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 #ifdef CONFIG_SCHED_HRTICK
225 * Use HR-timers to deliver accurate preemption points.
228 static void hrtick_clear(struct rq
*rq
)
230 if (hrtimer_active(&rq
->hrtick_timer
))
231 hrtimer_cancel(&rq
->hrtick_timer
);
235 * High-resolution timer tick.
236 * Runs from hardirq context with interrupts disabled.
238 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
240 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
243 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
247 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
250 return HRTIMER_NORESTART
;
255 static void __hrtick_restart(struct rq
*rq
)
257 struct hrtimer
*timer
= &rq
->hrtick_timer
;
259 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
263 * called from hardirq (IPI) context
265 static void __hrtick_start(void *arg
)
271 __hrtick_restart(rq
);
276 * Called to set the hrtick timer state.
278 * called with rq->lock held and irqs disabled
280 void hrtick_start(struct rq
*rq
, u64 delay
)
282 struct hrtimer
*timer
= &rq
->hrtick_timer
;
287 * Don't schedule slices shorter than 10000ns, that just
288 * doesn't make sense and can cause timer DoS.
290 delta
= max_t(s64
, delay
, 10000LL);
291 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
293 hrtimer_set_expires(timer
, time
);
296 __hrtick_restart(rq
);
298 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
303 * Called to set the hrtick timer state.
305 * called with rq->lock held and irqs disabled
307 void hrtick_start(struct rq
*rq
, u64 delay
)
310 * Don't schedule slices shorter than 10000ns, that just
311 * doesn't make sense. Rely on vruntime for fairness.
313 delay
= max_t(u64
, delay
, 10000LL);
314 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
315 HRTIMER_MODE_REL_PINNED_HARD
);
317 #endif /* CONFIG_SMP */
319 static void hrtick_rq_init(struct rq
*rq
)
322 rq
->hrtick_csd
.flags
= 0;
323 rq
->hrtick_csd
.func
= __hrtick_start
;
324 rq
->hrtick_csd
.info
= rq
;
327 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
328 rq
->hrtick_timer
.function
= hrtick
;
330 #else /* CONFIG_SCHED_HRTICK */
331 static inline void hrtick_clear(struct rq
*rq
)
335 static inline void hrtick_rq_init(struct rq
*rq
)
338 #endif /* CONFIG_SCHED_HRTICK */
341 * cmpxchg based fetch_or, macro so it works for different integer types
343 #define fetch_or(ptr, mask) \
345 typeof(ptr) _ptr = (ptr); \
346 typeof(mask) _mask = (mask); \
347 typeof(*_ptr) _old, _val = *_ptr; \
350 _old = cmpxchg(_ptr, _val, _val | _mask); \
358 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
360 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
361 * this avoids any races wrt polling state changes and thereby avoids
364 static bool set_nr_and_not_polling(struct task_struct
*p
)
366 struct thread_info
*ti
= task_thread_info(p
);
367 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
371 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
373 * If this returns true, then the idle task promises to call
374 * sched_ttwu_pending() and reschedule soon.
376 static bool set_nr_if_polling(struct task_struct
*p
)
378 struct thread_info
*ti
= task_thread_info(p
);
379 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
382 if (!(val
& _TIF_POLLING_NRFLAG
))
384 if (val
& _TIF_NEED_RESCHED
)
386 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
395 static bool set_nr_and_not_polling(struct task_struct
*p
)
397 set_tsk_need_resched(p
);
402 static bool set_nr_if_polling(struct task_struct
*p
)
409 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
411 struct wake_q_node
*node
= &task
->wake_q
;
414 * Atomically grab the task, if ->wake_q is !nil already it means
415 * its already queued (either by us or someone else) and will get the
416 * wakeup due to that.
418 * In order to ensure that a pending wakeup will observe our pending
419 * state, even in the failed case, an explicit smp_mb() must be used.
421 smp_mb__before_atomic();
422 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
426 * The head is context local, there can be no concurrency.
429 head
->lastp
= &node
->next
;
434 * wake_q_add() - queue a wakeup for 'later' waking.
435 * @head: the wake_q_head to add @task to
436 * @task: the task to queue for 'later' wakeup
438 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
439 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
442 * This function must be used as-if it were wake_up_process(); IOW the task
443 * must be ready to be woken at this location.
445 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
447 if (__wake_q_add(head
, task
))
448 get_task_struct(task
);
452 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
453 * @head: the wake_q_head to add @task to
454 * @task: the task to queue for 'later' wakeup
456 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
457 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
460 * This function must be used as-if it were wake_up_process(); IOW the task
461 * must be ready to be woken at this location.
463 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
464 * that already hold reference to @task can call the 'safe' version and trust
465 * wake_q to do the right thing depending whether or not the @task is already
468 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
470 if (!__wake_q_add(head
, task
))
471 put_task_struct(task
);
474 void wake_up_q(struct wake_q_head
*head
)
476 struct wake_q_node
*node
= head
->first
;
478 while (node
!= WAKE_Q_TAIL
) {
479 struct task_struct
*task
;
481 task
= container_of(node
, struct task_struct
, wake_q
);
483 /* Task can safely be re-inserted now: */
485 task
->wake_q
.next
= NULL
;
488 * wake_up_process() executes a full barrier, which pairs with
489 * the queueing in wake_q_add() so as not to miss wakeups.
491 wake_up_process(task
);
492 put_task_struct(task
);
497 * resched_curr - mark rq's current task 'to be rescheduled now'.
499 * On UP this means the setting of the need_resched flag, on SMP it
500 * might also involve a cross-CPU call to trigger the scheduler on
503 void resched_curr(struct rq
*rq
)
505 struct task_struct
*curr
= rq
->curr
;
508 lockdep_assert_held(&rq
->lock
);
510 if (test_tsk_need_resched(curr
))
515 if (cpu
== smp_processor_id()) {
516 set_tsk_need_resched(curr
);
517 set_preempt_need_resched();
521 if (set_nr_and_not_polling(curr
))
522 smp_send_reschedule(cpu
);
524 trace_sched_wake_idle_without_ipi(cpu
);
527 void resched_cpu(int cpu
)
529 struct rq
*rq
= cpu_rq(cpu
);
532 raw_spin_lock_irqsave(&rq
->lock
, flags
);
533 if (cpu_online(cpu
) || cpu
== smp_processor_id())
535 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
539 #ifdef CONFIG_NO_HZ_COMMON
541 * In the semi idle case, use the nearest busy CPU for migrating timers
542 * from an idle CPU. This is good for power-savings.
544 * We don't do similar optimization for completely idle system, as
545 * selecting an idle CPU will add more delays to the timers than intended
546 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
548 int get_nohz_timer_target(void)
550 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
551 struct sched_domain
*sd
;
553 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
560 for_each_domain(cpu
, sd
) {
561 for_each_cpu_and(i
, sched_domain_span(sd
),
562 housekeeping_cpumask(HK_FLAG_TIMER
)) {
573 if (default_cpu
== -1)
574 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
582 * When add_timer_on() enqueues a timer into the timer wheel of an
583 * idle CPU then this timer might expire before the next timer event
584 * which is scheduled to wake up that CPU. In case of a completely
585 * idle system the next event might even be infinite time into the
586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587 * leaves the inner idle loop so the newly added timer is taken into
588 * account when the CPU goes back to idle and evaluates the timer
589 * wheel for the next timer event.
591 static void wake_up_idle_cpu(int cpu
)
593 struct rq
*rq
= cpu_rq(cpu
);
595 if (cpu
== smp_processor_id())
598 if (set_nr_and_not_polling(rq
->idle
))
599 smp_send_reschedule(cpu
);
601 trace_sched_wake_idle_without_ipi(cpu
);
604 static bool wake_up_full_nohz_cpu(int cpu
)
607 * We just need the target to call irq_exit() and re-evaluate
608 * the next tick. The nohz full kick at least implies that.
609 * If needed we can still optimize that later with an
612 if (cpu_is_offline(cpu
))
613 return true; /* Don't try to wake offline CPUs. */
614 if (tick_nohz_full_cpu(cpu
)) {
615 if (cpu
!= smp_processor_id() ||
616 tick_nohz_tick_stopped())
617 tick_nohz_full_kick_cpu(cpu
);
625 * Wake up the specified CPU. If the CPU is going offline, it is the
626 * caller's responsibility to deal with the lost wakeup, for example,
627 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
629 void wake_up_nohz_cpu(int cpu
)
631 if (!wake_up_full_nohz_cpu(cpu
))
632 wake_up_idle_cpu(cpu
);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu
= smp_processor_id();
639 if (!(atomic_read(nohz_flags(cpu
)) & NOHZ_KICK_MASK
))
642 if (idle_cpu(cpu
) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 atomic_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(struct rq
*rq
)
667 /* Deadline tasks, even if single, need the tick */
668 if (rq
->dl
.dl_nr_running
)
672 * If there are more than one RR tasks, we need the tick to effect the
673 * actual RR behaviour.
675 if (rq
->rt
.rr_nr_running
) {
676 if (rq
->rt
.rr_nr_running
== 1)
683 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
684 * forced preemption between FIFO tasks.
686 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
691 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
692 * if there's more than one we need the tick for involuntary
695 if (rq
->nr_running
> 1)
700 #endif /* CONFIG_NO_HZ_FULL */
701 #endif /* CONFIG_SMP */
703 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
704 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
706 * Iterate task_group tree rooted at *from, calling @down when first entering a
707 * node and @up when leaving it for the final time.
709 * Caller must hold rcu_lock or sufficient equivalent.
711 int walk_tg_tree_from(struct task_group
*from
,
712 tg_visitor down
, tg_visitor up
, void *data
)
714 struct task_group
*parent
, *child
;
720 ret
= (*down
)(parent
, data
);
723 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
730 ret
= (*up
)(parent
, data
);
731 if (ret
|| parent
== from
)
735 parent
= parent
->parent
;
742 int tg_nop(struct task_group
*tg
, void *data
)
748 static void set_load_weight(struct task_struct
*p
, bool update_load
)
750 int prio
= p
->static_prio
- MAX_RT_PRIO
;
751 struct load_weight
*load
= &p
->se
.load
;
754 * SCHED_IDLE tasks get minimal weight:
756 if (task_has_idle_policy(p
)) {
757 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
758 load
->inv_weight
= WMULT_IDLEPRIO
;
763 * SCHED_OTHER tasks have to update their load when changing their
766 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
767 reweight_task(p
, prio
);
769 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
770 load
->inv_weight
= sched_prio_to_wmult
[prio
];
774 #ifdef CONFIG_UCLAMP_TASK
776 * Serializes updates of utilization clamp values
778 * The (slow-path) user-space triggers utilization clamp value updates which
779 * can require updates on (fast-path) scheduler's data structures used to
780 * support enqueue/dequeue operations.
781 * While the per-CPU rq lock protects fast-path update operations, user-space
782 * requests are serialized using a mutex to reduce the risk of conflicting
783 * updates or API abuses.
785 static DEFINE_MUTEX(uclamp_mutex
);
787 /* Max allowed minimum utilization */
788 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
790 /* Max allowed maximum utilization */
791 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
793 /* All clamps are required to be less or equal than these values */
794 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
796 /* Integer rounded range for each bucket */
797 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
799 #define for_each_clamp_id(clamp_id) \
800 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
802 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
804 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
807 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value
)
809 return UCLAMP_BUCKET_DELTA
* uclamp_bucket_id(clamp_value
);
812 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
814 if (clamp_id
== UCLAMP_MIN
)
816 return SCHED_CAPACITY_SCALE
;
819 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
820 unsigned int value
, bool user_defined
)
822 uc_se
->value
= value
;
823 uc_se
->bucket_id
= uclamp_bucket_id(value
);
824 uc_se
->user_defined
= user_defined
;
827 static inline unsigned int
828 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
829 unsigned int clamp_value
)
832 * Avoid blocked utilization pushing up the frequency when we go
833 * idle (which drops the max-clamp) by retaining the last known
836 if (clamp_id
== UCLAMP_MAX
) {
837 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
841 return uclamp_none(UCLAMP_MIN
);
844 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
845 unsigned int clamp_value
)
847 /* Reset max-clamp retention only on idle exit */
848 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
851 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
855 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
856 unsigned int clamp_value
)
858 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
859 int bucket_id
= UCLAMP_BUCKETS
- 1;
862 * Since both min and max clamps are max aggregated, find the
863 * top most bucket with tasks in.
865 for ( ; bucket_id
>= 0; bucket_id
--) {
866 if (!bucket
[bucket_id
].tasks
)
868 return bucket
[bucket_id
].value
;
871 /* No tasks -- default clamp values */
872 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
875 static inline struct uclamp_se
876 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
878 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
879 #ifdef CONFIG_UCLAMP_TASK_GROUP
880 struct uclamp_se uc_max
;
883 * Tasks in autogroups or root task group will be
884 * restricted by system defaults.
886 if (task_group_is_autogroup(task_group(p
)))
888 if (task_group(p
) == &root_task_group
)
891 uc_max
= task_group(p
)->uclamp
[clamp_id
];
892 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
900 * The effective clamp bucket index of a task depends on, by increasing
902 * - the task specific clamp value, when explicitly requested from userspace
903 * - the task group effective clamp value, for tasks not either in the root
904 * group or in an autogroup
905 * - the system default clamp value, defined by the sysadmin
907 static inline struct uclamp_se
908 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
910 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
911 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
913 /* System default restrictions always apply */
914 if (unlikely(uc_req
.value
> uc_max
.value
))
920 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
922 struct uclamp_se uc_eff
;
924 /* Task currently refcounted: use back-annotated (effective) value */
925 if (p
->uclamp
[clamp_id
].active
)
926 return (unsigned long)p
->uclamp
[clamp_id
].value
;
928 uc_eff
= uclamp_eff_get(p
, clamp_id
);
930 return (unsigned long)uc_eff
.value
;
934 * When a task is enqueued on a rq, the clamp bucket currently defined by the
935 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
936 * updates the rq's clamp value if required.
938 * Tasks can have a task-specific value requested from user-space, track
939 * within each bucket the maximum value for tasks refcounted in it.
940 * This "local max aggregation" allows to track the exact "requested" value
941 * for each bucket when all its RUNNABLE tasks require the same clamp.
943 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
944 enum uclamp_id clamp_id
)
946 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
947 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
948 struct uclamp_bucket
*bucket
;
950 lockdep_assert_held(&rq
->lock
);
952 /* Update task effective clamp */
953 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
955 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
957 uc_se
->active
= true;
959 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
962 * Local max aggregation: rq buckets always track the max
963 * "requested" clamp value of its RUNNABLE tasks.
965 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
966 bucket
->value
= uc_se
->value
;
968 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
969 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
973 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
974 * is released. If this is the last task reference counting the rq's max
975 * active clamp value, then the rq's clamp value is updated.
977 * Both refcounted tasks and rq's cached clamp values are expected to be
978 * always valid. If it's detected they are not, as defensive programming,
979 * enforce the expected state and warn.
981 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
982 enum uclamp_id clamp_id
)
984 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
985 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
986 struct uclamp_bucket
*bucket
;
987 unsigned int bkt_clamp
;
988 unsigned int rq_clamp
;
990 lockdep_assert_held(&rq
->lock
);
992 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
993 SCHED_WARN_ON(!bucket
->tasks
);
994 if (likely(bucket
->tasks
))
996 uc_se
->active
= false;
999 * Keep "local max aggregation" simple and accept to (possibly)
1000 * overboost some RUNNABLE tasks in the same bucket.
1001 * The rq clamp bucket value is reset to its base value whenever
1002 * there are no more RUNNABLE tasks refcounting it.
1004 if (likely(bucket
->tasks
))
1007 rq_clamp
= READ_ONCE(uc_rq
->value
);
1009 * Defensive programming: this should never happen. If it happens,
1010 * e.g. due to future modification, warn and fixup the expected value.
1012 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1013 if (bucket
->value
>= rq_clamp
) {
1014 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1015 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1019 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1021 enum uclamp_id clamp_id
;
1023 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1026 for_each_clamp_id(clamp_id
)
1027 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1029 /* Reset clamp idle holding when there is one RUNNABLE task */
1030 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1031 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1034 static inline void uclamp_rq_dec(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_dec_id(rq
, p
, clamp_id
);
1046 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1052 * Lock the task and the rq where the task is (or was) queued.
1054 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1055 * price to pay to safely serialize util_{min,max} updates with
1056 * enqueues, dequeues and migration operations.
1057 * This is the same locking schema used by __set_cpus_allowed_ptr().
1059 rq
= task_rq_lock(p
, &rf
);
1062 * Setting the clamp bucket is serialized by task_rq_lock().
1063 * If the task is not yet RUNNABLE and its task_struct is not
1064 * affecting a valid clamp bucket, the next time it's enqueued,
1065 * it will already see the updated clamp bucket value.
1067 if (p
->uclamp
[clamp_id
].active
) {
1068 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1069 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1072 task_rq_unlock(rq
, p
, &rf
);
1075 #ifdef CONFIG_UCLAMP_TASK_GROUP
1077 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1078 unsigned int clamps
)
1080 enum uclamp_id clamp_id
;
1081 struct css_task_iter it
;
1082 struct task_struct
*p
;
1084 css_task_iter_start(css
, 0, &it
);
1085 while ((p
= css_task_iter_next(&it
))) {
1086 for_each_clamp_id(clamp_id
) {
1087 if ((0x1 << clamp_id
) & clamps
)
1088 uclamp_update_active(p
, clamp_id
);
1091 css_task_iter_end(&it
);
1094 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1095 static void uclamp_update_root_tg(void)
1097 struct task_group
*tg
= &root_task_group
;
1099 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1100 sysctl_sched_uclamp_util_min
, false);
1101 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1102 sysctl_sched_uclamp_util_max
, false);
1105 cpu_util_update_eff(&root_task_group
.css
);
1109 static void uclamp_update_root_tg(void) { }
1112 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1113 void __user
*buffer
, size_t *lenp
,
1116 bool update_root_tg
= false;
1117 int old_min
, old_max
;
1120 mutex_lock(&uclamp_mutex
);
1121 old_min
= sysctl_sched_uclamp_util_min
;
1122 old_max
= sysctl_sched_uclamp_util_max
;
1124 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1130 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1131 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
) {
1136 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1137 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1138 sysctl_sched_uclamp_util_min
, false);
1139 update_root_tg
= true;
1141 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1142 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1143 sysctl_sched_uclamp_util_max
, false);
1144 update_root_tg
= true;
1148 uclamp_update_root_tg();
1151 * We update all RUNNABLE tasks only when task groups are in use.
1152 * Otherwise, keep it simple and do just a lazy update at each next
1153 * task enqueue time.
1159 sysctl_sched_uclamp_util_min
= old_min
;
1160 sysctl_sched_uclamp_util_max
= old_max
;
1162 mutex_unlock(&uclamp_mutex
);
1167 static int uclamp_validate(struct task_struct
*p
,
1168 const struct sched_attr
*attr
)
1170 unsigned int lower_bound
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1171 unsigned int upper_bound
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1173 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
)
1174 lower_bound
= attr
->sched_util_min
;
1175 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
)
1176 upper_bound
= attr
->sched_util_max
;
1178 if (lower_bound
> upper_bound
)
1180 if (upper_bound
> SCHED_CAPACITY_SCALE
)
1186 static void __setscheduler_uclamp(struct task_struct
*p
,
1187 const struct sched_attr
*attr
)
1189 enum uclamp_id clamp_id
;
1192 * On scheduling class change, reset to default clamps for tasks
1193 * without a task-specific value.
