1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp
);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp
);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp
);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp
);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug
unsigned int sysctl_sched_features
=
64 * Number of tasks to iterate in a single balance run.
65 * Limited because this is done with IRQs disabled.
67 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
70 * period over which we measure -rt task CPU usage in us.
73 unsigned int sysctl_sched_rt_period
= 1000000;
75 __read_mostly
int scheduler_running
;
78 * part of the period that we allow rt tasks to run in us.
81 int sysctl_sched_rt_runtime
= 950000;
85 * Serialization rules:
91 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
94 * rq2->lock where: rq1 < rq2
98 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
99 * local CPU's rq->lock, it optionally removes the task from the runqueue and
100 * always looks at the local rq data structures to find the most eligible task
103 * Task enqueue is also under rq->lock, possibly taken from another CPU.
104 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
105 * the local CPU to avoid bouncing the runqueue state around [ see
106 * ttwu_queue_wakelist() ]
108 * Task wakeup, specifically wakeups that involve migration, are horribly
109 * complicated to avoid having to take two rq->locks.
113 * System-calls and anything external will use task_rq_lock() which acquires
114 * both p->pi_lock and rq->lock. As a consequence the state they change is
115 * stable while holding either lock:
117 * - sched_setaffinity()/
118 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
119 * - set_user_nice(): p->se.load, p->*prio
120 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
121 * p->se.load, p->rt_priority,
122 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
123 * - sched_setnuma(): p->numa_preferred_nid
124 * - sched_move_task()/
125 * cpu_cgroup_fork(): p->sched_task_group
126 * - uclamp_update_active() p->uclamp*
128 * p->state <- TASK_*:
130 * is changed locklessly using set_current_state(), __set_current_state() or
131 * set_special_state(), see their respective comments, or by
132 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
135 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
137 * is set by activate_task() and cleared by deactivate_task(), under
138 * rq->lock. Non-zero indicates the task is runnable, the special
139 * ON_RQ_MIGRATING state is used for migration without holding both
140 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
142 * p->on_cpu <- { 0, 1 }:
144 * is set by prepare_task() and cleared by finish_task() such that it will be
145 * set before p is scheduled-in and cleared after p is scheduled-out, both
146 * under rq->lock. Non-zero indicates the task is running on its CPU.
148 * [ The astute reader will observe that it is possible for two tasks on one
149 * CPU to have ->on_cpu = 1 at the same time. ]
151 * task_cpu(p): is changed by set_task_cpu(), the rules are:
153 * - Don't call set_task_cpu() on a blocked task:
155 * We don't care what CPU we're not running on, this simplifies hotplug,
156 * the CPU assignment of blocked tasks isn't required to be valid.
158 * - for try_to_wake_up(), called under p->pi_lock:
160 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
162 * - for migration called under rq->lock:
163 * [ see task_on_rq_migrating() in task_rq_lock() ]
165 * o move_queued_task()
168 * - for migration called under double_rq_lock():
170 * o __migrate_swap_task()
171 * o push_rt_task() / pull_rt_task()
172 * o push_dl_task() / pull_dl_task()
173 * o dl_task_offline_migration()
178 * __task_rq_lock - lock the rq @p resides on.
180 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
185 lockdep_assert_held(&p
->pi_lock
);
189 raw_spin_lock(&rq
->lock
);
190 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
194 raw_spin_unlock(&rq
->lock
);
196 while (unlikely(task_on_rq_migrating(p
)))
202 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
204 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
205 __acquires(p
->pi_lock
)
211 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
213 raw_spin_lock(&rq
->lock
);
215 * move_queued_task() task_rq_lock()
218 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
219 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
220 * [S] ->cpu = new_cpu [L] task_rq()
224 * If we observe the old CPU in task_rq_lock(), the acquire of
225 * the old rq->lock will fully serialize against the stores.
227 * If we observe the new CPU in task_rq_lock(), the address
228 * dependency headed by '[L] rq = task_rq()' and the acquire
229 * will pair with the WMB to ensure we then also see migrating.
231 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
235 raw_spin_unlock(&rq
->lock
);
236 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
238 while (unlikely(task_on_rq_migrating(p
)))
244 * RQ-clock updating methods:
247 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
250 * In theory, the compile should just see 0 here, and optimize out the call
251 * to sched_rt_avg_update. But I don't trust it...
253 s64 __maybe_unused steal
= 0, irq_delta
= 0;
255 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
256 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
259 * Since irq_time is only updated on {soft,}irq_exit, we might run into
260 * this case when a previous update_rq_clock() happened inside a
263 * When this happens, we stop ->clock_task and only update the
264 * prev_irq_time stamp to account for the part that fit, so that a next
265 * update will consume the rest. This ensures ->clock_task is
268 * It does however cause some slight miss-attribution of {soft,}irq
269 * time, a more accurate solution would be to update the irq_time using
270 * the current rq->clock timestamp, except that would require using
273 if (irq_delta
> delta
)
276 rq
->prev_irq_time
+= irq_delta
;
279 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
280 if (static_key_false((¶virt_steal_rq_enabled
))) {
281 steal
= paravirt_steal_clock(cpu_of(rq
));
282 steal
-= rq
->prev_steal_time_rq
;
284 if (unlikely(steal
> delta
))
287 rq
->prev_steal_time_rq
+= steal
;
292 rq
->clock_task
+= delta
;
294 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
295 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
296 update_irq_load_avg(rq
, irq_delta
+ steal
);
298 update_rq_clock_pelt(rq
, delta
);
301 void update_rq_clock(struct rq
*rq
)
305 lockdep_assert_held(&rq
->lock
);
307 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
310 #ifdef CONFIG_SCHED_DEBUG
311 if (sched_feat(WARN_DOUBLE_CLOCK
))
312 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
313 rq
->clock_update_flags
|= RQCF_UPDATED
;
316 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
320 update_rq_clock_task(rq
, delta
);
324 rq_csd_init(struct rq
*rq
, call_single_data_t
*csd
, smp_call_func_t func
)
331 #ifdef CONFIG_SCHED_HRTICK
333 * Use HR-timers to deliver accurate preemption points.
336 static void hrtick_clear(struct rq
*rq
)
338 if (hrtimer_active(&rq
->hrtick_timer
))
339 hrtimer_cancel(&rq
->hrtick_timer
);
343 * High-resolution timer tick.
344 * Runs from hardirq context with interrupts disabled.
346 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
348 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
351 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
355 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
358 return HRTIMER_NORESTART
;
363 static void __hrtick_restart(struct rq
*rq
)
365 struct hrtimer
*timer
= &rq
->hrtick_timer
;
367 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
371 * called from hardirq (IPI) context
373 static void __hrtick_start(void *arg
)
379 __hrtick_restart(rq
);
384 * Called to set the hrtick timer state.
386 * called with rq->lock held and irqs disabled
388 void hrtick_start(struct rq
*rq
, u64 delay
)
390 struct hrtimer
*timer
= &rq
->hrtick_timer
;
395 * Don't schedule slices shorter than 10000ns, that just
396 * doesn't make sense and can cause timer DoS.
398 delta
= max_t(s64
, delay
, 10000LL);
399 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
401 hrtimer_set_expires(timer
, time
);
404 __hrtick_restart(rq
);
406 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq
*rq
, u64 delay
)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay
= max_t(u64
, delay
, 10000LL);
422 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
423 HRTIMER_MODE_REL_PINNED_HARD
);
426 #endif /* CONFIG_SMP */
428 static void hrtick_rq_init(struct rq
*rq
)
431 rq_csd_init(rq
, &rq
->hrtick_csd
, __hrtick_start
);
433 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
434 rq
->hrtick_timer
.function
= hrtick
;
436 #else /* CONFIG_SCHED_HRTICK */
437 static inline void hrtick_clear(struct rq
*rq
)
441 static inline void hrtick_rq_init(struct rq
*rq
)
444 #endif /* CONFIG_SCHED_HRTICK */
447 * cmpxchg based fetch_or, macro so it works for different integer types
449 #define fetch_or(ptr, mask) \
451 typeof(ptr) _ptr = (ptr); \
452 typeof(mask) _mask = (mask); \
453 typeof(*_ptr) _old, _val = *_ptr; \
456 _old = cmpxchg(_ptr, _val, _val | _mask); \
464 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
466 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
467 * this avoids any races wrt polling state changes and thereby avoids
470 static bool set_nr_and_not_polling(struct task_struct
*p
)
472 struct thread_info
*ti
= task_thread_info(p
);
473 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
477 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
479 * If this returns true, then the idle task promises to call
480 * sched_ttwu_pending() and reschedule soon.
482 static bool set_nr_if_polling(struct task_struct
*p
)
484 struct thread_info
*ti
= task_thread_info(p
);
485 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
488 if (!(val
& _TIF_POLLING_NRFLAG
))
490 if (val
& _TIF_NEED_RESCHED
)
492 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
501 static bool set_nr_and_not_polling(struct task_struct
*p
)
503 set_tsk_need_resched(p
);
508 static bool set_nr_if_polling(struct task_struct
*p
)
515 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
517 struct wake_q_node
*node
= &task
->wake_q
;
520 * Atomically grab the task, if ->wake_q is !nil already it means
521 * it's already queued (either by us or someone else) and will get the
522 * wakeup due to that.
524 * In order to ensure that a pending wakeup will observe our pending
525 * state, even in the failed case, an explicit smp_mb() must be used.
527 smp_mb__before_atomic();
528 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
532 * The head is context local, there can be no concurrency.
535 head
->lastp
= &node
->next
;
540 * wake_q_add() - queue a wakeup for 'later' waking.
541 * @head: the wake_q_head to add @task to
542 * @task: the task to queue for 'later' wakeup
544 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
545 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
548 * This function must be used as-if it were wake_up_process(); IOW the task
549 * must be ready to be woken at this location.
551 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
553 if (__wake_q_add(head
, task
))
554 get_task_struct(task
);
558 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
559 * @head: the wake_q_head to add @task to
560 * @task: the task to queue for 'later' wakeup
562 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
563 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
566 * This function must be used as-if it were wake_up_process(); IOW the task
567 * must be ready to be woken at this location.
569 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
570 * that already hold reference to @task can call the 'safe' version and trust
571 * wake_q to do the right thing depending whether or not the @task is already
574 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
576 if (!__wake_q_add(head
, task
))
577 put_task_struct(task
);
580 void wake_up_q(struct wake_q_head
*head
)
582 struct wake_q_node
*node
= head
->first
;
584 while (node
!= WAKE_Q_TAIL
) {
585 struct task_struct
*task
;
587 task
= container_of(node
, struct task_struct
, wake_q
);
589 /* Task can safely be re-inserted now: */
591 task
->wake_q
.next
= NULL
;
594 * wake_up_process() executes a full barrier, which pairs with
595 * the queueing in wake_q_add() so as not to miss wakeups.
597 wake_up_process(task
);
598 put_task_struct(task
);
603 * resched_curr - mark rq's current task 'to be rescheduled now'.
605 * On UP this means the setting of the need_resched flag, on SMP it
606 * might also involve a cross-CPU call to trigger the scheduler on
609 void resched_curr(struct rq
*rq
)
611 struct task_struct
*curr
= rq
->curr
;
614 lockdep_assert_held(&rq
->lock
);
616 if (test_tsk_need_resched(curr
))
621 if (cpu
== smp_processor_id()) {
622 set_tsk_need_resched(curr
);
623 set_preempt_need_resched();
627 if (set_nr_and_not_polling(curr
))
628 smp_send_reschedule(cpu
);
630 trace_sched_wake_idle_without_ipi(cpu
);
633 void resched_cpu(int cpu
)
635 struct rq
*rq
= cpu_rq(cpu
);
638 raw_spin_lock_irqsave(&rq
->lock
, flags
);
639 if (cpu_online(cpu
) || cpu
== smp_processor_id())
641 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
645 #ifdef CONFIG_NO_HZ_COMMON
647 * In the semi idle case, use the nearest busy CPU for migrating timers
648 * from an idle CPU. This is good for power-savings.
650 * We don't do similar optimization for completely idle system, as
651 * selecting an idle CPU will add more delays to the timers than intended
652 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
654 int get_nohz_timer_target(void)
656 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
657 struct sched_domain
*sd
;
659 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
666 for_each_domain(cpu
, sd
) {
667 for_each_cpu_and(i
, sched_domain_span(sd
),
668 housekeeping_cpumask(HK_FLAG_TIMER
)) {
679 if (default_cpu
== -1)
680 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
688 * When add_timer_on() enqueues a timer into the timer wheel of an
689 * idle CPU then this timer might expire before the next timer event
690 * which is scheduled to wake up that CPU. In case of a completely
691 * idle system the next event might even be infinite time into the
692 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
693 * leaves the inner idle loop so the newly added timer is taken into
694 * account when the CPU goes back to idle and evaluates the timer
695 * wheel for the next timer event.
697 static void wake_up_idle_cpu(int cpu
)
699 struct rq
*rq
= cpu_rq(cpu
);
701 if (cpu
== smp_processor_id())
704 if (set_nr_and_not_polling(rq
->idle
))
705 smp_send_reschedule(cpu
);
707 trace_sched_wake_idle_without_ipi(cpu
);
710 static bool wake_up_full_nohz_cpu(int cpu
)
713 * We just need the target to call irq_exit() and re-evaluate
714 * the next tick. The nohz full kick at least implies that.
715 * If needed we can still optimize that later with an
718 if (cpu_is_offline(cpu
))
719 return true; /* Don't try to wake offline CPUs. */
720 if (tick_nohz_full_cpu(cpu
)) {
721 if (cpu
!= smp_processor_id() ||
722 tick_nohz_tick_stopped())
723 tick_nohz_full_kick_cpu(cpu
);
731 * Wake up the specified CPU. If the CPU is going offline, it is the
732 * caller's responsibility to deal with the lost wakeup, for example,
733 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
735 void wake_up_nohz_cpu(int cpu
)
737 if (!wake_up_full_nohz_cpu(cpu
))
738 wake_up_idle_cpu(cpu
);
741 static void nohz_csd_func(void *info
)
743 struct rq
*rq
= info
;
744 int cpu
= cpu_of(rq
);
748 * Release the rq::nohz_csd.
750 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
751 WARN_ON(!(flags
& NOHZ_KICK_MASK
));
753 rq
->idle_balance
= idle_cpu(cpu
);
754 if (rq
->idle_balance
&& !need_resched()) {
755 rq
->nohz_idle_balance
= flags
;
756 raise_softirq_irqoff(SCHED_SOFTIRQ
);
760 #endif /* CONFIG_NO_HZ_COMMON */
762 #ifdef CONFIG_NO_HZ_FULL
763 bool sched_can_stop_tick(struct rq
*rq
)
767 /* Deadline tasks, even if single, need the tick */
768 if (rq
->dl
.dl_nr_running
)
772 * If there are more than one RR tasks, we need the tick to affect the
773 * actual RR behaviour.
775 if (rq
->rt
.rr_nr_running
) {
776 if (rq
->rt
.rr_nr_running
== 1)
783 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
784 * forced preemption between FIFO tasks.
786 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
791 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
792 * if there's more than one we need the tick for involuntary
795 if (rq
->nr_running
> 1)
800 #endif /* CONFIG_NO_HZ_FULL */
801 #endif /* CONFIG_SMP */
803 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
804 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
806 * Iterate task_group tree rooted at *from, calling @down when first entering a
807 * node and @up when leaving it for the final time.
809 * Caller must hold rcu_lock or sufficient equivalent.
811 int walk_tg_tree_from(struct task_group
*from
,
812 tg_visitor down
, tg_visitor up
, void *data
)
814 struct task_group
*parent
, *child
;
820 ret
= (*down
)(parent
, data
);
823 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
830 ret
= (*up
)(parent
, data
);
831 if (ret
|| parent
== from
)
835 parent
= parent
->parent
;
842 int tg_nop(struct task_group
*tg
, void *data
)
848 static void set_load_weight(struct task_struct
*p
, bool update_load
)
850 int prio
= p
->static_prio
- MAX_RT_PRIO
;
851 struct load_weight
*load
= &p
->se
.load
;
854 * SCHED_IDLE tasks get minimal weight:
856 if (task_has_idle_policy(p
)) {
857 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
858 load
->inv_weight
= WMULT_IDLEPRIO
;
863 * SCHED_OTHER tasks have to update their load when changing their
866 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
867 reweight_task(p
, prio
);
869 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
870 load
->inv_weight
= sched_prio_to_wmult
[prio
];
874 #ifdef CONFIG_UCLAMP_TASK
876 * Serializes updates of utilization clamp values
878 * The (slow-path) user-space triggers utilization clamp value updates which
879 * can require updates on (fast-path) scheduler's data structures used to
880 * support enqueue/dequeue operations.
881 * While the per-CPU rq lock protects fast-path update operations, user-space
882 * requests are serialized using a mutex to reduce the risk of conflicting
883 * updates or API abuses.
885 static DEFINE_MUTEX(uclamp_mutex
);
887 /* Max allowed minimum utilization */
888 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
890 /* Max allowed maximum utilization */
891 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
894 * By default RT tasks run at the maximum performance point/capacity of the
895 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
896 * SCHED_CAPACITY_SCALE.
898 * This knob allows admins to change the default behavior when uclamp is being
899 * used. In battery powered devices, particularly, running at the maximum
900 * capacity and frequency will increase energy consumption and shorten the
903 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
905 * This knob will not override the system default sched_util_clamp_min defined
908 unsigned int sysctl_sched_uclamp_util_min_rt_default
= SCHED_CAPACITY_SCALE
;
910 /* All clamps are required to be less or equal than these values */
911 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
914 * This static key is used to reduce the uclamp overhead in the fast path. It
915 * primarily disables the call to uclamp_rq_{inc, dec}() in
916 * enqueue/dequeue_task().
918 * This allows users to continue to enable uclamp in their kernel config with
919 * minimum uclamp overhead in the fast path.
921 * As soon as userspace modifies any of the uclamp knobs, the static key is
922 * enabled, since we have an actual users that make use of uclamp
925 * The knobs that would enable this static key are:
927 * * A task modifying its uclamp value with sched_setattr().
928 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
929 * * An admin modifying the cgroup cpu.uclamp.{min, max}
931 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used
);
933 /* Integer rounded range for each bucket */
934 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
936 #define for_each_clamp_id(clamp_id) \
937 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
939 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
941 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
944 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
946 if (clamp_id
== UCLAMP_MIN
)
948 return SCHED_CAPACITY_SCALE
;
951 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
952 unsigned int value
, bool user_defined
)
954 uc_se
->value
= value
;
955 uc_se
->bucket_id
= uclamp_bucket_id(value
);
956 uc_se
->user_defined
= user_defined
;
959 static inline unsigned int
960 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
961 unsigned int clamp_value
)
964 * Avoid blocked utilization pushing up the frequency when we go
965 * idle (which drops the max-clamp) by retaining the last known
968 if (clamp_id
== UCLAMP_MAX
) {
969 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
973 return uclamp_none(UCLAMP_MIN
);
976 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
977 unsigned int clamp_value
)
979 /* Reset max-clamp retention only on idle exit */
980 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
983 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
987 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
988 unsigned int clamp_value
)
990 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
991 int bucket_id
= UCLAMP_BUCKETS
- 1;
994 * Since both min and max clamps are max aggregated, find the
995 * top most bucket with tasks in.
997 for ( ; bucket_id
>= 0; bucket_id
--) {
998 if (!bucket
[bucket_id
].tasks
)
1000 return bucket
[bucket_id
].value
;
1003 /* No tasks -- default clamp values */
1004 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
1007 static void __uclamp_update_util_min_rt_default(struct task_struct
*p
)
1009 unsigned int default_util_min
;
1010 struct uclamp_se
*uc_se
;
1012 lockdep_assert_held(&p
->pi_lock
);
1014 uc_se
= &p
->uclamp_req
[UCLAMP_MIN
];
1016 /* Only sync if user didn't override the default */
1017 if (uc_se
->user_defined
)
1020 default_util_min
= sysctl_sched_uclamp_util_min_rt_default
;
1021 uclamp_se_set(uc_se
, default_util_min
, false);
1024 static void uclamp_update_util_min_rt_default(struct task_struct
*p
)
1032 /* Protect updates to p->uclamp_* */
1033 rq
= task_rq_lock(p
, &rf
);
1034 __uclamp_update_util_min_rt_default(p
);
1035 task_rq_unlock(rq
, p
, &rf
);
1038 static void uclamp_sync_util_min_rt_default(void)
1040 struct task_struct
*g
, *p
;
1043 * copy_process() sysctl_uclamp
1044 * uclamp_min_rt = X;
1045 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1046 * // link thread smp_mb__after_spinlock()
1047 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1048 * sched_post_fork() for_each_process_thread()
1049 * __uclamp_sync_rt() __uclamp_sync_rt()
1051 * Ensures that either sched_post_fork() will observe the new
1052 * uclamp_min_rt or for_each_process_thread() will observe the new
1055 read_lock(&tasklist_lock
);
1056 smp_mb__after_spinlock();
1057 read_unlock(&tasklist_lock
);
1060 for_each_process_thread(g
, p
)
1061 uclamp_update_util_min_rt_default(p
);
1065 static inline struct uclamp_se
1066 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
1068 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
1069 #ifdef CONFIG_UCLAMP_TASK_GROUP
1070 struct uclamp_se uc_max
;
1073 * Tasks in autogroups or root task group will be
1074 * restricted by system defaults.
1076 if (task_group_is_autogroup(task_group(p
)))
1078 if (task_group(p
) == &root_task_group
)
1081 uc_max
= task_group(p
)->uclamp
[clamp_id
];
1082 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
1090 * The effective clamp bucket index of a task depends on, by increasing
1092 * - the task specific clamp value, when explicitly requested from userspace
1093 * - the task group effective clamp value, for tasks not either in the root
1094 * group or in an autogroup
1095 * - the system default clamp value, defined by the sysadmin
1097 static inline struct uclamp_se
1098 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
1100 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
1101 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
1103 /* System default restrictions always apply */
1104 if (unlikely(uc_req
.value
> uc_max
.value
))
1110 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
1112 struct uclamp_se uc_eff
;
1114 /* Task currently refcounted: use back-annotated (effective) value */
1115 if (p
->uclamp
[clamp_id
].active
)
1116 return (unsigned long)p
->uclamp
[clamp_id
].value
;
1118 uc_eff
= uclamp_eff_get(p
, clamp_id
);
1120 return (unsigned long)uc_eff
.value
;
1124 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1125 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1126 * updates the rq's clamp value if required.
1128 * Tasks can have a task-specific value requested from user-space, track
1129 * within each bucket the maximum value for tasks refcounted in it.
1130 * This "local max aggregation" allows to track the exact "requested" value
1131 * for each bucket when all its RUNNABLE tasks require the same clamp.
1133 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
1134 enum uclamp_id clamp_id
)
1136 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1137 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1138 struct uclamp_bucket
*bucket
;
1140 lockdep_assert_held(&rq
->lock
);
1142 /* Update task effective clamp */
1143 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
1145 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1147 uc_se
->active
= true;
1149 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
1152 * Local max aggregation: rq buckets always track the max
1153 * "requested" clamp value of its RUNNABLE tasks.
1155 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
1156 bucket
->value
= uc_se
->value
;
1158 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
1159 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
1163 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1164 * is released. If this is the last task reference counting the rq's max
1165 * active clamp value, then the rq's clamp value is updated.
1167 * Both refcounted tasks and rq's cached clamp values are expected to be
1168 * always valid. If it's detected they are not, as defensive programming,
1169 * enforce the expected state and warn.
1171 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
1172 enum uclamp_id clamp_id
)
1174 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1175 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1176 struct uclamp_bucket
*bucket
;
1177 unsigned int bkt_clamp
;
1178 unsigned int rq_clamp
;
1180 lockdep_assert_held(&rq
->lock
);
1183 * If sched_uclamp_used was enabled after task @p was enqueued,
1184 * we could end up with unbalanced call to uclamp_rq_dec_id().
1186 * In this case the uc_se->active flag should be false since no uclamp
1187 * accounting was performed at enqueue time and we can just return
1190 * Need to be careful of the following enqueue/dequeue ordering
1194 * // sched_uclamp_used gets enabled
1197 * // Must not decrement bucket->tasks here
1200 * where we could end up with stale data in uc_se and
1201 * bucket[uc_se->bucket_id].
1203 * The following check here eliminates the possibility of such race.
1205 if (unlikely(!uc_se
->active
))
1208 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1210 SCHED_WARN_ON(!bucket
->tasks
);
1211 if (likely(bucket
->tasks
))
1214 uc_se
->active
= false;
1217 * Keep "local max aggregation" simple and accept to (possibly)
1218 * overboost some RUNNABLE tasks in the same bucket.
1219 * The rq clamp bucket value is reset to its base value whenever
1220 * there are no more RUNNABLE tasks refcounting it.
1222 if (likely(bucket
->tasks
))
1225 rq_clamp
= READ_ONCE(uc_rq
->value
);
1227 * Defensive programming: this should never happen. If it happens,
1228 * e.g. due to future modification, warn and fixup the expected value.
1230 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1231 if (bucket
->value
>= rq_clamp
) {
1232 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1233 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1237 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1239 enum uclamp_id clamp_id
;
1242 * Avoid any overhead until uclamp is actually used by the userspace.
1244 * The condition is constructed such that a NOP is generated when
1245 * sched_uclamp_used is disabled.
1247 if (!static_branch_unlikely(&sched_uclamp_used
))
1250 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1253 for_each_clamp_id(clamp_id
)
1254 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1256 /* Reset clamp idle holding when there is one RUNNABLE task */
1257 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1258 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1261 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1263 enum uclamp_id clamp_id
;
1266 * Avoid any overhead until uclamp is actually used by the userspace.
