1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
11 #include <linux/nospec.h>
13 #include <linux/kcov.h>
14 #include <linux/scs.h>
16 #include <asm/switch_to.h>
19 #include "../workqueue_internal.h"
20 #include "../../fs/io-wq.h"
21 #include "../smpboot.h"
26 #define CREATE_TRACE_POINTS
27 #include <trace/events/sched.h>
30 * Export tracepoints that act as a bare tracehook (ie: have no trace event
31 * associated with them) to allow external modules to probe them.
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
40 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
42 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
44 * Debugging: various feature bits
46 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
47 * sysctl_sched_features, defined in sched.h, to allow constants propagation
48 * at compile time and compiler optimization based on features default.
50 #define SCHED_FEAT(name, enabled) \
51 (1UL << __SCHED_FEAT_##name) * enabled |
52 const_debug
unsigned int sysctl_sched_features
=
59 * Number of tasks to iterate in a single balance run.
60 * Limited because this is done with IRQs disabled.
62 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
65 * period over which we measure -rt task CPU usage in us.
68 unsigned int sysctl_sched_rt_period
= 1000000;
70 __read_mostly
int scheduler_running
;
73 * part of the period that we allow rt tasks to run in us.
76 int sysctl_sched_rt_runtime
= 950000;
79 * __task_rq_lock - lock the rq @p resides on.
81 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
86 lockdep_assert_held(&p
->pi_lock
);
90 raw_spin_lock(&rq
->lock
);
91 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
95 raw_spin_unlock(&rq
->lock
);
97 while (unlikely(task_on_rq_migrating(p
)))
103 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
105 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
106 __acquires(p
->pi_lock
)
112 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
114 raw_spin_lock(&rq
->lock
);
116 * move_queued_task() task_rq_lock()
119 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
120 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
121 * [S] ->cpu = new_cpu [L] task_rq()
125 * If we observe the old CPU in task_rq_lock(), the acquire of
126 * the old rq->lock will fully serialize against the stores.
128 * If we observe the new CPU in task_rq_lock(), the address
129 * dependency headed by '[L] rq = task_rq()' and the acquire
130 * will pair with the WMB to ensure we then also see migrating.
132 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
136 raw_spin_unlock(&rq
->lock
);
137 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
139 while (unlikely(task_on_rq_migrating(p
)))
145 * RQ-clock updating methods:
148 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
151 * In theory, the compile should just see 0 here, and optimize out the call
152 * to sched_rt_avg_update. But I don't trust it...
154 s64 __maybe_unused steal
= 0, irq_delta
= 0;
156 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
157 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
160 * Since irq_time is only updated on {soft,}irq_exit, we might run into
161 * this case when a previous update_rq_clock() happened inside a
164 * When this happens, we stop ->clock_task and only update the
165 * prev_irq_time stamp to account for the part that fit, so that a next
166 * update will consume the rest. This ensures ->clock_task is
169 * It does however cause some slight miss-attribution of {soft,}irq
170 * time, a more accurate solution would be to update the irq_time using
171 * the current rq->clock timestamp, except that would require using
174 if (irq_delta
> delta
)
177 rq
->prev_irq_time
+= irq_delta
;
180 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
181 if (static_key_false((¶virt_steal_rq_enabled
))) {
182 steal
= paravirt_steal_clock(cpu_of(rq
));
183 steal
-= rq
->prev_steal_time_rq
;
185 if (unlikely(steal
> delta
))
188 rq
->prev_steal_time_rq
+= steal
;
193 rq
->clock_task
+= delta
;
195 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
196 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
197 update_irq_load_avg(rq
, irq_delta
+ steal
);
199 update_rq_clock_pelt(rq
, delta
);
202 void update_rq_clock(struct rq
*rq
)
206 lockdep_assert_held(&rq
->lock
);
208 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
211 #ifdef CONFIG_SCHED_DEBUG
212 if (sched_feat(WARN_DOUBLE_CLOCK
))
213 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
214 rq
->clock_update_flags
|= RQCF_UPDATED
;
217 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
221 update_rq_clock_task(rq
, delta
);
225 rq_csd_init(struct rq
*rq
, call_single_data_t
*csd
, smp_call_func_t func
)
232 #ifdef CONFIG_SCHED_HRTICK
234 * Use HR-timers to deliver accurate preemption points.
237 static void hrtick_clear(struct rq
*rq
)
239 if (hrtimer_active(&rq
->hrtick_timer
))
240 hrtimer_cancel(&rq
->hrtick_timer
);
244 * High-resolution timer tick.
245 * Runs from hardirq context with interrupts disabled.
247 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
249 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
252 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
256 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
259 return HRTIMER_NORESTART
;
264 static void __hrtick_restart(struct rq
*rq
)
266 struct hrtimer
*timer
= &rq
->hrtick_timer
;
268 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
272 * called from hardirq (IPI) context
274 static void __hrtick_start(void *arg
)
280 __hrtick_restart(rq
);
285 * Called to set the hrtick timer state.
287 * called with rq->lock held and irqs disabled
289 void hrtick_start(struct rq
*rq
, u64 delay
)
291 struct hrtimer
*timer
= &rq
->hrtick_timer
;
296 * Don't schedule slices shorter than 10000ns, that just
297 * doesn't make sense and can cause timer DoS.
299 delta
= max_t(s64
, delay
, 10000LL);
300 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
302 hrtimer_set_expires(timer
, time
);
305 __hrtick_restart(rq
);
307 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
312 * Called to set the hrtick timer state.
314 * called with rq->lock held and irqs disabled
316 void hrtick_start(struct rq
*rq
, u64 delay
)
319 * Don't schedule slices shorter than 10000ns, that just
320 * doesn't make sense. Rely on vruntime for fairness.
322 delay
= max_t(u64
, delay
, 10000LL);
323 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
324 HRTIMER_MODE_REL_PINNED_HARD
);
327 #endif /* CONFIG_SMP */
329 static void hrtick_rq_init(struct rq
*rq
)
332 rq_csd_init(rq
, &rq
->hrtick_csd
, __hrtick_start
);
334 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
335 rq
->hrtick_timer
.function
= hrtick
;
337 #else /* CONFIG_SCHED_HRTICK */
338 static inline void hrtick_clear(struct rq
*rq
)
342 static inline void hrtick_rq_init(struct rq
*rq
)
345 #endif /* CONFIG_SCHED_HRTICK */
348 * cmpxchg based fetch_or, macro so it works for different integer types
350 #define fetch_or(ptr, mask) \
352 typeof(ptr) _ptr = (ptr); \
353 typeof(mask) _mask = (mask); \
354 typeof(*_ptr) _old, _val = *_ptr; \
357 _old = cmpxchg(_ptr, _val, _val | _mask); \
365 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
367 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
368 * this avoids any races wrt polling state changes and thereby avoids
371 static bool set_nr_and_not_polling(struct task_struct
*p
)
373 struct thread_info
*ti
= task_thread_info(p
);
374 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
378 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
380 * If this returns true, then the idle task promises to call
381 * sched_ttwu_pending() and reschedule soon.
383 static bool set_nr_if_polling(struct task_struct
*p
)
385 struct thread_info
*ti
= task_thread_info(p
);
386 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
389 if (!(val
& _TIF_POLLING_NRFLAG
))
391 if (val
& _TIF_NEED_RESCHED
)
393 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
402 static bool set_nr_and_not_polling(struct task_struct
*p
)
404 set_tsk_need_resched(p
);
409 static bool set_nr_if_polling(struct task_struct
*p
)
416 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
418 struct wake_q_node
*node
= &task
->wake_q
;
421 * Atomically grab the task, if ->wake_q is !nil already it means
422 * its already queued (either by us or someone else) and will get the
423 * wakeup due to that.
425 * In order to ensure that a pending wakeup will observe our pending
426 * state, even in the failed case, an explicit smp_mb() must be used.
428 smp_mb__before_atomic();
429 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
433 * The head is context local, there can be no concurrency.
436 head
->lastp
= &node
->next
;
441 * wake_q_add() - queue a wakeup for 'later' waking.
442 * @head: the wake_q_head to add @task to
443 * @task: the task to queue for 'later' wakeup
445 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
446 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
449 * This function must be used as-if it were wake_up_process(); IOW the task
450 * must be ready to be woken at this location.
452 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
454 if (__wake_q_add(head
, task
))
455 get_task_struct(task
);
459 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
460 * @head: the wake_q_head to add @task to
461 * @task: the task to queue for 'later' wakeup
463 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
464 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
467 * This function must be used as-if it were wake_up_process(); IOW the task
468 * must be ready to be woken at this location.
470 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
471 * that already hold reference to @task can call the 'safe' version and trust
472 * wake_q to do the right thing depending whether or not the @task is already
475 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
477 if (!__wake_q_add(head
, task
))
478 put_task_struct(task
);
481 void wake_up_q(struct wake_q_head
*head
)
483 struct wake_q_node
*node
= head
->first
;
485 while (node
!= WAKE_Q_TAIL
) {
486 struct task_struct
*task
;
488 task
= container_of(node
, struct task_struct
, wake_q
);
490 /* Task can safely be re-inserted now: */
492 task
->wake_q
.next
= NULL
;
495 * wake_up_process() executes a full barrier, which pairs with
496 * the queueing in wake_q_add() so as not to miss wakeups.
498 wake_up_process(task
);
499 put_task_struct(task
);
504 * resched_curr - mark rq's current task 'to be rescheduled now'.
506 * On UP this means the setting of the need_resched flag, on SMP it
507 * might also involve a cross-CPU call to trigger the scheduler on
510 void resched_curr(struct rq
*rq
)
512 struct task_struct
*curr
= rq
->curr
;
515 lockdep_assert_held(&rq
->lock
);
517 if (test_tsk_need_resched(curr
))
522 if (cpu
== smp_processor_id()) {
523 set_tsk_need_resched(curr
);
524 set_preempt_need_resched();
528 if (set_nr_and_not_polling(curr
))
529 smp_send_reschedule(cpu
);
531 trace_sched_wake_idle_without_ipi(cpu
);
534 void resched_cpu(int cpu
)
536 struct rq
*rq
= cpu_rq(cpu
);
539 raw_spin_lock_irqsave(&rq
->lock
, flags
);
540 if (cpu_online(cpu
) || cpu
== smp_processor_id())
542 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
546 #ifdef CONFIG_NO_HZ_COMMON
548 * In the semi idle case, use the nearest busy CPU for migrating timers
549 * from an idle CPU. This is good for power-savings.
551 * We don't do similar optimization for completely idle system, as
552 * selecting an idle CPU will add more delays to the timers than intended
553 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
555 int get_nohz_timer_target(void)
557 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
558 struct sched_domain
*sd
;
560 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
567 for_each_domain(cpu
, sd
) {
568 for_each_cpu_and(i
, sched_domain_span(sd
),
569 housekeeping_cpumask(HK_FLAG_TIMER
)) {
580 if (default_cpu
== -1)
581 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
589 * When add_timer_on() enqueues a timer into the timer wheel of an
590 * idle CPU then this timer might expire before the next timer event
591 * which is scheduled to wake up that CPU. In case of a completely
592 * idle system the next event might even be infinite time into the
593 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
594 * leaves the inner idle loop so the newly added timer is taken into
595 * account when the CPU goes back to idle and evaluates the timer
596 * wheel for the next timer event.
598 static void wake_up_idle_cpu(int cpu
)
600 struct rq
*rq
= cpu_rq(cpu
);
602 if (cpu
== smp_processor_id())
605 if (set_nr_and_not_polling(rq
->idle
))
606 smp_send_reschedule(cpu
);
608 trace_sched_wake_idle_without_ipi(cpu
);
611 static bool wake_up_full_nohz_cpu(int cpu
)
614 * We just need the target to call irq_exit() and re-evaluate
615 * the next tick. The nohz full kick at least implies that.
616 * If needed we can still optimize that later with an
619 if (cpu_is_offline(cpu
))
620 return true; /* Don't try to wake offline CPUs. */
621 if (tick_nohz_full_cpu(cpu
)) {
622 if (cpu
!= smp_processor_id() ||
623 tick_nohz_tick_stopped())
624 tick_nohz_full_kick_cpu(cpu
);
632 * Wake up the specified CPU. If the CPU is going offline, it is the
633 * caller's responsibility to deal with the lost wakeup, for example,
634 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
636 void wake_up_nohz_cpu(int cpu
)
638 if (!wake_up_full_nohz_cpu(cpu
))
639 wake_up_idle_cpu(cpu
);
642 static void nohz_csd_func(void *info
)
644 struct rq
*rq
= info
;
645 int cpu
= cpu_of(rq
);
649 * Release the rq::nohz_csd.
651 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
652 WARN_ON(!(flags
& NOHZ_KICK_MASK
));
654 rq
->idle_balance
= idle_cpu(cpu
);
655 if (rq
->idle_balance
&& !need_resched()) {
656 rq
->nohz_idle_balance
= flags
;
657 raise_softirq_irqoff(SCHED_SOFTIRQ
);
661 #endif /* CONFIG_NO_HZ_COMMON */
663 #ifdef CONFIG_NO_HZ_FULL
664 bool sched_can_stop_tick(struct rq
*rq
)
668 /* Deadline tasks, even if single, need the tick */
669 if (rq
->dl
.dl_nr_running
)
673 * If there are more than one RR tasks, we need the tick to effect the
674 * actual RR behaviour.
676 if (rq
->rt
.rr_nr_running
) {
677 if (rq
->rt
.rr_nr_running
== 1)
684 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
685 * forced preemption between FIFO tasks.
687 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
692 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
693 * if there's more than one we need the tick for involuntary
696 if (rq
->nr_running
> 1)
701 #endif /* CONFIG_NO_HZ_FULL */
702 #endif /* CONFIG_SMP */
704 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
705 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
707 * Iterate task_group tree rooted at *from, calling @down when first entering a
708 * node and @up when leaving it for the final time.
710 * Caller must hold rcu_lock or sufficient equivalent.
712 int walk_tg_tree_from(struct task_group
*from
,
713 tg_visitor down
, tg_visitor up
, void *data
)
715 struct task_group
*parent
, *child
;
721 ret
= (*down
)(parent
, data
);
724 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
731 ret
= (*up
)(parent
, data
);
732 if (ret
|| parent
== from
)
736 parent
= parent
->parent
;
743 int tg_nop(struct task_group
*tg
, void *data
)
749 static void set_load_weight(struct task_struct
*p
, bool update_load
)
751 int prio
= p
->static_prio
- MAX_RT_PRIO
;
752 struct load_weight
*load
= &p
->se
.load
;
755 * SCHED_IDLE tasks get minimal weight:
757 if (task_has_idle_policy(p
)) {
758 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
759 load
->inv_weight
= WMULT_IDLEPRIO
;
764 * SCHED_OTHER tasks have to update their load when changing their
767 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
768 reweight_task(p
, prio
);
770 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
771 load
->inv_weight
= sched_prio_to_wmult
[prio
];
775 #ifdef CONFIG_UCLAMP_TASK
777 * Serializes updates of utilization clamp values
779 * The (slow-path) user-space triggers utilization clamp value updates which
780 * can require updates on (fast-path) scheduler's data structures used to
781 * support enqueue/dequeue operations.
782 * While the per-CPU rq lock protects fast-path update operations, user-space
783 * requests are serialized using a mutex to reduce the risk of conflicting
784 * updates or API abuses.
786 static DEFINE_MUTEX(uclamp_mutex
);
788 /* Max allowed minimum utilization */
789 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
791 /* Max allowed maximum utilization */
792 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
794 /* All clamps are required to be less or equal than these values */
795 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
797 /* Integer rounded range for each bucket */
798 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
800 #define for_each_clamp_id(clamp_id) \
801 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
803 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
805 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
808 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value
)
810 return UCLAMP_BUCKET_DELTA
* uclamp_bucket_id(clamp_value
);
813 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
815 if (clamp_id
== UCLAMP_MIN
)
817 return SCHED_CAPACITY_SCALE
;
820 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
821 unsigned int value
, bool user_defined
)
823 uc_se
->value
= value
;
824 uc_se
->bucket_id
= uclamp_bucket_id(value
);
825 uc_se
->user_defined
= user_defined
;
828 static inline unsigned int
829 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
830 unsigned int clamp_value
)
833 * Avoid blocked utilization pushing up the frequency when we go
834 * idle (which drops the max-clamp) by retaining the last known
837 if (clamp_id
== UCLAMP_MAX
) {
838 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
842 return uclamp_none(UCLAMP_MIN
);
845 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
846 unsigned int clamp_value
)
848 /* Reset max-clamp retention only on idle exit */
849 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
852 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
856 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
857 unsigned int clamp_value
)
859 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
860 int bucket_id
= UCLAMP_BUCKETS
- 1;
863 * Since both min and max clamps are max aggregated, find the
864 * top most bucket with tasks in.
866 for ( ; bucket_id
>= 0; bucket_id
--) {
867 if (!bucket
[bucket_id
].tasks
)
869 return bucket
[bucket_id
].value
;
872 /* No tasks -- default clamp values */
873 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
876 static inline struct uclamp_se
877 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
879 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
880 #ifdef CONFIG_UCLAMP_TASK_GROUP
881 struct uclamp_se uc_max
;
884 * Tasks in autogroups or root task group will be
885 * restricted by system defaults.
887 if (task_group_is_autogroup(task_group(p
)))
889 if (task_group(p
) == &root_task_group
)
892 uc_max
= task_group(p
)->uclamp
[clamp_id
];
893 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
901 * The effective clamp bucket index of a task depends on, by increasing
903 * - the task specific clamp value, when explicitly requested from userspace
904 * - the task group effective clamp value, for tasks not either in the root
905 * group or in an autogroup
906 * - the system default clamp value, defined by the sysadmin
908 static inline struct uclamp_se
909 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
911 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
912 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
914 /* System default restrictions always apply */
915 if (unlikely(uc_req
.value
> uc_max
.value
))
921 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
923 struct uclamp_se uc_eff
;
925 /* Task currently refcounted: use back-annotated (effective) value */
926 if (p
->uclamp
[clamp_id
].active
)
927 return (unsigned long)p
->uclamp
[clamp_id
].value
;
929 uc_eff
= uclamp_eff_get(p
, clamp_id
);
931 return (unsigned long)uc_eff
.value
;
935 * When a task is enqueued on a rq, the clamp bucket currently defined by the
936 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
937 * updates the rq's clamp value if required.
939 * Tasks can have a task-specific value requested from user-space, track
940 * within each bucket the maximum value for tasks refcounted in it.
941 * This "local max aggregation" allows to track the exact "requested" value
942 * for each bucket when all its RUNNABLE tasks require the same clamp.
944 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
945 enum uclamp_id clamp_id
)
947 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
948 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
949 struct uclamp_bucket
*bucket
;
951 lockdep_assert_held(&rq
->lock
);
953 /* Update task effective clamp */
954 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
956 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
958 uc_se
->active
= true;
960 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
963 * Local max aggregation: rq buckets always track the max
964 * "requested" clamp value of its RUNNABLE tasks.
966 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
967 bucket
->value
= uc_se
->value
;
969 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
970 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
974 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
975 * is released. If this is the last task reference counting the rq's max
976 * active clamp value, then the rq's clamp value is updated.
978 * Both refcounted tasks and rq's cached clamp values are expected to be
979 * always valid. If it's detected they are not, as defensive programming,
980 * enforce the expected state and warn.
982 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
983 enum uclamp_id clamp_id
)
985 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
986 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
987 struct uclamp_bucket
*bucket
;
988 unsigned int bkt_clamp
;
989 unsigned int rq_clamp
;
991 lockdep_assert_held(&rq
->lock
);
993 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
994 SCHED_WARN_ON(!bucket
->tasks
);
995 if (likely(bucket
->tasks
))
997 uc_se
->active
= false;
1000 * Keep "local max aggregation" simple and accept to (possibly)
1001 * overboost some RUNNABLE tasks in the same bucket.
1002 * The rq clamp bucket value is reset to its base value whenever
1003 * there are no more RUNNABLE tasks refcounting it.
1005 if (likely(bucket
->tasks
))
1008 rq_clamp
= READ_ONCE(uc_rq
->value
);
1010 * Defensive programming: this should never happen. If it happens,
1011 * e.g. due to future modification, warn and fixup the expected value.
1013 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1014 if (bucket
->value
>= rq_clamp
) {
1015 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1016 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1020 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1022 enum uclamp_id clamp_id
;
1024 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1027 for_each_clamp_id(clamp_id
)
1028 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1030 /* Reset clamp idle holding when there is one RUNNABLE task */
1031 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1032 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1035 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1037 enum uclamp_id clamp_id
;
1039 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1042 for_each_clamp_id(clamp_id
)
1043 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1047 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1053 * Lock the task and the rq where the task is (or was) queued.
1055 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1056 * price to pay to safely serialize util_{min,max} updates with
1057 * enqueues, dequeues and migration operations.
1058 * This is the same locking schema used by __set_cpus_allowed_ptr().
1060 rq
= task_rq_lock(p
, &rf
);
1063 * Setting the clamp bucket is serialized by task_rq_lock().
1064 * If the task is not yet RUNNABLE and its task_struct is not
1065 * affecting a valid clamp bucket, the next time it's enqueued,
1066 * it will already see the updated clamp bucket value.
