free_cpus_allowed
;
4726 if (!check_same_owner(p
)) {
4728 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4730 goto out_free_new_mask
;
4735 retval
= security_task_setscheduler(p
);
4737 goto out_free_new_mask
;
4740 cpuset_cpus_allowed(p
, cpus_allowed
);
4741 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4744 * Since bandwidth control happens on root_domain basis,
4745 * if admission test is enabled, we only admit -deadline
4746 * tasks allowed to run on all the CPUs in the task's
4750 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4752 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4755 goto out_free_new_mask
;
4761 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4764 cpuset_cpus_allowed(p
, cpus_allowed
);
4765 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4767 * We must have raced with a concurrent cpuset
4768 * update. Just reset the cpus_allowed to the
4769 * cpuset's cpus_allowed
4771 cpumask_copy(new_mask
, cpus_allowed
);
4776 free_cpumask_var(new_mask
);
4777 out_free_cpus_allowed
:
4778 free_cpumask_var(cpus_allowed
);
4784 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4785 struct cpumask
*new_mask
)
4787 if (len
< cpumask_size())
4788 cpumask_clear(new_mask
);
4789 else if (len
> cpumask_size())
4790 len
= cpumask_size();
4792 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4796 * sys_sched_setaffinity - set the CPU affinity of a process
4797 * @pid: pid of the process
4798 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4799 * @user_mask_ptr: user-space pointer to the new CPU mask
4801 * Return: 0 on success. An error code otherwise.
4803 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4804 unsigned long __user
*, user_mask_ptr
)
4806 cpumask_var_t new_mask
;
4809 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4812 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4814 retval
= sched_setaffinity(pid
, new_mask
);
4815 free_cpumask_var(new_mask
);
4819 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4821 struct task_struct
*p
;
4822 unsigned long flags
;
4828 p
= find_process_by_pid(pid
);
4832 retval
= security_task_getscheduler(p
);
4836 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4837 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4838 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4847 * sys_sched_getaffinity - get the CPU affinity of a process
4848 * @pid: pid of the process
4849 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4850 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4852 * Return: size of CPU mask copied to user_mask_ptr on success. An
4853 * error code otherwise.
4855 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4856 unsigned long __user
*, user_mask_ptr
)
4861 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4863 if (len
& (sizeof(unsigned long)-1))
4866 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4869 ret
= sched_getaffinity(pid
, mask
);
4871 size_t retlen
= min_t(size_t, len
, cpumask_size());
4873 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4878 free_cpumask_var(mask
);
4884 * sys_sched_yield - yield the current processor to other threads.
4886 * This function yields the current CPU to other tasks. If there are no
4887 * other threads running on this CPU then this function will return.
4891 SYSCALL_DEFINE0(sched_yield
)
4896 local_irq_disable();
4900 schedstat_inc(rq
->yld_count
);
4901 current
->sched_class
->yield_task(rq
);
4904 * Since we are going to call schedule() anyway, there's
4905 * no need to preempt or enable interrupts:
4909 sched_preempt_enable_no_resched();
4916 #ifndef CONFIG_PREEMPT
4917 int __sched
_cond_resched(void)
4919 if (should_resched(0)) {
4920 preempt_schedule_common();
4925 EXPORT_SYMBOL(_cond_resched
);
4929 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4930 * call schedule, and on return reacquire the lock.
4932 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4933 * operations here to prevent schedule() from being called twice (once via
4934 * spin_unlock(), once by hand).
4936 int __cond_resched_lock(spinlock_t
*lock
)
4938 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4941 lockdep_assert_held(lock
);
4943 if (spin_needbreak(lock
) || resched
) {
4946 preempt_schedule_common();
4954 EXPORT_SYMBOL(__cond_resched_lock
);
4956 int __sched
__cond_resched_softirq(void)
4958 BUG_ON(!in_softirq());
4960 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4962 preempt_schedule_common();
4968 EXPORT_SYMBOL(__cond_resched_softirq
);
4971 * yield - yield the current processor to other threads.
4973 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4975 * The scheduler is at all times free to pick the calling task as the most
4976 * eligible task to run, if removing the yield() call from your code breaks
4977 * it, its already broken.
4979 * Typical broken usage is:
4984 * where one assumes that yield() will let 'the other' process run that will
4985 * make event true. If the current task is a SCHED_FIFO task that will never
4986 * happen. Never use yield() as a progress guarantee!!
4988 * If you want to use yield() to wait for something, use wait_event().
4989 * If you want to use yield() to be 'nice' for others, use cond_resched().
4990 * If you still want to use yield(), do not!
4992 void __sched
yield(void)
4994 set_current_state(TASK_RUNNING
);
4997 EXPORT_SYMBOL(yield
);
5000 * yield_to - yield the current processor to another thread in
5001 * your thread group, or accelerate that thread toward the
5002 * processor it's on.
5004 * @preempt: whether task preemption is allowed or not
5006 * It's the caller's job to ensure that the target task struct
5007 * can't go away on us before we can do any checks.
5010 * true (>0) if we indeed boosted the target task.
5011 * false (0) if we failed to boost the target.
5012 * -ESRCH if there's no task to yield to.
5014 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5016 struct task_struct
*curr
= current
;
5017 struct rq
*rq
, *p_rq
;
5018 unsigned long flags
;
5021 local_irq_save(flags
);
5027 * If we're the only runnable task on the rq and target rq also
5028 * has only one task, there's absolutely no point in yielding.
5030 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5035 double_rq_lock(rq
, p_rq
);
5036 if (task_rq(p
) != p_rq
) {
5037 double_rq_unlock(rq
, p_rq
);
5041 if (!curr
->sched_class
->yield_to_task
)
5044 if (curr
->sched_class
!= p
->sched_class
)
5047 if (task_running(p_rq
, p
) || p
->state
)
5050 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5052 schedstat_inc(rq
->yld_count
);
5054 * Make p's CPU reschedule; pick_next_entity takes care of
5057 if (preempt
&& rq
!= p_rq
)
5062 double_rq_unlock(rq
, p_rq
);
5064 local_irq_restore(flags
);
5071 EXPORT_SYMBOL_GPL(yield_to
);
5073 int io_schedule_prepare(void)
5075 int old_iowait
= current
->in_iowait
;
5077 current
->in_iowait
= 1;
5078 blk_schedule_flush_plug(current
);
5083 void io_schedule_finish(int token
)
5085 current
->in_iowait
= token
;
5089 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5090 * that process accounting knows that this is a task in IO wait state.