1268 * The condition is constructed such that a NOP is generated when
1269 * sched_uclamp_used is disabled.
1271 if (!static_branch_unlikely(&sched_uclamp_used
))
1274 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1277 for_each_clamp_id(clamp_id
)
1278 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1282 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1288 * Lock the task and the rq where the task is (or was) queued.
1290 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1291 * price to pay to safely serialize util_{min,max} updates with
1292 * enqueues, dequeues and migration operations.
1293 * This is the same locking schema used by __set_cpus_allowed_ptr().
1295 rq
= task_rq_lock(p
, &rf
);
1298 * Setting the clamp bucket is serialized by task_rq_lock().
1299 * If the task is not yet RUNNABLE and its task_struct is not
1300 * affecting a valid clamp bucket, the next time it's enqueued,
1301 * it will already see the updated clamp bucket value.
1303 if (p
->uclamp
[clamp_id
].active
) {
1304 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1305 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1308 task_rq_unlock(rq
, p
, &rf
);
1311 #ifdef CONFIG_UCLAMP_TASK_GROUP
1313 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1314 unsigned int clamps
)
1316 enum uclamp_id clamp_id
;
1317 struct css_task_iter it
;
1318 struct task_struct
*p
;
1320 css_task_iter_start(css
, 0, &it
);
1321 while ((p
= css_task_iter_next(&it
))) {
1322 for_each_clamp_id(clamp_id
) {
1323 if ((0x1 << clamp_id
) & clamps
)
1324 uclamp_update_active(p
, clamp_id
);
1327 css_task_iter_end(&it
);
1330 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1331 static void uclamp_update_root_tg(void)
1333 struct task_group
*tg
= &root_task_group
;
1335 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1336 sysctl_sched_uclamp_util_min
, false);
1337 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1338 sysctl_sched_uclamp_util_max
, false);
1341 cpu_util_update_eff(&root_task_group
.css
);
1345 static void uclamp_update_root_tg(void) { }
1348 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1349 void *buffer
, size_t *lenp
, loff_t
*ppos
)
1351 bool update_root_tg
= false;
1352 int old_min
, old_max
, old_min_rt
;
1355 mutex_lock(&uclamp_mutex
);
1356 old_min
= sysctl_sched_uclamp_util_min
;
1357 old_max
= sysctl_sched_uclamp_util_max
;
1358 old_min_rt
= sysctl_sched_uclamp_util_min_rt_default
;
1360 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1366 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1367 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
||
1368 sysctl_sched_uclamp_util_min_rt_default
> SCHED_CAPACITY_SCALE
) {
1374 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1375 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1376 sysctl_sched_uclamp_util_min
, false);
1377 update_root_tg
= true;
1379 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1380 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1381 sysctl_sched_uclamp_util_max
, false);
1382 update_root_tg
= true;
1385 if (update_root_tg
) {
1386 static_branch_enable(&sched_uclamp_used
);
1387 uclamp_update_root_tg();
1390 if (old_min_rt
!= sysctl_sched_uclamp_util_min_rt_default
) {
1391 static_branch_enable(&sched_uclamp_used
);
1392 uclamp_sync_util_min_rt_default();
1396 * We update all RUNNABLE tasks only when task groups are in use.
1397 * Otherwise, keep it simple and do just a lazy update at each next
1398 * task enqueue time.
1404 sysctl_sched_uclamp_util_min
= old_min
;
1405 sysctl_sched_uclamp_util_max
= old_max
;
1406 sysctl_sched_uclamp_util_min_rt_default
= old_min_rt
;
1408 mutex_unlock(&uclamp_mutex
);
1413 static int uclamp_validate(struct task_struct
*p
,
1414 const struct sched_attr
*attr
)
1416 int util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1417 int util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1419 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1420 util_min
= attr
->sched_util_min
;
1422 if (util_min
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1426 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1427 util_max
= attr
->sched_util_max
;
1429 if (util_max
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1433 if (util_min
!= -1 && util_max
!= -1 && util_min
> util_max
)
1437 * We have valid uclamp attributes; make sure uclamp is enabled.
1439 * We need to do that here, because enabling static branches is a
1440 * blocking operation which obviously cannot be done while holding
1443 static_branch_enable(&sched_uclamp_used
);
1448 static bool uclamp_reset(const struct sched_attr
*attr
,
1449 enum uclamp_id clamp_id
,
1450 struct uclamp_se
*uc_se
)
1452 /* Reset on sched class change for a non user-defined clamp value. */
1453 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)) &&
1454 !uc_se
->user_defined
)
1457 /* Reset on sched_util_{min,max} == -1. */
1458 if (clamp_id
== UCLAMP_MIN
&&
1459 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1460 attr
->sched_util_min
== -1) {
1464 if (clamp_id
== UCLAMP_MAX
&&
1465 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1466 attr
->sched_util_max
== -1) {
1473 static void __setscheduler_uclamp(struct task_struct
*p
,
1474 const struct sched_attr
*attr
)
1476 enum uclamp_id clamp_id
;
1478 for_each_clamp_id(clamp_id
) {
1479 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1482 if (!uclamp_reset(attr
, clamp_id
, uc_se
))
1486 * RT by default have a 100% boost value that could be modified
1489 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1490 value
= sysctl_sched_uclamp_util_min_rt_default
;
1492 value
= uclamp_none(clamp_id
);
1494 uclamp_se_set(uc_se
, value
, false);
1498 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1501 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1502 attr
->sched_util_min
!= -1) {
1503 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1504 attr
->sched_util_min
, true);
1507 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1508 attr
->sched_util_max
!= -1) {
1509 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1510 attr
->sched_util_max
, true);
1514 static void uclamp_fork(struct task_struct
*p
)
1516 enum uclamp_id clamp_id
;
1519 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1520 * as the task is still at its early fork stages.
1522 for_each_clamp_id(clamp_id
)
1523 p
->uclamp
[clamp_id
].active
= false;
1525 if (likely(!p
->sched_reset_on_fork
))
1528 for_each_clamp_id(clamp_id
) {
1529 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1530 uclamp_none(clamp_id
), false);
1534 static void uclamp_post_fork(struct task_struct
*p
)
1536 uclamp_update_util_min_rt_default(p
);
1539 static void __init
init_uclamp_rq(struct rq
*rq
)
1541 enum uclamp_id clamp_id
;
1542 struct uclamp_rq
*uc_rq
= rq
->uclamp
;
1544 for_each_clamp_id(clamp_id
) {
1545 uc_rq
[clamp_id
] = (struct uclamp_rq
) {
1546 .value
= uclamp_none(clamp_id
)
1550 rq
->uclamp_flags
= 0;
1553 static void __init
init_uclamp(void)
1555 struct uclamp_se uc_max
= {};
1556 enum uclamp_id clamp_id
;
1559 for_each_possible_cpu(cpu
)
1560 init_uclamp_rq(cpu_rq(cpu
));
1562 for_each_clamp_id(clamp_id
) {
1563 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1564 uclamp_none(clamp_id
), false);
1567 /* System defaults allow max clamp values for both indexes */
1568 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1569 for_each_clamp_id(clamp_id
) {
1570 uclamp_default
[clamp_id
] = uc_max
;
1571 #ifdef CONFIG_UCLAMP_TASK_GROUP
1572 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1573 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1578 #else /* CONFIG_UCLAMP_TASK */
1579 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1580 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1581 static inline int uclamp_validate(struct task_struct
*p
,
1582 const struct sched_attr
*attr
)
1586 static void __setscheduler_uclamp(struct task_struct
*p
,
1587 const struct sched_attr
*attr
) { }
1588 static inline void uclamp_fork(struct task_struct
*p
) { }
1589 static inline void uclamp_post_fork(struct task_struct
*p
) { }
1590 static inline void init_uclamp(void) { }
1591 #endif /* CONFIG_UCLAMP_TASK */
1593 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1595 if (!(flags
& ENQUEUE_NOCLOCK
))
1596 update_rq_clock(rq
);
1598 if (!(flags
& ENQUEUE_RESTORE
)) {
1599 sched_info_queued(rq
, p
);
1600 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1603 uclamp_rq_inc(rq
, p
);
1604 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1607 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1609 if (!(flags
& DEQUEUE_NOCLOCK
))
1610 update_rq_clock(rq
);
1612 if (!(flags
& DEQUEUE_SAVE
)) {
1613 sched_info_dequeued(rq
, p
);
1614 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1617 uclamp_rq_dec(rq
, p
);
1618 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1621 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1623 enqueue_task(rq
, p
, flags
);
1625 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1628 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1630 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1632 dequeue_task(rq
, p
, flags
);
1636 * __normal_prio - return the priority that is based on the static prio
1638 static inline int __normal_prio(struct task_struct
*p
)
1640 return p
->static_prio
;
1644 * Calculate the expected normal priority: i.e. priority
1645 * without taking RT-inheritance into account. Might be
1646 * boosted by interactivity modifiers. Changes upon fork,
1647 * setprio syscalls, and whenever the interactivity
1648 * estimator recalculates.
1650 static inline int normal_prio(struct task_struct
*p
)
1654 if (task_has_dl_policy(p
))
1655 prio
= MAX_DL_PRIO
-1;
1656 else if (task_has_rt_policy(p
))
1657 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1659 prio
= __normal_prio(p
);
1664 * Calculate the current priority, i.e. the priority
1665 * taken into account by the scheduler. This value might
1666 * be boosted by RT tasks, or might be boosted by
1667 * interactivity modifiers. Will be RT if the task got
1668 * RT-boosted. If not then it returns p->normal_prio.
1670 static int effective_prio(struct task_struct
*p
)
1672 p
->normal_prio
= normal_prio(p
);
1674 * If we are RT tasks or we were boosted to RT priority,
1675 * keep the priority unchanged. Otherwise, update priority
1676 * to the normal priority:
1678 if (!rt_prio(p
->prio
))
1679 return p
->normal_prio
;
1684 * task_curr - is this task currently executing on a CPU?
1685 * @p: the task in question.
1687 * Return: 1 if the task is currently executing. 0 otherwise.
1689 inline int task_curr(const struct task_struct
*p
)
1691 return cpu_curr(task_cpu(p
)) == p
;
1695 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1696 * use the balance_callback list if you want balancing.
1698 * this means any call to check_class_changed() must be followed by a call to
1699 * balance_callback().
1701 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1702 const struct sched_class
*prev_class
,
1705 if (prev_class
!= p
->sched_class
) {
1706 if (prev_class
->switched_from
)
1707 prev_class
->switched_from(rq
, p
);
1709 p
->sched_class
->switched_to(rq
, p
);
1710 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1711 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1714 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1716 if (p
->sched_class
== rq
->curr
->sched_class
)
1717 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1718 else if (p
->sched_class
> rq
->curr
->sched_class
)
1722 * A queue event has occurred, and we're going to schedule. In
1723 * this case, we can save a useless back to back clock update.
1725 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1726 rq_clock_skip_update(rq
);
1732 __do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
);
1734 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1735 const struct cpumask
*new_mask
,
1738 static void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
)
1740 if (likely(!p
->migration_disabled
))
1743 if (p
->cpus_ptr
!= &p
->cpus_mask
)
1747 * Violates locking rules! see comment in __do_set_cpus_allowed().
1749 __do_set_cpus_allowed(p
, cpumask_of(rq
->cpu
), SCA_MIGRATE_DISABLE
);
1752 void migrate_disable(void)
1754 struct task_struct
*p
= current
;
1756 if (p
->migration_disabled
) {
1757 p
->migration_disabled
++;
1762 this_rq()->nr_pinned
++;
1763 p
->migration_disabled
= 1;
1766 EXPORT_SYMBOL_GPL(migrate_disable
);
1768 void migrate_enable(void)
1770 struct task_struct
*p
= current
;
1772 if (p
->migration_disabled
> 1) {
1773 p
->migration_disabled
--;
1778 * Ensure stop_task runs either before or after this, and that
1779 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
1782 if (p
->cpus_ptr
!= &p
->cpus_mask
)
1783 __set_cpus_allowed_ptr(p
, &p
->cpus_mask
, SCA_MIGRATE_ENABLE
);
1785 * Mustn't clear migration_disabled() until cpus_ptr points back at the
1786 * regular cpus_mask, otherwise things that race (eg.
1787 * select_fallback_rq) get confused.
1790 p
->migration_disabled
= 0;
1791 this_rq()->nr_pinned
--;
1794 EXPORT_SYMBOL_GPL(migrate_enable
);
1796 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
1798 return rq
->nr_pinned
;
1802 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1803 * __set_cpus_allowed_ptr() and select_fallback_rq().
1805 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1807 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1810 if (is_per_cpu_kthread(p
) || is_migration_disabled(p
))
1811 return cpu_online(cpu
);
1813 return cpu_active(cpu
);
1817 * This is how migration works:
1819 * 1) we invoke migration_cpu_stop() on the target CPU using
1821 * 2) stopper starts to run (implicitly forcing the migrated thread
1823 * 3) it checks whether the migrated task is still in the wrong runqueue.
1824 * 4) if it's in the wrong runqueue then the migration thread removes
1825 * it and puts it into the right queue.
1826 * 5) stopper completes and stop_one_cpu() returns and the migration
1831 * move_queued_task - move a queued task to new rq.
1833 * Returns (locked) new rq. Old rq's lock is released.
1835 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1836 struct task_struct
*p
, int new_cpu
)
1838 lockdep_assert_held(&rq
->lock
);
1840 deactivate_task(rq
, p
, DEQUEUE_NOCLOCK
);
1841 set_task_cpu(p
, new_cpu
);
1844 rq
= cpu_rq(new_cpu
);
1847 BUG_ON(task_cpu(p
) != new_cpu
);
1848 activate_task(rq
, p
, 0);
1849 check_preempt_curr(rq
, p
, 0);
1854 struct migration_arg
{
1855 struct task_struct
*task
;
1857 struct set_affinity_pending
*pending
;
1860 struct set_affinity_pending
{
1862 struct completion done
;
1863 struct cpu_stop_work stop_work
;
1864 struct migration_arg arg
;
1868 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1869 * this because either it can't run here any more (set_cpus_allowed()
1870 * away from this CPU, or CPU going down), or because we're
1871 * attempting to rebalance this task on exec (sched_exec).
1873 * So we race with normal scheduler movements, but that's OK, as long
1874 * as the task is no longer on this CPU.
1876 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1877 struct task_struct
*p
, int dest_cpu
)
1879 /* Affinity changed (again). */
1880 if (!is_cpu_allowed(p
, dest_cpu
))
1883 update_rq_clock(rq
);
1884 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1890 * migration_cpu_stop - this will be executed by a highprio stopper thread
1891 * and performs thread migration by bumping thread off CPU then
1892 * 'pushing' onto another runqueue.
1894 static int migration_cpu_stop(void *data
)
1896 struct set_affinity_pending
*pending
;
1897 struct migration_arg
*arg
= data
;
1898 struct task_struct
*p
= arg
->task
;
1899 int dest_cpu
= arg
->dest_cpu
;
1900 struct rq
*rq
= this_rq();
1901 bool complete
= false;
1905 * The original target CPU might have gone down and we might
1906 * be on another CPU but it doesn't matter.
1908 local_irq_save(rf
.flags
);
1910 * We need to explicitly wake pending tasks before running
1911 * __migrate_task() such that we will not miss enforcing cpus_ptr
1912 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1914 flush_smp_call_function_from_idle();
1916 raw_spin_lock(&p
->pi_lock
);
1919 pending
= p
->migration_pending
;
1921 * If task_rq(p) != rq, it cannot be migrated here, because we're
1922 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1923 * we're holding p->pi_lock.
1925 if (task_rq(p
) == rq
) {
1926 if (is_migration_disabled(p
))
1930 p
->migration_pending
= NULL
;
1934 /* migrate_enable() -- we must not race against SCA */
1937 * When this was migrate_enable() but we no longer
1938 * have a @pending, a concurrent SCA 'fixed' things
1939 * and we should be valid again. Nothing to do.
1942 WARN_ON_ONCE(!cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
));
1946 dest_cpu
= cpumask_any_distribute(&p
->cpus_mask
);
1949 if (task_on_rq_queued(p
))
1950 rq
= __migrate_task(rq
, &rf
, p
, dest_cpu
);
1952 p
->wake_cpu
= dest_cpu
;
1954 } else if (dest_cpu
< 0 || pending
) {
1956 * This happens when we get migrated between migrate_enable()'s
1957 * preempt_enable() and scheduling the stopper task. At that
1958 * point we're a regular task again and not current anymore.
1960 * A !PREEMPT kernel has a giant hole here, which makes it far
1965 * The task moved before the stopper got to run. We're holding
1966 * ->pi_lock, so the allowed mask is stable - if it got
1967 * somewhere allowed, we're done.
1969 if (pending
&& cpumask_test_cpu(task_cpu(p
), p
->cpus_ptr
)) {
1970 p
->migration_pending
= NULL
;
1976 * When this was migrate_enable() but we no longer have an
1977 * @pending, a concurrent SCA 'fixed' things and we should be
1978 * valid again. Nothing to do.
1981 WARN_ON_ONCE(!cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
));
1986 * When migrate_enable() hits a rq mis-match we can't reliably
1987 * determine is_migration_disabled() and so have to chase after
1990 task_rq_unlock(rq
, p
, &rf
);
1991 stop_one_cpu_nowait(task_cpu(p
), migration_cpu_stop
,
1992 &pending
->arg
, &pending
->stop_work
);
1996 task_rq_unlock(rq
, p
, &rf
);
1999 complete_all(&pending
->done
);
2001 /* For pending->{arg,stop_work} */
2002 pending
= arg
->pending
;
2003 if (pending
&& refcount_dec_and_test(&pending
->refs
))
2004 wake_up_var(&pending
->refs
);
2009 int push_cpu_stop(void *arg
)
2011 struct rq
*lowest_rq
= NULL
, *rq
= this_rq();
2012 struct task_struct
*p
= arg
;
2014 raw_spin_lock_irq(&p
->pi_lock
);
2015 raw_spin_lock(&rq
->lock
);
2017 if (task_rq(p
) != rq
)
2020 if (is_migration_disabled(p
)) {
2021 p
->migration_flags
|= MDF_PUSH
;
2025 p
->migration_flags
&= ~MDF_PUSH
;
2027 if (p
->sched_class
->find_lock_rq
)
2028 lowest_rq
= p
->sched_class
->find_lock_rq(p
, rq
);
2033 // XXX validate p is still the highest prio task
2034 if (task_rq(p
) == rq
) {
2035 deactivate_task(rq
, p
, 0);
2036 set_task_cpu(p
, lowest_rq
->cpu
);
2037 activate_task(lowest_rq
, p
, 0);
2038 resched_curr(lowest_rq
);
2041 double_unlock_balance(rq
, lowest_rq
);
2044 rq
->push_busy
= false;
2045 raw_spin_unlock(&rq
->lock
);
2046 raw_spin_unlock_irq(&p
->pi_lock
);
2053 * sched_class::set_cpus_allowed must do the below, but is not required to
2054 * actually call this function.
2056 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
)
2058 if (flags
& (SCA_MIGRATE_ENABLE
| SCA_MIGRATE_DISABLE
)) {
2059 p
->cpus_ptr
= new_mask
;
2063 cpumask_copy(&p
->cpus_mask
, new_mask
);
2064 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
2068 __do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
)
2070 struct rq
*rq
= task_rq(p
);
2071 bool queued
, running
;
2074 * This here violates the locking rules for affinity, since we're only
2075 * supposed to change these variables while holding both rq->lock and
2078 * HOWEVER, it magically works, because ttwu() is the only code that
2079 * accesses these variables under p->pi_lock and only does so after
2080 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2081 * before finish_task().
2083 * XXX do further audits, this smells like something putrid.
2085 if (flags
& SCA_MIGRATE_DISABLE
)
2086 SCHED_WARN_ON(!p
->on_cpu
);
2088 lockdep_assert_held(&p
->pi_lock
);
2090 queued
= task_on_rq_queued(p
);
2091 running
= task_current(rq
, p
);
2095 * Because __kthread_bind() calls this on blocked tasks without
2098 lockdep_assert_held(&rq
->lock
);
2099 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
2102 put_prev_task(rq
, p
);
2104 p
->sched_class
->set_cpus_allowed(p
, new_mask
, flags
);
2107 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
2109 set_next_task(rq
, p
);
2112 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
2114 __do_set_cpus_allowed(p
, new_mask
, 0);
2118 * This function is wildly self concurrent; here be dragons.
2121 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2122 * designated task is enqueued on an allowed CPU. If that task is currently
2123 * running, we have to kick it out using the CPU stopper.
2125 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2128 * Initial conditions: P0->cpus_mask = [0, 1]
2132 * migrate_disable();
2134 * set_cpus_allowed_ptr(P0, [1]);
2136 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2137 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2138 * This means we need the following scheme:
2142 * migrate_disable();
2144 * set_cpus_allowed_ptr(P0, [1]);
2148 * __set_cpus_allowed_ptr();
2149 * <wakes local stopper>
2150 * `--> <woken on migration completion>
2152 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2153 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2154 * task p are serialized by p->pi_lock, which we can leverage: the one that
2155 * should come into effect at the end of the Migrate-Disable region is the last
2156 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2157 * but we still need to properly signal those waiting tasks at the appropriate
2160 * This is implemented using struct set_affinity_pending. The first
2161 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2162 * setup an instance of that struct and install it on the targeted task_struct.
2163 * Any and all further callers will reuse that instance. Those then wait for
2164 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2165 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2168 * (1) In the cases covered above. There is one more where the completion is
2169 * signaled within affine_move_task() itself: when a subsequent affinity request
2170 * cancels the need for an active migration. Consider:
2172 * Initial conditions: P0->cpus_mask = [0, 1]
2176 * migrate_disable();
2178 * set_cpus_allowed_ptr(P0, [1]);
2180 * set_cpus_allowed_ptr(P0, [0, 1]);
2181 * <signal completion>
2184 * Note that the above is safe vs a concurrent migrate_enable(), as any
2185 * pending affinity completion is preceded by an uninstallation of
2186 * p->migration_pending done with p->pi_lock held.
2188 static int affine_move_task(struct rq
*rq
, struct task_struct
*p
, struct rq_flags
*rf
,
2189 int dest_cpu
, unsigned int flags
)
2191 struct set_affinity_pending my_pending
= { }, *pending
= NULL
;
2192 struct migration_arg arg
= {
2194 .dest_cpu
= dest_cpu
,
2196 bool complete
= false;
2198 /* Can the task run on the task's current CPU? If so, we're done */
2199 if (cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
)) {
2200 struct task_struct
*push_task
= NULL
;
2202 if ((flags
& SCA_MIGRATE_ENABLE
) &&
2203 (p
->migration_flags
& MDF_PUSH
) && !rq
->push_busy
) {
2204 rq
->push_busy
= true;
2205 push_task
= get_task_struct(p
);
2208 pending
= p
->migration_pending
;
2210 refcount_inc(&pending
->refs
);
2211 p
->migration_pending
= NULL
;
2214 task_rq_unlock(rq
, p
, rf
);
2217 stop_one_cpu_nowait(rq
->cpu
, push_cpu_stop
,
2227 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2228 /* serialized by p->pi_lock */
2229 if (!p
->migration_pending
) {
2230 /* Install the request */
2231 refcount_set(&my_pending
.refs
, 1);
2232 init_completion(&my_pending
.done
);
2233 p
->migration_pending
= &my_pending
;
2235 pending
= p
->migration_pending
;
2236 refcount_inc(&pending
->refs
);
2239 pending
= p
->migration_pending
;
2241 * - !MIGRATE_ENABLE:
2242 * we'll have installed a pending if there wasn't one already.
2245 * we're here because the current CPU isn't matching anymore,
2246 * the only way that can happen is because of a concurrent
2247 * set_cpus_allowed_ptr() call, which should then still be
2248 * pending completion.
2250 * Either way, we really should have a @pending here.
2252 if (WARN_ON_ONCE(!pending
)) {
2253 task_rq_unlock(rq
, p
, rf
);
2257 if (flags
& SCA_MIGRATE_ENABLE
) {
2259 refcount_inc(&pending
->refs
); /* pending->{arg,stop_work} */
2260 p
->migration_flags
&= ~MDF_PUSH
;
2261 task_rq_unlock(rq
, p
, rf
);
2263 pending
->arg
= (struct migration_arg
) {
2269 stop_one_cpu_nowait(cpu_of(rq
), migration_cpu_stop
,
2270 &pending
->arg
, &pending
->stop_work
);
2275 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
2277 * Lessen races (and headaches) by delegating
2278 * is_migration_disabled(p) checks to the stopper, which will
2279 * run on the same CPU as said p.
2281 task_rq_unlock(rq
, p
, rf
);
2282 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
2286 if (!is_migration_disabled(p
)) {
2287 if (task_on_rq_queued(p
))
2288 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
2290 p
->migration_pending
= NULL
;
2293 task_rq_unlock(rq
, p
, rf
);
2297 complete_all(&pending
->done
);
2300 wait_for_completion(&pending
->done
);
2302 if (refcount_dec_and_test(&pending
->refs
))
2303 wake_up_var(&pending
->refs
);
2306 * Block the original owner of &pending until all subsequent callers
2307 * have seen the completion and decremented the refcount
2309 wait_var_event(&my_pending
.refs
, !refcount_read(&my_pending
.refs
));
2315 * Change a given task's CPU affinity. Migrate the thread to a
2316 * proper CPU and schedule it away if the CPU it's executing on
2317 * is removed from the allowed bitmask.