1068 if (p
->uclamp
[clamp_id
].active
) {
1069 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1070 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1073 task_rq_unlock(rq
, p
, &rf
);
1076 #ifdef CONFIG_UCLAMP_TASK_GROUP
1078 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1079 unsigned int clamps
)
1081 enum uclamp_id clamp_id
;
1082 struct css_task_iter it
;
1083 struct task_struct
*p
;
1085 css_task_iter_start(css
, 0, &it
);
1086 while ((p
= css_task_iter_next(&it
))) {
1087 for_each_clamp_id(clamp_id
) {
1088 if ((0x1 << clamp_id
) & clamps
)
1089 uclamp_update_active(p
, clamp_id
);
1092 css_task_iter_end(&it
);
1095 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1096 static void uclamp_update_root_tg(void)
1098 struct task_group
*tg
= &root_task_group
;
1100 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1101 sysctl_sched_uclamp_util_min
, false);
1102 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1103 sysctl_sched_uclamp_util_max
, false);
1106 cpu_util_update_eff(&root_task_group
.css
);
1110 static void uclamp_update_root_tg(void) { }
1113 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1114 void __user
*buffer
, size_t *lenp
,
1117 bool update_root_tg
= false;
1118 int old_min
, old_max
;
1121 mutex_lock(&uclamp_mutex
);
1122 old_min
= sysctl_sched_uclamp_util_min
;
1123 old_max
= sysctl_sched_uclamp_util_max
;
1125 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1131 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1132 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
) {
1137 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1138 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1139 sysctl_sched_uclamp_util_min
, false);
1140 update_root_tg
= true;
1142 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1143 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1144 sysctl_sched_uclamp_util_max
, false);
1145 update_root_tg
= true;
1149 uclamp_update_root_tg();
1152 * We update all RUNNABLE tasks only when task groups are in use.
1153 * Otherwise, keep it simple and do just a lazy update at each next
1154 * task enqueue time.
1160 sysctl_sched_uclamp_util_min
= old_min
;
1161 sysctl_sched_uclamp_util_max
= old_max
;
1163 mutex_unlock(&uclamp_mutex
);
1168 static int uclamp_validate(struct task_struct
*p
,
1169 const struct sched_attr
*attr
)
1171 unsigned int lower_bound
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1172 unsigned int upper_bound
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1174 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
)
1175 lower_bound
= attr
->sched_util_min
;
1176 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
)
1177 upper_bound
= attr
->sched_util_max
;
1179 if (lower_bound
> upper_bound
)
1181 if (upper_bound
> SCHED_CAPACITY_SCALE
)
1187 static void __setscheduler_uclamp(struct task_struct
*p
,
1188 const struct sched_attr
*attr
)
1190 enum uclamp_id clamp_id
;
1193 * On scheduling class change, reset to default clamps for tasks
1194 * without a task-specific value.
1196 for_each_clamp_id(clamp_id
) {
1197 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1198 unsigned int clamp_value
= uclamp_none(clamp_id
);
1200 /* Keep using defined clamps across class changes */
1201 if (uc_se
->user_defined
)
1204 /* By default, RT tasks always get 100% boost */
1205 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1206 clamp_value
= uclamp_none(UCLAMP_MAX
);
1208 uclamp_se_set(uc_se
, clamp_value
, false);
1211 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1214 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1215 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1216 attr
->sched_util_min
, true);
1219 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1220 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1221 attr
->sched_util_max
, true);
1225 static void uclamp_fork(struct task_struct
*p
)
1227 enum uclamp_id clamp_id
;
1229 for_each_clamp_id(clamp_id
)
1230 p
->uclamp
[clamp_id
].active
= false;
1232 if (likely(!p
->sched_reset_on_fork
))
1235 for_each_clamp_id(clamp_id
) {
1236 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1237 uclamp_none(clamp_id
), false);
1241 static void __init
init_uclamp(void)
1243 struct uclamp_se uc_max
= {};
1244 enum uclamp_id clamp_id
;
1247 mutex_init(&uclamp_mutex
);
1249 for_each_possible_cpu(cpu
) {
1250 memset(&cpu_rq(cpu
)->uclamp
, 0,
1251 sizeof(struct uclamp_rq
)*UCLAMP_CNT
);
1252 cpu_rq(cpu
)->uclamp_flags
= 0;
1255 for_each_clamp_id(clamp_id
) {
1256 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1257 uclamp_none(clamp_id
), false);
1260 /* System defaults allow max clamp values for both indexes */
1261 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1262 for_each_clamp_id(clamp_id
) {
1263 uclamp_default
[clamp_id
] = uc_max
;
1264 #ifdef CONFIG_UCLAMP_TASK_GROUP
1265 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1266 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1271 #else /* CONFIG_UCLAMP_TASK */
1272 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1273 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1274 static inline int uclamp_validate(struct task_struct
*p
,
1275 const struct sched_attr
*attr
)
1279 static void __setscheduler_uclamp(struct task_struct
*p
,
1280 const struct sched_attr
*attr
) { }
1281 static inline void uclamp_fork(struct task_struct
*p
) { }
1282 static inline void init_uclamp(void) { }
1283 #endif /* CONFIG_UCLAMP_TASK */
1285 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1287 if (!(flags
& ENQUEUE_NOCLOCK
))
1288 update_rq_clock(rq
);
1290 if (!(flags
& ENQUEUE_RESTORE
)) {
1291 sched_info_queued(rq
, p
);
1292 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1295 uclamp_rq_inc(rq
, p
);
1296 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1299 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1301 if (!(flags
& DEQUEUE_NOCLOCK
))
1302 update_rq_clock(rq
);
1304 if (!(flags
& DEQUEUE_SAVE
)) {
1305 sched_info_dequeued(rq
, p
);
1306 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1309 uclamp_rq_dec(rq
, p
);
1310 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1313 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1315 if (task_contributes_to_load(p
))
1316 rq
->nr_uninterruptible
--;
1318 enqueue_task(rq
, p
, flags
);
1320 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1323 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1325 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1327 if (task_contributes_to_load(p
))
1328 rq
->nr_uninterruptible
++;
1330 dequeue_task(rq
, p
, flags
);
1334 * __normal_prio - return the priority that is based on the static prio
1336 static inline int __normal_prio(struct task_struct
*p
)
1338 return p
->static_prio
;
1342 * Calculate the expected normal priority: i.e. priority
1343 * without taking RT-inheritance into account. Might be
1344 * boosted by interactivity modifiers. Changes upon fork,
1345 * setprio syscalls, and whenever the interactivity
1346 * estimator recalculates.
1348 static inline int normal_prio(struct task_struct
*p
)
1352 if (task_has_dl_policy(p
))
1353 prio
= MAX_DL_PRIO
-1;
1354 else if (task_has_rt_policy(p
))
1355 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1357 prio
= __normal_prio(p
);
1362 * Calculate the current priority, i.e. the priority
1363 * taken into account by the scheduler. This value might
1364 * be boosted by RT tasks, or might be boosted by
1365 * interactivity modifiers. Will be RT if the task got
1366 * RT-boosted. If not then it returns p->normal_prio.
1368 static int effective_prio(struct task_struct
*p
)
1370 p
->normal_prio
= normal_prio(p
);
1372 * If we are RT tasks or we were boosted to RT priority,
1373 * keep the priority unchanged. Otherwise, update priority
1374 * to the normal priority:
1376 if (!rt_prio(p
->prio
))
1377 return p
->normal_prio
;
1382 * task_curr - is this task currently executing on a CPU?
1383 * @p: the task in question.
1385 * Return: 1 if the task is currently executing. 0 otherwise.
1387 inline int task_curr(const struct task_struct
*p
)
1389 return cpu_curr(task_cpu(p
)) == p
;
1393 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1394 * use the balance_callback list if you want balancing.
1396 * this means any call to check_class_changed() must be followed by a call to
1397 * balance_callback().
1399 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1400 const struct sched_class
*prev_class
,
1403 if (prev_class
!= p
->sched_class
) {
1404 if (prev_class
->switched_from
)
1405 prev_class
->switched_from(rq
, p
);
1407 p
->sched_class
->switched_to(rq
, p
);
1408 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1409 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1412 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1414 const struct sched_class
*class;
1416 if (p
->sched_class
== rq
->curr
->sched_class
) {
1417 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1419 for_each_class(class) {
1420 if (class == rq
->curr
->sched_class
)
1422 if (class == p
->sched_class
) {
1430 * A queue event has occurred, and we're going to schedule. In
1431 * this case, we can save a useless back to back clock update.
1433 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1434 rq_clock_skip_update(rq
);
1440 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1441 * __set_cpus_allowed_ptr() and select_fallback_rq().
1443 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1445 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1448 if (is_per_cpu_kthread(p
))
1449 return cpu_online(cpu
);
1451 return cpu_active(cpu
);
1455 * This is how migration works:
1457 * 1) we invoke migration_cpu_stop() on the target CPU using
1459 * 2) stopper starts to run (implicitly forcing the migrated thread
1461 * 3) it checks whether the migrated task is still in the wrong runqueue.
1462 * 4) if it's in the wrong runqueue then the migration thread removes
1463 * it and puts it into the right queue.
1464 * 5) stopper completes and stop_one_cpu() returns and the migration
1469 * move_queued_task - move a queued task to new rq.
1471 * Returns (locked) new rq. Old rq's lock is released.
1473 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1474 struct task_struct
*p
, int new_cpu
)
1476 lockdep_assert_held(&rq
->lock
);
1478 WRITE_ONCE(p
->on_rq
, TASK_ON_RQ_MIGRATING
);
1479 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
1480 set_task_cpu(p
, new_cpu
);
1483 rq
= cpu_rq(new_cpu
);
1486 BUG_ON(task_cpu(p
) != new_cpu
);
1487 enqueue_task(rq
, p
, 0);
1488 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1489 check_preempt_curr(rq
, p
, 0);
1494 struct migration_arg
{
1495 struct task_struct
*task
;
1500 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1501 * this because either it can't run here any more (set_cpus_allowed()
1502 * away from this CPU, or CPU going down), or because we're
1503 * attempting to rebalance this task on exec (sched_exec).
1505 * So we race with normal scheduler movements, but that's OK, as long
1506 * as the task is no longer on this CPU.
1508 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1509 struct task_struct
*p
, int dest_cpu
)
1511 /* Affinity changed (again). */
1512 if (!is_cpu_allowed(p
, dest_cpu
))
1515 update_rq_clock(rq
);
1516 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1522 * migration_cpu_stop - this will be executed by a highprio stopper thread
1523 * and performs thread migration by bumping thread off CPU then
1524 * 'pushing' onto another runqueue.
1526 static int migration_cpu_stop(void *data
)
1528 struct migration_arg
*arg
= data
;
1529 struct task_struct
*p
= arg
->task
;
1530 struct rq
*rq
= this_rq();
1534 * The original target CPU might have gone down and we might
1535 * be on another CPU but it doesn't matter.
1537 local_irq_disable();
1539 * We need to explicitly wake pending tasks before running
1540 * __migrate_task() such that we will not miss enforcing cpus_ptr
1541 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1543 flush_smp_call_function_from_idle();
1545 raw_spin_lock(&p
->pi_lock
);
1548 * If task_rq(p) != rq, it cannot be migrated here, because we're
1549 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1550 * we're holding p->pi_lock.
1552 if (task_rq(p
) == rq
) {
1553 if (task_on_rq_queued(p
))
1554 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1556 p
->wake_cpu
= arg
->dest_cpu
;
1559 raw_spin_unlock(&p
->pi_lock
);
1566 * sched_class::set_cpus_allowed must do the below, but is not required to
1567 * actually call this function.
1569 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1571 cpumask_copy(&p
->cpus_mask
, new_mask
);
1572 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1575 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1577 struct rq
*rq
= task_rq(p
);
1578 bool queued
, running
;
1580 lockdep_assert_held(&p
->pi_lock
);
1582 queued
= task_on_rq_queued(p
);
1583 running
= task_current(rq
, p
);
1587 * Because __kthread_bind() calls this on blocked tasks without
1590 lockdep_assert_held(&rq
->lock
);
1591 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1594 put_prev_task(rq
, p
);
1596 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1599 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1601 set_next_task(rq
, p
);
1605 * Change a given task's CPU affinity. Migrate the thread to a
1606 * proper CPU and schedule it away if the CPU it's executing on
1607 * is removed from the allowed bitmask.
1609 * NOTE: the caller must have a valid reference to the task, the
1610 * task must not exit() & deallocate itself prematurely. The
1611 * call is not atomic; no spinlocks may be held.
1613 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1614 const struct cpumask
*new_mask
, bool check
)
1616 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1617 unsigned int dest_cpu
;
1622 rq
= task_rq_lock(p
, &rf
);
1623 update_rq_clock(rq
);
1625 if (p
->flags
& PF_KTHREAD
) {
1627 * Kernel threads are allowed on online && !active CPUs
1629 cpu_valid_mask
= cpu_online_mask
;
1633 * Must re-check here, to close a race against __kthread_bind(),
1634 * sched_setaffinity() is not guaranteed to observe the flag.
1636 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1641 if (cpumask_equal(p
->cpus_ptr
, new_mask
))
1645 * Picking a ~random cpu helps in cases where we are changing affinity
1646 * for groups of tasks (ie. cpuset), so that load balancing is not
1647 * immediately required to distribute the tasks within their new mask.
1649 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
1650 if (dest_cpu
>= nr_cpu_ids
) {
1655 do_set_cpus_allowed(p
, new_mask
);
1657 if (p
->flags
& PF_KTHREAD
) {
1659 * For kernel threads that do indeed end up on online &&
1660 * !active we want to ensure they are strict per-CPU threads.
1662 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1663 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1664 p
->nr_cpus_allowed
!= 1);
1667 /* Can the task run on the task's current CPU? If so, we're done */
1668 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1671 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1672 struct migration_arg arg
= { p
, dest_cpu
};
1673 /* Need help from migration thread: drop lock and wait. */
1674 task_rq_unlock(rq
, p
, &rf
);
1675 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1677 } else if (task_on_rq_queued(p
)) {
1679 * OK, since we're going to drop the lock immediately
1680 * afterwards anyway.
1682 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1685 task_rq_unlock(rq
, p
, &rf
);
1690 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1692 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1694 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1696 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1698 #ifdef CONFIG_SCHED_DEBUG
1700 * We should never call set_task_cpu() on a blocked task,
1701 * ttwu() will sort out the placement.
1703 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1707 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1708 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1709 * time relying on p->on_rq.
1711 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1712 p
->sched_class
== &fair_sched_class
&&
1713 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1715 #ifdef CONFIG_LOCKDEP
1717 * The caller should hold either p->pi_lock or rq->lock, when changing
1718 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1720 * sched_move_task() holds both and thus holding either pins the cgroup,
1723 * Furthermore, all task_rq users should acquire both locks, see
1726 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1727 lockdep_is_held(&task_rq(p
)->lock
)));
1730 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1732 WARN_ON_ONCE(!cpu_online(new_cpu
));
1735 trace_sched_migrate_task(p
, new_cpu
);
1737 if (task_cpu(p
) != new_cpu
) {
1738 if (p
->sched_class
->migrate_task_rq
)
1739 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1740 p
->se
.nr_migrations
++;
1742 perf_event_task_migrate(p
);
1745 __set_task_cpu(p
, new_cpu
);
1748 #ifdef CONFIG_NUMA_BALANCING
1749 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1751 if (task_on_rq_queued(p
)) {
1752 struct rq
*src_rq
, *dst_rq
;
1753 struct rq_flags srf
, drf
;
1755 src_rq
= task_rq(p
);
1756 dst_rq
= cpu_rq(cpu
);
1758 rq_pin_lock(src_rq
, &srf
);
1759 rq_pin_lock(dst_rq
, &drf
);
1761 deactivate_task(src_rq
, p
, 0);
1762 set_task_cpu(p
, cpu
);
1763 activate_task(dst_rq
, p
, 0);
1764 check_preempt_curr(dst_rq
, p
, 0);
1766 rq_unpin_lock(dst_rq
, &drf
);
1767 rq_unpin_lock(src_rq
, &srf
);
1771 * Task isn't running anymore; make it appear like we migrated
1772 * it before it went to sleep. This means on wakeup we make the
1773 * previous CPU our target instead of where it really is.
1779 struct migration_swap_arg
{
1780 struct task_struct
*src_task
, *dst_task
;
1781 int src_cpu
, dst_cpu
;
1784 static int migrate_swap_stop(void *data
)
1786 struct migration_swap_arg
*arg
= data
;
1787 struct rq
*src_rq
, *dst_rq
;
1790 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1793 src_rq
= cpu_rq(arg
->src_cpu
);
1794 dst_rq
= cpu_rq(arg
->dst_cpu
);
1796 double_raw_lock(&arg
->src_task
->pi_lock
,
1797 &arg
->dst_task
->pi_lock
);
1798 double_rq_lock(src_rq
, dst_rq
);
1800 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1803 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1806 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
1809 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
1812 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1813 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1818 double_rq_unlock(src_rq
, dst_rq
);
1819 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1820 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1826 * Cross migrate two tasks
1828 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
1829 int target_cpu
, int curr_cpu
)
1831 struct migration_swap_arg arg
;
1834 arg
= (struct migration_swap_arg
){
1836 .src_cpu
= curr_cpu
,
1838 .dst_cpu
= target_cpu
,
1841 if (arg
.src_cpu
== arg
.dst_cpu
)
1845 * These three tests are all lockless; this is OK since all of them
1846 * will be re-checked with proper locks held further down the line.
1848 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1851 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
1854 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
1857 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1858 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1863 #endif /* CONFIG_NUMA_BALANCING */
1866 * wait_task_inactive - wait for a thread to unschedule.
1868 * If @match_state is nonzero, it's the @p->state value just checked and
1869 * not expected to change. If it changes, i.e. @p might have woken up,
1870 * then return zero. When we succeed in waiting for @p to be off its CPU,
1871 * we return a positive number (its total switch count). If a second call
1872 * a short while later returns the same number, the caller can be sure that
1873 * @p has remained unscheduled the whole time.
1875 * The caller must ensure that the task *will* unschedule sometime soon,
1876 * else this function might spin for a *long* time. This function can't
1877 * be called with interrupts off, or it may introduce deadlock with
1878 * smp_call_function() if an IPI is sent by the same process we are
1879 * waiting to become inactive.
1881 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1883 int running
, queued
;
1890 * We do the initial early heuristics without holding
1891 * any task-queue locks at all. We'll only try to get
1892 * the runqueue lock when things look like they will
1898 * If the task is actively running on another CPU
1899 * still, just relax and busy-wait without holding
1902 * NOTE! Since we don't hold any locks, it's not
1903 * even sure that "rq" stays as the right runqueue!
1904 * But we don't care, since "task_running()" will
1905 * return false if the runqueue has changed and p
1906 * is actually now running somewhere else!
1908 while (task_running(rq
, p
)) {
1909 if (match_state
&& unlikely(p
->state
!= match_state
))
1915 * Ok, time to look more closely! We need the rq
1916 * lock now, to be *sure*. If we're wrong, we'll
1917 * just go back and repeat.
1919 rq
= task_rq_lock(p
, &rf
);
1920 trace_sched_wait_task(p
);
1921 running
= task_running(rq
, p
);
1922 queued
= task_on_rq_queued(p
);
1924 if (!match_state
|| p
->state
== match_state
)
1925 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1926 task_rq_unlock(rq
, p
, &rf
);
1929 * If it changed from the expected state, bail out now.
1931 if (unlikely(!ncsw
))
1935 * Was it really running after all now that we
1936 * checked with the proper locks actually held?
1938 * Oops. Go back and try again..
1940 if (unlikely(running
)) {
1946 * It's not enough that it's not actively running,
1947 * it must be off the runqueue _entirely_, and not
1950 * So if it was still runnable (but just not actively
1951 * running right now), it's preempted, and we should
1952 * yield - it could be a while.
1954 if (unlikely(queued
)) {
1955 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1957 set_current_state(TASK_UNINTERRUPTIBLE
);
1958 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1963 * Ahh, all good. It wasn't running, and it wasn't
1964 * runnable, which means that it will never become
1965 * running in the future either. We're all done!
1974 * kick_process - kick a running thread to enter/exit the kernel
1975 * @p: the to-be-kicked thread
1977 * Cause a process which is running on another CPU to enter
1978 * kernel-mode, without any delay. (to get signals handled.)
1980 * NOTE: this function doesn't have to take the runqueue lock,
1981 * because all it wants to ensure is that the remote task enters
1982 * the kernel. If the IPI races and the task has been migrated
1983 * to another CPU then no harm is done and the purpose has been
1986 void kick_process(struct task_struct
*p
)
1992 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1993 smp_send_reschedule(cpu
);
1996 EXPORT_SYMBOL_GPL(kick_process
);
1999 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2001 * A few notes on cpu_active vs cpu_online:
2003 * - cpu_active must be a subset of cpu_online
2005 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2006 * see __set_cpus_allowed_ptr(). At this point the newly online
2007 * CPU isn't yet part of the sched domains, and balancing will not
2010 * - on CPU-down we clear cpu_active() to mask the sched domains and
2011 * avoid the load balancer to place new tasks on the to be removed
2012 * CPU. Existing tasks will remain running there and will be taken
2015 * This means that fallback selection must not select !active CPUs.
2016 * And can assume that any active CPU must be online. Conversely
2017 * select_task_rq() below may allow selection of !active CPUs in order
2018 * to satisfy the above rules.
2020 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2022 int nid
= cpu_to_node(cpu
);
2023 const struct cpumask
*nodemask
= NULL
;
2024 enum { cpuset
, possible
, fail
} state
= cpuset
;
2028 * If the node that the CPU is on has been offlined, cpu_to_node()
2029 * will return -1. There is no CPU on the node, and we should
2030 * select the CPU on the other node.
