1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
11 #include <linux/nospec.h>
13 #include <linux/kcov.h>
15 #include <asm/switch_to.h>
18 #include "../workqueue_internal.h"
19 #include "../smpboot.h"
23 #define CREATE_TRACE_POINTS
24 #include <trace/events/sched.h>
27 * Export tracepoints that act as a bare tracehook (ie: have no trace event
28 * associated with them) to allow external modules to probe them.
30 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
31 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
32 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
37 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
39 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
41 * Debugging: various feature bits
43 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
44 * sysctl_sched_features, defined in sched.h, to allow constants propagation
45 * at compile time and compiler optimization based on features default.
47 #define SCHED_FEAT(name, enabled) \
48 (1UL << __SCHED_FEAT_##name) * enabled |
49 const_debug
unsigned int sysctl_sched_features
=
56 * Number of tasks to iterate in a single balance run.
57 * Limited because this is done with IRQs disabled.
59 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
62 * period over which we measure -rt task CPU usage in us.
65 unsigned int sysctl_sched_rt_period
= 1000000;
67 __read_mostly
int scheduler_running
;
70 * part of the period that we allow rt tasks to run in us.
73 int sysctl_sched_rt_runtime
= 950000;
76 * __task_rq_lock - lock the rq @p resides on.
78 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
83 lockdep_assert_held(&p
->pi_lock
);
87 raw_spin_lock(&rq
->lock
);
88 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
92 raw_spin_unlock(&rq
->lock
);
94 while (unlikely(task_on_rq_migrating(p
)))
100 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
102 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
103 __acquires(p
->pi_lock
)
109 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
111 raw_spin_lock(&rq
->lock
);
113 * move_queued_task() task_rq_lock()
116 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
117 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
118 * [S] ->cpu = new_cpu [L] task_rq()
122 * If we observe the old CPU in task_rq_lock(), the acquire of
123 * the old rq->lock will fully serialize against the stores.
125 * If we observe the new CPU in task_rq_lock(), the address
126 * dependency headed by '[L] rq = task_rq()' and the acquire
127 * will pair with the WMB to ensure we then also see migrating.
129 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
133 raw_spin_unlock(&rq
->lock
);
134 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
136 while (unlikely(task_on_rq_migrating(p
)))
142 * RQ-clock updating methods:
145 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
148 * In theory, the compile should just see 0 here, and optimize out the call
149 * to sched_rt_avg_update. But I don't trust it...
151 s64 __maybe_unused steal
= 0, irq_delta
= 0;
153 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
154 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
157 * Since irq_time is only updated on {soft,}irq_exit, we might run into
158 * this case when a previous update_rq_clock() happened inside a
161 * When this happens, we stop ->clock_task and only update the
162 * prev_irq_time stamp to account for the part that fit, so that a next
163 * update will consume the rest. This ensures ->clock_task is
166 * It does however cause some slight miss-attribution of {soft,}irq
167 * time, a more accurate solution would be to update the irq_time using
168 * the current rq->clock timestamp, except that would require using
171 if (irq_delta
> delta
)
174 rq
->prev_irq_time
+= irq_delta
;
177 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
178 if (static_key_false((¶virt_steal_rq_enabled
))) {
179 steal
= paravirt_steal_clock(cpu_of(rq
));
180 steal
-= rq
->prev_steal_time_rq
;
182 if (unlikely(steal
> delta
))
185 rq
->prev_steal_time_rq
+= steal
;
190 rq
->clock_task
+= delta
;
192 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
193 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
194 update_irq_load_avg(rq
, irq_delta
+ steal
);
196 update_rq_clock_pelt(rq
, delta
);
199 void update_rq_clock(struct rq
*rq
)
203 lockdep_assert_held(&rq
->lock
);
205 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
208 #ifdef CONFIG_SCHED_DEBUG
209 if (sched_feat(WARN_DOUBLE_CLOCK
))
210 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
211 rq
->clock_update_flags
|= RQCF_UPDATED
;
214 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
218 update_rq_clock_task(rq
, delta
);
222 #ifdef CONFIG_SCHED_HRTICK
224 * Use HR-timers to deliver accurate preemption points.
227 static void hrtick_clear(struct rq
*rq
)
229 if (hrtimer_active(&rq
->hrtick_timer
))
230 hrtimer_cancel(&rq
->hrtick_timer
);
234 * High-resolution timer tick.
235 * Runs from hardirq context with interrupts disabled.
237 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
239 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
242 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
246 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
249 return HRTIMER_NORESTART
;
254 static void __hrtick_restart(struct rq
*rq
)
256 struct hrtimer
*timer
= &rq
->hrtick_timer
;
258 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
262 * called from hardirq (IPI) context
264 static void __hrtick_start(void *arg
)
270 __hrtick_restart(rq
);
271 rq
->hrtick_csd_pending
= 0;
276 * Called to set the hrtick timer state.
278 * called with rq->lock held and irqs disabled
280 void hrtick_start(struct rq
*rq
, u64 delay
)
282 struct hrtimer
*timer
= &rq
->hrtick_timer
;
287 * Don't schedule slices shorter than 10000ns, that just
288 * doesn't make sense and can cause timer DoS.
290 delta
= max_t(s64
, delay
, 10000LL);
291 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
293 hrtimer_set_expires(timer
, time
);
295 if (rq
== this_rq()) {
296 __hrtick_restart(rq
);
297 } else if (!rq
->hrtick_csd_pending
) {
298 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
299 rq
->hrtick_csd_pending
= 1;
305 * Called to set the hrtick timer state.
307 * called with rq->lock held and irqs disabled
309 void hrtick_start(struct rq
*rq
, u64 delay
)
312 * Don't schedule slices shorter than 10000ns, that just
313 * doesn't make sense. Rely on vruntime for fairness.
315 delay
= max_t(u64
, delay
, 10000LL);
316 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
317 HRTIMER_MODE_REL_PINNED_HARD
);
319 #endif /* CONFIG_SMP */
321 static void hrtick_rq_init(struct rq
*rq
)
324 rq
->hrtick_csd_pending
= 0;
326 rq
->hrtick_csd
.flags
= 0;
327 rq
->hrtick_csd
.func
= __hrtick_start
;
328 rq
->hrtick_csd
.info
= rq
;
331 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
332 rq
->hrtick_timer
.function
= hrtick
;
334 #else /* CONFIG_SCHED_HRTICK */
335 static inline void hrtick_clear(struct rq
*rq
)
339 static inline void hrtick_rq_init(struct rq
*rq
)
342 #endif /* CONFIG_SCHED_HRTICK */
345 * cmpxchg based fetch_or, macro so it works for different integer types
347 #define fetch_or(ptr, mask) \
349 typeof(ptr) _ptr = (ptr); \
350 typeof(mask) _mask = (mask); \
351 typeof(*_ptr) _old, _val = *_ptr; \
354 _old = cmpxchg(_ptr, _val, _val | _mask); \
362 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
364 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
365 * this avoids any races wrt polling state changes and thereby avoids
368 static bool set_nr_and_not_polling(struct task_struct
*p
)
370 struct thread_info
*ti
= task_thread_info(p
);
371 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
375 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
377 * If this returns true, then the idle task promises to call
378 * sched_ttwu_pending() and reschedule soon.
380 static bool set_nr_if_polling(struct task_struct
*p
)
382 struct thread_info
*ti
= task_thread_info(p
);
383 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
386 if (!(val
& _TIF_POLLING_NRFLAG
))
388 if (val
& _TIF_NEED_RESCHED
)
390 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
399 static bool set_nr_and_not_polling(struct task_struct
*p
)
401 set_tsk_need_resched(p
);
406 static bool set_nr_if_polling(struct task_struct
*p
)
413 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
415 struct wake_q_node
*node
= &task
->wake_q
;
418 * Atomically grab the task, if ->wake_q is !nil already it means
419 * its already queued (either by us or someone else) and will get the
420 * wakeup due to that.
422 * In order to ensure that a pending wakeup will observe our pending
423 * state, even in the failed case, an explicit smp_mb() must be used.
425 smp_mb__before_atomic();
426 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
430 * The head is context local, there can be no concurrency.
433 head
->lastp
= &node
->next
;
438 * wake_q_add() - queue a wakeup for 'later' waking.
439 * @head: the wake_q_head to add @task to
440 * @task: the task to queue for 'later' wakeup
442 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
443 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
446 * This function must be used as-if it were wake_up_process(); IOW the task
447 * must be ready to be woken at this location.
449 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
451 if (__wake_q_add(head
, task
))
452 get_task_struct(task
);
456 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
457 * @head: the wake_q_head to add @task to
458 * @task: the task to queue for 'later' wakeup
460 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
461 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
464 * This function must be used as-if it were wake_up_process(); IOW the task
465 * must be ready to be woken at this location.
467 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
468 * that already hold reference to @task can call the 'safe' version and trust
469 * wake_q to do the right thing depending whether or not the @task is already
472 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
474 if (!__wake_q_add(head
, task
))
475 put_task_struct(task
);
478 void wake_up_q(struct wake_q_head
*head
)
480 struct wake_q_node
*node
= head
->first
;
482 while (node
!= WAKE_Q_TAIL
) {
483 struct task_struct
*task
;
485 task
= container_of(node
, struct task_struct
, wake_q
);
487 /* Task can safely be re-inserted now: */
489 task
->wake_q
.next
= NULL
;
492 * wake_up_process() executes a full barrier, which pairs with
493 * the queueing in wake_q_add() so as not to miss wakeups.
495 wake_up_process(task
);
496 put_task_struct(task
);
501 * resched_curr - mark rq's current task 'to be rescheduled now'.
503 * On UP this means the setting of the need_resched flag, on SMP it
504 * might also involve a cross-CPU call to trigger the scheduler on
507 void resched_curr(struct rq
*rq
)
509 struct task_struct
*curr
= rq
->curr
;
512 lockdep_assert_held(&rq
->lock
);
514 if (test_tsk_need_resched(curr
))
519 if (cpu
== smp_processor_id()) {
520 set_tsk_need_resched(curr
);
521 set_preempt_need_resched();
525 if (set_nr_and_not_polling(curr
))
526 smp_send_reschedule(cpu
);
528 trace_sched_wake_idle_without_ipi(cpu
);
531 void resched_cpu(int cpu
)
533 struct rq
*rq
= cpu_rq(cpu
);
536 raw_spin_lock_irqsave(&rq
->lock
, flags
);
537 if (cpu_online(cpu
) || cpu
== smp_processor_id())
539 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
543 #ifdef CONFIG_NO_HZ_COMMON
545 * In the semi idle case, use the nearest busy CPU for migrating timers
546 * from an idle CPU. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle CPU will add more delays to the timers than intended
550 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int i
, cpu
= smp_processor_id();
555 struct sched_domain
*sd
;
557 if (!idle_cpu(cpu
) && housekeeping_cpu(cpu
, HK_FLAG_TIMER
))
561 for_each_domain(cpu
, sd
) {
562 for_each_cpu(i
, sched_domain_span(sd
)) {
566 if (!idle_cpu(i
) && housekeeping_cpu(i
, HK_FLAG_TIMER
)) {
573 if (!housekeeping_cpu(cpu
, HK_FLAG_TIMER
))
574 cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 static void wake_up_idle_cpu(int cpu
)
592 struct rq
*rq
= cpu_rq(cpu
);
594 if (cpu
== smp_processor_id())
597 if (set_nr_and_not_polling(rq
->idle
))
598 smp_send_reschedule(cpu
);
600 trace_sched_wake_idle_without_ipi(cpu
);
603 static bool wake_up_full_nohz_cpu(int cpu
)
606 * We just need the target to call irq_exit() and re-evaluate
607 * the next tick. The nohz full kick at least implies that.
608 * If needed we can still optimize that later with an
611 if (cpu_is_offline(cpu
))
612 return true; /* Don't try to wake offline CPUs. */
613 if (tick_nohz_full_cpu(cpu
)) {
614 if (cpu
!= smp_processor_id() ||
615 tick_nohz_tick_stopped())
616 tick_nohz_full_kick_cpu(cpu
);
624 * Wake up the specified CPU. If the CPU is going offline, it is the
625 * caller's responsibility to deal with the lost wakeup, for example,
626 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
628 void wake_up_nohz_cpu(int cpu
)
630 if (!wake_up_full_nohz_cpu(cpu
))
631 wake_up_idle_cpu(cpu
);
634 static inline bool got_nohz_idle_kick(void)
636 int cpu
= smp_processor_id();
638 if (!(atomic_read(nohz_flags(cpu
)) & NOHZ_KICK_MASK
))
641 if (idle_cpu(cpu
) && !need_resched())
645 * We can't run Idle Load Balance on this CPU for this time so we
646 * cancel it and clear NOHZ_BALANCE_KICK
648 atomic_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
652 #else /* CONFIG_NO_HZ_COMMON */
654 static inline bool got_nohz_idle_kick(void)
659 #endif /* CONFIG_NO_HZ_COMMON */
661 #ifdef CONFIG_NO_HZ_FULL
662 bool sched_can_stop_tick(struct rq
*rq
)
666 /* Deadline tasks, even if single, need the tick */
667 if (rq
->dl
.dl_nr_running
)
671 * If there are more than one RR tasks, we need the tick to effect the
672 * actual RR behaviour.
674 if (rq
->rt
.rr_nr_running
) {
675 if (rq
->rt
.rr_nr_running
== 1)
682 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
683 * forced preemption between FIFO tasks.
685 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
690 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
691 * if there's more than one we need the tick for involuntary
694 if (rq
->nr_running
> 1)
699 #endif /* CONFIG_NO_HZ_FULL */
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group
*from
,
711 tg_visitor down
, tg_visitor up
, void *data
)
713 struct task_group
*parent
, *child
;
719 ret
= (*down
)(parent
, data
);
722 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
729 ret
= (*up
)(parent
, data
);
730 if (ret
|| parent
== from
)
734 parent
= parent
->parent
;
741 int tg_nop(struct task_group
*tg
, void *data
)
747 static void set_load_weight(struct task_struct
*p
, bool update_load
)
749 int prio
= p
->static_prio
- MAX_RT_PRIO
;
750 struct load_weight
*load
= &p
->se
.load
;
753 * SCHED_IDLE tasks get minimal weight:
755 if (task_has_idle_policy(p
)) {
756 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
757 load
->inv_weight
= WMULT_IDLEPRIO
;
758 p
->se
.runnable_weight
= load
->weight
;
763 * SCHED_OTHER tasks have to update their load when changing their
766 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
767 reweight_task(p
, prio
);
769 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
770 load
->inv_weight
= sched_prio_to_wmult
[prio
];
771 p
->se
.runnable_weight
= load
->weight
;
775 #ifdef CONFIG_UCLAMP_TASK
777 * Serializes updates of utilization clamp values
779 * The (slow-path) user-space triggers utilization clamp value updates which
780 * can require updates on (fast-path) scheduler's data structures used to
781 * support enqueue/dequeue operations.
782 * While the per-CPU rq lock protects fast-path update operations, user-space
783 * requests are serialized using a mutex to reduce the risk of conflicting
784 * updates or API abuses.
786 static DEFINE_MUTEX(uclamp_mutex
);
788 /* Max allowed minimum utilization */
789 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
791 /* Max allowed maximum utilization */
792 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
794 /* All clamps are required to be less or equal than these values */
795 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
797 /* Integer rounded range for each bucket */
798 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
800 #define for_each_clamp_id(clamp_id) \
801 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
803 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
805 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
808 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value
)
810 return UCLAMP_BUCKET_DELTA
* uclamp_bucket_id(clamp_value
);
813 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
815 if (clamp_id
== UCLAMP_MIN
)
817 return SCHED_CAPACITY_SCALE
;
820 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
821 unsigned int value
, bool user_defined
)
823 uc_se
->value
= value
;
824 uc_se
->bucket_id
= uclamp_bucket_id(value
);
825 uc_se
->user_defined
= user_defined
;
828 static inline unsigned int
829 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
830 unsigned int clamp_value
)
833 * Avoid blocked utilization pushing up the frequency when we go
834 * idle (which drops the max-clamp) by retaining the last known
837 if (clamp_id
== UCLAMP_MAX
) {
838 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
842 return uclamp_none(UCLAMP_MIN
);
845 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
846 unsigned int clamp_value
)
848 /* Reset max-clamp retention only on idle exit */
849 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
852 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
856 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
857 unsigned int clamp_value
)
859 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
860 int bucket_id
= UCLAMP_BUCKETS
- 1;
863 * Since both min and max clamps are max aggregated, find the
864 * top most bucket with tasks in.
866 for ( ; bucket_id
>= 0; bucket_id
--) {
867 if (!bucket
[bucket_id
].tasks
)
869 return bucket
[bucket_id
].value
;
872 /* No tasks -- default clamp values */
873 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
876 static inline struct uclamp_se
877 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
879 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
880 #ifdef CONFIG_UCLAMP_TASK_GROUP
881 struct uclamp_se uc_max
;
884 * Tasks in autogroups or root task group will be
885 * restricted by system defaults.
887 if (task_group_is_autogroup(task_group(p
)))
889 if (task_group(p
) == &root_task_group
)
892 uc_max
= task_group(p
)->uclamp
[clamp_id
];
893 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
901 * The effective clamp bucket index of a task depends on, by increasing
903 * - the task specific clamp value, when explicitly requested from userspace
904 * - the task group effective clamp value, for tasks not either in the root
905 * group or in an autogroup
906 * - the system default clamp value, defined by the sysadmin
908 static inline struct uclamp_se
909 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
911 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
912 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
914 /* System default restrictions always apply */
915 if (unlikely(uc_req
.value
> uc_max
.value
))
921 unsigned int uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
923 struct uclamp_se uc_eff
;
925 /* Task currently refcounted: use back-annotated (effective) value */
926 if (p
->uclamp
[clamp_id
].active
)
927 return p
->uclamp
[clamp_id
].value
;
929 uc_eff
= uclamp_eff_get(p
, clamp_id
);
935 * When a task is enqueued on a rq, the clamp bucket currently defined by the
936 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
937 * updates the rq's clamp value if required.
939 * Tasks can have a task-specific value requested from user-space, track
940 * within each bucket the maximum value for tasks refcounted in it.
941 * This "local max aggregation" allows to track the exact "requested" value
942 * for each bucket when all its RUNNABLE tasks require the same clamp.
944 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
945 enum uclamp_id clamp_id
)
947 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
948 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
949 struct uclamp_bucket
*bucket
;
951 lockdep_assert_held(&rq
->lock
);
953 /* Update task effective clamp */
954 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
956 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
958 uc_se
->active
= true;
960 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
963 * Local max aggregation: rq buckets always track the max
964 * "requested" clamp value of its RUNNABLE tasks.
966 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
967 bucket
->value
= uc_se
->value
;
969 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
970 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
974 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
975 * is released. If this is the last task reference counting the rq's max
976 * active clamp value, then the rq's clamp value is updated.
978 * Both refcounted tasks and rq's cached clamp values are expected to be
979 * always valid. If it's detected they are not, as defensive programming,
980 * enforce the expected state and warn.
982 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
983 enum uclamp_id clamp_id
)
985 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
986 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
987 struct uclamp_bucket
*bucket
;
988 unsigned int bkt_clamp
;
989 unsigned int rq_clamp
;
991 lockdep_assert_held(&rq
->lock
);
993 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
994 SCHED_WARN_ON(!bucket
->tasks
);
995 if (likely(bucket
->tasks
))
997 uc_se
->active
= false;
1000 * Keep "local max aggregation" simple and accept to (possibly)
1001 * overboost some RUNNABLE tasks in the same bucket.
1002 * The rq clamp bucket value is reset to its base value whenever
1003 * there are no more RUNNABLE tasks refcounting it.
1005 if (likely(bucket
->tasks
))
1008 rq_clamp
= READ_ONCE(uc_rq
->value
);
1010 * Defensive programming: this should never happen. If it happens,
1011 * e.g. due to future modification, warn and fixup the expected value.
1013 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1014 if (bucket
->value
>= rq_clamp
) {
1015 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1016 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1020 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1022 enum uclamp_id clamp_id
;
1024 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1027 for_each_clamp_id(clamp_id
)
1028 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1030 /* Reset clamp idle holding when there is one RUNNABLE task */
1031 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1032 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1035 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1037 enum uclamp_id clamp_id
;
1039 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1042 for_each_clamp_id(clamp_id
)
1043 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1047 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1053 * Lock the task and the rq where the task is (or was) queued.
1055 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1056 * price to pay to safely serialize util_{min,max} updates with
1057 * enqueues, dequeues and migration operations.
1058 * This is the same locking schema used by __set_cpus_allowed_ptr().
1060 rq
= task_rq_lock(p
, &rf
);
1063 * Setting the clamp bucket is serialized by task_rq_lock().
1064 * If the task is not yet RUNNABLE and its task_struct is not
1065 * affecting a valid clamp bucket, the next time it's enqueued,
1066 * it will already see the updated clamp bucket value.