5092 long __sched
io_schedule_timeout(long timeout
)
5097 token
= io_schedule_prepare();
5098 ret
= schedule_timeout(timeout
);
5099 io_schedule_finish(token
);
5103 EXPORT_SYMBOL(io_schedule_timeout
);
5105 void io_schedule(void)
5109 token
= io_schedule_prepare();
5111 io_schedule_finish(token
);
5113 EXPORT_SYMBOL(io_schedule
);
5116 * sys_sched_get_priority_max - return maximum RT priority.
5117 * @policy: scheduling class.
5119 * Return: On success, this syscall returns the maximum
5120 * rt_priority that can be used by a given scheduling class.
5121 * On failure, a negative error code is returned.
5123 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5130 ret
= MAX_USER_RT_PRIO
-1;
5132 case SCHED_DEADLINE
:
5143 * sys_sched_get_priority_min - return minimum RT priority.
5144 * @policy: scheduling class.
5146 * Return: On success, this syscall returns the minimum
5147 * rt_priority that can be used by a given scheduling class.
5148 * On failure, a negative error code is returned.
5150 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5159 case SCHED_DEADLINE
:
5169 * sys_sched_rr_get_interval - return the default timeslice of a process.
5170 * @pid: pid of the process.
5171 * @interval: userspace pointer to the timeslice value.
5173 * this syscall writes the default timeslice value of a given process
5174 * into the user-space timespec buffer. A value of '0' means infinity.
5176 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5179 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5180 struct timespec __user
*, interval
)
5182 struct task_struct
*p
;
5183 unsigned int time_slice
;
5194 p
= find_process_by_pid(pid
);
5198 retval
= security_task_getscheduler(p
);
5202 rq
= task_rq_lock(p
, &rf
);
5204 if (p
->sched_class
->get_rr_interval
)
5205 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5206 task_rq_unlock(rq
, p
, &rf
);
5209 jiffies_to_timespec(time_slice
, &t
);
5210 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5218 void sched_show_task(struct task_struct
*p
)
5220 unsigned long free
= 0;
5223 if (!try_get_task_stack(p
))
5226 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5228 if (p
->state
== TASK_RUNNING
)
5229 printk(KERN_CONT
" running task ");
5230 #ifdef CONFIG_DEBUG_STACK_USAGE
5231 free
= stack_not_used(p
);
5236 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5238 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5239 task_pid_nr(p
), ppid
,
5240 (unsigned long)task_thread_info(p
)->flags
);
5242 print_worker_info(KERN_INFO
, p
);
5243 show_stack(p
, NULL
);
5248 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
5250 /* no filter, everything matches */
5254 /* filter, but doesn't match */
5255 if (!(p
->state
& state_filter
))
5259 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5262 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
5269 void show_state_filter(unsigned long state_filter
)
5271 struct task_struct
*g
, *p
;
5273 #if BITS_PER_LONG == 32
5275 " task PC stack pid father\n");
5278 " task PC stack pid father\n");
5281 for_each_process_thread(g
, p
) {
5283 * reset the NMI-timeout, listing all files on a slow
5284 * console might take a lot of time:
5285 * Also, reset softlockup watchdogs on all CPUs, because
5286 * another CPU might be blocked waiting for us to process
5289 touch_nmi_watchdog();
5290 touch_all_softlockup_watchdogs();
5291 if (state_filter_match(state_filter
, p
))
5295 #ifdef CONFIG_SCHED_DEBUG
5297 sysrq_sched_debug_show();
5301 * Only show locks if all tasks are dumped:
5304 debug_show_all_locks();
5308 * init_idle - set up an idle thread for a given CPU
5309 * @idle: task in question
5310 * @cpu: CPU the idle task belongs to
5312 * NOTE: this function does not set the idle thread's NEED_RESCHED
5313 * flag, to make booting more robust.
5315 void init_idle(struct task_struct
*idle
, int cpu
)
5317 struct rq
*rq
= cpu_rq(cpu
);
5318 unsigned long flags
;
5320 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5321 raw_spin_lock(&rq
->lock
);
5323 __sched_fork(0, idle
);
5324 idle
->state
= TASK_RUNNING
;
5325 idle
->se
.exec_start
= sched_clock();
5326 idle
->flags
|= PF_IDLE
;
5328 kasan_unpoison_task_stack(idle
);
5332 * Its possible that init_idle() gets called multiple times on a task,
5333 * in that case do_set_cpus_allowed() will not do the right thing.
5335 * And since this is boot we can forgo the serialization.
5337 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5340 * We're having a chicken and egg problem, even though we are
5341 * holding rq->lock, the CPU isn't yet set to this CPU so the
5342 * lockdep check in task_group() will fail.
5344 * Similar case to sched_fork(). / Alternatively we could
5345 * use task_rq_lock() here and obtain the other rq->lock.
5350 __set_task_cpu(idle
, cpu
);
5353 rq
->curr
= rq
->idle
= idle
;
5354 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5358 raw_spin_unlock(&rq
->lock
);
5359 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5361 /* Set the preempt count _outside_ the spinlocks! */
5362 init_idle_preempt_count(idle
, cpu
);
5365 * The idle tasks have their own, simple scheduling class:
5367 idle
->sched_class
= &idle_sched_class
;
5368 ftrace_graph_init_idle_task(idle
, cpu
);
5369 vtime_init_idle(idle
, cpu
);
5371 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5377 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5378 const struct cpumask
*trial
)
5382 if (!cpumask_weight(cur
))
5385 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
5390 int task_can_attach(struct task_struct
*p
,
5391 const struct cpumask
*cs_cpus_allowed
)
5396 * Kthreads which disallow setaffinity shouldn't be moved
5397 * to a new cpuset; we don't want to change their CPU
5398 * affinity and isolating such threads by their set of
5399 * allowed nodes is unnecessary. Thus, cpusets are not
5400 * applicable for such threads. This prevents checking for
5401 * success of set_cpus_allowed_ptr() on all attached tasks
5402 * before cpus_allowed may be changed.