1195 for_each_clamp_id(clamp_id
) {
1196 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1197 unsigned int clamp_value
= uclamp_none(clamp_id
);
1199 /* Keep using defined clamps across class changes */
1200 if (uc_se
->user_defined
)
1203 /* By default, RT tasks always get 100% boost */
1204 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1205 clamp_value
= uclamp_none(UCLAMP_MAX
);
1207 uclamp_se_set(uc_se
, clamp_value
, false);
1210 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1213 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1214 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1215 attr
->sched_util_min
, true);
1218 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1219 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1220 attr
->sched_util_max
, true);
1224 static void uclamp_fork(struct task_struct
*p
)
1226 enum uclamp_id clamp_id
;
1228 for_each_clamp_id(clamp_id
)
1229 p
->uclamp
[clamp_id
].active
= false;
1231 if (likely(!p
->sched_reset_on_fork
))
1234 for_each_clamp_id(clamp_id
) {
1235 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1236 uclamp_none(clamp_id
), false);
1240 static void __init
init_uclamp(void)
1242 struct uclamp_se uc_max
= {};
1243 enum uclamp_id clamp_id
;
1246 mutex_init(&uclamp_mutex
);
1248 for_each_possible_cpu(cpu
) {
1249 memset(&cpu_rq(cpu
)->uclamp
, 0,
1250 sizeof(struct uclamp_rq
)*UCLAMP_CNT
);
1251 cpu_rq(cpu
)->uclamp_flags
= 0;
1254 for_each_clamp_id(clamp_id
) {
1255 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1256 uclamp_none(clamp_id
), false);
1259 /* System defaults allow max clamp values for both indexes */
1260 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1261 for_each_clamp_id(clamp_id
) {
1262 uclamp_default
[clamp_id
] = uc_max
;
1263 #ifdef CONFIG_UCLAMP_TASK_GROUP
1264 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1265 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1270 #else /* CONFIG_UCLAMP_TASK */
1271 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1272 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1273 static inline int uclamp_validate(struct task_struct
*p
,
1274 const struct sched_attr
*attr
)
1278 static void __setscheduler_uclamp(struct task_struct
*p
,
1279 const struct sched_attr
*attr
) { }
1280 static inline void uclamp_fork(struct task_struct
*p
) { }
1281 static inline void init_uclamp(void) { }
1282 #endif /* CONFIG_UCLAMP_TASK */
1284 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1286 if (!(flags
& ENQUEUE_NOCLOCK
))
1287 update_rq_clock(rq
);
1289 if (!(flags
& ENQUEUE_RESTORE
)) {
1290 sched_info_queued(rq
, p
);
1291 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1294 uclamp_rq_inc(rq
, p
);
1295 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1298 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1300 if (!(flags
& DEQUEUE_NOCLOCK
))
1301 update_rq_clock(rq
);
1303 if (!(flags
& DEQUEUE_SAVE
)) {
1304 sched_info_dequeued(rq
, p
);
1305 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1308 uclamp_rq_dec(rq
, p
);
1309 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1312 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1314 if (task_contributes_to_load(p
))
1315 rq
->nr_uninterruptible
--;
1317 enqueue_task(rq
, p
, flags
);
1319 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1322 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1324 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1326 if (task_contributes_to_load(p
))
1327 rq
->nr_uninterruptible
++;
1329 dequeue_task(rq
, p
, flags
);
1333 * __normal_prio - return the priority that is based on the static prio
1335 static inline int __normal_prio(struct task_struct
*p
)
1337 return p
->static_prio
;
1341 * Calculate the expected normal priority: i.e. priority
1342 * without taking RT-inheritance into account. Might be
1343 * boosted by interactivity modifiers. Changes upon fork,
1344 * setprio syscalls, and whenever the interactivity
1345 * estimator recalculates.
1347 static inline int normal_prio(struct task_struct
*p
)
1351 if (task_has_dl_policy(p
))
1352 prio
= MAX_DL_PRIO
-1;
1353 else if (task_has_rt_policy(p
))
1354 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1356 prio
= __normal_prio(p
);
1361 * Calculate the current priority, i.e. the priority
1362 * taken into account by the scheduler. This value might
1363 * be boosted by RT tasks, or might be boosted by
1364 * interactivity modifiers. Will be RT if the task got
1365 * RT-boosted. If not then it returns p->normal_prio.
1367 static int effective_prio(struct task_struct
*p
)
1369 p
->normal_prio
= normal_prio(p
);
1371 * If we are RT tasks or we were boosted to RT priority,
1372 * keep the priority unchanged. Otherwise, update priority
1373 * to the normal priority:
1375 if (!rt_prio(p
->prio
))
1376 return p
->normal_prio
;
1381 * task_curr - is this task currently executing on a CPU?
1382 * @p: the task in question.
1384 * Return: 1 if the task is currently executing. 0 otherwise.
1386 inline int task_curr(const struct task_struct
*p
)
1388 return cpu_curr(task_cpu(p
)) == p
;
1392 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1393 * use the balance_callback list if you want balancing.
1395 * this means any call to check_class_changed() must be followed by a call to
1396 * balance_callback().
1398 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1399 const struct sched_class
*prev_class
,
1402 if (prev_class
!= p
->sched_class
) {
1403 if (prev_class
->switched_from
)
1404 prev_class
->switched_from(rq
, p
);
1406 p
->sched_class
->switched_to(rq
, p
);
1407 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1408 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1411 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1413 const struct sched_class
*class;
1415 if (p
->sched_class
== rq
->curr
->sched_class
) {
1416 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1418 for_each_class(class) {
1419 if (class == rq
->curr
->sched_class
)
1421 if (class == p
->sched_class
) {
1429 * A queue event has occurred, and we're going to schedule. In
1430 * this case, we can save a useless back to back clock update.
1432 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1433 rq_clock_skip_update(rq
);
1439 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1440 * __set_cpus_allowed_ptr() and select_fallback_rq().
1442 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1444 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1447 if (is_per_cpu_kthread(p
))
1448 return cpu_online(cpu
);
1450 return cpu_active(cpu
);
1454 * This is how migration works:
1456 * 1) we invoke migration_cpu_stop() on the target CPU using
1458 * 2) stopper starts to run (implicitly forcing the migrated thread
1460 * 3) it checks whether the migrated task is still in the wrong runqueue.
1461 * 4) if it's in the wrong runqueue then the migration thread removes
1462 * it and puts it into the right queue.
1463 * 5) stopper completes and stop_one_cpu() returns and the migration
1468 * move_queued_task - move a queued task to new rq.
1470 * Returns (locked) new rq. Old rq's lock is released.
1472 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1473 struct task_struct
*p
, int new_cpu
)
1475 lockdep_assert_held(&rq
->lock
);
1477 WRITE_ONCE(p
->on_rq
, TASK_ON_RQ_MIGRATING
);
1478 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
1479 set_task_cpu(p
, new_cpu
);
1482 rq
= cpu_rq(new_cpu
);
1485 BUG_ON(task_cpu(p
) != new_cpu
);
1486 enqueue_task(rq
, p
, 0);
1487 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1488 check_preempt_curr(rq
, p
, 0);
1493 struct migration_arg
{
1494 struct task_struct
*task
;
1499 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1500 * this because either it can't run here any more (set_cpus_allowed()
1501 * away from this CPU, or CPU going down), or because we're
1502 * attempting to rebalance this task on exec (sched_exec).
1504 * So we race with normal scheduler movements, but that's OK, as long
1505 * as the task is no longer on this CPU.
1507 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1508 struct task_struct
*p
, int dest_cpu
)
1510 /* Affinity changed (again). */
1511 if (!is_cpu_allowed(p
, dest_cpu
))
1514 update_rq_clock(rq
);
1515 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1521 * migration_cpu_stop - this will be executed by a highprio stopper thread
1522 * and performs thread migration by bumping thread off CPU then
1523 * 'pushing' onto another runqueue.
1525 static int migration_cpu_stop(void *data
)
1527 struct migration_arg
*arg
= data
;
1528 struct task_struct
*p
= arg
->task
;
1529 struct rq
*rq
= this_rq();
1533 * The original target CPU might have gone down and we might
1534 * be on another CPU but it doesn't matter.
1536 local_irq_disable();
1538 * We need to explicitly wake pending tasks before running
1539 * __migrate_task() such that we will not miss enforcing cpus_ptr
1540 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1542 sched_ttwu_pending();
1544 raw_spin_lock(&p
->pi_lock
);
1547 * If task_rq(p) != rq, it cannot be migrated here, because we're
1548 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1549 * we're holding p->pi_lock.
1551 if (task_rq(p
) == rq
) {
1552 if (task_on_rq_queued(p
))
1553 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1555 p
->wake_cpu
= arg
->dest_cpu
;
1558 raw_spin_unlock(&p
->pi_lock
);
1565 * sched_class::set_cpus_allowed must do the below, but is not required to
1566 * actually call this function.
1568 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1570 cpumask_copy(&p
->cpus_mask
, new_mask
);
1571 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1574 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1576 struct rq
*rq
= task_rq(p
);
1577 bool queued
, running
;
1579 lockdep_assert_held(&p
->pi_lock
);
1581 queued
= task_on_rq_queued(p
);
1582 running
= task_current(rq
, p
);
1586 * Because __kthread_bind() calls this on blocked tasks without
1589 lockdep_assert_held(&rq
->lock
);
1590 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1593 put_prev_task(rq
, p
);
1595 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1598 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1600 set_next_task(rq
, p
);
1604 * Change a given task's CPU affinity. Migrate the thread to a
1605 * proper CPU and schedule it away if the CPU it's executing on
1606 * is removed from the allowed bitmask.
1608 * NOTE: the caller must have a valid reference to the task, the
1609 * task must not exit() & deallocate itself prematurely. The
1610 * call is not atomic; no spinlocks may be held.
1612 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1613 const struct cpumask
*new_mask
, bool check
)
1615 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1616 unsigned int dest_cpu
;
1621 rq
= task_rq_lock(p
, &rf
);
1622 update_rq_clock(rq
);
1624 if (p
->flags
& PF_KTHREAD
) {
1626 * Kernel threads are allowed on online && !active CPUs
1628 cpu_valid_mask
= cpu_online_mask
;
1632 * Must re-check here, to close a race against __kthread_bind(),
1633 * sched_setaffinity() is not guaranteed to observe the flag.
1635 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1640 if (cpumask_equal(p
->cpus_ptr
, new_mask
))
1644 * Picking a ~random cpu helps in cases where we are changing affinity
1645 * for groups of tasks (ie. cpuset), so that load balancing is not
1646 * immediately required to distribute the tasks within their new mask.
1648 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
1649 if (dest_cpu
>= nr_cpu_ids
) {
1654 do_set_cpus_allowed(p
, new_mask
);
1656 if (p
->flags
& PF_KTHREAD
) {
1658 * For kernel threads that do indeed end up on online &&
1659 * !active we want to ensure they are strict per-CPU threads.
1661 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1662 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1663 p
->nr_cpus_allowed
!= 1);
1666 /* Can the task run on the task's current CPU? If so, we're done */
1667 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1670 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1671 struct migration_arg arg
= { p
, dest_cpu
};
1672 /* Need help from migration thread: drop lock and wait. */
1673 task_rq_unlock(rq
, p
, &rf
);
1674 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1676 } else if (task_on_rq_queued(p
)) {
1678 * OK, since we're going to drop the lock immediately
1679 * afterwards anyway.
1681 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1684 task_rq_unlock(rq
, p
, &rf
);
1689 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1691 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1693 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1695 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1697 #ifdef CONFIG_SCHED_DEBUG
1699 * We should never call set_task_cpu() on a blocked task,
1700 * ttwu() will sort out the placement.
1702 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1706 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1707 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1708 * time relying on p->on_rq.
1710 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1711 p
->sched_class
== &fair_sched_class
&&
1712 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1714 #ifdef CONFIG_LOCKDEP
1716 * The caller should hold either p->pi_lock or rq->lock, when changing
1717 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1719 * sched_move_task() holds both and thus holding either pins the cgroup,
1722 * Furthermore, all task_rq users should acquire both locks, see
1725 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1726 lockdep_is_held(&task_rq(p
)->lock
)));
1729 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1731 WARN_ON_ONCE(!cpu_online(new_cpu
));
1734 trace_sched_migrate_task(p
, new_cpu
);
1736 if (task_cpu(p
) != new_cpu
) {
1737 if (p
->sched_class
->migrate_task_rq
)
1738 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1739 p
->se
.nr_migrations
++;
1741 perf_event_task_migrate(p
);
1744 __set_task_cpu(p
, new_cpu
);
1747 #ifdef CONFIG_NUMA_BALANCING
1748 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1750 if (task_on_rq_queued(p
)) {
1751 struct rq
*src_rq
, *dst_rq
;
1752 struct rq_flags srf
, drf
;
1754 src_rq
= task_rq(p
);
1755 dst_rq
= cpu_rq(cpu
);
1757 rq_pin_lock(src_rq
, &srf
);
1758 rq_pin_lock(dst_rq
, &drf
);
1760 deactivate_task(src_rq
, p
, 0);
1761 set_task_cpu(p
, cpu
);
1762 activate_task(dst_rq
, p
, 0);
1763 check_preempt_curr(dst_rq
, p
, 0);
1765 rq_unpin_lock(dst_rq
, &drf
);
1766 rq_unpin_lock(src_rq
, &srf
);
1770 * Task isn't running anymore; make it appear like we migrated
1771 * it before it went to sleep. This means on wakeup we make the
1772 * previous CPU our target instead of where it really is.
1778 struct migration_swap_arg
{
1779 struct task_struct
*src_task
, *dst_task
;
1780 int src_cpu
, dst_cpu
;
1783 static int migrate_swap_stop(void *data
)
1785 struct migration_swap_arg
*arg
= data
;
1786 struct rq
*src_rq
, *dst_rq
;
1789 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1792 src_rq
= cpu_rq(arg
->src_cpu
);
1793 dst_rq
= cpu_rq(arg
->dst_cpu
);
1795 double_raw_lock(&arg
->src_task
->pi_lock
,
1796 &arg
->dst_task
->pi_lock
);
1797 double_rq_lock(src_rq
, dst_rq
);
1799 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1802 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1805 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
1808 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
1811 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1812 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1817 double_rq_unlock(src_rq
, dst_rq
);
1818 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1819 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1825 * Cross migrate two tasks
1827 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
1828 int target_cpu
, int curr_cpu
)
1830 struct migration_swap_arg arg
;
1833 arg
= (struct migration_swap_arg
){
1835 .src_cpu
= curr_cpu
,
1837 .dst_cpu
= target_cpu
,
1840 if (arg
.src_cpu
== arg
.dst_cpu
)
1844 * These three tests are all lockless; this is OK since all of them
1845 * will be re-checked with proper locks held further down the line.
1847 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1850 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
1853 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
1856 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1857 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1862 #endif /* CONFIG_NUMA_BALANCING */
1865 * wait_task_inactive - wait for a thread to unschedule.
1867 * If @match_state is nonzero, it's the @p->state value just checked and
1868 * not expected to change. If it changes, i.e. @p might have woken up,
1869 * then return zero. When we succeed in waiting for @p to be off its CPU,
1870 * we return a positive number (its total switch count). If a second call
1871 * a short while later returns the same number, the caller can be sure that
1872 * @p has remained unscheduled the whole time.
1874 * The caller must ensure that the task *will* unschedule sometime soon,
1875 * else this function might spin for a *long* time. This function can't
1876 * be called with interrupts off, or it may introduce deadlock with
1877 * smp_call_function() if an IPI is sent by the same process we are
1878 * waiting to become inactive.
1880 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1882 int running
, queued
;
1889 * We do the initial early heuristics without holding
1890 * any task-queue locks at all. We'll only try to get
1891 * the runqueue lock when things look like they will
1897 * If the task is actively running on another CPU
1898 * still, just relax and busy-wait without holding
1901 * NOTE! Since we don't hold any locks, it's not
1902 * even sure that "rq" stays as the right runqueue!
1903 * But we don't care, since "task_running()" will
1904 * return false if the runqueue has changed and p
1905 * is actually now running somewhere else!
1907 while (task_running(rq
, p
)) {
1908 if (match_state
&& unlikely(p
->state
!= match_state
))
1914 * Ok, time to look more closely! We need the rq
1915 * lock now, to be *sure*. If we're wrong, we'll
1916 * just go back and repeat.
1918 rq
= task_rq_lock(p
, &rf
);
1919 trace_sched_wait_task(p
);
1920 running
= task_running(rq
, p
);
1921 queued
= task_on_rq_queued(p
);
1923 if (!match_state
|| p
->state
== match_state
)
1924 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1925 task_rq_unlock(rq
, p
, &rf
);
1928 * If it changed from the expected state, bail out now.
1930 if (unlikely(!ncsw
))
1934 * Was it really running after all now that we
1935 * checked with the proper locks actually held?
1937 * Oops. Go back and try again..
1939 if (unlikely(running
)) {
1945 * It's not enough that it's not actively running,
1946 * it must be off the runqueue _entirely_, and not
1949 * So if it was still runnable (but just not actively
1950 * running right now), it's preempted, and we should
1951 * yield - it could be a while.
1953 if (unlikely(queued
)) {
1954 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1956 set_current_state(TASK_UNINTERRUPTIBLE
);
1957 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1962 * Ahh, all good. It wasn't running, and it wasn't
1963 * runnable, which means that it will never become
1964 * running in the future either. We're all done!
1973 * kick_process - kick a running thread to enter/exit the kernel
1974 * @p: the to-be-kicked thread
1976 * Cause a process which is running on another CPU to enter
1977 * kernel-mode, without any delay. (to get signals handled.)
1979 * NOTE: this function doesn't have to take the runqueue lock,
1980 * because all it wants to ensure is that the remote task enters
1981 * the kernel. If the IPI races and the task has been migrated
1982 * to another CPU then no harm is done and the purpose has been
1985 void kick_process(struct task_struct
*p
)
1991 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1992 smp_send_reschedule(cpu
);
1995 EXPORT_SYMBOL_GPL(kick_process
);
1998 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2000 * A few notes on cpu_active vs cpu_online:
2002 * - cpu_active must be a subset of cpu_online
2004 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2005 * see __set_cpus_allowed_ptr(). At this point the newly online
2006 * CPU isn't yet part of the sched domains, and balancing will not
2009 * - on CPU-down we clear cpu_active() to mask the sched domains and
2010 * avoid the load balancer to place new tasks on the to be removed
2011 * CPU. Existing tasks will remain running there and will be taken
2014 * This means that fallback selection must not select !active CPUs.
2015 * And can assume that any active CPU must be online. Conversely
2016 * select_task_rq() below may allow selection of !active CPUs in order
2017 * to satisfy the above rules.
2019 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2021 int nid
= cpu_to_node(cpu
);
2022 const struct cpumask
*nodemask
= NULL
;
2023 enum { cpuset
, possible
, fail
} state
= cpuset
;
2027 * If the node that the CPU is on has been offlined, cpu_to_node()
2028 * will return -1. There is no CPU on the node, and we should
2029 * select the CPU on the other node.
2032 nodemask
= cpumask_of_node(nid
);
2034 /* Look for allowed, online CPU in same node. */
2035 for_each_cpu(dest_cpu
, nodemask
) {
2036 if (!cpu_active(dest_cpu
))
2038 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2044 /* Any allowed, online CPU? */
2045 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2046 if (!is_cpu_allowed(p
, dest_cpu
))
2052 /* No more Mr. Nice Guy. */
2055 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2056 cpuset_cpus_allowed_fallback(p
);
2062 do_set_cpus_allowed(p
, cpu_possible_mask
);
2073 if (state
!= cpuset
) {
2075 * Don't tell them about moving exiting tasks or
2076 * kernel threads (both mm NULL), since they never
2079 if (p
->mm
&& printk_ratelimit()) {
2080 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2081 task_pid_nr(p
), p
->comm
, cpu
);
2089 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2092 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
2094 lockdep_assert_held(&p
->pi_lock
);
2096 if (p
->nr_cpus_allowed
> 1)
2097 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
2099 cpu
= cpumask_any(p
->cpus_ptr
);
2102 * In order not to call set_task_cpu() on a blocking task we need
2103 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2106 * Since this is common to all placement strategies, this lives here.
2108 * [ this allows ->select_task() to simply return task_cpu(p) and
2109 * not worry about this generic constraint ]
2111 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2112 cpu
= select_fallback_rq(task_cpu(p
), p
);
2117 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2119 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2120 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2124 * Make it appear like a SCHED_FIFO task, its something
2125 * userspace knows about and won't get confused about.
2127 * Also, it will make PI more or less work without too
2128 * much confusion -- but then, stop work should not
2129 * rely on PI working anyway.
2131 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2133 stop
->sched_class
= &stop_sched_class
;
2136 cpu_rq(cpu
)->stop
= stop
;
2140 * Reset it back to a normal scheduling class so that
2141 * it can die in pieces.
2143 old_stop
->sched_class
= &rt_sched_class
;
2149 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2150 const struct cpumask
*new_mask
, bool check
)
2152 return set_cpus_allowed_ptr(p
, new_mask
);
2155 #endif /* CONFIG_SMP */
2158 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2162 if (!schedstat_enabled())
2168 if (cpu
== rq
->cpu
) {
2169 __schedstat_inc(rq
->ttwu_local
);
2170 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2172 struct sched_domain
*sd
;
2174 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2176 for_each_domain(rq
->cpu
, sd
) {
2177 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2178 __schedstat_inc(sd
->ttwu_wake_remote
);
2185 if (wake_flags
& WF_MIGRATED
)
2186 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2187 #endif /* CONFIG_SMP */
2189 __schedstat_inc(rq
->ttwu_count
);
2190 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2192 if (wake_flags
& WF_SYNC
)
2193 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2197 * Mark the task runnable and perform wakeup-preemption.
2199 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2200 struct rq_flags
*rf
)
2202 check_preempt_curr(rq
, p
, wake_flags
);
2203 p
->state
= TASK_RUNNING
;
2204 trace_sched_wakeup(p
);
2207 if (p
->sched_class
->task_woken
) {
2209 * Our task @p is fully woken up and running; so its safe to
2210 * drop the rq->lock, hereafter rq is only used for statistics.