2319 * NOTE: the caller must have a valid reference to the task, the
2320 * task must not exit() & deallocate itself prematurely. The
2321 * call is not atomic; no spinlocks may be held.
2323 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
2324 const struct cpumask
*new_mask
,
2327 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
2328 unsigned int dest_cpu
;
2333 rq
= task_rq_lock(p
, &rf
);
2334 update_rq_clock(rq
);
2336 if (p
->flags
& PF_KTHREAD
|| is_migration_disabled(p
)) {
2338 * Kernel threads are allowed on online && !active CPUs.
2340 * Specifically, migration_disabled() tasks must not fail the
2341 * cpumask_any_and_distribute() pick below, esp. so on
2342 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2343 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2345 cpu_valid_mask
= cpu_online_mask
;
2349 * Must re-check here, to close a race against __kthread_bind(),
2350 * sched_setaffinity() is not guaranteed to observe the flag.
2352 if ((flags
& SCA_CHECK
) && (p
->flags
& PF_NO_SETAFFINITY
)) {
2357 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2358 if (cpumask_equal(&p
->cpus_mask
, new_mask
))
2361 if (WARN_ON_ONCE(p
== current
&&
2362 is_migration_disabled(p
) &&
2363 !cpumask_test_cpu(task_cpu(p
), new_mask
))) {
2370 * Picking a ~random cpu helps in cases where we are changing affinity
2371 * for groups of tasks (ie. cpuset), so that load balancing is not
2372 * immediately required to distribute the tasks within their new mask.
2374 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
2375 if (dest_cpu
>= nr_cpu_ids
) {
2380 __do_set_cpus_allowed(p
, new_mask
, flags
);
2382 if (p
->flags
& PF_KTHREAD
) {
2384 * For kernel threads that do indeed end up on online &&
2385 * !active we want to ensure they are strict per-CPU threads.
2387 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
2388 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
2389 p
->nr_cpus_allowed
!= 1);
2392 return affine_move_task(rq
, p
, &rf
, dest_cpu
, flags
);
2395 task_rq_unlock(rq
, p
, &rf
);
2400 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
2402 return __set_cpus_allowed_ptr(p
, new_mask
, 0);
2404 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
2406 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2408 #ifdef CONFIG_SCHED_DEBUG
2410 * We should never call set_task_cpu() on a blocked task,
2411 * ttwu() will sort out the placement.
2413 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2417 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2418 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2419 * time relying on p->on_rq.
2421 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
2422 p
->sched_class
== &fair_sched_class
&&
2423 (p
->on_rq
&& !task_on_rq_migrating(p
)));
2425 #ifdef CONFIG_LOCKDEP
2427 * The caller should hold either p->pi_lock or rq->lock, when changing
2428 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2430 * sched_move_task() holds both and thus holding either pins the cgroup,
2433 * Furthermore, all task_rq users should acquire both locks, see
2436 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2437 lockdep_is_held(&task_rq(p
)->lock
)));
2440 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2442 WARN_ON_ONCE(!cpu_online(new_cpu
));
2444 WARN_ON_ONCE(is_migration_disabled(p
));
2447 trace_sched_migrate_task(p
, new_cpu
);
2449 if (task_cpu(p
) != new_cpu
) {
2450 if (p
->sched_class
->migrate_task_rq
)
2451 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
2452 p
->se
.nr_migrations
++;
2454 perf_event_task_migrate(p
);
2457 __set_task_cpu(p
, new_cpu
);
2460 #ifdef CONFIG_NUMA_BALANCING
2461 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
2463 if (task_on_rq_queued(p
)) {
2464 struct rq
*src_rq
, *dst_rq
;
2465 struct rq_flags srf
, drf
;
2467 src_rq
= task_rq(p
);
2468 dst_rq
= cpu_rq(cpu
);
2470 rq_pin_lock(src_rq
, &srf
);
2471 rq_pin_lock(dst_rq
, &drf
);
2473 deactivate_task(src_rq
, p
, 0);
2474 set_task_cpu(p
, cpu
);
2475 activate_task(dst_rq
, p
, 0);
2476 check_preempt_curr(dst_rq
, p
, 0);
2478 rq_unpin_lock(dst_rq
, &drf
);
2479 rq_unpin_lock(src_rq
, &srf
);
2483 * Task isn't running anymore; make it appear like we migrated
2484 * it before it went to sleep. This means on wakeup we make the
2485 * previous CPU our target instead of where it really is.
2491 struct migration_swap_arg
{
2492 struct task_struct
*src_task
, *dst_task
;
2493 int src_cpu
, dst_cpu
;
2496 static int migrate_swap_stop(void *data
)
2498 struct migration_swap_arg
*arg
= data
;
2499 struct rq
*src_rq
, *dst_rq
;
2502 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
2505 src_rq
= cpu_rq(arg
->src_cpu
);
2506 dst_rq
= cpu_rq(arg
->dst_cpu
);
2508 double_raw_lock(&arg
->src_task
->pi_lock
,
2509 &arg
->dst_task
->pi_lock
);
2510 double_rq_lock(src_rq
, dst_rq
);
2512 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
2515 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
2518 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
2521 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
2524 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
2525 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
2530 double_rq_unlock(src_rq
, dst_rq
);
2531 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
2532 raw_spin_unlock(&arg
->src_task
->pi_lock
);
2538 * Cross migrate two tasks
2540 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
2541 int target_cpu
, int curr_cpu
)
2543 struct migration_swap_arg arg
;
2546 arg
= (struct migration_swap_arg
){
2548 .src_cpu
= curr_cpu
,
2550 .dst_cpu
= target_cpu
,
2553 if (arg
.src_cpu
== arg
.dst_cpu
)
2557 * These three tests are all lockless; this is OK since all of them
2558 * will be re-checked with proper locks held further down the line.
2560 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
2563 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
2566 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
2569 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
2570 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
2575 #endif /* CONFIG_NUMA_BALANCING */
2578 * wait_task_inactive - wait for a thread to unschedule.
2580 * If @match_state is nonzero, it's the @p->state value just checked and
2581 * not expected to change. If it changes, i.e. @p might have woken up,
2582 * then return zero. When we succeed in waiting for @p to be off its CPU,
2583 * we return a positive number (its total switch count). If a second call
2584 * a short while later returns the same number, the caller can be sure that
2585 * @p has remained unscheduled the whole time.
2587 * The caller must ensure that the task *will* unschedule sometime soon,
2588 * else this function might spin for a *long* time. This function can't
2589 * be called with interrupts off, or it may introduce deadlock with
2590 * smp_call_function() if an IPI is sent by the same process we are
2591 * waiting to become inactive.
2593 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2595 int running
, queued
;
2602 * We do the initial early heuristics without holding
2603 * any task-queue locks at all. We'll only try to get
2604 * the runqueue lock when things look like they will
2610 * If the task is actively running on another CPU
2611 * still, just relax and busy-wait without holding
2614 * NOTE! Since we don't hold any locks, it's not
2615 * even sure that "rq" stays as the right runqueue!
2616 * But we don't care, since "task_running()" will
2617 * return false if the runqueue has changed and p
2618 * is actually now running somewhere else!
2620 while (task_running(rq
, p
)) {
2621 if (match_state
&& unlikely(p
->state
!= match_state
))
2627 * Ok, time to look more closely! We need the rq
2628 * lock now, to be *sure*. If we're wrong, we'll
2629 * just go back and repeat.
2631 rq
= task_rq_lock(p
, &rf
);
2632 trace_sched_wait_task(p
);
2633 running
= task_running(rq
, p
);
2634 queued
= task_on_rq_queued(p
);
2636 if (!match_state
|| p
->state
== match_state
)
2637 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2638 task_rq_unlock(rq
, p
, &rf
);
2641 * If it changed from the expected state, bail out now.
2643 if (unlikely(!ncsw
))
2647 * Was it really running after all now that we
2648 * checked with the proper locks actually held?
2650 * Oops. Go back and try again..
2652 if (unlikely(running
)) {
2658 * It's not enough that it's not actively running,
2659 * it must be off the runqueue _entirely_, and not
2662 * So if it was still runnable (but just not actively
2663 * running right now), it's preempted, and we should
2664 * yield - it could be a while.
2666 if (unlikely(queued
)) {
2667 ktime_t to
= NSEC_PER_SEC
/ HZ
;
2669 set_current_state(TASK_UNINTERRUPTIBLE
);
2670 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2675 * Ahh, all good. It wasn't running, and it wasn't
2676 * runnable, which means that it will never become
2677 * running in the future either. We're all done!
2686 * kick_process - kick a running thread to enter/exit the kernel
2687 * @p: the to-be-kicked thread
2689 * Cause a process which is running on another CPU to enter
2690 * kernel-mode, without any delay. (to get signals handled.)
2692 * NOTE: this function doesn't have to take the runqueue lock,
2693 * because all it wants to ensure is that the remote task enters
2694 * the kernel. If the IPI races and the task has been migrated
2695 * to another CPU then no harm is done and the purpose has been
2698 void kick_process(struct task_struct
*p
)
2704 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2705 smp_send_reschedule(cpu
);
2708 EXPORT_SYMBOL_GPL(kick_process
);
2711 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2713 * A few notes on cpu_active vs cpu_online:
2715 * - cpu_active must be a subset of cpu_online
2717 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2718 * see __set_cpus_allowed_ptr(). At this point the newly online
2719 * CPU isn't yet part of the sched domains, and balancing will not
2722 * - on CPU-down we clear cpu_active() to mask the sched domains and
2723 * avoid the load balancer to place new tasks on the to be removed
2724 * CPU. Existing tasks will remain running there and will be taken
2727 * This means that fallback selection must not select !active CPUs.
2728 * And can assume that any active CPU must be online. Conversely
2729 * select_task_rq() below may allow selection of !active CPUs in order
2730 * to satisfy the above rules.
2732 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2734 int nid
= cpu_to_node(cpu
);
2735 const struct cpumask
*nodemask
= NULL
;
2736 enum { cpuset
, possible
, fail
} state
= cpuset
;
2740 * If the node that the CPU is on has been offlined, cpu_to_node()
2741 * will return -1. There is no CPU on the node, and we should
2742 * select the CPU on the other node.
2745 nodemask
= cpumask_of_node(nid
);
2747 /* Look for allowed, online CPU in same node. */
2748 for_each_cpu(dest_cpu
, nodemask
) {
2749 if (!cpu_active(dest_cpu
))
2751 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2757 /* Any allowed, online CPU? */
2758 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2759 if (!is_cpu_allowed(p
, dest_cpu
))
2765 /* No more Mr. Nice Guy. */
2768 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2769 cpuset_cpus_allowed_fallback(p
);
2776 * XXX When called from select_task_rq() we only
2777 * hold p->pi_lock and again violate locking order.
2779 * More yuck to audit.
2781 do_set_cpus_allowed(p
, cpu_possible_mask
);
2792 if (state
!= cpuset
) {
2794 * Don't tell them about moving exiting tasks or
2795 * kernel threads (both mm NULL), since they never
2798 if (p
->mm
&& printk_ratelimit()) {
2799 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2800 task_pid_nr(p
), p
->comm
, cpu
);
2808 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2811 int select_task_rq(struct task_struct
*p
, int cpu
, int wake_flags
)
2813 lockdep_assert_held(&p
->pi_lock
);
2815 if (p
->nr_cpus_allowed
> 1 && !is_migration_disabled(p
))
2816 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, wake_flags
);
2818 cpu
= cpumask_any(p
->cpus_ptr
);
2821 * In order not to call set_task_cpu() on a blocking task we need
2822 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2825 * Since this is common to all placement strategies, this lives here.
2827 * [ this allows ->select_task() to simply return task_cpu(p) and
2828 * not worry about this generic constraint ]
2830 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2831 cpu
= select_fallback_rq(task_cpu(p
), p
);
2836 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2838 static struct lock_class_key stop_pi_lock
;
2839 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2840 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2844 * Make it appear like a SCHED_FIFO task, its something
2845 * userspace knows about and won't get confused about.
2847 * Also, it will make PI more or less work without too
2848 * much confusion -- but then, stop work should not
2849 * rely on PI working anyway.
2851 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2853 stop
->sched_class
= &stop_sched_class
;
2856 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
2857 * adjust the effective priority of a task. As a result,
2858 * rt_mutex_setprio() can trigger (RT) balancing operations,
2859 * which can then trigger wakeups of the stop thread to push
2860 * around the current task.
2862 * The stop task itself will never be part of the PI-chain, it
2863 * never blocks, therefore that ->pi_lock recursion is safe.
2864 * Tell lockdep about this by placing the stop->pi_lock in its
2867 lockdep_set_class(&stop
->pi_lock
, &stop_pi_lock
);
2870 cpu_rq(cpu
)->stop
= stop
;
2874 * Reset it back to a normal scheduling class so that
2875 * it can die in pieces.
2877 old_stop
->sched_class
= &rt_sched_class
;
2881 #else /* CONFIG_SMP */
2883 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2884 const struct cpumask
*new_mask
,
2887 return set_cpus_allowed_ptr(p
, new_mask
);
2890 static inline void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
) { }
2892 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
2897 #endif /* !CONFIG_SMP */
2900 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2904 if (!schedstat_enabled())
2910 if (cpu
== rq
->cpu
) {
2911 __schedstat_inc(rq
->ttwu_local
);
2912 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2914 struct sched_domain
*sd
;
2916 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2918 for_each_domain(rq
->cpu
, sd
) {
2919 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2920 __schedstat_inc(sd
->ttwu_wake_remote
);
2927 if (wake_flags
& WF_MIGRATED
)
2928 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2929 #endif /* CONFIG_SMP */
2931 __schedstat_inc(rq
->ttwu_count
);
2932 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2934 if (wake_flags
& WF_SYNC
)
2935 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2939 * Mark the task runnable and perform wakeup-preemption.
2941 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2942 struct rq_flags
*rf
)
2944 check_preempt_curr(rq
, p
, wake_flags
);
2945 p
->state
= TASK_RUNNING
;
2946 trace_sched_wakeup(p
);
2949 if (p
->sched_class
->task_woken
) {
2951 * Our task @p is fully woken up and running; so it's safe to
2952 * drop the rq->lock, hereafter rq is only used for statistics.
2954 rq_unpin_lock(rq
, rf
);
2955 p
->sched_class
->task_woken(rq
, p
);
2956 rq_repin_lock(rq
, rf
);
2959 if (rq
->idle_stamp
) {
2960 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2961 u64 max
= 2*rq
->max_idle_balance_cost
;
2963 update_avg(&rq
->avg_idle
, delta
);
2965 if (rq
->avg_idle
> max
)
2974 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2975 struct rq_flags
*rf
)
2977 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2979 lockdep_assert_held(&rq
->lock
);
2981 if (p
->sched_contributes_to_load
)
2982 rq
->nr_uninterruptible
--;
2985 if (wake_flags
& WF_MIGRATED
)
2986 en_flags
|= ENQUEUE_MIGRATED
;
2989 activate_task(rq
, p
, en_flags
);
2990 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2994 * Consider @p being inside a wait loop:
2997 * set_current_state(TASK_UNINTERRUPTIBLE);
3004 * __set_current_state(TASK_RUNNING);
3006 * between set_current_state() and schedule(). In this case @p is still
3007 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3010 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3011 * then schedule() must still happen and p->state can be changed to
3012 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3013 * need to do a full wakeup with enqueue.
3015 * Returns: %true when the wakeup is done,
3018 static int ttwu_runnable(struct task_struct
*p
, int wake_flags
)
3024 rq
= __task_rq_lock(p
, &rf
);
3025 if (task_on_rq_queued(p
)) {
3026 /* check_preempt_curr() may use rq clock */
3027 update_rq_clock(rq
);
3028 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
3031 __task_rq_unlock(rq
, &rf
);
3037 void sched_ttwu_pending(void *arg
)
3039 struct llist_node
*llist
= arg
;
3040 struct rq
*rq
= this_rq();
3041 struct task_struct
*p
, *t
;
3048 * rq::ttwu_pending racy indication of out-standing wakeups.
3049 * Races such that false-negatives are possible, since they
3050 * are shorter lived that false-positives would be.
3052 WRITE_ONCE(rq
->ttwu_pending
, 0);
3054 rq_lock_irqsave(rq
, &rf
);
3055 update_rq_clock(rq
);
3057 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
.llist
) {
3058 if (WARN_ON_ONCE(p
->on_cpu
))
3059 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3061 if (WARN_ON_ONCE(task_cpu(p
) != cpu_of(rq
)))
3062 set_task_cpu(p
, cpu_of(rq
));
3064 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
3067 rq_unlock_irqrestore(rq
, &rf
);
3070 void send_call_function_single_ipi(int cpu
)
3072 struct rq
*rq
= cpu_rq(cpu
);
3074 if (!set_nr_if_polling(rq
->idle
))
3075 arch_send_call_function_single_ipi(cpu
);
3077 trace_sched_wake_idle_without_ipi(cpu
);
3081 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3082 * necessary. The wakee CPU on receipt of the IPI will queue the task
3083 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3084 * of the wakeup instead of the waker.
3086 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3088 struct rq
*rq
= cpu_rq(cpu
);
3090 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
3092 WRITE_ONCE(rq
->ttwu_pending
, 1);
3093 __smp_call_single_queue(cpu
, &p
->wake_entry
.llist
);
3096 void wake_up_if_idle(int cpu
)
3098 struct rq
*rq
= cpu_rq(cpu
);
3103 if (!is_idle_task(rcu_dereference(rq
->curr
)))
3106 if (set_nr_if_polling(rq
->idle
)) {
3107 trace_sched_wake_idle_without_ipi(cpu
);
3109 rq_lock_irqsave(rq
, &rf
);
3110 if (is_idle_task(rq
->curr
))
3111 smp_send_reschedule(cpu
);
3112 /* Else CPU is not idle, do nothing here: */
3113 rq_unlock_irqrestore(rq
, &rf
);
3120 bool cpus_share_cache(int this_cpu
, int that_cpu
)
3122 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
3125 static inline bool ttwu_queue_cond(int cpu
, int wake_flags
)
3128 * If the CPU does not share cache, then queue the task on the
3129 * remote rqs wakelist to avoid accessing remote data.
3131 if (!cpus_share_cache(smp_processor_id(), cpu
))
3135 * If the task is descheduling and the only running task on the
3136 * CPU then use the wakelist to offload the task activation to
3137 * the soon-to-be-idle CPU as the current CPU is likely busy.
3138 * nr_running is checked to avoid unnecessary task stacking.
3140 if ((wake_flags
& WF_ON_CPU
) && cpu_rq(cpu
)->nr_running
<= 1)
3146 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3148 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(cpu
, wake_flags
)) {
3149 if (WARN_ON_ONCE(cpu
== smp_processor_id()))
3152 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
3153 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
3160 #else /* !CONFIG_SMP */
3162 static inline bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3167 #endif /* CONFIG_SMP */
3169 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
3171 struct rq
*rq
= cpu_rq(cpu
);
3174 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
3178 update_rq_clock(rq
);
3179 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
3184 * Notes on Program-Order guarantees on SMP systems.
3188 * The basic program-order guarantee on SMP systems is that when a task [t]
3189 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3190 * execution on its new CPU [c1].
3192 * For migration (of runnable tasks) this is provided by the following means:
3194 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3195 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3196 * rq(c1)->lock (if not at the same time, then in that order).
3197 * C) LOCK of the rq(c1)->lock scheduling in task
3199 * Release/acquire chaining guarantees that B happens after A and C after B.
3200 * Note: the CPU doing B need not be c0 or c1
3209 * UNLOCK rq(0)->lock
3211 * LOCK rq(0)->lock // orders against CPU0
3213 * UNLOCK rq(0)->lock
3217 * UNLOCK rq(1)->lock
3219 * LOCK rq(1)->lock // orders against CPU2
3222 * UNLOCK rq(1)->lock
3225 * BLOCKING -- aka. SLEEP + WAKEUP
3227 * For blocking we (obviously) need to provide the same guarantee as for
3228 * migration. However the means are completely different as there is no lock
3229 * chain to provide order. Instead we do:
3231 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3232 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3236 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3238 * LOCK rq(0)->lock LOCK X->pi_lock
3241 * smp_store_release(X->on_cpu, 0);
3243 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3249 * X->state = RUNNING
3250 * UNLOCK rq(2)->lock
3252 * LOCK rq(2)->lock // orders against CPU1
3255 * UNLOCK rq(2)->lock
3258 * UNLOCK rq(0)->lock
3261 * However, for wakeups there is a second guarantee we must provide, namely we
3262 * must ensure that CONDITION=1 done by the caller can not be reordered with
3263 * accesses to the task state; see try_to_wake_up() and set_current_state().
3267 * try_to_wake_up - wake up a thread
3268 * @p: the thread to be awakened
3269 * @state: the mask of task states that can be woken
3270 * @wake_flags: wake modifier flags (WF_*)
3272 * Conceptually does:
3274 * If (@state & @p->state) @p->state = TASK_RUNNING.
3276 * If the task was not queued/runnable, also place it back on a runqueue.
3278 * This function is atomic against schedule() which would dequeue the task.
3280 * It issues a full memory barrier before accessing @p->state, see the comment
3281 * with set_current_state().
3283 * Uses p->pi_lock to serialize against concurrent wake-ups.
3285 * Relies on p->pi_lock stabilizing:
3288 * - p->sched_task_group
3289 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3291 * Tries really hard to only take one task_rq(p)->lock for performance.
3292 * Takes rq->lock in:
3293 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3294 * - ttwu_queue() -- new rq, for enqueue of the task;
3295 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3297 * As a consequence we race really badly with just about everything. See the
3298 * many memory barriers and their comments for details.
3300 * Return: %true if @p->state changes (an actual wakeup was done),
3304 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
3306 unsigned long flags
;
3307 int cpu
, success
= 0;
3312 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3313 * == smp_processor_id()'. Together this means we can special
3314 * case the whole 'p->on_rq && ttwu_runnable()' case below
3315 * without taking any locks.
3318 * - we rely on Program-Order guarantees for all the ordering,
3319 * - we're serialized against set_special_state() by virtue of
3320 * it disabling IRQs (this allows not taking ->pi_lock).
3322 if (!(p
->state
& state
))
3326 trace_sched_waking(p
);
3327 p
->state
= TASK_RUNNING
;
3328 trace_sched_wakeup(p
);
3333 * If we are going to wake up a thread waiting for CONDITION we
3334 * need to ensure that CONDITION=1 done by the caller can not be
3335 * reordered with p->state check below. This pairs with smp_store_mb()
3336 * in set_current_state() that the waiting thread does.
3338 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3339 smp_mb__after_spinlock();
3340 if (!(p
->state
& state
))
3343 trace_sched_waking(p
);
3345 /* We're going to change ->state: */
3349 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3350 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3351 * in smp_cond_load_acquire() below.
3353 * sched_ttwu_pending() try_to_wake_up()
3354 * STORE p->on_rq = 1 LOAD p->state
3357 * __schedule() (switch to task 'p')
3358 * LOCK rq->lock smp_rmb();
3359 * smp_mb__after_spinlock();
3363 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3365 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3366 * __schedule(). See the comment for smp_mb__after_spinlock().
3368 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3371 if (READ_ONCE(p
->on_rq
) && ttwu_runnable(p
, wake_flags
))
3375 delayacct_blkio_end(p
);
3376 atomic_dec(&task_rq(p
)->nr_iowait
);
3381 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3382 * possible to, falsely, observe p->on_cpu == 0.
3384 * One must be running (->on_cpu == 1) in order to remove oneself
3385 * from the runqueue.
3387 * __schedule() (switch to task 'p') try_to_wake_up()
3388 * STORE p->on_cpu = 1 LOAD p->on_rq
3391 * __schedule() (put 'p' to sleep)
3392 * LOCK rq->lock smp_rmb();
3393 * smp_mb__after_spinlock();
3394 * STORE p->on_rq = 0 LOAD p->on_cpu
3396 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3397 * __schedule(). See the comment for smp_mb__after_spinlock().
3399 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3400 * schedule()'s deactivate_task() has 'happened' and p will no longer
3401 * care about it's own p->state. See the comment in __schedule().
3403 smp_acquire__after_ctrl_dep();
3406 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3407 * == 0), which means we need to do an enqueue, change p->state to
3408 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3409 * enqueue, such as ttwu_queue_wakelist().
3411 p
->state
= TASK_WAKING
;
3414 * If the owning (remote) CPU is still in the middle of schedule() with
3415 * this task as prev, considering queueing p on the remote CPUs wake_list
3416 * which potentially sends an IPI instead of spinning on p->on_cpu to
3417 * let the waker make forward progress. This is safe because IRQs are
3418 * disabled and the IPI will deliver after on_cpu is cleared.
3420 * Ensure we load task_cpu(p) after p->on_cpu:
3422 * set_task_cpu(p, cpu);
3423 * STORE p->cpu = @cpu
3424 * __schedule() (switch to task 'p')
3426 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3427 * STORE p->on_cpu = 1 LOAD p->cpu
3429 * to ensure we observe the correct CPU on which the task is currently
3432 if (smp_load_acquire(&p
->on_cpu
) &&
3433 ttwu_queue_wakelist(p
, task_cpu(p
), wake_flags
| WF_ON_CPU
))
3437 * If the owning (remote) CPU is still in the middle of schedule() with
3438 * this task as prev, wait until it's done referencing the task.
3440 * Pairs with the smp_store_release() in finish_task().
3442 * This ensures that tasks getting woken will be fully ordered against
3443 * their previous state and preserve Program Order.
