1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp
);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp
);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp
);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp
);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp
);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp
);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp
);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp
);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp
);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp
);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug
unsigned int sysctl_sched_features
=
63 * Print a warning if need_resched is set for the given duration (if
64 * LATENCY_WARN is enabled).
66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
69 __read_mostly
int sysctl_resched_latency_warn_ms
= 100;
70 __read_mostly
int sysctl_resched_latency_warn_once
= 1;
71 #endif /* CONFIG_SCHED_DEBUG */
74 * Number of tasks to iterate in a single balance run.
75 * Limited because this is done with IRQs disabled.
77 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
80 * period over which we measure -rt task CPU usage in us.
83 unsigned int sysctl_sched_rt_period
= 1000000;
85 __read_mostly
int scheduler_running
;
88 * part of the period that we allow rt tasks to run in us.
91 int sysctl_sched_rt_runtime
= 950000;
95 * Serialization rules:
101 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
104 * rq2->lock where: rq1 < rq2
108 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
109 * local CPU's rq->lock, it optionally removes the task from the runqueue and
110 * always looks at the local rq data structures to find the most eligible task
113 * Task enqueue is also under rq->lock, possibly taken from another CPU.
114 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
115 * the local CPU to avoid bouncing the runqueue state around [ see
116 * ttwu_queue_wakelist() ]
118 * Task wakeup, specifically wakeups that involve migration, are horribly
119 * complicated to avoid having to take two rq->locks.
123 * System-calls and anything external will use task_rq_lock() which acquires
124 * both p->pi_lock and rq->lock. As a consequence the state they change is
125 * stable while holding either lock:
127 * - sched_setaffinity()/
128 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
129 * - set_user_nice(): p->se.load, p->*prio
130 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
131 * p->se.load, p->rt_priority,
132 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
133 * - sched_setnuma(): p->numa_preferred_nid
134 * - sched_move_task()/
135 * cpu_cgroup_fork(): p->sched_task_group
136 * - uclamp_update_active() p->uclamp*
138 * p->state <- TASK_*:
140 * is changed locklessly using set_current_state(), __set_current_state() or
141 * set_special_state(), see their respective comments, or by
142 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
145 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
147 * is set by activate_task() and cleared by deactivate_task(), under
148 * rq->lock. Non-zero indicates the task is runnable, the special
149 * ON_RQ_MIGRATING state is used for migration without holding both
150 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
152 * p->on_cpu <- { 0, 1 }:
154 * is set by prepare_task() and cleared by finish_task() such that it will be
155 * set before p is scheduled-in and cleared after p is scheduled-out, both
156 * under rq->lock. Non-zero indicates the task is running on its CPU.
158 * [ The astute reader will observe that it is possible for two tasks on one
159 * CPU to have ->on_cpu = 1 at the same time. ]
161 * task_cpu(p): is changed by set_task_cpu(), the rules are:
163 * - Don't call set_task_cpu() on a blocked task:
165 * We don't care what CPU we're not running on, this simplifies hotplug,
166 * the CPU assignment of blocked tasks isn't required to be valid.
168 * - for try_to_wake_up(), called under p->pi_lock:
170 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
172 * - for migration called under rq->lock:
173 * [ see task_on_rq_migrating() in task_rq_lock() ]
175 * o move_queued_task()
178 * - for migration called under double_rq_lock():
180 * o __migrate_swap_task()
181 * o push_rt_task() / pull_rt_task()
182 * o push_dl_task() / pull_dl_task()
183 * o dl_task_offline_migration()
188 * __task_rq_lock - lock the rq @p resides on.
190 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
195 lockdep_assert_held(&p
->pi_lock
);
199 raw_spin_lock(&rq
->lock
);
200 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
204 raw_spin_unlock(&rq
->lock
);
206 while (unlikely(task_on_rq_migrating(p
)))
212 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
214 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
215 __acquires(p
->pi_lock
)
221 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
223 raw_spin_lock(&rq
->lock
);
225 * move_queued_task() task_rq_lock()
228 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
229 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
230 * [S] ->cpu = new_cpu [L] task_rq()
234 * If we observe the old CPU in task_rq_lock(), the acquire of
235 * the old rq->lock will fully serialize against the stores.
237 * If we observe the new CPU in task_rq_lock(), the address
238 * dependency headed by '[L] rq = task_rq()' and the acquire
239 * will pair with the WMB to ensure we then also see migrating.
241 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
245 raw_spin_unlock(&rq
->lock
);
246 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
248 while (unlikely(task_on_rq_migrating(p
)))
254 * RQ-clock updating methods:
257 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
260 * In theory, the compile should just see 0 here, and optimize out the call
261 * to sched_rt_avg_update. But I don't trust it...
263 s64 __maybe_unused steal
= 0, irq_delta
= 0;
265 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
266 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
269 * Since irq_time is only updated on {soft,}irq_exit, we might run into
270 * this case when a previous update_rq_clock() happened inside a
273 * When this happens, we stop ->clock_task and only update the
274 * prev_irq_time stamp to account for the part that fit, so that a next
275 * update will consume the rest. This ensures ->clock_task is
278 * It does however cause some slight miss-attribution of {soft,}irq
279 * time, a more accurate solution would be to update the irq_time using
280 * the current rq->clock timestamp, except that would require using
283 if (irq_delta
> delta
)
286 rq
->prev_irq_time
+= irq_delta
;
289 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
290 if (static_key_false((¶virt_steal_rq_enabled
))) {
291 steal
= paravirt_steal_clock(cpu_of(rq
));
292 steal
-= rq
->prev_steal_time_rq
;
294 if (unlikely(steal
> delta
))
297 rq
->prev_steal_time_rq
+= steal
;
302 rq
->clock_task
+= delta
;
304 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
305 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
306 update_irq_load_avg(rq
, irq_delta
+ steal
);
308 update_rq_clock_pelt(rq
, delta
);
311 void update_rq_clock(struct rq
*rq
)
315 lockdep_assert_held(&rq
->lock
);
317 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
320 #ifdef CONFIG_SCHED_DEBUG
321 if (sched_feat(WARN_DOUBLE_CLOCK
))
322 SCHED_WARN_ON(rq
->clock_update_flags
& RQCF_UPDATED
);
323 rq
->clock_update_flags
|= RQCF_UPDATED
;
326 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
330 update_rq_clock_task(rq
, delta
);
333 #ifdef CONFIG_SCHED_HRTICK
335 * Use HR-timers to deliver accurate preemption points.
338 static void hrtick_clear(struct rq
*rq
)
340 if (hrtimer_active(&rq
->hrtick_timer
))
341 hrtimer_cancel(&rq
->hrtick_timer
);
345 * High-resolution timer tick.
346 * Runs from hardirq context with interrupts disabled.
348 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
350 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
353 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
357 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
360 return HRTIMER_NORESTART
;
365 static void __hrtick_restart(struct rq
*rq
)
367 struct hrtimer
*timer
= &rq
->hrtick_timer
;
368 ktime_t time
= rq
->hrtick_time
;
370 hrtimer_start(timer
, time
, HRTIMER_MODE_ABS_PINNED_HARD
);
374 * called from hardirq (IPI) context
376 static void __hrtick_start(void *arg
)
382 __hrtick_restart(rq
);
387 * Called to set the hrtick timer state.
389 * called with rq->lock held and irqs disabled
391 void hrtick_start(struct rq
*rq
, u64 delay
)
393 struct hrtimer
*timer
= &rq
->hrtick_timer
;
397 * Don't schedule slices shorter than 10000ns, that just
398 * doesn't make sense and can cause timer DoS.
400 delta
= max_t(s64
, delay
, 10000LL);
401 rq
->hrtick_time
= ktime_add_ns(timer
->base
->get_time(), delta
);
404 __hrtick_restart(rq
);
406 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq
*rq
, u64 delay
)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay
= max_t(u64
, delay
, 10000LL);
422 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
423 HRTIMER_MODE_REL_PINNED_HARD
);
426 #endif /* CONFIG_SMP */
428 static void hrtick_rq_init(struct rq
*rq
)
431 INIT_CSD(&rq
->hrtick_csd
, __hrtick_start
, rq
);
433 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
434 rq
->hrtick_timer
.function
= hrtick
;
436 #else /* CONFIG_SCHED_HRTICK */
437 static inline void hrtick_clear(struct rq
*rq
)
441 static inline void hrtick_rq_init(struct rq
*rq
)
444 #endif /* CONFIG_SCHED_HRTICK */
447 * cmpxchg based fetch_or, macro so it works for different integer types
449 #define fetch_or(ptr, mask) \
451 typeof(ptr) _ptr = (ptr); \
452 typeof(mask) _mask = (mask); \
453 typeof(*_ptr) _old, _val = *_ptr; \
456 _old = cmpxchg(_ptr, _val, _val | _mask); \
464 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
466 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
467 * this avoids any races wrt polling state changes and thereby avoids
470 static bool set_nr_and_not_polling(struct task_struct
*p
)
472 struct thread_info
*ti
= task_thread_info(p
);
473 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
477 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
479 * If this returns true, then the idle task promises to call
480 * sched_ttwu_pending() and reschedule soon.
482 static bool set_nr_if_polling(struct task_struct
*p
)
484 struct thread_info
*ti
= task_thread_info(p
);
485 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
488 if (!(val
& _TIF_POLLING_NRFLAG
))
490 if (val
& _TIF_NEED_RESCHED
)
492 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
501 static bool set_nr_and_not_polling(struct task_struct
*p
)
503 set_tsk_need_resched(p
);
508 static bool set_nr_if_polling(struct task_struct
*p
)
515 static bool __wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
517 struct wake_q_node
*node
= &task
->wake_q
;
520 * Atomically grab the task, if ->wake_q is !nil already it means
521 * it's already queued (either by us or someone else) and will get the
522 * wakeup due to that.
524 * In order to ensure that a pending wakeup will observe our pending
525 * state, even in the failed case, an explicit smp_mb() must be used.
527 smp_mb__before_atomic();
528 if (unlikely(cmpxchg_relaxed(&node
->next
, NULL
, WAKE_Q_TAIL
)))
532 * The head is context local, there can be no concurrency.
535 head
->lastp
= &node
->next
;
540 * wake_q_add() - queue a wakeup for 'later' waking.
541 * @head: the wake_q_head to add @task to
542 * @task: the task to queue for 'later' wakeup
544 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
545 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
548 * This function must be used as-if it were wake_up_process(); IOW the task
549 * must be ready to be woken at this location.
551 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
553 if (__wake_q_add(head
, task
))
554 get_task_struct(task
);
558 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
559 * @head: the wake_q_head to add @task to
560 * @task: the task to queue for 'later' wakeup
562 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
563 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
566 * This function must be used as-if it were wake_up_process(); IOW the task
567 * must be ready to be woken at this location.
569 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
570 * that already hold reference to @task can call the 'safe' version and trust
571 * wake_q to do the right thing depending whether or not the @task is already
574 void wake_q_add_safe(struct wake_q_head
*head
, struct task_struct
*task
)
576 if (!__wake_q_add(head
, task
))
577 put_task_struct(task
);
580 void wake_up_q(struct wake_q_head
*head
)
582 struct wake_q_node
*node
= head
->first
;
584 while (node
!= WAKE_Q_TAIL
) {
585 struct task_struct
*task
;
587 task
= container_of(node
, struct task_struct
, wake_q
);
589 /* Task can safely be re-inserted now: */
591 task
->wake_q
.next
= NULL
;
594 * wake_up_process() executes a full barrier, which pairs with
595 * the queueing in wake_q_add() so as not to miss wakeups.
597 wake_up_process(task
);
598 put_task_struct(task
);
603 * resched_curr - mark rq's current task 'to be rescheduled now'.
605 * On UP this means the setting of the need_resched flag, on SMP it
606 * might also involve a cross-CPU call to trigger the scheduler on
609 void resched_curr(struct rq
*rq
)
611 struct task_struct
*curr
= rq
->curr
;
614 lockdep_assert_held(&rq
->lock
);
616 if (test_tsk_need_resched(curr
))
621 if (cpu
== smp_processor_id()) {
622 set_tsk_need_resched(curr
);
623 set_preempt_need_resched();
627 if (set_nr_and_not_polling(curr
))
628 smp_send_reschedule(cpu
);
630 trace_sched_wake_idle_without_ipi(cpu
);
633 void resched_cpu(int cpu
)
635 struct rq
*rq
= cpu_rq(cpu
);
638 raw_spin_lock_irqsave(&rq
->lock
, flags
);
639 if (cpu_online(cpu
) || cpu
== smp_processor_id())
641 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
645 #ifdef CONFIG_NO_HZ_COMMON
647 * In the semi idle case, use the nearest busy CPU for migrating timers
648 * from an idle CPU. This is good for power-savings.
650 * We don't do similar optimization for completely idle system, as
651 * selecting an idle CPU will add more delays to the timers than intended
652 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
654 int get_nohz_timer_target(void)
656 int i
, cpu
= smp_processor_id(), default_cpu
= -1;
657 struct sched_domain
*sd
;
659 if (housekeeping_cpu(cpu
, HK_FLAG_TIMER
)) {
666 for_each_domain(cpu
, sd
) {
667 for_each_cpu_and(i
, sched_domain_span(sd
),
668 housekeeping_cpumask(HK_FLAG_TIMER
)) {
679 if (default_cpu
== -1)
680 default_cpu
= housekeeping_any_cpu(HK_FLAG_TIMER
);
688 * When add_timer_on() enqueues a timer into the timer wheel of an
689 * idle CPU then this timer might expire before the next timer event
690 * which is scheduled to wake up that CPU. In case of a completely
691 * idle system the next event might even be infinite time into the
692 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
693 * leaves the inner idle loop so the newly added timer is taken into
694 * account when the CPU goes back to idle and evaluates the timer
695 * wheel for the next timer event.
697 static void wake_up_idle_cpu(int cpu
)
699 struct rq
*rq
= cpu_rq(cpu
);
701 if (cpu
== smp_processor_id())
704 if (set_nr_and_not_polling(rq
->idle
))
705 smp_send_reschedule(cpu
);
707 trace_sched_wake_idle_without_ipi(cpu
);
710 static bool wake_up_full_nohz_cpu(int cpu
)
713 * We just need the target to call irq_exit() and re-evaluate
714 * the next tick. The nohz full kick at least implies that.
715 * If needed we can still optimize that later with an
718 if (cpu_is_offline(cpu
))
719 return true; /* Don't try to wake offline CPUs. */
720 if (tick_nohz_full_cpu(cpu
)) {
721 if (cpu
!= smp_processor_id() ||
722 tick_nohz_tick_stopped())
723 tick_nohz_full_kick_cpu(cpu
);
731 * Wake up the specified CPU. If the CPU is going offline, it is the
732 * caller's responsibility to deal with the lost wakeup, for example,
733 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
735 void wake_up_nohz_cpu(int cpu
)
737 if (!wake_up_full_nohz_cpu(cpu
))
738 wake_up_idle_cpu(cpu
);
741 static void nohz_csd_func(void *info
)
743 struct rq
*rq
= info
;
744 int cpu
= cpu_of(rq
);
748 * Release the rq::nohz_csd.
750 flags
= atomic_fetch_andnot(NOHZ_KICK_MASK
| NOHZ_NEWILB_KICK
, nohz_flags(cpu
));
751 WARN_ON(!(flags
& NOHZ_KICK_MASK
));
753 rq
->idle_balance
= idle_cpu(cpu
);
754 if (rq
->idle_balance
&& !need_resched()) {
755 rq
->nohz_idle_balance
= flags
;
756 raise_softirq_irqoff(SCHED_SOFTIRQ
);
760 #endif /* CONFIG_NO_HZ_COMMON */
762 #ifdef CONFIG_NO_HZ_FULL
763 bool sched_can_stop_tick(struct rq
*rq
)
767 /* Deadline tasks, even if single, need the tick */
768 if (rq
->dl
.dl_nr_running
)
772 * If there are more than one RR tasks, we need the tick to affect the
773 * actual RR behaviour.
775 if (rq
->rt
.rr_nr_running
) {
776 if (rq
->rt
.rr_nr_running
== 1)
783 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
784 * forced preemption between FIFO tasks.
786 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
791 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
792 * if there's more than one we need the tick for involuntary
795 if (rq
->nr_running
> 1)
800 #endif /* CONFIG_NO_HZ_FULL */
801 #endif /* CONFIG_SMP */
803 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
804 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
806 * Iterate task_group tree rooted at *from, calling @down when first entering a
807 * node and @up when leaving it for the final time.
809 * Caller must hold rcu_lock or sufficient equivalent.
811 int walk_tg_tree_from(struct task_group
*from
,
812 tg_visitor down
, tg_visitor up
, void *data
)
814 struct task_group
*parent
, *child
;
820 ret
= (*down
)(parent
, data
);
823 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
830 ret
= (*up
)(parent
, data
);
831 if (ret
|| parent
== from
)
835 parent
= parent
->parent
;
842 int tg_nop(struct task_group
*tg
, void *data
)
848 static void set_load_weight(struct task_struct
*p
, bool update_load
)
850 int prio
= p
->static_prio
- MAX_RT_PRIO
;
851 struct load_weight
*load
= &p
->se
.load
;
854 * SCHED_IDLE tasks get minimal weight:
856 if (task_has_idle_policy(p
)) {
857 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
858 load
->inv_weight
= WMULT_IDLEPRIO
;
863 * SCHED_OTHER tasks have to update their load when changing their
866 if (update_load
&& p
->sched_class
== &fair_sched_class
) {
867 reweight_task(p
, prio
);
869 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
870 load
->inv_weight
= sched_prio_to_wmult
[prio
];
874 #ifdef CONFIG_UCLAMP_TASK
876 * Serializes updates of utilization clamp values
878 * The (slow-path) user-space triggers utilization clamp value updates which
879 * can require updates on (fast-path) scheduler's data structures used to
880 * support enqueue/dequeue operations.
881 * While the per-CPU rq lock protects fast-path update operations, user-space
882 * requests are serialized using a mutex to reduce the risk of conflicting
883 * updates or API abuses.
885 static DEFINE_MUTEX(uclamp_mutex
);
887 /* Max allowed minimum utilization */
888 unsigned int sysctl_sched_uclamp_util_min
= SCHED_CAPACITY_SCALE
;
890 /* Max allowed maximum utilization */
891 unsigned int sysctl_sched_uclamp_util_max
= SCHED_CAPACITY_SCALE
;
894 * By default RT tasks run at the maximum performance point/capacity of the
895 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
896 * SCHED_CAPACITY_SCALE.
898 * This knob allows admins to change the default behavior when uclamp is being
899 * used. In battery powered devices, particularly, running at the maximum
900 * capacity and frequency will increase energy consumption and shorten the
903 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
905 * This knob will not override the system default sched_util_clamp_min defined
908 unsigned int sysctl_sched_uclamp_util_min_rt_default
= SCHED_CAPACITY_SCALE
;
910 /* All clamps are required to be less or equal than these values */
911 static struct uclamp_se uclamp_default
[UCLAMP_CNT
];
914 * This static key is used to reduce the uclamp overhead in the fast path. It
915 * primarily disables the call to uclamp_rq_{inc, dec}() in
916 * enqueue/dequeue_task().
918 * This allows users to continue to enable uclamp in their kernel config with
919 * minimum uclamp overhead in the fast path.
921 * As soon as userspace modifies any of the uclamp knobs, the static key is
922 * enabled, since we have an actual users that make use of uclamp
925 * The knobs that would enable this static key are:
927 * * A task modifying its uclamp value with sched_setattr().
928 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
929 * * An admin modifying the cgroup cpu.uclamp.{min, max}
931 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used
);
933 /* Integer rounded range for each bucket */
934 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
936 #define for_each_clamp_id(clamp_id) \
937 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
939 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value
)
941 return clamp_value
/ UCLAMP_BUCKET_DELTA
;
944 static inline unsigned int uclamp_none(enum uclamp_id clamp_id
)
946 if (clamp_id
== UCLAMP_MIN
)
948 return SCHED_CAPACITY_SCALE
;
951 static inline void uclamp_se_set(struct uclamp_se
*uc_se
,
952 unsigned int value
, bool user_defined
)
954 uc_se
->value
= value
;
955 uc_se
->bucket_id
= uclamp_bucket_id(value
);
956 uc_se
->user_defined
= user_defined
;
959 static inline unsigned int
960 uclamp_idle_value(struct rq
*rq
, enum uclamp_id clamp_id
,
961 unsigned int clamp_value
)
964 * Avoid blocked utilization pushing up the frequency when we go
965 * idle (which drops the max-clamp) by retaining the last known
968 if (clamp_id
== UCLAMP_MAX
) {
969 rq
->uclamp_flags
|= UCLAMP_FLAG_IDLE
;
973 return uclamp_none(UCLAMP_MIN
);
976 static inline void uclamp_idle_reset(struct rq
*rq
, enum uclamp_id clamp_id
,
977 unsigned int clamp_value
)
979 /* Reset max-clamp retention only on idle exit */
980 if (!(rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
))
983 WRITE_ONCE(rq
->uclamp
[clamp_id
].value
, clamp_value
);
987 unsigned int uclamp_rq_max_value(struct rq
*rq
, enum uclamp_id clamp_id
,
988 unsigned int clamp_value
)
990 struct uclamp_bucket
*bucket
= rq
->uclamp
[clamp_id
].bucket
;
991 int bucket_id
= UCLAMP_BUCKETS
- 1;
994 * Since both min and max clamps are max aggregated, find the
995 * top most bucket with tasks in.
997 for ( ; bucket_id
>= 0; bucket_id
--) {
998 if (!bucket
[bucket_id
].tasks
)
1000 return bucket
[bucket_id
].value
;
1003 /* No tasks -- default clamp values */
1004 return uclamp_idle_value(rq
, clamp_id
, clamp_value
);
1007 static void __uclamp_update_util_min_rt_default(struct task_struct
*p
)
1009 unsigned int default_util_min
;
1010 struct uclamp_se
*uc_se
;
1012 lockdep_assert_held(&p
->pi_lock
);
1014 uc_se
= &p
->uclamp_req
[UCLAMP_MIN
];
1016 /* Only sync if user didn't override the default */
1017 if (uc_se
->user_defined
)
1020 default_util_min
= sysctl_sched_uclamp_util_min_rt_default
;
1021 uclamp_se_set(uc_se
, default_util_min
, false);
1024 static void uclamp_update_util_min_rt_default(struct task_struct
*p
)
1032 /* Protect updates to p->uclamp_* */
1033 rq
= task_rq_lock(p
, &rf
);
1034 __uclamp_update_util_min_rt_default(p
);
1035 task_rq_unlock(rq
, p
, &rf
);
1038 static void uclamp_sync_util_min_rt_default(void)
1040 struct task_struct
*g
, *p
;
1043 * copy_process() sysctl_uclamp
1044 * uclamp_min_rt = X;
1045 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1046 * // link thread smp_mb__after_spinlock()
1047 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1048 * sched_post_fork() for_each_process_thread()
1049 * __uclamp_sync_rt() __uclamp_sync_rt()
1051 * Ensures that either sched_post_fork() will observe the new
1052 * uclamp_min_rt or for_each_process_thread() will observe the new
1055 read_lock(&tasklist_lock
);
1056 smp_mb__after_spinlock();
1057 read_unlock(&tasklist_lock
);
1060 for_each_process_thread(g
, p
)
1061 uclamp_update_util_min_rt_default(p
);
1065 static inline struct uclamp_se
1066 uclamp_tg_restrict(struct task_struct
*p
, enum uclamp_id clamp_id
)
1068 struct uclamp_se uc_req
= p
->uclamp_req
[clamp_id
];
1069 #ifdef CONFIG_UCLAMP_TASK_GROUP
1070 struct uclamp_se uc_max
;
1073 * Tasks in autogroups or root task group will be
1074 * restricted by system defaults.
1076 if (task_group_is_autogroup(task_group(p
)))
1078 if (task_group(p
) == &root_task_group
)
1081 uc_max
= task_group(p
)->uclamp
[clamp_id
];
1082 if (uc_req
.value
> uc_max
.value
|| !uc_req
.user_defined
)
1090 * The effective clamp bucket index of a task depends on, by increasing
1092 * - the task specific clamp value, when explicitly requested from userspace
1093 * - the task group effective clamp value, for tasks not either in the root
1094 * group or in an autogroup
1095 * - the system default clamp value, defined by the sysadmin
1097 static inline struct uclamp_se
1098 uclamp_eff_get(struct task_struct
*p
, enum uclamp_id clamp_id
)
1100 struct uclamp_se uc_req
= uclamp_tg_restrict(p
, clamp_id
);
1101 struct uclamp_se uc_max
= uclamp_default
[clamp_id
];
1103 /* System default restrictions always apply */
1104 if (unlikely(uc_req
.value
> uc_max
.value
))
1110 unsigned long uclamp_eff_value(struct task_struct
*p
, enum uclamp_id clamp_id
)
1112 struct uclamp_se uc_eff
;
1114 /* Task currently refcounted: use back-annotated (effective) value */
1115 if (p
->uclamp
[clamp_id
].active
)
1116 return (unsigned long)p
->uclamp
[clamp_id
].value
;
1118 uc_eff
= uclamp_eff_get(p
, clamp_id
);
1120 return (unsigned long)uc_eff
.value
;
1124 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1125 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1126 * updates the rq's clamp value if required.
1128 * Tasks can have a task-specific value requested from user-space, track
1129 * within each bucket the maximum value for tasks refcounted in it.
1130 * This "local max aggregation" allows to track the exact "requested" value
1131 * for each bucket when all its RUNNABLE tasks require the same clamp.
1133 static inline void uclamp_rq_inc_id(struct rq
*rq
, struct task_struct
*p
,
1134 enum uclamp_id clamp_id
)
1136 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1137 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1138 struct uclamp_bucket
*bucket
;
1140 lockdep_assert_held(&rq
->lock
);
1142 /* Update task effective clamp */
1143 p
->uclamp
[clamp_id
] = uclamp_eff_get(p
, clamp_id
);
1145 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1147 uc_se
->active
= true;
1149 uclamp_idle_reset(rq
, clamp_id
, uc_se
->value
);
1152 * Local max aggregation: rq buckets always track the max
1153 * "requested" clamp value of its RUNNABLE tasks.
1155 if (bucket
->tasks
== 1 || uc_se
->value
> bucket
->value
)
1156 bucket
->value
= uc_se
->value
;
1158 if (uc_se
->value
> READ_ONCE(uc_rq
->value
))
1159 WRITE_ONCE(uc_rq
->value
, uc_se
->value
);
1163 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1164 * is released. If this is the last task reference counting the rq's max
1165 * active clamp value, then the rq's clamp value is updated.
1167 * Both refcounted tasks and rq's cached clamp values are expected to be
1168 * always valid. If it's detected they are not, as defensive programming,
1169 * enforce the expected state and warn.
1171 static inline void uclamp_rq_dec_id(struct rq
*rq
, struct task_struct
*p
,
1172 enum uclamp_id clamp_id
)
1174 struct uclamp_rq
*uc_rq
= &rq
->uclamp
[clamp_id
];
1175 struct uclamp_se
*uc_se
= &p
->uclamp
[clamp_id
];
1176 struct uclamp_bucket
*bucket
;
1177 unsigned int bkt_clamp
;
1178 unsigned int rq_clamp
;
1180 lockdep_assert_held(&rq
->lock
);
1183 * If sched_uclamp_used was enabled after task @p was enqueued,
1184 * we could end up with unbalanced call to uclamp_rq_dec_id().
1186 * In this case the uc_se->active flag should be false since no uclamp
1187 * accounting was performed at enqueue time and we can just return
1190 * Need to be careful of the following enqueue/dequeue ordering
1194 * // sched_uclamp_used gets enabled
1197 * // Must not decrement bucket->tasks here
1200 * where we could end up with stale data in uc_se and
1201 * bucket[uc_se->bucket_id].
1203 * The following check here eliminates the possibility of such race.
1205 if (unlikely(!uc_se
->active
))
1208 bucket
= &uc_rq
->bucket
[uc_se
->bucket_id
];
1210 SCHED_WARN_ON(!bucket
->tasks
);
1211 if (likely(bucket
->tasks
))
1214 uc_se
->active
= false;
1217 * Keep "local max aggregation" simple and accept to (possibly)
1218 * overboost some RUNNABLE tasks in the same bucket.
1219 * The rq clamp bucket value is reset to its base value whenever
1220 * there are no more RUNNABLE tasks refcounting it.
1222 if (likely(bucket
->tasks
))
1225 rq_clamp
= READ_ONCE(uc_rq
->value
);
1227 * Defensive programming: this should never happen. If it happens,
1228 * e.g. due to future modification, warn and fixup the expected value.
1230 SCHED_WARN_ON(bucket
->value
> rq_clamp
);
1231 if (bucket
->value
>= rq_clamp
) {
1232 bkt_clamp
= uclamp_rq_max_value(rq
, clamp_id
, uc_se
->value
);
1233 WRITE_ONCE(uc_rq
->value
, bkt_clamp
);
1237 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
)
1239 enum uclamp_id clamp_id
;
1242 * Avoid any overhead until uclamp is actually used by the userspace.
1244 * The condition is constructed such that a NOP is generated when
1245 * sched_uclamp_used is disabled.
1247 if (!static_branch_unlikely(&sched_uclamp_used
))
1250 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1253 for_each_clamp_id(clamp_id
)
1254 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1256 /* Reset clamp idle holding when there is one RUNNABLE task */
1257 if (rq
->uclamp_flags
& UCLAMP_FLAG_IDLE
)
1258 rq
->uclamp_flags
&= ~UCLAMP_FLAG_IDLE
;
1261 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
)
1263 enum uclamp_id clamp_id
;
1266 * Avoid any overhead until uclamp is actually used by the userspace.
1268 * The condition is constructed such that a NOP is generated when
1269 * sched_uclamp_used is disabled.
