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 "../../fs/io-wq.h"
20 #include "../smpboot.h"
24 #define CREATE_TRACE_POINTS
25 #include <trace/events/sched.h>
28 * Export tracepoints that act as a bare tracehook (ie: have no trace event
29 * associated with them) to allow external modules to probe them.
31 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
32 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
38 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
40 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
42 * Debugging: various feature bits
44 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
45 * sysctl_sched_features, defined in sched.h, to allow constants propagation
46 * at compile time and compiler optimization based on features default.
48 #define SCHED_FEAT(name, enabled) \
49 (1UL << __SCHED_FEAT_##name) * enabled |
50 const_debug
unsigned int sysctl_sched_features
=
57 * Number of tasks to iterate in a single balance run.
58 * Limited because this is done with IRQs disabled.
60 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
63 * period over which we measure -rt task CPU usage in us.
66 unsigned int sysctl_sched_rt_period
= 1000000;
68 __read_mostly
int scheduler_running
;
71 * part of the period that we allow rt tasks to run in us.
74 int sysctl_sched_rt_runtime
= 950000;
77 * __task_rq_lock - lock the rq @p resides on.
79 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
84 lockdep_assert_held(&p
->pi_lock
);
88 raw_spin_lock(&rq
->lock
);
89 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
93 raw_spin_unlock(&rq
->lock
);
95 while (unlikely(task_on_rq_migrating(p
)))
101 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
103 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
104 __acquires(p
->pi_lock
)
110 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
112 raw_spin_lock(&rq
->lock
);
114 * move_queued_task() task_rq_lock()
117 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
118 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
119 * [S] ->cpu = new_cpu [L] task_rq()
123 * If we observe the old CPU in task_rq_lock(), the acquire of
124 * the old rq->lock will fully serialize against the stores.
126 * If we observe the new CPU in task_rq_lock(), the address
127 * dependency headed by '[L] rq = task_rq()' and the acquire
128 * will pair with the WMB to ensure we then also see migrating.
130 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
134 raw_spin_unlock(&rq
->lock
);
135 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
137 while (unlikely(task_on_rq_migrating(p
)))
143 * RQ-clock updating methods:
146 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
149 * In theory, the compile should just see 0 here, and optimize out the call
150 * to sched_rt_avg_update. But I don't trust it...
152 s64 __maybe_unused steal
= 0, irq_delta
= 0;
154 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
155 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
158 * Since irq_time is only updated on {soft,}irq_exit, we might run into
159 * this case when a previous update_rq_clock() happened inside a
162 * When this happens, we stop ->clock_task and only update the
163 * prev_irq_time stamp to account for the part that fit, so that a next
164 * update will consume the rest. This ensures ->clock_task is
167 * It does however cause some slight miss-attribution of {soft,}irq
168 * time, a more accurate solution would be to update the irq_time using
169 * the current rq->clock timestamp, except that would require using
172 if (irq_delta
> delta
)
175 rq
->prev_irq_time
+= irq_delta
;
178 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
179 if (static_key_false((¶virt_steal_rq_enabled
))) {
180 steal
= paravirt_steal_clock(cpu_of(rq
));
181 steal
-= rq
->prev_steal_time_rq
;
183 if (unlikely(steal
> delta
))
186 rq
->prev_steal_time_rq
+= steal
;
191 rq
->clock_task
+= delta
;
193 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
194 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
195 update_irq_load_avg(rq
, irq_delta
+ steal
);
197 update_rq_clock_pelt(rq
, delta
);
200 void update_rq_clock(struct rq
*rq
)
204 lockdep_assert_held(&rq
->lock
);
206 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
209 #ifdef CONFIG_SCHED_DEBUG
210 if (sched_feat(WARN_DOUBLE_CLOCK
))
211 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
212 rq
->clock_update_flags
|= RQCF_UPDATED
;
215 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
219 update_rq_clock_task(rq
, delta
);
223 rq_csd_init(struct rq
*rq
, call_single_data_t
*csd
, smp_call_func_t func
)
230 #ifdef CONFIG_SCHED_HRTICK
232 * Use HR-timers to deliver accurate preemption points.
235 static void hrtick_clear(struct rq
*rq
)
237 if (hrtimer_active(&rq
->hrtick_timer
))
238 hrtimer_cancel(&rq
->hrtick_timer
);
242 * High-resolution timer tick.
243 * Runs from hardirq context with interrupts disabled.
245 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
247 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
250 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
254 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
257 return HRTIMER_NORESTART
;
262 static void __hrtick_restart(struct rq
*rq
)
264 struct hrtimer
*timer
= &rq
->hrtick_timer
;
266 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
270 * called from hardirq (IPI) context
272 static void __hrtick_start(void *arg
)
278 __hrtick_restart(rq
);
283 * Called to set the hrtick timer state.
285 * called with rq->lock held and irqs disabled
287 void hrtick_start(struct rq
*rq
, u64 delay
)
289 struct hrtimer
*timer
= &rq
->hrtick_timer
;
294 * Don't schedule slices shorter than 10000ns, that just
295 * doesn't make sense and can cause timer DoS.
297 delta
= max_t(s64
, delay
, 10000LL);
298 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
300 hrtimer_set_expires(timer
, time
);
303 __hrtick_restart(rq
);
305 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
310 * Called to set the hrtick timer state.
312 * called with rq->lock held and irqs disabled
314 void hrtick_start(struct rq
*rq
, u64 delay
)
317 * Don't schedule slices shorter than 10000ns, that just
318 * doesn't make sense. Rely on vruntime for fairness.
320 delay
= max_t(u64
, delay
, 10000LL);
321 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
322 HRTIMER_MODE_REL_PINNED_HARD
);
325 #endif /* CONFIG_SMP */
327 static void hrtick_rq_init(struct rq
*rq
)
330 rq_csd_init(rq
, &rq
->hrtick_csd
, __hrtick_start
);
332 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
333 rq
->hrtick_timer
.function
= hrtick
;
335 #else /* CONFIG_SCHED_HRTICK */
336 static inline void hrtick_clear(struct rq
*rq
)
340 static inline void hrtick_rq_init(struct rq
*rq
)
343 #endif /* CONFIG_SCHED_HRTICK */
346 * cmpxchg based fetch_or, macro so it works for different integer types
348 #define fetch_or(ptr, mask) \
350 typeof(ptr) _ptr = (ptr); \
351 typeof(mask) _mask = (mask); \
352 typeof(*_ptr) _old, _val = *_ptr; \
355 _old = cmpxchg(_ptr, _val, _val | _mask); \
363 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
365 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
366 * this avoids any races wrt polling state changes and thereby avoids
369 static bool set_nr_and_not_polling(struct task_struct
*p
)
371 struct thread_info
*ti
= task_thread_info(p
);
372 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
376 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
378 * If this returns true, then the idle task promises to call
379 * sched_ttwu_pending() and reschedule soon.
381 static bool set_nr_if_polling(struct task_struct
*p
)
383 struct thread_info
*ti
= task_thread_info(p
);
384 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
387 if (!(val
& _TIF_POLLING_NRFLAG
))
389 if (val
& _TIF_NEED_RESCHED
)
391 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
400 static bool set_nr_and_not_polling(struct task_struct
*p
)
402 set_tsk_need_resched(p
);
407 static bool set_nr_if_polling(struct task_struct
*p
)
414 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
416 struct wake_q_node
*node
= &task
->wake_q
;
419 * Atomically grab the task, if ->wake_q is !nil already it means
420 * its already queued (either by us or someone else) and will get the
421 * wakeup due to that.
423 * In order to ensure that a pending wakeup will observe our pending
424 * state, even in the failed case, an explicit smp_mb() must be used.
426 smp_mb__before_atomic();
427 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
431 * The head is context local, there can be no concurrency.
434 head
->lastp
= &node
->next
;
439 * wake_q_add() - queue a wakeup for 'later' waking.
440 * @head: the wake_q_head to add @task to
441 * @task: the task to queue for 'later' wakeup
443 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
444 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
447 * This function must be used as-if it were wake_up_process(); IOW the task
448 * must be ready to be woken at this location.
450 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
452 if (__wake_q_add(head
, task
))
453 get_task_struct(task
);
457 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
458 * @head: the wake_q_head to add @task to
459 * @task: the task to queue for 'later' wakeup
461 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
462 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
465 * This function must be used as-if it were wake_up_process(); IOW the task
466 * must be ready to be woken at this location.
468 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
469 * that already hold reference to @task can call the 'safe' version and trust
470 * wake_q to do the right thing depending whether or not the @task is already
473 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
475 if (!__wake_q_add(head
, task
))
476 put_task_struct(task
);
479 void wake_up_q(struct wake_q_head
*head
)
481 struct wake_q_node
*node
= head
->first
;
483 while (node
!= WAKE_Q_TAIL
) {
484 struct task_struct
*task
;
486 task
= container_of(node
, struct task_struct
, wake_q
);
488 /* Task can safely be re-inserted now: */
490 task
->wake_q
.next
= NULL
;
493 * wake_up_process() executes a full barrier, which pairs with
494 * the queueing in wake_q_add() so as not to miss wakeups.
496 wake_up_process(task
);
497 put_task_struct(task
);
502 * resched_curr - mark rq's current task 'to be rescheduled now'.
504 * On UP this means the setting of the need_resched flag, on SMP it
505 * might also involve a cross-CPU call to trigger the scheduler on
508 void resched_curr(struct rq
*rq
)
510 struct task_struct
*curr
= rq
->curr
;
513 lockdep_assert_held(&rq
->lock
);
515 if (test_tsk_need_resched(curr
))
520 if (cpu
== smp_processor_id()) {
521 set_tsk_need_resched(curr
);
522 set_preempt_need_resched();
526 if (set_nr_and_not_polling(curr
))
527 smp_send_reschedule(cpu
);
529 trace_sched_wake_idle_without_ipi(cpu
);
532 void resched_cpu(int cpu
)
534 struct rq
*rq
= cpu_rq(cpu
);
537 raw_spin_lock_irqsave(&rq
->lock
, flags
);
538 if (cpu_online(cpu
) || cpu
== smp_processor_id())
540 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
544 #ifdef CONFIG_NO_HZ_COMMON
546 * In the semi idle case, use the nearest busy CPU for migrating timers
547 * from an idle CPU. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle CPU will add more delays to the timers than intended
551 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
556 struct sched_domain
*sd
;
558 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
565 for_each_domain(cpu
, sd
) {
566 for_each_cpu_and(i
, sched_domain_span(sd
),
567 housekeeping_cpumask(HK_FLAG_TIMER
)) {
578 if (default_cpu
== -1)
579 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
587 * When add_timer_on() enqueues a timer into the timer wheel of an
588 * idle CPU then this timer might expire before the next timer event
589 * which is scheduled to wake up that CPU. In case of a completely
590 * idle system the next event might even be infinite time into the
591 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
592 * leaves the inner idle loop so the newly added timer is taken into
593 * account when the CPU goes back to idle and evaluates the timer
594 * wheel for the next timer event.
596 static void wake_up_idle_cpu(int cpu
)
598 struct rq
*rq
= cpu_rq(cpu
);
600 if (cpu
== smp_processor_id())
603 if (set_nr_and_not_polling(rq
->idle
))
604 smp_send_reschedule(cpu
);
606 trace_sched_wake_idle_without_ipi(cpu
);
609 static bool wake_up_full_nohz_cpu(int cpu
)
612 * We just need the target to call irq_exit() and re-evaluate
613 * the next tick. The nohz full kick at least implies that.
614 * If needed we can still optimize that later with an
617 if (cpu_is_offline(cpu
))
618 return true; /* Don't try to wake offline CPUs. */
619 if (tick_nohz_full_cpu(cpu
)) {
620 if (cpu
!= smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 tick_nohz_full_kick_cpu(cpu
);
630 * Wake up the specified CPU. If the CPU is going offline, it is the
631 * caller's responsibility to deal with the lost wakeup, for example,
632 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
634 void wake_up_nohz_cpu(int cpu
)
636 if (!wake_up_full_nohz_cpu(cpu
))
637 wake_up_idle_cpu(cpu
);
640 static inline bool got_nohz_idle_kick(void)
642 int cpu
= smp_processor_id();
644 if (!(atomic_read(nohz_flags(cpu
)) & NOHZ_KICK_MASK
))
647 if (idle_cpu(cpu
) && !need_resched())
651 * We can't run Idle Load Balance on this CPU for this time so we
652 * cancel it and clear NOHZ_BALANCE_KICK
654 atomic_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
658 static void nohz_csd_func(void *info
)
660 struct rq
*rq
= info
;
662 if (got_nohz_idle_kick()) {
663 rq
->idle_balance
= 1;
664 raise_softirq_irqoff(SCHED_SOFTIRQ
);
668 #else /* CONFIG_NO_HZ_COMMON */
670 static inline bool got_nohz_idle_kick(void)
675 #endif /* CONFIG_NO_HZ_COMMON */
677 #ifdef CONFIG_NO_HZ_FULL
678 bool sched_can_stop_tick(struct rq
*rq
)
682 /* Deadline tasks, even if single, need the tick */
683 if (rq
->dl
.dl_nr_running
)
687 * If there are more than one RR tasks, we need the tick to effect the
688 * actual RR behaviour.
690 if (rq
->rt
.rr_nr_running
) {
691 if (rq
->rt
.rr_nr_running
== 1)
698 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
699 * forced preemption between FIFO tasks.
701 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
706 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
707 * if there's more than one we need the tick for involuntary
710 if (rq
->nr_running
> 1)
715 #endif /* CONFIG_NO_HZ_FULL */
716 #endif /* CONFIG_SMP */
718 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
719 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
721 * Iterate task_group tree rooted at *from, calling @down when first entering a
722 * node and @up when leaving it for the final time.
724 * Caller must hold rcu_lock or sufficient equivalent.
726 int walk_tg_tree_from(struct task_group
*from
,
727 tg_visitor down
, tg_visitor up
, void *data
)
729 struct task_group
*parent
, *child
;
735 ret
= (*down
)(parent
, data
);
738 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
745 ret
= (*up
)(parent
, data
);
746 if (ret
|| parent
== from
)
750 parent
= parent
->parent
;
757 int tg_nop(struct task_group
*tg
, void *data
)
763 static void set_load_weight(struct task_struct
*p
, bool update_load
)
765 int prio
= p
->static_prio
- MAX_RT_PRIO
;
766 struct load_weight
*load
= &p
->se
.load
;
769 * SCHED_IDLE tasks get minimal weight:
771 if (task_has_idle_policy(p
)) {
772 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
773 load
->inv_weight
= WMULT_IDLEPRIO
;
778 * SCHED_OTHER tasks have to update their load when changing their
781 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
782 reweight_task(p
, prio
);
784 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
785 load
->inv_weight
= sched_prio_to_wmult
[prio
];
789 #ifdef CONFIG_UCLAMP_TASK
791 * Serializes updates of utilization clamp values
793 * The (slow-path) user-space triggers utilization clamp value updates which
794 * can require updates on (fast-path) scheduler's data structures used to
795 * support enqueue/dequeue operations.
796 * While the per-CPU rq lock protects fast-path update operations, user-space
797 * requests are serialized using a mutex to reduce the risk of conflicting
798 * updates or API abuses.
800 static DEFINE_MUTEX(uclamp_mutex
);
802 /* Max allowed minimum utilization */
803 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
805 /* Max allowed maximum utilization */
806 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
808 /* All clamps are required to be less or equal than these values */
809 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
811 /* Integer rounded range for each bucket */
812 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
814 #define for_each_clamp_id(clamp_id) \
815 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
817 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
819 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
822 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value
)
824 return UCLAMP_BUCKET_DELTA
* uclamp_bucket_id(clamp_value
);
827 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
829 if (clamp_id
== UCLAMP_MIN
)
831 return SCHED_CAPACITY_SCALE
;
834 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
835 unsigned int value
, bool user_defined
)
837 uc_se
->value
= value
;
838 uc_se
->bucket_id
= uclamp_bucket_id(value
);
839 uc_se
->user_defined
= user_defined
;
842 static inline unsigned int
843 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
844 unsigned int clamp_value
)
847 * Avoid blocked utilization pushing up the frequency when we go
848 * idle (which drops the max-clamp) by retaining the last known
851 if (clamp_id
== UCLAMP_MAX
) {
852 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
856 return uclamp_none(UCLAMP_MIN
);
859 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
860 unsigned int clamp_value
)
862 /* Reset max-clamp retention only on idle exit */
863 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
866 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
870 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
871 unsigned int clamp_value
)
873 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
874 int bucket_id
= UCLAMP_BUCKETS
- 1;
877 * Since both min and max clamps are max aggregated, find the
878 * top most bucket with tasks in.
880 for ( ; bucket_id
>= 0; bucket_id
--) {
881 if (!bucket
[bucket_id
].tasks
)
883 return bucket
[bucket_id
].value
;
886 /* No tasks -- default clamp values */
887 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
890 static inline struct uclamp_se
891 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
893 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
894 #ifdef CONFIG_UCLAMP_TASK_GROUP
895 struct uclamp_se uc_max
;
898 * Tasks in autogroups or root task group will be
899 * restricted by system defaults.
901 if (task_group_is_autogroup(task_group(p
)))
903 if (task_group(p
) == &root_task_group
)
906 uc_max
= task_group(p
)->uclamp
[clamp_id
];
907 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
915 * The effective clamp bucket index of a task depends on, by increasing
917 * - the task specific clamp value, when explicitly requested from userspace
918 * - the task group effective clamp value, for tasks not either in the root
919 * group or in an autogroup
920 * - the system default clamp value, defined by the sysadmin
922 static inline struct uclamp_se
923 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
925 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
926 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
928 /* System default restrictions always apply */
929 if (unlikely(uc_req
.value
> uc_max
.value
))
935 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
937 struct uclamp_se uc_eff
;
939 /* Task currently refcounted: use back-annotated (effective) value */
940 if (p
->uclamp
[clamp_id
].active
)
941 return (unsigned long)p
->uclamp
[clamp_id
].value
;
943 uc_eff
= uclamp_eff_get(p
, clamp_id
);
945 return (unsigned long)uc_eff
.value
;
949 * When a task is enqueued on a rq, the clamp bucket currently defined by the
950 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
951 * updates the rq's clamp value if required.
953 * Tasks can have a task-specific value requested from user-space, track
954 * within each bucket the maximum value for tasks refcounted in it.
955 * This "local max aggregation" allows to track the exact "requested" value
956 * for each bucket when all its RUNNABLE tasks require the same clamp.
958 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
959 enum uclamp_id clamp_id
)
961 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
962 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
963 struct uclamp_bucket
*bucket
;
965 lockdep_assert_held(&rq
->lock
);
967 /* Update task effective clamp */
968 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
970 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
972 uc_se
->active
= true;
974 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
977 * Local max aggregation: rq buckets always track the max
978 * "requested" clamp value of its RUNNABLE tasks.
980 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
981 bucket
->value
= uc_se
->value
;
983 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
984 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
988 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
989 * is released. If this is the last task reference counting the rq's max
990 * active clamp value, then the rq's clamp value is updated.
992 * Both refcounted tasks and rq's cached clamp values are expected to be
993 * always valid. If it's detected they are not, as defensive programming,
994 * enforce the expected state and warn.
996 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
997 enum uclamp_id clamp_id
)
999 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1000 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1001 struct uclamp_bucket
*bucket
;
1002 unsigned int bkt_clamp
;
1003 unsigned int rq_clamp
;
1005 lockdep_assert_held(&rq
->lock
);
1007 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1008 SCHED_WARN_ON(!bucket
->tasks
);
1009 if (likely(bucket
->tasks
))
1011 uc_se
->active
= false;
1014 * Keep "local max aggregation" simple and accept to (possibly)
1015 * overboost some RUNNABLE tasks in the same bucket.
1016 * The rq clamp bucket value is reset to its base value whenever
1017 * there are no more RUNNABLE tasks refcounting it.
1019 if (likely(bucket
->tasks
))
1022 rq_clamp
= READ_ONCE(uc_rq
->value
);
1024 * Defensive programming: this should never happen. If it happens,
1025 * e.g. due to future modification, warn and fixup the expected value.
1027 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1028 if (bucket
->value
>= rq_clamp
) {
1029 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1030 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1034 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1036 enum uclamp_id clamp_id
;
1038 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1041 for_each_clamp_id(clamp_id
)
1042 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1044 /* Reset clamp idle holding when there is one RUNNABLE task */
1045 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1046 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1049 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1051 enum uclamp_id clamp_id
;
1053 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1056 for_each_clamp_id(clamp_id
)
1057 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1061 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1067 * Lock the task and the rq where the task is (or was) queued.
1069 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1070 * price to pay to safely serialize util_{min,max} updates with
1071 * enqueues, dequeues and migration operations.
1072 * This is the same locking schema used by __set_cpus_allowed_ptr().
1074 rq
= task_rq_lock(p
, &rf
);
1077 * Setting the clamp bucket is serialized by task_rq_lock().
1078 * If the task is not yet RUNNABLE and its task_struct is not
1079 * affecting a valid clamp bucket, the next time it's enqueued,
1080 * it will already see the updated clamp bucket value.