3445 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3447 cpu
= select_task_rq(p
, p
->wake_cpu
, wake_flags
| WF_TTWU
);
3448 if (task_cpu(p
) != cpu
) {
3449 wake_flags
|= WF_MIGRATED
;
3450 psi_ttwu_dequeue(p
);
3451 set_task_cpu(p
, cpu
);
3455 #endif /* CONFIG_SMP */
3457 ttwu_queue(p
, cpu
, wake_flags
);
3459 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3462 ttwu_stat(p
, task_cpu(p
), wake_flags
);
3469 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3470 * @p: Process for which the function is to be invoked.
3471 * @func: Function to invoke.
3472 * @arg: Argument to function.
3474 * If the specified task can be quickly locked into a definite state
3475 * (either sleeping or on a given runqueue), arrange to keep it in that
3476 * state while invoking @func(@arg). This function can use ->on_rq and
3477 * task_curr() to work out what the state is, if required. Given that
3478 * @func can be invoked with a runqueue lock held, it had better be quite
3482 * @false if the task slipped out from under the locks.
3483 * @true if the task was locked onto a runqueue or is sleeping.
3484 * However, @func can override this by returning @false.
3486 bool try_invoke_on_locked_down_task(struct task_struct
*p
, bool (*func
)(struct task_struct
*t
, void *arg
), void *arg
)
3492 lockdep_assert_irqs_enabled();
3493 raw_spin_lock_irq(&p
->pi_lock
);
3495 rq
= __task_rq_lock(p
, &rf
);
3496 if (task_rq(p
) == rq
)
3505 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3510 raw_spin_unlock_irq(&p
->pi_lock
);
3515 * wake_up_process - Wake up a specific process
3516 * @p: The process to be woken up.
3518 * Attempt to wake up the nominated process and move it to the set of runnable
3521 * Return: 1 if the process was woken up, 0 if it was already running.
3523 * This function executes a full memory barrier before accessing the task state.
3525 int wake_up_process(struct task_struct
*p
)
3527 return try_to_wake_up(p
, TASK_NORMAL
, 0);
3529 EXPORT_SYMBOL(wake_up_process
);
3531 int wake_up_state(struct task_struct
*p
, unsigned int state
)
3533 return try_to_wake_up(p
, state
, 0);
3537 * Perform scheduler related setup for a newly forked process p.
3538 * p is forked by current.
3540 * __sched_fork() is basic setup used by init_idle() too:
3542 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
3547 p
->se
.exec_start
= 0;
3548 p
->se
.sum_exec_runtime
= 0;
3549 p
->se
.prev_sum_exec_runtime
= 0;
3550 p
->se
.nr_migrations
= 0;
3552 INIT_LIST_HEAD(&p
->se
.group_node
);
3554 #ifdef CONFIG_FAIR_GROUP_SCHED
3555 p
->se
.cfs_rq
= NULL
;
3558 #ifdef CONFIG_SCHEDSTATS
3559 /* Even if schedstat is disabled, there should not be garbage */
3560 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
3563 RB_CLEAR_NODE(&p
->dl
.rb_node
);
3564 init_dl_task_timer(&p
->dl
);
3565 init_dl_inactive_task_timer(&p
->dl
);
3566 __dl_clear_params(p
);
3568 INIT_LIST_HEAD(&p
->rt
.run_list
);
3570 p
->rt
.time_slice
= sched_rr_timeslice
;
3574 #ifdef CONFIG_PREEMPT_NOTIFIERS
3575 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
3578 #ifdef CONFIG_COMPACTION
3579 p
->capture_control
= NULL
;
3581 init_numa_balancing(clone_flags
, p
);
3583 p
->wake_entry
.u_flags
= CSD_TYPE_TTWU
;
3584 p
->migration_pending
= NULL
;
3588 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
3590 #ifdef CONFIG_NUMA_BALANCING
3592 void set_numabalancing_state(bool enabled
)
3595 static_branch_enable(&sched_numa_balancing
);
3597 static_branch_disable(&sched_numa_balancing
);
3600 #ifdef CONFIG_PROC_SYSCTL
3601 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
3602 void *buffer
, size_t *lenp
, loff_t
*ppos
)
3606 int state
= static_branch_likely(&sched_numa_balancing
);
3608 if (write
&& !capable(CAP_SYS_ADMIN
))
3613 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
3617 set_numabalancing_state(state
);
3623 #ifdef CONFIG_SCHEDSTATS
3625 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
3626 static bool __initdata __sched_schedstats
= false;
3628 static void set_schedstats(bool enabled
)
3631 static_branch_enable(&sched_schedstats
);
3633 static_branch_disable(&sched_schedstats
);
3636 void force_schedstat_enabled(void)
3638 if (!schedstat_enabled()) {
3639 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3640 static_branch_enable(&sched_schedstats
);
3644 static int __init
setup_schedstats(char *str
)
3651 * This code is called before jump labels have been set up, so we can't
3652 * change the static branch directly just yet. Instead set a temporary
3653 * variable so init_schedstats() can do it later.
3655 if (!strcmp(str
, "enable")) {
3656 __sched_schedstats
= true;
3658 } else if (!strcmp(str
, "disable")) {
3659 __sched_schedstats
= false;
3664 pr_warn("Unable to parse schedstats=\n");
3668 __setup("schedstats=", setup_schedstats
);
3670 static void __init
init_schedstats(void)
3672 set_schedstats(__sched_schedstats
);
3675 #ifdef CONFIG_PROC_SYSCTL
3676 int sysctl_schedstats(struct ctl_table
*table
, int write
, void *buffer
,
3677 size_t *lenp
, loff_t
*ppos
)
3681 int state
= static_branch_likely(&sched_schedstats
);
3683 if (write
&& !capable(CAP_SYS_ADMIN
))
3688 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
3692 set_schedstats(state
);
3695 #endif /* CONFIG_PROC_SYSCTL */
3696 #else /* !CONFIG_SCHEDSTATS */
3697 static inline void init_schedstats(void) {}
3698 #endif /* CONFIG_SCHEDSTATS */
3701 * fork()/clone()-time setup:
3703 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
3705 unsigned long flags
;
3707 __sched_fork(clone_flags
, p
);
3709 * We mark the process as NEW here. This guarantees that
3710 * nobody will actually run it, and a signal or other external
3711 * event cannot wake it up and insert it on the runqueue either.
3713 p
->state
= TASK_NEW
;
3716 * Make sure we do not leak PI boosting priority to the child.
3718 p
->prio
= current
->normal_prio
;
3723 * Revert to default priority/policy on fork if requested.
3725 if (unlikely(p
->sched_reset_on_fork
)) {
3726 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3727 p
->policy
= SCHED_NORMAL
;
3728 p
->static_prio
= NICE_TO_PRIO(0);
3730 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
3731 p
->static_prio
= NICE_TO_PRIO(0);
3733 p
->prio
= p
->normal_prio
= __normal_prio(p
);
3734 set_load_weight(p
, false);
3737 * We don't need the reset flag anymore after the fork. It has
3738 * fulfilled its duty:
3740 p
->sched_reset_on_fork
= 0;
3743 if (dl_prio(p
->prio
))
3745 else if (rt_prio(p
->prio
))
3746 p
->sched_class
= &rt_sched_class
;
3748 p
->sched_class
= &fair_sched_class
;
3750 init_entity_runnable_average(&p
->se
);
3753 * The child is not yet in the pid-hash so no cgroup attach races,
3754 * and the cgroup is pinned to this child due to cgroup_fork()
3755 * is ran before sched_fork().
3757 * Silence PROVE_RCU.
3759 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3762 * We're setting the CPU for the first time, we don't migrate,
3763 * so use __set_task_cpu().
3765 __set_task_cpu(p
, smp_processor_id());
3766 if (p
->sched_class
->task_fork
)
3767 p
->sched_class
->task_fork(p
);
3768 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3770 #ifdef CONFIG_SCHED_INFO
3771 if (likely(sched_info_on()))
3772 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
3774 #if defined(CONFIG_SMP)
3777 init_task_preempt_count(p
);
3779 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
3780 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
3785 void sched_post_fork(struct task_struct
*p
)
3787 uclamp_post_fork(p
);
3790 unsigned long to_ratio(u64 period
, u64 runtime
)
3792 if (runtime
== RUNTIME_INF
)
3796 * Doing this here saves a lot of checks in all
3797 * the calling paths, and returning zero seems
3798 * safe for them anyway.
3803 return div64_u64(runtime
<< BW_SHIFT
, period
);
3807 * wake_up_new_task - wake up a newly created task for the first time.
3809 * This function will do some initial scheduler statistics housekeeping
3810 * that must be done for every newly created context, then puts the task
3811 * on the runqueue and wakes it.
3813 void wake_up_new_task(struct task_struct
*p
)
3818 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
3819 p
->state
= TASK_RUNNING
;
3822 * Fork balancing, do it here and not earlier because:
3823 * - cpus_ptr can change in the fork path
3824 * - any previously selected CPU might disappear through hotplug
3826 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3827 * as we're not fully set-up yet.
3829 p
->recent_used_cpu
= task_cpu(p
);
3831 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), WF_FORK
));
3833 rq
= __task_rq_lock(p
, &rf
);
3834 update_rq_clock(rq
);
3835 post_init_entity_util_avg(p
);
3837 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
3838 trace_sched_wakeup_new(p
);
3839 check_preempt_curr(rq
, p
, WF_FORK
);
3841 if (p
->sched_class
->task_woken
) {
3843 * Nothing relies on rq->lock after this, so it's fine to
3846 rq_unpin_lock(rq
, &rf
);
3847 p
->sched_class
->task_woken(rq
, p
);
3848 rq_repin_lock(rq
, &rf
);
3851 task_rq_unlock(rq
, p
, &rf
);
3854 #ifdef CONFIG_PREEMPT_NOTIFIERS
3856 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
3858 void preempt_notifier_inc(void)
3860 static_branch_inc(&preempt_notifier_key
);
3862 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
3864 void preempt_notifier_dec(void)
3866 static_branch_dec(&preempt_notifier_key
);
3868 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3871 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3872 * @notifier: notifier struct to register
3874 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3876 if (!static_branch_unlikely(&preempt_notifier_key
))
3877 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3879 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3881 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3884 * preempt_notifier_unregister - no longer interested in preemption notifications
3885 * @notifier: notifier struct to unregister
3887 * This is *not* safe to call from within a preemption notifier.
3889 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3891 hlist_del(¬ifier
->link
);
3893 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3895 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3897 struct preempt_notifier
*notifier
;
3899 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3900 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3903 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3905 if (static_branch_unlikely(&preempt_notifier_key
))
3906 __fire_sched_in_preempt_notifiers(curr
);
3910 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3911 struct task_struct
*next
)
3913 struct preempt_notifier
*notifier
;
3915 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3916 notifier
->ops
->sched_out(notifier
, next
);
3919 static __always_inline
void
3920 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3921 struct task_struct
*next
)
3923 if (static_branch_unlikely(&preempt_notifier_key
))
3924 __fire_sched_out_preempt_notifiers(curr
, next
);
3927 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3929 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3934 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3935 struct task_struct
*next
)
3939 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3941 static inline void prepare_task(struct task_struct
*next
)
3945 * Claim the task as running, we do this before switching to it
3946 * such that any running task will have this set.
3948 * See the ttwu() WF_ON_CPU case and its ordering comment.
3950 WRITE_ONCE(next
->on_cpu
, 1);
3954 static inline void finish_task(struct task_struct
*prev
)
3958 * This must be the very last reference to @prev from this CPU. After
3959 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3960 * must ensure this doesn't happen until the switch is completely
3963 * In particular, the load of prev->state in finish_task_switch() must
3964 * happen before this.
3966 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3968 smp_store_release(&prev
->on_cpu
, 0);
3974 static void do_balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
3976 void (*func
)(struct rq
*rq
);
3977 struct callback_head
*next
;
3979 lockdep_assert_held(&rq
->lock
);
3982 func
= (void (*)(struct rq
*))head
->func
;
3991 static inline struct callback_head
*splice_balance_callbacks(struct rq
*rq
)
3993 struct callback_head
*head
= rq
->balance_callback
;
3995 lockdep_assert_held(&rq
->lock
);
3997 rq
->balance_callback
= NULL
;
3998 rq
->balance_flags
&= ~BALANCE_WORK
;
4004 static void __balance_callbacks(struct rq
*rq
)
4006 do_balance_callbacks(rq
, splice_balance_callbacks(rq
));
4009 static inline void balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
4011 unsigned long flags
;
4013 if (unlikely(head
)) {
4014 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4015 do_balance_callbacks(rq
, head
);
4016 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4020 static void balance_push(struct rq
*rq
);
4022 static inline void balance_switch(struct rq
*rq
)
4024 if (likely(!rq
->balance_flags
))
4027 if (rq
->balance_flags
& BALANCE_PUSH
) {
4032 __balance_callbacks(rq
);
4037 static inline void __balance_callbacks(struct rq
*rq
)
4041 static inline struct callback_head
*splice_balance_callbacks(struct rq
*rq
)
4046 static inline void balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
4050 static inline void balance_switch(struct rq
*rq
)
4057 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
4060 * Since the runqueue lock will be released by the next
4061 * task (which is an invalid locking op but in the case
4062 * of the scheduler it's an obvious special-case), so we
4063 * do an early lockdep release here:
4065 rq_unpin_lock(rq
, rf
);
4066 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
4067 #ifdef CONFIG_DEBUG_SPINLOCK
4068 /* this is a valid case when another task releases the spinlock */
4069 rq
->lock
.owner
= next
;
4073 static inline void finish_lock_switch(struct rq
*rq
)
4076 * If we are tracking spinlock dependencies then we have to
4077 * fix up the runqueue lock - which gets 'carried over' from
4078 * prev into current:
4080 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
4082 raw_spin_unlock_irq(&rq
->lock
);
4086 * NOP if the arch has not defined these:
4089 #ifndef prepare_arch_switch
4090 # define prepare_arch_switch(next) do { } while (0)
4093 #ifndef finish_arch_post_lock_switch
4094 # define finish_arch_post_lock_switch() do { } while (0)
4098 * prepare_task_switch - prepare to switch tasks
4099 * @rq: the runqueue preparing to switch
4100 * @prev: the current task that is being switched out
4101 * @next: the task we are going to switch to.
4103 * This is called with the rq lock held and interrupts off. It must
4104 * be paired with a subsequent finish_task_switch after the context
4107 * prepare_task_switch sets up locking and calls architecture specific
4111 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
4112 struct task_struct
*next
)
4114 kcov_prepare_switch(prev
);
4115 sched_info_switch(rq
, prev
, next
);
4116 perf_event_task_sched_out(prev
, next
);
4118 fire_sched_out_preempt_notifiers(prev
, next
);
4120 prepare_arch_switch(next
);
4124 * finish_task_switch - clean up after a task-switch
4125 * @prev: the thread we just switched away from.
4127 * finish_task_switch must be called after the context switch, paired
4128 * with a prepare_task_switch call before the context switch.
4129 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4130 * and do any other architecture-specific cleanup actions.
4132 * Note that we may have delayed dropping an mm in context_switch(). If
4133 * so, we finish that here outside of the runqueue lock. (Doing it
4134 * with the lock held can cause deadlocks; see schedule() for
4137 * The context switch have flipped the stack from under us and restored the
4138 * local variables which were saved when this task called schedule() in the
4139 * past. prev == current is still correct but we need to recalculate this_rq
4140 * because prev may have moved to another CPU.
4142 static struct rq
*finish_task_switch(struct task_struct
*prev
)
4143 __releases(rq
->lock
)
4145 struct rq
*rq
= this_rq();
4146 struct mm_struct
*mm
= rq
->prev_mm
;
4150 * The previous task will have left us with a preempt_count of 2
4151 * because it left us after:
4154 * preempt_disable(); // 1
4156 * raw_spin_lock_irq(&rq->lock) // 2
4158 * Also, see FORK_PREEMPT_COUNT.
4160 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
4161 "corrupted preempt_count: %s/%d/0x%x\n",
4162 current
->comm
, current
->pid
, preempt_count()))
4163 preempt_count_set(FORK_PREEMPT_COUNT
);
4168 * A task struct has one reference for the use as "current".
4169 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4170 * schedule one last time. The schedule call will never return, and
4171 * the scheduled task must drop that reference.
4173 * We must observe prev->state before clearing prev->on_cpu (in
4174 * finish_task), otherwise a concurrent wakeup can get prev
4175 * running on another CPU and we could rave with its RUNNING -> DEAD
4176 * transition, resulting in a double drop.
4178 prev_state
= prev
->state
;
4179 vtime_task_switch(prev
);
4180 perf_event_task_sched_in(prev
, current
);
4182 finish_lock_switch(rq
);
4183 finish_arch_post_lock_switch();
4184 kcov_finish_switch(current
);
4186 fire_sched_in_preempt_notifiers(current
);
4188 * When switching through a kernel thread, the loop in
4189 * membarrier_{private,global}_expedited() may have observed that
4190 * kernel thread and not issued an IPI. It is therefore possible to
4191 * schedule between user->kernel->user threads without passing though
4192 * switch_mm(). Membarrier requires a barrier after storing to
4193 * rq->curr, before returning to userspace, so provide them here:
4195 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4196 * provided by mmdrop(),
4197 * - a sync_core for SYNC_CORE.
4200 membarrier_mm_sync_core_before_usermode(mm
);
4203 if (unlikely(prev_state
== TASK_DEAD
)) {
4204 if (prev
->sched_class
->task_dead
)
4205 prev
->sched_class
->task_dead(prev
);
4208 * Remove function-return probe instances associated with this
4209 * task and put them back on the free list.
4211 kprobe_flush_task(prev
);
4213 /* Task is done with its stack. */
4214 put_task_stack(prev
);
4216 put_task_struct_rcu_user(prev
);
4219 tick_nohz_task_switch();
4224 * schedule_tail - first thing a freshly forked thread must call.
4225 * @prev: the thread we just switched away from.
4227 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
4228 __releases(rq
->lock
)
4233 * New tasks start with FORK_PREEMPT_COUNT, see there and
4234 * finish_task_switch() for details.
4236 * finish_task_switch() will drop rq->lock() and lower preempt_count
4237 * and the preempt_enable() will end up enabling preemption (on
4238 * PREEMPT_COUNT kernels).
4241 rq
= finish_task_switch(prev
);
4244 if (current
->set_child_tid
)
4245 put_user(task_pid_vnr(current
), current
->set_child_tid
);
4247 calculate_sigpending();
4251 * context_switch - switch to the new MM and the new thread's register state.
4253 static __always_inline
struct rq
*
4254 context_switch(struct rq
*rq
, struct task_struct
*prev
,
4255 struct task_struct
*next
, struct rq_flags
*rf
)
4257 prepare_task_switch(rq
, prev
, next
);
4260 * For paravirt, this is coupled with an exit in switch_to to
4261 * combine the page table reload and the switch backend into
4264 arch_start_context_switch(prev
);
4267 * kernel -> kernel lazy + transfer active
4268 * user -> kernel lazy + mmgrab() active
4270 * kernel -> user switch + mmdrop() active
4271 * user -> user switch
4273 if (!next
->mm
) { // to kernel
4274 enter_lazy_tlb(prev
->active_mm
, next
);
4276 next
->active_mm
= prev
->active_mm
;
4277 if (prev
->mm
) // from user
4278 mmgrab(prev
->active_mm
);
4280 prev
->active_mm
= NULL
;
4282 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
4284 * sys_membarrier() requires an smp_mb() between setting
4285 * rq->curr / membarrier_switch_mm() and returning to userspace.
4287 * The below provides this either through switch_mm(), or in
4288 * case 'prev->active_mm == next->mm' through
4289 * finish_task_switch()'s mmdrop().
4291 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
4293 if (!prev
->mm
) { // from kernel
4294 /* will mmdrop() in finish_task_switch(). */
4295 rq
->prev_mm
= prev
->active_mm
;
4296 prev
->active_mm
= NULL
;
4300 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4302 prepare_lock_switch(rq
, next
, rf
);
4304 /* Here we just switch the register state and the stack. */
4305 switch_to(prev
, next
, prev
);
4308 return finish_task_switch(prev
);
4312 * nr_running and nr_context_switches:
4314 * externally visible scheduler statistics: current number of runnable
4315 * threads, total number of context switches performed since bootup.
4317 unsigned long nr_running(void)
4319 unsigned long i
, sum
= 0;
4321 for_each_online_cpu(i
)
4322 sum
+= cpu_rq(i
)->nr_running
;
4328 * Check if only the current task is running on the CPU.
4330 * Caution: this function does not check that the caller has disabled
4331 * preemption, thus the result might have a time-of-check-to-time-of-use
4332 * race. The caller is responsible to use it correctly, for example:
4334 * - from a non-preemptible section (of course)
4336 * - from a thread that is bound to a single CPU
4338 * - in a loop with very short iterations (e.g. a polling loop)
4340 bool single_task_running(void)
4342 return raw_rq()->nr_running
== 1;
4344 EXPORT_SYMBOL(single_task_running
);
4346 unsigned long long nr_context_switches(void)
4349 unsigned long long sum
= 0;
4351 for_each_possible_cpu(i
)
4352 sum
+= cpu_rq(i
)->nr_switches
;
4358 * Consumers of these two interfaces, like for example the cpuidle menu
4359 * governor, are using nonsensical data. Preferring shallow idle state selection
4360 * for a CPU that has IO-wait which might not even end up running the task when
4361 * it does become runnable.
4364 unsigned long nr_iowait_cpu(int cpu
)
4366 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
4370 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4372 * The idea behind IO-wait account is to account the idle time that we could
4373 * have spend running if it were not for IO. That is, if we were to improve the
4374 * storage performance, we'd have a proportional reduction in IO-wait time.
4376 * This all works nicely on UP, where, when a task blocks on IO, we account
4377 * idle time as IO-wait, because if the storage were faster, it could've been
4378 * running and we'd not be idle.
4380 * This has been extended to SMP, by doing the same for each CPU. This however
4383 * Imagine for instance the case where two tasks block on one CPU, only the one
4384 * CPU will have IO-wait accounted, while the other has regular idle. Even
4385 * though, if the storage were faster, both could've ran at the same time,
4386 * utilising both CPUs.
4388 * This means, that when looking globally, the current IO-wait accounting on
4389 * SMP is a lower bound, by reason of under accounting.
4391 * Worse, since the numbers are provided per CPU, they are sometimes
4392 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4393 * associated with any one particular CPU, it can wake to another CPU than it
4394 * blocked on. This means the per CPU IO-wait number is meaningless.
4396 * Task CPU affinities can make all that even more 'interesting'.
4399 unsigned long nr_iowait(void)
4401 unsigned long i
, sum
= 0;
4403 for_each_possible_cpu(i
)
4404 sum
+= nr_iowait_cpu(i
);
4412 * sched_exec - execve() is a valuable balancing opportunity, because at
4413 * this point the task has the smallest effective memory and cache footprint.
4415 void sched_exec(void)
4417 struct task_struct
*p
= current
;
4418 unsigned long flags
;
4421 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4422 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), WF_EXEC
);
4423 if (dest_cpu
== smp_processor_id())
4426 if (likely(cpu_active(dest_cpu
))) {
4427 struct migration_arg arg
= { p
, dest_cpu
};
4429 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4430 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
4434 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4439 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4440 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
4442 EXPORT_PER_CPU_SYMBOL(kstat
);
4443 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
4446 * The function fair_sched_class.update_curr accesses the struct curr
4447 * and its field curr->exec_start; when called from task_sched_runtime(),
4448 * we observe a high rate of cache misses in practice.
4449 * Prefetching this data results in improved performance.
4451 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
4453 #ifdef CONFIG_FAIR_GROUP_SCHED
4454 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
4456 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
4459 prefetch(&curr
->exec_start
);
4463 * Return accounted runtime for the task.
4464 * In case the task is currently running, return the runtime plus current's
4465 * pending runtime that have not been accounted yet.
4467 unsigned long long task_sched_runtime(struct task_struct
*p
)
4473 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4475 * 64-bit doesn't need locks to atomically read a 64-bit value.
4476 * So we have a optimization chance when the task's delta_exec is 0.
4477 * Reading ->on_cpu is racy, but this is ok.
4479 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4480 * If we race with it entering CPU, unaccounted time is 0. This is
4481 * indistinguishable from the read occurring a few cycles earlier.
4482 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4483 * been accounted, so we're correct here as well.
4485 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
4486 return p
->se
.sum_exec_runtime
;
4489 rq
= task_rq_lock(p
, &rf
);
4491 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4492 * project cycles that may never be accounted to this
4493 * thread, breaking clock_gettime().
4495 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
4496 prefetch_curr_exec_start(p
);
4497 update_rq_clock(rq
);
4498 p
->sched_class
->update_curr(rq
);
4500 ns
= p
->se
.sum_exec_runtime
;
4501 task_rq_unlock(rq
, p
, &rf
);
4507 * This function gets called by the timer code, with HZ frequency.
4508 * We call it with interrupts disabled.