1068 if (p
->uclamp
[clamp_id
].active
) {
1069 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1070 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1073 task_rq_unlock(rq
, p
, &rf
);
1076 #ifdef CONFIG_UCLAMP_TASK_GROUP
1078 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1079 unsigned int clamps
)
1081 enum uclamp_id clamp_id
;
1082 struct css_task_iter it
;
1083 struct task_struct
*p
;
1085 css_task_iter_start(css
, 0, &it
);
1086 while ((p
= css_task_iter_next(&it
))) {
1087 for_each_clamp_id(clamp_id
) {
1088 if ((0x1 << clamp_id
) & clamps
)
1089 uclamp_update_active(p
, clamp_id
);
1092 css_task_iter_end(&it
);
1095 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1096 static void uclamp_update_root_tg(void)
1098 struct task_group
*tg
= &root_task_group
;
1100 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1101 sysctl_sched_uclamp_util_min
, false);
1102 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1103 sysctl_sched_uclamp_util_max
, false);
1106 cpu_util_update_eff(&root_task_group
.css
);
1110 static void uclamp_update_root_tg(void) { }
1113 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1114 void __user
*buffer
, size_t *lenp
,
1117 bool update_root_tg
= false;
1118 int old_min
, old_max
;
1121 mutex_lock(&uclamp_mutex
);
1122 old_min
= sysctl_sched_uclamp_util_min
;
1123 old_max
= sysctl_sched_uclamp_util_max
;
1125 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1131 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1132 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
) {
1137 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1138 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1139 sysctl_sched_uclamp_util_min
, false);
1140 update_root_tg
= true;
1142 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1143 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1144 sysctl_sched_uclamp_util_max
, false);
1145 update_root_tg
= true;
1149 uclamp_update_root_tg();
1152 * We update all RUNNABLE tasks only when task groups are in use.
1153 * Otherwise, keep it simple and do just a lazy update at each next
1154 * task enqueue time.
1160 sysctl_sched_uclamp_util_min
= old_min
;
1161 sysctl_sched_uclamp_util_max
= old_max
;
1163 mutex_unlock(&uclamp_mutex
);
1168 static int uclamp_validate(struct task_struct
*p
,
1169 const struct sched_attr
*attr
)
1171 unsigned int lower_bound
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1172 unsigned int upper_bound
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1174 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
)
1175 lower_bound
= attr
->sched_util_min
;
1176 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
)
1177 upper_bound
= attr
->sched_util_max
;
1179 if (lower_bound
> upper_bound
)
1181 if (upper_bound
> SCHED_CAPACITY_SCALE
)
1187 static void __setscheduler_uclamp(struct task_struct
*p
,
1188 const struct sched_attr
*attr
)
1190 enum uclamp_id clamp_id
;
1193 * On scheduling class change, reset to default clamps for tasks
1194 * without a task-specific value.
1196 for_each_clamp_id(clamp_id
) {
1197 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1198 unsigned int clamp_value
= uclamp_none(clamp_id
);
1200 /* Keep using defined clamps across class changes */
1201 if (uc_se
->user_defined
)
1204 /* By default, RT tasks always get 100% boost */
1205 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1206 clamp_value
= uclamp_none(UCLAMP_MAX
);
1208 uclamp_se_set(uc_se
, clamp_value
, false);
1211 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1214 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1215 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1216 attr
->sched_util_min
, true);
1219 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1220 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1221 attr
->sched_util_max
, true);
1225 static void uclamp_fork(struct task_struct
*p
)
1227 enum uclamp_id clamp_id
;
1229 for_each_clamp_id(clamp_id
)
1230 p
->uclamp
[clamp_id
].active
= false;
1232 if (likely(!p
->sched_reset_on_fork
))
1235 for_each_clamp_id(clamp_id
) {
1236 unsigned int clamp_value
= uclamp_none(clamp_id
);
1238 /* By default, RT tasks always get 100% boost */
1239 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1240 clamp_value
= uclamp_none(UCLAMP_MAX
);
1242 uclamp_se_set(&p
->uclamp_req
[clamp_id
], clamp_value
, false);
1246 static void __init
init_uclamp(void)
1248 struct uclamp_se uc_max
= {};
1249 enum uclamp_id clamp_id
;
1252 mutex_init(&uclamp_mutex
);
1254 for_each_possible_cpu(cpu
) {
1255 memset(&cpu_rq(cpu
)->uclamp
, 0,
1256 sizeof(struct uclamp_rq
)*UCLAMP_CNT
);
1257 cpu_rq(cpu
)->uclamp_flags
= 0;
1260 for_each_clamp_id(clamp_id
) {
1261 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1262 uclamp_none(clamp_id
), false);
1265 /* System defaults allow max clamp values for both indexes */
1266 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1267 for_each_clamp_id(clamp_id
) {
1268 uclamp_default
[clamp_id
] = uc_max
;
1269 #ifdef CONFIG_UCLAMP_TASK_GROUP
1270 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1271 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1276 #else /* CONFIG_UCLAMP_TASK */
1277 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1278 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1279 static inline int uclamp_validate(struct task_struct
*p
,
1280 const struct sched_attr
*attr
)
1284 static void __setscheduler_uclamp(struct task_struct
*p
,
1285 const struct sched_attr
*attr
) { }
1286 static inline void uclamp_fork(struct task_struct
*p
) { }
1287 static inline void init_uclamp(void) { }
1288 #endif /* CONFIG_UCLAMP_TASK */
1290 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1292 if (!(flags
& ENQUEUE_NOCLOCK
))
1293 update_rq_clock(rq
);
1295 if (!(flags
& ENQUEUE_RESTORE
)) {
1296 sched_info_queued(rq
, p
);
1297 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1300 uclamp_rq_inc(rq
, p
);
1301 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1304 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1306 if (!(flags
& DEQUEUE_NOCLOCK
))
1307 update_rq_clock(rq
);
1309 if (!(flags
& DEQUEUE_SAVE
)) {
1310 sched_info_dequeued(rq
, p
);
1311 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1314 uclamp_rq_dec(rq
, p
);
1315 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1318 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1320 if (task_contributes_to_load(p
))
1321 rq
->nr_uninterruptible
--;
1323 enqueue_task(rq
, p
, flags
);
1325 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1328 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1330 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1332 if (task_contributes_to_load(p
))
1333 rq
->nr_uninterruptible
++;
1335 dequeue_task(rq
, p
, flags
);
1339 * __normal_prio - return the priority that is based on the static prio
1341 static inline int __normal_prio(struct task_struct
*p
)
1343 return p
->static_prio
;
1347 * Calculate the expected normal priority: i.e. priority
1348 * without taking RT-inheritance into account. Might be
1349 * boosted by interactivity modifiers. Changes upon fork,
1350 * setprio syscalls, and whenever the interactivity
1351 * estimator recalculates.
1353 static inline int normal_prio(struct task_struct
*p
)
1357 if (task_has_dl_policy(p
))
1358 prio
= MAX_DL_PRIO
-1;
1359 else if (task_has_rt_policy(p
))
1360 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1362 prio
= __normal_prio(p
);
1367 * Calculate the current priority, i.e. the priority
1368 * taken into account by the scheduler. This value might
1369 * be boosted by RT tasks, or might be boosted by
1370 * interactivity modifiers. Will be RT if the task got
1371 * RT-boosted. If not then it returns p->normal_prio.
1373 static int effective_prio(struct task_struct
*p
)
1375 p
->normal_prio
= normal_prio(p
);
1377 * If we are RT tasks or we were boosted to RT priority,
1378 * keep the priority unchanged. Otherwise, update priority
1379 * to the normal priority:
1381 if (!rt_prio(p
->prio
))
1382 return p
->normal_prio
;
1387 * task_curr - is this task currently executing on a CPU?
1388 * @p: the task in question.
1390 * Return: 1 if the task is currently executing. 0 otherwise.
1392 inline int task_curr(const struct task_struct
*p
)
1394 return cpu_curr(task_cpu(p
)) == p
;
1398 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1399 * use the balance_callback list if you want balancing.
1401 * this means any call to check_class_changed() must be followed by a call to
1402 * balance_callback().
1404 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1405 const struct sched_class
*prev_class
,
1408 if (prev_class
!= p
->sched_class
) {
1409 if (prev_class
->switched_from
)
1410 prev_class
->switched_from(rq
, p
);
1412 p
->sched_class
->switched_to(rq
, p
);
1413 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1414 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1417 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1419 const struct sched_class
*class;
1421 if (p
->sched_class
== rq
->curr
->sched_class
) {
1422 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1424 for_each_class(class) {
1425 if (class == rq
->curr
->sched_class
)
1427 if (class == p
->sched_class
) {
1435 * A queue event has occurred, and we're going to schedule. In
1436 * this case, we can save a useless back to back clock update.
1438 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1439 rq_clock_skip_update(rq
);
1444 static inline bool is_per_cpu_kthread(struct task_struct
*p
)
1446 if (!(p
->flags
& PF_KTHREAD
))
1449 if (p
->nr_cpus_allowed
!= 1)
1456 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1457 * __set_cpus_allowed_ptr() and select_fallback_rq().
1459 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1461 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1464 if (is_per_cpu_kthread(p
))
1465 return cpu_online(cpu
);
1467 return cpu_active(cpu
);
1471 * This is how migration works:
1473 * 1) we invoke migration_cpu_stop() on the target CPU using
1475 * 2) stopper starts to run (implicitly forcing the migrated thread
1477 * 3) it checks whether the migrated task is still in the wrong runqueue.
1478 * 4) if it's in the wrong runqueue then the migration thread removes
1479 * it and puts it into the right queue.
1480 * 5) stopper completes and stop_one_cpu() returns and the migration
1485 * move_queued_task - move a queued task to new rq.
1487 * Returns (locked) new rq. Old rq's lock is released.
1489 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1490 struct task_struct
*p
, int new_cpu
)
1492 lockdep_assert_held(&rq
->lock
);
1494 WRITE_ONCE(p
->on_rq
, TASK_ON_RQ_MIGRATING
);
1495 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
1496 set_task_cpu(p
, new_cpu
);
1499 rq
= cpu_rq(new_cpu
);
1502 BUG_ON(task_cpu(p
) != new_cpu
);
1503 enqueue_task(rq
, p
, 0);
1504 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1505 check_preempt_curr(rq
, p
, 0);
1510 struct migration_arg
{
1511 struct task_struct
*task
;
1516 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1517 * this because either it can't run here any more (set_cpus_allowed()
1518 * away from this CPU, or CPU going down), or because we're
1519 * attempting to rebalance this task on exec (sched_exec).
1521 * So we race with normal scheduler movements, but that's OK, as long
1522 * as the task is no longer on this CPU.
1524 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1525 struct task_struct
*p
, int dest_cpu
)
1527 /* Affinity changed (again). */
1528 if (!is_cpu_allowed(p
, dest_cpu
))
1531 update_rq_clock(rq
);
1532 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1538 * migration_cpu_stop - this will be executed by a highprio stopper thread
1539 * and performs thread migration by bumping thread off CPU then
1540 * 'pushing' onto another runqueue.
1542 static int migration_cpu_stop(void *data
)
1544 struct migration_arg
*arg
= data
;
1545 struct task_struct
*p
= arg
->task
;
1546 struct rq
*rq
= this_rq();
1550 * The original target CPU might have gone down and we might
1551 * be on another CPU but it doesn't matter.
1553 local_irq_disable();
1555 * We need to explicitly wake pending tasks before running
1556 * __migrate_task() such that we will not miss enforcing cpus_ptr
1557 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1559 sched_ttwu_pending();
1561 raw_spin_lock(&p
->pi_lock
);
1564 * If task_rq(p) != rq, it cannot be migrated here, because we're
1565 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1566 * we're holding p->pi_lock.
1568 if (task_rq(p
) == rq
) {
1569 if (task_on_rq_queued(p
))
1570 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1572 p
->wake_cpu
= arg
->dest_cpu
;
1575 raw_spin_unlock(&p
->pi_lock
);
1582 * sched_class::set_cpus_allowed must do the below, but is not required to
1583 * actually call this function.
1585 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1587 cpumask_copy(&p
->cpus_mask
, new_mask
);
1588 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1591 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1593 struct rq
*rq
= task_rq(p
);
1594 bool queued
, running
;
1596 lockdep_assert_held(&p
->pi_lock
);
1598 queued
= task_on_rq_queued(p
);
1599 running
= task_current(rq
, p
);
1603 * Because __kthread_bind() calls this on blocked tasks without
1606 lockdep_assert_held(&rq
->lock
);
1607 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1610 put_prev_task(rq
, p
);
1612 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1615 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1617 set_next_task(rq
, p
);
1621 * Change a given task's CPU affinity. Migrate the thread to a
1622 * proper CPU and schedule it away if the CPU it's executing on
1623 * is removed from the allowed bitmask.
1625 * NOTE: the caller must have a valid reference to the task, the
1626 * task must not exit() & deallocate itself prematurely. The
1627 * call is not atomic; no spinlocks may be held.
1629 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1630 const struct cpumask
*new_mask
, bool check
)
1632 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1633 unsigned int dest_cpu
;
1638 rq
= task_rq_lock(p
, &rf
);
1639 update_rq_clock(rq
);
1641 if (p
->flags
& PF_KTHREAD
) {
1643 * Kernel threads are allowed on online && !active CPUs
1645 cpu_valid_mask
= cpu_online_mask
;
1649 * Must re-check here, to close a race against __kthread_bind(),
1650 * sched_setaffinity() is not guaranteed to observe the flag.
1652 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1657 if (cpumask_equal(p
->cpus_ptr
, new_mask
))
1660 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1661 if (dest_cpu
>= nr_cpu_ids
) {
1666 do_set_cpus_allowed(p
, new_mask
);
1668 if (p
->flags
& PF_KTHREAD
) {
1670 * For kernel threads that do indeed end up on online &&
1671 * !active we want to ensure they are strict per-CPU threads.
1673 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1674 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1675 p
->nr_cpus_allowed
!= 1);
1678 /* Can the task run on the task's current CPU? If so, we're done */
1679 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1682 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1683 struct migration_arg arg
= { p
, dest_cpu
};
1684 /* Need help from migration thread: drop lock and wait. */
1685 task_rq_unlock(rq
, p
, &rf
);
1686 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1688 } else if (task_on_rq_queued(p
)) {
1690 * OK, since we're going to drop the lock immediately
1691 * afterwards anyway.
1693 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1696 task_rq_unlock(rq
, p
, &rf
);
1701 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1703 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1705 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1707 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1709 #ifdef CONFIG_SCHED_DEBUG
1711 * We should never call set_task_cpu() on a blocked task,
1712 * ttwu() will sort out the placement.
1714 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1718 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1719 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1720 * time relying on p->on_rq.
1722 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1723 p
->sched_class
== &fair_sched_class
&&
1724 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1726 #ifdef CONFIG_LOCKDEP
1728 * The caller should hold either p->pi_lock or rq->lock, when changing
1729 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1731 * sched_move_task() holds both and thus holding either pins the cgroup,
1734 * Furthermore, all task_rq users should acquire both locks, see
1737 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1738 lockdep_is_held(&task_rq(p
)->lock
)));
1741 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1743 WARN_ON_ONCE(!cpu_online(new_cpu
));
1746 trace_sched_migrate_task(p
, new_cpu
);
1748 if (task_cpu(p
) != new_cpu
) {
1749 if (p
->sched_class
->migrate_task_rq
)
1750 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1751 p
->se
.nr_migrations
++;
1753 perf_event_task_migrate(p
);
1756 __set_task_cpu(p
, new_cpu
);
1759 #ifdef CONFIG_NUMA_BALANCING
1760 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1762 if (task_on_rq_queued(p
)) {
1763 struct rq
*src_rq
, *dst_rq
;
1764 struct rq_flags srf
, drf
;
1766 src_rq
= task_rq(p
);
1767 dst_rq
= cpu_rq(cpu
);
1769 rq_pin_lock(src_rq
, &srf
);
1770 rq_pin_lock(dst_rq
, &drf
);
1772 deactivate_task(src_rq
, p
, 0);
1773 set_task_cpu(p
, cpu
);
1774 activate_task(dst_rq
, p
, 0);
1775 check_preempt_curr(dst_rq
, p
, 0);
1777 rq_unpin_lock(dst_rq
, &drf
);
1778 rq_unpin_lock(src_rq
, &srf
);
1782 * Task isn't running anymore; make it appear like we migrated
1783 * it before it went to sleep. This means on wakeup we make the
1784 * previous CPU our target instead of where it really is.
1790 struct migration_swap_arg
{
1791 struct task_struct
*src_task
, *dst_task
;
1792 int src_cpu
, dst_cpu
;
1795 static int migrate_swap_stop(void *data
)
1797 struct migration_swap_arg
*arg
= data
;
1798 struct rq
*src_rq
, *dst_rq
;
1801 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1804 src_rq
= cpu_rq(arg
->src_cpu
);
1805 dst_rq
= cpu_rq(arg
->dst_cpu
);
1807 double_raw_lock(&arg
->src_task
->pi_lock
,
1808 &arg
->dst_task
->pi_lock
);
1809 double_rq_lock(src_rq
, dst_rq
);
1811 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1814 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1817 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
1820 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
1823 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1824 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1829 double_rq_unlock(src_rq
, dst_rq
);
1830 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1831 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1837 * Cross migrate two tasks
1839 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
1840 int target_cpu
, int curr_cpu
)
1842 struct migration_swap_arg arg
;
1845 arg
= (struct migration_swap_arg
){
1847 .src_cpu
= curr_cpu
,
1849 .dst_cpu
= target_cpu
,
1852 if (arg
.src_cpu
== arg
.dst_cpu
)
1856 * These three tests are all lockless; this is OK since all of them
1857 * will be re-checked with proper locks held further down the line.
1859 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1862 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
1865 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
1868 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1869 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1874 #endif /* CONFIG_NUMA_BALANCING */
1877 * wait_task_inactive - wait for a thread to unschedule.
1879 * If @match_state is nonzero, it's the @p->state value just checked and
1880 * not expected to change. If it changes, i.e. @p might have woken up,
1881 * then return zero. When we succeed in waiting for @p to be off its CPU,
1882 * we return a positive number (its total switch count). If a second call
1883 * a short while later returns the same number, the caller can be sure that
1884 * @p has remained unscheduled the whole time.
1886 * The caller must ensure that the task *will* unschedule sometime soon,
1887 * else this function might spin for a *long* time. This function can't
1888 * be called with interrupts off, or it may introduce deadlock with
1889 * smp_call_function() if an IPI is sent by the same process we are
1890 * waiting to become inactive.
1892 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1894 int running
, queued
;
1901 * We do the initial early heuristics without holding
1902 * any task-queue locks at all. We'll only try to get
1903 * the runqueue lock when things look like they will
1909 * If the task is actively running on another CPU
1910 * still, just relax and busy-wait without holding
1913 * NOTE! Since we don't hold any locks, it's not
1914 * even sure that "rq" stays as the right runqueue!
1915 * But we don't care, since "task_running()" will
1916 * return false if the runqueue has changed and p
1917 * is actually now running somewhere else!
1919 while (task_running(rq
, p
)) {
1920 if (match_state
&& unlikely(p
->state
!= match_state
))
1926 * Ok, time to look more closely! We need the rq
1927 * lock now, to be *sure*. If we're wrong, we'll
1928 * just go back and repeat.
1930 rq
= task_rq_lock(p
, &rf
);
1931 trace_sched_wait_task(p
);
1932 running
= task_running(rq
, p
);
1933 queued
= task_on_rq_queued(p
);
1935 if (!match_state
|| p
->state
== match_state
)
1936 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1937 task_rq_unlock(rq
, p
, &rf
);
1940 * If it changed from the expected state, bail out now.
1942 if (unlikely(!ncsw
))
1946 * Was it really running after all now that we
1947 * checked with the proper locks actually held?
1949 * Oops. Go back and try again..
1951 if (unlikely(running
)) {
1957 * It's not enough that it's not actively running,
1958 * it must be off the runqueue _entirely_, and not
1961 * So if it was still runnable (but just not actively
1962 * running right now), it's preempted, and we should
1963 * yield - it could be a while.
1965 if (unlikely(queued
)) {
1966 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1968 set_current_state(TASK_UNINTERRUPTIBLE
);
1969 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1974 * Ahh, all good. It wasn't running, and it wasn't
1975 * runnable, which means that it will never become
1976 * running in the future either. We're all done!
1985 * kick_process - kick a running thread to enter/exit the kernel
1986 * @p: the to-be-kicked thread
1988 * Cause a process which is running on another CPU to enter
1989 * kernel-mode, without any delay. (to get signals handled.)
1991 * NOTE: this function doesn't have to take the runqueue lock,
1992 * because all it wants to ensure is that the remote task enters
1993 * the kernel. If the IPI races and the task has been migrated
1994 * to another CPU then no harm is done and the purpose has been
1997 void kick_process(struct task_struct
*p
)
2003 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2004 smp_send_reschedule(cpu
);
2007 EXPORT_SYMBOL_GPL(kick_process
);
2010 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2012 * A few notes on cpu_active vs cpu_online:
2014 * - cpu_active must be a subset of cpu_online
2016 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2017 * see __set_cpus_allowed_ptr(). At this point the newly online
2018 * CPU isn't yet part of the sched domains, and balancing will not
2021 * - on CPU-down we clear cpu_active() to mask the sched domains and
2022 * avoid the load balancer to place new tasks on the to be removed
2023 * CPU. Existing tasks will remain running there and will be taken
2026 * This means that fallback selection must not select !active CPUs.
2027 * And can assume that any active CPU must be online. Conversely
2028 * select_task_rq() below may allow selection of !active CPUs in order
2029 * to satisfy the above rules.
2031 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2033 int nid
= cpu_to_node(cpu
);
2034 const struct cpumask
*nodemask
= NULL
;
2035 enum { cpuset
, possible
, fail
} state
= cpuset
;
2039 * If the node that the CPU is on has been offlined, cpu_to_node()
2040 * will return -1. There is no CPU on the node, and we should
2041 * select the CPU on the other node.
