1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
11 #include <linux/nospec.h>
13 #include <linux/kcov.h>
15 #include <asm/switch_to.h>
18 #include "../workqueue_internal.h"
19 #include "../../fs/io-wq.h"
20 #include "../smpboot.h"
24 #define CREATE_TRACE_POINTS
25 #include <trace/events/sched.h>
28 * Export tracepoints that act as a bare tracehook (ie: have no trace event
29 * associated with them) to allow external modules to probe them.
31 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
32 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
38 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
40 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
42 * Debugging: various feature bits
44 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
45 * sysctl_sched_features, defined in sched.h, to allow constants propagation
46 * at compile time and compiler optimization based on features default.
48 #define SCHED_FEAT(name, enabled) \
49 (1UL << __SCHED_FEAT_##name) * enabled |
50 const_debug
unsigned int sysctl_sched_features
=
57 * Number of tasks to iterate in a single balance run.
58 * Limited because this is done with IRQs disabled.
60 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
63 * period over which we measure -rt task CPU usage in us.
66 unsigned int sysctl_sched_rt_period
= 1000000;
68 __read_mostly
int scheduler_running
;
71 * part of the period that we allow rt tasks to run in us.
74 int sysctl_sched_rt_runtime
= 950000;
77 * __task_rq_lock - lock the rq @p resides on.
79 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
84 lockdep_assert_held(&p
->pi_lock
);
88 raw_spin_lock(&rq
->lock
);
89 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
93 raw_spin_unlock(&rq
->lock
);
95 while (unlikely(task_on_rq_migrating(p
)))
101 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
103 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
104 __acquires(p
->pi_lock
)
110 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
112 raw_spin_lock(&rq
->lock
);
114 * move_queued_task() task_rq_lock()
117 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
118 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
119 * [S] ->cpu = new_cpu [L] task_rq()
123 * If we observe the old CPU in task_rq_lock(), the acquire of
124 * the old rq->lock will fully serialize against the stores.
126 * If we observe the new CPU in task_rq_lock(), the address
127 * dependency headed by '[L] rq = task_rq()' and the acquire
128 * will pair with the WMB to ensure we then also see migrating.
130 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
134 raw_spin_unlock(&rq
->lock
);
135 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
137 while (unlikely(task_on_rq_migrating(p
)))
143 * RQ-clock updating methods:
146 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
149 * In theory, the compile should just see 0 here, and optimize out the call
150 * to sched_rt_avg_update. But I don't trust it...
152 s64 __maybe_unused steal
= 0, irq_delta
= 0;
154 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
155 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
158 * Since irq_time is only updated on {soft,}irq_exit, we might run into
159 * this case when a previous update_rq_clock() happened inside a
162 * When this happens, we stop ->clock_task and only update the
163 * prev_irq_time stamp to account for the part that fit, so that a next
164 * update will consume the rest. This ensures ->clock_task is
167 * It does however cause some slight miss-attribution of {soft,}irq
168 * time, a more accurate solution would be to update the irq_time using
169 * the current rq->clock timestamp, except that would require using
172 if (irq_delta
> delta
)
175 rq
->prev_irq_time
+= irq_delta
;
178 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
179 if (static_key_false((¶virt_steal_rq_enabled
))) {
180 steal
= paravirt_steal_clock(cpu_of(rq
));
181 steal
-= rq
->prev_steal_time_rq
;
183 if (unlikely(steal
> delta
))
186 rq
->prev_steal_time_rq
+= steal
;
191 rq
->clock_task
+= delta
;
193 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
194 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
195 update_irq_load_avg(rq
, irq_delta
+ steal
);
197 update_rq_clock_pelt(rq
, delta
);
200 void update_rq_clock(struct rq
*rq
)
204 lockdep_assert_held(&rq
->lock
);
206 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
209 #ifdef CONFIG_SCHED_DEBUG
210 if (sched_feat(WARN_DOUBLE_CLOCK
))
211 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
212 rq
->clock_update_flags
|= RQCF_UPDATED
;
215 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
219 update_rq_clock_task(rq
, delta
);
223 rq_csd_init(struct rq
*rq
, call_single_data_t
*csd
, smp_call_func_t func
)
230 #ifdef CONFIG_SCHED_HRTICK
232 * Use HR-timers to deliver accurate preemption points.
235 static void hrtick_clear(struct rq
*rq
)
237 if (hrtimer_active(&rq
->hrtick_timer
))
238 hrtimer_cancel(&rq
->hrtick_timer
);
242 * High-resolution timer tick.
243 * Runs from hardirq context with interrupts disabled.
245 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
247 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
250 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
254 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
257 return HRTIMER_NORESTART
;
262 static void __hrtick_restart(struct rq
*rq
)
264 struct hrtimer
*timer
= &rq
->hrtick_timer
;
266 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED_HARD
);
270 * called from hardirq (IPI) context
272 static void __hrtick_start(void *arg
)
278 __hrtick_restart(rq
);
283 * Called to set the hrtick timer state.
285 * called with rq->lock held and irqs disabled
287 void hrtick_start(struct rq
*rq
, u64 delay
)
289 struct hrtimer
*timer
= &rq
->hrtick_timer
;
294 * Don't schedule slices shorter than 10000ns, that just
295 * doesn't make sense and can cause timer DoS.
297 delta
= max_t(s64
, delay
, 10000LL);
298 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
300 hrtimer_set_expires(timer
, time
);
303 __hrtick_restart(rq
);
305 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
310 * Called to set the hrtick timer state.
312 * called with rq->lock held and irqs disabled
314 void hrtick_start(struct rq
*rq
, u64 delay
)
317 * Don't schedule slices shorter than 10000ns, that just
318 * doesn't make sense. Rely on vruntime for fairness.
320 delay
= max_t(u64
, delay
, 10000LL);
321 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
322 HRTIMER_MODE_REL_PINNED_HARD
);
325 #endif /* CONFIG_SMP */
327 static void hrtick_rq_init(struct rq
*rq
)
330 rq_csd_init(rq
, &rq
->hrtick_csd
, __hrtick_start
);
332 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
333 rq
->hrtick_timer
.function
= hrtick
;
335 #else /* CONFIG_SCHED_HRTICK */
336 static inline void hrtick_clear(struct rq
*rq
)
340 static inline void hrtick_rq_init(struct rq
*rq
)
343 #endif /* CONFIG_SCHED_HRTICK */
346 * cmpxchg based fetch_or, macro so it works for different integer types
348 #define fetch_or(ptr, mask) \
350 typeof(ptr) _ptr = (ptr); \
351 typeof(mask) _mask = (mask); \
352 typeof(*_ptr) _old, _val = *_ptr; \
355 _old = cmpxchg(_ptr, _val, _val | _mask); \
363 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
365 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
366 * this avoids any races wrt polling state changes and thereby avoids
369 static bool set_nr_and_not_polling(struct task_struct
*p
)
371 struct thread_info
*ti
= task_thread_info(p
);
372 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
376 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
378 * If this returns true, then the idle task promises to call
379 * sched_ttwu_pending() and reschedule soon.
381 static bool set_nr_if_polling(struct task_struct
*p
)
383 struct thread_info
*ti
= task_thread_info(p
);
384 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
387 if (!(val
& _TIF_POLLING_NRFLAG
))
389 if (val
& _TIF_NEED_RESCHED
)
391 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
400 static bool set_nr_and_not_polling(struct task_struct
*p
)
402 set_tsk_need_resched(p
);
407 static bool set_nr_if_polling(struct task_struct
*p
)
414 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
416 struct wake_q_node
*node
= &task
->wake_q
;
419 * Atomically grab the task, if ->wake_q is !nil already it means
420 * its already queued (either by us or someone else) and will get the
421 * wakeup due to that.
423 * In order to ensure that a pending wakeup will observe our pending
424 * state, even in the failed case, an explicit smp_mb() must be used.
426 smp_mb__before_atomic();
427 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
431 * The head is context local, there can be no concurrency.
434 head
->lastp
= &node
->next
;
439 * wake_q_add() - queue a wakeup for 'later' waking.
440 * @head: the wake_q_head to add @task to
441 * @task: the task to queue for 'later' wakeup
443 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
444 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
447 * This function must be used as-if it were wake_up_process(); IOW the task
448 * must be ready to be woken at this location.
450 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
452 if (__wake_q_add(head
, task
))
453 get_task_struct(task
);
457 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
458 * @head: the wake_q_head to add @task to
459 * @task: the task to queue for 'later' wakeup
461 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
462 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
465 * This function must be used as-if it were wake_up_process(); IOW the task
466 * must be ready to be woken at this location.
468 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
469 * that already hold reference to @task can call the 'safe' version and trust
470 * wake_q to do the right thing depending whether or not the @task is already
473 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
475 if (!__wake_q_add(head
, task
))
476 put_task_struct(task
);
479 void wake_up_q(struct wake_q_head
*head
)
481 struct wake_q_node
*node
= head
->first
;
483 while (node
!= WAKE_Q_TAIL
) {
484 struct task_struct
*task
;
486 task
= container_of(node
, struct task_struct
, wake_q
);
488 /* Task can safely be re-inserted now: */
490 task
->wake_q
.next
= NULL
;
493 * wake_up_process() executes a full barrier, which pairs with
494 * the queueing in wake_q_add() so as not to miss wakeups.
496 wake_up_process(task
);
497 put_task_struct(task
);
502 * resched_curr - mark rq's current task 'to be rescheduled now'.
504 * On UP this means the setting of the need_resched flag, on SMP it
505 * might also involve a cross-CPU call to trigger the scheduler on
508 void resched_curr(struct rq
*rq
)
510 struct task_struct
*curr
= rq
->curr
;
513 lockdep_assert_held(&rq
->lock
);
515 if (test_tsk_need_resched(curr
))
520 if (cpu
== smp_processor_id()) {
521 set_tsk_need_resched(curr
);
522 set_preempt_need_resched();
526 if (set_nr_and_not_polling(curr
))
527 smp_send_reschedule(cpu
);
529 trace_sched_wake_idle_without_ipi(cpu
);
532 void resched_cpu(int cpu
)
534 struct rq
*rq
= cpu_rq(cpu
);
537 raw_spin_lock_irqsave(&rq
->lock
, flags
);
538 if (cpu_online(cpu
) || cpu
== smp_processor_id())
540 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
544 #ifdef CONFIG_NO_HZ_COMMON
546 * In the semi idle case, use the nearest busy CPU for migrating timers
547 * from an idle CPU. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle CPU will add more delays to the timers than intended
551 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
556 struct sched_domain
*sd
;
558 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
565 for_each_domain(cpu
, sd
) {
566 for_each_cpu_and(i
, sched_domain_span(sd
),
567 housekeeping_cpumask(HK_FLAG_TIMER
)) {
578 if (default_cpu
== -1)
579 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
587 * When add_timer_on() enqueues a timer into the timer wheel of an
588 * idle CPU then this timer might expire before the next timer event
589 * which is scheduled to wake up that CPU. In case of a completely
590 * idle system the next event might even be infinite time into the
591 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
592 * leaves the inner idle loop so the newly added timer is taken into
593 * account when the CPU goes back to idle and evaluates the timer
594 * wheel for the next timer event.
596 static void wake_up_idle_cpu(int cpu
)
598 struct rq
*rq
= cpu_rq(cpu
);
600 if (cpu
== smp_processor_id())
603 if (set_nr_and_not_polling(rq
->idle
))
604 smp_send_reschedule(cpu
);
606 trace_sched_wake_idle_without_ipi(cpu
);
609 static bool wake_up_full_nohz_cpu(int cpu
)
612 * We just need the target to call irq_exit() and re-evaluate
613 * the next tick. The nohz full kick at least implies that.
614 * If needed we can still optimize that later with an
617 if (cpu_is_offline(cpu
))
618 return true; /* Don't try to wake offline CPUs. */
619 if (tick_nohz_full_cpu(cpu
)) {
620 if (cpu
!= smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 tick_nohz_full_kick_cpu(cpu
);
630 * Wake up the specified CPU. If the CPU is going offline, it is the
631 * caller's responsibility to deal with the lost wakeup, for example,
632 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
634 void wake_up_nohz_cpu(int cpu
)
636 if (!wake_up_full_nohz_cpu(cpu
))
637 wake_up_idle_cpu(cpu
);
640 static void nohz_csd_func(void *info
)
642 struct rq
*rq
= info
;
643 int cpu
= cpu_of(rq
);
647 * Release the rq::nohz_csd.
649 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
, nohz_flags(cpu
));
650 WARN_ON(!(flags
& NOHZ_KICK_MASK
));
652 rq
->idle_balance
= idle_cpu(cpu
);
653 if (rq
->idle_balance
&& !need_resched()) {
654 rq
->nohz_idle_balance
= flags
;
655 raise_softirq_irqoff(SCHED_SOFTIRQ
);
659 #endif /* CONFIG_NO_HZ_COMMON */
661 #ifdef CONFIG_NO_HZ_FULL
662 bool sched_can_stop_tick(struct rq
*rq
)
666 /* Deadline tasks, even if single, need the tick */
667 if (rq
->dl
.dl_nr_running
)
671 * If there are more than one RR tasks, we need the tick to effect the
672 * actual RR behaviour.
674 if (rq
->rt
.rr_nr_running
) {
675 if (rq
->rt
.rr_nr_running
== 1)
682 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
683 * forced preemption between FIFO tasks.
685 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
690 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
691 * if there's more than one we need the tick for involuntary
694 if (rq
->nr_running
> 1)
699 #endif /* CONFIG_NO_HZ_FULL */
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group
*from
,
711 tg_visitor down
, tg_visitor up
, void *data
)
713 struct task_group
*parent
, *child
;
719 ret
= (*down
)(parent
, data
);
722 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
729 ret
= (*up
)(parent
, data
);
730 if (ret
|| parent
== from
)
734 parent
= parent
->parent
;
741 int tg_nop(struct task_group
*tg
, void *data
)
747 static void set_load_weight(struct task_struct
*p
, bool update_load
)
749 int prio
= p
->static_prio
- MAX_RT_PRIO
;
750 struct load_weight
*load
= &p
->se
.load
;
753 * SCHED_IDLE tasks get minimal weight:
755 if (task_has_idle_policy(p
)) {
756 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
757 load
->inv_weight
= WMULT_IDLEPRIO
;
762 * SCHED_OTHER tasks have to update their load when changing their
765 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
766 reweight_task(p
, prio
);
768 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
769 load
->inv_weight
= sched_prio_to_wmult
[prio
];
773 #ifdef CONFIG_UCLAMP_TASK
775 * Serializes updates of utilization clamp values
777 * The (slow-path) user-space triggers utilization clamp value updates which
778 * can require updates on (fast-path) scheduler's data structures used to
779 * support enqueue/dequeue operations.
780 * While the per-CPU rq lock protects fast-path update operations, user-space
781 * requests are serialized using a mutex to reduce the risk of conflicting
782 * updates or API abuses.
784 static DEFINE_MUTEX(uclamp_mutex
);
786 /* Max allowed minimum utilization */
787 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
789 /* Max allowed maximum utilization */
790 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
792 /* All clamps are required to be less or equal than these values */
793 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
795 /* Integer rounded range for each bucket */
796 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
798 #define for_each_clamp_id(clamp_id) \
799 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
801 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
803 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
806 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value
)
808 return UCLAMP_BUCKET_DELTA
* uclamp_bucket_id(clamp_value
);
811 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
813 if (clamp_id
== UCLAMP_MIN
)
815 return SCHED_CAPACITY_SCALE
;
818 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
819 unsigned int value
, bool user_defined
)
821 uc_se
->value
= value
;
822 uc_se
->bucket_id
= uclamp_bucket_id(value
);
823 uc_se
->user_defined
= user_defined
;
826 static inline unsigned int
827 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
828 unsigned int clamp_value
)
831 * Avoid blocked utilization pushing up the frequency when we go
832 * idle (which drops the max-clamp) by retaining the last known
835 if (clamp_id
== UCLAMP_MAX
) {
836 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
840 return uclamp_none(UCLAMP_MIN
);
843 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
844 unsigned int clamp_value
)
846 /* Reset max-clamp retention only on idle exit */
847 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
850 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
854 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
855 unsigned int clamp_value
)
857 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
858 int bucket_id
= UCLAMP_BUCKETS
- 1;
861 * Since both min and max clamps are max aggregated, find the
862 * top most bucket with tasks in.
864 for ( ; bucket_id
>= 0; bucket_id
--) {
865 if (!bucket
[bucket_id
].tasks
)
867 return bucket
[bucket_id
].value
;
870 /* No tasks -- default clamp values */
871 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
874 static inline struct uclamp_se
875 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
877 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
878 #ifdef CONFIG_UCLAMP_TASK_GROUP
879 struct uclamp_se uc_max
;
882 * Tasks in autogroups or root task group will be
883 * restricted by system defaults.
885 if (task_group_is_autogroup(task_group(p
)))
887 if (task_group(p
) == &root_task_group
)
890 uc_max
= task_group(p
)->uclamp
[clamp_id
];
891 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
899 * The effective clamp bucket index of a task depends on, by increasing
901 * - the task specific clamp value, when explicitly requested from userspace
902 * - the task group effective clamp value, for tasks not either in the root
903 * group or in an autogroup
904 * - the system default clamp value, defined by the sysadmin
906 static inline struct uclamp_se
907 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
909 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
910 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
912 /* System default restrictions always apply */
913 if (unlikely(uc_req
.value
> uc_max
.value
))
919 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
921 struct uclamp_se uc_eff
;
923 /* Task currently refcounted: use back-annotated (effective) value */
924 if (p
->uclamp
[clamp_id
].active
)
925 return (unsigned long)p
->uclamp
[clamp_id
].value
;
927 uc_eff
= uclamp_eff_get(p
, clamp_id
);
929 return (unsigned long)uc_eff
.value
;
933 * When a task is enqueued on a rq, the clamp bucket currently defined by the
934 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
935 * updates the rq's clamp value if required.
937 * Tasks can have a task-specific value requested from user-space, track
938 * within each bucket the maximum value for tasks refcounted in it.
939 * This "local max aggregation" allows to track the exact "requested" value
940 * for each bucket when all its RUNNABLE tasks require the same clamp.
942 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
943 enum uclamp_id clamp_id
)
945 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
946 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
947 struct uclamp_bucket
*bucket
;
949 lockdep_assert_held(&rq
->lock
);
951 /* Update task effective clamp */
952 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
954 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
956 uc_se
->active
= true;
958 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
961 * Local max aggregation: rq buckets always track the max
962 * "requested" clamp value of its RUNNABLE tasks.
964 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
965 bucket
->value
= uc_se
->value
;
967 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
968 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
972 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
973 * is released. If this is the last task reference counting the rq's max
974 * active clamp value, then the rq's clamp value is updated.
976 * Both refcounted tasks and rq's cached clamp values are expected to be
977 * always valid. If it's detected they are not, as defensive programming,
978 * enforce the expected state and warn.
980 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
981 enum uclamp_id clamp_id
)
983 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
984 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
985 struct uclamp_bucket
*bucket
;
986 unsigned int bkt_clamp
;
987 unsigned int rq_clamp
;
989 lockdep_assert_held(&rq
->lock
);
991 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
992 SCHED_WARN_ON(!bucket
->tasks
);
993 if (likely(bucket
->tasks
))
995 uc_se
->active
= false;
998 * Keep "local max aggregation" simple and accept to (possibly)
999 * overboost some RUNNABLE tasks in the same bucket.
1000 * The rq clamp bucket value is reset to its base value whenever
1001 * there are no more RUNNABLE tasks refcounting it.
1003 if (likely(bucket
->tasks
))
1006 rq_clamp
= READ_ONCE(uc_rq
->value
);
1008 * Defensive programming: this should never happen. If it happens,
1009 * e.g. due to future modification, warn and fixup the expected value.
1011 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1012 if (bucket
->value
>= rq_clamp
) {
1013 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1014 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1018 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1020 enum uclamp_id clamp_id
;
1022 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1025 for_each_clamp_id(clamp_id
)
1026 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1028 /* Reset clamp idle holding when there is one RUNNABLE task */
1029 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1030 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1033 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1035 enum uclamp_id clamp_id
;
1037 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1040 for_each_clamp_id(clamp_id
)
1041 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1045 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1051 * Lock the task and the rq where the task is (or was) queued.
1053 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1054 * price to pay to safely serialize util_{min,max} updates with
1055 * enqueues, dequeues and migration operations.
1056 * This is the same locking schema used by __set_cpus_allowed_ptr().
1058 rq
= task_rq_lock(p
, &rf
);
1061 * Setting the clamp bucket is serialized by task_rq_lock().
1062 * If the task is not yet RUNNABLE and its task_struct is not
1063 * affecting a valid clamp bucket, the next time it's enqueued,
1064 * it will already see the updated clamp bucket value.
1066 if (p
->uclamp
[clamp_id
].active
) {
1067 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1068 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1071 task_rq_unlock(rq
, p
, &rf
);
1074 #ifdef CONFIG_UCLAMP_TASK_GROUP
1076 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1077 unsigned int clamps
)
1079 enum uclamp_id clamp_id
;
1080 struct css_task_iter it
;
1081 struct task_struct
*p
;
1083 css_task_iter_start(css
, 0, &it
);
1084 while ((p
= css_task_iter_next(&it
))) {
1085 for_each_clamp_id(clamp_id
) {
1086 if ((0x1 << clamp_id
) & clamps
)
1087 uclamp_update_active(p
, clamp_id
);
1090 css_task_iter_end(&it
);
1093 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1094 static void uclamp_update_root_tg(void)
1096 struct task_group
*tg
= &root_task_group
;
1098 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1099 sysctl_sched_uclamp_util_min
, false);
1100 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1101 sysctl_sched_uclamp_util_max
, false);
1104 cpu_util_update_eff(&root_task_group
.css
);
1108 static void uclamp_update_root_tg(void) { }
1111 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1112 void __user
*buffer
, size_t *lenp
,
1115 bool update_root_tg
= false;
1116 int old_min
, old_max
;
1119 mutex_lock(&uclamp_mutex
);
1120 old_min
= sysctl_sched_uclamp_util_min
;
1121 old_max
= sysctl_sched_uclamp_util_max
;
1123 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1129 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1130 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
) {
1135 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1136 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1137 sysctl_sched_uclamp_util_min
, false);
1138 update_root_tg
= true;
1140 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1141 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1142 sysctl_sched_uclamp_util_max
, false);
1143 update_root_tg
= true;
1147 uclamp_update_root_tg();
1150 * We update all RUNNABLE tasks only when task groups are in use.