5404 if (p
->flags
& PF_NO_SETAFFINITY
) {
5409 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5411 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
5417 bool sched_smp_initialized __read_mostly
;
5419 #ifdef CONFIG_NUMA_BALANCING
5420 /* Migrate current task p to target_cpu */
5421 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5423 struct migration_arg arg
= { p
, target_cpu
};
5424 int curr_cpu
= task_cpu(p
);
5426 if (curr_cpu
== target_cpu
)
5429 if (!cpumask_test_cpu(target_cpu
, &p
->cpus_allowed
))
5432 /* TODO: This is not properly updating schedstats */
5434 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5435 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5439 * Requeue a task on a given node and accurately track the number of NUMA
5440 * tasks on the runqueues
5442 void sched_setnuma(struct task_struct
*p
, int nid
)
5444 bool queued
, running
;
5448 rq
= task_rq_lock(p
, &rf
);
5449 queued
= task_on_rq_queued(p
);
5450 running
= task_current(rq
, p
);
5453 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5455 put_prev_task(rq
, p
);
5457 p
->numa_preferred_nid
= nid
;
5460 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5462 set_curr_task(rq
, p
);
5463 task_rq_unlock(rq
, p
, &rf
);
5465 #endif /* CONFIG_NUMA_BALANCING */
5467 #ifdef CONFIG_HOTPLUG_CPU
5469 * Ensure that the idle task is using init_mm right before its CPU goes
5472 void idle_task_exit(void)
5474 struct mm_struct
*mm
= current
->active_mm
;
5476 BUG_ON(cpu_online(smp_processor_id()));
5478 if (mm
!= &init_mm
) {
5479 switch_mm(mm
, &init_mm
, current
);
5480 finish_arch_post_lock_switch();
5486 * Since this CPU is going 'away' for a while, fold any nr_active delta
5487 * we might have. Assumes we're called after migrate_tasks() so that the
5488 * nr_active count is stable. We need to take the teardown thread which
5489 * is calling this into account, so we hand in adjust = 1 to the load
5492 * Also see the comment "Global load-average calculations".
5494 static void calc_load_migrate(struct rq
*rq
)
5496 long delta
= calc_load_fold_active(rq
, 1);
5498 atomic_long_add(delta
, &calc_load_tasks
);
5501 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5505 static const struct sched_class fake_sched_class
= {
5506 .put_prev_task
= put_prev_task_fake
,
5509 static struct task_struct fake_task
= {
5511 * Avoid pull_{rt,dl}_task()
5513 .prio
= MAX_PRIO
+ 1,
5514 .sched_class
= &fake_sched_class
,
5518 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5519 * try_to_wake_up()->select_task_rq().
5521 * Called with rq->lock held even though we'er in stop_machine() and
5522 * there's no concurrency possible, we hold the required locks anyway
5523 * because of lock validation efforts.
5525 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
5527 struct rq
*rq
= dead_rq
;
5528 struct task_struct
*next
, *stop
= rq
->stop
;
5529 struct rq_flags orf
= *rf
;
5533 * Fudge the rq selection such that the below task selection loop
5534 * doesn't get stuck on the currently eligible stop task.
5536 * We're currently inside stop_machine() and the rq is either stuck
5537 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5538 * either way we should never end up calling schedule() until we're
5544 * put_prev_task() and pick_next_task() sched
5545 * class method both need to have an up-to-date
5546 * value of rq->clock[_task]
5548 update_rq_clock(rq
);
5552 * There's this thread running, bail when that's the only
5555 if (rq
->nr_running
== 1)
5559 * pick_next_task() assumes pinned rq->lock:
5561 next
= pick_next_task(rq
, &fake_task
, rf
);
5563 put_prev_task(rq
, next
);
5566 * Rules for changing task_struct::cpus_allowed are holding
5567 * both pi_lock and rq->lock, such that holding either
5568 * stabilizes the mask.
5570 * Drop rq->lock is not quite as disastrous as it usually is
5571 * because !cpu_active at this point, which means load-balance
5572 * will not interfere. Also, stop-machine.
5575 raw_spin_lock(&next
->pi_lock
);
5579 * Since we're inside stop-machine, _nothing_ should have
5580 * changed the task, WARN if weird stuff happened, because in
5581 * that case the above rq->lock drop is a fail too.
5583 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5584 raw_spin_unlock(&next
->pi_lock
);
5588 /* Find suitable destination for @next, with force if needed. */
5589 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5590 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
5591 if (rq
!= dead_rq
) {
5597 raw_spin_unlock(&next
->pi_lock
);
5602 #endif /* CONFIG_HOTPLUG_CPU */
5604 void set_rq_online(struct rq
*rq
)
5607 const struct sched_class
*class;
5609 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5612 for_each_class(class) {
5613 if (class->rq_online
)
5614 class->rq_online(rq
);
5619 void set_rq_offline(struct rq
*rq
)
5622 const struct sched_class
*class;
5624 for_each_class(class) {
5625 if (class->rq_offline
)
5626 class->rq_offline(rq
);
5629 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5634 static void set_cpu_rq_start_time(unsigned int cpu
)
5636 struct rq
*rq
= cpu_rq(cpu
);
5638 rq
->age_stamp
= sched_clock_cpu(cpu
);
5642 * used to mark begin/end of suspend/resume:
5644 static int num_cpus_frozen
;
5647 * Update cpusets according to cpu_active mask. If cpusets are
5648 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5649 * around partition_sched_domains().
5651 * If we come here as part of a suspend/resume, don't touch cpusets because we
5652 * want to restore it back to its original state upon resume anyway.