1082 if (p
->uclamp
[clamp_id
].active
) {
1083 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1084 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1087 task_rq_unlock(rq
, p
, &rf
);
1090 #ifdef CONFIG_UCLAMP_TASK_GROUP
1092 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1093 unsigned int clamps
)
1095 enum uclamp_id clamp_id
;
1096 struct css_task_iter it
;
1097 struct task_struct
*p
;
1099 css_task_iter_start(css
, 0, &it
);
1100 while ((p
= css_task_iter_next(&it
))) {
1101 for_each_clamp_id(clamp_id
) {
1102 if ((0x1 << clamp_id
) & clamps
)
1103 uclamp_update_active(p
, clamp_id
);
1106 css_task_iter_end(&it
);
1109 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1110 static void uclamp_update_root_tg(void)
1112 struct task_group
*tg
= &root_task_group
;
1114 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1115 sysctl_sched_uclamp_util_min
, false);
1116 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1117 sysctl_sched_uclamp_util_max
, false);
1120 cpu_util_update_eff(&root_task_group
.css
);
1124 static void uclamp_update_root_tg(void) { }
1127 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1128 void __user
*buffer
, size_t *lenp
,
1131 bool update_root_tg
= false;
1132 int old_min
, old_max
;
1135 mutex_lock(&uclamp_mutex
);
1136 old_min
= sysctl_sched_uclamp_util_min
;
1137 old_max
= sysctl_sched_uclamp_util_max
;
1139 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1145 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1146 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
) {
1151 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1152 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1153 sysctl_sched_uclamp_util_min
, false);
1154 update_root_tg
= true;
1156 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1157 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1158 sysctl_sched_uclamp_util_max
, false);
1159 update_root_tg
= true;
1163 uclamp_update_root_tg();
1166 * We update all RUNNABLE tasks only when task groups are in use.
1167 * Otherwise, keep it simple and do just a lazy update at each next
1168 * task enqueue time.
1174 sysctl_sched_uclamp_util_min
= old_min
;
1175 sysctl_sched_uclamp_util_max
= old_max
;
1177 mutex_unlock(&uclamp_mutex
);
1182 static int uclamp_validate(struct task_struct
*p
,
1183 const struct sched_attr
*attr
)
1185 unsigned int lower_bound
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1186 unsigned int upper_bound
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1188 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
)
1189 lower_bound
= attr
->sched_util_min
;
1190 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
)
1191 upper_bound
= attr
->sched_util_max
;
1193 if (lower_bound
> upper_bound
)
1195 if (upper_bound
> SCHED_CAPACITY_SCALE
)
1201 static void __setscheduler_uclamp(struct task_struct
*p
,
1202 const struct sched_attr
*attr
)
1204 enum uclamp_id clamp_id
;
1207 * On scheduling class change, reset to default clamps for tasks
1208 * without a task-specific value.
1210 for_each_clamp_id(clamp_id
) {
1211 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1212 unsigned int clamp_value
= uclamp_none(clamp_id
);
1214 /* Keep using defined clamps across class changes */
1215 if (uc_se
->user_defined
)
1218 /* By default, RT tasks always get 100% boost */
1219 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1220 clamp_value
= uclamp_none(UCLAMP_MAX
);
1222 uclamp_se_set(uc_se
, clamp_value
, false);
1225 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1228 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1229 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1230 attr
->sched_util_min
, true);
1233 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1234 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1235 attr
->sched_util_max
, true);
1239 static void uclamp_fork(struct task_struct
*p
)
1241 enum uclamp_id clamp_id
;
1243 for_each_clamp_id(clamp_id
)
1244 p
->uclamp
[clamp_id
].active
= false;
1246 if (likely(!p
->sched_reset_on_fork
))
1249 for_each_clamp_id(clamp_id
) {
1250 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1251 uclamp_none(clamp_id
), false);
1255 static void __init
init_uclamp(void)
1257 struct uclamp_se uc_max
= {};
1258 enum uclamp_id clamp_id
;
1261 mutex_init(&uclamp_mutex
);
1263 for_each_possible_cpu(cpu
) {
1264 memset(&cpu_rq(cpu
)->uclamp
, 0,
1265 sizeof(struct uclamp_rq
)*UCLAMP_CNT
);
1266 cpu_rq(cpu
)->uclamp_flags
= 0;
1269 for_each_clamp_id(clamp_id
) {
1270 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1271 uclamp_none(clamp_id
), false);
1274 /* System defaults allow max clamp values for both indexes */
1275 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1276 for_each_clamp_id(clamp_id
) {
1277 uclamp_default
[clamp_id
] = uc_max
;
1278 #ifdef CONFIG_UCLAMP_TASK_GROUP
1279 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1280 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1285 #else /* CONFIG_UCLAMP_TASK */
1286 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1287 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1288 static inline int uclamp_validate(struct task_struct
*p
,
1289 const struct sched_attr
*attr
)
1293 static void __setscheduler_uclamp(struct task_struct
*p
,
1294 const struct sched_attr
*attr
) { }
1295 static inline void uclamp_fork(struct task_struct
*p
) { }
1296 static inline void init_uclamp(void) { }
1297 #endif /* CONFIG_UCLAMP_TASK */
1299 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1301 if (!(flags
& ENQUEUE_NOCLOCK
))
1302 update_rq_clock(rq
);
1304 if (!(flags
& ENQUEUE_RESTORE
)) {
1305 sched_info_queued(rq
, p
);
1306 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1309 uclamp_rq_inc(rq
, p
);
1310 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1313 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1315 if (!(flags
& DEQUEUE_NOCLOCK
))
1316 update_rq_clock(rq
);
1318 if (!(flags
& DEQUEUE_SAVE
)) {
1319 sched_info_dequeued(rq
, p
);
1320 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1323 uclamp_rq_dec(rq
, p
);
1324 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1327 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1329 if (task_contributes_to_load(p
))
1330 rq
->nr_uninterruptible
--;
1332 enqueue_task(rq
, p
, flags
);
1334 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1337 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1339 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1341 if (task_contributes_to_load(p
))
1342 rq
->nr_uninterruptible
++;
1344 dequeue_task(rq
, p
, flags
);
1348 * __normal_prio - return the priority that is based on the static prio
1350 static inline int __normal_prio(struct task_struct
*p
)
1352 return p
->static_prio
;
1356 * Calculate the expected normal priority: i.e. priority
1357 * without taking RT-inheritance into account. Might be
1358 * boosted by interactivity modifiers. Changes upon fork,
1359 * setprio syscalls, and whenever the interactivity
1360 * estimator recalculates.
1362 static inline int normal_prio(struct task_struct
*p
)
1366 if (task_has_dl_policy(p
))
1367 prio
= MAX_DL_PRIO
-1;
1368 else if (task_has_rt_policy(p
))
1369 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1371 prio
= __normal_prio(p
);
1376 * Calculate the current priority, i.e. the priority
1377 * taken into account by the scheduler. This value might
1378 * be boosted by RT tasks, or might be boosted by
1379 * interactivity modifiers. Will be RT if the task got
1380 * RT-boosted. If not then it returns p->normal_prio.
1382 static int effective_prio(struct task_struct
*p
)
1384 p
->normal_prio
= normal_prio(p
);
1386 * If we are RT tasks or we were boosted to RT priority,
1387 * keep the priority unchanged. Otherwise, update priority
1388 * to the normal priority:
1390 if (!rt_prio(p
->prio
))
1391 return p
->normal_prio
;
1396 * task_curr - is this task currently executing on a CPU?
1397 * @p: the task in question.
1399 * Return: 1 if the task is currently executing. 0 otherwise.
1401 inline int task_curr(const struct task_struct
*p
)
1403 return cpu_curr(task_cpu(p
)) == p
;
1407 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1408 * use the balance_callback list if you want balancing.
1410 * this means any call to check_class_changed() must be followed by a call to
1411 * balance_callback().
1413 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1414 const struct sched_class
*prev_class
,
1417 if (prev_class
!= p
->sched_class
) {
1418 if (prev_class
->switched_from
)
1419 prev_class
->switched_from(rq
, p
);
1421 p
->sched_class
->switched_to(rq
, p
);
1422 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1423 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1426 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1428 const struct sched_class
*class;
1430 if (p
->sched_class
== rq
->curr
->sched_class
) {
1431 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1433 for_each_class(class) {
1434 if (class == rq
->curr
->sched_class
)
1436 if (class == p
->sched_class
) {
1444 * A queue event has occurred, and we're going to schedule. In
1445 * this case, we can save a useless back to back clock update.
1447 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1448 rq_clock_skip_update(rq
);
1454 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1455 * __set_cpus_allowed_ptr() and select_fallback_rq().
1457 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1459 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1462 if (is_per_cpu_kthread(p
))
1463 return cpu_online(cpu
);
1465 return cpu_active(cpu
);
1469 * This is how migration works:
1471 * 1) we invoke migration_cpu_stop() on the target CPU using
1473 * 2) stopper starts to run (implicitly forcing the migrated thread
1475 * 3) it checks whether the migrated task is still in the wrong runqueue.
1476 * 4) if it's in the wrong runqueue then the migration thread removes
1477 * it and puts it into the right queue.
1478 * 5) stopper completes and stop_one_cpu() returns and the migration
1483 * move_queued_task - move a queued task to new rq.
1485 * Returns (locked) new rq. Old rq's lock is released.
1487 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1488 struct task_struct
*p
, int new_cpu
)
1490 lockdep_assert_held(&rq
->lock
);
1492 WRITE_ONCE(p
->on_rq
, TASK_ON_RQ_MIGRATING
);
1493 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
1494 set_task_cpu(p
, new_cpu
);
1497 rq
= cpu_rq(new_cpu
);
1500 BUG_ON(task_cpu(p
) != new_cpu
);
1501 enqueue_task(rq
, p
, 0);
1502 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1503 check_preempt_curr(rq
, p
, 0);
1508 struct migration_arg
{
1509 struct task_struct
*task
;
1514 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1515 * this because either it can't run here any more (set_cpus_allowed()
1516 * away from this CPU, or CPU going down), or because we're
1517 * attempting to rebalance this task on exec (sched_exec).
1519 * So we race with normal scheduler movements, but that's OK, as long
1520 * as the task is no longer on this CPU.
1522 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1523 struct task_struct
*p
, int dest_cpu
)
1525 /* Affinity changed (again). */
1526 if (!is_cpu_allowed(p
, dest_cpu
))
1529 update_rq_clock(rq
);
1530 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1536 * migration_cpu_stop - this will be executed by a highprio stopper thread
1537 * and performs thread migration by bumping thread off CPU then
1538 * 'pushing' onto another runqueue.
1540 static int migration_cpu_stop(void *data
)
1542 struct migration_arg
*arg
= data
;
1543 struct task_struct
*p
= arg
->task
;
1544 struct rq
*rq
= this_rq();
1548 * The original target CPU might have gone down and we might
1549 * be on another CPU but it doesn't matter.
1551 local_irq_disable();
1553 * We need to explicitly wake pending tasks before running
1554 * __migrate_task() such that we will not miss enforcing cpus_ptr
1555 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1557 sched_ttwu_pending();
1559 raw_spin_lock(&p
->pi_lock
);
1562 * If task_rq(p) != rq, it cannot be migrated here, because we're
1563 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1564 * we're holding p->pi_lock.
1566 if (task_rq(p
) == rq
) {
1567 if (task_on_rq_queued(p
))
1568 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1570 p
->wake_cpu
= arg
->dest_cpu
;
1573 raw_spin_unlock(&p
->pi_lock
);
1580 * sched_class::set_cpus_allowed must do the below, but is not required to
1581 * actually call this function.
1583 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1585 cpumask_copy(&p
->cpus_mask
, new_mask
);
1586 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1589 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1591 struct rq
*rq
= task_rq(p
);
1592 bool queued
, running
;
1594 lockdep_assert_held(&p
->pi_lock
);
1596 queued
= task_on_rq_queued(p
);
1597 running
= task_current(rq
, p
);
1601 * Because __kthread_bind() calls this on blocked tasks without
1604 lockdep_assert_held(&rq
->lock
);
1605 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1608 put_prev_task(rq
, p
);
1610 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1613 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1615 set_next_task(rq
, p
);
1619 * Change a given task's CPU affinity. Migrate the thread to a
1620 * proper CPU and schedule it away if the CPU it's executing on
1621 * is removed from the allowed bitmask.
1623 * NOTE: the caller must have a valid reference to the task, the
1624 * task must not exit() & deallocate itself prematurely. The
1625 * call is not atomic; no spinlocks may be held.
1627 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1628 const struct cpumask
*new_mask
, bool check
)
1630 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1631 unsigned int dest_cpu
;
1636 rq
= task_rq_lock(p
, &rf
);
1637 update_rq_clock(rq
);
1639 if (p
->flags
& PF_KTHREAD
) {
1641 * Kernel threads are allowed on online && !active CPUs
1643 cpu_valid_mask
= cpu_online_mask
;
1647 * Must re-check here, to close a race against __kthread_bind(),
1648 * sched_setaffinity() is not guaranteed to observe the flag.
1650 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1655 if (cpumask_equal(p
->cpus_ptr
, new_mask
))
1659 * Picking a ~random cpu helps in cases where we are changing affinity
1660 * for groups of tasks (ie. cpuset), so that load balancing is not
1661 * immediately required to distribute the tasks within their new mask.
1663 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
1664 if (dest_cpu
>= nr_cpu_ids
) {
1669 do_set_cpus_allowed(p
, new_mask
);
1671 if (p
->flags
& PF_KTHREAD
) {
1673 * For kernel threads that do indeed end up on online &&
1674 * !active we want to ensure they are strict per-CPU threads.
1676 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1677 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1678 p
->nr_cpus_allowed
!= 1);
1681 /* Can the task run on the task's current CPU? If so, we're done */
1682 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1685 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1686 struct migration_arg arg
= { p
, dest_cpu
};
1687 /* Need help from migration thread: drop lock and wait. */
1688 task_rq_unlock(rq
, p
, &rf
);
1689 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1691 } else if (task_on_rq_queued(p
)) {
1693 * OK, since we're going to drop the lock immediately
1694 * afterwards anyway.
1696 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1699 task_rq_unlock(rq
, p
, &rf
);
1704 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1706 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1708 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1710 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1712 #ifdef CONFIG_SCHED_DEBUG
1714 * We should never call set_task_cpu() on a blocked task,
1715 * ttwu() will sort out the placement.
1717 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1721 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1722 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1723 * time relying on p->on_rq.
1725 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1726 p
->sched_class
== &fair_sched_class
&&
1727 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1729 #ifdef CONFIG_LOCKDEP
1731 * The caller should hold either p->pi_lock or rq->lock, when changing
1732 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1734 * sched_move_task() holds both and thus holding either pins the cgroup,
1737 * Furthermore, all task_rq users should acquire both locks, see
1740 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1741 lockdep_is_held(&task_rq(p
)->lock
)));
1744 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1746 WARN_ON_ONCE(!cpu_online(new_cpu
));
1749 trace_sched_migrate_task(p
, new_cpu
);
1751 if (task_cpu(p
) != new_cpu
) {
1752 if (p
->sched_class
->migrate_task_rq
)
1753 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1754 p
->se
.nr_migrations
++;
1756 perf_event_task_migrate(p
);
1759 __set_task_cpu(p
, new_cpu
);
1762 #ifdef CONFIG_NUMA_BALANCING
1763 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1765 if (task_on_rq_queued(p
)) {
1766 struct rq
*src_rq
, *dst_rq
;
1767 struct rq_flags srf
, drf
;
1769 src_rq
= task_rq(p
);
1770 dst_rq
= cpu_rq(cpu
);
1772 rq_pin_lock(src_rq
, &srf
);
1773 rq_pin_lock(dst_rq
, &drf
);
1775 deactivate_task(src_rq
, p
, 0);
1776 set_task_cpu(p
, cpu
);
1777 activate_task(dst_rq
, p
, 0);
1778 check_preempt_curr(dst_rq
, p
, 0);
1780 rq_unpin_lock(dst_rq
, &drf
);
1781 rq_unpin_lock(src_rq
, &srf
);
1785 * Task isn't running anymore; make it appear like we migrated
1786 * it before it went to sleep. This means on wakeup we make the
1787 * previous CPU our target instead of where it really is.
1793 struct migration_swap_arg
{
1794 struct task_struct
*src_task
, *dst_task
;
1795 int src_cpu
, dst_cpu
;
1798 static int migrate_swap_stop(void *data
)
1800 struct migration_swap_arg
*arg
= data
;
1801 struct rq
*src_rq
, *dst_rq
;
1804 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1807 src_rq
= cpu_rq(arg
->src_cpu
);
1808 dst_rq
= cpu_rq(arg
->dst_cpu
);
1810 double_raw_lock(&arg
->src_task
->pi_lock
,
1811 &arg
->dst_task
->pi_lock
);
1812 double_rq_lock(src_rq
, dst_rq
);
1814 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1817 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1820 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
1823 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
1826 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1827 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1832 double_rq_unlock(src_rq
, dst_rq
);
1833 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1834 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1840 * Cross migrate two tasks
1842 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
1843 int target_cpu
, int curr_cpu
)
1845 struct migration_swap_arg arg
;
1848 arg
= (struct migration_swap_arg
){
1850 .src_cpu
= curr_cpu
,
1852 .dst_cpu
= target_cpu
,
1855 if (arg
.src_cpu
== arg
.dst_cpu
)
1859 * These three tests are all lockless; this is OK since all of them
1860 * will be re-checked with proper locks held further down the line.
1862 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1865 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
1868 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
1871 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1872 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1877 #endif /* CONFIG_NUMA_BALANCING */
1880 * wait_task_inactive - wait for a thread to unschedule.
1882 * If @match_state is nonzero, it's the @p->state value just checked and
1883 * not expected to change. If it changes, i.e. @p might have woken up,
1884 * then return zero. When we succeed in waiting for @p to be off its CPU,
1885 * we return a positive number (its total switch count). If a second call
1886 * a short while later returns the same number, the caller can be sure that
1887 * @p has remained unscheduled the whole time.
1889 * The caller must ensure that the task *will* unschedule sometime soon,
1890 * else this function might spin for a *long* time. This function can't
1891 * be called with interrupts off, or it may introduce deadlock with
1892 * smp_call_function() if an IPI is sent by the same process we are
1893 * waiting to become inactive.
1895 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1897 int running
, queued
;
1904 * We do the initial early heuristics without holding
1905 * any task-queue locks at all. We'll only try to get
1906 * the runqueue lock when things look like they will
1912 * If the task is actively running on another CPU
1913 * still, just relax and busy-wait without holding
1916 * NOTE! Since we don't hold any locks, it's not
1917 * even sure that "rq" stays as the right runqueue!
1918 * But we don't care, since "task_running()" will
1919 * return false if the runqueue has changed and p
1920 * is actually now running somewhere else!
1922 while (task_running(rq
, p
)) {
1923 if (match_state
&& unlikely(p
->state
!= match_state
))
1929 * Ok, time to look more closely! We need the rq
1930 * lock now, to be *sure*. If we're wrong, we'll
1931 * just go back and repeat.
1933 rq
= task_rq_lock(p
, &rf
);
1934 trace_sched_wait_task(p
);
1935 running
= task_running(rq
, p
);
1936 queued
= task_on_rq_queued(p
);
1938 if (!match_state
|| p
->state
== match_state
)
1939 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1940 task_rq_unlock(rq
, p
, &rf
);
1943 * If it changed from the expected state, bail out now.
1945 if (unlikely(!ncsw
))
1949 * Was it really running after all now that we
1950 * checked with the proper locks actually held?
1952 * Oops. Go back and try again..
1954 if (unlikely(running
)) {
1960 * It's not enough that it's not actively running,
1961 * it must be off the runqueue _entirely_, and not
1964 * So if it was still runnable (but just not actively
1965 * running right now), it's preempted, and we should
1966 * yield - it could be a while.
1968 if (unlikely(queued
)) {
1969 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1971 set_current_state(TASK_UNINTERRUPTIBLE
);
1972 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1977 * Ahh, all good. It wasn't running, and it wasn't
1978 * runnable, which means that it will never become
1979 * running in the future either. We're all done!
1988 * kick_process - kick a running thread to enter/exit the kernel
1989 * @p: the to-be-kicked thread
1991 * Cause a process which is running on another CPU to enter
1992 * kernel-mode, without any delay. (to get signals handled.)
1994 * NOTE: this function doesn't have to take the runqueue lock,
1995 * because all it wants to ensure is that the remote task enters
1996 * the kernel. If the IPI races and the task has been migrated
1997 * to another CPU then no harm is done and the purpose has been
2000 void kick_process(struct task_struct
*p
)
2006 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2007 smp_send_reschedule(cpu
);
2010 EXPORT_SYMBOL_GPL(kick_process
);
2013 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2015 * A few notes on cpu_active vs cpu_online:
2017 * - cpu_active must be a subset of cpu_online
2019 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2020 * see __set_cpus_allowed_ptr(). At this point the newly online
2021 * CPU isn't yet part of the sched domains, and balancing will not
2024 * - on CPU-down we clear cpu_active() to mask the sched domains and
2025 * avoid the load balancer to place new tasks on the to be removed
2026 * CPU. Existing tasks will remain running there and will be taken
2029 * This means that fallback selection must not select !active CPUs.
2030 * And can assume that any active CPU must be online. Conversely
2031 * select_task_rq() below may allow selection of !active CPUs in order
2032 * to satisfy the above rules.
2034 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2036 int nid
= cpu_to_node(cpu
);
2037 const struct cpumask
*nodemask
= NULL
;
2038 enum { cpuset
, possible
, fail
} state
= cpuset
;
2042 * If the node that the CPU is on has been offlined, cpu_to_node()
2043 * will return -1. There is no CPU on the node, and we should
2044 * select the CPU on the other node.