2044 nodemask
= cpumask_of_node(nid
);
2046 /* Look for allowed, online CPU in same node. */
2047 for_each_cpu(dest_cpu
, nodemask
) {
2048 if (!cpu_active(dest_cpu
))
2050 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2056 /* Any allowed, online CPU? */
2057 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2058 if (!is_cpu_allowed(p
, dest_cpu
))
2064 /* No more Mr. Nice Guy. */
2067 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2068 cpuset_cpus_allowed_fallback(p
);
2074 do_set_cpus_allowed(p
, cpu_possible_mask
);
2085 if (state
!= cpuset
) {
2087 * Don't tell them about moving exiting tasks or
2088 * kernel threads (both mm NULL), since they never
2091 if (p
->mm
&& printk_ratelimit()) {
2092 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2093 task_pid_nr(p
), p
->comm
, cpu
);
2101 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2104 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
2106 lockdep_assert_held(&p
->pi_lock
);
2108 if (p
->nr_cpus_allowed
> 1)
2109 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
2111 cpu
= cpumask_any(p
->cpus_ptr
);
2114 * In order not to call set_task_cpu() on a blocking task we need
2115 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2118 * Since this is common to all placement strategies, this lives here.
2120 * [ this allows ->select_task() to simply return task_cpu(p) and
2121 * not worry about this generic constraint ]
2123 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2124 cpu
= select_fallback_rq(task_cpu(p
), p
);
2129 static void update_avg(u64
*avg
, u64 sample
)
2131 s64 diff
= sample
- *avg
;
2135 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2137 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2138 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2142 * Make it appear like a SCHED_FIFO task, its something
2143 * userspace knows about and won't get confused about.
2145 * Also, it will make PI more or less work without too
2146 * much confusion -- but then, stop work should not
2147 * rely on PI working anyway.
2149 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2151 stop
->sched_class
= &stop_sched_class
;
2154 cpu_rq(cpu
)->stop
= stop
;
2158 * Reset it back to a normal scheduling class so that
2159 * it can die in pieces.
2161 old_stop
->sched_class
= &rt_sched_class
;
2167 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2168 const struct cpumask
*new_mask
, bool check
)
2170 return set_cpus_allowed_ptr(p
, new_mask
);
2173 #endif /* CONFIG_SMP */
2176 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2180 if (!schedstat_enabled())
2186 if (cpu
== rq
->cpu
) {
2187 __schedstat_inc(rq
->ttwu_local
);
2188 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2190 struct sched_domain
*sd
;
2192 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2194 for_each_domain(rq
->cpu
, sd
) {
2195 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2196 __schedstat_inc(sd
->ttwu_wake_remote
);
2203 if (wake_flags
& WF_MIGRATED
)
2204 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2205 #endif /* CONFIG_SMP */
2207 __schedstat_inc(rq
->ttwu_count
);
2208 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2210 if (wake_flags
& WF_SYNC
)
2211 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2215 * Mark the task runnable and perform wakeup-preemption.
2217 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2218 struct rq_flags
*rf
)
2220 check_preempt_curr(rq
, p
, wake_flags
);
2221 p
->state
= TASK_RUNNING
;
2222 trace_sched_wakeup(p
);
2225 if (p
->sched_class
->task_woken
) {
2227 * Our task @p is fully woken up and running; so its safe to
2228 * drop the rq->lock, hereafter rq is only used for statistics.
2230 rq_unpin_lock(rq
, rf
);
2231 p
->sched_class
->task_woken(rq
, p
);
2232 rq_repin_lock(rq
, rf
);
2235 if (rq
->idle_stamp
) {
2236 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2237 u64 max
= 2*rq
->max_idle_balance_cost
;
2239 update_avg(&rq
->avg_idle
, delta
);
2241 if (rq
->avg_idle
> max
)
2250 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2251 struct rq_flags
*rf
)
2253 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2255 lockdep_assert_held(&rq
->lock
);
2258 if (p
->sched_contributes_to_load
)
2259 rq
->nr_uninterruptible
--;
2261 if (wake_flags
& WF_MIGRATED
)
2262 en_flags
|= ENQUEUE_MIGRATED
;
2265 activate_task(rq
, p
, en_flags
);
2266 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2270 * Called in case the task @p isn't fully descheduled from its runqueue,
2271 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2272 * since all we need to do is flip p->state to TASK_RUNNING, since
2273 * the task is still ->on_rq.
2275 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2281 rq
= __task_rq_lock(p
, &rf
);
2282 if (task_on_rq_queued(p
)) {
2283 /* check_preempt_curr() may use rq clock */
2284 update_rq_clock(rq
);
2285 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
2288 __task_rq_unlock(rq
, &rf
);
2294 void sched_ttwu_pending(void)
2296 struct rq
*rq
= this_rq();
2297 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
2298 struct task_struct
*p
, *t
;
2304 rq_lock_irqsave(rq
, &rf
);
2305 update_rq_clock(rq
);
2307 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
2308 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
2310 rq_unlock_irqrestore(rq
, &rf
);
2313 void scheduler_ipi(void)
2316 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2317 * TIF_NEED_RESCHED remotely (for the first time) will also send
2320 preempt_fold_need_resched();
2322 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
2326 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2327 * traditionally all their work was done from the interrupt return
2328 * path. Now that we actually do some work, we need to make sure
2331 * Some archs already do call them, luckily irq_enter/exit nest
2334 * Arguably we should visit all archs and update all handlers,
2335 * however a fair share of IPIs are still resched only so this would
2336 * somewhat pessimize the simple resched case.
2339 sched_ttwu_pending();
2342 * Check if someone kicked us for doing the nohz idle load balance.
2344 if (unlikely(got_nohz_idle_kick())) {
2345 this_rq()->idle_balance
= 1;
2346 raise_softirq_irqoff(SCHED_SOFTIRQ
);
2351 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
2353 struct rq
*rq
= cpu_rq(cpu
);
2355 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
2357 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
2358 if (!set_nr_if_polling(rq
->idle
))
2359 smp_send_reschedule(cpu
);
2361 trace_sched_wake_idle_without_ipi(cpu
);
2365 void wake_up_if_idle(int cpu
)
2367 struct rq
*rq
= cpu_rq(cpu
);
2372 if (!is_idle_task(rcu_dereference(rq
->curr
)))
2375 if (set_nr_if_polling(rq
->idle
)) {
2376 trace_sched_wake_idle_without_ipi(cpu
);
2378 rq_lock_irqsave(rq
, &rf
);
2379 if (is_idle_task(rq
->curr
))
2380 smp_send_reschedule(cpu
);
2381 /* Else CPU is not idle, do nothing here: */
2382 rq_unlock_irqrestore(rq
, &rf
);
2389 bool cpus_share_cache(int this_cpu
, int that_cpu
)
2391 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
2393 #endif /* CONFIG_SMP */
2395 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
2397 struct rq
*rq
= cpu_rq(cpu
);
2400 #if defined(CONFIG_SMP)
2401 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
2402 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
2403 ttwu_queue_remote(p
, cpu
, wake_flags
);
2409 update_rq_clock(rq
);
2410 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
2415 * Notes on Program-Order guarantees on SMP systems.
2419 * The basic program-order guarantee on SMP systems is that when a task [t]
2420 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2421 * execution on its new CPU [c1].
2423 * For migration (of runnable tasks) this is provided by the following means:
2425 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2426 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2427 * rq(c1)->lock (if not at the same time, then in that order).
2428 * C) LOCK of the rq(c1)->lock scheduling in task
2430 * Release/acquire chaining guarantees that B happens after A and C after B.
2431 * Note: the CPU doing B need not be c0 or c1
2440 * UNLOCK rq(0)->lock
2442 * LOCK rq(0)->lock // orders against CPU0
2444 * UNLOCK rq(0)->lock
2448 * UNLOCK rq(1)->lock
2450 * LOCK rq(1)->lock // orders against CPU2
2453 * UNLOCK rq(1)->lock
2456 * BLOCKING -- aka. SLEEP + WAKEUP
2458 * For blocking we (obviously) need to provide the same guarantee as for
2459 * migration. However the means are completely different as there is no lock
2460 * chain to provide order. Instead we do:
2462 * 1) smp_store_release(X->on_cpu, 0)
2463 * 2) smp_cond_load_acquire(!X->on_cpu)
2467 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2469 * LOCK rq(0)->lock LOCK X->pi_lock
2472 * smp_store_release(X->on_cpu, 0);
2474 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2480 * X->state = RUNNING
2481 * UNLOCK rq(2)->lock
2483 * LOCK rq(2)->lock // orders against CPU1
2486 * UNLOCK rq(2)->lock
2489 * UNLOCK rq(0)->lock
2492 * However, for wakeups there is a second guarantee we must provide, namely we
2493 * must ensure that CONDITION=1 done by the caller can not be reordered with
2494 * accesses to the task state; see try_to_wake_up() and set_current_state().
2498 * try_to_wake_up - wake up a thread
2499 * @p: the thread to be awakened
2500 * @state: the mask of task states that can be woken
2501 * @wake_flags: wake modifier flags (WF_*)
2503 * If (@state & @p->state) @p->state = TASK_RUNNING.
2505 * If the task was not queued/runnable, also place it back on a runqueue.
2507 * Atomic against schedule() which would dequeue a task, also see
2508 * set_current_state().
2510 * This function executes a full memory barrier before accessing the task
2511 * state; see set_current_state().
2513 * Return: %true if @p->state changes (an actual wakeup was done),
2517 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2519 unsigned long flags
;
2520 int cpu
, success
= 0;
2525 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2526 * == smp_processor_id()'. Together this means we can special
2527 * case the whole 'p->on_rq && ttwu_remote()' case below
2528 * without taking any locks.
2531 * - we rely on Program-Order guarantees for all the ordering,
2532 * - we're serialized against set_special_state() by virtue of
2533 * it disabling IRQs (this allows not taking ->pi_lock).
2535 if (!(p
->state
& state
))
2540 trace_sched_waking(p
);
2541 p
->state
= TASK_RUNNING
;
2542 trace_sched_wakeup(p
);
2547 * If we are going to wake up a thread waiting for CONDITION we
2548 * need to ensure that CONDITION=1 done by the caller can not be
2549 * reordered with p->state check below. This pairs with mb() in
2550 * set_current_state() the waiting thread does.
2552 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2553 smp_mb__after_spinlock();
2554 if (!(p
->state
& state
))
2557 trace_sched_waking(p
);
2559 /* We're going to change ->state: */
2564 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2565 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2566 * in smp_cond_load_acquire() below.
2568 * sched_ttwu_pending() try_to_wake_up()
2569 * STORE p->on_rq = 1 LOAD p->state
2572 * __schedule() (switch to task 'p')
2573 * LOCK rq->lock smp_rmb();
2574 * smp_mb__after_spinlock();
2578 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2580 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2581 * __schedule(). See the comment for smp_mb__after_spinlock().
2584 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2589 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2590 * possible to, falsely, observe p->on_cpu == 0.
2592 * One must be running (->on_cpu == 1) in order to remove oneself
2593 * from the runqueue.
2595 * __schedule() (switch to task 'p') try_to_wake_up()
2596 * STORE p->on_cpu = 1 LOAD p->on_rq
2599 * __schedule() (put 'p' to sleep)
2600 * LOCK rq->lock smp_rmb();
2601 * smp_mb__after_spinlock();
2602 * STORE p->on_rq = 0 LOAD p->on_cpu
2604 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2605 * __schedule(). See the comment for smp_mb__after_spinlock().
2610 * If the owning (remote) CPU is still in the middle of schedule() with
2611 * this task as prev, wait until its done referencing the task.
2613 * Pairs with the smp_store_release() in finish_task().
2615 * This ensures that tasks getting woken will be fully ordered against
2616 * their previous state and preserve Program Order.
2618 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2620 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2621 p
->state
= TASK_WAKING
;
2624 delayacct_blkio_end(p
);
2625 atomic_dec(&task_rq(p
)->nr_iowait
);
2628 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2629 if (task_cpu(p
) != cpu
) {
2630 wake_flags
|= WF_MIGRATED
;
2631 psi_ttwu_dequeue(p
);
2632 set_task_cpu(p
, cpu
);
2635 #else /* CONFIG_SMP */
2638 delayacct_blkio_end(p
);
2639 atomic_dec(&task_rq(p
)->nr_iowait
);
2642 #endif /* CONFIG_SMP */
2644 ttwu_queue(p
, cpu
, wake_flags
);
2646 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2649 ttwu_stat(p
, cpu
, wake_flags
);
2656 * wake_up_process - Wake up a specific process
2657 * @p: The process to be woken up.
2659 * Attempt to wake up the nominated process and move it to the set of runnable
2662 * Return: 1 if the process was woken up, 0 if it was already running.
2664 * This function executes a full memory barrier before accessing the task state.
2666 int wake_up_process(struct task_struct
*p
)
2668 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2670 EXPORT_SYMBOL(wake_up_process
);
2672 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2674 return try_to_wake_up(p
, state
, 0);
2678 * Perform scheduler related setup for a newly forked process p.
2679 * p is forked by current.
2681 * __sched_fork() is basic setup used by init_idle() too:
2683 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2688 p
->se
.exec_start
= 0;
2689 p
->se
.sum_exec_runtime
= 0;
2690 p
->se
.prev_sum_exec_runtime
= 0;
2691 p
->se
.nr_migrations
= 0;
2693 INIT_LIST_HEAD(&p
->se
.group_node
);
2695 #ifdef CONFIG_FAIR_GROUP_SCHED
2696 p
->se
.cfs_rq
= NULL
;
2699 #ifdef CONFIG_SCHEDSTATS
2700 /* Even if schedstat is disabled, there should not be garbage */
2701 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2704 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2705 init_dl_task_timer(&p
->dl
);
2706 init_dl_inactive_task_timer(&p
->dl
);
2707 __dl_clear_params(p
);
2709 INIT_LIST_HEAD(&p
->rt
.run_list
);
2711 p
->rt
.time_slice
= sched_rr_timeslice
;
2715 #ifdef CONFIG_PREEMPT_NOTIFIERS
2716 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2719 #ifdef CONFIG_COMPACTION
2720 p
->capture_control
= NULL
;
2722 init_numa_balancing(clone_flags
, p
);
2725 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2727 #ifdef CONFIG_NUMA_BALANCING
2729 void set_numabalancing_state(bool enabled
)
2732 static_branch_enable(&sched_numa_balancing
);
2734 static_branch_disable(&sched_numa_balancing
);
2737 #ifdef CONFIG_PROC_SYSCTL
2738 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2739 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2743 int state
= static_branch_likely(&sched_numa_balancing
);
2745 if (write
&& !capable(CAP_SYS_ADMIN
))
2750 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2754 set_numabalancing_state(state
);
2760 #ifdef CONFIG_SCHEDSTATS
2762 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2763 static bool __initdata __sched_schedstats
= false;
2765 static void set_schedstats(bool enabled
)
2768 static_branch_enable(&sched_schedstats
);
2770 static_branch_disable(&sched_schedstats
);
2773 void force_schedstat_enabled(void)
2775 if (!schedstat_enabled()) {
2776 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2777 static_branch_enable(&sched_schedstats
);
2781 static int __init
setup_schedstats(char *str
)
2788 * This code is called before jump labels have been set up, so we can't
2789 * change the static branch directly just yet. Instead set a temporary
2790 * variable so init_schedstats() can do it later.
2792 if (!strcmp(str
, "enable")) {
2793 __sched_schedstats
= true;
2795 } else if (!strcmp(str
, "disable")) {
2796 __sched_schedstats
= false;
2801 pr_warn("Unable to parse schedstats=\n");
2805 __setup("schedstats=", setup_schedstats
);
2807 static void __init
init_schedstats(void)
2809 set_schedstats(__sched_schedstats
);
2812 #ifdef CONFIG_PROC_SYSCTL
2813 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2814 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2818 int state
= static_branch_likely(&sched_schedstats
);
2820 if (write
&& !capable(CAP_SYS_ADMIN
))
2825 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2829 set_schedstats(state
);
2832 #endif /* CONFIG_PROC_SYSCTL */
2833 #else /* !CONFIG_SCHEDSTATS */
2834 static inline void init_schedstats(void) {}
2835 #endif /* CONFIG_SCHEDSTATS */
2838 * fork()/clone()-time setup:
2840 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2842 unsigned long flags
;
2844 __sched_fork(clone_flags
, p
);
2846 * We mark the process as NEW here. This guarantees that
2847 * nobody will actually run it, and a signal or other external
2848 * event cannot wake it up and insert it on the runqueue either.
2850 p
->state
= TASK_NEW
;
2853 * Make sure we do not leak PI boosting priority to the child.
2855 p
->prio
= current
->normal_prio
;
2860 * Revert to default priority/policy on fork if requested.
2862 if (unlikely(p
->sched_reset_on_fork
)) {
2863 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2864 p
->policy
= SCHED_NORMAL
;
2865 p
->static_prio
= NICE_TO_PRIO(0);
2867 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2868 p
->static_prio
= NICE_TO_PRIO(0);
2870 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2871 set_load_weight(p
, false);
2874 * We don't need the reset flag anymore after the fork. It has
2875 * fulfilled its duty:
2877 p
->sched_reset_on_fork
= 0;
2880 if (dl_prio(p
->prio
))
2882 else if (rt_prio(p
->prio
))
2883 p
->sched_class
= &rt_sched_class
;
2885 p
->sched_class
= &fair_sched_class
;
2887 init_entity_runnable_average(&p
->se
);
2890 * The child is not yet in the pid-hash so no cgroup attach races,
2891 * and the cgroup is pinned to this child due to cgroup_fork()
2892 * is ran before sched_fork().
2894 * Silence PROVE_RCU.
2896 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2898 * We're setting the CPU for the first time, we don't migrate,
2899 * so use __set_task_cpu().
2901 __set_task_cpu(p
, smp_processor_id());
2902 if (p
->sched_class
->task_fork
)
2903 p
->sched_class
->task_fork(p
);
2904 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2906 #ifdef CONFIG_SCHED_INFO
2907 if (likely(sched_info_on()))
2908 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2910 #if defined(CONFIG_SMP)
2913 init_task_preempt_count(p
);
2915 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2916 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2921 unsigned long to_ratio(u64 period
, u64 runtime
)
2923 if (runtime
== RUNTIME_INF
)
2927 * Doing this here saves a lot of checks in all
2928 * the calling paths, and returning zero seems
2929 * safe for them anyway.
2934 return div64_u64(runtime
<< BW_SHIFT
, period
);
2938 * wake_up_new_task - wake up a newly created task for the first time.
2940 * This function will do some initial scheduler statistics housekeeping
2941 * that must be done for every newly created context, then puts the task
2942 * on the runqueue and wakes it.
2944 void wake_up_new_task(struct task_struct
*p
)
2949 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2950 p
->state
= TASK_RUNNING
;
2953 * Fork balancing, do it here and not earlier because:
2954 * - cpus_ptr can change in the fork path
2955 * - any previously selected CPU might disappear through hotplug
2957 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2958 * as we're not fully set-up yet.
2960 p
->recent_used_cpu
= task_cpu(p
);
2961 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2963 rq
= __task_rq_lock(p
, &rf
);
2964 update_rq_clock(rq
);
2965 post_init_entity_util_avg(p
);
2967 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2968 trace_sched_wakeup_new(p
);
2969 check_preempt_curr(rq
, p
, WF_FORK
);
2971 if (p
->sched_class
->task_woken
) {
2973 * Nothing relies on rq->lock after this, so its fine to
2976 rq_unpin_lock(rq
, &rf
);
2977 p
->sched_class
->task_woken(rq
, p
);
2978 rq_repin_lock(rq
, &rf
);
2981 task_rq_unlock(rq
, p
, &rf
);
2984 #ifdef CONFIG_PREEMPT_NOTIFIERS
2986 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
2988 void preempt_notifier_inc(void)
2990 static_branch_inc(&preempt_notifier_key
);
2992 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2994 void preempt_notifier_dec(void)
2996 static_branch_dec(&preempt_notifier_key
);
2998 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3001 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3002 * @notifier: notifier struct to register
3004 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3006 if (!static_branch_unlikely(&preempt_notifier_key
))
3007 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3009 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3011 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3014 * preempt_notifier_unregister - no longer interested in preemption notifications
3015 * @notifier: notifier struct to unregister
3017 * This is *not* safe to call from within a preemption notifier.
3019 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3021 hlist_del(¬ifier
->link
);
3023 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3025 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3027 struct preempt_notifier
*notifier
;
3029 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3030 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3033 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3035 if (static_branch_unlikely(&preempt_notifier_key
))
3036 __fire_sched_in_preempt_notifiers(curr
);
3040 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3041 struct task_struct
*next
)
3043 struct preempt_notifier
*notifier
;
3045 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3046 notifier
->ops
->sched_out(notifier
, next
);
3049 static __always_inline
void
3050 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3051 struct task_struct
*next
)
3053 if (static_branch_unlikely(&preempt_notifier_key
))
3054 __fire_sched_out_preempt_notifiers(curr
, next
);
3057 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3059 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3064 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3065 struct task_struct
*next
)
3069 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3071 static inline void prepare_task(struct task_struct
*next
)
3075 * Claim the task as running, we do this before switching to it
3076 * such that any running task will have this set.