1271 if (!static_branch_unlikely(&sched_uclamp_used
))
1274 if (unlikely(!p
->sched_class
->uclamp_enabled
))
1277 for_each_clamp_id(clamp_id
)
1278 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1282 uclamp_update_active(struct task_struct
*p
, enum uclamp_id clamp_id
)
1288 * Lock the task and the rq where the task is (or was) queued.
1290 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1291 * price to pay to safely serialize util_{min,max} updates with
1292 * enqueues, dequeues and migration operations.
1293 * This is the same locking schema used by __set_cpus_allowed_ptr().
1295 rq
= task_rq_lock(p
, &rf
);
1298 * Setting the clamp bucket is serialized by task_rq_lock().
1299 * If the task is not yet RUNNABLE and its task_struct is not
1300 * affecting a valid clamp bucket, the next time it's enqueued,
1301 * it will already see the updated clamp bucket value.
1303 if (p
->uclamp
[clamp_id
].active
) {
1304 uclamp_rq_dec_id(rq
, p
, clamp_id
);
1305 uclamp_rq_inc_id(rq
, p
, clamp_id
);
1308 task_rq_unlock(rq
, p
, &rf
);
1311 #ifdef CONFIG_UCLAMP_TASK_GROUP
1313 uclamp_update_active_tasks(struct cgroup_subsys_state
*css
,
1314 unsigned int clamps
)
1316 enum uclamp_id clamp_id
;
1317 struct css_task_iter it
;
1318 struct task_struct
*p
;
1320 css_task_iter_start(css
, 0, &it
);
1321 while ((p
= css_task_iter_next(&it
))) {
1322 for_each_clamp_id(clamp_id
) {
1323 if ((0x1 << clamp_id
) & clamps
)
1324 uclamp_update_active(p
, clamp_id
);
1327 css_task_iter_end(&it
);
1330 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
);
1331 static void uclamp_update_root_tg(void)
1333 struct task_group
*tg
= &root_task_group
;
1335 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MIN
],
1336 sysctl_sched_uclamp_util_min
, false);
1337 uclamp_se_set(&tg
->uclamp_req
[UCLAMP_MAX
],
1338 sysctl_sched_uclamp_util_max
, false);
1341 cpu_util_update_eff(&root_task_group
.css
);
1345 static void uclamp_update_root_tg(void) { }
1348 int sysctl_sched_uclamp_handler(struct ctl_table
*table
, int write
,
1349 void *buffer
, size_t *lenp
, loff_t
*ppos
)
1351 bool update_root_tg
= false;
1352 int old_min
, old_max
, old_min_rt
;
1355 mutex_lock(&uclamp_mutex
);
1356 old_min
= sysctl_sched_uclamp_util_min
;
1357 old_max
= sysctl_sched_uclamp_util_max
;
1358 old_min_rt
= sysctl_sched_uclamp_util_min_rt_default
;
1360 result
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
1366 if (sysctl_sched_uclamp_util_min
> sysctl_sched_uclamp_util_max
||
1367 sysctl_sched_uclamp_util_max
> SCHED_CAPACITY_SCALE
||
1368 sysctl_sched_uclamp_util_min_rt_default
> SCHED_CAPACITY_SCALE
) {
1374 if (old_min
!= sysctl_sched_uclamp_util_min
) {
1375 uclamp_se_set(&uclamp_default
[UCLAMP_MIN
],
1376 sysctl_sched_uclamp_util_min
, false);
1377 update_root_tg
= true;
1379 if (old_max
!= sysctl_sched_uclamp_util_max
) {
1380 uclamp_se_set(&uclamp_default
[UCLAMP_MAX
],
1381 sysctl_sched_uclamp_util_max
, false);
1382 update_root_tg
= true;
1385 if (update_root_tg
) {
1386 static_branch_enable(&sched_uclamp_used
);
1387 uclamp_update_root_tg();
1390 if (old_min_rt
!= sysctl_sched_uclamp_util_min_rt_default
) {
1391 static_branch_enable(&sched_uclamp_used
);
1392 uclamp_sync_util_min_rt_default();
1396 * We update all RUNNABLE tasks only when task groups are in use.
1397 * Otherwise, keep it simple and do just a lazy update at each next
1398 * task enqueue time.
1404 sysctl_sched_uclamp_util_min
= old_min
;
1405 sysctl_sched_uclamp_util_max
= old_max
;
1406 sysctl_sched_uclamp_util_min_rt_default
= old_min_rt
;
1408 mutex_unlock(&uclamp_mutex
);
1413 static int uclamp_validate(struct task_struct
*p
,
1414 const struct sched_attr
*attr
)
1416 int util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
1417 int util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
1419 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
) {
1420 util_min
= attr
->sched_util_min
;
1422 if (util_min
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1426 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
) {
1427 util_max
= attr
->sched_util_max
;
1429 if (util_max
+ 1 > SCHED_CAPACITY_SCALE
+ 1)
1433 if (util_min
!= -1 && util_max
!= -1 && util_min
> util_max
)
1437 * We have valid uclamp attributes; make sure uclamp is enabled.
1439 * We need to do that here, because enabling static branches is a
1440 * blocking operation which obviously cannot be done while holding
1443 static_branch_enable(&sched_uclamp_used
);
1448 static bool uclamp_reset(const struct sched_attr
*attr
,
1449 enum uclamp_id clamp_id
,
1450 struct uclamp_se
*uc_se
)
1452 /* Reset on sched class change for a non user-defined clamp value. */
1453 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)) &&
1454 !uc_se
->user_defined
)
1457 /* Reset on sched_util_{min,max} == -1. */
1458 if (clamp_id
== UCLAMP_MIN
&&
1459 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1460 attr
->sched_util_min
== -1) {
1464 if (clamp_id
== UCLAMP_MAX
&&
1465 attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1466 attr
->sched_util_max
== -1) {
1473 static void __setscheduler_uclamp(struct task_struct
*p
,
1474 const struct sched_attr
*attr
)
1476 enum uclamp_id clamp_id
;
1478 for_each_clamp_id(clamp_id
) {
1479 struct uclamp_se
*uc_se
= &p
->uclamp_req
[clamp_id
];
1482 if (!uclamp_reset(attr
, clamp_id
, uc_se
))
1486 * RT by default have a 100% boost value that could be modified
1489 if (unlikely(rt_task(p
) && clamp_id
== UCLAMP_MIN
))
1490 value
= sysctl_sched_uclamp_util_min_rt_default
;
1492 value
= uclamp_none(clamp_id
);
1494 uclamp_se_set(uc_se
, value
, false);
1498 if (likely(!(attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)))
1501 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MIN
&&
1502 attr
->sched_util_min
!= -1) {
1503 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MIN
],
1504 attr
->sched_util_min
, true);
1507 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP_MAX
&&
1508 attr
->sched_util_max
!= -1) {
1509 uclamp_se_set(&p
->uclamp_req
[UCLAMP_MAX
],
1510 attr
->sched_util_max
, true);
1514 static void uclamp_fork(struct task_struct
*p
)
1516 enum uclamp_id clamp_id
;
1519 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1520 * as the task is still at its early fork stages.
1522 for_each_clamp_id(clamp_id
)
1523 p
->uclamp
[clamp_id
].active
= false;
1525 if (likely(!p
->sched_reset_on_fork
))
1528 for_each_clamp_id(clamp_id
) {
1529 uclamp_se_set(&p
->uclamp_req
[clamp_id
],
1530 uclamp_none(clamp_id
), false);
1534 static void uclamp_post_fork(struct task_struct
*p
)
1536 uclamp_update_util_min_rt_default(p
);
1539 static void __init
init_uclamp_rq(struct rq
*rq
)
1541 enum uclamp_id clamp_id
;
1542 struct uclamp_rq
*uc_rq
= rq
->uclamp
;
1544 for_each_clamp_id(clamp_id
) {
1545 uc_rq
[clamp_id
] = (struct uclamp_rq
) {
1546 .value
= uclamp_none(clamp_id
)
1550 rq
->uclamp_flags
= 0;
1553 static void __init
init_uclamp(void)
1555 struct uclamp_se uc_max
= {};
1556 enum uclamp_id clamp_id
;
1559 for_each_possible_cpu(cpu
)
1560 init_uclamp_rq(cpu_rq(cpu
));
1562 for_each_clamp_id(clamp_id
) {
1563 uclamp_se_set(&init_task
.uclamp_req
[clamp_id
],
1564 uclamp_none(clamp_id
), false);
1567 /* System defaults allow max clamp values for both indexes */
1568 uclamp_se_set(&uc_max
, uclamp_none(UCLAMP_MAX
), false);
1569 for_each_clamp_id(clamp_id
) {
1570 uclamp_default
[clamp_id
] = uc_max
;
1571 #ifdef CONFIG_UCLAMP_TASK_GROUP
1572 root_task_group
.uclamp_req
[clamp_id
] = uc_max
;
1573 root_task_group
.uclamp
[clamp_id
] = uc_max
;
1578 #else /* CONFIG_UCLAMP_TASK */
1579 static inline void uclamp_rq_inc(struct rq
*rq
, struct task_struct
*p
) { }
1580 static inline void uclamp_rq_dec(struct rq
*rq
, struct task_struct
*p
) { }
1581 static inline int uclamp_validate(struct task_struct
*p
,
1582 const struct sched_attr
*attr
)
1586 static void __setscheduler_uclamp(struct task_struct
*p
,
1587 const struct sched_attr
*attr
) { }
1588 static inline void uclamp_fork(struct task_struct
*p
) { }
1589 static inline void uclamp_post_fork(struct task_struct
*p
) { }
1590 static inline void init_uclamp(void) { }
1591 #endif /* CONFIG_UCLAMP_TASK */
1593 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1595 if (!(flags
& ENQUEUE_NOCLOCK
))
1596 update_rq_clock(rq
);
1598 if (!(flags
& ENQUEUE_RESTORE
)) {
1599 sched_info_queued(rq
, p
);
1600 psi_enqueue(p
, flags
& ENQUEUE_WAKEUP
);
1603 uclamp_rq_inc(rq
, p
);
1604 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1607 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1609 if (!(flags
& DEQUEUE_NOCLOCK
))
1610 update_rq_clock(rq
);
1612 if (!(flags
& DEQUEUE_SAVE
)) {
1613 sched_info_dequeued(rq
, p
);
1614 psi_dequeue(p
, flags
& DEQUEUE_SLEEP
);
1617 uclamp_rq_dec(rq
, p
);
1618 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1621 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1623 enqueue_task(rq
, p
, flags
);
1625 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1628 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1630 p
->on_rq
= (flags
& DEQUEUE_SLEEP
) ? 0 : TASK_ON_RQ_MIGRATING
;
1632 dequeue_task(rq
, p
, flags
);
1636 * __normal_prio - return the priority that is based on the static prio
1638 static inline int __normal_prio(struct task_struct
*p
)
1640 return p
->static_prio
;
1644 * Calculate the expected normal priority: i.e. priority
1645 * without taking RT-inheritance into account. Might be
1646 * boosted by interactivity modifiers. Changes upon fork,
1647 * setprio syscalls, and whenever the interactivity
1648 * estimator recalculates.
1650 static inline int normal_prio(struct task_struct
*p
)
1654 if (task_has_dl_policy(p
))
1655 prio
= MAX_DL_PRIO
-1;
1656 else if (task_has_rt_policy(p
))
1657 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1659 prio
= __normal_prio(p
);
1664 * Calculate the current priority, i.e. the priority
1665 * taken into account by the scheduler. This value might
1666 * be boosted by RT tasks, or might be boosted by
1667 * interactivity modifiers. Will be RT if the task got
1668 * RT-boosted. If not then it returns p->normal_prio.
1670 static int effective_prio(struct task_struct
*p
)
1672 p
->normal_prio
= normal_prio(p
);
1674 * If we are RT tasks or we were boosted to RT priority,
1675 * keep the priority unchanged. Otherwise, update priority
1676 * to the normal priority:
1678 if (!rt_prio(p
->prio
))
1679 return p
->normal_prio
;
1684 * task_curr - is this task currently executing on a CPU?
1685 * @p: the task in question.
1687 * Return: 1 if the task is currently executing. 0 otherwise.
1689 inline int task_curr(const struct task_struct
*p
)
1691 return cpu_curr(task_cpu(p
)) == p
;
1695 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1696 * use the balance_callback list if you want balancing.
1698 * this means any call to check_class_changed() must be followed by a call to
1699 * balance_callback().
1701 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1702 const struct sched_class
*prev_class
,
1705 if (prev_class
!= p
->sched_class
) {
1706 if (prev_class
->switched_from
)
1707 prev_class
->switched_from(rq
, p
);
1709 p
->sched_class
->switched_to(rq
, p
);
1710 } else if (oldprio
!= p
->prio
|| dl_task(p
))
1711 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
1714 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
1716 if (p
->sched_class
== rq
->curr
->sched_class
)
1717 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
1718 else if (p
->sched_class
> rq
->curr
->sched_class
)
1722 * A queue event has occurred, and we're going to schedule. In
1723 * this case, we can save a useless back to back clock update.
1725 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
1726 rq_clock_skip_update(rq
);
1732 __do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
);
1734 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1735 const struct cpumask
*new_mask
,
1738 static void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
)
1740 if (likely(!p
->migration_disabled
))
1743 if (p
->cpus_ptr
!= &p
->cpus_mask
)
1747 * Violates locking rules! see comment in __do_set_cpus_allowed().
1749 __do_set_cpus_allowed(p
, cpumask_of(rq
->cpu
), SCA_MIGRATE_DISABLE
);
1752 void migrate_disable(void)
1754 struct task_struct
*p
= current
;
1756 if (p
->migration_disabled
) {
1757 p
->migration_disabled
++;
1762 this_rq()->nr_pinned
++;
1763 p
->migration_disabled
= 1;
1766 EXPORT_SYMBOL_GPL(migrate_disable
);
1768 void migrate_enable(void)
1770 struct task_struct
*p
= current
;
1772 if (p
->migration_disabled
> 1) {
1773 p
->migration_disabled
--;
1778 * Ensure stop_task runs either before or after this, and that
1779 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
1782 if (p
->cpus_ptr
!= &p
->cpus_mask
)
1783 __set_cpus_allowed_ptr(p
, &p
->cpus_mask
, SCA_MIGRATE_ENABLE
);
1785 * Mustn't clear migration_disabled() until cpus_ptr points back at the
1786 * regular cpus_mask, otherwise things that race (eg.
1787 * select_fallback_rq) get confused.
1790 p
->migration_disabled
= 0;
1791 this_rq()->nr_pinned
--;
1794 EXPORT_SYMBOL_GPL(migrate_enable
);
1796 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
1798 return rq
->nr_pinned
;
1802 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1803 * __set_cpus_allowed_ptr() and select_fallback_rq().
1805 static inline bool is_cpu_allowed(struct task_struct
*p
, int cpu
)
1807 /* When not in the task's cpumask, no point in looking further. */
1808 if (!cpumask_test_cpu(cpu
, p
->cpus_ptr
))
1811 /* migrate_disabled() must be allowed to finish. */
1812 if (is_migration_disabled(p
))
1813 return cpu_online(cpu
);
1815 /* Non kernel threads are not allowed during either online or offline. */
1816 if (!(p
->flags
& PF_KTHREAD
))
1817 return cpu_active(cpu
);
1819 /* KTHREAD_IS_PER_CPU is always allowed. */
1820 if (kthread_is_per_cpu(p
))
1821 return cpu_online(cpu
);
1823 /* Regular kernel threads don't get to stay during offline. */
1827 /* But are allowed during online. */
1828 return cpu_online(cpu
);
1832 * This is how migration works:
1834 * 1) we invoke migration_cpu_stop() on the target CPU using
1836 * 2) stopper starts to run (implicitly forcing the migrated thread
1838 * 3) it checks whether the migrated task is still in the wrong runqueue.
1839 * 4) if it's in the wrong runqueue then the migration thread removes
1840 * it and puts it into the right queue.
1841 * 5) stopper completes and stop_one_cpu() returns and the migration
1846 * move_queued_task - move a queued task to new rq.
1848 * Returns (locked) new rq. Old rq's lock is released.
1850 static struct rq
*move_queued_task(struct rq
*rq
, struct rq_flags
*rf
,
1851 struct task_struct
*p
, int new_cpu
)
1853 lockdep_assert_held(&rq
->lock
);
1855 deactivate_task(rq
, p
, DEQUEUE_NOCLOCK
);
1856 set_task_cpu(p
, new_cpu
);
1859 rq
= cpu_rq(new_cpu
);
1862 BUG_ON(task_cpu(p
) != new_cpu
);
1863 activate_task(rq
, p
, 0);
1864 check_preempt_curr(rq
, p
, 0);
1869 struct migration_arg
{
1870 struct task_struct
*task
;
1872 struct set_affinity_pending
*pending
;
1876 * @refs: number of wait_for_completion()
1877 * @stop_pending: is @stop_work in use
1879 struct set_affinity_pending
{
1881 unsigned int stop_pending
;
1882 struct completion done
;
1883 struct cpu_stop_work stop_work
;
1884 struct migration_arg arg
;
1888 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1889 * this because either it can't run here any more (set_cpus_allowed()
1890 * away from this CPU, or CPU going down), or because we're
1891 * attempting to rebalance this task on exec (sched_exec).
1893 * So we race with normal scheduler movements, but that's OK, as long
1894 * as the task is no longer on this CPU.
1896 static struct rq
*__migrate_task(struct rq
*rq
, struct rq_flags
*rf
,
1897 struct task_struct
*p
, int dest_cpu
)
1899 /* Affinity changed (again). */
1900 if (!is_cpu_allowed(p
, dest_cpu
))
1903 update_rq_clock(rq
);
1904 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
1910 * migration_cpu_stop - this will be executed by a highprio stopper thread
1911 * and performs thread migration by bumping thread off CPU then
1912 * 'pushing' onto another runqueue.
1914 static int migration_cpu_stop(void *data
)
1916 struct migration_arg
*arg
= data
;
1917 struct set_affinity_pending
*pending
= arg
->pending
;
1918 struct task_struct
*p
= arg
->task
;
1919 int dest_cpu
= arg
->dest_cpu
;
1920 struct rq
*rq
= this_rq();
1921 bool complete
= false;
1925 * The original target CPU might have gone down and we might
1926 * be on another CPU but it doesn't matter.
1928 local_irq_save(rf
.flags
);
1930 * We need to explicitly wake pending tasks before running
1931 * __migrate_task() such that we will not miss enforcing cpus_ptr
1932 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1934 flush_smp_call_function_from_idle();
1936 raw_spin_lock(&p
->pi_lock
);
1940 * If we were passed a pending, then ->stop_pending was set, thus
1941 * p->migration_pending must have remained stable.
1943 WARN_ON_ONCE(pending
&& pending
!= p
->migration_pending
);
1946 * If task_rq(p) != rq, it cannot be migrated here, because we're
1947 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1948 * we're holding p->pi_lock.
1950 if (task_rq(p
) == rq
) {
1951 if (is_migration_disabled(p
))
1955 p
->migration_pending
= NULL
;
1960 if (cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
))
1963 dest_cpu
= cpumask_any_distribute(&p
->cpus_mask
);
1966 if (task_on_rq_queued(p
))
1967 rq
= __migrate_task(rq
, &rf
, p
, dest_cpu
);
1969 p
->wake_cpu
= dest_cpu
;
1972 * XXX __migrate_task() can fail, at which point we might end
1973 * up running on a dodgy CPU, AFAICT this can only happen
1974 * during CPU hotplug, at which point we'll get pushed out
1975 * anyway, so it's probably not a big deal.
1978 } else if (pending
) {
1980 * This happens when we get migrated between migrate_enable()'s
1981 * preempt_enable() and scheduling the stopper task. At that
1982 * point we're a regular task again and not current anymore.
1984 * A !PREEMPT kernel has a giant hole here, which makes it far
1989 * The task moved before the stopper got to run. We're holding
1990 * ->pi_lock, so the allowed mask is stable - if it got
1991 * somewhere allowed, we're done.
1993 if (cpumask_test_cpu(task_cpu(p
), p
->cpus_ptr
)) {
1994 p
->migration_pending
= NULL
;
2000 * When migrate_enable() hits a rq mis-match we can't reliably
2001 * determine is_migration_disabled() and so have to chase after
2004 WARN_ON_ONCE(!pending
->stop_pending
);
2005 task_rq_unlock(rq
, p
, &rf
);
2006 stop_one_cpu_nowait(task_cpu(p
), migration_cpu_stop
,
2007 &pending
->arg
, &pending
->stop_work
);
2012 pending
->stop_pending
= false;
2013 task_rq_unlock(rq
, p
, &rf
);
2016 complete_all(&pending
->done
);
2021 int push_cpu_stop(void *arg
)
2023 struct rq
*lowest_rq
= NULL
, *rq
= this_rq();
2024 struct task_struct
*p
= arg
;
2026 raw_spin_lock_irq(&p
->pi_lock
);
2027 raw_spin_lock(&rq
->lock
);
2029 if (task_rq(p
) != rq
)
2032 if (is_migration_disabled(p
)) {
2033 p
->migration_flags
|= MDF_PUSH
;
2037 p
->migration_flags
&= ~MDF_PUSH
;
2039 if (p
->sched_class
->find_lock_rq
)
2040 lowest_rq
= p
->sched_class
->find_lock_rq(p
, rq
);
2045 // XXX validate p is still the highest prio task
2046 if (task_rq(p
) == rq
) {
2047 deactivate_task(rq
, p
, 0);
2048 set_task_cpu(p
, lowest_rq
->cpu
);
2049 activate_task(lowest_rq
, p
, 0);
2050 resched_curr(lowest_rq
);
2053 double_unlock_balance(rq
, lowest_rq
);
2056 rq
->push_busy
= false;
2057 raw_spin_unlock(&rq
->lock
);
2058 raw_spin_unlock_irq(&p
->pi_lock
);
2065 * sched_class::set_cpus_allowed must do the below, but is not required to
2066 * actually call this function.
2068 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
)
2070 if (flags
& (SCA_MIGRATE_ENABLE
| SCA_MIGRATE_DISABLE
)) {
2071 p
->cpus_ptr
= new_mask
;
2075 cpumask_copy(&p
->cpus_mask
, new_mask
);
2076 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
2080 __do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
, u32 flags
)
2082 struct rq
*rq
= task_rq(p
);
2083 bool queued
, running
;
2086 * This here violates the locking rules for affinity, since we're only
2087 * supposed to change these variables while holding both rq->lock and
2090 * HOWEVER, it magically works, because ttwu() is the only code that
2091 * accesses these variables under p->pi_lock and only does so after
2092 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2093 * before finish_task().
2095 * XXX do further audits, this smells like something putrid.
2097 if (flags
& SCA_MIGRATE_DISABLE
)
2098 SCHED_WARN_ON(!p
->on_cpu
);
2100 lockdep_assert_held(&p
->pi_lock
);
2102 queued
= task_on_rq_queued(p
);
2103 running
= task_current(rq
, p
);
2107 * Because __kthread_bind() calls this on blocked tasks without
2110 lockdep_assert_held(&rq
->lock
);
2111 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
2114 put_prev_task(rq
, p
);
2116 p
->sched_class
->set_cpus_allowed(p
, new_mask
, flags
);
2119 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
2121 set_next_task(rq
, p
);
2124 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
2126 __do_set_cpus_allowed(p
, new_mask
, 0);
2130 * This function is wildly self concurrent; here be dragons.
2133 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2134 * designated task is enqueued on an allowed CPU. If that task is currently
2135 * running, we have to kick it out using the CPU stopper.
2137 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2140 * Initial conditions: P0->cpus_mask = [0, 1]
2144 * migrate_disable();
2146 * set_cpus_allowed_ptr(P0, [1]);
2148 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2149 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2150 * This means we need the following scheme:
2154 * migrate_disable();
2156 * set_cpus_allowed_ptr(P0, [1]);
2160 * __set_cpus_allowed_ptr();
2161 * <wakes local stopper>
2162 * `--> <woken on migration completion>
2164 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2165 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2166 * task p are serialized by p->pi_lock, which we can leverage: the one that
2167 * should come into effect at the end of the Migrate-Disable region is the last
2168 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2169 * but we still need to properly signal those waiting tasks at the appropriate
2172 * This is implemented using struct set_affinity_pending. The first
2173 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2174 * setup an instance of that struct and install it on the targeted task_struct.
2175 * Any and all further callers will reuse that instance. Those then wait for
2176 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2177 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2180 * (1) In the cases covered above. There is one more where the completion is
2181 * signaled within affine_move_task() itself: when a subsequent affinity request
2182 * occurs after the stopper bailed out due to the targeted task still being
2183 * Migrate-Disable. Consider:
2185 * Initial conditions: P0->cpus_mask = [0, 1]
2189 * migrate_disable();
2191 * set_cpus_allowed_ptr(P0, [1]);
2194 * migration_cpu_stop()
2195 * is_migration_disabled()
2197 * set_cpus_allowed_ptr(P0, [0, 1]);
2198 * <signal completion>
2201 * Note that the above is safe vs a concurrent migrate_enable(), as any
2202 * pending affinity completion is preceded by an uninstallation of
2203 * p->migration_pending done with p->pi_lock held.
2205 static int affine_move_task(struct rq
*rq
, struct task_struct
*p
, struct rq_flags
*rf
,
2206 int dest_cpu
, unsigned int flags
)
2208 struct set_affinity_pending my_pending
= { }, *pending
= NULL
;
2209 bool stop_pending
, complete
= false;
2211 /* Can the task run on the task's current CPU? If so, we're done */
2212 if (cpumask_test_cpu(task_cpu(p
), &p
->cpus_mask
)) {
2213 struct task_struct
*push_task
= NULL
;
2215 if ((flags
& SCA_MIGRATE_ENABLE
) &&
2216 (p
->migration_flags
& MDF_PUSH
) && !rq
->push_busy
) {
2217 rq
->push_busy
= true;
2218 push_task
= get_task_struct(p
);
2222 * If there are pending waiters, but no pending stop_work,
2223 * then complete now.
2225 pending
= p
->migration_pending
;
2226 if (pending
&& !pending
->stop_pending
) {
2227 p
->migration_pending
= NULL
;
2231 task_rq_unlock(rq
, p
, rf
);
2234 stop_one_cpu_nowait(rq
->cpu
, push_cpu_stop
,
2239 complete_all(&pending
->done
);
2244 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2245 /* serialized by p->pi_lock */
2246 if (!p
->migration_pending
) {
2247 /* Install the request */
2248 refcount_set(&my_pending
.refs
, 1);
2249 init_completion(&my_pending
.done
);
2250 my_pending
.arg
= (struct migration_arg
) {
2252 .dest_cpu
= -1, /* any */
2253 .pending
= &my_pending
,
2256 p
->migration_pending
= &my_pending
;
2258 pending
= p
->migration_pending
;
2259 refcount_inc(&pending
->refs
);
2262 pending
= p
->migration_pending
;
2264 * - !MIGRATE_ENABLE:
2265 * we'll have installed a pending if there wasn't one already.
2268 * we're here because the current CPU isn't matching anymore,
2269 * the only way that can happen is because of a concurrent
2270 * set_cpus_allowed_ptr() call, which should then still be
2271 * pending completion.
2273 * Either way, we really should have a @pending here.
2275 if (WARN_ON_ONCE(!pending
)) {
2276 task_rq_unlock(rq
, p
, rf
);
2280 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
2282 * MIGRATE_ENABLE gets here because 'p == current', but for
2283 * anything else we cannot do is_migration_disabled(), punt
2284 * and have the stopper function handle it all race-free.
2286 stop_pending
= pending
->stop_pending
;
2288 pending
->stop_pending
= true;
2290 if (flags
& SCA_MIGRATE_ENABLE
)
2291 p
->migration_flags
&= ~MDF_PUSH
;
2293 task_rq_unlock(rq
, p
, rf
);
2295 if (!stop_pending
) {
2296 stop_one_cpu_nowait(cpu_of(rq
), migration_cpu_stop
,
2297 &pending
->arg
, &pending
->stop_work
);
2300 if (flags
& SCA_MIGRATE_ENABLE
)
2304 if (!is_migration_disabled(p
)) {
2305 if (task_on_rq_queued(p
))
2306 rq
= move_queued_task(rq
, rf
, p
, dest_cpu
);
2308 if (!pending
->stop_pending
) {
2309 p
->migration_pending
= NULL
;
2313 task_rq_unlock(rq
, p
, rf
);
2316 complete_all(&pending
->done
);
2319 wait_for_completion(&pending
->done
);
2321 if (refcount_dec_and_test(&pending
->refs
))
2322 wake_up_var(&pending
->refs
); /* No UaF, just an address */
2325 * Block the original owner of &pending until all subsequent callers
2326 * have seen the completion and decremented the refcount
2328 wait_var_event(&my_pending
.refs
, !refcount_read(&my_pending
.refs
));
2331 WARN_ON_ONCE(my_pending
.stop_pending
);
2337 * Change a given task's CPU affinity. Migrate the thread to a
2338 * proper CPU and schedule it away if the CPU it's executing on
2339 * is removed from the allowed bitmask.
2341 * NOTE: the caller must have a valid reference to the task, the
2342 * task must not exit() & deallocate itself prematurely. The
2343 * call is not atomic; no spinlocks may be held.
2345 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
2346 const struct cpumask
*new_mask
,
2349 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
2350 unsigned int dest_cpu
;
2355 rq
= task_rq_lock(p
, &rf
);
2356 update_rq_clock(rq
);
2358 if (p
->flags
& PF_KTHREAD
|| is_migration_disabled(p
)) {
2360 * Kernel threads are allowed on online && !active CPUs,
2361 * however, during cpu-hot-unplug, even these might get pushed
2362 * away if not KTHREAD_IS_PER_CPU.