5654 static void cpuset_cpu_active(void)
5656 if (cpuhp_tasks_frozen
) {
5658 * num_cpus_frozen tracks how many CPUs are involved in suspend
5659 * resume sequence. As long as this is not the last online
5660 * operation in the resume sequence, just build a single sched
5661 * domain, ignoring cpusets.
5663 partition_sched_domains(1, NULL
, NULL
);
5664 if (--num_cpus_frozen
)
5667 * This is the last CPU online operation. So fall through and
5668 * restore the original sched domains by considering the
5669 * cpuset configurations.
5671 cpuset_force_rebuild();
5673 cpuset_update_active_cpus();
5676 static int cpuset_cpu_inactive(unsigned int cpu
)
5678 if (!cpuhp_tasks_frozen
) {
5679 if (dl_cpu_busy(cpu
))
5681 cpuset_update_active_cpus();
5684 partition_sched_domains(1, NULL
, NULL
);
5689 int sched_cpu_activate(unsigned int cpu
)
5691 struct rq
*rq
= cpu_rq(cpu
);
5694 #ifdef CONFIG_SCHED_SMT
5696 * When going up, increment the number of cores with SMT present.
5698 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
5699 static_branch_inc_cpuslocked(&sched_smt_present
);
5701 set_cpu_active(cpu
, true);
5703 if (sched_smp_initialized
) {
5704 sched_domains_numa_masks_set(cpu
);
5705 cpuset_cpu_active();
5709 * Put the rq online, if not already. This happens:
5711 * 1) In the early boot process, because we build the real domains
5712 * after all CPUs have been brought up.
5714 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5717 rq_lock_irqsave(rq
, &rf
);
5719 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5722 rq_unlock_irqrestore(rq
, &rf
);
5724 update_max_interval();
5729 int sched_cpu_deactivate(unsigned int cpu
)
5733 set_cpu_active(cpu
, false);
5735 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5736 * users of this state to go away such that all new such users will
5739 * Do sync before park smpboot threads to take care the rcu boost case.
5741 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
5743 #ifdef CONFIG_SCHED_SMT
5745 * When going down, decrement the number of cores with SMT present.
5747 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
5748 static_branch_dec_cpuslocked(&sched_smt_present
);
5751 if (!sched_smp_initialized
)
5754 ret
= cpuset_cpu_inactive(cpu
);
5756 set_cpu_active(cpu
, true);
5759 sched_domains_numa_masks_clear(cpu
);
5763 static void sched_rq_cpu_starting(unsigned int cpu
)
5765 struct rq
*rq
= cpu_rq(cpu
);
5767 rq
->calc_load_update
= calc_load_update
;
5768 update_max_interval();
5771 int sched_cpu_starting(unsigned int cpu
)
5773 set_cpu_rq_start_time(cpu
);
5774 sched_rq_cpu_starting(cpu
);
5778 #ifdef CONFIG_HOTPLUG_CPU
5779 int sched_cpu_dying(unsigned int cpu
)
5781 struct rq
*rq
= cpu_rq(cpu
);
5784 /* Handle pending wakeups and then migrate everything off */
5785 sched_ttwu_pending();
5787 rq_lock_irqsave(rq
, &rf
);
5789 walt_migrate_sync_cpu(cpu
);
5792 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5795 migrate_tasks(rq
, &rf
);
5796 BUG_ON(rq
->nr_running
!= 1);
5797 rq_unlock_irqrestore(rq
, &rf
);
5799 calc_load_migrate(rq
);
5800 update_max_interval();
5801 nohz_balance_exit_idle(cpu
);
5807 void __init
sched_init_smp(void)
5809 cpumask_var_t non_isolated_cpus
;
5811 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
5816 * There's no userspace yet to cause hotplug operations; hence all the
5817 * CPU masks are stable and all blatant races in the below code cannot
5818 * happen. The hotplug lock is nevertheless taken to satisfy lockdep,
5819 * but there won't be any contention on it.
5822 mutex_lock(&sched_domains_mutex
);
5823 sched_init_domains(cpu_active_mask
);
5824 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
5825 if (cpumask_empty(non_isolated_cpus
))
5826 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
5827 mutex_unlock(&sched_domains_mutex
);
5830 /* Move init over to a non-isolated CPU */
5831 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
5833 sched_init_granularity();
5834 free_cpumask_var(non_isolated_cpus
);
5836 init_sched_rt_class();
5837 init_sched_dl_class();
5839 sched_smp_initialized
= true;
5842 static int __init
migration_init(void)
5844 sched_rq_cpu_starting(smp_processor_id());
5847 early_initcall(migration_init
);
5850 void __init
sched_init_smp(void)
5852 sched_init_granularity();
5854 #endif /* CONFIG_SMP */
5856 int in_sched_functions(unsigned long addr
)
5858 return in_lock_functions(addr
) ||
5859 (addr
>= (unsigned long)__sched_text_start
5860 && addr
< (unsigned long)__sched_text_end
);
5863 #ifdef CONFIG_CGROUP_SCHED
5865 * Default task group.
5866 * Every task in system belongs to this group at bootup.
