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
);
323 #ifdef CONFIG_SCHED_HRTICK
325 * Use HR-timers to deliver accurate preemption points.
328 static void hrtick_clear(struct rq
*rq
)
330 if (hrtimer_active(&rq
->hrtick_timer
))
331 hrtimer_cancel(&rq
->hrtick_timer
);
335 * High-resolution timer tick.
336 * Runs from hardirq context with interrupts disabled.
338 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
340 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
343 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
347 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
350 return HRTIMER_NORESTART
;
355 static void __hrtick_restart(struct rq
*rq
)
357 struct hrtimer
*timer
= &rq
->hrtick_timer
;
359 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
363 * called from hardirq (IPI) context
365 static void __hrtick_start(void *arg
)
371 __hrtick_restart(rq
);
376 * Called to set the hrtick timer state.
378 * called with rq->lock held and irqs disabled
380 void hrtick_start(struct rq
*rq
, u64 delay
)
382 struct hrtimer
*timer
= &rq
->hrtick_timer
;
387 * Don't schedule slices shorter than 10000ns, that just
388 * doesn't make sense and can cause timer DoS.
390 delta
= max_t(s64
, delay
, 10000LL);
391 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
393 hrtimer_set_expires(timer
, time
);
396 __hrtick_restart(rq
);
398 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
403 * Called to set the hrtick timer state.
405 * called with rq->lock held and irqs disabled
407 void hrtick_start(struct rq
*rq
, u64 delay
)
410 * Don't schedule slices shorter than 10000ns, that just
411 * doesn't make sense. Rely on vruntime for fairness.
413 delay
= max_t(u64
, delay
, 10000LL);
414 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
415 HRTIMER_MODE_REL_PINNED_HARD
);
418 #endif /* CONFIG_SMP */
420 static void hrtick_rq_init(struct rq
*rq
)
423 INIT_CSD(&rq
->hrtick_csd
, __hrtick_start
, rq
);
425 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
426 rq
->hrtick_timer
.function
= hrtick
;
428 #else /* CONFIG_SCHED_HRTICK */
429 static inline void hrtick_clear(struct rq
*rq
)
433 static inline void hrtick_rq_init(struct rq
*rq
)
436 #endif /* CONFIG_SCHED_HRTICK */
439 * cmpxchg based fetch_or, macro so it works for different integer types
441 #define fetch_or(ptr, mask) \
443 typeof(ptr) _ptr = (ptr); \
444 typeof(mask) _mask = (mask); \
445 typeof(*_ptr) _old, _val = *_ptr; \
448 _old = cmpxchg(_ptr, _val, _val | _mask); \
456 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
458 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
459 * this avoids any races wrt polling state changes and thereby avoids
462 static bool set_nr_and_not_polling(struct task_struct
*p
)
464 struct thread_info
*ti
= task_thread_info(p
);
465 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
469 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
471 * If this returns true, then the idle task promises to call
472 * sched_ttwu_pending() and reschedule soon.
474 static bool set_nr_if_polling(struct task_struct
*p
)
476 struct thread_info
*ti
= task_thread_info(p
);
477 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
480 if (!(val
& _TIF_POLLING_NRFLAG
))
482 if (val
& _TIF_NEED_RESCHED
)
484 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
493 static bool set_nr_and_not_polling(struct task_struct
*p
)
495 set_tsk_need_resched(p
);
500 static bool set_nr_if_polling(struct task_struct
*p
)
507 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
509 struct wake_q_node
*node
= &task
->wake_q
;
512 * Atomically grab the task, if ->wake_q is !nil already it means
513 * it's already queued (either by us or someone else) and will get the
514 * wakeup due to that.
516 * In order to ensure that a pending wakeup will observe our pending
517 * state, even in the failed case, an explicit smp_mb() must be used.
519 smp_mb__before_atomic();
520 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
524 * The head is context local, there can be no concurrency.
527 head
->lastp
= &node
->next
;
532 * wake_q_add() - queue a wakeup for 'later' waking.
533 * @head: the wake_q_head to add @task to
534 * @task: the task to queue for 'later' wakeup
536 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
537 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
540 * This function must be used as-if it were wake_up_process(); IOW the task
541 * must be ready to be woken at this location.
543 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
545 if (__wake_q_add(head
, task
))
546 get_task_struct(task
);
550 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
551 * @head: the wake_q_head to add @task to
552 * @task: the task to queue for 'later' wakeup
554 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
555 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
558 * This function must be used as-if it were wake_up_process(); IOW the task
559 * must be ready to be woken at this location.
561 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
562 * that already hold reference to @task can call the 'safe' version and trust
563 * wake_q to do the right thing depending whether or not the @task is already
566 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
568 if (!__wake_q_add(head
, task
))
569 put_task_struct(task
);
572 void wake_up_q(struct wake_q_head
*head
)
574 struct wake_q_node
*node
= head
->first
;
576 while (node
!= WAKE_Q_TAIL
) {
577 struct task_struct
*task
;
579 task
= container_of(node
, struct task_struct
, wake_q
);
581 /* Task can safely be re-inserted now: */
583 task
->wake_q
.next
= NULL
;
586 * wake_up_process() executes a full barrier, which pairs with
587 * the queueing in wake_q_add() so as not to miss wakeups.
589 wake_up_process(task
);
590 put_task_struct(task
);
595 * resched_curr - mark rq's current task 'to be rescheduled now'.
597 * On UP this means the setting of the need_resched flag, on SMP it
598 * might also involve a cross-CPU call to trigger the scheduler on
601 void resched_curr(struct rq
*rq
)
603 struct task_struct
*curr
= rq
->curr
;
606 lockdep_assert_held(&rq
->lock
);
608 if (test_tsk_need_resched(curr
))
613 if (cpu
== smp_processor_id()) {
614 set_tsk_need_resched(curr
);
615 set_preempt_need_resched();
619 if (set_nr_and_not_polling(curr
))
620 smp_send_reschedule(cpu
);
622 trace_sched_wake_idle_without_ipi(cpu
);
625 void resched_cpu(int cpu
)
627 struct rq
*rq
= cpu_rq(cpu
);
630 raw_spin_lock_irqsave(&rq
->lock
, flags
);
631 if (cpu_online(cpu
) || cpu
== smp_processor_id())
633 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
637 #ifdef CONFIG_NO_HZ_COMMON
639 * In the semi idle case, use the nearest busy CPU for migrating timers
640 * from an idle CPU. This is good for power-savings.
642 * We don't do similar optimization for completely idle system, as
643 * selecting an idle CPU will add more delays to the timers than intended
644 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
646 int get_nohz_timer_target(void)
648 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
649 struct sched_domain
*sd
;
651 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
658 for_each_domain(cpu
, sd
) {
659 for_each_cpu_and(i
, sched_domain_span(sd
),
660 housekeeping_cpumask(HK_FLAG_TIMER
)) {
671 if (default_cpu
== -1)
672 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
680 * When add_timer_on() enqueues a timer into the timer wheel of an
681 * idle CPU then this timer might expire before the next timer event
682 * which is scheduled to wake up that CPU. In case of a completely
683 * idle system the next event might even be infinite time into the
684 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
685 * leaves the inner idle loop so the newly added timer is taken into
686 * account when the CPU goes back to idle and evaluates the timer
687 * wheel for the next timer event.
689 static void wake_up_idle_cpu(int cpu
)
691 struct rq
*rq
= cpu_rq(cpu
);
693 if (cpu
== smp_processor_id())
696 if (set_nr_and_not_polling(rq
->idle
))
697 smp_send_reschedule(cpu
);
699 trace_sched_wake_idle_without_ipi(cpu
);
702 static bool wake_up_full_nohz_cpu(int cpu
)
705 * We just need the target to call irq_exit() and re-evaluate
706 * the next tick. The nohz full kick at least implies that.
707 * If needed we can still optimize that later with an
710 if (cpu_is_offline(cpu
))
711 return true; /* Don't try to wake offline CPUs. */
712 if (tick_nohz_full_cpu(cpu
)) {
713 if (cpu
!= smp_processor_id() ||
714 tick_nohz_tick_stopped())
715 tick_nohz_full_kick_cpu(cpu
);
723 * Wake up the specified CPU. If the CPU is going offline, it is the
724 * caller's responsibility to deal with the lost wakeup, for example,
725 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
727 void wake_up_nohz_cpu(int cpu
)
729 if (!wake_up_full_nohz_cpu(cpu
))
730 wake_up_idle_cpu(cpu
);
733 static void nohz_csd_func(void *info
)
735 struct rq
*rq
= info
;
736 int cpu
= cpu_of(rq
);
740 * Release the rq::nohz_csd.
742 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
743 WARN_ON(!(flags
& NOHZ_KICK_MASK
));
745 rq
->idle_balance
= idle_cpu(cpu
);
746 if (rq
->idle_balance
&& !need_resched()) {
747 rq
->nohz_idle_balance
= flags
;
748 raise_softirq_irqoff(SCHED_SOFTIRQ
);
752 #endif /* CONFIG_NO_HZ_COMMON */
754 #ifdef CONFIG_NO_HZ_FULL
755 bool sched_can_stop_tick(struct rq
*rq
)
759 /* Deadline tasks, even if single, need the tick */
760 if (rq
->dl
.dl_nr_running
)
764 * If there are more than one RR tasks, we need the tick to affect the
765 * actual RR behaviour.
767 if (rq
->rt
.rr_nr_running
) {
768 if (rq
->rt
.rr_nr_running
== 1)
775 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
776 * forced preemption between FIFO tasks.
778 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
783 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
784 * if there's more than one we need the tick for involuntary
787 if (rq
->nr_running
> 1)
792 #endif /* CONFIG_NO_HZ_FULL */
793 #endif /* CONFIG_SMP */
795 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
796 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
798 * Iterate task_group tree rooted at *from, calling @down when first entering a
799 * node and @up when leaving it for the final time.
801 * Caller must hold rcu_lock or sufficient equivalent.
803 int walk_tg_tree_from(struct task_group
*from
,
804 tg_visitor down
, tg_visitor up
, void *data
)
806 struct task_group
*parent
, *child
;
812 ret
= (*down
)(parent
, data
);
815 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
822 ret
= (*up
)(parent
, data
);
823 if (ret
|| parent
== from
)
827 parent
= parent
->parent
;
834 int tg_nop(struct task_group
*tg
, void *data
)
840 static void set_load_weight(struct task_struct
*p
, bool update_load
)
842 int prio
= p
->static_prio
- MAX_RT_PRIO
;
843 struct load_weight
*load
= &p
->se
.load
;
846 * SCHED_IDLE tasks get minimal weight:
848 if (task_has_idle_policy(p
)) {
849 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
850 load
->inv_weight
= WMULT_IDLEPRIO
;
855 * SCHED_OTHER tasks have to update their load when changing their
858 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
859 reweight_task(p
, prio
);
861 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
862 load
->inv_weight
= sched_prio_to_wmult
[prio
];
866 #ifdef CONFIG_UCLAMP_TASK
868 * Serializes updates of utilization clamp values
870 * The (slow-path) user-space triggers utilization clamp value updates which
871 * can require updates on (fast-path) scheduler's data structures used to
872 * support enqueue/dequeue operations.
873 * While the per-CPU rq lock protects fast-path update operations, user-space
874 * requests are serialized using a mutex to reduce the risk of conflicting
875 * updates or API abuses.
877 static DEFINE_MUTEX(uclamp_mutex
);
879 /* Max allowed minimum utilization */
880 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
882 /* Max allowed maximum utilization */
883 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
886 * By default RT tasks run at the maximum performance point/capacity of the
887 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
888 * SCHED_CAPACITY_SCALE.
890 * This knob allows admins to change the default behavior when uclamp is being
891 * used. In battery powered devices, particularly, running at the maximum
892 * capacity and frequency will increase energy consumption and shorten the
895 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
897 * This knob will not override the system default sched_util_clamp_min defined
900 unsigned int sysctl_sched_uclamp_util_min_rt_default
= SCHED_CAPACITY_SCALE
;
902 /* All clamps are required to be less or equal than these values */
903 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
906 * This static key is used to reduce the uclamp overhead in the fast path. It
907 * primarily disables the call to uclamp_rq_{inc, dec}() in
908 * enqueue/dequeue_task().
910 * This allows users to continue to enable uclamp in their kernel config with
911 * minimum uclamp overhead in the fast path.
913 * As soon as userspace modifies any of the uclamp knobs, the static key is
914 * enabled, since we have an actual users that make use of uclamp
917 * The knobs that would enable this static key are:
919 * * A task modifying its uclamp value with sched_setattr().
920 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
921 * * An admin modifying the cgroup cpu.uclamp.{min, max}
923 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used
);
925 /* Integer rounded range for each bucket */
926 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
928 #define for_each_clamp_id(clamp_id) \
929 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
931 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
933 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
936 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
938 if (clamp_id
== UCLAMP_MIN
)
940 return SCHED_CAPACITY_SCALE
;
943 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
944 unsigned int value
, bool user_defined
)
946 uc_se
->value
= value
;
947 uc_se
->bucket_id
= uclamp_bucket_id(value
);
948 uc_se
->user_defined
= user_defined
;
951 static inline unsigned int
952 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
953 unsigned int clamp_value
)
956 * Avoid blocked utilization pushing up the frequency when we go
957 * idle (which drops the max-clamp) by retaining the last known
960 if (clamp_id
== UCLAMP_MAX
) {
961 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
965 return uclamp_none(UCLAMP_MIN
);
968 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
969 unsigned int clamp_value
)
971 /* Reset max-clamp retention only on idle exit */
972 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
975 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
979 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
980 unsigned int clamp_value
)
982 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
983 int bucket_id
= UCLAMP_BUCKETS
- 1;
986 * Since both min and max clamps are max aggregated, find the
987 * top most bucket with tasks in.
989 for ( ; bucket_id
>= 0; bucket_id
--) {
990 if (!bucket
[bucket_id
].tasks
)
992 return bucket
[bucket_id
].value
;
995 /* No tasks -- default clamp values */
996 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
999 static void __uclamp_update_util_min_rt_default(struct task_struct
*p
)
1001 unsigned int default_util_min
;
1002 struct uclamp_se
*uc_se
;
1004 lockdep_assert_held(&p
->pi_lock
);
1006 uc_se
= &p
->uclamp_req
[UCLAMP_MIN
];
1008 /* Only sync if user didn't override the default */
1009 if (uc_se
->user_defined
)
1012 default_util_min
= sysctl_sched_uclamp_util_min_rt_default
;
1013 uclamp_se_set(uc_se
, default_util_min
, false);
1016 static void uclamp_update_util_min_rt_default(struct task_struct
*p
)
1024 /* Protect updates to p->uclamp_* */
1025 rq
= task_rq_lock(p
, &rf
);
1026 __uclamp_update_util_min_rt_default(p
);
1027 task_rq_unlock(rq
, p
, &rf
);
1030 static void uclamp_sync_util_min_rt_default(void)
1032 struct task_struct
*g
, *p
;
1035 * copy_process() sysctl_uclamp
1036 * uclamp_min_rt = X;
1037 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1038 * // link thread smp_mb__after_spinlock()
1039 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1040 * sched_post_fork() for_each_process_thread()
1041 * __uclamp_sync_rt() __uclamp_sync_rt()
1043 * Ensures that either sched_post_fork() will observe the new
1044 * uclamp_min_rt or for_each_process_thread() will observe the new
1047 read_lock(&tasklist_lock
);
1048 smp_mb__after_spinlock();
1049 read_unlock(&tasklist_lock
);
1052 for_each_process_thread(g
, p
)
1053 uclamp_update_util_min_rt_default(p
);
1057 static inline struct uclamp_se
1058 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
1060 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
1061 #ifdef CONFIG_UCLAMP_TASK_GROUP
1062 struct uclamp_se uc_max
;
1065 * Tasks in autogroups or root task group will be
1066 * restricted by system defaults.
1068 if (task_group_is_autogroup(task_group(p
)))
1070 if (task_group(p
) == &root_task_group
)
1073 uc_max
= task_group(p
)->uclamp
[clamp_id
];
1074 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
1082 * The effective clamp bucket index of a task depends on, by increasing
1084 * - the task specific clamp value, when explicitly requested from userspace
1085 * - the task group effective clamp value, for tasks not either in the root
1086 * group or in an autogroup
1087 * - the system default clamp value, defined by the sysadmin
1089 static inline struct uclamp_se
1090 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
1092 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
1093 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
1095 /* System default restrictions always apply */
1096 if (unlikely(uc_req
.value
> uc_max
.value
))
1102 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
1104 struct uclamp_se uc_eff
;
1106 /* Task currently refcounted: use back-annotated (effective) value */
1107 if (p
->uclamp
[clamp_id
].active
)
1108 return (unsigned long)p
->uclamp
[clamp_id
].value
;
1110 uc_eff
= uclamp_eff_get(p
, clamp_id
);
1112 return (unsigned long)uc_eff
.value
;
1116 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1117 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1118 * updates the rq's clamp value if required.
1120 * Tasks can have a task-specific value requested from user-space, track
1121 * within each bucket the maximum value for tasks refcounted in it.
1122 * This "local max aggregation" allows to track the exact "requested" value
1123 * for each bucket when all its RUNNABLE tasks require the same clamp.
1125 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
1126 enum uclamp_id clamp_id
)
1128 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1129 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1130 struct uclamp_bucket
*bucket
;
1132 lockdep_assert_held(&rq
->lock
);
1134 /* Update task effective clamp */
1135 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
1137 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1139 uc_se
->active
= true;
1141 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
1144 * Local max aggregation: rq buckets always track the max
1145 * "requested" clamp value of its RUNNABLE tasks.
1147 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
1148 bucket
->value
= uc_se
->value
;
1150 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
1151 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
1155 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1156 * is released. If this is the last task reference counting the rq's max
1157 * active clamp value, then the rq's clamp value is updated.
1159 * Both refcounted tasks and rq's cached clamp values are expected to be
1160 * always valid. If it's detected they are not, as defensive programming,
1161 * enforce the expected state and warn.
1163 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
1164 enum uclamp_id clamp_id
)
1166 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1167 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1168 struct uclamp_bucket
*bucket
;
1169 unsigned int bkt_clamp
;
1170 unsigned int rq_clamp
;
1172 lockdep_assert_held(&rq
->lock
);
1175 * If sched_uclamp_used was enabled after task @p was enqueued,
1176 * we could end up with unbalanced call to uclamp_rq_dec_id().
1178 * In this case the uc_se->active flag should be false since no uclamp
1179 * accounting was performed at enqueue time and we can just return
1182 * Need to be careful of the following enqueue/dequeue ordering
1186 * // sched_uclamp_used gets enabled
1189 * // Must not decrement bucket->tasks here
1192 * where we could end up with stale data in uc_se and
1193 * bucket[uc_se->bucket_id].
1195 * The following check here eliminates the possibility of such race.
1197 if (unlikely(!uc_se
->active
))
1200 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1202 SCHED_WARN_ON(!bucket
->tasks
);
1203 if (likely(bucket
->tasks
))
1206 uc_se
->active
= false;
1209 * Keep "local max aggregation" simple and accept to (possibly)
1210 * overboost some RUNNABLE tasks in the same bucket.
1211 * The rq clamp bucket value is reset to its base value whenever
1212 * there are no more RUNNABLE tasks refcounting it.
1214 if (likely(bucket
->tasks
))
1217 rq_clamp
= READ_ONCE(uc_rq
->value
);
1219 * Defensive programming: this should never happen. If it happens,
1220 * e.g. due to future modification, warn and fixup the expected value.
1222 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1223 if (bucket
->value
>= rq_clamp
) {
1224 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1225 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1229 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1231 enum uclamp_id clamp_id
;
1234 * Avoid any overhead until uclamp is actually used by the userspace.
1236 * The condition is constructed such that a NOP is generated when
1237 * sched_uclamp_used is disabled.
1239 if (!static_branch_unlikely(&sched_uclamp_used
))
1242 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1245 for_each_clamp_id(clamp_id
)
1246 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1248 /* Reset clamp idle holding when there is one RUNNABLE task */
1249 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1250 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1253 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1255 enum uclamp_id clamp_id
;
1258 * Avoid any overhead until uclamp is actually used by the userspace.
1260 * The condition is constructed such that a NOP is generated when
1261 * sched_uclamp_used is disabled.
1263 if (!static_branch_unlikely(&sched_uclamp_used
))
1266 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1269 for_each_clamp_id(clamp_id
)
1270 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1274 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1280 * Lock the task and the rq where the task is (or was) queued.
1282 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1283 * price to pay to safely serialize util_{min,max} updates with
1284 * enqueues, dequeues and migration operations.
1285 * This is the same locking schema used by __set_cpus_allowed_ptr().
1287 rq
= task_rq_lock(p
, &rf
);
1290 * Setting the clamp bucket is serialized by task_rq_lock().
1291 * If the task is not yet RUNNABLE and its task_struct is not
1292 * affecting a valid clamp bucket, the next time it's enqueued,
1293 * it will already see the updated clamp bucket value.
1295 if (p
->uclamp
[clamp_id
].active
) {
1296 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1297 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1300 task_rq_unlock(rq
, p
, &rf
);
1303 #ifdef CONFIG_UCLAMP_TASK_GROUP
1305 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1306 unsigned int clamps
)
1308 enum uclamp_id clamp_id
;
1309 struct css_task_iter it
;
1310 struct task_struct
*p
;
1312 css_task_iter_start(css
, 0, &it
);
1313 while ((p
= css_task_iter_next(&it
))) {
1314 for_each_clamp_id(clamp_id
) {
1315 if ((0x1 << clamp_id
) & clamps
)
1316 uclamp_update_active(p
, clamp_id
);
1319 css_task_iter_end(&it
);
1322 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1323 static void uclamp_update_root_tg(void)
1325 struct task_group
*tg
= &root_task_group
;
1327 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1328 sysctl_sched_uclamp_util_min
, false);
1329 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1330 sysctl_sched_uclamp_util_max
, false);
1333 cpu_util_update_eff(&root_task_group
.css
);
1337 static void uclamp_update_root_tg(void) { }
1340 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1341 void *buffer
, size_t *lenp
, loff_t
*ppos
)
1343 bool update_root_tg
= false;
1344 int old_min
, old_max
, old_min_rt
;
1347 mutex_lock(&uclamp_mutex
);
1348 old_min
= sysctl_sched_uclamp_util_min
;
1349 old_max
= sysctl_sched_uclamp_util_max
;
1350 old_min_rt
= sysctl_sched_uclamp_util_min_rt_default
;
1352 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1358 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1359 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
||
1360 sysctl_sched_uclamp_util_min_rt_default
> SCHED_CAPACITY_SCALE
) {
1366 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1367 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1368 sysctl_sched_uclamp_util_min
, false);
1369 update_root_tg
= true;
1371 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1372 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1373 sysctl_sched_uclamp_util_max
, false);
1374 update_root_tg
= true;
1377 if (update_root_tg
) {
1378 static_branch_enable(&sched_uclamp_used
);
1379 uclamp_update_root_tg();
1382 if (old_min_rt
!= sysctl_sched_uclamp_util_min_rt_default
) {
1383 static_branch_enable(&sched_uclamp_used
);
1384 uclamp_sync_util_min_rt_default();
1388 * We update all RUNNABLE tasks only when task groups are in use.
1389 * Otherwise, keep it simple and do just a lazy update at each next
1390 * task enqueue time.
1396 sysctl_sched_uclamp_util_min
= old_min
;
1397 sysctl_sched_uclamp_util_max
= old_max
;
1398 sysctl_sched_uclamp_util_min_rt_default
= old_min_rt
;
1400 mutex_unlock(&uclamp_mutex
);
1405 static int uclamp_validate(struct task_struct
*p
,
1406 const struct sched_attr
*attr
)
1408 int util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1409 int util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1411 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1412 util_min
= attr
->sched_util_min
;
1414 if (util_min
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1418 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1419 util_max
= attr
->sched_util_max
;
1421 if (util_max
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1425 if (util_min
!= -1 && util_max
!= -1 && util_min
> util_max
)
1429 * We have valid uclamp attributes; make sure uclamp is enabled.
1431 * We need to do that here, because enabling static branches is a
1432 * blocking operation which obviously cannot be done while holding
1435 static_branch_enable(&sched_uclamp_used
);
1440 static bool uclamp_reset(const struct sched_attr
*attr
,
1441 enum uclamp_id clamp_id
,
1442 struct uclamp_se
*uc_se
)
1444 /* Reset on sched class change for a non user-defined clamp value. */
1445 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)) &&
1446 !uc_se
->user_defined
)
1449 /* Reset on sched_util_{min,max} == -1. */
1450 if (clamp_id
== UCLAMP_MIN
&&
1451 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1452 attr
->sched_util_min
== -1) {
1456 if (clamp_id
== UCLAMP_MAX
&&
1457 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1458 attr
->sched_util_max
== -1) {
1465 static void __setscheduler_uclamp(struct task_struct
*p
,
1466 const struct sched_attr
*attr
)
1468 enum uclamp_id clamp_id
;
1470 for_each_clamp_id(clamp_id
) {
1471 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1474 if (!uclamp_reset(attr
, clamp_id
, uc_se
))
1478 * RT by default have a 100% boost value that could be modified
1481 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1482 value
= sysctl_sched_uclamp_util_min_rt_default
;
1484 value
= uclamp_none(clamp_id
);
1486 uclamp_se_set(uc_se
, value
, false);
1490 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1493 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1494 attr
->sched_util_min
!= -1) {
1495 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1496 attr
->sched_util_min
, true);
1499 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1500 attr
->sched_util_max
!= -1) {
1501 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1502 attr
->sched_util_max
, true);
1506 static void uclamp_fork(struct task_struct
*p
)
1508 enum uclamp_id clamp_id
;
1511 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1512 * as the task is still at its early fork stages.