2364 * Specifically, migration_disabled() tasks must not fail the
2365 * cpumask_any_and_distribute() pick below, esp. so on
2366 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2367 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2369 cpu_valid_mask
= cpu_online_mask
;
2373 * Must re-check here, to close a race against __kthread_bind(),
2374 * sched_setaffinity() is not guaranteed to observe the flag.
2376 if ((flags
& SCA_CHECK
) && (p
->flags
& PF_NO_SETAFFINITY
)) {
2381 if (!(flags
& SCA_MIGRATE_ENABLE
)) {
2382 if (cpumask_equal(&p
->cpus_mask
, new_mask
))
2385 if (WARN_ON_ONCE(p
== current
&&
2386 is_migration_disabled(p
) &&
2387 !cpumask_test_cpu(task_cpu(p
), new_mask
))) {
2394 * Picking a ~random cpu helps in cases where we are changing affinity
2395 * for groups of tasks (ie. cpuset), so that load balancing is not
2396 * immediately required to distribute the tasks within their new mask.
2398 dest_cpu
= cpumask_any_and_distribute(cpu_valid_mask
, new_mask
);
2399 if (dest_cpu
>= nr_cpu_ids
) {
2404 __do_set_cpus_allowed(p
, new_mask
, flags
);
2406 return affine_move_task(rq
, p
, &rf
, dest_cpu
, flags
);
2409 task_rq_unlock(rq
, p
, &rf
);
2414 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
2416 return __set_cpus_allowed_ptr(p
, new_mask
, 0);
2418 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
2420 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2422 #ifdef CONFIG_SCHED_DEBUG
2424 * We should never call set_task_cpu() on a blocked task,
2425 * ttwu() will sort out the placement.
2427 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2431 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2432 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2433 * time relying on p->on_rq.
2435 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
2436 p
->sched_class
== &fair_sched_class
&&
2437 (p
->on_rq
&& !task_on_rq_migrating(p
)));
2439 #ifdef CONFIG_LOCKDEP
2441 * The caller should hold either p->pi_lock or rq->lock, when changing
2442 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2444 * sched_move_task() holds both and thus holding either pins the cgroup,
2447 * Furthermore, all task_rq users should acquire both locks, see
2450 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2451 lockdep_is_held(&task_rq(p
)->lock
)));
2454 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2456 WARN_ON_ONCE(!cpu_online(new_cpu
));
2458 WARN_ON_ONCE(is_migration_disabled(p
));
2461 trace_sched_migrate_task(p
, new_cpu
);
2463 if (task_cpu(p
) != new_cpu
) {
2464 if (p
->sched_class
->migrate_task_rq
)
2465 p
->sched_class
->migrate_task_rq(p
, new_cpu
);
2466 p
->se
.nr_migrations
++;
2468 perf_event_task_migrate(p
);
2471 __set_task_cpu(p
, new_cpu
);
2474 #ifdef CONFIG_NUMA_BALANCING
2475 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
2477 if (task_on_rq_queued(p
)) {
2478 struct rq
*src_rq
, *dst_rq
;
2479 struct rq_flags srf
, drf
;
2481 src_rq
= task_rq(p
);
2482 dst_rq
= cpu_rq(cpu
);
2484 rq_pin_lock(src_rq
, &srf
);
2485 rq_pin_lock(dst_rq
, &drf
);
2487 deactivate_task(src_rq
, p
, 0);
2488 set_task_cpu(p
, cpu
);
2489 activate_task(dst_rq
, p
, 0);
2490 check_preempt_curr(dst_rq
, p
, 0);
2492 rq_unpin_lock(dst_rq
, &drf
);
2493 rq_unpin_lock(src_rq
, &srf
);
2497 * Task isn't running anymore; make it appear like we migrated
2498 * it before it went to sleep. This means on wakeup we make the
2499 * previous CPU our target instead of where it really is.
2505 struct migration_swap_arg
{
2506 struct task_struct
*src_task
, *dst_task
;
2507 int src_cpu
, dst_cpu
;
2510 static int migrate_swap_stop(void *data
)
2512 struct migration_swap_arg
*arg
= data
;
2513 struct rq
*src_rq
, *dst_rq
;
2516 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
2519 src_rq
= cpu_rq(arg
->src_cpu
);
2520 dst_rq
= cpu_rq(arg
->dst_cpu
);
2522 double_raw_lock(&arg
->src_task
->pi_lock
,
2523 &arg
->dst_task
->pi_lock
);
2524 double_rq_lock(src_rq
, dst_rq
);
2526 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
2529 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
2532 if (!cpumask_test_cpu(arg
->dst_cpu
, arg
->src_task
->cpus_ptr
))
2535 if (!cpumask_test_cpu(arg
->src_cpu
, arg
->dst_task
->cpus_ptr
))
2538 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
2539 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
2544 double_rq_unlock(src_rq
, dst_rq
);
2545 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
2546 raw_spin_unlock(&arg
->src_task
->pi_lock
);
2552 * Cross migrate two tasks
2554 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
,
2555 int target_cpu
, int curr_cpu
)
2557 struct migration_swap_arg arg
;
2560 arg
= (struct migration_swap_arg
){
2562 .src_cpu
= curr_cpu
,
2564 .dst_cpu
= target_cpu
,
2567 if (arg
.src_cpu
== arg
.dst_cpu
)
2571 * These three tests are all lockless; this is OK since all of them
2572 * will be re-checked with proper locks held further down the line.
2574 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
2577 if (!cpumask_test_cpu(arg
.dst_cpu
, arg
.src_task
->cpus_ptr
))
2580 if (!cpumask_test_cpu(arg
.src_cpu
, arg
.dst_task
->cpus_ptr
))
2583 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
2584 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
2589 #endif /* CONFIG_NUMA_BALANCING */
2592 * wait_task_inactive - wait for a thread to unschedule.
2594 * If @match_state is nonzero, it's the @p->state value just checked and
2595 * not expected to change. If it changes, i.e. @p might have woken up,
2596 * then return zero. When we succeed in waiting for @p to be off its CPU,
2597 * we return a positive number (its total switch count). If a second call
2598 * a short while later returns the same number, the caller can be sure that
2599 * @p has remained unscheduled the whole time.
2601 * The caller must ensure that the task *will* unschedule sometime soon,
2602 * else this function might spin for a *long* time. This function can't
2603 * be called with interrupts off, or it may introduce deadlock with
2604 * smp_call_function() if an IPI is sent by the same process we are
2605 * waiting to become inactive.
2607 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2609 int running
, queued
;
2616 * We do the initial early heuristics without holding
2617 * any task-queue locks at all. We'll only try to get
2618 * the runqueue lock when things look like they will
2624 * If the task is actively running on another CPU
2625 * still, just relax and busy-wait without holding
2628 * NOTE! Since we don't hold any locks, it's not
2629 * even sure that "rq" stays as the right runqueue!
2630 * But we don't care, since "task_running()" will
2631 * return false if the runqueue has changed and p
2632 * is actually now running somewhere else!
2634 while (task_running(rq
, p
)) {
2635 if (match_state
&& unlikely(p
->state
!= match_state
))
2641 * Ok, time to look more closely! We need the rq
2642 * lock now, to be *sure*. If we're wrong, we'll
2643 * just go back and repeat.
2645 rq
= task_rq_lock(p
, &rf
);
2646 trace_sched_wait_task(p
);
2647 running
= task_running(rq
, p
);
2648 queued
= task_on_rq_queued(p
);
2650 if (!match_state
|| p
->state
== match_state
)
2651 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2652 task_rq_unlock(rq
, p
, &rf
);
2655 * If it changed from the expected state, bail out now.
2657 if (unlikely(!ncsw
))
2661 * Was it really running after all now that we
2662 * checked with the proper locks actually held?
2664 * Oops. Go back and try again..
2666 if (unlikely(running
)) {
2672 * It's not enough that it's not actively running,
2673 * it must be off the runqueue _entirely_, and not
2676 * So if it was still runnable (but just not actively
2677 * running right now), it's preempted, and we should
2678 * yield - it could be a while.
2680 if (unlikely(queued
)) {
2681 ktime_t to
= NSEC_PER_SEC
/ HZ
;
2683 set_current_state(TASK_UNINTERRUPTIBLE
);
2684 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2689 * Ahh, all good. It wasn't running, and it wasn't
2690 * runnable, which means that it will never become
2691 * running in the future either. We're all done!
2700 * kick_process - kick a running thread to enter/exit the kernel
2701 * @p: the to-be-kicked thread
2703 * Cause a process which is running on another CPU to enter
2704 * kernel-mode, without any delay. (to get signals handled.)
2706 * NOTE: this function doesn't have to take the runqueue lock,
2707 * because all it wants to ensure is that the remote task enters
2708 * the kernel. If the IPI races and the task has been migrated
2709 * to another CPU then no harm is done and the purpose has been
2712 void kick_process(struct task_struct
*p
)
2718 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2719 smp_send_reschedule(cpu
);
2722 EXPORT_SYMBOL_GPL(kick_process
);
2725 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2727 * A few notes on cpu_active vs cpu_online:
2729 * - cpu_active must be a subset of cpu_online
2731 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2732 * see __set_cpus_allowed_ptr(). At this point the newly online
2733 * CPU isn't yet part of the sched domains, and balancing will not
2736 * - on CPU-down we clear cpu_active() to mask the sched domains and
2737 * avoid the load balancer to place new tasks on the to be removed
2738 * CPU. Existing tasks will remain running there and will be taken
2741 * This means that fallback selection must not select !active CPUs.
2742 * And can assume that any active CPU must be online. Conversely
2743 * select_task_rq() below may allow selection of !active CPUs in order
2744 * to satisfy the above rules.
2746 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2748 int nid
= cpu_to_node(cpu
);
2749 const struct cpumask
*nodemask
= NULL
;
2750 enum { cpuset
, possible
, fail
} state
= cpuset
;
2754 * If the node that the CPU is on has been offlined, cpu_to_node()
2755 * will return -1. There is no CPU on the node, and we should
2756 * select the CPU on the other node.
2759 nodemask
= cpumask_of_node(nid
);
2761 /* Look for allowed, online CPU in same node. */
2762 for_each_cpu(dest_cpu
, nodemask
) {
2763 if (!cpu_active(dest_cpu
))
2765 if (cpumask_test_cpu(dest_cpu
, p
->cpus_ptr
))
2771 /* Any allowed, online CPU? */
2772 for_each_cpu(dest_cpu
, p
->cpus_ptr
) {
2773 if (!is_cpu_allowed(p
, dest_cpu
))
2779 /* No more Mr. Nice Guy. */
2782 if (IS_ENABLED(CONFIG_CPUSETS
)) {
2783 cpuset_cpus_allowed_fallback(p
);
2790 * XXX When called from select_task_rq() we only
2791 * hold p->pi_lock and again violate locking order.
2793 * More yuck to audit.
2795 do_set_cpus_allowed(p
, cpu_possible_mask
);
2806 if (state
!= cpuset
) {
2808 * Don't tell them about moving exiting tasks or
2809 * kernel threads (both mm NULL), since they never
2812 if (p
->mm
&& printk_ratelimit()) {
2813 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2814 task_pid_nr(p
), p
->comm
, cpu
);
2822 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2825 int select_task_rq(struct task_struct
*p
, int cpu
, int wake_flags
)
2827 lockdep_assert_held(&p
->pi_lock
);
2829 if (p
->nr_cpus_allowed
> 1 && !is_migration_disabled(p
))
2830 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, wake_flags
);
2832 cpu
= cpumask_any(p
->cpus_ptr
);
2835 * In order not to call set_task_cpu() on a blocking task we need
2836 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2839 * Since this is common to all placement strategies, this lives here.
2841 * [ this allows ->select_task() to simply return task_cpu(p) and
2842 * not worry about this generic constraint ]
2844 if (unlikely(!is_cpu_allowed(p
, cpu
)))
2845 cpu
= select_fallback_rq(task_cpu(p
), p
);
2850 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2852 static struct lock_class_key stop_pi_lock
;
2853 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2854 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2858 * Make it appear like a SCHED_FIFO task, its something
2859 * userspace knows about and won't get confused about.
2861 * Also, it will make PI more or less work without too
2862 * much confusion -- but then, stop work should not
2863 * rely on PI working anyway.
2865 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2867 stop
->sched_class
= &stop_sched_class
;
2870 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
2871 * adjust the effective priority of a task. As a result,
2872 * rt_mutex_setprio() can trigger (RT) balancing operations,
2873 * which can then trigger wakeups of the stop thread to push
2874 * around the current task.
2876 * The stop task itself will never be part of the PI-chain, it
2877 * never blocks, therefore that ->pi_lock recursion is safe.
2878 * Tell lockdep about this by placing the stop->pi_lock in its
2881 lockdep_set_class(&stop
->pi_lock
, &stop_pi_lock
);
2884 cpu_rq(cpu
)->stop
= stop
;
2888 * Reset it back to a normal scheduling class so that
2889 * it can die in pieces.
2891 old_stop
->sched_class
= &rt_sched_class
;
2895 #else /* CONFIG_SMP */
2897 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
2898 const struct cpumask
*new_mask
,
2901 return set_cpus_allowed_ptr(p
, new_mask
);
2904 static inline void migrate_disable_switch(struct rq
*rq
, struct task_struct
*p
) { }
2906 static inline bool rq_has_pinned_tasks(struct rq
*rq
)
2911 #endif /* !CONFIG_SMP */
2914 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2918 if (!schedstat_enabled())
2924 if (cpu
== rq
->cpu
) {
2925 __schedstat_inc(rq
->ttwu_local
);
2926 __schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
2928 struct sched_domain
*sd
;
2930 __schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
2932 for_each_domain(rq
->cpu
, sd
) {
2933 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2934 __schedstat_inc(sd
->ttwu_wake_remote
);
2941 if (wake_flags
& WF_MIGRATED
)
2942 __schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
2943 #endif /* CONFIG_SMP */
2945 __schedstat_inc(rq
->ttwu_count
);
2946 __schedstat_inc(p
->se
.statistics
.nr_wakeups
);
2948 if (wake_flags
& WF_SYNC
)
2949 __schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
2953 * Mark the task runnable and perform wakeup-preemption.
2955 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2956 struct rq_flags
*rf
)
2958 check_preempt_curr(rq
, p
, wake_flags
);
2959 p
->state
= TASK_RUNNING
;
2960 trace_sched_wakeup(p
);
2963 if (p
->sched_class
->task_woken
) {
2965 * Our task @p is fully woken up and running; so it's safe to
2966 * drop the rq->lock, hereafter rq is only used for statistics.
2968 rq_unpin_lock(rq
, rf
);
2969 p
->sched_class
->task_woken(rq
, p
);
2970 rq_repin_lock(rq
, rf
);
2973 if (rq
->idle_stamp
) {
2974 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
2975 u64 max
= 2*rq
->max_idle_balance_cost
;
2977 update_avg(&rq
->avg_idle
, delta
);
2979 if (rq
->avg_idle
> max
)
2988 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
2989 struct rq_flags
*rf
)
2991 int en_flags
= ENQUEUE_WAKEUP
| ENQUEUE_NOCLOCK
;
2993 lockdep_assert_held(&rq
->lock
);
2995 if (p
->sched_contributes_to_load
)
2996 rq
->nr_uninterruptible
--;
2999 if (wake_flags
& WF_MIGRATED
)
3000 en_flags
|= ENQUEUE_MIGRATED
;
3004 delayacct_blkio_end(p
);
3005 atomic_dec(&task_rq(p
)->nr_iowait
);
3008 activate_task(rq
, p
, en_flags
);
3009 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
3013 * Consider @p being inside a wait loop:
3016 * set_current_state(TASK_UNINTERRUPTIBLE);
3023 * __set_current_state(TASK_RUNNING);
3025 * between set_current_state() and schedule(). In this case @p is still
3026 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3029 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3030 * then schedule() must still happen and p->state can be changed to
3031 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3032 * need to do a full wakeup with enqueue.
3034 * Returns: %true when the wakeup is done,
3037 static int ttwu_runnable(struct task_struct
*p
, int wake_flags
)
3043 rq
= __task_rq_lock(p
, &rf
);
3044 if (task_on_rq_queued(p
)) {
3045 /* check_preempt_curr() may use rq clock */
3046 update_rq_clock(rq
);
3047 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
3050 __task_rq_unlock(rq
, &rf
);
3056 void sched_ttwu_pending(void *arg
)
3058 struct llist_node
*llist
= arg
;
3059 struct rq
*rq
= this_rq();
3060 struct task_struct
*p
, *t
;
3067 * rq::ttwu_pending racy indication of out-standing wakeups.
3068 * Races such that false-negatives are possible, since they
3069 * are shorter lived that false-positives would be.
3071 WRITE_ONCE(rq
->ttwu_pending
, 0);
3073 rq_lock_irqsave(rq
, &rf
);
3074 update_rq_clock(rq
);
3076 llist_for_each_entry_safe(p
, t
, llist
, wake_entry
.llist
) {
3077 if (WARN_ON_ONCE(p
->on_cpu
))
3078 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3080 if (WARN_ON_ONCE(task_cpu(p
) != cpu_of(rq
)))
3081 set_task_cpu(p
, cpu_of(rq
));
3083 ttwu_do_activate(rq
, p
, p
->sched_remote_wakeup
? WF_MIGRATED
: 0, &rf
);
3086 rq_unlock_irqrestore(rq
, &rf
);
3089 void send_call_function_single_ipi(int cpu
)
3091 struct rq
*rq
= cpu_rq(cpu
);
3093 if (!set_nr_if_polling(rq
->idle
))
3094 arch_send_call_function_single_ipi(cpu
);
3096 trace_sched_wake_idle_without_ipi(cpu
);
3100 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3101 * necessary. The wakee CPU on receipt of the IPI will queue the task
3102 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3103 * of the wakeup instead of the waker.
3105 static void __ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3107 struct rq
*rq
= cpu_rq(cpu
);
3109 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
3111 WRITE_ONCE(rq
->ttwu_pending
, 1);
3112 __smp_call_single_queue(cpu
, &p
->wake_entry
.llist
);
3115 void wake_up_if_idle(int cpu
)
3117 struct rq
*rq
= cpu_rq(cpu
);
3122 if (!is_idle_task(rcu_dereference(rq
->curr
)))
3125 if (set_nr_if_polling(rq
->idle
)) {
3126 trace_sched_wake_idle_without_ipi(cpu
);
3128 rq_lock_irqsave(rq
, &rf
);
3129 if (is_idle_task(rq
->curr
))
3130 smp_send_reschedule(cpu
);
3131 /* Else CPU is not idle, do nothing here: */
3132 rq_unlock_irqrestore(rq
, &rf
);
3139 bool cpus_share_cache(int this_cpu
, int that_cpu
)
3141 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
3144 static inline bool ttwu_queue_cond(int cpu
, int wake_flags
)
3147 * Do not complicate things with the async wake_list while the CPU is
3150 if (!cpu_active(cpu
))
3154 * If the CPU does not share cache, then queue the task on the
3155 * remote rqs wakelist to avoid accessing remote data.
3157 if (!cpus_share_cache(smp_processor_id(), cpu
))
3161 * If the task is descheduling and the only running task on the
3162 * CPU then use the wakelist to offload the task activation to
3163 * the soon-to-be-idle CPU as the current CPU is likely busy.
3164 * nr_running is checked to avoid unnecessary task stacking.
3166 if ((wake_flags
& WF_ON_CPU
) && cpu_rq(cpu
)->nr_running
<= 1)
3172 static bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3174 if (sched_feat(TTWU_QUEUE
) && ttwu_queue_cond(cpu
, wake_flags
)) {
3175 if (WARN_ON_ONCE(cpu
== smp_processor_id()))
3178 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
3179 __ttwu_queue_wakelist(p
, cpu
, wake_flags
);
3186 #else /* !CONFIG_SMP */
3188 static inline bool ttwu_queue_wakelist(struct task_struct
*p
, int cpu
, int wake_flags
)
3193 #endif /* CONFIG_SMP */
3195 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
3197 struct rq
*rq
= cpu_rq(cpu
);
3200 if (ttwu_queue_wakelist(p
, cpu
, wake_flags
))
3204 update_rq_clock(rq
);
3205 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
3210 * Notes on Program-Order guarantees on SMP systems.
3214 * The basic program-order guarantee on SMP systems is that when a task [t]
3215 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3216 * execution on its new CPU [c1].
3218 * For migration (of runnable tasks) this is provided by the following means:
3220 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3221 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3222 * rq(c1)->lock (if not at the same time, then in that order).
3223 * C) LOCK of the rq(c1)->lock scheduling in task
3225 * Release/acquire chaining guarantees that B happens after A and C after B.
3226 * Note: the CPU doing B need not be c0 or c1
3235 * UNLOCK rq(0)->lock
3237 * LOCK rq(0)->lock // orders against CPU0
3239 * UNLOCK rq(0)->lock
3243 * UNLOCK rq(1)->lock
3245 * LOCK rq(1)->lock // orders against CPU2
3248 * UNLOCK rq(1)->lock
3251 * BLOCKING -- aka. SLEEP + WAKEUP
3253 * For blocking we (obviously) need to provide the same guarantee as for
3254 * migration. However the means are completely different as there is no lock
3255 * chain to provide order. Instead we do:
3257 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3258 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3262 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3264 * LOCK rq(0)->lock LOCK X->pi_lock
3267 * smp_store_release(X->on_cpu, 0);
3269 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3275 * X->state = RUNNING
3276 * UNLOCK rq(2)->lock
3278 * LOCK rq(2)->lock // orders against CPU1
3281 * UNLOCK rq(2)->lock
3284 * UNLOCK rq(0)->lock
3287 * However, for wakeups there is a second guarantee we must provide, namely we
3288 * must ensure that CONDITION=1 done by the caller can not be reordered with
3289 * accesses to the task state; see try_to_wake_up() and set_current_state().
3293 * try_to_wake_up - wake up a thread
3294 * @p: the thread to be awakened
3295 * @state: the mask of task states that can be woken
3296 * @wake_flags: wake modifier flags (WF_*)
3298 * Conceptually does:
3300 * If (@state & @p->state) @p->state = TASK_RUNNING.
3302 * If the task was not queued/runnable, also place it back on a runqueue.
3304 * This function is atomic against schedule() which would dequeue the task.
3306 * It issues a full memory barrier before accessing @p->state, see the comment
3307 * with set_current_state().
3309 * Uses p->pi_lock to serialize against concurrent wake-ups.
3311 * Relies on p->pi_lock stabilizing:
3314 * - p->sched_task_group
3315 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3317 * Tries really hard to only take one task_rq(p)->lock for performance.
3318 * Takes rq->lock in:
3319 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3320 * - ttwu_queue() -- new rq, for enqueue of the task;
3321 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3323 * As a consequence we race really badly with just about everything. See the
3324 * many memory barriers and their comments for details.
3326 * Return: %true if @p->state changes (an actual wakeup was done),
3330 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
3332 unsigned long flags
;
3333 int cpu
, success
= 0;
3338 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3339 * == smp_processor_id()'. Together this means we can special
3340 * case the whole 'p->on_rq && ttwu_runnable()' case below
3341 * without taking any locks.
3344 * - we rely on Program-Order guarantees for all the ordering,
3345 * - we're serialized against set_special_state() by virtue of
3346 * it disabling IRQs (this allows not taking ->pi_lock).
3348 if (!(p
->state
& state
))
3352 trace_sched_waking(p
);
3353 p
->state
= TASK_RUNNING
;
3354 trace_sched_wakeup(p
);
3359 * If we are going to wake up a thread waiting for CONDITION we
3360 * need to ensure that CONDITION=1 done by the caller can not be
3361 * reordered with p->state check below. This pairs with smp_store_mb()
3362 * in set_current_state() that the waiting thread does.
3364 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3365 smp_mb__after_spinlock();
3366 if (!(p
->state
& state
))
3369 trace_sched_waking(p
);
3371 /* We're going to change ->state: */
3375 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3376 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3377 * in smp_cond_load_acquire() below.
3379 * sched_ttwu_pending() try_to_wake_up()
3380 * STORE p->on_rq = 1 LOAD p->state
3383 * __schedule() (switch to task 'p')
3384 * LOCK rq->lock smp_rmb();
3385 * smp_mb__after_spinlock();
3389 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3391 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3392 * __schedule(). See the comment for smp_mb__after_spinlock().
3394 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3397 if (READ_ONCE(p
->on_rq
) && ttwu_runnable(p
, wake_flags
))
3402 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3403 * possible to, falsely, observe p->on_cpu == 0.
3405 * One must be running (->on_cpu == 1) in order to remove oneself
3406 * from the runqueue.
3408 * __schedule() (switch to task 'p') try_to_wake_up()
3409 * STORE p->on_cpu = 1 LOAD p->on_rq
3412 * __schedule() (put 'p' to sleep)
3413 * LOCK rq->lock smp_rmb();
3414 * smp_mb__after_spinlock();
3415 * STORE p->on_rq = 0 LOAD p->on_cpu
3417 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3418 * __schedule(). See the comment for smp_mb__after_spinlock().
3420 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3421 * schedule()'s deactivate_task() has 'happened' and p will no longer
3422 * care about it's own p->state. See the comment in __schedule().
3424 smp_acquire__after_ctrl_dep();
3427 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3428 * == 0), which means we need to do an enqueue, change p->state to
3429 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3430 * enqueue, such as ttwu_queue_wakelist().
3432 p
->state
= TASK_WAKING
;
3435 * If the owning (remote) CPU is still in the middle of schedule() with
3436 * this task as prev, considering queueing p on the remote CPUs wake_list
3437 * which potentially sends an IPI instead of spinning on p->on_cpu to
3438 * let the waker make forward progress. This is safe because IRQs are
3439 * disabled and the IPI will deliver after on_cpu is cleared.
3441 * Ensure we load task_cpu(p) after p->on_cpu:
3443 * set_task_cpu(p, cpu);
3444 * STORE p->cpu = @cpu
3445 * __schedule() (switch to task 'p')
3447 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3448 * STORE p->on_cpu = 1 LOAD p->cpu
3450 * to ensure we observe the correct CPU on which the task is currently
3453 if (smp_load_acquire(&p
->on_cpu
) &&
3454 ttwu_queue_wakelist(p
, task_cpu(p
), wake_flags
| WF_ON_CPU
))
3458 * If the owning (remote) CPU is still in the middle of schedule() with
3459 * this task as prev, wait until it's done referencing the task.
3461 * Pairs with the smp_store_release() in finish_task().
3463 * This ensures that tasks getting woken will be fully ordered against
3464 * their previous state and preserve Program Order.
3466 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
3468 cpu
= select_task_rq(p
, p
->wake_cpu
, wake_flags
| WF_TTWU
);
3469 if (task_cpu(p
) != cpu
) {
3471 delayacct_blkio_end(p
);
3472 atomic_dec(&task_rq(p
)->nr_iowait
);
3475 wake_flags
|= WF_MIGRATED
;
3476 psi_ttwu_dequeue(p
);
3477 set_task_cpu(p
, cpu
);
3481 #endif /* CONFIG_SMP */
3483 ttwu_queue(p
, cpu
, wake_flags
);
3485 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3488 ttwu_stat(p
, task_cpu(p
), wake_flags
);
3495 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3496 * @p: Process for which the function is to be invoked, can be @current.
3497 * @func: Function to invoke.
3498 * @arg: Argument to function.
3500 * If the specified task can be quickly locked into a definite state
3501 * (either sleeping or on a given runqueue), arrange to keep it in that
3502 * state while invoking @func(@arg). This function can use ->on_rq and
3503 * task_curr() to work out what the state is, if required. Given that
3504 * @func can be invoked with a runqueue lock held, it had better be quite
3508 * @false if the task slipped out from under the locks.
3509 * @true if the task was locked onto a runqueue or is sleeping.
3510 * However, @func can override this by returning @false.
3512 bool try_invoke_on_locked_down_task(struct task_struct
*p
, bool (*func
)(struct task_struct
*t
, void *arg
), void *arg
)
3518 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
3520 rq
= __task_rq_lock(p
, &rf
);
3521 if (task_rq(p
) == rq
)
3530 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3535 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
.flags
);
3540 * wake_up_process - Wake up a specific process
3541 * @p: The process to be woken up.
3543 * Attempt to wake up the nominated process and move it to the set of runnable
3546 * Return: 1 if the process was woken up, 0 if it was already running.
3548 * This function executes a full memory barrier before accessing the task state.
3550 int wake_up_process(struct task_struct
*p
)
3552 return try_to_wake_up(p
, TASK_NORMAL
, 0);
3554 EXPORT_SYMBOL(wake_up_process
);
3556 int wake_up_state(struct task_struct
*p
, unsigned int state
)
3558 return try_to_wake_up(p
, state
, 0);
3562 * Perform scheduler related setup for a newly forked process p.
3563 * p is forked by current.