2033 nodemask
= cpumask_of_node(nid
);
2035 /* Look for allowed, online CPU in same node. */
2036 for_each_cpu(dest_cpu
, nodemask
) {
2037 if (!cpu_active(dest_cpu
))
2039 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2045 /* Any allowed, online CPU? */
2046 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2047 if (!is_cpu_allowed(p
, dest_cpu
))
2053 /* No more Mr. Nice Guy. */
2056 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2057 cpuset_cpus_allowed_fallback(p
);
2063 do_set_cpus_allowed(p
, cpu_possible_mask
);
2074 if (state
!= cpuset
) {
2076 * Don't tell them about moving exiting tasks or
2077 * kernel threads (both mm NULL), since they never
2080 if (p
->mm
&& printk_ratelimit()) {
2081 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2082 task_pid_nr(p
), p
->comm
, cpu
);
2090 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2093 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
2095 lockdep_assert_held(&p
->pi_lock
);
2097 if (p
->nr_cpus_allowed
> 1)
2098 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
2100 cpu
= cpumask_any(p
->cpus_ptr
);
2103 * In order not to call set_task_cpu() on a blocking task we need
2104 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2107 * Since this is common to all placement strategies, this lives here.
2109 * [ this allows ->select_task() to simply return task_cpu(p) and
2110 * not worry about this generic constraint ]
2112 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2113 cpu
= select_fallback_rq(task_cpu(p
), p
);
2118 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2120 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2121 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2125 * Make it appear like a SCHED_FIFO task, its something
2126 * userspace knows about and won't get confused about.
2128 * Also, it will make PI more or less work without too
2129 * much confusion -- but then, stop work should not
2130 * rely on PI working anyway.
2132 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2134 stop
->sched_class
= &stop_sched_class
;
2137 cpu_rq(cpu
)->stop
= stop
;
2141 * Reset it back to a normal scheduling class so that
2142 * it can die in pieces.
2144 old_stop
->sched_class
= &rt_sched_class
;
2150 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2151 const struct cpumask
*new_mask
, bool check
)
2153 return set_cpus_allowed_ptr(p
, new_mask
);
2156 #endif /* CONFIG_SMP */
2159 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2163 if (!schedstat_enabled())
2169 if (cpu
== rq
->cpu
) {
2170 __schedstat_inc(rq
->ttwu_local
);
2171 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2173 struct sched_domain
*sd
;
2175 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2177 for_each_domain(rq
->cpu
, sd
) {
2178 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2179 __schedstat_inc(sd
->ttwu_wake_remote
);
2186 if (wake_flags
& WF_MIGRATED
)
2187 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2188 #endif /* CONFIG_SMP */
2190 __schedstat_inc(rq
->ttwu_count
);
2191 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2193 if (wake_flags
& WF_SYNC
)
2194 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2198 * Mark the task runnable and perform wakeup-preemption.
2200 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2201 struct rq_flags
*rf
)
2203 check_preempt_curr(rq
, p
, wake_flags
);
2204 p
->state
= TASK_RUNNING
;
2205 trace_sched_wakeup(p
);
2208 if (p
->sched_class
->task_woken
) {
2210 * Our task @p is fully woken up and running; so its safe to
2211 * drop the rq->lock, hereafter rq is only used for statistics.
2213 rq_unpin_lock(rq
, rf
);
2214 p
->sched_class
->task_woken(rq
, p
);
2215 rq_repin_lock(rq
, rf
);
2218 if (rq
->idle_stamp
) {
2219 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2220 u64 max
= 2*rq
->max_idle_balance_cost
;
2222 update_avg(&rq
->avg_idle
, delta
);
2224 if (rq
->avg_idle
> max
)
2233 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2234 struct rq_flags
*rf
)
2236 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2238 lockdep_assert_held(&rq
->lock
);
2241 if (p
->sched_contributes_to_load
)
2242 rq
->nr_uninterruptible
--;
2244 if (wake_flags
& WF_MIGRATED
)
2245 en_flags
|= ENQUEUE_MIGRATED
;
2248 activate_task(rq
, p
, en_flags
);
2249 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2253 * Called in case the task @p isn't fully descheduled from its runqueue,
2254 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2255 * since all we need to do is flip p->state to TASK_RUNNING, since
2256 * the task is still ->on_rq.
2258 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2264 rq
= __task_rq_lock(p
, &rf
);
2265 if (task_on_rq_queued(p
)) {
2266 /* check_preempt_curr() may use rq clock */
2267 update_rq_clock(rq
);
2268 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
2271 __task_rq_unlock(rq
, &rf
);
2277 void sched_ttwu_pending(void *arg
)
2279 struct llist_node
*llist
= arg
;
2280 struct rq
*rq
= this_rq();
2281 struct task_struct
*p
, *t
;
2288 * rq::ttwu_pending racy indication of out-standing wakeups.
2289 * Races such that false-negatives are possible, since they
2290 * are shorter lived that false-positives would be.
2292 WRITE_ONCE(rq
->ttwu_pending
, 0);
2294 rq_lock_irqsave(rq
, &rf
);
2295 update_rq_clock(rq
);
2297 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
2298 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
2300 rq_unlock_irqrestore(rq
, &rf
);
2303 void send_call_function_single_ipi(int cpu
)
2305 struct rq
*rq
= cpu_rq(cpu
);
2307 if (!set_nr_if_polling(rq
->idle
))
2308 arch_send_call_function_single_ipi(cpu
);
2310 trace_sched_wake_idle_without_ipi(cpu
);
2314 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2315 * necessary. The wakee CPU on receipt of the IPI will queue the task
2316 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2317 * of the wakeup instead of the waker.
2319 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
2321 struct rq
*rq
= cpu_rq(cpu
);
2323 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
2325 WRITE_ONCE(rq
->ttwu_pending
, 1);
2326 __smp_call_single_queue(cpu
, &p
->wake_entry
);
2329 void wake_up_if_idle(int cpu
)
2331 struct rq
*rq
= cpu_rq(cpu
);
2336 if (!is_idle_task(rcu_dereference(rq
->curr
)))
2339 if (set_nr_if_polling(rq
->idle
)) {
2340 trace_sched_wake_idle_without_ipi(cpu
);
2342 rq_lock_irqsave(rq
, &rf
);
2343 if (is_idle_task(rq
->curr
))
2344 smp_send_reschedule(cpu
);
2345 /* Else CPU is not idle, do nothing here: */
2346 rq_unlock_irqrestore(rq
, &rf
);
2353 bool cpus_share_cache(int this_cpu
, int that_cpu
)
2355 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
2358 static inline bool ttwu_queue_cond(int cpu
, int wake_flags
)
2361 * If the CPU does not share cache, then queue the task on the
2362 * remote rqs wakelist to avoid accessing remote data.
2364 if (!cpus_share_cache(smp_processor_id(), cpu
))
2368 * If the task is descheduling and the only running task on the
2369 * CPU then use the wakelist to offload the task activation to
2370 * the soon-to-be-idle CPU as the current CPU is likely busy.
2371 * nr_running is checked to avoid unnecessary task stacking.
2373 if ((wake_flags
& WF_ON_RQ
) && cpu_rq(cpu
)->nr_running
<= 1)
2379 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
2381 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(cpu
, wake_flags
)) {
2382 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
2383 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
2389 #endif /* CONFIG_SMP */
2391 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
2393 struct rq
*rq
= cpu_rq(cpu
);
2396 #if defined(CONFIG_SMP)
2397 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
2402 update_rq_clock(rq
);
2403 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
2408 * Notes on Program-Order guarantees on SMP systems.
2412 * The basic program-order guarantee on SMP systems is that when a task [t]
2413 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2414 * execution on its new CPU [c1].
2416 * For migration (of runnable tasks) this is provided by the following means:
2418 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2419 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2420 * rq(c1)->lock (if not at the same time, then in that order).
2421 * C) LOCK of the rq(c1)->lock scheduling in task
2423 * Release/acquire chaining guarantees that B happens after A and C after B.
2424 * Note: the CPU doing B need not be c0 or c1
2433 * UNLOCK rq(0)->lock
2435 * LOCK rq(0)->lock // orders against CPU0
2437 * UNLOCK rq(0)->lock
2441 * UNLOCK rq(1)->lock
2443 * LOCK rq(1)->lock // orders against CPU2
2446 * UNLOCK rq(1)->lock
2449 * BLOCKING -- aka. SLEEP + WAKEUP
2451 * For blocking we (obviously) need to provide the same guarantee as for
2452 * migration. However the means are completely different as there is no lock
2453 * chain to provide order. Instead we do:
2455 * 1) smp_store_release(X->on_cpu, 0)
2456 * 2) smp_cond_load_acquire(!X->on_cpu)
2460 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2462 * LOCK rq(0)->lock LOCK X->pi_lock
2465 * smp_store_release(X->on_cpu, 0);
2467 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2473 * X->state = RUNNING
2474 * UNLOCK rq(2)->lock
2476 * LOCK rq(2)->lock // orders against CPU1
2479 * UNLOCK rq(2)->lock
2482 * UNLOCK rq(0)->lock
2485 * However, for wakeups there is a second guarantee we must provide, namely we
2486 * must ensure that CONDITION=1 done by the caller can not be reordered with
2487 * accesses to the task state; see try_to_wake_up() and set_current_state().
2491 * try_to_wake_up - wake up a thread
2492 * @p: the thread to be awakened
2493 * @state: the mask of task states that can be woken
2494 * @wake_flags: wake modifier flags (WF_*)
2496 * If (@state & @p->state) @p->state = TASK_RUNNING.
2498 * If the task was not queued/runnable, also place it back on a runqueue.
2500 * Atomic against schedule() which would dequeue a task, also see
2501 * set_current_state().
2503 * This function executes a full memory barrier before accessing the task
2504 * state; see set_current_state().
2506 * Return: %true if @p->state changes (an actual wakeup was done),
2510 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2512 unsigned long flags
;
2513 int cpu
, success
= 0;
2518 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2519 * == smp_processor_id()'. Together this means we can special
2520 * case the whole 'p->on_rq && ttwu_remote()' case below
2521 * without taking any locks.
2524 * - we rely on Program-Order guarantees for all the ordering,
2525 * - we're serialized against set_special_state() by virtue of
2526 * it disabling IRQs (this allows not taking ->pi_lock).
2528 if (!(p
->state
& state
))
2533 trace_sched_waking(p
);
2534 p
->state
= TASK_RUNNING
;
2535 trace_sched_wakeup(p
);
2540 * If we are going to wake up a thread waiting for CONDITION we
2541 * need to ensure that CONDITION=1 done by the caller can not be
2542 * reordered with p->state check below. This pairs with mb() in
2543 * set_current_state() the waiting thread does.
2545 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2546 smp_mb__after_spinlock();
2547 if (!(p
->state
& state
))
2550 trace_sched_waking(p
);
2552 /* We're going to change ->state: */
2557 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2558 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2559 * in smp_cond_load_acquire() below.
2561 * sched_ttwu_pending() try_to_wake_up()
2562 * STORE p->on_rq = 1 LOAD p->state
2565 * __schedule() (switch to task 'p')
2566 * LOCK rq->lock smp_rmb();
2567 * smp_mb__after_spinlock();
2571 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2573 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2574 * __schedule(). See the comment for smp_mb__after_spinlock().
2576 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2579 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2583 delayacct_blkio_end(p
);
2584 atomic_dec(&task_rq(p
)->nr_iowait
);
2588 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2589 p
->state
= TASK_WAKING
;
2592 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2593 * possible to, falsely, observe p->on_cpu == 0.
2595 * One must be running (->on_cpu == 1) in order to remove oneself
2596 * from the runqueue.
2598 * __schedule() (switch to task 'p') try_to_wake_up()
2599 * STORE p->on_cpu = 1 LOAD p->on_rq
2602 * __schedule() (put 'p' to sleep)
2603 * LOCK rq->lock smp_rmb();
2604 * smp_mb__after_spinlock();
2605 * STORE p->on_rq = 0 LOAD p->on_cpu
2607 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2608 * __schedule(). See the comment for smp_mb__after_spinlock().
2613 * If the owning (remote) CPU is still in the middle of schedule() with
2614 * this task as prev, considering queueing p on the remote CPUs wake_list
2615 * which potentially sends an IPI instead of spinning on p->on_cpu to
2616 * let the waker make forward progress. This is safe because IRQs are
2617 * disabled and the IPI will deliver after on_cpu is cleared.
2619 if (READ_ONCE(p
->on_cpu
) && ttwu_queue_wakelist(p
, cpu
, wake_flags
| WF_ON_RQ
))
2623 * If the owning (remote) CPU is still in the middle of schedule() with
2624 * this task as prev, wait until its done referencing the task.
2626 * Pairs with the smp_store_release() in finish_task().
2628 * This ensures that tasks getting woken will be fully ordered against
2629 * their previous state and preserve Program Order.
2631 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2633 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2634 if (task_cpu(p
) != cpu
) {
2635 wake_flags
|= WF_MIGRATED
;
2636 psi_ttwu_dequeue(p
);
2637 set_task_cpu(p
, cpu
);
2639 #endif /* CONFIG_SMP */
2641 ttwu_queue(p
, cpu
, wake_flags
);
2643 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2646 ttwu_stat(p
, cpu
, wake_flags
);
2653 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2654 * @p: Process for which the function is to be invoked.
2655 * @func: Function to invoke.
2656 * @arg: Argument to function.
2658 * If the specified task can be quickly locked into a definite state
2659 * (either sleeping or on a given runqueue), arrange to keep it in that
2660 * state while invoking @func(@arg). This function can use ->on_rq and
2661 * task_curr() to work out what the state is, if required. Given that
2662 * @func can be invoked with a runqueue lock held, it had better be quite
2666 * @false if the task slipped out from under the locks.
2667 * @true if the task was locked onto a runqueue or is sleeping.
2668 * However, @func can override this by returning @false.
2670 bool try_invoke_on_locked_down_task(struct task_struct
*p
, bool (*func
)(struct task_struct
*t
, void *arg
), void *arg
)
2676 lockdep_assert_irqs_enabled();
2677 raw_spin_lock_irq(&p
->pi_lock
);
2679 rq
= __task_rq_lock(p
, &rf
);
2680 if (task_rq(p
) == rq
)
2689 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
2694 raw_spin_unlock_irq(&p
->pi_lock
);
2699 * wake_up_process - Wake up a specific process
2700 * @p: The process to be woken up.
2702 * Attempt to wake up the nominated process and move it to the set of runnable
2705 * Return: 1 if the process was woken up, 0 if it was already running.
2707 * This function executes a full memory barrier before accessing the task state.
2709 int wake_up_process(struct task_struct
*p
)
2711 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2713 EXPORT_SYMBOL(wake_up_process
);
2715 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2717 return try_to_wake_up(p
, state
, 0);
2721 * Perform scheduler related setup for a newly forked process p.
2722 * p is forked by current.
2724 * __sched_fork() is basic setup used by init_idle() too:
2726 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2731 p
->se
.exec_start
= 0;
2732 p
->se
.sum_exec_runtime
= 0;
2733 p
->se
.prev_sum_exec_runtime
= 0;
2734 p
->se
.nr_migrations
= 0;
2736 INIT_LIST_HEAD(&p
->se
.group_node
);
2738 #ifdef CONFIG_FAIR_GROUP_SCHED
2739 p
->se
.cfs_rq
= NULL
;
2742 #ifdef CONFIG_SCHEDSTATS
2743 /* Even if schedstat is disabled, there should not be garbage */
2744 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2747 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2748 init_dl_task_timer(&p
->dl
);
2749 init_dl_inactive_task_timer(&p
->dl
);
2750 __dl_clear_params(p
);
2752 INIT_LIST_HEAD(&p
->rt
.run_list
);
2754 p
->rt
.time_slice
= sched_rr_timeslice
;
2758 #ifdef CONFIG_PREEMPT_NOTIFIERS
2759 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2762 #ifdef CONFIG_COMPACTION
2763 p
->capture_control
= NULL
;
2765 init_numa_balancing(clone_flags
, p
);
2767 p
->wake_entry_type
= CSD_TYPE_TTWU
;
2771 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2773 #ifdef CONFIG_NUMA_BALANCING
2775 void set_numabalancing_state(bool enabled
)
2778 static_branch_enable(&sched_numa_balancing
);
2780 static_branch_disable(&sched_numa_balancing
);
2783 #ifdef CONFIG_PROC_SYSCTL
2784 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2785 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2789 int state
= static_branch_likely(&sched_numa_balancing
);
2791 if (write
&& !capable(CAP_SYS_ADMIN
))
2796 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2800 set_numabalancing_state(state
);
2806 #ifdef CONFIG_SCHEDSTATS
2808 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2809 static bool __initdata __sched_schedstats
= false;
2811 static void set_schedstats(bool enabled
)
2814 static_branch_enable(&sched_schedstats
);
2816 static_branch_disable(&sched_schedstats
);
2819 void force_schedstat_enabled(void)
2821 if (!schedstat_enabled()) {
2822 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2823 static_branch_enable(&sched_schedstats
);
2827 static int __init
setup_schedstats(char *str
)
2834 * This code is called before jump labels have been set up, so we can't
2835 * change the static branch directly just yet. Instead set a temporary
2836 * variable so init_schedstats() can do it later.
2838 if (!strcmp(str
, "enable")) {
2839 __sched_schedstats
= true;
2841 } else if (!strcmp(str
, "disable")) {
2842 __sched_schedstats
= false;
2847 pr_warn("Unable to parse schedstats=\n");
2851 __setup("schedstats=", setup_schedstats
);
2853 static void __init
init_schedstats(void)
2855 set_schedstats(__sched_schedstats
);
2858 #ifdef CONFIG_PROC_SYSCTL
2859 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2860 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2864 int state
= static_branch_likely(&sched_schedstats
);
2866 if (write
&& !capable(CAP_SYS_ADMIN
))
2871 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2875 set_schedstats(state
);
2878 #endif /* CONFIG_PROC_SYSCTL */
2879 #else /* !CONFIG_SCHEDSTATS */
2880 static inline void init_schedstats(void) {}
2881 #endif /* CONFIG_SCHEDSTATS */
2884 * fork()/clone()-time setup:
2886 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2888 unsigned long flags
;
2890 __sched_fork(clone_flags
, p
);
2892 * We mark the process as NEW here. This guarantees that
2893 * nobody will actually run it, and a signal or other external
2894 * event cannot wake it up and insert it on the runqueue either.
2896 p
->state
= TASK_NEW
;
2899 * Make sure we do not leak PI boosting priority to the child.
2901 p
->prio
= current
->normal_prio
;
2906 * Revert to default priority/policy on fork if requested.
2908 if (unlikely(p
->sched_reset_on_fork
)) {
2909 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2910 p
->policy
= SCHED_NORMAL
;
2911 p
->static_prio
= NICE_TO_PRIO(0);
2913 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2914 p
->static_prio
= NICE_TO_PRIO(0);
2916 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2917 set_load_weight(p
, false);
2920 * We don't need the reset flag anymore after the fork. It has
2921 * fulfilled its duty:
2923 p
->sched_reset_on_fork
= 0;
2926 if (dl_prio(p
->prio
))
2928 else if (rt_prio(p
->prio
))
2929 p
->sched_class
= &rt_sched_class
;
2931 p
->sched_class
= &fair_sched_class
;
2933 init_entity_runnable_average(&p
->se
);
2936 * The child is not yet in the pid-hash so no cgroup attach races,
2937 * and the cgroup is pinned to this child due to cgroup_fork()
2938 * is ran before sched_fork().
2940 * Silence PROVE_RCU.
2942 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2944 * We're setting the CPU for the first time, we don't migrate,
2945 * so use __set_task_cpu().
2947 __set_task_cpu(p
, smp_processor_id());
2948 if (p
->sched_class
->task_fork
)
2949 p
->sched_class
->task_fork(p
);
2950 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2952 #ifdef CONFIG_SCHED_INFO
2953 if (likely(sched_info_on()))
2954 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2956 #if defined(CONFIG_SMP)
2959 init_task_preempt_count(p
);
2961 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2962 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2967 unsigned long to_ratio(u64 period
, u64 runtime
)
2969 if (runtime
== RUNTIME_INF
)
2973 * Doing this here saves a lot of checks in all
2974 * the calling paths, and returning zero seems
2975 * safe for them anyway.
2980 return div64_u64(runtime
<< BW_SHIFT
, period
);
2984 * wake_up_new_task - wake up a newly created task for the first time.
2986 * This function will do some initial scheduler statistics housekeeping
2987 * that must be done for every newly created context, then puts the task
2988 * on the runqueue and wakes it.
2990 void wake_up_new_task(struct task_struct
*p
)
2995 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2996 p
->state
= TASK_RUNNING
;
2999 * Fork balancing, do it here and not earlier because:
3000 * - cpus_ptr can change in the fork path
3001 * - any previously selected CPU might disappear through hotplug
3003 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3004 * as we're not fully set-up yet.
3006 p
->recent_used_cpu
= task_cpu(p
);
3007 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
3009 rq
= __task_rq_lock(p
, &rf
);
3010 update_rq_clock(rq
);
3011 post_init_entity_util_avg(p
);
3013 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
3014 trace_sched_wakeup_new(p
);
3015 check_preempt_curr(rq
, p
, WF_FORK
);
3017 if (p
->sched_class
->task_woken
) {
3019 * Nothing relies on rq->lock after this, so its fine to
3022 rq_unpin_lock(rq
, &rf
);
3023 p
->sched_class
->task_woken(rq
, p
);
3024 rq_repin_lock(rq
, &rf
);
3027 task_rq_unlock(rq
, p
, &rf
);
3030 #ifdef CONFIG_PREEMPT_NOTIFIERS
3032 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
3034 void preempt_notifier_inc(void)
3036 static_branch_inc(&preempt_notifier_key
);
3038 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
3040 void preempt_notifier_dec(void)
3042 static_branch_dec(&preempt_notifier_key
);
3044 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3047 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3048 * @notifier: notifier struct to register
3050 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3052 if (!static_branch_unlikely(&preempt_notifier_key
))
3053 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3055 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3057 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3060 * preempt_notifier_unregister - no longer interested in preemption notifications
3061 * @notifier: notifier struct to unregister
3063 * This is *not* safe to call from within a preemption notifier.