4510 void scheduler_tick(void)
4512 int cpu
= smp_processor_id();
4513 struct rq
*rq
= cpu_rq(cpu
);
4514 struct task_struct
*curr
= rq
->curr
;
4516 unsigned long thermal_pressure
;
4518 arch_scale_freq_tick();
4523 update_rq_clock(rq
);
4524 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
4525 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
4526 curr
->sched_class
->task_tick(rq
, curr
, 0);
4527 calc_global_load_tick(rq
);
4532 perf_event_task_tick();
4535 rq
->idle_balance
= idle_cpu(cpu
);
4536 trigger_load_balance(rq
);
4540 #ifdef CONFIG_NO_HZ_FULL
4545 struct delayed_work work
;
4547 /* Values for ->state, see diagram below. */
4548 #define TICK_SCHED_REMOTE_OFFLINE 0
4549 #define TICK_SCHED_REMOTE_OFFLINING 1
4550 #define TICK_SCHED_REMOTE_RUNNING 2
4553 * State diagram for ->state:
4556 * TICK_SCHED_REMOTE_OFFLINE
4559 * | | sched_tick_remote()
4562 * +--TICK_SCHED_REMOTE_OFFLINING
4565 * sched_tick_start() | | sched_tick_stop()
4568 * TICK_SCHED_REMOTE_RUNNING
4571 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4572 * and sched_tick_start() are happy to leave the state in RUNNING.
4575 static struct tick_work __percpu
*tick_work_cpu
;
4577 static void sched_tick_remote(struct work_struct
*work
)
4579 struct delayed_work
*dwork
= to_delayed_work(work
);
4580 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
4581 int cpu
= twork
->cpu
;
4582 struct rq
*rq
= cpu_rq(cpu
);
4583 struct task_struct
*curr
;
4589 * Handle the tick only if it appears the remote CPU is running in full
4590 * dynticks mode. The check is racy by nature, but missing a tick or
4591 * having one too much is no big deal because the scheduler tick updates
4592 * statistics and checks timeslices in a time-independent way, regardless
4593 * of when exactly it is running.
4595 if (!tick_nohz_tick_stopped_cpu(cpu
))
4598 rq_lock_irq(rq
, &rf
);
4600 if (cpu_is_offline(cpu
))
4603 update_rq_clock(rq
);
4605 if (!is_idle_task(curr
)) {
4607 * Make sure the next tick runs within a reasonable
4610 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
4611 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
4613 curr
->sched_class
->task_tick(rq
, curr
, 0);
4615 calc_load_nohz_remote(rq
);
4617 rq_unlock_irq(rq
, &rf
);
4621 * Run the remote tick once per second (1Hz). This arbitrary
4622 * frequency is large enough to avoid overload but short enough
4623 * to keep scheduler internal stats reasonably up to date. But
4624 * first update state to reflect hotplug activity if required.
4626 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
4627 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
4628 if (os
== TICK_SCHED_REMOTE_RUNNING
)
4629 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
4632 static void sched_tick_start(int cpu
)
4635 struct tick_work
*twork
;
4637 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
4640 WARN_ON_ONCE(!tick_work_cpu
);
4642 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
4643 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
4644 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
4645 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
4647 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
4648 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
4652 #ifdef CONFIG_HOTPLUG_CPU
4653 static void sched_tick_stop(int cpu
)
4655 struct tick_work
*twork
;
4658 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
4661 WARN_ON_ONCE(!tick_work_cpu
);
4663 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
4664 /* There cannot be competing actions, but don't rely on stop-machine. */
4665 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
4666 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
4667 /* Don't cancel, as this would mess up the state machine. */
4669 #endif /* CONFIG_HOTPLUG_CPU */
4671 int __init
sched_tick_offload_init(void)
4673 tick_work_cpu
= alloc_percpu(struct tick_work
);
4674 BUG_ON(!tick_work_cpu
);
4678 #else /* !CONFIG_NO_HZ_FULL */
4679 static inline void sched_tick_start(int cpu
) { }
4680 static inline void sched_tick_stop(int cpu
) { }
4683 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4684 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4686 * If the value passed in is equal to the current preempt count
4687 * then we just disabled preemption. Start timing the latency.
4689 static inline void preempt_latency_start(int val
)
4691 if (preempt_count() == val
) {
4692 unsigned long ip
= get_lock_parent_ip();
4693 #ifdef CONFIG_DEBUG_PREEMPT
4694 current
->preempt_disable_ip
= ip
;
4696 trace_preempt_off(CALLER_ADDR0
, ip
);
4700 void preempt_count_add(int val
)
4702 #ifdef CONFIG_DEBUG_PREEMPT
4706 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4709 __preempt_count_add(val
);
4710 #ifdef CONFIG_DEBUG_PREEMPT
4712 * Spinlock count overflowing soon?
4714 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4717 preempt_latency_start(val
);
4719 EXPORT_SYMBOL(preempt_count_add
);
4720 NOKPROBE_SYMBOL(preempt_count_add
);
4723 * If the value passed in equals to the current preempt count
4724 * then we just enabled preemption. Stop timing the latency.
4726 static inline void preempt_latency_stop(int val
)
4728 if (preempt_count() == val
)
4729 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
4732 void preempt_count_sub(int val
)
4734 #ifdef CONFIG_DEBUG_PREEMPT
4738 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4741 * Is the spinlock portion underflowing?
4743 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4744 !(preempt_count() & PREEMPT_MASK
)))
4748 preempt_latency_stop(val
);
4749 __preempt_count_sub(val
);
4751 EXPORT_SYMBOL(preempt_count_sub
);
4752 NOKPROBE_SYMBOL(preempt_count_sub
);
4755 static inline void preempt_latency_start(int val
) { }
4756 static inline void preempt_latency_stop(int val
) { }
4759 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
4761 #ifdef CONFIG_DEBUG_PREEMPT
4762 return p
->preempt_disable_ip
;
4769 * Print scheduling while atomic bug:
4771 static noinline
void __schedule_bug(struct task_struct
*prev
)
4773 /* Save this before calling printk(), since that will clobber it */
4774 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
4776 if (oops_in_progress
)
4779 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4780 prev
->comm
, prev
->pid
, preempt_count());
4782 debug_show_held_locks(prev
);
4784 if (irqs_disabled())
4785 print_irqtrace_events(prev
);
4786 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
4787 && in_atomic_preempt_off()) {
4788 pr_err("Preemption disabled at:");
4789 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
4792 panic("scheduling while atomic\n");
4795 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
4799 * Various schedule()-time debugging checks and statistics:
4801 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
4803 #ifdef CONFIG_SCHED_STACK_END_CHECK
4804 if (task_stack_end_corrupted(prev
))
4805 panic("corrupted stack end detected inside scheduler\n");
4807 if (task_scs_end_corrupted(prev
))
4808 panic("corrupted shadow stack detected inside scheduler\n");
4811 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4812 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
4813 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4814 prev
->comm
, prev
->pid
, prev
->non_block_count
);
4816 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
4820 if (unlikely(in_atomic_preempt_off())) {
4821 __schedule_bug(prev
);
4822 preempt_count_set(PREEMPT_DISABLED
);
4826 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4828 schedstat_inc(this_rq()->sched_count
);
4831 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
4832 struct rq_flags
*rf
)
4835 const struct sched_class
*class;
4837 * We must do the balancing pass before put_prev_task(), such
4838 * that when we release the rq->lock the task is in the same
4839 * state as before we took rq->lock.
4841 * We can terminate the balance pass as soon as we know there is
4842 * a runnable task of @class priority or higher.
4844 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
4845 if (class->balance(rq
, prev
, rf
))
4850 put_prev_task(rq
, prev
);
4854 * Pick up the highest-prio task:
4856 static inline struct task_struct
*
4857 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
4859 const struct sched_class
*class;
4860 struct task_struct
*p
;
4863 * Optimization: we know that if all tasks are in the fair class we can
4864 * call that function directly, but only if the @prev task wasn't of a
4865 * higher scheduling class, because otherwise those lose the
4866 * opportunity to pull in more work from other CPUs.
4868 if (likely(prev
->sched_class
<= &fair_sched_class
&&
4869 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
4871 p
= pick_next_task_fair(rq
, prev
, rf
);
4872 if (unlikely(p
== RETRY_TASK
))
4875 /* Assumes fair_sched_class->next == idle_sched_class */
4877 put_prev_task(rq
, prev
);
4878 p
= pick_next_task_idle(rq
);
4885 put_prev_task_balance(rq
, prev
, rf
);
4887 for_each_class(class) {
4888 p
= class->pick_next_task(rq
);
4893 /* The idle class should always have a runnable task: */
4898 * __schedule() is the main scheduler function.
4900 * The main means of driving the scheduler and thus entering this function are:
4902 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4904 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4905 * paths. For example, see arch/x86/entry_64.S.
4907 * To drive preemption between tasks, the scheduler sets the flag in timer
4908 * interrupt handler scheduler_tick().
4910 * 3. Wakeups don't really cause entry into schedule(). They add a
4911 * task to the run-queue and that's it.
4913 * Now, if the new task added to the run-queue preempts the current
4914 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4915 * called on the nearest possible occasion:
4917 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4919 * - in syscall or exception context, at the next outmost
4920 * preempt_enable(). (this might be as soon as the wake_up()'s
4923 * - in IRQ context, return from interrupt-handler to
4924 * preemptible context
4926 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4929 * - cond_resched() call
4930 * - explicit schedule() call
4931 * - return from syscall or exception to user-space
4932 * - return from interrupt-handler to user-space
4934 * WARNING: must be called with preemption disabled!
4936 static void __sched notrace
__schedule(bool preempt
)
4938 struct task_struct
*prev
, *next
;
4939 unsigned long *switch_count
;
4940 unsigned long prev_state
;
4945 cpu
= smp_processor_id();
4949 schedule_debug(prev
, preempt
);
4951 if (sched_feat(HRTICK
))
4954 local_irq_disable();
4955 rcu_note_context_switch(preempt
);
4958 * Make sure that signal_pending_state()->signal_pending() below
4959 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4960 * done by the caller to avoid the race with signal_wake_up():
4962 * __set_current_state(@state) signal_wake_up()
4963 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4964 * wake_up_state(p, state)
4965 * LOCK rq->lock LOCK p->pi_state
4966 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4967 * if (signal_pending_state()) if (p->state & @state)
4969 * Also, the membarrier system call requires a full memory barrier
4970 * after coming from user-space, before storing to rq->curr.
4973 smp_mb__after_spinlock();
4975 /* Promote REQ to ACT */
4976 rq
->clock_update_flags
<<= 1;
4977 update_rq_clock(rq
);
4979 switch_count
= &prev
->nivcsw
;
4982 * We must load prev->state once (task_struct::state is volatile), such
4985 * - we form a control dependency vs deactivate_task() below.
4986 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
4988 prev_state
= prev
->state
;
4989 if (!preempt
&& prev_state
) {
4990 if (signal_pending_state(prev_state
, prev
)) {
4991 prev
->state
= TASK_RUNNING
;
4993 prev
->sched_contributes_to_load
=
4994 (prev_state
& TASK_UNINTERRUPTIBLE
) &&
4995 !(prev_state
& TASK_NOLOAD
) &&
4996 !(prev
->flags
& PF_FROZEN
);
4998 if (prev
->sched_contributes_to_load
)
4999 rq
->nr_uninterruptible
++;
5002 * __schedule() ttwu()
5003 * prev_state = prev->state; if (p->on_rq && ...)
5004 * if (prev_state) goto out;
5005 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5006 * p->state = TASK_WAKING
5008 * Where __schedule() and ttwu() have matching control dependencies.
5010 * After this, schedule() must not care about p->state any more.
5012 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
5014 if (prev
->in_iowait
) {
5015 atomic_inc(&rq
->nr_iowait
);
5016 delayacct_blkio_start();
5019 switch_count
= &prev
->nvcsw
;
5022 next
= pick_next_task(rq
, prev
, &rf
);
5023 clear_tsk_need_resched(prev
);
5024 clear_preempt_need_resched();
5026 if (likely(prev
!= next
)) {
5029 * RCU users of rcu_dereference(rq->curr) may not see
5030 * changes to task_struct made by pick_next_task().
5032 RCU_INIT_POINTER(rq
->curr
, next
);
5034 * The membarrier system call requires each architecture
5035 * to have a full memory barrier after updating
5036 * rq->curr, before returning to user-space.
5038 * Here are the schemes providing that barrier on the
5039 * various architectures:
5040 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5041 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5042 * - finish_lock_switch() for weakly-ordered
5043 * architectures where spin_unlock is a full barrier,
5044 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5045 * is a RELEASE barrier),
5049 migrate_disable_switch(rq
, prev
);
5050 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
5052 trace_sched_switch(preempt
, prev
, next
);
5054 /* Also unlocks the rq: */
5055 rq
= context_switch(rq
, prev
, next
, &rf
);
5057 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
5059 rq_unpin_lock(rq
, &rf
);
5060 __balance_callbacks(rq
);
5061 raw_spin_unlock_irq(&rq
->lock
);
5065 void __noreturn
do_task_dead(void)
5067 /* Causes final put_task_struct in finish_task_switch(): */
5068 set_special_state(TASK_DEAD
);
5070 /* Tell freezer to ignore us: */
5071 current
->flags
|= PF_NOFREEZE
;
5076 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5081 static inline void sched_submit_work(struct task_struct
*tsk
)
5083 unsigned int task_flags
;
5088 task_flags
= tsk
->flags
;
5090 * If a worker went to sleep, notify and ask workqueue whether
5091 * it wants to wake up a task to maintain concurrency.
5092 * As this function is called inside the schedule() context,
5093 * we disable preemption to avoid it calling schedule() again
5094 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5097 if (task_flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
5099 if (task_flags
& PF_WQ_WORKER
)
5100 wq_worker_sleeping(tsk
);
5102 io_wq_worker_sleeping(tsk
);
5103 preempt_enable_no_resched();
5106 if (tsk_is_pi_blocked(tsk
))
5110 * If we are going to sleep and we have plugged IO queued,
5111 * make sure to submit it to avoid deadlocks.
5113 if (blk_needs_flush_plug(tsk
))
5114 blk_schedule_flush_plug(tsk
);
5117 static void sched_update_worker(struct task_struct
*tsk
)
5119 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
5120 if (tsk
->flags
& PF_WQ_WORKER
)
5121 wq_worker_running(tsk
);
5123 io_wq_worker_running(tsk
);
5127 asmlinkage __visible
void __sched
schedule(void)
5129 struct task_struct
*tsk
= current
;
5131 sched_submit_work(tsk
);
5135 sched_preempt_enable_no_resched();
5136 } while (need_resched());
5137 sched_update_worker(tsk
);
5139 EXPORT_SYMBOL(schedule
);
5142 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
5143 * state (have scheduled out non-voluntarily) by making sure that all
5144 * tasks have either left the run queue or have gone into user space.
5145 * As idle tasks do not do either, they must not ever be preempted
5146 * (schedule out non-voluntarily).
5148 * schedule_idle() is similar to schedule_preempt_disable() except that it
5149 * never enables preemption because it does not call sched_submit_work().
5151 void __sched
schedule_idle(void)
5154 * As this skips calling sched_submit_work(), which the idle task does
5155 * regardless because that function is a nop when the task is in a
5156 * TASK_RUNNING state, make sure this isn't used someplace that the
5157 * current task can be in any other state. Note, idle is always in the
5158 * TASK_RUNNING state.
5160 WARN_ON_ONCE(current
->state
);
5163 } while (need_resched());
5166 #ifdef CONFIG_CONTEXT_TRACKING
5167 asmlinkage __visible
void __sched
schedule_user(void)
5170 * If we come here after a random call to set_need_resched(),
5171 * or we have been woken up remotely but the IPI has not yet arrived,
5172 * we haven't yet exited the RCU idle mode. Do it here manually until
5173 * we find a better solution.
5175 * NB: There are buggy callers of this function. Ideally we
5176 * should warn if prev_state != CONTEXT_USER, but that will trigger
5177 * too frequently to make sense yet.
5179 enum ctx_state prev_state
= exception_enter();
5181 exception_exit(prev_state
);
5186 * schedule_preempt_disabled - called with preemption disabled
5188 * Returns with preemption disabled. Note: preempt_count must be 1
5190 void __sched
schedule_preempt_disabled(void)
5192 sched_preempt_enable_no_resched();
5197 static void __sched notrace
preempt_schedule_common(void)
5201 * Because the function tracer can trace preempt_count_sub()
5202 * and it also uses preempt_enable/disable_notrace(), if
5203 * NEED_RESCHED is set, the preempt_enable_notrace() called
5204 * by the function tracer will call this function again and
5205 * cause infinite recursion.
5207 * Preemption must be disabled here before the function
5208 * tracer can trace. Break up preempt_disable() into two
5209 * calls. One to disable preemption without fear of being
5210 * traced. The other to still record the preemption latency,
5211 * which can also be traced by the function tracer.
5213 preempt_disable_notrace();
5214 preempt_latency_start(1);
5216 preempt_latency_stop(1);
5217 preempt_enable_no_resched_notrace();
5220 * Check again in case we missed a preemption opportunity
5221 * between schedule and now.
5223 } while (need_resched());
5226 #ifdef CONFIG_PREEMPTION
5228 * This is the entry point to schedule() from in-kernel preemption
5229 * off of preempt_enable.
5231 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
5234 * If there is a non-zero preempt_count or interrupts are disabled,
5235 * we do not want to preempt the current task. Just return..
5237 if (likely(!preemptible()))
5240 preempt_schedule_common();
5242 NOKPROBE_SYMBOL(preempt_schedule
);
5243 EXPORT_SYMBOL(preempt_schedule
);
5246 * preempt_schedule_notrace - preempt_schedule called by tracing
5248 * The tracing infrastructure uses preempt_enable_notrace to prevent
5249 * recursion and tracing preempt enabling caused by the tracing
5250 * infrastructure itself. But as tracing can happen in areas coming
5251 * from userspace or just about to enter userspace, a preempt enable
5252 * can occur before user_exit() is called. This will cause the scheduler
5253 * to be called when the system is still in usermode.
5255 * To prevent this, the preempt_enable_notrace will use this function
5256 * instead of preempt_schedule() to exit user context if needed before
5257 * calling the scheduler.
5259 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
5261 enum ctx_state prev_ctx
;
5263 if (likely(!preemptible()))
5268 * Because the function tracer can trace preempt_count_sub()
5269 * and it also uses preempt_enable/disable_notrace(), if
5270 * NEED_RESCHED is set, the preempt_enable_notrace() called
5271 * by the function tracer will call this function again and
5272 * cause infinite recursion.
5274 * Preemption must be disabled here before the function
5275 * tracer can trace. Break up preempt_disable() into two
5276 * calls. One to disable preemption without fear of being
5277 * traced. The other to still record the preemption latency,
5278 * which can also be traced by the function tracer.
5280 preempt_disable_notrace();
5281 preempt_latency_start(1);
5283 * Needs preempt disabled in case user_exit() is traced
5284 * and the tracer calls preempt_enable_notrace() causing
5285 * an infinite recursion.
5287 prev_ctx
= exception_enter();
5289 exception_exit(prev_ctx
);
5291 preempt_latency_stop(1);
5292 preempt_enable_no_resched_notrace();
5293 } while (need_resched());
5295 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
5297 #endif /* CONFIG_PREEMPTION */
5300 * This is the entry point to schedule() from kernel preemption
5301 * off of irq context.
5302 * Note, that this is called and return with irqs disabled. This will
5303 * protect us against recursive calling from irq.
5305 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
5307 enum ctx_state prev_state
;
5309 /* Catch callers which need to be fixed */
5310 BUG_ON(preempt_count() || !irqs_disabled());
5312 prev_state
= exception_enter();
5318 local_irq_disable();
5319 sched_preempt_enable_no_resched();
5320 } while (need_resched());
5322 exception_exit(prev_state
);
5325 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
5328 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG
) && wake_flags
& ~WF_SYNC
);
5329 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5331 EXPORT_SYMBOL(default_wake_function
);
5333 #ifdef CONFIG_RT_MUTEXES
5335 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
5338 prio
= min(prio
, pi_task
->prio
);
5343 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
5345 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
5347 return __rt_effective_prio(pi_task
, prio
);
5351 * rt_mutex_setprio - set the current priority of a task
5353 * @pi_task: donor task
5355 * This function changes the 'effective' priority of a task. It does
5356 * not touch ->normal_prio like __setscheduler().
5358 * Used by the rt_mutex code to implement priority inheritance
5359 * logic. Call site only calls if the priority of the task changed.
5361 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
5363 int prio
, oldprio
, queued
, running
, queue_flag
=
5364 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
5365 const struct sched_class
*prev_class
;
5369 /* XXX used to be waiter->prio, not waiter->task->prio */
5370 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
5373 * If nothing changed; bail early.
5375 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
5378 rq
= __task_rq_lock(p
, &rf
);
5379 update_rq_clock(rq
);
5381 * Set under pi_lock && rq->lock, such that the value can be used under
5384 * Note that there is loads of tricky to make this pointer cache work
5385 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
5386 * ensure a task is de-boosted (pi_task is set to NULL) before the
5387 * task is allowed to run again (and can exit). This ensures the pointer
5388 * points to a blocked task -- which guarantees the task is present.
5390 p
->pi_top_task
= pi_task
;
5393 * For FIFO/RR we only need to set prio, if that matches we're done.
5395 if (prio
== p
->prio
&& !dl_prio(prio
))
5399 * Idle task boosting is a nono in general. There is one
5400 * exception, when PREEMPT_RT and NOHZ is active:
5402 * The idle task calls get_next_timer_interrupt() and holds
5403 * the timer wheel base->lock on the CPU and another CPU wants
5404 * to access the timer (probably to cancel it). We can safely
5405 * ignore the boosting request, as the idle CPU runs this code
5406 * with interrupts disabled and will complete the lock
5407 * protected section without being interrupted. So there is no
5408 * real need to boost.
5410 if (unlikely(p
== rq
->idle
)) {
5411 WARN_ON(p
!= rq
->curr
);
5412 WARN_ON(p
->pi_blocked_on
);
5416 trace_sched_pi_setprio(p
, pi_task
);
5419 if (oldprio
== prio
)
5420 queue_flag
&= ~DEQUEUE_MOVE
;
5422 prev_class
= p
->sched_class
;
5423 queued
= task_on_rq_queued(p
);
5424 running
= task_current(rq
, p
);
5426 dequeue_task(rq
, p
, queue_flag
);
5428 put_prev_task(rq
, p
);
5431 * Boosting condition are:
5432 * 1. -rt task is running and holds mutex A
5433 * --> -dl task blocks on mutex A
5435 * 2. -dl task is running and holds mutex A
5436 * --> -dl task blocks on mutex A and could preempt the
5439 if (dl_prio(prio
)) {
5440 if (!dl_prio(p
->normal_prio
) ||
5441 (pi_task
&& dl_prio(pi_task
->prio
) &&
5442 dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
5443 p
->dl
.dl_boosted
= 1;
5444 queue_flag
|= ENQUEUE_REPLENISH
;
5446 p
->dl
.dl_boosted
= 0;
5447 p
->sched_class
= &dl_sched_class
;
5448 } else if (rt_prio(prio
)) {
5449 if (dl_prio(oldprio
))
5450 p
->dl
.dl_boosted
= 0;
5452 queue_flag
|= ENQUEUE_HEAD
;
5453 p
->sched_class
= &rt_sched_class
;
5455 if (dl_prio(oldprio
))
5456 p
->dl
.dl_boosted
= 0;
5457 if (rt_prio(oldprio
))
5459 p
->sched_class
= &fair_sched_class
;
5465 enqueue_task(rq
, p
, queue_flag
);
5467 set_next_task(rq
, p
);
5469 check_class_changed(rq
, p
, prev_class
, oldprio
);
5471 /* Avoid rq from going away on us: */
5474 rq_unpin_lock(rq
, &rf
);
5475 __balance_callbacks(rq
);
5476 raw_spin_unlock(&rq
->lock
);
5481 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
5487 void set_user_nice(struct task_struct
*p
, long nice
)
5489 bool queued
, running
;
5494 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
5497 * We have to be careful, if called from sys_setpriority(),
5498 * the task might be in the middle of scheduling on another CPU.
5500 rq
= task_rq_lock(p
, &rf
);
5501 update_rq_clock(rq
);
5504 * The RT priorities are set via sched_setscheduler(), but we still
5505 * allow the 'normal' nice value to be set - but as expected
5506 * it won't have any effect on scheduling until the task is
5507 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5509 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
5510 p
->static_prio
= NICE_TO_PRIO(nice
);
5513 queued
= task_on_rq_queued(p
);
5514 running
= task_current(rq
, p
);
5516 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
5518 put_prev_task(rq
, p
);
5520 p
->static_prio
= NICE_TO_PRIO(nice
);
5521 set_load_weight(p
, true);
5523 p
->prio
= effective_prio(p
);
5526 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5528 set_next_task(rq
, p
);
5531 * If the task increased its priority or is running and
5532 * lowered its priority, then reschedule its CPU:
5534 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
5537 task_rq_unlock(rq
, p
, &rf
);
5539 EXPORT_SYMBOL(set_user_nice
);
5542 * can_nice - check if a task can reduce its nice value
5546 int can_nice(const struct task_struct
*p
, const int nice
)
5548 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5549 int nice_rlim
= nice_to_rlimit(nice
);
5551 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
5552 capable(CAP_SYS_NICE
));
5555 #ifdef __ARCH_WANT_SYS_NICE
5558 * sys_nice - change the priority of the current process.
5559 * @increment: priority increment
5561 * sys_setpriority is a more generic, but much slower function that
5562 * does similar things.
5564 SYSCALL_DEFINE1(nice
, int, increment
)
5569 * Setpriority might change our priority at the same moment.
5570 * We don't have to worry. Conceptually one call occurs first
5571 * and we have a single winner.
5573 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
5574 nice
= task_nice(current
) + increment
;
5576 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
5577 if (increment
< 0 && !can_nice(current
, nice
))
5580 retval
= security_task_setnice(current
, nice
);
5584 set_user_nice(current
, nice
);
5591 * task_prio - return the priority value of a given task.