2212 rq_unpin_lock(rq
, rf
);
2213 p
->sched_class
->task_woken(rq
, p
);
2214 rq_repin_lock(rq
, rf
);
2217 if (rq
->idle_stamp
) {
2218 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2219 u64 max
= 2*rq
->max_idle_balance_cost
;
2221 update_avg(&rq
->avg_idle
, delta
);
2223 if (rq
->avg_idle
> max
)
2232 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2233 struct rq_flags
*rf
)
2235 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2237 lockdep_assert_held(&rq
->lock
);
2240 if (p
->sched_contributes_to_load
)
2241 rq
->nr_uninterruptible
--;
2243 if (wake_flags
& WF_MIGRATED
)
2244 en_flags
|= ENQUEUE_MIGRATED
;
2247 activate_task(rq
, p
, en_flags
);
2248 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2252 * Called in case the task @p isn't fully descheduled from its runqueue,
2253 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2254 * since all we need to do is flip p->state to TASK_RUNNING, since
2255 * the task is still ->on_rq.
2257 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2263 rq
= __task_rq_lock(p
, &rf
);
2264 if (task_on_rq_queued(p
)) {
2265 /* check_preempt_curr() may use rq clock */
2266 update_rq_clock(rq
);
2267 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
2270 __task_rq_unlock(rq
, &rf
);
2276 void sched_ttwu_pending(void)
2278 struct rq
*rq
= this_rq();
2279 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
2280 struct task_struct
*p
, *t
;
2286 rq_lock_irqsave(rq
, &rf
);
2287 update_rq_clock(rq
);
2289 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
2290 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
2292 rq_unlock_irqrestore(rq
, &rf
);
2295 void scheduler_ipi(void)
2298 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2299 * TIF_NEED_RESCHED remotely (for the first time) will also send
2302 preempt_fold_need_resched();
2304 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
2308 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2309 * traditionally all their work was done from the interrupt return
2310 * path. Now that we actually do some work, we need to make sure
2313 * Some archs already do call them, luckily irq_enter/exit nest
2316 * Arguably we should visit all archs and update all handlers,
2317 * however a fair share of IPIs are still resched only so this would
2318 * somewhat pessimize the simple resched case.
2321 sched_ttwu_pending();
2324 * Check if someone kicked us for doing the nohz idle load balance.
2326 if (unlikely(got_nohz_idle_kick())) {
2327 this_rq()->idle_balance
= 1;
2328 raise_softirq_irqoff(SCHED_SOFTIRQ
);
2333 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
2335 struct rq
*rq
= cpu_rq(cpu
);
2337 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
2339 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
2340 if (!set_nr_if_polling(rq
->idle
))
2341 smp_send_reschedule(cpu
);
2343 trace_sched_wake_idle_without_ipi(cpu
);
2347 void wake_up_if_idle(int cpu
)
2349 struct rq
*rq
= cpu_rq(cpu
);
2354 if (!is_idle_task(rcu_dereference(rq
->curr
)))
2357 if (set_nr_if_polling(rq
->idle
)) {
2358 trace_sched_wake_idle_without_ipi(cpu
);
2360 rq_lock_irqsave(rq
, &rf
);
2361 if (is_idle_task(rq
->curr
))
2362 smp_send_reschedule(cpu
);
2363 /* Else CPU is not idle, do nothing here: */
2364 rq_unlock_irqrestore(rq
, &rf
);
2371 bool cpus_share_cache(int this_cpu
, int that_cpu
)
2373 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
2375 #endif /* CONFIG_SMP */
2377 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
2379 struct rq
*rq
= cpu_rq(cpu
);
2382 #if defined(CONFIG_SMP)
2383 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
2384 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
2385 ttwu_queue_remote(p
, cpu
, wake_flags
);
2391 update_rq_clock(rq
);
2392 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
2397 * Notes on Program-Order guarantees on SMP systems.
2401 * The basic program-order guarantee on SMP systems is that when a task [t]
2402 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2403 * execution on its new CPU [c1].
2405 * For migration (of runnable tasks) this is provided by the following means:
2407 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2408 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2409 * rq(c1)->lock (if not at the same time, then in that order).
2410 * C) LOCK of the rq(c1)->lock scheduling in task
2412 * Release/acquire chaining guarantees that B happens after A and C after B.
2413 * Note: the CPU doing B need not be c0 or c1
2422 * UNLOCK rq(0)->lock
2424 * LOCK rq(0)->lock // orders against CPU0
2426 * UNLOCK rq(0)->lock
2430 * UNLOCK rq(1)->lock
2432 * LOCK rq(1)->lock // orders against CPU2
2435 * UNLOCK rq(1)->lock
2438 * BLOCKING -- aka. SLEEP + WAKEUP
2440 * For blocking we (obviously) need to provide the same guarantee as for
2441 * migration. However the means are completely different as there is no lock
2442 * chain to provide order. Instead we do:
2444 * 1) smp_store_release(X->on_cpu, 0)
2445 * 2) smp_cond_load_acquire(!X->on_cpu)
2449 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2451 * LOCK rq(0)->lock LOCK X->pi_lock
2454 * smp_store_release(X->on_cpu, 0);
2456 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2462 * X->state = RUNNING
2463 * UNLOCK rq(2)->lock
2465 * LOCK rq(2)->lock // orders against CPU1
2468 * UNLOCK rq(2)->lock
2471 * UNLOCK rq(0)->lock
2474 * However, for wakeups there is a second guarantee we must provide, namely we
2475 * must ensure that CONDITION=1 done by the caller can not be reordered with
2476 * accesses to the task state; see try_to_wake_up() and set_current_state().
2480 * try_to_wake_up - wake up a thread
2481 * @p: the thread to be awakened
2482 * @state: the mask of task states that can be woken
2483 * @wake_flags: wake modifier flags (WF_*)
2485 * If (@state & @p->state) @p->state = TASK_RUNNING.
2487 * If the task was not queued/runnable, also place it back on a runqueue.
2489 * Atomic against schedule() which would dequeue a task, also see
2490 * set_current_state().
2492 * This function executes a full memory barrier before accessing the task
2493 * state; see set_current_state().
2495 * Return: %true if @p->state changes (an actual wakeup was done),
2499 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2501 unsigned long flags
;
2502 int cpu
, success
= 0;
2507 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2508 * == smp_processor_id()'. Together this means we can special
2509 * case the whole 'p->on_rq && ttwu_remote()' case below
2510 * without taking any locks.
2513 * - we rely on Program-Order guarantees for all the ordering,
2514 * - we're serialized against set_special_state() by virtue of
2515 * it disabling IRQs (this allows not taking ->pi_lock).
2517 if (!(p
->state
& state
))
2522 trace_sched_waking(p
);
2523 p
->state
= TASK_RUNNING
;
2524 trace_sched_wakeup(p
);
2529 * If we are going to wake up a thread waiting for CONDITION we
2530 * need to ensure that CONDITION=1 done by the caller can not be
2531 * reordered with p->state check below. This pairs with mb() in
2532 * set_current_state() the waiting thread does.
2534 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2535 smp_mb__after_spinlock();
2536 if (!(p
->state
& state
))
2539 trace_sched_waking(p
);
2541 /* We're going to change ->state: */
2546 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2547 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2548 * in smp_cond_load_acquire() below.
2550 * sched_ttwu_pending() try_to_wake_up()
2551 * STORE p->on_rq = 1 LOAD p->state
2554 * __schedule() (switch to task 'p')
2555 * LOCK rq->lock smp_rmb();
2556 * smp_mb__after_spinlock();
2560 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2562 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2563 * __schedule(). See the comment for smp_mb__after_spinlock().
2565 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2568 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2573 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2574 * possible to, falsely, observe p->on_cpu == 0.
2576 * One must be running (->on_cpu == 1) in order to remove oneself
2577 * from the runqueue.
2579 * __schedule() (switch to task 'p') try_to_wake_up()
2580 * STORE p->on_cpu = 1 LOAD p->on_rq
2583 * __schedule() (put 'p' to sleep)
2584 * LOCK rq->lock smp_rmb();
2585 * smp_mb__after_spinlock();
2586 * STORE p->on_rq = 0 LOAD p->on_cpu
2588 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2589 * __schedule(). See the comment for smp_mb__after_spinlock().
2594 * If the owning (remote) CPU is still in the middle of schedule() with
2595 * this task as prev, wait until its done referencing the task.
2597 * Pairs with the smp_store_release() in finish_task().
2599 * This ensures that tasks getting woken will be fully ordered against
2600 * their previous state and preserve Program Order.
2602 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2604 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2605 p
->state
= TASK_WAKING
;
2608 delayacct_blkio_end(p
);
2609 atomic_dec(&task_rq(p
)->nr_iowait
);
2612 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2613 if (task_cpu(p
) != cpu
) {
2614 wake_flags
|= WF_MIGRATED
;
2615 psi_ttwu_dequeue(p
);
2616 set_task_cpu(p
, cpu
);
2619 #else /* CONFIG_SMP */
2622 delayacct_blkio_end(p
);
2623 atomic_dec(&task_rq(p
)->nr_iowait
);
2626 #endif /* CONFIG_SMP */
2628 ttwu_queue(p
, cpu
, wake_flags
);
2630 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2633 ttwu_stat(p
, cpu
, wake_flags
);
2640 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2641 * @p: Process for which the function is to be invoked.
2642 * @func: Function to invoke.
2643 * @arg: Argument to function.
2645 * If the specified task can be quickly locked into a definite state
2646 * (either sleeping or on a given runqueue), arrange to keep it in that
2647 * state while invoking @func(@arg). This function can use ->on_rq and
2648 * task_curr() to work out what the state is, if required. Given that
2649 * @func can be invoked with a runqueue lock held, it had better be quite
2653 * @false if the task slipped out from under the locks.
2654 * @true if the task was locked onto a runqueue or is sleeping.
2655 * However, @func can override this by returning @false.
2657 bool try_invoke_on_locked_down_task(struct task_struct
*p
, bool (*func
)(struct task_struct
*t
, void *arg
), void *arg
)
2663 lockdep_assert_irqs_enabled();
2664 raw_spin_lock_irq(&p
->pi_lock
);
2666 rq
= __task_rq_lock(p
, &rf
);
2667 if (task_rq(p
) == rq
)
2676 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
2681 raw_spin_unlock_irq(&p
->pi_lock
);
2686 * wake_up_process - Wake up a specific process
2687 * @p: The process to be woken up.
2689 * Attempt to wake up the nominated process and move it to the set of runnable
2692 * Return: 1 if the process was woken up, 0 if it was already running.
2694 * This function executes a full memory barrier before accessing the task state.
2696 int wake_up_process(struct task_struct
*p
)
2698 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2700 EXPORT_SYMBOL(wake_up_process
);
2702 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2704 return try_to_wake_up(p
, state
, 0);
2708 * Perform scheduler related setup for a newly forked process p.
2709 * p is forked by current.
2711 * __sched_fork() is basic setup used by init_idle() too:
2713 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2718 p
->se
.exec_start
= 0;
2719 p
->se
.sum_exec_runtime
= 0;
2720 p
->se
.prev_sum_exec_runtime
= 0;
2721 p
->se
.nr_migrations
= 0;
2723 INIT_LIST_HEAD(&p
->se
.group_node
);
2725 #ifdef CONFIG_FAIR_GROUP_SCHED
2726 p
->se
.cfs_rq
= NULL
;
2729 #ifdef CONFIG_SCHEDSTATS
2730 /* Even if schedstat is disabled, there should not be garbage */
2731 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2734 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2735 init_dl_task_timer(&p
->dl
);
2736 init_dl_inactive_task_timer(&p
->dl
);
2737 __dl_clear_params(p
);
2739 INIT_LIST_HEAD(&p
->rt
.run_list
);
2741 p
->rt
.time_slice
= sched_rr_timeslice
;
2745 #ifdef CONFIG_PREEMPT_NOTIFIERS
2746 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2749 #ifdef CONFIG_COMPACTION
2750 p
->capture_control
= NULL
;
2752 init_numa_balancing(clone_flags
, p
);
2755 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2757 #ifdef CONFIG_NUMA_BALANCING
2759 void set_numabalancing_state(bool enabled
)
2762 static_branch_enable(&sched_numa_balancing
);
2764 static_branch_disable(&sched_numa_balancing
);
2767 #ifdef CONFIG_PROC_SYSCTL
2768 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2769 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2773 int state
= static_branch_likely(&sched_numa_balancing
);
2775 if (write
&& !capable(CAP_SYS_ADMIN
))
2780 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2784 set_numabalancing_state(state
);
2790 #ifdef CONFIG_SCHEDSTATS
2792 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2793 static bool __initdata __sched_schedstats
= false;
2795 static void set_schedstats(bool enabled
)
2798 static_branch_enable(&sched_schedstats
);
2800 static_branch_disable(&sched_schedstats
);
2803 void force_schedstat_enabled(void)
2805 if (!schedstat_enabled()) {
2806 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2807 static_branch_enable(&sched_schedstats
);
2811 static int __init
setup_schedstats(char *str
)
2818 * This code is called before jump labels have been set up, so we can't
2819 * change the static branch directly just yet. Instead set a temporary
2820 * variable so init_schedstats() can do it later.
2822 if (!strcmp(str
, "enable")) {
2823 __sched_schedstats
= true;
2825 } else if (!strcmp(str
, "disable")) {
2826 __sched_schedstats
= false;
2831 pr_warn("Unable to parse schedstats=\n");
2835 __setup("schedstats=", setup_schedstats
);
2837 static void __init
init_schedstats(void)
2839 set_schedstats(__sched_schedstats
);
2842 #ifdef CONFIG_PROC_SYSCTL
2843 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2844 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2848 int state
= static_branch_likely(&sched_schedstats
);
2850 if (write
&& !capable(CAP_SYS_ADMIN
))
2855 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2859 set_schedstats(state
);
2862 #endif /* CONFIG_PROC_SYSCTL */
2863 #else /* !CONFIG_SCHEDSTATS */
2864 static inline void init_schedstats(void) {}
2865 #endif /* CONFIG_SCHEDSTATS */
2868 * fork()/clone()-time setup:
2870 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2872 unsigned long flags
;
2874 __sched_fork(clone_flags
, p
);
2876 * We mark the process as NEW here. This guarantees that
2877 * nobody will actually run it, and a signal or other external
2878 * event cannot wake it up and insert it on the runqueue either.
2880 p
->state
= TASK_NEW
;
2883 * Make sure we do not leak PI boosting priority to the child.
2885 p
->prio
= current
->normal_prio
;
2890 * Revert to default priority/policy on fork if requested.
2892 if (unlikely(p
->sched_reset_on_fork
)) {
2893 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2894 p
->policy
= SCHED_NORMAL
;
2895 p
->static_prio
= NICE_TO_PRIO(0);
2897 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2898 p
->static_prio
= NICE_TO_PRIO(0);
2900 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2901 set_load_weight(p
, false);
2904 * We don't need the reset flag anymore after the fork. It has
2905 * fulfilled its duty:
2907 p
->sched_reset_on_fork
= 0;
2910 if (dl_prio(p
->prio
))
2912 else if (rt_prio(p
->prio
))
2913 p
->sched_class
= &rt_sched_class
;
2915 p
->sched_class
= &fair_sched_class
;
2917 init_entity_runnable_average(&p
->se
);
2920 * The child is not yet in the pid-hash so no cgroup attach races,
2921 * and the cgroup is pinned to this child due to cgroup_fork()
2922 * is ran before sched_fork().
2924 * Silence PROVE_RCU.
2926 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2928 * We're setting the CPU for the first time, we don't migrate,
2929 * so use __set_task_cpu().
2931 __set_task_cpu(p
, smp_processor_id());
2932 if (p
->sched_class
->task_fork
)
2933 p
->sched_class
->task_fork(p
);
2934 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2936 #ifdef CONFIG_SCHED_INFO
2937 if (likely(sched_info_on()))
2938 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2940 #if defined(CONFIG_SMP)
2943 init_task_preempt_count(p
);
2945 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2946 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2951 unsigned long to_ratio(u64 period
, u64 runtime
)
2953 if (runtime
== RUNTIME_INF
)
2957 * Doing this here saves a lot of checks in all
2958 * the calling paths, and returning zero seems
2959 * safe for them anyway.
2964 return div64_u64(runtime
<< BW_SHIFT
, period
);
2968 * wake_up_new_task - wake up a newly created task for the first time.
2970 * This function will do some initial scheduler statistics housekeeping
2971 * that must be done for every newly created context, then puts the task
2972 * on the runqueue and wakes it.
2974 void wake_up_new_task(struct task_struct
*p
)
2979 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2980 p
->state
= TASK_RUNNING
;
2983 * Fork balancing, do it here and not earlier because:
2984 * - cpus_ptr can change in the fork path
2985 * - any previously selected CPU might disappear through hotplug
2987 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2988 * as we're not fully set-up yet.
2990 p
->recent_used_cpu
= task_cpu(p
);
2991 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2993 rq
= __task_rq_lock(p
, &rf
);
2994 update_rq_clock(rq
);
2995 post_init_entity_util_avg(p
);
2997 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
2998 trace_sched_wakeup_new(p
);
2999 check_preempt_curr(rq
, p
, WF_FORK
);
3001 if (p
->sched_class
->task_woken
) {
3003 * Nothing relies on rq->lock after this, so its fine to
3006 rq_unpin_lock(rq
, &rf
);
3007 p
->sched_class
->task_woken(rq
, p
);
3008 rq_repin_lock(rq
, &rf
);
3011 task_rq_unlock(rq
, p
, &rf
);
3014 #ifdef CONFIG_PREEMPT_NOTIFIERS
3016 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
3018 void preempt_notifier_inc(void)
3020 static_branch_inc(&preempt_notifier_key
);
3022 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
3024 void preempt_notifier_dec(void)
3026 static_branch_dec(&preempt_notifier_key
);
3028 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3031 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3032 * @notifier: notifier struct to register
3034 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3036 if (!static_branch_unlikely(&preempt_notifier_key
))
3037 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3039 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3041 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3044 * preempt_notifier_unregister - no longer interested in preemption notifications
3045 * @notifier: notifier struct to unregister
3047 * This is *not* safe to call from within a preemption notifier.
3049 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3051 hlist_del(¬ifier
->link
);
3053 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3055 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3057 struct preempt_notifier
*notifier
;
3059 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3060 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3063 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3065 if (static_branch_unlikely(&preempt_notifier_key
))
3066 __fire_sched_in_preempt_notifiers(curr
);
3070 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3071 struct task_struct
*next
)
3073 struct preempt_notifier
*notifier
;
3075 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3076 notifier
->ops
->sched_out(notifier
, next
);
3079 static __always_inline
void
3080 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3081 struct task_struct
*next
)
3083 if (static_branch_unlikely(&preempt_notifier_key
))
3084 __fire_sched_out_preempt_notifiers(curr
, next
);
3087 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3089 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3094 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3095 struct task_struct
*next
)
3099 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3101 static inline void prepare_task(struct task_struct
*next
)
3105 * Claim the task as running, we do this before switching to it
3106 * such that any running task will have this set.
3112 static inline void finish_task(struct task_struct
*prev
)
3116 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3117 * We must ensure this doesn't happen until the switch is completely
3120 * In particular, the load of prev->state in finish_task_switch() must
3121 * happen before this.
3123 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3125 smp_store_release(&prev
->on_cpu
, 0);
3130 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
3133 * Since the runqueue lock will be released by the next
3134 * task (which is an invalid locking op but in the case
3135 * of the scheduler it's an obvious special-case), so we
3136 * do an early lockdep release here:
3138 rq_unpin_lock(rq
, rf
);
3139 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
3140 #ifdef CONFIG_DEBUG_SPINLOCK
3141 /* this is a valid case when another task releases the spinlock */
3142 rq
->lock
.owner
= next
;
3146 static inline void finish_lock_switch(struct rq
*rq
)
3149 * If we are tracking spinlock dependencies then we have to
3150 * fix up the runqueue lock - which gets 'carried over' from
3151 * prev into current:
3153 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
3154 raw_spin_unlock_irq(&rq
->lock
);
3158 * NOP if the arch has not defined these:
3161 #ifndef prepare_arch_switch
3162 # define prepare_arch_switch(next) do { } while (0)
3165 #ifndef finish_arch_post_lock_switch
3166 # define finish_arch_post_lock_switch() do { } while (0)
3170 * prepare_task_switch - prepare to switch tasks
3171 * @rq: the runqueue preparing to switch
3172 * @prev: the current task that is being switched out
3173 * @next: the task we are going to switch to.
3175 * This is called with the rq lock held and interrupts off. It must
3176 * be paired with a subsequent finish_task_switch after the context
3179 * prepare_task_switch sets up locking and calls architecture specific
3183 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3184 struct task_struct
*next
)
3186 kcov_prepare_switch(prev
);
3187 sched_info_switch(rq
, prev
, next
);
3188 perf_event_task_sched_out(prev
, next
);
3190 fire_sched_out_preempt_notifiers(prev
, next
);
3192 prepare_arch_switch(next
);
3196 * finish_task_switch - clean up after a task-switch
3197 * @prev: the thread we just switched away from.
3199 * finish_task_switch must be called after the context switch, paired
3200 * with a prepare_task_switch call before the context switch.
3201 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3202 * and do any other architecture-specific cleanup actions.