1514 for_each_clamp_id(clamp_id
)
1515 p
->uclamp
[clamp_id
].active
= false;
1517 if (likely(!p
->sched_reset_on_fork
))
1520 for_each_clamp_id(clamp_id
) {
1521 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1522 uclamp_none(clamp_id
), false);
1526 static void uclamp_post_fork(struct task_struct
*p
)
1528 uclamp_update_util_min_rt_default(p
);
1531 static void __init
init_uclamp_rq(struct rq
*rq
)
1533 enum uclamp_id clamp_id
;
1534 struct uclamp_rq
*uc_rq
= rq
->uclamp
;
1536 for_each_clamp_id(clamp_id
) {
1537 uc_rq
[clamp_id
] = (struct uclamp_rq
) {
1538 .value
= uclamp_none(clamp_id
)
1542 rq
->uclamp_flags
= 0;
1545 static void __init
init_uclamp(void)
1547 struct uclamp_se uc_max
= {};
1548 enum uclamp_id clamp_id
;
1551 for_each_possible_cpu(cpu
)
1552 init_uclamp_rq(cpu_rq(cpu
));
1554 for_each_clamp_id(clamp_id
) {
1555 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1556 uclamp_none(clamp_id
), false);
1559 /* System defaults allow max clamp values for both indexes */
1560 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1561 for_each_clamp_id(clamp_id
) {
1562 uclamp_default
[clamp_id
] = uc_max
;
1563 #ifdef CONFIG_UCLAMP_TASK_GROUP
1564 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1565 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1570 #else /* CONFIG_UCLAMP_TASK */
1571 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1572 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1573 static inline int uclamp_validate(struct task_struct
*p
,
1574 const struct sched_attr
*attr
)
1578 static void __setscheduler_uclamp(struct task_struct
*p
,
1579 const struct sched_attr
*attr
) { }
1580 static inline void uclamp_fork(struct task_struct
*p
) { }
1581 static inline void uclamp_post_fork(struct task_struct
*p
) { }
1582 static inline void init_uclamp(void) { }
1583 #endif /* CONFIG_UCLAMP_TASK */
1585 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1587 if (!(flags
& ENQUEUE_NOCLOCK
))
1588 update_rq_clock(rq
);
1590 if (!(flags
& ENQUEUE_RESTORE
)) {
1591 sched_info_queued(rq
, p
);
1592 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1595 uclamp_rq_inc(rq
, p
);
1596 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1599 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1601 if (!(flags
& DEQUEUE_NOCLOCK
))
1602 update_rq_clock(rq
);
1604 if (!(flags
& DEQUEUE_SAVE
)) {
1605 sched_info_dequeued(rq
, p
);
1606 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1609 uclamp_rq_dec(rq
, p
);
1610 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1613 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1615 enqueue_task(rq
, p
, flags
);
1617 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1620 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1622 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1624 dequeue_task(rq
, p
, flags
);
1628 * __normal_prio - return the priority that is based on the static prio
1630 static inline int __normal_prio(struct task_struct
*p
)
1632 return p
->static_prio
;
1636 * Calculate the expected normal priority: i.e. priority
1637 * without taking RT-inheritance into account. Might be
1638 * boosted by interactivity modifiers. Changes upon fork,
1639 * setprio syscalls, and whenever the interactivity
1640 * estimator recalculates.
1642 static inline int normal_prio(struct task_struct
*p
)
1646 if (task_has_dl_policy(p
))
1647 prio
= MAX_DL_PRIO
-1;
1648 else if (task_has_rt_policy(p
))
1649 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1651 prio
= __normal_prio(p
);
1656 * Calculate the current priority, i.e. the priority
1657 * taken into account by the scheduler. This value might
1658 * be boosted by RT tasks, or might be boosted by
1659 * interactivity modifiers. Will be RT if the task got
1660 * RT-boosted. If not then it returns p->normal_prio.
1662 static int effective_prio(struct task_struct
*p
)
1664 p
->normal_prio
= normal_prio(p
);
1666 * If we are RT tasks or we were boosted to RT priority,
1667 * keep the priority unchanged. Otherwise, update priority
1668 * to the normal priority:
1670 if (!rt_prio(p
->prio
))
1671 return p
->normal_prio
;
1676 * task_curr - is this task currently executing on a CPU?
1677 * @p: the task in question.
1679 * Return: 1 if the task is currently executing. 0 otherwise.
1681 inline int task_curr(const struct task_struct
*p
)
1683 return cpu_curr(task_cpu(p
)) == p
;
1687 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1688 * use the balance_callback list if you want balancing.
1690 * this means any call to check_class_changed() must be followed by a call to
1691 * balance_callback().
1693 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1694 const struct sched_class
*prev_class
,
1697 if (prev_class
!= p
->sched_class
) {
1698 if (prev_class
->switched_from
)
1699 prev_class
->switched_from(rq
, p
);
1701 p
->sched_class
->switched_to(rq
, p
);
1702 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1703 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1706 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1708 if (p
->sched_class
== rq
->curr
->sched_class
)
1709 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1710 else if (p
->sched_class
> rq
->curr
->sched_class
)
1714 * A queue event has occurred, and we're going to schedule. In
1715 * this case, we can save a useless back to back clock update.
1717 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1718 rq_clock_skip_update(rq
);
1724 __do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
);
1726 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1727 const struct cpumask
*new_mask
,
1730 static void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
)
1732 if (likely(!p
->migration_disabled
))
1735 if (p
->cpus_ptr
!= &p
->cpus_mask
)
1739 * Violates locking rules! see comment in __do_set_cpus_allowed().
1741 __do_set_cpus_allowed(p
, cpumask_of(rq
->cpu
), SCA_MIGRATE_DISABLE
);
1744 void migrate_disable(void)
1746 struct task_struct
*p
= current
;
1748 if (p
->migration_disabled
) {
1749 p
->migration_disabled
++;
1754 this_rq()->nr_pinned
++;
1755 p
->migration_disabled
= 1;
1758 EXPORT_SYMBOL_GPL(migrate_disable
);
1760 void migrate_enable(void)
1762 struct task_struct
*p
= current
;
1764 if (p
->migration_disabled
> 1) {
1765 p
->migration_disabled
--;
1770 * Ensure stop_task runs either before or after this, and that
1771 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
1774 if (p
->cpus_ptr
!= &p
->cpus_mask
)
1775 __set_cpus_allowed_ptr(p
, &p
->cpus_mask
, SCA_MIGRATE_ENABLE
);
1777 * Mustn't clear migration_disabled() until cpus_ptr points back at the
1778 * regular cpus_mask, otherwise things that race (eg.
1779 * select_fallback_rq) get confused.
1782 p
->migration_disabled
= 0;
1783 this_rq()->nr_pinned
--;
1786 EXPORT_SYMBOL_GPL(migrate_enable
);
1788 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
1790 return rq
->nr_pinned
;
1794 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1795 * __set_cpus_allowed_ptr() and select_fallback_rq().
1797 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1799 /* When not in the task's cpumask, no point in looking further. */
1800 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1803 /* migrate_disabled() must be allowed to finish. */
1804 if (is_migration_disabled(p
))
1805 return cpu_online(cpu
);
1807 /* Non kernel threads are not allowed during either online or offline. */
1808 if (!(p
->flags
& PF_KTHREAD
))
1809 return cpu_active(cpu
);
1811 /* KTHREAD_IS_PER_CPU is always allowed. */
1812 if (kthread_is_per_cpu(p
))
1813 return cpu_online(cpu
);
1815 /* Regular kernel threads don't get to stay during offline. */
1816 if (cpu_rq(cpu
)->balance_push
)
1819 /* But are allowed during online. */
1820 return cpu_online(cpu
);
1824 * This is how migration works:
1826 * 1) we invoke migration_cpu_stop() on the target CPU using
1828 * 2) stopper starts to run (implicitly forcing the migrated thread
1830 * 3) it checks whether the migrated task is still in the wrong runqueue.
1831 * 4) if it's in the wrong runqueue then the migration thread removes
1832 * it and puts it into the right queue.
1833 * 5) stopper completes and stop_one_cpu() returns and the migration
1838 * move_queued_task - move a queued task to new rq.
1840 * Returns (locked) new rq. Old rq's lock is released.
1842 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1843 struct task_struct
*p
, int new_cpu
)
1845 lockdep_assert_held(&rq
->lock
);
1847 deactivate_task(rq
, p
, DEQUEUE_NOCLOCK
);
1848 set_task_cpu(p
, new_cpu
);
1851 rq
= cpu_rq(new_cpu
);
1854 BUG_ON(task_cpu(p
) != new_cpu
);
1855 activate_task(rq
, p
, 0);
1856 check_preempt_curr(rq
, p
, 0);
1861 struct migration_arg
{
1862 struct task_struct
*task
;
1864 struct set_affinity_pending
*pending
;
1867 struct set_affinity_pending
{
1869 struct completion done
;
1870 struct cpu_stop_work stop_work
;
1871 struct migration_arg arg
;
1875 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1876 * this because either it can't run here any more (set_cpus_allowed()
1877 * away from this CPU, or CPU going down), or because we're
1878 * attempting to rebalance this task on exec (sched_exec).
1880 * So we race with normal scheduler movements, but that's OK, as long
1881 * as the task is no longer on this CPU.
1883 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1884 struct task_struct
*p
, int dest_cpu
)
1886 /* Affinity changed (again). */
1887 if (!is_cpu_allowed(p
, dest_cpu
))
1890 update_rq_clock(rq
);
1891 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1897 * migration_cpu_stop - this will be executed by a highprio stopper thread
1898 * and performs thread migration by bumping thread off CPU then
1899 * 'pushing' onto another runqueue.
1901 static int migration_cpu_stop(void *data
)
1903 struct set_affinity_pending
*pending
;
1904 struct migration_arg
*arg
= data
;
1905 struct task_struct
*p
= arg
->task
;
1906 int dest_cpu
= arg
->dest_cpu
;
1907 struct rq
*rq
= this_rq();
1908 bool complete
= false;
1912 * The original target CPU might have gone down and we might
1913 * be on another CPU but it doesn't matter.
1915 local_irq_save(rf
.flags
);
1917 * We need to explicitly wake pending tasks before running
1918 * __migrate_task() such that we will not miss enforcing cpus_ptr
1919 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1921 flush_smp_call_function_from_idle();
1923 raw_spin_lock(&p
->pi_lock
);
1926 pending
= p
->migration_pending
;
1928 * If task_rq(p) != rq, it cannot be migrated here, because we're
1929 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1930 * we're holding p->pi_lock.
1932 if (task_rq(p
) == rq
) {
1933 if (is_migration_disabled(p
))
1937 p
->migration_pending
= NULL
;
1941 /* migrate_enable() -- we must not race against SCA */
1944 * When this was migrate_enable() but we no longer
1945 * have a @pending, a concurrent SCA 'fixed' things
1946 * and we should be valid again. Nothing to do.
1949 WARN_ON_ONCE(!cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
));
1953 dest_cpu
= cpumask_any_distribute(&p
->cpus_mask
);
1956 if (task_on_rq_queued(p
))
1957 rq
= __migrate_task(rq
, &rf
, p
, dest_cpu
);
1959 p
->wake_cpu
= dest_cpu
;
1961 } else if (dest_cpu
< 0 || pending
) {
1963 * This happens when we get migrated between migrate_enable()'s
1964 * preempt_enable() and scheduling the stopper task. At that
1965 * point we're a regular task again and not current anymore.
1967 * A !PREEMPT kernel has a giant hole here, which makes it far
1972 * The task moved before the stopper got to run. We're holding
1973 * ->pi_lock, so the allowed mask is stable - if it got
1974 * somewhere allowed, we're done.
1976 if (pending
&& cpumask_test_cpu(task_cpu(p
), p
->cpus_ptr
)) {
1977 p
->migration_pending
= NULL
;
1983 * When this was migrate_enable() but we no longer have an
1984 * @pending, a concurrent SCA 'fixed' things and we should be
1985 * valid again. Nothing to do.
1988 WARN_ON_ONCE(!cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
));
1993 * When migrate_enable() hits a rq mis-match we can't reliably
1994 * determine is_migration_disabled() and so have to chase after
1997 task_rq_unlock(rq
, p
, &rf
);
1998 stop_one_cpu_nowait(task_cpu(p
), migration_cpu_stop
,
1999 &pending
->arg
, &pending
->stop_work
);
2003 task_rq_unlock(rq
, p
, &rf
);
2006 complete_all(&pending
->done
);
2008 /* For pending->{arg,stop_work} */
2009 pending
= arg
->pending
;
2010 if (pending
&& refcount_dec_and_test(&pending
->refs
))
2011 wake_up_var(&pending
->refs
);
2016 int push_cpu_stop(void *arg
)
2018 struct rq
*lowest_rq
= NULL
, *rq
= this_rq();
2019 struct task_struct
*p
= arg
;
2021 raw_spin_lock_irq(&p
->pi_lock
);
2022 raw_spin_lock(&rq
->lock
);
2024 if (task_rq(p
) != rq
)
2027 if (is_migration_disabled(p
)) {
2028 p
->migration_flags
|= MDF_PUSH
;
2032 p
->migration_flags
&= ~MDF_PUSH
;
2034 if (p
->sched_class
->find_lock_rq
)
2035 lowest_rq
= p
->sched_class
->find_lock_rq(p
, rq
);
2040 // XXX validate p is still the highest prio task
2041 if (task_rq(p
) == rq
) {
2042 deactivate_task(rq
, p
, 0);
2043 set_task_cpu(p
, lowest_rq
->cpu
);
2044 activate_task(lowest_rq
, p
, 0);
2045 resched_curr(lowest_rq
);
2048 double_unlock_balance(rq
, lowest_rq
);
2051 rq
->push_busy
= false;
2052 raw_spin_unlock(&rq
->lock
);
2053 raw_spin_unlock_irq(&p
->pi_lock
);
2060 * sched_class::set_cpus_allowed must do the below, but is not required to
2061 * actually call this function.
2063 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
)
2065 if (flags
& (SCA_MIGRATE_ENABLE
| SCA_MIGRATE_DISABLE
)) {
2066 p
->cpus_ptr
= new_mask
;
2070 cpumask_copy(&p
->cpus_mask
, new_mask
);
2071 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
2075 __do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
)
2077 struct rq
*rq
= task_rq(p
);
2078 bool queued
, running
;
2081 * This here violates the locking rules for affinity, since we're only
2082 * supposed to change these variables while holding both rq->lock and
2085 * HOWEVER, it magically works, because ttwu() is the only code that
2086 * accesses these variables under p->pi_lock and only does so after
2087 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2088 * before finish_task().
2090 * XXX do further audits, this smells like something putrid.
2092 if (flags
& SCA_MIGRATE_DISABLE
)
2093 SCHED_WARN_ON(!p
->on_cpu
);
2095 lockdep_assert_held(&p
->pi_lock
);
2097 queued
= task_on_rq_queued(p
);
2098 running
= task_current(rq
, p
);
2102 * Because __kthread_bind() calls this on blocked tasks without
2105 lockdep_assert_held(&rq
->lock
);
2106 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
2109 put_prev_task(rq
, p
);
2111 p
->sched_class
->set_cpus_allowed(p
, new_mask
, flags
);
2114 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
2116 set_next_task(rq
, p
);
2119 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
2121 __do_set_cpus_allowed(p
, new_mask
, 0);
2125 * This function is wildly self concurrent; here be dragons.
2128 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2129 * designated task is enqueued on an allowed CPU. If that task is currently
2130 * running, we have to kick it out using the CPU stopper.
2132 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2135 * Initial conditions: P0->cpus_mask = [0, 1]
2139 * migrate_disable();
2141 * set_cpus_allowed_ptr(P0, [1]);
2143 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2144 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2145 * This means we need the following scheme:
2149 * migrate_disable();
2151 * set_cpus_allowed_ptr(P0, [1]);
2155 * __set_cpus_allowed_ptr();
2156 * <wakes local stopper>
2157 * `--> <woken on migration completion>
2159 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2160 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2161 * task p are serialized by p->pi_lock, which we can leverage: the one that
2162 * should come into effect at the end of the Migrate-Disable region is the last
2163 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2164 * but we still need to properly signal those waiting tasks at the appropriate
2167 * This is implemented using struct set_affinity_pending. The first
2168 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2169 * setup an instance of that struct and install it on the targeted task_struct.
2170 * Any and all further callers will reuse that instance. Those then wait for
2171 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2172 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2175 * (1) In the cases covered above. There is one more where the completion is
2176 * signaled within affine_move_task() itself: when a subsequent affinity request
2177 * cancels the need for an active migration. Consider:
2179 * Initial conditions: P0->cpus_mask = [0, 1]
2183 * migrate_disable();
2185 * set_cpus_allowed_ptr(P0, [1]);
2187 * set_cpus_allowed_ptr(P0, [0, 1]);
2188 * <signal completion>
2191 * Note that the above is safe vs a concurrent migrate_enable(), as any
2192 * pending affinity completion is preceded by an uninstallation of
2193 * p->migration_pending done with p->pi_lock held.
2195 static int affine_move_task(struct rq
*rq
, struct task_struct
*p
, struct rq_flags
*rf
,
2196 int dest_cpu
, unsigned int flags
)
2198 struct set_affinity_pending my_pending
= { }, *pending
= NULL
;
2199 struct migration_arg arg
= {
2201 .dest_cpu
= dest_cpu
,
2203 bool complete
= false;
2205 /* Can the task run on the task's current CPU? If so, we're done */
2206 if (cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
)) {
2207 struct task_struct
*push_task
= NULL
;
2209 if ((flags
& SCA_MIGRATE_ENABLE
) &&
2210 (p
->migration_flags
& MDF_PUSH
) && !rq
->push_busy
) {
2211 rq
->push_busy
= true;
2212 push_task
= get_task_struct(p
);
2215 pending
= p
->migration_pending
;
2217 refcount_inc(&pending
->refs
);
2218 p
->migration_pending
= NULL
;
2221 task_rq_unlock(rq
, p
, rf
);
2224 stop_one_cpu_nowait(rq
->cpu
, push_cpu_stop
,
2234 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2235 /* serialized by p->pi_lock */
2236 if (!p
->migration_pending
) {
2237 /* Install the request */
2238 refcount_set(&my_pending
.refs
, 1);
2239 init_completion(&my_pending
.done
);
2240 p
->migration_pending
= &my_pending
;
2242 pending
= p
->migration_pending
;
2243 refcount_inc(&pending
->refs
);
2246 pending
= p
->migration_pending
;
2248 * - !MIGRATE_ENABLE:
2249 * we'll have installed a pending if there wasn't one already.
2252 * we're here because the current CPU isn't matching anymore,
2253 * the only way that can happen is because of a concurrent
2254 * set_cpus_allowed_ptr() call, which should then still be
2255 * pending completion.
2257 * Either way, we really should have a @pending here.
2259 if (WARN_ON_ONCE(!pending
)) {
2260 task_rq_unlock(rq
, p
, rf
);
2264 if (flags
& SCA_MIGRATE_ENABLE
) {
2266 refcount_inc(&pending
->refs
); /* pending->{arg,stop_work} */
2267 p
->migration_flags
&= ~MDF_PUSH
;
2268 task_rq_unlock(rq
, p
, rf
);
2270 pending
->arg
= (struct migration_arg
) {
2276 stop_one_cpu_nowait(cpu_of(rq
), migration_cpu_stop
,
2277 &pending
->arg
, &pending
->stop_work
);
2282 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
2284 * Lessen races (and headaches) by delegating
2285 * is_migration_disabled(p) checks to the stopper, which will
2286 * run on the same CPU as said p.
2288 task_rq_unlock(rq
, p
, rf
);
2289 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
2293 if (!is_migration_disabled(p
)) {
2294 if (task_on_rq_queued(p
))
2295 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
2297 p
->migration_pending
= NULL
;
2300 task_rq_unlock(rq
, p
, rf
);
2304 complete_all(&pending
->done
);
2307 wait_for_completion(&pending
->done
);
2309 if (refcount_dec_and_test(&pending
->refs
))
2310 wake_up_var(&pending
->refs
);
2313 * Block the original owner of &pending until all subsequent callers
2314 * have seen the completion and decremented the refcount
2316 wait_var_event(&my_pending
.refs
, !refcount_read(&my_pending
.refs
));
2322 * Change a given task's CPU affinity. Migrate the thread to a
2323 * proper CPU and schedule it away if the CPU it's executing on
2324 * is removed from the allowed bitmask.
2326 * NOTE: the caller must have a valid reference to the task, the
2327 * task must not exit() & deallocate itself prematurely. The
2328 * call is not atomic; no spinlocks may be held.
2330 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
2331 const struct cpumask
*new_mask
,
2334 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
2335 unsigned int dest_cpu
;
2340 rq
= task_rq_lock(p
, &rf
);
2341 update_rq_clock(rq
);
2343 if (p
->flags
& PF_KTHREAD
|| is_migration_disabled(p
)) {
2345 * Kernel threads are allowed on online && !active CPUs,
2346 * however, during cpu-hot-unplug, even these might get pushed
2347 * away if not KTHREAD_IS_PER_CPU.
2349 * Specifically, migration_disabled() tasks must not fail the
2350 * cpumask_any_and_distribute() pick below, esp. so on
2351 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2352 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2354 cpu_valid_mask
= cpu_online_mask
;
2358 * Must re-check here, to close a race against __kthread_bind(),
2359 * sched_setaffinity() is not guaranteed to observe the flag.
2361 if ((flags
& SCA_CHECK
) && (p
->flags
& PF_NO_SETAFFINITY
)) {
2366 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2367 if (cpumask_equal(&p
->cpus_mask
, new_mask
))
2370 if (WARN_ON_ONCE(p
== current
&&
2371 is_migration_disabled(p
) &&
2372 !cpumask_test_cpu(task_cpu(p
), new_mask
))) {
2379 * Picking a ~random cpu helps in cases where we are changing affinity
2380 * for groups of tasks (ie. cpuset), so that load balancing is not
2381 * immediately required to distribute the tasks within their new mask.
2383 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
2384 if (dest_cpu
>= nr_cpu_ids
) {
2389 __do_set_cpus_allowed(p
, new_mask
, flags
);
2391 return affine_move_task(rq
, p
, &rf
, dest_cpu
, flags
);
2394 task_rq_unlock(rq
, p
, &rf
);
2399 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
2401 return __set_cpus_allowed_ptr(p
, new_mask
, 0);
2403 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
2405 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2407 #ifdef CONFIG_SCHED_DEBUG
2409 * We should never call set_task_cpu() on a blocked task,
2410 * ttwu() will sort out the placement.
2412 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2416 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2417 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2418 * time relying on p->on_rq.
2420 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
2421 p
->sched_class
== &fair_sched_class
&&
2422 (p
->on_rq
&& !task_on_rq_migrating(p
)));
2424 #ifdef CONFIG_LOCKDEP
2426 * The caller should hold either p->pi_lock or rq->lock, when changing
2427 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2429 * sched_move_task() holds both and thus holding either pins the cgroup,
2432 * Furthermore, all task_rq users should acquire both locks, see
2435 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2436 lockdep_is_held(&task_rq(p
)->lock
)));
2439 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2441 WARN_ON_ONCE(!cpu_online(new_cpu
));
2443 WARN_ON_ONCE(is_migration_disabled(p
));
2446 trace_sched_migrate_task(p
, new_cpu
);
2448 if (task_cpu(p
) != new_cpu
) {
2449 if (p
->sched_class
->migrate_task_rq
)
2450 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
2451 p
->se
.nr_migrations
++;
2453 perf_event_task_migrate(p
);
2456 __set_task_cpu(p
, new_cpu
);
2459 #ifdef CONFIG_NUMA_BALANCING
2460 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
2462 if (task_on_rq_queued(p
)) {
2463 struct rq
*src_rq
, *dst_rq
;
2464 struct rq_flags srf
, drf
;
2466 src_rq
= task_rq(p
);
2467 dst_rq
= cpu_rq(cpu
);
2469 rq_pin_lock(src_rq
, &srf
);
2470 rq_pin_lock(dst_rq
, &drf
);
2472 deactivate_task(src_rq
, p
, 0);
2473 set_task_cpu(p
, cpu
);
2474 activate_task(dst_rq
, p
, 0);
2475 check_preempt_curr(dst_rq
, p
, 0);
2477 rq_unpin_lock(dst_rq
, &drf
);
2478 rq_unpin_lock(src_rq
, &srf
);
2482 * Task isn't running anymore; make it appear like we migrated
2483 * it before it went to sleep. This means on wakeup we make the
2484 * previous CPU our target instead of where it really is.