5592 * @p: the task in question.
5594 * Return: The priority value as seen by users in /proc.
5595 * RT tasks are offset by -200. Normal tasks are centered
5596 * around 0, value goes from -16 to +15.
5598 int task_prio(const struct task_struct
*p
)
5600 return p
->prio
- MAX_RT_PRIO
;
5604 * idle_cpu - is a given CPU idle currently?
5605 * @cpu: the processor in question.
5607 * Return: 1 if the CPU is currently idle. 0 otherwise.
5609 int idle_cpu(int cpu
)
5611 struct rq
*rq
= cpu_rq(cpu
);
5613 if (rq
->curr
!= rq
->idle
)
5620 if (rq
->ttwu_pending
)
5628 * available_idle_cpu - is a given CPU idle for enqueuing work.
5629 * @cpu: the CPU in question.
5631 * Return: 1 if the CPU is currently idle. 0 otherwise.
5633 int available_idle_cpu(int cpu
)
5638 if (vcpu_is_preempted(cpu
))
5645 * idle_task - return the idle task for a given CPU.
5646 * @cpu: the processor in question.
5648 * Return: The idle task for the CPU @cpu.
5650 struct task_struct
*idle_task(int cpu
)
5652 return cpu_rq(cpu
)->idle
;
5656 * find_process_by_pid - find a process with a matching PID value.
5657 * @pid: the pid in question.
5659 * The task of @pid, if found. %NULL otherwise.
5661 static struct task_struct
*find_process_by_pid(pid_t pid
)
5663 return pid
? find_task_by_vpid(pid
) : current
;
5667 * sched_setparam() passes in -1 for its policy, to let the functions
5668 * it calls know not to change it.
5670 #define SETPARAM_POLICY -1
5672 static void __setscheduler_params(struct task_struct
*p
,
5673 const struct sched_attr
*attr
)
5675 int policy
= attr
->sched_policy
;
5677 if (policy
== SETPARAM_POLICY
)
5682 if (dl_policy(policy
))
5683 __setparam_dl(p
, attr
);
5684 else if (fair_policy(policy
))
5685 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
5688 * __sched_setscheduler() ensures attr->sched_priority == 0 when
5689 * !rt_policy. Always setting this ensures that things like
5690 * getparam()/getattr() don't report silly values for !rt tasks.
5692 p
->rt_priority
= attr
->sched_priority
;
5693 p
->normal_prio
= normal_prio(p
);
5694 set_load_weight(p
, true);
5697 /* Actually do priority change: must hold pi & rq lock. */
5698 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
5699 const struct sched_attr
*attr
, bool keep_boost
)
5702 * If params can't change scheduling class changes aren't allowed
5705 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
5708 __setscheduler_params(p
, attr
);
5711 * Keep a potential priority boosting if called from
5712 * sched_setscheduler().
5714 p
->prio
= normal_prio(p
);
5716 p
->prio
= rt_effective_prio(p
, p
->prio
);
5718 if (dl_prio(p
->prio
))
5719 p
->sched_class
= &dl_sched_class
;
5720 else if (rt_prio(p
->prio
))
5721 p
->sched_class
= &rt_sched_class
;
5723 p
->sched_class
= &fair_sched_class
;
5727 * Check the target process has a UID that matches the current process's:
5729 static bool check_same_owner(struct task_struct
*p
)
5731 const struct cred
*cred
= current_cred(), *pcred
;
5735 pcred
= __task_cred(p
);
5736 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
5737 uid_eq(cred
->euid
, pcred
->uid
));
5742 static int __sched_setscheduler(struct task_struct
*p
,
5743 const struct sched_attr
*attr
,
5746 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
5747 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
5748 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
5749 int new_effective_prio
, policy
= attr
->sched_policy
;
5750 const struct sched_class
*prev_class
;
5751 struct callback_head
*head
;
5754 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
5757 /* The pi code expects interrupts enabled */
5758 BUG_ON(pi
&& in_interrupt());
5760 /* Double check policy once rq lock held: */
5762 reset_on_fork
= p
->sched_reset_on_fork
;
5763 policy
= oldpolicy
= p
->policy
;
5765 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
5767 if (!valid_policy(policy
))
5771 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
5775 * Valid priorities for SCHED_FIFO and SCHED_RR are
5776 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5777 * SCHED_BATCH and SCHED_IDLE is 0.
5779 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5780 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
5782 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
5783 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
5787 * Allow unprivileged RT tasks to decrease priority:
5789 if (user
&& !capable(CAP_SYS_NICE
)) {
5790 if (fair_policy(policy
)) {
5791 if (attr
->sched_nice
< task_nice(p
) &&
5792 !can_nice(p
, attr
->sched_nice
))
5796 if (rt_policy(policy
)) {
5797 unsigned long rlim_rtprio
=
5798 task_rlimit(p
, RLIMIT_RTPRIO
);
5800 /* Can't set/change the rt policy: */
5801 if (policy
!= p
->policy
&& !rlim_rtprio
)
5804 /* Can't increase priority: */
5805 if (attr
->sched_priority
> p
->rt_priority
&&
5806 attr
->sched_priority
> rlim_rtprio
)
5811 * Can't set/change SCHED_DEADLINE policy at all for now
5812 * (safest behavior); in the future we would like to allow
5813 * unprivileged DL tasks to increase their relative deadline
5814 * or reduce their runtime (both ways reducing utilization)
5816 if (dl_policy(policy
))
5820 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5821 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5823 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
5824 if (!can_nice(p
, task_nice(p
)))
5828 /* Can't change other user's priorities: */
5829 if (!check_same_owner(p
))
5832 /* Normal users shall not reset the sched_reset_on_fork flag: */
5833 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
5838 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
5841 retval
= security_task_setscheduler(p
);
5846 /* Update task specific "requested" clamps */
5847 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
5848 retval
= uclamp_validate(p
, attr
);
5857 * Make sure no PI-waiters arrive (or leave) while we are
5858 * changing the priority of the task:
5860 * To be able to change p->policy safely, the appropriate
5861 * runqueue lock must be held.
5863 rq
= task_rq_lock(p
, &rf
);
5864 update_rq_clock(rq
);
5867 * Changing the policy of the stop threads its a very bad idea:
5869 if (p
== rq
->stop
) {
5875 * If not changing anything there's no need to proceed further,
5876 * but store a possible modification of reset_on_fork.
5878 if (unlikely(policy
== p
->policy
)) {
5879 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
5881 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
5883 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
5885 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
5888 p
->sched_reset_on_fork
= reset_on_fork
;
5895 #ifdef CONFIG_RT_GROUP_SCHED
5897 * Do not allow realtime tasks into groups that have no runtime
5900 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5901 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5902 !task_group_is_autogroup(task_group(p
))) {
5908 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
5909 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
5910 cpumask_t
*span
= rq
->rd
->span
;
5913 * Don't allow tasks with an affinity mask smaller than
5914 * the entire root_domain to become SCHED_DEADLINE. We
5915 * will also fail if there's no bandwidth available.
5917 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
5918 rq
->rd
->dl_bw
.bw
== 0) {
5926 /* Re-check policy now with rq lock held: */
5927 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5928 policy
= oldpolicy
= -1;
5929 task_rq_unlock(rq
, p
, &rf
);
5931 cpuset_read_unlock();
5936 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5937 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5940 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
5945 p
->sched_reset_on_fork
= reset_on_fork
;
5950 * Take priority boosted tasks into account. If the new
5951 * effective priority is unchanged, we just store the new
5952 * normal parameters and do not touch the scheduler class and
5953 * the runqueue. This will be done when the task deboost
5956 new_effective_prio
= rt_effective_prio(p
, newprio
);
5957 if (new_effective_prio
== oldprio
)
5958 queue_flags
&= ~DEQUEUE_MOVE
;
5961 queued
= task_on_rq_queued(p
);
5962 running
= task_current(rq
, p
);
5964 dequeue_task(rq
, p
, queue_flags
);
5966 put_prev_task(rq
, p
);
5968 prev_class
= p
->sched_class
;
5970 __setscheduler(rq
, p
, attr
, pi
);
5971 __setscheduler_uclamp(p
, attr
);
5975 * We enqueue to tail when the priority of a task is
5976 * increased (user space view).
5978 if (oldprio
< p
->prio
)
5979 queue_flags
|= ENQUEUE_HEAD
;
5981 enqueue_task(rq
, p
, queue_flags
);
5984 set_next_task(rq
, p
);
5986 check_class_changed(rq
, p
, prev_class
, oldprio
);
5988 /* Avoid rq from going away on us: */
5990 head
= splice_balance_callbacks(rq
);
5991 task_rq_unlock(rq
, p
, &rf
);
5994 cpuset_read_unlock();
5995 rt_mutex_adjust_pi(p
);
5998 /* Run balance callbacks after we've adjusted the PI chain: */
5999 balance_callbacks(rq
, head
);
6005 task_rq_unlock(rq
, p
, &rf
);
6007 cpuset_read_unlock();
6011 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
6012 const struct sched_param
*param
, bool check
)
6014 struct sched_attr attr
= {
6015 .sched_policy
= policy
,
6016 .sched_priority
= param
->sched_priority
,
6017 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
6020 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
6021 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
6022 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
6023 policy
&= ~SCHED_RESET_ON_FORK
;
6024 attr
.sched_policy
= policy
;
6027 return __sched_setscheduler(p
, &attr
, check
, true);
6030 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6031 * @p: the task in question.
6032 * @policy: new policy.
6033 * @param: structure containing the new RT priority.
6035 * Use sched_set_fifo(), read its comment.
6037 * Return: 0 on success. An error code otherwise.
6039 * NOTE that the task may be already dead.
6041 int sched_setscheduler(struct task_struct
*p
, int policy
,
6042 const struct sched_param
*param
)
6044 return _sched_setscheduler(p
, policy
, param
, true);
6047 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
6049 return __sched_setscheduler(p
, attr
, true, true);
6052 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
6054 return __sched_setscheduler(p
, attr
, false, true);
6058 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6059 * @p: the task in question.
6060 * @policy: new policy.
6061 * @param: structure containing the new RT priority.
6063 * Just like sched_setscheduler, only don't bother checking if the
6064 * current context has permission. For example, this is needed in
6065 * stop_machine(): we create temporary high priority worker threads,
6066 * but our caller might not have that capability.
6068 * Return: 0 on success. An error code otherwise.
6070 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6071 const struct sched_param
*param
)
6073 return _sched_setscheduler(p
, policy
, param
, false);
6077 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
6078 * incapable of resource management, which is the one thing an OS really should
6081 * This is of course the reason it is limited to privileged users only.
6083 * Worse still; it is fundamentally impossible to compose static priority
6084 * workloads. You cannot take two correctly working static prio workloads
6085 * and smash them together and still expect them to work.
6087 * For this reason 'all' FIFO tasks the kernel creates are basically at:
6091 * The administrator _MUST_ configure the system, the kernel simply doesn't
6092 * know enough information to make a sensible choice.
6094 void sched_set_fifo(struct task_struct
*p
)
6096 struct sched_param sp
= { .sched_priority
= MAX_RT_PRIO
/ 2 };
6097 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
6099 EXPORT_SYMBOL_GPL(sched_set_fifo
);
6102 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
6104 void sched_set_fifo_low(struct task_struct
*p
)
6106 struct sched_param sp
= { .sched_priority
= 1 };
6107 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
6109 EXPORT_SYMBOL_GPL(sched_set_fifo_low
);
6111 void sched_set_normal(struct task_struct
*p
, int nice
)
6113 struct sched_attr attr
= {
6114 .sched_policy
= SCHED_NORMAL
,
6117 WARN_ON_ONCE(sched_setattr_nocheck(p
, &attr
) != 0);
6119 EXPORT_SYMBOL_GPL(sched_set_normal
);
6122 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6124 struct sched_param lparam
;
6125 struct task_struct
*p
;
6128 if (!param
|| pid
< 0)
6130 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6135 p
= find_process_by_pid(pid
);
6141 retval
= sched_setscheduler(p
, policy
, &lparam
);
6149 * Mimics kernel/events/core.c perf_copy_attr().
6151 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
6156 /* Zero the full structure, so that a short copy will be nice: */
6157 memset(attr
, 0, sizeof(*attr
));
6159 ret
= get_user(size
, &uattr
->size
);
6163 /* ABI compatibility quirk: */
6165 size
= SCHED_ATTR_SIZE_VER0
;
6166 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
6169 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
6176 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
6177 size
< SCHED_ATTR_SIZE_VER1
)
6181 * XXX: Do we want to be lenient like existing syscalls; or do we want
6182 * to be strict and return an error on out-of-bounds values?
6184 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
6189 put_user(sizeof(*attr
), &uattr
->size
);
6194 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6195 * @pid: the pid in question.
6196 * @policy: new policy.
6197 * @param: structure containing the new RT priority.
6199 * Return: 0 on success. An error code otherwise.
6201 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
6206 return do_sched_setscheduler(pid
, policy
, param
);
6210 * sys_sched_setparam - set/change the RT priority of a thread
6211 * @pid: the pid in question.
6212 * @param: structure containing the new RT priority.
6214 * Return: 0 on success. An error code otherwise.
6216 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6218 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
6222 * sys_sched_setattr - same as above, but with extended sched_attr
6223 * @pid: the pid in question.
6224 * @uattr: structure containing the extended parameters.
6225 * @flags: for future extension.
6227 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
6228 unsigned int, flags
)
6230 struct sched_attr attr
;
6231 struct task_struct
*p
;
6234 if (!uattr
|| pid
< 0 || flags
)
6237 retval
= sched_copy_attr(uattr
, &attr
);
6241 if ((int)attr
.sched_policy
< 0)
6243 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
6244 attr
.sched_policy
= SETPARAM_POLICY
;
6248 p
= find_process_by_pid(pid
);
6254 retval
= sched_setattr(p
, &attr
);
6262 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6263 * @pid: the pid in question.
6265 * Return: On success, the policy of the thread. Otherwise, a negative error
6268 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6270 struct task_struct
*p
;
6278 p
= find_process_by_pid(pid
);
6280 retval
= security_task_getscheduler(p
);
6283 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6290 * sys_sched_getparam - get the RT priority of a thread
6291 * @pid: the pid in question.
6292 * @param: structure containing the RT priority.
6294 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
6297 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6299 struct sched_param lp
= { .sched_priority
= 0 };
6300 struct task_struct
*p
;
6303 if (!param
|| pid
< 0)
6307 p
= find_process_by_pid(pid
);
6312 retval
= security_task_getscheduler(p
);
6316 if (task_has_rt_policy(p
))
6317 lp
.sched_priority
= p
->rt_priority
;
6321 * This one might sleep, we cannot do it with a spinlock held ...
6323 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6333 * Copy the kernel size attribute structure (which might be larger
6334 * than what user-space knows about) to user-space.
6336 * Note that all cases are valid: user-space buffer can be larger or
6337 * smaller than the kernel-space buffer. The usual case is that both
6338 * have the same size.
6341 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
6342 struct sched_attr
*kattr
,
6345 unsigned int ksize
= sizeof(*kattr
);
6347 if (!access_ok(uattr
, usize
))
6351 * sched_getattr() ABI forwards and backwards compatibility:
6353 * If usize == ksize then we just copy everything to user-space and all is good.
6355 * If usize < ksize then we only copy as much as user-space has space for,
6356 * this keeps ABI compatibility as well. We skip the rest.
6358 * If usize > ksize then user-space is using a newer version of the ABI,
6359 * which part the kernel doesn't know about. Just ignore it - tooling can
6360 * detect the kernel's knowledge of attributes from the attr->size value
6361 * which is set to ksize in this case.
6363 kattr
->size
= min(usize
, ksize
);
6365 if (copy_to_user(uattr
, kattr
, kattr
->size
))
6372 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
6373 * @pid: the pid in question.
6374 * @uattr: structure containing the extended parameters.
6375 * @usize: sizeof(attr) for fwd/bwd comp.
6376 * @flags: for future extension.
6378 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
6379 unsigned int, usize
, unsigned int, flags
)
6381 struct sched_attr kattr
= { };
6382 struct task_struct
*p
;
6385 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
6386 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
6390 p
= find_process_by_pid(pid
);
6395 retval
= security_task_getscheduler(p
);
6399 kattr
.sched_policy
= p
->policy
;
6400 if (p
->sched_reset_on_fork
)
6401 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
6402 if (task_has_dl_policy(p
))
6403 __getparam_dl(p
, &kattr
);
6404 else if (task_has_rt_policy(p
))
6405 kattr
.sched_priority
= p
->rt_priority
;
6407 kattr
.sched_nice
= task_nice(p
);
6409 #ifdef CONFIG_UCLAMP_TASK
6411 * This could race with another potential updater, but this is fine
6412 * because it'll correctly read the old or the new value. We don't need
6413 * to guarantee who wins the race as long as it doesn't return garbage.
6415 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
6416 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
6421 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
6428 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6430 cpumask_var_t cpus_allowed
, new_mask
;
6431 struct task_struct
*p
;
6436 p
= find_process_by_pid(pid
);
6442 /* Prevent p going away */
6446 if (p
->flags
& PF_NO_SETAFFINITY
) {
6450 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6454 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6456 goto out_free_cpus_allowed
;
6459 if (!check_same_owner(p
)) {
6461 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
6463 goto out_free_new_mask
;
6468 retval
= security_task_setscheduler(p
);
6470 goto out_free_new_mask
;
6473 cpuset_cpus_allowed(p
, cpus_allowed
);
6474 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6477 * Since bandwidth control happens on root_domain basis,
6478 * if admission test is enabled, we only admit -deadline
6479 * tasks allowed to run on all the CPUs in the task's
6483 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
6485 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
6488 goto out_free_new_mask
;
6494 retval
= __set_cpus_allowed_ptr(p
, new_mask
, SCA_CHECK
);
6497 cpuset_cpus_allowed(p
, cpus_allowed
);
6498 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6500 * We must have raced with a concurrent cpuset
6501 * update. Just reset the cpus_allowed to the
6502 * cpuset's cpus_allowed
6504 cpumask_copy(new_mask
, cpus_allowed
);
6509 free_cpumask_var(new_mask
);
6510 out_free_cpus_allowed
:
6511 free_cpumask_var(cpus_allowed
);
6517 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6518 struct cpumask
*new_mask
)
6520 if (len
< cpumask_size())
6521 cpumask_clear(new_mask
);
6522 else if (len
> cpumask_size())
6523 len
= cpumask_size();
6525 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6529 * sys_sched_setaffinity - set the CPU affinity of a process
6530 * @pid: pid of the process
6531 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6532 * @user_mask_ptr: user-space pointer to the new CPU mask
6534 * Return: 0 on success. An error code otherwise.
6536 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6537 unsigned long __user
*, user_mask_ptr
)
6539 cpumask_var_t new_mask
;
6542 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6545 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6547 retval
= sched_setaffinity(pid
, new_mask
);
6548 free_cpumask_var(new_mask
);
6552 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6554 struct task_struct
*p
;
6555 unsigned long flags
;
6561 p
= find_process_by_pid(pid
);
6565 retval
= security_task_getscheduler(p
);
6569 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
6570 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
6571 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6580 * sys_sched_getaffinity - get the CPU affinity of a process
6581 * @pid: pid of the process
6582 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6583 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6585 * Return: size of CPU mask copied to user_mask_ptr on success. An
6586 * error code otherwise.
6588 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6589 unsigned long __user
*, user_mask_ptr
)
6594 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
6596 if (len
& (sizeof(unsigned long)-1))
6599 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6602 ret
= sched_getaffinity(pid
, mask
);
6604 unsigned int retlen
= min(len
, cpumask_size());
6606 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
6611 free_cpumask_var(mask
);
6617 * sys_sched_yield - yield the current processor to other threads.
6619 * This function yields the current CPU to other tasks. If there are no
6620 * other threads running on this CPU then this function will return.
6624 static void do_sched_yield(void)
6629 rq
= this_rq_lock_irq(&rf
);
6631 schedstat_inc(rq
->yld_count
);
6632 current
->sched_class
->yield_task(rq
);
6635 rq_unlock_irq(rq
, &rf
);
6636 sched_preempt_enable_no_resched();
6641 SYSCALL_DEFINE0(sched_yield
)
6647 #ifndef CONFIG_PREEMPTION
6648 int __sched
_cond_resched(void)
6650 if (should_resched(0)) {
6651 preempt_schedule_common();
6657 EXPORT_SYMBOL(_cond_resched
);
6661 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6662 * call schedule, and on return reacquire the lock.
6664 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
6665 * operations here to prevent schedule() from being called twice (once via
6666 * spin_unlock(), once by hand).
6668 int __cond_resched_lock(spinlock_t
*lock
)
6670 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
6673 lockdep_assert_held(lock
);
6675 if (spin_needbreak(lock
) || resched
) {
6678 preempt_schedule_common();
6686 EXPORT_SYMBOL(__cond_resched_lock
);
6689 * yield - yield the current processor to other threads.
6691 * Do not ever use this function, there's a 99% chance you're doing it wrong.
6693 * The scheduler is at all times free to pick the calling task as the most
6694 * eligible task to run, if removing the yield() call from your code breaks
6695 * it, it's already broken.
6697 * Typical broken usage is:
6702 * where one assumes that yield() will let 'the other' process run that will
6703 * make event true. If the current task is a SCHED_FIFO task that will never
6704 * happen. Never use yield() as a progress guarantee!!
6706 * If you want to use yield() to wait for something, use wait_event().
6707 * If you want to use yield() to be 'nice' for others, use cond_resched().
6708 * If you still want to use yield(), do not!
6710 void __sched
yield(void)
6712 set_current_state(TASK_RUNNING
);
6715 EXPORT_SYMBOL(yield
);
6718 * yield_to - yield the current processor to another thread in
6719 * your thread group, or accelerate that thread toward the
6720 * processor it's on.
6722 * @preempt: whether task preemption is allowed or not
6724 * It's the caller's job to ensure that the target task struct
6725 * can't go away on us before we can do any checks.
6728 * true (>0) if we indeed boosted the target task.
6729 * false (0) if we failed to boost the target.
6730 * -ESRCH if there's no task to yield to.
6732 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
6734 struct task_struct
*curr
= current
;
6735 struct rq
*rq
, *p_rq
;
6736 unsigned long flags
;
6739 local_irq_save(flags
);
6745 * If we're the only runnable task on the rq and target rq also
6746 * has only one task, there's absolutely no point in yielding.
6748 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
6753 double_rq_lock(rq
, p_rq
);
6754 if (task_rq(p
) != p_rq
) {
6755 double_rq_unlock(rq
, p_rq
);
6759 if (!curr
->sched_class
->yield_to_task
)
6762 if (curr
->sched_class
!= p
->sched_class
)
6765 if (task_running(p_rq
, p
) || p
->state
)
6768 yielded
= curr
->sched_class
->yield_to_task(rq
, p
);
6770 schedstat_inc(rq
->yld_count
);
6772 * Make p's CPU reschedule; pick_next_entity takes care of
6775 if (preempt
&& rq
!= p_rq
)
6780 double_rq_unlock(rq
, p_rq
);
6782 local_irq_restore(flags
);
6789 EXPORT_SYMBOL_GPL(yield_to
);
6791 int io_schedule_prepare(void)
6793 int old_iowait
= current
->in_iowait
;
6795 current
->in_iowait
= 1;
6796 blk_schedule_flush_plug(current
);
6801 void io_schedule_finish(int token
)
6803 current
->in_iowait
= token
;
6807 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6808 * that process accounting knows that this is a task in IO wait state.
6810 long __sched
io_schedule_timeout(long timeout
)
6815 token
= io_schedule_prepare();
6816 ret
= schedule_timeout(timeout
);
6817 io_schedule_finish(token
);
6821 EXPORT_SYMBOL(io_schedule_timeout
);
6823 void __sched
io_schedule(void)
6827 token
= io_schedule_prepare();
6829 io_schedule_finish(token
);
6831 EXPORT_SYMBOL(io_schedule
);
6834 * sys_sched_get_priority_max - return maximum RT priority.
6835 * @policy: scheduling class.
6837 * Return: On success, this syscall returns the maximum
6838 * rt_priority that can be used by a given scheduling class.
6839 * On failure, a negative error code is returned.
6841 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6848 ret
= MAX_USER_RT_PRIO
-1;
6850 case SCHED_DEADLINE
:
6861 * sys_sched_get_priority_min - return minimum RT priority.
6862 * @policy: scheduling class.
6864 * Return: On success, this syscall returns the minimum
6865 * rt_priority that can be used by a given scheduling class.
6866 * On failure, a negative error code is returned.
6868 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6877 case SCHED_DEADLINE
:
6886 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
6888 struct task_struct
*p
;
6889 unsigned int time_slice
;
6899 p
= find_process_by_pid(pid
);
6903 retval
= security_task_getscheduler(p
);
6907 rq
= task_rq_lock(p
, &rf
);
6909 if (p
->sched_class
->get_rr_interval
)
6910 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
6911 task_rq_unlock(rq
, p
, &rf
);
6914 jiffies_to_timespec64(time_slice
, t
);
6923 * sys_sched_rr_get_interval - return the default timeslice of a process.
6924 * @pid: pid of the process.
6925 * @interval: userspace pointer to the timeslice value.
6927 * this syscall writes the default timeslice value of a given process
6928 * into the user-space timespec buffer. A value of '0' means infinity.