2047 nodemask
= cpumask_of_node(nid
);
2049 /* Look for allowed, online CPU in same node. */
2050 for_each_cpu(dest_cpu
, nodemask
) {
2051 if (!cpu_active(dest_cpu
))
2053 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2059 /* Any allowed, online CPU? */
2060 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2061 if (!is_cpu_allowed(p
, dest_cpu
))
2067 /* No more Mr. Nice Guy. */
2070 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2071 cpuset_cpus_allowed_fallback(p
);
2077 do_set_cpus_allowed(p
, cpu_possible_mask
);
2088 if (state
!= cpuset
) {
2090 * Don't tell them about moving exiting tasks or
2091 * kernel threads (both mm NULL), since they never
2094 if (p
->mm
&& printk_ratelimit()) {
2095 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2096 task_pid_nr(p
), p
->comm
, cpu
);
2104 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2107 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
2109 lockdep_assert_held(&p
->pi_lock
);
2111 if (p
->nr_cpus_allowed
> 1)
2112 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
2114 cpu
= cpumask_any(p
->cpus_ptr
);
2117 * In order not to call set_task_cpu() on a blocking task we need
2118 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2121 * Since this is common to all placement strategies, this lives here.
2123 * [ this allows ->select_task() to simply return task_cpu(p) and
2124 * not worry about this generic constraint ]
2126 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2127 cpu
= select_fallback_rq(task_cpu(p
), p
);
2132 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2134 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2135 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2139 * Make it appear like a SCHED_FIFO task, its something
2140 * userspace knows about and won't get confused about.
2142 * Also, it will make PI more or less work without too
2143 * much confusion -- but then, stop work should not
2144 * rely on PI working anyway.
2146 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2148 stop
->sched_class
= &stop_sched_class
;
2151 cpu_rq(cpu
)->stop
= stop
;
2155 * Reset it back to a normal scheduling class so that
2156 * it can die in pieces.
2158 old_stop
->sched_class
= &rt_sched_class
;
2164 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2165 const struct cpumask
*new_mask
, bool check
)
2167 return set_cpus_allowed_ptr(p
, new_mask
);
2170 #endif /* CONFIG_SMP */
2173 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2177 if (!schedstat_enabled())
2183 if (cpu
== rq
->cpu
) {
2184 __schedstat_inc(rq
->ttwu_local
);
2185 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2187 struct sched_domain
*sd
;
2189 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2191 for_each_domain(rq
->cpu
, sd
) {
2192 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2193 __schedstat_inc(sd
->ttwu_wake_remote
);
2200 if (wake_flags
& WF_MIGRATED
)
2201 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2202 #endif /* CONFIG_SMP */
2204 __schedstat_inc(rq
->ttwu_count
);
2205 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2207 if (wake_flags
& WF_SYNC
)
2208 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2212 * Mark the task runnable and perform wakeup-preemption.
2214 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2215 struct rq_flags
*rf
)
2217 check_preempt_curr(rq
, p
, wake_flags
);
2218 p
->state
= TASK_RUNNING
;
2219 trace_sched_wakeup(p
);
2222 if (p
->sched_class
->task_woken
) {
2224 * Our task @p is fully woken up and running; so its safe to
2225 * drop the rq->lock, hereafter rq is only used for statistics.
2227 rq_unpin_lock(rq
, rf
);
2228 p
->sched_class
->task_woken(rq
, p
);
2229 rq_repin_lock(rq
, rf
);
2232 if (rq
->idle_stamp
) {
2233 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2234 u64 max
= 2*rq
->max_idle_balance_cost
;
2236 update_avg(&rq
->avg_idle
, delta
);
2238 if (rq
->avg_idle
> max
)
2247 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2248 struct rq_flags
*rf
)
2250 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2252 lockdep_assert_held(&rq
->lock
);
2255 if (p
->sched_contributes_to_load
)
2256 rq
->nr_uninterruptible
--;
2258 if (wake_flags
& WF_MIGRATED
)
2259 en_flags
|= ENQUEUE_MIGRATED
;
2262 activate_task(rq
, p
, en_flags
);
2263 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2267 * Called in case the task @p isn't fully descheduled from its runqueue,
2268 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2269 * since all we need to do is flip p->state to TASK_RUNNING, since
2270 * the task is still ->on_rq.
2272 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2278 rq
= __task_rq_lock(p
, &rf
);
2279 if (task_on_rq_queued(p
)) {
2280 /* check_preempt_curr() may use rq clock */
2281 update_rq_clock(rq
);
2282 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
2285 __task_rq_unlock(rq
, &rf
);
2291 void sched_ttwu_pending(void)
2293 struct rq
*rq
= this_rq();
2294 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
2295 struct task_struct
*p
, *t
;
2301 rq_lock_irqsave(rq
, &rf
);
2302 update_rq_clock(rq
);
2304 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
2305 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
2307 rq_unlock_irqrestore(rq
, &rf
);
2310 static void wake_csd_func(void *info
)
2312 sched_ttwu_pending();
2316 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2317 * necessary. The wakee CPU on receipt of the IPI will queue the task
2318 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2319 * of the wakeup instead of the waker.
2321 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
2323 struct rq
*rq
= cpu_rq(cpu
);
2325 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
2327 if (llist_add(&p
->wake_entry
, &rq
->wake_list
)) {
2328 if (!set_nr_if_polling(rq
->idle
))
2329 smp_call_function_single_async(cpu
, &rq
->wake_csd
);
2331 trace_sched_wake_idle_without_ipi(cpu
);
2335 void wake_up_if_idle(int cpu
)
2337 struct rq
*rq
= cpu_rq(cpu
);
2342 if (!is_idle_task(rcu_dereference(rq
->curr
)))
2345 if (set_nr_if_polling(rq
->idle
)) {
2346 trace_sched_wake_idle_without_ipi(cpu
);
2348 rq_lock_irqsave(rq
, &rf
);
2349 if (is_idle_task(rq
->curr
))
2350 smp_send_reschedule(cpu
);
2351 /* Else CPU is not idle, do nothing here: */
2352 rq_unlock_irqrestore(rq
, &rf
);
2359 bool cpus_share_cache(int this_cpu
, int that_cpu
)
2361 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
2364 static inline bool ttwu_queue_cond(int cpu
, int wake_flags
)
2367 * If the CPU does not share cache, then queue the task on the
2368 * remote rqs wakelist to avoid accessing remote data.
2370 if (!cpus_share_cache(smp_processor_id(), cpu
))
2374 * If the task is descheduling and the only running task on the
2375 * CPU then use the wakelist to offload the task activation to
2376 * the soon-to-be-idle CPU as the current CPU is likely busy.
2377 * nr_running is checked to avoid unnecessary task stacking.
2379 if ((wake_flags
& WF_ON_RQ
) && cpu_rq(cpu
)->nr_running
<= 1)
2385 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
2387 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(cpu
, wake_flags
)) {
2388 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
2389 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
2395 #endif /* CONFIG_SMP */
2397 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
2399 struct rq
*rq
= cpu_rq(cpu
);
2402 #if defined(CONFIG_SMP)
2403 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
2408 update_rq_clock(rq
);
2409 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
2414 * Notes on Program-Order guarantees on SMP systems.
2418 * The basic program-order guarantee on SMP systems is that when a task [t]
2419 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2420 * execution on its new CPU [c1].
2422 * For migration (of runnable tasks) this is provided by the following means:
2424 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2425 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2426 * rq(c1)->lock (if not at the same time, then in that order).
2427 * C) LOCK of the rq(c1)->lock scheduling in task
2429 * Release/acquire chaining guarantees that B happens after A and C after B.
2430 * Note: the CPU doing B need not be c0 or c1
2439 * UNLOCK rq(0)->lock
2441 * LOCK rq(0)->lock // orders against CPU0
2443 * UNLOCK rq(0)->lock
2447 * UNLOCK rq(1)->lock
2449 * LOCK rq(1)->lock // orders against CPU2
2452 * UNLOCK rq(1)->lock
2455 * BLOCKING -- aka. SLEEP + WAKEUP
2457 * For blocking we (obviously) need to provide the same guarantee as for
2458 * migration. However the means are completely different as there is no lock
2459 * chain to provide order. Instead we do:
2461 * 1) smp_store_release(X->on_cpu, 0)
2462 * 2) smp_cond_load_acquire(!X->on_cpu)
2466 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2468 * LOCK rq(0)->lock LOCK X->pi_lock
2471 * smp_store_release(X->on_cpu, 0);
2473 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2479 * X->state = RUNNING
2480 * UNLOCK rq(2)->lock
2482 * LOCK rq(2)->lock // orders against CPU1
2485 * UNLOCK rq(2)->lock
2488 * UNLOCK rq(0)->lock
2491 * However, for wakeups there is a second guarantee we must provide, namely we
2492 * must ensure that CONDITION=1 done by the caller can not be reordered with
2493 * accesses to the task state; see try_to_wake_up() and set_current_state().
2497 * try_to_wake_up - wake up a thread
2498 * @p: the thread to be awakened
2499 * @state: the mask of task states that can be woken
2500 * @wake_flags: wake modifier flags (WF_*)
2502 * If (@state & @p->state) @p->state = TASK_RUNNING.
2504 * If the task was not queued/runnable, also place it back on a runqueue.
2506 * Atomic against schedule() which would dequeue a task, also see
2507 * set_current_state().
2509 * This function executes a full memory barrier before accessing the task
2510 * state; see set_current_state().
2512 * Return: %true if @p->state changes (an actual wakeup was done),
2516 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2518 unsigned long flags
;
2519 int cpu
, success
= 0;
2524 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2525 * == smp_processor_id()'. Together this means we can special
2526 * case the whole 'p->on_rq && ttwu_remote()' case below
2527 * without taking any locks.
2530 * - we rely on Program-Order guarantees for all the ordering,
2531 * - we're serialized against set_special_state() by virtue of
2532 * it disabling IRQs (this allows not taking ->pi_lock).
2534 if (!(p
->state
& state
))
2539 trace_sched_waking(p
);
2540 p
->state
= TASK_RUNNING
;
2541 trace_sched_wakeup(p
);
2546 * If we are going to wake up a thread waiting for CONDITION we
2547 * need to ensure that CONDITION=1 done by the caller can not be
2548 * reordered with p->state check below. This pairs with mb() in
2549 * set_current_state() the waiting thread does.
2551 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2552 smp_mb__after_spinlock();
2553 if (!(p
->state
& state
))
2556 trace_sched_waking(p
);
2558 /* We're going to change ->state: */
2563 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2564 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2565 * in smp_cond_load_acquire() below.
2567 * sched_ttwu_pending() try_to_wake_up()
2568 * STORE p->on_rq = 1 LOAD p->state
2571 * __schedule() (switch to task 'p')
2572 * LOCK rq->lock smp_rmb();
2573 * smp_mb__after_spinlock();
2577 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2579 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2580 * __schedule(). See the comment for smp_mb__after_spinlock().
2583 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2587 delayacct_blkio_end(p
);
2588 atomic_dec(&task_rq(p
)->nr_iowait
);
2592 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2593 p
->state
= TASK_WAKING
;
2596 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2597 * possible to, falsely, observe p->on_cpu == 0.
2599 * One must be running (->on_cpu == 1) in order to remove oneself
2600 * from the runqueue.
2602 * __schedule() (switch to task 'p') try_to_wake_up()
2603 * STORE p->on_cpu = 1 LOAD p->on_rq
2606 * __schedule() (put 'p' to sleep)
2607 * LOCK rq->lock smp_rmb();
2608 * smp_mb__after_spinlock();
2609 * STORE p->on_rq = 0 LOAD p->on_cpu
2611 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2612 * __schedule(). See the comment for smp_mb__after_spinlock().
2617 * If the owning (remote) CPU is still in the middle of schedule() with
2618 * this task as prev, considering queueing p on the remote CPUs wake_list
2619 * which potentially sends an IPI instead of spinning on p->on_cpu to
2620 * let the waker make forward progress. This is safe because IRQs are
2621 * disabled and the IPI will deliver after on_cpu is cleared.
2623 if (READ_ONCE(p
->on_cpu
) && ttwu_queue_wakelist(p
, cpu
, wake_flags
| WF_ON_RQ
))
2627 * If the owning (remote) CPU is still in the middle of schedule() with
2628 * this task as prev, wait until its done referencing the task.
2630 * Pairs with the smp_store_release() in finish_task().
2632 * This ensures that tasks getting woken will be fully ordered against
2633 * their previous state and preserve Program Order.
2635 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2637 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2638 if (task_cpu(p
) != cpu
) {
2639 wake_flags
|= WF_MIGRATED
;
2640 psi_ttwu_dequeue(p
);
2641 set_task_cpu(p
, cpu
);
2643 #endif /* CONFIG_SMP */
2645 ttwu_queue(p
, cpu
, wake_flags
);
2647 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2650 ttwu_stat(p
, cpu
, wake_flags
);
2657 * wake_up_process - Wake up a specific process
2658 * @p: The process to be woken up.
2660 * Attempt to wake up the nominated process and move it to the set of runnable
2663 * Return: 1 if the process was woken up, 0 if it was already running.
2665 * This function executes a full memory barrier before accessing the task state.
2667 int wake_up_process(struct task_struct
*p
)
2669 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2671 EXPORT_SYMBOL(wake_up_process
);
2673 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2675 return try_to_wake_up(p
, state
, 0);
2679 * Perform scheduler related setup for a newly forked process p.
2680 * p is forked by current.
2682 * __sched_fork() is basic setup used by init_idle() too:
2684 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2689 p
->se
.exec_start
= 0;
2690 p
->se
.sum_exec_runtime
= 0;
2691 p
->se
.prev_sum_exec_runtime
= 0;
2692 p
->se
.nr_migrations
= 0;
2694 INIT_LIST_HEAD(&p
->se
.group_node
);
2696 #ifdef CONFIG_FAIR_GROUP_SCHED
2697 p
->se
.cfs_rq
= NULL
;
2700 #ifdef CONFIG_SCHEDSTATS
2701 /* Even if schedstat is disabled, there should not be garbage */
2702 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2705 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2706 init_dl_task_timer(&p
->dl
);
2707 init_dl_inactive_task_timer(&p
->dl
);
2708 __dl_clear_params(p
);
2710 INIT_LIST_HEAD(&p
->rt
.run_list
);
2712 p
->rt
.time_slice
= sched_rr_timeslice
;
2716 #ifdef CONFIG_PREEMPT_NOTIFIERS
2717 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2720 #ifdef CONFIG_COMPACTION
2721 p
->capture_control
= NULL
;
2723 init_numa_balancing(clone_flags
, p
);
2726 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2728 #ifdef CONFIG_NUMA_BALANCING
2730 void set_numabalancing_state(bool enabled
)
2733 static_branch_enable(&sched_numa_balancing
);
2735 static_branch_disable(&sched_numa_balancing
);
2738 #ifdef CONFIG_PROC_SYSCTL
2739 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2740 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2744 int state
= static_branch_likely(&sched_numa_balancing
);
2746 if (write
&& !capable(CAP_SYS_ADMIN
))
2751 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2755 set_numabalancing_state(state
);
2761 #ifdef CONFIG_SCHEDSTATS
2763 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2764 static bool __initdata __sched_schedstats
= false;
2766 static void set_schedstats(bool enabled
)
2769 static_branch_enable(&sched_schedstats
);
2771 static_branch_disable(&sched_schedstats
);
2774 void force_schedstat_enabled(void)
2776 if (!schedstat_enabled()) {
2777 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2778 static_branch_enable(&sched_schedstats
);
2782 static int __init
setup_schedstats(char *str
)
2789 * This code is called before jump labels have been set up, so we can't
2790 * change the static branch directly just yet. Instead set a temporary
2791 * variable so init_schedstats() can do it later.
2793 if (!strcmp(str
, "enable")) {
2794 __sched_schedstats
= true;
2796 } else if (!strcmp(str
, "disable")) {
2797 __sched_schedstats
= false;
2802 pr_warn("Unable to parse schedstats=\n");
2806 __setup("schedstats=", setup_schedstats
);
2808 static void __init
init_schedstats(void)
2810 set_schedstats(__sched_schedstats
);
2813 #ifdef CONFIG_PROC_SYSCTL
2814 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2815 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2819 int state
= static_branch_likely(&sched_schedstats
);
2821 if (write
&& !capable(CAP_SYS_ADMIN
))
2826 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2830 set_schedstats(state
);
2833 #endif /* CONFIG_PROC_SYSCTL */
2834 #else /* !CONFIG_SCHEDSTATS */
2835 static inline void init_schedstats(void) {}
2836 #endif /* CONFIG_SCHEDSTATS */
2839 * fork()/clone()-time setup:
2841 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2843 unsigned long flags
;
2845 __sched_fork(clone_flags
, p
);
2847 * We mark the process as NEW here. This guarantees that
2848 * nobody will actually run it, and a signal or other external
2849 * event cannot wake it up and insert it on the runqueue either.
2851 p
->state
= TASK_NEW
;
2854 * Make sure we do not leak PI boosting priority to the child.
2856 p
->prio
= current
->normal_prio
;
2861 * Revert to default priority/policy on fork if requested.
2863 if (unlikely(p
->sched_reset_on_fork
)) {
2864 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2865 p
->policy
= SCHED_NORMAL
;
2866 p
->static_prio
= NICE_TO_PRIO(0);
2868 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2869 p
->static_prio
= NICE_TO_PRIO(0);
2871 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2872 set_load_weight(p
, false);
2875 * We don't need the reset flag anymore after the fork. It has
2876 * fulfilled its duty:
2878 p
->sched_reset_on_fork
= 0;
2881 if (dl_prio(p
->prio
))
2883 else if (rt_prio(p
->prio
))
2884 p
->sched_class
= &rt_sched_class
;
2886 p
->sched_class
= &fair_sched_class
;
2888 init_entity_runnable_average(&p
->se
);
2891 * The child is not yet in the pid-hash so no cgroup attach races,
2892 * and the cgroup is pinned to this child due to cgroup_fork()
2893 * is ran before sched_fork().
2895 * Silence PROVE_RCU.
2897 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2899 * We're setting the CPU for the first time, we don't migrate,
2900 * so use __set_task_cpu().
2902 __set_task_cpu(p
, smp_processor_id());
2903 if (p
->sched_class
->task_fork
)
2904 p
->sched_class
->task_fork(p
);
2905 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2907 #ifdef CONFIG_SCHED_INFO
2908 if (likely(sched_info_on()))
2909 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2911 #if defined(CONFIG_SMP)
2914 init_task_preempt_count(p
);
2916 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2917 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2922 unsigned long to_ratio(u64 period
, u64 runtime
)
2924 if (runtime
== RUNTIME_INF
)
2928 * Doing this here saves a lot of checks in all
2929 * the calling paths, and returning zero seems
2930 * safe for them anyway.
2935 return div64_u64(runtime
<< BW_SHIFT
, period
);
2939 * wake_up_new_task - wake up a newly created task for the first time.
2941 * This function will do some initial scheduler statistics housekeeping
2942 * that must be done for every newly created context, then puts the task
2943 * on the runqueue and wakes it.
2945 void wake_up_new_task(struct task_struct
*p
)
2950 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2951 p
->state
= TASK_RUNNING
;
2954 * Fork balancing, do it here and not earlier because:
2955 * - cpus_ptr can change in the fork path
2956 * - any previously selected CPU might disappear through hotplug
2958 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2959 * as we're not fully set-up yet.
2961 p
->recent_used_cpu
= task_cpu(p
);
2962 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2964 rq
= __task_rq_lock(p
, &rf
);
2965 update_rq_clock(rq
);
2966 post_init_entity_util_avg(p
);
2968 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2969 trace_sched_wakeup_new(p
);
2970 check_preempt_curr(rq
, p
, WF_FORK
);
2972 if (p
->sched_class
->task_woken
) {
2974 * Nothing relies on rq->lock after this, so its fine to
2977 rq_unpin_lock(rq
, &rf
);
2978 p
->sched_class
->task_woken(rq
, p
);
2979 rq_repin_lock(rq
, &rf
);
2982 task_rq_unlock(rq
, p
, &rf
);
2985 #ifdef CONFIG_PREEMPT_NOTIFIERS
2987 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
2989 void preempt_notifier_inc(void)
2991 static_branch_inc(&preempt_notifier_key
);
2993 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2995 void preempt_notifier_dec(void)
2997 static_branch_dec(&preempt_notifier_key
);
2999 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3002 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3003 * @notifier: notifier struct to register
3005 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3007 if (!static_branch_unlikely(&preempt_notifier_key
))
3008 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3010 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3012 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3015 * preempt_notifier_unregister - no longer interested in preemption notifications
3016 * @notifier: notifier struct to unregister
3018 * This is *not* safe to call from within a preemption notifier.
3020 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3022 hlist_del(¬ifier
->link
);
3024 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3026 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3028 struct preempt_notifier
*notifier
;
3030 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3031 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3034 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3036 if (static_branch_unlikely(&preempt_notifier_key
))
3037 __fire_sched_in_preempt_notifiers(curr
);
3041 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3042 struct task_struct
*next
)
3044 struct preempt_notifier
*notifier
;
3046 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3047 notifier
->ops
->sched_out(notifier
, next
);
3050 static __always_inline
void
3051 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3052 struct task_struct
*next
)
3054 if (static_branch_unlikely(&preempt_notifier_key
))
3055 __fire_sched_out_preempt_notifiers(curr
, next
);
3058 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3060 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3065 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3066 struct task_struct
*next
)
3070 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3072 static inline void prepare_task(struct task_struct
*next
)
3076 * Claim the task as running, we do this before switching to it
3077 * such that any running task will have this set.
3083 static inline void finish_task(struct task_struct
*prev
)
3087 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3088 * We must ensure this doesn't happen until the switch is completely
3091 * In particular, the load of prev->state in finish_task_switch() must
3092 * happen before this.