2490 struct migration_swap_arg
{
2491 struct task_struct
*src_task
, *dst_task
;
2492 int src_cpu
, dst_cpu
;
2495 static int migrate_swap_stop(void *data
)
2497 struct migration_swap_arg
*arg
= data
;
2498 struct rq
*src_rq
, *dst_rq
;
2501 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
2504 src_rq
= cpu_rq(arg
->src_cpu
);
2505 dst_rq
= cpu_rq(arg
->dst_cpu
);
2507 double_raw_lock(&arg
->src_task
->pi_lock
,
2508 &arg
->dst_task
->pi_lock
);
2509 double_rq_lock(src_rq
, dst_rq
);
2511 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
2514 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
2517 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
2520 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
2523 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
2524 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
2529 double_rq_unlock(src_rq
, dst_rq
);
2530 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
2531 raw_spin_unlock(&arg
->src_task
->pi_lock
);
2537 * Cross migrate two tasks
2539 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
2540 int target_cpu
, int curr_cpu
)
2542 struct migration_swap_arg arg
;
2545 arg
= (struct migration_swap_arg
){
2547 .src_cpu
= curr_cpu
,
2549 .dst_cpu
= target_cpu
,
2552 if (arg
.src_cpu
== arg
.dst_cpu
)
2556 * These three tests are all lockless; this is OK since all of them
2557 * will be re-checked with proper locks held further down the line.
2559 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
2562 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
2565 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
2568 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
2569 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
2574 #endif /* CONFIG_NUMA_BALANCING */
2577 * wait_task_inactive - wait for a thread to unschedule.
2579 * If @match_state is nonzero, it's the @p->state value just checked and
2580 * not expected to change. If it changes, i.e. @p might have woken up,
2581 * then return zero. When we succeed in waiting for @p to be off its CPU,
2582 * we return a positive number (its total switch count). If a second call
2583 * a short while later returns the same number, the caller can be sure that
2584 * @p has remained unscheduled the whole time.
2586 * The caller must ensure that the task *will* unschedule sometime soon,
2587 * else this function might spin for a *long* time. This function can't
2588 * be called with interrupts off, or it may introduce deadlock with
2589 * smp_call_function() if an IPI is sent by the same process we are
2590 * waiting to become inactive.
2592 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2594 int running
, queued
;
2601 * We do the initial early heuristics without holding
2602 * any task-queue locks at all. We'll only try to get
2603 * the runqueue lock when things look like they will
2609 * If the task is actively running on another CPU
2610 * still, just relax and busy-wait without holding
2613 * NOTE! Since we don't hold any locks, it's not
2614 * even sure that "rq" stays as the right runqueue!
2615 * But we don't care, since "task_running()" will
2616 * return false if the runqueue has changed and p
2617 * is actually now running somewhere else!
2619 while (task_running(rq
, p
)) {
2620 if (match_state
&& unlikely(p
->state
!= match_state
))
2626 * Ok, time to look more closely! We need the rq
2627 * lock now, to be *sure*. If we're wrong, we'll
2628 * just go back and repeat.
2630 rq
= task_rq_lock(p
, &rf
);
2631 trace_sched_wait_task(p
);
2632 running
= task_running(rq
, p
);
2633 queued
= task_on_rq_queued(p
);
2635 if (!match_state
|| p
->state
== match_state
)
2636 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2637 task_rq_unlock(rq
, p
, &rf
);
2640 * If it changed from the expected state, bail out now.
2642 if (unlikely(!ncsw
))
2646 * Was it really running after all now that we
2647 * checked with the proper locks actually held?
2649 * Oops. Go back and try again..
2651 if (unlikely(running
)) {
2657 * It's not enough that it's not actively running,
2658 * it must be off the runqueue _entirely_, and not
2661 * So if it was still runnable (but just not actively
2662 * running right now), it's preempted, and we should
2663 * yield - it could be a while.
2665 if (unlikely(queued
)) {
2666 ktime_t to
= NSEC_PER_SEC
/ HZ
;
2668 set_current_state(TASK_UNINTERRUPTIBLE
);
2669 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2674 * Ahh, all good. It wasn't running, and it wasn't
2675 * runnable, which means that it will never become
2676 * running in the future either. We're all done!
2685 * kick_process - kick a running thread to enter/exit the kernel
2686 * @p: the to-be-kicked thread
2688 * Cause a process which is running on another CPU to enter
2689 * kernel-mode, without any delay. (to get signals handled.)
2691 * NOTE: this function doesn't have to take the runqueue lock,
2692 * because all it wants to ensure is that the remote task enters
2693 * the kernel. If the IPI races and the task has been migrated
2694 * to another CPU then no harm is done and the purpose has been
2697 void kick_process(struct task_struct
*p
)
2703 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2704 smp_send_reschedule(cpu
);
2707 EXPORT_SYMBOL_GPL(kick_process
);
2710 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2712 * A few notes on cpu_active vs cpu_online:
2714 * - cpu_active must be a subset of cpu_online
2716 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2717 * see __set_cpus_allowed_ptr(). At this point the newly online
2718 * CPU isn't yet part of the sched domains, and balancing will not
2721 * - on CPU-down we clear cpu_active() to mask the sched domains and
2722 * avoid the load balancer to place new tasks on the to be removed
2723 * CPU. Existing tasks will remain running there and will be taken
2726 * This means that fallback selection must not select !active CPUs.
2727 * And can assume that any active CPU must be online. Conversely
2728 * select_task_rq() below may allow selection of !active CPUs in order
2729 * to satisfy the above rules.
2731 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2733 int nid
= cpu_to_node(cpu
);
2734 const struct cpumask
*nodemask
= NULL
;
2735 enum { cpuset
, possible
, fail
} state
= cpuset
;
2739 * If the node that the CPU is on has been offlined, cpu_to_node()
2740 * will return -1. There is no CPU on the node, and we should
2741 * select the CPU on the other node.
2744 nodemask
= cpumask_of_node(nid
);
2746 /* Look for allowed, online CPU in same node. */
2747 for_each_cpu(dest_cpu
, nodemask
) {
2748 if (!cpu_active(dest_cpu
))
2750 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2756 /* Any allowed, online CPU? */
2757 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2758 if (!is_cpu_allowed(p
, dest_cpu
))
2764 /* No more Mr. Nice Guy. */
2767 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2768 cpuset_cpus_allowed_fallback(p
);
2775 * XXX When called from select_task_rq() we only
2776 * hold p->pi_lock and again violate locking order.
2778 * More yuck to audit.
2780 do_set_cpus_allowed(p
, cpu_possible_mask
);
2791 if (state
!= cpuset
) {
2793 * Don't tell them about moving exiting tasks or
2794 * kernel threads (both mm NULL), since they never
2797 if (p
->mm
&& printk_ratelimit()) {
2798 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2799 task_pid_nr(p
), p
->comm
, cpu
);
2807 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2810 int select_task_rq(struct task_struct
*p
, int cpu
, int wake_flags
)
2812 lockdep_assert_held(&p
->pi_lock
);
2814 if (p
->nr_cpus_allowed
> 1 && !is_migration_disabled(p
))
2815 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, wake_flags
);
2817 cpu
= cpumask_any(p
->cpus_ptr
);
2820 * In order not to call set_task_cpu() on a blocking task we need
2821 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2824 * Since this is common to all placement strategies, this lives here.
2826 * [ this allows ->select_task() to simply return task_cpu(p) and
2827 * not worry about this generic constraint ]
2829 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2830 cpu
= select_fallback_rq(task_cpu(p
), p
);
2835 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2837 static struct lock_class_key stop_pi_lock
;
2838 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2839 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2843 * Make it appear like a SCHED_FIFO task, its something
2844 * userspace knows about and won't get confused about.
2846 * Also, it will make PI more or less work without too
2847 * much confusion -- but then, stop work should not
2848 * rely on PI working anyway.
2850 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2852 stop
->sched_class
= &stop_sched_class
;
2855 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
2856 * adjust the effective priority of a task. As a result,
2857 * rt_mutex_setprio() can trigger (RT) balancing operations,
2858 * which can then trigger wakeups of the stop thread to push
2859 * around the current task.
2861 * The stop task itself will never be part of the PI-chain, it
2862 * never blocks, therefore that ->pi_lock recursion is safe.
2863 * Tell lockdep about this by placing the stop->pi_lock in its
2866 lockdep_set_class(&stop
->pi_lock
, &stop_pi_lock
);
2869 cpu_rq(cpu
)->stop
= stop
;
2873 * Reset it back to a normal scheduling class so that
2874 * it can die in pieces.
2876 old_stop
->sched_class
= &rt_sched_class
;
2880 #else /* CONFIG_SMP */
2882 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2883 const struct cpumask
*new_mask
,
2886 return set_cpus_allowed_ptr(p
, new_mask
);
2889 static inline void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
) { }
2891 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
2896 #endif /* !CONFIG_SMP */
2899 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2903 if (!schedstat_enabled())
2909 if (cpu
== rq
->cpu
) {
2910 __schedstat_inc(rq
->ttwu_local
);
2911 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2913 struct sched_domain
*sd
;
2915 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2917 for_each_domain(rq
->cpu
, sd
) {
2918 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2919 __schedstat_inc(sd
->ttwu_wake_remote
);
2926 if (wake_flags
& WF_MIGRATED
)
2927 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2928 #endif /* CONFIG_SMP */
2930 __schedstat_inc(rq
->ttwu_count
);
2931 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2933 if (wake_flags
& WF_SYNC
)
2934 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2938 * Mark the task runnable and perform wakeup-preemption.
2940 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2941 struct rq_flags
*rf
)
2943 check_preempt_curr(rq
, p
, wake_flags
);
2944 p
->state
= TASK_RUNNING
;
2945 trace_sched_wakeup(p
);
2948 if (p
->sched_class
->task_woken
) {
2950 * Our task @p is fully woken up and running; so it's safe to
2951 * drop the rq->lock, hereafter rq is only used for statistics.
2953 rq_unpin_lock(rq
, rf
);
2954 p
->sched_class
->task_woken(rq
, p
);
2955 rq_repin_lock(rq
, rf
);
2958 if (rq
->idle_stamp
) {
2959 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2960 u64 max
= 2*rq
->max_idle_balance_cost
;
2962 update_avg(&rq
->avg_idle
, delta
);
2964 if (rq
->avg_idle
> max
)
2973 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2974 struct rq_flags
*rf
)
2976 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2978 lockdep_assert_held(&rq
->lock
);
2980 if (p
->sched_contributes_to_load
)
2981 rq
->nr_uninterruptible
--;
2984 if (wake_flags
& WF_MIGRATED
)
2985 en_flags
|= ENQUEUE_MIGRATED
;
2989 delayacct_blkio_end(p
);
2990 atomic_dec(&task_rq(p
)->nr_iowait
);
2993 activate_task(rq
, p
, en_flags
);
2994 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2998 * Consider @p being inside a wait loop:
3001 * set_current_state(TASK_UNINTERRUPTIBLE);
3008 * __set_current_state(TASK_RUNNING);
3010 * between set_current_state() and schedule(). In this case @p is still
3011 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3014 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3015 * then schedule() must still happen and p->state can be changed to
3016 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3017 * need to do a full wakeup with enqueue.
3019 * Returns: %true when the wakeup is done,
3022 static int ttwu_runnable(struct task_struct
*p
, int wake_flags
)
3028 rq
= __task_rq_lock(p
, &rf
);
3029 if (task_on_rq_queued(p
)) {
3030 /* check_preempt_curr() may use rq clock */
3031 update_rq_clock(rq
);
3032 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
3035 __task_rq_unlock(rq
, &rf
);
3041 void sched_ttwu_pending(void *arg
)
3043 struct llist_node
*llist
= arg
;
3044 struct rq
*rq
= this_rq();
3045 struct task_struct
*p
, *t
;
3052 * rq::ttwu_pending racy indication of out-standing wakeups.
3053 * Races such that false-negatives are possible, since they
3054 * are shorter lived that false-positives would be.
3056 WRITE_ONCE(rq
->ttwu_pending
, 0);
3058 rq_lock_irqsave(rq
, &rf
);
3059 update_rq_clock(rq
);
3061 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
.llist
) {
3062 if (WARN_ON_ONCE(p
->on_cpu
))
3063 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3065 if (WARN_ON_ONCE(task_cpu(p
) != cpu_of(rq
)))
3066 set_task_cpu(p
, cpu_of(rq
));
3068 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
3071 rq_unlock_irqrestore(rq
, &rf
);
3074 void send_call_function_single_ipi(int cpu
)
3076 struct rq
*rq
= cpu_rq(cpu
);
3078 if (!set_nr_if_polling(rq
->idle
))
3079 arch_send_call_function_single_ipi(cpu
);
3081 trace_sched_wake_idle_without_ipi(cpu
);
3085 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3086 * necessary. The wakee CPU on receipt of the IPI will queue the task
3087 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3088 * of the wakeup instead of the waker.
3090 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3092 struct rq
*rq
= cpu_rq(cpu
);
3094 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
3096 WRITE_ONCE(rq
->ttwu_pending
, 1);
3097 __smp_call_single_queue(cpu
, &p
->wake_entry
.llist
);
3100 void wake_up_if_idle(int cpu
)
3102 struct rq
*rq
= cpu_rq(cpu
);
3107 if (!is_idle_task(rcu_dereference(rq
->curr
)))
3110 if (set_nr_if_polling(rq
->idle
)) {
3111 trace_sched_wake_idle_without_ipi(cpu
);
3113 rq_lock_irqsave(rq
, &rf
);
3114 if (is_idle_task(rq
->curr
))
3115 smp_send_reschedule(cpu
);
3116 /* Else CPU is not idle, do nothing here: */
3117 rq_unlock_irqrestore(rq
, &rf
);
3124 bool cpus_share_cache(int this_cpu
, int that_cpu
)
3126 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
3129 static inline bool ttwu_queue_cond(int cpu
, int wake_flags
)
3132 * Do not complicate things with the async wake_list while the CPU is
3135 if (!cpu_active(cpu
))
3139 * If the CPU does not share cache, then queue the task on the
3140 * remote rqs wakelist to avoid accessing remote data.
3142 if (!cpus_share_cache(smp_processor_id(), cpu
))
3146 * If the task is descheduling and the only running task on the
3147 * CPU then use the wakelist to offload the task activation to
3148 * the soon-to-be-idle CPU as the current CPU is likely busy.
3149 * nr_running is checked to avoid unnecessary task stacking.
3151 if ((wake_flags
& WF_ON_CPU
) && cpu_rq(cpu
)->nr_running
<= 1)
3157 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3159 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(cpu
, wake_flags
)) {
3160 if (WARN_ON_ONCE(cpu
== smp_processor_id()))
3163 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
3164 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
3171 #else /* !CONFIG_SMP */
3173 static inline bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3178 #endif /* CONFIG_SMP */
3180 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
3182 struct rq
*rq
= cpu_rq(cpu
);
3185 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
3189 update_rq_clock(rq
);
3190 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
3195 * Notes on Program-Order guarantees on SMP systems.
3199 * The basic program-order guarantee on SMP systems is that when a task [t]
3200 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3201 * execution on its new CPU [c1].
3203 * For migration (of runnable tasks) this is provided by the following means:
3205 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3206 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3207 * rq(c1)->lock (if not at the same time, then in that order).
3208 * C) LOCK of the rq(c1)->lock scheduling in task
3210 * Release/acquire chaining guarantees that B happens after A and C after B.
3211 * Note: the CPU doing B need not be c0 or c1
3220 * UNLOCK rq(0)->lock
3222 * LOCK rq(0)->lock // orders against CPU0
3224 * UNLOCK rq(0)->lock
3228 * UNLOCK rq(1)->lock
3230 * LOCK rq(1)->lock // orders against CPU2
3233 * UNLOCK rq(1)->lock
3236 * BLOCKING -- aka. SLEEP + WAKEUP
3238 * For blocking we (obviously) need to provide the same guarantee as for
3239 * migration. However the means are completely different as there is no lock
3240 * chain to provide order. Instead we do:
3242 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3243 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3247 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3249 * LOCK rq(0)->lock LOCK X->pi_lock
3252 * smp_store_release(X->on_cpu, 0);
3254 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3260 * X->state = RUNNING
3261 * UNLOCK rq(2)->lock
3263 * LOCK rq(2)->lock // orders against CPU1
3266 * UNLOCK rq(2)->lock
3269 * UNLOCK rq(0)->lock
3272 * However, for wakeups there is a second guarantee we must provide, namely we
3273 * must ensure that CONDITION=1 done by the caller can not be reordered with
3274 * accesses to the task state; see try_to_wake_up() and set_current_state().
3278 * try_to_wake_up - wake up a thread
3279 * @p: the thread to be awakened
3280 * @state: the mask of task states that can be woken
3281 * @wake_flags: wake modifier flags (WF_*)
3283 * Conceptually does:
3285 * If (@state & @p->state) @p->state = TASK_RUNNING.
3287 * If the task was not queued/runnable, also place it back on a runqueue.
3289 * This function is atomic against schedule() which would dequeue the task.
3291 * It issues a full memory barrier before accessing @p->state, see the comment
3292 * with set_current_state().
3294 * Uses p->pi_lock to serialize against concurrent wake-ups.
3296 * Relies on p->pi_lock stabilizing:
3299 * - p->sched_task_group
3300 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3302 * Tries really hard to only take one task_rq(p)->lock for performance.
3303 * Takes rq->lock in:
3304 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3305 * - ttwu_queue() -- new rq, for enqueue of the task;
3306 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3308 * As a consequence we race really badly with just about everything. See the
3309 * many memory barriers and their comments for details.
3311 * Return: %true if @p->state changes (an actual wakeup was done),
3315 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
3317 unsigned long flags
;
3318 int cpu
, success
= 0;
3323 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3324 * == smp_processor_id()'. Together this means we can special
3325 * case the whole 'p->on_rq && ttwu_runnable()' case below
3326 * without taking any locks.
3329 * - we rely on Program-Order guarantees for all the ordering,
3330 * - we're serialized against set_special_state() by virtue of
3331 * it disabling IRQs (this allows not taking ->pi_lock).
3333 if (!(p
->state
& state
))
3337 trace_sched_waking(p
);
3338 p
->state
= TASK_RUNNING
;
3339 trace_sched_wakeup(p
);
3344 * If we are going to wake up a thread waiting for CONDITION we
3345 * need to ensure that CONDITION=1 done by the caller can not be
3346 * reordered with p->state check below. This pairs with smp_store_mb()
3347 * in set_current_state() that the waiting thread does.
3349 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3350 smp_mb__after_spinlock();
3351 if (!(p
->state
& state
))
3354 trace_sched_waking(p
);
3356 /* We're going to change ->state: */
3360 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3361 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3362 * in smp_cond_load_acquire() below.
3364 * sched_ttwu_pending() try_to_wake_up()
3365 * STORE p->on_rq = 1 LOAD p->state
3368 * __schedule() (switch to task 'p')
3369 * LOCK rq->lock smp_rmb();
3370 * smp_mb__after_spinlock();
3374 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3376 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3377 * __schedule(). See the comment for smp_mb__after_spinlock().
3379 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3382 if (READ_ONCE(p
->on_rq
) && ttwu_runnable(p
, wake_flags
))
3387 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3388 * possible to, falsely, observe p->on_cpu == 0.
3390 * One must be running (->on_cpu == 1) in order to remove oneself
3391 * from the runqueue.
3393 * __schedule() (switch to task 'p') try_to_wake_up()
3394 * STORE p->on_cpu = 1 LOAD p->on_rq
3397 * __schedule() (put 'p' to sleep)
3398 * LOCK rq->lock smp_rmb();
3399 * smp_mb__after_spinlock();
3400 * STORE p->on_rq = 0 LOAD p->on_cpu
3402 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3403 * __schedule(). See the comment for smp_mb__after_spinlock().
3405 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3406 * schedule()'s deactivate_task() has 'happened' and p will no longer
3407 * care about it's own p->state. See the comment in __schedule().
3409 smp_acquire__after_ctrl_dep();
3412 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3413 * == 0), which means we need to do an enqueue, change p->state to
3414 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3415 * enqueue, such as ttwu_queue_wakelist().
3417 p
->state
= TASK_WAKING
;
3420 * If the owning (remote) CPU is still in the middle of schedule() with
3421 * this task as prev, considering queueing p on the remote CPUs wake_list
3422 * which potentially sends an IPI instead of spinning on p->on_cpu to
3423 * let the waker make forward progress. This is safe because IRQs are
3424 * disabled and the IPI will deliver after on_cpu is cleared.
3426 * Ensure we load task_cpu(p) after p->on_cpu:
3428 * set_task_cpu(p, cpu);
3429 * STORE p->cpu = @cpu
3430 * __schedule() (switch to task 'p')
3432 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3433 * STORE p->on_cpu = 1 LOAD p->cpu
3435 * to ensure we observe the correct CPU on which the task is currently
3438 if (smp_load_acquire(&p
->on_cpu
) &&
3439 ttwu_queue_wakelist(p
, task_cpu(p
), wake_flags
| WF_ON_CPU
))
3443 * If the owning (remote) CPU is still in the middle of schedule() with
3444 * this task as prev, wait until it's done referencing the task.
3446 * Pairs with the smp_store_release() in finish_task().
3448 * This ensures that tasks getting woken will be fully ordered against
3449 * their previous state and preserve Program Order.
3451 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3453 cpu
= select_task_rq(p
, p
->wake_cpu
, wake_flags
| WF_TTWU
);
3454 if (task_cpu(p
) != cpu
) {
3456 delayacct_blkio_end(p
);
3457 atomic_dec(&task_rq(p
)->nr_iowait
);
3460 wake_flags
|= WF_MIGRATED
;
3461 psi_ttwu_dequeue(p
);
3462 set_task_cpu(p
, cpu
);
3466 #endif /* CONFIG_SMP */
3468 ttwu_queue(p
, cpu
, wake_flags
);
3470 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3473 ttwu_stat(p
, task_cpu(p
), wake_flags
);
3480 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3481 * @p: Process for which the function is to be invoked, can be @current.
3482 * @func: Function to invoke.
3483 * @arg: Argument to function.
3485 * If the specified task can be quickly locked into a definite state
3486 * (either sleeping or on a given runqueue), arrange to keep it in that
3487 * state while invoking @func(@arg). This function can use ->on_rq and
3488 * task_curr() to work out what the state is, if required. Given that
3489 * @func can be invoked with a runqueue lock held, it had better be quite
3493 * @false if the task slipped out from under the locks.
3494 * @true if the task was locked onto a runqueue or is sleeping.
3495 * However, @func can override this by returning @false.
3497 bool try_invoke_on_locked_down_task(struct task_struct
*p
, bool (*func
)(struct task_struct
*t
, void *arg
), void *arg
)
3503 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
3505 rq
= __task_rq_lock(p
, &rf
);
3506 if (task_rq(p
) == rq
)
3515 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3520 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
.flags
);
3525 * wake_up_process - Wake up a specific process
3526 * @p: The process to be woken up.
3528 * Attempt to wake up the nominated process and move it to the set of runnable
3531 * Return: 1 if the process was woken up, 0 if it was already running.
3533 * This function executes a full memory barrier before accessing the task state.
3535 int wake_up_process(struct task_struct
*p
)
3537 return try_to_wake_up(p
, TASK_NORMAL
, 0);
3539 EXPORT_SYMBOL(wake_up_process
);
3541 int wake_up_state(struct task_struct
*p
, unsigned int state
)
3543 return try_to_wake_up(p
, state
, 0);
3547 * Perform scheduler related setup for a newly forked process p.
3548 * p is forked by current.
3550 * __sched_fork() is basic setup used by init_idle() too:
3552 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
3557 p
->se
.exec_start
= 0;
3558 p
->se
.sum_exec_runtime
= 0;
3559 p
->se
.prev_sum_exec_runtime
= 0;
3560 p
->se
.nr_migrations
= 0;
3562 INIT_LIST_HEAD(&p
->se
.group_node
);
3564 #ifdef CONFIG_FAIR_GROUP_SCHED
3565 p
->se
.cfs_rq
= NULL
;
3568 #ifdef CONFIG_SCHEDSTATS
3569 /* Even if schedstat is disabled, there should not be garbage */
3570 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
3573 RB_CLEAR_NODE(&p
->dl
.rb_node
);
3574 init_dl_task_timer(&p
->dl
);
3575 init_dl_inactive_task_timer(&p
->dl
);
3576 __dl_clear_params(p
);
3578 INIT_LIST_HEAD(&p
->rt
.run_list
);
3580 p
->rt
.time_slice
= sched_rr_timeslice
;
3584 #ifdef CONFIG_PREEMPT_NOTIFIERS
3585 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
3588 #ifdef CONFIG_COMPACTION
3589 p
->capture_control
= NULL
;
3591 init_numa_balancing(clone_flags
, p
);
3593 p
->wake_entry
.u_flags
= CSD_TYPE_TTWU
;
3594 p
->migration_pending
= NULL
;
3598 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
3600 #ifdef CONFIG_NUMA_BALANCING
3602 void set_numabalancing_state(bool enabled
)
3605 static_branch_enable(&sched_numa_balancing
);
3607 static_branch_disable(&sched_numa_balancing
);
3610 #ifdef CONFIG_PROC_SYSCTL
3611 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
3612 void *buffer
, size_t *lenp
, loff_t
*ppos
)
3616 int state
= static_branch_likely(&sched_numa_balancing
);
3618 if (write
&& !capable(CAP_SYS_ADMIN
))
3623 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
3627 set_numabalancing_state(state
);
3633 #ifdef CONFIG_SCHEDSTATS
3635 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
3636 static bool __initdata __sched_schedstats
= false;
3638 static void set_schedstats(bool enabled
)
3641 static_branch_enable(&sched_schedstats
);
3643 static_branch_disable(&sched_schedstats
);
3646 void force_schedstat_enabled(void)
3648 if (!schedstat_enabled()) {
3649 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3650 static_branch_enable(&sched_schedstats
);
3654 static int __init
setup_schedstats(char *str
)
3661 * This code is called before jump labels have been set up, so we can't
3662 * change the static branch directly just yet. Instead set a temporary
3663 * variable so init_schedstats() can do it later.