3082 static inline void finish_task(struct task_struct
*prev
)
3086 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3087 * We must ensure this doesn't happen until the switch is completely
3090 * In particular, the load of prev->state in finish_task_switch() must
3091 * happen before this.
3093 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3095 smp_store_release(&prev
->on_cpu
, 0);
3100 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
3103 * Since the runqueue lock will be released by the next
3104 * task (which is an invalid locking op but in the case
3105 * of the scheduler it's an obvious special-case), so we
3106 * do an early lockdep release here:
3108 rq_unpin_lock(rq
, rf
);
3109 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3110 #ifdef CONFIG_DEBUG_SPINLOCK
3111 /* this is a valid case when another task releases the spinlock */
3112 rq
->lock
.owner
= next
;
3116 static inline void finish_lock_switch(struct rq
*rq
)
3119 * If we are tracking spinlock dependencies then we have to
3120 * fix up the runqueue lock - which gets 'carried over' from
3121 * prev into current:
3123 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
3124 raw_spin_unlock_irq(&rq
->lock
);
3128 * NOP if the arch has not defined these:
3131 #ifndef prepare_arch_switch
3132 # define prepare_arch_switch(next) do { } while (0)
3135 #ifndef finish_arch_post_lock_switch
3136 # define finish_arch_post_lock_switch() do { } while (0)
3140 * prepare_task_switch - prepare to switch tasks
3141 * @rq: the runqueue preparing to switch
3142 * @prev: the current task that is being switched out
3143 * @next: the task we are going to switch to.
3145 * This is called with the rq lock held and interrupts off. It must
3146 * be paired with a subsequent finish_task_switch after the context
3149 * prepare_task_switch sets up locking and calls architecture specific
3153 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3154 struct task_struct
*next
)
3156 kcov_prepare_switch(prev
);
3157 sched_info_switch(rq
, prev
, next
);
3158 perf_event_task_sched_out(prev
, next
);
3160 fire_sched_out_preempt_notifiers(prev
, next
);
3162 prepare_arch_switch(next
);
3166 * finish_task_switch - clean up after a task-switch
3167 * @prev: the thread we just switched away from.
3169 * finish_task_switch must be called after the context switch, paired
3170 * with a prepare_task_switch call before the context switch.
3171 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3172 * and do any other architecture-specific cleanup actions.
3174 * Note that we may have delayed dropping an mm in context_switch(). If
3175 * so, we finish that here outside of the runqueue lock. (Doing it
3176 * with the lock held can cause deadlocks; see schedule() for
3179 * The context switch have flipped the stack from under us and restored the
3180 * local variables which were saved when this task called schedule() in the
3181 * past. prev == current is still correct but we need to recalculate this_rq
3182 * because prev may have moved to another CPU.
3184 static struct rq
*finish_task_switch(struct task_struct
*prev
)
3185 __releases(rq
->lock
)
3187 struct rq
*rq
= this_rq();
3188 struct mm_struct
*mm
= rq
->prev_mm
;
3192 * The previous task will have left us with a preempt_count of 2
3193 * because it left us after:
3196 * preempt_disable(); // 1
3198 * raw_spin_lock_irq(&rq->lock) // 2
3200 * Also, see FORK_PREEMPT_COUNT.
3202 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
3203 "corrupted preempt_count: %s/%d/0x%x\n",
3204 current
->comm
, current
->pid
, preempt_count()))
3205 preempt_count_set(FORK_PREEMPT_COUNT
);
3210 * A task struct has one reference for the use as "current".
3211 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3212 * schedule one last time. The schedule call will never return, and
3213 * the scheduled task must drop that reference.
3215 * We must observe prev->state before clearing prev->on_cpu (in
3216 * finish_task), otherwise a concurrent wakeup can get prev
3217 * running on another CPU and we could rave with its RUNNING -> DEAD
3218 * transition, resulting in a double drop.
3220 prev_state
= prev
->state
;
3221 vtime_task_switch(prev
);
3222 perf_event_task_sched_in(prev
, current
);
3224 finish_lock_switch(rq
);
3225 finish_arch_post_lock_switch();
3226 kcov_finish_switch(current
);
3228 fire_sched_in_preempt_notifiers(current
);
3230 * When switching through a kernel thread, the loop in
3231 * membarrier_{private,global}_expedited() may have observed that
3232 * kernel thread and not issued an IPI. It is therefore possible to
3233 * schedule between user->kernel->user threads without passing though
3234 * switch_mm(). Membarrier requires a barrier after storing to
3235 * rq->curr, before returning to userspace, so provide them here:
3237 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3238 * provided by mmdrop(),
3239 * - a sync_core for SYNC_CORE.
3242 membarrier_mm_sync_core_before_usermode(mm
);
3245 if (unlikely(prev_state
== TASK_DEAD
)) {
3246 if (prev
->sched_class
->task_dead
)
3247 prev
->sched_class
->task_dead(prev
);
3250 * Remove function-return probe instances associated with this
3251 * task and put them back on the free list.
3253 kprobe_flush_task(prev
);
3255 /* Task is done with its stack. */
3256 put_task_stack(prev
);
3258 put_task_struct_rcu_user(prev
);
3261 tick_nohz_task_switch();
3267 /* rq->lock is NOT held, but preemption is disabled */
3268 static void __balance_callback(struct rq
*rq
)
3270 struct callback_head
*head
, *next
;
3271 void (*func
)(struct rq
*rq
);
3272 unsigned long flags
;
3274 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3275 head
= rq
->balance_callback
;
3276 rq
->balance_callback
= NULL
;
3278 func
= (void (*)(struct rq
*))head
->func
;
3285 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3288 static inline void balance_callback(struct rq
*rq
)
3290 if (unlikely(rq
->balance_callback
))
3291 __balance_callback(rq
);
3296 static inline void balance_callback(struct rq
*rq
)
3303 * schedule_tail - first thing a freshly forked thread must call.
3304 * @prev: the thread we just switched away from.
3306 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
3307 __releases(rq
->lock
)
3312 * New tasks start with FORK_PREEMPT_COUNT, see there and
3313 * finish_task_switch() for details.
3315 * finish_task_switch() will drop rq->lock() and lower preempt_count
3316 * and the preempt_enable() will end up enabling preemption (on
3317 * PREEMPT_COUNT kernels).
3320 rq
= finish_task_switch(prev
);
3321 balance_callback(rq
);
3324 if (current
->set_child_tid
)
3325 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3327 calculate_sigpending();
3331 * context_switch - switch to the new MM and the new thread's register state.
3333 static __always_inline
struct rq
*
3334 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3335 struct task_struct
*next
, struct rq_flags
*rf
)
3337 prepare_task_switch(rq
, prev
, next
);
3340 * For paravirt, this is coupled with an exit in switch_to to
3341 * combine the page table reload and the switch backend into
3344 arch_start_context_switch(prev
);
3347 * kernel -> kernel lazy + transfer active
3348 * user -> kernel lazy + mmgrab() active
3350 * kernel -> user switch + mmdrop() active
3351 * user -> user switch
3353 if (!next
->mm
) { // to kernel
3354 enter_lazy_tlb(prev
->active_mm
, next
);
3356 next
->active_mm
= prev
->active_mm
;
3357 if (prev
->mm
) // from user
3358 mmgrab(prev
->active_mm
);
3360 prev
->active_mm
= NULL
;
3362 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
3364 * sys_membarrier() requires an smp_mb() between setting
3365 * rq->curr / membarrier_switch_mm() and returning to userspace.
3367 * The below provides this either through switch_mm(), or in
3368 * case 'prev->active_mm == next->mm' through
3369 * finish_task_switch()'s mmdrop().
3371 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
3373 if (!prev
->mm
) { // from kernel
3374 /* will mmdrop() in finish_task_switch(). */
3375 rq
->prev_mm
= prev
->active_mm
;
3376 prev
->active_mm
= NULL
;
3380 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3382 prepare_lock_switch(rq
, next
, rf
);
3384 /* Here we just switch the register state and the stack. */
3385 switch_to(prev
, next
, prev
);
3388 return finish_task_switch(prev
);
3392 * nr_running and nr_context_switches:
3394 * externally visible scheduler statistics: current number of runnable
3395 * threads, total number of context switches performed since bootup.
3397 unsigned long nr_running(void)
3399 unsigned long i
, sum
= 0;
3401 for_each_online_cpu(i
)
3402 sum
+= cpu_rq(i
)->nr_running
;
3408 * Check if only the current task is running on the CPU.
3410 * Caution: this function does not check that the caller has disabled
3411 * preemption, thus the result might have a time-of-check-to-time-of-use
3412 * race. The caller is responsible to use it correctly, for example:
3414 * - from a non-preemptible section (of course)
3416 * - from a thread that is bound to a single CPU
3418 * - in a loop with very short iterations (e.g. a polling loop)
3420 bool single_task_running(void)
3422 return raw_rq()->nr_running
== 1;
3424 EXPORT_SYMBOL(single_task_running
);
3426 unsigned long long nr_context_switches(void)
3429 unsigned long long sum
= 0;
3431 for_each_possible_cpu(i
)
3432 sum
+= cpu_rq(i
)->nr_switches
;
3438 * Consumers of these two interfaces, like for example the cpuidle menu
3439 * governor, are using nonsensical data. Preferring shallow idle state selection
3440 * for a CPU that has IO-wait which might not even end up running the task when
3441 * it does become runnable.
3444 unsigned long nr_iowait_cpu(int cpu
)
3446 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
3450 * IO-wait accounting, and how its mostly bollocks (on SMP).
3452 * The idea behind IO-wait account is to account the idle time that we could
3453 * have spend running if it were not for IO. That is, if we were to improve the
3454 * storage performance, we'd have a proportional reduction in IO-wait time.
3456 * This all works nicely on UP, where, when a task blocks on IO, we account
3457 * idle time as IO-wait, because if the storage were faster, it could've been
3458 * running and we'd not be idle.
3460 * This has been extended to SMP, by doing the same for each CPU. This however
3463 * Imagine for instance the case where two tasks block on one CPU, only the one
3464 * CPU will have IO-wait accounted, while the other has regular idle. Even
3465 * though, if the storage were faster, both could've ran at the same time,
3466 * utilising both CPUs.
3468 * This means, that when looking globally, the current IO-wait accounting on
3469 * SMP is a lower bound, by reason of under accounting.
3471 * Worse, since the numbers are provided per CPU, they are sometimes
3472 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3473 * associated with any one particular CPU, it can wake to another CPU than it
3474 * blocked on. This means the per CPU IO-wait number is meaningless.
3476 * Task CPU affinities can make all that even more 'interesting'.
3479 unsigned long nr_iowait(void)
3481 unsigned long i
, sum
= 0;
3483 for_each_possible_cpu(i
)
3484 sum
+= nr_iowait_cpu(i
);
3492 * sched_exec - execve() is a valuable balancing opportunity, because at
3493 * this point the task has the smallest effective memory and cache footprint.
3495 void sched_exec(void)
3497 struct task_struct
*p
= current
;
3498 unsigned long flags
;
3501 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3502 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
3503 if (dest_cpu
== smp_processor_id())
3506 if (likely(cpu_active(dest_cpu
))) {
3507 struct migration_arg arg
= { p
, dest_cpu
};
3509 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3510 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3514 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3519 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3520 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3522 EXPORT_PER_CPU_SYMBOL(kstat
);
3523 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3526 * The function fair_sched_class.update_curr accesses the struct curr
3527 * and its field curr->exec_start; when called from task_sched_runtime(),
3528 * we observe a high rate of cache misses in practice.
3529 * Prefetching this data results in improved performance.
3531 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3533 #ifdef CONFIG_FAIR_GROUP_SCHED
3534 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3536 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3539 prefetch(&curr
->exec_start
);
3543 * Return accounted runtime for the task.
3544 * In case the task is currently running, return the runtime plus current's
3545 * pending runtime that have not been accounted yet.
3547 unsigned long long task_sched_runtime(struct task_struct
*p
)
3553 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3555 * 64-bit doesn't need locks to atomically read a 64-bit value.
3556 * So we have a optimization chance when the task's delta_exec is 0.
3557 * Reading ->on_cpu is racy, but this is ok.
3559 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3560 * If we race with it entering CPU, unaccounted time is 0. This is
3561 * indistinguishable from the read occurring a few cycles earlier.
3562 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3563 * been accounted, so we're correct here as well.
3565 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3566 return p
->se
.sum_exec_runtime
;
3569 rq
= task_rq_lock(p
, &rf
);
3571 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3572 * project cycles that may never be accounted to this
3573 * thread, breaking clock_gettime().
3575 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3576 prefetch_curr_exec_start(p
);
3577 update_rq_clock(rq
);
3578 p
->sched_class
->update_curr(rq
);
3580 ns
= p
->se
.sum_exec_runtime
;
3581 task_rq_unlock(rq
, p
, &rf
);
3587 * This function gets called by the timer code, with HZ frequency.
3588 * We call it with interrupts disabled.
3590 void scheduler_tick(void)
3592 int cpu
= smp_processor_id();
3593 struct rq
*rq
= cpu_rq(cpu
);
3594 struct task_struct
*curr
= rq
->curr
;
3601 update_rq_clock(rq
);
3602 curr
->sched_class
->task_tick(rq
, curr
, 0);
3603 calc_global_load_tick(rq
);
3608 perf_event_task_tick();
3611 rq
->idle_balance
= idle_cpu(cpu
);
3612 trigger_load_balance(rq
);
3616 #ifdef CONFIG_NO_HZ_FULL
3621 struct delayed_work work
;
3623 /* Values for ->state, see diagram below. */
3624 #define TICK_SCHED_REMOTE_OFFLINE 0
3625 #define TICK_SCHED_REMOTE_OFFLINING 1
3626 #define TICK_SCHED_REMOTE_RUNNING 2
3629 * State diagram for ->state:
3632 * TICK_SCHED_REMOTE_OFFLINE
3635 * | | sched_tick_remote()
3638 * +--TICK_SCHED_REMOTE_OFFLINING
3641 * sched_tick_start() | | sched_tick_stop()
3644 * TICK_SCHED_REMOTE_RUNNING
3647 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3648 * and sched_tick_start() are happy to leave the state in RUNNING.
3651 static struct tick_work __percpu
*tick_work_cpu
;
3653 static void sched_tick_remote(struct work_struct
*work
)
3655 struct delayed_work
*dwork
= to_delayed_work(work
);
3656 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3657 int cpu
= twork
->cpu
;
3658 struct rq
*rq
= cpu_rq(cpu
);
3659 struct task_struct
*curr
;
3665 * Handle the tick only if it appears the remote CPU is running in full
3666 * dynticks mode. The check is racy by nature, but missing a tick or
3667 * having one too much is no big deal because the scheduler tick updates
3668 * statistics and checks timeslices in a time-independent way, regardless
3669 * of when exactly it is running.
3671 if (!tick_nohz_tick_stopped_cpu(cpu
))
3674 rq_lock_irq(rq
, &rf
);
3676 if (cpu_is_offline(cpu
))
3680 update_rq_clock(rq
);
3682 if (!is_idle_task(curr
)) {
3684 * Make sure the next tick runs within a reasonable
3687 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3688 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3690 curr
->sched_class
->task_tick(rq
, curr
, 0);
3692 calc_load_nohz_remote(rq
);
3694 rq_unlock_irq(rq
, &rf
);
3698 * Run the remote tick once per second (1Hz). This arbitrary
3699 * frequency is large enough to avoid overload but short enough
3700 * to keep scheduler internal stats reasonably up to date. But
3701 * first update state to reflect hotplug activity if required.
3703 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
3704 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
3705 if (os
== TICK_SCHED_REMOTE_RUNNING
)
3706 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3709 static void sched_tick_start(int cpu
)
3712 struct tick_work
*twork
;
3714 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3717 WARN_ON_ONCE(!tick_work_cpu
);
3719 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3720 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
3721 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
3722 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
3724 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3725 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3729 #ifdef CONFIG_HOTPLUG_CPU
3730 static void sched_tick_stop(int cpu
)
3732 struct tick_work
*twork
;
3735 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3738 WARN_ON_ONCE(!tick_work_cpu
);
3740 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3741 /* There cannot be competing actions, but don't rely on stop-machine. */
3742 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
3743 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
3744 /* Don't cancel, as this would mess up the state machine. */
3746 #endif /* CONFIG_HOTPLUG_CPU */
3748 int __init
sched_tick_offload_init(void)
3750 tick_work_cpu
= alloc_percpu(struct tick_work
);
3751 BUG_ON(!tick_work_cpu
);
3755 #else /* !CONFIG_NO_HZ_FULL */
3756 static inline void sched_tick_start(int cpu
) { }
3757 static inline void sched_tick_stop(int cpu
) { }
3760 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3761 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3763 * If the value passed in is equal to the current preempt count
3764 * then we just disabled preemption. Start timing the latency.
3766 static inline void preempt_latency_start(int val
)
3768 if (preempt_count() == val
) {
3769 unsigned long ip
= get_lock_parent_ip();
3770 #ifdef CONFIG_DEBUG_PREEMPT
3771 current
->preempt_disable_ip
= ip
;
3773 trace_preempt_off(CALLER_ADDR0
, ip
);
3777 void preempt_count_add(int val
)
3779 #ifdef CONFIG_DEBUG_PREEMPT
3783 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3786 __preempt_count_add(val
);
3787 #ifdef CONFIG_DEBUG_PREEMPT
3789 * Spinlock count overflowing soon?
3791 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3794 preempt_latency_start(val
);
3796 EXPORT_SYMBOL(preempt_count_add
);
3797 NOKPROBE_SYMBOL(preempt_count_add
);
3800 * If the value passed in equals to the current preempt count
3801 * then we just enabled preemption. Stop timing the latency.
3803 static inline void preempt_latency_stop(int val
)
3805 if (preempt_count() == val
)
3806 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3809 void preempt_count_sub(int val
)
3811 #ifdef CONFIG_DEBUG_PREEMPT
3815 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3818 * Is the spinlock portion underflowing?
3820 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3821 !(preempt_count() & PREEMPT_MASK
)))
3825 preempt_latency_stop(val
);
3826 __preempt_count_sub(val
);
3828 EXPORT_SYMBOL(preempt_count_sub
);
3829 NOKPROBE_SYMBOL(preempt_count_sub
);
3832 static inline void preempt_latency_start(int val
) { }
3833 static inline void preempt_latency_stop(int val
) { }
3836 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3838 #ifdef CONFIG_DEBUG_PREEMPT
3839 return p
->preempt_disable_ip
;
3846 * Print scheduling while atomic bug:
3848 static noinline
void __schedule_bug(struct task_struct
*prev
)
3850 /* Save this before calling printk(), since that will clobber it */
3851 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3853 if (oops_in_progress
)
3856 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3857 prev
->comm
, prev
->pid
, preempt_count());
3859 debug_show_held_locks(prev
);
3861 if (irqs_disabled())
3862 print_irqtrace_events(prev
);
3863 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3864 && in_atomic_preempt_off()) {
3865 pr_err("Preemption disabled at:");
3866 print_ip_sym(preempt_disable_ip
);
3870 panic("scheduling while atomic\n");
3873 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3877 * Various schedule()-time debugging checks and statistics:
3879 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
3881 #ifdef CONFIG_SCHED_STACK_END_CHECK
3882 if (task_stack_end_corrupted(prev
))
3883 panic("corrupted stack end detected inside scheduler\n");
3886 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3887 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
3888 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3889 prev
->comm
, prev
->pid
, prev
->non_block_count
);
3891 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3895 if (unlikely(in_atomic_preempt_off())) {
3896 __schedule_bug(prev
);
3897 preempt_count_set(PREEMPT_DISABLED
);
3901 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3903 schedstat_inc(this_rq()->sched_count
);
3907 * Pick up the highest-prio task:
3909 static inline struct task_struct
*
3910 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3912 const struct sched_class
*class;
3913 struct task_struct
*p
;
3916 * Optimization: we know that if all tasks are in the fair class we can
3917 * call that function directly, but only if the @prev task wasn't of a
3918 * higher scheduling class, because otherwise those loose the
3919 * opportunity to pull in more work from other CPUs.
3921 if (likely((prev
->sched_class
== &idle_sched_class
||
3922 prev
->sched_class
== &fair_sched_class
) &&
3923 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3925 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3926 if (unlikely(p
== RETRY_TASK
))
3929 /* Assumes fair_sched_class->next == idle_sched_class */
3931 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3939 * We must do the balancing pass before put_next_task(), such
3940 * that when we release the rq->lock the task is in the same
3941 * state as before we took rq->lock.
3943 * We can terminate the balance pass as soon as we know there is
3944 * a runnable task of @class priority or higher.
3946 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
3947 if (class->balance(rq
, prev
, rf
))
3952 put_prev_task(rq
, prev
);
3954 for_each_class(class) {
3955 p
= class->pick_next_task(rq
, NULL
, NULL
);
3960 /* The idle class should always have a runnable task: */
3965 * __schedule() is the main scheduler function.
3967 * The main means of driving the scheduler and thus entering this function are:
3969 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3971 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3972 * paths. For example, see arch/x86/entry_64.S.
3974 * To drive preemption between tasks, the scheduler sets the flag in timer
3975 * interrupt handler scheduler_tick().
3977 * 3. Wakeups don't really cause entry into schedule(). They add a
3978 * task to the run-queue and that's it.