1151 * Otherwise, keep it simple and do just a lazy update at each next
1152 * task enqueue time.
1158 sysctl_sched_uclamp_util_min
= old_min
;
1159 sysctl_sched_uclamp_util_max
= old_max
;
1161 mutex_unlock(&uclamp_mutex
);
1166 static int uclamp_validate(struct task_struct
*p
,
1167 const struct sched_attr
*attr
)
1169 unsigned int lower_bound
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1170 unsigned int upper_bound
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1172 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
)
1173 lower_bound
= attr
->sched_util_min
;
1174 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
)
1175 upper_bound
= attr
->sched_util_max
;
1177 if (lower_bound
> upper_bound
)
1179 if (upper_bound
> SCHED_CAPACITY_SCALE
)
1185 static void __setscheduler_uclamp(struct task_struct
*p
,
1186 const struct sched_attr
*attr
)
1188 enum uclamp_id clamp_id
;
1191 * On scheduling class change, reset to default clamps for tasks
1192 * without a task-specific value.
1194 for_each_clamp_id(clamp_id
) {
1195 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1196 unsigned int clamp_value
= uclamp_none(clamp_id
);
1198 /* Keep using defined clamps across class changes */
1199 if (uc_se
->user_defined
)
1202 /* By default, RT tasks always get 100% boost */
1203 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1204 clamp_value
= uclamp_none(UCLAMP_MAX
);
1206 uclamp_se_set(uc_se
, clamp_value
, false);
1209 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1212 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1213 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1214 attr
->sched_util_min
, true);
1217 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1218 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1219 attr
->sched_util_max
, true);
1223 static void uclamp_fork(struct task_struct
*p
)
1225 enum uclamp_id clamp_id
;
1227 for_each_clamp_id(clamp_id
)
1228 p
->uclamp
[clamp_id
].active
= false;
1230 if (likely(!p
->sched_reset_on_fork
))
1233 for_each_clamp_id(clamp_id
) {
1234 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1235 uclamp_none(clamp_id
), false);
1239 static void __init
init_uclamp(void)
1241 struct uclamp_se uc_max
= {};
1242 enum uclamp_id clamp_id
;
1245 mutex_init(&uclamp_mutex
);
1247 for_each_possible_cpu(cpu
) {
1248 memset(&cpu_rq(cpu
)->uclamp
, 0,
1249 sizeof(struct uclamp_rq
)*UCLAMP_CNT
);
1250 cpu_rq(cpu
)->uclamp_flags
= 0;
1253 for_each_clamp_id(clamp_id
) {
1254 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1255 uclamp_none(clamp_id
), false);
1258 /* System defaults allow max clamp values for both indexes */
1259 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1260 for_each_clamp_id(clamp_id
) {
1261 uclamp_default
[clamp_id
] = uc_max
;
1262 #ifdef CONFIG_UCLAMP_TASK_GROUP
1263 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1264 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1269 #else /* CONFIG_UCLAMP_TASK */
1270 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1271 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1272 static inline int uclamp_validate(struct task_struct
*p
,
1273 const struct sched_attr
*attr
)
1277 static void __setscheduler_uclamp(struct task_struct
*p
,
1278 const struct sched_attr
*attr
) { }
1279 static inline void uclamp_fork(struct task_struct
*p
) { }
1280 static inline void init_uclamp(void) { }
1281 #endif /* CONFIG_UCLAMP_TASK */
1283 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1285 if (!(flags
& ENQUEUE_NOCLOCK
))
1286 update_rq_clock(rq
);
1288 if (!(flags
& ENQUEUE_RESTORE
)) {
1289 sched_info_queued(rq
, p
);
1290 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1293 uclamp_rq_inc(rq
, p
);
1294 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1297 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1299 if (!(flags
& DEQUEUE_NOCLOCK
))
1300 update_rq_clock(rq
);
1302 if (!(flags
& DEQUEUE_SAVE
)) {
1303 sched_info_dequeued(rq
, p
);
1304 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1307 uclamp_rq_dec(rq
, p
);
1308 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1311 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1313 if (task_contributes_to_load(p
))
1314 rq
->nr_uninterruptible
--;
1316 enqueue_task(rq
, p
, flags
);
1318 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1321 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1323 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1325 if (task_contributes_to_load(p
))
1326 rq
->nr_uninterruptible
++;
1328 dequeue_task(rq
, p
, flags
);
1332 * __normal_prio - return the priority that is based on the static prio
1334 static inline int __normal_prio(struct task_struct
*p
)
1336 return p
->static_prio
;
1340 * Calculate the expected normal priority: i.e. priority
1341 * without taking RT-inheritance into account. Might be
1342 * boosted by interactivity modifiers. Changes upon fork,
1343 * setprio syscalls, and whenever the interactivity
1344 * estimator recalculates.
1346 static inline int normal_prio(struct task_struct
*p
)
1350 if (task_has_dl_policy(p
))
1351 prio
= MAX_DL_PRIO
-1;
1352 else if (task_has_rt_policy(p
))
1353 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1355 prio
= __normal_prio(p
);
1360 * Calculate the current priority, i.e. the priority
1361 * taken into account by the scheduler. This value might
1362 * be boosted by RT tasks, or might be boosted by
1363 * interactivity modifiers. Will be RT if the task got
1364 * RT-boosted. If not then it returns p->normal_prio.
1366 static int effective_prio(struct task_struct
*p
)
1368 p
->normal_prio
= normal_prio(p
);
1370 * If we are RT tasks or we were boosted to RT priority,
1371 * keep the priority unchanged. Otherwise, update priority
1372 * to the normal priority:
1374 if (!rt_prio(p
->prio
))
1375 return p
->normal_prio
;
1380 * task_curr - is this task currently executing on a CPU?
1381 * @p: the task in question.
1383 * Return: 1 if the task is currently executing. 0 otherwise.
1385 inline int task_curr(const struct task_struct
*p
)
1387 return cpu_curr(task_cpu(p
)) == p
;
1391 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1392 * use the balance_callback list if you want balancing.
1394 * this means any call to check_class_changed() must be followed by a call to
1395 * balance_callback().
1397 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1398 const struct sched_class
*prev_class
,
1401 if (prev_class
!= p
->sched_class
) {
1402 if (prev_class
->switched_from
)
1403 prev_class
->switched_from(rq
, p
);
1405 p
->sched_class
->switched_to(rq
, p
);
1406 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1407 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1410 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1412 const struct sched_class
*class;
1414 if (p
->sched_class
== rq
->curr
->sched_class
) {
1415 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1417 for_each_class(class) {
1418 if (class == rq
->curr
->sched_class
)
1420 if (class == p
->sched_class
) {
1428 * A queue event has occurred, and we're going to schedule. In
1429 * this case, we can save a useless back to back clock update.
1431 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1432 rq_clock_skip_update(rq
);
1438 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1439 * __set_cpus_allowed_ptr() and select_fallback_rq().
1441 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1443 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1446 if (is_per_cpu_kthread(p
))
1447 return cpu_online(cpu
);
1449 return cpu_active(cpu
);
1453 * This is how migration works:
1455 * 1) we invoke migration_cpu_stop() on the target CPU using
1457 * 2) stopper starts to run (implicitly forcing the migrated thread
1459 * 3) it checks whether the migrated task is still in the wrong runqueue.
1460 * 4) if it's in the wrong runqueue then the migration thread removes
1461 * it and puts it into the right queue.
1462 * 5) stopper completes and stop_one_cpu() returns and the migration
1467 * move_queued_task - move a queued task to new rq.
1469 * Returns (locked) new rq. Old rq's lock is released.
1471 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1472 struct task_struct
*p
, int new_cpu
)
1474 lockdep_assert_held(&rq
->lock
);
1476 WRITE_ONCE(p
->on_rq
, TASK_ON_RQ_MIGRATING
);
1477 dequeue_task(rq
, p
, DEQUEUE_NOCLOCK
);
1478 set_task_cpu(p
, new_cpu
);
1481 rq
= cpu_rq(new_cpu
);
1484 BUG_ON(task_cpu(p
) != new_cpu
);
1485 enqueue_task(rq
, p
, 0);
1486 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1487 check_preempt_curr(rq
, p
, 0);
1492 struct migration_arg
{
1493 struct task_struct
*task
;
1498 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1499 * this because either it can't run here any more (set_cpus_allowed()
1500 * away from this CPU, or CPU going down), or because we're
1501 * attempting to rebalance this task on exec (sched_exec).
1503 * So we race with normal scheduler movements, but that's OK, as long
1504 * as the task is no longer on this CPU.
1506 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1507 struct task_struct
*p
, int dest_cpu
)
1509 /* Affinity changed (again). */
1510 if (!is_cpu_allowed(p
, dest_cpu
))
1513 update_rq_clock(rq
);
1514 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1520 * migration_cpu_stop - this will be executed by a highprio stopper thread
1521 * and performs thread migration by bumping thread off CPU then
1522 * 'pushing' onto another runqueue.
1524 static int migration_cpu_stop(void *data
)
1526 struct migration_arg
*arg
= data
;
1527 struct task_struct
*p
= arg
->task
;
1528 struct rq
*rq
= this_rq();
1532 * The original target CPU might have gone down and we might
1533 * be on another CPU but it doesn't matter.
1535 local_irq_disable();
1537 * We need to explicitly wake pending tasks before running
1538 * __migrate_task() such that we will not miss enforcing cpus_ptr
1539 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1541 sched_ttwu_pending();
1543 raw_spin_lock(&p
->pi_lock
);
1546 * If task_rq(p) != rq, it cannot be migrated here, because we're
1547 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1548 * we're holding p->pi_lock.
1550 if (task_rq(p
) == rq
) {
1551 if (task_on_rq_queued(p
))
1552 rq
= __migrate_task(rq
, &rf
, p
, arg
->dest_cpu
);
1554 p
->wake_cpu
= arg
->dest_cpu
;
1557 raw_spin_unlock(&p
->pi_lock
);
1564 * sched_class::set_cpus_allowed must do the below, but is not required to
1565 * actually call this function.
1567 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1569 cpumask_copy(&p
->cpus_mask
, new_mask
);
1570 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1573 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1575 struct rq
*rq
= task_rq(p
);
1576 bool queued
, running
;
1578 lockdep_assert_held(&p
->pi_lock
);
1580 queued
= task_on_rq_queued(p
);
1581 running
= task_current(rq
, p
);
1585 * Because __kthread_bind() calls this on blocked tasks without
1588 lockdep_assert_held(&rq
->lock
);
1589 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
1592 put_prev_task(rq
, p
);
1594 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1597 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
1599 set_next_task(rq
, p
);
1603 * Change a given task's CPU affinity. Migrate the thread to a
1604 * proper CPU and schedule it away if the CPU it's executing on
1605 * is removed from the allowed bitmask.
1607 * NOTE: the caller must have a valid reference to the task, the
1608 * task must not exit() & deallocate itself prematurely. The
1609 * call is not atomic; no spinlocks may be held.
1611 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1612 const struct cpumask
*new_mask
, bool check
)
1614 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1615 unsigned int dest_cpu
;
1620 rq
= task_rq_lock(p
, &rf
);
1621 update_rq_clock(rq
);
1623 if (p
->flags
& PF_KTHREAD
) {
1625 * Kernel threads are allowed on online && !active CPUs
1627 cpu_valid_mask
= cpu_online_mask
;
1631 * Must re-check here, to close a race against __kthread_bind(),
1632 * sched_setaffinity() is not guaranteed to observe the flag.
1634 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1639 if (cpumask_equal(p
->cpus_ptr
, new_mask
))
1643 * Picking a ~random cpu helps in cases where we are changing affinity
1644 * for groups of tasks (ie. cpuset), so that load balancing is not
1645 * immediately required to distribute the tasks within their new mask.
1647 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
1648 if (dest_cpu
>= nr_cpu_ids
) {
1653 do_set_cpus_allowed(p
, new_mask
);
1655 if (p
->flags
& PF_KTHREAD
) {
1657 * For kernel threads that do indeed end up on online &&
1658 * !active we want to ensure they are strict per-CPU threads.
1660 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1661 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1662 p
->nr_cpus_allowed
!= 1);
1665 /* Can the task run on the task's current CPU? If so, we're done */
1666 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1669 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1670 struct migration_arg arg
= { p
, dest_cpu
};
1671 /* Need help from migration thread: drop lock and wait. */
1672 task_rq_unlock(rq
, p
, &rf
);
1673 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1675 } else if (task_on_rq_queued(p
)) {
1677 * OK, since we're going to drop the lock immediately
1678 * afterwards anyway.
1680 rq
= move_queued_task(rq
, &rf
, p
, dest_cpu
);
1683 task_rq_unlock(rq
, p
, &rf
);
1688 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1690 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1692 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1694 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1696 #ifdef CONFIG_SCHED_DEBUG
1698 * We should never call set_task_cpu() on a blocked task,
1699 * ttwu() will sort out the placement.
1701 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1705 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1706 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1707 * time relying on p->on_rq.
1709 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1710 p
->sched_class
== &fair_sched_class
&&
1711 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1713 #ifdef CONFIG_LOCKDEP
1715 * The caller should hold either p->pi_lock or rq->lock, when changing
1716 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1718 * sched_move_task() holds both and thus holding either pins the cgroup,
1721 * Furthermore, all task_rq users should acquire both locks, see
1724 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1725 lockdep_is_held(&task_rq(p
)->lock
)));
1728 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1730 WARN_ON_ONCE(!cpu_online(new_cpu
));
1733 trace_sched_migrate_task(p
, new_cpu
);
1735 if (task_cpu(p
) != new_cpu
) {
1736 if (p
->sched_class
->migrate_task_rq
)
1737 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
1738 p
->se
.nr_migrations
++;
1740 perf_event_task_migrate(p
);
1743 __set_task_cpu(p
, new_cpu
);
1746 #ifdef CONFIG_NUMA_BALANCING
1747 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1749 if (task_on_rq_queued(p
)) {
1750 struct rq
*src_rq
, *dst_rq
;
1751 struct rq_flags srf
, drf
;
1753 src_rq
= task_rq(p
);
1754 dst_rq
= cpu_rq(cpu
);
1756 rq_pin_lock(src_rq
, &srf
);
1757 rq_pin_lock(dst_rq
, &drf
);
1759 deactivate_task(src_rq
, p
, 0);
1760 set_task_cpu(p
, cpu
);
1761 activate_task(dst_rq
, p
, 0);
1762 check_preempt_curr(dst_rq
, p
, 0);
1764 rq_unpin_lock(dst_rq
, &drf
);
1765 rq_unpin_lock(src_rq
, &srf
);
1769 * Task isn't running anymore; make it appear like we migrated
1770 * it before it went to sleep. This means on wakeup we make the
1771 * previous CPU our target instead of where it really is.
1777 struct migration_swap_arg
{
1778 struct task_struct
*src_task
, *dst_task
;
1779 int src_cpu
, dst_cpu
;
1782 static int migrate_swap_stop(void *data
)
1784 struct migration_swap_arg
*arg
= data
;
1785 struct rq
*src_rq
, *dst_rq
;
1788 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1791 src_rq
= cpu_rq(arg
->src_cpu
);
1792 dst_rq
= cpu_rq(arg
->dst_cpu
);
1794 double_raw_lock(&arg
->src_task
->pi_lock
,
1795 &arg
->dst_task
->pi_lock
);
1796 double_rq_lock(src_rq
, dst_rq
);
1798 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1801 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1804 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
1807 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
1810 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1811 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1816 double_rq_unlock(src_rq
, dst_rq
);
1817 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1818 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1824 * Cross migrate two tasks
1826 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
1827 int target_cpu
, int curr_cpu
)
1829 struct migration_swap_arg arg
;
1832 arg
= (struct migration_swap_arg
){
1834 .src_cpu
= curr_cpu
,
1836 .dst_cpu
= target_cpu
,
1839 if (arg
.src_cpu
== arg
.dst_cpu
)
1843 * These three tests are all lockless; this is OK since all of them
1844 * will be re-checked with proper locks held further down the line.
1846 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1849 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
1852 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
1855 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1856 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1861 #endif /* CONFIG_NUMA_BALANCING */
1864 * wait_task_inactive - wait for a thread to unschedule.
1866 * If @match_state is nonzero, it's the @p->state value just checked and
1867 * not expected to change. If it changes, i.e. @p might have woken up,
1868 * then return zero. When we succeed in waiting for @p to be off its CPU,
1869 * we return a positive number (its total switch count). If a second call
1870 * a short while later returns the same number, the caller can be sure that
1871 * @p has remained unscheduled the whole time.
1873 * The caller must ensure that the task *will* unschedule sometime soon,
1874 * else this function might spin for a *long* time. This function can't
1875 * be called with interrupts off, or it may introduce deadlock with
1876 * smp_call_function() if an IPI is sent by the same process we are
1877 * waiting to become inactive.
1879 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1881 int running
, queued
;
1888 * We do the initial early heuristics without holding
1889 * any task-queue locks at all. We'll only try to get
1890 * the runqueue lock when things look like they will
1896 * If the task is actively running on another CPU
1897 * still, just relax and busy-wait without holding
1900 * NOTE! Since we don't hold any locks, it's not
1901 * even sure that "rq" stays as the right runqueue!
1902 * But we don't care, since "task_running()" will
1903 * return false if the runqueue has changed and p
1904 * is actually now running somewhere else!
1906 while (task_running(rq
, p
)) {
1907 if (match_state
&& unlikely(p
->state
!= match_state
))
1913 * Ok, time to look more closely! We need the rq
1914 * lock now, to be *sure*. If we're wrong, we'll
1915 * just go back and repeat.
1917 rq
= task_rq_lock(p
, &rf
);
1918 trace_sched_wait_task(p
);
1919 running
= task_running(rq
, p
);
1920 queued
= task_on_rq_queued(p
);
1922 if (!match_state
|| p
->state
== match_state
)
1923 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1924 task_rq_unlock(rq
, p
, &rf
);
1927 * If it changed from the expected state, bail out now.
1929 if (unlikely(!ncsw
))
1933 * Was it really running after all now that we
1934 * checked with the proper locks actually held?
1936 * Oops. Go back and try again..
1938 if (unlikely(running
)) {
1944 * It's not enough that it's not actively running,
1945 * it must be off the runqueue _entirely_, and not
1948 * So if it was still runnable (but just not actively
1949 * running right now), it's preempted, and we should
1950 * yield - it could be a while.
1952 if (unlikely(queued
)) {
1953 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1955 set_current_state(TASK_UNINTERRUPTIBLE
);
1956 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1961 * Ahh, all good. It wasn't running, and it wasn't
1962 * runnable, which means that it will never become
1963 * running in the future either. We're all done!
1972 * kick_process - kick a running thread to enter/exit the kernel
1973 * @p: the to-be-kicked thread
1975 * Cause a process which is running on another CPU to enter
1976 * kernel-mode, without any delay. (to get signals handled.)
1978 * NOTE: this function doesn't have to take the runqueue lock,
1979 * because all it wants to ensure is that the remote task enters
1980 * the kernel. If the IPI races and the task has been migrated
1981 * to another CPU then no harm is done and the purpose has been
1984 void kick_process(struct task_struct
*p
)
1990 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1991 smp_send_reschedule(cpu
);
1994 EXPORT_SYMBOL_GPL(kick_process
);
1997 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
1999 * A few notes on cpu_active vs cpu_online:
2001 * - cpu_active must be a subset of cpu_online
2003 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2004 * see __set_cpus_allowed_ptr(). At this point the newly online
2005 * CPU isn't yet part of the sched domains, and balancing will not
2008 * - on CPU-down we clear cpu_active() to mask the sched domains and
2009 * avoid the load balancer to place new tasks on the to be removed
2010 * CPU. Existing tasks will remain running there and will be taken
2013 * This means that fallback selection must not select !active CPUs.
2014 * And can assume that any active CPU must be online. Conversely
2015 * select_task_rq() below may allow selection of !active CPUs in order
2016 * to satisfy the above rules.
2018 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2020 int nid
= cpu_to_node(cpu
);
2021 const struct cpumask
*nodemask
= NULL
;
2022 enum { cpuset
, possible
, fail
} state
= cpuset
;
2026 * If the node that the CPU is on has been offlined, cpu_to_node()
2027 * will return -1. There is no CPU on the node, and we should
2028 * select the CPU on the other node.
2031 nodemask
= cpumask_of_node(nid
);
2033 /* Look for allowed, online CPU in same node. */
2034 for_each_cpu(dest_cpu
, nodemask
) {
2035 if (!cpu_active(dest_cpu
))
2037 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2043 /* Any allowed, online CPU? */
2044 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2045 if (!is_cpu_allowed(p
, dest_cpu
))
2051 /* No more Mr. Nice Guy. */
2054 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2055 cpuset_cpus_allowed_fallback(p
);
2061 do_set_cpus_allowed(p
, cpu_possible_mask
);
2072 if (state
!= cpuset
) {
2074 * Don't tell them about moving exiting tasks or
2075 * kernel threads (both mm NULL), since they never
2078 if (p
->mm
&& printk_ratelimit()) {
2079 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2080 task_pid_nr(p
), p
->comm
, cpu
);
2088 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2091 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
2093 lockdep_assert_held(&p
->pi_lock
);
2095 if (p
->nr_cpus_allowed
> 1)
2096 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
2098 cpu
= cpumask_any(p
->cpus_ptr
);
2101 * In order not to call set_task_cpu() on a blocking task we need
2102 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2105 * Since this is common to all placement strategies, this lives here.