3065 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3067 hlist_del(¬ifier
->link
);
3069 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3071 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3073 struct preempt_notifier
*notifier
;
3075 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3076 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3079 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3081 if (static_branch_unlikely(&preempt_notifier_key
))
3082 __fire_sched_in_preempt_notifiers(curr
);
3086 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3087 struct task_struct
*next
)
3089 struct preempt_notifier
*notifier
;
3091 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3092 notifier
->ops
->sched_out(notifier
, next
);
3095 static __always_inline
void
3096 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3097 struct task_struct
*next
)
3099 if (static_branch_unlikely(&preempt_notifier_key
))
3100 __fire_sched_out_preempt_notifiers(curr
, next
);
3103 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3105 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3110 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3111 struct task_struct
*next
)
3115 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3117 static inline void prepare_task(struct task_struct
*next
)
3121 * Claim the task as running, we do this before switching to it
3122 * such that any running task will have this set.
3128 static inline void finish_task(struct task_struct
*prev
)
3132 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3133 * We must ensure this doesn't happen until the switch is completely
3136 * In particular, the load of prev->state in finish_task_switch() must
3137 * happen before this.
3139 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3141 smp_store_release(&prev
->on_cpu
, 0);
3146 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
3149 * Since the runqueue lock will be released by the next
3150 * task (which is an invalid locking op but in the case
3151 * of the scheduler it's an obvious special-case), so we
3152 * do an early lockdep release here:
3154 rq_unpin_lock(rq
, rf
);
3155 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
3156 #ifdef CONFIG_DEBUG_SPINLOCK
3157 /* this is a valid case when another task releases the spinlock */
3158 rq
->lock
.owner
= next
;
3162 static inline void finish_lock_switch(struct rq
*rq
)
3165 * If we are tracking spinlock dependencies then we have to
3166 * fix up the runqueue lock - which gets 'carried over' from
3167 * prev into current:
3169 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
3170 raw_spin_unlock_irq(&rq
->lock
);
3174 * NOP if the arch has not defined these:
3177 #ifndef prepare_arch_switch
3178 # define prepare_arch_switch(next) do { } while (0)
3181 #ifndef finish_arch_post_lock_switch
3182 # define finish_arch_post_lock_switch() do { } while (0)
3186 * prepare_task_switch - prepare to switch tasks
3187 * @rq: the runqueue preparing to switch
3188 * @prev: the current task that is being switched out
3189 * @next: the task we are going to switch to.
3191 * This is called with the rq lock held and interrupts off. It must
3192 * be paired with a subsequent finish_task_switch after the context
3195 * prepare_task_switch sets up locking and calls architecture specific
3199 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3200 struct task_struct
*next
)
3202 kcov_prepare_switch(prev
);
3203 sched_info_switch(rq
, prev
, next
);
3204 perf_event_task_sched_out(prev
, next
);
3206 fire_sched_out_preempt_notifiers(prev
, next
);
3208 prepare_arch_switch(next
);
3212 * finish_task_switch - clean up after a task-switch
3213 * @prev: the thread we just switched away from.
3215 * finish_task_switch must be called after the context switch, paired
3216 * with a prepare_task_switch call before the context switch.
3217 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3218 * and do any other architecture-specific cleanup actions.
3220 * Note that we may have delayed dropping an mm in context_switch(). If
3221 * so, we finish that here outside of the runqueue lock. (Doing it
3222 * with the lock held can cause deadlocks; see schedule() for
3225 * The context switch have flipped the stack from under us and restored the
3226 * local variables which were saved when this task called schedule() in the
3227 * past. prev == current is still correct but we need to recalculate this_rq
3228 * because prev may have moved to another CPU.
3230 static struct rq
*finish_task_switch(struct task_struct
*prev
)
3231 __releases(rq
->lock
)
3233 struct rq
*rq
= this_rq();
3234 struct mm_struct
*mm
= rq
->prev_mm
;
3238 * The previous task will have left us with a preempt_count of 2
3239 * because it left us after:
3242 * preempt_disable(); // 1
3244 * raw_spin_lock_irq(&rq->lock) // 2
3246 * Also, see FORK_PREEMPT_COUNT.
3248 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
3249 "corrupted preempt_count: %s/%d/0x%x\n",
3250 current
->comm
, current
->pid
, preempt_count()))
3251 preempt_count_set(FORK_PREEMPT_COUNT
);
3256 * A task struct has one reference for the use as "current".
3257 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3258 * schedule one last time. The schedule call will never return, and
3259 * the scheduled task must drop that reference.
3261 * We must observe prev->state before clearing prev->on_cpu (in
3262 * finish_task), otherwise a concurrent wakeup can get prev
3263 * running on another CPU and we could rave with its RUNNING -> DEAD
3264 * transition, resulting in a double drop.
3266 prev_state
= prev
->state
;
3267 vtime_task_switch(prev
);
3268 perf_event_task_sched_in(prev
, current
);
3270 finish_lock_switch(rq
);
3271 finish_arch_post_lock_switch();
3272 kcov_finish_switch(current
);
3274 fire_sched_in_preempt_notifiers(current
);
3276 * When switching through a kernel thread, the loop in
3277 * membarrier_{private,global}_expedited() may have observed that
3278 * kernel thread and not issued an IPI. It is therefore possible to
3279 * schedule between user->kernel->user threads without passing though
3280 * switch_mm(). Membarrier requires a barrier after storing to
3281 * rq->curr, before returning to userspace, so provide them here:
3283 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3284 * provided by mmdrop(),
3285 * - a sync_core for SYNC_CORE.
3288 membarrier_mm_sync_core_before_usermode(mm
);
3291 if (unlikely(prev_state
== TASK_DEAD
)) {
3292 if (prev
->sched_class
->task_dead
)
3293 prev
->sched_class
->task_dead(prev
);
3296 * Remove function-return probe instances associated with this
3297 * task and put them back on the free list.
3299 kprobe_flush_task(prev
);
3301 /* Task is done with its stack. */
3302 put_task_stack(prev
);
3304 put_task_struct_rcu_user(prev
);
3307 tick_nohz_task_switch();
3313 /* rq->lock is NOT held, but preemption is disabled */
3314 static void __balance_callback(struct rq
*rq
)
3316 struct callback_head
*head
, *next
;
3317 void (*func
)(struct rq
*rq
);
3318 unsigned long flags
;
3320 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3321 head
= rq
->balance_callback
;
3322 rq
->balance_callback
= NULL
;
3324 func
= (void (*)(struct rq
*))head
->func
;
3331 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3334 static inline void balance_callback(struct rq
*rq
)
3336 if (unlikely(rq
->balance_callback
))
3337 __balance_callback(rq
);
3342 static inline void balance_callback(struct rq
*rq
)
3349 * schedule_tail - first thing a freshly forked thread must call.
3350 * @prev: the thread we just switched away from.
3352 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
3353 __releases(rq
->lock
)
3358 * New tasks start with FORK_PREEMPT_COUNT, see there and
3359 * finish_task_switch() for details.
3361 * finish_task_switch() will drop rq->lock() and lower preempt_count
3362 * and the preempt_enable() will end up enabling preemption (on
3363 * PREEMPT_COUNT kernels).
3366 rq
= finish_task_switch(prev
);
3367 balance_callback(rq
);
3370 if (current
->set_child_tid
)
3371 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3373 calculate_sigpending();
3377 * context_switch - switch to the new MM and the new thread's register state.
3379 static __always_inline
struct rq
*
3380 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3381 struct task_struct
*next
, struct rq_flags
*rf
)
3383 prepare_task_switch(rq
, prev
, next
);
3386 * For paravirt, this is coupled with an exit in switch_to to
3387 * combine the page table reload and the switch backend into
3390 arch_start_context_switch(prev
);
3393 * kernel -> kernel lazy + transfer active
3394 * user -> kernel lazy + mmgrab() active
3396 * kernel -> user switch + mmdrop() active
3397 * user -> user switch
3399 if (!next
->mm
) { // to kernel
3400 enter_lazy_tlb(prev
->active_mm
, next
);
3402 next
->active_mm
= prev
->active_mm
;
3403 if (prev
->mm
) // from user
3404 mmgrab(prev
->active_mm
);
3406 prev
->active_mm
= NULL
;
3408 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
3410 * sys_membarrier() requires an smp_mb() between setting
3411 * rq->curr / membarrier_switch_mm() and returning to userspace.
3413 * The below provides this either through switch_mm(), or in
3414 * case 'prev->active_mm == next->mm' through
3415 * finish_task_switch()'s mmdrop().
3417 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
3419 if (!prev
->mm
) { // from kernel
3420 /* will mmdrop() in finish_task_switch(). */
3421 rq
->prev_mm
= prev
->active_mm
;
3422 prev
->active_mm
= NULL
;
3426 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3428 prepare_lock_switch(rq
, next
, rf
);
3430 /* Here we just switch the register state and the stack. */
3431 switch_to(prev
, next
, prev
);
3434 return finish_task_switch(prev
);
3438 * nr_running and nr_context_switches:
3440 * externally visible scheduler statistics: current number of runnable
3441 * threads, total number of context switches performed since bootup.
3443 unsigned long nr_running(void)
3445 unsigned long i
, sum
= 0;
3447 for_each_online_cpu(i
)
3448 sum
+= cpu_rq(i
)->nr_running
;
3454 * Check if only the current task is running on the CPU.
3456 * Caution: this function does not check that the caller has disabled
3457 * preemption, thus the result might have a time-of-check-to-time-of-use
3458 * race. The caller is responsible to use it correctly, for example:
3460 * - from a non-preemptible section (of course)
3462 * - from a thread that is bound to a single CPU
3464 * - in a loop with very short iterations (e.g. a polling loop)
3466 bool single_task_running(void)
3468 return raw_rq()->nr_running
== 1;
3470 EXPORT_SYMBOL(single_task_running
);
3472 unsigned long long nr_context_switches(void)
3475 unsigned long long sum
= 0;
3477 for_each_possible_cpu(i
)
3478 sum
+= cpu_rq(i
)->nr_switches
;
3484 * Consumers of these two interfaces, like for example the cpuidle menu
3485 * governor, are using nonsensical data. Preferring shallow idle state selection
3486 * for a CPU that has IO-wait which might not even end up running the task when
3487 * it does become runnable.
3490 unsigned long nr_iowait_cpu(int cpu
)
3492 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
3496 * IO-wait accounting, and how its mostly bollocks (on SMP).
3498 * The idea behind IO-wait account is to account the idle time that we could
3499 * have spend running if it were not for IO. That is, if we were to improve the
3500 * storage performance, we'd have a proportional reduction in IO-wait time.
3502 * This all works nicely on UP, where, when a task blocks on IO, we account
3503 * idle time as IO-wait, because if the storage were faster, it could've been
3504 * running and we'd not be idle.
3506 * This has been extended to SMP, by doing the same for each CPU. This however
3509 * Imagine for instance the case where two tasks block on one CPU, only the one
3510 * CPU will have IO-wait accounted, while the other has regular idle. Even
3511 * though, if the storage were faster, both could've ran at the same time,
3512 * utilising both CPUs.
3514 * This means, that when looking globally, the current IO-wait accounting on
3515 * SMP is a lower bound, by reason of under accounting.
3517 * Worse, since the numbers are provided per CPU, they are sometimes
3518 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3519 * associated with any one particular CPU, it can wake to another CPU than it
3520 * blocked on. This means the per CPU IO-wait number is meaningless.
3522 * Task CPU affinities can make all that even more 'interesting'.
3525 unsigned long nr_iowait(void)
3527 unsigned long i
, sum
= 0;
3529 for_each_possible_cpu(i
)
3530 sum
+= nr_iowait_cpu(i
);
3538 * sched_exec - execve() is a valuable balancing opportunity, because at
3539 * this point the task has the smallest effective memory and cache footprint.
3541 void sched_exec(void)
3543 struct task_struct
*p
= current
;
3544 unsigned long flags
;
3547 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3548 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
3549 if (dest_cpu
== smp_processor_id())
3552 if (likely(cpu_active(dest_cpu
))) {
3553 struct migration_arg arg
= { p
, dest_cpu
};
3555 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3556 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3560 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3565 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3566 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3568 EXPORT_PER_CPU_SYMBOL(kstat
);
3569 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3572 * The function fair_sched_class.update_curr accesses the struct curr
3573 * and its field curr->exec_start; when called from task_sched_runtime(),
3574 * we observe a high rate of cache misses in practice.
3575 * Prefetching this data results in improved performance.
3577 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3579 #ifdef CONFIG_FAIR_GROUP_SCHED
3580 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3582 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3585 prefetch(&curr
->exec_start
);
3589 * Return accounted runtime for the task.
3590 * In case the task is currently running, return the runtime plus current's
3591 * pending runtime that have not been accounted yet.
3593 unsigned long long task_sched_runtime(struct task_struct
*p
)
3599 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3601 * 64-bit doesn't need locks to atomically read a 64-bit value.
3602 * So we have a optimization chance when the task's delta_exec is 0.
3603 * Reading ->on_cpu is racy, but this is ok.
3605 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3606 * If we race with it entering CPU, unaccounted time is 0. This is
3607 * indistinguishable from the read occurring a few cycles earlier.
3608 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3609 * been accounted, so we're correct here as well.
3611 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3612 return p
->se
.sum_exec_runtime
;
3615 rq
= task_rq_lock(p
, &rf
);
3617 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3618 * project cycles that may never be accounted to this
3619 * thread, breaking clock_gettime().
3621 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3622 prefetch_curr_exec_start(p
);
3623 update_rq_clock(rq
);
3624 p
->sched_class
->update_curr(rq
);
3626 ns
= p
->se
.sum_exec_runtime
;
3627 task_rq_unlock(rq
, p
, &rf
);
3632 DEFINE_PER_CPU(unsigned long, thermal_pressure
);
3634 void arch_set_thermal_pressure(struct cpumask
*cpus
,
3635 unsigned long th_pressure
)
3639 for_each_cpu(cpu
, cpus
)
3640 WRITE_ONCE(per_cpu(thermal_pressure
, cpu
), th_pressure
);
3644 * This function gets called by the timer code, with HZ frequency.
3645 * We call it with interrupts disabled.
3647 void scheduler_tick(void)
3649 int cpu
= smp_processor_id();
3650 struct rq
*rq
= cpu_rq(cpu
);
3651 struct task_struct
*curr
= rq
->curr
;
3653 unsigned long thermal_pressure
;
3655 arch_scale_freq_tick();
3660 update_rq_clock(rq
);
3661 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
3662 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
3663 curr
->sched_class
->task_tick(rq
, curr
, 0);
3664 calc_global_load_tick(rq
);
3669 perf_event_task_tick();
3672 rq
->idle_balance
= idle_cpu(cpu
);
3673 trigger_load_balance(rq
);
3677 #ifdef CONFIG_NO_HZ_FULL
3682 struct delayed_work work
;
3684 /* Values for ->state, see diagram below. */
3685 #define TICK_SCHED_REMOTE_OFFLINE 0
3686 #define TICK_SCHED_REMOTE_OFFLINING 1
3687 #define TICK_SCHED_REMOTE_RUNNING 2
3690 * State diagram for ->state:
3693 * TICK_SCHED_REMOTE_OFFLINE
3696 * | | sched_tick_remote()
3699 * +--TICK_SCHED_REMOTE_OFFLINING
3702 * sched_tick_start() | | sched_tick_stop()
3705 * TICK_SCHED_REMOTE_RUNNING
3708 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3709 * and sched_tick_start() are happy to leave the state in RUNNING.
3712 static struct tick_work __percpu
*tick_work_cpu
;
3714 static void sched_tick_remote(struct work_struct
*work
)
3716 struct delayed_work
*dwork
= to_delayed_work(work
);
3717 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3718 int cpu
= twork
->cpu
;
3719 struct rq
*rq
= cpu_rq(cpu
);
3720 struct task_struct
*curr
;
3726 * Handle the tick only if it appears the remote CPU is running in full
3727 * dynticks mode. The check is racy by nature, but missing a tick or
3728 * having one too much is no big deal because the scheduler tick updates
3729 * statistics and checks timeslices in a time-independent way, regardless
3730 * of when exactly it is running.
3732 if (!tick_nohz_tick_stopped_cpu(cpu
))
3735 rq_lock_irq(rq
, &rf
);
3737 if (cpu_is_offline(cpu
))
3740 update_rq_clock(rq
);
3742 if (!is_idle_task(curr
)) {
3744 * Make sure the next tick runs within a reasonable
3747 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3748 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3750 curr
->sched_class
->task_tick(rq
, curr
, 0);
3752 calc_load_nohz_remote(rq
);
3754 rq_unlock_irq(rq
, &rf
);
3758 * Run the remote tick once per second (1Hz). This arbitrary
3759 * frequency is large enough to avoid overload but short enough
3760 * to keep scheduler internal stats reasonably up to date. But
3761 * first update state to reflect hotplug activity if required.
3763 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
3764 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
3765 if (os
== TICK_SCHED_REMOTE_RUNNING
)
3766 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3769 static void sched_tick_start(int cpu
)
3772 struct tick_work
*twork
;
3774 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3777 WARN_ON_ONCE(!tick_work_cpu
);
3779 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3780 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
3781 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
3782 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
3784 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3785 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3789 #ifdef CONFIG_HOTPLUG_CPU
3790 static void sched_tick_stop(int cpu
)
3792 struct tick_work
*twork
;
3795 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3798 WARN_ON_ONCE(!tick_work_cpu
);
3800 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3801 /* There cannot be competing actions, but don't rely on stop-machine. */
3802 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
3803 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
3804 /* Don't cancel, as this would mess up the state machine. */
3806 #endif /* CONFIG_HOTPLUG_CPU */
3808 int __init
sched_tick_offload_init(void)
3810 tick_work_cpu
= alloc_percpu(struct tick_work
);
3811 BUG_ON(!tick_work_cpu
);
3815 #else /* !CONFIG_NO_HZ_FULL */
3816 static inline void sched_tick_start(int cpu
) { }
3817 static inline void sched_tick_stop(int cpu
) { }
3820 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3821 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3823 * If the value passed in is equal to the current preempt count
3824 * then we just disabled preemption. Start timing the latency.
3826 static inline void preempt_latency_start(int val
)
3828 if (preempt_count() == val
) {
3829 unsigned long ip
= get_lock_parent_ip();
3830 #ifdef CONFIG_DEBUG_PREEMPT
3831 current
->preempt_disable_ip
= ip
;
3833 trace_preempt_off(CALLER_ADDR0
, ip
);
3837 void preempt_count_add(int val
)
3839 #ifdef CONFIG_DEBUG_PREEMPT
3843 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3846 __preempt_count_add(val
);
3847 #ifdef CONFIG_DEBUG_PREEMPT
3849 * Spinlock count overflowing soon?
3851 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3854 preempt_latency_start(val
);
3856 EXPORT_SYMBOL(preempt_count_add
);
3857 NOKPROBE_SYMBOL(preempt_count_add
);
3860 * If the value passed in equals to the current preempt count
3861 * then we just enabled preemption. Stop timing the latency.
3863 static inline void preempt_latency_stop(int val
)
3865 if (preempt_count() == val
)
3866 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3869 void preempt_count_sub(int val
)
3871 #ifdef CONFIG_DEBUG_PREEMPT
3875 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3878 * Is the spinlock portion underflowing?
3880 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3881 !(preempt_count() & PREEMPT_MASK
)))
3885 preempt_latency_stop(val
);
3886 __preempt_count_sub(val
);
3888 EXPORT_SYMBOL(preempt_count_sub
);
3889 NOKPROBE_SYMBOL(preempt_count_sub
);
3892 static inline void preempt_latency_start(int val
) { }
3893 static inline void preempt_latency_stop(int val
) { }
3896 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3898 #ifdef CONFIG_DEBUG_PREEMPT
3899 return p
->preempt_disable_ip
;
3906 * Print scheduling while atomic bug:
3908 static noinline
void __schedule_bug(struct task_struct
*prev
)
3910 /* Save this before calling printk(), since that will clobber it */
3911 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3913 if (oops_in_progress
)
3916 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3917 prev
->comm
, prev
->pid
, preempt_count());
3919 debug_show_held_locks(prev
);
3921 if (irqs_disabled())
3922 print_irqtrace_events(prev
);
3923 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3924 && in_atomic_preempt_off()) {
3925 pr_err("Preemption disabled at:");
3926 print_ip_sym(preempt_disable_ip
);
3930 panic("scheduling while atomic\n");
3933 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3937 * Various schedule()-time debugging checks and statistics:
3939 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
3941 #ifdef CONFIG_SCHED_STACK_END_CHECK
3942 if (task_stack_end_corrupted(prev
))
3943 panic("corrupted stack end detected inside scheduler\n");
3945 if (task_scs_end_corrupted(prev
))
3946 panic("corrupted shadow stack detected inside scheduler\n");
3949 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3950 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
3951 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3952 prev
->comm
, prev
->pid
, prev
->non_block_count
);
3954 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3958 if (unlikely(in_atomic_preempt_off())) {
3959 __schedule_bug(prev
);
3960 preempt_count_set(PREEMPT_DISABLED
);
3964 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3966 schedstat_inc(this_rq()->sched_count
);
3969 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
3970 struct rq_flags
*rf
)
3973 const struct sched_class
*class;
3975 * We must do the balancing pass before put_prev_task(), such
3976 * that when we release the rq->lock the task is in the same
3977 * state as before we took rq->lock.
3979 * We can terminate the balance pass as soon as we know there is
3980 * a runnable task of @class priority or higher.
3982 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
3983 if (class->balance(rq
, prev
, rf
))
3988 put_prev_task(rq
, prev
);
3992 * Pick up the highest-prio task:
3994 static inline struct task_struct
*
3995 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3997 const struct sched_class
*class;
3998 struct task_struct
*p
;
4001 * Optimization: we know that if all tasks are in the fair class we can
4002 * call that function directly, but only if the @prev task wasn't of a
4003 * higher scheduling class, because otherwise those loose the
4004 * opportunity to pull in more work from other CPUs.