3665 if (!strcmp(str
, "enable")) {
3666 __sched_schedstats
= true;
3668 } else if (!strcmp(str
, "disable")) {
3669 __sched_schedstats
= false;
3674 pr_warn("Unable to parse schedstats=\n");
3678 __setup("schedstats=", setup_schedstats
);
3680 static void __init
init_schedstats(void)
3682 set_schedstats(__sched_schedstats
);
3685 #ifdef CONFIG_PROC_SYSCTL
3686 int sysctl_schedstats(struct ctl_table
*table
, int write
, void *buffer
,
3687 size_t *lenp
, loff_t
*ppos
)
3691 int state
= static_branch_likely(&sched_schedstats
);
3693 if (write
&& !capable(CAP_SYS_ADMIN
))
3698 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
3702 set_schedstats(state
);
3705 #endif /* CONFIG_PROC_SYSCTL */
3706 #else /* !CONFIG_SCHEDSTATS */
3707 static inline void init_schedstats(void) {}
3708 #endif /* CONFIG_SCHEDSTATS */
3711 * fork()/clone()-time setup:
3713 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
3715 unsigned long flags
;
3717 __sched_fork(clone_flags
, p
);
3719 * We mark the process as NEW here. This guarantees that
3720 * nobody will actually run it, and a signal or other external
3721 * event cannot wake it up and insert it on the runqueue either.
3723 p
->state
= TASK_NEW
;
3726 * Make sure we do not leak PI boosting priority to the child.
3728 p
->prio
= current
->normal_prio
;
3733 * Revert to default priority/policy on fork if requested.
3735 if (unlikely(p
->sched_reset_on_fork
)) {
3736 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3737 p
->policy
= SCHED_NORMAL
;
3738 p
->static_prio
= NICE_TO_PRIO(0);
3740 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
3741 p
->static_prio
= NICE_TO_PRIO(0);
3743 p
->prio
= p
->normal_prio
= __normal_prio(p
);
3744 set_load_weight(p
, false);
3747 * We don't need the reset flag anymore after the fork. It has
3748 * fulfilled its duty:
3750 p
->sched_reset_on_fork
= 0;
3753 if (dl_prio(p
->prio
))
3755 else if (rt_prio(p
->prio
))
3756 p
->sched_class
= &rt_sched_class
;
3758 p
->sched_class
= &fair_sched_class
;
3760 init_entity_runnable_average(&p
->se
);
3763 * The child is not yet in the pid-hash so no cgroup attach races,
3764 * and the cgroup is pinned to this child due to cgroup_fork()
3765 * is ran before sched_fork().
3767 * Silence PROVE_RCU.
3769 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3772 * We're setting the CPU for the first time, we don't migrate,
3773 * so use __set_task_cpu().
3775 __set_task_cpu(p
, smp_processor_id());
3776 if (p
->sched_class
->task_fork
)
3777 p
->sched_class
->task_fork(p
);
3778 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3780 #ifdef CONFIG_SCHED_INFO
3781 if (likely(sched_info_on()))
3782 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
3784 #if defined(CONFIG_SMP)
3787 init_task_preempt_count(p
);
3789 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
3790 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
3795 void sched_post_fork(struct task_struct
*p
)
3797 uclamp_post_fork(p
);
3800 unsigned long to_ratio(u64 period
, u64 runtime
)
3802 if (runtime
== RUNTIME_INF
)
3806 * Doing this here saves a lot of checks in all
3807 * the calling paths, and returning zero seems
3808 * safe for them anyway.
3813 return div64_u64(runtime
<< BW_SHIFT
, period
);
3817 * wake_up_new_task - wake up a newly created task for the first time.
3819 * This function will do some initial scheduler statistics housekeeping
3820 * that must be done for every newly created context, then puts the task
3821 * on the runqueue and wakes it.
3823 void wake_up_new_task(struct task_struct
*p
)
3828 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
3829 p
->state
= TASK_RUNNING
;
3832 * Fork balancing, do it here and not earlier because:
3833 * - cpus_ptr can change in the fork path
3834 * - any previously selected CPU might disappear through hotplug
3836 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3837 * as we're not fully set-up yet.
3839 p
->recent_used_cpu
= task_cpu(p
);
3841 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), WF_FORK
));
3843 rq
= __task_rq_lock(p
, &rf
);
3844 update_rq_clock(rq
);
3845 post_init_entity_util_avg(p
);
3847 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
3848 trace_sched_wakeup_new(p
);
3849 check_preempt_curr(rq
, p
, WF_FORK
);
3851 if (p
->sched_class
->task_woken
) {
3853 * Nothing relies on rq->lock after this, so it's fine to
3856 rq_unpin_lock(rq
, &rf
);
3857 p
->sched_class
->task_woken(rq
, p
);
3858 rq_repin_lock(rq
, &rf
);
3861 task_rq_unlock(rq
, p
, &rf
);
3864 #ifdef CONFIG_PREEMPT_NOTIFIERS
3866 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
3868 void preempt_notifier_inc(void)
3870 static_branch_inc(&preempt_notifier_key
);
3872 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
3874 void preempt_notifier_dec(void)
3876 static_branch_dec(&preempt_notifier_key
);
3878 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3881 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3882 * @notifier: notifier struct to register
3884 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3886 if (!static_branch_unlikely(&preempt_notifier_key
))
3887 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3889 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3891 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3894 * preempt_notifier_unregister - no longer interested in preemption notifications
3895 * @notifier: notifier struct to unregister
3897 * This is *not* safe to call from within a preemption notifier.
3899 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3901 hlist_del(¬ifier
->link
);
3903 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3905 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3907 struct preempt_notifier
*notifier
;
3909 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3910 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3913 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3915 if (static_branch_unlikely(&preempt_notifier_key
))
3916 __fire_sched_in_preempt_notifiers(curr
);
3920 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3921 struct task_struct
*next
)
3923 struct preempt_notifier
*notifier
;
3925 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3926 notifier
->ops
->sched_out(notifier
, next
);
3929 static __always_inline
void
3930 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3931 struct task_struct
*next
)
3933 if (static_branch_unlikely(&preempt_notifier_key
))
3934 __fire_sched_out_preempt_notifiers(curr
, next
);
3937 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3939 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3944 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3945 struct task_struct
*next
)
3949 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3951 static inline void prepare_task(struct task_struct
*next
)
3955 * Claim the task as running, we do this before switching to it
3956 * such that any running task will have this set.
3958 * See the ttwu() WF_ON_CPU case and its ordering comment.
3960 WRITE_ONCE(next
->on_cpu
, 1);
3964 static inline void finish_task(struct task_struct
*prev
)
3968 * This must be the very last reference to @prev from this CPU. After
3969 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3970 * must ensure this doesn't happen until the switch is completely
3973 * In particular, the load of prev->state in finish_task_switch() must
3974 * happen before this.
3976 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3978 smp_store_release(&prev
->on_cpu
, 0);
3984 static void do_balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
3986 void (*func
)(struct rq
*rq
);
3987 struct callback_head
*next
;
3989 lockdep_assert_held(&rq
->lock
);
3992 func
= (void (*)(struct rq
*))head
->func
;
4001 static void balance_push(struct rq
*rq
);
4003 struct callback_head balance_push_callback
= {
4005 .func
= (void (*)(struct callback_head
*))balance_push
,
4008 static inline struct callback_head
*splice_balance_callbacks(struct rq
*rq
)
4010 struct callback_head
*head
= rq
->balance_callback
;
4012 lockdep_assert_held(&rq
->lock
);
4014 rq
->balance_callback
= NULL
;
4019 static void __balance_callbacks(struct rq
*rq
)
4021 do_balance_callbacks(rq
, splice_balance_callbacks(rq
));
4024 static inline void balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
4026 unsigned long flags
;
4028 if (unlikely(head
)) {
4029 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4030 do_balance_callbacks(rq
, head
);
4031 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
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
)
4053 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
4056 * Since the runqueue lock will be released by the next
4057 * task (which is an invalid locking op but in the case
4058 * of the scheduler it's an obvious special-case), so we
4059 * do an early lockdep release here:
4061 rq_unpin_lock(rq
, rf
);
4062 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
4063 #ifdef CONFIG_DEBUG_SPINLOCK
4064 /* this is a valid case when another task releases the spinlock */
4065 rq
->lock
.owner
= next
;
4069 static inline void finish_lock_switch(struct rq
*rq
)
4072 * If we are tracking spinlock dependencies then we have to
4073 * fix up the runqueue lock - which gets 'carried over' from
4074 * prev into current:
4076 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
4077 __balance_callbacks(rq
);
4078 raw_spin_unlock_irq(&rq
->lock
);
4082 * NOP if the arch has not defined these:
4085 #ifndef prepare_arch_switch
4086 # define prepare_arch_switch(next) do { } while (0)
4089 #ifndef finish_arch_post_lock_switch
4090 # define finish_arch_post_lock_switch() do { } while (0)
4093 static inline void kmap_local_sched_out(void)
4095 #ifdef CONFIG_KMAP_LOCAL
4096 if (unlikely(current
->kmap_ctrl
.idx
))
4097 __kmap_local_sched_out();
4101 static inline void kmap_local_sched_in(void)
4103 #ifdef CONFIG_KMAP_LOCAL
4104 if (unlikely(current
->kmap_ctrl
.idx
))
4105 __kmap_local_sched_in();
4110 * prepare_task_switch - prepare to switch tasks
4111 * @rq: the runqueue preparing to switch
4112 * @prev: the current task that is being switched out
4113 * @next: the task we are going to switch to.
4115 * This is called with the rq lock held and interrupts off. It must
4116 * be paired with a subsequent finish_task_switch after the context
4119 * prepare_task_switch sets up locking and calls architecture specific
4123 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
4124 struct task_struct
*next
)
4126 kcov_prepare_switch(prev
);
4127 sched_info_switch(rq
, prev
, next
);
4128 perf_event_task_sched_out(prev
, next
);
4130 fire_sched_out_preempt_notifiers(prev
, next
);
4131 kmap_local_sched_out();
4133 prepare_arch_switch(next
);
4137 * finish_task_switch - clean up after a task-switch
4138 * @prev: the thread we just switched away from.
4140 * finish_task_switch must be called after the context switch, paired
4141 * with a prepare_task_switch call before the context switch.
4142 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4143 * and do any other architecture-specific cleanup actions.
4145 * Note that we may have delayed dropping an mm in context_switch(). If
4146 * so, we finish that here outside of the runqueue lock. (Doing it
4147 * with the lock held can cause deadlocks; see schedule() for
4150 * The context switch have flipped the stack from under us and restored the
4151 * local variables which were saved when this task called schedule() in the
4152 * past. prev == current is still correct but we need to recalculate this_rq
4153 * because prev may have moved to another CPU.
4155 static struct rq
*finish_task_switch(struct task_struct
*prev
)
4156 __releases(rq
->lock
)
4158 struct rq
*rq
= this_rq();
4159 struct mm_struct
*mm
= rq
->prev_mm
;
4163 * The previous task will have left us with a preempt_count of 2
4164 * because it left us after:
4167 * preempt_disable(); // 1
4169 * raw_spin_lock_irq(&rq->lock) // 2
4171 * Also, see FORK_PREEMPT_COUNT.
4173 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
4174 "corrupted preempt_count: %s/%d/0x%x\n",
4175 current
->comm
, current
->pid
, preempt_count()))
4176 preempt_count_set(FORK_PREEMPT_COUNT
);
4181 * A task struct has one reference for the use as "current".
4182 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4183 * schedule one last time. The schedule call will never return, and
4184 * the scheduled task must drop that reference.
4186 * We must observe prev->state before clearing prev->on_cpu (in
4187 * finish_task), otherwise a concurrent wakeup can get prev
4188 * running on another CPU and we could rave with its RUNNING -> DEAD
4189 * transition, resulting in a double drop.
4191 prev_state
= prev
->state
;
4192 vtime_task_switch(prev
);
4193 perf_event_task_sched_in(prev
, current
);
4195 finish_lock_switch(rq
);
4196 finish_arch_post_lock_switch();
4197 kcov_finish_switch(current
);
4199 * kmap_local_sched_out() is invoked with rq::lock held and
4200 * interrupts disabled. There is no requirement for that, but the
4201 * sched out code does not have an interrupt enabled section.
4202 * Restoring the maps on sched in does not require interrupts being
4205 kmap_local_sched_in();
4207 fire_sched_in_preempt_notifiers(current
);
4209 * When switching through a kernel thread, the loop in
4210 * membarrier_{private,global}_expedited() may have observed that
4211 * kernel thread and not issued an IPI. It is therefore possible to
4212 * schedule between user->kernel->user threads without passing though
4213 * switch_mm(). Membarrier requires a barrier after storing to
4214 * rq->curr, before returning to userspace, so provide them here:
4216 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4217 * provided by mmdrop(),
4218 * - a sync_core for SYNC_CORE.
4221 membarrier_mm_sync_core_before_usermode(mm
);
4224 if (unlikely(prev_state
== TASK_DEAD
)) {
4225 if (prev
->sched_class
->task_dead
)
4226 prev
->sched_class
->task_dead(prev
);
4229 * Remove function-return probe instances associated with this
4230 * task and put them back on the free list.
4232 kprobe_flush_task(prev
);
4234 /* Task is done with its stack. */
4235 put_task_stack(prev
);
4237 put_task_struct_rcu_user(prev
);
4240 tick_nohz_task_switch();
4245 * schedule_tail - first thing a freshly forked thread must call.
4246 * @prev: the thread we just switched away from.
4248 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
4249 __releases(rq
->lock
)
4254 * New tasks start with FORK_PREEMPT_COUNT, see there and
4255 * finish_task_switch() for details.
4257 * finish_task_switch() will drop rq->lock() and lower preempt_count
4258 * and the preempt_enable() will end up enabling preemption (on
4259 * PREEMPT_COUNT kernels).
4262 rq
= finish_task_switch(prev
);
4265 if (current
->set_child_tid
)
4266 put_user(task_pid_vnr(current
), current
->set_child_tid
);
4268 calculate_sigpending();
4272 * context_switch - switch to the new MM and the new thread's register state.
4274 static __always_inline
struct rq
*
4275 context_switch(struct rq
*rq
, struct task_struct
*prev
,
4276 struct task_struct
*next
, struct rq_flags
*rf
)
4278 prepare_task_switch(rq
, prev
, next
);
4281 * For paravirt, this is coupled with an exit in switch_to to
4282 * combine the page table reload and the switch backend into
4285 arch_start_context_switch(prev
);
4288 * kernel -> kernel lazy + transfer active
4289 * user -> kernel lazy + mmgrab() active
4291 * kernel -> user switch + mmdrop() active
4292 * user -> user switch
4294 if (!next
->mm
) { // to kernel
4295 enter_lazy_tlb(prev
->active_mm
, next
);
4297 next
->active_mm
= prev
->active_mm
;
4298 if (prev
->mm
) // from user
4299 mmgrab(prev
->active_mm
);
4301 prev
->active_mm
= NULL
;
4303 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
4305 * sys_membarrier() requires an smp_mb() between setting
4306 * rq->curr / membarrier_switch_mm() and returning to userspace.
4308 * The below provides this either through switch_mm(), or in
4309 * case 'prev->active_mm == next->mm' through
4310 * finish_task_switch()'s mmdrop().
4312 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
4314 if (!prev
->mm
) { // from kernel
4315 /* will mmdrop() in finish_task_switch(). */
4316 rq
->prev_mm
= prev
->active_mm
;
4317 prev
->active_mm
= NULL
;
4321 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4323 prepare_lock_switch(rq
, next
, rf
);
4325 /* Here we just switch the register state and the stack. */
4326 switch_to(prev
, next
, prev
);
4329 return finish_task_switch(prev
);
4333 * nr_running and nr_context_switches:
4335 * externally visible scheduler statistics: current number of runnable
4336 * threads, total number of context switches performed since bootup.
4338 unsigned long nr_running(void)
4340 unsigned long i
, sum
= 0;
4342 for_each_online_cpu(i
)
4343 sum
+= cpu_rq(i
)->nr_running
;
4349 * Check if only the current task is running on the CPU.
4351 * Caution: this function does not check that the caller has disabled
4352 * preemption, thus the result might have a time-of-check-to-time-of-use
4353 * race. The caller is responsible to use it correctly, for example:
4355 * - from a non-preemptible section (of course)
4357 * - from a thread that is bound to a single CPU
4359 * - in a loop with very short iterations (e.g. a polling loop)
4361 bool single_task_running(void)
4363 return raw_rq()->nr_running
== 1;
4365 EXPORT_SYMBOL(single_task_running
);
4367 unsigned long long nr_context_switches(void)
4370 unsigned long long sum
= 0;
4372 for_each_possible_cpu(i
)
4373 sum
+= cpu_rq(i
)->nr_switches
;
4379 * Consumers of these two interfaces, like for example the cpuidle menu
4380 * governor, are using nonsensical data. Preferring shallow idle state selection
4381 * for a CPU that has IO-wait which might not even end up running the task when
4382 * it does become runnable.
4385 unsigned long nr_iowait_cpu(int cpu
)
4387 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
4391 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4393 * The idea behind IO-wait account is to account the idle time that we could
4394 * have spend running if it were not for IO. That is, if we were to improve the
4395 * storage performance, we'd have a proportional reduction in IO-wait time.
4397 * This all works nicely on UP, where, when a task blocks on IO, we account
4398 * idle time as IO-wait, because if the storage were faster, it could've been
4399 * running and we'd not be idle.
4401 * This has been extended to SMP, by doing the same for each CPU. This however
4404 * Imagine for instance the case where two tasks block on one CPU, only the one
4405 * CPU will have IO-wait accounted, while the other has regular idle. Even
4406 * though, if the storage were faster, both could've ran at the same time,
4407 * utilising both CPUs.
4409 * This means, that when looking globally, the current IO-wait accounting on
4410 * SMP is a lower bound, by reason of under accounting.
4412 * Worse, since the numbers are provided per CPU, they are sometimes
4413 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4414 * associated with any one particular CPU, it can wake to another CPU than it
4415 * blocked on. This means the per CPU IO-wait number is meaningless.
4417 * Task CPU affinities can make all that even more 'interesting'.
4420 unsigned long nr_iowait(void)
4422 unsigned long i
, sum
= 0;
4424 for_each_possible_cpu(i
)
4425 sum
+= nr_iowait_cpu(i
);
4433 * sched_exec - execve() is a valuable balancing opportunity, because at
4434 * this point the task has the smallest effective memory and cache footprint.
4436 void sched_exec(void)
4438 struct task_struct
*p
= current
;
4439 unsigned long flags
;
4442 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4443 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), WF_EXEC
);
4444 if (dest_cpu
== smp_processor_id())
4447 if (likely(cpu_active(dest_cpu
))) {
4448 struct migration_arg arg
= { p
, dest_cpu
};
4450 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4451 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
4455 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4460 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4461 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
4463 EXPORT_PER_CPU_SYMBOL(kstat
);
4464 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
4467 * The function fair_sched_class.update_curr accesses the struct curr
4468 * and its field curr->exec_start; when called from task_sched_runtime(),
4469 * we observe a high rate of cache misses in practice.
4470 * Prefetching this data results in improved performance.
4472 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
4474 #ifdef CONFIG_FAIR_GROUP_SCHED
4475 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
4477 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
4480 prefetch(&curr
->exec_start
);
4484 * Return accounted runtime for the task.
4485 * In case the task is currently running, return the runtime plus current's
4486 * pending runtime that have not been accounted yet.
4488 unsigned long long task_sched_runtime(struct task_struct
*p
)
4494 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4496 * 64-bit doesn't need locks to atomically read a 64-bit value.
4497 * So we have a optimization chance when the task's delta_exec is 0.
4498 * Reading ->on_cpu is racy, but this is ok.
4500 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4501 * If we race with it entering CPU, unaccounted time is 0. This is
4502 * indistinguishable from the read occurring a few cycles earlier.
4503 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4504 * been accounted, so we're correct here as well.
4506 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
4507 return p
->se
.sum_exec_runtime
;
4510 rq
= task_rq_lock(p
, &rf
);
4512 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4513 * project cycles that may never be accounted to this
4514 * thread, breaking clock_gettime().
4516 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
4517 prefetch_curr_exec_start(p
);
4518 update_rq_clock(rq
);
4519 p
->sched_class
->update_curr(rq
);
4521 ns
= p
->se
.sum_exec_runtime
;
4522 task_rq_unlock(rq
, p
, &rf
);
4528 * This function gets called by the timer code, with HZ frequency.
4529 * We call it with interrupts disabled.
4531 void scheduler_tick(void)
4533 int cpu
= smp_processor_id();
4534 struct rq
*rq
= cpu_rq(cpu
);
4535 struct task_struct
*curr
= rq
->curr
;
4537 unsigned long thermal_pressure
;
4539 arch_scale_freq_tick();
4544 update_rq_clock(rq
);
4545 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
4546 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
4547 curr
->sched_class
->task_tick(rq
, curr
, 0);
4548 calc_global_load_tick(rq
);
4553 perf_event_task_tick();
4556 rq
->idle_balance
= idle_cpu(cpu
);
4557 trigger_load_balance(rq
);
4561 #ifdef CONFIG_NO_HZ_FULL
4566 struct delayed_work work
;
4568 /* Values for ->state, see diagram below. */
4569 #define TICK_SCHED_REMOTE_OFFLINE 0
4570 #define TICK_SCHED_REMOTE_OFFLINING 1
4571 #define TICK_SCHED_REMOTE_RUNNING 2
4574 * State diagram for ->state:
4577 * TICK_SCHED_REMOTE_OFFLINE
4580 * | | sched_tick_remote()
4583 * +--TICK_SCHED_REMOTE_OFFLINING
4586 * sched_tick_start() | | sched_tick_stop()
4589 * TICK_SCHED_REMOTE_RUNNING
4592 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4593 * and sched_tick_start() are happy to leave the state in RUNNING.
4596 static struct tick_work __percpu
*tick_work_cpu
;
4598 static void sched_tick_remote(struct work_struct
*work
)
4600 struct delayed_work
*dwork
= to_delayed_work(work
);
4601 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
4602 int cpu
= twork
->cpu
;
4603 struct rq
*rq
= cpu_rq(cpu
);
4604 struct task_struct
*curr
;
4610 * Handle the tick only if it appears the remote CPU is running in full
4611 * dynticks mode. The check is racy by nature, but missing a tick or
4612 * having one too much is no big deal because the scheduler tick updates
4613 * statistics and checks timeslices in a time-independent way, regardless
4614 * of when exactly it is running.
4616 if (!tick_nohz_tick_stopped_cpu(cpu
))
4619 rq_lock_irq(rq
, &rf
);
4621 if (cpu_is_offline(cpu
))
4624 update_rq_clock(rq
);
4626 if (!is_idle_task(curr
)) {
4628 * Make sure the next tick runs within a reasonable
4631 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
4632 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
4634 curr
->sched_class
->task_tick(rq
, curr
, 0);
4636 calc_load_nohz_remote(rq
);
4638 rq_unlock_irq(rq
, &rf
);
4642 * Run the remote tick once per second (1Hz). This arbitrary
4643 * frequency is large enough to avoid overload but short enough
4644 * to keep scheduler internal stats reasonably up to date. But
4645 * first update state to reflect hotplug activity if required.
4647 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
4648 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
4649 if (os
== TICK_SCHED_REMOTE_RUNNING
)
4650 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
4653 static void sched_tick_start(int cpu
)
4656 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 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
4665 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
4666 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
4668 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
4669 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
4673 #ifdef CONFIG_HOTPLUG_CPU
4674 static void sched_tick_stop(int cpu
)
4676 struct tick_work
*twork
;
4679 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
4682 WARN_ON_ONCE(!tick_work_cpu
);
4684 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
4685 /* There cannot be competing actions, but don't rely on stop-machine. */
4686 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
4687 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
4688 /* Don't cancel, as this would mess up the state machine. */
4690 #endif /* CONFIG_HOTPLUG_CPU */
4692 int __init
sched_tick_offload_init(void)
4694 tick_work_cpu
= alloc_percpu(struct tick_work
);
4695 BUG_ON(!tick_work_cpu
);
4699 #else /* !CONFIG_NO_HZ_FULL */
4700 static inline void sched_tick_start(int cpu
) { }
4701 static inline void sched_tick_stop(int cpu
) { }
4704 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4705 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4707 * If the value passed in is equal to the current preempt count
4708 * then we just disabled preemption. Start timing the latency.