2107 * [ this allows ->select_task() to simply return task_cpu(p) and
2108 * not worry about this generic constraint ]
2110 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2111 cpu
= select_fallback_rq(task_cpu(p
), p
);
2116 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2118 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2119 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2123 * Make it appear like a SCHED_FIFO task, its something
2124 * userspace knows about and won't get confused about.
2126 * Also, it will make PI more or less work without too
2127 * much confusion -- but then, stop work should not
2128 * rely on PI working anyway.
2130 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2132 stop
->sched_class
= &stop_sched_class
;
2135 cpu_rq(cpu
)->stop
= stop
;
2139 * Reset it back to a normal scheduling class so that
2140 * it can die in pieces.
2142 old_stop
->sched_class
= &rt_sched_class
;
2148 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2149 const struct cpumask
*new_mask
, bool check
)
2151 return set_cpus_allowed_ptr(p
, new_mask
);
2154 #endif /* CONFIG_SMP */
2157 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2161 if (!schedstat_enabled())
2167 if (cpu
== rq
->cpu
) {
2168 __schedstat_inc(rq
->ttwu_local
);
2169 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2171 struct sched_domain
*sd
;
2173 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2175 for_each_domain(rq
->cpu
, sd
) {
2176 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2177 __schedstat_inc(sd
->ttwu_wake_remote
);
2184 if (wake_flags
& WF_MIGRATED
)
2185 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2186 #endif /* CONFIG_SMP */
2188 __schedstat_inc(rq
->ttwu_count
);
2189 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2191 if (wake_flags
& WF_SYNC
)
2192 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2196 * Mark the task runnable and perform wakeup-preemption.
2198 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2199 struct rq_flags
*rf
)
2201 check_preempt_curr(rq
, p
, wake_flags
);
2202 p
->state
= TASK_RUNNING
;
2203 trace_sched_wakeup(p
);
2206 if (p
->sched_class
->task_woken
) {
2208 * Our task @p is fully woken up and running; so its safe to
2209 * drop the rq->lock, hereafter rq is only used for statistics.
2211 rq_unpin_lock(rq
, rf
);
2212 p
->sched_class
->task_woken(rq
, p
);
2213 rq_repin_lock(rq
, rf
);
2216 if (rq
->idle_stamp
) {
2217 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2218 u64 max
= 2*rq
->max_idle_balance_cost
;
2220 update_avg(&rq
->avg_idle
, delta
);
2222 if (rq
->avg_idle
> max
)
2231 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2232 struct rq_flags
*rf
)
2234 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2236 lockdep_assert_held(&rq
->lock
);
2239 if (p
->sched_contributes_to_load
)
2240 rq
->nr_uninterruptible
--;
2242 if (wake_flags
& WF_MIGRATED
)
2243 en_flags
|= ENQUEUE_MIGRATED
;
2246 activate_task(rq
, p
, en_flags
);
2247 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
2251 * Called in case the task @p isn't fully descheduled from its runqueue,
2252 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2253 * since all we need to do is flip p->state to TASK_RUNNING, since
2254 * the task is still ->on_rq.
2256 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2262 rq
= __task_rq_lock(p
, &rf
);
2263 if (task_on_rq_queued(p
)) {
2264 /* check_preempt_curr() may use rq clock */
2265 update_rq_clock(rq
);
2266 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
2269 __task_rq_unlock(rq
, &rf
);
2275 void sched_ttwu_pending(void)
2277 struct rq
*rq
= this_rq();
2278 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
2279 struct task_struct
*p
, *t
;
2285 rq_lock_irqsave(rq
, &rf
);
2286 update_rq_clock(rq
);
2288 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
)
2289 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
2291 rq_unlock_irqrestore(rq
, &rf
);
2294 static void wake_csd_func(void *info
)
2296 sched_ttwu_pending();
2300 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2301 * necessary. The wakee CPU on receipt of the IPI will queue the task
2302 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2303 * of the wakeup instead of the waker.
2305 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
2307 struct rq
*rq
= cpu_rq(cpu
);
2309 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
2311 if (llist_add(&p
->wake_entry
, &rq
->wake_list
)) {
2312 if (!set_nr_if_polling(rq
->idle
))
2313 smp_call_function_single_async(cpu
, &rq
->wake_csd
);
2315 trace_sched_wake_idle_without_ipi(cpu
);
2319 void wake_up_if_idle(int cpu
)
2321 struct rq
*rq
= cpu_rq(cpu
);
2326 if (!is_idle_task(rcu_dereference(rq
->curr
)))
2329 if (set_nr_if_polling(rq
->idle
)) {
2330 trace_sched_wake_idle_without_ipi(cpu
);
2332 rq_lock_irqsave(rq
, &rf
);
2333 if (is_idle_task(rq
->curr
))
2334 smp_send_reschedule(cpu
);
2335 /* Else CPU is not idle, do nothing here: */
2336 rq_unlock_irqrestore(rq
, &rf
);
2343 bool cpus_share_cache(int this_cpu
, int that_cpu
)
2345 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
2348 static inline bool ttwu_queue_cond(int cpu
, int wake_flags
)
2351 * If the CPU does not share cache, then queue the task on the
2352 * remote rqs wakelist to avoid accessing remote data.
2354 if (!cpus_share_cache(smp_processor_id(), cpu
))
2358 * If the task is descheduling and the only running task on the
2359 * CPU then use the wakelist to offload the task activation to
2360 * the soon-to-be-idle CPU as the current CPU is likely busy.
2361 * nr_running is checked to avoid unnecessary task stacking.
2363 if ((wake_flags
& WF_ON_RQ
) && cpu_rq(cpu
)->nr_running
<= 1)
2369 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
2371 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(cpu
, wake_flags
)) {
2372 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
2373 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
2379 #endif /* CONFIG_SMP */
2381 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
2383 struct rq
*rq
= cpu_rq(cpu
);
2386 #if defined(CONFIG_SMP)
2387 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
2392 update_rq_clock(rq
);
2393 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
2398 * Notes on Program-Order guarantees on SMP systems.
2402 * The basic program-order guarantee on SMP systems is that when a task [t]
2403 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2404 * execution on its new CPU [c1].
2406 * For migration (of runnable tasks) this is provided by the following means:
2408 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2409 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2410 * rq(c1)->lock (if not at the same time, then in that order).
2411 * C) LOCK of the rq(c1)->lock scheduling in task
2413 * Release/acquire chaining guarantees that B happens after A and C after B.
2414 * Note: the CPU doing B need not be c0 or c1
2423 * UNLOCK rq(0)->lock
2425 * LOCK rq(0)->lock // orders against CPU0
2427 * UNLOCK rq(0)->lock
2431 * UNLOCK rq(1)->lock
2433 * LOCK rq(1)->lock // orders against CPU2
2436 * UNLOCK rq(1)->lock
2439 * BLOCKING -- aka. SLEEP + WAKEUP
2441 * For blocking we (obviously) need to provide the same guarantee as for
2442 * migration. However the means are completely different as there is no lock
2443 * chain to provide order. Instead we do:
2445 * 1) smp_store_release(X->on_cpu, 0)
2446 * 2) smp_cond_load_acquire(!X->on_cpu)
2450 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2452 * LOCK rq(0)->lock LOCK X->pi_lock
2455 * smp_store_release(X->on_cpu, 0);
2457 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2463 * X->state = RUNNING
2464 * UNLOCK rq(2)->lock
2466 * LOCK rq(2)->lock // orders against CPU1
2469 * UNLOCK rq(2)->lock
2472 * UNLOCK rq(0)->lock
2475 * However, for wakeups there is a second guarantee we must provide, namely we
2476 * must ensure that CONDITION=1 done by the caller can not be reordered with
2477 * accesses to the task state; see try_to_wake_up() and set_current_state().
2481 * try_to_wake_up - wake up a thread
2482 * @p: the thread to be awakened
2483 * @state: the mask of task states that can be woken
2484 * @wake_flags: wake modifier flags (WF_*)
2486 * If (@state & @p->state) @p->state = TASK_RUNNING.
2488 * If the task was not queued/runnable, also place it back on a runqueue.
2490 * Atomic against schedule() which would dequeue a task, also see
2491 * set_current_state().
2493 * This function executes a full memory barrier before accessing the task
2494 * state; see set_current_state().
2496 * Return: %true if @p->state changes (an actual wakeup was done),
2500 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2502 unsigned long flags
;
2503 int cpu
, success
= 0;
2508 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2509 * == smp_processor_id()'. Together this means we can special
2510 * case the whole 'p->on_rq && ttwu_remote()' case below
2511 * without taking any locks.
2514 * - we rely on Program-Order guarantees for all the ordering,
2515 * - we're serialized against set_special_state() by virtue of
2516 * it disabling IRQs (this allows not taking ->pi_lock).
2518 if (!(p
->state
& state
))
2523 trace_sched_waking(p
);
2524 p
->state
= TASK_RUNNING
;
2525 trace_sched_wakeup(p
);
2530 * If we are going to wake up a thread waiting for CONDITION we
2531 * need to ensure that CONDITION=1 done by the caller can not be
2532 * reordered with p->state check below. This pairs with mb() in
2533 * set_current_state() the waiting thread does.
2535 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2536 smp_mb__after_spinlock();
2537 if (!(p
->state
& state
))
2540 trace_sched_waking(p
);
2542 /* We're going to change ->state: */
2547 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2548 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2549 * in smp_cond_load_acquire() below.
2551 * sched_ttwu_pending() try_to_wake_up()
2552 * STORE p->on_rq = 1 LOAD p->state
2555 * __schedule() (switch to task 'p')
2556 * LOCK rq->lock smp_rmb();
2557 * smp_mb__after_spinlock();
2561 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2563 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2564 * __schedule(). See the comment for smp_mb__after_spinlock().
2566 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2569 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2573 delayacct_blkio_end(p
);
2574 atomic_dec(&task_rq(p
)->nr_iowait
);
2578 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2579 p
->state
= TASK_WAKING
;
2582 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2583 * possible to, falsely, observe p->on_cpu == 0.
2585 * One must be running (->on_cpu == 1) in order to remove oneself
2586 * from the runqueue.
2588 * __schedule() (switch to task 'p') try_to_wake_up()
2589 * STORE p->on_cpu = 1 LOAD p->on_rq
2592 * __schedule() (put 'p' to sleep)
2593 * LOCK rq->lock smp_rmb();
2594 * smp_mb__after_spinlock();
2595 * STORE p->on_rq = 0 LOAD p->on_cpu
2597 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2598 * __schedule(). See the comment for smp_mb__after_spinlock().
2603 * If the owning (remote) CPU is still in the middle of schedule() with
2604 * this task as prev, considering queueing p on the remote CPUs wake_list
2605 * which potentially sends an IPI instead of spinning on p->on_cpu to
2606 * let the waker make forward progress. This is safe because IRQs are
2607 * disabled and the IPI will deliver after on_cpu is cleared.
2609 if (READ_ONCE(p
->on_cpu
) && ttwu_queue_wakelist(p
, cpu
, wake_flags
| WF_ON_RQ
))
2613 * If the owning (remote) CPU is still in the middle of schedule() with
2614 * this task as prev, wait until its done referencing the task.
2616 * Pairs with the smp_store_release() in finish_task().
2618 * This ensures that tasks getting woken will be fully ordered against
2619 * their previous state and preserve Program Order.
2621 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2623 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2624 if (task_cpu(p
) != cpu
) {
2625 wake_flags
|= WF_MIGRATED
;
2626 psi_ttwu_dequeue(p
);
2627 set_task_cpu(p
, cpu
);
2629 #endif /* CONFIG_SMP */
2631 ttwu_queue(p
, cpu
, wake_flags
);
2633 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2636 ttwu_stat(p
, cpu
, wake_flags
);
2643 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2644 * @p: Process for which the function is to be invoked.
2645 * @func: Function to invoke.
2646 * @arg: Argument to function.
2648 * If the specified task can be quickly locked into a definite state
2649 * (either sleeping or on a given runqueue), arrange to keep it in that
2650 * state while invoking @func(@arg). This function can use ->on_rq and
2651 * task_curr() to work out what the state is, if required. Given that
2652 * @func can be invoked with a runqueue lock held, it had better be quite
2656 * @false if the task slipped out from under the locks.
2657 * @true if the task was locked onto a runqueue or is sleeping.
2658 * However, @func can override this by returning @false.
2660 bool try_invoke_on_locked_down_task(struct task_struct
*p
, bool (*func
)(struct task_struct
*t
, void *arg
), void *arg
)
2666 lockdep_assert_irqs_enabled();
2667 raw_spin_lock_irq(&p
->pi_lock
);
2669 rq
= __task_rq_lock(p
, &rf
);
2670 if (task_rq(p
) == rq
)
2679 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
2684 raw_spin_unlock_irq(&p
->pi_lock
);
2689 * wake_up_process - Wake up a specific process
2690 * @p: The process to be woken up.
2692 * Attempt to wake up the nominated process and move it to the set of runnable
2695 * Return: 1 if the process was woken up, 0 if it was already running.
2697 * This function executes a full memory barrier before accessing the task state.
2699 int wake_up_process(struct task_struct
*p
)
2701 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2703 EXPORT_SYMBOL(wake_up_process
);
2705 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2707 return try_to_wake_up(p
, state
, 0);
2711 * Perform scheduler related setup for a newly forked process p.
2712 * p is forked by current.
2714 * __sched_fork() is basic setup used by init_idle() too:
2716 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2721 p
->se
.exec_start
= 0;
2722 p
->se
.sum_exec_runtime
= 0;
2723 p
->se
.prev_sum_exec_runtime
= 0;
2724 p
->se
.nr_migrations
= 0;
2726 INIT_LIST_HEAD(&p
->se
.group_node
);
2728 #ifdef CONFIG_FAIR_GROUP_SCHED
2729 p
->se
.cfs_rq
= NULL
;
2732 #ifdef CONFIG_SCHEDSTATS
2733 /* Even if schedstat is disabled, there should not be garbage */
2734 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2737 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2738 init_dl_task_timer(&p
->dl
);
2739 init_dl_inactive_task_timer(&p
->dl
);
2740 __dl_clear_params(p
);
2742 INIT_LIST_HEAD(&p
->rt
.run_list
);
2744 p
->rt
.time_slice
= sched_rr_timeslice
;
2748 #ifdef CONFIG_PREEMPT_NOTIFIERS
2749 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2752 #ifdef CONFIG_COMPACTION
2753 p
->capture_control
= NULL
;
2755 init_numa_balancing(clone_flags
, p
);
2758 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2760 #ifdef CONFIG_NUMA_BALANCING
2762 void set_numabalancing_state(bool enabled
)
2765 static_branch_enable(&sched_numa_balancing
);
2767 static_branch_disable(&sched_numa_balancing
);
2770 #ifdef CONFIG_PROC_SYSCTL
2771 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2772 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2776 int state
= static_branch_likely(&sched_numa_balancing
);
2778 if (write
&& !capable(CAP_SYS_ADMIN
))
2783 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2787 set_numabalancing_state(state
);
2793 #ifdef CONFIG_SCHEDSTATS
2795 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2796 static bool __initdata __sched_schedstats
= false;
2798 static void set_schedstats(bool enabled
)
2801 static_branch_enable(&sched_schedstats
);
2803 static_branch_disable(&sched_schedstats
);
2806 void force_schedstat_enabled(void)
2808 if (!schedstat_enabled()) {
2809 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2810 static_branch_enable(&sched_schedstats
);
2814 static int __init
setup_schedstats(char *str
)
2821 * This code is called before jump labels have been set up, so we can't
2822 * change the static branch directly just yet. Instead set a temporary
2823 * variable so init_schedstats() can do it later.
2825 if (!strcmp(str
, "enable")) {
2826 __sched_schedstats
= true;
2828 } else if (!strcmp(str
, "disable")) {
2829 __sched_schedstats
= false;
2834 pr_warn("Unable to parse schedstats=\n");
2838 __setup("schedstats=", setup_schedstats
);
2840 static void __init
init_schedstats(void)
2842 set_schedstats(__sched_schedstats
);
2845 #ifdef CONFIG_PROC_SYSCTL
2846 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2847 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2851 int state
= static_branch_likely(&sched_schedstats
);
2853 if (write
&& !capable(CAP_SYS_ADMIN
))
2858 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2862 set_schedstats(state
);
2865 #endif /* CONFIG_PROC_SYSCTL */
2866 #else /* !CONFIG_SCHEDSTATS */
2867 static inline void init_schedstats(void) {}
2868 #endif /* CONFIG_SCHEDSTATS */
2871 * fork()/clone()-time setup:
2873 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2875 unsigned long flags
;
2877 __sched_fork(clone_flags
, p
);
2879 * We mark the process as NEW here. This guarantees that
2880 * nobody will actually run it, and a signal or other external
2881 * event cannot wake it up and insert it on the runqueue either.
2883 p
->state
= TASK_NEW
;
2886 * Make sure we do not leak PI boosting priority to the child.
2888 p
->prio
= current
->normal_prio
;
2893 * Revert to default priority/policy on fork if requested.
2895 if (unlikely(p
->sched_reset_on_fork
)) {
2896 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2897 p
->policy
= SCHED_NORMAL
;
2898 p
->static_prio
= NICE_TO_PRIO(0);
2900 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2901 p
->static_prio
= NICE_TO_PRIO(0);
2903 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2904 set_load_weight(p
, false);
2907 * We don't need the reset flag anymore after the fork. It has
2908 * fulfilled its duty:
2910 p
->sched_reset_on_fork
= 0;
2913 if (dl_prio(p
->prio
))
2915 else if (rt_prio(p
->prio
))
2916 p
->sched_class
= &rt_sched_class
;
2918 p
->sched_class
= &fair_sched_class
;
2920 init_entity_runnable_average(&p
->se
);
2923 * The child is not yet in the pid-hash so no cgroup attach races,
2924 * and the cgroup is pinned to this child due to cgroup_fork()
2925 * is ran before sched_fork().
2927 * Silence PROVE_RCU.
2929 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2931 * We're setting the CPU for the first time, we don't migrate,
2932 * so use __set_task_cpu().
2934 __set_task_cpu(p
, smp_processor_id());
2935 if (p
->sched_class
->task_fork
)
2936 p
->sched_class
->task_fork(p
);
2937 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2939 #ifdef CONFIG_SCHED_INFO
2940 if (likely(sched_info_on()))
2941 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2943 #if defined(CONFIG_SMP)
2946 init_task_preempt_count(p
);
2948 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2949 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2954 unsigned long to_ratio(u64 period
, u64 runtime
)
2956 if (runtime
== RUNTIME_INF
)
2960 * Doing this here saves a lot of checks in all
2961 * the calling paths, and returning zero seems
2962 * safe for them anyway.
2967 return div64_u64(runtime
<< BW_SHIFT
, period
);
2971 * wake_up_new_task - wake up a newly created task for the first time.
2973 * This function will do some initial scheduler statistics housekeeping
2974 * that must be done for every newly created context, then puts the task
2975 * on the runqueue and wakes it.
2977 void wake_up_new_task(struct task_struct
*p
)
2982 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2983 p
->state
= TASK_RUNNING
;
2986 * Fork balancing, do it here and not earlier because:
2987 * - cpus_ptr can change in the fork path
2988 * - any previously selected CPU might disappear through hotplug
2990 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2991 * as we're not fully set-up yet.
2993 p
->recent_used_cpu
= task_cpu(p
);
2994 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2996 rq
= __task_rq_lock(p
, &rf
);
2997 update_rq_clock(rq
);
2998 post_init_entity_util_avg(p
);
3000 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
3001 trace_sched_wakeup_new(p
);
3002 check_preempt_curr(rq
, p
, WF_FORK
);
3004 if (p
->sched_class
->task_woken
) {
3006 * Nothing relies on rq->lock after this, so its fine to
3009 rq_unpin_lock(rq
, &rf
);
3010 p
->sched_class
->task_woken(rq
, p
);
3011 rq_repin_lock(rq
, &rf
);
3014 task_rq_unlock(rq
, p
, &rf
);
3017 #ifdef CONFIG_PREEMPT_NOTIFIERS
3019 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
3021 void preempt_notifier_inc(void)
3023 static_branch_inc(&preempt_notifier_key
);
3025 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
3027 void preempt_notifier_dec(void)
3029 static_branch_dec(&preempt_notifier_key
);
3031 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3034 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3035 * @notifier: notifier struct to register
3037 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3039 if (!static_branch_unlikely(&preempt_notifier_key
))
3040 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3042 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3044 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3047 * preempt_notifier_unregister - no longer interested in preemption notifications
3048 * @notifier: notifier struct to unregister
3050 * This is *not* safe to call from within a preemption notifier.
3052 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3054 hlist_del(¬ifier
->link
);
3056 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3058 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3060 struct preempt_notifier
*notifier
;
3062 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3063 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3066 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3068 if (static_branch_unlikely(&preempt_notifier_key
))
3069 __fire_sched_in_preempt_notifiers(curr
);
3073 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3074 struct task_struct
*next
)
3076 struct preempt_notifier
*notifier
;
3078 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3079 notifier
->ops
->sched_out(notifier
, next
);
3082 static __always_inline
void
3083 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3084 struct task_struct
*next
)
3086 if (static_branch_unlikely(&preempt_notifier_key
))
3087 __fire_sched_out_preempt_notifiers(curr
, next
);
3090 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3092 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3097 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3098 struct task_struct
*next
)
3102 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3104 static inline void prepare_task(struct task_struct
*next
)
3108 * Claim the task as running, we do this before switching to it
3109 * such that any running task will have this set.