5868 struct task_group root_task_group
;
5869 LIST_HEAD(task_groups
);
5871 /* Cacheline aligned slab cache for task_group */
5872 static struct kmem_cache
*task_group_cache __read_mostly
;
5875 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5876 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5878 void __init
sched_init(void)
5881 unsigned long alloc_size
= 0, ptr
;
5886 #ifdef CONFIG_FAIR_GROUP_SCHED
5887 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5889 #ifdef CONFIG_RT_GROUP_SCHED
5890 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5893 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
5895 #ifdef CONFIG_FAIR_GROUP_SCHED
5896 root_task_group
.se
= (struct sched_entity
**)ptr
;
5897 ptr
+= nr_cpu_ids
* sizeof(void **);
5899 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
5900 ptr
+= nr_cpu_ids
* sizeof(void **);
5902 #endif /* CONFIG_FAIR_GROUP_SCHED */
5903 #ifdef CONFIG_RT_GROUP_SCHED
5904 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
5905 ptr
+= nr_cpu_ids
* sizeof(void **);
5907 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
5908 ptr
+= nr_cpu_ids
* sizeof(void **);
5910 #endif /* CONFIG_RT_GROUP_SCHED */
5912 #ifdef CONFIG_CPUMASK_OFFSTACK
5913 for_each_possible_cpu(i
) {
5914 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5915 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5916 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5917 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5919 #endif /* CONFIG_CPUMASK_OFFSTACK */
5921 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
5922 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
5925 init_defrootdomain();
5928 #ifdef CONFIG_RT_GROUP_SCHED
5929 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
5930 global_rt_period(), global_rt_runtime());
5931 #endif /* CONFIG_RT_GROUP_SCHED */
5933 #ifdef CONFIG_CGROUP_SCHED
5934 task_group_cache
= KMEM_CACHE(task_group
, 0);
5936 list_add(&root_task_group
.list
, &task_groups
);
5937 INIT_LIST_HEAD(&root_task_group
.children
);
5938 INIT_LIST_HEAD(&root_task_group
.siblings
);
5939 autogroup_init(&init_task
);
5940 #endif /* CONFIG_CGROUP_SCHED */
5942 for_each_possible_cpu(i
) {
5946 raw_spin_lock_init(&rq
->lock
);
5948 rq
->calc_load_active
= 0;
5949 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
5950 init_cfs_rq(&rq
->cfs
);
5951 init_rt_rq(&rq
->rt
);
5952 init_dl_rq(&rq
->dl
);
5953 #ifdef CONFIG_FAIR_GROUP_SCHED
5954 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
5955 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
5956 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
5958 * How much CPU bandwidth does root_task_group get?
5960 * In case of task-groups formed thr' the cgroup filesystem, it
5961 * gets 100% of the CPU resources in the system. This overall
5962 * system CPU resource is divided among the tasks of
5963 * root_task_group and its child task-groups in a fair manner,
5964 * based on each entity's (task or task-group's) weight
5965 * (se->load.weight).
5967 * In other words, if root_task_group has 10 tasks of weight
5968 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5969 * then A0's share of the CPU resource is:
5971 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
5973 * We achieve this by letting root_task_group's tasks sit
5974 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
5976 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
5977 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
5978 #endif /* CONFIG_FAIR_GROUP_SCHED */
5980 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
5981 #ifdef CONFIG_RT_GROUP_SCHED
5982 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
5985 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
5986 rq
->cpu_load
[j
] = 0;
5991 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
5992 rq
->balance_callback
= NULL
;
5993 rq
->active_balance
= 0;
5994 rq
->next_balance
= jiffies
;
5999 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6000 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6001 #ifdef CONFIG_SCHED_WALT
6002 rq
->cur_irqload
= 0;
6003 rq
->avg_irqload
= 0;
6007 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6009 rq_attach_root(rq
, &def_root_domain
);
6010 #ifdef CONFIG_NO_HZ_COMMON
6011 rq
->last_load_update_tick
= jiffies
;
6012 rq
->last_blocked_load_update_tick
= jiffies
;
6015 #ifdef CONFIG_NO_HZ_FULL
6016 rq
->last_sched_tick
= 0;
6018 #endif /* CONFIG_SMP */
6020 atomic_set(&rq
->nr_iowait
, 0);
6023 set_load_weight(&init_task
);
6028 * The boot idle thread does lazy MMU switching as well:
6031 enter_lazy_tlb(&init_mm
, current
);
6034 * Make us the idle thread. Technically, schedule() should not be
6035 * called from this thread, however somewhere below it might be,
6036 * but because we are the idle thread, we just pick up running again
6037 * when this runqueue becomes "idle".
6039 init_idle(current
, smp_processor_id());
6041 calc_load_update
= jiffies
+ LOAD_FREQ
;
6044 /* May be allocated at isolcpus cmdline parse time */
6045 if (cpu_isolated_map
== NULL
)
6046 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6047 idle_thread_set_boot_cpu();
6048 set_cpu_rq_start_time(smp_processor_id());
6050 init_sched_fair_class();
6054 scheduler_running
= 1;
6057 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6058 static inline int preempt_count_equals(int preempt_offset
)
6060 int nested
= preempt_count() + rcu_preempt_depth();
6062 return (nested
== preempt_offset
);
6065 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6068 * Blocking primitives will set (and therefore destroy) current->state,
6069 * since we will exit with TASK_RUNNING make sure we enter with it,
6070 * otherwise we will destroy state.
6072 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6073 "do not call blocking ops when !TASK_RUNNING; "
6074 "state=%lx set at [<%p>] %pS\n",
6076 (void *)current
->task_state_change
,
6077 (void *)current
->task_state_change
);
6079 ___might_sleep(file
, line
, preempt_offset
);
6081 EXPORT_SYMBOL(__might_sleep
);
6083 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6085 /* Ratelimiting timestamp: */
6086 static unsigned long prev_jiffy
;
6088 unsigned long preempt_disable_ip
;
6090 /* WARN_ON_ONCE() by default, no rate limit required: */
6093 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6094 !is_idle_task(current
)) ||
6095 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6099 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6101 prev_jiffy
= jiffies
;
6103 /* Save this before calling printk(), since that will clobber it: */
6104 preempt_disable_ip
= get_preempt_disable_ip(current
);
6107 "BUG: sleeping function called from invalid context at %s:%d\n",
6110 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6111 in_atomic(), irqs_disabled(),
6112 current
->pid
, current
->comm
);
6114 if (task_stack_end_corrupted(current
))
6115 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6117 debug_show_held_locks(current
);
6118 if (irqs_disabled())
6119 print_irqtrace_events(current
);
6120 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6121 && !preempt_count_equals(preempt_offset
)) {
6122 pr_err("Preemption disabled at:");
6123 print_ip_sym(preempt_disable_ip
);
6127 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6129 EXPORT_SYMBOL(___might_sleep
);
6132 #ifdef CONFIG_MAGIC_SYSRQ
6133 void normalize_rt_tasks(void)
6135 struct task_struct
*g
, *p
;
6136 struct sched_attr attr
= {
6137 .sched_policy
= SCHED_NORMAL
,
6140 read_lock(&tasklist_lock
);
6141 for_each_process_thread(g
, p
) {
6143 * Only normalize user tasks:
6145 if (p
->flags
& PF_KTHREAD
)
6148 p
->se
.exec_start
= 0;
6149 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6150 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6151 schedstat_set(p
->se
.statistics
.block_start
, 0);
6153 if (!dl_task(p
) && !rt_task(p
)) {
6155 * Renice negative nice level userspace
6158 if (task_nice(p
) < 0)
6159 set_user_nice(p
, 0);
6163 __sched_setscheduler(p
, &attr
, false, false);
6165 read_unlock(&tasklist_lock
);
6168 #endif /* CONFIG_MAGIC_SYSRQ */
6170 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6172 * These functions are only useful for the IA64 MCA handling, or kdb.