3094 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3096 smp_store_release(&prev
->on_cpu
, 0);
3101 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
3104 * Since the runqueue lock will be released by the next
3105 * task (which is an invalid locking op but in the case
3106 * of the scheduler it's an obvious special-case), so we
3107 * do an early lockdep release here:
3109 rq_unpin_lock(rq
, rf
);
3110 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
3111 #ifdef CONFIG_DEBUG_SPINLOCK
3112 /* this is a valid case when another task releases the spinlock */
3113 rq
->lock
.owner
= next
;
3117 static inline void finish_lock_switch(struct rq
*rq
)
3120 * If we are tracking spinlock dependencies then we have to
3121 * fix up the runqueue lock - which gets 'carried over' from
3122 * prev into current:
3124 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
3125 raw_spin_unlock_irq(&rq
->lock
);
3129 * NOP if the arch has not defined these:
3132 #ifndef prepare_arch_switch
3133 # define prepare_arch_switch(next) do { } while (0)
3136 #ifndef finish_arch_post_lock_switch
3137 # define finish_arch_post_lock_switch() do { } while (0)
3141 * prepare_task_switch - prepare to switch tasks
3142 * @rq: the runqueue preparing to switch
3143 * @prev: the current task that is being switched out
3144 * @next: the task we are going to switch to.
3146 * This is called with the rq lock held and interrupts off. It must
3147 * be paired with a subsequent finish_task_switch after the context
3150 * prepare_task_switch sets up locking and calls architecture specific
3154 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3155 struct task_struct
*next
)
3157 kcov_prepare_switch(prev
);
3158 sched_info_switch(rq
, prev
, next
);
3159 perf_event_task_sched_out(prev
, next
);
3161 fire_sched_out_preempt_notifiers(prev
, next
);
3163 prepare_arch_switch(next
);
3167 * finish_task_switch - clean up after a task-switch
3168 * @prev: the thread we just switched away from.
3170 * finish_task_switch must be called after the context switch, paired
3171 * with a prepare_task_switch call before the context switch.
3172 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3173 * and do any other architecture-specific cleanup actions.
3175 * Note that we may have delayed dropping an mm in context_switch(). If
3176 * so, we finish that here outside of the runqueue lock. (Doing it
3177 * with the lock held can cause deadlocks; see schedule() for
3180 * The context switch have flipped the stack from under us and restored the
3181 * local variables which were saved when this task called schedule() in the
3182 * past. prev == current is still correct but we need to recalculate this_rq
3183 * because prev may have moved to another CPU.
3185 static struct rq
*finish_task_switch(struct task_struct
*prev
)
3186 __releases(rq
->lock
)
3188 struct rq
*rq
= this_rq();
3189 struct mm_struct
*mm
= rq
->prev_mm
;
3193 * The previous task will have left us with a preempt_count of 2
3194 * because it left us after:
3197 * preempt_disable(); // 1
3199 * raw_spin_lock_irq(&rq->lock) // 2
3201 * Also, see FORK_PREEMPT_COUNT.
3203 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
3204 "corrupted preempt_count: %s/%d/0x%x\n",
3205 current
->comm
, current
->pid
, preempt_count()))
3206 preempt_count_set(FORK_PREEMPT_COUNT
);
3211 * A task struct has one reference for the use as "current".
3212 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3213 * schedule one last time. The schedule call will never return, and
3214 * the scheduled task must drop that reference.
3216 * We must observe prev->state before clearing prev->on_cpu (in
3217 * finish_task), otherwise a concurrent wakeup can get prev
3218 * running on another CPU and we could rave with its RUNNING -> DEAD
3219 * transition, resulting in a double drop.
3221 prev_state
= prev
->state
;
3222 vtime_task_switch(prev
);
3223 perf_event_task_sched_in(prev
, current
);
3225 finish_lock_switch(rq
);
3226 finish_arch_post_lock_switch();
3227 kcov_finish_switch(current
);
3229 fire_sched_in_preempt_notifiers(current
);
3231 * When switching through a kernel thread, the loop in
3232 * membarrier_{private,global}_expedited() may have observed that
3233 * kernel thread and not issued an IPI. It is therefore possible to
3234 * schedule between user->kernel->user threads without passing though
3235 * switch_mm(). Membarrier requires a barrier after storing to
3236 * rq->curr, before returning to userspace, so provide them here:
3238 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3239 * provided by mmdrop(),
3240 * - a sync_core for SYNC_CORE.
3243 membarrier_mm_sync_core_before_usermode(mm
);
3246 if (unlikely(prev_state
== TASK_DEAD
)) {
3247 if (prev
->sched_class
->task_dead
)
3248 prev
->sched_class
->task_dead(prev
);
3251 * Remove function-return probe instances associated with this
3252 * task and put them back on the free list.
3254 kprobe_flush_task(prev
);
3256 /* Task is done with its stack. */
3257 put_task_stack(prev
);
3259 put_task_struct_rcu_user(prev
);
3262 tick_nohz_task_switch();
3268 /* rq->lock is NOT held, but preemption is disabled */
3269 static void __balance_callback(struct rq
*rq
)
3271 struct callback_head
*head
, *next
;
3272 void (*func
)(struct rq
*rq
);
3273 unsigned long flags
;
3275 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3276 head
= rq
->balance_callback
;
3277 rq
->balance_callback
= NULL
;
3279 func
= (void (*)(struct rq
*))head
->func
;
3286 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3289 static inline void balance_callback(struct rq
*rq
)
3291 if (unlikely(rq
->balance_callback
))
3292 __balance_callback(rq
);
3297 static inline void balance_callback(struct rq
*rq
)
3304 * schedule_tail - first thing a freshly forked thread must call.
3305 * @prev: the thread we just switched away from.
3307 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
3308 __releases(rq
->lock
)
3313 * New tasks start with FORK_PREEMPT_COUNT, see there and
3314 * finish_task_switch() for details.
3316 * finish_task_switch() will drop rq->lock() and lower preempt_count
3317 * and the preempt_enable() will end up enabling preemption (on
3318 * PREEMPT_COUNT kernels).
3321 rq
= finish_task_switch(prev
);
3322 balance_callback(rq
);
3325 if (current
->set_child_tid
)
3326 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3328 calculate_sigpending();
3332 * context_switch - switch to the new MM and the new thread's register state.
3334 static __always_inline
struct rq
*
3335 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3336 struct task_struct
*next
, struct rq_flags
*rf
)
3338 prepare_task_switch(rq
, prev
, next
);
3341 * For paravirt, this is coupled with an exit in switch_to to
3342 * combine the page table reload and the switch backend into
3345 arch_start_context_switch(prev
);
3348 * kernel -> kernel lazy + transfer active
3349 * user -> kernel lazy + mmgrab() active
3351 * kernel -> user switch + mmdrop() active
3352 * user -> user switch
3354 if (!next
->mm
) { // to kernel
3355 enter_lazy_tlb(prev
->active_mm
, next
);
3357 next
->active_mm
= prev
->active_mm
;
3358 if (prev
->mm
) // from user
3359 mmgrab(prev
->active_mm
);
3361 prev
->active_mm
= NULL
;
3363 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
3365 * sys_membarrier() requires an smp_mb() between setting
3366 * rq->curr / membarrier_switch_mm() and returning to userspace.
3368 * The below provides this either through switch_mm(), or in
3369 * case 'prev->active_mm == next->mm' through
3370 * finish_task_switch()'s mmdrop().
3372 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
3374 if (!prev
->mm
) { // from kernel
3375 /* will mmdrop() in finish_task_switch(). */
3376 rq
->prev_mm
= prev
->active_mm
;
3377 prev
->active_mm
= NULL
;
3381 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3383 prepare_lock_switch(rq
, next
, rf
);
3385 /* Here we just switch the register state and the stack. */
3386 switch_to(prev
, next
, prev
);
3389 return finish_task_switch(prev
);
3393 * nr_running and nr_context_switches:
3395 * externally visible scheduler statistics: current number of runnable
3396 * threads, total number of context switches performed since bootup.
3398 unsigned long nr_running(void)
3400 unsigned long i
, sum
= 0;
3402 for_each_online_cpu(i
)
3403 sum
+= cpu_rq(i
)->nr_running
;
3409 * Check if only the current task is running on the CPU.
3411 * Caution: this function does not check that the caller has disabled
3412 * preemption, thus the result might have a time-of-check-to-time-of-use
3413 * race. The caller is responsible to use it correctly, for example:
3415 * - from a non-preemptible section (of course)
3417 * - from a thread that is bound to a single CPU
3419 * - in a loop with very short iterations (e.g. a polling loop)
3421 bool single_task_running(void)
3423 return raw_rq()->nr_running
== 1;
3425 EXPORT_SYMBOL(single_task_running
);
3427 unsigned long long nr_context_switches(void)
3430 unsigned long long sum
= 0;
3432 for_each_possible_cpu(i
)
3433 sum
+= cpu_rq(i
)->nr_switches
;
3439 * Consumers of these two interfaces, like for example the cpuidle menu
3440 * governor, are using nonsensical data. Preferring shallow idle state selection
3441 * for a CPU that has IO-wait which might not even end up running the task when
3442 * it does become runnable.
3445 unsigned long nr_iowait_cpu(int cpu
)
3447 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
3451 * IO-wait accounting, and how its mostly bollocks (on SMP).
3453 * The idea behind IO-wait account is to account the idle time that we could
3454 * have spend running if it were not for IO. That is, if we were to improve the
3455 * storage performance, we'd have a proportional reduction in IO-wait time.
3457 * This all works nicely on UP, where, when a task blocks on IO, we account
3458 * idle time as IO-wait, because if the storage were faster, it could've been
3459 * running and we'd not be idle.
3461 * This has been extended to SMP, by doing the same for each CPU. This however
3464 * Imagine for instance the case where two tasks block on one CPU, only the one
3465 * CPU will have IO-wait accounted, while the other has regular idle. Even
3466 * though, if the storage were faster, both could've ran at the same time,
3467 * utilising both CPUs.
3469 * This means, that when looking globally, the current IO-wait accounting on
3470 * SMP is a lower bound, by reason of under accounting.
3472 * Worse, since the numbers are provided per CPU, they are sometimes
3473 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3474 * associated with any one particular CPU, it can wake to another CPU than it
3475 * blocked on. This means the per CPU IO-wait number is meaningless.
3477 * Task CPU affinities can make all that even more 'interesting'.
3480 unsigned long nr_iowait(void)
3482 unsigned long i
, sum
= 0;
3484 for_each_possible_cpu(i
)
3485 sum
+= nr_iowait_cpu(i
);
3493 * sched_exec - execve() is a valuable balancing opportunity, because at
3494 * this point the task has the smallest effective memory and cache footprint.
3496 void sched_exec(void)
3498 struct task_struct
*p
= current
;
3499 unsigned long flags
;
3502 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3503 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
3504 if (dest_cpu
== smp_processor_id())
3507 if (likely(cpu_active(dest_cpu
))) {
3508 struct migration_arg arg
= { p
, dest_cpu
};
3510 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3511 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3515 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3520 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3521 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3523 EXPORT_PER_CPU_SYMBOL(kstat
);
3524 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3527 * The function fair_sched_class.update_curr accesses the struct curr
3528 * and its field curr->exec_start; when called from task_sched_runtime(),
3529 * we observe a high rate of cache misses in practice.
3530 * Prefetching this data results in improved performance.
3532 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3534 #ifdef CONFIG_FAIR_GROUP_SCHED
3535 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3537 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3540 prefetch(&curr
->exec_start
);
3544 * Return accounted runtime for the task.
3545 * In case the task is currently running, return the runtime plus current's
3546 * pending runtime that have not been accounted yet.
3548 unsigned long long task_sched_runtime(struct task_struct
*p
)
3554 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3556 * 64-bit doesn't need locks to atomically read a 64-bit value.
3557 * So we have a optimization chance when the task's delta_exec is 0.
3558 * Reading ->on_cpu is racy, but this is ok.
3560 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3561 * If we race with it entering CPU, unaccounted time is 0. This is
3562 * indistinguishable from the read occurring a few cycles earlier.
3563 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3564 * been accounted, so we're correct here as well.
3566 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3567 return p
->se
.sum_exec_runtime
;
3570 rq
= task_rq_lock(p
, &rf
);
3572 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3573 * project cycles that may never be accounted to this
3574 * thread, breaking clock_gettime().
3576 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3577 prefetch_curr_exec_start(p
);
3578 update_rq_clock(rq
);
3579 p
->sched_class
->update_curr(rq
);
3581 ns
= p
->se
.sum_exec_runtime
;
3582 task_rq_unlock(rq
, p
, &rf
);
3587 DEFINE_PER_CPU(unsigned long, thermal_pressure
);
3589 void arch_set_thermal_pressure(struct cpumask
*cpus
,
3590 unsigned long th_pressure
)
3594 for_each_cpu(cpu
, cpus
)
3595 WRITE_ONCE(per_cpu(thermal_pressure
, cpu
), th_pressure
);
3599 * This function gets called by the timer code, with HZ frequency.
3600 * We call it with interrupts disabled.
3602 void scheduler_tick(void)
3604 int cpu
= smp_processor_id();
3605 struct rq
*rq
= cpu_rq(cpu
);
3606 struct task_struct
*curr
= rq
->curr
;
3608 unsigned long thermal_pressure
;
3610 arch_scale_freq_tick();
3615 update_rq_clock(rq
);
3616 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
3617 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
3618 curr
->sched_class
->task_tick(rq
, curr
, 0);
3619 calc_global_load_tick(rq
);
3624 perf_event_task_tick();
3627 rq
->idle_balance
= idle_cpu(cpu
);
3628 trigger_load_balance(rq
);
3632 #ifdef CONFIG_NO_HZ_FULL
3637 struct delayed_work work
;
3639 /* Values for ->state, see diagram below. */
3640 #define TICK_SCHED_REMOTE_OFFLINE 0
3641 #define TICK_SCHED_REMOTE_OFFLINING 1
3642 #define TICK_SCHED_REMOTE_RUNNING 2
3645 * State diagram for ->state:
3648 * TICK_SCHED_REMOTE_OFFLINE
3651 * | | sched_tick_remote()
3654 * +--TICK_SCHED_REMOTE_OFFLINING
3657 * sched_tick_start() | | sched_tick_stop()
3660 * TICK_SCHED_REMOTE_RUNNING
3663 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3664 * and sched_tick_start() are happy to leave the state in RUNNING.
3667 static struct tick_work __percpu
*tick_work_cpu
;
3669 static void sched_tick_remote(struct work_struct
*work
)
3671 struct delayed_work
*dwork
= to_delayed_work(work
);
3672 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3673 int cpu
= twork
->cpu
;
3674 struct rq
*rq
= cpu_rq(cpu
);
3675 struct task_struct
*curr
;
3681 * Handle the tick only if it appears the remote CPU is running in full
3682 * dynticks mode. The check is racy by nature, but missing a tick or
3683 * having one too much is no big deal because the scheduler tick updates
3684 * statistics and checks timeslices in a time-independent way, regardless
3685 * of when exactly it is running.
3687 if (!tick_nohz_tick_stopped_cpu(cpu
))
3690 rq_lock_irq(rq
, &rf
);
3692 if (cpu_is_offline(cpu
))
3695 update_rq_clock(rq
);
3697 if (!is_idle_task(curr
)) {
3699 * Make sure the next tick runs within a reasonable
3702 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3703 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3705 curr
->sched_class
->task_tick(rq
, curr
, 0);
3707 calc_load_nohz_remote(rq
);
3709 rq_unlock_irq(rq
, &rf
);
3713 * Run the remote tick once per second (1Hz). This arbitrary
3714 * frequency is large enough to avoid overload but short enough
3715 * to keep scheduler internal stats reasonably up to date. But
3716 * first update state to reflect hotplug activity if required.
3718 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
3719 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
3720 if (os
== TICK_SCHED_REMOTE_RUNNING
)
3721 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3724 static void sched_tick_start(int cpu
)
3727 struct tick_work
*twork
;
3729 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3732 WARN_ON_ONCE(!tick_work_cpu
);
3734 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3735 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
3736 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
3737 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
3739 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3740 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3744 #ifdef CONFIG_HOTPLUG_CPU
3745 static void sched_tick_stop(int cpu
)
3747 struct tick_work
*twork
;
3750 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3753 WARN_ON_ONCE(!tick_work_cpu
);
3755 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3756 /* There cannot be competing actions, but don't rely on stop-machine. */
3757 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
3758 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
3759 /* Don't cancel, as this would mess up the state machine. */
3761 #endif /* CONFIG_HOTPLUG_CPU */
3763 int __init
sched_tick_offload_init(void)
3765 tick_work_cpu
= alloc_percpu(struct tick_work
);
3766 BUG_ON(!tick_work_cpu
);
3770 #else /* !CONFIG_NO_HZ_FULL */
3771 static inline void sched_tick_start(int cpu
) { }
3772 static inline void sched_tick_stop(int cpu
) { }
3775 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3776 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3778 * If the value passed in is equal to the current preempt count
3779 * then we just disabled preemption. Start timing the latency.
3781 static inline void preempt_latency_start(int val
)
3783 if (preempt_count() == val
) {
3784 unsigned long ip
= get_lock_parent_ip();
3785 #ifdef CONFIG_DEBUG_PREEMPT
3786 current
->preempt_disable_ip
= ip
;
3788 trace_preempt_off(CALLER_ADDR0
, ip
);
3792 void preempt_count_add(int val
)
3794 #ifdef CONFIG_DEBUG_PREEMPT
3798 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3801 __preempt_count_add(val
);
3802 #ifdef CONFIG_DEBUG_PREEMPT
3804 * Spinlock count overflowing soon?
3806 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3809 preempt_latency_start(val
);
3811 EXPORT_SYMBOL(preempt_count_add
);
3812 NOKPROBE_SYMBOL(preempt_count_add
);
3815 * If the value passed in equals to the current preempt count
3816 * then we just enabled preemption. Stop timing the latency.
3818 static inline void preempt_latency_stop(int val
)
3820 if (preempt_count() == val
)
3821 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3824 void preempt_count_sub(int val
)
3826 #ifdef CONFIG_DEBUG_PREEMPT
3830 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3833 * Is the spinlock portion underflowing?
3835 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3836 !(preempt_count() & PREEMPT_MASK
)))
3840 preempt_latency_stop(val
);
3841 __preempt_count_sub(val
);
3843 EXPORT_SYMBOL(preempt_count_sub
);
3844 NOKPROBE_SYMBOL(preempt_count_sub
);
3847 static inline void preempt_latency_start(int val
) { }
3848 static inline void preempt_latency_stop(int val
) { }
3851 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3853 #ifdef CONFIG_DEBUG_PREEMPT
3854 return p
->preempt_disable_ip
;
3861 * Print scheduling while atomic bug:
3863 static noinline
void __schedule_bug(struct task_struct
*prev
)
3865 /* Save this before calling printk(), since that will clobber it */
3866 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3868 if (oops_in_progress
)
3871 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3872 prev
->comm
, prev
->pid
, preempt_count());
3874 debug_show_held_locks(prev
);
3876 if (irqs_disabled())
3877 print_irqtrace_events(prev
);
3878 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3879 && in_atomic_preempt_off()) {
3880 pr_err("Preemption disabled at:");
3881 print_ip_sym(preempt_disable_ip
);
3885 panic("scheduling while atomic\n");
3888 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3892 * Various schedule()-time debugging checks and statistics:
3894 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
3896 #ifdef CONFIG_SCHED_STACK_END_CHECK
3897 if (task_stack_end_corrupted(prev
))
3898 panic("corrupted stack end detected inside scheduler\n");
3901 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3902 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
3903 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3904 prev
->comm
, prev
->pid
, prev
->non_block_count
);
3906 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3910 if (unlikely(in_atomic_preempt_off())) {
3911 __schedule_bug(prev
);
3912 preempt_count_set(PREEMPT_DISABLED
);
3916 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3918 schedstat_inc(this_rq()->sched_count
);
3921 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
3922 struct rq_flags
*rf
)
3925 const struct sched_class
*class;
3927 * We must do the balancing pass before put_prev_task(), such
3928 * that when we release the rq->lock the task is in the same
3929 * state as before we took rq->lock.
3931 * We can terminate the balance pass as soon as we know there is
3932 * a runnable task of @class priority or higher.
3934 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
3935 if (class->balance(rq
, prev
, rf
))
3940 put_prev_task(rq
, prev
);
3944 * Pick up the highest-prio task:
3946 static inline struct task_struct
*
3947 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3949 const struct sched_class
*class;
3950 struct task_struct
*p
;
3953 * Optimization: we know that if all tasks are in the fair class we can
3954 * call that function directly, but only if the @prev task wasn't of a
3955 * higher scheduling class, because otherwise those loose the
3956 * opportunity to pull in more work from other CPUs.
3958 if (likely((prev
->sched_class
== &idle_sched_class
||
3959 prev
->sched_class
== &fair_sched_class
) &&
3960 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3962 p
= pick_next_task_fair(rq
, prev
, rf
);
3963 if (unlikely(p
== RETRY_TASK
))
3966 /* Assumes fair_sched_class->next == idle_sched_class */
3968 put_prev_task(rq
, prev
);
3969 p
= pick_next_task_idle(rq
);
3976 put_prev_task_balance(rq
, prev
, rf
);
3978 for_each_class(class) {
3979 p
= class->pick_next_task(rq
);
3984 /* The idle class should always have a runnable task: */
3989 * __schedule() is the main scheduler function.
3991 * The main means of driving the scheduler and thus entering this function are:
3993 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3995 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3996 * paths. For example, see arch/x86/entry_64.S.