6930 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6933 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6934 struct __kernel_timespec __user
*, interval
)
6936 struct timespec64 t
;
6937 int retval
= sched_rr_get_interval(pid
, &t
);
6940 retval
= put_timespec64(&t
, interval
);
6945 #ifdef CONFIG_COMPAT_32BIT_TIME
6946 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
6947 struct old_timespec32 __user
*, interval
)
6949 struct timespec64 t
;
6950 int retval
= sched_rr_get_interval(pid
, &t
);
6953 retval
= put_old_timespec32(&t
, interval
);
6958 void sched_show_task(struct task_struct
*p
)
6960 unsigned long free
= 0;
6963 if (!try_get_task_stack(p
))
6966 pr_info("task:%-15.15s state:%c", p
->comm
, task_state_to_char(p
));
6968 if (p
->state
== TASK_RUNNING
)
6969 pr_cont(" running task ");
6970 #ifdef CONFIG_DEBUG_STACK_USAGE
6971 free
= stack_not_used(p
);
6976 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
6978 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
6979 free
, task_pid_nr(p
), ppid
,
6980 (unsigned long)task_thread_info(p
)->flags
);
6982 print_worker_info(KERN_INFO
, p
);
6983 print_stop_info(KERN_INFO
, p
);
6984 show_stack(p
, NULL
, KERN_INFO
);
6987 EXPORT_SYMBOL_GPL(sched_show_task
);
6990 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
6992 /* no filter, everything matches */
6996 /* filter, but doesn't match */
6997 if (!(p
->state
& state_filter
))
7001 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
7004 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
7011 void show_state_filter(unsigned long state_filter
)
7013 struct task_struct
*g
, *p
;
7016 for_each_process_thread(g
, p
) {
7018 * reset the NMI-timeout, listing all files on a slow
7019 * console might take a lot of time:
7020 * Also, reset softlockup watchdogs on all CPUs, because
7021 * another CPU might be blocked waiting for us to process
7024 touch_nmi_watchdog();
7025 touch_all_softlockup_watchdogs();
7026 if (state_filter_match(state_filter
, p
))
7030 #ifdef CONFIG_SCHED_DEBUG
7032 sysrq_sched_debug_show();
7036 * Only show locks if all tasks are dumped:
7039 debug_show_all_locks();
7043 * init_idle - set up an idle thread for a given CPU
7044 * @idle: task in question
7045 * @cpu: CPU the idle task belongs to
7047 * NOTE: this function does not set the idle thread's NEED_RESCHED
7048 * flag, to make booting more robust.
7050 void init_idle(struct task_struct
*idle
, int cpu
)
7052 struct rq
*rq
= cpu_rq(cpu
);
7053 unsigned long flags
;
7055 __sched_fork(0, idle
);
7057 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
7058 raw_spin_lock(&rq
->lock
);
7060 idle
->state
= TASK_RUNNING
;
7061 idle
->se
.exec_start
= sched_clock();
7062 idle
->flags
|= PF_IDLE
;
7064 scs_task_reset(idle
);
7065 kasan_unpoison_task_stack(idle
);
7069 * It's possible that init_idle() gets called multiple times on a task,
7070 * in that case do_set_cpus_allowed() will not do the right thing.
7072 * And since this is boot we can forgo the serialization.
7074 set_cpus_allowed_common(idle
, cpumask_of(cpu
), 0);
7077 * We're having a chicken and egg problem, even though we are
7078 * holding rq->lock, the CPU isn't yet set to this CPU so the
7079 * lockdep check in task_group() will fail.
7081 * Similar case to sched_fork(). / Alternatively we could
7082 * use task_rq_lock() here and obtain the other rq->lock.
7087 __set_task_cpu(idle
, cpu
);
7091 rcu_assign_pointer(rq
->curr
, idle
);
7092 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
7096 raw_spin_unlock(&rq
->lock
);
7097 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
7099 /* Set the preempt count _outside_ the spinlocks! */
7100 init_idle_preempt_count(idle
, cpu
);
7103 * The idle tasks have their own, simple scheduling class:
7105 idle
->sched_class
= &idle_sched_class
;
7106 ftrace_graph_init_idle_task(idle
, cpu
);
7107 vtime_init_idle(idle
, cpu
);
7109 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
7115 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
7116 const struct cpumask
*trial
)
7120 if (!cpumask_weight(cur
))
7123 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
7128 int task_can_attach(struct task_struct
*p
,
7129 const struct cpumask
*cs_cpus_allowed
)
7134 * Kthreads which disallow setaffinity shouldn't be moved
7135 * to a new cpuset; we don't want to change their CPU
7136 * affinity and isolating such threads by their set of
7137 * allowed nodes is unnecessary. Thus, cpusets are not
7138 * applicable for such threads. This prevents checking for
7139 * success of set_cpus_allowed_ptr() on all attached tasks
7140 * before cpus_mask may be changed.
7142 if (p
->flags
& PF_NO_SETAFFINITY
) {
7147 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
7149 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
7155 bool sched_smp_initialized __read_mostly
;
7157 #ifdef CONFIG_NUMA_BALANCING
7158 /* Migrate current task p to target_cpu */
7159 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
7161 struct migration_arg arg
= { p
, target_cpu
};
7162 int curr_cpu
= task_cpu(p
);
7164 if (curr_cpu
== target_cpu
)
7167 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
7170 /* TODO: This is not properly updating schedstats */
7172 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
7173 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
7177 * Requeue a task on a given node and accurately track the number of NUMA
7178 * tasks on the runqueues
7180 void sched_setnuma(struct task_struct
*p
, int nid
)
7182 bool queued
, running
;
7186 rq
= task_rq_lock(p
, &rf
);
7187 queued
= task_on_rq_queued(p
);
7188 running
= task_current(rq
, p
);
7191 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
7193 put_prev_task(rq
, p
);
7195 p
->numa_preferred_nid
= nid
;
7198 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
7200 set_next_task(rq
, p
);
7201 task_rq_unlock(rq
, p
, &rf
);
7203 #endif /* CONFIG_NUMA_BALANCING */
7205 #ifdef CONFIG_HOTPLUG_CPU
7207 * Ensure that the idle task is using init_mm right before its CPU goes
7210 void idle_task_exit(void)
7212 struct mm_struct
*mm
= current
->active_mm
;
7214 BUG_ON(cpu_online(smp_processor_id()));
7215 BUG_ON(current
!= this_rq()->idle
);
7217 if (mm
!= &init_mm
) {
7218 switch_mm(mm
, &init_mm
, current
);
7219 finish_arch_post_lock_switch();
7222 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
7225 static int __balance_push_cpu_stop(void *arg
)
7227 struct task_struct
*p
= arg
;
7228 struct rq
*rq
= this_rq();
7232 raw_spin_lock_irq(&p
->pi_lock
);
7235 update_rq_clock(rq
);
7237 if (task_rq(p
) == rq
&& task_on_rq_queued(p
)) {
7238 cpu
= select_fallback_rq(rq
->cpu
, p
);
7239 rq
= __migrate_task(rq
, &rf
, p
, cpu
);
7243 raw_spin_unlock_irq(&p
->pi_lock
);
7250 static DEFINE_PER_CPU(struct cpu_stop_work
, push_work
);
7253 * Ensure we only run per-cpu kthreads once the CPU goes !active.
7255 static void balance_push(struct rq
*rq
)
7257 struct task_struct
*push_task
= rq
->curr
;
7259 lockdep_assert_held(&rq
->lock
);
7260 SCHED_WARN_ON(rq
->cpu
!= smp_processor_id());
7263 * Both the cpu-hotplug and stop task are in this case and are
7264 * required to complete the hotplug process.
7266 if (is_per_cpu_kthread(push_task
) || is_migration_disabled(push_task
)) {
7268 * If this is the idle task on the outgoing CPU try to wake
7269 * up the hotplug control thread which might wait for the
7270 * last task to vanish. The rcuwait_active() check is
7271 * accurate here because the waiter is pinned on this CPU
7272 * and can't obviously be running in parallel.
7274 * On RT kernels this also has to check whether there are
7275 * pinned and scheduled out tasks on the runqueue. They
7276 * need to leave the migrate disabled section first.
7278 if (!rq
->nr_running
&& !rq_has_pinned_tasks(rq
) &&
7279 rcuwait_active(&rq
->hotplug_wait
)) {
7280 raw_spin_unlock(&rq
->lock
);
7281 rcuwait_wake_up(&rq
->hotplug_wait
);
7282 raw_spin_lock(&rq
->lock
);
7287 get_task_struct(push_task
);
7289 * Temporarily drop rq->lock such that we can wake-up the stop task.
7290 * Both preemption and IRQs are still disabled.
7292 raw_spin_unlock(&rq
->lock
);
7293 stop_one_cpu_nowait(rq
->cpu
, __balance_push_cpu_stop
, push_task
,
7294 this_cpu_ptr(&push_work
));
7296 * At this point need_resched() is true and we'll take the loop in
7297 * schedule(). The next pick is obviously going to be the stop task
7298 * which is_per_cpu_kthread() and will push this task away.
7300 raw_spin_lock(&rq
->lock
);
7303 static void balance_push_set(int cpu
, bool on
)
7305 struct rq
*rq
= cpu_rq(cpu
);
7308 rq_lock_irqsave(rq
, &rf
);
7310 rq
->balance_flags
|= BALANCE_PUSH
;
7312 rq
->balance_flags
&= ~BALANCE_PUSH
;
7313 rq_unlock_irqrestore(rq
, &rf
);
7317 * Invoked from a CPUs hotplug control thread after the CPU has been marked
7318 * inactive. All tasks which are not per CPU kernel threads are either
7319 * pushed off this CPU now via balance_push() or placed on a different CPU
7320 * during wakeup. Wait until the CPU is quiescent.
7322 static void balance_hotplug_wait(void)
7324 struct rq
*rq
= this_rq();
7326 rcuwait_wait_event(&rq
->hotplug_wait
,
7327 rq
->nr_running
== 1 && !rq_has_pinned_tasks(rq
),
7328 TASK_UNINTERRUPTIBLE
);
7333 static inline void balance_push(struct rq
*rq
)
7337 static inline void balance_push_set(int cpu
, bool on
)
7341 static inline void balance_hotplug_wait(void)
7345 #endif /* CONFIG_HOTPLUG_CPU */
7347 void set_rq_online(struct rq
*rq
)
7350 const struct sched_class
*class;
7352 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7355 for_each_class(class) {
7356 if (class->rq_online
)
7357 class->rq_online(rq
);
7362 void set_rq_offline(struct rq
*rq
)
7365 const struct sched_class
*class;
7367 for_each_class(class) {
7368 if (class->rq_offline
)
7369 class->rq_offline(rq
);
7372 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7378 * used to mark begin/end of suspend/resume:
7380 static int num_cpus_frozen
;
7383 * Update cpusets according to cpu_active mask. If cpusets are
7384 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7385 * around partition_sched_domains().
7387 * If we come here as part of a suspend/resume, don't touch cpusets because we
7388 * want to restore it back to its original state upon resume anyway.
7390 static void cpuset_cpu_active(void)
7392 if (cpuhp_tasks_frozen
) {
7394 * num_cpus_frozen tracks how many CPUs are involved in suspend
7395 * resume sequence. As long as this is not the last online
7396 * operation in the resume sequence, just build a single sched
7397 * domain, ignoring cpusets.
7399 partition_sched_domains(1, NULL
, NULL
);
7400 if (--num_cpus_frozen
)
7403 * This is the last CPU online operation. So fall through and
7404 * restore the original sched domains by considering the
7405 * cpuset configurations.
7407 cpuset_force_rebuild();
7409 cpuset_update_active_cpus();
7412 static int cpuset_cpu_inactive(unsigned int cpu
)
7414 if (!cpuhp_tasks_frozen
) {
7415 if (dl_cpu_busy(cpu
))
7417 cpuset_update_active_cpus();
7420 partition_sched_domains(1, NULL
, NULL
);
7425 int sched_cpu_activate(unsigned int cpu
)
7427 struct rq
*rq
= cpu_rq(cpu
);
7430 balance_push_set(cpu
, false);
7432 #ifdef CONFIG_SCHED_SMT
7434 * When going up, increment the number of cores with SMT present.
7436 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
7437 static_branch_inc_cpuslocked(&sched_smt_present
);
7439 set_cpu_active(cpu
, true);
7441 if (sched_smp_initialized
) {
7442 sched_domains_numa_masks_set(cpu
);
7443 cpuset_cpu_active();
7447 * Put the rq online, if not already. This happens:
7449 * 1) In the early boot process, because we build the real domains
7450 * after all CPUs have been brought up.
7452 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7455 rq_lock_irqsave(rq
, &rf
);
7457 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7460 rq_unlock_irqrestore(rq
, &rf
);
7465 int sched_cpu_deactivate(unsigned int cpu
)
7467 struct rq
*rq
= cpu_rq(cpu
);
7471 set_cpu_active(cpu
, false);
7473 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7474 * users of this state to go away such that all new such users will
7477 * Do sync before park smpboot threads to take care the rcu boost case.
7481 balance_push_set(cpu
, true);
7483 rq_lock_irqsave(rq
, &rf
);
7485 update_rq_clock(rq
);
7486 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7489 rq_unlock_irqrestore(rq
, &rf
);
7491 #ifdef CONFIG_SCHED_SMT
7493 * When going down, decrement the number of cores with SMT present.
7495 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
7496 static_branch_dec_cpuslocked(&sched_smt_present
);
7499 if (!sched_smp_initialized
)
7502 ret
= cpuset_cpu_inactive(cpu
);
7504 balance_push_set(cpu
, false);
7505 set_cpu_active(cpu
, true);
7508 sched_domains_numa_masks_clear(cpu
);
7512 static void sched_rq_cpu_starting(unsigned int cpu
)
7514 struct rq
*rq
= cpu_rq(cpu
);
7516 rq
->calc_load_update
= calc_load_update
;
7517 update_max_interval();
7520 int sched_cpu_starting(unsigned int cpu
)
7522 sched_rq_cpu_starting(cpu
);
7523 sched_tick_start(cpu
);
7527 #ifdef CONFIG_HOTPLUG_CPU
7530 * Invoked immediately before the stopper thread is invoked to bring the
7531 * CPU down completely. At this point all per CPU kthreads except the
7532 * hotplug thread (current) and the stopper thread (inactive) have been
7533 * either parked or have been unbound from the outgoing CPU. Ensure that
7534 * any of those which might be on the way out are gone.
7536 * If after this point a bound task is being woken on this CPU then the
7537 * responsible hotplug callback has failed to do it's job.
7538 * sched_cpu_dying() will catch it with the appropriate fireworks.
7540 int sched_cpu_wait_empty(unsigned int cpu
)
7542 balance_hotplug_wait();
7547 * Since this CPU is going 'away' for a while, fold any nr_active delta we
7548 * might have. Called from the CPU stopper task after ensuring that the
7549 * stopper is the last running task on the CPU, so nr_active count is
7550 * stable. We need to take the teardown thread which is calling this into
7551 * account, so we hand in adjust = 1 to the load calculation.
7553 * Also see the comment "Global load-average calculations".
7555 static void calc_load_migrate(struct rq
*rq
)
7557 long delta
= calc_load_fold_active(rq
, 1);
7560 atomic_long_add(delta
, &calc_load_tasks
);
7563 int sched_cpu_dying(unsigned int cpu
)
7565 struct rq
*rq
= cpu_rq(cpu
);
7568 /* Handle pending wakeups and then migrate everything off */
7569 sched_tick_stop(cpu
);
7571 rq_lock_irqsave(rq
, &rf
);
7572 BUG_ON(rq
->nr_running
!= 1 || rq_has_pinned_tasks(rq
));
7573 rq_unlock_irqrestore(rq
, &rf
);
7575 calc_load_migrate(rq
);
7576 update_max_interval();
7577 nohz_balance_exit_idle(rq
);
7583 void __init
sched_init_smp(void)
7588 * There's no userspace yet to cause hotplug operations; hence all the
7589 * CPU masks are stable and all blatant races in the below code cannot
7592 mutex_lock(&sched_domains_mutex
);
7593 sched_init_domains(cpu_active_mask
);
7594 mutex_unlock(&sched_domains_mutex
);
7596 /* Move init over to a non-isolated CPU */
7597 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
7599 sched_init_granularity();
7601 init_sched_rt_class();
7602 init_sched_dl_class();
7604 sched_smp_initialized
= true;
7607 static int __init
migration_init(void)
7609 sched_cpu_starting(smp_processor_id());
7612 early_initcall(migration_init
);
7615 void __init
sched_init_smp(void)
7617 sched_init_granularity();
7619 #endif /* CONFIG_SMP */
7621 int in_sched_functions(unsigned long addr
)
7623 return in_lock_functions(addr
) ||
7624 (addr
>= (unsigned long)__sched_text_start
7625 && addr
< (unsigned long)__sched_text_end
);
7628 #ifdef CONFIG_CGROUP_SCHED
7630 * Default task group.
7631 * Every task in system belongs to this group at bootup.
7633 struct task_group root_task_group
;
7634 LIST_HEAD(task_groups
);
7636 /* Cacheline aligned slab cache for task_group */
7637 static struct kmem_cache
*task_group_cache __read_mostly
;
7640 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7641 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
7643 void __init
sched_init(void)
7645 unsigned long ptr
= 0;
7648 /* Make sure the linker didn't screw up */
7649 BUG_ON(&idle_sched_class
+ 1 != &fair_sched_class
||
7650 &fair_sched_class
+ 1 != &rt_sched_class
||
7651 &rt_sched_class
+ 1 != &dl_sched_class
);
7653 BUG_ON(&dl_sched_class
+ 1 != &stop_sched_class
);
7658 #ifdef CONFIG_FAIR_GROUP_SCHED
7659 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
7661 #ifdef CONFIG_RT_GROUP_SCHED
7662 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
7665 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
7667 #ifdef CONFIG_FAIR_GROUP_SCHED
7668 root_task_group
.se
= (struct sched_entity
**)ptr
;
7669 ptr
+= nr_cpu_ids
* sizeof(void **);
7671 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7672 ptr
+= nr_cpu_ids
* sizeof(void **);
7674 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7675 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7676 #endif /* CONFIG_FAIR_GROUP_SCHED */
7677 #ifdef CONFIG_RT_GROUP_SCHED
7678 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7679 ptr
+= nr_cpu_ids
* sizeof(void **);
7681 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7682 ptr
+= nr_cpu_ids
* sizeof(void **);
7684 #endif /* CONFIG_RT_GROUP_SCHED */
7686 #ifdef CONFIG_CPUMASK_OFFSTACK
7687 for_each_possible_cpu(i
) {
7688 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7689 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7690 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7691 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7693 #endif /* CONFIG_CPUMASK_OFFSTACK */
7695 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
7696 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
7699 init_defrootdomain();
7702 #ifdef CONFIG_RT_GROUP_SCHED
7703 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7704 global_rt_period(), global_rt_runtime());
7705 #endif /* CONFIG_RT_GROUP_SCHED */
7707 #ifdef CONFIG_CGROUP_SCHED
7708 task_group_cache
= KMEM_CACHE(task_group
, 0);
7710 list_add(&root_task_group
.list
, &task_groups
);
7711 INIT_LIST_HEAD(&root_task_group
.children
);
7712 INIT_LIST_HEAD(&root_task_group
.siblings
);
7713 autogroup_init(&init_task
);
7714 #endif /* CONFIG_CGROUP_SCHED */
7716 for_each_possible_cpu(i
) {
7720 raw_spin_lock_init(&rq
->lock
);
7722 rq
->calc_load_active
= 0;
7723 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7724 init_cfs_rq(&rq
->cfs
);
7725 init_rt_rq(&rq
->rt
);
7726 init_dl_rq(&rq
->dl
);
7727 #ifdef CONFIG_FAIR_GROUP_SCHED
7728 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7729 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
7731 * How much CPU bandwidth does root_task_group get?
7733 * In case of task-groups formed thr' the cgroup filesystem, it
7734 * gets 100% of the CPU resources in the system. This overall
7735 * system CPU resource is divided among the tasks of
7736 * root_task_group and its child task-groups in a fair manner,
7737 * based on each entity's (task or task-group's) weight
7738 * (se->load.weight).
7740 * In other words, if root_task_group has 10 tasks of weight
7741 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7742 * then A0's share of the CPU resource is:
7744 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7746 * We achieve this by letting root_task_group's tasks sit
7747 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7749 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7750 #endif /* CONFIG_FAIR_GROUP_SCHED */
7752 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7753 #ifdef CONFIG_RT_GROUP_SCHED
7754 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7759 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7760 rq
->balance_callback
= NULL
;
7761 rq
->active_balance
= 0;
7762 rq
->next_balance
= jiffies
;
7767 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7768 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7770 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7772 rq_attach_root(rq
, &def_root_domain
);
7773 #ifdef CONFIG_NO_HZ_COMMON
7774 rq
->last_blocked_load_update_tick
= jiffies
;
7775 atomic_set(&rq
->nohz_flags
, 0);
7777 rq_csd_init(rq
, &rq
->nohz_csd
, nohz_csd_func
);
7779 #ifdef CONFIG_HOTPLUG_CPU
7780 rcuwait_init(&rq
->hotplug_wait
);
7782 #endif /* CONFIG_SMP */
7784 atomic_set(&rq
->nr_iowait
, 0);
7787 set_load_weight(&init_task
, false);
7790 * The boot idle thread does lazy MMU switching as well:
7793 enter_lazy_tlb(&init_mm
, current
);
7796 * Make us the idle thread. Technically, schedule() should not be
7797 * called from this thread, however somewhere below it might be,
7798 * but because we are the idle thread, we just pick up running again
7799 * when this runqueue becomes "idle".
7801 init_idle(current
, smp_processor_id());
7803 calc_load_update
= jiffies
+ LOAD_FREQ
;
7806 idle_thread_set_boot_cpu();
7808 init_sched_fair_class();
7816 scheduler_running
= 1;
7819 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7820 static inline int preempt_count_equals(int preempt_offset
)
7822 int nested
= preempt_count() + rcu_preempt_depth();
7824 return (nested
== preempt_offset
);
7827 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7830 * Blocking primitives will set (and therefore destroy) current->state,
7831 * since we will exit with TASK_RUNNING make sure we enter with it,
7832 * otherwise we will destroy state.
7834 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7835 "do not call blocking ops when !TASK_RUNNING; "
7836 "state=%lx set at [<%p>] %pS\n",
7838 (void *)current
->task_state_change
,
7839 (void *)current
->task_state_change
);
7841 ___might_sleep(file
, line
, preempt_offset
);
7843 EXPORT_SYMBOL(__might_sleep
);
7845 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7847 /* Ratelimiting timestamp: */
7848 static unsigned long prev_jiffy
;
7850 unsigned long preempt_disable_ip
;
7852 /* WARN_ON_ONCE() by default, no rate limit required: */
7855 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7856 !is_idle_task(current
) && !current
->non_block_count
) ||
7857 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
7861 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7863 prev_jiffy
= jiffies
;
7865 /* Save this before calling printk(), since that will clobber it: */
7866 preempt_disable_ip
= get_preempt_disable_ip(current
);
7869 "BUG: sleeping function called from invalid context at %s:%d\n",
7872 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
7873 in_atomic(), irqs_disabled(), current
->non_block_count
,
7874 current
->pid
, current
->comm
);
7876 if (task_stack_end_corrupted(current
))
7877 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7879 debug_show_held_locks(current
);
7880 if (irqs_disabled())
7881 print_irqtrace_events(current
);
7882 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
7883 && !preempt_count_equals(preempt_offset
)) {
7884 pr_err("Preemption disabled at:");
7885 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
7888 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7890 EXPORT_SYMBOL(___might_sleep
);
7892 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
7894 static unsigned long prev_jiffy
;
7896 if (irqs_disabled())
7899 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
7902 if (preempt_count() > preempt_offset
)
7905 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7907 prev_jiffy
= jiffies
;
7909 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
7910 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7911 in_atomic(), irqs_disabled(),
7912 current
->pid
, current
->comm
);
7914 debug_show_held_locks(current
);
7916 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7918 EXPORT_SYMBOL_GPL(__cant_sleep
);
7921 void __cant_migrate(const char *file
, int line
)
7923 static unsigned long prev_jiffy
;
7925 if (irqs_disabled())
7928 if (is_migration_disabled(current
))
7931 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
7934 if (preempt_count() > 0)
7937 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7939 prev_jiffy
= jiffies
;
7941 pr_err("BUG: assuming non migratable context at %s:%d\n", file
, line
);
7942 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
7943 in_atomic(), irqs_disabled(), is_migration_disabled(current
),
7944 current
->pid
, current
->comm
);
7946 debug_show_held_locks(current
);
7948 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7950 EXPORT_SYMBOL_GPL(__cant_migrate
);
7954 #ifdef CONFIG_MAGIC_SYSRQ
7955 void normalize_rt_tasks(void)
7957 struct task_struct
*g
, *p
;
7958 struct sched_attr attr
= {
7959 .sched_policy
= SCHED_NORMAL
,
7962 read_lock(&tasklist_lock
);
7963 for_each_process_thread(g
, p
) {
7965 * Only normalize user tasks:
7967 if (p
->flags
& PF_KTHREAD
)
7970 p
->se
.exec_start
= 0;
7971 schedstat_set(p
->se
.statistics
.wait_start
, 0);
7972 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
7973 schedstat_set(p
->se
.statistics
.block_start
, 0);
7975 if (!dl_task(p
) && !rt_task(p
)) {
7977 * Renice negative nice level userspace
7980 if (task_nice(p
) < 0)
7981 set_user_nice(p
, 0);
7985 __sched_setscheduler(p
, &attr
, false, false);
7987 read_unlock(&tasklist_lock
);
7990 #endif /* CONFIG_MAGIC_SYSRQ */
7992 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7994 * These functions are only useful for the IA64 MCA handling, or kdb.