3204 * Note that we may have delayed dropping an mm in context_switch(). If
3205 * so, we finish that here outside of the runqueue lock. (Doing it
3206 * with the lock held can cause deadlocks; see schedule() for
3209 * The context switch have flipped the stack from under us and restored the
3210 * local variables which were saved when this task called schedule() in the
3211 * past. prev == current is still correct but we need to recalculate this_rq
3212 * because prev may have moved to another CPU.
3214 static struct rq
*finish_task_switch(struct task_struct
*prev
)
3215 __releases(rq
->lock
)
3217 struct rq
*rq
= this_rq();
3218 struct mm_struct
*mm
= rq
->prev_mm
;
3222 * The previous task will have left us with a preempt_count of 2
3223 * because it left us after:
3226 * preempt_disable(); // 1
3228 * raw_spin_lock_irq(&rq->lock) // 2
3230 * Also, see FORK_PREEMPT_COUNT.
3232 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
3233 "corrupted preempt_count: %s/%d/0x%x\n",
3234 current
->comm
, current
->pid
, preempt_count()))
3235 preempt_count_set(FORK_PREEMPT_COUNT
);
3240 * A task struct has one reference for the use as "current".
3241 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3242 * schedule one last time. The schedule call will never return, and
3243 * the scheduled task must drop that reference.
3245 * We must observe prev->state before clearing prev->on_cpu (in
3246 * finish_task), otherwise a concurrent wakeup can get prev
3247 * running on another CPU and we could rave with its RUNNING -> DEAD
3248 * transition, resulting in a double drop.
3250 prev_state
= prev
->state
;
3251 vtime_task_switch(prev
);
3252 perf_event_task_sched_in(prev
, current
);
3254 finish_lock_switch(rq
);
3255 finish_arch_post_lock_switch();
3256 kcov_finish_switch(current
);
3258 fire_sched_in_preempt_notifiers(current
);
3260 * When switching through a kernel thread, the loop in
3261 * membarrier_{private,global}_expedited() may have observed that
3262 * kernel thread and not issued an IPI. It is therefore possible to
3263 * schedule between user->kernel->user threads without passing though
3264 * switch_mm(). Membarrier requires a barrier after storing to
3265 * rq->curr, before returning to userspace, so provide them here:
3267 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3268 * provided by mmdrop(),
3269 * - a sync_core for SYNC_CORE.
3272 membarrier_mm_sync_core_before_usermode(mm
);
3275 if (unlikely(prev_state
== TASK_DEAD
)) {
3276 if (prev
->sched_class
->task_dead
)
3277 prev
->sched_class
->task_dead(prev
);
3280 * Remove function-return probe instances associated with this
3281 * task and put them back on the free list.
3283 kprobe_flush_task(prev
);
3285 /* Task is done with its stack. */
3286 put_task_stack(prev
);
3288 put_task_struct_rcu_user(prev
);
3291 tick_nohz_task_switch();
3297 /* rq->lock is NOT held, but preemption is disabled */
3298 static void __balance_callback(struct rq
*rq
)
3300 struct callback_head
*head
, *next
;
3301 void (*func
)(struct rq
*rq
);
3302 unsigned long flags
;
3304 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3305 head
= rq
->balance_callback
;
3306 rq
->balance_callback
= NULL
;
3308 func
= (void (*)(struct rq
*))head
->func
;
3315 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3318 static inline void balance_callback(struct rq
*rq
)
3320 if (unlikely(rq
->balance_callback
))
3321 __balance_callback(rq
);
3326 static inline void balance_callback(struct rq
*rq
)
3333 * schedule_tail - first thing a freshly forked thread must call.
3334 * @prev: the thread we just switched away from.
3336 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
3337 __releases(rq
->lock
)
3342 * New tasks start with FORK_PREEMPT_COUNT, see there and
3343 * finish_task_switch() for details.
3345 * finish_task_switch() will drop rq->lock() and lower preempt_count
3346 * and the preempt_enable() will end up enabling preemption (on
3347 * PREEMPT_COUNT kernels).
3350 rq
= finish_task_switch(prev
);
3351 balance_callback(rq
);
3354 if (current
->set_child_tid
)
3355 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3357 calculate_sigpending();
3361 * context_switch - switch to the new MM and the new thread's register state.
3363 static __always_inline
struct rq
*
3364 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3365 struct task_struct
*next
, struct rq_flags
*rf
)
3367 prepare_task_switch(rq
, prev
, next
);
3370 * For paravirt, this is coupled with an exit in switch_to to
3371 * combine the page table reload and the switch backend into
3374 arch_start_context_switch(prev
);
3377 * kernel -> kernel lazy + transfer active
3378 * user -> kernel lazy + mmgrab() active
3380 * kernel -> user switch + mmdrop() active
3381 * user -> user switch
3383 if (!next
->mm
) { // to kernel
3384 enter_lazy_tlb(prev
->active_mm
, next
);
3386 next
->active_mm
= prev
->active_mm
;
3387 if (prev
->mm
) // from user
3388 mmgrab(prev
->active_mm
);
3390 prev
->active_mm
= NULL
;
3392 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
3394 * sys_membarrier() requires an smp_mb() between setting
3395 * rq->curr / membarrier_switch_mm() and returning to userspace.
3397 * The below provides this either through switch_mm(), or in
3398 * case 'prev->active_mm == next->mm' through
3399 * finish_task_switch()'s mmdrop().
3401 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
3403 if (!prev
->mm
) { // from kernel
3404 /* will mmdrop() in finish_task_switch(). */
3405 rq
->prev_mm
= prev
->active_mm
;
3406 prev
->active_mm
= NULL
;
3410 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3412 prepare_lock_switch(rq
, next
, rf
);
3414 /* Here we just switch the register state and the stack. */
3415 switch_to(prev
, next
, prev
);
3418 return finish_task_switch(prev
);
3422 * nr_running and nr_context_switches:
3424 * externally visible scheduler statistics: current number of runnable
3425 * threads, total number of context switches performed since bootup.
3427 unsigned long nr_running(void)
3429 unsigned long i
, sum
= 0;
3431 for_each_online_cpu(i
)
3432 sum
+= cpu_rq(i
)->nr_running
;
3438 * Check if only the current task is running on the CPU.
3440 * Caution: this function does not check that the caller has disabled
3441 * preemption, thus the result might have a time-of-check-to-time-of-use
3442 * race. The caller is responsible to use it correctly, for example:
3444 * - from a non-preemptible section (of course)
3446 * - from a thread that is bound to a single CPU
3448 * - in a loop with very short iterations (e.g. a polling loop)
3450 bool single_task_running(void)
3452 return raw_rq()->nr_running
== 1;
3454 EXPORT_SYMBOL(single_task_running
);
3456 unsigned long long nr_context_switches(void)
3459 unsigned long long sum
= 0;
3461 for_each_possible_cpu(i
)
3462 sum
+= cpu_rq(i
)->nr_switches
;
3468 * Consumers of these two interfaces, like for example the cpuidle menu
3469 * governor, are using nonsensical data. Preferring shallow idle state selection
3470 * for a CPU that has IO-wait which might not even end up running the task when
3471 * it does become runnable.
3474 unsigned long nr_iowait_cpu(int cpu
)
3476 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
3480 * IO-wait accounting, and how its mostly bollocks (on SMP).
3482 * The idea behind IO-wait account is to account the idle time that we could
3483 * have spend running if it were not for IO. That is, if we were to improve the
3484 * storage performance, we'd have a proportional reduction in IO-wait time.
3486 * This all works nicely on UP, where, when a task blocks on IO, we account
3487 * idle time as IO-wait, because if the storage were faster, it could've been
3488 * running and we'd not be idle.
3490 * This has been extended to SMP, by doing the same for each CPU. This however
3493 * Imagine for instance the case where two tasks block on one CPU, only the one
3494 * CPU will have IO-wait accounted, while the other has regular idle. Even
3495 * though, if the storage were faster, both could've ran at the same time,
3496 * utilising both CPUs.
3498 * This means, that when looking globally, the current IO-wait accounting on
3499 * SMP is a lower bound, by reason of under accounting.
3501 * Worse, since the numbers are provided per CPU, they are sometimes
3502 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3503 * associated with any one particular CPU, it can wake to another CPU than it
3504 * blocked on. This means the per CPU IO-wait number is meaningless.
3506 * Task CPU affinities can make all that even more 'interesting'.
3509 unsigned long nr_iowait(void)
3511 unsigned long i
, sum
= 0;
3513 for_each_possible_cpu(i
)
3514 sum
+= nr_iowait_cpu(i
);
3522 * sched_exec - execve() is a valuable balancing opportunity, because at
3523 * this point the task has the smallest effective memory and cache footprint.
3525 void sched_exec(void)
3527 struct task_struct
*p
= current
;
3528 unsigned long flags
;
3531 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3532 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
3533 if (dest_cpu
== smp_processor_id())
3536 if (likely(cpu_active(dest_cpu
))) {
3537 struct migration_arg arg
= { p
, dest_cpu
};
3539 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3540 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3544 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3549 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3550 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3552 EXPORT_PER_CPU_SYMBOL(kstat
);
3553 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3556 * The function fair_sched_class.update_curr accesses the struct curr
3557 * and its field curr->exec_start; when called from task_sched_runtime(),
3558 * we observe a high rate of cache misses in practice.
3559 * Prefetching this data results in improved performance.
3561 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3563 #ifdef CONFIG_FAIR_GROUP_SCHED
3564 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3566 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3569 prefetch(&curr
->exec_start
);
3573 * Return accounted runtime for the task.
3574 * In case the task is currently running, return the runtime plus current's
3575 * pending runtime that have not been accounted yet.
3577 unsigned long long task_sched_runtime(struct task_struct
*p
)
3583 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3585 * 64-bit doesn't need locks to atomically read a 64-bit value.
3586 * So we have a optimization chance when the task's delta_exec is 0.
3587 * Reading ->on_cpu is racy, but this is ok.
3589 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3590 * If we race with it entering CPU, unaccounted time is 0. This is
3591 * indistinguishable from the read occurring a few cycles earlier.
3592 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3593 * been accounted, so we're correct here as well.
3595 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3596 return p
->se
.sum_exec_runtime
;
3599 rq
= task_rq_lock(p
, &rf
);
3601 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3602 * project cycles that may never be accounted to this
3603 * thread, breaking clock_gettime().
3605 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3606 prefetch_curr_exec_start(p
);
3607 update_rq_clock(rq
);
3608 p
->sched_class
->update_curr(rq
);
3610 ns
= p
->se
.sum_exec_runtime
;
3611 task_rq_unlock(rq
, p
, &rf
);
3616 DEFINE_PER_CPU(unsigned long, thermal_pressure
);
3618 void arch_set_thermal_pressure(struct cpumask
*cpus
,
3619 unsigned long th_pressure
)
3623 for_each_cpu(cpu
, cpus
)
3624 WRITE_ONCE(per_cpu(thermal_pressure
, cpu
), th_pressure
);
3628 * This function gets called by the timer code, with HZ frequency.
3629 * We call it with interrupts disabled.
3631 void scheduler_tick(void)
3633 int cpu
= smp_processor_id();
3634 struct rq
*rq
= cpu_rq(cpu
);
3635 struct task_struct
*curr
= rq
->curr
;
3637 unsigned long thermal_pressure
;
3639 arch_scale_freq_tick();
3644 update_rq_clock(rq
);
3645 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
3646 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
3647 curr
->sched_class
->task_tick(rq
, curr
, 0);
3648 calc_global_load_tick(rq
);
3653 perf_event_task_tick();
3656 rq
->idle_balance
= idle_cpu(cpu
);
3657 trigger_load_balance(rq
);
3661 #ifdef CONFIG_NO_HZ_FULL
3666 struct delayed_work work
;
3668 /* Values for ->state, see diagram below. */
3669 #define TICK_SCHED_REMOTE_OFFLINE 0
3670 #define TICK_SCHED_REMOTE_OFFLINING 1
3671 #define TICK_SCHED_REMOTE_RUNNING 2
3674 * State diagram for ->state:
3677 * TICK_SCHED_REMOTE_OFFLINE
3680 * | | sched_tick_remote()
3683 * +--TICK_SCHED_REMOTE_OFFLINING
3686 * sched_tick_start() | | sched_tick_stop()
3689 * TICK_SCHED_REMOTE_RUNNING
3692 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3693 * and sched_tick_start() are happy to leave the state in RUNNING.
3696 static struct tick_work __percpu
*tick_work_cpu
;
3698 static void sched_tick_remote(struct work_struct
*work
)
3700 struct delayed_work
*dwork
= to_delayed_work(work
);
3701 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3702 int cpu
= twork
->cpu
;
3703 struct rq
*rq
= cpu_rq(cpu
);
3704 struct task_struct
*curr
;
3710 * Handle the tick only if it appears the remote CPU is running in full
3711 * dynticks mode. The check is racy by nature, but missing a tick or
3712 * having one too much is no big deal because the scheduler tick updates
3713 * statistics and checks timeslices in a time-independent way, regardless
3714 * of when exactly it is running.
3716 if (!tick_nohz_tick_stopped_cpu(cpu
))
3719 rq_lock_irq(rq
, &rf
);
3721 if (cpu_is_offline(cpu
))
3724 update_rq_clock(rq
);
3726 if (!is_idle_task(curr
)) {
3728 * Make sure the next tick runs within a reasonable
3731 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3732 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3734 curr
->sched_class
->task_tick(rq
, curr
, 0);
3736 calc_load_nohz_remote(rq
);
3738 rq_unlock_irq(rq
, &rf
);
3742 * Run the remote tick once per second (1Hz). This arbitrary
3743 * frequency is large enough to avoid overload but short enough
3744 * to keep scheduler internal stats reasonably up to date. But
3745 * first update state to reflect hotplug activity if required.
3747 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
3748 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
3749 if (os
== TICK_SCHED_REMOTE_RUNNING
)
3750 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3753 static void sched_tick_start(int cpu
)
3756 struct tick_work
*twork
;
3758 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3761 WARN_ON_ONCE(!tick_work_cpu
);
3763 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3764 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
3765 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
3766 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
3768 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3769 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3773 #ifdef CONFIG_HOTPLUG_CPU
3774 static void sched_tick_stop(int cpu
)
3776 struct tick_work
*twork
;
3779 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3782 WARN_ON_ONCE(!tick_work_cpu
);
3784 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3785 /* There cannot be competing actions, but don't rely on stop-machine. */
3786 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
3787 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
3788 /* Don't cancel, as this would mess up the state machine. */
3790 #endif /* CONFIG_HOTPLUG_CPU */
3792 int __init
sched_tick_offload_init(void)
3794 tick_work_cpu
= alloc_percpu(struct tick_work
);
3795 BUG_ON(!tick_work_cpu
);
3799 #else /* !CONFIG_NO_HZ_FULL */
3800 static inline void sched_tick_start(int cpu
) { }
3801 static inline void sched_tick_stop(int cpu
) { }
3804 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3805 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3807 * If the value passed in is equal to the current preempt count
3808 * then we just disabled preemption. Start timing the latency.
3810 static inline void preempt_latency_start(int val
)
3812 if (preempt_count() == val
) {
3813 unsigned long ip
= get_lock_parent_ip();
3814 #ifdef CONFIG_DEBUG_PREEMPT
3815 current
->preempt_disable_ip
= ip
;
3817 trace_preempt_off(CALLER_ADDR0
, ip
);
3821 void preempt_count_add(int val
)
3823 #ifdef CONFIG_DEBUG_PREEMPT
3827 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3830 __preempt_count_add(val
);
3831 #ifdef CONFIG_DEBUG_PREEMPT
3833 * Spinlock count overflowing soon?
3835 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3838 preempt_latency_start(val
);
3840 EXPORT_SYMBOL(preempt_count_add
);
3841 NOKPROBE_SYMBOL(preempt_count_add
);
3844 * If the value passed in equals to the current preempt count
3845 * then we just enabled preemption. Stop timing the latency.
3847 static inline void preempt_latency_stop(int val
)
3849 if (preempt_count() == val
)
3850 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3853 void preempt_count_sub(int val
)
3855 #ifdef CONFIG_DEBUG_PREEMPT
3859 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3862 * Is the spinlock portion underflowing?
3864 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3865 !(preempt_count() & PREEMPT_MASK
)))
3869 preempt_latency_stop(val
);
3870 __preempt_count_sub(val
);
3872 EXPORT_SYMBOL(preempt_count_sub
);
3873 NOKPROBE_SYMBOL(preempt_count_sub
);
3876 static inline void preempt_latency_start(int val
) { }
3877 static inline void preempt_latency_stop(int val
) { }
3880 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3882 #ifdef CONFIG_DEBUG_PREEMPT
3883 return p
->preempt_disable_ip
;
3890 * Print scheduling while atomic bug:
3892 static noinline
void __schedule_bug(struct task_struct
*prev
)
3894 /* Save this before calling printk(), since that will clobber it */
3895 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3897 if (oops_in_progress
)
3900 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3901 prev
->comm
, prev
->pid
, preempt_count());
3903 debug_show_held_locks(prev
);
3905 if (irqs_disabled())
3906 print_irqtrace_events(prev
);
3907 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3908 && in_atomic_preempt_off()) {
3909 pr_err("Preemption disabled at:");
3910 print_ip_sym(preempt_disable_ip
);
3914 panic("scheduling while atomic\n");
3917 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3921 * Various schedule()-time debugging checks and statistics:
3923 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
3925 #ifdef CONFIG_SCHED_STACK_END_CHECK
3926 if (task_stack_end_corrupted(prev
))
3927 panic("corrupted stack end detected inside scheduler\n");
3930 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3931 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
3932 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3933 prev
->comm
, prev
->pid
, prev
->non_block_count
);
3935 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3939 if (unlikely(in_atomic_preempt_off())) {
3940 __schedule_bug(prev
);
3941 preempt_count_set(PREEMPT_DISABLED
);
3945 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3947 schedstat_inc(this_rq()->sched_count
);
3951 * Pick up the highest-prio task:
3953 static inline struct task_struct
*
3954 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3956 const struct sched_class
*class;
3957 struct task_struct
*p
;
3960 * Optimization: we know that if all tasks are in the fair class we can
3961 * call that function directly, but only if the @prev task wasn't of a
3962 * higher scheduling class, because otherwise those loose the
3963 * opportunity to pull in more work from other CPUs.
3965 if (likely((prev
->sched_class
== &idle_sched_class
||
3966 prev
->sched_class
== &fair_sched_class
) &&
3967 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3969 p
= pick_next_task_fair(rq
, prev
, rf
);
3970 if (unlikely(p
== RETRY_TASK
))
3973 /* Assumes fair_sched_class->next == idle_sched_class */
3975 put_prev_task(rq
, prev
);
3976 p
= pick_next_task_idle(rq
);
3985 * We must do the balancing pass before put_next_task(), such
3986 * that when we release the rq->lock the task is in the same
3987 * state as before we took rq->lock.
3989 * We can terminate the balance pass as soon as we know there is
3990 * a runnable task of @class priority or higher.
3992 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
3993 if (class->balance(rq
, prev
, rf
))
3998 put_prev_task(rq
, prev
);
4000 for_each_class(class) {
4001 p
= class->pick_next_task(rq
);
4006 /* The idle class should always have a runnable task: */
4011 * __schedule() is the main scheduler function.
4013 * The main means of driving the scheduler and thus entering this function are:
4015 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4017 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4018 * paths. For example, see arch/x86/entry_64.S.
4020 * To drive preemption between tasks, the scheduler sets the flag in timer
4021 * interrupt handler scheduler_tick().
4023 * 3. Wakeups don't really cause entry into schedule(). They add a
4024 * task to the run-queue and that's it.
4026 * Now, if the new task added to the run-queue preempts the current
4027 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4028 * called on the nearest possible occasion:
4030 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4032 * - in syscall or exception context, at the next outmost
4033 * preempt_enable(). (this might be as soon as the wake_up()'s
4036 * - in IRQ context, return from interrupt-handler to
4037 * preemptible context
4039 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4042 * - cond_resched() call
4043 * - explicit schedule() call
4044 * - return from syscall or exception to user-space
4045 * - return from interrupt-handler to user-space
4047 * WARNING: must be called with preemption disabled!
4049 static void __sched notrace
__schedule(bool preempt
)
4051 struct task_struct
*prev
, *next
;
4052 unsigned long *switch_count
;
4057 cpu
= smp_processor_id();
4061 schedule_debug(prev
, preempt
);
4063 if (sched_feat(HRTICK
))
4066 local_irq_disable();
4067 rcu_note_context_switch(preempt
);
4070 * Make sure that signal_pending_state()->signal_pending() below
4071 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4072 * done by the caller to avoid the race with signal_wake_up().
4074 * The membarrier system call requires a full memory barrier
4075 * after coming from user-space, before storing to rq->curr.