3115 static inline void finish_task(struct task_struct
*prev
)
3119 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3120 * We must ensure this doesn't happen until the switch is completely
3123 * In particular, the load of prev->state in finish_task_switch() must
3124 * happen before this.
3126 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3128 smp_store_release(&prev
->on_cpu
, 0);
3133 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
3136 * Since the runqueue lock will be released by the next
3137 * task (which is an invalid locking op but in the case
3138 * of the scheduler it's an obvious special-case), so we
3139 * do an early lockdep release here:
3141 rq_unpin_lock(rq
, rf
);
3142 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
3143 #ifdef CONFIG_DEBUG_SPINLOCK
3144 /* this is a valid case when another task releases the spinlock */
3145 rq
->lock
.owner
= next
;
3149 static inline void finish_lock_switch(struct rq
*rq
)
3152 * If we are tracking spinlock dependencies then we have to
3153 * fix up the runqueue lock - which gets 'carried over' from
3154 * prev into current:
3156 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
3157 raw_spin_unlock_irq(&rq
->lock
);
3161 * NOP if the arch has not defined these:
3164 #ifndef prepare_arch_switch
3165 # define prepare_arch_switch(next) do { } while (0)
3168 #ifndef finish_arch_post_lock_switch
3169 # define finish_arch_post_lock_switch() do { } while (0)
3173 * prepare_task_switch - prepare to switch tasks
3174 * @rq: the runqueue preparing to switch
3175 * @prev: the current task that is being switched out
3176 * @next: the task we are going to switch to.
3178 * This is called with the rq lock held and interrupts off. It must
3179 * be paired with a subsequent finish_task_switch after the context
3182 * prepare_task_switch sets up locking and calls architecture specific
3186 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3187 struct task_struct
*next
)
3189 kcov_prepare_switch(prev
);
3190 sched_info_switch(rq
, prev
, next
);
3191 perf_event_task_sched_out(prev
, next
);
3193 fire_sched_out_preempt_notifiers(prev
, next
);
3195 prepare_arch_switch(next
);
3199 * finish_task_switch - clean up after a task-switch
3200 * @prev: the thread we just switched away from.
3202 * finish_task_switch must be called after the context switch, paired
3203 * with a prepare_task_switch call before the context switch.
3204 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3205 * and do any other architecture-specific cleanup actions.
3207 * Note that we may have delayed dropping an mm in context_switch(). If
3208 * so, we finish that here outside of the runqueue lock. (Doing it
3209 * with the lock held can cause deadlocks; see schedule() for
3212 * The context switch have flipped the stack from under us and restored the
3213 * local variables which were saved when this task called schedule() in the
3214 * past. prev == current is still correct but we need to recalculate this_rq
3215 * because prev may have moved to another CPU.
3217 static struct rq
*finish_task_switch(struct task_struct
*prev
)
3218 __releases(rq
->lock
)
3220 struct rq
*rq
= this_rq();
3221 struct mm_struct
*mm
= rq
->prev_mm
;
3225 * The previous task will have left us with a preempt_count of 2
3226 * because it left us after:
3229 * preempt_disable(); // 1
3231 * raw_spin_lock_irq(&rq->lock) // 2
3233 * Also, see FORK_PREEMPT_COUNT.
3235 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
3236 "corrupted preempt_count: %s/%d/0x%x\n",
3237 current
->comm
, current
->pid
, preempt_count()))
3238 preempt_count_set(FORK_PREEMPT_COUNT
);
3243 * A task struct has one reference for the use as "current".
3244 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3245 * schedule one last time. The schedule call will never return, and
3246 * the scheduled task must drop that reference.
3248 * We must observe prev->state before clearing prev->on_cpu (in
3249 * finish_task), otherwise a concurrent wakeup can get prev
3250 * running on another CPU and we could rave with its RUNNING -> DEAD
3251 * transition, resulting in a double drop.
3253 prev_state
= prev
->state
;
3254 vtime_task_switch(prev
);
3255 perf_event_task_sched_in(prev
, current
);
3257 finish_lock_switch(rq
);
3258 finish_arch_post_lock_switch();
3259 kcov_finish_switch(current
);
3261 fire_sched_in_preempt_notifiers(current
);
3263 * When switching through a kernel thread, the loop in
3264 * membarrier_{private,global}_expedited() may have observed that
3265 * kernel thread and not issued an IPI. It is therefore possible to
3266 * schedule between user->kernel->user threads without passing though
3267 * switch_mm(). Membarrier requires a barrier after storing to
3268 * rq->curr, before returning to userspace, so provide them here:
3270 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3271 * provided by mmdrop(),
3272 * - a sync_core for SYNC_CORE.
3275 membarrier_mm_sync_core_before_usermode(mm
);
3278 if (unlikely(prev_state
== TASK_DEAD
)) {
3279 if (prev
->sched_class
->task_dead
)
3280 prev
->sched_class
->task_dead(prev
);
3283 * Remove function-return probe instances associated with this
3284 * task and put them back on the free list.
3286 kprobe_flush_task(prev
);
3288 /* Task is done with its stack. */
3289 put_task_stack(prev
);
3291 put_task_struct_rcu_user(prev
);
3294 tick_nohz_task_switch();
3300 /* rq->lock is NOT held, but preemption is disabled */
3301 static void __balance_callback(struct rq
*rq
)
3303 struct callback_head
*head
, *next
;
3304 void (*func
)(struct rq
*rq
);
3305 unsigned long flags
;
3307 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3308 head
= rq
->balance_callback
;
3309 rq
->balance_callback
= NULL
;
3311 func
= (void (*)(struct rq
*))head
->func
;
3318 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3321 static inline void balance_callback(struct rq
*rq
)
3323 if (unlikely(rq
->balance_callback
))
3324 __balance_callback(rq
);
3329 static inline void balance_callback(struct rq
*rq
)
3336 * schedule_tail - first thing a freshly forked thread must call.
3337 * @prev: the thread we just switched away from.
3339 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
3340 __releases(rq
->lock
)
3345 * New tasks start with FORK_PREEMPT_COUNT, see there and
3346 * finish_task_switch() for details.
3348 * finish_task_switch() will drop rq->lock() and lower preempt_count
3349 * and the preempt_enable() will end up enabling preemption (on
3350 * PREEMPT_COUNT kernels).
3353 rq
= finish_task_switch(prev
);
3354 balance_callback(rq
);
3357 if (current
->set_child_tid
)
3358 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3360 calculate_sigpending();
3364 * context_switch - switch to the new MM and the new thread's register state.
3366 static __always_inline
struct rq
*
3367 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3368 struct task_struct
*next
, struct rq_flags
*rf
)
3370 prepare_task_switch(rq
, prev
, next
);
3373 * For paravirt, this is coupled with an exit in switch_to to
3374 * combine the page table reload and the switch backend into
3377 arch_start_context_switch(prev
);
3380 * kernel -> kernel lazy + transfer active
3381 * user -> kernel lazy + mmgrab() active
3383 * kernel -> user switch + mmdrop() active
3384 * user -> user switch
3386 if (!next
->mm
) { // to kernel
3387 enter_lazy_tlb(prev
->active_mm
, next
);
3389 next
->active_mm
= prev
->active_mm
;
3390 if (prev
->mm
) // from user
3391 mmgrab(prev
->active_mm
);
3393 prev
->active_mm
= NULL
;
3395 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
3397 * sys_membarrier() requires an smp_mb() between setting
3398 * rq->curr / membarrier_switch_mm() and returning to userspace.
3400 * The below provides this either through switch_mm(), or in
3401 * case 'prev->active_mm == next->mm' through
3402 * finish_task_switch()'s mmdrop().
3404 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
3406 if (!prev
->mm
) { // from kernel
3407 /* will mmdrop() in finish_task_switch(). */
3408 rq
->prev_mm
= prev
->active_mm
;
3409 prev
->active_mm
= NULL
;
3413 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3415 prepare_lock_switch(rq
, next
, rf
);
3417 /* Here we just switch the register state and the stack. */
3418 switch_to(prev
, next
, prev
);
3421 return finish_task_switch(prev
);
3425 * nr_running and nr_context_switches:
3427 * externally visible scheduler statistics: current number of runnable
3428 * threads, total number of context switches performed since bootup.
3430 unsigned long nr_running(void)
3432 unsigned long i
, sum
= 0;
3434 for_each_online_cpu(i
)
3435 sum
+= cpu_rq(i
)->nr_running
;
3441 * Check if only the current task is running on the CPU.
3443 * Caution: this function does not check that the caller has disabled
3444 * preemption, thus the result might have a time-of-check-to-time-of-use
3445 * race. The caller is responsible to use it correctly, for example:
3447 * - from a non-preemptible section (of course)
3449 * - from a thread that is bound to a single CPU
3451 * - in a loop with very short iterations (e.g. a polling loop)
3453 bool single_task_running(void)
3455 return raw_rq()->nr_running
== 1;
3457 EXPORT_SYMBOL(single_task_running
);
3459 unsigned long long nr_context_switches(void)
3462 unsigned long long sum
= 0;
3464 for_each_possible_cpu(i
)
3465 sum
+= cpu_rq(i
)->nr_switches
;
3471 * Consumers of these two interfaces, like for example the cpuidle menu
3472 * governor, are using nonsensical data. Preferring shallow idle state selection
3473 * for a CPU that has IO-wait which might not even end up running the task when
3474 * it does become runnable.
3477 unsigned long nr_iowait_cpu(int cpu
)
3479 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
3483 * IO-wait accounting, and how its mostly bollocks (on SMP).
3485 * The idea behind IO-wait account is to account the idle time that we could
3486 * have spend running if it were not for IO. That is, if we were to improve the
3487 * storage performance, we'd have a proportional reduction in IO-wait time.
3489 * This all works nicely on UP, where, when a task blocks on IO, we account
3490 * idle time as IO-wait, because if the storage were faster, it could've been
3491 * running and we'd not be idle.
3493 * This has been extended to SMP, by doing the same for each CPU. This however
3496 * Imagine for instance the case where two tasks block on one CPU, only the one
3497 * CPU will have IO-wait accounted, while the other has regular idle. Even
3498 * though, if the storage were faster, both could've ran at the same time,
3499 * utilising both CPUs.
3501 * This means, that when looking globally, the current IO-wait accounting on
3502 * SMP is a lower bound, by reason of under accounting.
3504 * Worse, since the numbers are provided per CPU, they are sometimes
3505 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3506 * associated with any one particular CPU, it can wake to another CPU than it
3507 * blocked on. This means the per CPU IO-wait number is meaningless.
3509 * Task CPU affinities can make all that even more 'interesting'.
3512 unsigned long nr_iowait(void)
3514 unsigned long i
, sum
= 0;
3516 for_each_possible_cpu(i
)
3517 sum
+= nr_iowait_cpu(i
);
3525 * sched_exec - execve() is a valuable balancing opportunity, because at
3526 * this point the task has the smallest effective memory and cache footprint.
3528 void sched_exec(void)
3530 struct task_struct
*p
= current
;
3531 unsigned long flags
;
3534 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3535 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
3536 if (dest_cpu
== smp_processor_id())
3539 if (likely(cpu_active(dest_cpu
))) {
3540 struct migration_arg arg
= { p
, dest_cpu
};
3542 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3543 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3547 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3552 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3553 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3555 EXPORT_PER_CPU_SYMBOL(kstat
);
3556 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3559 * The function fair_sched_class.update_curr accesses the struct curr
3560 * and its field curr->exec_start; when called from task_sched_runtime(),
3561 * we observe a high rate of cache misses in practice.
3562 * Prefetching this data results in improved performance.
3564 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3566 #ifdef CONFIG_FAIR_GROUP_SCHED
3567 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3569 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3572 prefetch(&curr
->exec_start
);
3576 * Return accounted runtime for the task.
3577 * In case the task is currently running, return the runtime plus current's
3578 * pending runtime that have not been accounted yet.
3580 unsigned long long task_sched_runtime(struct task_struct
*p
)
3586 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3588 * 64-bit doesn't need locks to atomically read a 64-bit value.
3589 * So we have a optimization chance when the task's delta_exec is 0.
3590 * Reading ->on_cpu is racy, but this is ok.
3592 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3593 * If we race with it entering CPU, unaccounted time is 0. This is
3594 * indistinguishable from the read occurring a few cycles earlier.
3595 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3596 * been accounted, so we're correct here as well.
3598 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3599 return p
->se
.sum_exec_runtime
;
3602 rq
= task_rq_lock(p
, &rf
);
3604 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3605 * project cycles that may never be accounted to this
3606 * thread, breaking clock_gettime().
3608 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3609 prefetch_curr_exec_start(p
);
3610 update_rq_clock(rq
);
3611 p
->sched_class
->update_curr(rq
);
3613 ns
= p
->se
.sum_exec_runtime
;
3614 task_rq_unlock(rq
, p
, &rf
);
3619 DEFINE_PER_CPU(unsigned long, thermal_pressure
);
3621 void arch_set_thermal_pressure(struct cpumask
*cpus
,
3622 unsigned long th_pressure
)
3626 for_each_cpu(cpu
, cpus
)
3627 WRITE_ONCE(per_cpu(thermal_pressure
, cpu
), th_pressure
);
3631 * This function gets called by the timer code, with HZ frequency.
3632 * We call it with interrupts disabled.
3634 void scheduler_tick(void)
3636 int cpu
= smp_processor_id();
3637 struct rq
*rq
= cpu_rq(cpu
);
3638 struct task_struct
*curr
= rq
->curr
;
3640 unsigned long thermal_pressure
;
3642 arch_scale_freq_tick();
3647 update_rq_clock(rq
);
3648 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
3649 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
3650 curr
->sched_class
->task_tick(rq
, curr
, 0);
3651 calc_global_load_tick(rq
);
3656 perf_event_task_tick();
3659 rq
->idle_balance
= idle_cpu(cpu
);
3660 trigger_load_balance(rq
);
3664 #ifdef CONFIG_NO_HZ_FULL
3669 struct delayed_work work
;
3671 /* Values for ->state, see diagram below. */
3672 #define TICK_SCHED_REMOTE_OFFLINE 0
3673 #define TICK_SCHED_REMOTE_OFFLINING 1
3674 #define TICK_SCHED_REMOTE_RUNNING 2
3677 * State diagram for ->state:
3680 * TICK_SCHED_REMOTE_OFFLINE
3683 * | | sched_tick_remote()
3686 * +--TICK_SCHED_REMOTE_OFFLINING
3689 * sched_tick_start() | | sched_tick_stop()
3692 * TICK_SCHED_REMOTE_RUNNING
3695 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3696 * and sched_tick_start() are happy to leave the state in RUNNING.
3699 static struct tick_work __percpu
*tick_work_cpu
;
3701 static void sched_tick_remote(struct work_struct
*work
)
3703 struct delayed_work
*dwork
= to_delayed_work(work
);
3704 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
3705 int cpu
= twork
->cpu
;
3706 struct rq
*rq
= cpu_rq(cpu
);
3707 struct task_struct
*curr
;
3713 * Handle the tick only if it appears the remote CPU is running in full
3714 * dynticks mode. The check is racy by nature, but missing a tick or
3715 * having one too much is no big deal because the scheduler tick updates
3716 * statistics and checks timeslices in a time-independent way, regardless
3717 * of when exactly it is running.
3719 if (!tick_nohz_tick_stopped_cpu(cpu
))
3722 rq_lock_irq(rq
, &rf
);
3724 if (cpu_is_offline(cpu
))
3727 update_rq_clock(rq
);
3729 if (!is_idle_task(curr
)) {
3731 * Make sure the next tick runs within a reasonable
3734 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
3735 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
3737 curr
->sched_class
->task_tick(rq
, curr
, 0);
3739 calc_load_nohz_remote(rq
);
3741 rq_unlock_irq(rq
, &rf
);
3745 * Run the remote tick once per second (1Hz). This arbitrary
3746 * frequency is large enough to avoid overload but short enough
3747 * to keep scheduler internal stats reasonably up to date. But
3748 * first update state to reflect hotplug activity if required.
3750 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
3751 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
3752 if (os
== TICK_SCHED_REMOTE_RUNNING
)
3753 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
3756 static void sched_tick_start(int cpu
)
3759 struct tick_work
*twork
;
3761 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3764 WARN_ON_ONCE(!tick_work_cpu
);
3766 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3767 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
3768 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
3769 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
3771 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
3772 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
3776 #ifdef CONFIG_HOTPLUG_CPU
3777 static void sched_tick_stop(int cpu
)
3779 struct tick_work
*twork
;
3782 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
3785 WARN_ON_ONCE(!tick_work_cpu
);
3787 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
3788 /* There cannot be competing actions, but don't rely on stop-machine. */
3789 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
3790 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
3791 /* Don't cancel, as this would mess up the state machine. */
3793 #endif /* CONFIG_HOTPLUG_CPU */
3795 int __init
sched_tick_offload_init(void)
3797 tick_work_cpu
= alloc_percpu(struct tick_work
);
3798 BUG_ON(!tick_work_cpu
);
3802 #else /* !CONFIG_NO_HZ_FULL */
3803 static inline void sched_tick_start(int cpu
) { }
3804 static inline void sched_tick_stop(int cpu
) { }
3807 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3808 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3810 * If the value passed in is equal to the current preempt count
3811 * then we just disabled preemption. Start timing the latency.
3813 static inline void preempt_latency_start(int val
)
3815 if (preempt_count() == val
) {
3816 unsigned long ip
= get_lock_parent_ip();
3817 #ifdef CONFIG_DEBUG_PREEMPT
3818 current
->preempt_disable_ip
= ip
;
3820 trace_preempt_off(CALLER_ADDR0
, ip
);
3824 void preempt_count_add(int val
)
3826 #ifdef CONFIG_DEBUG_PREEMPT
3830 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3833 __preempt_count_add(val
);
3834 #ifdef CONFIG_DEBUG_PREEMPT
3836 * Spinlock count overflowing soon?
3838 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3841 preempt_latency_start(val
);
3843 EXPORT_SYMBOL(preempt_count_add
);
3844 NOKPROBE_SYMBOL(preempt_count_add
);
3847 * If the value passed in equals to the current preempt count
3848 * then we just enabled preemption. Stop timing the latency.
3850 static inline void preempt_latency_stop(int val
)
3852 if (preempt_count() == val
)
3853 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3856 void preempt_count_sub(int val
)
3858 #ifdef CONFIG_DEBUG_PREEMPT
3862 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3865 * Is the spinlock portion underflowing?
3867 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3868 !(preempt_count() & PREEMPT_MASK
)))
3872 preempt_latency_stop(val
);
3873 __preempt_count_sub(val
);
3875 EXPORT_SYMBOL(preempt_count_sub
);
3876 NOKPROBE_SYMBOL(preempt_count_sub
);
3879 static inline void preempt_latency_start(int val
) { }
3880 static inline void preempt_latency_stop(int val
) { }
3883 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
3885 #ifdef CONFIG_DEBUG_PREEMPT
3886 return p
->preempt_disable_ip
;
3893 * Print scheduling while atomic bug:
3895 static noinline
void __schedule_bug(struct task_struct
*prev
)
3897 /* Save this before calling printk(), since that will clobber it */
3898 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3900 if (oops_in_progress
)
3903 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3904 prev
->comm
, prev
->pid
, preempt_count());
3906 debug_show_held_locks(prev
);
3908 if (irqs_disabled())
3909 print_irqtrace_events(prev
);
3910 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3911 && in_atomic_preempt_off()) {
3912 pr_err("Preemption disabled at:");
3913 print_ip_sym(preempt_disable_ip
);
3917 panic("scheduling while atomic\n");
3920 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3924 * Various schedule()-time debugging checks and statistics:
3926 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
3928 #ifdef CONFIG_SCHED_STACK_END_CHECK
3929 if (task_stack_end_corrupted(prev
))
3930 panic("corrupted stack end detected inside scheduler\n");
3933 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3934 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
3935 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3936 prev
->comm
, prev
->pid
, prev
->non_block_count
);
3938 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3942 if (unlikely(in_atomic_preempt_off())) {
3943 __schedule_bug(prev
);
3944 preempt_count_set(PREEMPT_DISABLED
);
3948 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3950 schedstat_inc(this_rq()->sched_count
);
3953 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
3954 struct rq_flags
*rf
)
3957 const struct sched_class
*class;
3959 * We must do the balancing pass before put_prev_task(), such
3960 * that when we release the rq->lock the task is in the same
3961 * state as before we took rq->lock.
3963 * We can terminate the balance pass as soon as we know there is
3964 * a runnable task of @class priority or higher.
3966 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
3967 if (class->balance(rq
, prev
, rf
))
3972 put_prev_task(rq
, prev
);
3976 * Pick up the highest-prio task:
3978 static inline struct task_struct
*
3979 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3981 const struct sched_class
*class;
3982 struct task_struct
*p
;
3985 * Optimization: we know that if all tasks are in the fair class we can
3986 * call that function directly, but only if the @prev task wasn't of a
3987 * higher scheduling class, because otherwise those loose the
3988 * opportunity to pull in more work from other CPUs.
3990 if (likely((prev
->sched_class
== &idle_sched_class
||
3991 prev
->sched_class
== &fair_sched_class
) &&
3992 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3994 p
= pick_next_task_fair(rq
, prev
, rf
);
3995 if (unlikely(p
== RETRY_TASK
))
3998 /* Assumes fair_sched_class->next == idle_sched_class */
4000 put_prev_task(rq
, prev
);
4001 p
= pick_next_task_idle(rq
);
4008 put_prev_task_balance(rq
, prev
, rf
);
4010 for_each_class(class) {
4011 p
= class->pick_next_task(rq
);
4016 /* The idle class should always have a runnable task: */
4021 * __schedule() is the main scheduler function.