4006 if (likely((prev
->sched_class
== &idle_sched_class
||
4007 prev
->sched_class
== &fair_sched_class
) &&
4008 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
4010 p
= pick_next_task_fair(rq
, prev
, rf
);
4011 if (unlikely(p
== RETRY_TASK
))
4014 /* Assumes fair_sched_class->next == idle_sched_class */
4016 put_prev_task(rq
, prev
);
4017 p
= pick_next_task_idle(rq
);
4024 put_prev_task_balance(rq
, prev
, rf
);
4026 for_each_class(class) {
4027 p
= class->pick_next_task(rq
);
4032 /* The idle class should always have a runnable task: */
4037 * __schedule() is the main scheduler function.
4039 * The main means of driving the scheduler and thus entering this function are:
4041 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4043 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4044 * paths. For example, see arch/x86/entry_64.S.
4046 * To drive preemption between tasks, the scheduler sets the flag in timer
4047 * interrupt handler scheduler_tick().
4049 * 3. Wakeups don't really cause entry into schedule(). They add a
4050 * task to the run-queue and that's it.
4052 * Now, if the new task added to the run-queue preempts the current
4053 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4054 * called on the nearest possible occasion:
4056 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4058 * - in syscall or exception context, at the next outmost
4059 * preempt_enable(). (this might be as soon as the wake_up()'s
4062 * - in IRQ context, return from interrupt-handler to
4063 * preemptible context
4065 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4068 * - cond_resched() call
4069 * - explicit schedule() call
4070 * - return from syscall or exception to user-space
4071 * - return from interrupt-handler to user-space
4073 * WARNING: must be called with preemption disabled!
4075 static void __sched notrace
__schedule(bool preempt
)
4077 struct task_struct
*prev
, *next
;
4078 unsigned long *switch_count
;
4083 cpu
= smp_processor_id();
4087 schedule_debug(prev
, preempt
);
4089 if (sched_feat(HRTICK
))
4092 local_irq_disable();
4093 rcu_note_context_switch(preempt
);
4096 * Make sure that signal_pending_state()->signal_pending() below
4097 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4098 * done by the caller to avoid the race with signal_wake_up().
4100 * The membarrier system call requires a full memory barrier
4101 * after coming from user-space, before storing to rq->curr.
4104 smp_mb__after_spinlock();
4106 /* Promote REQ to ACT */
4107 rq
->clock_update_flags
<<= 1;
4108 update_rq_clock(rq
);
4110 switch_count
= &prev
->nivcsw
;
4111 if (!preempt
&& prev
->state
) {
4112 if (signal_pending_state(prev
->state
, prev
)) {
4113 prev
->state
= TASK_RUNNING
;
4115 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
4117 if (prev
->in_iowait
) {
4118 atomic_inc(&rq
->nr_iowait
);
4119 delayacct_blkio_start();
4122 switch_count
= &prev
->nvcsw
;
4125 next
= pick_next_task(rq
, prev
, &rf
);
4126 clear_tsk_need_resched(prev
);
4127 clear_preempt_need_resched();
4129 if (likely(prev
!= next
)) {
4132 * RCU users of rcu_dereference(rq->curr) may not see
4133 * changes to task_struct made by pick_next_task().
4135 RCU_INIT_POINTER(rq
->curr
, next
);
4137 * The membarrier system call requires each architecture
4138 * to have a full memory barrier after updating
4139 * rq->curr, before returning to user-space.
4141 * Here are the schemes providing that barrier on the
4142 * various architectures:
4143 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4144 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4145 * - finish_lock_switch() for weakly-ordered
4146 * architectures where spin_unlock is a full barrier,
4147 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4148 * is a RELEASE barrier),
4152 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
4154 trace_sched_switch(preempt
, prev
, next
);
4156 /* Also unlocks the rq: */
4157 rq
= context_switch(rq
, prev
, next
, &rf
);
4159 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4160 rq_unlock_irq(rq
, &rf
);
4163 balance_callback(rq
);
4166 void __noreturn
do_task_dead(void)
4168 /* Causes final put_task_struct in finish_task_switch(): */
4169 set_special_state(TASK_DEAD
);
4171 /* Tell freezer to ignore us: */
4172 current
->flags
|= PF_NOFREEZE
;
4177 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4182 static inline void sched_submit_work(struct task_struct
*tsk
)
4188 * If a worker went to sleep, notify and ask workqueue whether
4189 * it wants to wake up a task to maintain concurrency.
4190 * As this function is called inside the schedule() context,
4191 * we disable preemption to avoid it calling schedule() again
4192 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4195 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4197 if (tsk
->flags
& PF_WQ_WORKER
)
4198 wq_worker_sleeping(tsk
);
4200 io_wq_worker_sleeping(tsk
);
4201 preempt_enable_no_resched();
4204 if (tsk_is_pi_blocked(tsk
))
4208 * If we are going to sleep and we have plugged IO queued,
4209 * make sure to submit it to avoid deadlocks.
4211 if (blk_needs_flush_plug(tsk
))
4212 blk_schedule_flush_plug(tsk
);
4215 static void sched_update_worker(struct task_struct
*tsk
)
4217 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4218 if (tsk
->flags
& PF_WQ_WORKER
)
4219 wq_worker_running(tsk
);
4221 io_wq_worker_running(tsk
);
4225 asmlinkage __visible
void __sched
schedule(void)
4227 struct task_struct
*tsk
= current
;
4229 sched_submit_work(tsk
);
4233 sched_preempt_enable_no_resched();
4234 } while (need_resched());
4235 sched_update_worker(tsk
);
4237 EXPORT_SYMBOL(schedule
);
4240 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4241 * state (have scheduled out non-voluntarily) by making sure that all
4242 * tasks have either left the run queue or have gone into user space.
4243 * As idle tasks do not do either, they must not ever be preempted
4244 * (schedule out non-voluntarily).
4246 * schedule_idle() is similar to schedule_preempt_disable() except that it
4247 * never enables preemption because it does not call sched_submit_work().
4249 void __sched
schedule_idle(void)
4252 * As this skips calling sched_submit_work(), which the idle task does
4253 * regardless because that function is a nop when the task is in a
4254 * TASK_RUNNING state, make sure this isn't used someplace that the
4255 * current task can be in any other state. Note, idle is always in the
4256 * TASK_RUNNING state.
4258 WARN_ON_ONCE(current
->state
);
4261 } while (need_resched());
4264 #ifdef CONFIG_CONTEXT_TRACKING
4265 asmlinkage __visible
void __sched
schedule_user(void)
4268 * If we come here after a random call to set_need_resched(),
4269 * or we have been woken up remotely but the IPI has not yet arrived,
4270 * we haven't yet exited the RCU idle mode. Do it here manually until
4271 * we find a better solution.
4273 * NB: There are buggy callers of this function. Ideally we
4274 * should warn if prev_state != CONTEXT_USER, but that will trigger
4275 * too frequently to make sense yet.
4277 enum ctx_state prev_state
= exception_enter();
4279 exception_exit(prev_state
);
4284 * schedule_preempt_disabled - called with preemption disabled
4286 * Returns with preemption disabled. Note: preempt_count must be 1
4288 void __sched
schedule_preempt_disabled(void)
4290 sched_preempt_enable_no_resched();
4295 static void __sched notrace
preempt_schedule_common(void)
4299 * Because the function tracer can trace preempt_count_sub()
4300 * and it also uses preempt_enable/disable_notrace(), if
4301 * NEED_RESCHED is set, the preempt_enable_notrace() called
4302 * by the function tracer will call this function again and
4303 * cause infinite recursion.
4305 * Preemption must be disabled here before the function
4306 * tracer can trace. Break up preempt_disable() into two
4307 * calls. One to disable preemption without fear of being
4308 * traced. The other to still record the preemption latency,
4309 * which can also be traced by the function tracer.
4311 preempt_disable_notrace();
4312 preempt_latency_start(1);
4314 preempt_latency_stop(1);
4315 preempt_enable_no_resched_notrace();
4318 * Check again in case we missed a preemption opportunity
4319 * between schedule and now.
4321 } while (need_resched());
4324 #ifdef CONFIG_PREEMPTION
4326 * This is the entry point to schedule() from in-kernel preemption
4327 * off of preempt_enable.
4329 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
4332 * If there is a non-zero preempt_count or interrupts are disabled,
4333 * we do not want to preempt the current task. Just return..
4335 if (likely(!preemptible()))
4338 preempt_schedule_common();
4340 NOKPROBE_SYMBOL(preempt_schedule
);
4341 EXPORT_SYMBOL(preempt_schedule
);
4344 * preempt_schedule_notrace - preempt_schedule called by tracing
4346 * The tracing infrastructure uses preempt_enable_notrace to prevent
4347 * recursion and tracing preempt enabling caused by the tracing
4348 * infrastructure itself. But as tracing can happen in areas coming
4349 * from userspace or just about to enter userspace, a preempt enable
4350 * can occur before user_exit() is called. This will cause the scheduler
4351 * to be called when the system is still in usermode.
4353 * To prevent this, the preempt_enable_notrace will use this function
4354 * instead of preempt_schedule() to exit user context if needed before
4355 * calling the scheduler.
4357 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
4359 enum ctx_state prev_ctx
;
4361 if (likely(!preemptible()))
4366 * Because the function tracer can trace preempt_count_sub()
4367 * and it also uses preempt_enable/disable_notrace(), if
4368 * NEED_RESCHED is set, the preempt_enable_notrace() called
4369 * by the function tracer will call this function again and
4370 * cause infinite recursion.
4372 * Preemption must be disabled here before the function
4373 * tracer can trace. Break up preempt_disable() into two
4374 * calls. One to disable preemption without fear of being
4375 * traced. The other to still record the preemption latency,
4376 * which can also be traced by the function tracer.
4378 preempt_disable_notrace();
4379 preempt_latency_start(1);
4381 * Needs preempt disabled in case user_exit() is traced
4382 * and the tracer calls preempt_enable_notrace() causing
4383 * an infinite recursion.
4385 prev_ctx
= exception_enter();
4387 exception_exit(prev_ctx
);
4389 preempt_latency_stop(1);
4390 preempt_enable_no_resched_notrace();
4391 } while (need_resched());
4393 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
4395 #endif /* CONFIG_PREEMPTION */
4398 * This is the entry point to schedule() from kernel preemption
4399 * off of irq context.
4400 * Note, that this is called and return with irqs disabled. This will
4401 * protect us against recursive calling from irq.
4403 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
4405 enum ctx_state prev_state
;
4407 /* Catch callers which need to be fixed */
4408 BUG_ON(preempt_count() || !irqs_disabled());
4410 prev_state
= exception_enter();
4416 local_irq_disable();
4417 sched_preempt_enable_no_resched();
4418 } while (need_resched());
4420 exception_exit(prev_state
);
4423 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
4426 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4428 EXPORT_SYMBOL(default_wake_function
);
4430 #ifdef CONFIG_RT_MUTEXES
4432 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
4435 prio
= min(prio
, pi_task
->prio
);
4440 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4442 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
4444 return __rt_effective_prio(pi_task
, prio
);
4448 * rt_mutex_setprio - set the current priority of a task
4450 * @pi_task: donor task
4452 * This function changes the 'effective' priority of a task. It does
4453 * not touch ->normal_prio like __setscheduler().
4455 * Used by the rt_mutex code to implement priority inheritance
4456 * logic. Call site only calls if the priority of the task changed.
4458 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
4460 int prio
, oldprio
, queued
, running
, queue_flag
=
4461 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4462 const struct sched_class
*prev_class
;
4466 /* XXX used to be waiter->prio, not waiter->task->prio */
4467 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
4470 * If nothing changed; bail early.
4472 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
4475 rq
= __task_rq_lock(p
, &rf
);
4476 update_rq_clock(rq
);
4478 * Set under pi_lock && rq->lock, such that the value can be used under
4481 * Note that there is loads of tricky to make this pointer cache work
4482 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4483 * ensure a task is de-boosted (pi_task is set to NULL) before the
4484 * task is allowed to run again (and can exit). This ensures the pointer
4485 * points to a blocked task -- which guaratees the task is present.
4487 p
->pi_top_task
= pi_task
;
4490 * For FIFO/RR we only need to set prio, if that matches we're done.
4492 if (prio
== p
->prio
&& !dl_prio(prio
))
4496 * Idle task boosting is a nono in general. There is one
4497 * exception, when PREEMPT_RT and NOHZ is active:
4499 * The idle task calls get_next_timer_interrupt() and holds
4500 * the timer wheel base->lock on the CPU and another CPU wants
4501 * to access the timer (probably to cancel it). We can safely
4502 * ignore the boosting request, as the idle CPU runs this code
4503 * with interrupts disabled and will complete the lock
4504 * protected section without being interrupted. So there is no
4505 * real need to boost.
4507 if (unlikely(p
== rq
->idle
)) {
4508 WARN_ON(p
!= rq
->curr
);
4509 WARN_ON(p
->pi_blocked_on
);
4513 trace_sched_pi_setprio(p
, pi_task
);
4516 if (oldprio
== prio
)
4517 queue_flag
&= ~DEQUEUE_MOVE
;
4519 prev_class
= p
->sched_class
;
4520 queued
= task_on_rq_queued(p
);
4521 running
= task_current(rq
, p
);
4523 dequeue_task(rq
, p
, queue_flag
);
4525 put_prev_task(rq
, p
);
4528 * Boosting condition are:
4529 * 1. -rt task is running and holds mutex A
4530 * --> -dl task blocks on mutex A
4532 * 2. -dl task is running and holds mutex A
4533 * --> -dl task blocks on mutex A and could preempt the
4536 if (dl_prio(prio
)) {
4537 if (!dl_prio(p
->normal_prio
) ||
4538 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
4539 p
->dl
.dl_boosted
= 1;
4540 queue_flag
|= ENQUEUE_REPLENISH
;
4542 p
->dl
.dl_boosted
= 0;
4543 p
->sched_class
= &dl_sched_class
;
4544 } else if (rt_prio(prio
)) {
4545 if (dl_prio(oldprio
))
4546 p
->dl
.dl_boosted
= 0;
4548 queue_flag
|= ENQUEUE_HEAD
;
4549 p
->sched_class
= &rt_sched_class
;
4551 if (dl_prio(oldprio
))
4552 p
->dl
.dl_boosted
= 0;
4553 if (rt_prio(oldprio
))
4555 p
->sched_class
= &fair_sched_class
;
4561 enqueue_task(rq
, p
, queue_flag
);
4563 set_next_task(rq
, p
);
4565 check_class_changed(rq
, p
, prev_class
, oldprio
);
4567 /* Avoid rq from going away on us: */
4569 __task_rq_unlock(rq
, &rf
);
4571 balance_callback(rq
);
4575 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4581 void set_user_nice(struct task_struct
*p
, long nice
)
4583 bool queued
, running
;
4588 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
4591 * We have to be careful, if called from sys_setpriority(),
4592 * the task might be in the middle of scheduling on another CPU.
4594 rq
= task_rq_lock(p
, &rf
);
4595 update_rq_clock(rq
);
4598 * The RT priorities are set via sched_setscheduler(), but we still
4599 * allow the 'normal' nice value to be set - but as expected
4600 * it wont have any effect on scheduling until the task is
4601 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4603 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
4604 p
->static_prio
= NICE_TO_PRIO(nice
);
4607 queued
= task_on_rq_queued(p
);
4608 running
= task_current(rq
, p
);
4610 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
4612 put_prev_task(rq
, p
);
4614 p
->static_prio
= NICE_TO_PRIO(nice
);
4615 set_load_weight(p
, true);
4617 p
->prio
= effective_prio(p
);
4620 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
4622 set_next_task(rq
, p
);
4625 * If the task increased its priority or is running and
4626 * lowered its priority, then reschedule its CPU:
4628 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
4631 task_rq_unlock(rq
, p
, &rf
);
4633 EXPORT_SYMBOL(set_user_nice
);
4636 * can_nice - check if a task can reduce its nice value
4640 int can_nice(const struct task_struct
*p
, const int nice
)
4642 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4643 int nice_rlim
= nice_to_rlimit(nice
);
4645 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4646 capable(CAP_SYS_NICE
));
4649 #ifdef __ARCH_WANT_SYS_NICE
4652 * sys_nice - change the priority of the current process.
4653 * @increment: priority increment
4655 * sys_setpriority is a more generic, but much slower function that
4656 * does similar things.
4658 SYSCALL_DEFINE1(nice
, int, increment
)
4663 * Setpriority might change our priority at the same moment.
4664 * We don't have to worry. Conceptually one call occurs first
4665 * and we have a single winner.
4667 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
4668 nice
= task_nice(current
) + increment
;
4670 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
4671 if (increment
< 0 && !can_nice(current
, nice
))
4674 retval
= security_task_setnice(current
, nice
);
4678 set_user_nice(current
, nice
);
4685 * task_prio - return the priority value of a given task.
4686 * @p: the task in question.
4688 * Return: The priority value as seen by users in /proc.
4689 * RT tasks are offset by -200. Normal tasks are centered
4690 * around 0, value goes from -16 to +15.
4692 int task_prio(const struct task_struct
*p
)
4694 return p
->prio
- MAX_RT_PRIO
;
4698 * idle_cpu - is a given CPU idle currently?
4699 * @cpu: the processor in question.
4701 * Return: 1 if the CPU is currently idle. 0 otherwise.
4703 int idle_cpu(int cpu
)
4705 struct rq
*rq
= cpu_rq(cpu
);
4707 if (rq
->curr
!= rq
->idle
)
4714 if (rq
->ttwu_pending
)
4722 * available_idle_cpu - is a given CPU idle for enqueuing work.
4723 * @cpu: the CPU in question.
4725 * Return: 1 if the CPU is currently idle. 0 otherwise.
4727 int available_idle_cpu(int cpu
)
4732 if (vcpu_is_preempted(cpu
))
4739 * idle_task - return the idle task for a given CPU.
4740 * @cpu: the processor in question.
4742 * Return: The idle task for the CPU @cpu.
4744 struct task_struct
*idle_task(int cpu
)
4746 return cpu_rq(cpu
)->idle
;
4750 * find_process_by_pid - find a process with a matching PID value.
4751 * @pid: the pid in question.
4753 * The task of @pid, if found. %NULL otherwise.
4755 static struct task_struct
*find_process_by_pid(pid_t pid
)
4757 return pid
? find_task_by_vpid(pid
) : current
;
4761 * sched_setparam() passes in -1 for its policy, to let the functions
4762 * it calls know not to change it.
4764 #define SETPARAM_POLICY -1
4766 static void __setscheduler_params(struct task_struct
*p
,
4767 const struct sched_attr
*attr
)
4769 int policy
= attr
->sched_policy
;
4771 if (policy
== SETPARAM_POLICY
)
4776 if (dl_policy(policy
))
4777 __setparam_dl(p
, attr
);
4778 else if (fair_policy(policy
))
4779 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4782 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4783 * !rt_policy. Always setting this ensures that things like
4784 * getparam()/getattr() don't report silly values for !rt tasks.
4786 p
->rt_priority
= attr
->sched_priority
;
4787 p
->normal_prio
= normal_prio(p
);
4788 set_load_weight(p
, true);
4791 /* Actually do priority change: must hold pi & rq lock. */
4792 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4793 const struct sched_attr
*attr
, bool keep_boost
)
4796 * If params can't change scheduling class changes aren't allowed
4799 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
4802 __setscheduler_params(p
, attr
);
4805 * Keep a potential priority boosting if called from
4806 * sched_setscheduler().
4808 p
->prio
= normal_prio(p
);
4810 p
->prio
= rt_effective_prio(p
, p
->prio
);
4812 if (dl_prio(p
->prio
))
4813 p
->sched_class
= &dl_sched_class
;
4814 else if (rt_prio(p
->prio
))
4815 p
->sched_class
= &rt_sched_class
;
4817 p
->sched_class
= &fair_sched_class
;
4821 * Check the target process has a UID that matches the current process's:
4823 static bool check_same_owner(struct task_struct
*p
)
4825 const struct cred
*cred
= current_cred(), *pcred
;
4829 pcred
= __task_cred(p
);
4830 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4831 uid_eq(cred
->euid
, pcred
->uid
));
4836 static int __sched_setscheduler(struct task_struct
*p
,
4837 const struct sched_attr
*attr
,
4840 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4841 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4842 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4843 int new_effective_prio
, policy
= attr
->sched_policy
;
4844 const struct sched_class
*prev_class
;
4847 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4850 /* The pi code expects interrupts enabled */
4851 BUG_ON(pi
&& in_interrupt());
4853 /* Double check policy once rq lock held: */
4855 reset_on_fork
= p
->sched_reset_on_fork
;
4856 policy
= oldpolicy
= p
->policy
;
4858 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4860 if (!valid_policy(policy
))
4864 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4868 * Valid priorities for SCHED_FIFO and SCHED_RR are
4869 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4870 * SCHED_BATCH and SCHED_IDLE is 0.