6174 * They can only be called when the whole system has been
6175 * stopped - every CPU needs to be quiescent, and no scheduling
6176 * activity can take place. Using them for anything else would
6177 * be a serious bug, and as a result, they aren't even visible
6178 * under any other configuration.
6182 * curr_task - return the current task for a given CPU.
6183 * @cpu: the processor in question.
6185 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6187 * Return: The current task for @cpu.
6189 struct task_struct
*curr_task(int cpu
)
6191 return cpu_curr(cpu
);
6194 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6198 * set_curr_task - set the current task for a given CPU.
6199 * @cpu: the processor in question.
6200 * @p: the task pointer to set.
6202 * Description: This function must only be used when non-maskable interrupts
6203 * are serviced on a separate stack. It allows the architecture to switch the
6204 * notion of the current task on a CPU in a non-blocking manner. This function
6205 * must be called with all CPU's synchronized, and interrupts disabled, the
6206 * and caller must save the original value of the current task (see
6207 * curr_task() above) and restore that value before reenabling interrupts and
6208 * re-starting the system.
6210 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6212 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6219 #ifdef CONFIG_CGROUP_SCHED
6220 /* task_group_lock serializes the addition/removal of task groups */
6221 static DEFINE_SPINLOCK(task_group_lock
);
6223 static void sched_free_group(struct task_group
*tg
)
6225 free_fair_sched_group(tg
);
6226 free_rt_sched_group(tg
);
6228 kmem_cache_free(task_group_cache
, tg
);
6231 /* allocate runqueue etc for a new task group */
6232 struct task_group
*sched_create_group(struct task_group
*parent
)
6234 struct task_group
*tg
;
6236 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6238 return ERR_PTR(-ENOMEM
);
6240 if (!alloc_fair_sched_group(tg
, parent
))
6243 if (!alloc_rt_sched_group(tg
, parent
))
6249 sched_free_group(tg
);
6250 return ERR_PTR(-ENOMEM
);
6253 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6255 unsigned long flags
;
6257 spin_lock_irqsave(&task_group_lock
, flags
);
6258 list_add_rcu(&tg
->list
, &task_groups
);
6260 /* Root should already exist: */
6263 tg
->parent
= parent
;
6264 INIT_LIST_HEAD(&tg
->children
);
6265 list_add_rcu(&tg
->siblings
, &parent
->children
);
6266 spin_unlock_irqrestore(&task_group_lock
, flags
);
6268 online_fair_sched_group(tg
);
6271 /* rcu callback to free various structures associated with a task group */
6272 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6274 /* Now it should be safe to free those cfs_rqs: */
6275 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6278 void sched_destroy_group(struct task_group
*tg
)
6280 /* Wait for possible concurrent references to cfs_rqs complete: */
6281 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6284 void sched_offline_group(struct task_group
*tg
)
6286 unsigned long flags
;
6288 /* End participation in shares distribution: */
6289 unregister_fair_sched_group(tg
);
6291 spin_lock_irqsave(&task_group_lock
, flags
);
6292 list_del_rcu(&tg
->list
);
6293 list_del_rcu(&tg
->siblings
);
6294 spin_unlock_irqrestore(&task_group_lock
, flags
);
6297 static void sched_change_group(struct task_struct
*tsk
, int type
)
6299 struct task_group
*tg
;
6302 * All callers are synchronized by task_rq_lock(); we do not use RCU
6303 * which is pointless here. Thus, we pass "true" to task_css_check()
6304 * to prevent lockdep warnings.
6306 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
6307 struct task_group
, css
);
6308 tg
= autogroup_task_group(tsk
, tg
);
6309 tsk
->sched_task_group
= tg
;
6311 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
6312 if (tsk
->sched_class
->task_change_group
)
6313 tsk
->sched_class
->task_change_group(tsk
, type
);
6316 set_task_rq(tsk
, task_cpu(tsk
));
6320 * Change task's runqueue when it moves between groups.
6322 * The caller of this function should have put the task in its new group by
6323 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6326 void sched_move_task(struct task_struct
*tsk
)
6328 int queued
, running
, queue_flags
=
6329 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
6333 rq
= task_rq_lock(tsk
, &rf
);
6334 update_rq_clock(rq
);
6336 running
= task_current(rq
, tsk
);
6337 queued
= task_on_rq_queued(tsk
);
6340 dequeue_task(rq
, tsk
, queue_flags
);
6342 put_prev_task(rq
, tsk
);
6344 sched_change_group(tsk
, TASK_MOVE_GROUP
);
6347 enqueue_task(rq
, tsk
, queue_flags
);
6349 set_curr_task(rq
, tsk
);
6351 task_rq_unlock(rq
, tsk
, &rf
);
6354 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
6356 return css
? container_of(css
, struct task_group
, css
) : NULL
;
6359 static struct cgroup_subsys_state
*
6360 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6362 struct task_group
*parent
= css_tg(parent_css
);
6363 struct task_group
*tg
;
6366 /* This is early initialization for the top cgroup */
6367 return &root_task_group
.css
;
6370 tg
= sched_create_group(parent
);
6372 return ERR_PTR(-ENOMEM
);
6377 /* Expose task group only after completing cgroup initialization */
6378 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
6380 struct task_group
*tg
= css_tg(css
);
6381 struct task_group
*parent
= css_tg(css
->parent
);
6384 sched_online_group(tg
, parent
);
6388 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
6390 struct task_group
*tg
= css_tg(css
);
6392 sched_offline_group(tg
);
6395 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
6397 struct task_group
*tg
= css_tg(css
);
6400 * Relies on the RCU grace period between css_released() and this.