4710 static inline void preempt_latency_start(int val
)
4712 if (preempt_count() == val
) {
4713 unsigned long ip
= get_lock_parent_ip();
4714 #ifdef CONFIG_DEBUG_PREEMPT
4715 current
->preempt_disable_ip
= ip
;
4717 trace_preempt_off(CALLER_ADDR0
, ip
);
4721 void preempt_count_add(int val
)
4723 #ifdef CONFIG_DEBUG_PREEMPT
4727 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4730 __preempt_count_add(val
);
4731 #ifdef CONFIG_DEBUG_PREEMPT
4733 * Spinlock count overflowing soon?
4735 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4738 preempt_latency_start(val
);
4740 EXPORT_SYMBOL(preempt_count_add
);
4741 NOKPROBE_SYMBOL(preempt_count_add
);
4744 * If the value passed in equals to the current preempt count
4745 * then we just enabled preemption. Stop timing the latency.
4747 static inline void preempt_latency_stop(int val
)
4749 if (preempt_count() == val
)
4750 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
4753 void preempt_count_sub(int val
)
4755 #ifdef CONFIG_DEBUG_PREEMPT
4759 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4762 * Is the spinlock portion underflowing?
4764 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4765 !(preempt_count() & PREEMPT_MASK
)))
4769 preempt_latency_stop(val
);
4770 __preempt_count_sub(val
);
4772 EXPORT_SYMBOL(preempt_count_sub
);
4773 NOKPROBE_SYMBOL(preempt_count_sub
);
4776 static inline void preempt_latency_start(int val
) { }
4777 static inline void preempt_latency_stop(int val
) { }
4780 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
4782 #ifdef CONFIG_DEBUG_PREEMPT
4783 return p
->preempt_disable_ip
;
4790 * Print scheduling while atomic bug:
4792 static noinline
void __schedule_bug(struct task_struct
*prev
)
4794 /* Save this before calling printk(), since that will clobber it */
4795 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
4797 if (oops_in_progress
)
4800 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4801 prev
->comm
, prev
->pid
, preempt_count());
4803 debug_show_held_locks(prev
);
4805 if (irqs_disabled())
4806 print_irqtrace_events(prev
);
4807 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
4808 && in_atomic_preempt_off()) {
4809 pr_err("Preemption disabled at:");
4810 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
4813 panic("scheduling while atomic\n");
4816 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
4820 * Various schedule()-time debugging checks and statistics:
4822 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
4824 #ifdef CONFIG_SCHED_STACK_END_CHECK
4825 if (task_stack_end_corrupted(prev
))
4826 panic("corrupted stack end detected inside scheduler\n");
4828 if (task_scs_end_corrupted(prev
))
4829 panic("corrupted shadow stack detected inside scheduler\n");
4832 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4833 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
4834 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4835 prev
->comm
, prev
->pid
, prev
->non_block_count
);
4837 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
4841 if (unlikely(in_atomic_preempt_off())) {
4842 __schedule_bug(prev
);
4843 preempt_count_set(PREEMPT_DISABLED
);
4846 SCHED_WARN_ON(ct_state() == CONTEXT_USER
);
4848 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4850 schedstat_inc(this_rq()->sched_count
);
4853 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
4854 struct rq_flags
*rf
)
4857 const struct sched_class
*class;
4859 * We must do the balancing pass before put_prev_task(), such
4860 * that when we release the rq->lock the task is in the same
4861 * state as before we took rq->lock.
4863 * We can terminate the balance pass as soon as we know there is
4864 * a runnable task of @class priority or higher.
4866 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
4867 if (class->balance(rq
, prev
, rf
))
4872 put_prev_task(rq
, prev
);
4876 * Pick up the highest-prio task:
4878 static inline struct task_struct
*
4879 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
4881 const struct sched_class
*class;
4882 struct task_struct
*p
;
4885 * Optimization: we know that if all tasks are in the fair class we can
4886 * call that function directly, but only if the @prev task wasn't of a
4887 * higher scheduling class, because otherwise those lose the
4888 * opportunity to pull in more work from other CPUs.
4890 if (likely(prev
->sched_class
<= &fair_sched_class
&&
4891 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
4893 p
= pick_next_task_fair(rq
, prev
, rf
);
4894 if (unlikely(p
== RETRY_TASK
))
4897 /* Assumes fair_sched_class->next == idle_sched_class */
4899 put_prev_task(rq
, prev
);
4900 p
= pick_next_task_idle(rq
);
4907 put_prev_task_balance(rq
, prev
, rf
);
4909 for_each_class(class) {
4910 p
= class->pick_next_task(rq
);
4915 /* The idle class should always have a runnable task: */
4920 * __schedule() is the main scheduler function.
4922 * The main means of driving the scheduler and thus entering this function are:
4924 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4926 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4927 * paths. For example, see arch/x86/entry_64.S.
4929 * To drive preemption between tasks, the scheduler sets the flag in timer
4930 * interrupt handler scheduler_tick().
4932 * 3. Wakeups don't really cause entry into schedule(). They add a
4933 * task to the run-queue and that's it.
4935 * Now, if the new task added to the run-queue preempts the current
4936 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4937 * called on the nearest possible occasion:
4939 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4941 * - in syscall or exception context, at the next outmost
4942 * preempt_enable(). (this might be as soon as the wake_up()'s
4945 * - in IRQ context, return from interrupt-handler to
4946 * preemptible context
4948 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4951 * - cond_resched() call
4952 * - explicit schedule() call
4953 * - return from syscall or exception to user-space
4954 * - return from interrupt-handler to user-space
4956 * WARNING: must be called with preemption disabled!
4958 static void __sched notrace
__schedule(bool preempt
)
4960 struct task_struct
*prev
, *next
;
4961 unsigned long *switch_count
;
4962 unsigned long prev_state
;
4967 cpu
= smp_processor_id();
4971 schedule_debug(prev
, preempt
);
4973 if (sched_feat(HRTICK
))
4976 local_irq_disable();
4977 rcu_note_context_switch(preempt
);
4980 * Make sure that signal_pending_state()->signal_pending() below
4981 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4982 * done by the caller to avoid the race with signal_wake_up():
4984 * __set_current_state(@state) signal_wake_up()
4985 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4986 * wake_up_state(p, state)
4987 * LOCK rq->lock LOCK p->pi_state
4988 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4989 * if (signal_pending_state()) if (p->state & @state)
4991 * Also, the membarrier system call requires a full memory barrier
4992 * after coming from user-space, before storing to rq->curr.
4995 smp_mb__after_spinlock();
4997 /* Promote REQ to ACT */
4998 rq
->clock_update_flags
<<= 1;
4999 update_rq_clock(rq
);
5001 switch_count
= &prev
->nivcsw
;
5004 * We must load prev->state once (task_struct::state is volatile), such
5007 * - we form a control dependency vs deactivate_task() below.
5008 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5010 prev_state
= prev
->state
;
5011 if (!preempt
&& prev_state
) {
5012 if (signal_pending_state(prev_state
, prev
)) {
5013 prev
->state
= TASK_RUNNING
;
5015 prev
->sched_contributes_to_load
=
5016 (prev_state
& TASK_UNINTERRUPTIBLE
) &&
5017 !(prev_state
& TASK_NOLOAD
) &&
5018 !(prev
->flags
& PF_FROZEN
);
5020 if (prev
->sched_contributes_to_load
)
5021 rq
->nr_uninterruptible
++;
5024 * __schedule() ttwu()
5025 * prev_state = prev->state; if (p->on_rq && ...)
5026 * if (prev_state) goto out;
5027 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5028 * p->state = TASK_WAKING
5030 * Where __schedule() and ttwu() have matching control dependencies.
5032 * After this, schedule() must not care about p->state any more.
5034 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
5036 if (prev
->in_iowait
) {
5037 atomic_inc(&rq
->nr_iowait
);
5038 delayacct_blkio_start();
5041 switch_count
= &prev
->nvcsw
;
5044 next
= pick_next_task(rq
, prev
, &rf
);
5045 clear_tsk_need_resched(prev
);
5046 clear_preempt_need_resched();
5048 if (likely(prev
!= next
)) {
5051 * RCU users of rcu_dereference(rq->curr) may not see
5052 * changes to task_struct made by pick_next_task().
5054 RCU_INIT_POINTER(rq
->curr
, next
);
5056 * The membarrier system call requires each architecture
5057 * to have a full memory barrier after updating
5058 * rq->curr, before returning to user-space.
5060 * Here are the schemes providing that barrier on the
5061 * various architectures:
5062 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5063 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5064 * - finish_lock_switch() for weakly-ordered
5065 * architectures where spin_unlock is a full barrier,
5066 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5067 * is a RELEASE barrier),
5071 migrate_disable_switch(rq
, prev
);
5072 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
5074 trace_sched_switch(preempt
, prev
, next
);
5076 /* Also unlocks the rq: */
5077 rq
= context_switch(rq
, prev
, next
, &rf
);
5079 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
5081 rq_unpin_lock(rq
, &rf
);
5082 __balance_callbacks(rq
);
5083 raw_spin_unlock_irq(&rq
->lock
);
5087 void __noreturn
do_task_dead(void)
5089 /* Causes final put_task_struct in finish_task_switch(): */
5090 set_special_state(TASK_DEAD
);
5092 /* Tell freezer to ignore us: */
5093 current
->flags
|= PF_NOFREEZE
;
5098 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5103 static inline void sched_submit_work(struct task_struct
*tsk
)
5105 unsigned int task_flags
;
5110 task_flags
= tsk
->flags
;
5112 * If a worker went to sleep, notify and ask workqueue whether
5113 * it wants to wake up a task to maintain concurrency.
5114 * As this function is called inside the schedule() context,
5115 * we disable preemption to avoid it calling schedule() again
5116 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5119 if (task_flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
5121 if (task_flags
& PF_WQ_WORKER
)
5122 wq_worker_sleeping(tsk
);
5124 io_wq_worker_sleeping(tsk
);
5125 preempt_enable_no_resched();
5128 if (tsk_is_pi_blocked(tsk
))
5132 * If we are going to sleep and we have plugged IO queued,
5133 * make sure to submit it to avoid deadlocks.
5135 if (blk_needs_flush_plug(tsk
))
5136 blk_schedule_flush_plug(tsk
);
5139 static void sched_update_worker(struct task_struct
*tsk
)
5141 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
5142 if (tsk
->flags
& PF_WQ_WORKER
)
5143 wq_worker_running(tsk
);
5145 io_wq_worker_running(tsk
);
5149 asmlinkage __visible
void __sched
schedule(void)
5151 struct task_struct
*tsk
= current
;
5153 sched_submit_work(tsk
);
5157 sched_preempt_enable_no_resched();
5158 } while (need_resched());
5159 sched_update_worker(tsk
);
5161 EXPORT_SYMBOL(schedule
);
5164 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
5165 * state (have scheduled out non-voluntarily) by making sure that all
5166 * tasks have either left the run queue or have gone into user space.
5167 * As idle tasks do not do either, they must not ever be preempted
5168 * (schedule out non-voluntarily).
5170 * schedule_idle() is similar to schedule_preempt_disable() except that it
5171 * never enables preemption because it does not call sched_submit_work().
5173 void __sched
schedule_idle(void)
5176 * As this skips calling sched_submit_work(), which the idle task does
5177 * regardless because that function is a nop when the task is in a
5178 * TASK_RUNNING state, make sure this isn't used someplace that the
5179 * current task can be in any other state. Note, idle is always in the
5180 * TASK_RUNNING state.
5182 WARN_ON_ONCE(current
->state
);
5185 } while (need_resched());
5188 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
5189 asmlinkage __visible
void __sched
schedule_user(void)
5192 * If we come here after a random call to set_need_resched(),
5193 * or we have been woken up remotely but the IPI has not yet arrived,
5194 * we haven't yet exited the RCU idle mode. Do it here manually until
5195 * we find a better solution.
5197 * NB: There are buggy callers of this function. Ideally we
5198 * should warn if prev_state != CONTEXT_USER, but that will trigger
5199 * too frequently to make sense yet.
5201 enum ctx_state prev_state
= exception_enter();
5203 exception_exit(prev_state
);
5208 * schedule_preempt_disabled - called with preemption disabled
5210 * Returns with preemption disabled. Note: preempt_count must be 1
5212 void __sched
schedule_preempt_disabled(void)
5214 sched_preempt_enable_no_resched();
5219 static void __sched notrace
preempt_schedule_common(void)
5223 * Because the function tracer can trace preempt_count_sub()
5224 * and it also uses preempt_enable/disable_notrace(), if
5225 * NEED_RESCHED is set, the preempt_enable_notrace() called
5226 * by the function tracer will call this function again and
5227 * cause infinite recursion.
5229 * Preemption must be disabled here before the function
5230 * tracer can trace. Break up preempt_disable() into two
5231 * calls. One to disable preemption without fear of being
5232 * traced. The other to still record the preemption latency,
5233 * which can also be traced by the function tracer.
5235 preempt_disable_notrace();
5236 preempt_latency_start(1);
5238 preempt_latency_stop(1);
5239 preempt_enable_no_resched_notrace();
5242 * Check again in case we missed a preemption opportunity
5243 * between schedule and now.
5245 } while (need_resched());
5248 #ifdef CONFIG_PREEMPTION
5250 * This is the entry point to schedule() from in-kernel preemption
5251 * off of preempt_enable.
5253 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
5256 * If there is a non-zero preempt_count or interrupts are disabled,
5257 * we do not want to preempt the current task. Just return..
5259 if (likely(!preemptible()))
5262 preempt_schedule_common();
5264 NOKPROBE_SYMBOL(preempt_schedule
);
5265 EXPORT_SYMBOL(preempt_schedule
);
5268 * preempt_schedule_notrace - preempt_schedule called by tracing
5270 * The tracing infrastructure uses preempt_enable_notrace to prevent
5271 * recursion and tracing preempt enabling caused by the tracing
5272 * infrastructure itself. But as tracing can happen in areas coming
5273 * from userspace or just about to enter userspace, a preempt enable
5274 * can occur before user_exit() is called. This will cause the scheduler
5275 * to be called when the system is still in usermode.
5277 * To prevent this, the preempt_enable_notrace will use this function
5278 * instead of preempt_schedule() to exit user context if needed before
5279 * calling the scheduler.
5281 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
5283 enum ctx_state prev_ctx
;
5285 if (likely(!preemptible()))
5290 * Because the function tracer can trace preempt_count_sub()
5291 * and it also uses preempt_enable/disable_notrace(), if
5292 * NEED_RESCHED is set, the preempt_enable_notrace() called
5293 * by the function tracer will call this function again and
5294 * cause infinite recursion.
5296 * Preemption must be disabled here before the function
5297 * tracer can trace. Break up preempt_disable() into two
5298 * calls. One to disable preemption without fear of being
5299 * traced. The other to still record the preemption latency,
5300 * which can also be traced by the function tracer.
5302 preempt_disable_notrace();
5303 preempt_latency_start(1);
5305 * Needs preempt disabled in case user_exit() is traced
5306 * and the tracer calls preempt_enable_notrace() causing
5307 * an infinite recursion.
5309 prev_ctx
= exception_enter();
5311 exception_exit(prev_ctx
);
5313 preempt_latency_stop(1);
5314 preempt_enable_no_resched_notrace();
5315 } while (need_resched());
5317 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
5319 #endif /* CONFIG_PREEMPTION */
5322 * This is the entry point to schedule() from kernel preemption
5323 * off of irq context.
5324 * Note, that this is called and return with irqs disabled. This will
5325 * protect us against recursive calling from irq.
5327 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
5329 enum ctx_state prev_state
;
5331 /* Catch callers which need to be fixed */
5332 BUG_ON(preempt_count() || !irqs_disabled());
5334 prev_state
= exception_enter();
5340 local_irq_disable();
5341 sched_preempt_enable_no_resched();
5342 } while (need_resched());
5344 exception_exit(prev_state
);
5347 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
5350 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG
) && wake_flags
& ~WF_SYNC
);
5351 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5353 EXPORT_SYMBOL(default_wake_function
);
5355 #ifdef CONFIG_RT_MUTEXES
5357 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
5360 prio
= min(prio
, pi_task
->prio
);
5365 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
5367 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
5369 return __rt_effective_prio(pi_task
, prio
);
5373 * rt_mutex_setprio - set the current priority of a task
5375 * @pi_task: donor task
5377 * This function changes the 'effective' priority of a task. It does
5378 * not touch ->normal_prio like __setscheduler().
5380 * Used by the rt_mutex code to implement priority inheritance
5381 * logic. Call site only calls if the priority of the task changed.
5383 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
5385 int prio
, oldprio
, queued
, running
, queue_flag
=
5386 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
5387 const struct sched_class
*prev_class
;
5391 /* XXX used to be waiter->prio, not waiter->task->prio */
5392 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
5395 * If nothing changed; bail early.
5397 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
5400 rq
= __task_rq_lock(p
, &rf
);
5401 update_rq_clock(rq
);
5403 * Set under pi_lock && rq->lock, such that the value can be used under
5406 * Note that there is loads of tricky to make this pointer cache work
5407 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
5408 * ensure a task is de-boosted (pi_task is set to NULL) before the
5409 * task is allowed to run again (and can exit). This ensures the pointer
5410 * points to a blocked task -- which guarantees the task is present.
5412 p
->pi_top_task
= pi_task
;
5415 * For FIFO/RR we only need to set prio, if that matches we're done.
5417 if (prio
== p
->prio
&& !dl_prio(prio
))
5421 * Idle task boosting is a nono in general. There is one
5422 * exception, when PREEMPT_RT and NOHZ is active:
5424 * The idle task calls get_next_timer_interrupt() and holds
5425 * the timer wheel base->lock on the CPU and another CPU wants
5426 * to access the timer (probably to cancel it). We can safely
5427 * ignore the boosting request, as the idle CPU runs this code
5428 * with interrupts disabled and will complete the lock
5429 * protected section without being interrupted. So there is no
5430 * real need to boost.
5432 if (unlikely(p
== rq
->idle
)) {
5433 WARN_ON(p
!= rq
->curr
);
5434 WARN_ON(p
->pi_blocked_on
);
5438 trace_sched_pi_setprio(p
, pi_task
);
5441 if (oldprio
== prio
)
5442 queue_flag
&= ~DEQUEUE_MOVE
;
5444 prev_class
= p
->sched_class
;
5445 queued
= task_on_rq_queued(p
);
5446 running
= task_current(rq
, p
);
5448 dequeue_task(rq
, p
, queue_flag
);
5450 put_prev_task(rq
, p
);
5453 * Boosting condition are:
5454 * 1. -rt task is running and holds mutex A
5455 * --> -dl task blocks on mutex A
5457 * 2. -dl task is running and holds mutex A
5458 * --> -dl task blocks on mutex A and could preempt the
5461 if (dl_prio(prio
)) {
5462 if (!dl_prio(p
->normal_prio
) ||
5463 (pi_task
&& dl_prio(pi_task
->prio
) &&
5464 dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
5465 p
->dl
.pi_se
= pi_task
->dl
.pi_se
;
5466 queue_flag
|= ENQUEUE_REPLENISH
;
5468 p
->dl
.pi_se
= &p
->dl
;
5470 p
->sched_class
= &dl_sched_class
;
5471 } else if (rt_prio(prio
)) {
5472 if (dl_prio(oldprio
))
5473 p
->dl
.pi_se
= &p
->dl
;
5475 queue_flag
|= ENQUEUE_HEAD
;
5476 p
->sched_class
= &rt_sched_class
;
5478 if (dl_prio(oldprio
))
5479 p
->dl
.pi_se
= &p
->dl
;
5480 if (rt_prio(oldprio
))
5482 p
->sched_class
= &fair_sched_class
;
5488 enqueue_task(rq
, p
, queue_flag
);
5490 set_next_task(rq
, p
);
5492 check_class_changed(rq
, p
, prev_class
, oldprio
);
5494 /* Avoid rq from going away on us: */
5497 rq_unpin_lock(rq
, &rf
);
5498 __balance_callbacks(rq
);
5499 raw_spin_unlock(&rq
->lock
);
5504 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
5510 void set_user_nice(struct task_struct
*p
, long nice
)
5512 bool queued
, running
;
5517 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
5520 * We have to be careful, if called from sys_setpriority(),
5521 * the task might be in the middle of scheduling on another CPU.
5523 rq
= task_rq_lock(p
, &rf
);
5524 update_rq_clock(rq
);
5527 * The RT priorities are set via sched_setscheduler(), but we still
5528 * allow the 'normal' nice value to be set - but as expected
5529 * it won't have any effect on scheduling until the task is
5530 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5532 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
5533 p
->static_prio
= NICE_TO_PRIO(nice
);
5536 queued
= task_on_rq_queued(p
);
5537 running
= task_current(rq
, p
);
5539 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
5541 put_prev_task(rq
, p
);
5543 p
->static_prio
= NICE_TO_PRIO(nice
);
5544 set_load_weight(p
, true);
5546 p
->prio
= effective_prio(p
);
5549 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5551 set_next_task(rq
, p
);
5554 * If the task increased its priority or is running and
5555 * lowered its priority, then reschedule its CPU:
5557 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
5560 task_rq_unlock(rq
, p
, &rf
);
5562 EXPORT_SYMBOL(set_user_nice
);
5565 * can_nice - check if a task can reduce its nice value
5569 int can_nice(const struct task_struct
*p
, const int nice
)
5571 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5572 int nice_rlim
= nice_to_rlimit(nice
);
5574 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
5575 capable(CAP_SYS_NICE
));
5578 #ifdef __ARCH_WANT_SYS_NICE
5581 * sys_nice - change the priority of the current process.
5582 * @increment: priority increment
5584 * sys_setpriority is a more generic, but much slower function that
5585 * does similar things.
5587 SYSCALL_DEFINE1(nice
, int, increment
)
5592 * Setpriority might change our priority at the same moment.
5593 * We don't have to worry. Conceptually one call occurs first
5594 * and we have a single winner.
5596 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
5597 nice
= task_nice(current
) + increment
;
5599 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
5600 if (increment
< 0 && !can_nice(current
, nice
))
5603 retval
= security_task_setnice(current
, nice
);
5607 set_user_nice(current
, nice
);
5614 * task_prio - return the priority value of a given task.
5615 * @p: the task in question.
5617 * Return: The priority value as seen by users in /proc.
5618 * RT tasks are offset by -200. Normal tasks are centered
5619 * around 0, value goes from -16 to +15.
5621 int task_prio(const struct task_struct
*p
)
5623 return p
->prio
- MAX_RT_PRIO
;
5627 * idle_cpu - is a given CPU idle currently?
5628 * @cpu: the processor in question.
5630 * Return: 1 if the CPU is currently idle. 0 otherwise.
5632 int idle_cpu(int cpu
)
5634 struct rq
*rq
= cpu_rq(cpu
);
5636 if (rq
->curr
!= rq
->idle
)
5643 if (rq
->ttwu_pending
)
5651 * available_idle_cpu - is a given CPU idle for enqueuing work.
5652 * @cpu: the CPU in question.
5654 * Return: 1 if the CPU is currently idle. 0 otherwise.
5656 int available_idle_cpu(int cpu
)
5661 if (vcpu_is_preempted(cpu
))
5668 * idle_task - return the idle task for a given CPU.
5669 * @cpu: the processor in question.
5671 * Return: The idle task for the CPU @cpu.
5673 struct task_struct
*idle_task(int cpu
)
5675 return cpu_rq(cpu
)->idle
;
5679 * find_process_by_pid - find a process with a matching PID value.
5680 * @pid: the pid in question.
5682 * The task of @pid, if found. %NULL otherwise.
5684 static struct task_struct
*find_process_by_pid(pid_t pid
)
5686 return pid
? find_task_by_vpid(pid
) : current
;
5690 * sched_setparam() passes in -1 for its policy, to let the functions
5691 * it calls know not to change it.
5693 #define SETPARAM_POLICY -1
5695 static void __setscheduler_params(struct task_struct
*p
,
5696 const struct sched_attr
*attr
)
5698 int policy
= attr
->sched_policy
;
5700 if (policy
== SETPARAM_POLICY
)
5705 if (dl_policy(policy
))
5706 __setparam_dl(p
, attr
);
5707 else if (fair_policy(policy
))
5708 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
5711 * __sched_setscheduler() ensures attr->sched_priority == 0 when
5712 * !rt_policy. Always setting this ensures that things like
5713 * getparam()/getattr() don't report silly values for !rt tasks.
5715 p
->rt_priority
= attr
->sched_priority
;
5716 p
->normal_prio
= normal_prio(p
);
5717 set_load_weight(p
, true);
5720 /* Actually do priority change: must hold pi & rq lock. */
5721 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
5722 const struct sched_attr
*attr
, bool keep_boost
)
5725 * If params can't change scheduling class changes aren't allowed
5728 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
5731 __setscheduler_params(p
, attr
);
5734 * Keep a potential priority boosting if called from
5735 * sched_setscheduler().