4023 * The main means of driving the scheduler and thus entering this function are:
4025 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4027 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4028 * paths. For example, see arch/x86/entry_64.S.
4030 * To drive preemption between tasks, the scheduler sets the flag in timer
4031 * interrupt handler scheduler_tick().
4033 * 3. Wakeups don't really cause entry into schedule(). They add a
4034 * task to the run-queue and that's it.
4036 * Now, if the new task added to the run-queue preempts the current
4037 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4038 * called on the nearest possible occasion:
4040 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4042 * - in syscall or exception context, at the next outmost
4043 * preempt_enable(). (this might be as soon as the wake_up()'s
4046 * - in IRQ context, return from interrupt-handler to
4047 * preemptible context
4049 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4052 * - cond_resched() call
4053 * - explicit schedule() call
4054 * - return from syscall or exception to user-space
4055 * - return from interrupt-handler to user-space
4057 * WARNING: must be called with preemption disabled!
4059 static void __sched notrace
__schedule(bool preempt
)
4061 struct task_struct
*prev
, *next
;
4062 unsigned long *switch_count
;
4067 cpu
= smp_processor_id();
4071 schedule_debug(prev
, preempt
);
4073 if (sched_feat(HRTICK
))
4076 local_irq_disable();
4077 rcu_note_context_switch(preempt
);
4080 * Make sure that signal_pending_state()->signal_pending() below
4081 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4082 * done by the caller to avoid the race with signal_wake_up().
4084 * The membarrier system call requires a full memory barrier
4085 * after coming from user-space, before storing to rq->curr.
4088 smp_mb__after_spinlock();
4090 /* Promote REQ to ACT */
4091 rq
->clock_update_flags
<<= 1;
4092 update_rq_clock(rq
);
4094 switch_count
= &prev
->nivcsw
;
4095 if (!preempt
&& prev
->state
) {
4096 if (signal_pending_state(prev
->state
, prev
)) {
4097 prev
->state
= TASK_RUNNING
;
4099 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
4101 if (prev
->in_iowait
) {
4102 atomic_inc(&rq
->nr_iowait
);
4103 delayacct_blkio_start();
4106 switch_count
= &prev
->nvcsw
;
4109 next
= pick_next_task(rq
, prev
, &rf
);
4110 clear_tsk_need_resched(prev
);
4111 clear_preempt_need_resched();
4113 if (likely(prev
!= next
)) {
4116 * RCU users of rcu_dereference(rq->curr) may not see
4117 * changes to task_struct made by pick_next_task().
4119 RCU_INIT_POINTER(rq
->curr
, next
);
4121 * The membarrier system call requires each architecture
4122 * to have a full memory barrier after updating
4123 * rq->curr, before returning to user-space.
4125 * Here are the schemes providing that barrier on the
4126 * various architectures:
4127 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4128 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4129 * - finish_lock_switch() for weakly-ordered
4130 * architectures where spin_unlock is a full barrier,
4131 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4132 * is a RELEASE barrier),
4136 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
4138 trace_sched_switch(preempt
, prev
, next
);
4140 /* Also unlocks the rq: */
4141 rq
= context_switch(rq
, prev
, next
, &rf
);
4143 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4144 rq_unlock_irq(rq
, &rf
);
4147 balance_callback(rq
);
4150 void __noreturn
do_task_dead(void)
4152 /* Causes final put_task_struct in finish_task_switch(): */
4153 set_special_state(TASK_DEAD
);
4155 /* Tell freezer to ignore us: */
4156 current
->flags
|= PF_NOFREEZE
;
4161 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4166 static inline void sched_submit_work(struct task_struct
*tsk
)
4172 * If a worker went to sleep, notify and ask workqueue whether
4173 * it wants to wake up a task to maintain concurrency.
4174 * As this function is called inside the schedule() context,
4175 * we disable preemption to avoid it calling schedule() again
4176 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4179 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4181 if (tsk
->flags
& PF_WQ_WORKER
)
4182 wq_worker_sleeping(tsk
);
4184 io_wq_worker_sleeping(tsk
);
4185 preempt_enable_no_resched();
4188 if (tsk_is_pi_blocked(tsk
))
4192 * If we are going to sleep and we have plugged IO queued,
4193 * make sure to submit it to avoid deadlocks.
4195 if (blk_needs_flush_plug(tsk
))
4196 blk_schedule_flush_plug(tsk
);
4199 static void sched_update_worker(struct task_struct
*tsk
)
4201 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
4202 if (tsk
->flags
& PF_WQ_WORKER
)
4203 wq_worker_running(tsk
);
4205 io_wq_worker_running(tsk
);
4209 asmlinkage __visible
void __sched
schedule(void)
4211 struct task_struct
*tsk
= current
;
4213 sched_submit_work(tsk
);
4217 sched_preempt_enable_no_resched();
4218 } while (need_resched());
4219 sched_update_worker(tsk
);
4221 EXPORT_SYMBOL(schedule
);
4224 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4225 * state (have scheduled out non-voluntarily) by making sure that all
4226 * tasks have either left the run queue or have gone into user space.
4227 * As idle tasks do not do either, they must not ever be preempted
4228 * (schedule out non-voluntarily).
4230 * schedule_idle() is similar to schedule_preempt_disable() except that it
4231 * never enables preemption because it does not call sched_submit_work().
4233 void __sched
schedule_idle(void)
4236 * As this skips calling sched_submit_work(), which the idle task does
4237 * regardless because that function is a nop when the task is in a
4238 * TASK_RUNNING state, make sure this isn't used someplace that the
4239 * current task can be in any other state. Note, idle is always in the
4240 * TASK_RUNNING state.
4242 WARN_ON_ONCE(current
->state
);
4245 } while (need_resched());
4248 #ifdef CONFIG_CONTEXT_TRACKING
4249 asmlinkage __visible
void __sched
schedule_user(void)
4252 * If we come here after a random call to set_need_resched(),
4253 * or we have been woken up remotely but the IPI has not yet arrived,
4254 * we haven't yet exited the RCU idle mode. Do it here manually until
4255 * we find a better solution.
4257 * NB: There are buggy callers of this function. Ideally we
4258 * should warn if prev_state != CONTEXT_USER, but that will trigger
4259 * too frequently to make sense yet.
4261 enum ctx_state prev_state
= exception_enter();
4263 exception_exit(prev_state
);
4268 * schedule_preempt_disabled - called with preemption disabled
4270 * Returns with preemption disabled. Note: preempt_count must be 1
4272 void __sched
schedule_preempt_disabled(void)
4274 sched_preempt_enable_no_resched();
4279 static void __sched notrace
preempt_schedule_common(void)
4283 * Because the function tracer can trace preempt_count_sub()
4284 * and it also uses preempt_enable/disable_notrace(), if
4285 * NEED_RESCHED is set, the preempt_enable_notrace() called
4286 * by the function tracer will call this function again and
4287 * cause infinite recursion.
4289 * Preemption must be disabled here before the function
4290 * tracer can trace. Break up preempt_disable() into two
4291 * calls. One to disable preemption without fear of being
4292 * traced. The other to still record the preemption latency,
4293 * which can also be traced by the function tracer.
4295 preempt_disable_notrace();
4296 preempt_latency_start(1);
4298 preempt_latency_stop(1);
4299 preempt_enable_no_resched_notrace();
4302 * Check again in case we missed a preemption opportunity
4303 * between schedule and now.
4305 } while (need_resched());
4308 #ifdef CONFIG_PREEMPTION
4310 * This is the entry point to schedule() from in-kernel preemption
4311 * off of preempt_enable.
4313 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
4316 * If there is a non-zero preempt_count or interrupts are disabled,
4317 * we do not want to preempt the current task. Just return..
4319 if (likely(!preemptible()))
4322 preempt_schedule_common();
4324 NOKPROBE_SYMBOL(preempt_schedule
);
4325 EXPORT_SYMBOL(preempt_schedule
);
4328 * preempt_schedule_notrace - preempt_schedule called by tracing
4330 * The tracing infrastructure uses preempt_enable_notrace to prevent
4331 * recursion and tracing preempt enabling caused by the tracing
4332 * infrastructure itself. But as tracing can happen in areas coming
4333 * from userspace or just about to enter userspace, a preempt enable
4334 * can occur before user_exit() is called. This will cause the scheduler
4335 * to be called when the system is still in usermode.
4337 * To prevent this, the preempt_enable_notrace will use this function
4338 * instead of preempt_schedule() to exit user context if needed before
4339 * calling the scheduler.
4341 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
4343 enum ctx_state prev_ctx
;
4345 if (likely(!preemptible()))
4350 * Because the function tracer can trace preempt_count_sub()
4351 * and it also uses preempt_enable/disable_notrace(), if
4352 * NEED_RESCHED is set, the preempt_enable_notrace() called
4353 * by the function tracer will call this function again and
4354 * cause infinite recursion.
4356 * Preemption must be disabled here before the function
4357 * tracer can trace. Break up preempt_disable() into two
4358 * calls. One to disable preemption without fear of being
4359 * traced. The other to still record the preemption latency,
4360 * which can also be traced by the function tracer.
4362 preempt_disable_notrace();
4363 preempt_latency_start(1);
4365 * Needs preempt disabled in case user_exit() is traced
4366 * and the tracer calls preempt_enable_notrace() causing
4367 * an infinite recursion.
4369 prev_ctx
= exception_enter();
4371 exception_exit(prev_ctx
);
4373 preempt_latency_stop(1);
4374 preempt_enable_no_resched_notrace();
4375 } while (need_resched());
4377 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
4379 #endif /* CONFIG_PREEMPTION */
4382 * This is the entry point to schedule() from kernel preemption
4383 * off of irq context.
4384 * Note, that this is called and return with irqs disabled. This will
4385 * protect us against recursive calling from irq.
4387 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
4389 enum ctx_state prev_state
;
4391 /* Catch callers which need to be fixed */
4392 BUG_ON(preempt_count() || !irqs_disabled());
4394 prev_state
= exception_enter();
4400 local_irq_disable();
4401 sched_preempt_enable_no_resched();
4402 } while (need_resched());
4404 exception_exit(prev_state
);
4407 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
4410 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4412 EXPORT_SYMBOL(default_wake_function
);
4414 #ifdef CONFIG_RT_MUTEXES
4416 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
4419 prio
= min(prio
, pi_task
->prio
);
4424 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4426 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
4428 return __rt_effective_prio(pi_task
, prio
);
4432 * rt_mutex_setprio - set the current priority of a task
4434 * @pi_task: donor task
4436 * This function changes the 'effective' priority of a task. It does
4437 * not touch ->normal_prio like __setscheduler().
4439 * Used by the rt_mutex code to implement priority inheritance
4440 * logic. Call site only calls if the priority of the task changed.
4442 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
4444 int prio
, oldprio
, queued
, running
, queue_flag
=
4445 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4446 const struct sched_class
*prev_class
;
4450 /* XXX used to be waiter->prio, not waiter->task->prio */
4451 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
4454 * If nothing changed; bail early.
4456 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
4459 rq
= __task_rq_lock(p
, &rf
);
4460 update_rq_clock(rq
);
4462 * Set under pi_lock && rq->lock, such that the value can be used under
4465 * Note that there is loads of tricky to make this pointer cache work
4466 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4467 * ensure a task is de-boosted (pi_task is set to NULL) before the
4468 * task is allowed to run again (and can exit). This ensures the pointer
4469 * points to a blocked task -- which guaratees the task is present.
4471 p
->pi_top_task
= pi_task
;
4474 * For FIFO/RR we only need to set prio, if that matches we're done.
4476 if (prio
== p
->prio
&& !dl_prio(prio
))
4480 * Idle task boosting is a nono in general. There is one
4481 * exception, when PREEMPT_RT and NOHZ is active:
4483 * The idle task calls get_next_timer_interrupt() and holds
4484 * the timer wheel base->lock on the CPU and another CPU wants
4485 * to access the timer (probably to cancel it). We can safely
4486 * ignore the boosting request, as the idle CPU runs this code
4487 * with interrupts disabled and will complete the lock
4488 * protected section without being interrupted. So there is no
4489 * real need to boost.
4491 if (unlikely(p
== rq
->idle
)) {
4492 WARN_ON(p
!= rq
->curr
);
4493 WARN_ON(p
->pi_blocked_on
);
4497 trace_sched_pi_setprio(p
, pi_task
);
4500 if (oldprio
== prio
)
4501 queue_flag
&= ~DEQUEUE_MOVE
;
4503 prev_class
= p
->sched_class
;
4504 queued
= task_on_rq_queued(p
);
4505 running
= task_current(rq
, p
);
4507 dequeue_task(rq
, p
, queue_flag
);
4509 put_prev_task(rq
, p
);
4512 * Boosting condition are:
4513 * 1. -rt task is running and holds mutex A
4514 * --> -dl task blocks on mutex A
4516 * 2. -dl task is running and holds mutex A
4517 * --> -dl task blocks on mutex A and could preempt the
4520 if (dl_prio(prio
)) {
4521 if (!dl_prio(p
->normal_prio
) ||
4522 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
4523 p
->dl
.dl_boosted
= 1;
4524 queue_flag
|= ENQUEUE_REPLENISH
;
4526 p
->dl
.dl_boosted
= 0;
4527 p
->sched_class
= &dl_sched_class
;
4528 } else if (rt_prio(prio
)) {
4529 if (dl_prio(oldprio
))
4530 p
->dl
.dl_boosted
= 0;
4532 queue_flag
|= ENQUEUE_HEAD
;
4533 p
->sched_class
= &rt_sched_class
;
4535 if (dl_prio(oldprio
))
4536 p
->dl
.dl_boosted
= 0;
4537 if (rt_prio(oldprio
))
4539 p
->sched_class
= &fair_sched_class
;
4545 enqueue_task(rq
, p
, queue_flag
);
4547 set_next_task(rq
, p
);
4549 check_class_changed(rq
, p
, prev_class
, oldprio
);
4551 /* Avoid rq from going away on us: */
4553 __task_rq_unlock(rq
, &rf
);
4555 balance_callback(rq
);
4559 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
4565 void set_user_nice(struct task_struct
*p
, long nice
)
4567 bool queued
, running
;
4572 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
4575 * We have to be careful, if called from sys_setpriority(),
4576 * the task might be in the middle of scheduling on another CPU.
4578 rq
= task_rq_lock(p
, &rf
);
4579 update_rq_clock(rq
);
4582 * The RT priorities are set via sched_setscheduler(), but we still
4583 * allow the 'normal' nice value to be set - but as expected
4584 * it wont have any effect on scheduling until the task is
4585 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4587 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
4588 p
->static_prio
= NICE_TO_PRIO(nice
);
4591 queued
= task_on_rq_queued(p
);
4592 running
= task_current(rq
, p
);
4594 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
4596 put_prev_task(rq
, p
);
4598 p
->static_prio
= NICE_TO_PRIO(nice
);
4599 set_load_weight(p
, true);
4601 p
->prio
= effective_prio(p
);
4604 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
4606 set_next_task(rq
, p
);
4609 * If the task increased its priority or is running and
4610 * lowered its priority, then reschedule its CPU:
4612 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
4615 task_rq_unlock(rq
, p
, &rf
);
4617 EXPORT_SYMBOL(set_user_nice
);
4620 * can_nice - check if a task can reduce its nice value
4624 int can_nice(const struct task_struct
*p
, const int nice
)
4626 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4627 int nice_rlim
= nice_to_rlimit(nice
);
4629 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4630 capable(CAP_SYS_NICE
));
4633 #ifdef __ARCH_WANT_SYS_NICE
4636 * sys_nice - change the priority of the current process.
4637 * @increment: priority increment
4639 * sys_setpriority is a more generic, but much slower function that
4640 * does similar things.
4642 SYSCALL_DEFINE1(nice
, int, increment
)
4647 * Setpriority might change our priority at the same moment.
4648 * We don't have to worry. Conceptually one call occurs first
4649 * and we have a single winner.
4651 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
4652 nice
= task_nice(current
) + increment
;
4654 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
4655 if (increment
< 0 && !can_nice(current
, nice
))
4658 retval
= security_task_setnice(current
, nice
);
4662 set_user_nice(current
, nice
);
4669 * task_prio - return the priority value of a given task.
4670 * @p: the task in question.
4672 * Return: The priority value as seen by users in /proc.
4673 * RT tasks are offset by -200. Normal tasks are centered
4674 * around 0, value goes from -16 to +15.
4676 int task_prio(const struct task_struct
*p
)
4678 return p
->prio
- MAX_RT_PRIO
;
4682 * idle_cpu - is a given CPU idle currently?
4683 * @cpu: the processor in question.
4685 * Return: 1 if the CPU is currently idle. 0 otherwise.
4687 int idle_cpu(int cpu
)
4689 struct rq
*rq
= cpu_rq(cpu
);
4691 if (rq
->curr
!= rq
->idle
)
4698 if (!llist_empty(&rq
->wake_list
))
4706 * available_idle_cpu - is a given CPU idle for enqueuing work.
4707 * @cpu: the CPU in question.
4709 * Return: 1 if the CPU is currently idle. 0 otherwise.
4711 int available_idle_cpu(int cpu
)
4716 if (vcpu_is_preempted(cpu
))
4723 * idle_task - return the idle task for a given CPU.
4724 * @cpu: the processor in question.
4726 * Return: The idle task for the CPU @cpu.
4728 struct task_struct
*idle_task(int cpu
)
4730 return cpu_rq(cpu
)->idle
;
4734 * find_process_by_pid - find a process with a matching PID value.
4735 * @pid: the pid in question.
4737 * The task of @pid, if found. %NULL otherwise.
4739 static struct task_struct
*find_process_by_pid(pid_t pid
)
4741 return pid
? find_task_by_vpid(pid
) : current
;
4745 * sched_setparam() passes in -1 for its policy, to let the functions
4746 * it calls know not to change it.
4748 #define SETPARAM_POLICY -1
4750 static void __setscheduler_params(struct task_struct
*p
,
4751 const struct sched_attr
*attr
)
4753 int policy
= attr
->sched_policy
;
4755 if (policy
== SETPARAM_POLICY
)
4760 if (dl_policy(policy
))
4761 __setparam_dl(p
, attr
);
4762 else if (fair_policy(policy
))
4763 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
4766 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4767 * !rt_policy. Always setting this ensures that things like
4768 * getparam()/getattr() don't report silly values for !rt tasks.
4770 p
->rt_priority
= attr
->sched_priority
;
4771 p
->normal_prio
= normal_prio(p
);
4772 set_load_weight(p
, true);
4775 /* Actually do priority change: must hold pi & rq lock. */
4776 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
4777 const struct sched_attr
*attr
, bool keep_boost
)
4780 * If params can't change scheduling class changes aren't allowed
4783 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
4786 __setscheduler_params(p
, attr
);
4789 * Keep a potential priority boosting if called from
4790 * sched_setscheduler().
4792 p
->prio
= normal_prio(p
);
4794 p
->prio
= rt_effective_prio(p
, p
->prio
);
4796 if (dl_prio(p
->prio
))
4797 p
->sched_class
= &dl_sched_class
;
4798 else if (rt_prio(p
->prio
))
4799 p
->sched_class
= &rt_sched_class
;
4801 p
->sched_class
= &fair_sched_class
;
4805 * Check the target process has a UID that matches the current process's:
4807 static bool check_same_owner(struct task_struct
*p
)
4809 const struct cred
*cred
= current_cred(), *pcred
;
4813 pcred
= __task_cred(p
);
4814 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4815 uid_eq(cred
->euid
, pcred
->uid
));
4820 static int __sched_setscheduler(struct task_struct
*p
,
4821 const struct sched_attr
*attr
,
4824 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4825 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4826 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4827 int new_effective_prio
, policy
= attr
->sched_policy
;
4828 const struct sched_class
*prev_class
;
4831 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
4834 /* The pi code expects interrupts enabled */
4835 BUG_ON(pi
&& in_interrupt());
4837 /* Double check policy once rq lock held: */
4839 reset_on_fork
= p
->sched_reset_on_fork
;
4840 policy
= oldpolicy
= p
->policy
;
4842 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4844 if (!valid_policy(policy
))
4848 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
4852 * Valid priorities for SCHED_FIFO and SCHED_RR are
4853 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4854 * SCHED_BATCH and SCHED_IDLE is 0.
4856 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4857 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4859 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4860 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4864 * Allow unprivileged RT tasks to decrease priority:
4866 if (user
&& !capable(CAP_SYS_NICE
)) {
4867 if (fair_policy(policy
)) {
4868 if (attr
->sched_nice
< task_nice(p
) &&
4869 !can_nice(p
, attr
->sched_nice
))
4873 if (rt_policy(policy
)) {
4874 unsigned long rlim_rtprio
=
4875 task_rlimit(p
, RLIMIT_RTPRIO
);
4877 /* Can't set/change the rt policy: */
4878 if (policy
!= p
->policy
&& !rlim_rtprio
)
4881 /* Can't increase priority: */
4882 if (attr
->sched_priority
> p
->rt_priority
&&
4883 attr
->sched_priority
> rlim_rtprio
)
4888 * Can't set/change SCHED_DEADLINE policy at all for now
4889 * (safest behavior); in the future we would like to allow
4890 * unprivileged DL tasks to increase their relative deadline
4891 * or reduce their runtime (both ways reducing utilization)
4893 if (dl_policy(policy
))
4897 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4898 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4900 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
4901 if (!can_nice(p
, task_nice(p
)))
4905 /* Can't change other user's priorities: */
4906 if (!check_same_owner(p
))
4909 /* Normal users shall not reset the sched_reset_on_fork flag: */
4910 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4915 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
4918 retval
= security_task_setscheduler(p
);
4923 /* Update task specific "requested" clamps */
4924 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
4925 retval
= uclamp_validate(p
, attr
);
4934 * Make sure no PI-waiters arrive (or leave) while we are
4935 * changing the priority of the task:
4937 * To be able to change p->policy safely, the appropriate
4938 * runqueue lock must be held.