3980 * Now, if the new task added to the run-queue preempts the current
3981 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3982 * called on the nearest possible occasion:
3984 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
3986 * - in syscall or exception context, at the next outmost
3987 * preempt_enable(). (this might be as soon as the wake_up()'s
3990 * - in IRQ context, return from interrupt-handler to
3991 * preemptible context
3993 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
3996 * - cond_resched() call
3997 * - explicit schedule() call
3998 * - return from syscall or exception to user-space
3999 * - return from interrupt-handler to user-space
4001 * WARNING: must be called with preemption disabled!
4003 static void __sched notrace
__schedule(bool preempt
)
4005 struct task_struct
*prev
, *next
;
4006 unsigned long *switch_count
;
4011 cpu
= smp_processor_id();
4015 schedule_debug(prev
, preempt
);
4017 if (sched_feat(HRTICK
))
4020 local_irq_disable();
4021 rcu_note_context_switch(preempt
);
4024 * Make sure that signal_pending_state()->signal_pending() below
4025 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4026 * done by the caller to avoid the race with signal_wake_up().
4028 * The membarrier system call requires a full memory barrier
4029 * after coming from user-space, before storing to rq->curr.
4032 smp_mb__after_spinlock();
4034 /* Promote REQ to ACT */
4035 rq
->clock_update_flags
<<= 1;
4036 update_rq_clock(rq
);
4038 switch_count
= &prev
->nivcsw
;
4039 if (!preempt
&& prev
->state
) {
4040 if (signal_pending_state(prev
->state
, prev
)) {
4041 prev
->state
= TASK_RUNNING
;
4043 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
4045 if (prev
->in_iowait
) {
4046 atomic_inc(&rq
->nr_iowait
);
4047 delayacct_blkio_start();
4050 switch_count
= &prev
->nvcsw
;
4053 next
= pick_next_task(rq
, prev
, &rf
);
4054 clear_tsk_need_resched(prev
);
4055 clear_preempt_need_resched();
4057 if (likely(prev
!= next
)) {
4060 * RCU users of rcu_dereference(rq->curr) may not see
4061 * changes to task_struct made by pick_next_task().
4063 RCU_INIT_POINTER(rq
->curr
, next
);
4065 * The membarrier system call requires each architecture
4066 * to have a full memory barrier after updating
4067 * rq->curr, before returning to user-space.
4069 * Here are the schemes providing that barrier on the
4070 * various architectures:
4071 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4072 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4073 * - finish_lock_switch() for weakly-ordered
4074 * architectures where spin_unlock is a full barrier,
4075 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4076 * is a RELEASE barrier),
4080 trace_sched_switch(preempt
, prev
, next
);
4082 /* Also unlocks the rq: */
4083 rq
= context_switch(rq
, prev
, next
, &rf
);
4085 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4086 rq_unlock_irq(rq
, &rf
);
4089 balance_callback(rq
);
4092 void __noreturn
do_task_dead(void)
4094 /* Causes final put_task_struct in finish_task_switch(): */
4095 set_special_state(TASK_DEAD
);
4097 /* Tell freezer to ignore us: */
4098 current
->flags
|= PF_NOFREEZE
;
4103 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4108 static inline void sched_submit_work(struct task_struct
*tsk
)
4114 * If a worker went to sleep, notify and ask workqueue whether
4115 * it wants to wake up a task to maintain concurrency.
4116 * As this function is called inside the schedule() context,
4117 * we disable preemption to avoid it calling schedule() again
4118 * in the possible wakeup of a kworker.
4120 if (tsk
->flags
& PF_WQ_WORKER
) {
4122 wq_worker_sleeping(tsk
);
4123 preempt_enable_no_resched();
4126 if (tsk_is_pi_blocked(tsk
))
4130 * If we are going to sleep and we have plugged IO queued,
4131 * make sure to submit it to avoid deadlocks.
4133 if (blk_needs_flush_plug(tsk
))
4134 blk_schedule_flush_plug(tsk
);
4137 static void sched_update_worker(struct task_struct
*tsk
)
4139 if (tsk
->flags
& PF_WQ_WORKER
)
4140 wq_worker_running(tsk
);
4143 asmlinkage __visible
void __sched
schedule(void)
4145 struct task_struct
*tsk
= current
;
4147 sched_submit_work(tsk
);
4151 sched_preempt_enable_no_resched();
4152 } while (need_resched());
4153 sched_update_worker(tsk
);
4155 EXPORT_SYMBOL(schedule
);
4158 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4159 * state (have scheduled out non-voluntarily) by making sure that all
4160 * tasks have either left the run queue or have gone into user space.
4161 * As idle tasks do not do either, they must not ever be preempted
4162 * (schedule out non-voluntarily).
4164 * schedule_idle() is similar to schedule_preempt_disable() except that it
4165 * never enables preemption because it does not call sched_submit_work().
4167 void __sched
schedule_idle(void)
4170 * As this skips calling sched_submit_work(), which the idle task does
4171 * regardless because that function is a nop when the task is in a
4172 * TASK_RUNNING state, make sure this isn't used someplace that the
4173 * current task can be in any other state. Note, idle is always in the
4174 * TASK_RUNNING state.
4176 WARN_ON_ONCE(current
->state
);
4179 } while (need_resched());
4182 #ifdef CONFIG_CONTEXT_TRACKING
4183 asmlinkage __visible
void __sched
schedule_user(void)
4186 * If we come here after a random call to set_need_resched(),
4187 * or we have been woken up remotely but the IPI has not yet arrived,
4188 * we haven't yet exited the RCU idle mode. Do it here manually until
4189 * we find a better solution.
4191 * NB: There are buggy callers of this function. Ideally we
4192 * should warn if prev_state != CONTEXT_USER, but that will trigger
4193 * too frequently to make sense yet.
4195 enum ctx_state prev_state
= exception_enter();
4197 exception_exit(prev_state
);
4202 * schedule_preempt_disabled - called with preemption disabled
4204 * Returns with preemption disabled. Note: preempt_count must be 1
4206 void __sched
schedule_preempt_disabled(void)
4208 sched_preempt_enable_no_resched();
4213 static void __sched notrace
preempt_schedule_common(void)
4217 * Because the function tracer can trace preempt_count_sub()
4218 * and it also uses preempt_enable/disable_notrace(), if
4219 * NEED_RESCHED is set, the preempt_enable_notrace() called
4220 * by the function tracer will call this function again and
4221 * cause infinite recursion.
4223 * Preemption must be disabled here before the function
4224 * tracer can trace. Break up preempt_disable() into two
4225 * calls. One to disable preemption without fear of being
4226 * traced. The other to still record the preemption latency,
4227 * which can also be traced by the function tracer.
4229 preempt_disable_notrace();
4230 preempt_latency_start(1);
4232 preempt_latency_stop(1);
4233 preempt_enable_no_resched_notrace();
4236 * Check again in case we missed a preemption opportunity
4237 * between schedule and now.
4239 } while (need_resched());
4242 #ifdef CONFIG_PREEMPTION
4244 * This is the entry point to schedule() from in-kernel preemption
4245 * off of preempt_enable.
4247 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
4250 * If there is a non-zero preempt_count or interrupts are disabled,
4251 * we do not want to preempt the current task. Just return..
4253 if (likely(!preemptible()))
4256 preempt_schedule_common();
4258 NOKPROBE_SYMBOL(preempt_schedule
);
4259 EXPORT_SYMBOL(preempt_schedule
);
4262 * preempt_schedule_notrace - preempt_schedule called by tracing
4264 * The tracing infrastructure uses preempt_enable_notrace to prevent
4265 * recursion and tracing preempt enabling caused by the tracing
4266 * infrastructure itself. But as tracing can happen in areas coming
4267 * from userspace or just about to enter userspace, a preempt enable
4268 * can occur before user_exit() is called. This will cause the scheduler
4269 * to be called when the system is still in usermode.
4271 * To prevent this, the preempt_enable_notrace will use this function
4272 * instead of preempt_schedule() to exit user context if needed before
4273 * calling the scheduler.
4275 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
4277 enum ctx_state prev_ctx
;
4279 if (likely(!preemptible()))
4284 * Because the function tracer can trace preempt_count_sub()
4285 * and it also uses preempt_enable/disable_notrace(), if
4286 * NEED_RESCHED is set, the preempt_enable_notrace() called
4287 * by the function tracer will call this function again and
4288 * cause infinite recursion.
4290 * Preemption must be disabled here before the function
4291 * tracer can trace. Break up preempt_disable() into two
4292 * calls. One to disable preemption without fear of being
4293 * traced. The other to still record the preemption latency,
4294 * which can also be traced by the function tracer.
4296 preempt_disable_notrace();
4297 preempt_latency_start(1);
4299 * Needs preempt disabled in case user_exit() is traced
4300 * and the tracer calls preempt_enable_notrace() causing
4301 * an infinite recursion.
4303 prev_ctx
= exception_enter();
4305 exception_exit(prev_ctx
);
4307 preempt_latency_stop(1);
4308 preempt_enable_no_resched_notrace();
4309 } while (need_resched());
4311 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
4313 #endif /* CONFIG_PREEMPTION */
4316 * This is the entry point to schedule() from kernel preemption
4317 * off of irq context.
4318 * Note, that this is called and return with irqs disabled. This will
4319 * protect us against recursive calling from irq.
4321 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
4323 enum ctx_state prev_state
;
4325 /* Catch callers which need to be fixed */
4326 BUG_ON(preempt_count() || !irqs_disabled());
4328 prev_state
= exception_enter();
4334 local_irq_disable();
4335 sched_preempt_enable_no_resched();
4336 } while (need_resched());
4338 exception_exit(prev_state
);
4341 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
4344 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4346 EXPORT_SYMBOL(default_wake_function
);
4348 #ifdef CONFIG_RT_MUTEXES
4350 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
4353 prio
= min(prio
, pi_task
->prio
);
4358 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4360 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
4362 return __rt_effective_prio(pi_task
, prio
);
4366 * rt_mutex_setprio - set the current priority of a task
4368 * @pi_task: donor task
4370 * This function changes the 'effective' priority of a task. It does
4371 * not touch ->normal_prio like __setscheduler().
4373 * Used by the rt_mutex code to implement priority inheritance
4374 * logic. Call site only calls if the priority of the task changed.
4376 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
4378 int prio
, oldprio
, queued
, running
, queue_flag
=
4379 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4380 const struct sched_class
*prev_class
;
4384 /* XXX used to be waiter->prio, not waiter->task->prio */
4385 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
4388 * If nothing changed; bail early.
4390 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
4393 rq
= __task_rq_lock(p
, &rf
);
4394 update_rq_clock(rq
);
4396 * Set under pi_lock && rq->lock, such that the value can be used under
4399 * Note that there is loads of tricky to make this pointer cache work
4400 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4401 * ensure a task is de-boosted (pi_task is set to NULL) before the
4402 * task is allowed to run again (and can exit). This ensures the pointer
4403 * points to a blocked task -- which guaratees the task is present.
4405 p
->pi_top_task
= pi_task
;
4408 * For FIFO/RR we only need to set prio, if that matches we're done.
4410 if (prio
== p
->prio
&& !dl_prio(prio
))
4414 * Idle task boosting is a nono in general. There is one
4415 * exception, when PREEMPT_RT and NOHZ is active:
4417 * The idle task calls get_next_timer_interrupt() and holds
4418 * the timer wheel base->lock on the CPU and another CPU wants
4419 * to access the timer (probably to cancel it). We can safely
4420 * ignore the boosting request, as the idle CPU runs this code
4421 * with interrupts disabled and will complete the lock
4422 * protected section without being interrupted. So there is no
4423 * real need to boost.
4425 if (unlikely(p
== rq
->idle
)) {
4426 WARN_ON(p
!= rq
->curr
);
4427 WARN_ON(p
->pi_blocked_on
);
4431 trace_sched_pi_setprio(p
, pi_task
);
4434 if (oldprio
== prio
)
4435 queue_flag
&= ~DEQUEUE_MOVE
;
4437 prev_class
= p
->sched_class
;
4438 queued
= task_on_rq_queued(p
);
4439 running
= task_current(rq
, p
);
4441 dequeue_task(rq
, p
, queue_flag
);
4443 put_prev_task(rq
, p
);
4446 * Boosting condition are:
4447 * 1. -rt task is running and holds mutex A
4448 * --> -dl task blocks on mutex A
4450 * 2. -dl task is running and holds mutex A
4451 * --> -dl task blocks on mutex A and could preempt the
4454 if (dl_prio(prio
)) {
4455 if (!dl_prio(p
->normal_prio
) ||
4456 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
4457 p
->dl
.dl_boosted
= 1;
4458 queue_flag
|= ENQUEUE_REPLENISH
;
4460 p
->dl
.dl_boosted
= 0;
4461 p
->sched_class
= &dl_sched_class
;
4462 } else if (rt_prio(prio
)) {
4463 if (dl_prio(oldprio
))
4464 p
->dl
.dl_boosted
= 0;
4466 queue_flag
|= ENQUEUE_HEAD
;
4467 p
->sched_class
= &rt_sched_class
;
4469 if (dl_prio(oldprio
))
4470 p
->dl
.dl_boosted
= 0;
4471 if (rt_prio(oldprio
))
4473 p
->sched_class
= &fair_sched_class
;
4479 enqueue_task(rq
, p
, queue_flag
);
4481 set_next_task(rq
, p
);
4483 check_class_changed(rq
, p
, prev_class
, oldprio
);
4485 /* Avoid rq from going away on us: */
4487 __task_rq_unlock(rq
, &rf
);
4489 balance_callback(rq
);
4493 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4499 void set_user_nice(struct task_struct
*p
, long nice
)
4501 bool queued
, running
;
4502 int old_prio
, delta
;
4506 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
4509 * We have to be careful, if called from sys_setpriority(),
4510 * the task might be in the middle of scheduling on another CPU.
4512 rq
= task_rq_lock(p
, &rf
);
4513 update_rq_clock(rq
);
4516 * The RT priorities are set via sched_setscheduler(), but we still
4517 * allow the 'normal' nice value to be set - but as expected
4518 * it wont have any effect on scheduling until the task is
4519 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4521 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
4522 p
->static_prio
= NICE_TO_PRIO(nice
);
4525 queued
= task_on_rq_queued(p
);
4526 running
= task_current(rq
, p
);
4528 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
4530 put_prev_task(rq
, p
);
4532 p
->static_prio
= NICE_TO_PRIO(nice
);
4533 set_load_weight(p
, true);
4535 p
->prio
= effective_prio(p
);
4536 delta
= p
->prio
- old_prio
;
4539 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
4541 * If the task increased its priority or is running and
4542 * lowered its priority, then reschedule its CPU:
4544 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4548 set_next_task(rq
, p
);
4550 task_rq_unlock(rq
, p
, &rf
);
4552 EXPORT_SYMBOL(set_user_nice
);
4555 * can_nice - check if a task can reduce its nice value
4559 int can_nice(const struct task_struct
*p
, const int nice
)
4561 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4562 int nice_rlim
= nice_to_rlimit(nice
);
4564 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4565 capable(CAP_SYS_NICE
));
4567 EXPORT_SYMBOL(can_nice
);
4569 #ifdef __ARCH_WANT_SYS_NICE
4572 * sys_nice - change the priority of the current process.
4573 * @increment: priority increment
4575 * sys_setpriority is a more generic, but much slower function that
4576 * does similar things.
4578 SYSCALL_DEFINE1(nice
, int, increment
)
4583 * Setpriority might change our priority at the same moment.
4584 * We don't have to worry. Conceptually one call occurs first
4585 * and we have a single winner.
4587 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
4588 nice
= task_nice(current
) + increment
;
4590 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
4591 if (increment
< 0 && !can_nice(current
, nice
))
4594 retval
= security_task_setnice(current
, nice
);
4598 set_user_nice(current
, nice
);
4605 * task_prio - return the priority value of a given task.
4606 * @p: the task in question.
4608 * Return: The priority value as seen by users in /proc.
4609 * RT tasks are offset by -200. Normal tasks are centered
4610 * around 0, value goes from -16 to +15.
4612 int task_prio(const struct task_struct
*p
)
4614 return p
->prio
- MAX_RT_PRIO
;
4618 * idle_cpu - is a given CPU idle currently?
4619 * @cpu: the processor in question.
4621 * Return: 1 if the CPU is currently idle. 0 otherwise.
4623 int idle_cpu(int cpu
)
4625 struct rq
*rq
= cpu_rq(cpu
);
4627 if (rq
->curr
!= rq
->idle
)
4634 if (!llist_empty(&rq
->wake_list
))
4642 * available_idle_cpu - is a given CPU idle for enqueuing work.
4643 * @cpu: the CPU in question.
4645 * Return: 1 if the CPU is currently idle. 0 otherwise.
4647 int available_idle_cpu(int cpu
)
4652 if (vcpu_is_preempted(cpu
))
4659 * idle_task - return the idle task for a given CPU.
4660 * @cpu: the processor in question.
4662 * Return: The idle task for the CPU @cpu.
4664 struct task_struct
*idle_task(int cpu
)
4666 return cpu_rq(cpu
)->idle
;
4670 * find_process_by_pid - find a process with a matching PID value.
4671 * @pid: the pid in question.
4673 * The task of @pid, if found. %NULL otherwise.
4675 static struct task_struct
*find_process_by_pid(pid_t pid
)
4677 return pid
? find_task_by_vpid(pid
) : current
;
4681 * sched_setparam() passes in -1 for its policy, to let the functions
4682 * it calls know not to change it.
4684 #define SETPARAM_POLICY -1
4686 static void __setscheduler_params(struct task_struct
*p
,
4687 const struct sched_attr
*attr
)
4689 int policy
= attr
->sched_policy
;
4691 if (policy
== SETPARAM_POLICY
)
4696 if (dl_policy(policy
))
4697 __setparam_dl(p
, attr
);
4698 else if (fair_policy(policy
))
4699 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4702 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4703 * !rt_policy. Always setting this ensures that things like
4704 * getparam()/getattr() don't report silly values for !rt tasks.
4706 p
->rt_priority
= attr
->sched_priority
;
4707 p
->normal_prio
= normal_prio(p
);
4708 set_load_weight(p
, true);
4711 /* Actually do priority change: must hold pi & rq lock. */
4712 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4713 const struct sched_attr
*attr
, bool keep_boost
)
4716 * If params can't change scheduling class changes aren't allowed
4719 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
4722 __setscheduler_params(p
, attr
);
4725 * Keep a potential priority boosting if called from
4726 * sched_setscheduler().
4728 p
->prio
= normal_prio(p
);
4730 p
->prio
= rt_effective_prio(p
, p
->prio
);
4732 if (dl_prio(p
->prio
))
4733 p
->sched_class
= &dl_sched_class
;
4734 else if (rt_prio(p
->prio
))
4735 p
->sched_class
= &rt_sched_class
;
4737 p
->sched_class
= &fair_sched_class
;
4741 * Check the target process has a UID that matches the current process's:
4743 static bool check_same_owner(struct task_struct
*p
)
4745 const struct cred
*cred
= current_cred(), *pcred
;
4749 pcred
= __task_cred(p
);
4750 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4751 uid_eq(cred
->euid
, pcred
->uid
));
4756 static int __sched_setscheduler(struct task_struct
*p
,
4757 const struct sched_attr
*attr
,
4760 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4761 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4762 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4763 int new_effective_prio
, policy
= attr
->sched_policy
;
4764 const struct sched_class
*prev_class
;
4767 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4770 /* The pi code expects interrupts enabled */
4771 BUG_ON(pi
&& in_interrupt());
4773 /* Double check policy once rq lock held: */
4775 reset_on_fork
= p
->sched_reset_on_fork
;
4776 policy
= oldpolicy
= p
->policy
;
4778 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4780 if (!valid_policy(policy
))
4784 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4788 * Valid priorities for SCHED_FIFO and SCHED_RR are
4789 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4790 * SCHED_BATCH and SCHED_IDLE is 0.
4792 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4793 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4795 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4796 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4800 * Allow unprivileged RT tasks to decrease priority:
4802 if (user
&& !capable(CAP_SYS_NICE
)) {
4803 if (fair_policy(policy
)) {
4804 if (attr
->sched_nice
< task_nice(p
) &&
4805 !can_nice(p
, attr
->sched_nice
))
4809 if (rt_policy(policy
)) {
4810 unsigned long rlim_rtprio
=
4811 task_rlimit(p
, RLIMIT_RTPRIO
);
4813 /* Can't set/change the rt policy: */
4814 if (policy
!= p
->policy
&& !rlim_rtprio
)
4817 /* Can't increase priority: */
4818 if (attr
->sched_priority
> p
->rt_priority
&&
4819 attr
->sched_priority
> rlim_rtprio
)
4824 * Can't set/change SCHED_DEADLINE policy at all for now
4825 * (safest behavior); in the future we would like to allow
4826 * unprivileged DL tasks to increase their relative deadline
4827 * or reduce their runtime (both ways reducing utilization)
4829 if (dl_policy(policy
))
4833 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4834 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4836 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
4837 if (!can_nice(p
, task_nice(p
)))
4841 /* Can't change other user's priorities: */
4842 if (!check_same_owner(p
))
4845 /* Normal users shall not reset the sched_reset_on_fork flag: */
4846 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4851 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4854 retval
= security_task_setscheduler(p
);
4859 /* Update task specific "requested" clamps */
4860 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
4861 retval
= uclamp_validate(p
, attr
);
4870 * Make sure no PI-waiters arrive (or leave) while we are
4871 * changing the priority of the task:
4873 * To be able to change p->policy safely, the appropriate
4874 * runqueue lock must be held.