3565 * __sched_fork() is basic setup used by init_idle() too:
3567 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
3572 p
->se
.exec_start
= 0;
3573 p
->se
.sum_exec_runtime
= 0;
3574 p
->se
.prev_sum_exec_runtime
= 0;
3575 p
->se
.nr_migrations
= 0;
3577 INIT_LIST_HEAD(&p
->se
.group_node
);
3579 #ifdef CONFIG_FAIR_GROUP_SCHED
3580 p
->se
.cfs_rq
= NULL
;
3583 #ifdef CONFIG_SCHEDSTATS
3584 /* Even if schedstat is disabled, there should not be garbage */
3585 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
3588 RB_CLEAR_NODE(&p
->dl
.rb_node
);
3589 init_dl_task_timer(&p
->dl
);
3590 init_dl_inactive_task_timer(&p
->dl
);
3591 __dl_clear_params(p
);
3593 INIT_LIST_HEAD(&p
->rt
.run_list
);
3595 p
->rt
.time_slice
= sched_rr_timeslice
;
3599 #ifdef CONFIG_PREEMPT_NOTIFIERS
3600 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
3603 #ifdef CONFIG_COMPACTION
3604 p
->capture_control
= NULL
;
3606 init_numa_balancing(clone_flags
, p
);
3608 p
->wake_entry
.u_flags
= CSD_TYPE_TTWU
;
3609 p
->migration_pending
= NULL
;
3613 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
3615 #ifdef CONFIG_NUMA_BALANCING
3617 void set_numabalancing_state(bool enabled
)
3620 static_branch_enable(&sched_numa_balancing
);
3622 static_branch_disable(&sched_numa_balancing
);
3625 #ifdef CONFIG_PROC_SYSCTL
3626 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
3627 void *buffer
, size_t *lenp
, loff_t
*ppos
)
3631 int state
= static_branch_likely(&sched_numa_balancing
);
3633 if (write
&& !capable(CAP_SYS_ADMIN
))
3638 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
3642 set_numabalancing_state(state
);
3648 #ifdef CONFIG_SCHEDSTATS
3650 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
3651 static bool __initdata __sched_schedstats
= false;
3653 static void set_schedstats(bool enabled
)
3656 static_branch_enable(&sched_schedstats
);
3658 static_branch_disable(&sched_schedstats
);
3661 void force_schedstat_enabled(void)
3663 if (!schedstat_enabled()) {
3664 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3665 static_branch_enable(&sched_schedstats
);
3669 static int __init
setup_schedstats(char *str
)
3676 * This code is called before jump labels have been set up, so we can't
3677 * change the static branch directly just yet. Instead set a temporary
3678 * variable so init_schedstats() can do it later.
3680 if (!strcmp(str
, "enable")) {
3681 __sched_schedstats
= true;
3683 } else if (!strcmp(str
, "disable")) {
3684 __sched_schedstats
= false;
3689 pr_warn("Unable to parse schedstats=\n");
3693 __setup("schedstats=", setup_schedstats
);
3695 static void __init
init_schedstats(void)
3697 set_schedstats(__sched_schedstats
);
3700 #ifdef CONFIG_PROC_SYSCTL
3701 int sysctl_schedstats(struct ctl_table
*table
, int write
, void *buffer
,
3702 size_t *lenp
, loff_t
*ppos
)
3706 int state
= static_branch_likely(&sched_schedstats
);
3708 if (write
&& !capable(CAP_SYS_ADMIN
))
3713 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
3717 set_schedstats(state
);
3720 #endif /* CONFIG_PROC_SYSCTL */
3721 #else /* !CONFIG_SCHEDSTATS */
3722 static inline void init_schedstats(void) {}
3723 #endif /* CONFIG_SCHEDSTATS */
3726 * fork()/clone()-time setup:
3728 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
3730 unsigned long flags
;
3732 __sched_fork(clone_flags
, p
);
3734 * We mark the process as NEW here. This guarantees that
3735 * nobody will actually run it, and a signal or other external
3736 * event cannot wake it up and insert it on the runqueue either.
3738 p
->state
= TASK_NEW
;
3741 * Make sure we do not leak PI boosting priority to the child.
3743 p
->prio
= current
->normal_prio
;
3748 * Revert to default priority/policy on fork if requested.
3750 if (unlikely(p
->sched_reset_on_fork
)) {
3751 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3752 p
->policy
= SCHED_NORMAL
;
3753 p
->static_prio
= NICE_TO_PRIO(0);
3755 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
3756 p
->static_prio
= NICE_TO_PRIO(0);
3758 p
->prio
= p
->normal_prio
= __normal_prio(p
);
3759 set_load_weight(p
, false);
3762 * We don't need the reset flag anymore after the fork. It has
3763 * fulfilled its duty:
3765 p
->sched_reset_on_fork
= 0;
3768 if (dl_prio(p
->prio
))
3770 else if (rt_prio(p
->prio
))
3771 p
->sched_class
= &rt_sched_class
;
3773 p
->sched_class
= &fair_sched_class
;
3775 init_entity_runnable_average(&p
->se
);
3778 * The child is not yet in the pid-hash so no cgroup attach races,
3779 * and the cgroup is pinned to this child due to cgroup_fork()
3780 * is ran before sched_fork().
3782 * Silence PROVE_RCU.
3784 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3787 * We're setting the CPU for the first time, we don't migrate,
3788 * so use __set_task_cpu().
3790 __set_task_cpu(p
, smp_processor_id());
3791 if (p
->sched_class
->task_fork
)
3792 p
->sched_class
->task_fork(p
);
3793 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3795 #ifdef CONFIG_SCHED_INFO
3796 if (likely(sched_info_on()))
3797 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
3799 #if defined(CONFIG_SMP)
3802 init_task_preempt_count(p
);
3804 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
3805 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
3810 void sched_post_fork(struct task_struct
*p
)
3812 uclamp_post_fork(p
);
3815 unsigned long to_ratio(u64 period
, u64 runtime
)
3817 if (runtime
== RUNTIME_INF
)
3821 * Doing this here saves a lot of checks in all
3822 * the calling paths, and returning zero seems
3823 * safe for them anyway.
3828 return div64_u64(runtime
<< BW_SHIFT
, period
);
3832 * wake_up_new_task - wake up a newly created task for the first time.
3834 * This function will do some initial scheduler statistics housekeeping
3835 * that must be done for every newly created context, then puts the task
3836 * on the runqueue and wakes it.
3838 void wake_up_new_task(struct task_struct
*p
)
3843 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
3844 p
->state
= TASK_RUNNING
;
3847 * Fork balancing, do it here and not earlier because:
3848 * - cpus_ptr can change in the fork path
3849 * - any previously selected CPU might disappear through hotplug
3851 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3852 * as we're not fully set-up yet.
3854 p
->recent_used_cpu
= task_cpu(p
);
3856 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), WF_FORK
));
3858 rq
= __task_rq_lock(p
, &rf
);
3859 update_rq_clock(rq
);
3860 post_init_entity_util_avg(p
);
3862 activate_task(rq
, p
, ENQUEUE_NOCLOCK
);
3863 trace_sched_wakeup_new(p
);
3864 check_preempt_curr(rq
, p
, WF_FORK
);
3866 if (p
->sched_class
->task_woken
) {
3868 * Nothing relies on rq->lock after this, so it's fine to
3871 rq_unpin_lock(rq
, &rf
);
3872 p
->sched_class
->task_woken(rq
, p
);
3873 rq_repin_lock(rq
, &rf
);
3876 task_rq_unlock(rq
, p
, &rf
);
3879 #ifdef CONFIG_PREEMPT_NOTIFIERS
3881 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key
);
3883 void preempt_notifier_inc(void)
3885 static_branch_inc(&preempt_notifier_key
);
3887 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
3889 void preempt_notifier_dec(void)
3891 static_branch_dec(&preempt_notifier_key
);
3893 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
3896 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3897 * @notifier: notifier struct to register
3899 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3901 if (!static_branch_unlikely(&preempt_notifier_key
))
3902 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3904 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3906 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3909 * preempt_notifier_unregister - no longer interested in preemption notifications
3910 * @notifier: notifier struct to unregister
3912 * This is *not* safe to call from within a preemption notifier.
3914 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3916 hlist_del(¬ifier
->link
);
3918 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3920 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3922 struct preempt_notifier
*notifier
;
3924 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3925 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3928 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3930 if (static_branch_unlikely(&preempt_notifier_key
))
3931 __fire_sched_in_preempt_notifiers(curr
);
3935 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3936 struct task_struct
*next
)
3938 struct preempt_notifier
*notifier
;
3940 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
3941 notifier
->ops
->sched_out(notifier
, next
);
3944 static __always_inline
void
3945 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3946 struct task_struct
*next
)
3948 if (static_branch_unlikely(&preempt_notifier_key
))
3949 __fire_sched_out_preempt_notifiers(curr
, next
);
3952 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3954 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3959 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3960 struct task_struct
*next
)
3964 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3966 static inline void prepare_task(struct task_struct
*next
)
3970 * Claim the task as running, we do this before switching to it
3971 * such that any running task will have this set.
3973 * See the ttwu() WF_ON_CPU case and its ordering comment.
3975 WRITE_ONCE(next
->on_cpu
, 1);
3979 static inline void finish_task(struct task_struct
*prev
)
3983 * This must be the very last reference to @prev from this CPU. After
3984 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3985 * must ensure this doesn't happen until the switch is completely
3988 * In particular, the load of prev->state in finish_task_switch() must
3989 * happen before this.
3991 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3993 smp_store_release(&prev
->on_cpu
, 0);
3999 static void do_balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
4001 void (*func
)(struct rq
*rq
);
4002 struct callback_head
*next
;
4004 lockdep_assert_held(&rq
->lock
);
4007 func
= (void (*)(struct rq
*))head
->func
;
4016 static void balance_push(struct rq
*rq
);
4018 struct callback_head balance_push_callback
= {
4020 .func
= (void (*)(struct callback_head
*))balance_push
,
4023 static inline struct callback_head
*splice_balance_callbacks(struct rq
*rq
)
4025 struct callback_head
*head
= rq
->balance_callback
;
4027 lockdep_assert_held(&rq
->lock
);
4029 rq
->balance_callback
= NULL
;
4034 static void __balance_callbacks(struct rq
*rq
)
4036 do_balance_callbacks(rq
, splice_balance_callbacks(rq
));
4039 static inline void balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
4041 unsigned long flags
;
4043 if (unlikely(head
)) {
4044 raw_spin_lock_irqsave(&rq
->lock
, flags
);
4045 do_balance_callbacks(rq
, head
);
4046 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
4052 static inline void __balance_callbacks(struct rq
*rq
)
4056 static inline struct callback_head
*splice_balance_callbacks(struct rq
*rq
)
4061 static inline void balance_callbacks(struct rq
*rq
, struct callback_head
*head
)
4068 prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
, struct rq_flags
*rf
)
4071 * Since the runqueue lock will be released by the next
4072 * task (which is an invalid locking op but in the case
4073 * of the scheduler it's an obvious special-case), so we
4074 * do an early lockdep release here:
4076 rq_unpin_lock(rq
, rf
);
4077 spin_release(&rq
->lock
.dep_map
, _THIS_IP_
);
4078 #ifdef CONFIG_DEBUG_SPINLOCK
4079 /* this is a valid case when another task releases the spinlock */
4080 rq
->lock
.owner
= next
;
4084 static inline void finish_lock_switch(struct rq
*rq
)
4087 * If we are tracking spinlock dependencies then we have to
4088 * fix up the runqueue lock - which gets 'carried over' from
4089 * prev into current:
4091 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
4092 __balance_callbacks(rq
);
4093 raw_spin_unlock_irq(&rq
->lock
);
4097 * NOP if the arch has not defined these:
4100 #ifndef prepare_arch_switch
4101 # define prepare_arch_switch(next) do { } while (0)
4104 #ifndef finish_arch_post_lock_switch
4105 # define finish_arch_post_lock_switch() do { } while (0)
4108 static inline void kmap_local_sched_out(void)
4110 #ifdef CONFIG_KMAP_LOCAL
4111 if (unlikely(current
->kmap_ctrl
.idx
))
4112 __kmap_local_sched_out();
4116 static inline void kmap_local_sched_in(void)
4118 #ifdef CONFIG_KMAP_LOCAL
4119 if (unlikely(current
->kmap_ctrl
.idx
))
4120 __kmap_local_sched_in();
4125 * prepare_task_switch - prepare to switch tasks
4126 * @rq: the runqueue preparing to switch
4127 * @prev: the current task that is being switched out
4128 * @next: the task we are going to switch to.
4130 * This is called with the rq lock held and interrupts off. It must
4131 * be paired with a subsequent finish_task_switch after the context
4134 * prepare_task_switch sets up locking and calls architecture specific
4138 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
4139 struct task_struct
*next
)
4141 kcov_prepare_switch(prev
);
4142 sched_info_switch(rq
, prev
, next
);
4143 perf_event_task_sched_out(prev
, next
);
4145 fire_sched_out_preempt_notifiers(prev
, next
);
4146 kmap_local_sched_out();
4148 prepare_arch_switch(next
);
4152 * finish_task_switch - clean up after a task-switch
4153 * @prev: the thread we just switched away from.
4155 * finish_task_switch must be called after the context switch, paired
4156 * with a prepare_task_switch call before the context switch.
4157 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4158 * and do any other architecture-specific cleanup actions.
4160 * Note that we may have delayed dropping an mm in context_switch(). If
4161 * so, we finish that here outside of the runqueue lock. (Doing it
4162 * with the lock held can cause deadlocks; see schedule() for
4165 * The context switch have flipped the stack from under us and restored the
4166 * local variables which were saved when this task called schedule() in the
4167 * past. prev == current is still correct but we need to recalculate this_rq
4168 * because prev may have moved to another CPU.
4170 static struct rq
*finish_task_switch(struct task_struct
*prev
)
4171 __releases(rq
->lock
)
4173 struct rq
*rq
= this_rq();
4174 struct mm_struct
*mm
= rq
->prev_mm
;
4178 * The previous task will have left us with a preempt_count of 2
4179 * because it left us after:
4182 * preempt_disable(); // 1
4184 * raw_spin_lock_irq(&rq->lock) // 2
4186 * Also, see FORK_PREEMPT_COUNT.
4188 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
4189 "corrupted preempt_count: %s/%d/0x%x\n",
4190 current
->comm
, current
->pid
, preempt_count()))
4191 preempt_count_set(FORK_PREEMPT_COUNT
);
4196 * A task struct has one reference for the use as "current".
4197 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4198 * schedule one last time. The schedule call will never return, and
4199 * the scheduled task must drop that reference.
4201 * We must observe prev->state before clearing prev->on_cpu (in
4202 * finish_task), otherwise a concurrent wakeup can get prev
4203 * running on another CPU and we could rave with its RUNNING -> DEAD
4204 * transition, resulting in a double drop.
4206 prev_state
= prev
->state
;
4207 vtime_task_switch(prev
);
4208 perf_event_task_sched_in(prev
, current
);
4210 finish_lock_switch(rq
);
4211 finish_arch_post_lock_switch();
4212 kcov_finish_switch(current
);
4214 * kmap_local_sched_out() is invoked with rq::lock held and
4215 * interrupts disabled. There is no requirement for that, but the
4216 * sched out code does not have an interrupt enabled section.
4217 * Restoring the maps on sched in does not require interrupts being
4220 kmap_local_sched_in();
4222 fire_sched_in_preempt_notifiers(current
);
4224 * When switching through a kernel thread, the loop in
4225 * membarrier_{private,global}_expedited() may have observed that
4226 * kernel thread and not issued an IPI. It is therefore possible to
4227 * schedule between user->kernel->user threads without passing though
4228 * switch_mm(). Membarrier requires a barrier after storing to
4229 * rq->curr, before returning to userspace, so provide them here:
4231 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4232 * provided by mmdrop(),
4233 * - a sync_core for SYNC_CORE.
4236 membarrier_mm_sync_core_before_usermode(mm
);
4239 if (unlikely(prev_state
== TASK_DEAD
)) {
4240 if (prev
->sched_class
->task_dead
)
4241 prev
->sched_class
->task_dead(prev
);
4244 * Remove function-return probe instances associated with this
4245 * task and put them back on the free list.
4247 kprobe_flush_task(prev
);
4249 /* Task is done with its stack. */
4250 put_task_stack(prev
);
4252 put_task_struct_rcu_user(prev
);
4255 tick_nohz_task_switch();
4260 * schedule_tail - first thing a freshly forked thread must call.
4261 * @prev: the thread we just switched away from.
4263 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
4264 __releases(rq
->lock
)
4267 * New tasks start with FORK_PREEMPT_COUNT, see there and
4268 * finish_task_switch() for details.
4270 * finish_task_switch() will drop rq->lock() and lower preempt_count
4271 * and the preempt_enable() will end up enabling preemption (on
4272 * PREEMPT_COUNT kernels).
4275 finish_task_switch(prev
);
4278 if (current
->set_child_tid
)
4279 put_user(task_pid_vnr(current
), current
->set_child_tid
);
4281 calculate_sigpending();
4285 * context_switch - switch to the new MM and the new thread's register state.
4287 static __always_inline
struct rq
*
4288 context_switch(struct rq
*rq
, struct task_struct
*prev
,
4289 struct task_struct
*next
, struct rq_flags
*rf
)
4291 prepare_task_switch(rq
, prev
, next
);
4294 * For paravirt, this is coupled with an exit in switch_to to
4295 * combine the page table reload and the switch backend into
4298 arch_start_context_switch(prev
);
4301 * kernel -> kernel lazy + transfer active
4302 * user -> kernel lazy + mmgrab() active
4304 * kernel -> user switch + mmdrop() active
4305 * user -> user switch
4307 if (!next
->mm
) { // to kernel
4308 enter_lazy_tlb(prev
->active_mm
, next
);
4310 next
->active_mm
= prev
->active_mm
;
4311 if (prev
->mm
) // from user
4312 mmgrab(prev
->active_mm
);
4314 prev
->active_mm
= NULL
;
4316 membarrier_switch_mm(rq
, prev
->active_mm
, next
->mm
);
4318 * sys_membarrier() requires an smp_mb() between setting
4319 * rq->curr / membarrier_switch_mm() and returning to userspace.
4321 * The below provides this either through switch_mm(), or in
4322 * case 'prev->active_mm == next->mm' through
4323 * finish_task_switch()'s mmdrop().
4325 switch_mm_irqs_off(prev
->active_mm
, next
->mm
, next
);
4327 if (!prev
->mm
) { // from kernel
4328 /* will mmdrop() in finish_task_switch(). */
4329 rq
->prev_mm
= prev
->active_mm
;
4330 prev
->active_mm
= NULL
;
4334 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
4336 prepare_lock_switch(rq
, next
, rf
);
4338 /* Here we just switch the register state and the stack. */
4339 switch_to(prev
, next
, prev
);
4342 return finish_task_switch(prev
);
4346 * nr_running and nr_context_switches:
4348 * externally visible scheduler statistics: current number of runnable
4349 * threads, total number of context switches performed since bootup.
4351 unsigned long nr_running(void)
4353 unsigned long i
, sum
= 0;
4355 for_each_online_cpu(i
)
4356 sum
+= cpu_rq(i
)->nr_running
;
4362 * Check if only the current task is running on the CPU.
4364 * Caution: this function does not check that the caller has disabled
4365 * preemption, thus the result might have a time-of-check-to-time-of-use
4366 * race. The caller is responsible to use it correctly, for example:
4368 * - from a non-preemptible section (of course)
4370 * - from a thread that is bound to a single CPU
4372 * - in a loop with very short iterations (e.g. a polling loop)
4374 bool single_task_running(void)
4376 return raw_rq()->nr_running
== 1;
4378 EXPORT_SYMBOL(single_task_running
);
4380 unsigned long long nr_context_switches(void)
4383 unsigned long long sum
= 0;
4385 for_each_possible_cpu(i
)
4386 sum
+= cpu_rq(i
)->nr_switches
;
4392 * Consumers of these two interfaces, like for example the cpuidle menu
4393 * governor, are using nonsensical data. Preferring shallow idle state selection
4394 * for a CPU that has IO-wait which might not even end up running the task when
4395 * it does become runnable.
4398 unsigned long nr_iowait_cpu(int cpu
)
4400 return atomic_read(&cpu_rq(cpu
)->nr_iowait
);
4404 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4406 * The idea behind IO-wait account is to account the idle time that we could
4407 * have spend running if it were not for IO. That is, if we were to improve the
4408 * storage performance, we'd have a proportional reduction in IO-wait time.
4410 * This all works nicely on UP, where, when a task blocks on IO, we account
4411 * idle time as IO-wait, because if the storage were faster, it could've been
4412 * running and we'd not be idle.
4414 * This has been extended to SMP, by doing the same for each CPU. This however
4417 * Imagine for instance the case where two tasks block on one CPU, only the one
4418 * CPU will have IO-wait accounted, while the other has regular idle. Even
4419 * though, if the storage were faster, both could've ran at the same time,
4420 * utilising both CPUs.
4422 * This means, that when looking globally, the current IO-wait accounting on
4423 * SMP is a lower bound, by reason of under accounting.
4425 * Worse, since the numbers are provided per CPU, they are sometimes
4426 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4427 * associated with any one particular CPU, it can wake to another CPU than it
4428 * blocked on. This means the per CPU IO-wait number is meaningless.
4430 * Task CPU affinities can make all that even more 'interesting'.
4433 unsigned long nr_iowait(void)
4435 unsigned long i
, sum
= 0;
4437 for_each_possible_cpu(i
)
4438 sum
+= nr_iowait_cpu(i
);
4446 * sched_exec - execve() is a valuable balancing opportunity, because at
4447 * this point the task has the smallest effective memory and cache footprint.
4449 void sched_exec(void)
4451 struct task_struct
*p
= current
;
4452 unsigned long flags
;
4455 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4456 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), WF_EXEC
);
4457 if (dest_cpu
== smp_processor_id())
4460 if (likely(cpu_active(dest_cpu
))) {
4461 struct migration_arg arg
= { p
, dest_cpu
};
4463 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4464 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
4468 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4473 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4474 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
4476 EXPORT_PER_CPU_SYMBOL(kstat
);
4477 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
4480 * The function fair_sched_class.update_curr accesses the struct curr
4481 * and its field curr->exec_start; when called from task_sched_runtime(),
4482 * we observe a high rate of cache misses in practice.
4483 * Prefetching this data results in improved performance.
4485 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
4487 #ifdef CONFIG_FAIR_GROUP_SCHED
4488 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
4490 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
4493 prefetch(&curr
->exec_start
);
4497 * Return accounted runtime for the task.
4498 * In case the task is currently running, return the runtime plus current's
4499 * pending runtime that have not been accounted yet.
4501 unsigned long long task_sched_runtime(struct task_struct
*p
)
4507 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4509 * 64-bit doesn't need locks to atomically read a 64-bit value.
4510 * So we have a optimization chance when the task's delta_exec is 0.
4511 * Reading ->on_cpu is racy, but this is ok.
4513 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4514 * If we race with it entering CPU, unaccounted time is 0. This is
4515 * indistinguishable from the read occurring a few cycles earlier.
4516 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4517 * been accounted, so we're correct here as well.
4519 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
4520 return p
->se
.sum_exec_runtime
;
4523 rq
= task_rq_lock(p
, &rf
);
4525 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4526 * project cycles that may never be accounted to this
4527 * thread, breaking clock_gettime().
4529 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
4530 prefetch_curr_exec_start(p
);
4531 update_rq_clock(rq
);
4532 p
->sched_class
->update_curr(rq
);
4534 ns
= p
->se
.sum_exec_runtime
;
4535 task_rq_unlock(rq
, p
, &rf
);
4540 #ifdef CONFIG_SCHED_DEBUG
4541 static u64
cpu_resched_latency(struct rq
*rq
)
4543 int latency_warn_ms
= READ_ONCE(sysctl_resched_latency_warn_ms
);
4544 u64 resched_latency
, now
= rq_clock(rq
);
4545 static bool warned_once
;
4547 if (sysctl_resched_latency_warn_once
&& warned_once
)
4550 if (!need_resched() || !latency_warn_ms
)
4553 if (system_state
== SYSTEM_BOOTING
)
4556 if (!rq
->last_seen_need_resched_ns
) {
4557 rq
->last_seen_need_resched_ns
= now
;
4558 rq
->ticks_without_resched
= 0;
4562 rq
->ticks_without_resched
++;
4563 resched_latency
= now
- rq
->last_seen_need_resched_ns
;
4564 if (resched_latency
<= latency_warn_ms
* NSEC_PER_MSEC
)
4569 return resched_latency
;
4572 static int __init
setup_resched_latency_warn_ms(char *str
)
4576 if ((kstrtol(str
, 0, &val
))) {
4577 pr_warn("Unable to set resched_latency_warn_ms\n");
4581 sysctl_resched_latency_warn_ms
= val
;
4584 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms
);
4586 static inline u64
cpu_resched_latency(struct rq
*rq
) { return 0; }
4587 #endif /* CONFIG_SCHED_DEBUG */
4590 * This function gets called by the timer code, with HZ frequency.
4591 * We call it with interrupts disabled.
4593 void scheduler_tick(void)
4595 int cpu
= smp_processor_id();
4596 struct rq
*rq
= cpu_rq(cpu
);
4597 struct task_struct
*curr
= rq
->curr
;
4599 unsigned long thermal_pressure
;
4600 u64 resched_latency
;
4602 arch_scale_freq_tick();
4607 update_rq_clock(rq
);
4608 thermal_pressure
= arch_scale_thermal_pressure(cpu_of(rq
));
4609 update_thermal_load_avg(rq_clock_thermal(rq
), rq
, thermal_pressure
);
4610 curr
->sched_class
->task_tick(rq
, curr
, 0);
4611 if (sched_feat(LATENCY_WARN
))
4612 resched_latency
= cpu_resched_latency(rq
);
4613 calc_global_load_tick(rq
);
4617 if (sched_feat(LATENCY_WARN
) && resched_latency
)
4618 resched_latency_warn(cpu
, resched_latency
);
4620 perf_event_task_tick();
4623 rq
->idle_balance
= idle_cpu(cpu
);
4624 trigger_load_balance(rq
);
4628 #ifdef CONFIG_NO_HZ_FULL
4633 struct delayed_work work
;
4635 /* Values for ->state, see diagram below. */
4636 #define TICK_SCHED_REMOTE_OFFLINE 0
4637 #define TICK_SCHED_REMOTE_OFFLINING 1
4638 #define TICK_SCHED_REMOTE_RUNNING 2
4641 * State diagram for ->state:
4644 * TICK_SCHED_REMOTE_OFFLINE
4647 * | | sched_tick_remote()
4650 * +--TICK_SCHED_REMOTE_OFFLINING
4653 * sched_tick_start() | | sched_tick_stop()
4656 * TICK_SCHED_REMOTE_RUNNING
4659 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4660 * and sched_tick_start() are happy to leave the state in RUNNING.
4663 static struct tick_work __percpu
*tick_work_cpu
;
4665 static void sched_tick_remote(struct work_struct
*work
)
4667 struct delayed_work
*dwork
= to_delayed_work(work
);
4668 struct tick_work
*twork
= container_of(dwork
, struct tick_work
, work
);
4669 int cpu
= twork
->cpu
;
4670 struct rq
*rq
= cpu_rq(cpu
);
4671 struct task_struct
*curr
;
4677 * Handle the tick only if it appears the remote CPU is running in full
4678 * dynticks mode. The check is racy by nature, but missing a tick or
4679 * having one too much is no big deal because the scheduler tick updates
4680 * statistics and checks timeslices in a time-independent way, regardless
4681 * of when exactly it is running.
4683 if (!tick_nohz_tick_stopped_cpu(cpu
))
4686 rq_lock_irq(rq
, &rf
);
4688 if (cpu_is_offline(cpu
))
4691 update_rq_clock(rq
);
4693 if (!is_idle_task(curr
)) {
4695 * Make sure the next tick runs within a reasonable
4698 delta
= rq_clock_task(rq
) - curr
->se
.exec_start
;
4699 WARN_ON_ONCE(delta
> (u64
)NSEC_PER_SEC
* 3);
4701 curr
->sched_class
->task_tick(rq
, curr
, 0);
4703 calc_load_nohz_remote(rq
);
4705 rq_unlock_irq(rq
, &rf
);
4709 * Run the remote tick once per second (1Hz). This arbitrary
4710 * frequency is large enough to avoid overload but short enough
4711 * to keep scheduler internal stats reasonably up to date. But
4712 * first update state to reflect hotplug activity if required.
4714 os
= atomic_fetch_add_unless(&twork
->state
, -1, TICK_SCHED_REMOTE_RUNNING
);
4715 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_OFFLINE
);
4716 if (os
== TICK_SCHED_REMOTE_RUNNING
)
4717 queue_delayed_work(system_unbound_wq
, dwork
, HZ
);
4720 static void sched_tick_start(int cpu
)
4723 struct tick_work
*twork
;
4725 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
4728 WARN_ON_ONCE(!tick_work_cpu
);
4730 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
4731 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_RUNNING
);
4732 WARN_ON_ONCE(os
== TICK_SCHED_REMOTE_RUNNING
);
4733 if (os
== TICK_SCHED_REMOTE_OFFLINE
) {
4735 INIT_DELAYED_WORK(&twork
->work
, sched_tick_remote
);
4736 queue_delayed_work(system_unbound_wq
, &twork
->work
, HZ
);
4740 #ifdef CONFIG_HOTPLUG_CPU
4741 static void sched_tick_stop(int cpu
)
4743 struct tick_work
*twork
;
4746 if (housekeeping_cpu(cpu
, HK_FLAG_TICK
))
4749 WARN_ON_ONCE(!tick_work_cpu
);
4751 twork
= per_cpu_ptr(tick_work_cpu
, cpu
);
4752 /* There cannot be competing actions, but don't rely on stop-machine. */
4753 os
= atomic_xchg(&twork
->state
, TICK_SCHED_REMOTE_OFFLINING
);
4754 WARN_ON_ONCE(os
!= TICK_SCHED_REMOTE_RUNNING
);
4755 /* Don't cancel, as this would mess up the state machine. */
4757 #endif /* CONFIG_HOTPLUG_CPU */
4759 int __init
sched_tick_offload_init(void)
4761 tick_work_cpu
= alloc_percpu(struct tick_work
);
4762 BUG_ON(!tick_work_cpu
);
4766 #else /* !CONFIG_NO_HZ_FULL */
4767 static inline void sched_tick_start(int cpu
) { }
4768 static inline void sched_tick_stop(int cpu
) { }
4771 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4772 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4774 * If the value passed in is equal to the current preempt count
4775 * then we just disabled preemption. Start timing the latency.