4940 rq
= task_rq_lock(p
, &rf
);
4941 update_rq_clock(rq
);
4944 * Changing the policy of the stop threads its a very bad idea:
4946 if (p
== rq
->stop
) {
4952 * If not changing anything there's no need to proceed further,
4953 * but store a possible modification of reset_on_fork.
4955 if (unlikely(policy
== p
->policy
)) {
4956 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4958 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4960 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4962 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
4965 p
->sched_reset_on_fork
= reset_on_fork
;
4972 #ifdef CONFIG_RT_GROUP_SCHED
4974 * Do not allow realtime tasks into groups that have no runtime
4977 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4978 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4979 !task_group_is_autogroup(task_group(p
))) {
4985 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
4986 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
4987 cpumask_t
*span
= rq
->rd
->span
;
4990 * Don't allow tasks with an affinity mask smaller than
4991 * the entire root_domain to become SCHED_DEADLINE. We
4992 * will also fail if there's no bandwidth available.
4994 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
4995 rq
->rd
->dl_bw
.bw
== 0) {
5003 /* Re-check policy now with rq lock held: */
5004 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5005 policy
= oldpolicy
= -1;
5006 task_rq_unlock(rq
, p
, &rf
);
5008 cpuset_read_unlock();
5013 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5014 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5017 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
5022 p
->sched_reset_on_fork
= reset_on_fork
;
5027 * Take priority boosted tasks into account. If the new
5028 * effective priority is unchanged, we just store the new
5029 * normal parameters and do not touch the scheduler class and
5030 * the runqueue. This will be done when the task deboost
5033 new_effective_prio
= rt_effective_prio(p
, newprio
);
5034 if (new_effective_prio
== oldprio
)
5035 queue_flags
&= ~DEQUEUE_MOVE
;
5038 queued
= task_on_rq_queued(p
);
5039 running
= task_current(rq
, p
);
5041 dequeue_task(rq
, p
, queue_flags
);
5043 put_prev_task(rq
, p
);
5045 prev_class
= p
->sched_class
;
5047 __setscheduler(rq
, p
, attr
, pi
);
5048 __setscheduler_uclamp(p
, attr
);
5052 * We enqueue to tail when the priority of a task is
5053 * increased (user space view).
5055 if (oldprio
< p
->prio
)
5056 queue_flags
|= ENQUEUE_HEAD
;
5058 enqueue_task(rq
, p
, queue_flags
);
5061 set_next_task(rq
, p
);
5063 check_class_changed(rq
, p
, prev_class
, oldprio
);
5065 /* Avoid rq from going away on us: */
5067 task_rq_unlock(rq
, p
, &rf
);
5070 cpuset_read_unlock();
5071 rt_mutex_adjust_pi(p
);
5074 /* Run balance callbacks after we've adjusted the PI chain: */
5075 balance_callback(rq
);
5081 task_rq_unlock(rq
, p
, &rf
);
5083 cpuset_read_unlock();
5087 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
5088 const struct sched_param
*param
, bool check
)
5090 struct sched_attr attr
= {
5091 .sched_policy
= policy
,
5092 .sched_priority
= param
->sched_priority
,
5093 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
5096 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5097 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
5098 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5099 policy
&= ~SCHED_RESET_ON_FORK
;
5100 attr
.sched_policy
= policy
;
5103 return __sched_setscheduler(p
, &attr
, check
, true);
5106 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5107 * @p: the task in question.
5108 * @policy: new policy.
5109 * @param: structure containing the new RT priority.
5111 * Return: 0 on success. An error code otherwise.
5113 * NOTE that the task may be already dead.
5115 int sched_setscheduler(struct task_struct
*p
, int policy
,
5116 const struct sched_param
*param
)
5118 return _sched_setscheduler(p
, policy
, param
, true);
5120 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5122 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
5124 return __sched_setscheduler(p
, attr
, true, true);
5126 EXPORT_SYMBOL_GPL(sched_setattr
);
5128 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
5130 return __sched_setscheduler(p
, attr
, false, true);
5134 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5135 * @p: the task in question.
5136 * @policy: new policy.
5137 * @param: structure containing the new RT priority.
5139 * Just like sched_setscheduler, only don't bother checking if the
5140 * current context has permission. For example, this is needed in
5141 * stop_machine(): we create temporary high priority worker threads,
5142 * but our caller might not have that capability.
5144 * Return: 0 on success. An error code otherwise.
5146 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5147 const struct sched_param
*param
)
5149 return _sched_setscheduler(p
, policy
, param
, false);
5151 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
5154 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5156 struct sched_param lparam
;
5157 struct task_struct
*p
;
5160 if (!param
|| pid
< 0)
5162 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5167 p
= find_process_by_pid(pid
);
5173 retval
= sched_setscheduler(p
, policy
, &lparam
);
5181 * Mimics kernel/events/core.c perf_copy_attr().
5183 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
5188 /* Zero the full structure, so that a short copy will be nice: */
5189 memset(attr
, 0, sizeof(*attr
));
5191 ret
= get_user(size
, &uattr
->size
);
5195 /* ABI compatibility quirk: */
5197 size
= SCHED_ATTR_SIZE_VER0
;
5198 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
5201 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
5208 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
5209 size
< SCHED_ATTR_SIZE_VER1
)
5213 * XXX: Do we want to be lenient like existing syscalls; or do we want
5214 * to be strict and return an error on out-of-bounds values?
5216 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
5221 put_user(sizeof(*attr
), &uattr
->size
);
5226 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5227 * @pid: the pid in question.
5228 * @policy: new policy.
5229 * @param: structure containing the new RT priority.
5231 * Return: 0 on success. An error code otherwise.
5233 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
5238 return do_sched_setscheduler(pid
, policy
, param
);
5242 * sys_sched_setparam - set/change the RT priority of a thread
5243 * @pid: the pid in question.
5244 * @param: structure containing the new RT priority.
5246 * Return: 0 on success. An error code otherwise.
5248 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5250 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
5254 * sys_sched_setattr - same as above, but with extended sched_attr
5255 * @pid: the pid in question.
5256 * @uattr: structure containing the extended parameters.
5257 * @flags: for future extension.
5259 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5260 unsigned int, flags
)
5262 struct sched_attr attr
;
5263 struct task_struct
*p
;
5266 if (!uattr
|| pid
< 0 || flags
)
5269 retval
= sched_copy_attr(uattr
, &attr
);
5273 if ((int)attr
.sched_policy
< 0)
5275 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
5276 attr
.sched_policy
= SETPARAM_POLICY
;
5280 p
= find_process_by_pid(pid
);
5286 retval
= sched_setattr(p
, &attr
);
5294 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5295 * @pid: the pid in question.
5297 * Return: On success, the policy of the thread. Otherwise, a negative error
5300 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5302 struct task_struct
*p
;
5310 p
= find_process_by_pid(pid
);
5312 retval
= security_task_getscheduler(p
);
5315 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5322 * sys_sched_getparam - get the RT priority of a thread
5323 * @pid: the pid in question.
5324 * @param: structure containing the RT priority.
5326 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5329 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5331 struct sched_param lp
= { .sched_priority
= 0 };
5332 struct task_struct
*p
;
5335 if (!param
|| pid
< 0)
5339 p
= find_process_by_pid(pid
);
5344 retval
= security_task_getscheduler(p
);
5348 if (task_has_rt_policy(p
))
5349 lp
.sched_priority
= p
->rt_priority
;
5353 * This one might sleep, we cannot do it with a spinlock held ...
5355 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5365 * Copy the kernel size attribute structure (which might be larger
5366 * than what user-space knows about) to user-space.
5368 * Note that all cases are valid: user-space buffer can be larger or
5369 * smaller than the kernel-space buffer. The usual case is that both
5370 * have the same size.
5373 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
5374 struct sched_attr
*kattr
,
5377 unsigned int ksize
= sizeof(*kattr
);
5379 if (!access_ok(uattr
, usize
))
5383 * sched_getattr() ABI forwards and backwards compatibility:
5385 * If usize == ksize then we just copy everything to user-space and all is good.
5387 * If usize < ksize then we only copy as much as user-space has space for,
5388 * this keeps ABI compatibility as well. We skip the rest.
5390 * If usize > ksize then user-space is using a newer version of the ABI,
5391 * which part the kernel doesn't know about. Just ignore it - tooling can
5392 * detect the kernel's knowledge of attributes from the attr->size value
5393 * which is set to ksize in this case.
5395 kattr
->size
= min(usize
, ksize
);
5397 if (copy_to_user(uattr
, kattr
, kattr
->size
))
5404 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5405 * @pid: the pid in question.
5406 * @uattr: structure containing the extended parameters.
5407 * @usize: sizeof(attr) for fwd/bwd comp.
5408 * @flags: for future extension.
5410 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
5411 unsigned int, usize
, unsigned int, flags
)
5413 struct sched_attr kattr
= { };
5414 struct task_struct
*p
;
5417 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
5418 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
5422 p
= find_process_by_pid(pid
);
5427 retval
= security_task_getscheduler(p
);
5431 kattr
.sched_policy
= p
->policy
;
5432 if (p
->sched_reset_on_fork
)
5433 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
5434 if (task_has_dl_policy(p
))
5435 __getparam_dl(p
, &kattr
);
5436 else if (task_has_rt_policy(p
))
5437 kattr
.sched_priority
= p
->rt_priority
;
5439 kattr
.sched_nice
= task_nice(p
);
5441 #ifdef CONFIG_UCLAMP_TASK
5442 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
5443 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
5448 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
5455 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5457 cpumask_var_t cpus_allowed
, new_mask
;
5458 struct task_struct
*p
;
5463 p
= find_process_by_pid(pid
);
5469 /* Prevent p going away */
5473 if (p
->flags
& PF_NO_SETAFFINITY
) {
5477 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5481 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5483 goto out_free_cpus_allowed
;
5486 if (!check_same_owner(p
)) {
5488 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
5490 goto out_free_new_mask
;
5495 retval
= security_task_setscheduler(p
);
5497 goto out_free_new_mask
;
5500 cpuset_cpus_allowed(p
, cpus_allowed
);
5501 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5504 * Since bandwidth control happens on root_domain basis,
5505 * if admission test is enabled, we only admit -deadline
5506 * tasks allowed to run on all the CPUs in the task's
5510 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
5512 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
5515 goto out_free_new_mask
;
5521 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
5524 cpuset_cpus_allowed(p
, cpus_allowed
);
5525 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5527 * We must have raced with a concurrent cpuset
5528 * update. Just reset the cpus_allowed to the
5529 * cpuset's cpus_allowed
5531 cpumask_copy(new_mask
, cpus_allowed
);
5536 free_cpumask_var(new_mask
);
5537 out_free_cpus_allowed
:
5538 free_cpumask_var(cpus_allowed
);
5544 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5545 struct cpumask
*new_mask
)
5547 if (len
< cpumask_size())
5548 cpumask_clear(new_mask
);
5549 else if (len
> cpumask_size())
5550 len
= cpumask_size();
5552 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5556 * sys_sched_setaffinity - set the CPU affinity of a process
5557 * @pid: pid of the process
5558 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5559 * @user_mask_ptr: user-space pointer to the new CPU mask
5561 * Return: 0 on success. An error code otherwise.
5563 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5564 unsigned long __user
*, user_mask_ptr
)
5566 cpumask_var_t new_mask
;
5569 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5572 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5574 retval
= sched_setaffinity(pid
, new_mask
);
5575 free_cpumask_var(new_mask
);
5579 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5581 struct task_struct
*p
;
5582 unsigned long flags
;
5588 p
= find_process_by_pid(pid
);
5592 retval
= security_task_getscheduler(p
);
5596 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5597 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
5598 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5607 * sys_sched_getaffinity - get the CPU affinity of a process
5608 * @pid: pid of the process
5609 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5610 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5612 * Return: size of CPU mask copied to user_mask_ptr on success. An
5613 * error code otherwise.
5615 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5616 unsigned long __user
*, user_mask_ptr
)
5621 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5623 if (len
& (sizeof(unsigned long)-1))
5626 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5629 ret
= sched_getaffinity(pid
, mask
);
5631 unsigned int retlen
= min(len
, cpumask_size());
5633 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5638 free_cpumask_var(mask
);
5644 * sys_sched_yield - yield the current processor to other threads.
5646 * This function yields the current CPU to other tasks. If there are no
5647 * other threads running on this CPU then this function will return.
5651 static void do_sched_yield(void)
5656 rq
= this_rq_lock_irq(&rf
);
5658 schedstat_inc(rq
->yld_count
);
5659 current
->sched_class
->yield_task(rq
);
5662 * Since we are going to call schedule() anyway, there's
5663 * no need to preempt or enable interrupts:
5667 sched_preempt_enable_no_resched();
5672 SYSCALL_DEFINE0(sched_yield
)
5678 #ifndef CONFIG_PREEMPTION
5679 int __sched
_cond_resched(void)
5681 if (should_resched(0)) {
5682 preempt_schedule_common();
5688 EXPORT_SYMBOL(_cond_resched
);
5692 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5693 * call schedule, and on return reacquire the lock.
5695 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5696 * operations here to prevent schedule() from being called twice (once via
5697 * spin_unlock(), once by hand).
5699 int __cond_resched_lock(spinlock_t
*lock
)
5701 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
5704 lockdep_assert_held(lock
);
5706 if (spin_needbreak(lock
) || resched
) {
5709 preempt_schedule_common();
5717 EXPORT_SYMBOL(__cond_resched_lock
);
5720 * yield - yield the current processor to other threads.
5722 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5724 * The scheduler is at all times free to pick the calling task as the most
5725 * eligible task to run, if removing the yield() call from your code breaks
5726 * it, its already broken.
5728 * Typical broken usage is:
5733 * where one assumes that yield() will let 'the other' process run that will
5734 * make event true. If the current task is a SCHED_FIFO task that will never
5735 * happen. Never use yield() as a progress guarantee!!
5737 * If you want to use yield() to wait for something, use wait_event().
5738 * If you want to use yield() to be 'nice' for others, use cond_resched().
5739 * If you still want to use yield(), do not!
5741 void __sched
yield(void)
5743 set_current_state(TASK_RUNNING
);
5746 EXPORT_SYMBOL(yield
);
5749 * yield_to - yield the current processor to another thread in
5750 * your thread group, or accelerate that thread toward the
5751 * processor it's on.
5753 * @preempt: whether task preemption is allowed or not
5755 * It's the caller's job to ensure that the target task struct
5756 * can't go away on us before we can do any checks.
5759 * true (>0) if we indeed boosted the target task.
5760 * false (0) if we failed to boost the target.
5761 * -ESRCH if there's no task to yield to.
5763 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5765 struct task_struct
*curr
= current
;
5766 struct rq
*rq
, *p_rq
;
5767 unsigned long flags
;
5770 local_irq_save(flags
);
5776 * If we're the only runnable task on the rq and target rq also
5777 * has only one task, there's absolutely no point in yielding.
5779 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5784 double_rq_lock(rq
, p_rq
);
5785 if (task_rq(p
) != p_rq
) {
5786 double_rq_unlock(rq
, p_rq
);
5790 if (!curr
->sched_class
->yield_to_task
)
5793 if (curr
->sched_class
!= p
->sched_class
)
5796 if (task_running(p_rq
, p
) || p
->state
)
5799 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5801 schedstat_inc(rq
->yld_count
);
5803 * Make p's CPU reschedule; pick_next_entity takes care of
5806 if (preempt
&& rq
!= p_rq
)
5811 double_rq_unlock(rq
, p_rq
);
5813 local_irq_restore(flags
);
5820 EXPORT_SYMBOL_GPL(yield_to
);
5822 int io_schedule_prepare(void)
5824 int old_iowait
= current
->in_iowait
;
5826 current
->in_iowait
= 1;
5827 blk_schedule_flush_plug(current
);
5832 void io_schedule_finish(int token
)
5834 current
->in_iowait
= token
;
5838 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5839 * that process accounting knows that this is a task in IO wait state.
5841 long __sched
io_schedule_timeout(long timeout
)
5846 token
= io_schedule_prepare();
5847 ret
= schedule_timeout(timeout
);
5848 io_schedule_finish(token
);
5852 EXPORT_SYMBOL(io_schedule_timeout
);
5854 void __sched
io_schedule(void)
5858 token
= io_schedule_prepare();
5860 io_schedule_finish(token
);
5862 EXPORT_SYMBOL(io_schedule
);
5865 * sys_sched_get_priority_max - return maximum RT priority.
5866 * @policy: scheduling class.
5868 * Return: On success, this syscall returns the maximum
5869 * rt_priority that can be used by a given scheduling class.
5870 * On failure, a negative error code is returned.
5872 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5879 ret
= MAX_USER_RT_PRIO
-1;
5881 case SCHED_DEADLINE
:
5892 * sys_sched_get_priority_min - return minimum RT priority.
5893 * @policy: scheduling class.
5895 * Return: On success, this syscall returns the minimum
5896 * rt_priority that can be used by a given scheduling class.
5897 * On failure, a negative error code is returned.
5899 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5908 case SCHED_DEADLINE
:
5917 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
5919 struct task_struct
*p
;
5920 unsigned int time_slice
;
5930 p
= find_process_by_pid(pid
);
5934 retval
= security_task_getscheduler(p
);
5938 rq
= task_rq_lock(p
, &rf
);
5940 if (p
->sched_class
->get_rr_interval
)
5941 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5942 task_rq_unlock(rq
, p
, &rf
);
5945 jiffies_to_timespec64(time_slice
, t
);
5954 * sys_sched_rr_get_interval - return the default timeslice of a process.
5955 * @pid: pid of the process.
5956 * @interval: userspace pointer to the timeslice value.
5958 * this syscall writes the default timeslice value of a given process
5959 * into the user-space timespec buffer. A value of '0' means infinity.
5961 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5964 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5965 struct __kernel_timespec __user
*, interval
)
5967 struct timespec64 t
;
5968 int retval
= sched_rr_get_interval(pid
, &t
);
5971 retval
= put_timespec64(&t
, interval
);
5976 #ifdef CONFIG_COMPAT_32BIT_TIME
5977 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
5978 struct old_timespec32 __user
*, interval
)
5980 struct timespec64 t
;
5981 int retval
= sched_rr_get_interval(pid
, &t
);
5984 retval
= put_old_timespec32(&t
, interval
);
5989 void sched_show_task(struct task_struct
*p
)
5991 unsigned long free
= 0;
5994 if (!try_get_task_stack(p
))
5997 printk(KERN_INFO
"%-15.15s %c", p
->comm
, task_state_to_char(p
));
5999 if (p
->state
== TASK_RUNNING
)
6000 printk(KERN_CONT
" running task ");
6001 #ifdef CONFIG_DEBUG_STACK_USAGE
6002 free
= stack_not_used(p
);
6007 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
6009 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6010 task_pid_nr(p
), ppid
,
6011 (unsigned long)task_thread_info(p
)->flags
);
6013 print_worker_info(KERN_INFO
, p
);
6014 show_stack(p
, NULL
);
6017 EXPORT_SYMBOL_GPL(sched_show_task
);
6020 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
6022 /* no filter, everything matches */
6026 /* filter, but doesn't match */
6027 if (!(p
->state
& state_filter
))
6031 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6034 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
6041 void show_state_filter(unsigned long state_filter
)
6043 struct task_struct
*g
, *p
;
6045 #if BITS_PER_LONG == 32
6047 " task PC stack pid father\n");
6050 " task PC stack pid father\n");
6053 for_each_process_thread(g
, p
) {
6055 * reset the NMI-timeout, listing all files on a slow
6056 * console might take a lot of time:
6057 * Also, reset softlockup watchdogs on all CPUs, because
6058 * another CPU might be blocked waiting for us to process
6061 touch_nmi_watchdog();
6062 touch_all_softlockup_watchdogs();
6063 if (state_filter_match(state_filter
, p
))
6067 #ifdef CONFIG_SCHED_DEBUG
6069 sysrq_sched_debug_show();
6073 * Only show locks if all tasks are dumped:
6076 debug_show_all_locks();
6080 * init_idle - set up an idle thread for a given CPU
6081 * @idle: task in question
6082 * @cpu: CPU the idle task belongs to
6084 * NOTE: this function does not set the idle thread's NEED_RESCHED
6085 * flag, to make booting more robust.
6087 void init_idle(struct task_struct
*idle
, int cpu
)
6089 struct rq
*rq
= cpu_rq(cpu
);
6090 unsigned long flags
;
6092 __sched_fork(0, idle
);
6094 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
6095 raw_spin_lock(&rq
->lock
);
6097 idle
->state
= TASK_RUNNING
;
6098 idle
->se
.exec_start
= sched_clock();
6099 idle
->flags
|= PF_IDLE
;
6101 kasan_unpoison_task_stack(idle
);
6105 * Its possible that init_idle() gets called multiple times on a task,
6106 * in that case do_set_cpus_allowed() will not do the right thing.
6108 * And since this is boot we can forgo the serialization.
6110 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
6113 * We're having a chicken and egg problem, even though we are
6114 * holding rq->lock, the CPU isn't yet set to this CPU so the
6115 * lockdep check in task_group() will fail.
6117 * Similar case to sched_fork(). / Alternatively we could
6118 * use task_rq_lock() here and obtain the other rq->lock.