6402 sched_free_group(tg
);
6406 * This is called before wake_up_new_task(), therefore we really only
6407 * have to set its group bits, all the other stuff does not apply.
6409 static void cpu_cgroup_fork(struct task_struct
*task
)
6414 rq
= task_rq_lock(task
, &rf
);
6416 update_rq_clock(rq
);
6417 sched_change_group(task
, TASK_SET_GROUP
);
6419 task_rq_unlock(rq
, task
, &rf
);
6422 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
6424 struct task_struct
*task
;
6425 struct cgroup_subsys_state
*css
;
6428 cgroup_taskset_for_each(task
, css
, tset
) {
6429 #ifdef CONFIG_RT_GROUP_SCHED
6430 if (!sched_rt_can_attach(css_tg(css
), task
))
6433 /* We don't support RT-tasks being in separate groups */
6434 if (task
->sched_class
!= &fair_sched_class
)
6438 * Serialize against wake_up_new_task() such that if its
6439 * running, we're sure to observe its full state.
6441 raw_spin_lock_irq(&task
->pi_lock
);
6443 * Avoid calling sched_move_task() before wake_up_new_task()
6444 * has happened. This would lead to problems with PELT, due to
6445 * move wanting to detach+attach while we're not attached yet.
6447 if (task
->state
== TASK_NEW
)
6449 raw_spin_unlock_irq(&task
->pi_lock
);
6457 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
6459 struct task_struct
*task
;
6460 struct cgroup_subsys_state
*css
;
6462 cgroup_taskset_for_each(task
, css
, tset
)
6463 sched_move_task(task
);
6466 #ifdef CONFIG_FAIR_GROUP_SCHED
6467 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
6468 struct cftype
*cftype
, u64 shareval
)
6470 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
6473 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
6476 struct task_group
*tg
= css_tg(css
);
6478 return (u64
) scale_load_down(tg
->shares
);
6481 #ifdef CONFIG_CFS_BANDWIDTH
6482 static DEFINE_MUTEX(cfs_constraints_mutex
);
6484 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
6485 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
6487 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
6489 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
6491 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
6492 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6494 if (tg
== &root_task_group
)
6498 * Ensure we have at some amount of bandwidth every period. This is
6499 * to prevent reaching a state of large arrears when throttled via
6500 * entity_tick() resulting in prolonged exit starvation.
6502 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
6506 * Likewise, bound things on the otherside by preventing insane quota
6507 * periods. This also allows us to normalize in computing quota
6510 if (period
> max_cfs_quota_period
)
6514 * Prevent race between setting of cfs_rq->runtime_enabled and
6515 * unthrottle_offline_cfs_rqs().
6518 mutex_lock(&cfs_constraints_mutex
);
6519 ret
= __cfs_schedulable(tg
, period
, quota
);
6523 runtime_enabled
= quota
!= RUNTIME_INF
;
6524 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
6526 * If we need to toggle cfs_bandwidth_used, off->on must occur
6527 * before making related changes, and on->off must occur afterwards
6529 if (runtime_enabled
&& !runtime_was_enabled
)
6530 cfs_bandwidth_usage_inc();
6531 raw_spin_lock_irq(&cfs_b
->lock
);
6532 cfs_b
->period
= ns_to_ktime(period
);
6533 cfs_b
->quota
= quota
;
6535 __refill_cfs_bandwidth_runtime(cfs_b
);
6537 /* Restart the period timer (if active) to handle new period expiry: */
6538 if (runtime_enabled
)
6539 start_cfs_bandwidth(cfs_b
);
6541 raw_spin_unlock_irq(&cfs_b
->lock
);
6543 for_each_online_cpu(i
) {
6544 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
6545 struct rq
*rq
= cfs_rq
->rq
;
6548 rq_lock_irq(rq
, &rf
);
6549 cfs_rq
->runtime_enabled
= runtime_enabled
;
6550 cfs_rq
->runtime_remaining
= 0;
6552 if (cfs_rq
->throttled
)
6553 unthrottle_cfs_rq(cfs_rq
);
6554 rq_unlock_irq(rq
, &rf
);
6556 if (runtime_was_enabled
&& !runtime_enabled
)
6557 cfs_bandwidth_usage_dec();
6559 mutex_unlock(&cfs_constraints_mutex
);
6565 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
6569 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6570 if (cfs_quota_us
< 0)
6571 quota
= RUNTIME_INF
;
6573 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
6575 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6578 long tg_get_cfs_quota(struct task_group
*tg
)
6582 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
6585 quota_us
= tg
->cfs_bandwidth
.quota
;
6586 do_div(quota_us
, NSEC_PER_USEC
);
6591 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
6595 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
6596 quota
= tg
->cfs_bandwidth
.quota
;
6598 return tg_set_cfs_bandwidth(tg
, period
, quota
);
6601 long tg_get_cfs_period(struct task_group
*tg
)
6605 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
6606 do_div(cfs_period_us
, NSEC_PER_USEC
);
6608 return cfs_period_us
;
6611 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
6614 return tg_get_cfs_quota(css_tg(css
));
6617 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
6618 struct cftype
*cftype
, s64 cfs_quota_us
)
6620 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
6623 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
6626 return tg_get_cfs_period(css_tg(css
));
6629 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
6630 struct cftype
*cftype
, u64 cfs_period_us
)
6632 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
6635 struct cfs_schedulable_data
{
6636 struct task_group
*tg
;
6641 * normalize group quota/period to be quota/max_period
6642 * note: units are usecs
6644 static u64
normalize_cfs_quota(struct task_group
*tg
,
6645 struct cfs_schedulable_data
*d
)
6653 period
= tg_get_cfs_period(tg