3998 * To drive preemption between tasks, the scheduler sets the flag in timer
3999 * interrupt handler scheduler_tick().
4001 * 3. Wakeups don't really cause entry into schedule(). They add a
4002 * task to the run-queue and that's it.
4004 * Now, if the new task added to the run-queue preempts the current
4005 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4006 * called on the nearest possible occasion:
4008 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4010 * - in syscall or exception context, at the next outmost
4011 * preempt_enable(). (this might be as soon as the wake_up()'s
4014 * - in IRQ context, return from interrupt-handler to
4015 * preemptible context
4017 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4020 * - cond_resched() call
4021 * - explicit schedule() call
4022 * - return from syscall or exception to user-space
4023 * - return from interrupt-handler to user-space
4025 * WARNING: must be called with preemption disabled!
4027 static void __sched notrace
__schedule(bool preempt
)
4029 struct task_struct
*prev
, *next
;
4030 unsigned long *switch_count
;
4035 cpu
= smp_processor_id();
4039 schedule_debug(prev
, preempt
);
4041 if (sched_feat(HRTICK
))
4044 local_irq_disable();
4045 rcu_note_context_switch(preempt
);
4048 * Make sure that signal_pending_state()->signal_pending() below
4049 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4050 * done by the caller to avoid the race with signal_wake_up().
4052 * The membarrier system call requires a full memory barrier
4053 * after coming from user-space, before storing to rq->curr.
4056 smp_mb__after_spinlock();
4058 /* Promote REQ to ACT */
4059 rq
->clock_update_flags
<<= 1;
4060 update_rq_clock(rq
);
4062 switch_count
= &prev
->nivcsw
;
4063 if (!preempt
&& prev
->state
) {
4064 if (signal_pending_state(prev
->state
, prev
)) {
4065 prev
->state
= TASK_RUNNING
;
4067 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
4069 if (prev
->in_iowait
) {
4070 atomic_inc(&rq
->nr_iowait
);
4071 delayacct_blkio_start();
4074 switch_count
= &prev
->nvcsw
;
4077 next
= pick_next_task(rq
, prev
, &rf
);
4078 clear_tsk_need_resched(prev
);
4079 clear_preempt_need_resched();
4081 if (likely(prev
!= next
)) {
4084 * RCU users of rcu_dereference(rq->curr) may not see
4085 * changes to task_struct made by pick_next_task().
4087 RCU_INIT_POINTER(rq
->curr
, next
);
4089 * The membarrier system call requires each architecture
4090 * to have a full memory barrier after updating
4091 * rq->curr, before returning to user-space.
4093 * Here are the schemes providing that barrier on the
4094 * various architectures:
4095 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4096 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4097 * - finish_lock_switch() for weakly-ordered
4098 * architectures where spin_unlock is a full barrier,
4099 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4100 * is a RELEASE barrier),
4104 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
4106 trace_sched_switch(preempt
, prev
, next
);
4108 /* Also unlocks the rq: */
4109 rq
= context_switch(rq
, prev
, next
, &rf
);
4111 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4112 rq_unlock_irq(rq
, &rf
);
4115 balance_callback(rq
);
4118 void __noreturn
do_task_dead(void)
4120 /* Causes final put_task_struct in finish_task_switch(): */
4121 set_special_state(TASK_DEAD
);
4123 /* Tell freezer to ignore us: */
4124 current
->flags
|= PF_NOFREEZE
;
4129 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4134 static inline void sched_submit_work(struct task_struct
*tsk
)
4140 * If a worker went to sleep, notify and ask workqueue whether
4141 * it wants to wake up a task to maintain concurrency.
4142 * As this function is called inside the schedule() context,
4143 * we disable preemption to avoid it calling schedule() again
4144 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4147 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4149 if (tsk
->flags
& PF_WQ_WORKER
)
4150 wq_worker_sleeping(tsk
);
4152 io_wq_worker_sleeping(tsk
);
4153 preempt_enable_no_resched();
4156 if (tsk_is_pi_blocked(tsk
))
4160 * If we are going to sleep and we have plugged IO queued,
4161 * make sure to submit it to avoid deadlocks.
4163 if (blk_needs_flush_plug(tsk
))
4164 blk_schedule_flush_plug(tsk
);
4167 static void sched_update_worker(struct task_struct
*tsk
)
4169 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4170 if (tsk
->flags
& PF_WQ_WORKER
)
4171 wq_worker_running(tsk
);
4173 io_wq_worker_running(tsk
);
4177 asmlinkage __visible
void __sched
schedule(void)
4179 struct task_struct
*tsk
= current
;
4181 sched_submit_work(tsk
);
4185 sched_preempt_enable_no_resched();
4186 } while (need_resched());
4187 sched_update_worker(tsk
);
4189 EXPORT_SYMBOL(schedule
);
4192 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4193 * state (have scheduled out non-voluntarily) by making sure that all
4194 * tasks have either left the run queue or have gone into user space.
4195 * As idle tasks do not do either, they must not ever be preempted
4196 * (schedule out non-voluntarily).
4198 * schedule_idle() is similar to schedule_preempt_disable() except that it
4199 * never enables preemption because it does not call sched_submit_work().
4201 void __sched
schedule_idle(void)
4204 * As this skips calling sched_submit_work(), which the idle task does
4205 * regardless because that function is a nop when the task is in a
4206 * TASK_RUNNING state, make sure this isn't used someplace that the
4207 * current task can be in any other state. Note, idle is always in the
4208 * TASK_RUNNING state.
4210 WARN_ON_ONCE(current
->state
);
4213 } while (need_resched());
4216 #ifdef CONFIG_CONTEXT_TRACKING
4217 asmlinkage __visible
void __sched
schedule_user(void)
4220 * If we come here after a random call to set_need_resched(),
4221 * or we have been woken up remotely but the IPI has not yet arrived,
4222 * we haven't yet exited the RCU idle mode. Do it here manually until
4223 * we find a better solution.
4225 * NB: There are buggy callers of this function. Ideally we
4226 * should warn if prev_state != CONTEXT_USER, but that will trigger
4227 * too frequently to make sense yet.
4229 enum ctx_state prev_state
= exception_enter();
4231 exception_exit(prev_state
);
4236 * schedule_preempt_disabled - called with preemption disabled
4238 * Returns with preemption disabled. Note: preempt_count must be 1
4240 void __sched
schedule_preempt_disabled(void)
4242 sched_preempt_enable_no_resched();
4247 static void __sched notrace
preempt_schedule_common(void)
4251 * Because the function tracer can trace preempt_count_sub()
4252 * and it also uses preempt_enable/disable_notrace(), if
4253 * NEED_RESCHED is set, the preempt_enable_notrace() called
4254 * by the function tracer will call this function again and
4255 * cause infinite recursion.
4257 * Preemption must be disabled here before the function
4258 * tracer can trace. Break up preempt_disable() into two
4259 * calls. One to disable preemption without fear of being
4260 * traced. The other to still record the preemption latency,
4261 * which can also be traced by the function tracer.
4263 preempt_disable_notrace();
4264 preempt_latency_start(1);
4266 preempt_latency_stop(1);
4267 preempt_enable_no_resched_notrace();
4270 * Check again in case we missed a preemption opportunity
4271 * between schedule and now.
4273 } while (need_resched());
4276 #ifdef CONFIG_PREEMPTION
4278 * This is the entry point to schedule() from in-kernel preemption
4279 * off of preempt_enable.
4281 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
4284 * If there is a non-zero preempt_count or interrupts are disabled,
4285 * we do not want to preempt the current task. Just return..
4287 if (likely(!preemptible()))
4290 preempt_schedule_common();
4292 NOKPROBE_SYMBOL(preempt_schedule
);
4293 EXPORT_SYMBOL(preempt_schedule
);
4296 * preempt_schedule_notrace - preempt_schedule called by tracing
4298 * The tracing infrastructure uses preempt_enable_notrace to prevent
4299 * recursion and tracing preempt enabling caused by the tracing
4300 * infrastructure itself. But as tracing can happen in areas coming
4301 * from userspace or just about to enter userspace, a preempt enable
4302 * can occur before user_exit() is called. This will cause the scheduler
4303 * to be called when the system is still in usermode.
4305 * To prevent this, the preempt_enable_notrace will use this function
4306 * instead of preempt_schedule() to exit user context if needed before
4307 * calling the scheduler.
4309 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
4311 enum ctx_state prev_ctx
;
4313 if (likely(!preemptible()))
4318 * Because the function tracer can trace preempt_count_sub()
4319 * and it also uses preempt_enable/disable_notrace(), if
4320 * NEED_RESCHED is set, the preempt_enable_notrace() called
4321 * by the function tracer will call this function again and
4322 * cause infinite recursion.
4324 * Preemption must be disabled here before the function
4325 * tracer can trace. Break up preempt_disable() into two
4326 * calls. One to disable preemption without fear of being
4327 * traced. The other to still record the preemption latency,
4328 * which can also be traced by the function tracer.
4330 preempt_disable_notrace();
4331 preempt_latency_start(1);
4333 * Needs preempt disabled in case user_exit() is traced
4334 * and the tracer calls preempt_enable_notrace() causing
4335 * an infinite recursion.
4337 prev_ctx
= exception_enter();
4339 exception_exit(prev_ctx
);
4341 preempt_latency_stop(1);
4342 preempt_enable_no_resched_notrace();
4343 } while (need_resched());
4345 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
4347 #endif /* CONFIG_PREEMPTION */
4350 * This is the entry point to schedule() from kernel preemption
4351 * off of irq context.
4352 * Note, that this is called and return with irqs disabled. This will
4353 * protect us against recursive calling from irq.
4355 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
4357 enum ctx_state prev_state
;
4359 /* Catch callers which need to be fixed */
4360 BUG_ON(preempt_count() || !irqs_disabled());
4362 prev_state
= exception_enter();
4368 local_irq_disable();
4369 sched_preempt_enable_no_resched();
4370 } while (need_resched());
4372 exception_exit(prev_state
);
4375 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
4378 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4380 EXPORT_SYMBOL(default_wake_function
);
4382 #ifdef CONFIG_RT_MUTEXES
4384 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
4387 prio
= min(prio
, pi_task
->prio
);
4392 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4394 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
4396 return __rt_effective_prio(pi_task
, prio
);
4400 * rt_mutex_setprio - set the current priority of a task
4402 * @pi_task: donor task
4404 * This function changes the 'effective' priority of a task. It does
4405 * not touch ->normal_prio like __setscheduler().
4407 * Used by the rt_mutex code to implement priority inheritance
4408 * logic. Call site only calls if the priority of the task changed.
4410 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
4412 int prio
, oldprio
, queued
, running
, queue_flag
=
4413 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4414 const struct sched_class
*prev_class
;
4418 /* XXX used to be waiter->prio, not waiter->task->prio */
4419 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
4422 * If nothing changed; bail early.
4424 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
4427 rq
= __task_rq_lock(p
, &rf
);
4428 update_rq_clock(rq
);
4430 * Set under pi_lock && rq->lock, such that the value can be used under
4433 * Note that there is loads of tricky to make this pointer cache work
4434 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4435 * ensure a task is de-boosted (pi_task is set to NULL) before the
4436 * task is allowed to run again (and can exit). This ensures the pointer
4437 * points to a blocked task -- which guaratees the task is present.
4439 p
->pi_top_task
= pi_task
;
4442 * For FIFO/RR we only need to set prio, if that matches we're done.
4444 if (prio
== p
->prio
&& !dl_prio(prio
))
4448 * Idle task boosting is a nono in general. There is one
4449 * exception, when PREEMPT_RT and NOHZ is active:
4451 * The idle task calls get_next_timer_interrupt() and holds
4452 * the timer wheel base->lock on the CPU and another CPU wants
4453 * to access the timer (probably to cancel it). We can safely
4454 * ignore the boosting request, as the idle CPU runs this code
4455 * with interrupts disabled and will complete the lock
4456 * protected section without being interrupted. So there is no
4457 * real need to boost.
4459 if (unlikely(p
== rq
->idle
)) {
4460 WARN_ON(p
!= rq
->curr
);
4461 WARN_ON(p
->pi_blocked_on
);
4465 trace_sched_pi_setprio(p
, pi_task
);
4468 if (oldprio
== prio
)
4469 queue_flag
&= ~DEQUEUE_MOVE
;
4471 prev_class
= p
->sched_class
;
4472 queued
= task_on_rq_queued(p
);
4473 running
= task_current(rq
, p
);
4475 dequeue_task(rq
, p
, queue_flag
);
4477 put_prev_task(rq
, p
);
4480 * Boosting condition are:
4481 * 1. -rt task is running and holds mutex A
4482 * --> -dl task blocks on mutex A
4484 * 2. -dl task is running and holds mutex A
4485 * --> -dl task blocks on mutex A and could preempt the
4488 if (dl_prio(prio
)) {
4489 if (!dl_prio(p
->normal_prio
) ||
4490 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
4491 p
->dl
.dl_boosted
= 1;
4492 queue_flag
|= ENQUEUE_REPLENISH
;
4494 p
->dl
.dl_boosted
= 0;
4495 p
->sched_class
= &dl_sched_class
;
4496 } else if (rt_prio(prio
)) {
4497 if (dl_prio(oldprio
))
4498 p
->dl
.dl_boosted
= 0;
4500 queue_flag
|= ENQUEUE_HEAD
;
4501 p
->sched_class
= &rt_sched_class
;
4503 if (dl_prio(oldprio
))
4504 p
->dl
.dl_boosted
= 0;
4505 if (rt_prio(oldprio
))
4507 p
->sched_class
= &fair_sched_class
;
4513 enqueue_task(rq
, p
, queue_flag
);
4515 set_next_task(rq
, p
);
4517 check_class_changed(rq
, p
, prev_class
, oldprio
);
4519 /* Avoid rq from going away on us: */
4521 __task_rq_unlock(rq
, &rf
);
4523 balance_callback(rq
);
4527 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4533 void set_user_nice(struct task_struct
*p
, long nice
)
4535 bool queued
, running
;
4540 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
4543 * We have to be careful, if called from sys_setpriority(),
4544 * the task might be in the middle of scheduling on another CPU.
4546 rq
= task_rq_lock(p
, &rf
);
4547 update_rq_clock(rq
);
4550 * The RT priorities are set via sched_setscheduler(), but we still
4551 * allow the 'normal' nice value to be set - but as expected
4552 * it wont have any effect on scheduling until the task is
4553 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4555 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
4556 p
->static_prio
= NICE_TO_PRIO(nice
);
4559 queued
= task_on_rq_queued(p
);
4560 running
= task_current(rq
, p
);
4562 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
4564 put_prev_task(rq
, p
);
4566 p
->static_prio
= NICE_TO_PRIO(nice
);
4567 set_load_weight(p
, true);
4569 p
->prio
= effective_prio(p
);
4572 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
4574 set_next_task(rq
, p
);
4577 * If the task increased its priority or is running and
4578 * lowered its priority, then reschedule its CPU:
4580 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
4583 task_rq_unlock(rq
, p
, &rf
);
4585 EXPORT_SYMBOL(set_user_nice
);
4588 * can_nice - check if a task can reduce its nice value
4592 int can_nice(const struct task_struct
*p
, const int nice
)
4594 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4595 int nice_rlim
= nice_to_rlimit(nice
);
4597 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4598 capable(CAP_SYS_NICE
));
4601 #ifdef __ARCH_WANT_SYS_NICE
4604 * sys_nice - change the priority of the current process.
4605 * @increment: priority increment
4607 * sys_setpriority is a more generic, but much slower function that
4608 * does similar things.
4610 SYSCALL_DEFINE1(nice
, int, increment
)
4615 * Setpriority might change our priority at the same moment.
4616 * We don't have to worry. Conceptually one call occurs first
4617 * and we have a single winner.
4619 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
4620 nice
= task_nice(current
) + increment
;
4622 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
4623 if (increment
< 0 && !can_nice(current
, nice
))
4626 retval
= security_task_setnice(current
, nice
);
4630 set_user_nice(current
, nice
);
4637 * task_prio - return the priority value of a given task.
4638 * @p: the task in question.
4640 * Return: The priority value as seen by users in /proc.
4641 * RT tasks are offset by -200. Normal tasks are centered
4642 * around 0, value goes from -16 to +15.
4644 int task_prio(const struct task_struct
*p
)
4646 return p
->prio
- MAX_RT_PRIO
;
4650 * idle_cpu - is a given CPU idle currently?
4651 * @cpu: the processor in question.
4653 * Return: 1 if the CPU is currently idle. 0 otherwise.
4655 int idle_cpu(int cpu
)
4657 struct rq
*rq
= cpu_rq(cpu
);
4659 if (rq
->curr
!= rq
->idle
)
4666 if (!llist_empty(&rq
->wake_list
))
4674 * available_idle_cpu - is a given CPU idle for enqueuing work.
4675 * @cpu: the CPU in question.
4677 * Return: 1 if the CPU is currently idle. 0 otherwise.
4679 int available_idle_cpu(int cpu
)
4684 if (vcpu_is_preempted(cpu
))
4691 * idle_task - return the idle task for a given CPU.
4692 * @cpu: the processor in question.
4694 * Return: The idle task for the CPU @cpu.
4696 struct task_struct
*idle_task(int cpu
)
4698 return cpu_rq(cpu
)->idle
;
4702 * find_process_by_pid - find a process with a matching PID value.
4703 * @pid: the pid in question.
4705 * The task of @pid, if found. %NULL otherwise.
4707 static struct task_struct
*find_process_by_pid(pid_t pid
)
4709 return pid
? find_task_by_vpid(pid
) : current
;
4713 * sched_setparam() passes in -1 for its policy, to let the functions
4714 * it calls know not to change it.
4716 #define SETPARAM_POLICY -1
4718 static void __setscheduler_params(struct task_struct
*p
,
4719 const struct sched_attr
*attr
)
4721 int policy
= attr
->sched_policy
;
4723 if (policy
== SETPARAM_POLICY
)
4728 if (dl_policy(policy
))
4729 __setparam_dl(p
, attr
);
4730 else if (fair_policy(policy
))
4731 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4734 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4735 * !rt_policy. Always setting this ensures that things like
4736 * getparam()/getattr() don't report silly values for !rt tasks.
4738 p
->rt_priority
= attr
->sched_priority
;
4739 p
->normal_prio
= normal_prio(p
);
4740 set_load_weight(p
, true);
4743 /* Actually do priority change: must hold pi & rq lock. */
4744 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4745 const struct sched_attr
*attr
, bool keep_boost
)
4748 * If params can't change scheduling class changes aren't allowed
4751 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
4754 __setscheduler_params(p
, attr
);
4757 * Keep a potential priority boosting if called from
4758 * sched_setscheduler().
4760 p
->prio
= normal_prio(p
);
4762 p
->prio
= rt_effective_prio(p
, p
->prio
);
4764 if (dl_prio(p
->prio
))
4765 p
->sched_class
= &dl_sched_class
;
4766 else if (rt_prio(p
->prio
))
4767 p
->sched_class
= &rt_sched_class
;
4769 p
->sched_class
= &fair_sched_class
;
4773 * Check the target process has a UID that matches the current process's:
4775 static bool check_same_owner(struct task_struct
*p
)
4777 const struct cred
*cred
= current_cred(), *pcred
;
4781 pcred
= __task_cred(p
);
4782 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4783 uid_eq(cred
->euid
, pcred
->uid
));
4788 static int __sched_setscheduler(struct task_struct
*p
,
4789 const struct sched_attr
*attr
,
4792 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4793 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4794 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4795 int new_effective_prio
, policy
= attr
->sched_policy
;
4796 const struct sched_class
*prev_class
;
4799 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4802 /* The pi code expects interrupts enabled */
4803 BUG_ON(pi
&& in_interrupt());
4805 /* Double check policy once rq lock held: */
4807 reset_on_fork
= p
->sched_reset_on_fork
;
4808 policy
= oldpolicy
= p
->policy
;
4810 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4812 if (!valid_policy(policy
))
4816 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4820 * Valid priorities for SCHED_FIFO and SCHED_RR are
4821 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4822 * SCHED_BATCH and SCHED_IDLE is 0.
4824 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4825 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4827 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4828 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4832 * Allow unprivileged RT tasks to decrease priority:
4834 if (user
&& !capable(CAP_SYS_NICE
)) {
4835 if (fair_policy(policy
)) {
4836 if (attr
->sched_nice
< task_nice(p
) &&
4837 !can_nice(p
, attr
->sched_nice
))
4841 if (rt_policy(policy
)) {
4842 unsigned long rlim_rtprio
=
4843 task_rlimit(p
, RLIMIT_RTPRIO
);
4845 /* Can't set/change the rt policy: */
4846 if (policy
!= p
->policy
&& !rlim_rtprio
)
4849 /* Can't increase priority: */
4850 if (attr
->sched_priority
> p
->rt_priority
&&
4851 attr
->sched_priority
> rlim_rtprio
)
4856 * Can't set/change SCHED_DEADLINE policy at all for now
4857 * (safest behavior); in the future we would like to allow
4858 * unprivileged DL tasks to increase their relative deadline
4859 * or reduce their runtime (both ways reducing utilization)
4861 if (dl_policy(policy
))
4865 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4866 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4868 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
4869 if (!can_nice(p
, task_nice(p
)))
4873 /* Can't change other user's priorities: */
4874 if (!check_same_owner(p
))
4877 /* Normal users shall not reset the sched_reset_on_fork flag: */
4878 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4883 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4886 retval
= security_task_setscheduler(p
);
4891 /* Update task specific "requested" clamps */
4892 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
4893 retval
= uclamp_validate(p
, attr
);
4902 * Make sure no PI-waiters arrive (or leave) while we are
4903 * changing the priority of the task:
4905 * To be able to change p->policy safely, the appropriate
4906 * runqueue lock must be held.