6123 __set_task_cpu(idle
, cpu
);
6127 rcu_assign_pointer(rq
->curr
, idle
);
6128 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
6132 raw_spin_unlock(&rq
->lock
);
6133 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
6135 /* Set the preempt count _outside_ the spinlocks! */
6136 init_idle_preempt_count(idle
, cpu
);
6139 * The idle tasks have their own, simple scheduling class:
6141 idle
->sched_class
= &idle_sched_class
;
6142 ftrace_graph_init_idle_task(idle
, cpu
);
6143 vtime_init_idle(idle
, cpu
);
6145 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
6151 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
6152 const struct cpumask
*trial
)
6156 if (!cpumask_weight(cur
))
6159 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
6164 int task_can_attach(struct task_struct
*p
,
6165 const struct cpumask
*cs_cpus_allowed
)
6170 * Kthreads which disallow setaffinity shouldn't be moved
6171 * to a new cpuset; we don't want to change their CPU
6172 * affinity and isolating such threads by their set of
6173 * allowed nodes is unnecessary. Thus, cpusets are not
6174 * applicable for such threads. This prevents checking for
6175 * success of set_cpus_allowed_ptr() on all attached tasks
6176 * before cpus_mask may be changed.
6178 if (p
->flags
& PF_NO_SETAFFINITY
) {
6183 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
6185 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
6191 bool sched_smp_initialized __read_mostly
;
6193 #ifdef CONFIG_NUMA_BALANCING
6194 /* Migrate current task p to target_cpu */
6195 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
6197 struct migration_arg arg
= { p
, target_cpu
};
6198 int curr_cpu
= task_cpu(p
);
6200 if (curr_cpu
== target_cpu
)
6203 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
6206 /* TODO: This is not properly updating schedstats */
6208 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
6209 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
6213 * Requeue a task on a given node and accurately track the number of NUMA
6214 * tasks on the runqueues
6216 void sched_setnuma(struct task_struct
*p
, int nid
)
6218 bool queued
, running
;
6222 rq
= task_rq_lock(p
, &rf
);
6223 queued
= task_on_rq_queued(p
);
6224 running
= task_current(rq
, p
);
6227 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
6229 put_prev_task(rq
, p
);
6231 p
->numa_preferred_nid
= nid
;
6234 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
6236 set_next_task(rq
, p
);
6237 task_rq_unlock(rq
, p
, &rf
);
6239 #endif /* CONFIG_NUMA_BALANCING */
6241 #ifdef CONFIG_HOTPLUG_CPU
6243 * Ensure that the idle task is using init_mm right before its CPU goes
6246 void idle_task_exit(void)
6248 struct mm_struct
*mm
= current
->active_mm
;
6250 BUG_ON(cpu_online(smp_processor_id()));
6251 BUG_ON(current
!= this_rq()->idle
);
6253 if (mm
!= &init_mm
) {
6254 switch_mm(mm
, &init_mm
, current
);
6255 finish_arch_post_lock_switch();
6258 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6262 * Since this CPU is going 'away' for a while, fold any nr_active delta
6263 * we might have. Assumes we're called after migrate_tasks() so that the
6264 * nr_active count is stable. We need to take the teardown thread which
6265 * is calling this into account, so we hand in adjust = 1 to the load
6268 * Also see the comment "Global load-average calculations".
6270 static void calc_load_migrate(struct rq
*rq
)
6272 long delta
= calc_load_fold_active(rq
, 1);
6274 atomic_long_add(delta
, &calc_load_tasks
);
6277 static struct task_struct
*__pick_migrate_task(struct rq
*rq
)
6279 const struct sched_class
*class;
6280 struct task_struct
*next
;
6282 for_each_class(class) {
6283 next
= class->pick_next_task(rq
);
6285 next
->sched_class
->put_prev_task(rq
, next
);
6290 /* The idle class should always have a runnable task */
6295 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6296 * try_to_wake_up()->select_task_rq().
6298 * Called with rq->lock held even though we'er in stop_machine() and
6299 * there's no concurrency possible, we hold the required locks anyway
6300 * because of lock validation efforts.
6302 static void migrate_tasks(struct rq
*dead_rq
, struct rq_flags
*rf
)
6304 struct rq
*rq
= dead_rq
;
6305 struct task_struct
*next
, *stop
= rq
->stop
;
6306 struct rq_flags orf
= *rf
;
6310 * Fudge the rq selection such that the below task selection loop
6311 * doesn't get stuck on the currently eligible stop task.
6313 * We're currently inside stop_machine() and the rq is either stuck
6314 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6315 * either way we should never end up calling schedule() until we're
6321 * put_prev_task() and pick_next_task() sched
6322 * class method both need to have an up-to-date
6323 * value of rq->clock[_task]
6325 update_rq_clock(rq
);
6329 * There's this thread running, bail when that's the only
6332 if (rq
->nr_running
== 1)
6335 next
= __pick_migrate_task(rq
);
6338 * Rules for changing task_struct::cpus_mask are holding
6339 * both pi_lock and rq->lock, such that holding either
6340 * stabilizes the mask.
6342 * Drop rq->lock is not quite as disastrous as it usually is
6343 * because !cpu_active at this point, which means load-balance
6344 * will not interfere. Also, stop-machine.
6347 raw_spin_lock(&next
->pi_lock
);
6351 * Since we're inside stop-machine, _nothing_ should have
6352 * changed the task, WARN if weird stuff happened, because in
6353 * that case the above rq->lock drop is a fail too.
6355 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
6356 raw_spin_unlock(&next
->pi_lock
);
6360 /* Find suitable destination for @next, with force if needed. */
6361 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
6362 rq
= __migrate_task(rq
, rf
, next
, dest_cpu
);
6363 if (rq
!= dead_rq
) {
6369 raw_spin_unlock(&next
->pi_lock
);
6374 #endif /* CONFIG_HOTPLUG_CPU */
6376 void set_rq_online(struct rq
*rq
)
6379 const struct sched_class
*class;
6381 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6384 for_each_class(class) {
6385 if (class->rq_online
)
6386 class->rq_online(rq
);
6391 void set_rq_offline(struct rq
*rq
)
6394 const struct sched_class
*class;
6396 for_each_class(class) {
6397 if (class->rq_offline
)
6398 class->rq_offline(rq
);
6401 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6407 * used to mark begin/end of suspend/resume:
6409 static int num_cpus_frozen
;
6412 * Update cpusets according to cpu_active mask. If cpusets are
6413 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6414 * around partition_sched_domains().
6416 * If we come here as part of a suspend/resume, don't touch cpusets because we
6417 * want to restore it back to its original state upon resume anyway.
6419 static void cpuset_cpu_active(void)
6421 if (cpuhp_tasks_frozen
) {
6423 * num_cpus_frozen tracks how many CPUs are involved in suspend
6424 * resume sequence. As long as this is not the last online
6425 * operation in the resume sequence, just build a single sched
6426 * domain, ignoring cpusets.
6428 partition_sched_domains(1, NULL
, NULL
);
6429 if (--num_cpus_frozen
)
6432 * This is the last CPU online operation. So fall through and
6433 * restore the original sched domains by considering the
6434 * cpuset configurations.
6436 cpuset_force_rebuild();
6438 cpuset_update_active_cpus();
6441 static int cpuset_cpu_inactive(unsigned int cpu
)
6443 if (!cpuhp_tasks_frozen
) {
6444 if (dl_cpu_busy(cpu
))
6446 cpuset_update_active_cpus();
6449 partition_sched_domains(1, NULL
, NULL
);
6454 int sched_cpu_activate(unsigned int cpu
)
6456 struct rq
*rq
= cpu_rq(cpu
);
6459 #ifdef CONFIG_SCHED_SMT
6461 * When going up, increment the number of cores with SMT present.
6463 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6464 static_branch_inc_cpuslocked(&sched_smt_present
);
6466 set_cpu_active(cpu
, true);
6468 if (sched_smp_initialized
) {
6469 sched_domains_numa_masks_set(cpu
);
6470 cpuset_cpu_active();
6474 * Put the rq online, if not already. This happens:
6476 * 1) In the early boot process, because we build the real domains
6477 * after all CPUs have been brought up.
6479 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6482 rq_lock_irqsave(rq
, &rf
);
6484 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6487 rq_unlock_irqrestore(rq
, &rf
);
6492 int sched_cpu_deactivate(unsigned int cpu
)
6496 set_cpu_active(cpu
, false);
6498 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6499 * users of this state to go away such that all new such users will
6502 * Do sync before park smpboot threads to take care the rcu boost case.
6506 #ifdef CONFIG_SCHED_SMT
6508 * When going down, decrement the number of cores with SMT present.
6510 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
6511 static_branch_dec_cpuslocked(&sched_smt_present
);
6514 if (!sched_smp_initialized
)
6517 ret
= cpuset_cpu_inactive(cpu
);
6519 set_cpu_active(cpu
, true);
6522 sched_domains_numa_masks_clear(cpu
);
6526 static void sched_rq_cpu_starting(unsigned int cpu
)
6528 struct rq
*rq
= cpu_rq(cpu
);
6530 rq
->calc_load_update
= calc_load_update
;
6531 update_max_interval();
6534 int sched_cpu_starting(unsigned int cpu
)
6536 sched_rq_cpu_starting(cpu
);
6537 sched_tick_start(cpu
);
6541 #ifdef CONFIG_HOTPLUG_CPU
6542 int sched_cpu_dying(unsigned int cpu
)
6544 struct rq
*rq
= cpu_rq(cpu
);
6547 /* Handle pending wakeups and then migrate everything off */
6548 sched_ttwu_pending();
6549 sched_tick_stop(cpu
);
6551 rq_lock_irqsave(rq
, &rf
);
6553 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6556 migrate_tasks(rq
, &rf
);
6557 BUG_ON(rq
->nr_running
!= 1);
6558 rq_unlock_irqrestore(rq
, &rf
);
6560 calc_load_migrate(rq
);
6561 update_max_interval();
6562 nohz_balance_exit_idle(rq
);
6568 void __init
sched_init_smp(void)
6573 * There's no userspace yet to cause hotplug operations; hence all the
6574 * CPU masks are stable and all blatant races in the below code cannot
6577 mutex_lock(&sched_domains_mutex
);
6578 sched_init_domains(cpu_active_mask
);
6579 mutex_unlock(&sched_domains_mutex
);
6581 /* Move init over to a non-isolated CPU */
6582 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
6584 sched_init_granularity();
6586 init_sched_rt_class();
6587 init_sched_dl_class();
6589 sched_smp_initialized
= true;
6592 static int __init
migration_init(void)
6594 sched_cpu_starting(smp_processor_id());
6597 early_initcall(migration_init
);
6600 void __init
sched_init_smp(void)
6602 sched_init_granularity();
6604 #endif /* CONFIG_SMP */
6606 int in_sched_functions(unsigned long addr
)
6608 return in_lock_functions(addr
) ||
6609 (addr
>= (unsigned long)__sched_text_start
6610 && addr
< (unsigned long)__sched_text_end
);
6613 #ifdef CONFIG_CGROUP_SCHED
6615 * Default task group.
6616 * Every task in system belongs to this group at bootup.
6618 struct task_group root_task_group
;
6619 LIST_HEAD(task_groups
);
6621 /* Cacheline aligned slab cache for task_group */
6622 static struct kmem_cache
*task_group_cache __read_mostly
;
6625 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
6626 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
6628 void __init
sched_init(void)
6630 unsigned long ptr
= 0;
6635 #ifdef CONFIG_FAIR_GROUP_SCHED
6636 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6638 #ifdef CONFIG_RT_GROUP_SCHED
6639 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
6642 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
6644 #ifdef CONFIG_FAIR_GROUP_SCHED
6645 root_task_group
.se
= (struct sched_entity
**)ptr
;
6646 ptr
+= nr_cpu_ids
* sizeof(void **);
6648 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
6649 ptr
+= nr_cpu_ids
* sizeof(void **);
6651 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6652 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6653 #endif /* CONFIG_FAIR_GROUP_SCHED */
6654 #ifdef CONFIG_RT_GROUP_SCHED
6655 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
6656 ptr
+= nr_cpu_ids
* sizeof(void **);
6658 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
6659 ptr
+= nr_cpu_ids
* sizeof(void **);
6661 #endif /* CONFIG_RT_GROUP_SCHED */
6663 #ifdef CONFIG_CPUMASK_OFFSTACK
6664 for_each_possible_cpu(i
) {
6665 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6666 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6667 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
6668 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
6670 #endif /* CONFIG_CPUMASK_OFFSTACK */
6672 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
6673 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6676 init_defrootdomain();
6679 #ifdef CONFIG_RT_GROUP_SCHED
6680 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6681 global_rt_period(), global_rt_runtime());
6682 #endif /* CONFIG_RT_GROUP_SCHED */
6684 #ifdef CONFIG_CGROUP_SCHED
6685 task_group_cache
= KMEM_CACHE(task_group
, 0);
6687 list_add(&root_task_group
.list
, &task_groups
);
6688 INIT_LIST_HEAD(&root_task_group
.children
);
6689 INIT_LIST_HEAD(&root_task_group
.siblings
);
6690 autogroup_init(&init_task
);
6691 #endif /* CONFIG_CGROUP_SCHED */
6693 for_each_possible_cpu(i
) {
6697 raw_spin_lock_init(&rq
->lock
);
6699 rq
->calc_load_active
= 0;
6700 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6701 init_cfs_rq(&rq
->cfs
);
6702 init_rt_rq(&rq
->rt
);
6703 init_dl_rq(&rq
->dl
);
6704 #ifdef CONFIG_FAIR_GROUP_SCHED
6705 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6706 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6708 * How much CPU bandwidth does root_task_group get?
6710 * In case of task-groups formed thr' the cgroup filesystem, it
6711 * gets 100% of the CPU resources in the system. This overall
6712 * system CPU resource is divided among the tasks of
6713 * root_task_group and its child task-groups in a fair manner,
6714 * based on each entity's (task or task-group's) weight
6715 * (se->load.weight).
6717 * In other words, if root_task_group has 10 tasks of weight
6718 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6719 * then A0's share of the CPU resource is:
6721 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6723 * We achieve this by letting root_task_group's tasks sit
6724 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6726 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6727 #endif /* CONFIG_FAIR_GROUP_SCHED */
6729 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6730 #ifdef CONFIG_RT_GROUP_SCHED
6731 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6736 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6737 rq
->balance_callback
= NULL
;
6738 rq
->active_balance
= 0;
6739 rq
->next_balance
= jiffies
;
6744 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6745 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6747 rq_csd_init(rq
, &rq
->wake_csd
, wake_csd_func
);
6749 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6751 rq_attach_root(rq
, &def_root_domain
);
6752 #ifdef CONFIG_NO_HZ_COMMON
6753 rq
->last_blocked_load_update_tick
= jiffies
;
6754 atomic_set(&rq
->nohz_flags
, 0);
6756 rq_csd_init(rq
, &rq
->nohz_csd
, nohz_csd_func
);
6758 #endif /* CONFIG_SMP */
6760 atomic_set(&rq
->nr_iowait
, 0);
6763 set_load_weight(&init_task
, false);
6766 * The boot idle thread does lazy MMU switching as well:
6769 enter_lazy_tlb(&init_mm
, current
);
6772 * Make us the idle thread. Technically, schedule() should not be
6773 * called from this thread, however somewhere below it might be,
6774 * but because we are the idle thread, we just pick up running again
6775 * when this runqueue becomes "idle".
6777 init_idle(current
, smp_processor_id());
6779 calc_load_update
= jiffies
+ LOAD_FREQ
;
6782 idle_thread_set_boot_cpu();
6784 init_sched_fair_class();
6792 scheduler_running
= 1;
6795 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6796 static inline int preempt_count_equals(int preempt_offset
)
6798 int nested
= preempt_count() + rcu_preempt_depth();
6800 return (nested
== preempt_offset
);
6803 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6806 * Blocking primitives will set (and therefore destroy) current->state,
6807 * since we will exit with TASK_RUNNING make sure we enter with it,
6808 * otherwise we will destroy state.
6810 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6811 "do not call blocking ops when !TASK_RUNNING; "
6812 "state=%lx set at [<%p>] %pS\n",
6814 (void *)current
->task_state_change
,
6815 (void *)current
->task_state_change
);
6817 ___might_sleep(file
, line
, preempt_offset
);
6819 EXPORT_SYMBOL(__might_sleep
);
6821 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6823 /* Ratelimiting timestamp: */
6824 static unsigned long prev_jiffy
;
6826 unsigned long preempt_disable_ip
;
6828 /* WARN_ON_ONCE() by default, no rate limit required: */
6831 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6832 !is_idle_task(current
) && !current
->non_block_count
) ||
6833 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
6837 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6839 prev_jiffy
= jiffies
;
6841 /* Save this before calling printk(), since that will clobber it: */
6842 preempt_disable_ip
= get_preempt_disable_ip(current
);
6845 "BUG: sleeping function called from invalid context at %s:%d\n",
6848 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6849 in_atomic(), irqs_disabled(), current
->non_block_count
,
6850 current
->pid
, current
->comm
);
6852 if (task_stack_end_corrupted(current
))
6853 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6855 debug_show_held_locks(current
);
6856 if (irqs_disabled())
6857 print_irqtrace_events(current
);
6858 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6859 && !preempt_count_equals(preempt_offset
)) {
6860 pr_err("Preemption disabled at:");
6861 print_ip_sym(preempt_disable_ip
);
6865 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6867 EXPORT_SYMBOL(___might_sleep
);
6869 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
6871 static unsigned long prev_jiffy
;
6873 if (irqs_disabled())
6876 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
6879 if (preempt_count() > preempt_offset
)
6882 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6884 prev_jiffy
= jiffies
;
6886 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
6887 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6888 in_atomic(), irqs_disabled(),
6889 current
->pid
, current
->comm
);
6891 debug_show_held_locks(current
);
6893 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6895 EXPORT_SYMBOL_GPL(__cant_sleep
);
6898 #ifdef CONFIG_MAGIC_SYSRQ
6899 void normalize_rt_tasks(void)
6901 struct task_struct
*g
, *p
;
6902 struct sched_attr attr
= {
6903 .sched_policy
= SCHED_NORMAL
,
6906 read_lock(&tasklist_lock
);
6907 for_each_process_thread(g
, p
) {
6909 * Only normalize user tasks:
6911 if (p
->flags
& PF_KTHREAD
)
6914 p
->se
.exec_start
= 0;
6915 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6916 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6917 schedstat_set(p
->se
.statistics
.block_start
, 0);
6919 if (!dl_task(p
) && !rt_task(p
)) {
6921 * Renice negative nice level userspace
6924 if (task_nice(p
) < 0)
6925 set_user_nice(p
, 0);
6929 __sched_setscheduler(p
, &attr
, false, false);
6931 read_unlock(&tasklist_lock
);
6934 #endif /* CONFIG_MAGIC_SYSRQ */
6936 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6938 * These functions are only useful for the IA64 MCA handling, or kdb.
6940 * They can only be called when the whole system has been
6941 * stopped - every CPU needs to be quiescent, and no scheduling
6942 * activity can take place. Using them for anything else would
6943 * be a serious bug, and as a result, they aren't even visible
6944 * under any other configuration.
6948 * curr_task - return the current task for a given CPU.
6949 * @cpu: the processor in question.
6951 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6953 * Return: The current task for @cpu.
6955 struct task_struct
*curr_task(int cpu
)
6957 return cpu_curr(cpu
);
6960 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6964 * ia64_set_curr_task - set the current task for a given CPU.
6965 * @cpu: the processor in question.
6966 * @p: the task pointer to set.
6968 * Description: This function must only be used when non-maskable interrupts
6969 * are serviced on a separate stack. It allows the architecture to switch the
6970 * notion of the current task on a CPU in a non-blocking manner. This function
6971 * must be called with all CPU's synchronized, and interrupts disabled, the
6972 * and caller must save the original value of the current task (see
6973 * curr_task() above) and restore that value before reenabling interrupts and
6974 * re-starting the system.
6976 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6978 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6985 #ifdef CONFIG_CGROUP_SCHED
6986 /* task_group_lock serializes the addition/removal of task groups */
6987 static DEFINE_SPINLOCK(task_group_lock
);
6989 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
6990 struct task_group
*parent
)
6992 #ifdef CONFIG_UCLAMP_TASK_GROUP
6993 enum uclamp_id clamp_id
;
6995 for_each_clamp_id(clamp_id
) {
6996 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
6997 uclamp_none(clamp_id
), false);
6998 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
7003 static void sched_free_group(struct task_group
*tg
)
7005 free_fair_sched_group(tg
);
7006 free_rt_sched_group(tg
);
7008 kmem_cache_free(task_group_cache
, tg
);
7011 /* allocate runqueue etc for a new task group */
7012 struct task_group
*sched_create_group(struct task_group
*parent
)
7014 struct task_group
*tg
;
7016 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
7018 return ERR_PTR(-ENOMEM
);
7020 if (!alloc_fair_sched_group(tg
, parent
))
7023 if (!alloc_rt_sched_group(tg
, parent
))
7026 alloc_uclamp_sched_group(tg
, parent
);
7031 sched_free_group(tg
);
7032 return ERR_PTR(-ENOMEM
);
7035 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
7037 unsigned long flags
;
7039 spin_lock_irqsave(&task_group_lock
, flags
);
7040 list_add_rcu(&tg
->list
, &task_groups
);
7042 /* Root should already exist: */
7045 tg
->parent
= parent
;
7046 INIT_LIST_HEAD(&tg
->children
);
7047 list_add_rcu(&tg
->siblings
, &parent
->children
);
7048 spin_unlock_irqrestore(&task_group_lock
, flags
);
7050 online_fair_sched_group(tg
);
7053 /* rcu callback to free various structures associated with a task group */
7054 static void sched_free_group_rcu(struct rcu_head
*rhp
)
7056 /* Now it should be safe to free those cfs_rqs: */
7057 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
7060 void sched_destroy_group(struct task_group
*tg
)
7062 /* Wait for possible concurrent references to cfs_rqs complete: */
7063 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
7066 void sched_offline_group(struct task_group
*tg
)
7068 unsigned long flags
;
7070 /* End participation in shares distribution: */
7071 unregister_fair_sched_group(tg
);
7073 spin_lock_irqsave(&task_group_lock
, flags
);
7074 list_del_rcu(&tg
->list
);
7075 list_del_rcu(&tg
->siblings
);
7076 spin_unlock_irqrestore(&task_group_lock
, flags
);
7079 static void sched_change_group(struct task_struct
*tsk
, int type
)
7081 struct task_group
*tg
;
7084 * All callers are synchronized by task_rq_lock(); we do not use RCU
7085 * which is pointless here. Thus, we pass "true" to task_css_check()
7086 * to prevent lockdep warnings.