4876 rq
= task_rq_lock(p
, &rf
);
4877 update_rq_clock(rq
);
4880 * Changing the policy of the stop threads its a very bad idea:
4882 if (p
== rq
->stop
) {
4888 * If not changing anything there's no need to proceed further,
4889 * but store a possible modification of reset_on_fork.
4891 if (unlikely(policy
== p
->policy
)) {
4892 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4894 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4896 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4898 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
4901 p
->sched_reset_on_fork
= reset_on_fork
;
4908 #ifdef CONFIG_RT_GROUP_SCHED
4910 * Do not allow realtime tasks into groups that have no runtime
4913 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4914 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4915 !task_group_is_autogroup(task_group(p
))) {
4921 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
4922 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
4923 cpumask_t
*span
= rq
->rd
->span
;
4926 * Don't allow tasks with an affinity mask smaller than
4927 * the entire root_domain to become SCHED_DEADLINE. We
4928 * will also fail if there's no bandwidth available.
4930 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
4931 rq
->rd
->dl_bw
.bw
== 0) {
4939 /* Re-check policy now with rq lock held: */
4940 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4941 policy
= oldpolicy
= -1;
4942 task_rq_unlock(rq
, p
, &rf
);
4944 cpuset_read_unlock();
4949 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4950 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4953 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
4958 p
->sched_reset_on_fork
= reset_on_fork
;
4963 * Take priority boosted tasks into account. If the new
4964 * effective priority is unchanged, we just store the new
4965 * normal parameters and do not touch the scheduler class and
4966 * the runqueue. This will be done when the task deboost
4969 new_effective_prio
= rt_effective_prio(p
, newprio
);
4970 if (new_effective_prio
== oldprio
)
4971 queue_flags
&= ~DEQUEUE_MOVE
;
4974 queued
= task_on_rq_queued(p
);
4975 running
= task_current(rq
, p
);
4977 dequeue_task(rq
, p
, queue_flags
);
4979 put_prev_task(rq
, p
);
4981 prev_class
= p
->sched_class
;
4983 __setscheduler(rq
, p
, attr
, pi
);
4984 __setscheduler_uclamp(p
, attr
);
4988 * We enqueue to tail when the priority of a task is
4989 * increased (user space view).
4991 if (oldprio
< p
->prio
)
4992 queue_flags
|= ENQUEUE_HEAD
;
4994 enqueue_task(rq
, p
, queue_flags
);
4997 set_next_task(rq
, p
);
4999 check_class_changed(rq
, p
, prev_class
, oldprio
);
5001 /* Avoid rq from going away on us: */
5003 task_rq_unlock(rq
, p
, &rf
);
5006 cpuset_read_unlock();
5007 rt_mutex_adjust_pi(p
);
5010 /* Run balance callbacks after we've adjusted the PI chain: */
5011 balance_callback(rq
);
5017 task_rq_unlock(rq
, p
, &rf
);
5019 cpuset_read_unlock();
5023 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
5024 const struct sched_param
*param
, bool check
)
5026 struct sched_attr attr
= {
5027 .sched_policy
= policy
,
5028 .sched_priority
= param
->sched_priority
,
5029 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
5032 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5033 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
5034 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5035 policy
&= ~SCHED_RESET_ON_FORK
;
5036 attr
.sched_policy
= policy
;
5039 return __sched_setscheduler(p
, &attr
, check
, true);
5042 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5043 * @p: the task in question.
5044 * @policy: new policy.
5045 * @param: structure containing the new RT priority.
5047 * Return: 0 on success. An error code otherwise.
5049 * NOTE that the task may be already dead.
5051 int sched_setscheduler(struct task_struct
*p
, int policy
,
5052 const struct sched_param
*param
)
5054 return _sched_setscheduler(p
, policy
, param
, true);
5056 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5058 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
5060 return __sched_setscheduler(p
, attr
, true, true);
5062 EXPORT_SYMBOL_GPL(sched_setattr
);
5064 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
5066 return __sched_setscheduler(p
, attr
, false, true);
5070 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5071 * @p: the task in question.
5072 * @policy: new policy.
5073 * @param: structure containing the new RT priority.
5075 * Just like sched_setscheduler, only don't bother checking if the
5076 * current context has permission. For example, this is needed in
5077 * stop_machine(): we create temporary high priority worker threads,
5078 * but our caller might not have that capability.
5080 * Return: 0 on success. An error code otherwise.
5082 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5083 const struct sched_param
*param
)
5085 return _sched_setscheduler(p
, policy
, param
, false);
5087 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
5090 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5092 struct sched_param lparam
;
5093 struct task_struct
*p
;
5096 if (!param
|| pid
< 0)
5098 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5103 p
= find_process_by_pid(pid
);
5109 retval
= sched_setscheduler(p
, policy
, &lparam
);
5117 * Mimics kernel/events/core.c perf_copy_attr().
5119 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
5124 /* Zero the full structure, so that a short copy will be nice: */
5125 memset(attr
, 0, sizeof(*attr
));
5127 ret
= get_user(size
, &uattr
->size
);
5131 /* ABI compatibility quirk: */
5133 size
= SCHED_ATTR_SIZE_VER0
;
5134 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
5137 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
5144 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
5145 size
< SCHED_ATTR_SIZE_VER1
)
5149 * XXX: Do we want to be lenient like existing syscalls; or do we want
5150 * to be strict and return an error on out-of-bounds values?
5152 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
5157 put_user(sizeof(*attr
), &uattr
->size
);
5162 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5163 * @pid: the pid in question.
5164 * @policy: new policy.
5165 * @param: structure containing the new RT priority.
5167 * Return: 0 on success. An error code otherwise.
5169 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
5174 return do_sched_setscheduler(pid
, policy
, param
);
5178 * sys_sched_setparam - set/change the RT priority of a thread
5179 * @pid: the pid in question.
5180 * @param: structure containing the new RT priority.
5182 * Return: 0 on success. An error code otherwise.
5184 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5186 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
5190 * sys_sched_setattr - same as above, but with extended sched_attr
5191 * @pid: the pid in question.
5192 * @uattr: structure containing the extended parameters.
5193 * @flags: for future extension.
5195 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5196 unsigned int, flags
)
5198 struct sched_attr attr
;
5199 struct task_struct
*p
;
5202 if (!uattr
|| pid
< 0 || flags
)
5205 retval
= sched_copy_attr(uattr
, &attr
);
5209 if ((int)attr
.sched_policy
< 0)
5211 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
5212 attr
.sched_policy
= SETPARAM_POLICY
;
5216 p
= find_process_by_pid(pid
);
5222 retval
= sched_setattr(p
, &attr
);
5230 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5231 * @pid: the pid in question.
5233 * Return: On success, the policy of the thread. Otherwise, a negative error
5236 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5238 struct task_struct
*p
;
5246 p
= find_process_by_pid(pid
);
5248 retval
= security_task_getscheduler(p
);
5251 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5258 * sys_sched_getparam - get the RT priority of a thread
5259 * @pid: the pid in question.
5260 * @param: structure containing the RT priority.
5262 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5265 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5267 struct sched_param lp
= { .sched_priority
= 0 };
5268 struct task_struct
*p
;
5271 if (!param
|| pid
< 0)
5275 p
= find_process_by_pid(pid
);
5280 retval
= security_task_getscheduler(p
);
5284 if (task_has_rt_policy(p
))
5285 lp
.sched_priority
= p
->rt_priority
;
5289 * This one might sleep, we cannot do it with a spinlock held ...
5291 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5301 * Copy the kernel size attribute structure (which might be larger
5302 * than what user-space knows about) to user-space.
5304 * Note that all cases are valid: user-space buffer can be larger or
5305 * smaller than the kernel-space buffer. The usual case is that both
5306 * have the same size.
5309 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
5310 struct sched_attr
*kattr
,
5313 unsigned int ksize
= sizeof(*kattr
);
5315 if (!access_ok(uattr
, usize
))
5319 * sched_getattr() ABI forwards and backwards compatibility:
5321 * If usize == ksize then we just copy everything to user-space and all is good.
5323 * If usize < ksize then we only copy as much as user-space has space for,
5324 * this keeps ABI compatibility as well. We skip the rest.
5326 * If usize > ksize then user-space is using a newer version of the ABI,
5327 * which part the kernel doesn't know about. Just ignore it - tooling can
5328 * detect the kernel's knowledge of attributes from the attr->size value
5329 * which is set to ksize in this case.
5331 kattr
->size
= min(usize
, ksize
);
5333 if (copy_to_user(uattr
, kattr
, kattr
->size
))
5340 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5341 * @pid: the pid in question.
5342 * @uattr: structure containing the extended parameters.
5343 * @usize: sizeof(attr) for fwd/bwd comp.
5344 * @flags: for future extension.
5346 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5347 unsigned int, usize
, unsigned int, flags
)
5349 struct sched_attr kattr
= { };
5350 struct task_struct
*p
;
5353 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
5354 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
5358 p
= find_process_by_pid(pid
);
5363 retval
= security_task_getscheduler(p
);
5367 kattr
.sched_policy
= p
->policy
;
5368 if (p
->sched_reset_on_fork
)
5369 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5370 if (task_has_dl_policy(p
))
5371 __getparam_dl(p
, &kattr
);
5372 else if (task_has_rt_policy(p
))
5373 kattr
.sched_priority
= p
->rt_priority
;
5375 kattr
.sched_nice
= task_nice(p
);
5377 #ifdef CONFIG_UCLAMP_TASK
5378 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
5379 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
5384 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
5391 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5393 cpumask_var_t cpus_allowed
, new_mask
;
5394 struct task_struct
*p
;
5399 p
= find_process_by_pid(pid
);
5405 /* Prevent p going away */
5409 if (p
->flags
& PF_NO_SETAFFINITY
) {
5413 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5417 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5419 goto out_free_cpus_allowed
;
5422 if (!check_same_owner(p
)) {
5424 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
5426 goto out_free_new_mask
;
5431 retval
= security_task_setscheduler(p
);
5433 goto out_free_new_mask
;
5436 cpuset_cpus_allowed(p
, cpus_allowed
);
5437 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5440 * Since bandwidth control happens on root_domain basis,
5441 * if admission test is enabled, we only admit -deadline
5442 * tasks allowed to run on all the CPUs in the task's
5446 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
5448 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
5451 goto out_free_new_mask
;
5457 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
5460 cpuset_cpus_allowed(p
, cpus_allowed
);
5461 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5463 * We must have raced with a concurrent cpuset
5464 * update. Just reset the cpus_allowed to the
5465 * cpuset's cpus_allowed
5467 cpumask_copy(new_mask
, cpus_allowed
);
5472 free_cpumask_var(new_mask
);
5473 out_free_cpus_allowed
:
5474 free_cpumask_var(cpus_allowed
);
5480 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5481 struct cpumask
*new_mask
)
5483 if (len
< cpumask_size())
5484 cpumask_clear(new_mask
);
5485 else if (len
> cpumask_size())
5486 len
= cpumask_size();
5488 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5492 * sys_sched_setaffinity - set the CPU affinity of a process
5493 * @pid: pid of the process
5494 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5495 * @user_mask_ptr: user-space pointer to the new CPU mask
5497 * Return: 0 on success. An error code otherwise.
5499 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5500 unsigned long __user
*, user_mask_ptr
)
5502 cpumask_var_t new_mask
;
5505 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5508 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5510 retval
= sched_setaffinity(pid
, new_mask
);
5511 free_cpumask_var(new_mask
);
5515 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5517 struct task_struct
*p
;
5518 unsigned long flags
;
5524 p
= find_process_by_pid(pid
);
5528 retval
= security_task_getscheduler(p
);
5532 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5533 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
5534 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5543 * sys_sched_getaffinity - get the CPU affinity of a process
5544 * @pid: pid of the process
5545 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5546 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5548 * Return: size of CPU mask copied to user_mask_ptr on success. An
5549 * error code otherwise.
5551 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5552 unsigned long __user
*, user_mask_ptr
)
5557 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5559 if (len
& (sizeof(unsigned long)-1))
5562 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5565 ret
= sched_getaffinity(pid
, mask
);
5567 unsigned int retlen
= min(len
, cpumask_size());
5569 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5574 free_cpumask_var(mask
);
5580 * sys_sched_yield - yield the current processor to other threads.
5582 * This function yields the current CPU to other tasks. If there are no
5583 * other threads running on this CPU then this function will return.
5587 static void do_sched_yield(void)
5592 rq
= this_rq_lock_irq(&rf
);
5594 schedstat_inc(rq
->yld_count
);
5595 current
->sched_class
->yield_task(rq
);
5598 * Since we are going to call schedule() anyway, there's
5599 * no need to preempt or enable interrupts:
5603 sched_preempt_enable_no_resched();
5608 SYSCALL_DEFINE0(sched_yield
)
5614 #ifndef CONFIG_PREEMPTION
5615 int __sched
_cond_resched(void)
5617 if (should_resched(0)) {
5618 preempt_schedule_common();
5624 EXPORT_SYMBOL(_cond_resched
);
5628 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5629 * call schedule, and on return reacquire the lock.
5631 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5632 * operations here to prevent schedule() from being called twice (once via
5633 * spin_unlock(), once by hand).
5635 int __cond_resched_lock(spinlock_t
*lock
)
5637 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5640 lockdep_assert_held(lock
);
5642 if (spin_needbreak(lock
) || resched
) {
5645 preempt_schedule_common();
5653 EXPORT_SYMBOL(__cond_resched_lock
);
5656 * yield - yield the current processor to other threads.
5658 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5660 * The scheduler is at all times free to pick the calling task as the most
5661 * eligible task to run, if removing the yield() call from your code breaks
5662 * it, its already broken.
5664 * Typical broken usage is:
5669 * where one assumes that yield() will let 'the other' process run that will
5670 * make event true. If the current task is a SCHED_FIFO task that will never
5671 * happen. Never use yield() as a progress guarantee!!
5673 * If you want to use yield() to wait for something, use wait_event().
5674 * If you want to use yield() to be 'nice' for others, use cond_resched().
5675 * If you still want to use yield(), do not!
5677 void __sched
yield(void)
5679 set_current_state(TASK_RUNNING
);
5682 EXPORT_SYMBOL(yield
);
5685 * yield_to - yield the current processor to another thread in
5686 * your thread group, or accelerate that thread toward the
5687 * processor it's on.
5689 * @preempt: whether task preemption is allowed or not
5691 * It's the caller's job to ensure that the target task struct
5692 * can't go away on us before we can do any checks.
5695 * true (>0) if we indeed boosted the target task.
5696 * false (0) if we failed to boost the target.
5697 * -ESRCH if there's no task to yield to.
5699 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5701 struct task_struct
*curr
= current
;
5702 struct rq
*rq
, *p_rq
;
5703 unsigned long flags
;
5706 local_irq_save(flags
);
5712 * If we're the only runnable task on the rq and target rq also
5713 * has only one task, there's absolutely no point in yielding.
5715 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5720 double_rq_lock(rq
, p_rq
);
5721 if (task_rq(p
) != p_rq
) {
5722 double_rq_unlock(rq
, p_rq
);
5726 if (!curr
->sched_class
->yield_to_task
)
5729 if (curr
->sched_class
!= p
->sched_class
)
5732 if (task_running(p_rq
, p
) || p
->state
)
5735 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5737 schedstat_inc(rq
->yld_count
);
5739 * Make p's CPU reschedule; pick_next_entity takes care of
5742 if (preempt
&& rq
!= p_rq
)
5747 double_rq_unlock(rq
, p_rq
);
5749 local_irq_restore(flags
);
5756 EXPORT_SYMBOL_GPL(yield_to
);
5758 int io_schedule_prepare(void)
5760 int old_iowait
= current
->in_iowait
;
5762 current
->in_iowait
= 1;
5763 blk_schedule_flush_plug(current
);
5768 void io_schedule_finish(int token
)
5770 current
->in_iowait
= token
;
5774 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5775 * that process accounting knows that this is a task in IO wait state.
5777 long __sched
io_schedule_timeout(long timeout
)
5782 token
= io_schedule_prepare();
5783 ret
= schedule_timeout(timeout
);
5784 io_schedule_finish(token
);
5788 EXPORT_SYMBOL(io_schedule_timeout
);
5790 void __sched
io_schedule(void)
5794 token
= io_schedule_prepare();
5796 io_schedule_finish(token
);
5798 EXPORT_SYMBOL(io_schedule
);
5801 * sys_sched_get_priority_max - return maximum RT priority.
5802 * @policy: scheduling class.
5804 * Return: On success, this syscall returns the maximum
5805 * rt_priority that can be used by a given scheduling class.
5806 * On failure, a negative error code is returned.
5808 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5815 ret
= MAX_USER_RT_PRIO
-1;
5817 case SCHED_DEADLINE
:
5828 * sys_sched_get_priority_min - return minimum RT priority.
5829 * @policy: scheduling class.
5831 * Return: On success, this syscall returns the minimum
5832 * rt_priority that can be used by a given scheduling class.
5833 * On failure, a negative error code is returned.
5835 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5844 case SCHED_DEADLINE
:
5853 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5855 struct task_struct
*p
;
5856 unsigned int time_slice
;
5866 p
= find_process_by_pid(pid
);
5870 retval
= security_task_getscheduler(p
);
5874 rq
= task_rq_lock(p
, &rf
);
5876 if (p
->sched_class
->get_rr_interval
)
5877 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5878 task_rq_unlock(rq
, p
, &rf
);
5881 jiffies_to_timespec64(time_slice
, t
);
5890 * sys_sched_rr_get_interval - return the default timeslice of a process.
5891 * @pid: pid of the process.
5892 * @interval: userspace pointer to the timeslice value.
5894 * this syscall writes the default timeslice value of a given process
5895 * into the user-space timespec buffer. A value of '0' means infinity.
5897 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5900 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5901 struct __kernel_timespec __user
*, interval
)
5903 struct timespec64 t
;
5904 int retval
= sched_rr_get_interval(pid
, &t
);
5907 retval
= put_timespec64(&t
, interval
);
5912 #ifdef CONFIG_COMPAT_32BIT_TIME
5913 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
5914 struct old_timespec32 __user
*, interval
)
5916 struct timespec64 t
;
5917 int retval
= sched_rr_get_interval(pid
, &t
);
5920 retval
= put_old_timespec32(&t
, interval
);
5925 void sched_show_task(struct task_struct
*p
)
5927 unsigned long free
= 0;
5930 if (!try_get_task_stack(p
))
5933 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5935 if (p
->state
== TASK_RUNNING
)
5936 printk(KERN_CONT
" running task ");
5937 #ifdef CONFIG_DEBUG_STACK_USAGE
5938 free
= stack_not_used(p
);
5943 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5945 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5946 task_pid_nr(p
), ppid
,
5947 (unsigned long)task_thread_info(p
)->flags
);
5949 print_worker_info(KERN_INFO
, p
);
5950 show_stack(p
, NULL
);
5953 EXPORT_SYMBOL_GPL(sched_show_task
);
5956 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
5958 /* no filter, everything matches */
5962 /* filter, but doesn't match */
5963 if (!(p
->state
& state_filter
))
5967 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5970 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
5977 void show_state_filter(unsigned long state_filter
)
5979 struct task_struct
*g
, *p
;
5981 #if BITS_PER_LONG == 32
5983 " task PC stack pid father\n");
5986 " task PC stack pid father\n");
5989 for_each_process_thread(g
, p
) {
5991 * reset the NMI-timeout, listing all files on a slow
5992 * console might take a lot of time:
5993 * Also, reset softlockup watchdogs on all CPUs, because
5994 * another CPU might be blocked waiting for us to process
5997 touch_nmi_watchdog();
5998 touch_all_softlockup_watchdogs();
5999 if (state_filter_match(state_filter
, p
))
6003 #ifdef CONFIG_SCHED_DEBUG
6005 sysrq_sched_debug_show();
6009 * Only show locks if all tasks are dumped:
6012 debug_show_all_locks();
6016 * init_idle - set up an idle thread for a given CPU
6017 * @idle: task in question
6018 * @cpu: CPU the idle task belongs to
6020 * NOTE: this function does not set the idle thread's NEED_RESCHED
6021 * flag, to make booting more robust.
6023 void init_idle(struct task_struct
*idle
, int cpu
)
6025 struct rq
*rq
= cpu_rq(cpu
);
6026 unsigned long flags
;
6028 __sched_fork(0, idle
);
6030 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
6031 raw_spin_lock(&rq
->lock
);
6033 idle
->state
= TASK_RUNNING
;
6034 idle
->se
.exec_start
= sched_clock();
6035 idle
->flags
|= PF_IDLE
;
6037 kasan_unpoison_task_stack(idle
);
6041 * Its possible that init_idle() gets called multiple times on a task,
6042 * in that case do_set_cpus_allowed() will not do the right thing.
6044 * And since this is boot we can forgo the serialization.
6046 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
6049 * We're having a chicken and egg problem, even though we are
6050 * holding rq->lock, the CPU isn't yet set to this CPU so the
6051 * lockdep check in task_group() will fail.
6053 * Similar case to sched_fork(). / Alternatively we could
6054 * use task_rq_lock() here and obtain the other rq->lock.