5737 p
->prio
= normal_prio(p
);
5739 p
->prio
= rt_effective_prio(p
, p
->prio
);
5741 if (dl_prio(p
->prio
))
5742 p
->sched_class
= &dl_sched_class
;
5743 else if (rt_prio(p
->prio
))
5744 p
->sched_class
= &rt_sched_class
;
5746 p
->sched_class
= &fair_sched_class
;
5750 * Check the target process has a UID that matches the current process's:
5752 static bool check_same_owner(struct task_struct
*p
)
5754 const struct cred
*cred
= current_cred(), *pcred
;
5758 pcred
= __task_cred(p
);
5759 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
5760 uid_eq(cred
->euid
, pcred
->uid
));
5765 static int __sched_setscheduler(struct task_struct
*p
,
5766 const struct sched_attr
*attr
,
5769 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
5770 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
5771 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
5772 int new_effective_prio
, policy
= attr
->sched_policy
;
5773 const struct sched_class
*prev_class
;
5774 struct callback_head
*head
;
5777 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
5780 /* The pi code expects interrupts enabled */
5781 BUG_ON(pi
&& in_interrupt());
5783 /* Double check policy once rq lock held: */
5785 reset_on_fork
= p
->sched_reset_on_fork
;
5786 policy
= oldpolicy
= p
->policy
;
5788 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
5790 if (!valid_policy(policy
))
5794 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
5798 * Valid priorities for SCHED_FIFO and SCHED_RR are
5799 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5800 * SCHED_BATCH and SCHED_IDLE is 0.
5802 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5803 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
5805 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
5806 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
5810 * Allow unprivileged RT tasks to decrease priority:
5812 if (user
&& !capable(CAP_SYS_NICE
)) {
5813 if (fair_policy(policy
)) {
5814 if (attr
->sched_nice
< task_nice(p
) &&
5815 !can_nice(p
, attr
->sched_nice
))
5819 if (rt_policy(policy
)) {
5820 unsigned long rlim_rtprio
=
5821 task_rlimit(p
, RLIMIT_RTPRIO
);
5823 /* Can't set/change the rt policy: */
5824 if (policy
!= p
->policy
&& !rlim_rtprio
)
5827 /* Can't increase priority: */
5828 if (attr
->sched_priority
> p
->rt_priority
&&
5829 attr
->sched_priority
> rlim_rtprio
)
5834 * Can't set/change SCHED_DEADLINE policy at all for now
5835 * (safest behavior); in the future we would like to allow
5836 * unprivileged DL tasks to increase their relative deadline
5837 * or reduce their runtime (both ways reducing utilization)
5839 if (dl_policy(policy
))
5843 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5844 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5846 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
5847 if (!can_nice(p
, task_nice(p
)))
5851 /* Can't change other user's priorities: */
5852 if (!check_same_owner(p
))
5855 /* Normal users shall not reset the sched_reset_on_fork flag: */
5856 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
5861 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
5864 retval
= security_task_setscheduler(p
);
5869 /* Update task specific "requested" clamps */
5870 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
5871 retval
= uclamp_validate(p
, attr
);
5880 * Make sure no PI-waiters arrive (or leave) while we are
5881 * changing the priority of the task:
5883 * To be able to change p->policy safely, the appropriate
5884 * runqueue lock must be held.
5886 rq
= task_rq_lock(p
, &rf
);
5887 update_rq_clock(rq
);
5890 * Changing the policy of the stop threads its a very bad idea:
5892 if (p
== rq
->stop
) {
5898 * If not changing anything there's no need to proceed further,
5899 * but store a possible modification of reset_on_fork.
5901 if (unlikely(policy
== p
->policy
)) {
5902 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
5904 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
5906 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
5908 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
5911 p
->sched_reset_on_fork
= reset_on_fork
;
5918 #ifdef CONFIG_RT_GROUP_SCHED
5920 * Do not allow realtime tasks into groups that have no runtime
5923 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5924 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5925 !task_group_is_autogroup(task_group(p
))) {
5931 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
5932 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
5933 cpumask_t
*span
= rq
->rd
->span
;
5936 * Don't allow tasks with an affinity mask smaller than
5937 * the entire root_domain to become SCHED_DEADLINE. We
5938 * will also fail if there's no bandwidth available.
5940 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
5941 rq
->rd
->dl_bw
.bw
== 0) {
5949 /* Re-check policy now with rq lock held: */
5950 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5951 policy
= oldpolicy
= -1;
5952 task_rq_unlock(rq
, p
, &rf
);
5954 cpuset_read_unlock();
5959 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5960 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5963 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
5968 p
->sched_reset_on_fork
= reset_on_fork
;
5973 * Take priority boosted tasks into account. If the new
5974 * effective priority is unchanged, we just store the new
5975 * normal parameters and do not touch the scheduler class and
5976 * the runqueue. This will be done when the task deboost
5979 new_effective_prio
= rt_effective_prio(p
, newprio
);
5980 if (new_effective_prio
== oldprio
)
5981 queue_flags
&= ~DEQUEUE_MOVE
;
5984 queued
= task_on_rq_queued(p
);
5985 running
= task_current(rq
, p
);
5987 dequeue_task(rq
, p
, queue_flags
);
5989 put_prev_task(rq
, p
);
5991 prev_class
= p
->sched_class
;
5993 __setscheduler(rq
, p
, attr
, pi
);
5994 __setscheduler_uclamp(p
, attr
);
5998 * We enqueue to tail when the priority of a task is
5999 * increased (user space view).
6001 if (oldprio
< p
->prio
)
6002 queue_flags
|= ENQUEUE_HEAD
;
6004 enqueue_task(rq
, p
, queue_flags
);
6007 set_next_task(rq
, p
);
6009 check_class_changed(rq
, p
, prev_class
, oldprio
);
6011 /* Avoid rq from going away on us: */
6013 head
= splice_balance_callbacks(rq
);
6014 task_rq_unlock(rq
, p
, &rf
);
6017 cpuset_read_unlock();
6018 rt_mutex_adjust_pi(p
);
6021 /* Run balance callbacks after we've adjusted the PI chain: */
6022 balance_callbacks(rq
, head
);
6028 task_rq_unlock(rq
, p
, &rf
);
6030 cpuset_read_unlock();
6034 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
6035 const struct sched_param
*param
, bool check
)
6037 struct sched_attr attr
= {
6038 .sched_policy
= policy
,
6039 .sched_priority
= param
->sched_priority
,
6040 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
6043 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
6044 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
6045 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
6046 policy
&= ~SCHED_RESET_ON_FORK
;
6047 attr
.sched_policy
= policy
;
6050 return __sched_setscheduler(p
, &attr
, check
, true);
6053 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6054 * @p: the task in question.
6055 * @policy: new policy.
6056 * @param: structure containing the new RT priority.
6058 * Use sched_set_fifo(), read its comment.
6060 * Return: 0 on success. An error code otherwise.
6062 * NOTE that the task may be already dead.
6064 int sched_setscheduler(struct task_struct
*p
, int policy
,
6065 const struct sched_param
*param
)
6067 return _sched_setscheduler(p
, policy
, param
, true);
6070 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
6072 return __sched_setscheduler(p
, attr
, true, true);
6075 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
6077 return __sched_setscheduler(p
, attr
, false, true);
6081 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6082 * @p: the task in question.
6083 * @policy: new policy.
6084 * @param: structure containing the new RT priority.
6086 * Just like sched_setscheduler, only don't bother checking if the
6087 * current context has permission. For example, this is needed in
6088 * stop_machine(): we create temporary high priority worker threads,
6089 * but our caller might not have that capability.
6091 * Return: 0 on success. An error code otherwise.
6093 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6094 const struct sched_param
*param
)
6096 return _sched_setscheduler(p
, policy
, param
, false);
6100 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
6101 * incapable of resource management, which is the one thing an OS really should
6104 * This is of course the reason it is limited to privileged users only.
6106 * Worse still; it is fundamentally impossible to compose static priority
6107 * workloads. You cannot take two correctly working static prio workloads
6108 * and smash them together and still expect them to work.
6110 * For this reason 'all' FIFO tasks the kernel creates are basically at:
6114 * The administrator _MUST_ configure the system, the kernel simply doesn't
6115 * know enough information to make a sensible choice.
6117 void sched_set_fifo(struct task_struct
*p
)
6119 struct sched_param sp
= { .sched_priority
= MAX_RT_PRIO
/ 2 };
6120 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
6122 EXPORT_SYMBOL_GPL(sched_set_fifo
);
6125 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
6127 void sched_set_fifo_low(struct task_struct
*p
)
6129 struct sched_param sp
= { .sched_priority
= 1 };
6130 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
6132 EXPORT_SYMBOL_GPL(sched_set_fifo_low
);
6134 void sched_set_normal(struct task_struct
*p
, int nice
)
6136 struct sched_attr attr
= {
6137 .sched_policy
= SCHED_NORMAL
,
6140 WARN_ON_ONCE(sched_setattr_nocheck(p
, &attr
) != 0);
6142 EXPORT_SYMBOL_GPL(sched_set_normal
);
6145 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6147 struct sched_param lparam
;
6148 struct task_struct
*p
;
6151 if (!param
|| pid
< 0)
6153 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6158 p
= find_process_by_pid(pid
);
6164 retval
= sched_setscheduler(p
, policy
, &lparam
);
6172 * Mimics kernel/events/core.c perf_copy_attr().
6174 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
6179 /* Zero the full structure, so that a short copy will be nice: */
6180 memset(attr
, 0, sizeof(*attr
));
6182 ret
= get_user(size
, &uattr
->size
);
6186 /* ABI compatibility quirk: */
6188 size
= SCHED_ATTR_SIZE_VER0
;
6189 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
6192 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
6199 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
6200 size
< SCHED_ATTR_SIZE_VER1
)
6204 * XXX: Do we want to be lenient like existing syscalls; or do we want
6205 * to be strict and return an error on out-of-bounds values?
6207 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
6212 put_user(sizeof(*attr
), &uattr
->size
);
6217 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6218 * @pid: the pid in question.
6219 * @policy: new policy.
6220 * @param: structure containing the new RT priority.
6222 * Return: 0 on success. An error code otherwise.
6224 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
6229 return do_sched_setscheduler(pid
, policy
, param
);
6233 * sys_sched_setparam - set/change the RT priority of a thread
6234 * @pid: the pid in question.
6235 * @param: structure containing the new RT priority.
6237 * Return: 0 on success. An error code otherwise.
6239 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6241 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
6245 * sys_sched_setattr - same as above, but with extended sched_attr
6246 * @pid: the pid in question.
6247 * @uattr: structure containing the extended parameters.
6248 * @flags: for future extension.
6250 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
6251 unsigned int, flags
)
6253 struct sched_attr attr
;
6254 struct task_struct
*p
;
6257 if (!uattr
|| pid
< 0 || flags
)
6260 retval
= sched_copy_attr(uattr
, &attr
);
6264 if ((int)attr
.sched_policy
< 0)
6266 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
6267 attr
.sched_policy
= SETPARAM_POLICY
;
6271 p
= find_process_by_pid(pid
);
6277 retval
= sched_setattr(p
, &attr
);
6285 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6286 * @pid: the pid in question.
6288 * Return: On success, the policy of the thread. Otherwise, a negative error
6291 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6293 struct task_struct
*p
;
6301 p
= find_process_by_pid(pid
);
6303 retval
= security_task_getscheduler(p
);
6306 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6313 * sys_sched_getparam - get the RT priority of a thread
6314 * @pid: the pid in question.
6315 * @param: structure containing the RT priority.
6317 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
6320 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6322 struct sched_param lp
= { .sched_priority
= 0 };
6323 struct task_struct
*p
;
6326 if (!param
|| pid
< 0)
6330 p
= find_process_by_pid(pid
);
6335 retval
= security_task_getscheduler(p
);
6339 if (task_has_rt_policy(p
))
6340 lp
.sched_priority
= p
->rt_priority
;
6344 * This one might sleep, we cannot do it with a spinlock held ...
6346 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6356 * Copy the kernel size attribute structure (which might be larger
6357 * than what user-space knows about) to user-space.
6359 * Note that all cases are valid: user-space buffer can be larger or
6360 * smaller than the kernel-space buffer. The usual case is that both
6361 * have the same size.
6364 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
6365 struct sched_attr
*kattr
,
6368 unsigned int ksize
= sizeof(*kattr
);
6370 if (!access_ok(uattr
, usize
))
6374 * sched_getattr() ABI forwards and backwards compatibility:
6376 * If usize == ksize then we just copy everything to user-space and all is good.
6378 * If usize < ksize then we only copy as much as user-space has space for,
6379 * this keeps ABI compatibility as well. We skip the rest.
6381 * If usize > ksize then user-space is using a newer version of the ABI,
6382 * which part the kernel doesn't know about. Just ignore it - tooling can
6383 * detect the kernel's knowledge of attributes from the attr->size value
6384 * which is set to ksize in this case.
6386 kattr
->size
= min(usize
, ksize
);
6388 if (copy_to_user(uattr
, kattr
, kattr
->size
))
6395 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
6396 * @pid: the pid in question.
6397 * @uattr: structure containing the extended parameters.
6398 * @usize: sizeof(attr) for fwd/bwd comp.
6399 * @flags: for future extension.
6401 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
6402 unsigned int, usize
, unsigned int, flags
)
6404 struct sched_attr kattr
= { };
6405 struct task_struct
*p
;
6408 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
6409 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
6413 p
= find_process_by_pid(pid
);
6418 retval
= security_task_getscheduler(p
);
6422 kattr
.sched_policy
= p
->policy
;
6423 if (p
->sched_reset_on_fork
)
6424 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
6425 if (task_has_dl_policy(p
))
6426 __getparam_dl(p
, &kattr
);
6427 else if (task_has_rt_policy(p
))
6428 kattr
.sched_priority
= p
->rt_priority
;
6430 kattr
.sched_nice
= task_nice(p
);
6432 #ifdef CONFIG_UCLAMP_TASK
6434 * This could race with another potential updater, but this is fine
6435 * because it'll correctly read the old or the new value. We don't need
6436 * to guarantee who wins the race as long as it doesn't return garbage.
6438 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
6439 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
6444 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
6451 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6453 cpumask_var_t cpus_allowed
, new_mask
;
6454 struct task_struct
*p
;
6459 p
= find_process_by_pid(pid
);
6465 /* Prevent p going away */
6469 if (p
->flags
& PF_NO_SETAFFINITY
) {
6473 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6477 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6479 goto out_free_cpus_allowed
;
6482 if (!check_same_owner(p
)) {
6484 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
6486 goto out_free_new_mask
;
6491 retval
= security_task_setscheduler(p
);
6493 goto out_free_new_mask
;
6496 cpuset_cpus_allowed(p
, cpus_allowed
);
6497 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6500 * Since bandwidth control happens on root_domain basis,
6501 * if admission test is enabled, we only admit -deadline
6502 * tasks allowed to run on all the CPUs in the task's
6506 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
6508 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
6511 goto out_free_new_mask
;
6517 retval
= __set_cpus_allowed_ptr(p
, new_mask
, SCA_CHECK
);
6520 cpuset_cpus_allowed(p
, cpus_allowed
);
6521 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6523 * We must have raced with a concurrent cpuset
6524 * update. Just reset the cpus_allowed to the
6525 * cpuset's cpus_allowed
6527 cpumask_copy(new_mask
, cpus_allowed
);
6532 free_cpumask_var(new_mask
);
6533 out_free_cpus_allowed
:
6534 free_cpumask_var(cpus_allowed
);
6540 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6541 struct cpumask
*new_mask
)
6543 if (len
< cpumask_size())
6544 cpumask_clear(new_mask
);
6545 else if (len
> cpumask_size())
6546 len
= cpumask_size();
6548 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6552 * sys_sched_setaffinity - set the CPU affinity of a process
6553 * @pid: pid of the process
6554 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6555 * @user_mask_ptr: user-space pointer to the new CPU mask
6557 * Return: 0 on success. An error code otherwise.
6559 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6560 unsigned long __user
*, user_mask_ptr
)
6562 cpumask_var_t new_mask
;
6565 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6568 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6570 retval
= sched_setaffinity(pid
, new_mask
);
6571 free_cpumask_var(new_mask
);
6575 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6577 struct task_struct
*p
;
6578 unsigned long flags
;
6584 p
= find_process_by_pid(pid
);
6588 retval
= security_task_getscheduler(p
);
6592 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
6593 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
6594 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6603 * sys_sched_getaffinity - get the CPU affinity of a process
6604 * @pid: pid of the process
6605 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6606 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6608 * Return: size of CPU mask copied to user_mask_ptr on success. An
6609 * error code otherwise.
6611 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6612 unsigned long __user
*, user_mask_ptr
)
6617 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
6619 if (len
& (sizeof(unsigned long)-1))
6622 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6625 ret
= sched_getaffinity(pid
, mask
);
6627 unsigned int retlen
= min(len
, cpumask_size());
6629 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
6634 free_cpumask_var(mask
);
6639 static void do_sched_yield(void)
6644 rq
= this_rq_lock_irq(&rf
);
6646 schedstat_inc(rq
->yld_count
);
6647 current
->sched_class
->yield_task(rq
);
6650 rq_unlock_irq(rq
, &rf
);
6651 sched_preempt_enable_no_resched();
6657 * sys_sched_yield - yield the current processor to other threads.
6659 * This function yields the current CPU to other tasks. If there are no
6660 * other threads running on this CPU then this function will return.
6664 SYSCALL_DEFINE0(sched_yield
)
6670 #ifndef CONFIG_PREEMPTION
6671 int __sched
_cond_resched(void)
6673 if (should_resched(0)) {
6674 preempt_schedule_common();
6680 EXPORT_SYMBOL(_cond_resched
);
6684 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6685 * call schedule, and on return reacquire the lock.
6687 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
6688 * operations here to prevent schedule() from being called twice (once via
6689 * spin_unlock(), once by hand).
6691 int __cond_resched_lock(spinlock_t
*lock
)
6693 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
6696 lockdep_assert_held(lock
);
6698 if (spin_needbreak(lock
) || resched
) {
6701 preempt_schedule_common();
6709 EXPORT_SYMBOL(__cond_resched_lock
);
6712 * yield - yield the current processor to other threads.
6714 * Do not ever use this function, there's a 99% chance you're doing it wrong.
6716 * The scheduler is at all times free to pick the calling task as the most
6717 * eligible task to run, if removing the yield() call from your code breaks
6718 * it, it's already broken.
6720 * Typical broken usage is:
6725 * where one assumes that yield() will let 'the other' process run that will
6726 * make event true. If the current task is a SCHED_FIFO task that will never
6727 * happen. Never use yield() as a progress guarantee!!
6729 * If you want to use yield() to wait for something, use wait_event().
6730 * If you want to use yield() to be 'nice' for others, use cond_resched().
6731 * If you still want to use yield(), do not!
6733 void __sched
yield(void)
6735 set_current_state(TASK_RUNNING
);
6738 EXPORT_SYMBOL(yield
);
6741 * yield_to - yield the current processor to another thread in
6742 * your thread group, or accelerate that thread toward the
6743 * processor it's on.
6745 * @preempt: whether task preemption is allowed or not
6747 * It's the caller's job to ensure that the target task struct
6748 * can't go away on us before we can do any checks.
6751 * true (>0) if we indeed boosted the target task.
6752 * false (0) if we failed to boost the target.
6753 * -ESRCH if there's no task to yield to.
6755 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
6757 struct task_struct
*curr
= current
;
6758 struct rq
*rq
, *p_rq
;
6759 unsigned long flags
;
6762 local_irq_save(flags
);
6768 * If we're the only runnable task on the rq and target rq also
6769 * has only one task, there's absolutely no point in yielding.
6771 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
6776 double_rq_lock(rq
, p_rq
);
6777 if (task_rq(p
) != p_rq
) {
6778 double_rq_unlock(rq
, p_rq
);
6782 if (!curr
->sched_class
->yield_to_task
)
6785 if (curr
->sched_class
!= p
->sched_class
)
6788 if (task_running(p_rq
, p
) || p
->state
)
6791 yielded
= curr
->sched_class
->yield_to_task(rq
, p
);
6793 schedstat_inc(rq
->yld_count
);
6795 * Make p's CPU reschedule; pick_next_entity takes care of
6798 if (preempt
&& rq
!= p_rq
)
6803 double_rq_unlock(rq
, p_rq
);
6805 local_irq_restore(flags
);
6812 EXPORT_SYMBOL_GPL(yield_to
);
6814 int io_schedule_prepare(void)
6816 int old_iowait
= current
->in_iowait
;
6818 current
->in_iowait
= 1;
6819 blk_schedule_flush_plug(current
);
6824 void io_schedule_finish(int token
)
6826 current
->in_iowait
= token
;
6830 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6831 * that process accounting knows that this is a task in IO wait state.
6833 long __sched
io_schedule_timeout(long timeout
)
6838 token
= io_schedule_prepare();
6839 ret
= schedule_timeout(timeout
);
6840 io_schedule_finish(token
);
6844 EXPORT_SYMBOL(io_schedule_timeout
);
6846 void __sched
io_schedule(void)
6850 token
= io_schedule_prepare();
6852 io_schedule_finish(token
);
6854 EXPORT_SYMBOL(io_schedule
);
6857 * sys_sched_get_priority_max - return maximum RT priority.
6858 * @policy: scheduling class.
6860 * Return: On success, this syscall returns the maximum
6861 * rt_priority that can be used by a given scheduling class.
6862 * On failure, a negative error code is returned.
6864 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6871 ret
= MAX_USER_RT_PRIO
-1;
6873 case SCHED_DEADLINE
:
6884 * sys_sched_get_priority_min - return minimum RT priority.
6885 * @policy: scheduling class.
6887 * Return: On success, this syscall returns the minimum
6888 * rt_priority that can be used by a given scheduling class.
6889 * On failure, a negative error code is returned.
6891 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6900 case SCHED_DEADLINE
:
6909 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
6911 struct task_struct
*p
;
6912 unsigned int time_slice
;
6922 p
= find_process_by_pid(pid
);
6926 retval
= security_task_getscheduler(p
);
6930 rq
= task_rq_lock(p
, &rf
);
6932 if (p
->sched_class
->get_rr_interval
)
6933 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
6934 task_rq_unlock(rq
, p
, &rf
);
6937 jiffies_to_timespec64(time_slice
, t
);
6946 * sys_sched_rr_get_interval - return the default timeslice of a process.
6947 * @pid: pid of the process.
6948 * @interval: userspace pointer to the timeslice value.
6950 * this syscall writes the default timeslice value of a given process
6951 * into the user-space timespec buffer. A value of '0' means infinity.
6953 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6956 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6957 struct __kernel_timespec __user
*, interval
)
6959 struct timespec64 t
;
6960 int retval
= sched_rr_get_interval(pid
, &t
);
6963 retval
= put_timespec64(&t
, interval
);
6968 #ifdef CONFIG_COMPAT_32BIT_TIME
6969 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
6970 struct old_timespec32 __user
*, interval
)
6972 struct timespec64 t
;
6973 int retval
= sched_rr_get_interval(pid
, &t
);
6976 retval
= put_old_timespec32(&t
, interval
);
6981 void sched_show_task(struct task_struct
*p
)
6983 unsigned long free
= 0;
6986 if (!try_get_task_stack(p
))
6989 pr_info("task:%-15.15s state:%c", p
->comm
, task_state_to_char(p
));
6991 if (p
->state
== TASK_RUNNING
)
6992 pr_cont(" running task ");
6993 #ifdef CONFIG_DEBUG_STACK_USAGE
6994 free
= stack_not_used(p
);
6999 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
7001 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
7002 free
, task_pid_nr(p
), ppid
,
7003 (unsigned long)task_thread_info(p
)->flags
);
7005 print_worker_info(KERN_INFO
, p
);
7006 print_stop_info(KERN_INFO
, p
);
7007 show_stack(p
, NULL
, KERN_INFO
);
7010 EXPORT_SYMBOL_GPL(sched_show_task
);
7013 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
7015 /* no filter, everything matches */
7019 /* filter, but doesn't match */
7020 if (!(p
->state
& state_filter
))
7024 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
7027 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
7034 void show_state_filter(unsigned long state_filter
)
7036 struct task_struct
*g
, *p
;
7039 for_each_process_thread(g
, p
) {
7041 * reset the NMI-timeout, listing all files on a slow
7042 * console might take a lot of time:
7043 * Also, reset softlockup watchdogs on all CPUs, because
7044 * another CPU might be blocked waiting for us to process
7047 touch_nmi_watchdog();
7048 touch_all_softlockup_watchdogs();
7049 if (state_filter_match(state_filter
, p
))
7053 #ifdef CONFIG_SCHED_DEBUG
7055 sysrq_sched_debug_show();
7059 * Only show locks if all tasks are dumped:
7062 debug_show_all_locks();
7066 * init_idle - set up an idle thread for a given CPU
7067 * @idle: task in question
7068 * @cpu: CPU the idle task belongs to
7070 * NOTE: this function does not set the idle thread's NEED_RESCHED
7071 * flag, to make booting more robust.
7073 void init_idle(struct task_struct
*idle
, int cpu
)
7075 struct rq
*rq
= cpu_rq(cpu
);
7076 unsigned long flags
;
7078 __sched_fork(0, idle
);
7080 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
7081 raw_spin_lock(&rq
->lock
);
7083 idle
->state
= TASK_RUNNING
;
7084 idle
->se
.exec_start
= sched_clock();
7085 idle
->flags
|= PF_IDLE
;
7087 scs_task_reset(idle
);
7088 kasan_unpoison_task_stack(idle
);
7092 * It's possible that init_idle() gets called multiple times on a task,
7093 * in that case do_set_cpus_allowed() will not do the right thing.