4872 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4873 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4875 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4876 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4880 * Allow unprivileged RT tasks to decrease priority:
4882 if (user
&& !capable(CAP_SYS_NICE
)) {
4883 if (fair_policy(policy
)) {
4884 if (attr
->sched_nice
< task_nice(p
) &&
4885 !can_nice(p
, attr
->sched_nice
))
4889 if (rt_policy(policy
)) {
4890 unsigned long rlim_rtprio
=
4891 task_rlimit(p
, RLIMIT_RTPRIO
);
4893 /* Can't set/change the rt policy: */
4894 if (policy
!= p
->policy
&& !rlim_rtprio
)
4897 /* Can't increase priority: */
4898 if (attr
->sched_priority
> p
->rt_priority
&&
4899 attr
->sched_priority
> rlim_rtprio
)
4904 * Can't set/change SCHED_DEADLINE policy at all for now
4905 * (safest behavior); in the future we would like to allow
4906 * unprivileged DL tasks to increase their relative deadline
4907 * or reduce their runtime (both ways reducing utilization)
4909 if (dl_policy(policy
))
4913 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4914 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4916 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
4917 if (!can_nice(p
, task_nice(p
)))
4921 /* Can't change other user's priorities: */
4922 if (!check_same_owner(p
))
4925 /* Normal users shall not reset the sched_reset_on_fork flag: */
4926 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4931 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4934 retval
= security_task_setscheduler(p
);
4939 /* Update task specific "requested" clamps */
4940 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
4941 retval
= uclamp_validate(p
, attr
);
4950 * Make sure no PI-waiters arrive (or leave) while we are
4951 * changing the priority of the task:
4953 * To be able to change p->policy safely, the appropriate
4954 * runqueue lock must be held.
4956 rq
= task_rq_lock(p
, &rf
);
4957 update_rq_clock(rq
);
4960 * Changing the policy of the stop threads its a very bad idea:
4962 if (p
== rq
->stop
) {
4968 * If not changing anything there's no need to proceed further,
4969 * but store a possible modification of reset_on_fork.
4971 if (unlikely(policy
== p
->policy
)) {
4972 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4974 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4976 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4978 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
4981 p
->sched_reset_on_fork
= reset_on_fork
;
4988 #ifdef CONFIG_RT_GROUP_SCHED
4990 * Do not allow realtime tasks into groups that have no runtime
4993 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4994 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4995 !task_group_is_autogroup(task_group(p
))) {
5001 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
5002 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
5003 cpumask_t
*span
= rq
->rd
->span
;
5006 * Don't allow tasks with an affinity mask smaller than
5007 * the entire root_domain to become SCHED_DEADLINE. We
5008 * will also fail if there's no bandwidth available.
5010 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
5011 rq
->rd
->dl_bw
.bw
== 0) {
5019 /* Re-check policy now with rq lock held: */
5020 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5021 policy
= oldpolicy
= -1;
5022 task_rq_unlock(rq
, p
, &rf
);
5024 cpuset_read_unlock();
5029 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5030 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5033 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
5038 p
->sched_reset_on_fork
= reset_on_fork
;
5043 * Take priority boosted tasks into account. If the new
5044 * effective priority is unchanged, we just store the new
5045 * normal parameters and do not touch the scheduler class and
5046 * the runqueue. This will be done when the task deboost
5049 new_effective_prio
= rt_effective_prio(p
, newprio
);
5050 if (new_effective_prio
== oldprio
)
5051 queue_flags
&= ~DEQUEUE_MOVE
;
5054 queued
= task_on_rq_queued(p
);
5055 running
= task_current(rq
, p
);
5057 dequeue_task(rq
, p
, queue_flags
);
5059 put_prev_task(rq
, p
);
5061 prev_class
= p
->sched_class
;
5063 __setscheduler(rq
, p
, attr
, pi
);
5064 __setscheduler_uclamp(p
, attr
);
5068 * We enqueue to tail when the priority of a task is
5069 * increased (user space view).
5071 if (oldprio
< p
->prio
)
5072 queue_flags
|= ENQUEUE_HEAD
;
5074 enqueue_task(rq
, p
, queue_flags
);
5077 set_next_task(rq
, p
);
5079 check_class_changed(rq
, p
, prev_class
, oldprio
);
5081 /* Avoid rq from going away on us: */
5083 task_rq_unlock(rq
, p
, &rf
);
5086 cpuset_read_unlock();
5087 rt_mutex_adjust_pi(p
);
5090 /* Run balance callbacks after we've adjusted the PI chain: */
5091 balance_callback(rq
);
5097 task_rq_unlock(rq
, p
, &rf
);
5099 cpuset_read_unlock();
5103 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
5104 const struct sched_param
*param
, bool check
)
5106 struct sched_attr attr
= {
5107 .sched_policy
= policy
,
5108 .sched_priority
= param
->sched_priority
,
5109 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
5112 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5113 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
5114 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5115 policy
&= ~SCHED_RESET_ON_FORK
;
5116 attr
.sched_policy
= policy
;
5119 return __sched_setscheduler(p
, &attr
, check
, true);
5122 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5123 * @p: the task in question.
5124 * @policy: new policy.
5125 * @param: structure containing the new RT priority.
5127 * Return: 0 on success. An error code otherwise.
5129 * NOTE that the task may be already dead.
5131 int sched_setscheduler(struct task_struct
*p
, int policy
,
5132 const struct sched_param
*param
)
5134 return _sched_setscheduler(p
, policy
, param
, true);
5136 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5138 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
5140 return __sched_setscheduler(p
, attr
, true, true);
5142 EXPORT_SYMBOL_GPL(sched_setattr
);
5144 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
5146 return __sched_setscheduler(p
, attr
, false, true);
5150 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5151 * @p: the task in question.
5152 * @policy: new policy.
5153 * @param: structure containing the new RT priority.
5155 * Just like sched_setscheduler, only don't bother checking if the
5156 * current context has permission. For example, this is needed in
5157 * stop_machine(): we create temporary high priority worker threads,
5158 * but our caller might not have that capability.
5160 * Return: 0 on success. An error code otherwise.
5162 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5163 const struct sched_param
*param
)
5165 return _sched_setscheduler(p
, policy
, param
, false);
5167 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
5170 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5172 struct sched_param lparam
;
5173 struct task_struct
*p
;
5176 if (!param
|| pid
< 0)
5178 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5183 p
= find_process_by_pid(pid
);
5189 retval
= sched_setscheduler(p
, policy
, &lparam
);
5197 * Mimics kernel/events/core.c perf_copy_attr().
5199 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
5204 /* Zero the full structure, so that a short copy will be nice: */
5205 memset(attr
, 0, sizeof(*attr
));
5207 ret
= get_user(size
, &uattr
->size
);
5211 /* ABI compatibility quirk: */
5213 size
= SCHED_ATTR_SIZE_VER0
;
5214 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
5217 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
5224 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
5225 size
< SCHED_ATTR_SIZE_VER1
)
5229 * XXX: Do we want to be lenient like existing syscalls; or do we want
5230 * to be strict and return an error on out-of-bounds values?
5232 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
5237 put_user(sizeof(*attr
), &uattr
->size
);
5242 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5243 * @pid: the pid in question.
5244 * @policy: new policy.
5245 * @param: structure containing the new RT priority.
5247 * Return: 0 on success. An error code otherwise.
5249 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
5254 return do_sched_setscheduler(pid
, policy
, param
);
5258 * sys_sched_setparam - set/change the RT priority of a thread
5259 * @pid: the pid in question.
5260 * @param: structure containing the new RT priority.
5262 * Return: 0 on success. An error code otherwise.
5264 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5266 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
5270 * sys_sched_setattr - same as above, but with extended sched_attr
5271 * @pid: the pid in question.
5272 * @uattr: structure containing the extended parameters.
5273 * @flags: for future extension.
5275 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5276 unsigned int, flags
)
5278 struct sched_attr attr
;
5279 struct task_struct
*p
;
5282 if (!uattr
|| pid
< 0 || flags
)
5285 retval
= sched_copy_attr(uattr
, &attr
);
5289 if ((int)attr
.sched_policy
< 0)
5291 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
5292 attr
.sched_policy
= SETPARAM_POLICY
;
5296 p
= find_process_by_pid(pid
);
5302 retval
= sched_setattr(p
, &attr
);
5310 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5311 * @pid: the pid in question.
5313 * Return: On success, the policy of the thread. Otherwise, a negative error
5316 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5318 struct task_struct
*p
;
5326 p
= find_process_by_pid(pid
);
5328 retval
= security_task_getscheduler(p
);
5331 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5338 * sys_sched_getparam - get the RT priority of a thread
5339 * @pid: the pid in question.
5340 * @param: structure containing the RT priority.
5342 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5345 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5347 struct sched_param lp
= { .sched_priority
= 0 };
5348 struct task_struct
*p
;
5351 if (!param
|| pid
< 0)
5355 p
= find_process_by_pid(pid
);
5360 retval
= security_task_getscheduler(p
);
5364 if (task_has_rt_policy(p
))
5365 lp
.sched_priority
= p
->rt_priority
;
5369 * This one might sleep, we cannot do it with a spinlock held ...
5371 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5381 * Copy the kernel size attribute structure (which might be larger
5382 * than what user-space knows about) to user-space.
5384 * Note that all cases are valid: user-space buffer can be larger or
5385 * smaller than the kernel-space buffer. The usual case is that both
5386 * have the same size.
5389 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
5390 struct sched_attr
*kattr
,
5393 unsigned int ksize
= sizeof(*kattr
);
5395 if (!access_ok(uattr
, usize
))
5399 * sched_getattr() ABI forwards and backwards compatibility:
5401 * If usize == ksize then we just copy everything to user-space and all is good.
5403 * If usize < ksize then we only copy as much as user-space has space for,
5404 * this keeps ABI compatibility as well. We skip the rest.
5406 * If usize > ksize then user-space is using a newer version of the ABI,
5407 * which part the kernel doesn't know about. Just ignore it - tooling can
5408 * detect the kernel's knowledge of attributes from the attr->size value
5409 * which is set to ksize in this case.
5411 kattr
->size
= min(usize
, ksize
);
5413 if (copy_to_user(uattr
, kattr
, kattr
->size
))
5420 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5421 * @pid: the pid in question.
5422 * @uattr: structure containing the extended parameters.
5423 * @usize: sizeof(attr) for fwd/bwd comp.
5424 * @flags: for future extension.
5426 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5427 unsigned int, usize
, unsigned int, flags
)
5429 struct sched_attr kattr
= { };
5430 struct task_struct
*p
;
5433 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
5434 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
5438 p
= find_process_by_pid(pid
);
5443 retval
= security_task_getscheduler(p
);
5447 kattr
.sched_policy
= p
->policy
;
5448 if (p
->sched_reset_on_fork
)
5449 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5450 if (task_has_dl_policy(p
))
5451 __getparam_dl(p
, &kattr
);
5452 else if (task_has_rt_policy(p
))
5453 kattr
.sched_priority
= p
->rt_priority
;
5455 kattr
.sched_nice
= task_nice(p
);
5457 #ifdef CONFIG_UCLAMP_TASK
5458 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
5459 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
5464 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
5471 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5473 cpumask_var_t cpus_allowed
, new_mask
;
5474 struct task_struct
*p
;
5479 p
= find_process_by_pid(pid
);
5485 /* Prevent p going away */
5489 if (p
->flags
& PF_NO_SETAFFINITY
) {
5493 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5497 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5499 goto out_free_cpus_allowed
;
5502 if (!check_same_owner(p
)) {
5504 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
5506 goto out_free_new_mask
;
5511 retval
= security_task_setscheduler(p
);
5513 goto out_free_new_mask
;
5516 cpuset_cpus_allowed(p
, cpus_allowed
);
5517 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5520 * Since bandwidth control happens on root_domain basis,
5521 * if admission test is enabled, we only admit -deadline
5522 * tasks allowed to run on all the CPUs in the task's
5526 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
5528 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
5531 goto out_free_new_mask
;
5537 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
5540 cpuset_cpus_allowed(p
, cpus_allowed
);
5541 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5543 * We must have raced with a concurrent cpuset
5544 * update. Just reset the cpus_allowed to the
5545 * cpuset's cpus_allowed
5547 cpumask_copy(new_mask
, cpus_allowed
);
5552 free_cpumask_var(new_mask
);
5553 out_free_cpus_allowed
:
5554 free_cpumask_var(cpus_allowed
);
5560 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5561 struct cpumask
*new_mask
)
5563 if (len
< cpumask_size())
5564 cpumask_clear(new_mask
);
5565 else if (len
> cpumask_size())
5566 len
= cpumask_size();
5568 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5572 * sys_sched_setaffinity - set the CPU affinity of a process
5573 * @pid: pid of the process
5574 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5575 * @user_mask_ptr: user-space pointer to the new CPU mask
5577 * Return: 0 on success. An error code otherwise.
5579 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5580 unsigned long __user
*, user_mask_ptr
)
5582 cpumask_var_t new_mask
;
5585 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5588 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5590 retval
= sched_setaffinity(pid
, new_mask
);
5591 free_cpumask_var(new_mask
);
5595 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5597 struct task_struct
*p
;
5598 unsigned long flags
;
5604 p
= find_process_by_pid(pid
);
5608 retval
= security_task_getscheduler(p
);
5612 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5613 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
5614 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5623 * sys_sched_getaffinity - get the CPU affinity of a process
5624 * @pid: pid of the process
5625 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5626 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5628 * Return: size of CPU mask copied to user_mask_ptr on success. An
5629 * error code otherwise.
5631 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5632 unsigned long __user
*, user_mask_ptr
)
5637 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5639 if (len
& (sizeof(unsigned long)-1))
5642 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5645 ret
= sched_getaffinity(pid
, mask
);
5647 unsigned int retlen
= min(len
, cpumask_size());
5649 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5654 free_cpumask_var(mask
);
5660 * sys_sched_yield - yield the current processor to other threads.
5662 * This function yields the current CPU to other tasks. If there are no
5663 * other threads running on this CPU then this function will return.
5667 static void do_sched_yield(void)
5672 rq
= this_rq_lock_irq(&rf
);
5674 schedstat_inc(rq
->yld_count
);
5675 current
->sched_class
->yield_task(rq
);
5678 * Since we are going to call schedule() anyway, there's
5679 * no need to preempt or enable interrupts:
5683 sched_preempt_enable_no_resched();
5688 SYSCALL_DEFINE0(sched_yield
)
5694 #ifndef CONFIG_PREEMPTION
5695 int __sched
_cond_resched(void)
5697 if (should_resched(0)) {
5698 preempt_schedule_common();
5704 EXPORT_SYMBOL(_cond_resched
);
5708 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5709 * call schedule, and on return reacquire the lock.
5711 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5712 * operations here to prevent schedule() from being called twice (once via
5713 * spin_unlock(), once by hand).
5715 int __cond_resched_lock(spinlock_t
*lock
)
5717 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5720 lockdep_assert_held(lock
);
5722 if (spin_needbreak(lock
) || resched
) {
5725 preempt_schedule_common();
5733 EXPORT_SYMBOL(__cond_resched_lock
);
5736 * yield - yield the current processor to other threads.
5738 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5740 * The scheduler is at all times free to pick the calling task as the most
5741 * eligible task to run, if removing the yield() call from your code breaks
5742 * it, its already broken.
5744 * Typical broken usage is:
5749 * where one assumes that yield() will let 'the other' process run that will
5750 * make event true. If the current task is a SCHED_FIFO task that will never
5751 * happen. Never use yield() as a progress guarantee!!
5753 * If you want to use yield() to wait for something, use wait_event().
5754 * If you want to use yield() to be 'nice' for others, use cond_resched().
5755 * If you still want to use yield(), do not!
5757 void __sched
yield(void)
5759 set_current_state(TASK_RUNNING
);
5762 EXPORT_SYMBOL(yield
);
5765 * yield_to - yield the current processor to another thread in
5766 * your thread group, or accelerate that thread toward the
5767 * processor it's on.
5769 * @preempt: whether task preemption is allowed or not
5771 * It's the caller's job to ensure that the target task struct
5772 * can't go away on us before we can do any checks.
5775 * true (>0) if we indeed boosted the target task.
5776 * false (0) if we failed to boost the target.
5777 * -ESRCH if there's no task to yield to.
5779 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5781 struct task_struct
*curr
= current
;
5782 struct rq
*rq
, *p_rq
;
5783 unsigned long flags
;
5786 local_irq_save(flags
);
5792 * If we're the only runnable task on the rq and target rq also
5793 * has only one task, there's absolutely no point in yielding.
5795 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5800 double_rq_lock(rq
, p_rq
);
5801 if (task_rq(p
) != p_rq
) {
5802 double_rq_unlock(rq
, p_rq
);
5806 if (!curr
->sched_class
->yield_to_task
)
5809 if (curr
->sched_class
!= p
->sched_class
)
5812 if (task_running(p_rq
, p
) || p
->state
)
5815 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5817 schedstat_inc(rq
->yld_count
);
5819 * Make p's CPU reschedule; pick_next_entity takes care of
5822 if (preempt
&& rq
!= p_rq
)
5827 double_rq_unlock(rq
, p_rq
);
5829 local_irq_restore(flags
);
5836 EXPORT_SYMBOL_GPL(yield_to
);
5838 int io_schedule_prepare(void)
5840 int old_iowait
= current
->in_iowait
;
5842 current
->in_iowait
= 1;
5843 blk_schedule_flush_plug(current
);
5848 void io_schedule_finish(int token
)
5850 current
->in_iowait
= token
;
5854 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5855 * that process accounting knows that this is a task in IO wait state.
5857 long __sched
io_schedule_timeout(long timeout
)
5862 token
= io_schedule_prepare();
5863 ret
= schedule_timeout(timeout
);
5864 io_schedule_finish(token
);
5868 EXPORT_SYMBOL(io_schedule_timeout
);
5870 void __sched
io_schedule(void)
5874 token
= io_schedule_prepare();
5876 io_schedule_finish(token
);
5878 EXPORT_SYMBOL(io_schedule
);
5881 * sys_sched_get_priority_max - return maximum RT priority.
5882 * @policy: scheduling class.
5884 * Return: On success, this syscall returns the maximum
5885 * rt_priority that can be used by a given scheduling class.
5886 * On failure, a negative error code is returned.
5888 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5895 ret
= MAX_USER_RT_PRIO
-1;
5897 case SCHED_DEADLINE
:
5908 * sys_sched_get_priority_min - return minimum RT priority.
5909 * @policy: scheduling class.
5911 * Return: On success, this syscall returns the minimum
5912 * rt_priority that can be used by a given scheduling class.
5913 * On failure, a negative error code is returned.
5915 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5924 case SCHED_DEADLINE
:
5933 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5935 struct task_struct
*p
;
5936 unsigned int time_slice
;
5946 p
= find_process_by_pid(pid
);
5950 retval
= security_task_getscheduler(p
);
5954 rq
= task_rq_lock(p
, &rf
);
5956 if (p
->sched_class
->get_rr_interval
)
5957 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5958 task_rq_unlock(rq
, p
, &rf
);
5961 jiffies_to_timespec64(time_slice
, t
);
5970 * sys_sched_rr_get_interval - return the default timeslice of a process.
5971 * @pid: pid of the process.
5972 * @interval: userspace pointer to the timeslice value.
5974 * this syscall writes the default timeslice value of a given process
5975 * into the user-space timespec buffer. A value of '0' means infinity.
5977 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5980 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5981 struct __kernel_timespec __user
*, interval
)
5983 struct timespec64 t
;
5984 int retval
= sched_rr_get_interval(pid
, &t
);
5987 retval
= put_timespec64(&t
, interval
);
5992 #ifdef CONFIG_COMPAT_32BIT_TIME
5993 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
5994 struct old_timespec32 __user
*, interval
)
5996 struct timespec64 t
;
5997 int retval
= sched_rr_get_interval(pid
, &t
);
6000 retval
= put_old_timespec32(&t
, interval
);
6005 void sched_show_task(struct task_struct
*p
)
6007 unsigned long free
= 0;
6010 if (!try_get_task_stack(p
))
6013 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
6015 if (p
->state
== TASK_RUNNING
)
6016 printk(KERN_CONT
" running task ");
6017 #ifdef CONFIG_DEBUG_STACK_USAGE
6018 free
= stack_not_used(p
);
6023 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
6025 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6026 task_pid_nr(p
), ppid
,
6027 (unsigned long)task_thread_info(p
)->flags
);
6029 print_worker_info(KERN_INFO
, p
);
6030 show_stack(p
, NULL
);
6033 EXPORT_SYMBOL_GPL(sched_show_task
);
6036 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
6038 /* no filter, everything matches */
6042 /* filter, but doesn't match */
6043 if (!(p
->state
& state_filter
))
6047 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6050 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
6057 void show_state_filter(unsigned long state_filter
)
6059 struct task_struct
*g
, *p
;
6061 #if BITS_PER_LONG == 32
6063 " task PC stack pid father\n");
6066 " task PC stack pid father\n");
6069 for_each_process_thread(g
, p
) {
6071 * reset the NMI-timeout, listing all files on a slow
6072 * console might take a lot of time:
6073 * Also, reset softlockup watchdogs on all CPUs, because
6074 * another CPU might be blocked waiting for us to process
6077 touch_nmi_watchdog();
6078 touch_all_softlockup_watchdogs();
6079 if (state_filter_match(state_filter
, p
))
6083 #ifdef CONFIG_SCHED_DEBUG
6085 sysrq_sched_debug_show();
6089 * Only show locks if all tasks are dumped:
6092 debug_show_all_locks();
6096 * init_idle - set up an idle thread for a given CPU
6097 * @idle: task in question
6098 * @cpu: CPU the idle task belongs to
6100 * NOTE: this function does not set the idle thread's NEED_RESCHED
6101 * flag, to make booting more robust.
6103 void init_idle(struct task_struct
*idle
, int cpu
)
6105 struct rq
*rq
= cpu_rq(cpu
);
6106 unsigned long flags
;
6108 __sched_fork(0, idle
);
6110 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
6111 raw_spin_lock(&rq
->lock
);
6113 idle
->state
= TASK_RUNNING
;
6114 idle
->se
.exec_start
= sched_clock();
6115 idle
->flags
|= PF_IDLE
;
6117 scs_task_reset(idle
);
6118 kasan_unpoison_task_stack(idle
);
6122 * Its possible that init_idle() gets called multiple times on a task,
6123 * in that case do_set_cpus_allowed() will not do the right thing.
6125 * And since this is boot we can forgo the serialization.