4777 static inline void preempt_latency_start(int val
)
4779 if (preempt_count() == val
) {
4780 unsigned long ip
= get_lock_parent_ip();
4781 #ifdef CONFIG_DEBUG_PREEMPT
4782 current
->preempt_disable_ip
= ip
;
4784 trace_preempt_off(CALLER_ADDR0
, ip
);
4788 void preempt_count_add(int val
)
4790 #ifdef CONFIG_DEBUG_PREEMPT
4794 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4797 __preempt_count_add(val
);
4798 #ifdef CONFIG_DEBUG_PREEMPT
4800 * Spinlock count overflowing soon?
4802 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4805 preempt_latency_start(val
);
4807 EXPORT_SYMBOL(preempt_count_add
);
4808 NOKPROBE_SYMBOL(preempt_count_add
);
4811 * If the value passed in equals to the current preempt count
4812 * then we just enabled preemption. Stop timing the latency.
4814 static inline void preempt_latency_stop(int val
)
4816 if (preempt_count() == val
)
4817 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
4820 void preempt_count_sub(int val
)
4822 #ifdef CONFIG_DEBUG_PREEMPT
4826 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4829 * Is the spinlock portion underflowing?
4831 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4832 !(preempt_count() & PREEMPT_MASK
)))
4836 preempt_latency_stop(val
);
4837 __preempt_count_sub(val
);
4839 EXPORT_SYMBOL(preempt_count_sub
);
4840 NOKPROBE_SYMBOL(preempt_count_sub
);
4843 static inline void preempt_latency_start(int val
) { }
4844 static inline void preempt_latency_stop(int val
) { }
4847 static inline unsigned long get_preempt_disable_ip(struct task_struct
*p
)
4849 #ifdef CONFIG_DEBUG_PREEMPT
4850 return p
->preempt_disable_ip
;
4857 * Print scheduling while atomic bug:
4859 static noinline
void __schedule_bug(struct task_struct
*prev
)
4861 /* Save this before calling printk(), since that will clobber it */
4862 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
4864 if (oops_in_progress
)
4867 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4868 prev
->comm
, prev
->pid
, preempt_count());
4870 debug_show_held_locks(prev
);
4872 if (irqs_disabled())
4873 print_irqtrace_events(prev
);
4874 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
4875 && in_atomic_preempt_off()) {
4876 pr_err("Preemption disabled at:");
4877 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
4880 panic("scheduling while atomic\n");
4883 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
4887 * Various schedule()-time debugging checks and statistics:
4889 static inline void schedule_debug(struct task_struct
*prev
, bool preempt
)
4891 #ifdef CONFIG_SCHED_STACK_END_CHECK
4892 if (task_stack_end_corrupted(prev
))
4893 panic("corrupted stack end detected inside scheduler\n");
4895 if (task_scs_end_corrupted(prev
))
4896 panic("corrupted shadow stack detected inside scheduler\n");
4899 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4900 if (!preempt
&& prev
->state
&& prev
->non_block_count
) {
4901 printk(KERN_ERR
"BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4902 prev
->comm
, prev
->pid
, prev
->non_block_count
);
4904 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
4908 if (unlikely(in_atomic_preempt_off())) {
4909 __schedule_bug(prev
);
4910 preempt_count_set(PREEMPT_DISABLED
);
4913 SCHED_WARN_ON(ct_state() == CONTEXT_USER
);
4915 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4917 schedstat_inc(this_rq()->sched_count
);
4920 static void put_prev_task_balance(struct rq
*rq
, struct task_struct
*prev
,
4921 struct rq_flags
*rf
)
4924 const struct sched_class
*class;
4926 * We must do the balancing pass before put_prev_task(), such
4927 * that when we release the rq->lock the task is in the same
4928 * state as before we took rq->lock.
4930 * We can terminate the balance pass as soon as we know there is
4931 * a runnable task of @class priority or higher.
4933 for_class_range(class, prev
->sched_class
, &idle_sched_class
) {
4934 if (class->balance(rq
, prev
, rf
))
4939 put_prev_task(rq
, prev
);
4943 * Pick up the highest-prio task:
4945 static inline struct task_struct
*
4946 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
4948 const struct sched_class
*class;
4949 struct task_struct
*p
;
4952 * Optimization: we know that if all tasks are in the fair class we can
4953 * call that function directly, but only if the @prev task wasn't of a
4954 * higher scheduling class, because otherwise those lose the
4955 * opportunity to pull in more work from other CPUs.
4957 if (likely(prev
->sched_class
<= &fair_sched_class
&&
4958 rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
4960 p
= pick_next_task_fair(rq
, prev
, rf
);
4961 if (unlikely(p
== RETRY_TASK
))
4964 /* Assumes fair_sched_class->next == idle_sched_class */
4966 put_prev_task(rq
, prev
);
4967 p
= pick_next_task_idle(rq
);
4974 put_prev_task_balance(rq
, prev
, rf
);
4976 for_each_class(class) {
4977 p
= class->pick_next_task(rq
);
4982 /* The idle class should always have a runnable task: */
4987 * __schedule() is the main scheduler function.
4989 * The main means of driving the scheduler and thus entering this function are:
4991 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4993 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4994 * paths. For example, see arch/x86/entry_64.S.
4996 * To drive preemption between tasks, the scheduler sets the flag in timer
4997 * interrupt handler scheduler_tick().
4999 * 3. Wakeups don't really cause entry into schedule(). They add a
5000 * task to the run-queue and that's it.
5002 * Now, if the new task added to the run-queue preempts the current
5003 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
5004 * called on the nearest possible occasion:
5006 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
5008 * - in syscall or exception context, at the next outmost
5009 * preempt_enable(). (this might be as soon as the wake_up()'s
5012 * - in IRQ context, return from interrupt-handler to
5013 * preemptible context
5015 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
5018 * - cond_resched() call
5019 * - explicit schedule() call
5020 * - return from syscall or exception to user-space
5021 * - return from interrupt-handler to user-space
5023 * WARNING: must be called with preemption disabled!
5025 static void __sched notrace
__schedule(bool preempt
)
5027 struct task_struct
*prev
, *next
;
5028 unsigned long *switch_count
;
5029 unsigned long prev_state
;
5034 cpu
= smp_processor_id();
5038 schedule_debug(prev
, preempt
);
5040 if (sched_feat(HRTICK
) || sched_feat(HRTICK_DL
))
5043 local_irq_disable();
5044 rcu_note_context_switch(preempt
);
5047 * Make sure that signal_pending_state()->signal_pending() below
5048 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
5049 * done by the caller to avoid the race with signal_wake_up():
5051 * __set_current_state(@state) signal_wake_up()
5052 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
5053 * wake_up_state(p, state)
5054 * LOCK rq->lock LOCK p->pi_state
5055 * smp_mb__after_spinlock() smp_mb__after_spinlock()
5056 * if (signal_pending_state()) if (p->state & @state)
5058 * Also, the membarrier system call requires a full memory barrier
5059 * after coming from user-space, before storing to rq->curr.
5062 smp_mb__after_spinlock();
5064 /* Promote REQ to ACT */
5065 rq
->clock_update_flags
<<= 1;
5066 update_rq_clock(rq
);
5068 switch_count
= &prev
->nivcsw
;
5071 * We must load prev->state once (task_struct::state is volatile), such
5074 * - we form a control dependency vs deactivate_task() below.
5075 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5077 prev_state
= prev
->state
;
5078 if (!preempt
&& prev_state
) {
5079 if (signal_pending_state(prev_state
, prev
)) {
5080 prev
->state
= TASK_RUNNING
;
5082 prev
->sched_contributes_to_load
=
5083 (prev_state
& TASK_UNINTERRUPTIBLE
) &&
5084 !(prev_state
& TASK_NOLOAD
) &&
5085 !(prev
->flags
& PF_FROZEN
);
5087 if (prev
->sched_contributes_to_load
)
5088 rq
->nr_uninterruptible
++;
5091 * __schedule() ttwu()
5092 * prev_state = prev->state; if (p->on_rq && ...)
5093 * if (prev_state) goto out;
5094 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5095 * p->state = TASK_WAKING
5097 * Where __schedule() and ttwu() have matching control dependencies.
5099 * After this, schedule() must not care about p->state any more.
5101 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
| DEQUEUE_NOCLOCK
);
5103 if (prev
->in_iowait
) {
5104 atomic_inc(&rq
->nr_iowait
);
5105 delayacct_blkio_start();
5108 switch_count
= &prev
->nvcsw
;
5111 next
= pick_next_task(rq
, prev
, &rf
);
5112 clear_tsk_need_resched(prev
);
5113 clear_preempt_need_resched();
5114 #ifdef CONFIG_SCHED_DEBUG
5115 rq
->last_seen_need_resched_ns
= 0;
5118 if (likely(prev
!= next
)) {
5121 * RCU users of rcu_dereference(rq->curr) may not see
5122 * changes to task_struct made by pick_next_task().
5124 RCU_INIT_POINTER(rq
->curr
, next
);
5126 * The membarrier system call requires each architecture
5127 * to have a full memory barrier after updating
5128 * rq->curr, before returning to user-space.
5130 * Here are the schemes providing that barrier on the
5131 * various architectures:
5132 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5133 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5134 * - finish_lock_switch() for weakly-ordered
5135 * architectures where spin_unlock is a full barrier,
5136 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5137 * is a RELEASE barrier),
5141 migrate_disable_switch(rq
, prev
);
5142 psi_sched_switch(prev
, next
, !task_on_rq_queued(prev
));
5144 trace_sched_switch(preempt
, prev
, next
);
5146 /* Also unlocks the rq: */
5147 rq
= context_switch(rq
, prev
, next
, &rf
);
5149 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
5151 rq_unpin_lock(rq
, &rf
);
5152 __balance_callbacks(rq
);
5153 raw_spin_unlock_irq(&rq
->lock
);
5157 void __noreturn
do_task_dead(void)
5159 /* Causes final put_task_struct in finish_task_switch(): */
5160 set_special_state(TASK_DEAD
);
5162 /* Tell freezer to ignore us: */
5163 current
->flags
|= PF_NOFREEZE
;
5168 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5173 static inline void sched_submit_work(struct task_struct
*tsk
)
5175 unsigned int task_flags
;
5180 task_flags
= tsk
->flags
;
5182 * If a worker went to sleep, notify and ask workqueue whether
5183 * it wants to wake up a task to maintain concurrency.
5184 * As this function is called inside the schedule() context,
5185 * we disable preemption to avoid it calling schedule() again
5186 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5189 if (task_flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
5191 if (task_flags
& PF_WQ_WORKER
)
5192 wq_worker_sleeping(tsk
);
5194 io_wq_worker_sleeping(tsk
);
5195 preempt_enable_no_resched();
5198 if (tsk_is_pi_blocked(tsk
))
5202 * If we are going to sleep and we have plugged IO queued,
5203 * make sure to submit it to avoid deadlocks.
5205 if (blk_needs_flush_plug(tsk
))
5206 blk_schedule_flush_plug(tsk
);
5209 static void sched_update_worker(struct task_struct
*tsk
)
5211 if (tsk
->flags
& (PF_WQ_WORKER
| PF_IO_WORKER
)) {
5212 if (tsk
->flags
& PF_WQ_WORKER
)
5213 wq_worker_running(tsk
);
5215 io_wq_worker_running(tsk
);
5219 asmlinkage __visible
void __sched
schedule(void)
5221 struct task_struct
*tsk
= current
;
5223 sched_submit_work(tsk
);
5227 sched_preempt_enable_no_resched();
5228 } while (need_resched());
5229 sched_update_worker(tsk
);
5231 EXPORT_SYMBOL(schedule
);
5234 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
5235 * state (have scheduled out non-voluntarily) by making sure that all
5236 * tasks have either left the run queue or have gone into user space.
5237 * As idle tasks do not do either, they must not ever be preempted
5238 * (schedule out non-voluntarily).
5240 * schedule_idle() is similar to schedule_preempt_disable() except that it
5241 * never enables preemption because it does not call sched_submit_work().
5243 void __sched
schedule_idle(void)
5246 * As this skips calling sched_submit_work(), which the idle task does
5247 * regardless because that function is a nop when the task is in a
5248 * TASK_RUNNING state, make sure this isn't used someplace that the
5249 * current task can be in any other state. Note, idle is always in the
5250 * TASK_RUNNING state.
5252 WARN_ON_ONCE(current
->state
);
5255 } while (need_resched());
5258 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
5259 asmlinkage __visible
void __sched
schedule_user(void)
5262 * If we come here after a random call to set_need_resched(),
5263 * or we have been woken up remotely but the IPI has not yet arrived,
5264 * we haven't yet exited the RCU idle mode. Do it here manually until
5265 * we find a better solution.
5267 * NB: There are buggy callers of this function. Ideally we
5268 * should warn if prev_state != CONTEXT_USER, but that will trigger
5269 * too frequently to make sense yet.
5271 enum ctx_state prev_state
= exception_enter();
5273 exception_exit(prev_state
);
5278 * schedule_preempt_disabled - called with preemption disabled
5280 * Returns with preemption disabled. Note: preempt_count must be 1
5282 void __sched
schedule_preempt_disabled(void)
5284 sched_preempt_enable_no_resched();
5289 static void __sched notrace
preempt_schedule_common(void)
5293 * Because the function tracer can trace preempt_count_sub()
5294 * and it also uses preempt_enable/disable_notrace(), if
5295 * NEED_RESCHED is set, the preempt_enable_notrace() called
5296 * by the function tracer will call this function again and
5297 * cause infinite recursion.
5299 * Preemption must be disabled here before the function
5300 * tracer can trace. Break up preempt_disable() into two
5301 * calls. One to disable preemption without fear of being
5302 * traced. The other to still record the preemption latency,
5303 * which can also be traced by the function tracer.
5305 preempt_disable_notrace();
5306 preempt_latency_start(1);
5308 preempt_latency_stop(1);
5309 preempt_enable_no_resched_notrace();
5312 * Check again in case we missed a preemption opportunity
5313 * between schedule and now.
5315 } while (need_resched());
5318 #ifdef CONFIG_PREEMPTION
5320 * This is the entry point to schedule() from in-kernel preemption
5321 * off of preempt_enable.
5323 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
5326 * If there is a non-zero preempt_count or interrupts are disabled,
5327 * we do not want to preempt the current task. Just return..
5329 if (likely(!preemptible()))
5332 preempt_schedule_common();
5334 NOKPROBE_SYMBOL(preempt_schedule
);
5335 EXPORT_SYMBOL(preempt_schedule
);
5337 #ifdef CONFIG_PREEMPT_DYNAMIC
5338 DEFINE_STATIC_CALL(preempt_schedule
, __preempt_schedule_func
);
5339 EXPORT_STATIC_CALL_TRAMP(preempt_schedule
);
5344 * preempt_schedule_notrace - preempt_schedule called by tracing
5346 * The tracing infrastructure uses preempt_enable_notrace to prevent
5347 * recursion and tracing preempt enabling caused by the tracing
5348 * infrastructure itself. But as tracing can happen in areas coming
5349 * from userspace or just about to enter userspace, a preempt enable
5350 * can occur before user_exit() is called. This will cause the scheduler
5351 * to be called when the system is still in usermode.
5353 * To prevent this, the preempt_enable_notrace will use this function
5354 * instead of preempt_schedule() to exit user context if needed before
5355 * calling the scheduler.
5357 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
5359 enum ctx_state prev_ctx
;
5361 if (likely(!preemptible()))
5366 * Because the function tracer can trace preempt_count_sub()
5367 * and it also uses preempt_enable/disable_notrace(), if
5368 * NEED_RESCHED is set, the preempt_enable_notrace() called
5369 * by the function tracer will call this function again and
5370 * cause infinite recursion.
5372 * Preemption must be disabled here before the function
5373 * tracer can trace. Break up preempt_disable() into two
5374 * calls. One to disable preemption without fear of being
5375 * traced. The other to still record the preemption latency,
5376 * which can also be traced by the function tracer.
5378 preempt_disable_notrace();
5379 preempt_latency_start(1);
5381 * Needs preempt disabled in case user_exit() is traced
5382 * and the tracer calls preempt_enable_notrace() causing
5383 * an infinite recursion.
5385 prev_ctx
= exception_enter();
5387 exception_exit(prev_ctx
);
5389 preempt_latency_stop(1);
5390 preempt_enable_no_resched_notrace();
5391 } while (need_resched());
5393 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
5395 #ifdef CONFIG_PREEMPT_DYNAMIC
5396 DEFINE_STATIC_CALL(preempt_schedule_notrace
, __preempt_schedule_notrace_func
);
5397 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace
);
5400 #endif /* CONFIG_PREEMPTION */
5402 #ifdef CONFIG_PREEMPT_DYNAMIC
5404 #include <linux/entry-common.h>
5409 * SC:preempt_schedule
5410 * SC:preempt_schedule_notrace
5411 * SC:irqentry_exit_cond_resched
5415 * cond_resched <- __cond_resched
5416 * might_resched <- RET0
5417 * preempt_schedule <- NOP
5418 * preempt_schedule_notrace <- NOP
5419 * irqentry_exit_cond_resched <- NOP
5422 * cond_resched <- __cond_resched
5423 * might_resched <- __cond_resched
5424 * preempt_schedule <- NOP
5425 * preempt_schedule_notrace <- NOP
5426 * irqentry_exit_cond_resched <- NOP
5429 * cond_resched <- RET0
5430 * might_resched <- RET0
5431 * preempt_schedule <- preempt_schedule
5432 * preempt_schedule_notrace <- preempt_schedule_notrace
5433 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
5437 preempt_dynamic_none
= 0,
5438 preempt_dynamic_voluntary
,
5439 preempt_dynamic_full
,
5442 int preempt_dynamic_mode
= preempt_dynamic_full
;
5444 int sched_dynamic_mode(const char *str
)
5446 if (!strcmp(str
, "none"))
5447 return preempt_dynamic_none
;
5449 if (!strcmp(str
, "voluntary"))
5450 return preempt_dynamic_voluntary
;
5452 if (!strcmp(str
, "full"))
5453 return preempt_dynamic_full
;
5458 void sched_dynamic_update(int mode
)
5461 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
5462 * the ZERO state, which is invalid.
5464 static_call_update(cond_resched
, __cond_resched
);
5465 static_call_update(might_resched
, __cond_resched
);
5466 static_call_update(preempt_schedule
, __preempt_schedule_func
);
5467 static_call_update(preempt_schedule_notrace
, __preempt_schedule_notrace_func
);
5468 static_call_update(irqentry_exit_cond_resched
, irqentry_exit_cond_resched
);
5471 case preempt_dynamic_none
:
5472 static_call_update(cond_resched
, __cond_resched
);
5473 static_call_update(might_resched
, (void *)&__static_call_return0
);
5474 static_call_update(preempt_schedule
, NULL
);
5475 static_call_update(preempt_schedule_notrace
, NULL
);
5476 static_call_update(irqentry_exit_cond_resched
, NULL
);
5477 pr_info("Dynamic Preempt: none\n");
5480 case preempt_dynamic_voluntary
:
5481 static_call_update(cond_resched
, __cond_resched
);
5482 static_call_update(might_resched
, __cond_resched
);
5483 static_call_update(preempt_schedule
, NULL
);
5484 static_call_update(preempt_schedule_notrace
, NULL
);
5485 static_call_update(irqentry_exit_cond_resched
, NULL
);
5486 pr_info("Dynamic Preempt: voluntary\n");
5489 case preempt_dynamic_full
:
5490 static_call_update(cond_resched
, (void *)&__static_call_return0
);
5491 static_call_update(might_resched
, (void *)&__static_call_return0
);
5492 static_call_update(preempt_schedule
, __preempt_schedule_func
);
5493 static_call_update(preempt_schedule_notrace
, __preempt_schedule_notrace_func
);
5494 static_call_update(irqentry_exit_cond_resched
, irqentry_exit_cond_resched
);
5495 pr_info("Dynamic Preempt: full\n");
5499 preempt_dynamic_mode
= mode
;
5502 static int __init
setup_preempt_mode(char *str
)
5504 int mode
= sched_dynamic_mode(str
);
5506 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str
);
5510 sched_dynamic_update(mode
);
5513 __setup("preempt=", setup_preempt_mode
);
5515 #endif /* CONFIG_PREEMPT_DYNAMIC */
5518 * This is the entry point to schedule() from kernel preemption
5519 * off of irq context.
5520 * Note, that this is called and return with irqs disabled. This will
5521 * protect us against recursive calling from irq.
5523 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
5525 enum ctx_state prev_state
;
5527 /* Catch callers which need to be fixed */
5528 BUG_ON(preempt_count() || !irqs_disabled());
5530 prev_state
= exception_enter();
5536 local_irq_disable();
5537 sched_preempt_enable_no_resched();
5538 } while (need_resched());
5540 exception_exit(prev_state
);
5543 int default_wake_function(wait_queue_entry_t
*curr
, unsigned mode
, int wake_flags
,
5546 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG
) && wake_flags
& ~WF_SYNC
);
5547 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5549 EXPORT_SYMBOL(default_wake_function
);
5551 #ifdef CONFIG_RT_MUTEXES
5553 static inline int __rt_effective_prio(struct task_struct
*pi_task
, int prio
)
5556 prio
= min(prio
, pi_task
->prio
);
5561 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
5563 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
5565 return __rt_effective_prio(pi_task
, prio
);
5569 * rt_mutex_setprio - set the current priority of a task
5571 * @pi_task: donor task
5573 * This function changes the 'effective' priority of a task. It does
5574 * not touch ->normal_prio like __setscheduler().
5576 * Used by the rt_mutex code to implement priority inheritance
5577 * logic. Call site only calls if the priority of the task changed.
5579 void rt_mutex_setprio(struct task_struct
*p
, struct task_struct
*pi_task
)
5581 int prio
, oldprio
, queued
, running
, queue_flag
=
5582 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
5583 const struct sched_class
*prev_class
;
5587 /* XXX used to be waiter->prio, not waiter->task->prio */
5588 prio
= __rt_effective_prio(pi_task
, p
->normal_prio
);
5591 * If nothing changed; bail early.
5593 if (p
->pi_top_task
== pi_task
&& prio
== p
->prio
&& !dl_prio(prio
))
5596 rq
= __task_rq_lock(p
, &rf
);
5597 update_rq_clock(rq
);
5599 * Set under pi_lock && rq->lock, such that the value can be used under
5602 * Note that there is loads of tricky to make this pointer cache work
5603 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
5604 * ensure a task is de-boosted (pi_task is set to NULL) before the
5605 * task is allowed to run again (and can exit). This ensures the pointer
5606 * points to a blocked task -- which guarantees the task is present.
5608 p
->pi_top_task
= pi_task
;
5611 * For FIFO/RR we only need to set prio, if that matches we're done.
5613 if (prio
== p
->prio
&& !dl_prio(prio
))
5617 * Idle task boosting is a nono in general. There is one
5618 * exception, when PREEMPT_RT and NOHZ is active:
5620 * The idle task calls get_next_timer_interrupt() and holds
5621 * the timer wheel base->lock on the CPU and another CPU wants
5622 * to access the timer (probably to cancel it). We can safely
5623 * ignore the boosting request, as the idle CPU runs this code
5624 * with interrupts disabled and will complete the lock
5625 * protected section without being interrupted. So there is no
5626 * real need to boost.
5628 if (unlikely(p
== rq
->idle
)) {
5629 WARN_ON(p
!= rq
->curr
);
5630 WARN_ON(p
->pi_blocked_on
);
5634 trace_sched_pi_setprio(p
, pi_task
);
5637 if (oldprio
== prio
)
5638 queue_flag
&= ~DEQUEUE_MOVE
;
5640 prev_class
= p
->sched_class
;
5641 queued
= task_on_rq_queued(p
);
5642 running
= task_current(rq
, p
);
5644 dequeue_task(rq
, p
, queue_flag
);
5646 put_prev_task(rq
, p
);
5649 * Boosting condition are:
5650 * 1. -rt task is running and holds mutex A
5651 * --> -dl task blocks on mutex A
5653 * 2. -dl task is running and holds mutex A
5654 * --> -dl task blocks on mutex A and could preempt the
5657 if (dl_prio(prio
)) {
5658 if (!dl_prio(p
->normal_prio
) ||
5659 (pi_task
&& dl_prio(pi_task
->prio
) &&
5660 dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
5661 p
->dl
.pi_se
= pi_task
->dl
.pi_se
;
5662 queue_flag
|= ENQUEUE_REPLENISH
;
5664 p
->dl
.pi_se
= &p
->dl
;
5666 p
->sched_class
= &dl_sched_class
;
5667 } else if (rt_prio(prio
)) {
5668 if (dl_prio(oldprio
))
5669 p
->dl
.pi_se
= &p
->dl
;
5671 queue_flag
|= ENQUEUE_HEAD
;
5672 p
->sched_class
= &rt_sched_class
;
5674 if (dl_prio(oldprio
))
5675 p
->dl
.pi_se
= &p
->dl
;
5676 if (rt_prio(oldprio
))
5678 p
->sched_class
= &fair_sched_class
;
5684 enqueue_task(rq
, p
, queue_flag
);
5686 set_next_task(rq
, p
);
5688 check_class_changed(rq
, p
, prev_class
, oldprio
);
5690 /* Avoid rq from going away on us: */
5693 rq_unpin_lock(rq
, &rf
);
5694 __balance_callbacks(rq
);
5695 raw_spin_unlock(&rq
->lock
);
5700 static inline int rt_effective_prio(struct task_struct
*p
, int prio
)
5706 void set_user_nice(struct task_struct
*p
, long nice
)
5708 bool queued
, running
;
5713 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
5716 * We have to be careful, if called from sys_setpriority(),
5717 * the task might be in the middle of scheduling on another CPU.
5719 rq
= task_rq_lock(p
, &rf
);
5720 update_rq_clock(rq
);
5723 * The RT priorities are set via sched_setscheduler(), but we still
5724 * allow the 'normal' nice value to be set - but as expected
5725 * it won't have any effect on scheduling until the task is
5726 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5728 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
5729 p
->static_prio
= NICE_TO_PRIO(nice
);
5732 queued
= task_on_rq_queued(p
);
5733 running
= task_current(rq
, p
);
5735 dequeue_task(rq
, p
, DEQUEUE_SAVE
| DEQUEUE_NOCLOCK
);
5737 put_prev_task(rq
, p
);
5739 p
->static_prio
= NICE_TO_PRIO(nice
);
5740 set_load_weight(p
, true);
5742 p
->prio
= effective_prio(p
);
5745 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
5747 set_next_task(rq
, p
);
5750 * If the task increased its priority or is running and
5751 * lowered its priority, then reschedule its CPU:
5753 p
->sched_class
->prio_changed(rq
, p
, old_prio
);
5756 task_rq_unlock(rq
, p
, &rf
);
5758 EXPORT_SYMBOL(set_user_nice
);
5761 * can_nice - check if a task can reduce its nice value
5765 int can_nice(const struct task_struct
*p
, const int nice
)
5767 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5768 int nice_rlim
= nice_to_rlimit(nice
);
5770 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
5771 capable(CAP_SYS_NICE
));
5774 #ifdef __ARCH_WANT_SYS_NICE
5777 * sys_nice - change the priority of the current process.
5778 * @increment: priority increment
5780 * sys_setpriority is a more generic, but much slower function that
5781 * does similar things.
5783 SYSCALL_DEFINE1(nice
, int, increment
)
5788 * Setpriority might change our priority at the same moment.
5789 * We don't have to worry. Conceptually one call occurs first
5790 * and we have a single winner.
5792 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
5793 nice
= task_nice(current
) + increment
;
5795 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
5796 if (increment
< 0 && !can_nice(current
, nice
))
5799 retval
= security_task_setnice(current
, nice
);
5803 set_user_nice(current
, nice
);
5810 * task_prio - return the priority value of a given task.
5811 * @p: the task in question.
5813 * Return: The priority value as seen by users in /proc.
5815 * sched policy return value kernel prio user prio/nice
5817 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
5818 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
5819 * deadline -101 -1 0
5821 int task_prio(const struct task_struct
*p
)
5823 return p
->prio
- MAX_RT_PRIO
;
5827 * idle_cpu - is a given CPU idle currently?
5828 * @cpu: the processor in question.
5830 * Return: 1 if the CPU is currently idle. 0 otherwise.
5832 int idle_cpu(int cpu
)
5834 struct rq
*rq
= cpu_rq(cpu
);
5836 if (rq
->curr
!= rq
->idle
)
5843 if (rq
->ttwu_pending
)
5851 * available_idle_cpu - is a given CPU idle for enqueuing work.
5852 * @cpu: the CPU in question.
5854 * Return: 1 if the CPU is currently idle. 0 otherwise.
5856 int available_idle_cpu(int cpu
)
5861 if (vcpu_is_preempted(cpu
))
5868 * idle_task - return the idle task for a given CPU.
5869 * @cpu: the processor in question.
5871 * Return: The idle task for the CPU @cpu.
5873 struct task_struct
*idle_task(int cpu
)
5875 return cpu_rq(cpu
)->idle
;
5880 * This function computes an effective utilization for the given CPU, to be
5881 * used for frequency selection given the linear relation: f = u * f_max.
5883 * The scheduler tracks the following metrics:
5885 * cpu_util_{cfs,rt,dl,irq}()
5888 * Where the cfs,rt and dl util numbers are tracked with the same metric and
5889 * synchronized windows and are thus directly comparable.
5891 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
5892 * which excludes things like IRQ and steal-time. These latter are then accrued
5893 * in the irq utilization.
5895 * The DL bandwidth number otoh is not a measured metric but a value computed
5896 * based on the task model parameters and gives the minimal utilization
5897 * required to meet deadlines.