);
6654 quota
= tg_get_cfs_quota(tg
);
6657 /* note: these should typically be equivalent */
6658 if (quota
== RUNTIME_INF
|| quota
== -1)
6661 return to_ratio(period
, quota
);
6664 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
6666 struct cfs_schedulable_data
*d
= data
;
6667 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6668 s64 quota
= 0, parent_quota
= -1;
6671 quota
= RUNTIME_INF
;
6673 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
6675 quota
= normalize_cfs_quota(tg
, d
);
6676 parent_quota
= parent_b
->hierarchical_quota
;
6679 * Ensure max(child_quota) <= parent_quota, inherit when no
6682 if (quota
== RUNTIME_INF
)
6683 quota
= parent_quota
;
6684 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
6687 cfs_b
->hierarchical_quota
= quota
;
6692 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
6695 struct cfs_schedulable_data data
= {
6701 if (quota
!= RUNTIME_INF
) {
6702 do_div(data
.period
, NSEC_PER_USEC
);
6703 do_div(data
.quota
, NSEC_PER_USEC
);
6707 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
6713 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
6715 struct task_group
*tg
= css_tg(seq_css(sf
));
6716 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6718 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
6719 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
6720 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
6724 #endif /* CONFIG_CFS_BANDWIDTH */
6725 #endif /* CONFIG_FAIR_GROUP_SCHED */
6727 #ifdef CONFIG_RT_GROUP_SCHED
6728 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
6729 struct cftype
*cft
, s64 val
)
6731 return sched_group_set_rt_runtime(css_tg(css
), val
);
6734 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
6737 return sched_group_rt_runtime(css_tg(css
));
6740 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
6741 struct cftype
*cftype
, u64 rt_period_us
)
6743 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
6746 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
6749 return sched_group_rt_period(css_tg(css
));
6751 #endif /* CONFIG_RT_GROUP_SCHED */
6753 static struct cftype cpu_files
[] = {
6754 #ifdef CONFIG_FAIR_GROUP_SCHED
6757 .read_u64
= cpu_shares_read_u64
,
6758 .write_u64
= cpu_shares_write_u64
,
6761 #ifdef CONFIG_CFS_BANDWIDTH
6763 .name
= "cfs_quota_us",
6764 .read_s64
= cpu_cfs_quota_read_s64
,
6765 .write_s64
= cpu_cfs_quota_write_s64
,
6768 .name
= "cfs_period_us",
6769 .read_u64
= cpu_cfs_period_read_u64
,
6770 .write_u64
= cpu_cfs_period_write_u64
,
6774 .seq_show
= cpu_stats_show
,
6777 #ifdef CONFIG_RT_GROUP_SCHED
6779 .name
= "rt_runtime_us",
6780 .read_s64
= cpu_rt_runtime_read
,
6781 .write_s64
= cpu_rt_runtime_write
,
6784 .name
= "rt_period_us",
6785 .read_u64
= cpu_rt_period_read_uint
,
6786 .write_u64
= cpu_rt_period_write_uint
,
6792 struct cgroup_subsys cpu_cgrp_subsys
= {
6793 .css_alloc
= cpu_cgroup_css_alloc
,
6794 .css_online
= cpu_cgroup_css_online
,
6795 .css_released
= cpu_cgroup_css_released
,
6796 .css_free
= cpu_cgroup_css_free
,
6797 .fork
= cpu_cgroup_fork
,
6798 .can_attach
= cpu_cgroup_can_attach
,
6799 .attach
= cpu_cgroup_attach
,
6800 .legacy_cftypes
= cpu_files
,
6804 #endif /* CONFIG_CGROUP_SCHED */
6806 void dump_cpu_task(int cpu
)
6808 pr_info("Task dump for CPU %d:\n", cpu
);
6809 sched_show_task(cpu_curr(cpu
));
6813 * Nice levels are multiplicative, with a gentle 10% change for every
6814 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
6815 * nice 1, it will get ~10% less CPU time than another CPU-bound task
6816 * that remained on nice 0.
6818 * The "10% effect" is relative and cumulative: from _any_ nice level,
6819 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
6820 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
6821 * If a task goes up by ~10% and another task goes down by ~10% then
6822 * the relative distance between them is ~25%.)
6824 const int sched_prio_to_weight
[40] = {
6825 /* -20 */ 88761, 71755, 56483, 46273, 36291,
6826 /* -15 */ 29154, 23254, 18705, 14949, 11916,
6827 /* -10 */ 9548, 7620, 6100, 4904, 3906,
6828 /* -5 */ 3121, 2501, 1991, 1586, 1277,
6829 /* 0 */ 1024, 820, 655, 526, 423,
6830 /* 5 */ 335, 272, 215, 172, 137,
6831 /* 10 */ 110, 87, 70, 56, 45,
6832 /* 15 */ 36, 29, 23, 18, 15,
6836 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
6838 * In cases where the weight does not change often, we can use the
6839 * precalculated inverse to speed up arithmetics by turning divisions
6840 * into multiplications:
6842 const u32 sched_prio_to_wmult
[40] = {
6843 /* -20 */ 48388, 59856, 76040, 92818, 118348,
6844 /* -15 */ 147320, 184698, 229616, 287308, 360437,
6845 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
6846 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
6847 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
6848 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
6849 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
6850 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
6854 * RT Extension for 'prio_to_weight'
6856 const int rtprio_to_weight
[51] = {
6857 /* 0 */ 17222521, 15500269, 13950242, 12555218, 11299696,
6858 /* 10 */ 10169726, 9152754, 8237478, 7413730, 6672357,
6859 /* 20 */ 6005122, 5404609, 4864149, 4377734, 3939960,
6860 /* 30 */ 3545964, 3191368, 2872231, 2585008, 2326507,
6861 /* 40 */ 2093856, 1884471, 1696024, 1526421, 1373779,
6862 /* 50 */ 1236401, 1112761, 1001485, 901337, 811203,
6863 /* 60 */ 730083, 657074, 591367, 532230, 479007,
6864 /* 70 */ 431106, 387996, 349196, 314277, 282849,
6865 /* 80 */ 254564, 229108, 206197, 185577, 167019,
6866 /* 90 */ 150318, 135286, 121757, 109581, 98623,
6867 /* 100 for Fair class */ 88761,