7095 * And since this is boot we can forgo the serialization.
7097 set_cpus_allowed_common(idle
, cpumask_of(cpu
), 0);
7100 * We're having a chicken and egg problem, even though we are
7101 * holding rq->lock, the CPU isn't yet set to this CPU so the
7102 * lockdep check in task_group() will fail.
7104 * Similar case to sched_fork(). / Alternatively we could
7105 * use task_rq_lock() here and obtain the other rq->lock.
7110 __set_task_cpu(idle
, cpu
);
7114 rcu_assign_pointer(rq
->curr
, idle
);
7115 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
7119 raw_spin_unlock(&rq
->lock
);
7120 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
7122 /* Set the preempt count _outside_ the spinlocks! */
7123 init_idle_preempt_count(idle
, cpu
);
7126 * The idle tasks have their own, simple scheduling class:
7128 idle
->sched_class
= &idle_sched_class
;
7129 ftrace_graph_init_idle_task(idle
, cpu
);
7130 vtime_init_idle(idle
, cpu
);
7132 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
7138 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
7139 const struct cpumask
*trial
)
7143 if (!cpumask_weight(cur
))
7146 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
7151 int task_can_attach(struct task_struct
*p
,
7152 const struct cpumask
*cs_cpus_allowed
)
7157 * Kthreads which disallow setaffinity shouldn't be moved
7158 * to a new cpuset; we don't want to change their CPU
7159 * affinity and isolating such threads by their set of
7160 * allowed nodes is unnecessary. Thus, cpusets are not
7161 * applicable for such threads. This prevents checking for
7162 * success of set_cpus_allowed_ptr() on all attached tasks
7163 * before cpus_mask may be changed.
7165 if (p
->flags
& PF_NO_SETAFFINITY
) {
7170 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
7172 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
7178 bool sched_smp_initialized __read_mostly
;
7180 #ifdef CONFIG_NUMA_BALANCING
7181 /* Migrate current task p to target_cpu */
7182 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
7184 struct migration_arg arg
= { p
, target_cpu
};
7185 int curr_cpu
= task_cpu(p
);
7187 if (curr_cpu
== target_cpu
)
7190 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
7193 /* TODO: This is not properly updating schedstats */
7195 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
7196 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
7200 * Requeue a task on a given node and accurately track the number of NUMA
7201 * tasks on the runqueues
7203 void sched_setnuma(struct task_struct
*p
, int nid
)
7205 bool queued
, running
;
7209 rq
= task_rq_lock(p
, &rf
);
7210 queued
= task_on_rq_queued(p
);
7211 running
= task_current(rq
, p
);
7214 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
7216 put_prev_task(rq
, p
);
7218 p
->numa_preferred_nid
= nid
;
7221 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
7223 set_next_task(rq
, p
);
7224 task_rq_unlock(rq
, p
, &rf
);
7226 #endif /* CONFIG_NUMA_BALANCING */
7228 #ifdef CONFIG_HOTPLUG_CPU
7230 * Ensure that the idle task is using init_mm right before its CPU goes
7233 void idle_task_exit(void)
7235 struct mm_struct
*mm
= current
->active_mm
;
7237 BUG_ON(cpu_online(smp_processor_id()));
7238 BUG_ON(current
!= this_rq()->idle
);
7240 if (mm
!= &init_mm
) {
7241 switch_mm(mm
, &init_mm
, current
);
7242 finish_arch_post_lock_switch();
7245 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
7248 static int __balance_push_cpu_stop(void *arg
)
7250 struct task_struct
*p
= arg
;
7251 struct rq
*rq
= this_rq();
7255 raw_spin_lock_irq(&p
->pi_lock
);
7258 update_rq_clock(rq
);
7260 if (task_rq(p
) == rq
&& task_on_rq_queued(p
)) {
7261 cpu
= select_fallback_rq(rq
->cpu
, p
);
7262 rq
= __migrate_task(rq
, &rf
, p
, cpu
);
7266 raw_spin_unlock_irq(&p
->pi_lock
);
7273 static DEFINE_PER_CPU(struct cpu_stop_work
, push_work
);
7276 * Ensure we only run per-cpu kthreads once the CPU goes !active.
7278 static void balance_push(struct rq
*rq
)
7280 struct task_struct
*push_task
= rq
->curr
;
7282 lockdep_assert_held(&rq
->lock
);
7283 SCHED_WARN_ON(rq
->cpu
!= smp_processor_id());
7285 * Ensure the thing is persistent until balance_push_set(.on = false);
7287 rq
->balance_callback
= &balance_push_callback
;
7290 * Both the cpu-hotplug and stop task are in this case and are
7291 * required to complete the hotplug process.
7293 * XXX: the idle task does not match kthread_is_per_cpu() due to
7294 * histerical raisins.
7296 if (rq
->idle
== push_task
||
7297 ((push_task
->flags
& PF_KTHREAD
) && kthread_is_per_cpu(push_task
)) ||
7298 is_migration_disabled(push_task
)) {
7301 * If this is the idle task on the outgoing CPU try to wake
7302 * up the hotplug control thread which might wait for the
7303 * last task to vanish. The rcuwait_active() check is
7304 * accurate here because the waiter is pinned on this CPU
7305 * and can't obviously be running in parallel.
7307 * On RT kernels this also has to check whether there are
7308 * pinned and scheduled out tasks on the runqueue. They
7309 * need to leave the migrate disabled section first.
7311 if (!rq
->nr_running
&& !rq_has_pinned_tasks(rq
) &&
7312 rcuwait_active(&rq
->hotplug_wait
)) {
7313 raw_spin_unlock(&rq
->lock
);
7314 rcuwait_wake_up(&rq
->hotplug_wait
);
7315 raw_spin_lock(&rq
->lock
);
7320 get_task_struct(push_task
);
7322 * Temporarily drop rq->lock such that we can wake-up the stop task.
7323 * Both preemption and IRQs are still disabled.
7325 raw_spin_unlock(&rq
->lock
);
7326 stop_one_cpu_nowait(rq
->cpu
, __balance_push_cpu_stop
, push_task
,
7327 this_cpu_ptr(&push_work
));
7329 * At this point need_resched() is true and we'll take the loop in
7330 * schedule(). The next pick is obviously going to be the stop task
7331 * which kthread_is_per_cpu() and will push this task away.
7333 raw_spin_lock(&rq
->lock
);
7336 static void balance_push_set(int cpu
, bool on
)
7338 struct rq
*rq
= cpu_rq(cpu
);
7341 rq_lock_irqsave(rq
, &rf
);
7342 rq
->balance_push
= on
;
7344 WARN_ON_ONCE(rq
->balance_callback
);
7345 rq
->balance_callback
= &balance_push_callback
;
7346 } else if (rq
->balance_callback
== &balance_push_callback
) {
7347 rq
->balance_callback
= NULL
;
7349 rq_unlock_irqrestore(rq
, &rf
);
7353 * Invoked from a CPUs hotplug control thread after the CPU has been marked
7354 * inactive. All tasks which are not per CPU kernel threads are either
7355 * pushed off this CPU now via balance_push() or placed on a different CPU
7356 * during wakeup. Wait until the CPU is quiescent.
7358 static void balance_hotplug_wait(void)
7360 struct rq
*rq
= this_rq();
7362 rcuwait_wait_event(&rq
->hotplug_wait
,
7363 rq
->nr_running
== 1 && !rq_has_pinned_tasks(rq
),
7364 TASK_UNINTERRUPTIBLE
);
7369 static inline void balance_push(struct rq
*rq
)
7373 static inline void balance_push_set(int cpu
, bool on
)
7377 static inline void balance_hotplug_wait(void)
7381 #endif /* CONFIG_HOTPLUG_CPU */
7383 void set_rq_online(struct rq
*rq
)
7386 const struct sched_class
*class;
7388 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7391 for_each_class(class) {
7392 if (class->rq_online
)
7393 class->rq_online(rq
);
7398 void set_rq_offline(struct rq
*rq
)
7401 const struct sched_class
*class;
7403 for_each_class(class) {
7404 if (class->rq_offline
)
7405 class->rq_offline(rq
);
7408 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7414 * used to mark begin/end of suspend/resume:
7416 static int num_cpus_frozen
;
7419 * Update cpusets according to cpu_active mask. If cpusets are
7420 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7421 * around partition_sched_domains().
7423 * If we come here as part of a suspend/resume, don't touch cpusets because we
7424 * want to restore it back to its original state upon resume anyway.
7426 static void cpuset_cpu_active(void)
7428 if (cpuhp_tasks_frozen
) {
7430 * num_cpus_frozen tracks how many CPUs are involved in suspend
7431 * resume sequence. As long as this is not the last online
7432 * operation in the resume sequence, just build a single sched
7433 * domain, ignoring cpusets.
7435 partition_sched_domains(1, NULL
, NULL
);
7436 if (--num_cpus_frozen
)
7439 * This is the last CPU online operation. So fall through and
7440 * restore the original sched domains by considering the
7441 * cpuset configurations.
7443 cpuset_force_rebuild();
7445 cpuset_update_active_cpus();
7448 static int cpuset_cpu_inactive(unsigned int cpu
)
7450 if (!cpuhp_tasks_frozen
) {
7451 if (dl_cpu_busy(cpu
))
7453 cpuset_update_active_cpus();
7456 partition_sched_domains(1, NULL
, NULL
);
7461 int sched_cpu_activate(unsigned int cpu
)
7463 struct rq
*rq
= cpu_rq(cpu
);
7467 * Make sure that when the hotplug state machine does a roll-back
7468 * we clear balance_push. Ideally that would happen earlier...
7470 balance_push_set(cpu
, false);
7472 #ifdef CONFIG_SCHED_SMT
7474 * When going up, increment the number of cores with SMT present.
7476 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
7477 static_branch_inc_cpuslocked(&sched_smt_present
);
7479 set_cpu_active(cpu
, true);
7481 if (sched_smp_initialized
) {
7482 sched_domains_numa_masks_set(cpu
);
7483 cpuset_cpu_active();
7487 * Put the rq online, if not already. This happens:
7489 * 1) In the early boot process, because we build the real domains
7490 * after all CPUs have been brought up.
7492 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7495 rq_lock_irqsave(rq
, &rf
);
7497 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7500 rq_unlock_irqrestore(rq
, &rf
);
7505 int sched_cpu_deactivate(unsigned int cpu
)
7507 struct rq
*rq
= cpu_rq(cpu
);
7511 set_cpu_active(cpu
, false);
7514 * From this point forward, this CPU will refuse to run any task that
7515 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
7516 * push those tasks away until this gets cleared, see
7517 * sched_cpu_dying().
7519 balance_push_set(cpu
, true);
7522 * We've cleared cpu_active_mask / set balance_push, wait for all
7523 * preempt-disabled and RCU users of this state to go away such that
7524 * all new such users will observe it.
7526 * Specifically, we rely on ttwu to no longer target this CPU, see
7527 * ttwu_queue_cond() and is_cpu_allowed().
7529 * Do sync before park smpboot threads to take care the rcu boost case.
7533 rq_lock_irqsave(rq
, &rf
);
7535 update_rq_clock(rq
);
7536 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7539 rq_unlock_irqrestore(rq
, &rf
);
7541 #ifdef CONFIG_SCHED_SMT
7543 * When going down, decrement the number of cores with SMT present.
7545 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
7546 static_branch_dec_cpuslocked(&sched_smt_present
);
7549 if (!sched_smp_initialized
)
7552 ret
= cpuset_cpu_inactive(cpu
);
7554 balance_push_set(cpu
, false);
7555 set_cpu_active(cpu
, true);
7558 sched_domains_numa_masks_clear(cpu
);
7562 static void sched_rq_cpu_starting(unsigned int cpu
)
7564 struct rq
*rq
= cpu_rq(cpu
);
7566 rq
->calc_load_update
= calc_load_update
;
7567 update_max_interval();
7570 int sched_cpu_starting(unsigned int cpu
)
7572 sched_rq_cpu_starting(cpu
);
7573 sched_tick_start(cpu
);
7577 #ifdef CONFIG_HOTPLUG_CPU
7580 * Invoked immediately before the stopper thread is invoked to bring the
7581 * CPU down completely. At this point all per CPU kthreads except the
7582 * hotplug thread (current) and the stopper thread (inactive) have been
7583 * either parked or have been unbound from the outgoing CPU. Ensure that
7584 * any of those which might be on the way out are gone.
7586 * If after this point a bound task is being woken on this CPU then the
7587 * responsible hotplug callback has failed to do it's job.
7588 * sched_cpu_dying() will catch it with the appropriate fireworks.
7590 int sched_cpu_wait_empty(unsigned int cpu
)
7592 balance_hotplug_wait();
7597 * Since this CPU is going 'away' for a while, fold any nr_active delta we
7598 * might have. Called from the CPU stopper task after ensuring that the
7599 * stopper is the last running task on the CPU, so nr_active count is
7600 * stable. We need to take the teardown thread which is calling this into
7601 * account, so we hand in adjust = 1 to the load calculation.
7603 * Also see the comment "Global load-average calculations".
7605 static void calc_load_migrate(struct rq
*rq
)
7607 long delta
= calc_load_fold_active(rq
, 1);
7610 atomic_long_add(delta
, &calc_load_tasks
);
7613 static void dump_rq_tasks(struct rq
*rq
, const char *loglvl
)
7615 struct task_struct
*g
, *p
;
7616 int cpu
= cpu_of(rq
);
7618 lockdep_assert_held(&rq
->lock
);
7620 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl
, cpu
, rq
->nr_running
);
7621 for_each_process_thread(g
, p
) {
7622 if (task_cpu(p
) != cpu
)
7625 if (!task_on_rq_queued(p
))
7628 printk("%s\tpid: %d, name: %s\n", loglvl
, p
->pid
, p
->comm
);
7632 int sched_cpu_dying(unsigned int cpu
)
7634 struct rq
*rq
= cpu_rq(cpu
);
7637 /* Handle pending wakeups and then migrate everything off */
7638 sched_tick_stop(cpu
);
7640 rq_lock_irqsave(rq
, &rf
);
7641 if (rq
->nr_running
!= 1 || rq_has_pinned_tasks(rq
)) {
7642 WARN(true, "Dying CPU not properly vacated!");
7643 dump_rq_tasks(rq
, KERN_WARNING
);
7645 rq_unlock_irqrestore(rq
, &rf
);
7648 * Now that the CPU is offline, make sure we're welcome
7649 * to new tasks once we come back up.
7651 balance_push_set(cpu
, false);
7653 calc_load_migrate(rq
);
7654 update_max_interval();
7655 nohz_balance_exit_idle(rq
);
7661 void __init
sched_init_smp(void)
7666 * There's no userspace yet to cause hotplug operations; hence all the
7667 * CPU masks are stable and all blatant races in the below code cannot
7670 mutex_lock(&sched_domains_mutex
);
7671 sched_init_domains(cpu_active_mask
);
7672 mutex_unlock(&sched_domains_mutex
);
7674 /* Move init over to a non-isolated CPU */
7675 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
7677 sched_init_granularity();
7679 init_sched_rt_class();
7680 init_sched_dl_class();
7682 sched_smp_initialized
= true;
7685 static int __init
migration_init(void)
7687 sched_cpu_starting(smp_processor_id());
7690 early_initcall(migration_init
);
7693 void __init
sched_init_smp(void)
7695 sched_init_granularity();
7697 #endif /* CONFIG_SMP */
7699 int in_sched_functions(unsigned long addr
)
7701 return in_lock_functions(addr
) ||
7702 (addr
>= (unsigned long)__sched_text_start
7703 && addr
< (unsigned long)__sched_text_end
);
7706 #ifdef CONFIG_CGROUP_SCHED
7708 * Default task group.
7709 * Every task in system belongs to this group at bootup.
7711 struct task_group root_task_group
;
7712 LIST_HEAD(task_groups
);
7714 /* Cacheline aligned slab cache for task_group */
7715 static struct kmem_cache
*task_group_cache __read_mostly
;
7718 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
7719 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
7721 void __init
sched_init(void)
7723 unsigned long ptr
= 0;
7726 /* Make sure the linker didn't screw up */
7727 BUG_ON(&idle_sched_class
+ 1 != &fair_sched_class
||
7728 &fair_sched_class
+ 1 != &rt_sched_class
||
7729 &rt_sched_class
+ 1 != &dl_sched_class
);
7731 BUG_ON(&dl_sched_class
+ 1 != &stop_sched_class
);
7736 #ifdef CONFIG_FAIR_GROUP_SCHED
7737 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
7739 #ifdef CONFIG_RT_GROUP_SCHED
7740 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
7743 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
7745 #ifdef CONFIG_FAIR_GROUP_SCHED
7746 root_task_group
.se
= (struct sched_entity
**)ptr
;
7747 ptr
+= nr_cpu_ids
* sizeof(void **);
7749 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7750 ptr
+= nr_cpu_ids
* sizeof(void **);
7752 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
7753 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
7754 #endif /* CONFIG_FAIR_GROUP_SCHED */
7755 #ifdef CONFIG_RT_GROUP_SCHED
7756 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7757 ptr
+= nr_cpu_ids
* sizeof(void **);
7759 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7760 ptr
+= nr_cpu_ids
* sizeof(void **);
7762 #endif /* CONFIG_RT_GROUP_SCHED */
7764 #ifdef CONFIG_CPUMASK_OFFSTACK
7765 for_each_possible_cpu(i
) {
7766 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7767 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7768 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
7769 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
7771 #endif /* CONFIG_CPUMASK_OFFSTACK */
7773 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
7774 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
7777 init_defrootdomain();
7780 #ifdef CONFIG_RT_GROUP_SCHED
7781 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7782 global_rt_period(), global_rt_runtime());
7783 #endif /* CONFIG_RT_GROUP_SCHED */
7785 #ifdef CONFIG_CGROUP_SCHED
7786 task_group_cache
= KMEM_CACHE(task_group
, 0);
7788 list_add(&root_task_group
.list
, &task_groups
);
7789 INIT_LIST_HEAD(&root_task_group
.children
);
7790 INIT_LIST_HEAD(&root_task_group
.siblings
);
7791 autogroup_init(&init_task
);
7792 #endif /* CONFIG_CGROUP_SCHED */
7794 for_each_possible_cpu(i
) {
7798 raw_spin_lock_init(&rq
->lock
);
7800 rq
->calc_load_active
= 0;
7801 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7802 init_cfs_rq(&rq
->cfs
);
7803 init_rt_rq(&rq
->rt
);
7804 init_dl_rq(&rq
->dl
);
7805 #ifdef CONFIG_FAIR_GROUP_SCHED
7806 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7807 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
7809 * How much CPU bandwidth does root_task_group get?
7811 * In case of task-groups formed thr' the cgroup filesystem, it
7812 * gets 100% of the CPU resources in the system. This overall
7813 * system CPU resource is divided among the tasks of
7814 * root_task_group and its child task-groups in a fair manner,
7815 * based on each entity's (task or task-group's) weight
7816 * (se->load.weight).
7818 * In other words, if root_task_group has 10 tasks of weight
7819 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7820 * then A0's share of the CPU resource is:
7822 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7824 * We achieve this by letting root_task_group's tasks sit
7825 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7827 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7828 #endif /* CONFIG_FAIR_GROUP_SCHED */
7830 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7831 #ifdef CONFIG_RT_GROUP_SCHED
7832 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7837 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
7838 rq
->balance_callback
= NULL
;
7839 rq
->active_balance
= 0;
7840 rq
->next_balance
= jiffies
;
7845 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7846 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
7848 INIT_LIST_HEAD(&rq
->cfs_tasks
);
7850 rq_attach_root(rq
, &def_root_domain
);
7851 #ifdef CONFIG_NO_HZ_COMMON
7852 rq
->last_blocked_load_update_tick
= jiffies
;
7853 atomic_set(&rq
->nohz_flags
, 0);
7855 INIT_CSD(&rq
->nohz_csd
, nohz_csd_func
, rq
);
7857 #ifdef CONFIG_HOTPLUG_CPU
7858 rcuwait_init(&rq
->hotplug_wait
);
7860 #endif /* CONFIG_SMP */
7862 atomic_set(&rq
->nr_iowait
, 0);
7865 set_load_weight(&init_task
, false);
7868 * The boot idle thread does lazy MMU switching as well:
7871 enter_lazy_tlb(&init_mm
, current
);
7874 * Make us the idle thread. Technically, schedule() should not be
7875 * called from this thread, however somewhere below it might be,
7876 * but because we are the idle thread, we just pick up running again
7877 * when this runqueue becomes "idle".
7879 init_idle(current
, smp_processor_id());
7881 calc_load_update
= jiffies
+ LOAD_FREQ
;
7884 idle_thread_set_boot_cpu();
7886 init_sched_fair_class();
7894 scheduler_running
= 1;
7897 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7898 static inline int preempt_count_equals(int preempt_offset
)
7900 int nested
= preempt_count() + rcu_preempt_depth();
7902 return (nested
== preempt_offset
);
7905 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7908 * Blocking primitives will set (and therefore destroy) current->state,
7909 * since we will exit with TASK_RUNNING make sure we enter with it,
7910 * otherwise we will destroy state.
7912 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
7913 "do not call blocking ops when !TASK_RUNNING; "
7914 "state=%lx set at [<%p>] %pS\n",
7916 (void *)current
->task_state_change
,
7917 (void *)current
->task_state_change
);
7919 ___might_sleep(file
, line
, preempt_offset
);
7921 EXPORT_SYMBOL(__might_sleep
);
7923 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
7925 /* Ratelimiting timestamp: */
7926 static unsigned long prev_jiffy
;
7928 unsigned long preempt_disable_ip
;
7930 /* WARN_ON_ONCE() by default, no rate limit required: */
7933 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
7934 !is_idle_task(current
) && !current
->non_block_count
) ||
7935 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
7939 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7941 prev_jiffy
= jiffies
;
7943 /* Save this before calling printk(), since that will clobber it: */
7944 preempt_disable_ip
= get_preempt_disable_ip(current
);
7947 "BUG: sleeping function called from invalid context at %s:%d\n",
7950 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
7951 in_atomic(), irqs_disabled(), current
->non_block_count
,
7952 current
->pid
, current
->comm
);
7954 if (task_stack_end_corrupted(current
))
7955 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
7957 debug_show_held_locks(current
);
7958 if (irqs_disabled())
7959 print_irqtrace_events(current
);
7960 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
7961 && !preempt_count_equals(preempt_offset
)) {
7962 pr_err("Preemption disabled at:");
7963 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
7966 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7968 EXPORT_SYMBOL(___might_sleep
);
7970 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
7972 static unsigned long prev_jiffy
;
7974 if (irqs_disabled())
7977 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
7980 if (preempt_count() > preempt_offset
)
7983 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7985 prev_jiffy
= jiffies
;
7987 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
7988 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7989 in_atomic(), irqs_disabled(),
7990 current
->pid
, current
->comm
);
7992 debug_show_held_locks(current
);
7994 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
7996 EXPORT_SYMBOL_GPL(__cant_sleep
);
7999 void __cant_migrate(const char *file
, int line
)
8001 static unsigned long prev_jiffy
;
8003 if (irqs_disabled())
8006 if (is_migration_disabled(current
))
8009 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
8012 if (preempt_count() > 0)
8015 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8017 prev_jiffy
= jiffies
;
8019 pr_err("BUG: assuming non migratable context at %s:%d\n", file
, line
);
8020 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
8021 in_atomic(), irqs_disabled(), is_migration_disabled(current
),
8022 current
->pid
, current
->comm
);
8024 debug_show_held_locks(current
);
8026 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
8028 EXPORT_SYMBOL_GPL(__cant_migrate
);
8032 #ifdef CONFIG_MAGIC_SYSRQ
8033 void normalize_rt_tasks(void)
8035 struct task_struct
*g
, *p
;
8036 struct sched_attr attr
= {
8037 .sched_policy
= SCHED_NORMAL
,
8040 read_lock(&tasklist_lock
);
8041 for_each_process_thread(g
, p
) {
8043 * Only normalize user tasks:
8045 if (p
->flags
& PF_KTHREAD
)
8048 p
->se
.exec_start
= 0;
8049 schedstat_set(p
->se
.statistics
.wait_start
, 0);
8050 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
8051 schedstat_set(p
->se
.statistics
.block_start
, 0);
8053 if (!dl_task(p
) && !rt_task(p
)) {
8055 * Renice negative nice level userspace
8058 if (task_nice(p
) < 0)
8059 set_user_nice(p
, 0);
8063 __sched_setscheduler(p
, &attr
, false, false);
8065 read_unlock(&tasklist_lock
);
8068 #endif /* CONFIG_MAGIC_SYSRQ */
8070 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8072 * These functions are only useful for the IA64 MCA handling, or kdb.
8074 * They can only be called when the whole system has been
8075 * stopped - every CPU needs to be quiescent, and no scheduling
8076 * activity can take place. Using them for anything else would
8077 * be a serious bug, and as a result, they aren't even visible
8078 * under any other configuration.
8082 * curr_task - return the current task for a given CPU.
8083 * @cpu: the processor in question.
8085 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8087 * Return: The current task for @cpu.