5899 unsigned long effective_cpu_util(int cpu
, unsigned long util_cfs
,
5900 unsigned long max
, enum cpu_util_type type
,
5901 struct task_struct
*p
)
5903 unsigned long dl_util
, util
, irq
;
5904 struct rq
*rq
= cpu_rq(cpu
);
5906 if (!uclamp_is_used() &&
5907 type
== FREQUENCY_UTIL
&& rt_rq_is_runnable(&rq
->rt
)) {
5912 * Early check to see if IRQ/steal time saturates the CPU, can be
5913 * because of inaccuracies in how we track these -- see
5914 * update_irq_load_avg().
5916 irq
= cpu_util_irq(rq
);
5917 if (unlikely(irq
>= max
))
5921 * Because the time spend on RT/DL tasks is visible as 'lost' time to
5922 * CFS tasks and we use the same metric to track the effective
5923 * utilization (PELT windows are synchronized) we can directly add them
5924 * to obtain the CPU's actual utilization.
5926 * CFS and RT utilization can be boosted or capped, depending on
5927 * utilization clamp constraints requested by currently RUNNABLE
5929 * When there are no CFS RUNNABLE tasks, clamps are released and
5930 * frequency will be gracefully reduced with the utilization decay.
5932 util
= util_cfs
+ cpu_util_rt(rq
);
5933 if (type
== FREQUENCY_UTIL
)
5934 util
= uclamp_rq_util_with(rq
, util
, p
);
5936 dl_util
= cpu_util_dl(rq
);
5939 * For frequency selection we do not make cpu_util_dl() a permanent part
5940 * of this sum because we want to use cpu_bw_dl() later on, but we need
5941 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
5942 * that we select f_max when there is no idle time.
5944 * NOTE: numerical errors or stop class might cause us to not quite hit
5945 * saturation when we should -- something for later.
5947 if (util
+ dl_util
>= max
)
5951 * OTOH, for energy computation we need the estimated running time, so
5952 * include util_dl and ignore dl_bw.
5954 if (type
== ENERGY_UTIL
)
5958 * There is still idle time; further improve the number by using the
5959 * irq metric. Because IRQ/steal time is hidden from the task clock we
5960 * need to scale the task numbers:
5963 * U' = irq + --------- * U
5966 util
= scale_irq_capacity(util
, irq
, max
);
5970 * Bandwidth required by DEADLINE must always be granted while, for
5971 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
5972 * to gracefully reduce the frequency when no tasks show up for longer
5975 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
5976 * bw_dl as requested freq. However, cpufreq is not yet ready for such
5977 * an interface. So, we only do the latter for now.
5979 if (type
== FREQUENCY_UTIL
)
5980 util
+= cpu_bw_dl(rq
);
5982 return min(max
, util
);
5985 unsigned long sched_cpu_util(int cpu
, unsigned long max
)
5987 return effective_cpu_util(cpu
, cpu_util_cfs(cpu_rq(cpu
)), max
,
5990 #endif /* CONFIG_SMP */
5993 * find_process_by_pid - find a process with a matching PID value.
5994 * @pid: the pid in question.
5996 * The task of @pid, if found. %NULL otherwise.
5998 static struct task_struct
*find_process_by_pid(pid_t pid
)
6000 return pid
? find_task_by_vpid(pid
) : current
;
6004 * sched_setparam() passes in -1 for its policy, to let the functions
6005 * it calls know not to change it.
6007 #define SETPARAM_POLICY -1
6009 static void __setscheduler_params(struct task_struct
*p
,
6010 const struct sched_attr
*attr
)
6012 int policy
= attr
->sched_policy
;
6014 if (policy
== SETPARAM_POLICY
)
6019 if (dl_policy(policy
))
6020 __setparam_dl(p
, attr
);
6021 else if (fair_policy(policy
))
6022 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
6025 * __sched_setscheduler() ensures attr->sched_priority == 0 when
6026 * !rt_policy. Always setting this ensures that things like
6027 * getparam()/getattr() don't report silly values for !rt tasks.
6029 p
->rt_priority
= attr
->sched_priority
;
6030 p
->normal_prio
= normal_prio(p
);
6031 set_load_weight(p
, true);
6034 /* Actually do priority change: must hold pi & rq lock. */
6035 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
6036 const struct sched_attr
*attr
, bool keep_boost
)
6039 * If params can't change scheduling class changes aren't allowed
6042 if (attr
->sched_flags
& SCHED_FLAG_KEEP_PARAMS
)
6045 __setscheduler_params(p
, attr
);
6048 * Keep a potential priority boosting if called from
6049 * sched_setscheduler().
6051 p
->prio
= normal_prio(p
);
6053 p
->prio
= rt_effective_prio(p
, p
->prio
);
6055 if (dl_prio(p
->prio
))
6056 p
->sched_class
= &dl_sched_class
;
6057 else if (rt_prio(p
->prio
))
6058 p
->sched_class
= &rt_sched_class
;
6060 p
->sched_class
= &fair_sched_class
;
6064 * Check the target process has a UID that matches the current process's:
6066 static bool check_same_owner(struct task_struct
*p
)
6068 const struct cred
*cred
= current_cred(), *pcred
;
6072 pcred
= __task_cred(p
);
6073 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
6074 uid_eq(cred
->euid
, pcred
->uid
));
6079 static int __sched_setscheduler(struct task_struct
*p
,
6080 const struct sched_attr
*attr
,
6083 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
6084 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
6085 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
6086 int new_effective_prio
, policy
= attr
->sched_policy
;
6087 const struct sched_class
*prev_class
;
6088 struct callback_head
*head
;
6091 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
6094 /* The pi code expects interrupts enabled */
6095 BUG_ON(pi
&& in_interrupt());
6097 /* Double check policy once rq lock held: */
6099 reset_on_fork
= p
->sched_reset_on_fork
;
6100 policy
= oldpolicy
= p
->policy
;
6102 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
6104 if (!valid_policy(policy
))
6108 if (attr
->sched_flags
& ~(SCHED_FLAG_ALL
| SCHED_FLAG_SUGOV
))
6112 * Valid priorities for SCHED_FIFO and SCHED_RR are
6113 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6114 * SCHED_BATCH and SCHED_IDLE is 0.
6116 if (attr
->sched_priority
> MAX_RT_PRIO
-1)
6118 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
6119 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
6123 * Allow unprivileged RT tasks to decrease priority:
6125 if (user
&& !capable(CAP_SYS_NICE
)) {
6126 if (fair_policy(policy
)) {
6127 if (attr
->sched_nice
< task_nice(p
) &&
6128 !can_nice(p
, attr
->sched_nice
))
6132 if (rt_policy(policy
)) {
6133 unsigned long rlim_rtprio
=
6134 task_rlimit(p
, RLIMIT_RTPRIO
);
6136 /* Can't set/change the rt policy: */
6137 if (policy
!= p
->policy
&& !rlim_rtprio
)
6140 /* Can't increase priority: */
6141 if (attr
->sched_priority
> p
->rt_priority
&&
6142 attr
->sched_priority
> rlim_rtprio
)
6147 * Can't set/change SCHED_DEADLINE policy at all for now
6148 * (safest behavior); in the future we would like to allow
6149 * unprivileged DL tasks to increase their relative deadline
6150 * or reduce their runtime (both ways reducing utilization)
6152 if (dl_policy(policy
))
6156 * Treat SCHED_IDLE as nice 20. Only allow a switch to
6157 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6159 if (task_has_idle_policy(p
) && !idle_policy(policy
)) {
6160 if (!can_nice(p
, task_nice(p
)))
6164 /* Can't change other user's priorities: */
6165 if (!check_same_owner(p
))
6168 /* Normal users shall not reset the sched_reset_on_fork flag: */
6169 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6174 if (attr
->sched_flags
& SCHED_FLAG_SUGOV
)
6177 retval
= security_task_setscheduler(p
);
6182 /* Update task specific "requested" clamps */
6183 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) {
6184 retval
= uclamp_validate(p
, attr
);
6193 * Make sure no PI-waiters arrive (or leave) while we are
6194 * changing the priority of the task:
6196 * To be able to change p->policy safely, the appropriate
6197 * runqueue lock must be held.
6199 rq
= task_rq_lock(p
, &rf
);
6200 update_rq_clock(rq
);
6203 * Changing the policy of the stop threads its a very bad idea:
6205 if (p
== rq
->stop
) {
6211 * If not changing anything there's no need to proceed further,
6212 * but store a possible modification of reset_on_fork.
6214 if (unlikely(policy
== p
->policy
)) {
6215 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
6217 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
6219 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
6221 if (attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
)
6224 p
->sched_reset_on_fork
= reset_on_fork
;
6231 #ifdef CONFIG_RT_GROUP_SCHED
6233 * Do not allow realtime tasks into groups that have no runtime
6236 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6237 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
6238 !task_group_is_autogroup(task_group(p
))) {
6244 if (dl_bandwidth_enabled() && dl_policy(policy
) &&
6245 !(attr
->sched_flags
& SCHED_FLAG_SUGOV
)) {
6246 cpumask_t
*span
= rq
->rd
->span
;
6249 * Don't allow tasks with an affinity mask smaller than
6250 * the entire root_domain to become SCHED_DEADLINE. We
6251 * will also fail if there's no bandwidth available.
6253 if (!cpumask_subset(span
, p
->cpus_ptr
) ||
6254 rq
->rd
->dl_bw
.bw
== 0) {
6262 /* Re-check policy now with rq lock held: */
6263 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6264 policy
= oldpolicy
= -1;
6265 task_rq_unlock(rq
, p
, &rf
);
6267 cpuset_read_unlock();
6272 * If setscheduling to SCHED_DEADLINE (or changing the parameters
6273 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
6276 if ((dl_policy(policy
) || dl_task(p
)) && sched_dl_overflow(p
, policy
, attr
)) {
6281 p
->sched_reset_on_fork
= reset_on_fork
;
6286 * Take priority boosted tasks into account. If the new
6287 * effective priority is unchanged, we just store the new
6288 * normal parameters and do not touch the scheduler class and
6289 * the runqueue. This will be done when the task deboost
6292 new_effective_prio
= rt_effective_prio(p
, newprio
);
6293 if (new_effective_prio
== oldprio
)
6294 queue_flags
&= ~DEQUEUE_MOVE
;
6297 queued
= task_on_rq_queued(p
);
6298 running
= task_current(rq
, p
);
6300 dequeue_task(rq
, p
, queue_flags
);
6302 put_prev_task(rq
, p
);
6304 prev_class
= p
->sched_class
;
6306 __setscheduler(rq
, p
, attr
, pi
);
6307 __setscheduler_uclamp(p
, attr
);
6311 * We enqueue to tail when the priority of a task is
6312 * increased (user space view).
6314 if (oldprio
< p
->prio
)
6315 queue_flags
|= ENQUEUE_HEAD
;
6317 enqueue_task(rq
, p
, queue_flags
);
6320 set_next_task(rq
, p
);
6322 check_class_changed(rq
, p
, prev_class
, oldprio
);
6324 /* Avoid rq from going away on us: */
6326 head
= splice_balance_callbacks(rq
);
6327 task_rq_unlock(rq
, p
, &rf
);
6330 cpuset_read_unlock();
6331 rt_mutex_adjust_pi(p
);
6334 /* Run balance callbacks after we've adjusted the PI chain: */
6335 balance_callbacks(rq
, head
);
6341 task_rq_unlock(rq
, p
, &rf
);
6343 cpuset_read_unlock();
6347 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
6348 const struct sched_param
*param
, bool check
)
6350 struct sched_attr attr
= {
6351 .sched_policy
= policy
,
6352 .sched_priority
= param
->sched_priority
,
6353 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
6356 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
6357 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
6358 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
6359 policy
&= ~SCHED_RESET_ON_FORK
;
6360 attr
.sched_policy
= policy
;
6363 return __sched_setscheduler(p
, &attr
, check
, true);
6366 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6367 * @p: the task in question.
6368 * @policy: new policy.
6369 * @param: structure containing the new RT priority.
6371 * Use sched_set_fifo(), read its comment.
6373 * Return: 0 on success. An error code otherwise.
6375 * NOTE that the task may be already dead.
6377 int sched_setscheduler(struct task_struct
*p
, int policy
,
6378 const struct sched_param
*param
)
6380 return _sched_setscheduler(p
, policy
, param
, true);
6383 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
6385 return __sched_setscheduler(p
, attr
, true, true);
6388 int sched_setattr_nocheck(struct task_struct
*p
, const struct sched_attr
*attr
)
6390 return __sched_setscheduler(p
, attr
, false, true);
6392 EXPORT_SYMBOL_GPL(sched_setattr_nocheck
);
6395 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6396 * @p: the task in question.
6397 * @policy: new policy.
6398 * @param: structure containing the new RT priority.
6400 * Just like sched_setscheduler, only don't bother checking if the
6401 * current context has permission. For example, this is needed in
6402 * stop_machine(): we create temporary high priority worker threads,
6403 * but our caller might not have that capability.
6405 * Return: 0 on success. An error code otherwise.
6407 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6408 const struct sched_param
*param
)
6410 return _sched_setscheduler(p
, policy
, param
, false);
6414 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
6415 * incapable of resource management, which is the one thing an OS really should
6418 * This is of course the reason it is limited to privileged users only.
6420 * Worse still; it is fundamentally impossible to compose static priority
6421 * workloads. You cannot take two correctly working static prio workloads
6422 * and smash them together and still expect them to work.
6424 * For this reason 'all' FIFO tasks the kernel creates are basically at:
6428 * The administrator _MUST_ configure the system, the kernel simply doesn't
6429 * know enough information to make a sensible choice.
6431 void sched_set_fifo(struct task_struct
*p
)
6433 struct sched_param sp
= { .sched_priority
= MAX_RT_PRIO
/ 2 };
6434 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
6436 EXPORT_SYMBOL_GPL(sched_set_fifo
);
6439 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
6441 void sched_set_fifo_low(struct task_struct
*p
)
6443 struct sched_param sp
= { .sched_priority
= 1 };
6444 WARN_ON_ONCE(sched_setscheduler_nocheck(p
, SCHED_FIFO
, &sp
) != 0);
6446 EXPORT_SYMBOL_GPL(sched_set_fifo_low
);
6448 void sched_set_normal(struct task_struct
*p
, int nice
)
6450 struct sched_attr attr
= {
6451 .sched_policy
= SCHED_NORMAL
,
6454 WARN_ON_ONCE(sched_setattr_nocheck(p
, &attr
) != 0);
6456 EXPORT_SYMBOL_GPL(sched_set_normal
);
6459 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6461 struct sched_param lparam
;
6462 struct task_struct
*p
;
6465 if (!param
|| pid
< 0)
6467 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6472 p
= find_process_by_pid(pid
);
6478 retval
= sched_setscheduler(p
, policy
, &lparam
);
6486 * Mimics kernel/events/core.c perf_copy_attr().
6488 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
6493 /* Zero the full structure, so that a short copy will be nice: */
6494 memset(attr
, 0, sizeof(*attr
));
6496 ret
= get_user(size
, &uattr
->size
);
6500 /* ABI compatibility quirk: */
6502 size
= SCHED_ATTR_SIZE_VER0
;
6503 if (size
< SCHED_ATTR_SIZE_VER0
|| size
> PAGE_SIZE
)
6506 ret
= copy_struct_from_user(attr
, sizeof(*attr
), uattr
, size
);
6513 if ((attr
->sched_flags
& SCHED_FLAG_UTIL_CLAMP
) &&
6514 size
< SCHED_ATTR_SIZE_VER1
)
6518 * XXX: Do we want to be lenient like existing syscalls; or do we want
6519 * to be strict and return an error on out-of-bounds values?
6521 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
6526 put_user(sizeof(*attr
), &uattr
->size
);
6531 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6532 * @pid: the pid in question.
6533 * @policy: new policy.
6534 * @param: structure containing the new RT priority.
6536 * Return: 0 on success. An error code otherwise.
6538 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
6543 return do_sched_setscheduler(pid
, policy
, param
);
6547 * sys_sched_setparam - set/change the RT priority of a thread
6548 * @pid: the pid in question.
6549 * @param: structure containing the new RT priority.
6551 * Return: 0 on success. An error code otherwise.
6553 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6555 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
6559 * sys_sched_setattr - same as above, but with extended sched_attr
6560 * @pid: the pid in question.
6561 * @uattr: structure containing the extended parameters.
6562 * @flags: for future extension.
6564 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
6565 unsigned int, flags
)
6567 struct sched_attr attr
;
6568 struct task_struct
*p
;
6571 if (!uattr
|| pid
< 0 || flags
)
6574 retval
= sched_copy_attr(uattr
, &attr
);
6578 if ((int)attr
.sched_policy
< 0)
6580 if (attr
.sched_flags
& SCHED_FLAG_KEEP_POLICY
)
6581 attr
.sched_policy
= SETPARAM_POLICY
;
6585 p
= find_process_by_pid(pid
);
6591 retval
= sched_setattr(p
, &attr
);
6599 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6600 * @pid: the pid in question.
6602 * Return: On success, the policy of the thread. Otherwise, a negative error
6605 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6607 struct task_struct
*p
;
6615 p
= find_process_by_pid(pid
);
6617 retval
= security_task_getscheduler(p
);
6620 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6627 * sys_sched_getparam - get the RT priority of a thread
6628 * @pid: the pid in question.
6629 * @param: structure containing the RT priority.
6631 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
6634 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6636 struct sched_param lp
= { .sched_priority
= 0 };
6637 struct task_struct
*p
;
6640 if (!param
|| pid
< 0)
6644 p
= find_process_by_pid(pid
);
6649 retval
= security_task_getscheduler(p
);
6653 if (task_has_rt_policy(p
))
6654 lp
.sched_priority
= p
->rt_priority
;
6658 * This one might sleep, we cannot do it with a spinlock held ...
6660 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6670 * Copy the kernel size attribute structure (which might be larger
6671 * than what user-space knows about) to user-space.
6673 * Note that all cases are valid: user-space buffer can be larger or
6674 * smaller than the kernel-space buffer. The usual case is that both
6675 * have the same size.
6678 sched_attr_copy_to_user(struct sched_attr __user
*uattr
,
6679 struct sched_attr
*kattr
,
6682 unsigned int ksize
= sizeof(*kattr
);
6684 if (!access_ok(uattr
, usize
))
6688 * sched_getattr() ABI forwards and backwards compatibility:
6690 * If usize == ksize then we just copy everything to user-space and all is good.
6692 * If usize < ksize then we only copy as much as user-space has space for,
6693 * this keeps ABI compatibility as well. We skip the rest.
6695 * If usize > ksize then user-space is using a newer version of the ABI,
6696 * which part the kernel doesn't know about. Just ignore it - tooling can
6697 * detect the kernel's knowledge of attributes from the attr->size value
6698 * which is set to ksize in this case.
6700 kattr
->size
= min(usize
, ksize
);
6702 if (copy_to_user(uattr
, kattr
, kattr
->size
))
6709 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
6710 * @pid: the pid in question.
6711 * @uattr: structure containing the extended parameters.
6712 * @usize: sizeof(attr) for fwd/bwd comp.
6713 * @flags: for future extension.
6715 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
6716 unsigned int, usize
, unsigned int, flags
)
6718 struct sched_attr kattr
= { };
6719 struct task_struct
*p
;
6722 if (!uattr
|| pid
< 0 || usize
> PAGE_SIZE
||
6723 usize
< SCHED_ATTR_SIZE_VER0
|| flags
)
6727 p
= find_process_by_pid(pid
);
6732 retval
= security_task_getscheduler(p
);
6736 kattr
.sched_policy
= p
->policy
;
6737 if (p
->sched_reset_on_fork
)
6738 kattr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
6739 if (task_has_dl_policy(p
))
6740 __getparam_dl(p
, &kattr
);
6741 else if (task_has_rt_policy(p
))
6742 kattr
.sched_priority
= p
->rt_priority
;
6744 kattr
.sched_nice
= task_nice(p
);
6746 #ifdef CONFIG_UCLAMP_TASK
6748 * This could race with another potential updater, but this is fine
6749 * because it'll correctly read the old or the new value. We don't need
6750 * to guarantee who wins the race as long as it doesn't return garbage.
6752 kattr
.sched_util_min
= p
->uclamp_req
[UCLAMP_MIN
].value
;
6753 kattr
.sched_util_max
= p
->uclamp_req
[UCLAMP_MAX
].value
;
6758 return sched_attr_copy_to_user(uattr
, &kattr
, usize
);
6765 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6767 cpumask_var_t cpus_allowed
, new_mask
;
6768 struct task_struct
*p
;
6773 p
= find_process_by_pid(pid
);
6779 /* Prevent p going away */
6783 if (p
->flags
& PF_NO_SETAFFINITY
) {
6787 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6791 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6793 goto out_free_cpus_allowed
;
6796 if (!check_same_owner(p
)) {
6798 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
6800 goto out_free_new_mask
;
6805 retval
= security_task_setscheduler(p
);
6807 goto out_free_new_mask
;
6810 cpuset_cpus_allowed(p
, cpus_allowed
);
6811 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6814 * Since bandwidth control happens on root_domain basis,
6815 * if admission test is enabled, we only admit -deadline
6816 * tasks allowed to run on all the CPUs in the task's
6820 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
6822 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
6825 goto out_free_new_mask
;
6831 retval
= __set_cpus_allowed_ptr(p
, new_mask
, SCA_CHECK
);
6834 cpuset_cpus_allowed(p
, cpus_allowed
);
6835 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6837 * We must have raced with a concurrent cpuset
6838 * update. Just reset the cpus_allowed to the
6839 * cpuset's cpus_allowed
6841 cpumask_copy(new_mask
, cpus_allowed
);
6846 free_cpumask_var(new_mask
);
6847 out_free_cpus_allowed
:
6848 free_cpumask_var(cpus_allowed
);
6854 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6855 struct cpumask
*new_mask
)
6857 if (len
< cpumask_size())
6858 cpumask_clear(new_mask
);
6859 else if (len
> cpumask_size())
6860 len
= cpumask_size();
6862 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6866 * sys_sched_setaffinity - set the CPU affinity of a process
6867 * @pid: pid of the process
6868 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6869 * @user_mask_ptr: user-space pointer to the new CPU mask
6871 * Return: 0 on success. An error code otherwise.
6873 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6874 unsigned long __user
*, user_mask_ptr
)
6876 cpumask_var_t new_mask
;
6879 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6882 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6884 retval
= sched_setaffinity(pid
, new_mask
);
6885 free_cpumask_var(new_mask
);
6889 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6891 struct task_struct
*p
;
6892 unsigned long flags
;
6898 p
= find_process_by_pid(pid
);
6902 retval
= security_task_getscheduler(p
);
6906 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
6907 cpumask_and(mask
, &p
->cpus_mask
, cpu_active_mask
);
6908 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6917 * sys_sched_getaffinity - get the CPU affinity of a process
6918 * @pid: pid of the process
6919 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6920 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6922 * Return: size of CPU mask copied to user_mask_ptr on success. An
6923 * error code otherwise.
6925 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6926 unsigned long __user
*, user_mask_ptr
)
6931 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
6933 if (len
& (sizeof(unsigned long)-1))
6936 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6939 ret
= sched_getaffinity(pid
, mask
);
6941 unsigned int retlen
= min(len
, cpumask_size());
6943 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
6948 free_cpumask_var(mask
);
6953 static void do_sched_yield(void)
6958 rq
= this_rq_lock_irq(&rf
);
6960 schedstat_inc(rq
->yld_count
);
6961 current
->sched_class
->yield_task(rq
);
6964 rq_unlock_irq(rq
, &rf
);
6965 sched_preempt_enable_no_resched();
6971 * sys_sched_yield - yield the current processor to other threads.
6973 * This function yields the current CPU to other tasks. If there are no
6974 * other threads running on this CPU then this function will return.
6978 SYSCALL_DEFINE0(sched_yield
)
6984 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
6985 int __sched
__cond_resched(void)
6987 if (should_resched(0)) {
6988 preempt_schedule_common();
6991 #ifndef CONFIG_PREEMPT_RCU
6996 EXPORT_SYMBOL(__cond_resched
);
6999 #ifdef CONFIG_PREEMPT_DYNAMIC
7000 DEFINE_STATIC_CALL_RET0(cond_resched
, __cond_resched
);
7001 EXPORT_STATIC_CALL_TRAMP(cond_resched
);
7003 DEFINE_STATIC_CALL_RET0(might_resched
, __cond_resched
);
7004 EXPORT_STATIC_CALL_TRAMP(might_resched
);
7008 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7009 * call schedule, and on return reacquire the lock.
7011 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7012 * operations here to prevent schedule() from being called twice (once via
7013 * spin_unlock(), once by hand).
7015 int __cond_resched_lock(spinlock_t
*lock
)
7017 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
7020 lockdep_assert_held(lock
);
7022 if (spin_needbreak(lock
) || resched
) {
7025 preempt_schedule_common();
7033 EXPORT_SYMBOL(__cond_resched_lock
);
7035 int __cond_resched_rwlock_read(rwlock_t
*lock
)
7037 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
7040 lockdep_assert_held_read(lock
);
7042 if (rwlock_needbreak(lock
) || resched
) {
7045 preempt_schedule_common();
7053 EXPORT_SYMBOL(__cond_resched_rwlock_read
);
7055 int __cond_resched_rwlock_write(rwlock_t
*lock
)
7057 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
7060 lockdep_assert_held_write(lock
);
7062 if (rwlock_needbreak(lock
) || resched
) {
7065 preempt_schedule_common();
7073 EXPORT_SYMBOL(__cond_resched_rwlock_write
);
7076 * yield - yield the current processor to other threads.
7078 * Do not ever use this function, there's a 99% chance you're doing it wrong.
7080 * The scheduler is at all times free to pick the calling task as the most
7081 * eligible task to run, if removing the yield() call from your code breaks
7082 * it, it's already broken.
7084 * Typical broken usage is:
7089 * where one assumes that yield() will let 'the other' process run that will
7090 * make event true. If the current task is a SCHED_FIFO task that will never
7091 * happen. Never use yield() as a progress guarantee!!
7093 * If you want to use yield() to wait for something, use wait_event().
7094 * If you want to use yield() to be 'nice' for others, use cond_resched().
7095 * If you still want to use yield(), do not!
7097 void __sched
yield(void)
7099 set_current_state(TASK_RUNNING
);
7102 EXPORT_SYMBOL(yield
);
7105 * yield_to - yield the current processor to another thread in
7106 * your thread group, or accelerate that thread toward the
7107 * processor it's on.
7109 * @preempt: whether task preemption is allowed or not
7111 * It's the caller's job to ensure that the target task struct
7112 * can't go away on us before we can do any checks.
7115 * true (>0) if we indeed boosted the target task.
7116 * false (0) if we failed to boost the target.
7117 * -ESRCH if there's no task to yield to.
7119 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
7121 struct task_struct
*curr
= current
;
7122 struct rq
*rq
, *p_rq
;
7123 unsigned long flags
;
7126 local_irq_save(flags
);
7132 * If we're the only runnable task on the rq and target rq also
7133 * has only one task, there's absolutely no point in yielding.
7135 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
7140 double_rq_lock(rq
, p_rq
);
7141 if (task_rq(p
) != p_rq
) {
7142 double_rq_unlock(rq
, p_rq
);
7146 if (!curr
->sched_class
->yield_to_task
)
7149 if (curr
->sched_class
!= p
->sched_class
)
7152 if (task_running(p_rq
, p
) || p
->state
)
7155 yielded
= curr
->sched_class
->yield_to_task(rq
, p
);
7157 schedstat_inc(rq
->yld_count
);
7159 * Make p's CPU reschedule; pick_next_entity takes care of
7162 if (preempt
&& rq
!= p_rq
)
7167 double_rq_unlock(rq
, p_rq
);
7169 local_irq_restore(flags
);
7176 EXPORT_SYMBOL_GPL(yield_to
);
7178 int io_schedule_prepare(void)
7180 int old_iowait
= current
->in_iowait
;
7182 current
->in_iowait
= 1;
7183 blk_schedule_flush_plug(current
);
7188 void io_schedule_finish(int token
)
7190 current
->in_iowait
= token
;
7194 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7195 * that process accounting knows that this is a task in IO wait state.
7197 long __sched
io_schedule_timeout(long timeout
)
7202 token
= io_schedule_prepare();
7203 ret
= schedule_timeout(timeout
);
7204 io_schedule_finish(token
);
7208 EXPORT_SYMBOL(io_schedule_timeout
);
7210 void __sched
io_schedule(void)
7214 token
= io_schedule_prepare();
7216 io_schedule_finish(token
);
7218 EXPORT_SYMBOL(io_schedule
);
7221 * sys_sched_get_priority_max - return maximum RT priority.
7222 * @policy: scheduling class.
7224 * Return: On success, this syscall returns the maximum
7225 * rt_priority that can be used by a given scheduling class.
7226 * On failure, a negative error code is returned.
7228 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
7235 ret
= MAX_RT_PRIO
-1;
7237 case SCHED_DEADLINE
:
7248 * sys_sched_get_priority_min - return minimum RT priority.
7249 * @policy: scheduling class.
7251 * Return: On success, this syscall returns the minimum
7252 * rt_priority that can be used by a given scheduling class.
7253 * On failure, a negative error code is returned.
7255 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
7264 case SCHED_DEADLINE
:
7273 static int sched_rr_get_interval(pid_t pid
, struct timespec64
*t
)
7275 struct task_struct
*p
;
7276 unsigned int time_slice
;
7286 p
= find_process_by_pid(pid
);
7290 retval
= security_task_getscheduler(p
);
7294 rq
= task_rq_lock(p
, &rf
);
7296 if (p
->sched_class
->get_rr_interval
)
7297 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
7298 task_rq_unlock(rq
, p
, &rf
);
7301 jiffies_to_timespec64(time_slice
, t
);
7310 * sys_sched_rr_get_interval - return the default timeslice of a process.
7311 * @pid: pid of the process.
7312 * @interval: userspace pointer to the timeslice value.