4078 smp_mb__after_spinlock();
4080 /* Promote REQ to ACT */
4081 rq
->clock_update_flags
<<= 1;
4082 update_rq_clock(rq
);
4084 switch_count
= &prev
->nivcsw
;
4085 if (!preempt
&& prev
->state
) {
4086 if (signal_pending_state(prev
->state
, prev
)) {
4087 prev
->state
= TASK_RUNNING
;
4089 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
4091 if (prev
->in_iowait
) {
4092 atomic_inc(&rq
->nr_iowait
);
4093 delayacct_blkio_start();
4096 switch_count
= &prev
->nvcsw
;
4099 next
= pick_next_task(rq
, prev
, &rf
);
4100 clear_tsk_need_resched(prev
);
4101 clear_preempt_need_resched();
4103 if (likely(prev
!= next
)) {
4106 * RCU users of rcu_dereference(rq->curr) may not see
4107 * changes to task_struct made by pick_next_task().
4109 RCU_INIT_POINTER(rq
->curr
, next
);
4111 * The membarrier system call requires each architecture
4112 * to have a full memory barrier after updating
4113 * rq->curr, before returning to user-space.
4115 * Here are the schemes providing that barrier on the
4116 * various architectures:
4117 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4118 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4119 * - finish_lock_switch() for weakly-ordered
4120 * architectures where spin_unlock is a full barrier,
4121 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4122 * is a RELEASE barrier),
4126 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
4128 trace_sched_switch(preempt
, prev
, next
);
4130 /* Also unlocks the rq: */
4131 rq
= context_switch(rq
, prev
, next
, &rf
);
4133 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4134 rq_unlock_irq(rq
, &rf
);
4137 balance_callback(rq
);
4140 void __noreturn
do_task_dead(void)
4142 /* Causes final put_task_struct in finish_task_switch(): */
4143 set_special_state(TASK_DEAD
);
4145 /* Tell freezer to ignore us: */
4146 current
->flags
|= PF_NOFREEZE
;
4151 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4156 static inline void sched_submit_work(struct task_struct
*tsk
)
4162 * If a worker went to sleep, notify and ask workqueue whether
4163 * it wants to wake up a task to maintain concurrency.
4164 * As this function is called inside the schedule() context,
4165 * we disable preemption to avoid it calling schedule() again
4166 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4169 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4171 if (tsk
->flags
& PF_WQ_WORKER
)
4172 wq_worker_sleeping(tsk
);
4174 io_wq_worker_sleeping(tsk
);
4175 preempt_enable_no_resched();
4178 if (tsk_is_pi_blocked(tsk
))
4182 * If we are going to sleep and we have plugged IO queued,
4183 * make sure to submit it to avoid deadlocks.
4185 if (blk_needs_flush_plug(tsk
))
4186 blk_schedule_flush_plug(tsk
);
4189 static void sched_update_worker(struct task_struct
*tsk
)
4191 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4192 if (tsk
->flags
& PF_WQ_WORKER
)
4193 wq_worker_running(tsk
);
4195 io_wq_worker_running(tsk
);
4199 asmlinkage __visible
void __sched
schedule(void)
4201 struct task_struct
*tsk
= current
;
4203 sched_submit_work(tsk
);
4207 sched_preempt_enable_no_resched();
4208 } while (need_resched());
4209 sched_update_worker(tsk
);
4211 EXPORT_SYMBOL(schedule
);
4214 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4215 * state (have scheduled out non-voluntarily) by making sure that all
4216 * tasks have either left the run queue or have gone into user space.
4217 * As idle tasks do not do either, they must not ever be preempted
4218 * (schedule out non-voluntarily).
4220 * schedule_idle() is similar to schedule_preempt_disable() except that it
4221 * never enables preemption because it does not call sched_submit_work().
4223 void __sched
schedule_idle(void)
4226 * As this skips calling sched_submit_work(), which the idle task does
4227 * regardless because that function is a nop when the task is in a
4228 * TASK_RUNNING state, make sure this isn't used someplace that the
4229 * current task can be in any other state. Note, idle is always in the
4230 * TASK_RUNNING state.
4232 WARN_ON_ONCE(current
->state
);
4235 } while (need_resched());
4238 #ifdef CONFIG_CONTEXT_TRACKING
4239 asmlinkage __visible
void __sched
schedule_user(void)
4242 * If we come here after a random call to set_need_resched(),
4243 * or we have been woken up remotely but the IPI has not yet arrived,
4244 * we haven't yet exited the RCU idle mode. Do it here manually until
4245 * we find a better solution.
4247 * NB: There are buggy callers of this function. Ideally we
4248 * should warn if prev_state != CONTEXT_USER, but that will trigger
4249 * too frequently to make sense yet.
4251 enum ctx_state prev_state
= exception_enter();
4253 exception_exit(prev_state
);
4258 * schedule_preempt_disabled - called with preemption disabled
4260 * Returns with preemption disabled. Note: preempt_count must be 1
4262 void __sched
schedule_preempt_disabled(void)
4264 sched_preempt_enable_no_resched();
4269 static void __sched notrace
preempt_schedule_common(void)
4273 * Because the function tracer can trace preempt_count_sub()
4274 * and it also uses preempt_enable/disable_notrace(), if
4275 * NEED_RESCHED is set, the preempt_enable_notrace() called
4276 * by the function tracer will call this function again and
4277 * cause infinite recursion.
4279 * Preemption must be disabled here before the function
4280 * tracer can trace. Break up preempt_disable() into two
4281 * calls. One to disable preemption without fear of being
4282 * traced. The other to still record the preemption latency,
4283 * which can also be traced by the function tracer.
4285 preempt_disable_notrace();
4286 preempt_latency_start(1);
4288 preempt_latency_stop(1);
4289 preempt_enable_no_resched_notrace();
4292 * Check again in case we missed a preemption opportunity
4293 * between schedule and now.
4295 } while (need_resched());
4298 #ifdef CONFIG_PREEMPTION
4300 * This is the entry point to schedule() from in-kernel preemption
4301 * off of preempt_enable.
4303 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
4306 * If there is a non-zero preempt_count or interrupts are disabled,
4307 * we do not want to preempt the current task. Just return..
4309 if (likely(!preemptible()))
4312 preempt_schedule_common();
4314 NOKPROBE_SYMBOL(preempt_schedule
);
4315 EXPORT_SYMBOL(preempt_schedule
);
4318 * preempt_schedule_notrace - preempt_schedule called by tracing
4320 * The tracing infrastructure uses preempt_enable_notrace to prevent
4321 * recursion and tracing preempt enabling caused by the tracing
4322 * infrastructure itself. But as tracing can happen in areas coming
4323 * from userspace or just about to enter userspace, a preempt enable
4324 * can occur before user_exit() is called. This will cause the scheduler
4325 * to be called when the system is still in usermode.
4327 * To prevent this, the preempt_enable_notrace will use this function
4328 * instead of preempt_schedule() to exit user context if needed before
4329 * calling the scheduler.
4331 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
4333 enum ctx_state prev_ctx
;
4335 if (likely(!preemptible()))
4340 * Because the function tracer can trace preempt_count_sub()
4341 * and it also uses preempt_enable/disable_notrace(), if
4342 * NEED_RESCHED is set, the preempt_enable_notrace() called
4343 * by the function tracer will call this function again and
4344 * cause infinite recursion.
4346 * Preemption must be disabled here before the function
4347 * tracer can trace. Break up preempt_disable() into two
4348 * calls. One to disable preemption without fear of being
4349 * traced. The other to still record the preemption latency,
4350 * which can also be traced by the function tracer.
4352 preempt_disable_notrace();
4353 preempt_latency_start(1);
4355 * Needs preempt disabled in case user_exit() is traced
4356 * and the tracer calls preempt_enable_notrace() causing
4357 * an infinite recursion.
4359 prev_ctx
= exception_enter();
4361 exception_exit(prev_ctx
);
4363 preempt_latency_stop(1);
4364 preempt_enable_no_resched_notrace();
4365 } while (need_resched());
4367 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
4369 #endif /* CONFIG_PREEMPTION */
4372 * This is the entry point to schedule() from kernel preemption
4373 * off of irq context.
4374 * Note, that this is called and return with irqs disabled. This will
4375 * protect us against recursive calling from irq.
4377 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
4379 enum ctx_state prev_state
;
4381 /* Catch callers which need to be fixed */
4382 BUG_ON(preempt_count() || !irqs_disabled());
4384 prev_state
= exception_enter();
4390 local_irq_disable();
4391 sched_preempt_enable_no_resched();
4392 } while (need_resched());
4394 exception_exit(prev_state
);
4397 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
4400 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4402 EXPORT_SYMBOL(default_wake_function
);
4404 #ifdef CONFIG_RT_MUTEXES
4406 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
4409 prio
= min(prio
, pi_task
->prio
);
4414 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4416 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
4418 return __rt_effective_prio(pi_task
, prio
);
4422 * rt_mutex_setprio - set the current priority of a task
4424 * @pi_task: donor task
4426 * This function changes the 'effective' priority of a task. It does
4427 * not touch ->normal_prio like __setscheduler().
4429 * Used by the rt_mutex code to implement priority inheritance
4430 * logic. Call site only calls if the priority of the task changed.
4432 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
4434 int prio
, oldprio
, queued
, running
, queue_flag
=
4435 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4436 const struct sched_class
*prev_class
;
4440 /* XXX used to be waiter->prio, not waiter->task->prio */
4441 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
4444 * If nothing changed; bail early.
4446 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
4449 rq
= __task_rq_lock(p
, &rf
);
4450 update_rq_clock(rq
);
4452 * Set under pi_lock && rq->lock, such that the value can be used under
4455 * Note that there is loads of tricky to make this pointer cache work
4456 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4457 * ensure a task is de-boosted (pi_task is set to NULL) before the
4458 * task is allowed to run again (and can exit). This ensures the pointer
4459 * points to a blocked task -- which guaratees the task is present.
4461 p
->pi_top_task
= pi_task
;
4464 * For FIFO/RR we only need to set prio, if that matches we're done.
4466 if (prio
== p
->prio
&& !dl_prio(prio
))
4470 * Idle task boosting is a nono in general. There is one
4471 * exception, when PREEMPT_RT and NOHZ is active:
4473 * The idle task calls get_next_timer_interrupt() and holds
4474 * the timer wheel base->lock on the CPU and another CPU wants
4475 * to access the timer (probably to cancel it). We can safely
4476 * ignore the boosting request, as the idle CPU runs this code
4477 * with interrupts disabled and will complete the lock
4478 * protected section without being interrupted. So there is no
4479 * real need to boost.
4481 if (unlikely(p
== rq
->idle
)) {
4482 WARN_ON(p
!= rq
->curr
);
4483 WARN_ON(p
->pi_blocked_on
);
4487 trace_sched_pi_setprio(p
, pi_task
);
4490 if (oldprio
== prio
)
4491 queue_flag
&= ~DEQUEUE_MOVE
;
4493 prev_class
= p
->sched_class
;
4494 queued
= task_on_rq_queued(p
);
4495 running
= task_current(rq
, p
);
4497 dequeue_task(rq
, p
, queue_flag
);
4499 put_prev_task(rq
, p
);
4502 * Boosting condition are:
4503 * 1. -rt task is running and holds mutex A
4504 * --> -dl task blocks on mutex A
4506 * 2. -dl task is running and holds mutex A
4507 * --> -dl task blocks on mutex A and could preempt the
4510 if (dl_prio(prio
)) {
4511 if (!dl_prio(p
->normal_prio
) ||
4512 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
4513 p
->dl
.dl_boosted
= 1;
4514 queue_flag
|= ENQUEUE_REPLENISH
;
4516 p
->dl
.dl_boosted
= 0;
4517 p
->sched_class
= &dl_sched_class
;
4518 } else if (rt_prio(prio
)) {
4519 if (dl_prio(oldprio
))
4520 p
->dl
.dl_boosted
= 0;
4522 queue_flag
|= ENQUEUE_HEAD
;
4523 p
->sched_class
= &rt_sched_class
;
4525 if (dl_prio(oldprio
))
4526 p
->dl
.dl_boosted
= 0;
4527 if (rt_prio(oldprio
))
4529 p
->sched_class
= &fair_sched_class
;
4535 enqueue_task(rq
, p
, queue_flag
);
4537 set_next_task(rq
, p
);
4539 check_class_changed(rq
, p
, prev_class
, oldprio
);
4541 /* Avoid rq from going away on us: */
4543 __task_rq_unlock(rq
, &rf
);
4545 balance_callback(rq
);
4549 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4555 void set_user_nice(struct task_struct
*p
, long nice
)
4557 bool queued
, running
;
4562 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
4565 * We have to be careful, if called from sys_setpriority(),
4566 * the task might be in the middle of scheduling on another CPU.
4568 rq
= task_rq_lock(p
, &rf
);
4569 update_rq_clock(rq
);
4572 * The RT priorities are set via sched_setscheduler(), but we still
4573 * allow the 'normal' nice value to be set - but as expected
4574 * it wont have any effect on scheduling until the task is
4575 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4577 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
4578 p
->static_prio
= NICE_TO_PRIO(nice
);
4581 queued
= task_on_rq_queued(p
);
4582 running
= task_current(rq
, p
);
4584 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
4586 put_prev_task(rq
, p
);
4588 p
->static_prio
= NICE_TO_PRIO(nice
);
4589 set_load_weight(p
, true);
4591 p
->prio
= effective_prio(p
);
4594 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
4596 set_next_task(rq
, p
);
4599 * If the task increased its priority or is running and
4600 * lowered its priority, then reschedule its CPU:
4602 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
4605 task_rq_unlock(rq
, p
, &rf
);
4607 EXPORT_SYMBOL(set_user_nice
);
4610 * can_nice - check if a task can reduce its nice value
4614 int can_nice(const struct task_struct
*p
, const int nice
)
4616 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4617 int nice_rlim
= nice_to_rlimit(nice
);
4619 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4620 capable(CAP_SYS_NICE
));
4623 #ifdef __ARCH_WANT_SYS_NICE
4626 * sys_nice - change the priority of the current process.
4627 * @increment: priority increment
4629 * sys_setpriority is a more generic, but much slower function that
4630 * does similar things.
4632 SYSCALL_DEFINE1(nice
, int, increment
)
4637 * Setpriority might change our priority at the same moment.
4638 * We don't have to worry. Conceptually one call occurs first
4639 * and we have a single winner.
4641 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
4642 nice
= task_nice(current
) + increment
;
4644 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
4645 if (increment
< 0 && !can_nice(current
, nice
))
4648 retval
= security_task_setnice(current
, nice
);
4652 set_user_nice(current
, nice
);
4659 * task_prio - return the priority value of a given task.
4660 * @p: the task in question.
4662 * Return: The priority value as seen by users in /proc.
4663 * RT tasks are offset by -200. Normal tasks are centered
4664 * around 0, value goes from -16 to +15.
4666 int task_prio(const struct task_struct
*p
)
4668 return p
->prio
- MAX_RT_PRIO
;
4672 * idle_cpu - is a given CPU idle currently?
4673 * @cpu: the processor in question.
4675 * Return: 1 if the CPU is currently idle. 0 otherwise.
4677 int idle_cpu(int cpu
)
4679 struct rq
*rq
= cpu_rq(cpu
);
4681 if (rq
->curr
!= rq
->idle
)
4688 if (!llist_empty(&rq
->wake_list
))
4696 * available_idle_cpu - is a given CPU idle for enqueuing work.
4697 * @cpu: the CPU in question.
4699 * Return: 1 if the CPU is currently idle. 0 otherwise.
4701 int available_idle_cpu(int cpu
)
4706 if (vcpu_is_preempted(cpu
))
4713 * idle_task - return the idle task for a given CPU.
4714 * @cpu: the processor in question.
4716 * Return: The idle task for the CPU @cpu.
4718 struct task_struct
*idle_task(int cpu
)
4720 return cpu_rq(cpu
)->idle
;
4724 * find_process_by_pid - find a process with a matching PID value.
4725 * @pid: the pid in question.
4727 * The task of @pid, if found. %NULL otherwise.
4729 static struct task_struct
*find_process_by_pid(pid_t pid
)
4731 return pid
? find_task_by_vpid(pid
) : current
;
4735 * sched_setparam() passes in -1 for its policy, to let the functions
4736 * it calls know not to change it.
4738 #define SETPARAM_POLICY -1
4740 static void __setscheduler_params(struct task_struct
*p
,
4741 const struct sched_attr
*attr
)
4743 int policy
= attr
->sched_policy
;
4745 if (policy
== SETPARAM_POLICY
)
4750 if (dl_policy(policy
))
4751 __setparam_dl(p
, attr
);
4752 else if (fair_policy(policy
))
4753 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4756 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4757 * !rt_policy. Always setting this ensures that things like
4758 * getparam()/getattr() don't report silly values for !rt tasks.
4760 p
->rt_priority
= attr
->sched_priority
;
4761 p
->normal_prio
= normal_prio(p
);
4762 set_load_weight(p
, true);
4765 /* Actually do priority change: must hold pi & rq lock. */
4766 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4767 const struct sched_attr
*attr
, bool keep_boost
)
4770 * If params can't change scheduling class changes aren't allowed
4773 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
4776 __setscheduler_params(p
, attr
);
4779 * Keep a potential priority boosting if called from
4780 * sched_setscheduler().
4782 p
->prio
= normal_prio(p
);
4784 p
->prio
= rt_effective_prio(p
, p
->prio
);
4786 if (dl_prio(p
->prio
))
4787 p
->sched_class
= &dl_sched_class
;
4788 else if (rt_prio(p
->prio
))
4789 p
->sched_class
= &rt_sched_class
;
4791 p
->sched_class
= &fair_sched_class
;
4795 * Check the target process has a UID that matches the current process's:
4797 static bool check_same_owner(struct task_struct
*p
)
4799 const struct cred
*cred
= current_cred(), *pcred
;
4803 pcred
= __task_cred(p
);
4804 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4805 uid_eq(cred
->euid
, pcred
->uid
));
4810 static int __sched_setscheduler(struct task_struct
*p
,
4811 const struct sched_attr
*attr
,
4814 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4815 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4816 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4817 int new_effective_prio
, policy
= attr
->sched_policy
;
4818 const struct sched_class
*prev_class
;
4821 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4824 /* The pi code expects interrupts enabled */
4825 BUG_ON(pi
&& in_interrupt());
4827 /* Double check policy once rq lock held: */
4829 reset_on_fork
= p
->sched_reset_on_fork
;
4830 policy
= oldpolicy
= p
->policy
;
4832 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4834 if (!valid_policy(policy
))
4838 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4842 * Valid priorities for SCHED_FIFO and SCHED_RR are
4843 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4844 * SCHED_BATCH and SCHED_IDLE is 0.
4846 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4847 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4849 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4850 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4854 * Allow unprivileged RT tasks to decrease priority:
4856 if (user
&& !capable(CAP_SYS_NICE
)) {
4857 if (fair_policy(policy
)) {
4858 if (attr
->sched_nice
< task_nice(p
) &&
4859 !can_nice(p
, attr
->sched_nice
))
4863 if (rt_policy(policy
)) {
4864 unsigned long rlim_rtprio
=
4865 task_rlimit(p
, RLIMIT_RTPRIO
);
4867 /* Can't set/change the rt policy: */
4868 if (policy
!= p
->policy
&& !rlim_rtprio
)
4871 /* Can't increase priority: */
4872 if (attr
->sched_priority
> p
->rt_priority
&&
4873 attr
->sched_priority
> rlim_rtprio
)
4878 * Can't set/change SCHED_DEADLINE policy at all for now
4879 * (safest behavior); in the future we would like to allow
4880 * unprivileged DL tasks to increase their relative deadline
4881 * or reduce their runtime (both ways reducing utilization)
4883 if (dl_policy(policy
))
4887 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4888 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4890 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
4891 if (!can_nice(p
, task_nice(p
)))
4895 /* Can't change other user's priorities: */
4896 if (!check_same_owner(p
))
4899 /* Normal users shall not reset the sched_reset_on_fork flag: */
4900 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4905 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4908 retval
= security_task_setscheduler(p
);
4913 /* Update task specific "requested" clamps */
4914 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
4915 retval
= uclamp_validate(p
, attr
);
4924 * Make sure no PI-waiters arrive (or leave) while we are
4925 * changing the priority of the task:
4927 * To be able to change p->policy safely, the appropriate
4928 * runqueue lock must be held.
4930 rq
= task_rq_lock(p
, &rf
);
4931 update_rq_clock(rq
);
4934 * Changing the policy of the stop threads its a very bad idea:
4936 if (p
== rq
->stop
) {
4942 * If not changing anything there's no need to proceed further,
4943 * but store a possible modification of reset_on_fork.
4945 if (unlikely(policy
== p
->policy
)) {
4946 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4948 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4950 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4952 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
4955 p
->sched_reset_on_fork
= reset_on_fork
;
4962 #ifdef CONFIG_RT_GROUP_SCHED
4964 * Do not allow realtime tasks into groups that have no runtime
4967 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4968 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4969 !task_group_is_autogroup(task_group(p
))) {
4975 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
4976 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
4977 cpumask_t
*span
= rq
->rd
->span
;
4980 * Don't allow tasks with an affinity mask smaller than
4981 * the entire root_domain to become SCHED_DEADLINE. We
4982 * will also fail if there's no bandwidth available.