6127 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
6130 * We're having a chicken and egg problem, even though we are
6131 * holding rq->lock, the CPU isn't yet set to this CPU so the
6132 * lockdep check in task_group() will fail.
6134 * Similar case to sched_fork(). / Alternatively we could
6135 * use task_rq_lock() here and obtain the other rq->lock.
6140 __set_task_cpu(idle
, cpu
);
6144 rcu_assign_pointer(rq
->curr
, idle
);
6145 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
6149 raw_spin_unlock(&rq
->lock
);
6150 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
6152 /* Set the preempt count _outside_ the spinlocks! */
6153 init_idle_preempt_count(idle
, cpu
);
6156 * The idle tasks have their own, simple scheduling class:
6158 idle
->sched_class
= &idle_sched_class
;
6159 ftrace_graph_init_idle_task(idle
, cpu
);
6160 vtime_init_idle(idle
, cpu
);
6162 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
6168 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
6169 const struct cpumask
*trial
)
6173 if (!cpumask_weight(cur
))
6176 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
6181 int task_can_attach(struct task_struct
*p
,
6182 const struct cpumask
*cs_cpus_allowed
)
6187 * Kthreads which disallow setaffinity shouldn't be moved
6188 * to a new cpuset; we don't want to change their CPU
6189 * affinity and isolating such threads by their set of
6190 * allowed nodes is unnecessary. Thus, cpusets are not
6191 * applicable for such threads. This prevents checking for
6192 * success of set_cpus_allowed_ptr() on all attached tasks
6193 * before cpus_mask may be changed.
6195 if (p
->flags
& PF_NO_SETAFFINITY
) {
6200 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
6202 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
6208 bool sched_smp_initialized __read_mostly
;
6210 #ifdef CONFIG_NUMA_BALANCING
6211 /* Migrate current task p to target_cpu */
6212 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
6214 struct migration_arg arg
= { p
, target_cpu
};
6215 int curr_cpu
= task_cpu(p
);
6217 if (curr_cpu
== target_cpu
)
6220 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
6223 /* TODO: This is not properly updating schedstats */
6225 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
6226 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
6230 * Requeue a task on a given node and accurately track the number of NUMA
6231 * tasks on the runqueues
6233 void sched_setnuma(struct task_struct
*p
, int nid
)
6235 bool queued
, running
;
6239 rq
= task_rq_lock(p
, &rf
);
6240 queued
= task_on_rq_queued(p
);
6241 running
= task_current(rq
, p
);
6244 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
6246 put_prev_task(rq
, p
);
6248 p
->numa_preferred_nid
= nid
;
6251 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
6253 set_next_task(rq
, p
);
6254 task_rq_unlock(rq
, p
, &rf
);
6256 #endif /* CONFIG_NUMA_BALANCING */
6258 #ifdef CONFIG_HOTPLUG_CPU
6260 * Ensure that the idle task is using init_mm right before its CPU goes
6263 void idle_task_exit(void)
6265 struct mm_struct
*mm
= current
->active_mm
;
6267 BUG_ON(cpu_online(smp_processor_id()));
6268 BUG_ON(current
!= this_rq()->idle
);
6270 if (mm
!= &init_mm
) {
6271 switch_mm(mm
, &init_mm
, current
);
6272 finish_arch_post_lock_switch();
6275 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6279 * Since this CPU is going 'away' for a while, fold any nr_active delta
6280 * we might have. Assumes we're called after migrate_tasks() so that the
6281 * nr_active count is stable. We need to take the teardown thread which
6282 * is calling this into account, so we hand in adjust = 1 to the load
6285 * Also see the comment "Global load-average calculations".
6287 static void calc_load_migrate(struct rq
*rq
)
6289 long delta
= calc_load_fold_active(rq
, 1);
6291 atomic_long_add(delta
, &calc_load_tasks
);
6294 static struct task_struct
*__pick_migrate_task(struct rq
*rq
)
6296 const struct sched_class
*class;
6297 struct task_struct
*next
;
6299 for_each_class(class) {
6300 next
= class->pick_next_task(rq
);
6302 next
->sched_class
->put_prev_task(rq
, next
);
6307 /* The idle class should always have a runnable task */
6312 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6313 * try_to_wake_up()->select_task_rq().
6315 * Called with rq->lock held even though we'er in stop_machine() and
6316 * there's no concurrency possible, we hold the required locks anyway
6317 * because of lock validation efforts.
6319 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
6321 struct rq
*rq
= dead_rq
;
6322 struct task_struct
*next
, *stop
= rq
->stop
;
6323 struct rq_flags orf
= *rf
;
6327 * Fudge the rq selection such that the below task selection loop
6328 * doesn't get stuck on the currently eligible stop task.
6330 * We're currently inside stop_machine() and the rq is either stuck
6331 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6332 * either way we should never end up calling schedule() until we're
6338 * put_prev_task() and pick_next_task() sched
6339 * class method both need to have an up-to-date
6340 * value of rq->clock[_task]
6342 update_rq_clock(rq
);
6346 * There's this thread running, bail when that's the only
6349 if (rq
->nr_running
== 1)
6352 next
= __pick_migrate_task(rq
);
6355 * Rules for changing task_struct::cpus_mask are holding
6356 * both pi_lock and rq->lock, such that holding either
6357 * stabilizes the mask.
6359 * Drop rq->lock is not quite as disastrous as it usually is
6360 * because !cpu_active at this point, which means load-balance
6361 * will not interfere. Also, stop-machine.
6364 raw_spin_lock(&next
->pi_lock
);
6368 * Since we're inside stop-machine, _nothing_ should have
6369 * changed the task, WARN if weird stuff happened, because in
6370 * that case the above rq->lock drop is a fail too.
6372 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
6373 raw_spin_unlock(&next
->pi_lock
);
6377 /* Find suitable destination for @next, with force if needed. */
6378 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
6379 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
6380 if (rq
!= dead_rq
) {
6386 raw_spin_unlock(&next
->pi_lock
);
6391 #endif /* CONFIG_HOTPLUG_CPU */
6393 void set_rq_online(struct rq
*rq
)
6396 const struct sched_class
*class;
6398 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6401 for_each_class(class) {
6402 if (class->rq_online
)
6403 class->rq_online(rq
);
6408 void set_rq_offline(struct rq
*rq
)
6411 const struct sched_class
*class;
6413 for_each_class(class) {
6414 if (class->rq_offline
)
6415 class->rq_offline(rq
);
6418 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6424 * used to mark begin/end of suspend/resume:
6426 static int num_cpus_frozen
;
6429 * Update cpusets according to cpu_active mask. If cpusets are
6430 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6431 * around partition_sched_domains().
6433 * If we come here as part of a suspend/resume, don't touch cpusets because we
6434 * want to restore it back to its original state upon resume anyway.
6436 static void cpuset_cpu_active(void)
6438 if (cpuhp_tasks_frozen
) {
6440 * num_cpus_frozen tracks how many CPUs are involved in suspend
6441 * resume sequence. As long as this is not the last online
6442 * operation in the resume sequence, just build a single sched
6443 * domain, ignoring cpusets.
6445 partition_sched_domains(1, NULL
, NULL
);
6446 if (--num_cpus_frozen
)
6449 * This is the last CPU online operation. So fall through and
6450 * restore the original sched domains by considering the
6451 * cpuset configurations.
6453 cpuset_force_rebuild();
6455 cpuset_update_active_cpus();
6458 static int cpuset_cpu_inactive(unsigned int cpu
)
6460 if (!cpuhp_tasks_frozen
) {
6461 if (dl_cpu_busy(cpu
))
6463 cpuset_update_active_cpus();
6466 partition_sched_domains(1, NULL
, NULL
);
6471 int sched_cpu_activate(unsigned int cpu
)
6473 struct rq
*rq
= cpu_rq(cpu
);
6476 #ifdef CONFIG_SCHED_SMT
6478 * When going up, increment the number of cores with SMT present.
6480 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6481 static_branch_inc_cpuslocked(&sched_smt_present
);
6483 set_cpu_active(cpu
, true);
6485 if (sched_smp_initialized
) {
6486 sched_domains_numa_masks_set(cpu
);
6487 cpuset_cpu_active();
6491 * Put the rq online, if not already. This happens:
6493 * 1) In the early boot process, because we build the real domains
6494 * after all CPUs have been brought up.
6496 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6499 rq_lock_irqsave(rq
, &rf
);
6501 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6504 rq_unlock_irqrestore(rq
, &rf
);
6509 int sched_cpu_deactivate(unsigned int cpu
)
6513 set_cpu_active(cpu
, false);
6515 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6516 * users of this state to go away such that all new such users will
6519 * Do sync before park smpboot threads to take care the rcu boost case.
6523 #ifdef CONFIG_SCHED_SMT
6525 * When going down, decrement the number of cores with SMT present.
6527 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6528 static_branch_dec_cpuslocked(&sched_smt_present
);
6531 if (!sched_smp_initialized
)
6534 ret
= cpuset_cpu_inactive(cpu
);
6536 set_cpu_active(cpu
, true);
6539 sched_domains_numa_masks_clear(cpu
);
6543 static void sched_rq_cpu_starting(unsigned int cpu
)
6545 struct rq
*rq
= cpu_rq(cpu
);
6547 rq
->calc_load_update
= calc_load_update
;
6548 update_max_interval();
6551 int sched_cpu_starting(unsigned int cpu
)
6553 sched_rq_cpu_starting(cpu
);
6554 sched_tick_start(cpu
);
6558 #ifdef CONFIG_HOTPLUG_CPU
6559 int sched_cpu_dying(unsigned int cpu
)
6561 struct rq
*rq
= cpu_rq(cpu
);
6564 /* Handle pending wakeups and then migrate everything off */
6565 sched_tick_stop(cpu
);
6567 rq_lock_irqsave(rq
, &rf
);
6569 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6572 migrate_tasks(rq
, &rf
);
6573 BUG_ON(rq
->nr_running
!= 1);
6574 rq_unlock_irqrestore(rq
, &rf
);
6576 calc_load_migrate(rq
);
6577 update_max_interval();
6578 nohz_balance_exit_idle(rq
);
6584 void __init
sched_init_smp(void)
6589 * There's no userspace yet to cause hotplug operations; hence all the
6590 * CPU masks are stable and all blatant races in the below code cannot
6593 mutex_lock(&sched_domains_mutex
);
6594 sched_init_domains(cpu_active_mask
);
6595 mutex_unlock(&sched_domains_mutex
);
6597 /* Move init over to a non-isolated CPU */
6598 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
6600 sched_init_granularity();
6602 init_sched_rt_class();
6603 init_sched_dl_class();
6605 sched_smp_initialized
= true;
6608 static int __init
migration_init(void)
6610 sched_cpu_starting(smp_processor_id());
6613 early_initcall(migration_init
);
6616 void __init
sched_init_smp(void)
6618 sched_init_granularity();
6620 #endif /* CONFIG_SMP */
6622 int in_sched_functions(unsigned long addr
)
6624 return in_lock_functions(addr
) ||
6625 (addr
>= (unsigned long)__sched_text_start
6626 && addr
< (unsigned long)__sched_text_end
);
6629 #ifdef CONFIG_CGROUP_SCHED
6631 * Default task group.
6632 * Every task in system belongs to this group at bootup.
6634 struct task_group root_task_group
;
6635 LIST_HEAD(task_groups
);
6637 /* Cacheline aligned slab cache for task_group */
6638 static struct kmem_cache
*task_group_cache __read_mostly
;
6641 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6642 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
6644 void __init
sched_init(void)
6646 unsigned long ptr
= 0;
6651 #ifdef CONFIG_FAIR_GROUP_SCHED
6652 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6654 #ifdef CONFIG_RT_GROUP_SCHED
6655 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6658 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
6660 #ifdef CONFIG_FAIR_GROUP_SCHED
6661 root_task_group
.se
= (struct sched_entity
**)ptr
;
6662 ptr
+= nr_cpu_ids
* sizeof(void **);
6664 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6665 ptr
+= nr_cpu_ids
* sizeof(void **);
6667 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6668 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6669 #endif /* CONFIG_FAIR_GROUP_SCHED */
6670 #ifdef CONFIG_RT_GROUP_SCHED
6671 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6672 ptr
+= nr_cpu_ids
* sizeof(void **);
6674 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6675 ptr
+= nr_cpu_ids
* sizeof(void **);
6677 #endif /* CONFIG_RT_GROUP_SCHED */
6679 #ifdef CONFIG_CPUMASK_OFFSTACK
6680 for_each_possible_cpu(i
) {
6681 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6682 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6683 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6684 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6686 #endif /* CONFIG_CPUMASK_OFFSTACK */
6688 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
6689 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6692 init_defrootdomain();
6695 #ifdef CONFIG_RT_GROUP_SCHED
6696 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6697 global_rt_period(), global_rt_runtime());
6698 #endif /* CONFIG_RT_GROUP_SCHED */
6700 #ifdef CONFIG_CGROUP_SCHED
6701 task_group_cache
= KMEM_CACHE(task_group
, 0);
6703 list_add(&root_task_group
.list
, &task_groups
);
6704 INIT_LIST_HEAD(&root_task_group
.children
);
6705 INIT_LIST_HEAD(&root_task_group
.siblings
);
6706 autogroup_init(&init_task
);
6707 #endif /* CONFIG_CGROUP_SCHED */
6709 for_each_possible_cpu(i
) {
6713 raw_spin_lock_init(&rq
->lock
);
6715 rq
->calc_load_active
= 0;
6716 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6717 init_cfs_rq(&rq
->cfs
);
6718 init_rt_rq(&rq
->rt
);
6719 init_dl_rq(&rq
->dl
);
6720 #ifdef CONFIG_FAIR_GROUP_SCHED
6721 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6722 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6724 * How much CPU bandwidth does root_task_group get?
6726 * In case of task-groups formed thr' the cgroup filesystem, it
6727 * gets 100% of the CPU resources in the system. This overall
6728 * system CPU resource is divided among the tasks of
6729 * root_task_group and its child task-groups in a fair manner,
6730 * based on each entity's (task or task-group's) weight
6731 * (se->load.weight).
6733 * In other words, if root_task_group has 10 tasks of weight
6734 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6735 * then A0's share of the CPU resource is:
6737 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6739 * We achieve this by letting root_task_group's tasks sit
6740 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6742 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6743 #endif /* CONFIG_FAIR_GROUP_SCHED */
6745 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6746 #ifdef CONFIG_RT_GROUP_SCHED
6747 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6752 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6753 rq
->balance_callback
= NULL
;
6754 rq
->active_balance
= 0;
6755 rq
->next_balance
= jiffies
;
6760 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6761 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6763 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6765 rq_attach_root(rq
, &def_root_domain
);
6766 #ifdef CONFIG_NO_HZ_COMMON
6767 rq
->last_blocked_load_update_tick
= jiffies
;
6768 atomic_set(&rq
->nohz_flags
, 0);
6770 rq_csd_init(rq
, &rq
->nohz_csd
, nohz_csd_func
);
6772 #endif /* CONFIG_SMP */
6774 atomic_set(&rq
->nr_iowait
, 0);
6777 set_load_weight(&init_task
, false);
6780 * The boot idle thread does lazy MMU switching as well:
6783 enter_lazy_tlb(&init_mm
, current
);
6786 * Make us the idle thread. Technically, schedule() should not be
6787 * called from this thread, however somewhere below it might be,
6788 * but because we are the idle thread, we just pick up running again
6789 * when this runqueue becomes "idle".
6791 init_idle(current
, smp_processor_id());
6793 calc_load_update
= jiffies
+ LOAD_FREQ
;
6796 idle_thread_set_boot_cpu();
6798 init_sched_fair_class();
6806 scheduler_running
= 1;
6809 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6810 static inline int preempt_count_equals(int preempt_offset
)
6812 int nested
= preempt_count() + rcu_preempt_depth();
6814 return (nested
== preempt_offset
);
6817 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6820 * Blocking primitives will set (and therefore destroy) current->state,
6821 * since we will exit with TASK_RUNNING make sure we enter with it,
6822 * otherwise we will destroy state.
6824 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6825 "do not call blocking ops when !TASK_RUNNING; "
6826 "state=%lx set at [<%p>] %pS\n",
6828 (void *)current
->task_state_change
,
6829 (void *)current
->task_state_change
);
6831 ___might_sleep(file
, line
, preempt_offset
);
6833 EXPORT_SYMBOL(__might_sleep
);
6835 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6837 /* Ratelimiting timestamp: */
6838 static unsigned long prev_jiffy
;
6840 unsigned long preempt_disable_ip
;
6842 /* WARN_ON_ONCE() by default, no rate limit required: */
6845 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6846 !is_idle_task(current
) && !current
->non_block_count
) ||
6847 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6851 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6853 prev_jiffy
= jiffies
;
6855 /* Save this before calling printk(), since that will clobber it: */
6856 preempt_disable_ip
= get_preempt_disable_ip(current
);
6859 "BUG: sleeping function called from invalid context at %s:%d\n",
6862 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6863 in_atomic(), irqs_disabled(), current
->non_block_count
,
6864 current
->pid
, current
->comm
);
6866 if (task_stack_end_corrupted(current
))
6867 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6869 debug_show_held_locks(current
);
6870 if (irqs_disabled())
6871 print_irqtrace_events(current
);
6872 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6873 && !preempt_count_equals(preempt_offset
)) {
6874 pr_err("Preemption disabled at:");
6875 print_ip_sym(preempt_disable_ip
);
6879 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6881 EXPORT_SYMBOL(___might_sleep
);
6883 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
6885 static unsigned long prev_jiffy
;
6887 if (irqs_disabled())
6890 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
6893 if (preempt_count() > preempt_offset
)
6896 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6898 prev_jiffy
= jiffies
;
6900 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
6901 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6902 in_atomic(), irqs_disabled(),
6903 current
->pid
, current
->comm
);
6905 debug_show_held_locks(current
);
6907 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6909 EXPORT_SYMBOL_GPL(__cant_sleep
);
6912 #ifdef CONFIG_MAGIC_SYSRQ
6913 void normalize_rt_tasks(void)
6915 struct task_struct
*g
, *p
;
6916 struct sched_attr attr
= {
6917 .sched_policy
= SCHED_NORMAL
,
6920 read_lock(&tasklist_lock
);
6921 for_each_process_thread(g
, p
) {
6923 * Only normalize user tasks:
6925 if (p
->flags
& PF_KTHREAD
)
6928 p
->se
.exec_start
= 0;
6929 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6930 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6931 schedstat_set(p
->se
.statistics
.block_start
, 0);
6933 if (!dl_task(p
) && !rt_task(p
)) {
6935 * Renice negative nice level userspace
6938 if (task_nice(p
) < 0)
6939 set_user_nice(p
, 0);
6943 __sched_setscheduler(p
, &attr
, false, false);
6945 read_unlock(&tasklist_lock
);
6948 #endif /* CONFIG_MAGIC_SYSRQ */
6950 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6952 * These functions are only useful for the IA64 MCA handling, or kdb.
6954 * They can only be called when the whole system has been
6955 * stopped - every CPU needs to be quiescent, and no scheduling
6956 * activity can take place. Using them for anything else would
6957 * be a serious bug, and as a result, they aren't even visible
6958 * under any other configuration.
6962 * curr_task - return the current task for a given CPU.
6963 * @cpu: the processor in question.
6965 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6967 * Return: The current task for @cpu.
6969 struct task_struct
*curr_task(int cpu
)
6971 return cpu_curr(cpu
);
6974 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6978 * ia64_set_curr_task - set the current task for a given CPU.
6979 * @cpu: the processor in question.
6980 * @p: the task pointer to set.
6982 * Description: This function must only be used when non-maskable interrupts
6983 * are serviced on a separate stack. It allows the architecture to switch the
6984 * notion of the current task on a CPU in a non-blocking manner. This function
6985 * must be called with all CPU's synchronized, and interrupts disabled, the
6986 * and caller must save the original value of the current task (see
6987 * curr_task() above) and restore that value before reenabling interrupts and
6988 * re-starting the system.
6990 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6992 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6999 #ifdef CONFIG_CGROUP_SCHED
7000 /* task_group_lock serializes the addition/removal of task groups */
7001 static DEFINE_SPINLOCK(task_group_lock
);
7003 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
7004 struct task_group
*parent
)
7006 #ifdef CONFIG_UCLAMP_TASK_GROUP
7007 enum uclamp_id clamp_id
;
7009 for_each_clamp_id(clamp_id
) {
7010 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
7011 uclamp_none(clamp_id
), false);
7012 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
7017 static void sched_free_group(struct task_group
*tg
)
7019 free_fair_sched_group(tg
);
7020 free_rt_sched_group(tg
);
7022 kmem_cache_free(task_group_cache
, tg
);
7025 /* allocate runqueue etc for a new task group */
7026 struct task_group
*sched_create_group(struct task_group
*parent
)
7028 struct task_group
*tg
;
7030 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7032 return ERR_PTR(-ENOMEM
);
7034 if (!alloc_fair_sched_group(tg
, parent
))
7037 if (!alloc_rt_sched_group(tg
, parent
))
7040 alloc_uclamp_sched_group(tg
, parent
);
7045 sched_free_group(tg
);
7046 return ERR_PTR(-ENOMEM
);
7049 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7051 unsigned long flags
;
7053 spin_lock_irqsave(&task_group_lock
, flags
);
7054 list_add_rcu(&tg
->list
, &task_groups
);
7056 /* Root should already exist: */
7059 tg
->parent
= parent
;
7060 INIT_LIST_HEAD(&tg
->children
);
7061 list_add_rcu(&tg
->siblings
, &parent
->children
);
7062 spin_unlock_irqrestore(&task_group_lock
, flags
);
7064 online_fair_sched_group(tg
);
7067 /* rcu callback to free various structures associated with a task group */
7068 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7070 /* Now it should be safe to free those cfs_rqs: */
7071 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7074 void sched_destroy_group(struct task_group
*tg
)
7076 /* Wait for possible concurrent references to cfs_rqs complete: */
7077 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7080 void sched_offline_group(struct task_group
*tg
)
7082 unsigned long flags
;
7084 /* End participation in shares distribution: */
7085 unregister_fair_sched_group(tg
);
7087 spin_lock_irqsave(&task_group_lock
, flags
);
7088 list_del_rcu(&tg
->list
);
7089 list_del_rcu(&tg
->siblings
);
7090 spin_unlock_irqrestore(&task_group_lock
, flags
);
7093 static void sched_change_group(struct task_struct
*tsk
, int type
)
7095 struct task_group
*tg
;
7098 * All callers are synchronized by task_rq_lock(); we do not use RCU
7099 * which is pointless here. Thus, we pass "true" to task_css_check()
7100 * to prevent lockdep warnings.