7314 * this syscall writes the default timeslice value of a given process
7315 * into the user-space timespec buffer. A value of '0' means infinity.
7317 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
7320 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
7321 struct __kernel_timespec __user
*, interval
)
7323 struct timespec64 t
;
7324 int retval
= sched_rr_get_interval(pid
, &t
);
7327 retval
= put_timespec64(&t
, interval
);
7332 #ifdef CONFIG_COMPAT_32BIT_TIME
7333 SYSCALL_DEFINE2(sched_rr_get_interval_time32
, pid_t
, pid
,
7334 struct old_timespec32 __user
*, interval
)
7336 struct timespec64 t
;
7337 int retval
= sched_rr_get_interval(pid
, &t
);
7340 retval
= put_old_timespec32(&t
, interval
);
7345 void sched_show_task(struct task_struct
*p
)
7347 unsigned long free
= 0;
7350 if (!try_get_task_stack(p
))
7353 pr_info("task:%-15.15s state:%c", p
->comm
, task_state_to_char(p
));
7355 if (p
->state
== TASK_RUNNING
)
7356 pr_cont(" running task ");
7357 #ifdef CONFIG_DEBUG_STACK_USAGE
7358 free
= stack_not_used(p
);
7363 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
7365 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
7366 free
, task_pid_nr(p
), ppid
,
7367 (unsigned long)task_thread_info(p
)->flags
);
7369 print_worker_info(KERN_INFO
, p
);
7370 print_stop_info(KERN_INFO
, p
);
7371 show_stack(p
, NULL
, KERN_INFO
);
7374 EXPORT_SYMBOL_GPL(sched_show_task
);
7377 state_filter_match(unsigned long state_filter
, struct task_struct
*p
)
7379 /* no filter, everything matches */
7383 /* filter, but doesn't match */
7384 if (!(p
->state
& state_filter
))
7388 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
7391 if (state_filter
== TASK_UNINTERRUPTIBLE
&& p
->state
== TASK_IDLE
)
7398 void show_state_filter(unsigned long state_filter
)
7400 struct task_struct
*g
, *p
;
7403 for_each_process_thread(g
, p
) {
7405 * reset the NMI-timeout, listing all files on a slow
7406 * console might take a lot of time:
7407 * Also, reset softlockup watchdogs on all CPUs, because
7408 * another CPU might be blocked waiting for us to process
7411 touch_nmi_watchdog();
7412 touch_all_softlockup_watchdogs();
7413 if (state_filter_match(state_filter
, p
))
7417 #ifdef CONFIG_SCHED_DEBUG
7419 sysrq_sched_debug_show();
7423 * Only show locks if all tasks are dumped:
7426 debug_show_all_locks();
7430 * init_idle - set up an idle thread for a given CPU
7431 * @idle: task in question
7432 * @cpu: CPU the idle task belongs to
7434 * NOTE: this function does not set the idle thread's NEED_RESCHED
7435 * flag, to make booting more robust.
7437 void init_idle(struct task_struct
*idle
, int cpu
)
7439 struct rq
*rq
= cpu_rq(cpu
);
7440 unsigned long flags
;
7442 __sched_fork(0, idle
);
7444 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
7445 raw_spin_lock(&rq
->lock
);
7447 idle
->state
= TASK_RUNNING
;
7448 idle
->se
.exec_start
= sched_clock();
7449 idle
->flags
|= PF_IDLE
;
7451 scs_task_reset(idle
);
7452 kasan_unpoison_task_stack(idle
);
7456 * It's possible that init_idle() gets called multiple times on a task,
7457 * in that case do_set_cpus_allowed() will not do the right thing.
7459 * And since this is boot we can forgo the serialization.
7461 set_cpus_allowed_common(idle
, cpumask_of(cpu
), 0);
7464 * We're having a chicken and egg problem, even though we are
7465 * holding rq->lock, the CPU isn't yet set to this CPU so the
7466 * lockdep check in task_group() will fail.
7468 * Similar case to sched_fork(). / Alternatively we could
7469 * use task_rq_lock() here and obtain the other rq->lock.
7474 __set_task_cpu(idle
, cpu
);
7478 rcu_assign_pointer(rq
->curr
, idle
);
7479 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
7483 raw_spin_unlock(&rq
->lock
);
7484 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
7486 /* Set the preempt count _outside_ the spinlocks! */
7487 init_idle_preempt_count(idle
, cpu
);
7490 * The idle tasks have their own, simple scheduling class:
7492 idle
->sched_class
= &idle_sched_class
;
7493 ftrace_graph_init_idle_task(idle
, cpu
);
7494 vtime_init_idle(idle
, cpu
);
7496 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
7502 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
7503 const struct cpumask
*trial
)
7507 if (!cpumask_weight(cur
))
7510 ret
= dl_cpuset_cpumask_can_shrink(cur
, trial
);
7515 int task_can_attach(struct task_struct
*p
,
7516 const struct cpumask
*cs_cpus_allowed
)
7521 * Kthreads which disallow setaffinity shouldn't be moved
7522 * to a new cpuset; we don't want to change their CPU
7523 * affinity and isolating such threads by their set of
7524 * allowed nodes is unnecessary. Thus, cpusets are not
7525 * applicable for such threads. This prevents checking for
7526 * success of set_cpus_allowed_ptr() on all attached tasks
7527 * before cpus_mask may be changed.
7529 if (p
->flags
& PF_NO_SETAFFINITY
) {
7534 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
7536 ret
= dl_task_can_attach(p
, cs_cpus_allowed
);
7542 bool sched_smp_initialized __read_mostly
;
7544 #ifdef CONFIG_NUMA_BALANCING
7545 /* Migrate current task p to target_cpu */
7546 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
7548 struct migration_arg arg
= { p
, target_cpu
};
7549 int curr_cpu
= task_cpu(p
);
7551 if (curr_cpu
== target_cpu
)
7554 if (!cpumask_test_cpu(target_cpu
, p
->cpus_ptr
))
7557 /* TODO: This is not properly updating schedstats */
7559 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
7560 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
7564 * Requeue a task on a given node and accurately track the number of NUMA
7565 * tasks on the runqueues
7567 void sched_setnuma(struct task_struct
*p
, int nid
)
7569 bool queued
, running
;
7573 rq
= task_rq_lock(p
, &rf
);
7574 queued
= task_on_rq_queued(p
);
7575 running
= task_current(rq
, p
);
7578 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
7580 put_prev_task(rq
, p
);
7582 p
->numa_preferred_nid
= nid
;
7585 enqueue_task(rq
, p
, ENQUEUE_RESTORE
| ENQUEUE_NOCLOCK
);
7587 set_next_task(rq
, p
);
7588 task_rq_unlock(rq
, p
, &rf
);
7590 #endif /* CONFIG_NUMA_BALANCING */
7592 #ifdef CONFIG_HOTPLUG_CPU
7594 * Ensure that the idle task is using init_mm right before its CPU goes
7597 void idle_task_exit(void)
7599 struct mm_struct
*mm
= current
->active_mm
;
7601 BUG_ON(cpu_online(smp_processor_id()));
7602 BUG_ON(current
!= this_rq()->idle
);
7604 if (mm
!= &init_mm
) {
7605 switch_mm(mm
, &init_mm
, current
);
7606 finish_arch_post_lock_switch();
7609 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
7612 static int __balance_push_cpu_stop(void *arg
)
7614 struct task_struct
*p
= arg
;
7615 struct rq
*rq
= this_rq();
7619 raw_spin_lock_irq(&p
->pi_lock
);
7622 update_rq_clock(rq
);
7624 if (task_rq(p
) == rq
&& task_on_rq_queued(p
)) {
7625 cpu
= select_fallback_rq(rq
->cpu
, p
);
7626 rq
= __migrate_task(rq
, &rf
, p
, cpu
);
7630 raw_spin_unlock_irq(&p
->pi_lock
);
7637 static DEFINE_PER_CPU(struct cpu_stop_work
, push_work
);
7640 * Ensure we only run per-cpu kthreads once the CPU goes !active.
7642 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
7643 * effective when the hotplug motion is down.
7645 static void balance_push(struct rq
*rq
)
7647 struct task_struct
*push_task
= rq
->curr
;
7649 lockdep_assert_held(&rq
->lock
);
7650 SCHED_WARN_ON(rq
->cpu
!= smp_processor_id());
7653 * Ensure the thing is persistent until balance_push_set(.on = false);
7655 rq
->balance_callback
= &balance_push_callback
;
7658 * Only active while going offline.
7660 if (!cpu_dying(rq
->cpu
))
7664 * Both the cpu-hotplug and stop task are in this case and are
7665 * required to complete the hotplug process.
7667 * XXX: the idle task does not match kthread_is_per_cpu() due to
7668 * histerical raisins.
7670 if (rq
->idle
== push_task
||
7671 kthread_is_per_cpu(push_task
) ||
7672 is_migration_disabled(push_task
)) {
7675 * If this is the idle task on the outgoing CPU try to wake
7676 * up the hotplug control thread which might wait for the
7677 * last task to vanish. The rcuwait_active() check is
7678 * accurate here because the waiter is pinned on this CPU
7679 * and can't obviously be running in parallel.
7681 * On RT kernels this also has to check whether there are
7682 * pinned and scheduled out tasks on the runqueue. They
7683 * need to leave the migrate disabled section first.
7685 if (!rq
->nr_running
&& !rq_has_pinned_tasks(rq
) &&
7686 rcuwait_active(&rq
->hotplug_wait
)) {
7687 raw_spin_unlock(&rq
->lock
);
7688 rcuwait_wake_up(&rq
->hotplug_wait
);
7689 raw_spin_lock(&rq
->lock
);
7694 get_task_struct(push_task
);
7696 * Temporarily drop rq->lock such that we can wake-up the stop task.
7697 * Both preemption and IRQs are still disabled.
7699 raw_spin_unlock(&rq
->lock
);
7700 stop_one_cpu_nowait(rq
->cpu
, __balance_push_cpu_stop
, push_task
,
7701 this_cpu_ptr(&push_work
));
7703 * At this point need_resched() is true and we'll take the loop in
7704 * schedule(). The next pick is obviously going to be the stop task
7705 * which kthread_is_per_cpu() and will push this task away.
7707 raw_spin_lock(&rq
->lock
);
7710 static void balance_push_set(int cpu
, bool on
)
7712 struct rq
*rq
= cpu_rq(cpu
);
7715 rq_lock_irqsave(rq
, &rf
);
7717 WARN_ON_ONCE(rq
->balance_callback
);
7718 rq
->balance_callback
= &balance_push_callback
;
7719 } else if (rq
->balance_callback
== &balance_push_callback
) {
7720 rq
->balance_callback
= NULL
;
7722 rq_unlock_irqrestore(rq
, &rf
);
7726 * Invoked from a CPUs hotplug control thread after the CPU has been marked
7727 * inactive. All tasks which are not per CPU kernel threads are either
7728 * pushed off this CPU now via balance_push() or placed on a different CPU
7729 * during wakeup. Wait until the CPU is quiescent.
7731 static void balance_hotplug_wait(void)
7733 struct rq
*rq
= this_rq();
7735 rcuwait_wait_event(&rq
->hotplug_wait
,
7736 rq
->nr_running
== 1 && !rq_has_pinned_tasks(rq
),
7737 TASK_UNINTERRUPTIBLE
);
7742 static inline void balance_push(struct rq
*rq
)
7746 static inline void balance_push_set(int cpu
, bool on
)
7750 static inline void balance_hotplug_wait(void)
7754 #endif /* CONFIG_HOTPLUG_CPU */
7756 void set_rq_online(struct rq
*rq
)
7759 const struct sched_class
*class;
7761 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7764 for_each_class(class) {
7765 if (class->rq_online
)
7766 class->rq_online(rq
);
7771 void set_rq_offline(struct rq
*rq
)
7774 const struct sched_class
*class;
7776 for_each_class(class) {
7777 if (class->rq_offline
)
7778 class->rq_offline(rq
);
7781 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7787 * used to mark begin/end of suspend/resume:
7789 static int num_cpus_frozen
;
7792 * Update cpusets according to cpu_active mask. If cpusets are
7793 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7794 * around partition_sched_domains().
7796 * If we come here as part of a suspend/resume, don't touch cpusets because we
7797 * want to restore it back to its original state upon resume anyway.
7799 static void cpuset_cpu_active(void)
7801 if (cpuhp_tasks_frozen
) {
7803 * num_cpus_frozen tracks how many CPUs are involved in suspend
7804 * resume sequence. As long as this is not the last online
7805 * operation in the resume sequence, just build a single sched
7806 * domain, ignoring cpusets.
7808 partition_sched_domains(1, NULL
, NULL
);
7809 if (--num_cpus_frozen
)
7812 * This is the last CPU online operation. So fall through and
7813 * restore the original sched domains by considering the
7814 * cpuset configurations.
7816 cpuset_force_rebuild();
7818 cpuset_update_active_cpus();
7821 static int cpuset_cpu_inactive(unsigned int cpu
)
7823 if (!cpuhp_tasks_frozen
) {
7824 if (dl_cpu_busy(cpu
))
7826 cpuset_update_active_cpus();
7829 partition_sched_domains(1, NULL
, NULL
);
7834 int sched_cpu_activate(unsigned int cpu
)
7836 struct rq
*rq
= cpu_rq(cpu
);
7840 * Clear the balance_push callback and prepare to schedule
7843 balance_push_set(cpu
, false);
7845 #ifdef CONFIG_SCHED_SMT
7847 * When going up, increment the number of cores with SMT present.
7849 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
7850 static_branch_inc_cpuslocked(&sched_smt_present
);
7852 set_cpu_active(cpu
, true);
7854 if (sched_smp_initialized
) {
7855 sched_domains_numa_masks_set(cpu
);
7856 cpuset_cpu_active();
7860 * Put the rq online, if not already. This happens:
7862 * 1) In the early boot process, because we build the real domains
7863 * after all CPUs have been brought up.
7865 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7868 rq_lock_irqsave(rq
, &rf
);
7870 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7873 rq_unlock_irqrestore(rq
, &rf
);
7878 int sched_cpu_deactivate(unsigned int cpu
)
7880 struct rq
*rq
= cpu_rq(cpu
);
7885 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
7886 * load balancing when not active
7888 nohz_balance_exit_idle(rq
);
7890 set_cpu_active(cpu
, false);
7893 * From this point forward, this CPU will refuse to run any task that
7894 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
7895 * push those tasks away until this gets cleared, see
7896 * sched_cpu_dying().
7898 balance_push_set(cpu
, true);
7901 * We've cleared cpu_active_mask / set balance_push, wait for all
7902 * preempt-disabled and RCU users of this state to go away such that
7903 * all new such users will observe it.
7905 * Specifically, we rely on ttwu to no longer target this CPU, see
7906 * ttwu_queue_cond() and is_cpu_allowed().
7908 * Do sync before park smpboot threads to take care the rcu boost case.
7912 rq_lock_irqsave(rq
, &rf
);
7914 update_rq_clock(rq
);
7915 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7918 rq_unlock_irqrestore(rq
, &rf
);
7920 #ifdef CONFIG_SCHED_SMT
7922 * When going down, decrement the number of cores with SMT present.
7924 if (cpumask_weight(cpu_smt_mask(cpu
)) == 2)
7925 static_branch_dec_cpuslocked(&sched_smt_present
);
7928 if (!sched_smp_initialized
)
7931 ret
= cpuset_cpu_inactive(cpu
);
7933 balance_push_set(cpu
, false);
7934 set_cpu_active(cpu
, true);
7937 sched_domains_numa_masks_clear(cpu
);
7941 static void sched_rq_cpu_starting(unsigned int cpu
)
7943 struct rq
*rq
= cpu_rq(cpu
);
7945 rq
->calc_load_update
= calc_load_update
;
7946 update_max_interval();
7949 int sched_cpu_starting(unsigned int cpu
)
7951 sched_rq_cpu_starting(cpu
);
7952 sched_tick_start(cpu
);
7956 #ifdef CONFIG_HOTPLUG_CPU
7959 * Invoked immediately before the stopper thread is invoked to bring the
7960 * CPU down completely. At this point all per CPU kthreads except the
7961 * hotplug thread (current) and the stopper thread (inactive) have been
7962 * either parked or have been unbound from the outgoing CPU. Ensure that
7963 * any of those which might be on the way out are gone.
7965 * If after this point a bound task is being woken on this CPU then the
7966 * responsible hotplug callback has failed to do it's job.
7967 * sched_cpu_dying() will catch it with the appropriate fireworks.
7969 int sched_cpu_wait_empty(unsigned int cpu
)
7971 balance_hotplug_wait();
7976 * Since this CPU is going 'away' for a while, fold any nr_active delta we
7977 * might have. Called from the CPU stopper task after ensuring that the
7978 * stopper is the last running task on the CPU, so nr_active count is
7979 * stable. We need to take the teardown thread which is calling this into
7980 * account, so we hand in adjust = 1 to the load calculation.
7982 * Also see the comment "Global load-average calculations".
7984 static void calc_load_migrate(struct rq
*rq
)
7986 long delta
= calc_load_fold_active(rq
, 1);
7989 atomic_long_add(delta
, &calc_load_tasks
);
7992 static void dump_rq_tasks(struct rq
*rq
, const char *loglvl
)
7994 struct task_struct
*g
, *p
;
7995 int cpu
= cpu_of(rq
);
7997 lockdep_assert_held(&rq
->lock
);
7999 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl
, cpu
, rq
->nr_running
);
8000 for_each_process_thread(g
, p
) {
8001 if (task_cpu(p
) != cpu
)
8004 if (!task_on_rq_queued(p
))
8007 printk("%s\tpid: %d, name: %s\n", loglvl
, p
->pid
, p
->comm
);
8011 int sched_cpu_dying(unsigned int cpu
)
8013 struct rq
*rq
= cpu_rq(cpu
);
8016 /* Handle pending wakeups and then migrate everything off */
8017 sched_tick_stop(cpu
);
8019 rq_lock_irqsave(rq
, &rf
);
8020 if (rq
->nr_running
!= 1 || rq_has_pinned_tasks(rq
)) {
8021 WARN(true, "Dying CPU not properly vacated!");
8022 dump_rq_tasks(rq
, KERN_WARNING
);
8024 rq_unlock_irqrestore(rq
, &rf
);
8026 calc_load_migrate(rq
);
8027 update_max_interval();
8033 void __init
sched_init_smp(void)
8038 * There's no userspace yet to cause hotplug operations; hence all the
8039 * CPU masks are stable and all blatant races in the below code cannot
8042 mutex_lock(&sched_domains_mutex
);
8043 sched_init_domains(cpu_active_mask
);
8044 mutex_unlock(&sched_domains_mutex
);
8046 /* Move init over to a non-isolated CPU */
8047 if (set_cpus_allowed_ptr(current
, housekeeping_cpumask(HK_FLAG_DOMAIN
)) < 0)
8049 sched_init_granularity();
8051 init_sched_rt_class();
8052 init_sched_dl_class();
8054 sched_smp_initialized
= true;
8057 static int __init
migration_init(void)
8059 sched_cpu_starting(smp_processor_id());
8062 early_initcall(migration_init
);
8065 void __init
sched_init_smp(void)
8067 sched_init_granularity();
8069 #endif /* CONFIG_SMP */
8071 int in_sched_functions(unsigned long addr
)
8073 return in_lock_functions(addr
) ||
8074 (addr
>= (unsigned long)__sched_text_start
8075 && addr
< (unsigned long)__sched_text_end
);
8078 #ifdef CONFIG_CGROUP_SCHED
8080 * Default task group.
8081 * Every task in system belongs to this group at bootup.
8083 struct task_group root_task_group
;
8084 LIST_HEAD(task_groups
);
8086 /* Cacheline aligned slab cache for task_group */
8087 static struct kmem_cache
*task_group_cache __read_mostly
;
8090 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
8091 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
8093 void __init
sched_init(void)
8095 unsigned long ptr
= 0;
8098 /* Make sure the linker didn't screw up */
8099 BUG_ON(&idle_sched_class
+ 1 != &fair_sched_class
||
8100 &fair_sched_class
+ 1 != &rt_sched_class
||
8101 &rt_sched_class
+ 1 != &dl_sched_class
);
8103 BUG_ON(&dl_sched_class
+ 1 != &stop_sched_class
);
8108 #ifdef CONFIG_FAIR_GROUP_SCHED
8109 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
8111 #ifdef CONFIG_RT_GROUP_SCHED
8112 ptr
+= 2 * nr_cpu_ids
* sizeof(void **);
8115 ptr
= (unsigned long)kzalloc(ptr
, GFP_NOWAIT
);
8117 #ifdef CONFIG_FAIR_GROUP_SCHED
8118 root_task_group
.se
= (struct sched_entity
**)ptr
;
8119 ptr
+= nr_cpu_ids
* sizeof(void **);
8121 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8122 ptr
+= nr_cpu_ids
* sizeof(void **);
8124 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
8125 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
8126 #endif /* CONFIG_FAIR_GROUP_SCHED */
8127 #ifdef CONFIG_RT_GROUP_SCHED
8128 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8129 ptr
+= nr_cpu_ids
* sizeof(void **);
8131 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8132 ptr
+= nr_cpu_ids
* sizeof(void **);
8134 #endif /* CONFIG_RT_GROUP_SCHED */
8136 #ifdef CONFIG_CPUMASK_OFFSTACK
8137 for_each_possible_cpu(i
) {
8138 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
8139 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
8140 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
8141 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
8143 #endif /* CONFIG_CPUMASK_OFFSTACK */
8145 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
8146 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
8149 init_defrootdomain();
8152 #ifdef CONFIG_RT_GROUP_SCHED
8153 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8154 global_rt_period(), global_rt_runtime());
8155 #endif /* CONFIG_RT_GROUP_SCHED */
8157 #ifdef CONFIG_CGROUP_SCHED
8158 task_group_cache
= KMEM_CACHE(task_group
, 0);
8160 list_add(&root_task_group
.list
, &task_groups
);
8161 INIT_LIST_HEAD(&root_task_group
.children
);
8162 INIT_LIST_HEAD(&root_task_group
.siblings
);
8163 autogroup_init(&init_task
);
8164 #endif /* CONFIG_CGROUP_SCHED */
8166 for_each_possible_cpu(i
) {
8170 raw_spin_lock_init(&rq
->lock
);
8172 rq
->calc_load_active
= 0;
8173 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
8174 init_cfs_rq(&rq
->cfs
);
8175 init_rt_rq(&rq
->rt
);
8176 init_dl_rq(&rq
->dl
);
8177 #ifdef CONFIG_FAIR_GROUP_SCHED
8178 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8179 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
8181 * How much CPU bandwidth does root_task_group get?
8183 * In case of task-groups formed thr' the cgroup filesystem, it
8184 * gets 100% of the CPU resources in the system. This overall
8185 * system CPU resource is divided among the tasks of
8186 * root_task_group and its child task-groups in a fair manner,
8187 * based on each entity's (task or task-group's) weight
8188 * (se->load.weight).
8190 * In other words, if root_task_group has 10 tasks of weight
8191 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8192 * then A0's share of the CPU resource is:
8194 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8196 * We achieve this by letting root_task_group's tasks sit
8197 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8199 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
8200 #endif /* CONFIG_FAIR_GROUP_SCHED */
8202 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8203 #ifdef CONFIG_RT_GROUP_SCHED
8204 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
8209 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
8210 rq
->balance_callback
= &balance_push_callback
;
8211 rq
->active_balance
= 0;
8212 rq
->next_balance
= jiffies
;
8217 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8218 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
8220 INIT_LIST_HEAD(&rq
->cfs_tasks
);
8222 rq_attach_root(rq
, &def_root_domain
);
8223 #ifdef CONFIG_NO_HZ_COMMON
8224 rq
->last_blocked_load_update_tick
= jiffies
;
8225 atomic_set(&rq
->nohz_flags
, 0);
8227 INIT_CSD(&rq
->nohz_csd
, nohz_csd_func
, rq
);
8229 #ifdef CONFIG_HOTPLUG_CPU
8230 rcuwait_init(&rq
->hotplug_wait
);
8232 #endif /* CONFIG_SMP */
8234 atomic_set(&rq
->nr_iowait
, 0);
8237 set_load_weight(&init_task
, false);
8240 * The boot idle thread does lazy MMU switching as well:
8243 enter_lazy_tlb(&init_mm
, current
);
8246 * Make us the idle thread. Technically, schedule() should not be
8247 * called from this thread, however somewhere below it might be,
8248 * but because we are the idle thread, we just pick up running again
8249 * when this runqueue becomes "idle".
8251 init_idle(current
, smp_processor_id());
8253 calc_load_update
= jiffies
+ LOAD_FREQ
;
8256 idle_thread_set_boot_cpu();
8257 balance_push_set(smp_processor_id(), false);
8259 init_sched_fair_class();
8267 scheduler_running
= 1;
8270 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8271 static inline int preempt_count_equals(int preempt_offset
)
8273 int nested
= preempt_count() + rcu_preempt_depth();
8275 return (nested
== preempt_offset
);
8278 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8281 * Blocking primitives will set (and therefore destroy) current->state,
8282 * since we will exit with TASK_RUNNING make sure we enter with it,
8283 * otherwise we will destroy state.
8285 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
8286 "do not call blocking ops when !TASK_RUNNING; "
8287 "state=%lx set at [<%p>] %pS\n",
8289 (void *)current
->task_state_change
,
8290 (void *)current
->task_state_change
);
8292 ___might_sleep(file
, line
, preempt_offset
);
8294 EXPORT_SYMBOL(__might_sleep
);
8296 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
8298 /* Ratelimiting timestamp: */
8299 static unsigned long prev_jiffy
;
8301 unsigned long preempt_disable_ip
;
8303 /* WARN_ON_ONCE() by default, no rate limit required: */
8306 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
8307 !is_idle_task(current
) && !current
->non_block_count
) ||
8308 system_state
== SYSTEM_BOOTING
|| system_state
> SYSTEM_RUNNING
||
8312 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8314 prev_jiffy
= jiffies
;
8316 /* Save this before calling printk(), since that will clobber it: */
8317 preempt_disable_ip
= get_preempt_disable_ip(current
);
8320 "BUG: sleeping function called from invalid context at %s:%d\n",
8323 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
8324 in_atomic(), irqs_disabled(), current
->non_block_count
,
8325 current
->pid
, current
->comm
);
8327 if (task_stack_end_corrupted(current
))
8328 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
8330 debug_show_held_locks(current
);
8331 if (irqs_disabled())
8332 print_irqtrace_events(current
);
8333 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
8334 && !preempt_count_equals(preempt_offset
)) {
8335 pr_err("Preemption disabled at:");
8336 print_ip_sym(KERN_ERR
, preempt_disable_ip
);
8339 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
8341 EXPORT_SYMBOL(___might_sleep
);
8343 void __cant_sleep(const char *file
, int line
, int preempt_offset
)
8345 static unsigned long prev_jiffy
;
8347 if (irqs_disabled())
8350 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
8353 if (preempt_count() > preempt_offset
)
8356 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8358 prev_jiffy
= jiffies
;
8360 printk(KERN_ERR
"BUG: assuming atomic context at %s:%d\n", file
, line
);
8361 printk(KERN_ERR
"in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8362 in_atomic(), irqs_disabled(),
8363 current
->pid
, current
->comm
);
8365 debug_show_held_locks(current
);
8367 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
8369 EXPORT_SYMBOL_GPL(__cant_sleep
);
8372 void __cant_migrate(const char *file
, int line
)
8374 static unsigned long prev_jiffy
;
8376 if (irqs_disabled())
8379 if (is_migration_disabled(current
))
8382 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT
))
8385 if (preempt_count() > 0)
8388 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8390 prev_jiffy
= jiffies
;
8392 pr_err("BUG: assuming non migratable context at %s:%d\n", file
, line
);
8393 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
8394 in_atomic(), irqs_disabled(), is_migration_disabled(current
),
8395 current
->pid
, current
->comm
);
8397 debug_show_held_locks(current
);
8399 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
8401 EXPORT_SYMBOL_GPL(__cant_migrate
);
8405 #ifdef CONFIG_MAGIC_SYSRQ
8406 void normalize_rt_tasks(void)
8408 struct task_struct
*g
, *p
;
8409 struct sched_attr attr
= {
8410 .sched_policy
= SCHED_NORMAL
,
8413 read_lock(&tasklist_lock
);
8414 for_each_process_thread(g
, p
) {
8416 * Only normalize user tasks:
8418 if (p
->flags
& PF_KTHREAD
)
8421 p
->se
.exec_start
= 0;
8422 schedstat_set(p
->se
.statistics
.wait_start
, 0);
8423 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
8424 schedstat_set(p
->se
.statistics
.block_start
, 0);
8426 if (!dl_task(p
) && !rt_task(p
)) {
8428 * Renice negative nice level userspace
8431 if (task_nice(p
) < 0)
8432 set_user_nice(p
, 0);
8436 __sched_setscheduler(p
, &attr
, false, false);
8438 read_unlock(&tasklist_lock
);
8441 #endif /* CONFIG_MAGIC_SYSRQ */
8443 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8445 * These functions are only useful for the IA64 MCA handling, or kdb.
8447 * They can only be called when the whole system has been
8448 * stopped - every CPU needs to be quiescent, and no scheduling
8449 * activity can take place. Using them for anything else would
8450 * be a serious bug, and as a result, they aren't even visible
8451 * under any other configuration.
8455 * curr_task - return the current task for a given CPU.
8456 * @cpu: the processor in question.
8458 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8460 * Return: The current task for @cpu.
8462 struct task_struct
*curr_task(int cpu
)
8464 return cpu_curr(cpu
);
8467 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8471 * ia64_set_curr_task - set the current task for a given CPU.
8472 * @cpu: the processor in question.
8473 * @p: the task pointer to set.