6059 __set_task_cpu(idle
, cpu
);
6063 rcu_assign_pointer(rq
->curr
, idle
);
6064 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
6068 raw_spin_unlock(&rq
->lock
);
6069 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
6071 /* Set the preempt count _outside_ the spinlocks! */
6072 init_idle_preempt_count(idle
, cpu
);
6075 * The idle tasks have their own, simple scheduling class:
6077 idle
->sched_class
= &idle_sched_class
;
6078 ftrace_graph_init_idle_task(idle
, cpu
);
6079 vtime_init_idle(idle
, cpu
);
6081 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
6087 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
6088 const struct cpumask
*trial
)
6092 if (!cpumask_weight(cur
))
6095 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
6100 int task_can_attach(struct task_struct
*p
,
6101 const struct cpumask
*cs_cpus_allowed
)
6106 * Kthreads which disallow setaffinity shouldn't be moved
6107 * to a new cpuset; we don't want to change their CPU
6108 * affinity and isolating such threads by their set of
6109 * allowed nodes is unnecessary. Thus, cpusets are not
6110 * applicable for such threads. This prevents checking for
6111 * success of set_cpus_allowed_ptr() on all attached tasks
6112 * before cpus_mask may be changed.
6114 if (p
->flags
& PF_NO_SETAFFINITY
) {
6119 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
6121 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
6127 bool sched_smp_initialized __read_mostly
;
6129 #ifdef CONFIG_NUMA_BALANCING
6130 /* Migrate current task p to target_cpu */
6131 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
6133 struct migration_arg arg
= { p
, target_cpu
};
6134 int curr_cpu
= task_cpu(p
);
6136 if (curr_cpu
== target_cpu
)
6139 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
6142 /* TODO: This is not properly updating schedstats */
6144 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
6145 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
6149 * Requeue a task on a given node and accurately track the number of NUMA
6150 * tasks on the runqueues
6152 void sched_setnuma(struct task_struct
*p
, int nid
)
6154 bool queued
, running
;
6158 rq
= task_rq_lock(p
, &rf
);
6159 queued
= task_on_rq_queued(p
);
6160 running
= task_current(rq
, p
);
6163 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
6165 put_prev_task(rq
, p
);
6167 p
->numa_preferred_nid
= nid
;
6170 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
6172 set_next_task(rq
, p
);
6173 task_rq_unlock(rq
, p
, &rf
);
6175 #endif /* CONFIG_NUMA_BALANCING */
6177 #ifdef CONFIG_HOTPLUG_CPU
6179 * Ensure that the idle task is using init_mm right before its CPU goes
6182 void idle_task_exit(void)
6184 struct mm_struct
*mm
= current
->active_mm
;
6186 BUG_ON(cpu_online(smp_processor_id()));
6188 if (mm
!= &init_mm
) {
6189 switch_mm(mm
, &init_mm
, current
);
6190 current
->active_mm
= &init_mm
;
6191 finish_arch_post_lock_switch();
6197 * Since this CPU is going 'away' for a while, fold any nr_active delta
6198 * we might have. Assumes we're called after migrate_tasks() so that the
6199 * nr_active count is stable. We need to take the teardown thread which
6200 * is calling this into account, so we hand in adjust = 1 to the load
6203 * Also see the comment "Global load-average calculations".
6205 static void calc_load_migrate(struct rq
*rq
)
6207 long delta
= calc_load_fold_active(rq
, 1);
6209 atomic_long_add(delta
, &calc_load_tasks
);
6212 static struct task_struct
*__pick_migrate_task(struct rq
*rq
)
6214 const struct sched_class
*class;
6215 struct task_struct
*next
;
6217 for_each_class(class) {
6218 next
= class->pick_next_task(rq
, NULL
, NULL
);
6220 next
->sched_class
->put_prev_task(rq
, next
);
6225 /* The idle class should always have a runnable task */
6230 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6231 * try_to_wake_up()->select_task_rq().
6233 * Called with rq->lock held even though we'er in stop_machine() and
6234 * there's no concurrency possible, we hold the required locks anyway
6235 * because of lock validation efforts.
6237 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
6239 struct rq
*rq
= dead_rq
;
6240 struct task_struct
*next
, *stop
= rq
->stop
;
6241 struct rq_flags orf
= *rf
;
6245 * Fudge the rq selection such that the below task selection loop
6246 * doesn't get stuck on the currently eligible stop task.
6248 * We're currently inside stop_machine() and the rq is either stuck
6249 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6250 * either way we should never end up calling schedule() until we're
6256 * put_prev_task() and pick_next_task() sched
6257 * class method both need to have an up-to-date
6258 * value of rq->clock[_task]
6260 update_rq_clock(rq
);
6264 * There's this thread running, bail when that's the only
6267 if (rq
->nr_running
== 1)
6270 next
= __pick_migrate_task(rq
);
6273 * Rules for changing task_struct::cpus_mask are holding
6274 * both pi_lock and rq->lock, such that holding either
6275 * stabilizes the mask.
6277 * Drop rq->lock is not quite as disastrous as it usually is
6278 * because !cpu_active at this point, which means load-balance
6279 * will not interfere. Also, stop-machine.
6282 raw_spin_lock(&next
->pi_lock
);
6286 * Since we're inside stop-machine, _nothing_ should have
6287 * changed the task, WARN if weird stuff happened, because in
6288 * that case the above rq->lock drop is a fail too.
6290 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
6291 raw_spin_unlock(&next
->pi_lock
);
6295 /* Find suitable destination for @next, with force if needed. */
6296 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
6297 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
6298 if (rq
!= dead_rq
) {
6304 raw_spin_unlock(&next
->pi_lock
);
6309 #endif /* CONFIG_HOTPLUG_CPU */
6311 void set_rq_online(struct rq
*rq
)
6314 const struct sched_class
*class;
6316 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6319 for_each_class(class) {
6320 if (class->rq_online
)
6321 class->rq_online(rq
);
6326 void set_rq_offline(struct rq
*rq
)
6329 const struct sched_class
*class;
6331 for_each_class(class) {
6332 if (class->rq_offline
)
6333 class->rq_offline(rq
);
6336 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6342 * used to mark begin/end of suspend/resume:
6344 static int num_cpus_frozen
;
6347 * Update cpusets according to cpu_active mask. If cpusets are
6348 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6349 * around partition_sched_domains().
6351 * If we come here as part of a suspend/resume, don't touch cpusets because we
6352 * want to restore it back to its original state upon resume anyway.
6354 static void cpuset_cpu_active(void)
6356 if (cpuhp_tasks_frozen
) {
6358 * num_cpus_frozen tracks how many CPUs are involved in suspend
6359 * resume sequence. As long as this is not the last online
6360 * operation in the resume sequence, just build a single sched
6361 * domain, ignoring cpusets.
6363 partition_sched_domains(1, NULL
, NULL
);
6364 if (--num_cpus_frozen
)
6367 * This is the last CPU online operation. So fall through and
6368 * restore the original sched domains by considering the
6369 * cpuset configurations.
6371 cpuset_force_rebuild();
6373 cpuset_update_active_cpus();
6376 static int cpuset_cpu_inactive(unsigned int cpu
)
6378 if (!cpuhp_tasks_frozen
) {
6379 if (dl_cpu_busy(cpu
))
6381 cpuset_update_active_cpus();
6384 partition_sched_domains(1, NULL
, NULL
);
6389 int sched_cpu_activate(unsigned int cpu
)
6391 struct rq
*rq
= cpu_rq(cpu
);
6394 #ifdef CONFIG_SCHED_SMT
6396 * When going up, increment the number of cores with SMT present.
6398 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6399 static_branch_inc_cpuslocked(&sched_smt_present
);
6401 set_cpu_active(cpu
, true);
6403 if (sched_smp_initialized
) {
6404 sched_domains_numa_masks_set(cpu
);
6405 cpuset_cpu_active();
6409 * Put the rq online, if not already. This happens:
6411 * 1) In the early boot process, because we build the real domains
6412 * after all CPUs have been brought up.
6414 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6417 rq_lock_irqsave(rq
, &rf
);
6419 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6422 rq_unlock_irqrestore(rq
, &rf
);
6427 int sched_cpu_deactivate(unsigned int cpu
)
6431 set_cpu_active(cpu
, false);
6433 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6434 * users of this state to go away such that all new such users will
6437 * Do sync before park smpboot threads to take care the rcu boost case.
6441 #ifdef CONFIG_SCHED_SMT
6443 * When going down, decrement the number of cores with SMT present.
6445 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6446 static_branch_dec_cpuslocked(&sched_smt_present
);
6449 if (!sched_smp_initialized
)
6452 ret
= cpuset_cpu_inactive(cpu
);
6454 set_cpu_active(cpu
, true);
6457 sched_domains_numa_masks_clear(cpu
);
6461 static void sched_rq_cpu_starting(unsigned int cpu
)
6463 struct rq
*rq
= cpu_rq(cpu
);
6465 rq
->calc_load_update
= calc_load_update
;
6466 update_max_interval();
6469 int sched_cpu_starting(unsigned int cpu
)
6471 sched_rq_cpu_starting(cpu
);
6472 sched_tick_start(cpu
);
6476 #ifdef CONFIG_HOTPLUG_CPU
6477 int sched_cpu_dying(unsigned int cpu
)
6479 struct rq
*rq
= cpu_rq(cpu
);
6482 /* Handle pending wakeups and then migrate everything off */
6483 sched_ttwu_pending();
6484 sched_tick_stop(cpu
);
6486 rq_lock_irqsave(rq
, &rf
);
6488 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6491 migrate_tasks(rq
, &rf
);
6492 BUG_ON(rq
->nr_running
!= 1);
6493 rq_unlock_irqrestore(rq
, &rf
);
6495 calc_load_migrate(rq
);
6496 update_max_interval();
6497 nohz_balance_exit_idle(rq
);
6503 void __init
sched_init_smp(void)
6508 * There's no userspace yet to cause hotplug operations; hence all the
6509 * CPU masks are stable and all blatant races in the below code cannot
6512 mutex_lock(&sched_domains_mutex
);
6513 sched_init_domains(cpu_active_mask
);
6514 mutex_unlock(&sched_domains_mutex
);
6516 /* Move init over to a non-isolated CPU */
6517 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
6519 sched_init_granularity();
6521 init_sched_rt_class();
6522 init_sched_dl_class();
6524 sched_smp_initialized
= true;
6527 static int __init
migration_init(void)
6529 sched_cpu_starting(smp_processor_id());
6532 early_initcall(migration_init
);
6535 void __init
sched_init_smp(void)
6537 sched_init_granularity();
6539 #endif /* CONFIG_SMP */
6541 int in_sched_functions(unsigned long addr
)
6543 return in_lock_functions(addr
) ||
6544 (addr
>= (unsigned long)__sched_text_start
6545 && addr
< (unsigned long)__sched_text_end
);
6548 #ifdef CONFIG_CGROUP_SCHED
6550 * Default task group.
6551 * Every task in system belongs to this group at bootup.
6553 struct task_group root_task_group
;
6554 LIST_HEAD(task_groups
);
6556 /* Cacheline aligned slab cache for task_group */
6557 static struct kmem_cache
*task_group_cache __read_mostly
;
6560 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6561 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
6563 void __init
sched_init(void)
6565 unsigned long ptr
= 0;
6570 #ifdef CONFIG_FAIR_GROUP_SCHED
6571 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6573 #ifdef CONFIG_RT_GROUP_SCHED
6574 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6577 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
6579 #ifdef CONFIG_FAIR_GROUP_SCHED
6580 root_task_group
.se
= (struct sched_entity
**)ptr
;
6581 ptr
+= nr_cpu_ids
* sizeof(void **);
6583 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6584 ptr
+= nr_cpu_ids
* sizeof(void **);
6586 #endif /* CONFIG_FAIR_GROUP_SCHED */
6587 #ifdef CONFIG_RT_GROUP_SCHED
6588 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6589 ptr
+= nr_cpu_ids
* sizeof(void **);
6591 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6592 ptr
+= nr_cpu_ids
* sizeof(void **);
6594 #endif /* CONFIG_RT_GROUP_SCHED */
6596 #ifdef CONFIG_CPUMASK_OFFSTACK
6597 for_each_possible_cpu(i
) {
6598 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6599 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6600 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6601 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6603 #endif /* CONFIG_CPUMASK_OFFSTACK */
6605 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
6606 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6609 init_defrootdomain();
6612 #ifdef CONFIG_RT_GROUP_SCHED
6613 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6614 global_rt_period(), global_rt_runtime());
6615 #endif /* CONFIG_RT_GROUP_SCHED */
6617 #ifdef CONFIG_CGROUP_SCHED
6618 task_group_cache
= KMEM_CACHE(task_group
, 0);
6620 list_add(&root_task_group
.list
, &task_groups
);
6621 INIT_LIST_HEAD(&root_task_group
.children
);
6622 INIT_LIST_HEAD(&root_task_group
.siblings
);
6623 autogroup_init(&init_task
);
6624 #endif /* CONFIG_CGROUP_SCHED */
6626 for_each_possible_cpu(i
) {
6630 raw_spin_lock_init(&rq
->lock
);
6632 rq
->calc_load_active
= 0;
6633 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6634 init_cfs_rq(&rq
->cfs
);
6635 init_rt_rq(&rq
->rt
);
6636 init_dl_rq(&rq
->dl
);
6637 #ifdef CONFIG_FAIR_GROUP_SCHED
6638 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6639 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6640 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6642 * How much CPU bandwidth does root_task_group get?
6644 * In case of task-groups formed thr' the cgroup filesystem, it
6645 * gets 100% of the CPU resources in the system. This overall
6646 * system CPU resource is divided among the tasks of
6647 * root_task_group and its child task-groups in a fair manner,
6648 * based on each entity's (task or task-group's) weight
6649 * (se->load.weight).
6651 * In other words, if root_task_group has 10 tasks of weight
6652 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6653 * then A0's share of the CPU resource is:
6655 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6657 * We achieve this by letting root_task_group's tasks sit
6658 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6660 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6661 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6662 #endif /* CONFIG_FAIR_GROUP_SCHED */
6664 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6665 #ifdef CONFIG_RT_GROUP_SCHED
6666 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6671 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6672 rq
->balance_callback
= NULL
;
6673 rq
->active_balance
= 0;
6674 rq
->next_balance
= jiffies
;
6679 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6680 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6682 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6684 rq_attach_root(rq
, &def_root_domain
);
6685 #ifdef CONFIG_NO_HZ_COMMON
6686 rq
->last_load_update_tick
= jiffies
;
6687 rq
->last_blocked_load_update_tick
= jiffies
;
6688 atomic_set(&rq
->nohz_flags
, 0);
6690 #endif /* CONFIG_SMP */
6692 atomic_set(&rq
->nr_iowait
, 0);
6695 set_load_weight(&init_task
, false);
6698 * The boot idle thread does lazy MMU switching as well:
6701 enter_lazy_tlb(&init_mm
, current
);
6704 * Make us the idle thread. Technically, schedule() should not be
6705 * called from this thread, however somewhere below it might be,
6706 * but because we are the idle thread, we just pick up running again
6707 * when this runqueue becomes "idle".
6709 init_idle(current
, smp_processor_id());
6711 calc_load_update
= jiffies
+ LOAD_FREQ
;
6714 idle_thread_set_boot_cpu();
6716 init_sched_fair_class();
6724 scheduler_running
= 1;
6727 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6728 static inline int preempt_count_equals(int preempt_offset
)
6730 int nested
= preempt_count() + rcu_preempt_depth();
6732 return (nested
== preempt_offset
);
6735 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6738 * Blocking primitives will set (and therefore destroy) current->state,
6739 * since we will exit with TASK_RUNNING make sure we enter with it,
6740 * otherwise we will destroy state.
6742 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6743 "do not call blocking ops when !TASK_RUNNING; "
6744 "state=%lx set at [<%p>] %pS\n",
6746 (void *)current
->task_state_change
,
6747 (void *)current
->task_state_change
);
6749 ___might_sleep(file
, line
, preempt_offset
);
6751 EXPORT_SYMBOL(__might_sleep
);
6753 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6755 /* Ratelimiting timestamp: */
6756 static unsigned long prev_jiffy
;
6758 unsigned long preempt_disable_ip
;
6760 /* WARN_ON_ONCE() by default, no rate limit required: */
6763 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6764 !is_idle_task(current
) && !current
->non_block_count
) ||
6765 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6769 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6771 prev_jiffy
= jiffies
;
6773 /* Save this before calling printk(), since that will clobber it: */
6774 preempt_disable_ip
= get_preempt_disable_ip(current
);
6777 "BUG: sleeping function called from invalid context at %s:%d\n",
6780 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6781 in_atomic(), irqs_disabled(), current
->non_block_count
,
6782 current
->pid
, current
->comm
);
6784 if (task_stack_end_corrupted(current
))
6785 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6787 debug_show_held_locks(current
);
6788 if (irqs_disabled())
6789 print_irqtrace_events(current
);
6790 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6791 && !preempt_count_equals(preempt_offset
)) {
6792 pr_err("Preemption disabled at:");
6793 print_ip_sym(preempt_disable_ip
);
6797 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6799 EXPORT_SYMBOL(___might_sleep
);
6801 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
6803 static unsigned long prev_jiffy
;
6805 if (irqs_disabled())
6808 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
6811 if (preempt_count() > preempt_offset
)
6814 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6816 prev_jiffy
= jiffies
;
6818 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
6819 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6820 in_atomic(), irqs_disabled(),
6821 current
->pid
, current
->comm
);
6823 debug_show_held_locks(current
);
6825 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6827 EXPORT_SYMBOL_GPL(__cant_sleep
);
6830 #ifdef CONFIG_MAGIC_SYSRQ
6831 void normalize_rt_tasks(void)
6833 struct task_struct
*g
, *p
;
6834 struct sched_attr attr
= {
6835 .sched_policy
= SCHED_NORMAL
,
6838 read_lock(&tasklist_lock
);
6839 for_each_process_thread(g
, p
) {
6841 * Only normalize user tasks:
6843 if (p
->flags
& PF_KTHREAD
)
6846 p
->se
.exec_start
= 0;
6847 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6848 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6849 schedstat_set(p
->se
.statistics
.block_start
, 0);
6851 if (!dl_task(p
) && !rt_task(p
)) {
6853 * Renice negative nice level userspace
6856 if (task_nice(p
) < 0)
6857 set_user_nice(p
, 0);
6861 __sched_setscheduler(p
, &attr
, false, false);
6863 read_unlock(&tasklist_lock
);
6866 #endif /* CONFIG_MAGIC_SYSRQ */
6868 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6870 * These functions are only useful for the IA64 MCA handling, or kdb.
6872 * They can only be called when the whole system has been
6873 * stopped - every CPU needs to be quiescent, and no scheduling
6874 * activity can take place. Using them for anything else would
6875 * be a serious bug, and as a result, they aren't even visible
6876 * under any other configuration.
6880 * curr_task - return the current task for a given CPU.
6881 * @cpu: the processor in question.
6883 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6885 * Return: The current task for @cpu.
6887 struct task_struct
*curr_task(int cpu
)
6889 return cpu_curr(cpu
);
6892 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6896 * ia64_set_curr_task - set the current task for a given CPU.
6897 * @cpu: the processor in question.
6898 * @p: the task pointer to set.
6900 * Description: This function must only be used when non-maskable interrupts
6901 * are serviced on a separate stack. It allows the architecture to switch the
6902 * notion of the current task on a CPU in a non-blocking manner. This function
6903 * must be called with all CPU's synchronized, and interrupts disabled, the
6904 * and caller must save the original value of the current task (see
6905 * curr_task() above) and restore that value before reenabling interrupts and
6906 * re-starting the system.
6908 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6910 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6917 #ifdef CONFIG_CGROUP_SCHED
6918 /* task_group_lock serializes the addition/removal of task groups */
6919 static DEFINE_SPINLOCK(task_group_lock
);
6921 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
6922 struct task_group
*parent
)
6924 #ifdef CONFIG_UCLAMP_TASK_GROUP
6925 enum uclamp_id clamp_id
;
6927 for_each_clamp_id(clamp_id
) {
6928 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
6929 uclamp_none(clamp_id
), false);
6930 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
6935 static void sched_free_group(struct task_group
*tg
)
6937 free_fair_sched_group(tg
);
6938 free_rt_sched_group(tg
);
6940 kmem_cache_free(task_group_cache
, tg
);
6943 /* allocate runqueue etc for a new task group */
6944 struct task_group
*sched_create_group(struct task_group
*parent
)
6946 struct task_group
*tg
;
6948 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6950 return ERR_PTR(-ENOMEM
);
6952 if (!alloc_fair_sched_group(tg
, parent
))
6955 if (!alloc_rt_sched_group(tg
, parent
))
6958 alloc_uclamp_sched_group(tg
, parent
);
6963 sched_free_group(tg
);
6964 return ERR_PTR(-ENOMEM
);
6967 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6969 unsigned long flags
;
6971 spin_lock_irqsave(&task_group_lock
, flags
);
6972 list_add_rcu(&tg
->list
, &task_groups
);
6974 /* Root should already exist: */
6977 tg
->parent
= parent
;
6978 INIT_LIST_HEAD(&tg
->children
);
6979 list_add_rcu(&tg
->siblings
, &parent
->children
);
6980 spin_unlock_irqrestore(&task_group_lock
, flags
);
6982 online_fair_sched_group(tg
);
6985 /* rcu callback to free various structures associated with a task group */
6986 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6988 /* Now it should be safe to free those cfs_rqs: */
6989 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6992 void sched_destroy_group(struct task_group
*tg
)
6994 /* Wait for possible concurrent references to cfs_rqs complete: */
6995 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6998 void sched_offline_group(struct task_group
*tg
)
7000 unsigned long flags
;
7002 /* End participation in shares distribution: */
7003 unregister_fair_sched_group(tg
);
7005 spin_lock_irqsave(&task_group_lock
, flags
);
7006 list_del_rcu(&tg
->list
);
7007 list_del_rcu(&tg
->siblings
);
7008 spin_unlock_irqrestore(&task_group_lock
, flags
);
7011 static void sched_change_group(struct task_struct
*tsk
, int type
)
7013 struct task_group
*tg
;
7016 * All callers are synchronized by task_rq_lock(); we do not use RCU
7017 * which is pointless here. Thus, we pass "true" to task_css_check()
7018 * to prevent lockdep warnings.