4984 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
4985 rq
->rd
->dl_bw
.bw
== 0) {
4993 /* Re-check policy now with rq lock held: */
4994 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4995 policy
= oldpolicy
= -1;
4996 task_rq_unlock(rq
, p
, &rf
);
4998 cpuset_read_unlock();
5003 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5004 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5007 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
5012 p
->sched_reset_on_fork
= reset_on_fork
;
5017 * Take priority boosted tasks into account. If the new
5018 * effective priority is unchanged, we just store the new
5019 * normal parameters and do not touch the scheduler class and
5020 * the runqueue. This will be done when the task deboost
5023 new_effective_prio
= rt_effective_prio(p
, newprio
);
5024 if (new_effective_prio
== oldprio
)
5025 queue_flags
&= ~DEQUEUE_MOVE
;
5028 queued
= task_on_rq_queued(p
);
5029 running
= task_current(rq
, p
);
5031 dequeue_task(rq
, p
, queue_flags
);
5033 put_prev_task(rq
, p
);
5035 prev_class
= p
->sched_class
;
5037 __setscheduler(rq
, p
, attr
, pi
);
5038 __setscheduler_uclamp(p
, attr
);
5042 * We enqueue to tail when the priority of a task is
5043 * increased (user space view).
5045 if (oldprio
< p
->prio
)
5046 queue_flags
|= ENQUEUE_HEAD
;
5048 enqueue_task(rq
, p
, queue_flags
);
5051 set_next_task(rq
, p
);
5053 check_class_changed(rq
, p
, prev_class
, oldprio
);
5055 /* Avoid rq from going away on us: */
5057 task_rq_unlock(rq
, p
, &rf
);
5060 cpuset_read_unlock();
5061 rt_mutex_adjust_pi(p
);
5064 /* Run balance callbacks after we've adjusted the PI chain: */
5065 balance_callback(rq
);
5071 task_rq_unlock(rq
, p
, &rf
);
5073 cpuset_read_unlock();
5077 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
5078 const struct sched_param
*param
, bool check
)
5080 struct sched_attr attr
= {
5081 .sched_policy
= policy
,
5082 .sched_priority
= param
->sched_priority
,
5083 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
5086 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5087 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
5088 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5089 policy
&= ~SCHED_RESET_ON_FORK
;
5090 attr
.sched_policy
= policy
;
5093 return __sched_setscheduler(p
, &attr
, check
, true);
5096 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5097 * @p: the task in question.
5098 * @policy: new policy.
5099 * @param: structure containing the new RT priority.
5101 * Return: 0 on success. An error code otherwise.
5103 * NOTE that the task may be already dead.
5105 int sched_setscheduler(struct task_struct
*p
, int policy
,
5106 const struct sched_param
*param
)
5108 return _sched_setscheduler(p
, policy
, param
, true);
5110 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5112 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
5114 return __sched_setscheduler(p
, attr
, true, true);
5116 EXPORT_SYMBOL_GPL(sched_setattr
);
5118 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
5120 return __sched_setscheduler(p
, attr
, false, true);
5124 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5125 * @p: the task in question.
5126 * @policy: new policy.
5127 * @param: structure containing the new RT priority.
5129 * Just like sched_setscheduler, only don't bother checking if the
5130 * current context has permission. For example, this is needed in
5131 * stop_machine(): we create temporary high priority worker threads,
5132 * but our caller might not have that capability.
5134 * Return: 0 on success. An error code otherwise.
5136 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5137 const struct sched_param
*param
)
5139 return _sched_setscheduler(p
, policy
, param
, false);
5141 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
5144 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5146 struct sched_param lparam
;
5147 struct task_struct
*p
;
5150 if (!param
|| pid
< 0)
5152 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5157 p
= find_process_by_pid(pid
);
5163 retval
= sched_setscheduler(p
, policy
, &lparam
);
5171 * Mimics kernel/events/core.c perf_copy_attr().
5173 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
5178 /* Zero the full structure, so that a short copy will be nice: */
5179 memset(attr
, 0, sizeof(*attr
));
5181 ret
= get_user(size
, &uattr
->size
);
5185 /* ABI compatibility quirk: */
5187 size
= SCHED_ATTR_SIZE_VER0
;
5188 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
5191 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
5198 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
5199 size
< SCHED_ATTR_SIZE_VER1
)
5203 * XXX: Do we want to be lenient like existing syscalls; or do we want
5204 * to be strict and return an error on out-of-bounds values?
5206 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
5211 put_user(sizeof(*attr
), &uattr
->size
);
5216 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5217 * @pid: the pid in question.
5218 * @policy: new policy.
5219 * @param: structure containing the new RT priority.
5221 * Return: 0 on success. An error code otherwise.
5223 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
5228 return do_sched_setscheduler(pid
, policy
, param
);
5232 * sys_sched_setparam - set/change the RT priority of a thread
5233 * @pid: the pid in question.
5234 * @param: structure containing the new RT priority.
5236 * Return: 0 on success. An error code otherwise.
5238 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5240 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
5244 * sys_sched_setattr - same as above, but with extended sched_attr
5245 * @pid: the pid in question.
5246 * @uattr: structure containing the extended parameters.
5247 * @flags: for future extension.
5249 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5250 unsigned int, flags
)
5252 struct sched_attr attr
;
5253 struct task_struct
*p
;
5256 if (!uattr
|| pid
< 0 || flags
)
5259 retval
= sched_copy_attr(uattr
, &attr
);
5263 if ((int)attr
.sched_policy
< 0)
5265 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
5266 attr
.sched_policy
= SETPARAM_POLICY
;
5270 p
= find_process_by_pid(pid
);
5276 retval
= sched_setattr(p
, &attr
);
5284 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5285 * @pid: the pid in question.
5287 * Return: On success, the policy of the thread. Otherwise, a negative error
5290 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5292 struct task_struct
*p
;
5300 p
= find_process_by_pid(pid
);
5302 retval
= security_task_getscheduler(p
);
5305 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5312 * sys_sched_getparam - get the RT priority of a thread
5313 * @pid: the pid in question.
5314 * @param: structure containing the RT priority.
5316 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5319 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5321 struct sched_param lp
= { .sched_priority
= 0 };
5322 struct task_struct
*p
;
5325 if (!param
|| pid
< 0)
5329 p
= find_process_by_pid(pid
);
5334 retval
= security_task_getscheduler(p
);
5338 if (task_has_rt_policy(p
))
5339 lp
.sched_priority
= p
->rt_priority
;
5343 * This one might sleep, we cannot do it with a spinlock held ...
5345 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5355 * Copy the kernel size attribute structure (which might be larger
5356 * than what user-space knows about) to user-space.
5358 * Note that all cases are valid: user-space buffer can be larger or
5359 * smaller than the kernel-space buffer. The usual case is that both
5360 * have the same size.
5363 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
5364 struct sched_attr
*kattr
,
5367 unsigned int ksize
= sizeof(*kattr
);
5369 if (!access_ok(uattr
, usize
))
5373 * sched_getattr() ABI forwards and backwards compatibility:
5375 * If usize == ksize then we just copy everything to user-space and all is good.
5377 * If usize < ksize then we only copy as much as user-space has space for,
5378 * this keeps ABI compatibility as well. We skip the rest.
5380 * If usize > ksize then user-space is using a newer version of the ABI,
5381 * which part the kernel doesn't know about. Just ignore it - tooling can
5382 * detect the kernel's knowledge of attributes from the attr->size value
5383 * which is set to ksize in this case.
5385 kattr
->size
= min(usize
, ksize
);
5387 if (copy_to_user(uattr
, kattr
, kattr
->size
))
5394 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5395 * @pid: the pid in question.
5396 * @uattr: structure containing the extended parameters.
5397 * @usize: sizeof(attr) for fwd/bwd comp.
5398 * @flags: for future extension.
5400 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5401 unsigned int, usize
, unsigned int, flags
)
5403 struct sched_attr kattr
= { };
5404 struct task_struct
*p
;
5407 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
5408 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
5412 p
= find_process_by_pid(pid
);
5417 retval
= security_task_getscheduler(p
);
5421 kattr
.sched_policy
= p
->policy
;
5422 if (p
->sched_reset_on_fork
)
5423 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5424 if (task_has_dl_policy(p
))
5425 __getparam_dl(p
, &kattr
);
5426 else if (task_has_rt_policy(p
))
5427 kattr
.sched_priority
= p
->rt_priority
;
5429 kattr
.sched_nice
= task_nice(p
);
5431 #ifdef CONFIG_UCLAMP_TASK
5432 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
5433 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
5438 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
5445 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5447 cpumask_var_t cpus_allowed
, new_mask
;
5448 struct task_struct
*p
;
5453 p
= find_process_by_pid(pid
);
5459 /* Prevent p going away */
5463 if (p
->flags
& PF_NO_SETAFFINITY
) {
5467 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5471 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5473 goto out_free_cpus_allowed
;
5476 if (!check_same_owner(p
)) {
5478 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
5480 goto out_free_new_mask
;
5485 retval
= security_task_setscheduler(p
);
5487 goto out_free_new_mask
;
5490 cpuset_cpus_allowed(p
, cpus_allowed
);
5491 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5494 * Since bandwidth control happens on root_domain basis,
5495 * if admission test is enabled, we only admit -deadline
5496 * tasks allowed to run on all the CPUs in the task's
5500 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
5502 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
5505 goto out_free_new_mask
;
5511 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
5514 cpuset_cpus_allowed(p
, cpus_allowed
);
5515 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5517 * We must have raced with a concurrent cpuset
5518 * update. Just reset the cpus_allowed to the
5519 * cpuset's cpus_allowed
5521 cpumask_copy(new_mask
, cpus_allowed
);
5526 free_cpumask_var(new_mask
);
5527 out_free_cpus_allowed
:
5528 free_cpumask_var(cpus_allowed
);
5534 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5535 struct cpumask
*new_mask
)
5537 if (len
< cpumask_size())
5538 cpumask_clear(new_mask
);
5539 else if (len
> cpumask_size())
5540 len
= cpumask_size();
5542 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5546 * sys_sched_setaffinity - set the CPU affinity of a process
5547 * @pid: pid of the process
5548 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5549 * @user_mask_ptr: user-space pointer to the new CPU mask
5551 * Return: 0 on success. An error code otherwise.
5553 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5554 unsigned long __user
*, user_mask_ptr
)
5556 cpumask_var_t new_mask
;
5559 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5562 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5564 retval
= sched_setaffinity(pid
, new_mask
);
5565 free_cpumask_var(new_mask
);
5569 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5571 struct task_struct
*p
;
5572 unsigned long flags
;
5578 p
= find_process_by_pid(pid
);
5582 retval
= security_task_getscheduler(p
);
5586 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5587 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
5588 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5597 * sys_sched_getaffinity - get the CPU affinity of a process
5598 * @pid: pid of the process
5599 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5600 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5602 * Return: size of CPU mask copied to user_mask_ptr on success. An
5603 * error code otherwise.
5605 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5606 unsigned long __user
*, user_mask_ptr
)
5611 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5613 if (len
& (sizeof(unsigned long)-1))
5616 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5619 ret
= sched_getaffinity(pid
, mask
);
5621 unsigned int retlen
= min(len
, cpumask_size());
5623 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5628 free_cpumask_var(mask
);
5634 * sys_sched_yield - yield the current processor to other threads.
5636 * This function yields the current CPU to other tasks. If there are no
5637 * other threads running on this CPU then this function will return.
5641 static void do_sched_yield(void)
5646 rq
= this_rq_lock_irq(&rf
);
5648 schedstat_inc(rq
->yld_count
);
5649 current
->sched_class
->yield_task(rq
);
5652 * Since we are going to call schedule() anyway, there's
5653 * no need to preempt or enable interrupts:
5657 sched_preempt_enable_no_resched();
5662 SYSCALL_DEFINE0(sched_yield
)
5668 #ifndef CONFIG_PREEMPTION
5669 int __sched
_cond_resched(void)
5671 if (should_resched(0)) {
5672 preempt_schedule_common();
5678 EXPORT_SYMBOL(_cond_resched
);
5682 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5683 * call schedule, and on return reacquire the lock.
5685 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5686 * operations here to prevent schedule() from being called twice (once via
5687 * spin_unlock(), once by hand).
5689 int __cond_resched_lock(spinlock_t
*lock
)
5691 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5694 lockdep_assert_held(lock
);
5696 if (spin_needbreak(lock
) || resched
) {
5699 preempt_schedule_common();
5707 EXPORT_SYMBOL(__cond_resched_lock
);
5710 * yield - yield the current processor to other threads.
5712 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5714 * The scheduler is at all times free to pick the calling task as the most
5715 * eligible task to run, if removing the yield() call from your code breaks
5716 * it, its already broken.
5718 * Typical broken usage is:
5723 * where one assumes that yield() will let 'the other' process run that will
5724 * make event true. If the current task is a SCHED_FIFO task that will never
5725 * happen. Never use yield() as a progress guarantee!!
5727 * If you want to use yield() to wait for something, use wait_event().
5728 * If you want to use yield() to be 'nice' for others, use cond_resched().
5729 * If you still want to use yield(), do not!
5731 void __sched
yield(void)
5733 set_current_state(TASK_RUNNING
);
5736 EXPORT_SYMBOL(yield
);
5739 * yield_to - yield the current processor to another thread in
5740 * your thread group, or accelerate that thread toward the
5741 * processor it's on.
5743 * @preempt: whether task preemption is allowed or not
5745 * It's the caller's job to ensure that the target task struct
5746 * can't go away on us before we can do any checks.
5749 * true (>0) if we indeed boosted the target task.
5750 * false (0) if we failed to boost the target.
5751 * -ESRCH if there's no task to yield to.
5753 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5755 struct task_struct
*curr
= current
;
5756 struct rq
*rq
, *p_rq
;
5757 unsigned long flags
;
5760 local_irq_save(flags
);
5766 * If we're the only runnable task on the rq and target rq also
5767 * has only one task, there's absolutely no point in yielding.
5769 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5774 double_rq_lock(rq
, p_rq
);
5775 if (task_rq(p
) != p_rq
) {
5776 double_rq_unlock(rq
, p_rq
);
5780 if (!curr
->sched_class
->yield_to_task
)
5783 if (curr
->sched_class
!= p
->sched_class
)
5786 if (task_running(p_rq
, p
) || p
->state
)
5789 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5791 schedstat_inc(rq
->yld_count
);
5793 * Make p's CPU reschedule; pick_next_entity takes care of
5796 if (preempt
&& rq
!= p_rq
)
5801 double_rq_unlock(rq
, p_rq
);
5803 local_irq_restore(flags
);
5810 EXPORT_SYMBOL_GPL(yield_to
);
5812 int io_schedule_prepare(void)
5814 int old_iowait
= current
->in_iowait
;
5816 current
->in_iowait
= 1;
5817 blk_schedule_flush_plug(current
);
5822 void io_schedule_finish(int token
)
5824 current
->in_iowait
= token
;
5828 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5829 * that process accounting knows that this is a task in IO wait state.
5831 long __sched
io_schedule_timeout(long timeout
)
5836 token
= io_schedule_prepare();
5837 ret
= schedule_timeout(timeout
);
5838 io_schedule_finish(token
);
5842 EXPORT_SYMBOL(io_schedule_timeout
);
5844 void __sched
io_schedule(void)
5848 token
= io_schedule_prepare();
5850 io_schedule_finish(token
);
5852 EXPORT_SYMBOL(io_schedule
);
5855 * sys_sched_get_priority_max - return maximum RT priority.
5856 * @policy: scheduling class.
5858 * Return: On success, this syscall returns the maximum
5859 * rt_priority that can be used by a given scheduling class.
5860 * On failure, a negative error code is returned.
5862 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5869 ret
= MAX_USER_RT_PRIO
-1;
5871 case SCHED_DEADLINE
:
5882 * sys_sched_get_priority_min - return minimum RT priority.
5883 * @policy: scheduling class.
5885 * Return: On success, this syscall returns the minimum
5886 * rt_priority that can be used by a given scheduling class.
5887 * On failure, a negative error code is returned.
5889 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5898 case SCHED_DEADLINE
:
5907 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5909 struct task_struct
*p
;
5910 unsigned int time_slice
;
5920 p
= find_process_by_pid(pid
);
5924 retval
= security_task_getscheduler(p
);
5928 rq
= task_rq_lock(p
, &rf
);
5930 if (p
->sched_class
->get_rr_interval
)
5931 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5932 task_rq_unlock(rq
, p
, &rf
);
5935 jiffies_to_timespec64(time_slice
, t
);
5944 * sys_sched_rr_get_interval - return the default timeslice of a process.
5945 * @pid: pid of the process.
5946 * @interval: userspace pointer to the timeslice value.
5948 * this syscall writes the default timeslice value of a given process
5949 * into the user-space timespec buffer. A value of '0' means infinity.
5951 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5954 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5955 struct __kernel_timespec __user
*, interval
)
5957 struct timespec64 t
;
5958 int retval
= sched_rr_get_interval(pid
, &t
);
5961 retval
= put_timespec64(&t
, interval
);
5966 #ifdef CONFIG_COMPAT_32BIT_TIME
5967 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
5968 struct old_timespec32 __user
*, interval
)
5970 struct timespec64 t
;
5971 int retval
= sched_rr_get_interval(pid
, &t
);
5974 retval
= put_old_timespec32(&t
, interval
);
5979 void sched_show_task(struct task_struct
*p
)
5981 unsigned long free
= 0;
5984 if (!try_get_task_stack(p
))
5987 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5989 if (p
->state
== TASK_RUNNING
)
5990 printk(KERN_CONT
" running task ");
5991 #ifdef CONFIG_DEBUG_STACK_USAGE
5992 free
= stack_not_used(p
);
5997 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5999 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6000 task_pid_nr(p
), ppid
,
6001 (unsigned long)task_thread_info(p
)->flags
);
6003 print_worker_info(KERN_INFO
, p
);
6004 show_stack(p
, NULL
);
6007 EXPORT_SYMBOL_GPL(sched_show_task
);
6010 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
6012 /* no filter, everything matches */
6016 /* filter, but doesn't match */
6017 if (!(p
->state
& state_filter
))
6021 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6024 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
6031 void show_state_filter(unsigned long state_filter
)
6033 struct task_struct
*g
, *p
;
6035 #if BITS_PER_LONG == 32
6037 " task PC stack pid father\n");
6040 " task PC stack pid father\n");
6043 for_each_process_thread(g
, p
) {
6045 * reset the NMI-timeout, listing all files on a slow
6046 * console might take a lot of time:
6047 * Also, reset softlockup watchdogs on all CPUs, because
6048 * another CPU might be blocked waiting for us to process
6051 touch_nmi_watchdog();
6052 touch_all_softlockup_watchdogs();
6053 if (state_filter_match(state_filter
, p
))
6057 #ifdef CONFIG_SCHED_DEBUG
6059 sysrq_sched_debug_show();
6063 * Only show locks if all tasks are dumped:
6066 debug_show_all_locks();
6070 * init_idle - set up an idle thread for a given CPU
6071 * @idle: task in question
6072 * @cpu: CPU the idle task belongs to
6074 * NOTE: this function does not set the idle thread's NEED_RESCHED
6075 * flag, to make booting more robust.
6077 void init_idle(struct task_struct
*idle
, int cpu
)
6079 struct rq
*rq
= cpu_rq(cpu
);
6080 unsigned long flags
;
6082 __sched_fork(0, idle
);
6084 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
6085 raw_spin_lock(&rq
->lock
);
6087 idle
->state
= TASK_RUNNING
;
6088 idle
->se
.exec_start
= sched_clock();
6089 idle
->flags
|= PF_IDLE
;
6091 kasan_unpoison_task_stack(idle
);
6095 * Its possible that init_idle() gets called multiple times on a task,
6096 * in that case do_set_cpus_allowed() will not do the right thing.
6098 * And since this is boot we can forgo the serialization.
6100 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
6103 * We're having a chicken and egg problem, even though we are
6104 * holding rq->lock, the CPU isn't yet set to this CPU so the
6105 * lockdep check in task_group() will fail.
6107 * Similar case to sched_fork(). / Alternatively we could
6108 * use task_rq_lock() here and obtain the other rq->lock.
6113 __set_task_cpu(idle
, cpu
);
6117 rcu_assign_pointer(rq
->curr
, idle
);
6118 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
6122 raw_spin_unlock(&rq
->lock
);
6123 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
6125 /* Set the preempt count _outside_ the spinlocks! */
6126 init_idle_preempt_count(idle
, cpu
);
6129 * The idle tasks have their own, simple scheduling class:
6131 idle
->sched_class
= &idle_sched_class
;
6132 ftrace_graph_init_idle_task(idle
, cpu
);
6133 vtime_init_idle(idle
, cpu
);
6135 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
6141 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
6142 const struct cpumask
*trial
)
6146 if (!cpumask_weight(cur
))
6149 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
6154 int task_can_attach(struct task_struct
*p
,
6155 const struct cpumask
*cs_cpus_allowed
)
6160 * Kthreads which disallow setaffinity shouldn't be moved
6161 * to a new cpuset; we don't want to change their CPU
6162 * affinity and isolating such threads by their set of
6163 * allowed nodes is unnecessary. Thus, cpusets are not
6164 * applicable for such threads. This prevents checking for
6165 * success of set_cpus_allowed_ptr() on all attached tasks
6166 * before cpus_mask may be changed.