8475 * Description: This function must only be used when non-maskable interrupts
8476 * are serviced on a separate stack. It allows the architecture to switch the
8477 * notion of the current task on a CPU in a non-blocking manner. This function
8478 * must be called with all CPU's synchronized, and interrupts disabled, the
8479 * and caller must save the original value of the current task (see
8480 * curr_task() above) and restore that value before reenabling interrupts and
8481 * re-starting the system.
8483 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8485 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
8492 #ifdef CONFIG_CGROUP_SCHED
8493 /* task_group_lock serializes the addition/removal of task groups */
8494 static DEFINE_SPINLOCK(task_group_lock
);
8496 static inline void alloc_uclamp_sched_group(struct task_group
*tg
,
8497 struct task_group
*parent
)
8499 #ifdef CONFIG_UCLAMP_TASK_GROUP
8500 enum uclamp_id clamp_id
;
8502 for_each_clamp_id(clamp_id
) {
8503 uclamp_se_set(&tg
->uclamp_req
[clamp_id
],
8504 uclamp_none(clamp_id
), false);
8505 tg
->uclamp
[clamp_id
] = parent
->uclamp
[clamp_id
];
8510 static void sched_free_group(struct task_group
*tg
)
8512 free_fair_sched_group(tg
);
8513 free_rt_sched_group(tg
);
8515 kmem_cache_free(task_group_cache
, tg
);
8518 /* allocate runqueue etc for a new task group */
8519 struct task_group
*sched_create_group(struct task_group
*parent
)
8521 struct task_group
*tg
;
8523 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
8525 return ERR_PTR(-ENOMEM
);
8527 if (!alloc_fair_sched_group(tg
, parent
))
8530 if (!alloc_rt_sched_group(tg
, parent
))
8533 alloc_uclamp_sched_group(tg
, parent
);
8538 sched_free_group(tg
);
8539 return ERR_PTR(-ENOMEM
);
8542 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
8544 unsigned long flags
;
8546 spin_lock_irqsave(&task_group_lock
, flags
);
8547 list_add_rcu(&tg
->list
, &task_groups
);
8549 /* Root should already exist: */
8552 tg
->parent
= parent
;
8553 INIT_LIST_HEAD(&tg
->children
);
8554 list_add_rcu(&tg
->siblings
, &parent
->children
);
8555 spin_unlock_irqrestore(&task_group_lock
, flags
);
8557 online_fair_sched_group(tg
);
8560 /* rcu callback to free various structures associated with a task group */
8561 static void sched_free_group_rcu(struct rcu_head
*rhp
)
8563 /* Now it should be safe to free those cfs_rqs: */
8564 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
8567 void sched_destroy_group(struct task_group
*tg
)
8569 /* Wait for possible concurrent references to cfs_rqs complete: */
8570 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
8573 void sched_offline_group(struct task_group
*tg
)
8575 unsigned long flags
;
8577 /* End participation in shares distribution: */
8578 unregister_fair_sched_group(tg
);
8580 spin_lock_irqsave(&task_group_lock
, flags
);
8581 list_del_rcu(&tg
->list
);
8582 list_del_rcu(&tg
->siblings
);
8583 spin_unlock_irqrestore(&task_group_lock
, flags
);
8586 static void sched_change_group(struct task_struct
*tsk
, int type
)
8588 struct task_group
*tg
;
8591 * All callers are synchronized by task_rq_lock(); we do not use RCU
8592 * which is pointless here. Thus, we pass "true" to task_css_check()
8593 * to prevent lockdep warnings.
8595 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
8596 struct task_group
, css
);
8597 tg
= autogroup_task_group(tsk
, tg
);
8598 tsk
->sched_task_group
= tg
;
8600 #ifdef CONFIG_FAIR_GROUP_SCHED
8601 if (tsk
->sched_class
->task_change_group
)
8602 tsk
->sched_class
->task_change_group(tsk
, type
);
8605 set_task_rq(tsk
, task_cpu(tsk
));
8609 * Change task's runqueue when it moves between groups.
8611 * The caller of this function should have put the task in its new group by
8612 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8615 void sched_move_task(struct task_struct
*tsk
)
8617 int queued
, running
, queue_flags
=
8618 DEQUEUE_SAVE
| DEQUEUE_MOVE
| DEQUEUE_NOCLOCK
;
8622 rq
= task_rq_lock(tsk
, &rf
);
8623 update_rq_clock(rq
);
8625 running
= task_current(rq
, tsk
);
8626 queued
= task_on_rq_queued(tsk
);
8629 dequeue_task(rq
, tsk
, queue_flags
);
8631 put_prev_task(rq
, tsk
);
8633 sched_change_group(tsk
, TASK_MOVE_GROUP
);
8636 enqueue_task(rq
, tsk
, queue_flags
);
8638 set_next_task(rq
, tsk
);
8640 * After changing group, the running task may have joined a
8641 * throttled one but it's still the running task. Trigger a
8642 * resched to make sure that task can still run.
8647 task_rq_unlock(rq
, tsk
, &rf
);
8650 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
8652 return css
? container_of(css
, struct task_group
, css
) : NULL
;
8655 static struct cgroup_subsys_state
*
8656 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
8658 struct task_group
*parent
= css_tg(parent_css
);
8659 struct task_group
*tg
;
8662 /* This is early initialization for the top cgroup */
8663 return &root_task_group
.css
;
8666 tg
= sched_create_group(parent
);
8668 return ERR_PTR(-ENOMEM
);
8673 /* Expose task group only after completing cgroup initialization */
8674 static int cpu_cgroup_css_online(struct cgroup_subsys_state
*css
)
8676 struct task_group
*tg
= css_tg(css
);
8677 struct task_group
*parent
= css_tg(css
->parent
);
8680 sched_online_group(tg
, parent
);
8682 #ifdef CONFIG_UCLAMP_TASK_GROUP
8683 /* Propagate the effective uclamp value for the new group */
8684 cpu_util_update_eff(css
);
8690 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
8692 struct task_group
*tg
= css_tg(css
);
8694 sched_offline_group(tg
);
8697 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
8699 struct task_group
*tg
= css_tg(css
);
8702 * Relies on the RCU grace period between css_released() and this.
8704 sched_free_group(tg
);
8708 * This is called before wake_up_new_task(), therefore we really only
8709 * have to set its group bits, all the other stuff does not apply.
8711 static void cpu_cgroup_fork(struct task_struct
*task
)
8716 rq
= task_rq_lock(task
, &rf
);
8718 update_rq_clock(rq
);
8719 sched_change_group(task
, TASK_SET_GROUP
);
8721 task_rq_unlock(rq
, task
, &rf
);
8724 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
8726 struct task_struct
*task
;
8727 struct cgroup_subsys_state
*css
;
8730 cgroup_taskset_for_each(task
, css
, tset
) {
8731 #ifdef CONFIG_RT_GROUP_SCHED
8732 if (!sched_rt_can_attach(css_tg(css
), task
))
8736 * Serialize against wake_up_new_task() such that if it's
8737 * running, we're sure to observe its full state.
8739 raw_spin_lock_irq(&task
->pi_lock
);
8741 * Avoid calling sched_move_task() before wake_up_new_task()
8742 * has happened. This would lead to problems with PELT, due to
8743 * move wanting to detach+attach while we're not attached yet.
8745 if (task
->state
== TASK_NEW
)
8747 raw_spin_unlock_irq(&task
->pi_lock
);
8755 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
8757 struct task_struct
*task
;
8758 struct cgroup_subsys_state
*css
;
8760 cgroup_taskset_for_each(task
, css
, tset
)
8761 sched_move_task(task
);
8764 #ifdef CONFIG_UCLAMP_TASK_GROUP
8765 static void cpu_util_update_eff(struct cgroup_subsys_state
*css
)
8767 struct cgroup_subsys_state
*top_css
= css
;
8768 struct uclamp_se
*uc_parent
= NULL
;
8769 struct uclamp_se
*uc_se
= NULL
;
8770 unsigned int eff
[UCLAMP_CNT
];
8771 enum uclamp_id clamp_id
;
8772 unsigned int clamps
;
8774 css_for_each_descendant_pre(css
, top_css
) {
8775 uc_parent
= css_tg(css
)->parent
8776 ? css_tg(css
)->parent
->uclamp
: NULL
;
8778 for_each_clamp_id(clamp_id
) {
8779 /* Assume effective clamps matches requested clamps */
8780 eff
[clamp_id
] = css_tg(css
)->uclamp_req
[clamp_id
].value
;
8781 /* Cap effective clamps with parent's effective clamps */
8783 eff
[clamp_id
] > uc_parent
[clamp_id
].value
) {
8784 eff
[clamp_id
] = uc_parent
[clamp_id
].value
;
8787 /* Ensure protection is always capped by limit */
8788 eff
[UCLAMP_MIN
] = min(eff
[UCLAMP_MIN
], eff
[UCLAMP_MAX
]);
8790 /* Propagate most restrictive effective clamps */
8792 uc_se
= css_tg(css
)->uclamp
;
8793 for_each_clamp_id(clamp_id
) {
8794 if (eff
[clamp_id
] == uc_se
[clamp_id
].value
)
8796 uc_se
[clamp_id
].value
= eff
[clamp_id
];
8797 uc_se
[clamp_id
].bucket_id
= uclamp_bucket_id(eff
[clamp_id
]);
8798 clamps
|= (0x1 << clamp_id
);
8801 css
= css_rightmost_descendant(css
);
8805 /* Immediately update descendants RUNNABLE tasks */
8806 uclamp_update_active_tasks(css
, clamps
);
8811 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
8812 * C expression. Since there is no way to convert a macro argument (N) into a
8813 * character constant, use two levels of macros.
8815 #define _POW10(exp) ((unsigned int)1e##exp)
8816 #define POW10(exp) _POW10(exp)
8818 struct uclamp_request
{
8819 #define UCLAMP_PERCENT_SHIFT 2
8820 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
8826 static inline struct uclamp_request
8827 capacity_from_percent(char *buf
)
8829 struct uclamp_request req
= {
8830 .percent
= UCLAMP_PERCENT_SCALE
,
8831 .util
= SCHED_CAPACITY_SCALE
,
8836 if (strcmp(buf
, "max")) {
8837 req
.ret
= cgroup_parse_float(buf
, UCLAMP_PERCENT_SHIFT
,
8841 if ((u64
)req
.percent
> UCLAMP_PERCENT_SCALE
) {
8846 req
.util
= req
.percent
<< SCHED_CAPACITY_SHIFT
;
8847 req
.util
= DIV_ROUND_CLOSEST_ULL(req
.util
, UCLAMP_PERCENT_SCALE
);
8853 static ssize_t
cpu_uclamp_write(struct kernfs_open_file
*of
, char *buf
,
8854 size_t nbytes
, loff_t off
,
8855 enum uclamp_id clamp_id
)
8857 struct uclamp_request req
;
8858 struct task_group
*tg
;
8860 req
= capacity_from_percent(buf
);
8864 static_branch_enable(&sched_uclamp_used
);
8866 mutex_lock(&uclamp_mutex
);
8869 tg
= css_tg(of_css(of
));
8870 if (tg
->uclamp_req
[clamp_id
].value
!= req
.util
)
8871 uclamp_se_set(&tg
->uclamp_req
[clamp_id
], req
.util
, false);
8874 * Because of not recoverable conversion rounding we keep track of the
8875 * exact requested value
8877 tg
->uclamp_pct
[clamp_id
] = req
.percent
;
8879 /* Update effective clamps to track the most restrictive value */
8880 cpu_util_update_eff(of_css(of
));
8883 mutex_unlock(&uclamp_mutex
);
8888 static ssize_t
cpu_uclamp_min_write(struct kernfs_open_file
*of
,
8889 char *buf
, size_t nbytes
,
8892 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MIN
);
8895 static ssize_t
cpu_uclamp_max_write(struct kernfs_open_file
*of
,
8896 char *buf
, size_t nbytes
,
8899 return cpu_uclamp_write(of
, buf
, nbytes
, off
, UCLAMP_MAX
);
8902 static inline void cpu_uclamp_print(struct seq_file
*sf
,
8903 enum uclamp_id clamp_id
)
8905 struct task_group
*tg
;
8911 tg
= css_tg(seq_css(sf
));
8912 util_clamp
= tg
->uclamp_req
[clamp_id
].value
;
8915 if (util_clamp
== SCHED_CAPACITY_SCALE
) {
8916 seq_puts(sf
, "max\n");
8920 percent
= tg
->uclamp_pct
[clamp_id
];
8921 percent
= div_u64_rem(percent
, POW10(UCLAMP_PERCENT_SHIFT
), &rem
);
8922 seq_printf(sf
, "%llu.%0*u\n", percent
, UCLAMP_PERCENT_SHIFT
, rem
);
8925 static int cpu_uclamp_min_show(struct seq_file
*sf
, void *v
)
8927 cpu_uclamp_print(sf
, UCLAMP_MIN
);
8931 static int cpu_uclamp_max_show(struct seq_file
*sf
, void *v
)
8933 cpu_uclamp_print(sf
, UCLAMP_MAX
);
8936 #endif /* CONFIG_UCLAMP_TASK_GROUP */
8938 #ifdef CONFIG_FAIR_GROUP_SCHED
8939 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
8940 struct cftype
*cftype
, u64 shareval
)
8942 if (shareval
> scale_load_down(ULONG_MAX
))
8943 shareval
= MAX_SHARES
;
8944 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
8947 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
8950 struct task_group
*tg
= css_tg(css
);
8952 return (u64
) scale_load_down(tg
->shares
);
8955 #ifdef CONFIG_CFS_BANDWIDTH
8956 static DEFINE_MUTEX(cfs_constraints_mutex
);
8958 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
8959 static const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
8960 /* More than 203 days if BW_SHIFT equals 20. */
8961 static const u64 max_cfs_runtime
= MAX_BW
* NSEC_PER_USEC
;
8963 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
8965 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
8967 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
8968 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
8970 if (tg
== &root_task_group
)
8974 * Ensure we have at some amount of bandwidth every period. This is
8975 * to prevent reaching a state of large arrears when throttled via
8976 * entity_tick() resulting in prolonged exit starvation.
8978 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
8982 * Likewise, bound things on the other side by preventing insane quota
8983 * periods. This also allows us to normalize in computing quota
8986 if (period
> max_cfs_quota_period
)
8990 * Bound quota to defend quota against overflow during bandwidth shift.
8992 if (quota
!= RUNTIME_INF
&& quota
> max_cfs_runtime
)
8996 * Prevent race between setting of cfs_rq->runtime_enabled and
8997 * unthrottle_offline_cfs_rqs().
9000 mutex_lock(&cfs_constraints_mutex
);
9001 ret
= __cfs_schedulable(tg
, period
, quota
);
9005 runtime_enabled
= quota
!= RUNTIME_INF
;
9006 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
9008 * If we need to toggle cfs_bandwidth_used, off->on must occur
9009 * before making related changes, and on->off must occur afterwards
9011 if (runtime_enabled
&& !runtime_was_enabled
)
9012 cfs_bandwidth_usage_inc();
9013 raw_spin_lock_irq(&cfs_b
->lock
);
9014 cfs_b
->period
= ns_to_ktime(period
);
9015 cfs_b
->quota
= quota
;
9017 __refill_cfs_bandwidth_runtime(cfs_b
);
9019 /* Restart the period timer (if active) to handle new period expiry: */
9020 if (runtime_enabled
)
9021 start_cfs_bandwidth(cfs_b
);
9023 raw_spin_unlock_irq(&cfs_b
->lock
);
9025 for_each_online_cpu(i
) {
9026 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
9027 struct rq
*rq
= cfs_rq
->rq
;
9030 rq_lock_irq(rq
, &rf
);
9031 cfs_rq
->runtime_enabled
= runtime_enabled
;
9032 cfs_rq
->runtime_remaining
= 0;
9034 if (cfs_rq
->throttled
)
9035 unthrottle_cfs_rq(cfs_rq
);
9036 rq_unlock_irq(rq
, &rf
);
9038 if (runtime_was_enabled
&& !runtime_enabled
)
9039 cfs_bandwidth_usage_dec();
9041 mutex_unlock(&cfs_constraints_mutex
);
9047 static int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
9051 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
9052 if (cfs_quota_us
< 0)
9053 quota
= RUNTIME_INF
;
9054 else if ((u64
)cfs_quota_us
<= U64_MAX
/ NSEC_PER_USEC
)
9055 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
9059 return tg_set_cfs_bandwidth(tg
, period
, quota
);
9062 static long tg_get_cfs_quota(struct task_group
*tg
)
9066 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
9069 quota_us
= tg
->cfs_bandwidth
.quota
;
9070 do_div(quota_us
, NSEC_PER_USEC
);
9075 static int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
9079 if ((u64
)cfs_period_us
> U64_MAX
/ NSEC_PER_USEC
)
9082 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
9083 quota
= tg
->cfs_bandwidth
.quota
;
9085 return tg_set_cfs_bandwidth(tg
, period
, quota
);
9088 static long tg_get_cfs_period(struct task_group
*tg
)
9092 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
9093 do_div(cfs_period_us
, NSEC_PER_USEC
);
9095 return cfs_period_us
;
9098 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
9101 return tg_get_cfs_quota(css_tg(css
));
9104 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
9105 struct cftype
*cftype
, s64 cfs_quota_us
)
9107 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
9110 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
9113 return tg_get_cfs_period(css_tg(css
));
9116 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
9117 struct cftype
*cftype
, u64 cfs_period_us
)
9119 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
9122 struct cfs_schedulable_data
{
9123 struct task_group
*tg
;
9128 * normalize group quota/period to be quota/max_period
9129 * note: units are usecs
9131 static u64
normalize_cfs_quota(struct task_group
*tg
,
9132 struct cfs_schedulable_data
*d
)
9140 period
= tg_get_cfs_period(tg
);
9141 quota
= tg_get_cfs_quota(tg
);
9144 /* note: these should typically be equivalent */
9145 if (quota
== RUNTIME_INF
|| quota
== -1)
9148 return to_ratio(period
, quota
);
9151 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
9153 struct cfs_schedulable_data
*d
= data
;
9154 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
9155 s64 quota
= 0, parent_quota
= -1;
9158 quota
= RUNTIME_INF
;
9160 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
9162 quota
= normalize_cfs_quota(tg
, d
);
9163 parent_quota
= parent_b
->hierarchical_quota
;
9166 * Ensure max(child_quota) <= parent_quota. On cgroup2,
9167 * always take the min. On cgroup1, only inherit when no
9170 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys
)) {
9171 quota
= min(quota
, parent_quota
);
9173 if (quota
== RUNTIME_INF
)
9174 quota
= parent_quota
;
9175 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
9179 cfs_b
->hierarchical_quota
= quota
;
9184 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
9187 struct cfs_schedulable_data data
= {
9193 if (quota
!= RUNTIME_INF
) {
9194 do_div(data
.period
, NSEC_PER_USEC
);
9195 do_div(data
.quota
, NSEC_PER_USEC
);
9199 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
9205 static int cpu_cfs_stat_show(struct seq_file
*sf
, void *v
)
9207 struct task_group
*tg
= css_tg(seq_css(sf
));
9208 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
9210 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
9211 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
9212 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
9214 if (schedstat_enabled() && tg
!= &root_task_group
) {
9218 for_each_possible_cpu(i
)
9219 ws
+= schedstat_val(tg
->se
[i
]->statistics
.wait_sum
);
9221 seq_printf(sf
, "wait_sum %llu\n", ws
);
9226 #endif /* CONFIG_CFS_BANDWIDTH */
9227 #endif /* CONFIG_FAIR_GROUP_SCHED */
9229 #ifdef CONFIG_RT_GROUP_SCHED
9230 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
9231 struct cftype
*cft
, s64 val
)
9233 return sched_group_set_rt_runtime(css_tg(css
), val
);
9236 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
9239 return sched_group_rt_runtime(css_tg(css
));
9242 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
9243 struct cftype
*cftype
, u64 rt_period_us
)
9245 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
9248 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
9251 return sched_group_rt_period(css_tg(css
));
9253 #endif /* CONFIG_RT_GROUP_SCHED */
9255 static struct cftype cpu_legacy_files
[] = {
9256 #ifdef CONFIG_FAIR_GROUP_SCHED
9259 .read_u64
= cpu_shares_read_u64
,
9260 .write_u64
= cpu_shares_write_u64
,
9263 #ifdef CONFIG_CFS_BANDWIDTH
9265 .name
= "cfs_quota_us",
9266 .read_s64
= cpu_cfs_quota_read_s64
,
9267 .write_s64
= cpu_cfs_quota_write_s64
,
9270 .name
= "cfs_period_us",
9271 .read_u64
= cpu_cfs_period_read_u64
,
9272 .write_u64
= cpu_cfs_period_write_u64
,
9276 .seq_show
= cpu_cfs_stat_show
,
9279 #ifdef CONFIG_RT_GROUP_SCHED
9281 .name
= "rt_runtime_us",
9282 .read_s64
= cpu_rt_runtime_read
,
9283 .write_s64
= cpu_rt_runtime_write
,
9286 .name
= "rt_period_us",
9287 .read_u64
= cpu_rt_period_read_uint
,
9288 .write_u64
= cpu_rt_period_write_uint
,
9291 #ifdef CONFIG_UCLAMP_TASK_GROUP
9293 .name
= "uclamp.min",
9294 .flags
= CFTYPE_NOT_ON_ROOT
,
9295 .seq_show
= cpu_uclamp_min_show
,
9296 .write
= cpu_uclamp_min_write
,
9299 .name
= "uclamp.max",
9300 .flags
= CFTYPE_NOT_ON_ROOT
,
9301 .seq_show
= cpu_uclamp_max_show
,
9302 .write
= cpu_uclamp_max_write
,
9308 static int cpu_extra_stat_show(struct seq_file
*sf
,
9309 struct cgroup_subsys_state
*css
)
9311 #ifdef CONFIG_CFS_BANDWIDTH
9313 struct task_group
*tg
= css_tg(css
);
9314 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
9317 throttled_usec
= cfs_b
->throttled_time
;
9318 do_div(throttled_usec
, NSEC_PER_USEC
);
9320 seq_printf(sf
, "nr_periods %d\n"
9322 "throttled_usec %llu\n",
9323 cfs_b
->nr_periods
, cfs_b
->nr_throttled
,
9330 #ifdef CONFIG_FAIR_GROUP_SCHED
9331 static u64
cpu_weight_read_u64(struct cgroup_subsys_state
*css
,
9334 struct task_group
*tg
= css_tg(css
);
9335 u64 weight
= scale_load_down(tg
->shares
);
9337 return DIV_ROUND_CLOSEST_ULL(weight
* CGROUP_WEIGHT_DFL
, 1024);
9340 static int cpu_weight_write_u64(struct cgroup_subsys_state
*css
,
9341 struct cftype
*cft
, u64 weight
)
9344 * cgroup weight knobs should use the common MIN, DFL and MAX
9345 * values which are 1, 100 and 10000 respectively. While it loses
9346 * a bit of range on both ends, it maps pretty well onto the shares
9347 * value used by scheduler and the round-trip conversions preserve
9348 * the original value over the entire range.
9350 if (weight
< CGROUP_WEIGHT_MIN
|| weight
> CGROUP_WEIGHT_MAX
)
9353 weight
= DIV_ROUND_CLOSEST_ULL(weight
* 1024, CGROUP_WEIGHT_DFL
);
9355 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
9358 static s64
cpu_weight_nice_read_s64(struct cgroup_subsys_state
*css
,
9361 unsigned long weight
= scale_load_down(css_tg(css
)->shares
);
9362 int last_delta
= INT_MAX
;
9365 /* find the closest nice value to the current weight */
9366 for (prio
= 0; prio
< ARRAY_SIZE(sched_prio_to_weight
); prio
++) {
9367 delta
= abs(sched_prio_to_weight
[prio
] - weight
);
9368 if (delta
>= last_delta
)
9373 return PRIO_TO_NICE(prio
- 1 + MAX_RT_PRIO
);
9376 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state
*css
,
9377 struct cftype
*cft
, s64 nice
)
9379 unsigned long weight
;
9382 if (nice
< MIN_NICE
|| nice
> MAX_NICE
)
9385 idx
= NICE_TO_PRIO(nice
) - MAX_RT_PRIO
;
9386 idx
= array_index_nospec(idx
, 40);
9387 weight
= sched_prio_to_weight
[idx
];
9389 return sched_group_set_shares(css_tg(css
), scale_load(weight
));
9393 static void __maybe_unused
cpu_period_quota_print(struct seq_file
*sf
,
9394 long period
, long quota
)
9397 seq_puts(sf
, "max");
9399 seq_printf(sf
, "%ld", quota
);
9401 seq_printf(sf
, " %ld\n", period
);
9404 /* caller should put the current value in *@periodp before calling */
9405 static int __maybe_unused
cpu_period_quota_parse(char *buf
,
9406 u64
*periodp
, u64
*quotap
)
9408 char tok
[21]; /* U64_MAX */
9410 if (sscanf(buf
, "%20s %llu", tok
, periodp
) < 1)
9413 *periodp
*= NSEC_PER_USEC
;
9415 if (sscanf(tok
, "%llu", quotap
))
9416 *quotap
*= NSEC_PER_USEC
;
9417 else if (!strcmp(tok
, "max"))
9418 *quotap
= RUNTIME_INF
;
9425 #ifdef CONFIG_CFS_BANDWIDTH
9426 static int cpu_max_show(struct seq_file
*sf
, void *v
)
9428 struct task_group
*tg
= css_tg(seq_css(sf
));
9430 cpu_period_quota_print(sf
, tg_get_cfs_period(tg
), tg_get_cfs_quota(tg
));
9434 static ssize_t
cpu_max_write(struct kernfs_open_file
*of
,
9435 char *buf
, size_t nbytes
, loff_t off
)
9437 struct task_group
*tg
= css_tg(of_css(of
));
9438 u64 period
= tg_get_cfs_period(tg
);
9442 ret
= cpu_period_quota_parse(buf
, &period
, "a
);
9444 ret
= tg_set_cfs_bandwidth(tg
, period
, quota
);
9445 return ret
?: nbytes
;
9449 static struct cftype cpu_files
[] = {
9450 #ifdef CONFIG_FAIR_GROUP_SCHED
9453 .flags
= CFTYPE_NOT_ON_ROOT
,
9454 .read_u64
= cpu_weight_read_u64
,
9455 .write_u64
= cpu_weight_write_u64
,
9458 .name
= "weight.nice",
9459 .flags
= CFTYPE_NOT_ON_ROOT
,
9460 .read_s64
= cpu_weight_nice_read_s64
,
9461 .write_s64
= cpu_weight_nice_write_s64
,
9464 #ifdef CONFIG_CFS_BANDWIDTH
9467 .flags
= CFTYPE_NOT_ON_ROOT
,
9468 .seq_show
= cpu_max_show
,
9469 .write
= cpu_max_write
,
9472 #ifdef CONFIG_UCLAMP_TASK_GROUP
9474 .name
= "uclamp.min",
9475 .flags
= CFTYPE_NOT_ON_ROOT
,
9476 .seq_show
= cpu_uclamp_min_show
,
9477 .write
= cpu_uclamp_min_write
,
9480 .name
= "uclamp.max",
9481 .flags
= CFTYPE_NOT_ON_ROOT
,
9482 .seq_show
= cpu_uclamp_max_show
,
9483 .write
= cpu_uclamp_max_write
,
9489 struct cgroup_subsys cpu_cgrp_subsys
= {
9490 .css_alloc
= cpu_cgroup_css_alloc
,
9491 .css_online
= cpu_cgroup_css_online
,
9492 .css_released
= cpu_cgroup_css_released
,
9493 .css_free
= cpu_cgroup_css_free
,
9494 .css_extra_stat_show
= cpu_extra_stat_show
,
9495 .fork
= cpu_cgroup_fork
,
9496 .can_attach
= cpu_cgroup_can_attach
,
9497 .attach
= cpu_cgroup_attach
,
9498 .legacy_cftypes
= cpu_legacy_files
,
9499 .dfl_cftypes
= cpu_files
,
9504 #endif /* CONFIG_CGROUP_SCHED */
9506 void dump_cpu_task(int cpu
)
9508 pr_info("Task dump for CPU %d:\n", cpu
);
9509 sched_show_task(cpu_curr(cpu
));
9513 * Nice levels are multiplicative, with a gentle 10% change for every
9514 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
9515 * nice 1, it will get ~10% less CPU time than another CPU-bound task
9516 * that remained on nice 0.
9518 * The "10% effect" is relative and cumulative: from _any_ nice level,
9519 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
9520 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
9521 * If a task goes up by ~10% and another task goes down by ~10% then
9522 * the relative distance between them is ~25%.)
9524 const int sched_prio_to_weight
[40] = {
9525 /* -20 */ 88761, 71755, 56483, 46273, 36291,
9526 /* -15 */ 29154, 23254, 18705, 14949, 11916,
9527 /* -10 */ 9548, 7620, 6100, 4904, 3906,
9528 /* -5 */ 3121, 2501, 1991, 1586, 1277,
9529 /* 0 */ 1024, 820, 655, 526, 423,
9530 /* 5 */ 335, 272, 215, 172, 137,
9531 /* 10 */ 110, 87, 70, 56, 45,
9532 /* 15 */ 36, 29, 23, 18, 15,
9536 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
9538 * In cases where the weight does not change often, we can use the
9539 * precalculated inverse to speed up arithmetics by turning divisions
9540 * into multiplications:
9542 const u32 sched_prio_to_wmult
[40] = {
9543 /* -20 */ 48388, 59856, 76040, 92818, 118348,
9544 /* -15 */ 147320, 184698, 229616, 287308, 360437,
9545 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
9546 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
9547 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
9548 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
9549 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
9550 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
9553 void call_trace_sched_update_nr_running(struct rq
*rq
, int count
)
9555 trace_sched_update_nr_running_tp(rq
, count
);