4908 rq
= task_rq_lock(p
, &rf
);
4909 update_rq_clock(rq
);
4912 * Changing the policy of the stop threads its a very bad idea:
4914 if (p
== rq
->stop
) {
4920 * If not changing anything there's no need to proceed further,
4921 * but store a possible modification of reset_on_fork.
4923 if (unlikely(policy
== p
->policy
)) {
4924 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4926 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4928 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4930 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
4933 p
->sched_reset_on_fork
= reset_on_fork
;
4940 #ifdef CONFIG_RT_GROUP_SCHED
4942 * Do not allow realtime tasks into groups that have no runtime
4945 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4946 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4947 !task_group_is_autogroup(task_group(p
))) {
4953 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
4954 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
4955 cpumask_t
*span
= rq
->rd
->span
;
4958 * Don't allow tasks with an affinity mask smaller than
4959 * the entire root_domain to become SCHED_DEADLINE. We
4960 * will also fail if there's no bandwidth available.
4962 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
4963 rq
->rd
->dl_bw
.bw
== 0) {
4971 /* Re-check policy now with rq lock held: */
4972 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4973 policy
= oldpolicy
= -1;
4974 task_rq_unlock(rq
, p
, &rf
);
4976 cpuset_read_unlock();
4981 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4982 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4985 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
4990 p
->sched_reset_on_fork
= reset_on_fork
;
4995 * Take priority boosted tasks into account. If the new
4996 * effective priority is unchanged, we just store the new
4997 * normal parameters and do not touch the scheduler class and
4998 * the runqueue. This will be done when the task deboost
5001 new_effective_prio
= rt_effective_prio(p
, newprio
);
5002 if (new_effective_prio
== oldprio
)
5003 queue_flags
&= ~DEQUEUE_MOVE
;
5006 queued
= task_on_rq_queued(p
);
5007 running
= task_current(rq
, p
);
5009 dequeue_task(rq
, p
, queue_flags
);
5011 put_prev_task(rq
, p
);
5013 prev_class
= p
->sched_class
;
5015 __setscheduler(rq
, p
, attr
, pi
);
5016 __setscheduler_uclamp(p
, attr
);
5020 * We enqueue to tail when the priority of a task is
5021 * increased (user space view).
5023 if (oldprio
< p
->prio
)
5024 queue_flags
|= ENQUEUE_HEAD
;
5026 enqueue_task(rq
, p
, queue_flags
);
5029 set_next_task(rq
, p
);
5031 check_class_changed(rq
, p
, prev_class
, oldprio
);
5033 /* Avoid rq from going away on us: */
5035 task_rq_unlock(rq
, p
, &rf
);
5038 cpuset_read_unlock();
5039 rt_mutex_adjust_pi(p
);
5042 /* Run balance callbacks after we've adjusted the PI chain: */
5043 balance_callback(rq
);
5049 task_rq_unlock(rq
, p
, &rf
);
5051 cpuset_read_unlock();
5055 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
5056 const struct sched_param
*param
, bool check
)
5058 struct sched_attr attr
= {
5059 .sched_policy
= policy
,
5060 .sched_priority
= param
->sched_priority
,
5061 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
5064 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5065 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
5066 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5067 policy
&= ~SCHED_RESET_ON_FORK
;
5068 attr
.sched_policy
= policy
;
5071 return __sched_setscheduler(p
, &attr
, check
, true);
5074 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5075 * @p: the task in question.
5076 * @policy: new policy.
5077 * @param: structure containing the new RT priority.
5079 * Return: 0 on success. An error code otherwise.
5081 * NOTE that the task may be already dead.
5083 int sched_setscheduler(struct task_struct
*p
, int policy
,
5084 const struct sched_param
*param
)
5086 return _sched_setscheduler(p
, policy
, param
, true);
5088 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5090 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
5092 return __sched_setscheduler(p
, attr
, true, true);
5094 EXPORT_SYMBOL_GPL(sched_setattr
);
5096 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
5098 return __sched_setscheduler(p
, attr
, false, true);
5102 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5103 * @p: the task in question.
5104 * @policy: new policy.
5105 * @param: structure containing the new RT priority.
5107 * Just like sched_setscheduler, only don't bother checking if the
5108 * current context has permission. For example, this is needed in
5109 * stop_machine(): we create temporary high priority worker threads,
5110 * but our caller might not have that capability.
5112 * Return: 0 on success. An error code otherwise.
5114 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5115 const struct sched_param
*param
)
5117 return _sched_setscheduler(p
, policy
, param
, false);
5119 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
5122 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5124 struct sched_param lparam
;
5125 struct task_struct
*p
;
5128 if (!param
|| pid
< 0)
5130 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5135 p
= find_process_by_pid(pid
);
5141 retval
= sched_setscheduler(p
, policy
, &lparam
);
5149 * Mimics kernel/events/core.c perf_copy_attr().
5151 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
5156 /* Zero the full structure, so that a short copy will be nice: */
5157 memset(attr
, 0, sizeof(*attr
));
5159 ret
= get_user(size
, &uattr
->size
);
5163 /* ABI compatibility quirk: */
5165 size
= SCHED_ATTR_SIZE_VER0
;
5166 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
5169 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
5176 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
5177 size
< SCHED_ATTR_SIZE_VER1
)
5181 * XXX: Do we want to be lenient like existing syscalls; or do we want
5182 * to be strict and return an error on out-of-bounds values?
5184 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
5189 put_user(sizeof(*attr
), &uattr
->size
);
5194 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5195 * @pid: the pid in question.
5196 * @policy: new policy.
5197 * @param: structure containing the new RT priority.
5199 * Return: 0 on success. An error code otherwise.
5201 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
5206 return do_sched_setscheduler(pid
, policy
, param
);
5210 * sys_sched_setparam - set/change the RT priority of a thread
5211 * @pid: the pid in question.
5212 * @param: structure containing the new RT priority.
5214 * Return: 0 on success. An error code otherwise.
5216 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5218 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
5222 * sys_sched_setattr - same as above, but with extended sched_attr
5223 * @pid: the pid in question.
5224 * @uattr: structure containing the extended parameters.
5225 * @flags: for future extension.
5227 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5228 unsigned int, flags
)
5230 struct sched_attr attr
;
5231 struct task_struct
*p
;
5234 if (!uattr
|| pid
< 0 || flags
)
5237 retval
= sched_copy_attr(uattr
, &attr
);
5241 if ((int)attr
.sched_policy
< 0)
5243 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
5244 attr
.sched_policy
= SETPARAM_POLICY
;
5248 p
= find_process_by_pid(pid
);
5254 retval
= sched_setattr(p
, &attr
);
5262 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5263 * @pid: the pid in question.
5265 * Return: On success, the policy of the thread. Otherwise, a negative error
5268 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5270 struct task_struct
*p
;
5278 p
= find_process_by_pid(pid
);
5280 retval
= security_task_getscheduler(p
);
5283 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5290 * sys_sched_getparam - get the RT priority of a thread
5291 * @pid: the pid in question.
5292 * @param: structure containing the RT priority.
5294 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5297 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5299 struct sched_param lp
= { .sched_priority
= 0 };
5300 struct task_struct
*p
;
5303 if (!param
|| pid
< 0)
5307 p
= find_process_by_pid(pid
);
5312 retval
= security_task_getscheduler(p
);
5316 if (task_has_rt_policy(p
))
5317 lp
.sched_priority
= p
->rt_priority
;
5321 * This one might sleep, we cannot do it with a spinlock held ...
5323 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5333 * Copy the kernel size attribute structure (which might be larger
5334 * than what user-space knows about) to user-space.
5336 * Note that all cases are valid: user-space buffer can be larger or
5337 * smaller than the kernel-space buffer. The usual case is that both
5338 * have the same size.
5341 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
5342 struct sched_attr
*kattr
,
5345 unsigned int ksize
= sizeof(*kattr
);
5347 if (!access_ok(uattr
, usize
))
5351 * sched_getattr() ABI forwards and backwards compatibility:
5353 * If usize == ksize then we just copy everything to user-space and all is good.
5355 * If usize < ksize then we only copy as much as user-space has space for,
5356 * this keeps ABI compatibility as well. We skip the rest.
5358 * If usize > ksize then user-space is using a newer version of the ABI,
5359 * which part the kernel doesn't know about. Just ignore it - tooling can
5360 * detect the kernel's knowledge of attributes from the attr->size value
5361 * which is set to ksize in this case.
5363 kattr
->size
= min(usize
, ksize
);
5365 if (copy_to_user(uattr
, kattr
, kattr
->size
))
5372 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5373 * @pid: the pid in question.
5374 * @uattr: structure containing the extended parameters.
5375 * @usize: sizeof(attr) for fwd/bwd comp.
5376 * @flags: for future extension.
5378 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5379 unsigned int, usize
, unsigned int, flags
)
5381 struct sched_attr kattr
= { };
5382 struct task_struct
*p
;
5385 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
5386 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
5390 p
= find_process_by_pid(pid
);
5395 retval
= security_task_getscheduler(p
);
5399 kattr
.sched_policy
= p
->policy
;
5400 if (p
->sched_reset_on_fork
)
5401 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5402 if (task_has_dl_policy(p
))
5403 __getparam_dl(p
, &kattr
);
5404 else if (task_has_rt_policy(p
))
5405 kattr
.sched_priority
= p
->rt_priority
;
5407 kattr
.sched_nice
= task_nice(p
);
5409 #ifdef CONFIG_UCLAMP_TASK
5410 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
5411 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
5416 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
5423 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5425 cpumask_var_t cpus_allowed
, new_mask
;
5426 struct task_struct
*p
;
5431 p
= find_process_by_pid(pid
);
5437 /* Prevent p going away */
5441 if (p
->flags
& PF_NO_SETAFFINITY
) {
5445 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5449 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5451 goto out_free_cpus_allowed
;
5454 if (!check_same_owner(p
)) {
5456 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
5458 goto out_free_new_mask
;
5463 retval
= security_task_setscheduler(p
);
5465 goto out_free_new_mask
;
5468 cpuset_cpus_allowed(p
, cpus_allowed
);
5469 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5472 * Since bandwidth control happens on root_domain basis,
5473 * if admission test is enabled, we only admit -deadline
5474 * tasks allowed to run on all the CPUs in the task's
5478 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
5480 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
5483 goto out_free_new_mask
;
5489 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
5492 cpuset_cpus_allowed(p
, cpus_allowed
);
5493 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5495 * We must have raced with a concurrent cpuset
5496 * update. Just reset the cpus_allowed to the
5497 * cpuset's cpus_allowed
5499 cpumask_copy(new_mask
, cpus_allowed
);
5504 free_cpumask_var(new_mask
);
5505 out_free_cpus_allowed
:
5506 free_cpumask_var(cpus_allowed
);
5512 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5513 struct cpumask
*new_mask
)
5515 if (len
< cpumask_size())
5516 cpumask_clear(new_mask
);
5517 else if (len
> cpumask_size())
5518 len
= cpumask_size();
5520 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5524 * sys_sched_setaffinity - set the CPU affinity of a process
5525 * @pid: pid of the process
5526 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5527 * @user_mask_ptr: user-space pointer to the new CPU mask
5529 * Return: 0 on success. An error code otherwise.
5531 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5532 unsigned long __user
*, user_mask_ptr
)
5534 cpumask_var_t new_mask
;
5537 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5540 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5542 retval
= sched_setaffinity(pid
, new_mask
);
5543 free_cpumask_var(new_mask
);
5547 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5549 struct task_struct
*p
;
5550 unsigned long flags
;
5556 p
= find_process_by_pid(pid
);
5560 retval
= security_task_getscheduler(p
);
5564 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5565 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
5566 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5575 * sys_sched_getaffinity - get the CPU affinity of a process
5576 * @pid: pid of the process
5577 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5578 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5580 * Return: size of CPU mask copied to user_mask_ptr on success. An
5581 * error code otherwise.
5583 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5584 unsigned long __user
*, user_mask_ptr
)
5589 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5591 if (len
& (sizeof(unsigned long)-1))
5594 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5597 ret
= sched_getaffinity(pid
, mask
);
5599 unsigned int retlen
= min(len
, cpumask_size());
5601 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5606 free_cpumask_var(mask
);
5612 * sys_sched_yield - yield the current processor to other threads.
5614 * This function yields the current CPU to other tasks. If there are no
5615 * other threads running on this CPU then this function will return.
5619 static void do_sched_yield(void)
5624 rq
= this_rq_lock_irq(&rf
);
5626 schedstat_inc(rq
->yld_count
);
5627 current
->sched_class
->yield_task(rq
);
5630 * Since we are going to call schedule() anyway, there's
5631 * no need to preempt or enable interrupts:
5635 sched_preempt_enable_no_resched();
5640 SYSCALL_DEFINE0(sched_yield
)
5646 #ifndef CONFIG_PREEMPTION
5647 int __sched
_cond_resched(void)
5649 if (should_resched(0)) {
5650 preempt_schedule_common();
5656 EXPORT_SYMBOL(_cond_resched
);
5660 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5661 * call schedule, and on return reacquire the lock.
5663 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5664 * operations here to prevent schedule() from being called twice (once via
5665 * spin_unlock(), once by hand).
5667 int __cond_resched_lock(spinlock_t
*lock
)
5669 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5672 lockdep_assert_held(lock
);
5674 if (spin_needbreak(lock
) || resched
) {
5677 preempt_schedule_common();
5685 EXPORT_SYMBOL(__cond_resched_lock
);
5688 * yield - yield the current processor to other threads.
5690 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5692 * The scheduler is at all times free to pick the calling task as the most
5693 * eligible task to run, if removing the yield() call from your code breaks
5694 * it, its already broken.
5696 * Typical broken usage is:
5701 * where one assumes that yield() will let 'the other' process run that will
5702 * make event true. If the current task is a SCHED_FIFO task that will never
5703 * happen. Never use yield() as a progress guarantee!!
5705 * If you want to use yield() to wait for something, use wait_event().
5706 * If you want to use yield() to be 'nice' for others, use cond_resched().
5707 * If you still want to use yield(), do not!
5709 void __sched
yield(void)
5711 set_current_state(TASK_RUNNING
);
5714 EXPORT_SYMBOL(yield
);
5717 * yield_to - yield the current processor to another thread in
5718 * your thread group, or accelerate that thread toward the
5719 * processor it's on.
5721 * @preempt: whether task preemption is allowed or not
5723 * It's the caller's job to ensure that the target task struct
5724 * can't go away on us before we can do any checks.
5727 * true (>0) if we indeed boosted the target task.
5728 * false (0) if we failed to boost the target.
5729 * -ESRCH if there's no task to yield to.
5731 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5733 struct task_struct
*curr
= current
;
5734 struct rq
*rq
, *p_rq
;
5735 unsigned long flags
;
5738 local_irq_save(flags
);
5744 * If we're the only runnable task on the rq and target rq also
5745 * has only one task, there's absolutely no point in yielding.
5747 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5752 double_rq_lock(rq
, p_rq
);
5753 if (task_rq(p
) != p_rq
) {
5754 double_rq_unlock(rq
, p_rq
);
5758 if (!curr
->sched_class
->yield_to_task
)
5761 if (curr
->sched_class
!= p
->sched_class
)
5764 if (task_running(p_rq
, p
) || p
->state
)
5767 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5769 schedstat_inc(rq
->yld_count
);
5771 * Make p's CPU reschedule; pick_next_entity takes care of
5774 if (preempt
&& rq
!= p_rq
)
5779 double_rq_unlock(rq
, p_rq
);
5781 local_irq_restore(flags
);
5788 EXPORT_SYMBOL_GPL(yield_to
);
5790 int io_schedule_prepare(void)
5792 int old_iowait
= current
->in_iowait
;
5794 current
->in_iowait
= 1;
5795 blk_schedule_flush_plug(current
);
5800 void io_schedule_finish(int token
)
5802 current
->in_iowait
= token
;
5806 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5807 * that process accounting knows that this is a task in IO wait state.
5809 long __sched
io_schedule_timeout(long timeout
)
5814 token
= io_schedule_prepare();
5815 ret
= schedule_timeout(timeout
);
5816 io_schedule_finish(token
);
5820 EXPORT_SYMBOL(io_schedule_timeout
);
5822 void __sched
io_schedule(void)
5826 token
= io_schedule_prepare();
5828 io_schedule_finish(token
);
5830 EXPORT_SYMBOL(io_schedule
);
5833 * sys_sched_get_priority_max - return maximum RT priority.
5834 * @policy: scheduling class.
5836 * Return: On success, this syscall returns the maximum
5837 * rt_priority that can be used by a given scheduling class.
5838 * On failure, a negative error code is returned.
5840 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5847 ret
= MAX_USER_RT_PRIO
-1;
5849 case SCHED_DEADLINE
:
5860 * sys_sched_get_priority_min - return minimum RT priority.
5861 * @policy: scheduling class.
5863 * Return: On success, this syscall returns the minimum
5864 * rt_priority that can be used by a given scheduling class.
5865 * On failure, a negative error code is returned.
5867 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5876 case SCHED_DEADLINE
:
5885 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5887 struct task_struct
*p
;
5888 unsigned int time_slice
;
5898 p
= find_process_by_pid(pid
);
5902 retval
= security_task_getscheduler(p
);
5906 rq
= task_rq_lock(p
, &rf
);
5908 if (p
->sched_class
->get_rr_interval
)
5909 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5910 task_rq_unlock(rq
, p
, &rf
);
5913 jiffies_to_timespec64(time_slice
, t
);
5922 * sys_sched_rr_get_interval - return the default timeslice of a process.
5923 * @pid: pid of the process.
5924 * @interval: userspace pointer to the timeslice value.
5926 * this syscall writes the default timeslice value of a given process
5927 * into the user-space timespec buffer. A value of '0' means infinity.
5929 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5932 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5933 struct __kernel_timespec __user
*, interval
)
5935 struct timespec64 t
;
5936 int retval
= sched_rr_get_interval(pid
, &t
);
5939 retval
= put_timespec64(&t
, interval
);
5944 #ifdef CONFIG_COMPAT_32BIT_TIME
5945 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
5946 struct old_timespec32 __user
*, interval
)
5948 struct timespec64 t
;
5949 int retval
= sched_rr_get_interval(pid
, &t
);
5952 retval
= put_old_timespec32(&t
, interval
);
5957 void sched_show_task(struct task_struct
*p
)
5959 unsigned long free
= 0;
5962 if (!try_get_task_stack(p
))
5965 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5967 if (p
->state
== TASK_RUNNING
)
5968 printk(KERN_CONT
" running task ");
5969 #ifdef CONFIG_DEBUG_STACK_USAGE
5970 free
= stack_not_used(p
);
5975 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5977 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5978 task_pid_nr(p
), ppid
,
5979 (unsigned long)task_thread_info(p
)->flags
);
5981 print_worker_info(KERN_INFO
, p
);
5982 show_stack(p
, NULL
);
5985 EXPORT_SYMBOL_GPL(sched_show_task
);
5988 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
5990 /* no filter, everything matches */
5994 /* filter, but doesn't match */
5995 if (!(p
->state
& state_filter
))
5999 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6002 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
6009 void show_state_filter(unsigned long state_filter
)
6011 struct task_struct
*g
, *p
;
6013 #if BITS_PER_LONG == 32
6015 " task PC stack pid father\n");
6018 " task PC stack pid father\n");
6021 for_each_process_thread(g
, p
) {
6023 * reset the NMI-timeout, listing all files on a slow
6024 * console might take a lot of time:
6025 * Also, reset softlockup watchdogs on all CPUs, because
6026 * another CPU might be blocked waiting for us to process
6029 touch_nmi_watchdog();
6030 touch_all_softlockup_watchdogs();
6031 if (state_filter_match(state_filter
, p
))
6035 #ifdef CONFIG_SCHED_DEBUG
6037 sysrq_sched_debug_show();
6041 * Only show locks if all tasks are dumped:
6044 debug_show_all_locks();
6048 * init_idle - set up an idle thread for a given CPU
6049 * @idle: task in question
6050 * @cpu: CPU the idle task belongs to
6052 * NOTE: this function does not set the idle thread's NEED_RESCHED
6053 * flag, to make booting more robust.
6055 void init_idle(struct task_struct
*idle
, int cpu
)
6057 struct rq
*rq
= cpu_rq(cpu
);
6058 unsigned long flags
;
6060 __sched_fork(0, idle
);
6062 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
6063 raw_spin_lock(&rq
->lock
);
6065 idle
->state
= TASK_RUNNING
;
6066 idle
->se
.exec_start
= sched_clock();
6067 idle
->flags
|= PF_IDLE
;
6069 kasan_unpoison_task_stack(idle
);
6073 * Its possible that init_idle() gets called multiple times on a task,
6074 * in that case do_set_cpus_allowed() will not do the right thing.
6076 * And since this is boot we can forgo the serialization.
6078 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
6081 * We're having a chicken and egg problem, even though we are
6082 * holding rq->lock, the CPU isn't yet set to this CPU so the
6083 * lockdep check in task_group() will fail.
6085 * Similar case to sched_fork(). / Alternatively we could
6086 * use task_rq_lock() here and obtain the other rq->lock.