6168 if (p
->flags
& PF_NO_SETAFFINITY
) {
6173 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
6175 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
6181 bool sched_smp_initialized __read_mostly
;
6183 #ifdef CONFIG_NUMA_BALANCING
6184 /* Migrate current task p to target_cpu */
6185 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
6187 struct migration_arg arg
= { p
, target_cpu
};
6188 int curr_cpu
= task_cpu(p
);
6190 if (curr_cpu
== target_cpu
)
6193 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
6196 /* TODO: This is not properly updating schedstats */
6198 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
6199 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
6203 * Requeue a task on a given node and accurately track the number of NUMA
6204 * tasks on the runqueues
6206 void sched_setnuma(struct task_struct
*p
, int nid
)
6208 bool queued
, running
;
6212 rq
= task_rq_lock(p
, &rf
);
6213 queued
= task_on_rq_queued(p
);
6214 running
= task_current(rq
, p
);
6217 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
6219 put_prev_task(rq
, p
);
6221 p
->numa_preferred_nid
= nid
;
6224 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
6226 set_next_task(rq
, p
);
6227 task_rq_unlock(rq
, p
, &rf
);
6229 #endif /* CONFIG_NUMA_BALANCING */
6231 #ifdef CONFIG_HOTPLUG_CPU
6233 * Ensure that the idle task is using init_mm right before its CPU goes
6236 void idle_task_exit(void)
6238 struct mm_struct
*mm
= current
->active_mm
;
6240 BUG_ON(cpu_online(smp_processor_id()));
6242 if (mm
!= &init_mm
) {
6243 switch_mm(mm
, &init_mm
, current
);
6244 current
->active_mm
= &init_mm
;
6245 finish_arch_post_lock_switch();
6251 * Since this CPU is going 'away' for a while, fold any nr_active delta
6252 * we might have. Assumes we're called after migrate_tasks() so that the
6253 * nr_active count is stable. We need to take the teardown thread which
6254 * is calling this into account, so we hand in adjust = 1 to the load
6257 * Also see the comment "Global load-average calculations".
6259 static void calc_load_migrate(struct rq
*rq
)
6261 long delta
= calc_load_fold_active(rq
, 1);
6263 atomic_long_add(delta
, &calc_load_tasks
);
6266 static struct task_struct
*__pick_migrate_task(struct rq
*rq
)
6268 const struct sched_class
*class;
6269 struct task_struct
*next
;
6271 for_each_class(class) {
6272 next
= class->pick_next_task(rq
);
6274 next
->sched_class
->put_prev_task(rq
, next
);
6279 /* The idle class should always have a runnable task */
6284 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6285 * try_to_wake_up()->select_task_rq().
6287 * Called with rq->lock held even though we'er in stop_machine() and
6288 * there's no concurrency possible, we hold the required locks anyway
6289 * because of lock validation efforts.
6291 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
6293 struct rq
*rq
= dead_rq
;
6294 struct task_struct
*next
, *stop
= rq
->stop
;
6295 struct rq_flags orf
= *rf
;
6299 * Fudge the rq selection such that the below task selection loop
6300 * doesn't get stuck on the currently eligible stop task.
6302 * We're currently inside stop_machine() and the rq is either stuck
6303 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6304 * either way we should never end up calling schedule() until we're
6310 * put_prev_task() and pick_next_task() sched
6311 * class method both need to have an up-to-date
6312 * value of rq->clock[_task]
6314 update_rq_clock(rq
);
6318 * There's this thread running, bail when that's the only
6321 if (rq
->nr_running
== 1)
6324 next
= __pick_migrate_task(rq
);
6327 * Rules for changing task_struct::cpus_mask are holding
6328 * both pi_lock and rq->lock, such that holding either
6329 * stabilizes the mask.
6331 * Drop rq->lock is not quite as disastrous as it usually is
6332 * because !cpu_active at this point, which means load-balance
6333 * will not interfere. Also, stop-machine.
6336 raw_spin_lock(&next
->pi_lock
);
6340 * Since we're inside stop-machine, _nothing_ should have
6341 * changed the task, WARN if weird stuff happened, because in
6342 * that case the above rq->lock drop is a fail too.
6344 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
6345 raw_spin_unlock(&next
->pi_lock
);
6349 /* Find suitable destination for @next, with force if needed. */
6350 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
6351 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
6352 if (rq
!= dead_rq
) {
6358 raw_spin_unlock(&next
->pi_lock
);
6363 #endif /* CONFIG_HOTPLUG_CPU */
6365 void set_rq_online(struct rq
*rq
)
6368 const struct sched_class
*class;
6370 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6373 for_each_class(class) {
6374 if (class->rq_online
)
6375 class->rq_online(rq
);
6380 void set_rq_offline(struct rq
*rq
)
6383 const struct sched_class
*class;
6385 for_each_class(class) {
6386 if (class->rq_offline
)
6387 class->rq_offline(rq
);
6390 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6396 * used to mark begin/end of suspend/resume:
6398 static int num_cpus_frozen
;
6401 * Update cpusets according to cpu_active mask. If cpusets are
6402 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6403 * around partition_sched_domains().
6405 * If we come here as part of a suspend/resume, don't touch cpusets because we
6406 * want to restore it back to its original state upon resume anyway.
6408 static void cpuset_cpu_active(void)
6410 if (cpuhp_tasks_frozen
) {
6412 * num_cpus_frozen tracks how many CPUs are involved in suspend
6413 * resume sequence. As long as this is not the last online
6414 * operation in the resume sequence, just build a single sched
6415 * domain, ignoring cpusets.
6417 partition_sched_domains(1, NULL
, NULL
);
6418 if (--num_cpus_frozen
)
6421 * This is the last CPU online operation. So fall through and
6422 * restore the original sched domains by considering the
6423 * cpuset configurations.
6425 cpuset_force_rebuild();
6427 cpuset_update_active_cpus();
6430 static int cpuset_cpu_inactive(unsigned int cpu
)
6432 if (!cpuhp_tasks_frozen
) {
6433 if (dl_cpu_busy(cpu
))
6435 cpuset_update_active_cpus();
6438 partition_sched_domains(1, NULL
, NULL
);
6443 int sched_cpu_activate(unsigned int cpu
)
6445 struct rq
*rq
= cpu_rq(cpu
);
6448 #ifdef CONFIG_SCHED_SMT
6450 * When going up, increment the number of cores with SMT present.
6452 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6453 static_branch_inc_cpuslocked(&sched_smt_present
);
6455 set_cpu_active(cpu
, true);
6457 if (sched_smp_initialized
) {
6458 sched_domains_numa_masks_set(cpu
);
6459 cpuset_cpu_active();
6463 * Put the rq online, if not already. This happens:
6465 * 1) In the early boot process, because we build the real domains
6466 * after all CPUs have been brought up.
6468 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6471 rq_lock_irqsave(rq
, &rf
);
6473 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6476 rq_unlock_irqrestore(rq
, &rf
);
6481 int sched_cpu_deactivate(unsigned int cpu
)
6485 set_cpu_active(cpu
, false);
6487 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6488 * users of this state to go away such that all new such users will
6491 * Do sync before park smpboot threads to take care the rcu boost case.
6495 #ifdef CONFIG_SCHED_SMT
6497 * When going down, decrement the number of cores with SMT present.
6499 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6500 static_branch_dec_cpuslocked(&sched_smt_present
);
6503 if (!sched_smp_initialized
)
6506 ret
= cpuset_cpu_inactive(cpu
);
6508 set_cpu_active(cpu
, true);
6511 sched_domains_numa_masks_clear(cpu
);
6515 static void sched_rq_cpu_starting(unsigned int cpu
)
6517 struct rq
*rq
= cpu_rq(cpu
);
6519 rq
->calc_load_update
= calc_load_update
;
6520 update_max_interval();
6523 int sched_cpu_starting(unsigned int cpu
)
6525 sched_rq_cpu_starting(cpu
);
6526 sched_tick_start(cpu
);
6530 #ifdef CONFIG_HOTPLUG_CPU
6531 int sched_cpu_dying(unsigned int cpu
)
6533 struct rq
*rq
= cpu_rq(cpu
);
6536 /* Handle pending wakeups and then migrate everything off */
6537 sched_ttwu_pending();
6538 sched_tick_stop(cpu
);
6540 rq_lock_irqsave(rq
, &rf
);
6542 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6545 migrate_tasks(rq
, &rf
);
6546 BUG_ON(rq
->nr_running
!= 1);
6547 rq_unlock_irqrestore(rq
, &rf
);
6549 calc_load_migrate(rq
);
6550 update_max_interval();
6551 nohz_balance_exit_idle(rq
);
6557 void __init
sched_init_smp(void)
6562 * There's no userspace yet to cause hotplug operations; hence all the
6563 * CPU masks are stable and all blatant races in the below code cannot
6566 mutex_lock(&sched_domains_mutex
);
6567 sched_init_domains(cpu_active_mask
);
6568 mutex_unlock(&sched_domains_mutex
);
6570 /* Move init over to a non-isolated CPU */
6571 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
6573 sched_init_granularity();
6575 init_sched_rt_class();
6576 init_sched_dl_class();
6578 sched_smp_initialized
= true;
6581 static int __init
migration_init(void)
6583 sched_cpu_starting(smp_processor_id());
6586 early_initcall(migration_init
);
6589 void __init
sched_init_smp(void)
6591 sched_init_granularity();
6593 #endif /* CONFIG_SMP */
6595 int in_sched_functions(unsigned long addr
)
6597 return in_lock_functions(addr
) ||
6598 (addr
>= (unsigned long)__sched_text_start
6599 && addr
< (unsigned long)__sched_text_end
);
6602 #ifdef CONFIG_CGROUP_SCHED
6604 * Default task group.
6605 * Every task in system belongs to this group at bootup.
6607 struct task_group root_task_group
;
6608 LIST_HEAD(task_groups
);
6610 /* Cacheline aligned slab cache for task_group */
6611 static struct kmem_cache
*task_group_cache __read_mostly
;
6614 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6615 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
6617 void __init
sched_init(void)
6619 unsigned long ptr
= 0;
6624 #ifdef CONFIG_FAIR_GROUP_SCHED
6625 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6627 #ifdef CONFIG_RT_GROUP_SCHED
6628 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6631 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
6633 #ifdef CONFIG_FAIR_GROUP_SCHED
6634 root_task_group
.se
= (struct sched_entity
**)ptr
;
6635 ptr
+= nr_cpu_ids
* sizeof(void **);
6637 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6638 ptr
+= nr_cpu_ids
* sizeof(void **);
6640 #endif /* CONFIG_FAIR_GROUP_SCHED */
6641 #ifdef CONFIG_RT_GROUP_SCHED
6642 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6643 ptr
+= nr_cpu_ids
* sizeof(void **);
6645 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6646 ptr
+= nr_cpu_ids
* sizeof(void **);
6648 #endif /* CONFIG_RT_GROUP_SCHED */
6650 #ifdef CONFIG_CPUMASK_OFFSTACK
6651 for_each_possible_cpu(i
) {
6652 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6653 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6654 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6655 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6657 #endif /* CONFIG_CPUMASK_OFFSTACK */
6659 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
6660 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6663 init_defrootdomain();
6666 #ifdef CONFIG_RT_GROUP_SCHED
6667 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6668 global_rt_period(), global_rt_runtime());
6669 #endif /* CONFIG_RT_GROUP_SCHED */
6671 #ifdef CONFIG_CGROUP_SCHED
6672 task_group_cache
= KMEM_CACHE(task_group
, 0);
6674 list_add(&root_task_group
.list
, &task_groups
);
6675 INIT_LIST_HEAD(&root_task_group
.children
);
6676 INIT_LIST_HEAD(&root_task_group
.siblings
);
6677 autogroup_init(&init_task
);
6678 #endif /* CONFIG_CGROUP_SCHED */
6680 for_each_possible_cpu(i
) {
6684 raw_spin_lock_init(&rq
->lock
);
6686 rq
->calc_load_active
= 0;
6687 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6688 init_cfs_rq(&rq
->cfs
);
6689 init_rt_rq(&rq
->rt
);
6690 init_dl_rq(&rq
->dl
);
6691 #ifdef CONFIG_FAIR_GROUP_SCHED
6692 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6693 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6694 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6696 * How much CPU bandwidth does root_task_group get?
6698 * In case of task-groups formed thr' the cgroup filesystem, it
6699 * gets 100% of the CPU resources in the system. This overall
6700 * system CPU resource is divided among the tasks of
6701 * root_task_group and its child task-groups in a fair manner,
6702 * based on each entity's (task or task-group's) weight
6703 * (se->load.weight).
6705 * In other words, if root_task_group has 10 tasks of weight
6706 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6707 * then A0's share of the CPU resource is:
6709 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6711 * We achieve this by letting root_task_group's tasks sit
6712 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6714 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6715 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6716 #endif /* CONFIG_FAIR_GROUP_SCHED */
6718 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6719 #ifdef CONFIG_RT_GROUP_SCHED
6720 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6725 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6726 rq
->balance_callback
= NULL
;
6727 rq
->active_balance
= 0;
6728 rq
->next_balance
= jiffies
;
6733 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6734 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6736 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6738 rq_attach_root(rq
, &def_root_domain
);
6739 #ifdef CONFIG_NO_HZ_COMMON
6740 rq
->last_blocked_load_update_tick
= jiffies
;
6741 atomic_set(&rq
->nohz_flags
, 0);
6743 #endif /* CONFIG_SMP */
6745 atomic_set(&rq
->nr_iowait
, 0);
6748 set_load_weight(&init_task
, false);
6751 * The boot idle thread does lazy MMU switching as well:
6754 enter_lazy_tlb(&init_mm
, current
);
6757 * Make us the idle thread. Technically, schedule() should not be
6758 * called from this thread, however somewhere below it might be,
6759 * but because we are the idle thread, we just pick up running again
6760 * when this runqueue becomes "idle".
6762 init_idle(current
, smp_processor_id());
6764 calc_load_update
= jiffies
+ LOAD_FREQ
;
6767 idle_thread_set_boot_cpu();
6769 init_sched_fair_class();
6777 scheduler_running
= 1;
6780 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6781 static inline int preempt_count_equals(int preempt_offset
)
6783 int nested
= preempt_count() + rcu_preempt_depth();
6785 return (nested
== preempt_offset
);
6788 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6791 * Blocking primitives will set (and therefore destroy) current->state,
6792 * since we will exit with TASK_RUNNING make sure we enter with it,
6793 * otherwise we will destroy state.
6795 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6796 "do not call blocking ops when !TASK_RUNNING; "
6797 "state=%lx set at [<%p>] %pS\n",
6799 (void *)current
->task_state_change
,
6800 (void *)current
->task_state_change
);
6802 ___might_sleep(file
, line
, preempt_offset
);
6804 EXPORT_SYMBOL(__might_sleep
);
6806 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6808 /* Ratelimiting timestamp: */
6809 static unsigned long prev_jiffy
;
6811 unsigned long preempt_disable_ip
;
6813 /* WARN_ON_ONCE() by default, no rate limit required: */
6816 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6817 !is_idle_task(current
) && !current
->non_block_count
) ||
6818 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6822 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6824 prev_jiffy
= jiffies
;
6826 /* Save this before calling printk(), since that will clobber it: */
6827 preempt_disable_ip
= get_preempt_disable_ip(current
);
6830 "BUG: sleeping function called from invalid context at %s:%d\n",
6833 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6834 in_atomic(), irqs_disabled(), current
->non_block_count
,
6835 current
->pid
, current
->comm
);
6837 if (task_stack_end_corrupted(current
))
6838 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6840 debug_show_held_locks(current
);
6841 if (irqs_disabled())
6842 print_irqtrace_events(current
);
6843 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6844 && !preempt_count_equals(preempt_offset
)) {
6845 pr_err("Preemption disabled at:");
6846 print_ip_sym(preempt_disable_ip
);
6850 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6852 EXPORT_SYMBOL(___might_sleep
);
6854 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
6856 static unsigned long prev_jiffy
;
6858 if (irqs_disabled())
6861 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
6864 if (preempt_count() > preempt_offset
)
6867 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6869 prev_jiffy
= jiffies
;
6871 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
6872 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6873 in_atomic(), irqs_disabled(),
6874 current
->pid
, current
->comm
);
6876 debug_show_held_locks(current
);
6878 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6880 EXPORT_SYMBOL_GPL(__cant_sleep
);
6883 #ifdef CONFIG_MAGIC_SYSRQ
6884 void normalize_rt_tasks(void)
6886 struct task_struct
*g
, *p
;
6887 struct sched_attr attr
= {
6888 .sched_policy
= SCHED_NORMAL
,
6891 read_lock(&tasklist_lock
);
6892 for_each_process_thread(g
, p
) {
6894 * Only normalize user tasks:
6896 if (p
->flags
& PF_KTHREAD
)
6899 p
->se
.exec_start
= 0;
6900 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6901 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6902 schedstat_set(p
->se
.statistics
.block_start
, 0);
6904 if (!dl_task(p
) && !rt_task(p
)) {
6906 * Renice negative nice level userspace
6909 if (task_nice(p
) < 0)
6910 set_user_nice(p
, 0);
6914 __sched_setscheduler(p
, &attr
, false, false);
6916 read_unlock(&tasklist_lock
);
6919 #endif /* CONFIG_MAGIC_SYSRQ */
6921 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6923 * These functions are only useful for the IA64 MCA handling, or kdb.
6925 * They can only be called when the whole system has been
6926 * stopped - every CPU needs to be quiescent, and no scheduling
6927 * activity can take place. Using them for anything else would
6928 * be a serious bug, and as a result, they aren't even visible
6929 * under any other configuration.
6933 * curr_task - return the current task for a given CPU.
6934 * @cpu: the processor in question.
6936 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6938 * Return: The current task for @cpu.
6940 struct task_struct
*curr_task(int cpu
)
6942 return cpu_curr(cpu
);
6945 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6949 * ia64_set_curr_task - set the current task for a given CPU.
6950 * @cpu: the processor in question.
6951 * @p: the task pointer to set.
6953 * Description: This function must only be used when non-maskable interrupts
6954 * are serviced on a separate stack. It allows the architecture to switch the
6955 * notion of the current task on a CPU in a non-blocking manner. This function
6956 * must be called with all CPU's synchronized, and interrupts disabled, the
6957 * and caller must save the original value of the current task (see
6958 * curr_task() above) and restore that value before reenabling interrupts and
6959 * re-starting the system.
6961 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6963 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6970 #ifdef CONFIG_CGROUP_SCHED
6971 /* task_group_lock serializes the addition/removal of task groups */
6972 static DEFINE_SPINLOCK(task_group_lock
);
6974 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
6975 struct task_group
*parent
)
6977 #ifdef CONFIG_UCLAMP_TASK_GROUP
6978 enum uclamp_id clamp_id
;
6980 for_each_clamp_id(clamp_id
) {
6981 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
6982 uclamp_none(clamp_id
), false);
6983 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
6988 static void sched_free_group(struct task_group
*tg
)
6990 free_fair_sched_group(tg
);
6991 free_rt_sched_group(tg
);
6993 kmem_cache_free(task_group_cache
, tg
);
6996 /* allocate runqueue etc for a new task group */
6997 struct task_group
*sched_create_group(struct task_group
*parent
)
6999 struct task_group
*tg
;
7001 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7003 return ERR_PTR(-ENOMEM
);
7005 if (!alloc_fair_sched_group(tg
, parent
))
7008 if (!alloc_rt_sched_group(tg
, parent
))
7011 alloc_uclamp_sched_group(tg
, parent
);
7016 sched_free_group(tg
);
7017 return ERR_PTR(-ENOMEM
);
7020 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7022 unsigned long flags
;
7024 spin_lock_irqsave(&task_group_lock
, flags
);
7025 list_add_rcu(&tg
->list
, &task_groups
);
7027 /* Root should already exist: */
7030 tg
->parent
= parent
;
7031 INIT_LIST_HEAD(&tg
->children
);
7032 list_add_rcu(&tg
->siblings
, &parent
->children
);
7033 spin_unlock_irqrestore(&task_group_lock
, flags
);
7035 online_fair_sched_group(tg
);
7038 /* rcu callback to free various structures associated with a task group */
7039 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7041 /* Now it should be safe to free those cfs_rqs: */
7042 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7045 void sched_destroy_group(struct task_group
*tg
)
7047 /* Wait for possible concurrent references to cfs_rqs complete: */
7048 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7051 void sched_offline_group(struct task_group
*tg
)
7053 unsigned long flags
;
7055 /* End participation in shares distribution: */
7056 unregister_fair_sched_group(tg
);
7058 spin_lock_irqsave(&task_group_lock
, flags
);
7059 list_del_rcu(&tg
->list
);
7060 list_del_rcu(&tg
->siblings
);
7061 spin_unlock_irqrestore(&task_group_lock
, flags
);
7064 static void sched_change_group(struct task_struct
*tsk
, int type
)
7066 struct task_group
*tg
;
7069 * All callers are synchronized by task_rq_lock(); we do not use RCU
7070 * which is pointless here. Thus, we pass "true" to task_css_check()
7071 * to prevent lockdep warnings.