7996 * They can only be called when the whole system has been
7997 * stopped - every CPU needs to be quiescent, and no scheduling
7998 * activity can take place. Using them for anything else would
7999 * be a serious bug, and as a result, they aren't even visible
8000 * under any other configuration.
8004 * curr_task - return the current task for a given CPU.
8005 * @cpu: the processor in question.
8007 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8009 * Return: The current task for @cpu.
8011 struct task_struct
*curr_task(int cpu
)
8013 return cpu_curr(cpu
);
8016 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8020 * ia64_set_curr_task - set the current task for a given CPU.
8021 * @cpu: the processor in question.
8022 * @p: the task pointer to set.
8024 * Description: This function must only be used when non-maskable interrupts
8025 * are serviced on a separate stack. It allows the architecture to switch the
8026 * notion of the current task on a CPU in a non-blocking manner. This function
8027 * must be called with all CPU's synchronized, and interrupts disabled, the
8028 * and caller must save the original value of the current task (see
8029 * curr_task() above) and restore that value before reenabling interrupts and
8030 * re-starting the system.
8032 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8034 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
8041 #ifdef CONFIG_CGROUP_SCHED
8042 /* task_group_lock serializes the addition/removal of task groups */
8043 static DEFINE_SPINLOCK(task_group_lock
);
8045 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
8046 struct task_group
*parent
)
8048 #ifdef CONFIG_UCLAMP_TASK_GROUP
8049 enum uclamp_id clamp_id
;
8051 for_each_clamp_id(clamp_id
) {
8052 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
8053 uclamp_none(clamp_id
), false);
8054 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
8059 static void sched_free_group(struct task_group
*tg
)
8061 free_fair_sched_group(tg
);
8062 free_rt_sched_group(tg
);
8064 kmem_cache_free(task_group_cache
, tg
);
8067 /* allocate runqueue etc for a new task group */
8068 struct task_group
*sched_create_group(struct task_group
*parent
)
8070 struct task_group
*tg
;
8072 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
8074 return ERR_PTR(-ENOMEM
);
8076 if (!alloc_fair_sched_group(tg
, parent
))
8079 if (!alloc_rt_sched_group(tg
, parent
))
8082 alloc_uclamp_sched_group(tg
, parent
);
8087 sched_free_group(tg
);
8088 return ERR_PTR(-ENOMEM
);
8091 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
8093 unsigned long flags
;
8095 spin_lock_irqsave(&task_group_lock
, flags
);
8096 list_add_rcu(&tg
->list
, &task_groups
);
8098 /* Root should already exist: */
8101 tg
->parent
= parent
;
8102 INIT_LIST_HEAD(&tg
->children
);
8103 list_add_rcu(&tg
->siblings
, &parent
->children
);
8104 spin_unlock_irqrestore(&task_group_lock
, flags
);
8106 online_fair_sched_group(tg
);
8109 /* rcu callback to free various structures associated with a task group */
8110 static void sched_free_group_rcu(struct rcu_head
*rhp
)
8112 /* Now it should be safe to free those cfs_rqs: */
8113 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
8116 void sched_destroy_group(struct task_group
*tg
)
8118 /* Wait for possible concurrent references to cfs_rqs complete: */
8119 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
8122 void sched_offline_group(struct task_group
*tg
)
8124 unsigned long flags
;
8126 /* End participation in shares distribution: */
8127 unregister_fair_sched_group(tg
);
8129 spin_lock_irqsave(&task_group_lock
, flags
);
8130 list_del_rcu(&tg
->list
);
8131 list_del_rcu(&tg
->siblings
);
8132 spin_unlock_irqrestore(&task_group_lock
, flags
);
8135 static void sched_change_group(struct task_struct
*tsk
, int type
)
8137 struct task_group
*tg
;
8140 * All callers are synchronized by task_rq_lock(); we do not use RCU
8141 * which is pointless here. Thus, we pass "true" to task_css_check()
8142 * to prevent lockdep warnings.
8144 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
8145 struct task_group
, css
);
8146 tg
= autogroup_task_group(tsk
, tg
);
8147 tsk
->sched_task_group
= tg
;
8149 #ifdef CONFIG_FAIR_GROUP_SCHED
8150 if (tsk
->sched_class
->task_change_group
)
8151 tsk
->sched_class
->task_change_group(tsk
, type
);
8154 set_task_rq(tsk
, task_cpu(tsk
));
8158 * Change task's runqueue when it moves between groups.
8160 * The caller of this function should have put the task in its new group by
8161 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8164 void sched_move_task(struct task_struct
*tsk
)
8166 int queued
, running
, queue_flags
=
8167 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
8171 rq
= task_rq_lock(tsk
, &rf
);
8172 update_rq_clock(rq
);
8174 running
= task_current(rq
, tsk
);
8175 queued
= task_on_rq_queued(tsk
);
8178 dequeue_task(rq
, tsk
, queue_flags
);
8180 put_prev_task(rq
, tsk
);
8182 sched_change_group(tsk
, TASK_MOVE_GROUP
);
8185 enqueue_task(rq
, tsk
, queue_flags
);
8187 set_next_task(rq
, tsk
);
8189 * After changing group, the running task may have joined a
8190 * throttled one but it's still the running task. Trigger a
8191 * resched to make sure that task can still run.
8196 task_rq_unlock(rq
, tsk
, &rf
);
8199 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8201 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8204 static struct cgroup_subsys_state
*
8205 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8207 struct task_group
*parent
= css_tg(parent_css
);
8208 struct task_group
*tg
;
8211 /* This is early initialization for the top cgroup */
8212 return &root_task_group
.css
;
8215 tg
= sched_create_group(parent
);
8217 return ERR_PTR(-ENOMEM
);
8222 /* Expose task group only after completing cgroup initialization */
8223 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8225 struct task_group
*tg
= css_tg(css
);
8226 struct task_group
*parent
= css_tg(css
->parent
);
8229 sched_online_group(tg
, parent
);
8231 #ifdef CONFIG_UCLAMP_TASK_GROUP
8232 /* Propagate the effective uclamp value for the new group */
8233 cpu_util_update_eff(css
);
8239 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8241 struct task_group
*tg
= css_tg(css
);
8243 sched_offline_group(tg
);
8246 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8248 struct task_group
*tg
= css_tg(css
);
8251 * Relies on the RCU grace period between css_released() and this.
8253 sched_free_group(tg
);
8257 * This is called before wake_up_new_task(), therefore we really only
8258 * have to set its group bits, all the other stuff does not apply.
8260 static void cpu_cgroup_fork(struct task_struct
*task
)
8265 rq
= task_rq_lock(task
, &rf
);
8267 update_rq_clock(rq
);
8268 sched_change_group(task
, TASK_SET_GROUP
);
8270 task_rq_unlock(rq
, task
, &rf
);
8273 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8275 struct task_struct
*task
;
8276 struct cgroup_subsys_state
*css
;
8279 cgroup_taskset_for_each(task
, css
, tset
) {
8280 #ifdef CONFIG_RT_GROUP_SCHED
8281 if (!sched_rt_can_attach(css_tg(css
), task
))
8285 * Serialize against wake_up_new_task() such that if it's
8286 * running, we're sure to observe its full state.
8288 raw_spin_lock_irq(&task
->pi_lock
);
8290 * Avoid calling sched_move_task() before wake_up_new_task()
8291 * has happened. This would lead to problems with PELT, due to
8292 * move wanting to detach+attach while we're not attached yet.
8294 if (task
->state
== TASK_NEW
)
8296 raw_spin_unlock_irq(&task
->pi_lock
);
8304 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8306 struct task_struct
*task
;
8307 struct cgroup_subsys_state
*css
;
8309 cgroup_taskset_for_each(task
, css
, tset
)
8310 sched_move_task(task
);
8313 #ifdef CONFIG_UCLAMP_TASK_GROUP
8314 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
8316 struct cgroup_subsys_state
*top_css
= css
;
8317 struct uclamp_se
*uc_parent
= NULL
;
8318 struct uclamp_se
*uc_se
= NULL
;
8319 unsigned int eff
[UCLAMP_CNT
];
8320 enum uclamp_id clamp_id
;
8321 unsigned int clamps
;
8323 css_for_each_descendant_pre(css
, top_css
) {
8324 uc_parent
= css_tg(css
)->parent
8325 ? css_tg(css
)->parent
->uclamp
: NULL
;
8327 for_each_clamp_id(clamp_id
) {
8328 /* Assume effective clamps matches requested clamps */
8329 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
8330 /* Cap effective clamps with parent's effective clamps */
8332 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
8333 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
8336 /* Ensure protection is always capped by limit */
8337 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
8339 /* Propagate most restrictive effective clamps */
8341 uc_se
= css_tg(css
)->uclamp
;
8342 for_each_clamp_id(clamp_id
) {
8343 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
8345 uc_se
[clamp_id
].value
= eff
[clamp_id
];
8346 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
8347 clamps
|= (0x1 << clamp_id
);
8350 css
= css_rightmost_descendant(css
);
8354 /* Immediately update descendants RUNNABLE tasks */
8355 uclamp_update_active_tasks(css
, clamps
);
8360 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
8361 * C expression. Since there is no way to convert a macro argument (N) into a
8362 * character constant, use two levels of macros.
8364 #define _POW10(exp) ((unsigned int)1e##exp)
8365 #define POW10(exp) _POW10(exp)
8367 struct uclamp_request
{
8368 #define UCLAMP_PERCENT_SHIFT 2
8369 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
8375 static inline struct uclamp_request
8376 capacity_from_percent(char *buf
)
8378 struct uclamp_request req
= {
8379 .percent
= UCLAMP_PERCENT_SCALE
,
8380 .util
= SCHED_CAPACITY_SCALE
,
8385 if (strcmp(buf
, "max")) {
8386 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
8390 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
8395 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
8396 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
8402 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
8403 size_t nbytes
, loff_t off
,
8404 enum uclamp_id clamp_id
)
8406 struct uclamp_request req
;
8407 struct task_group
*tg
;
8409 req
= capacity_from_percent(buf
);
8413 static_branch_enable(&sched_uclamp_used
);
8415 mutex_lock(&uclamp_mutex
);
8418 tg
= css_tg(of_css(of
));
8419 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
8420 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
8423 * Because of not recoverable conversion rounding we keep track of the
8424 * exact requested value
8426 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
8428 /* Update effective clamps to track the most restrictive value */
8429 cpu_util_update_eff(of_css(of
));
8432 mutex_unlock(&uclamp_mutex
);
8437 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
8438 char *buf
, size_t nbytes
,
8441 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
8444 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
8445 char *buf
, size_t nbytes
,
8448 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
8451 static inline void cpu_uclamp_print(struct seq_file
*sf
,
8452 enum uclamp_id clamp_id
)
8454 struct task_group
*tg
;
8460 tg
= css_tg(seq_css(sf
));
8461 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
8464 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
8465 seq_puts(sf
, "max\n");
8469 percent
= tg
->uclamp_pct
[clamp_id
];
8470 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
8471 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
8474 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
8476 cpu_uclamp_print(sf
, UCLAMP_MIN
);
8480 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
8482 cpu_uclamp_print(sf
, UCLAMP_MAX
);
8485 #endif /* CONFIG_UCLAMP_TASK_GROUP */
8487 #ifdef CONFIG_FAIR_GROUP_SCHED
8488 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8489 struct cftype
*cftype
, u64 shareval
)
8491 if (shareval
> scale_load_down(ULONG_MAX
))
8492 shareval
= MAX_SHARES
;
8493 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8496 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8499 struct task_group
*tg
= css_tg(css
);
8501 return (u64
) scale_load_down(tg
->shares
);
8504 #ifdef CONFIG_CFS_BANDWIDTH
8505 static DEFINE_MUTEX(cfs_constraints_mutex
);
8507 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8508 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8509 /* More than 203 days if BW_SHIFT equals 20. */
8510 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
8512 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8514 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8516 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8517 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8519 if (tg
== &root_task_group
)
8523 * Ensure we have at some amount of bandwidth every period. This is
8524 * to prevent reaching a state of large arrears when throttled via
8525 * entity_tick() resulting in prolonged exit starvation.
8527 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8531 * Likewise, bound things on the otherside by preventing insane quota
8532 * periods. This also allows us to normalize in computing quota
8535 if (period
> max_cfs_quota_period
)
8539 * Bound quota to defend quota against overflow during bandwidth shift.
8541 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
8545 * Prevent race between setting of cfs_rq->runtime_enabled and
8546 * unthrottle_offline_cfs_rqs().
8549 mutex_lock(&cfs_constraints_mutex
);
8550 ret
= __cfs_schedulable(tg
, period
, quota
);
8554 runtime_enabled
= quota
!= RUNTIME_INF
;
8555 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8557 * If we need to toggle cfs_bandwidth_used, off->on must occur
8558 * before making related changes, and on->off must occur afterwards
8560 if (runtime_enabled
&& !runtime_was_enabled
)
8561 cfs_bandwidth_usage_inc();
8562 raw_spin_lock_irq(&cfs_b
->lock
);
8563 cfs_b
->period
= ns_to_ktime(period
);
8564 cfs_b
->quota
= quota
;
8566 __refill_cfs_bandwidth_runtime(cfs_b
);
8568 /* Restart the period timer (if active) to handle new period expiry: */
8569 if (runtime_enabled
)
8570 start_cfs_bandwidth(cfs_b
);
8572 raw_spin_unlock_irq(&cfs_b
->lock
);
8574 for_each_online_cpu(i
) {
8575 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8576 struct rq
*rq
= cfs_rq
->rq
;
8579 rq_lock_irq(rq
, &rf
);
8580 cfs_rq
->runtime_enabled
= runtime_enabled
;
8581 cfs_rq
->runtime_remaining
= 0;
8583 if (cfs_rq
->throttled
)
8584 unthrottle_cfs_rq(cfs_rq
);
8585 rq_unlock_irq(rq
, &rf
);
8587 if (runtime_was_enabled
&& !runtime_enabled
)
8588 cfs_bandwidth_usage_dec();
8590 mutex_unlock(&cfs_constraints_mutex
);
8596 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8600 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8601 if (cfs_quota_us
< 0)
8602 quota
= RUNTIME_INF
;
8603 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
8604 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8608 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8611 static long tg_get_cfs_quota(struct task_group
*tg
)
8615 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8618 quota_us
= tg
->cfs_bandwidth
.quota
;
8619 do_div(quota_us
, NSEC_PER_USEC
);
8624 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8628 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
8631 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8632 quota
= tg
->cfs_bandwidth
.quota
;
8634 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8637 static long tg_get_cfs_period(struct task_group
*tg
)
8641 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8642 do_div(cfs_period_us
, NSEC_PER_USEC
);
8644 return cfs_period_us
;
8647 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8650 return tg_get_cfs_quota(css_tg(css
));
8653 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8654 struct cftype
*cftype
, s64 cfs_quota_us
)
8656 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8659 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8662 return tg_get_cfs_period(css_tg(css
));
8665 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8666 struct cftype
*cftype
, u64 cfs_period_us
)
8668 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8671 struct cfs_schedulable_data
{
8672 struct task_group
*tg
;
8677 * normalize group quota/period to be quota/max_period
8678 * note: units are usecs
8680 static u64
normalize_cfs_quota(struct task_group
*tg
,
8681 struct cfs_schedulable_data
*d
)
8689 period
= tg_get_cfs_period(tg
);
8690 quota
= tg_get_cfs_quota(tg
);
8693 /* note: these should typically be equivalent */
8694 if (quota
== RUNTIME_INF
|| quota
== -1)
8697 return to_ratio(period
, quota
);
8700 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8702 struct cfs_schedulable_data
*d
= data
;
8703 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8704 s64 quota
= 0, parent_quota
= -1;
8707 quota
= RUNTIME_INF
;
8709 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8711 quota
= normalize_cfs_quota(tg
, d
);
8712 parent_quota
= parent_b
->hierarchical_quota
;
8715 * Ensure max(child_quota) <= parent_quota. On cgroup2,
8716 * always take the min. On cgroup1, only inherit when no
8719 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
8720 quota
= min(quota
, parent_quota
);
8722 if (quota
== RUNTIME_INF
)
8723 quota
= parent_quota
;
8724 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8728 cfs_b
->hierarchical_quota
= quota
;
8733 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8736 struct cfs_schedulable_data data
= {
8742 if (quota
!= RUNTIME_INF
) {
8743 do_div(data
.period
, NSEC_PER_USEC
);
8744 do_div(data
.quota
, NSEC_PER_USEC
);
8748 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8754 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
8756 struct task_group
*tg
= css_tg(seq_css(sf
));
8757 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8759 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8760 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8761 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8763 if (schedstat_enabled() && tg
!= &root_task_group
) {
8767 for_each_possible_cpu(i
)
8768 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
8770 seq_printf(sf
, "wait_sum %llu\n", ws
);
8775 #endif /* CONFIG_CFS_BANDWIDTH */
8776 #endif /* CONFIG_FAIR_GROUP_SCHED */
8778 #ifdef CONFIG_RT_GROUP_SCHED
8779 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8780 struct cftype
*cft
, s64 val
)
8782 return sched_group_set_rt_runtime(css_tg(css
), val
);
8785 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8788 return sched_group_rt_runtime(css_tg(css
));
8791 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8792 struct cftype
*cftype
, u64 rt_period_us
)
8794 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8797 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8800 return sched_group_rt_period(css_tg(css
));
8802 #endif /* CONFIG_RT_GROUP_SCHED */
8804 static struct cftype cpu_legacy_files
[] = {
8805 #ifdef CONFIG_FAIR_GROUP_SCHED
8808 .read_u64
= cpu_shares_read_u64
,
8809 .write_u64
= cpu_shares_write_u64
,
8812 #ifdef CONFIG_CFS_BANDWIDTH
8814 .name
= "cfs_quota_us",
8815 .read_s64
= cpu_cfs_quota_read_s64
,
8816 .write_s64
= cpu_cfs_quota_write_s64
,
8819 .name
= "cfs_period_us",
8820 .read_u64
= cpu_cfs_period_read_u64
,
8821 .write_u64
= cpu_cfs_period_write_u64
,
8825 .seq_show
= cpu_cfs_stat_show
,
8828 #ifdef CONFIG_RT_GROUP_SCHED
8830 .name
= "rt_runtime_us",
8831 .read_s64
= cpu_rt_runtime_read
,
8832 .write_s64
= cpu_rt_runtime_write
,
8835 .name
= "rt_period_us",
8836 .read_u64
= cpu_rt_period_read_uint
,
8837 .write_u64
= cpu_rt_period_write_uint
,
8840 #ifdef CONFIG_UCLAMP_TASK_GROUP
8842 .name
= "uclamp.min",
8843 .flags
= CFTYPE_NOT_ON_ROOT
,
8844 .seq_show
= cpu_uclamp_min_show
,
8845 .write
= cpu_uclamp_min_write
,
8848 .name
= "uclamp.max",
8849 .flags
= CFTYPE_NOT_ON_ROOT
,
8850 .seq_show
= cpu_uclamp_max_show
,
8851 .write
= cpu_uclamp_max_write
,
8857 static int cpu_extra_stat_show(struct seq_file
*sf
,
8858 struct cgroup_subsys_state
*css
)
8860 #ifdef CONFIG_CFS_BANDWIDTH
8862 struct task_group
*tg
= css_tg(css
);
8863 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8866 throttled_usec
= cfs_b
->throttled_time
;
8867 do_div(throttled_usec
, NSEC_PER_USEC
);
8869 seq_printf(sf
, "nr_periods %d\n"
8871 "throttled_usec %llu\n",
8872 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
8879 #ifdef CONFIG_FAIR_GROUP_SCHED
8880 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
8883 struct task_group
*tg
= css_tg(css
);
8884 u64 weight
= scale_load_down(tg
->shares
);
8886 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
8889 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
8890 struct cftype
*cft
, u64 weight
)
8893 * cgroup weight knobs should use the common MIN, DFL and MAX
8894 * values which are 1, 100 and 10000 respectively. While it loses
8895 * a bit of range on both ends, it maps pretty well onto the shares
8896 * value used by scheduler and the round-trip conversions preserve
8897 * the original value over the entire range.
8899 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
8902 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
8904 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
8907 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
8910 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
8911 int last_delta
= INT_MAX
;
8914 /* find the closest nice value to the current weight */
8915 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
8916 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
8917 if (delta
>= last_delta
)
8922 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
8925 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
8926 struct cftype
*cft
, s64 nice
)
8928 unsigned long weight
;
8931 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
8934 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
8935 idx
= array_index_nospec(idx
, 40);
8936 weight
= sched_prio_to_weight
[idx
];
8938 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
8942 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
8943 long period
, long quota
)
8946 seq_puts(sf
, "max");
8948 seq_printf(sf
, "%ld", quota
);
8950 seq_printf(sf
, " %ld\n", period
);
8953 /* caller should put the current value in *@periodp before calling */
8954 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
8955 u64
*periodp
, u64
*quotap
)
8957 char tok
[21]; /* U64_MAX */
8959 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
8962 *periodp
*= NSEC_PER_USEC
;
8964 if (sscanf(tok
, "%llu", quotap
))
8965 *quotap
*= NSEC_PER_USEC
;
8966 else if (!strcmp(tok
, "max"))
8967 *quotap
= RUNTIME_INF
;
8974 #ifdef CONFIG_CFS_BANDWIDTH
8975 static int cpu_max_show(struct seq_file
*sf
, void *v
)
8977 struct task_group
*tg
= css_tg(seq_css(sf
));
8979 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
8983 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
8984 char *buf
, size_t nbytes
, loff_t off
)
8986 struct task_group
*tg
= css_tg(of_css(of
));
8987 u64 period
= tg_get_cfs_period(tg
);
8991 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
8993 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
8994 return ret
?: nbytes
;
8998 static struct cftype cpu_files
[] = {
8999 #ifdef CONFIG_FAIR_GROUP_SCHED
9002 .flags
= CFTYPE_NOT_ON_ROOT
,
9003 .read_u64
= cpu_weight_read_u64
,
9004 .write_u64
= cpu_weight_write_u64
,
9007 .name
= "weight.nice",
9008 .flags
= CFTYPE_NOT_ON_ROOT
,
9009 .read_s64
= cpu_weight_nice_read_s64
,
9010 .write_s64
= cpu_weight_nice_write_s64
,
9013 #ifdef CONFIG_CFS_BANDWIDTH
9016 .flags
= CFTYPE_NOT_ON_ROOT
,
9017 .seq_show
= cpu_max_show
,
9018 .write
= cpu_max_write
,
9021 #ifdef CONFIG_UCLAMP_TASK_GROUP
9023 .name
= "uclamp.min",
9024 .flags
= CFTYPE_NOT_ON_ROOT
,
9025 .seq_show
= cpu_uclamp_min_show
,
9026 .write
= cpu_uclamp_min_write
,
9029 .name
= "uclamp.max",
9030 .flags
= CFTYPE_NOT_ON_ROOT
,
9031 .seq_show
= cpu_uclamp_max_show
,
9032 .write
= cpu_uclamp_max_write
,
9038 struct cgroup_subsys cpu_cgrp_subsys
= {
9039 .css_alloc
= cpu_cgroup_css_alloc
,
9040 .css_online
= cpu_cgroup_css_online
,
9041 .css_released
= cpu_cgroup_css_released
,
9042 .css_free
= cpu_cgroup_css_free
,
9043 .css_extra_stat_show
= cpu_extra_stat_show
,
9044 .fork
= cpu_cgroup_fork
,
9045 .can_attach
= cpu_cgroup_can_attach
,
9046 .attach
= cpu_cgroup_attach
,
9047 .legacy_cftypes
= cpu_legacy_files
,
9048 .dfl_cftypes
= cpu_files
,
9053 #endif /* CONFIG_CGROUP_SCHED */
9055 void dump_cpu_task(int cpu
)
9057 pr_info("Task dump for CPU %d:\n", cpu
);
9058 sched_show_task(cpu_curr(cpu
));
9062 * Nice levels are multiplicative, with a gentle 10% change for every
9063 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
9064 * nice 1, it will get ~10% less CPU time than another CPU-bound task
9065 * that remained on nice 0.
9067 * The "10% effect" is relative and cumulative: from _any_ nice level,
9068 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
9069 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
9070 * If a task goes up by ~10% and another task goes down by ~10% then
9071 * the relative distance between them is ~25%.)
9073 const int sched_prio_to_weight
[40] = {
9074 /* -20 */ 88761, 71755, 56483, 46273, 36291,
9075 /* -15 */ 29154, 23254, 18705, 14949, 11916,
9076 /* -10 */ 9548, 7620, 6100, 4904, 3906,
9077 /* -5 */ 3121, 2501, 1991, 1586, 1277,
9078 /* 0 */ 1024, 820, 655, 526, 423,
9079 /* 5 */ 335, 272, 215, 172, 137,
9080 /* 10 */ 110, 87, 70, 56, 45,
9081 /* 15 */ 36, 29, 23, 18, 15,
9085 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
9087 * In cases where the weight does not change often, we can use the
9088 * precalculated inverse to speed up arithmetics by turning divisions
9089 * into multiplications:
9091 const u32 sched_prio_to_wmult
[40] = {
9092 /* -20 */ 48388, 59856, 76040, 92818, 118348,
9093 /* -15 */ 147320, 184698, 229616, 287308, 360437,
9094 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
9095 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
9096 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
9097 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
9098 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
9099 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
9102 void call_trace_sched_update_nr_running(struct rq
*rq
, int count
)
9104 trace_sched_update_nr_running_tp(rq
, count
);