7088 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
7089 struct task_group
, css
);
7090 tg
= autogroup_task_group(tsk
, tg
);
7091 tsk
->sched_task_group
= tg
;
7093 #ifdef CONFIG_FAIR_GROUP_SCHED
7094 if (tsk
->sched_class
->task_change_group
)
7095 tsk
->sched_class
->task_change_group(tsk
, type
);
7098 set_task_rq(tsk
, task_cpu(tsk
));
7102 * Change task's runqueue when it moves between groups.
7104 * The caller of this function should have put the task in its new group by
7105 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7108 void sched_move_task(struct task_struct
*tsk
)
7110 int queued
, running
, queue_flags
=
7111 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
7115 rq
= task_rq_lock(tsk
, &rf
);
7116 update_rq_clock(rq
);
7118 running
= task_current(rq
, tsk
);
7119 queued
= task_on_rq_queued(tsk
);
7122 dequeue_task(rq
, tsk
, queue_flags
);
7124 put_prev_task(rq
, tsk
);
7126 sched_change_group(tsk
, TASK_MOVE_GROUP
);
7129 enqueue_task(rq
, tsk
, queue_flags
);
7131 set_next_task(rq
, tsk
);
7133 * After changing group, the running task may have joined a
7134 * throttled one but it's still the running task. Trigger a
7135 * resched to make sure that task can still run.
7140 task_rq_unlock(rq
, tsk
, &rf
);
7143 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
7145 return css
? container_of(css
, struct task_group
, css
) : NULL
;
7148 static struct cgroup_subsys_state
*
7149 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
7151 struct task_group
*parent
= css_tg(parent_css
);
7152 struct task_group
*tg
;
7155 /* This is early initialization for the top cgroup */
7156 return &root_task_group
.css
;
7159 tg
= sched_create_group(parent
);
7161 return ERR_PTR(-ENOMEM
);
7166 /* Expose task group only after completing cgroup initialization */
7167 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
7169 struct task_group
*tg
= css_tg(css
);
7170 struct task_group
*parent
= css_tg(css
->parent
);
7173 sched_online_group(tg
, parent
);
7175 #ifdef CONFIG_UCLAMP_TASK_GROUP
7176 /* Propagate the effective uclamp value for the new group */
7177 cpu_util_update_eff(css
);
7183 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
7185 struct task_group
*tg
= css_tg(css
);
7187 sched_offline_group(tg
);
7190 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
7192 struct task_group
*tg
= css_tg(css
);
7195 * Relies on the RCU grace period between css_released() and this.
7197 sched_free_group(tg
);
7201 * This is called before wake_up_new_task(), therefore we really only
7202 * have to set its group bits, all the other stuff does not apply.
7204 static void cpu_cgroup_fork(struct task_struct
*task
)
7209 rq
= task_rq_lock(task
, &rf
);
7211 update_rq_clock(rq
);
7212 sched_change_group(task
, TASK_SET_GROUP
);
7214 task_rq_unlock(rq
, task
, &rf
);
7217 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
7219 struct task_struct
*task
;
7220 struct cgroup_subsys_state
*css
;
7223 cgroup_taskset_for_each(task
, css
, tset
) {
7224 #ifdef CONFIG_RT_GROUP_SCHED
7225 if (!sched_rt_can_attach(css_tg(css
), task
))
7229 * Serialize against wake_up_new_task() such that if its
7230 * running, we're sure to observe its full state.
7232 raw_spin_lock_irq(&task
->pi_lock
);
7234 * Avoid calling sched_move_task() before wake_up_new_task()
7235 * has happened. This would lead to problems with PELT, due to
7236 * move wanting to detach+attach while we're not attached yet.
7238 if (task
->state
== TASK_NEW
)
7240 raw_spin_unlock_irq(&task
->pi_lock
);
7248 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
7250 struct task_struct
*task
;
7251 struct cgroup_subsys_state
*css
;
7253 cgroup_taskset_for_each(task
, css
, tset
)
7254 sched_move_task(task
);
7257 #ifdef CONFIG_UCLAMP_TASK_GROUP
7258 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
7260 struct cgroup_subsys_state
*top_css
= css
;
7261 struct uclamp_se
*uc_parent
= NULL
;
7262 struct uclamp_se
*uc_se
= NULL
;
7263 unsigned int eff
[UCLAMP_CNT
];
7264 enum uclamp_id clamp_id
;
7265 unsigned int clamps
;
7267 css_for_each_descendant_pre(css
, top_css
) {
7268 uc_parent
= css_tg(css
)->parent
7269 ? css_tg(css
)->parent
->uclamp
: NULL
;
7271 for_each_clamp_id(clamp_id
) {
7272 /* Assume effective clamps matches requested clamps */
7273 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
7274 /* Cap effective clamps with parent's effective clamps */
7276 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
7277 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
7280 /* Ensure protection is always capped by limit */
7281 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
7283 /* Propagate most restrictive effective clamps */
7285 uc_se
= css_tg(css
)->uclamp
;
7286 for_each_clamp_id(clamp_id
) {
7287 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
7289 uc_se
[clamp_id
].value
= eff
[clamp_id
];
7290 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
7291 clamps
|= (0x1 << clamp_id
);
7294 css
= css_rightmost_descendant(css
);
7298 /* Immediately update descendants RUNNABLE tasks */
7299 uclamp_update_active_tasks(css
, clamps
);
7304 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7305 * C expression. Since there is no way to convert a macro argument (N) into a
7306 * character constant, use two levels of macros.
7308 #define _POW10(exp) ((unsigned int)1e##exp)
7309 #define POW10(exp) _POW10(exp)
7311 struct uclamp_request
{
7312 #define UCLAMP_PERCENT_SHIFT 2
7313 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7319 static inline struct uclamp_request
7320 capacity_from_percent(char *buf
)
7322 struct uclamp_request req
= {
7323 .percent
= UCLAMP_PERCENT_SCALE
,
7324 .util
= SCHED_CAPACITY_SCALE
,
7329 if (strcmp(buf
, "max")) {
7330 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
7334 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
7339 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
7340 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
7346 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
7347 size_t nbytes
, loff_t off
,
7348 enum uclamp_id clamp_id
)
7350 struct uclamp_request req
;
7351 struct task_group
*tg
;
7353 req
= capacity_from_percent(buf
);
7357 mutex_lock(&uclamp_mutex
);
7360 tg
= css_tg(of_css(of
));
7361 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
7362 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
7365 * Because of not recoverable conversion rounding we keep track of the
7366 * exact requested value
7368 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
7370 /* Update effective clamps to track the most restrictive value */
7371 cpu_util_update_eff(of_css(of
));
7374 mutex_unlock(&uclamp_mutex
);
7379 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
7380 char *buf
, size_t nbytes
,
7383 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
7386 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
7387 char *buf
, size_t nbytes
,
7390 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
7393 static inline void cpu_uclamp_print(struct seq_file
*sf
,
7394 enum uclamp_id clamp_id
)
7396 struct task_group
*tg
;
7402 tg
= css_tg(seq_css(sf
));
7403 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
7406 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
7407 seq_puts(sf
, "max\n");
7411 percent
= tg
->uclamp_pct
[clamp_id
];
7412 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
7413 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
7416 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
7418 cpu_uclamp_print(sf
, UCLAMP_MIN
);
7422 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
7424 cpu_uclamp_print(sf
, UCLAMP_MAX
);
7427 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7429 #ifdef CONFIG_FAIR_GROUP_SCHED
7430 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
7431 struct cftype
*cftype
, u64 shareval
)
7433 if (shareval
> scale_load_down(ULONG_MAX
))
7434 shareval
= MAX_SHARES
;
7435 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
7438 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
7441 struct task_group
*tg
= css_tg(css
);
7443 return (u64
) scale_load_down(tg
->shares
);
7446 #ifdef CONFIG_CFS_BANDWIDTH
7447 static DEFINE_MUTEX(cfs_constraints_mutex
);
7449 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
7450 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
7451 /* More than 203 days if BW_SHIFT equals 20. */
7452 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
7454 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
7456 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
7458 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
7459 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7461 if (tg
== &root_task_group
)
7465 * Ensure we have at some amount of bandwidth every period. This is
7466 * to prevent reaching a state of large arrears when throttled via
7467 * entity_tick() resulting in prolonged exit starvation.
7469 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
7473 * Likewise, bound things on the otherside by preventing insane quota
7474 * periods. This also allows us to normalize in computing quota
7477 if (period
> max_cfs_quota_period
)
7481 * Bound quota to defend quota against overflow during bandwidth shift.
7483 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
7487 * Prevent race between setting of cfs_rq->runtime_enabled and
7488 * unthrottle_offline_cfs_rqs().
7491 mutex_lock(&cfs_constraints_mutex
);
7492 ret
= __cfs_schedulable(tg
, period
, quota
);
7496 runtime_enabled
= quota
!= RUNTIME_INF
;
7497 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
7499 * If we need to toggle cfs_bandwidth_used, off->on must occur
7500 * before making related changes, and on->off must occur afterwards
7502 if (runtime_enabled
&& !runtime_was_enabled
)
7503 cfs_bandwidth_usage_inc();
7504 raw_spin_lock_irq(&cfs_b
->lock
);
7505 cfs_b
->period
= ns_to_ktime(period
);
7506 cfs_b
->quota
= quota
;
7508 __refill_cfs_bandwidth_runtime(cfs_b
);
7510 /* Restart the period timer (if active) to handle new period expiry: */
7511 if (runtime_enabled
)
7512 start_cfs_bandwidth(cfs_b
);
7514 raw_spin_unlock_irq(&cfs_b
->lock
);
7516 for_each_online_cpu(i
) {
7517 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
7518 struct rq
*rq
= cfs_rq
->rq
;
7521 rq_lock_irq(rq
, &rf
);
7522 cfs_rq
->runtime_enabled
= runtime_enabled
;
7523 cfs_rq
->runtime_remaining
= 0;
7525 if (cfs_rq
->throttled
)
7526 unthrottle_cfs_rq(cfs_rq
);
7527 rq_unlock_irq(rq
, &rf
);
7529 if (runtime_was_enabled
&& !runtime_enabled
)
7530 cfs_bandwidth_usage_dec();
7532 mutex_unlock(&cfs_constraints_mutex
);
7538 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7542 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7543 if (cfs_quota_us
< 0)
7544 quota
= RUNTIME_INF
;
7545 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
7546 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7550 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7553 static long tg_get_cfs_quota(struct task_group
*tg
)
7557 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7560 quota_us
= tg
->cfs_bandwidth
.quota
;
7561 do_div(quota_us
, NSEC_PER_USEC
);
7566 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7570 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
7573 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7574 quota
= tg
->cfs_bandwidth
.quota
;
7576 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7579 static long tg_get_cfs_period(struct task_group
*tg
)
7583 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7584 do_div(cfs_period_us
, NSEC_PER_USEC
);
7586 return cfs_period_us
;
7589 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7592 return tg_get_cfs_quota(css_tg(css
));
7595 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7596 struct cftype
*cftype
, s64 cfs_quota_us
)
7598 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7601 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7604 return tg_get_cfs_period(css_tg(css
));
7607 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7608 struct cftype
*cftype
, u64 cfs_period_us
)
7610 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7613 struct cfs_schedulable_data
{
7614 struct task_group
*tg
;
7619 * normalize group quota/period to be quota/max_period
7620 * note: units are usecs
7622 static u64
normalize_cfs_quota(struct task_group
*tg
,
7623 struct cfs_schedulable_data
*d
)
7631 period
= tg_get_cfs_period(tg
);
7632 quota
= tg_get_cfs_quota(tg
);
7635 /* note: these should typically be equivalent */
7636 if (quota
== RUNTIME_INF
|| quota
== -1)
7639 return to_ratio(period
, quota
);
7642 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7644 struct cfs_schedulable_data
*d
= data
;
7645 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7646 s64 quota
= 0, parent_quota
= -1;
7649 quota
= RUNTIME_INF
;
7651 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7653 quota
= normalize_cfs_quota(tg
, d
);
7654 parent_quota
= parent_b
->hierarchical_quota
;
7657 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7658 * always take the min. On cgroup1, only inherit when no
7661 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
7662 quota
= min(quota
, parent_quota
);
7664 if (quota
== RUNTIME_INF
)
7665 quota
= parent_quota
;
7666 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7670 cfs_b
->hierarchical_quota
= quota
;
7675 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7678 struct cfs_schedulable_data data
= {
7684 if (quota
!= RUNTIME_INF
) {
7685 do_div(data
.period
, NSEC_PER_USEC
);
7686 do_div(data
.quota
, NSEC_PER_USEC
);
7690 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7696 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
7698 struct task_group
*tg
= css_tg(seq_css(sf
));
7699 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7701 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7702 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7703 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7705 if (schedstat_enabled() && tg
!= &root_task_group
) {
7709 for_each_possible_cpu(i
)
7710 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
7712 seq_printf(sf
, "wait_sum %llu\n", ws
);
7717 #endif /* CONFIG_CFS_BANDWIDTH */
7718 #endif /* CONFIG_FAIR_GROUP_SCHED */
7720 #ifdef CONFIG_RT_GROUP_SCHED
7721 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7722 struct cftype
*cft
, s64 val
)
7724 return sched_group_set_rt_runtime(css_tg(css
), val
);
7727 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7730 return sched_group_rt_runtime(css_tg(css
));
7733 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7734 struct cftype
*cftype
, u64 rt_period_us
)
7736 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7739 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7742 return sched_group_rt_period(css_tg(css
));
7744 #endif /* CONFIG_RT_GROUP_SCHED */
7746 static struct cftype cpu_legacy_files
[] = {
7747 #ifdef CONFIG_FAIR_GROUP_SCHED
7750 .read_u64
= cpu_shares_read_u64
,
7751 .write_u64
= cpu_shares_write_u64
,
7754 #ifdef CONFIG_CFS_BANDWIDTH
7756 .name
= "cfs_quota_us",
7757 .read_s64
= cpu_cfs_quota_read_s64
,
7758 .write_s64
= cpu_cfs_quota_write_s64
,
7761 .name
= "cfs_period_us",
7762 .read_u64
= cpu_cfs_period_read_u64
,
7763 .write_u64
= cpu_cfs_period_write_u64
,
7767 .seq_show
= cpu_cfs_stat_show
,
7770 #ifdef CONFIG_RT_GROUP_SCHED
7772 .name
= "rt_runtime_us",
7773 .read_s64
= cpu_rt_runtime_read
,
7774 .write_s64
= cpu_rt_runtime_write
,
7777 .name
= "rt_period_us",
7778 .read_u64
= cpu_rt_period_read_uint
,
7779 .write_u64
= cpu_rt_period_write_uint
,
7782 #ifdef CONFIG_UCLAMP_TASK_GROUP
7784 .name
= "uclamp.min",
7785 .flags
= CFTYPE_NOT_ON_ROOT
,
7786 .seq_show
= cpu_uclamp_min_show
,
7787 .write
= cpu_uclamp_min_write
,
7790 .name
= "uclamp.max",
7791 .flags
= CFTYPE_NOT_ON_ROOT
,
7792 .seq_show
= cpu_uclamp_max_show
,
7793 .write
= cpu_uclamp_max_write
,
7799 static int cpu_extra_stat_show(struct seq_file
*sf
,
7800 struct cgroup_subsys_state
*css
)
7802 #ifdef CONFIG_CFS_BANDWIDTH
7804 struct task_group
*tg
= css_tg(css
);
7805 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7808 throttled_usec
= cfs_b
->throttled_time
;
7809 do_div(throttled_usec
, NSEC_PER_USEC
);
7811 seq_printf(sf
, "nr_periods %d\n"
7813 "throttled_usec %llu\n",
7814 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
7821 #ifdef CONFIG_FAIR_GROUP_SCHED
7822 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
7825 struct task_group
*tg
= css_tg(css
);
7826 u64 weight
= scale_load_down(tg
->shares
);
7828 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
7831 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
7832 struct cftype
*cft
, u64 weight
)
7835 * cgroup weight knobs should use the common MIN, DFL and MAX
7836 * values which are 1, 100 and 10000 respectively. While it loses
7837 * a bit of range on both ends, it maps pretty well onto the shares
7838 * value used by scheduler and the round-trip conversions preserve
7839 * the original value over the entire range.
7841 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
7844 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
7846 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7849 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
7852 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
7853 int last_delta
= INT_MAX
;
7856 /* find the closest nice value to the current weight */
7857 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
7858 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
7859 if (delta
>= last_delta
)
7864 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
7867 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
7868 struct cftype
*cft
, s64 nice
)
7870 unsigned long weight
;
7873 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
7876 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
7877 idx
= array_index_nospec(idx
, 40);
7878 weight
= sched_prio_to_weight
[idx
];
7880 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
7884 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
7885 long period
, long quota
)
7888 seq_puts(sf
, "max");
7890 seq_printf(sf
, "%ld", quota
);
7892 seq_printf(sf
, " %ld\n", period
);
7895 /* caller should put the current value in *@periodp before calling */
7896 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
7897 u64
*periodp
, u64
*quotap
)
7899 char tok
[21]; /* U64_MAX */
7901 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
7904 *periodp
*= NSEC_PER_USEC
;
7906 if (sscanf(tok
, "%llu", quotap
))
7907 *quotap
*= NSEC_PER_USEC
;
7908 else if (!strcmp(tok
, "max"))
7909 *quotap
= RUNTIME_INF
;
7916 #ifdef CONFIG_CFS_BANDWIDTH
7917 static int cpu_max_show(struct seq_file
*sf
, void *v
)
7919 struct task_group
*tg
= css_tg(seq_css(sf
));
7921 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
7925 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
7926 char *buf
, size_t nbytes
, loff_t off
)
7928 struct task_group
*tg
= css_tg(of_css(of
));
7929 u64 period
= tg_get_cfs_period(tg
);
7933 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
7935 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
7936 return ret
?: nbytes
;
7940 static struct cftype cpu_files
[] = {
7941 #ifdef CONFIG_FAIR_GROUP_SCHED
7944 .flags
= CFTYPE_NOT_ON_ROOT
,
7945 .read_u64
= cpu_weight_read_u64
,
7946 .write_u64
= cpu_weight_write_u64
,
7949 .name
= "weight.nice",
7950 .flags
= CFTYPE_NOT_ON_ROOT
,
7951 .read_s64
= cpu_weight_nice_read_s64
,
7952 .write_s64
= cpu_weight_nice_write_s64
,
7955 #ifdef CONFIG_CFS_BANDWIDTH
7958 .flags
= CFTYPE_NOT_ON_ROOT
,
7959 .seq_show
= cpu_max_show
,
7960 .write
= cpu_max_write
,
7963 #ifdef CONFIG_UCLAMP_TASK_GROUP
7965 .name
= "uclamp.min",
7966 .flags
= CFTYPE_NOT_ON_ROOT
,
7967 .seq_show
= cpu_uclamp_min_show
,
7968 .write
= cpu_uclamp_min_write
,
7971 .name
= "uclamp.max",
7972 .flags
= CFTYPE_NOT_ON_ROOT
,
7973 .seq_show
= cpu_uclamp_max_show
,
7974 .write
= cpu_uclamp_max_write
,
7980 struct cgroup_subsys cpu_cgrp_subsys
= {
7981 .css_alloc
= cpu_cgroup_css_alloc
,
7982 .css_online
= cpu_cgroup_css_online
,
7983 .css_released
= cpu_cgroup_css_released
,
7984 .css_free
= cpu_cgroup_css_free
,
7985 .css_extra_stat_show
= cpu_extra_stat_show
,
7986 .fork
= cpu_cgroup_fork
,
7987 .can_attach
= cpu_cgroup_can_attach
,
7988 .attach
= cpu_cgroup_attach
,
7989 .legacy_cftypes
= cpu_legacy_files
,
7990 .dfl_cftypes
= cpu_files
,
7995 #endif /* CONFIG_CGROUP_SCHED */
7997 void dump_cpu_task(int cpu
)
7999 pr_info("Task dump for CPU %d:\n", cpu
);
8000 sched_show_task(cpu_curr(cpu
));
8004 * Nice levels are multiplicative, with a gentle 10% change for every
8005 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8006 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8007 * that remained on nice 0.
8009 * The "10% effect" is relative and cumulative: from _any_ nice level,
8010 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8011 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8012 * If a task goes up by ~10% and another task goes down by ~10% then
8013 * the relative distance between them is ~25%.)
8015 const int sched_prio_to_weight
[40] = {
8016 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8017 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8018 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8019 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8020 /* 0 */ 1024, 820, 655, 526, 423,
8021 /* 5 */ 335, 272, 215, 172, 137,
8022 /* 10 */ 110, 87, 70, 56, 45,
8023 /* 15 */ 36, 29, 23, 18, 15,
8027 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8029 * In cases where the weight does not change often, we can use the
8030 * precalculated inverse to speed up arithmetics by turning divisions
8031 * into multiplications:
8033 const u32 sched_prio_to_wmult
[40] = {
8034 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8035 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8036 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8037 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8038 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8039 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8040 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8041 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8044 #undef CREATE_TRACE_POINTS