7073 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7074 struct task_group
, css
);
7075 tg
= autogroup_task_group(tsk
, tg
);
7076 tsk
->sched_task_group
= tg
;
7078 #ifdef CONFIG_FAIR_GROUP_SCHED
7079 if (tsk
->sched_class
->task_change_group
)
7080 tsk
->sched_class
->task_change_group(tsk
, type
);
7083 set_task_rq(tsk
, task_cpu(tsk
));
7087 * Change task's runqueue when it moves between groups.
7089 * The caller of this function should have put the task in its new group by
7090 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7093 void sched_move_task(struct task_struct
*tsk
)
7095 int queued
, running
, queue_flags
=
7096 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7100 rq
= task_rq_lock(tsk
, &rf
);
7101 update_rq_clock(rq
);
7103 running
= task_current(rq
, tsk
);
7104 queued
= task_on_rq_queued(tsk
);
7107 dequeue_task(rq
, tsk
, queue_flags
);
7109 put_prev_task(rq
, tsk
);
7111 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7114 enqueue_task(rq
, tsk
, queue_flags
);
7116 set_next_task(rq
, tsk
);
7118 * After changing group, the running task may have joined a
7119 * throttled one but it's still the running task. Trigger a
7120 * resched to make sure that task can still run.
7125 task_rq_unlock(rq
, tsk
, &rf
);
7128 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7130 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7133 static struct cgroup_subsys_state
*
7134 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7136 struct task_group
*parent
= css_tg(parent_css
);
7137 struct task_group
*tg
;
7140 /* This is early initialization for the top cgroup */
7141 return &root_task_group
.css
;
7144 tg
= sched_create_group(parent
);
7146 return ERR_PTR(-ENOMEM
);
7151 /* Expose task group only after completing cgroup initialization */
7152 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7154 struct task_group
*tg
= css_tg(css
);
7155 struct task_group
*parent
= css_tg(css
->parent
);
7158 sched_online_group(tg
, parent
);
7160 #ifdef CONFIG_UCLAMP_TASK_GROUP
7161 /* Propagate the effective uclamp value for the new group */
7162 cpu_util_update_eff(css
);
7168 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
7170 struct task_group
*tg
= css_tg(css
);
7172 sched_offline_group(tg
);
7175 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7177 struct task_group
*tg
= css_tg(css
);
7180 * Relies on the RCU grace period between css_released() and this.
7182 sched_free_group(tg
);
7186 * This is called before wake_up_new_task(), therefore we really only
7187 * have to set its group bits, all the other stuff does not apply.
7189 static void cpu_cgroup_fork(struct task_struct
*task
)
7194 rq
= task_rq_lock(task
, &rf
);
7196 update_rq_clock(rq
);
7197 sched_change_group(task
, TASK_SET_GROUP
);
7199 task_rq_unlock(rq
, task
, &rf
);
7202 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
7204 struct task_struct
*task
;
7205 struct cgroup_subsys_state
*css
;
7208 cgroup_taskset_for_each(task
, css
, tset
) {
7209 #ifdef CONFIG_RT_GROUP_SCHED
7210 if (!sched_rt_can_attach(css_tg(css
), task
))
7214 * Serialize against wake_up_new_task() such that if its
7215 * running, we're sure to observe its full state.
7217 raw_spin_lock_irq(&task
->pi_lock
);
7219 * Avoid calling sched_move_task() before wake_up_new_task()
7220 * has happened. This would lead to problems with PELT, due to
7221 * move wanting to detach+attach while we're not attached yet.
7223 if (task
->state
== TASK_NEW
)
7225 raw_spin_unlock_irq(&task
->pi_lock
);
7233 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
7235 struct task_struct
*task
;
7236 struct cgroup_subsys_state
*css
;
7238 cgroup_taskset_for_each(task
, css
, tset
)
7239 sched_move_task(task
);
7242 #ifdef CONFIG_UCLAMP_TASK_GROUP
7243 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
7245 struct cgroup_subsys_state
*top_css
= css
;
7246 struct uclamp_se
*uc_parent
= NULL
;
7247 struct uclamp_se
*uc_se
= NULL
;
7248 unsigned int eff
[UCLAMP_CNT
];
7249 enum uclamp_id clamp_id
;
7250 unsigned int clamps
;
7252 css_for_each_descendant_pre(css
, top_css
) {
7253 uc_parent
= css_tg(css
)->parent
7254 ? css_tg(css
)->parent
->uclamp
: NULL
;
7256 for_each_clamp_id(clamp_id
) {
7257 /* Assume effective clamps matches requested clamps */
7258 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
7259 /* Cap effective clamps with parent's effective clamps */
7261 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
7262 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
7265 /* Ensure protection is always capped by limit */
7266 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
7268 /* Propagate most restrictive effective clamps */
7270 uc_se
= css_tg(css
)->uclamp
;
7271 for_each_clamp_id(clamp_id
) {
7272 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
7274 uc_se
[clamp_id
].value
= eff
[clamp_id
];
7275 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
7276 clamps
|= (0x1 << clamp_id
);
7279 css
= css_rightmost_descendant(css
);
7283 /* Immediately update descendants RUNNABLE tasks */
7284 uclamp_update_active_tasks(css
, clamps
);
7289 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7290 * C expression. Since there is no way to convert a macro argument (N) into a
7291 * character constant, use two levels of macros.
7293 #define _POW10(exp) ((unsigned int)1e##exp)
7294 #define POW10(exp) _POW10(exp)
7296 struct uclamp_request
{
7297 #define UCLAMP_PERCENT_SHIFT 2
7298 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7304 static inline struct uclamp_request
7305 capacity_from_percent(char *buf
)
7307 struct uclamp_request req
= {
7308 .percent
= UCLAMP_PERCENT_SCALE
,
7309 .util
= SCHED_CAPACITY_SCALE
,
7314 if (strcmp(buf
, "max")) {
7315 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
7319 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
7324 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
7325 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
7331 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
7332 size_t nbytes
, loff_t off
,
7333 enum uclamp_id clamp_id
)
7335 struct uclamp_request req
;
7336 struct task_group
*tg
;
7338 req
= capacity_from_percent(buf
);
7342 mutex_lock(&uclamp_mutex
);
7345 tg
= css_tg(of_css(of
));
7346 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
7347 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
7350 * Because of not recoverable conversion rounding we keep track of the
7351 * exact requested value
7353 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
7355 /* Update effective clamps to track the most restrictive value */
7356 cpu_util_update_eff(of_css(of
));
7359 mutex_unlock(&uclamp_mutex
);
7364 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
7365 char *buf
, size_t nbytes
,
7368 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
7371 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
7372 char *buf
, size_t nbytes
,
7375 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
7378 static inline void cpu_uclamp_print(struct seq_file
*sf
,
7379 enum uclamp_id clamp_id
)
7381 struct task_group
*tg
;
7387 tg
= css_tg(seq_css(sf
));
7388 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
7391 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
7392 seq_puts(sf
, "max\n");
7396 percent
= tg
->uclamp_pct
[clamp_id
];
7397 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
7398 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
7401 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
7403 cpu_uclamp_print(sf
, UCLAMP_MIN
);
7407 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
7409 cpu_uclamp_print(sf
, UCLAMP_MAX
);
7412 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7414 #ifdef CONFIG_FAIR_GROUP_SCHED
7415 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7416 struct cftype
*cftype
, u64 shareval
)
7418 if (shareval
> scale_load_down(ULONG_MAX
))
7419 shareval
= MAX_SHARES
;
7420 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7423 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7426 struct task_group
*tg
= css_tg(css
);
7428 return (u64
) scale_load_down(tg
->shares
);
7431 #ifdef CONFIG_CFS_BANDWIDTH
7432 static DEFINE_MUTEX(cfs_constraints_mutex
);
7434 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7435 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7437 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7439 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7441 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7442 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7444 if (tg
== &root_task_group
)
7448 * Ensure we have at some amount of bandwidth every period. This is
7449 * to prevent reaching a state of large arrears when throttled via
7450 * entity_tick() resulting in prolonged exit starvation.
7452 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7456 * Likewise, bound things on the otherside by preventing insane quota
7457 * periods. This also allows us to normalize in computing quota
7460 if (period
> max_cfs_quota_period
)
7464 * Prevent race between setting of cfs_rq->runtime_enabled and
7465 * unthrottle_offline_cfs_rqs().
7468 mutex_lock(&cfs_constraints_mutex
);
7469 ret
= __cfs_schedulable(tg
, period
, quota
);
7473 runtime_enabled
= quota
!= RUNTIME_INF
;
7474 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7476 * If we need to toggle cfs_bandwidth_used, off->on must occur
7477 * before making related changes, and on->off must occur afterwards
7479 if (runtime_enabled
&& !runtime_was_enabled
)
7480 cfs_bandwidth_usage_inc();
7481 raw_spin_lock_irq(&cfs_b
->lock
);
7482 cfs_b
->period
= ns_to_ktime(period
);
7483 cfs_b
->quota
= quota
;
7485 __refill_cfs_bandwidth_runtime(cfs_b
);
7487 /* Restart the period timer (if active) to handle new period expiry: */
7488 if (runtime_enabled
)
7489 start_cfs_bandwidth(cfs_b
);
7491 raw_spin_unlock_irq(&cfs_b
->lock
);
7493 for_each_online_cpu(i
) {
7494 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7495 struct rq
*rq
= cfs_rq
->rq
;
7498 rq_lock_irq(rq
, &rf
);
7499 cfs_rq
->runtime_enabled
= runtime_enabled
;
7500 cfs_rq
->runtime_remaining
= 0;
7502 if (cfs_rq
->throttled
)
7503 unthrottle_cfs_rq(cfs_rq
);
7504 rq_unlock_irq(rq
, &rf
);
7506 if (runtime_was_enabled
&& !runtime_enabled
)
7507 cfs_bandwidth_usage_dec();
7509 mutex_unlock(&cfs_constraints_mutex
);
7515 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7519 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7520 if (cfs_quota_us
< 0)
7521 quota
= RUNTIME_INF
;
7522 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
7523 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7527 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7530 static long tg_get_cfs_quota(struct task_group
*tg
)
7534 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7537 quota_us
= tg
->cfs_bandwidth
.quota
;
7538 do_div(quota_us
, NSEC_PER_USEC
);
7543 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7547 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
7550 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7551 quota
= tg
->cfs_bandwidth
.quota
;
7553 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7556 static long tg_get_cfs_period(struct task_group
*tg
)
7560 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7561 do_div(cfs_period_us
, NSEC_PER_USEC
);
7563 return cfs_period_us
;
7566 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7569 return tg_get_cfs_quota(css_tg(css
));
7572 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7573 struct cftype
*cftype
, s64 cfs_quota_us
)
7575 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7578 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7581 return tg_get_cfs_period(css_tg(css
));
7584 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7585 struct cftype
*cftype
, u64 cfs_period_us
)
7587 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7590 struct cfs_schedulable_data
{
7591 struct task_group
*tg
;
7596 * normalize group quota/period to be quota/max_period
7597 * note: units are usecs
7599 static u64
normalize_cfs_quota(struct task_group
*tg
,
7600 struct cfs_schedulable_data
*d
)
7608 period
= tg_get_cfs_period(tg
);
7609 quota
= tg_get_cfs_quota(tg
);
7612 /* note: these should typically be equivalent */
7613 if (quota
== RUNTIME_INF
|| quota
== -1)
7616 return to_ratio(period
, quota
);
7619 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7621 struct cfs_schedulable_data
*d
= data
;
7622 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7623 s64 quota
= 0, parent_quota
= -1;
7626 quota
= RUNTIME_INF
;
7628 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7630 quota
= normalize_cfs_quota(tg
, d
);
7631 parent_quota
= parent_b
->hierarchical_quota
;
7634 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7635 * always take the min. On cgroup1, only inherit when no
7638 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
7639 quota
= min(quota
, parent_quota
);
7641 if (quota
== RUNTIME_INF
)
7642 quota
= parent_quota
;
7643 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7647 cfs_b
->hierarchical_quota
= quota
;
7652 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7655 struct cfs_schedulable_data data
= {
7661 if (quota
!= RUNTIME_INF
) {
7662 do_div(data
.period
, NSEC_PER_USEC
);
7663 do_div(data
.quota
, NSEC_PER_USEC
);
7667 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7673 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
7675 struct task_group
*tg
= css_tg(seq_css(sf
));
7676 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7678 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7679 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7680 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7682 if (schedstat_enabled() && tg
!= &root_task_group
) {
7686 for_each_possible_cpu(i
)
7687 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
7689 seq_printf(sf
, "wait_sum %llu\n", ws
);
7694 #endif /* CONFIG_CFS_BANDWIDTH */
7695 #endif /* CONFIG_FAIR_GROUP_SCHED */
7697 #ifdef CONFIG_RT_GROUP_SCHED
7698 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7699 struct cftype
*cft
, s64 val
)
7701 return sched_group_set_rt_runtime(css_tg(css
), val
);
7704 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7707 return sched_group_rt_runtime(css_tg(css
));
7710 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7711 struct cftype
*cftype
, u64 rt_period_us
)
7713 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7716 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7719 return sched_group_rt_period(css_tg(css
));
7721 #endif /* CONFIG_RT_GROUP_SCHED */
7723 static struct cftype cpu_legacy_files
[] = {
7724 #ifdef CONFIG_FAIR_GROUP_SCHED
7727 .read_u64
= cpu_shares_read_u64
,
7728 .write_u64
= cpu_shares_write_u64
,
7731 #ifdef CONFIG_CFS_BANDWIDTH
7733 .name
= "cfs_quota_us",
7734 .read_s64
= cpu_cfs_quota_read_s64
,
7735 .write_s64
= cpu_cfs_quota_write_s64
,
7738 .name
= "cfs_period_us",
7739 .read_u64
= cpu_cfs_period_read_u64
,
7740 .write_u64
= cpu_cfs_period_write_u64
,
7744 .seq_show
= cpu_cfs_stat_show
,
7747 #ifdef CONFIG_RT_GROUP_SCHED
7749 .name
= "rt_runtime_us",
7750 .read_s64
= cpu_rt_runtime_read
,
7751 .write_s64
= cpu_rt_runtime_write
,
7754 .name
= "rt_period_us",
7755 .read_u64
= cpu_rt_period_read_uint
,
7756 .write_u64
= cpu_rt_period_write_uint
,
7759 #ifdef CONFIG_UCLAMP_TASK_GROUP
7761 .name
= "uclamp.min",
7762 .flags
= CFTYPE_NOT_ON_ROOT
,
7763 .seq_show
= cpu_uclamp_min_show
,
7764 .write
= cpu_uclamp_min_write
,
7767 .name
= "uclamp.max",
7768 .flags
= CFTYPE_NOT_ON_ROOT
,
7769 .seq_show
= cpu_uclamp_max_show
,
7770 .write
= cpu_uclamp_max_write
,
7776 static int cpu_extra_stat_show(struct seq_file
*sf
,
7777 struct cgroup_subsys_state
*css
)
7779 #ifdef CONFIG_CFS_BANDWIDTH
7781 struct task_group
*tg
= css_tg(css
);
7782 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7785 throttled_usec
= cfs_b
->throttled_time
;
7786 do_div(throttled_usec
, NSEC_PER_USEC
);
7788 seq_printf(sf
, "nr_periods %d\n"
7790 "throttled_usec %llu\n",
7791 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
7798 #ifdef CONFIG_FAIR_GROUP_SCHED
7799 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
7802 struct task_group
*tg
= css_tg(css
);
7803 u64 weight
= scale_load_down(tg
->shares
);
7805 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
7808 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
7809 struct cftype
*cft
, u64 weight
)
7812 * cgroup weight knobs should use the common MIN, DFL and MAX
7813 * values which are 1, 100 and 10000 respectively. While it loses
7814 * a bit of range on both ends, it maps pretty well onto the shares
7815 * value used by scheduler and the round-trip conversions preserve
7816 * the original value over the entire range.
7818 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
7821 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
7823 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7826 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
7829 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
7830 int last_delta
= INT_MAX
;
7833 /* find the closest nice value to the current weight */
7834 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
7835 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
7836 if (delta
>= last_delta
)
7841 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
7844 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
7845 struct cftype
*cft
, s64 nice
)
7847 unsigned long weight
;
7850 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
7853 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
7854 idx
= array_index_nospec(idx
, 40);
7855 weight
= sched_prio_to_weight
[idx
];
7857 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7861 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
7862 long period
, long quota
)
7865 seq_puts(sf
, "max");
7867 seq_printf(sf
, "%ld", quota
);
7869 seq_printf(sf
, " %ld\n", period
);
7872 /* caller should put the current value in *@periodp before calling */
7873 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
7874 u64
*periodp
, u64
*quotap
)
7876 char tok
[21]; /* U64_MAX */
7878 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
7881 *periodp
*= NSEC_PER_USEC
;
7883 if (sscanf(tok
, "%llu", quotap
))
7884 *quotap
*= NSEC_PER_USEC
;
7885 else if (!strcmp(tok
, "max"))
7886 *quotap
= RUNTIME_INF
;
7893 #ifdef CONFIG_CFS_BANDWIDTH
7894 static int cpu_max_show(struct seq_file
*sf
, void *v
)
7896 struct task_group
*tg
= css_tg(seq_css(sf
));
7898 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
7902 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
7903 char *buf
, size_t nbytes
, loff_t off
)
7905 struct task_group
*tg
= css_tg(of_css(of
));
7906 u64 period
= tg_get_cfs_period(tg
);
7910 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
7912 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
7913 return ret
?: nbytes
;
7917 static struct cftype cpu_files
[] = {
7918 #ifdef CONFIG_FAIR_GROUP_SCHED
7921 .flags
= CFTYPE_NOT_ON_ROOT
,
7922 .read_u64
= cpu_weight_read_u64
,
7923 .write_u64
= cpu_weight_write_u64
,
7926 .name
= "weight.nice",
7927 .flags
= CFTYPE_NOT_ON_ROOT
,
7928 .read_s64
= cpu_weight_nice_read_s64
,
7929 .write_s64
= cpu_weight_nice_write_s64
,
7932 #ifdef CONFIG_CFS_BANDWIDTH
7935 .flags
= CFTYPE_NOT_ON_ROOT
,
7936 .seq_show
= cpu_max_show
,
7937 .write
= cpu_max_write
,
7940 #ifdef CONFIG_UCLAMP_TASK_GROUP
7942 .name
= "uclamp.min",
7943 .flags
= CFTYPE_NOT_ON_ROOT
,
7944 .seq_show
= cpu_uclamp_min_show
,
7945 .write
= cpu_uclamp_min_write
,
7948 .name
= "uclamp.max",
7949 .flags
= CFTYPE_NOT_ON_ROOT
,
7950 .seq_show
= cpu_uclamp_max_show
,
7951 .write
= cpu_uclamp_max_write
,
7957 struct cgroup_subsys cpu_cgrp_subsys
= {
7958 .css_alloc
= cpu_cgroup_css_alloc
,
7959 .css_online
= cpu_cgroup_css_online
,
7960 .css_released
= cpu_cgroup_css_released
,
7961 .css_free
= cpu_cgroup_css_free
,
7962 .css_extra_stat_show
= cpu_extra_stat_show
,
7963 .fork
= cpu_cgroup_fork
,
7964 .can_attach
= cpu_cgroup_can_attach
,
7965 .attach
= cpu_cgroup_attach
,
7966 .legacy_cftypes
= cpu_legacy_files
,
7967 .dfl_cftypes
= cpu_files
,
7972 #endif /* CONFIG_CGROUP_SCHED */
7974 void dump_cpu_task(int cpu
)
7976 pr_info("Task dump for CPU %d:\n", cpu
);
7977 sched_show_task(cpu_curr(cpu
));
7981 * Nice levels are multiplicative, with a gentle 10% change for every
7982 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7983 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7984 * that remained on nice 0.
7986 * The "10% effect" is relative and cumulative: from _any_ nice level,
7987 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7988 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7989 * If a task goes up by ~10% and another task goes down by ~10% then
7990 * the relative distance between them is ~25%.)
7992 const int sched_prio_to_weight
[40] = {
7993 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7994 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7995 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7996 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7997 /* 0 */ 1024, 820, 655, 526, 423,
7998 /* 5 */ 335, 272, 215, 172, 137,
7999 /* 10 */ 110, 87, 70, 56, 45,
8000 /* 15 */ 36, 29, 23, 18, 15,
8004 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8006 * In cases where the weight does not change often, we can use the
8007 * precalculated inverse to speed up arithmetics by turning divisions
8008 * into multiplications:
8010 const u32 sched_prio_to_wmult
[40] = {
8011 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8012 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8013 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8014 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8015 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8016 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8017 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8018 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8021 #undef CREATE_TRACE_POINTS