7102 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7103 struct task_group
, css
);
7104 tg
= autogroup_task_group(tsk
, tg
);
7105 tsk
->sched_task_group
= tg
;
7107 #ifdef CONFIG_FAIR_GROUP_SCHED
7108 if (tsk
->sched_class
->task_change_group
)
7109 tsk
->sched_class
->task_change_group(tsk
, type
);
7112 set_task_rq(tsk
, task_cpu(tsk
));
7116 * Change task's runqueue when it moves between groups.
7118 * The caller of this function should have put the task in its new group by
7119 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7122 void sched_move_task(struct task_struct
*tsk
)
7124 int queued
, running
, queue_flags
=
7125 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7129 rq
= task_rq_lock(tsk
, &rf
);
7130 update_rq_clock(rq
);
7132 running
= task_current(rq
, tsk
);
7133 queued
= task_on_rq_queued(tsk
);
7136 dequeue_task(rq
, tsk
, queue_flags
);
7138 put_prev_task(rq
, tsk
);
7140 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7143 enqueue_task(rq
, tsk
, queue_flags
);
7145 set_next_task(rq
, tsk
);
7147 * After changing group, the running task may have joined a
7148 * throttled one but it's still the running task. Trigger a
7149 * resched to make sure that task can still run.
7154 task_rq_unlock(rq
, tsk
, &rf
);
7157 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7159 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7162 static struct cgroup_subsys_state
*
7163 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7165 struct task_group
*parent
= css_tg(parent_css
);
7166 struct task_group
*tg
;
7169 /* This is early initialization for the top cgroup */
7170 return &root_task_group
.css
;
7173 tg
= sched_create_group(parent
);
7175 return ERR_PTR(-ENOMEM
);
7180 /* Expose task group only after completing cgroup initialization */
7181 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7183 struct task_group
*tg
= css_tg(css
);
7184 struct task_group
*parent
= css_tg(css
->parent
);
7187 sched_online_group(tg
, parent
);
7189 #ifdef CONFIG_UCLAMP_TASK_GROUP
7190 /* Propagate the effective uclamp value for the new group */
7191 cpu_util_update_eff(css
);
7197 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
7199 struct task_group
*tg
= css_tg(css
);
7201 sched_offline_group(tg
);
7204 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7206 struct task_group
*tg
= css_tg(css
);
7209 * Relies on the RCU grace period between css_released() and this.
7211 sched_free_group(tg
);
7215 * This is called before wake_up_new_task(), therefore we really only
7216 * have to set its group bits, all the other stuff does not apply.
7218 static void cpu_cgroup_fork(struct task_struct
*task
)
7223 rq
= task_rq_lock(task
, &rf
);
7225 update_rq_clock(rq
);
7226 sched_change_group(task
, TASK_SET_GROUP
);
7228 task_rq_unlock(rq
, task
, &rf
);
7231 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
7233 struct task_struct
*task
;
7234 struct cgroup_subsys_state
*css
;
7237 cgroup_taskset_for_each(task
, css
, tset
) {
7238 #ifdef CONFIG_RT_GROUP_SCHED
7239 if (!sched_rt_can_attach(css_tg(css
), task
))
7243 * Serialize against wake_up_new_task() such that if its
7244 * running, we're sure to observe its full state.
7246 raw_spin_lock_irq(&task
->pi_lock
);
7248 * Avoid calling sched_move_task() before wake_up_new_task()
7249 * has happened. This would lead to problems with PELT, due to
7250 * move wanting to detach+attach while we're not attached yet.
7252 if (task
->state
== TASK_NEW
)
7254 raw_spin_unlock_irq(&task
->pi_lock
);
7262 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
7264 struct task_struct
*task
;
7265 struct cgroup_subsys_state
*css
;
7267 cgroup_taskset_for_each(task
, css
, tset
)
7268 sched_move_task(task
);
7271 #ifdef CONFIG_UCLAMP_TASK_GROUP
7272 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
7274 struct cgroup_subsys_state
*top_css
= css
;
7275 struct uclamp_se
*uc_parent
= NULL
;
7276 struct uclamp_se
*uc_se
= NULL
;
7277 unsigned int eff
[UCLAMP_CNT
];
7278 enum uclamp_id clamp_id
;
7279 unsigned int clamps
;
7281 css_for_each_descendant_pre(css
, top_css
) {
7282 uc_parent
= css_tg(css
)->parent
7283 ? css_tg(css
)->parent
->uclamp
: NULL
;
7285 for_each_clamp_id(clamp_id
) {
7286 /* Assume effective clamps matches requested clamps */
7287 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
7288 /* Cap effective clamps with parent's effective clamps */
7290 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
7291 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
7294 /* Ensure protection is always capped by limit */
7295 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
7297 /* Propagate most restrictive effective clamps */
7299 uc_se
= css_tg(css
)->uclamp
;
7300 for_each_clamp_id(clamp_id
) {
7301 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
7303 uc_se
[clamp_id
].value
= eff
[clamp_id
];
7304 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
7305 clamps
|= (0x1 << clamp_id
);
7308 css
= css_rightmost_descendant(css
);
7312 /* Immediately update descendants RUNNABLE tasks */
7313 uclamp_update_active_tasks(css
, clamps
);
7318 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7319 * C expression. Since there is no way to convert a macro argument (N) into a
7320 * character constant, use two levels of macros.
7322 #define _POW10(exp) ((unsigned int)1e##exp)
7323 #define POW10(exp) _POW10(exp)
7325 struct uclamp_request
{
7326 #define UCLAMP_PERCENT_SHIFT 2
7327 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7333 static inline struct uclamp_request
7334 capacity_from_percent(char *buf
)
7336 struct uclamp_request req
= {
7337 .percent
= UCLAMP_PERCENT_SCALE
,
7338 .util
= SCHED_CAPACITY_SCALE
,
7343 if (strcmp(buf
, "max")) {
7344 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
7348 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
7353 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
7354 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
7360 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
7361 size_t nbytes
, loff_t off
,
7362 enum uclamp_id clamp_id
)
7364 struct uclamp_request req
;
7365 struct task_group
*tg
;
7367 req
= capacity_from_percent(buf
);
7371 mutex_lock(&uclamp_mutex
);
7374 tg
= css_tg(of_css(of
));
7375 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
7376 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
7379 * Because of not recoverable conversion rounding we keep track of the
7380 * exact requested value
7382 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
7384 /* Update effective clamps to track the most restrictive value */
7385 cpu_util_update_eff(of_css(of
));
7388 mutex_unlock(&uclamp_mutex
);
7393 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
7394 char *buf
, size_t nbytes
,
7397 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
7400 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
7401 char *buf
, size_t nbytes
,
7404 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
7407 static inline void cpu_uclamp_print(struct seq_file
*sf
,
7408 enum uclamp_id clamp_id
)
7410 struct task_group
*tg
;
7416 tg
= css_tg(seq_css(sf
));
7417 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
7420 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
7421 seq_puts(sf
, "max\n");
7425 percent
= tg
->uclamp_pct
[clamp_id
];
7426 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
7427 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
7430 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
7432 cpu_uclamp_print(sf
, UCLAMP_MIN
);
7436 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
7438 cpu_uclamp_print(sf
, UCLAMP_MAX
);
7441 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7443 #ifdef CONFIG_FAIR_GROUP_SCHED
7444 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7445 struct cftype
*cftype
, u64 shareval
)
7447 if (shareval
> scale_load_down(ULONG_MAX
))
7448 shareval
= MAX_SHARES
;
7449 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7452 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7455 struct task_group
*tg
= css_tg(css
);
7457 return (u64
) scale_load_down(tg
->shares
);
7460 #ifdef CONFIG_CFS_BANDWIDTH
7461 static DEFINE_MUTEX(cfs_constraints_mutex
);
7463 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7464 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7465 /* More than 203 days if BW_SHIFT equals 20. */
7466 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
7468 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7470 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7472 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7473 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7475 if (tg
== &root_task_group
)
7479 * Ensure we have at some amount of bandwidth every period. This is
7480 * to prevent reaching a state of large arrears when throttled via
7481 * entity_tick() resulting in prolonged exit starvation.
7483 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7487 * Likewise, bound things on the otherside by preventing insane quota
7488 * periods. This also allows us to normalize in computing quota
7491 if (period
> max_cfs_quota_period
)
7495 * Bound quota to defend quota against overflow during bandwidth shift.
7497 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
7501 * Prevent race between setting of cfs_rq->runtime_enabled and
7502 * unthrottle_offline_cfs_rqs().
7505 mutex_lock(&cfs_constraints_mutex
);
7506 ret
= __cfs_schedulable(tg
, period
, quota
);
7510 runtime_enabled
= quota
!= RUNTIME_INF
;
7511 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7513 * If we need to toggle cfs_bandwidth_used, off->on must occur
7514 * before making related changes, and on->off must occur afterwards
7516 if (runtime_enabled
&& !runtime_was_enabled
)
7517 cfs_bandwidth_usage_inc();
7518 raw_spin_lock_irq(&cfs_b
->lock
);
7519 cfs_b
->period
= ns_to_ktime(period
);
7520 cfs_b
->quota
= quota
;
7522 __refill_cfs_bandwidth_runtime(cfs_b
);
7524 /* Restart the period timer (if active) to handle new period expiry: */
7525 if (runtime_enabled
)
7526 start_cfs_bandwidth(cfs_b
);
7528 raw_spin_unlock_irq(&cfs_b
->lock
);
7530 for_each_online_cpu(i
) {
7531 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7532 struct rq
*rq
= cfs_rq
->rq
;
7535 rq_lock_irq(rq
, &rf
);
7536 cfs_rq
->runtime_enabled
= runtime_enabled
;
7537 cfs_rq
->runtime_remaining
= 0;
7539 if (cfs_rq
->throttled
)
7540 unthrottle_cfs_rq(cfs_rq
);
7541 rq_unlock_irq(rq
, &rf
);
7543 if (runtime_was_enabled
&& !runtime_enabled
)
7544 cfs_bandwidth_usage_dec();
7546 mutex_unlock(&cfs_constraints_mutex
);
7552 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7556 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7557 if (cfs_quota_us
< 0)
7558 quota
= RUNTIME_INF
;
7559 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
7560 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7564 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7567 static long tg_get_cfs_quota(struct task_group
*tg
)
7571 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7574 quota_us
= tg
->cfs_bandwidth
.quota
;
7575 do_div(quota_us
, NSEC_PER_USEC
);
7580 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7584 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
7587 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7588 quota
= tg
->cfs_bandwidth
.quota
;
7590 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7593 static long tg_get_cfs_period(struct task_group
*tg
)
7597 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7598 do_div(cfs_period_us
, NSEC_PER_USEC
);
7600 return cfs_period_us
;
7603 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7606 return tg_get_cfs_quota(css_tg(css
));
7609 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7610 struct cftype
*cftype
, s64 cfs_quota_us
)
7612 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7615 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7618 return tg_get_cfs_period(css_tg(css
));
7621 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7622 struct cftype
*cftype
, u64 cfs_period_us
)
7624 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7627 struct cfs_schedulable_data
{
7628 struct task_group
*tg
;
7633 * normalize group quota/period to be quota/max_period
7634 * note: units are usecs
7636 static u64
normalize_cfs_quota(struct task_group
*tg
,
7637 struct cfs_schedulable_data
*d
)
7645 period
= tg_get_cfs_period(tg
);
7646 quota
= tg_get_cfs_quota(tg
);
7649 /* note: these should typically be equivalent */
7650 if (quota
== RUNTIME_INF
|| quota
== -1)
7653 return to_ratio(period
, quota
);
7656 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7658 struct cfs_schedulable_data
*d
= data
;
7659 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7660 s64 quota
= 0, parent_quota
= -1;
7663 quota
= RUNTIME_INF
;
7665 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7667 quota
= normalize_cfs_quota(tg
, d
);
7668 parent_quota
= parent_b
->hierarchical_quota
;
7671 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7672 * always take the min. On cgroup1, only inherit when no
7675 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
7676 quota
= min(quota
, parent_quota
);
7678 if (quota
== RUNTIME_INF
)
7679 quota
= parent_quota
;
7680 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7684 cfs_b
->hierarchical_quota
= quota
;
7689 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7692 struct cfs_schedulable_data data
= {
7698 if (quota
!= RUNTIME_INF
) {
7699 do_div(data
.period
, NSEC_PER_USEC
);
7700 do_div(data
.quota
, NSEC_PER_USEC
);
7704 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7710 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
7712 struct task_group
*tg
= css_tg(seq_css(sf
));
7713 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7715 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7716 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7717 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7719 if (schedstat_enabled() && tg
!= &root_task_group
) {
7723 for_each_possible_cpu(i
)
7724 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
7726 seq_printf(sf
, "wait_sum %llu\n", ws
);
7731 #endif /* CONFIG_CFS_BANDWIDTH */
7732 #endif /* CONFIG_FAIR_GROUP_SCHED */
7734 #ifdef CONFIG_RT_GROUP_SCHED
7735 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7736 struct cftype
*cft
, s64 val
)
7738 return sched_group_set_rt_runtime(css_tg(css
), val
);
7741 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7744 return sched_group_rt_runtime(css_tg(css
));
7747 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7748 struct cftype
*cftype
, u64 rt_period_us
)
7750 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7753 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7756 return sched_group_rt_period(css_tg(css
));
7758 #endif /* CONFIG_RT_GROUP_SCHED */
7760 static struct cftype cpu_legacy_files
[] = {
7761 #ifdef CONFIG_FAIR_GROUP_SCHED
7764 .read_u64
= cpu_shares_read_u64
,
7765 .write_u64
= cpu_shares_write_u64
,
7768 #ifdef CONFIG_CFS_BANDWIDTH
7770 .name
= "cfs_quota_us",
7771 .read_s64
= cpu_cfs_quota_read_s64
,
7772 .write_s64
= cpu_cfs_quota_write_s64
,
7775 .name
= "cfs_period_us",
7776 .read_u64
= cpu_cfs_period_read_u64
,
7777 .write_u64
= cpu_cfs_period_write_u64
,
7781 .seq_show
= cpu_cfs_stat_show
,
7784 #ifdef CONFIG_RT_GROUP_SCHED
7786 .name
= "rt_runtime_us",
7787 .read_s64
= cpu_rt_runtime_read
,
7788 .write_s64
= cpu_rt_runtime_write
,
7791 .name
= "rt_period_us",
7792 .read_u64
= cpu_rt_period_read_uint
,
7793 .write_u64
= cpu_rt_period_write_uint
,
7796 #ifdef CONFIG_UCLAMP_TASK_GROUP
7798 .name
= "uclamp.min",
7799 .flags
= CFTYPE_NOT_ON_ROOT
,
7800 .seq_show
= cpu_uclamp_min_show
,
7801 .write
= cpu_uclamp_min_write
,
7804 .name
= "uclamp.max",
7805 .flags
= CFTYPE_NOT_ON_ROOT
,
7806 .seq_show
= cpu_uclamp_max_show
,
7807 .write
= cpu_uclamp_max_write
,
7813 static int cpu_extra_stat_show(struct seq_file
*sf
,
7814 struct cgroup_subsys_state
*css
)
7816 #ifdef CONFIG_CFS_BANDWIDTH
7818 struct task_group
*tg
= css_tg(css
);
7819 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7822 throttled_usec
= cfs_b
->throttled_time
;
7823 do_div(throttled_usec
, NSEC_PER_USEC
);
7825 seq_printf(sf
, "nr_periods %d\n"
7827 "throttled_usec %llu\n",
7828 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
7835 #ifdef CONFIG_FAIR_GROUP_SCHED
7836 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
7839 struct task_group
*tg
= css_tg(css
);
7840 u64 weight
= scale_load_down(tg
->shares
);
7842 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
7845 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
7846 struct cftype
*cft
, u64 weight
)
7849 * cgroup weight knobs should use the common MIN, DFL and MAX
7850 * values which are 1, 100 and 10000 respectively. While it loses
7851 * a bit of range on both ends, it maps pretty well onto the shares
7852 * value used by scheduler and the round-trip conversions preserve
7853 * the original value over the entire range.
7855 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
7858 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
7860 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7863 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
7866 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
7867 int last_delta
= INT_MAX
;
7870 /* find the closest nice value to the current weight */
7871 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
7872 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
7873 if (delta
>= last_delta
)
7878 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
7881 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
7882 struct cftype
*cft
, s64 nice
)
7884 unsigned long weight
;
7887 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
7890 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
7891 idx
= array_index_nospec(idx
, 40);
7892 weight
= sched_prio_to_weight
[idx
];
7894 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7898 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
7899 long period
, long quota
)
7902 seq_puts(sf
, "max");
7904 seq_printf(sf
, "%ld", quota
);
7906 seq_printf(sf
, " %ld\n", period
);
7909 /* caller should put the current value in *@periodp before calling */
7910 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
7911 u64
*periodp
, u64
*quotap
)
7913 char tok
[21]; /* U64_MAX */
7915 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
7918 *periodp
*= NSEC_PER_USEC
;
7920 if (sscanf(tok
, "%llu", quotap
))
7921 *quotap
*= NSEC_PER_USEC
;
7922 else if (!strcmp(tok
, "max"))
7923 *quotap
= RUNTIME_INF
;
7930 #ifdef CONFIG_CFS_BANDWIDTH
7931 static int cpu_max_show(struct seq_file
*sf
, void *v
)
7933 struct task_group
*tg
= css_tg(seq_css(sf
));
7935 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
7939 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
7940 char *buf
, size_t nbytes
, loff_t off
)
7942 struct task_group
*tg
= css_tg(of_css(of
));
7943 u64 period
= tg_get_cfs_period(tg
);
7947 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
7949 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
7950 return ret
?: nbytes
;
7954 static struct cftype cpu_files
[] = {
7955 #ifdef CONFIG_FAIR_GROUP_SCHED
7958 .flags
= CFTYPE_NOT_ON_ROOT
,
7959 .read_u64
= cpu_weight_read_u64
,
7960 .write_u64
= cpu_weight_write_u64
,
7963 .name
= "weight.nice",
7964 .flags
= CFTYPE_NOT_ON_ROOT
,
7965 .read_s64
= cpu_weight_nice_read_s64
,
7966 .write_s64
= cpu_weight_nice_write_s64
,
7969 #ifdef CONFIG_CFS_BANDWIDTH
7972 .flags
= CFTYPE_NOT_ON_ROOT
,
7973 .seq_show
= cpu_max_show
,
7974 .write
= cpu_max_write
,
7977 #ifdef CONFIG_UCLAMP_TASK_GROUP
7979 .name
= "uclamp.min",
7980 .flags
= CFTYPE_NOT_ON_ROOT
,
7981 .seq_show
= cpu_uclamp_min_show
,
7982 .write
= cpu_uclamp_min_write
,
7985 .name
= "uclamp.max",
7986 .flags
= CFTYPE_NOT_ON_ROOT
,
7987 .seq_show
= cpu_uclamp_max_show
,
7988 .write
= cpu_uclamp_max_write
,
7994 struct cgroup_subsys cpu_cgrp_subsys
= {
7995 .css_alloc
= cpu_cgroup_css_alloc
,
7996 .css_online
= cpu_cgroup_css_online
,
7997 .css_released
= cpu_cgroup_css_released
,
7998 .css_free
= cpu_cgroup_css_free
,
7999 .css_extra_stat_show
= cpu_extra_stat_show
,
8000 .fork
= cpu_cgroup_fork
,
8001 .can_attach
= cpu_cgroup_can_attach
,
8002 .attach
= cpu_cgroup_attach
,
8003 .legacy_cftypes
= cpu_legacy_files
,
8004 .dfl_cftypes
= cpu_files
,
8009 #endif /* CONFIG_CGROUP_SCHED */
8011 void dump_cpu_task(int cpu
)
8013 pr_info("Task dump for CPU %d:\n", cpu
);
8014 sched_show_task(cpu_curr(cpu
));
8018 * Nice levels are multiplicative, with a gentle 10% change for every
8019 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8020 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8021 * that remained on nice 0.
8023 * The "10% effect" is relative and cumulative: from _any_ nice level,
8024 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8025 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8026 * If a task goes up by ~10% and another task goes down by ~10% then
8027 * the relative distance between them is ~25%.)
8029 const int sched_prio_to_weight
[40] = {
8030 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8031 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8032 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8033 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8034 /* 0 */ 1024, 820, 655, 526, 423,
8035 /* 5 */ 335, 272, 215, 172, 137,
8036 /* 10 */ 110, 87, 70, 56, 45,
8037 /* 15 */ 36, 29, 23, 18, 15,
8041 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8043 * In cases where the weight does not change often, we can use the
8044 * precalculated inverse to speed up arithmetics by turning divisions
8045 * into multiplications:
8047 const u32 sched_prio_to_wmult
[40] = {
8048 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8049 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8050 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8051 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8052 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8053 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8054 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8055 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8058 #undef CREATE_TRACE_POINTS