6091 __set_task_cpu(idle
, cpu
);
6095 rcu_assign_pointer(rq
->curr
, idle
);
6096 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
6100 raw_spin_unlock(&rq
->lock
);
6101 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
6103 /* Set the preempt count _outside_ the spinlocks! */
6104 init_idle_preempt_count(idle
, cpu
);
6107 * The idle tasks have their own, simple scheduling class:
6109 idle
->sched_class
= &idle_sched_class
;
6110 ftrace_graph_init_idle_task(idle
, cpu
);
6111 vtime_init_idle(idle
, cpu
);
6113 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
6119 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
6120 const struct cpumask
*trial
)
6124 if (!cpumask_weight(cur
))
6127 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
6132 int task_can_attach(struct task_struct
*p
,
6133 const struct cpumask
*cs_cpus_allowed
)
6138 * Kthreads which disallow setaffinity shouldn't be moved
6139 * to a new cpuset; we don't want to change their CPU
6140 * affinity and isolating such threads by their set of
6141 * allowed nodes is unnecessary. Thus, cpusets are not
6142 * applicable for such threads. This prevents checking for
6143 * success of set_cpus_allowed_ptr() on all attached tasks
6144 * before cpus_mask may be changed.
6146 if (p
->flags
& PF_NO_SETAFFINITY
) {
6151 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
6153 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
6159 bool sched_smp_initialized __read_mostly
;
6161 #ifdef CONFIG_NUMA_BALANCING
6162 /* Migrate current task p to target_cpu */
6163 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
6165 struct migration_arg arg
= { p
, target_cpu
};
6166 int curr_cpu
= task_cpu(p
);
6168 if (curr_cpu
== target_cpu
)
6171 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
6174 /* TODO: This is not properly updating schedstats */
6176 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
6177 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
6181 * Requeue a task on a given node and accurately track the number of NUMA
6182 * tasks on the runqueues
6184 void sched_setnuma(struct task_struct
*p
, int nid
)
6186 bool queued
, running
;
6190 rq
= task_rq_lock(p
, &rf
);
6191 queued
= task_on_rq_queued(p
);
6192 running
= task_current(rq
, p
);
6195 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
6197 put_prev_task(rq
, p
);
6199 p
->numa_preferred_nid
= nid
;
6202 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
6204 set_next_task(rq
, p
);
6205 task_rq_unlock(rq
, p
, &rf
);
6207 #endif /* CONFIG_NUMA_BALANCING */
6209 #ifdef CONFIG_HOTPLUG_CPU
6211 * Ensure that the idle task is using init_mm right before its CPU goes
6214 void idle_task_exit(void)
6216 struct mm_struct
*mm
= current
->active_mm
;
6218 BUG_ON(cpu_online(smp_processor_id()));
6219 BUG_ON(current
!= this_rq()->idle
);
6221 if (mm
!= &init_mm
) {
6222 switch_mm(mm
, &init_mm
, current
);
6223 finish_arch_post_lock_switch();
6226 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6230 * Since this CPU is going 'away' for a while, fold any nr_active delta
6231 * we might have. Assumes we're called after migrate_tasks() so that the
6232 * nr_active count is stable. We need to take the teardown thread which
6233 * is calling this into account, so we hand in adjust = 1 to the load
6236 * Also see the comment "Global load-average calculations".
6238 static void calc_load_migrate(struct rq
*rq
)
6240 long delta
= calc_load_fold_active(rq
, 1);
6242 atomic_long_add(delta
, &calc_load_tasks
);
6245 static struct task_struct
*__pick_migrate_task(struct rq
*rq
)
6247 const struct sched_class
*class;
6248 struct task_struct
*next
;
6250 for_each_class(class) {
6251 next
= class->pick_next_task(rq
);
6253 next
->sched_class
->put_prev_task(rq
, next
);
6258 /* The idle class should always have a runnable task */
6263 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6264 * try_to_wake_up()->select_task_rq().
6266 * Called with rq->lock held even though we'er in stop_machine() and
6267 * there's no concurrency possible, we hold the required locks anyway
6268 * because of lock validation efforts.
6270 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
6272 struct rq
*rq
= dead_rq
;
6273 struct task_struct
*next
, *stop
= rq
->stop
;
6274 struct rq_flags orf
= *rf
;
6278 * Fudge the rq selection such that the below task selection loop
6279 * doesn't get stuck on the currently eligible stop task.
6281 * We're currently inside stop_machine() and the rq is either stuck
6282 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6283 * either way we should never end up calling schedule() until we're
6289 * put_prev_task() and pick_next_task() sched
6290 * class method both need to have an up-to-date
6291 * value of rq->clock[_task]
6293 update_rq_clock(rq
);
6297 * There's this thread running, bail when that's the only
6300 if (rq
->nr_running
== 1)
6303 next
= __pick_migrate_task(rq
);
6306 * Rules for changing task_struct::cpus_mask are holding
6307 * both pi_lock and rq->lock, such that holding either
6308 * stabilizes the mask.
6310 * Drop rq->lock is not quite as disastrous as it usually is
6311 * because !cpu_active at this point, which means load-balance
6312 * will not interfere. Also, stop-machine.
6315 raw_spin_lock(&next
->pi_lock
);
6319 * Since we're inside stop-machine, _nothing_ should have
6320 * changed the task, WARN if weird stuff happened, because in
6321 * that case the above rq->lock drop is a fail too.
6323 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
6324 raw_spin_unlock(&next
->pi_lock
);
6328 /* Find suitable destination for @next, with force if needed. */
6329 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
6330 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
6331 if (rq
!= dead_rq
) {
6337 raw_spin_unlock(&next
->pi_lock
);
6342 #endif /* CONFIG_HOTPLUG_CPU */
6344 void set_rq_online(struct rq
*rq
)
6347 const struct sched_class
*class;
6349 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6352 for_each_class(class) {
6353 if (class->rq_online
)
6354 class->rq_online(rq
);
6359 void set_rq_offline(struct rq
*rq
)
6362 const struct sched_class
*class;
6364 for_each_class(class) {
6365 if (class->rq_offline
)
6366 class->rq_offline(rq
);
6369 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6375 * used to mark begin/end of suspend/resume:
6377 static int num_cpus_frozen
;
6380 * Update cpusets according to cpu_active mask. If cpusets are
6381 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6382 * around partition_sched_domains().
6384 * If we come here as part of a suspend/resume, don't touch cpusets because we
6385 * want to restore it back to its original state upon resume anyway.
6387 static void cpuset_cpu_active(void)
6389 if (cpuhp_tasks_frozen
) {
6391 * num_cpus_frozen tracks how many CPUs are involved in suspend
6392 * resume sequence. As long as this is not the last online
6393 * operation in the resume sequence, just build a single sched
6394 * domain, ignoring cpusets.
6396 partition_sched_domains(1, NULL
, NULL
);
6397 if (--num_cpus_frozen
)
6400 * This is the last CPU online operation. So fall through and
6401 * restore the original sched domains by considering the
6402 * cpuset configurations.
6404 cpuset_force_rebuild();
6406 cpuset_update_active_cpus();
6409 static int cpuset_cpu_inactive(unsigned int cpu
)
6411 if (!cpuhp_tasks_frozen
) {
6412 if (dl_cpu_busy(cpu
))
6414 cpuset_update_active_cpus();
6417 partition_sched_domains(1, NULL
, NULL
);
6422 int sched_cpu_activate(unsigned int cpu
)
6424 struct rq
*rq
= cpu_rq(cpu
);
6427 #ifdef CONFIG_SCHED_SMT
6429 * When going up, increment the number of cores with SMT present.
6431 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6432 static_branch_inc_cpuslocked(&sched_smt_present
);
6434 set_cpu_active(cpu
, true);
6436 if (sched_smp_initialized
) {
6437 sched_domains_numa_masks_set(cpu
);
6438 cpuset_cpu_active();
6442 * Put the rq online, if not already. This happens:
6444 * 1) In the early boot process, because we build the real domains
6445 * after all CPUs have been brought up.
6447 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6450 rq_lock_irqsave(rq
, &rf
);
6452 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6455 rq_unlock_irqrestore(rq
, &rf
);
6460 int sched_cpu_deactivate(unsigned int cpu
)
6464 set_cpu_active(cpu
, false);
6466 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6467 * users of this state to go away such that all new such users will
6470 * Do sync before park smpboot threads to take care the rcu boost case.
6474 #ifdef CONFIG_SCHED_SMT
6476 * When going down, decrement the number of cores with SMT present.
6478 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6479 static_branch_dec_cpuslocked(&sched_smt_present
);
6482 if (!sched_smp_initialized
)
6485 ret
= cpuset_cpu_inactive(cpu
);
6487 set_cpu_active(cpu
, true);
6490 sched_domains_numa_masks_clear(cpu
);
6494 static void sched_rq_cpu_starting(unsigned int cpu
)
6496 struct rq
*rq
= cpu_rq(cpu
);
6498 rq
->calc_load_update
= calc_load_update
;
6499 update_max_interval();
6502 int sched_cpu_starting(unsigned int cpu
)
6504 sched_rq_cpu_starting(cpu
);
6505 sched_tick_start(cpu
);
6509 #ifdef CONFIG_HOTPLUG_CPU
6510 int sched_cpu_dying(unsigned int cpu
)
6512 struct rq
*rq
= cpu_rq(cpu
);
6515 /* Handle pending wakeups and then migrate everything off */
6516 sched_ttwu_pending();
6517 sched_tick_stop(cpu
);
6519 rq_lock_irqsave(rq
, &rf
);
6521 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6524 migrate_tasks(rq
, &rf
);
6525 BUG_ON(rq
->nr_running
!= 1);
6526 rq_unlock_irqrestore(rq
, &rf
);
6528 calc_load_migrate(rq
);
6529 update_max_interval();
6530 nohz_balance_exit_idle(rq
);
6536 void __init
sched_init_smp(void)
6541 * There's no userspace yet to cause hotplug operations; hence all the
6542 * CPU masks are stable and all blatant races in the below code cannot
6545 mutex_lock(&sched_domains_mutex
);
6546 sched_init_domains(cpu_active_mask
);
6547 mutex_unlock(&sched_domains_mutex
);
6549 /* Move init over to a non-isolated CPU */
6550 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
6552 sched_init_granularity();
6554 init_sched_rt_class();
6555 init_sched_dl_class();
6557 sched_smp_initialized
= true;
6560 static int __init
migration_init(void)
6562 sched_cpu_starting(smp_processor_id());
6565 early_initcall(migration_init
);
6568 void __init
sched_init_smp(void)
6570 sched_init_granularity();
6572 #endif /* CONFIG_SMP */
6574 int in_sched_functions(unsigned long addr
)
6576 return in_lock_functions(addr
) ||
6577 (addr
>= (unsigned long)__sched_text_start
6578 && addr
< (unsigned long)__sched_text_end
);
6581 #ifdef CONFIG_CGROUP_SCHED
6583 * Default task group.
6584 * Every task in system belongs to this group at bootup.
6586 struct task_group root_task_group
;
6587 LIST_HEAD(task_groups
);
6589 /* Cacheline aligned slab cache for task_group */
6590 static struct kmem_cache
*task_group_cache __read_mostly
;
6593 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6594 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
6596 void __init
sched_init(void)
6598 unsigned long ptr
= 0;
6603 #ifdef CONFIG_FAIR_GROUP_SCHED
6604 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6606 #ifdef CONFIG_RT_GROUP_SCHED
6607 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6610 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
6612 #ifdef CONFIG_FAIR_GROUP_SCHED
6613 root_task_group
.se
= (struct sched_entity
**)ptr
;
6614 ptr
+= nr_cpu_ids
* sizeof(void **);
6616 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6617 ptr
+= nr_cpu_ids
* sizeof(void **);
6619 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6620 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6621 #endif /* CONFIG_FAIR_GROUP_SCHED */
6622 #ifdef CONFIG_RT_GROUP_SCHED
6623 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6624 ptr
+= nr_cpu_ids
* sizeof(void **);
6626 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6627 ptr
+= nr_cpu_ids
* sizeof(void **);
6629 #endif /* CONFIG_RT_GROUP_SCHED */
6631 #ifdef CONFIG_CPUMASK_OFFSTACK
6632 for_each_possible_cpu(i
) {
6633 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6634 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6635 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6636 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6638 #endif /* CONFIG_CPUMASK_OFFSTACK */
6640 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
6641 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6644 init_defrootdomain();
6647 #ifdef CONFIG_RT_GROUP_SCHED
6648 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6649 global_rt_period(), global_rt_runtime());
6650 #endif /* CONFIG_RT_GROUP_SCHED */
6652 #ifdef CONFIG_CGROUP_SCHED
6653 task_group_cache
= KMEM_CACHE(task_group
, 0);
6655 list_add(&root_task_group
.list
, &task_groups
);
6656 INIT_LIST_HEAD(&root_task_group
.children
);
6657 INIT_LIST_HEAD(&root_task_group
.siblings
);
6658 autogroup_init(&init_task
);
6659 #endif /* CONFIG_CGROUP_SCHED */
6661 for_each_possible_cpu(i
) {
6665 raw_spin_lock_init(&rq
->lock
);
6667 rq
->calc_load_active
= 0;
6668 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6669 init_cfs_rq(&rq
->cfs
);
6670 init_rt_rq(&rq
->rt
);
6671 init_dl_rq(&rq
->dl
);
6672 #ifdef CONFIG_FAIR_GROUP_SCHED
6673 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6674 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6676 * How much CPU bandwidth does root_task_group get?
6678 * In case of task-groups formed thr' the cgroup filesystem, it
6679 * gets 100% of the CPU resources in the system. This overall
6680 * system CPU resource is divided among the tasks of
6681 * root_task_group and its child task-groups in a fair manner,
6682 * based on each entity's (task or task-group's) weight
6683 * (se->load.weight).
6685 * In other words, if root_task_group has 10 tasks of weight
6686 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6687 * then A0's share of the CPU resource is:
6689 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6691 * We achieve this by letting root_task_group's tasks sit
6692 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6694 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6695 #endif /* CONFIG_FAIR_GROUP_SCHED */
6697 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6698 #ifdef CONFIG_RT_GROUP_SCHED
6699 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6704 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6705 rq
->balance_callback
= NULL
;
6706 rq
->active_balance
= 0;
6707 rq
->next_balance
= jiffies
;
6712 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6713 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6715 rq_csd_init(rq
, &rq
->wake_csd
, wake_csd_func
);
6717 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6719 rq_attach_root(rq
, &def_root_domain
);
6720 #ifdef CONFIG_NO_HZ_COMMON
6721 rq
->last_blocked_load_update_tick
= jiffies
;
6722 atomic_set(&rq
->nohz_flags
, 0);
6724 rq_csd_init(rq
, &rq
->nohz_csd
, nohz_csd_func
);
6726 #endif /* CONFIG_SMP */
6728 atomic_set(&rq
->nr_iowait
, 0);
6731 set_load_weight(&init_task
, false);
6734 * The boot idle thread does lazy MMU switching as well:
6737 enter_lazy_tlb(&init_mm
, current
);
6740 * Make us the idle thread. Technically, schedule() should not be
6741 * called from this thread, however somewhere below it might be,
6742 * but because we are the idle thread, we just pick up running again
6743 * when this runqueue becomes "idle".
6745 init_idle(current
, smp_processor_id());
6747 calc_load_update
= jiffies
+ LOAD_FREQ
;
6750 idle_thread_set_boot_cpu();
6752 init_sched_fair_class();
6760 scheduler_running
= 1;
6763 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6764 static inline int preempt_count_equals(int preempt_offset
)
6766 int nested
= preempt_count() + rcu_preempt_depth();
6768 return (nested
== preempt_offset
);
6771 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6774 * Blocking primitives will set (and therefore destroy) current->state,
6775 * since we will exit with TASK_RUNNING make sure we enter with it,
6776 * otherwise we will destroy state.
6778 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6779 "do not call blocking ops when !TASK_RUNNING; "
6780 "state=%lx set at [<%p>] %pS\n",
6782 (void *)current
->task_state_change
,
6783 (void *)current
->task_state_change
);
6785 ___might_sleep(file
, line
, preempt_offset
);
6787 EXPORT_SYMBOL(__might_sleep
);
6789 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6791 /* Ratelimiting timestamp: */
6792 static unsigned long prev_jiffy
;
6794 unsigned long preempt_disable_ip
;
6796 /* WARN_ON_ONCE() by default, no rate limit required: */
6799 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6800 !is_idle_task(current
) && !current
->non_block_count
) ||
6801 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6805 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6807 prev_jiffy
= jiffies
;
6809 /* Save this before calling printk(), since that will clobber it: */
6810 preempt_disable_ip
= get_preempt_disable_ip(current
);
6813 "BUG: sleeping function called from invalid context at %s:%d\n",
6816 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6817 in_atomic(), irqs_disabled(), current
->non_block_count
,
6818 current
->pid
, current
->comm
);
6820 if (task_stack_end_corrupted(current
))
6821 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6823 debug_show_held_locks(current
);
6824 if (irqs_disabled())
6825 print_irqtrace_events(current
);
6826 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6827 && !preempt_count_equals(preempt_offset
)) {
6828 pr_err("Preemption disabled at:");
6829 print_ip_sym(preempt_disable_ip
);
6833 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6835 EXPORT_SYMBOL(___might_sleep
);
6837 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
6839 static unsigned long prev_jiffy
;
6841 if (irqs_disabled())
6844 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
6847 if (preempt_count() > preempt_offset
)
6850 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6852 prev_jiffy
= jiffies
;
6854 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
6855 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6856 in_atomic(), irqs_disabled(),
6857 current
->pid
, current
->comm
);
6859 debug_show_held_locks(current
);
6861 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6863 EXPORT_SYMBOL_GPL(__cant_sleep
);
6866 #ifdef CONFIG_MAGIC_SYSRQ
6867 void normalize_rt_tasks(void)
6869 struct task_struct
*g
, *p
;
6870 struct sched_attr attr
= {
6871 .sched_policy
= SCHED_NORMAL
,
6874 read_lock(&tasklist_lock
);
6875 for_each_process_thread(g
, p
) {
6877 * Only normalize user tasks:
6879 if (p
->flags
& PF_KTHREAD
)
6882 p
->se
.exec_start
= 0;
6883 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6884 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6885 schedstat_set(p
->se
.statistics
.block_start
, 0);
6887 if (!dl_task(p
) && !rt_task(p
)) {
6889 * Renice negative nice level userspace
6892 if (task_nice(p
) < 0)
6893 set_user_nice(p
, 0);
6897 __sched_setscheduler(p
, &attr
, false, false);
6899 read_unlock(&tasklist_lock
);
6902 #endif /* CONFIG_MAGIC_SYSRQ */
6904 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6906 * These functions are only useful for the IA64 MCA handling, or kdb.
6908 * They can only be called when the whole system has been
6909 * stopped - every CPU needs to be quiescent, and no scheduling
6910 * activity can take place. Using them for anything else would
6911 * be a serious bug, and as a result, they aren't even visible
6912 * under any other configuration.
6916 * curr_task - return the current task for a given CPU.
6917 * @cpu: the processor in question.
6919 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6921 * Return: The current task for @cpu.
6923 struct task_struct
*curr_task(int cpu
)
6925 return cpu_curr(cpu
);
6928 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6932 * ia64_set_curr_task - set the current task for a given CPU.
6933 * @cpu: the processor in question.
6934 * @p: the task pointer to set.
6936 * Description: This function must only be used when non-maskable interrupts
6937 * are serviced on a separate stack. It allows the architecture to switch the
6938 * notion of the current task on a CPU in a non-blocking manner. This function
6939 * must be called with all CPU's synchronized, and interrupts disabled, the
6940 * and caller must save the original value of the current task (see
6941 * curr_task() above) and restore that value before reenabling interrupts and
6942 * re-starting the system.
6944 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6946 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6953 #ifdef CONFIG_CGROUP_SCHED
6954 /* task_group_lock serializes the addition/removal of task groups */
6955 static DEFINE_SPINLOCK(task_group_lock
);
6957 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
6958 struct task_group
*parent
)
6960 #ifdef CONFIG_UCLAMP_TASK_GROUP
6961 enum uclamp_id clamp_id
;
6963 for_each_clamp_id(clamp_id
) {
6964 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
6965 uclamp_none(clamp_id
), false);
6966 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
6971 static void sched_free_group(struct task_group
*tg
)
6973 free_fair_sched_group(tg
);
6974 free_rt_sched_group(tg
);
6976 kmem_cache_free(task_group_cache
, tg
);
6979 /* allocate runqueue etc for a new task group */
6980 struct task_group
*sched_create_group(struct task_group
*parent
)
6982 struct task_group
*tg
;
6984 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6986 return ERR_PTR(-ENOMEM
);
6988 if (!alloc_fair_sched_group(tg
, parent
))
6991 if (!alloc_rt_sched_group(tg
, parent
))
6994 alloc_uclamp_sched_group(tg
, parent
);
6999 sched_free_group(tg
);
7000 return ERR_PTR(-ENOMEM
);
7003 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7005 unsigned long flags
;
7007 spin_lock_irqsave(&task_group_lock
, flags
);
7008 list_add_rcu(&tg
->list
, &task_groups
);
7010 /* Root should already exist: */
7013 tg
->parent
= parent
;
7014 INIT_LIST_HEAD(&tg
->children
);
7015 list_add_rcu(&tg
->siblings
, &parent
->children
);
7016 spin_unlock_irqrestore(&task_group_lock
, flags
);
7018 online_fair_sched_group(tg
);
7021 /* rcu callback to free various structures associated with a task group */
7022 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7024 /* Now it should be safe to free those cfs_rqs: */
7025 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7028 void sched_destroy_group(struct task_group
*tg
)
7030 /* Wait for possible concurrent references to cfs_rqs complete: */
7031 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7034 void sched_offline_group(struct task_group
*tg
)
7036 unsigned long flags
;
7038 /* End participation in shares distribution: */
7039 unregister_fair_sched_group(tg
);
7041 spin_lock_irqsave(&task_group_lock
, flags
);
7042 list_del_rcu(&tg
->list
);
7043 list_del_rcu(&tg
->siblings
);
7044 spin_unlock_irqrestore(&task_group_lock
, flags
);
7047 static void sched_change_group(struct task_struct
*tsk
, int type
)
7049 struct task_group
*tg
;
7052 * All callers are synchronized by task_rq_lock(); we do not use RCU
7053 * which is pointless here. Thus, we pass "true" to task_css_check()
7054 * to prevent lockdep warnings.