7020 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7021 struct task_group
, css
);
7022 tg
= autogroup_task_group(tsk
, tg
);
7023 tsk
->sched_task_group
= tg
;
7025 #ifdef CONFIG_FAIR_GROUP_SCHED
7026 if (tsk
->sched_class
->task_change_group
)
7027 tsk
->sched_class
->task_change_group(tsk
, type
);
7030 set_task_rq(tsk
, task_cpu(tsk
));
7034 * Change task's runqueue when it moves between groups.
7036 * The caller of this function should have put the task in its new group by
7037 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7040 void sched_move_task(struct task_struct
*tsk
)
7042 int queued
, running
, queue_flags
=
7043 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7047 rq
= task_rq_lock(tsk
, &rf
);
7048 update_rq_clock(rq
);
7050 running
= task_current(rq
, tsk
);
7051 queued
= task_on_rq_queued(tsk
);
7054 dequeue_task(rq
, tsk
, queue_flags
);
7056 put_prev_task(rq
, tsk
);
7058 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7061 enqueue_task(rq
, tsk
, queue_flags
);
7063 set_next_task(rq
, tsk
);
7065 * After changing group, the running task may have joined a
7066 * throttled one but it's still the running task. Trigger a
7067 * resched to make sure that task can still run.
7072 task_rq_unlock(rq
, tsk
, &rf
);
7075 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7077 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7080 static struct cgroup_subsys_state
*
7081 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7083 struct task_group
*parent
= css_tg(parent_css
);
7084 struct task_group
*tg
;
7087 /* This is early initialization for the top cgroup */
7088 return &root_task_group
.css
;
7091 tg
= sched_create_group(parent
);
7093 return ERR_PTR(-ENOMEM
);
7098 /* Expose task group only after completing cgroup initialization */
7099 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7101 struct task_group
*tg
= css_tg(css
);
7102 struct task_group
*parent
= css_tg(css
->parent
);
7105 sched_online_group(tg
, parent
);
7107 #ifdef CONFIG_UCLAMP_TASK_GROUP
7108 /* Propagate the effective uclamp value for the new group */
7109 cpu_util_update_eff(css
);
7115 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
7117 struct task_group
*tg
= css_tg(css
);
7119 sched_offline_group(tg
);
7122 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7124 struct task_group
*tg
= css_tg(css
);
7127 * Relies on the RCU grace period between css_released() and this.
7129 sched_free_group(tg
);
7133 * This is called before wake_up_new_task(), therefore we really only
7134 * have to set its group bits, all the other stuff does not apply.
7136 static void cpu_cgroup_fork(struct task_struct
*task
)
7141 rq
= task_rq_lock(task
, &rf
);
7143 update_rq_clock(rq
);
7144 sched_change_group(task
, TASK_SET_GROUP
);
7146 task_rq_unlock(rq
, task
, &rf
);
7149 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
7151 struct task_struct
*task
;
7152 struct cgroup_subsys_state
*css
;
7155 cgroup_taskset_for_each(task
, css
, tset
) {
7156 #ifdef CONFIG_RT_GROUP_SCHED
7157 if (!sched_rt_can_attach(css_tg(css
), task
))
7161 * Serialize against wake_up_new_task() such that if its
7162 * running, we're sure to observe its full state.
7164 raw_spin_lock_irq(&task
->pi_lock
);
7166 * Avoid calling sched_move_task() before wake_up_new_task()
7167 * has happened. This would lead to problems with PELT, due to
7168 * move wanting to detach+attach while we're not attached yet.
7170 if (task
->state
== TASK_NEW
)
7172 raw_spin_unlock_irq(&task
->pi_lock
);
7180 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
7182 struct task_struct
*task
;
7183 struct cgroup_subsys_state
*css
;
7185 cgroup_taskset_for_each(task
, css
, tset
)
7186 sched_move_task(task
);
7189 #ifdef CONFIG_UCLAMP_TASK_GROUP
7190 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
7192 struct cgroup_subsys_state
*top_css
= css
;
7193 struct uclamp_se
*uc_parent
= NULL
;
7194 struct uclamp_se
*uc_se
= NULL
;
7195 unsigned int eff
[UCLAMP_CNT
];
7196 enum uclamp_id clamp_id
;
7197 unsigned int clamps
;
7199 css_for_each_descendant_pre(css
, top_css
) {
7200 uc_parent
= css_tg(css
)->parent
7201 ? css_tg(css
)->parent
->uclamp
: NULL
;
7203 for_each_clamp_id(clamp_id
) {
7204 /* Assume effective clamps matches requested clamps */
7205 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
7206 /* Cap effective clamps with parent's effective clamps */
7208 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
7209 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
7212 /* Ensure protection is always capped by limit */
7213 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
7215 /* Propagate most restrictive effective clamps */
7217 uc_se
= css_tg(css
)->uclamp
;
7218 for_each_clamp_id(clamp_id
) {
7219 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
7221 uc_se
[clamp_id
].value
= eff
[clamp_id
];
7222 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
7223 clamps
|= (0x1 << clamp_id
);
7226 css
= css_rightmost_descendant(css
);
7230 /* Immediately update descendants RUNNABLE tasks */
7231 uclamp_update_active_tasks(css
, clamps
);
7236 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7237 * C expression. Since there is no way to convert a macro argument (N) into a
7238 * character constant, use two levels of macros.
7240 #define _POW10(exp) ((unsigned int)1e##exp)
7241 #define POW10(exp) _POW10(exp)
7243 struct uclamp_request
{
7244 #define UCLAMP_PERCENT_SHIFT 2
7245 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7251 static inline struct uclamp_request
7252 capacity_from_percent(char *buf
)
7254 struct uclamp_request req
= {
7255 .percent
= UCLAMP_PERCENT_SCALE
,
7256 .util
= SCHED_CAPACITY_SCALE
,
7261 if (strcmp(buf
, "max")) {
7262 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
7266 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
7271 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
7272 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
7278 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
7279 size_t nbytes
, loff_t off
,
7280 enum uclamp_id clamp_id
)
7282 struct uclamp_request req
;
7283 struct task_group
*tg
;
7285 req
= capacity_from_percent(buf
);
7289 mutex_lock(&uclamp_mutex
);
7292 tg
= css_tg(of_css(of
));
7293 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
7294 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
7297 * Because of not recoverable conversion rounding we keep track of the
7298 * exact requested value
7300 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
7302 /* Update effective clamps to track the most restrictive value */
7303 cpu_util_update_eff(of_css(of
));
7306 mutex_unlock(&uclamp_mutex
);
7311 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
7312 char *buf
, size_t nbytes
,
7315 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
7318 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
7319 char *buf
, size_t nbytes
,
7322 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
7325 static inline void cpu_uclamp_print(struct seq_file
*sf
,
7326 enum uclamp_id clamp_id
)
7328 struct task_group
*tg
;
7334 tg
= css_tg(seq_css(sf
));
7335 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
7338 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
7339 seq_puts(sf
, "max\n");
7343 percent
= tg
->uclamp_pct
[clamp_id
];
7344 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
7345 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
7348 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
7350 cpu_uclamp_print(sf
, UCLAMP_MIN
);
7354 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
7356 cpu_uclamp_print(sf
, UCLAMP_MAX
);
7359 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7361 #ifdef CONFIG_FAIR_GROUP_SCHED
7362 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7363 struct cftype
*cftype
, u64 shareval
)
7365 if (shareval
> scale_load_down(ULONG_MAX
))
7366 shareval
= MAX_SHARES
;
7367 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7370 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7373 struct task_group
*tg
= css_tg(css
);
7375 return (u64
) scale_load_down(tg
->shares
);
7378 #ifdef CONFIG_CFS_BANDWIDTH
7379 static DEFINE_MUTEX(cfs_constraints_mutex
);
7381 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7382 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7384 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7386 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7388 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7389 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7391 if (tg
== &root_task_group
)
7395 * Ensure we have at some amount of bandwidth every period. This is
7396 * to prevent reaching a state of large arrears when throttled via
7397 * entity_tick() resulting in prolonged exit starvation.
7399 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7403 * Likewise, bound things on the otherside by preventing insane quota
7404 * periods. This also allows us to normalize in computing quota
7407 if (period
> max_cfs_quota_period
)
7411 * Prevent race between setting of cfs_rq->runtime_enabled and
7412 * unthrottle_offline_cfs_rqs().
7415 mutex_lock(&cfs_constraints_mutex
);
7416 ret
= __cfs_schedulable(tg
, period
, quota
);
7420 runtime_enabled
= quota
!= RUNTIME_INF
;
7421 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7423 * If we need to toggle cfs_bandwidth_used, off->on must occur
7424 * before making related changes, and on->off must occur afterwards
7426 if (runtime_enabled
&& !runtime_was_enabled
)
7427 cfs_bandwidth_usage_inc();
7428 raw_spin_lock_irq(&cfs_b
->lock
);
7429 cfs_b
->period
= ns_to_ktime(period
);
7430 cfs_b
->quota
= quota
;
7432 __refill_cfs_bandwidth_runtime(cfs_b
);
7434 /* Restart the period timer (if active) to handle new period expiry: */
7435 if (runtime_enabled
)
7436 start_cfs_bandwidth(cfs_b
);
7438 raw_spin_unlock_irq(&cfs_b
->lock
);
7440 for_each_online_cpu(i
) {
7441 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7442 struct rq
*rq
= cfs_rq
->rq
;
7445 rq_lock_irq(rq
, &rf
);
7446 cfs_rq
->runtime_enabled
= runtime_enabled
;
7447 cfs_rq
->runtime_remaining
= 0;
7449 if (cfs_rq
->throttled
)
7450 unthrottle_cfs_rq(cfs_rq
);
7451 rq_unlock_irq(rq
, &rf
);
7453 if (runtime_was_enabled
&& !runtime_enabled
)
7454 cfs_bandwidth_usage_dec();
7456 mutex_unlock(&cfs_constraints_mutex
);
7462 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7466 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7467 if (cfs_quota_us
< 0)
7468 quota
= RUNTIME_INF
;
7469 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
7470 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7474 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7477 static long tg_get_cfs_quota(struct task_group
*tg
)
7481 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7484 quota_us
= tg
->cfs_bandwidth
.quota
;
7485 do_div(quota_us
, NSEC_PER_USEC
);
7490 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7494 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
7497 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7498 quota
= tg
->cfs_bandwidth
.quota
;
7500 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7503 static long tg_get_cfs_period(struct task_group
*tg
)
7507 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7508 do_div(cfs_period_us
, NSEC_PER_USEC
);
7510 return cfs_period_us
;
7513 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7516 return tg_get_cfs_quota(css_tg(css
));
7519 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7520 struct cftype
*cftype
, s64 cfs_quota_us
)
7522 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7525 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7528 return tg_get_cfs_period(css_tg(css
));
7531 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7532 struct cftype
*cftype
, u64 cfs_period_us
)
7534 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7537 struct cfs_schedulable_data
{
7538 struct task_group
*tg
;
7543 * normalize group quota/period to be quota/max_period
7544 * note: units are usecs
7546 static u64
normalize_cfs_quota(struct task_group
*tg
,
7547 struct cfs_schedulable_data
*d
)
7555 period
= tg_get_cfs_period(tg
);
7556 quota
= tg_get_cfs_quota(tg
);
7559 /* note: these should typically be equivalent */
7560 if (quota
== RUNTIME_INF
|| quota
== -1)
7563 return to_ratio(period
, quota
);
7566 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7568 struct cfs_schedulable_data
*d
= data
;
7569 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7570 s64 quota
= 0, parent_quota
= -1;
7573 quota
= RUNTIME_INF
;
7575 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7577 quota
= normalize_cfs_quota(tg
, d
);
7578 parent_quota
= parent_b
->hierarchical_quota
;
7581 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7582 * always take the min. On cgroup1, only inherit when no
7585 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
7586 quota
= min(quota
, parent_quota
);
7588 if (quota
== RUNTIME_INF
)
7589 quota
= parent_quota
;
7590 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7594 cfs_b
->hierarchical_quota
= quota
;
7599 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7602 struct cfs_schedulable_data data
= {
7608 if (quota
!= RUNTIME_INF
) {
7609 do_div(data
.period
, NSEC_PER_USEC
);
7610 do_div(data
.quota
, NSEC_PER_USEC
);
7614 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7620 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
7622 struct task_group
*tg
= css_tg(seq_css(sf
));
7623 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7625 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7626 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7627 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7629 if (schedstat_enabled() && tg
!= &root_task_group
) {
7633 for_each_possible_cpu(i
)
7634 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
7636 seq_printf(sf
, "wait_sum %llu\n", ws
);
7641 #endif /* CONFIG_CFS_BANDWIDTH */
7642 #endif /* CONFIG_FAIR_GROUP_SCHED */
7644 #ifdef CONFIG_RT_GROUP_SCHED
7645 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7646 struct cftype
*cft
, s64 val
)
7648 return sched_group_set_rt_runtime(css_tg(css
), val
);
7651 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7654 return sched_group_rt_runtime(css_tg(css
));
7657 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7658 struct cftype
*cftype
, u64 rt_period_us
)
7660 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7663 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7666 return sched_group_rt_period(css_tg(css
));
7668 #endif /* CONFIG_RT_GROUP_SCHED */
7670 static struct cftype cpu_legacy_files
[] = {
7671 #ifdef CONFIG_FAIR_GROUP_SCHED
7674 .read_u64
= cpu_shares_read_u64
,
7675 .write_u64
= cpu_shares_write_u64
,
7678 #ifdef CONFIG_CFS_BANDWIDTH
7680 .name
= "cfs_quota_us",
7681 .read_s64
= cpu_cfs_quota_read_s64
,
7682 .write_s64
= cpu_cfs_quota_write_s64
,
7685 .name
= "cfs_period_us",
7686 .read_u64
= cpu_cfs_period_read_u64
,
7687 .write_u64
= cpu_cfs_period_write_u64
,
7691 .seq_show
= cpu_cfs_stat_show
,
7694 #ifdef CONFIG_RT_GROUP_SCHED
7696 .name
= "rt_runtime_us",
7697 .read_s64
= cpu_rt_runtime_read
,
7698 .write_s64
= cpu_rt_runtime_write
,
7701 .name
= "rt_period_us",
7702 .read_u64
= cpu_rt_period_read_uint
,
7703 .write_u64
= cpu_rt_period_write_uint
,
7706 #ifdef CONFIG_UCLAMP_TASK_GROUP
7708 .name
= "uclamp.min",
7709 .flags
= CFTYPE_NOT_ON_ROOT
,
7710 .seq_show
= cpu_uclamp_min_show
,
7711 .write
= cpu_uclamp_min_write
,
7714 .name
= "uclamp.max",
7715 .flags
= CFTYPE_NOT_ON_ROOT
,
7716 .seq_show
= cpu_uclamp_max_show
,
7717 .write
= cpu_uclamp_max_write
,
7723 static int cpu_extra_stat_show(struct seq_file
*sf
,
7724 struct cgroup_subsys_state
*css
)
7726 #ifdef CONFIG_CFS_BANDWIDTH
7728 struct task_group
*tg
= css_tg(css
);
7729 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7732 throttled_usec
= cfs_b
->throttled_time
;
7733 do_div(throttled_usec
, NSEC_PER_USEC
);
7735 seq_printf(sf
, "nr_periods %d\n"
7737 "throttled_usec %llu\n",
7738 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
7745 #ifdef CONFIG_FAIR_GROUP_SCHED
7746 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
7749 struct task_group
*tg
= css_tg(css
);
7750 u64 weight
= scale_load_down(tg
->shares
);
7752 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
7755 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
7756 struct cftype
*cft
, u64 weight
)
7759 * cgroup weight knobs should use the common MIN, DFL and MAX
7760 * values which are 1, 100 and 10000 respectively. While it loses
7761 * a bit of range on both ends, it maps pretty well onto the shares
7762 * value used by scheduler and the round-trip conversions preserve
7763 * the original value over the entire range.
7765 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
7768 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
7770 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7773 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
7776 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
7777 int last_delta
= INT_MAX
;
7780 /* find the closest nice value to the current weight */
7781 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
7782 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
7783 if (delta
>= last_delta
)
7788 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
7791 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
7792 struct cftype
*cft
, s64 nice
)
7794 unsigned long weight
;
7797 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
7800 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
7801 idx
= array_index_nospec(idx
, 40);
7802 weight
= sched_prio_to_weight
[idx
];
7804 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7808 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
7809 long period
, long quota
)
7812 seq_puts(sf
, "max");
7814 seq_printf(sf
, "%ld", quota
);
7816 seq_printf(sf
, " %ld\n", period
);
7819 /* caller should put the current value in *@periodp before calling */
7820 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
7821 u64
*periodp
, u64
*quotap
)
7823 char tok
[21]; /* U64_MAX */
7825 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
7828 *periodp
*= NSEC_PER_USEC
;
7830 if (sscanf(tok
, "%llu", quotap
))
7831 *quotap
*= NSEC_PER_USEC
;
7832 else if (!strcmp(tok
, "max"))
7833 *quotap
= RUNTIME_INF
;
7840 #ifdef CONFIG_CFS_BANDWIDTH
7841 static int cpu_max_show(struct seq_file
*sf
, void *v
)
7843 struct task_group
*tg
= css_tg(seq_css(sf
));
7845 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
7849 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
7850 char *buf
, size_t nbytes
, loff_t off
)
7852 struct task_group
*tg
= css_tg(of_css(of
));
7853 u64 period
= tg_get_cfs_period(tg
);
7857 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
7859 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
7860 return ret
?: nbytes
;
7864 static struct cftype cpu_files
[] = {
7865 #ifdef CONFIG_FAIR_GROUP_SCHED
7868 .flags
= CFTYPE_NOT_ON_ROOT
,
7869 .read_u64
= cpu_weight_read_u64
,
7870 .write_u64
= cpu_weight_write_u64
,
7873 .name
= "weight.nice",
7874 .flags
= CFTYPE_NOT_ON_ROOT
,
7875 .read_s64
= cpu_weight_nice_read_s64
,
7876 .write_s64
= cpu_weight_nice_write_s64
,
7879 #ifdef CONFIG_CFS_BANDWIDTH
7882 .flags
= CFTYPE_NOT_ON_ROOT
,
7883 .seq_show
= cpu_max_show
,
7884 .write
= cpu_max_write
,
7887 #ifdef CONFIG_UCLAMP_TASK_GROUP
7889 .name
= "uclamp.min",
7890 .flags
= CFTYPE_NOT_ON_ROOT
,
7891 .seq_show
= cpu_uclamp_min_show
,
7892 .write
= cpu_uclamp_min_write
,
7895 .name
= "uclamp.max",
7896 .flags
= CFTYPE_NOT_ON_ROOT
,
7897 .seq_show
= cpu_uclamp_max_show
,
7898 .write
= cpu_uclamp_max_write
,
7904 struct cgroup_subsys cpu_cgrp_subsys
= {
7905 .css_alloc
= cpu_cgroup_css_alloc
,
7906 .css_online
= cpu_cgroup_css_online
,
7907 .css_released
= cpu_cgroup_css_released
,
7908 .css_free
= cpu_cgroup_css_free
,
7909 .css_extra_stat_show
= cpu_extra_stat_show
,
7910 .fork
= cpu_cgroup_fork
,
7911 .can_attach
= cpu_cgroup_can_attach
,
7912 .attach
= cpu_cgroup_attach
,
7913 .legacy_cftypes
= cpu_legacy_files
,
7914 .dfl_cftypes
= cpu_files
,
7919 #endif /* CONFIG_CGROUP_SCHED */
7921 void dump_cpu_task(int cpu
)
7923 pr_info("Task dump for CPU %d:\n", cpu
);
7924 sched_show_task(cpu_curr(cpu
));
7928 * Nice levels are multiplicative, with a gentle 10% change for every
7929 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7930 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7931 * that remained on nice 0.
7933 * The "10% effect" is relative and cumulative: from _any_ nice level,
7934 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7935 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7936 * If a task goes up by ~10% and another task goes down by ~10% then
7937 * the relative distance between them is ~25%.)
7939 const int sched_prio_to_weight
[40] = {
7940 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7941 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7942 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7943 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7944 /* 0 */ 1024, 820, 655, 526, 423,
7945 /* 5 */ 335, 272, 215, 172, 137,
7946 /* 10 */ 110, 87, 70, 56, 45,
7947 /* 15 */ 36, 29, 23, 18, 15,
7951 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7953 * In cases where the weight does not change often, we can use the
7954 * precalculated inverse to speed up arithmetics by turning divisions
7955 * into multiplications:
7957 const u32 sched_prio_to_wmult
[40] = {
7958 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7959 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7960 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7961 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7962 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7963 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7964 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7965 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7968 #undef CREATE_TRACE_POINTS