8089 struct task_struct
*curr_task(int cpu
)
8091 return cpu_curr(cpu
);
8094 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8098 * ia64_set_curr_task - set the current task for a given CPU.
8099 * @cpu: the processor in question.
8100 * @p: the task pointer to set.
8102 * Description: This function must only be used when non-maskable interrupts
8103 * are serviced on a separate stack. It allows the architecture to switch the
8104 * notion of the current task on a CPU in a non-blocking manner. This function
8105 * must be called with all CPU's synchronized, and interrupts disabled, the
8106 * and caller must save the original value of the current task (see
8107 * curr_task() above) and restore that value before reenabling interrupts and
8108 * re-starting the system.
8110 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8112 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
8119 #ifdef CONFIG_CGROUP_SCHED
8120 /* task_group_lock serializes the addition/removal of task groups */
8121 static DEFINE_SPINLOCK(task_group_lock
);
8123 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
8124 struct task_group
*parent
)
8126 #ifdef CONFIG_UCLAMP_TASK_GROUP
8127 enum uclamp_id clamp_id
;
8129 for_each_clamp_id(clamp_id
) {
8130 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
8131 uclamp_none(clamp_id
), false);
8132 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
8137 static void sched_free_group(struct task_group
*tg
)
8139 free_fair_sched_group(tg
);
8140 free_rt_sched_group(tg
);
8142 kmem_cache_free(task_group_cache
, tg
);
8145 /* allocate runqueue etc for a new task group */
8146 struct task_group
*sched_create_group(struct task_group
*parent
)
8148 struct task_group
*tg
;
8150 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
8152 return ERR_PTR(-ENOMEM
);
8154 if (!alloc_fair_sched_group(tg
, parent
))
8157 if (!alloc_rt_sched_group(tg
, parent
))
8160 alloc_uclamp_sched_group(tg
, parent
);
8165 sched_free_group(tg
);
8166 return ERR_PTR(-ENOMEM
);
8169 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
8171 unsigned long flags
;
8173 spin_lock_irqsave(&task_group_lock
, flags
);
8174 list_add_rcu(&tg
->list
, &task_groups
);
8176 /* Root should already exist: */
8179 tg
->parent
= parent
;
8180 INIT_LIST_HEAD(&tg
->children
);
8181 list_add_rcu(&tg
->siblings
, &parent
->children
);
8182 spin_unlock_irqrestore(&task_group_lock
, flags
);
8184 online_fair_sched_group(tg
);
8187 /* rcu callback to free various structures associated with a task group */
8188 static void sched_free_group_rcu(struct rcu_head
*rhp
)
8190 /* Now it should be safe to free those cfs_rqs: */
8191 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
8194 void sched_destroy_group(struct task_group
*tg
)
8196 /* Wait for possible concurrent references to cfs_rqs complete: */
8197 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
8200 void sched_offline_group(struct task_group
*tg
)
8202 unsigned long flags
;
8204 /* End participation in shares distribution: */
8205 unregister_fair_sched_group(tg
);
8207 spin_lock_irqsave(&task_group_lock
, flags
);
8208 list_del_rcu(&tg
->list
);
8209 list_del_rcu(&tg
->siblings
);
8210 spin_unlock_irqrestore(&task_group_lock
, flags
);
8213 static void sched_change_group(struct task_struct
*tsk
, int type
)
8215 struct task_group
*tg
;
8218 * All callers are synchronized by task_rq_lock(); we do not use RCU
8219 * which is pointless here. Thus, we pass "true" to task_css_check()
8220 * to prevent lockdep warnings.
8222 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
8223 struct task_group
, css
);
8224 tg
= autogroup_task_group(tsk
, tg
);
8225 tsk
->sched_task_group
= tg
;
8227 #ifdef CONFIG_FAIR_GROUP_SCHED
8228 if (tsk
->sched_class
->task_change_group
)
8229 tsk
->sched_class
->task_change_group(tsk
, type
);
8232 set_task_rq(tsk
, task_cpu(tsk
));
8236 * Change task's runqueue when it moves between groups.
8238 * The caller of this function should have put the task in its new group by
8239 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8242 void sched_move_task(struct task_struct
*tsk
)
8244 int queued
, running
, queue_flags
=
8245 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
8249 rq
= task_rq_lock(tsk
, &rf
);
8250 update_rq_clock(rq
);
8252 running
= task_current(rq
, tsk
);
8253 queued
= task_on_rq_queued(tsk
);
8256 dequeue_task(rq
, tsk
, queue_flags
);
8258 put_prev_task(rq
, tsk
);
8260 sched_change_group(tsk
, TASK_MOVE_GROUP
);
8263 enqueue_task(rq
, tsk
, queue_flags
);
8265 set_next_task(rq
, tsk
);
8267 * After changing group, the running task may have joined a
8268 * throttled one but it's still the running task. Trigger a
8269 * resched to make sure that task can still run.
8274 task_rq_unlock(rq
, tsk
, &rf
);
8277 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8279 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8282 static struct cgroup_subsys_state
*
8283 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8285 struct task_group
*parent
= css_tg(parent_css
);
8286 struct task_group
*tg
;
8289 /* This is early initialization for the top cgroup */
8290 return &root_task_group
.css
;
8293 tg
= sched_create_group(parent
);
8295 return ERR_PTR(-ENOMEM
);
8300 /* Expose task group only after completing cgroup initialization */
8301 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8303 struct task_group
*tg
= css_tg(css
);
8304 struct task_group
*parent
= css_tg(css
->parent
);
8307 sched_online_group(tg
, parent
);
8309 #ifdef CONFIG_UCLAMP_TASK_GROUP
8310 /* Propagate the effective uclamp value for the new group */
8311 cpu_util_update_eff(css
);
8317 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8319 struct task_group
*tg
= css_tg(css
);
8321 sched_offline_group(tg
);
8324 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8326 struct task_group
*tg
= css_tg(css
);
8329 * Relies on the RCU grace period between css_released() and this.
8331 sched_free_group(tg
);
8335 * This is called before wake_up_new_task(), therefore we really only
8336 * have to set its group bits, all the other stuff does not apply.
8338 static void cpu_cgroup_fork(struct task_struct
*task
)
8343 rq
= task_rq_lock(task
, &rf
);
8345 update_rq_clock(rq
);
8346 sched_change_group(task
, TASK_SET_GROUP
);
8348 task_rq_unlock(rq
, task
, &rf
);
8351 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8353 struct task_struct
*task
;
8354 struct cgroup_subsys_state
*css
;
8357 cgroup_taskset_for_each(task
, css
, tset
) {
8358 #ifdef CONFIG_RT_GROUP_SCHED
8359 if (!sched_rt_can_attach(css_tg(css
), task
))
8363 * Serialize against wake_up_new_task() such that if it's
8364 * running, we're sure to observe its full state.
8366 raw_spin_lock_irq(&task
->pi_lock
);
8368 * Avoid calling sched_move_task() before wake_up_new_task()
8369 * has happened. This would lead to problems with PELT, due to
8370 * move wanting to detach+attach while we're not attached yet.
8372 if (task
->state
== TASK_NEW
)
8374 raw_spin_unlock_irq(&task
->pi_lock
);
8382 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8384 struct task_struct
*task
;
8385 struct cgroup_subsys_state
*css
;
8387 cgroup_taskset_for_each(task
, css
, tset
)
8388 sched_move_task(task
);
8391 #ifdef CONFIG_UCLAMP_TASK_GROUP
8392 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
8394 struct cgroup_subsys_state
*top_css
= css
;
8395 struct uclamp_se
*uc_parent
= NULL
;
8396 struct uclamp_se
*uc_se
= NULL
;
8397 unsigned int eff
[UCLAMP_CNT
];
8398 enum uclamp_id clamp_id
;
8399 unsigned int clamps
;
8401 css_for_each_descendant_pre(css
, top_css
) {
8402 uc_parent
= css_tg(css
)->parent
8403 ? css_tg(css
)->parent
->uclamp
: NULL
;
8405 for_each_clamp_id(clamp_id
) {
8406 /* Assume effective clamps matches requested clamps */
8407 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
8408 /* Cap effective clamps with parent's effective clamps */
8410 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
8411 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
8414 /* Ensure protection is always capped by limit */
8415 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
8417 /* Propagate most restrictive effective clamps */
8419 uc_se
= css_tg(css
)->uclamp
;
8420 for_each_clamp_id(clamp_id
) {
8421 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
8423 uc_se
[clamp_id
].value
= eff
[clamp_id
];
8424 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
8425 clamps
|= (0x1 << clamp_id
);
8428 css
= css_rightmost_descendant(css
);
8432 /* Immediately update descendants RUNNABLE tasks */
8433 uclamp_update_active_tasks(css
, clamps
);
8438 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
8439 * C expression. Since there is no way to convert a macro argument (N) into a
8440 * character constant, use two levels of macros.
8442 #define _POW10(exp) ((unsigned int)1e##exp)
8443 #define POW10(exp) _POW10(exp)
8445 struct uclamp_request
{
8446 #define UCLAMP_PERCENT_SHIFT 2
8447 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
8453 static inline struct uclamp_request
8454 capacity_from_percent(char *buf
)
8456 struct uclamp_request req
= {
8457 .percent
= UCLAMP_PERCENT_SCALE
,
8458 .util
= SCHED_CAPACITY_SCALE
,
8463 if (strcmp(buf
, "max")) {
8464 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
8468 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
8473 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
8474 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
8480 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
8481 size_t nbytes
, loff_t off
,
8482 enum uclamp_id clamp_id
)
8484 struct uclamp_request req
;
8485 struct task_group
*tg
;
8487 req
= capacity_from_percent(buf
);
8491 static_branch_enable(&sched_uclamp_used
);
8493 mutex_lock(&uclamp_mutex
);
8496 tg
= css_tg(of_css(of
));
8497 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
8498 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
8501 * Because of not recoverable conversion rounding we keep track of the
8502 * exact requested value
8504 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
8506 /* Update effective clamps to track the most restrictive value */
8507 cpu_util_update_eff(of_css(of
));
8510 mutex_unlock(&uclamp_mutex
);
8515 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
8516 char *buf
, size_t nbytes
,
8519 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
8522 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
8523 char *buf
, size_t nbytes
,
8526 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
8529 static inline void cpu_uclamp_print(struct seq_file
*sf
,
8530 enum uclamp_id clamp_id
)
8532 struct task_group
*tg
;
8538 tg
= css_tg(seq_css(sf
));
8539 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
8542 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
8543 seq_puts(sf
, "max\n");
8547 percent
= tg
->uclamp_pct
[clamp_id
];
8548 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
8549 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
8552 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
8554 cpu_uclamp_print(sf
, UCLAMP_MIN
);
8558 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
8560 cpu_uclamp_print(sf
, UCLAMP_MAX
);
8563 #endif /* CONFIG_UCLAMP_TASK_GROUP */
8565 #ifdef CONFIG_FAIR_GROUP_SCHED
8566 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8567 struct cftype
*cftype
, u64 shareval
)
8569 if (shareval
> scale_load_down(ULONG_MAX
))
8570 shareval
= MAX_SHARES
;
8571 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8574 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8577 struct task_group
*tg
= css_tg(css
);
8579 return (u64
) scale_load_down(tg
->shares
);
8582 #ifdef CONFIG_CFS_BANDWIDTH
8583 static DEFINE_MUTEX(cfs_constraints_mutex
);
8585 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8586 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8587 /* More than 203 days if BW_SHIFT equals 20. */
8588 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
8590 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8592 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8594 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8595 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8597 if (tg
== &root_task_group
)
8601 * Ensure we have at some amount of bandwidth every period. This is
8602 * to prevent reaching a state of large arrears when throttled via
8603 * entity_tick() resulting in prolonged exit starvation.
8605 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8609 * Likewise, bound things on the otherside by preventing insane quota
8610 * periods. This also allows us to normalize in computing quota
8613 if (period
> max_cfs_quota_period
)
8617 * Bound quota to defend quota against overflow during bandwidth shift.
8619 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
8623 * Prevent race between setting of cfs_rq->runtime_enabled and
8624 * unthrottle_offline_cfs_rqs().
8627 mutex_lock(&cfs_constraints_mutex
);
8628 ret
= __cfs_schedulable(tg
, period
, quota
);
8632 runtime_enabled
= quota
!= RUNTIME_INF
;
8633 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
8635 * If we need to toggle cfs_bandwidth_used, off->on must occur
8636 * before making related changes, and on->off must occur afterwards
8638 if (runtime_enabled
&& !runtime_was_enabled
)
8639 cfs_bandwidth_usage_inc();
8640 raw_spin_lock_irq(&cfs_b
->lock
);
8641 cfs_b
->period
= ns_to_ktime(period
);
8642 cfs_b
->quota
= quota
;
8644 __refill_cfs_bandwidth_runtime(cfs_b
);
8646 /* Restart the period timer (if active) to handle new period expiry: */
8647 if (runtime_enabled
)
8648 start_cfs_bandwidth(cfs_b
);
8650 raw_spin_unlock_irq(&cfs_b
->lock
);
8652 for_each_online_cpu(i
) {
8653 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
8654 struct rq
*rq
= cfs_rq
->rq
;
8657 rq_lock_irq(rq
, &rf
);
8658 cfs_rq
->runtime_enabled
= runtime_enabled
;
8659 cfs_rq
->runtime_remaining
= 0;
8661 if (cfs_rq
->throttled
)
8662 unthrottle_cfs_rq(cfs_rq
);
8663 rq_unlock_irq(rq
, &rf
);
8665 if (runtime_was_enabled
&& !runtime_enabled
)
8666 cfs_bandwidth_usage_dec();
8668 mutex_unlock(&cfs_constraints_mutex
);
8674 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
8678 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8679 if (cfs_quota_us
< 0)
8680 quota
= RUNTIME_INF
;
8681 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
8682 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
8686 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8689 static long tg_get_cfs_quota(struct task_group
*tg
)
8693 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
8696 quota_us
= tg
->cfs_bandwidth
.quota
;
8697 do_div(quota_us
, NSEC_PER_USEC
);
8702 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
8706 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
8709 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
8710 quota
= tg
->cfs_bandwidth
.quota
;
8712 return tg_set_cfs_bandwidth(tg
, period
, quota
);
8715 static long tg_get_cfs_period(struct task_group
*tg
)
8719 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
8720 do_div(cfs_period_us
, NSEC_PER_USEC
);
8722 return cfs_period_us
;
8725 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
8728 return tg_get_cfs_quota(css_tg(css
));
8731 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
8732 struct cftype
*cftype
, s64 cfs_quota_us
)
8734 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
8737 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
8740 return tg_get_cfs_period(css_tg(css
));
8743 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
8744 struct cftype
*cftype
, u64 cfs_period_us
)
8746 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
8749 struct cfs_schedulable_data
{
8750 struct task_group
*tg
;
8755 * normalize group quota/period to be quota/max_period
8756 * note: units are usecs
8758 static u64
normalize_cfs_quota(struct task_group
*tg
,
8759 struct cfs_schedulable_data
*d
)
8767 period
= tg_get_cfs_period(tg
);
8768 quota
= tg_get_cfs_quota(tg
);
8771 /* note: these should typically be equivalent */
8772 if (quota
== RUNTIME_INF
|| quota
== -1)
8775 return to_ratio(period
, quota
);
8778 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
8780 struct cfs_schedulable_data
*d
= data
;
8781 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8782 s64 quota
= 0, parent_quota
= -1;
8785 quota
= RUNTIME_INF
;
8787 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
8789 quota
= normalize_cfs_quota(tg
, d
);
8790 parent_quota
= parent_b
->hierarchical_quota
;
8793 * Ensure max(child_quota) <= parent_quota. On cgroup2,
8794 * always take the min. On cgroup1, only inherit when no
8797 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
8798 quota
= min(quota
, parent_quota
);
8800 if (quota
== RUNTIME_INF
)
8801 quota
= parent_quota
;
8802 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
8806 cfs_b
->hierarchical_quota
= quota
;
8811 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
8814 struct cfs_schedulable_data data
= {
8820 if (quota
!= RUNTIME_INF
) {
8821 do_div(data
.period
, NSEC_PER_USEC
);
8822 do_div(data
.quota
, NSEC_PER_USEC
);
8826 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
8832 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
8834 struct task_group
*tg
= css_tg(seq_css(sf
));
8835 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8837 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
8838 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
8839 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
8841 if (schedstat_enabled() && tg
!= &root_task_group
) {
8845 for_each_possible_cpu(i
)
8846 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
8848 seq_printf(sf
, "wait_sum %llu\n", ws
);
8853 #endif /* CONFIG_CFS_BANDWIDTH */
8854 #endif /* CONFIG_FAIR_GROUP_SCHED */
8856 #ifdef CONFIG_RT_GROUP_SCHED
8857 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
8858 struct cftype
*cft
, s64 val
)
8860 return sched_group_set_rt_runtime(css_tg(css
), val
);
8863 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
8866 return sched_group_rt_runtime(css_tg(css
));
8869 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
8870 struct cftype
*cftype
, u64 rt_period_us
)
8872 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
8875 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
8878 return sched_group_rt_period(css_tg(css
));
8880 #endif /* CONFIG_RT_GROUP_SCHED */
8882 static struct cftype cpu_legacy_files
[] = {
8883 #ifdef CONFIG_FAIR_GROUP_SCHED
8886 .read_u64
= cpu_shares_read_u64
,
8887 .write_u64
= cpu_shares_write_u64
,
8890 #ifdef CONFIG_CFS_BANDWIDTH
8892 .name
= "cfs_quota_us",
8893 .read_s64
= cpu_cfs_quota_read_s64
,
8894 .write_s64
= cpu_cfs_quota_write_s64
,
8897 .name
= "cfs_period_us",
8898 .read_u64
= cpu_cfs_period_read_u64
,
8899 .write_u64
= cpu_cfs_period_write_u64
,
8903 .seq_show
= cpu_cfs_stat_show
,
8906 #ifdef CONFIG_RT_GROUP_SCHED
8908 .name
= "rt_runtime_us",
8909 .read_s64
= cpu_rt_runtime_read
,
8910 .write_s64
= cpu_rt_runtime_write
,
8913 .name
= "rt_period_us",
8914 .read_u64
= cpu_rt_period_read_uint
,
8915 .write_u64
= cpu_rt_period_write_uint
,
8918 #ifdef CONFIG_UCLAMP_TASK_GROUP
8920 .name
= "uclamp.min",
8921 .flags
= CFTYPE_NOT_ON_ROOT
,
8922 .seq_show
= cpu_uclamp_min_show
,
8923 .write
= cpu_uclamp_min_write
,
8926 .name
= "uclamp.max",
8927 .flags
= CFTYPE_NOT_ON_ROOT
,
8928 .seq_show
= cpu_uclamp_max_show
,
8929 .write
= cpu_uclamp_max_write
,
8935 static int cpu_extra_stat_show(struct seq_file
*sf
,
8936 struct cgroup_subsys_state
*css
)
8938 #ifdef CONFIG_CFS_BANDWIDTH
8940 struct task_group
*tg
= css_tg(css
);
8941 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8944 throttled_usec
= cfs_b
->throttled_time
;
8945 do_div(throttled_usec
, NSEC_PER_USEC
);
8947 seq_printf(sf
, "nr_periods %d\n"
8949 "throttled_usec %llu\n",
8950 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
8957 #ifdef CONFIG_FAIR_GROUP_SCHED
8958 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
8961 struct task_group
*tg
= css_tg(css
);
8962 u64 weight
= scale_load_down(tg
->shares
);
8964 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
8967 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
8968 struct cftype
*cft
, u64 weight
)
8971 * cgroup weight knobs should use the common MIN, DFL and MAX
8972 * values which are 1, 100 and 10000 respectively. While it loses
8973 * a bit of range on both ends, it maps pretty well onto the shares
8974 * value used by scheduler and the round-trip conversions preserve
8975 * the original value over the entire range.
8977 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
8980 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
8982 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
8985 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
8988 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
8989 int last_delta
= INT_MAX
;
8992 /* find the closest nice value to the current weight */
8993 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
8994 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
8995 if (delta
>= last_delta
)
9000 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
9003 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
9004 struct cftype
*cft
, s64 nice
)
9006 unsigned long weight
;
9009 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
9012 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
9013 idx
= array_index_nospec(idx
, 40);
9014 weight
= sched_prio_to_weight
[idx
];
9016 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
9020 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
9021 long period
, long quota
)
9024 seq_puts(sf
, "max");
9026 seq_printf(sf
, "%ld", quota
);
9028 seq_printf(sf
, " %ld\n", period
);
9031 /* caller should put the current value in *@periodp before calling */
9032 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
9033 u64
*periodp
, u64
*quotap
)
9035 char tok
[21]; /* U64_MAX */
9037 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
9040 *periodp
*= NSEC_PER_USEC
;
9042 if (sscanf(tok
, "%llu", quotap
))
9043 *quotap
*= NSEC_PER_USEC
;
9044 else if (!strcmp(tok
, "max"))
9045 *quotap
= RUNTIME_INF
;
9052 #ifdef CONFIG_CFS_BANDWIDTH
9053 static int cpu_max_show(struct seq_file
*sf
, void *v
)
9055 struct task_group
*tg
= css_tg(seq_css(sf
));
9057 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
9061 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
9062 char *buf
, size_t nbytes
, loff_t off
)
9064 struct task_group
*tg
= css_tg(of_css(of
));
9065 u64 period
= tg_get_cfs_period(tg
);
9069 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
9071 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
9072 return ret
?: nbytes
;
9076 static struct cftype cpu_files
[] = {
9077 #ifdef CONFIG_FAIR_GROUP_SCHED
9080 .flags
= CFTYPE_NOT_ON_ROOT
,
9081 .read_u64
= cpu_weight_read_u64
,
9082 .write_u64
= cpu_weight_write_u64
,
9085 .name
= "weight.nice",
9086 .flags
= CFTYPE_NOT_ON_ROOT
,
9087 .read_s64
= cpu_weight_nice_read_s64
,
9088 .write_s64
= cpu_weight_nice_write_s64
,
9091 #ifdef CONFIG_CFS_BANDWIDTH
9094 .flags
= CFTYPE_NOT_ON_ROOT
,
9095 .seq_show
= cpu_max_show
,
9096 .write
= cpu_max_write
,
9099 #ifdef CONFIG_UCLAMP_TASK_GROUP
9101 .name
= "uclamp.min",
9102 .flags
= CFTYPE_NOT_ON_ROOT
,
9103 .seq_show
= cpu_uclamp_min_show
,
9104 .write
= cpu_uclamp_min_write
,
9107 .name
= "uclamp.max",
9108 .flags
= CFTYPE_NOT_ON_ROOT
,
9109 .seq_show
= cpu_uclamp_max_show
,
9110 .write
= cpu_uclamp_max_write
,
9116 struct cgroup_subsys cpu_cgrp_subsys
= {
9117 .css_alloc
= cpu_cgroup_css_alloc
,
9118 .css_online
= cpu_cgroup_css_online
,
9119 .css_released
= cpu_cgroup_css_released
,
9120 .css_free
= cpu_cgroup_css_free
,
9121 .css_extra_stat_show
= cpu_extra_stat_show
,
9122 .fork
= cpu_cgroup_fork
,
9123 .can_attach
= cpu_cgroup_can_attach
,
9124 .attach
= cpu_cgroup_attach
,
9125 .legacy_cftypes
= cpu_legacy_files
,
9126 .dfl_cftypes
= cpu_files
,
9131 #endif /* CONFIG_CGROUP_SCHED */
9133 void dump_cpu_task(int cpu
)
9135 pr_info("Task dump for CPU %d:\n", cpu
);
9136 sched_show_task(cpu_curr(cpu
));
9140 * Nice levels are multiplicative, with a gentle 10% change for every
9141 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
9142 * nice 1, it will get ~10% less CPU time than another CPU-bound task
9143 * that remained on nice 0.
9145 * The "10% effect" is relative and cumulative: from _any_ nice level,
9146 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
9147 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
9148 * If a task goes up by ~10% and another task goes down by ~10% then
9149 * the relative distance between them is ~25%.)
9151 const int sched_prio_to_weight
[40] = {
9152 /* -20 */ 88761, 71755, 56483, 46273, 36291,
9153 /* -15 */ 29154, 23254, 18705, 14949, 11916,
9154 /* -10 */ 9548, 7620, 6100, 4904, 3906,
9155 /* -5 */ 3121, 2501, 1991, 1586, 1277,
9156 /* 0 */ 1024, 820, 655, 526, 423,
9157 /* 5 */ 335, 272, 215, 172, 137,
9158 /* 10 */ 110, 87, 70, 56, 45,
9159 /* 15 */ 36, 29, 23, 18, 15,
9163 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
9165 * In cases where the weight does not change often, we can use the
9166 * precalculated inverse to speed up arithmetics by turning divisions
9167 * into multiplications:
9169 const u32 sched_prio_to_wmult
[40] = {
9170 /* -20 */ 48388, 59856, 76040, 92818, 118348,
9171 /* -15 */ 147320, 184698, 229616, 287308, 360437,
9172 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
9173 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
9174 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
9175 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
9176 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
9177 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
9180 void call_trace_sched_update_nr_running(struct rq
*rq
, int count
)
9182 trace_sched_update_nr_running_tp(rq
, count
);