7056 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7057 struct task_group
, css
);
7058 tg
= autogroup_task_group(tsk
, tg
);
7059 tsk
->sched_task_group
= tg
;
7061 #ifdef CONFIG_FAIR_GROUP_SCHED
7062 if (tsk
->sched_class
->task_change_group
)
7063 tsk
->sched_class
->task_change_group(tsk
, type
);
7066 set_task_rq(tsk
, task_cpu(tsk
));
7070 * Change task's runqueue when it moves between groups.
7072 * The caller of this function should have put the task in its new group by
7073 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7076 void sched_move_task(struct task_struct
*tsk
)
7078 int queued
, running
, queue_flags
=
7079 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7083 rq
= task_rq_lock(tsk
, &rf
);
7084 update_rq_clock(rq
);
7086 running
= task_current(rq
, tsk
);
7087 queued
= task_on_rq_queued(tsk
);
7090 dequeue_task(rq
, tsk
, queue_flags
);
7092 put_prev_task(rq
, tsk
);
7094 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7097 enqueue_task(rq
, tsk
, queue_flags
);
7099 set_next_task(rq
, tsk
);
7101 * After changing group, the running task may have joined a
7102 * throttled one but it's still the running task. Trigger a
7103 * resched to make sure that task can still run.
7108 task_rq_unlock(rq
, tsk
, &rf
);
7111 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7113 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7116 static struct cgroup_subsys_state
*
7117 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7119 struct task_group
*parent
= css_tg(parent_css
);
7120 struct task_group
*tg
;
7123 /* This is early initialization for the top cgroup */
7124 return &root_task_group
.css
;
7127 tg
= sched_create_group(parent
);
7129 return ERR_PTR(-ENOMEM
);
7134 /* Expose task group only after completing cgroup initialization */
7135 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7137 struct task_group
*tg
= css_tg(css
);
7138 struct task_group
*parent
= css_tg(css
->parent
);
7141 sched_online_group(tg
, parent
);
7143 #ifdef CONFIG_UCLAMP_TASK_GROUP
7144 /* Propagate the effective uclamp value for the new group */
7145 cpu_util_update_eff(css
);
7151 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
7153 struct task_group
*tg
= css_tg(css
);
7155 sched_offline_group(tg
);
7158 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7160 struct task_group
*tg
= css_tg(css
);
7163 * Relies on the RCU grace period between css_released() and this.
7165 sched_free_group(tg
);
7169 * This is called before wake_up_new_task(), therefore we really only
7170 * have to set its group bits, all the other stuff does not apply.
7172 static void cpu_cgroup_fork(struct task_struct
*task
)
7177 rq
= task_rq_lock(task
, &rf
);
7179 update_rq_clock(rq
);
7180 sched_change_group(task
, TASK_SET_GROUP
);
7182 task_rq_unlock(rq
, task
, &rf
);
7185 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
7187 struct task_struct
*task
;
7188 struct cgroup_subsys_state
*css
;
7191 cgroup_taskset_for_each(task
, css
, tset
) {
7192 #ifdef CONFIG_RT_GROUP_SCHED
7193 if (!sched_rt_can_attach(css_tg(css
), task
))
7197 * Serialize against wake_up_new_task() such that if its
7198 * running, we're sure to observe its full state.
7200 raw_spin_lock_irq(&task
->pi_lock
);
7202 * Avoid calling sched_move_task() before wake_up_new_task()
7203 * has happened. This would lead to problems with PELT, due to
7204 * move wanting to detach+attach while we're not attached yet.
7206 if (task
->state
== TASK_NEW
)
7208 raw_spin_unlock_irq(&task
->pi_lock
);
7216 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
7218 struct task_struct
*task
;
7219 struct cgroup_subsys_state
*css
;
7221 cgroup_taskset_for_each(task
, css
, tset
)
7222 sched_move_task(task
);
7225 #ifdef CONFIG_UCLAMP_TASK_GROUP
7226 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
7228 struct cgroup_subsys_state
*top_css
= css
;
7229 struct uclamp_se
*uc_parent
= NULL
;
7230 struct uclamp_se
*uc_se
= NULL
;
7231 unsigned int eff
[UCLAMP_CNT
];
7232 enum uclamp_id clamp_id
;
7233 unsigned int clamps
;
7235 css_for_each_descendant_pre(css
, top_css
) {
7236 uc_parent
= css_tg(css
)->parent
7237 ? css_tg(css
)->parent
->uclamp
: NULL
;
7239 for_each_clamp_id(clamp_id
) {
7240 /* Assume effective clamps matches requested clamps */
7241 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
7242 /* Cap effective clamps with parent's effective clamps */
7244 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
7245 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
7248 /* Ensure protection is always capped by limit */
7249 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
7251 /* Propagate most restrictive effective clamps */
7253 uc_se
= css_tg(css
)->uclamp
;
7254 for_each_clamp_id(clamp_id
) {
7255 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
7257 uc_se
[clamp_id
].value
= eff
[clamp_id
];
7258 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
7259 clamps
|= (0x1 << clamp_id
);
7262 css
= css_rightmost_descendant(css
);
7266 /* Immediately update descendants RUNNABLE tasks */
7267 uclamp_update_active_tasks(css
, clamps
);
7272 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7273 * C expression. Since there is no way to convert a macro argument (N) into a
7274 * character constant, use two levels of macros.
7276 #define _POW10(exp) ((unsigned int)1e##exp)
7277 #define POW10(exp) _POW10(exp)
7279 struct uclamp_request
{
7280 #define UCLAMP_PERCENT_SHIFT 2
7281 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7287 static inline struct uclamp_request
7288 capacity_from_percent(char *buf
)
7290 struct uclamp_request req
= {
7291 .percent
= UCLAMP_PERCENT_SCALE
,
7292 .util
= SCHED_CAPACITY_SCALE
,
7297 if (strcmp(buf
, "max")) {
7298 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
7302 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
7307 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
7308 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
7314 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
7315 size_t nbytes
, loff_t off
,
7316 enum uclamp_id clamp_id
)
7318 struct uclamp_request req
;
7319 struct task_group
*tg
;
7321 req
= capacity_from_percent(buf
);
7325 mutex_lock(&uclamp_mutex
);
7328 tg
= css_tg(of_css(of
));
7329 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
7330 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
7333 * Because of not recoverable conversion rounding we keep track of the
7334 * exact requested value
7336 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
7338 /* Update effective clamps to track the most restrictive value */
7339 cpu_util_update_eff(of_css(of
));
7342 mutex_unlock(&uclamp_mutex
);
7347 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
7348 char *buf
, size_t nbytes
,
7351 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
7354 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
7355 char *buf
, size_t nbytes
,
7358 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
7361 static inline void cpu_uclamp_print(struct seq_file
*sf
,
7362 enum uclamp_id clamp_id
)
7364 struct task_group
*tg
;
7370 tg
= css_tg(seq_css(sf
));
7371 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
7374 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
7375 seq_puts(sf
, "max\n");
7379 percent
= tg
->uclamp_pct
[clamp_id
];
7380 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
7381 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
7384 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
7386 cpu_uclamp_print(sf
, UCLAMP_MIN
);
7390 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
7392 cpu_uclamp_print(sf
, UCLAMP_MAX
);
7395 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7397 #ifdef CONFIG_FAIR_GROUP_SCHED
7398 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7399 struct cftype
*cftype
, u64 shareval
)
7401 if (shareval
> scale_load_down(ULONG_MAX
))
7402 shareval
= MAX_SHARES
;
7403 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7406 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7409 struct task_group
*tg
= css_tg(css
);
7411 return (u64
) scale_load_down(tg
->shares
);
7414 #ifdef CONFIG_CFS_BANDWIDTH
7415 static DEFINE_MUTEX(cfs_constraints_mutex
);
7417 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7418 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7419 /* More than 203 days if BW_SHIFT equals 20. */
7420 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
7422 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7424 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7426 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7427 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7429 if (tg
== &root_task_group
)
7433 * Ensure we have at some amount of bandwidth every period. This is
7434 * to prevent reaching a state of large arrears when throttled via
7435 * entity_tick() resulting in prolonged exit starvation.
7437 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7441 * Likewise, bound things on the otherside by preventing insane quota
7442 * periods. This also allows us to normalize in computing quota
7445 if (period
> max_cfs_quota_period
)
7449 * Bound quota to defend quota against overflow during bandwidth shift.
7451 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
7455 * Prevent race between setting of cfs_rq->runtime_enabled and
7456 * unthrottle_offline_cfs_rqs().
7459 mutex_lock(&cfs_constraints_mutex
);
7460 ret
= __cfs_schedulable(tg
, period
, quota
);
7464 runtime_enabled
= quota
!= RUNTIME_INF
;
7465 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7467 * If we need to toggle cfs_bandwidth_used, off->on must occur
7468 * before making related changes, and on->off must occur afterwards
7470 if (runtime_enabled
&& !runtime_was_enabled
)
7471 cfs_bandwidth_usage_inc();
7472 raw_spin_lock_irq(&cfs_b
->lock
);
7473 cfs_b
->period
= ns_to_ktime(period
);
7474 cfs_b
->quota
= quota
;
7476 __refill_cfs_bandwidth_runtime(cfs_b
);
7478 /* Restart the period timer (if active) to handle new period expiry: */
7479 if (runtime_enabled
)
7480 start_cfs_bandwidth(cfs_b
);
7482 raw_spin_unlock_irq(&cfs_b
->lock
);
7484 for_each_online_cpu(i
) {
7485 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7486 struct rq
*rq
= cfs_rq
->rq
;
7489 rq_lock_irq(rq
, &rf
);
7490 cfs_rq
->runtime_enabled
= runtime_enabled
;
7491 cfs_rq
->runtime_remaining
= 0;
7493 if (cfs_rq
->throttled
)
7494 unthrottle_cfs_rq(cfs_rq
);
7495 rq_unlock_irq(rq
, &rf
);
7497 if (runtime_was_enabled
&& !runtime_enabled
)
7498 cfs_bandwidth_usage_dec();
7500 mutex_unlock(&cfs_constraints_mutex
);
7506 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7510 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7511 if (cfs_quota_us
< 0)
7512 quota
= RUNTIME_INF
;
7513 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
7514 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7518 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7521 static long tg_get_cfs_quota(struct task_group
*tg
)
7525 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7528 quota_us
= tg
->cfs_bandwidth
.quota
;
7529 do_div(quota_us
, NSEC_PER_USEC
);
7534 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7538 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
7541 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7542 quota
= tg
->cfs_bandwidth
.quota
;
7544 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7547 static long tg_get_cfs_period(struct task_group
*tg
)
7551 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7552 do_div(cfs_period_us
, NSEC_PER_USEC
);
7554 return cfs_period_us
;
7557 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7560 return tg_get_cfs_quota(css_tg(css
));
7563 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7564 struct cftype
*cftype
, s64 cfs_quota_us
)
7566 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7569 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7572 return tg_get_cfs_period(css_tg(css
));
7575 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7576 struct cftype
*cftype
, u64 cfs_period_us
)
7578 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7581 struct cfs_schedulable_data
{
7582 struct task_group
*tg
;
7587 * normalize group quota/period to be quota/max_period
7588 * note: units are usecs
7590 static u64
normalize_cfs_quota(struct task_group
*tg
,
7591 struct cfs_schedulable_data
*d
)
7599 period
= tg_get_cfs_period(tg
);
7600 quota
= tg_get_cfs_quota(tg
);
7603 /* note: these should typically be equivalent */
7604 if (quota
== RUNTIME_INF
|| quota
== -1)
7607 return to_ratio(period
, quota
);
7610 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7612 struct cfs_schedulable_data
*d
= data
;
7613 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7614 s64 quota
= 0, parent_quota
= -1;
7617 quota
= RUNTIME_INF
;
7619 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7621 quota
= normalize_cfs_quota(tg
, d
);
7622 parent_quota
= parent_b
->hierarchical_quota
;
7625 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7626 * always take the min. On cgroup1, only inherit when no
7629 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
7630 quota
= min(quota
, parent_quota
);
7632 if (quota
== RUNTIME_INF
)
7633 quota
= parent_quota
;
7634 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7638 cfs_b
->hierarchical_quota
= quota
;
7643 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7646 struct cfs_schedulable_data data
= {
7652 if (quota
!= RUNTIME_INF
) {
7653 do_div(data
.period
, NSEC_PER_USEC
);
7654 do_div(data
.quota
, NSEC_PER_USEC
);
7658 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7664 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
7666 struct task_group
*tg
= css_tg(seq_css(sf
));
7667 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7669 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7670 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7671 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7673 if (schedstat_enabled() && tg
!= &root_task_group
) {
7677 for_each_possible_cpu(i
)
7678 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
7680 seq_printf(sf
, "wait_sum %llu\n", ws
);
7685 #endif /* CONFIG_CFS_BANDWIDTH */
7686 #endif /* CONFIG_FAIR_GROUP_SCHED */
7688 #ifdef CONFIG_RT_GROUP_SCHED
7689 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7690 struct cftype
*cft
, s64 val
)
7692 return sched_group_set_rt_runtime(css_tg(css
), val
);
7695 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7698 return sched_group_rt_runtime(css_tg(css
));
7701 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7702 struct cftype
*cftype
, u64 rt_period_us
)
7704 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7707 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7710 return sched_group_rt_period(css_tg(css
));
7712 #endif /* CONFIG_RT_GROUP_SCHED */
7714 static struct cftype cpu_legacy_files
[] = {
7715 #ifdef CONFIG_FAIR_GROUP_SCHED
7718 .read_u64
= cpu_shares_read_u64
,
7719 .write_u64
= cpu_shares_write_u64
,
7722 #ifdef CONFIG_CFS_BANDWIDTH
7724 .name
= "cfs_quota_us",
7725 .read_s64
= cpu_cfs_quota_read_s64
,
7726 .write_s64
= cpu_cfs_quota_write_s64
,
7729 .name
= "cfs_period_us",
7730 .read_u64
= cpu_cfs_period_read_u64
,
7731 .write_u64
= cpu_cfs_period_write_u64
,
7735 .seq_show
= cpu_cfs_stat_show
,
7738 #ifdef CONFIG_RT_GROUP_SCHED
7740 .name
= "rt_runtime_us",
7741 .read_s64
= cpu_rt_runtime_read
,
7742 .write_s64
= cpu_rt_runtime_write
,
7745 .name
= "rt_period_us",
7746 .read_u64
= cpu_rt_period_read_uint
,
7747 .write_u64
= cpu_rt_period_write_uint
,
7750 #ifdef CONFIG_UCLAMP_TASK_GROUP
7752 .name
= "uclamp.min",
7753 .flags
= CFTYPE_NOT_ON_ROOT
,
7754 .seq_show
= cpu_uclamp_min_show
,
7755 .write
= cpu_uclamp_min_write
,
7758 .name
= "uclamp.max",
7759 .flags
= CFTYPE_NOT_ON_ROOT
,
7760 .seq_show
= cpu_uclamp_max_show
,
7761 .write
= cpu_uclamp_max_write
,
7767 static int cpu_extra_stat_show(struct seq_file
*sf
,
7768 struct cgroup_subsys_state
*css
)
7770 #ifdef CONFIG_CFS_BANDWIDTH
7772 struct task_group
*tg
= css_tg(css
);
7773 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7776 throttled_usec
= cfs_b
->throttled_time
;
7777 do_div(throttled_usec
, NSEC_PER_USEC
);
7779 seq_printf(sf
, "nr_periods %d\n"
7781 "throttled_usec %llu\n",
7782 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
7789 #ifdef CONFIG_FAIR_GROUP_SCHED
7790 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
7793 struct task_group
*tg
= css_tg(css
);
7794 u64 weight
= scale_load_down(tg
->shares
);
7796 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
7799 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
7800 struct cftype
*cft
, u64 weight
)
7803 * cgroup weight knobs should use the common MIN, DFL and MAX
7804 * values which are 1, 100 and 10000 respectively. While it loses
7805 * a bit of range on both ends, it maps pretty well onto the shares
7806 * value used by scheduler and the round-trip conversions preserve
7807 * the original value over the entire range.
7809 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
7812 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
7814 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7817 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
7820 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
7821 int last_delta
= INT_MAX
;
7824 /* find the closest nice value to the current weight */
7825 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
7826 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
7827 if (delta
>= last_delta
)
7832 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
7835 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
7836 struct cftype
*cft
, s64 nice
)
7838 unsigned long weight
;
7841 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
7844 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
7845 idx
= array_index_nospec(idx
, 40);
7846 weight
= sched_prio_to_weight
[idx
];
7848 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7852 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
7853 long period
, long quota
)
7856 seq_puts(sf
, "max");
7858 seq_printf(sf
, "%ld", quota
);
7860 seq_printf(sf
, " %ld\n", period
);
7863 /* caller should put the current value in *@periodp before calling */
7864 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
7865 u64
*periodp
, u64
*quotap
)
7867 char tok
[21]; /* U64_MAX */
7869 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
7872 *periodp
*= NSEC_PER_USEC
;
7874 if (sscanf(tok
, "%llu", quotap
))
7875 *quotap
*= NSEC_PER_USEC
;
7876 else if (!strcmp(tok
, "max"))
7877 *quotap
= RUNTIME_INF
;
7884 #ifdef CONFIG_CFS_BANDWIDTH
7885 static int cpu_max_show(struct seq_file
*sf
, void *v
)
7887 struct task_group
*tg
= css_tg(seq_css(sf
));
7889 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
7893 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
7894 char *buf
, size_t nbytes
, loff_t off
)
7896 struct task_group
*tg
= css_tg(of_css(of
));
7897 u64 period
= tg_get_cfs_period(tg
);
7901 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
7903 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
7904 return ret
?: nbytes
;
7908 static struct cftype cpu_files
[] = {
7909 #ifdef CONFIG_FAIR_GROUP_SCHED
7912 .flags
= CFTYPE_NOT_ON_ROOT
,
7913 .read_u64
= cpu_weight_read_u64
,
7914 .write_u64
= cpu_weight_write_u64
,
7917 .name
= "weight.nice",
7918 .flags
= CFTYPE_NOT_ON_ROOT
,
7919 .read_s64
= cpu_weight_nice_read_s64
,
7920 .write_s64
= cpu_weight_nice_write_s64
,
7923 #ifdef CONFIG_CFS_BANDWIDTH
7926 .flags
= CFTYPE_NOT_ON_ROOT
,
7927 .seq_show
= cpu_max_show
,
7928 .write
= cpu_max_write
,
7931 #ifdef CONFIG_UCLAMP_TASK_GROUP
7933 .name
= "uclamp.min",
7934 .flags
= CFTYPE_NOT_ON_ROOT
,
7935 .seq_show
= cpu_uclamp_min_show
,
7936 .write
= cpu_uclamp_min_write
,
7939 .name
= "uclamp.max",
7940 .flags
= CFTYPE_NOT_ON_ROOT
,
7941 .seq_show
= cpu_uclamp_max_show
,
7942 .write
= cpu_uclamp_max_write
,
7948 struct cgroup_subsys cpu_cgrp_subsys
= {
7949 .css_alloc
= cpu_cgroup_css_alloc
,
7950 .css_online
= cpu_cgroup_css_online
,
7951 .css_released
= cpu_cgroup_css_released
,
7952 .css_free
= cpu_cgroup_css_free
,
7953 .css_extra_stat_show
= cpu_extra_stat_show
,
7954 .fork
= cpu_cgroup_fork
,
7955 .can_attach
= cpu_cgroup_can_attach
,
7956 .attach
= cpu_cgroup_attach
,
7957 .legacy_cftypes
= cpu_legacy_files
,
7958 .dfl_cftypes
= cpu_files
,
7963 #endif /* CONFIG_CGROUP_SCHED */
7965 void dump_cpu_task(int cpu
)
7967 pr_info("Task dump for CPU %d:\n", cpu
);
7968 sched_show_task(cpu_curr(cpu
));
7972 * Nice levels are multiplicative, with a gentle 10% change for every
7973 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7974 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7975 * that remained on nice 0.
7977 * The "10% effect" is relative and cumulative: from _any_ nice level,
7978 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7979 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7980 * If a task goes up by ~10% and another task goes down by ~10% then
7981 * the relative distance between them is ~25%.)
7983 const int sched_prio_to_weight
[40] = {
7984 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7985 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7986 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7987 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7988 /* 0 */ 1024, 820, 655, 526, 423,
7989 /* 5 */ 335, 272, 215, 172, 137,
7990 /* 10 */ 110, 87, 70, 56, 45,
7991 /* 15 */ 36, 29, 23, 18, 15,
7995 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7997 * In cases where the weight does not change often, we can use the
7998 * precalculated inverse to speed up arithmetics by turning divisions
7999 * into multiplications:
8001 const u32 sched_prio_to_wmult
[40] = {
8002 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8003 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8004 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8005 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8006 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8007 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8008 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8009 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8012 #undef CREATE_TRACE_POINTS