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;
87 #ifdef CONFIG_SCHED_CORE
89 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
91 /* kernel prio, less is more */
92 static inline int __task_prio(struct task_struct *p)
94 if (p->sched_class == &stop_sched_class) /* trumps deadline */
97 if (rt_prio(p->prio)) /* includes deadline */
98 return p->prio; /* [-1, 99] */
100 if (p->sched_class == &idle_sched_class)
101 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
103 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
113 /* real prio, less is less */
114 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
117 int pa = __task_prio(a), pb = __task_prio(b);
125 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
126 return !dl_time_before(a->dl.deadline, b->dl.deadline);
128 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
129 return cfs_prio_less(a, b, in_fi);
134 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
136 if (a->core_cookie < b->core_cookie)
139 if (a->core_cookie > b->core_cookie)
142 /* flip prio, so high prio is leftmost */
143 if (prio_less(b, a, task_rq(a)->core->core_forceidle))
149 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
151 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
153 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
156 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
158 const struct task_struct *p = __node_2_sc(node);
159 unsigned long cookie = (unsigned long)key;
161 if (cookie < p->core_cookie)
164 if (cookie > p->core_cookie)
170 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
172 rq->core->core_task_seq++;
177 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
180 void sched_core_dequeue(struct rq *rq, struct task_struct *p)
182 rq->core->core_task_seq++;
184 if (!sched_core_enqueued(p))
187 rb_erase(&p->core_node, &rq->core_tree);
188 RB_CLEAR_NODE(&p->core_node);
192 * Find left-most (aka, highest priority) task matching @cookie.
194 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
196 struct rb_node *node;
198 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
200 * The idle task always matches any cookie!
203 return idle_sched_class.pick_task(rq);
205 return __node_2_sc(node);
208 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
210 struct rb_node *node = &p->core_node;
212 node = rb_next(node);
216 p = container_of(node, struct task_struct, core_node);
217 if (p->core_cookie != cookie)
224 * Magic required such that:
226 * raw_spin_rq_lock(rq);
228 * raw_spin_rq_unlock(rq);
230 * ends up locking and unlocking the _same_ lock, and all CPUs
231 * always agree on what rq has what lock.
233 * XXX entirely possible to selectively enable cores, don't bother for now.
236 static DEFINE_MUTEX(sched_core_mutex);
237 static atomic_t sched_core_count;
238 static struct cpumask sched_core_mask;
240 static void sched_core_lock(int cpu, unsigned long *flags)
242 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
245 local_irq_save(*flags);
246 for_each_cpu(t, smt_mask)
247 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
250 static void sched_core_unlock(int cpu, unsigned long *flags)
252 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
255 for_each_cpu(t, smt_mask)
256 raw_spin_unlock(&cpu_rq(t)->__lock);
257 local_irq_restore(*flags);
260 static void __sched_core_flip(bool enabled)
268 * Toggle the online cores, one by one.
270 cpumask_copy(&sched_core_mask, cpu_online_mask);
271 for_each_cpu(cpu, &sched_core_mask) {
272 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
274 sched_core_lock(cpu, &flags);
276 for_each_cpu(t, smt_mask)
277 cpu_rq(t)->core_enabled = enabled;
279 sched_core_unlock(cpu, &flags);
281 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
285 * Toggle the offline CPUs.
287 cpumask_copy(&sched_core_mask, cpu_possible_mask);
288 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
290 for_each_cpu(cpu, &sched_core_mask)
291 cpu_rq(cpu)->core_enabled = enabled;
296 static void sched_core_assert_empty(void)
300 for_each_possible_cpu(cpu)
301 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
304 static void __sched_core_enable(void)
306 static_branch_enable(&__sched_core_enabled);
308 * Ensure all previous instances of raw_spin_rq_*lock() have finished
309 * and future ones will observe !sched_core_disabled().
312 __sched_core_flip(true);
313 sched_core_assert_empty();
316 static void __sched_core_disable(void)
318 sched_core_assert_empty();
319 __sched_core_flip(false);
320 static_branch_disable(&__sched_core_enabled);
323 void sched_core_get(void)
325 if (atomic_inc_not_zero(&sched_core_count))
328 mutex_lock(&sched_core_mutex);
329 if (!atomic_read(&sched_core_count))
330 __sched_core_enable();
332 smp_mb__before_atomic();
333 atomic_inc(&sched_core_count);
334 mutex_unlock(&sched_core_mutex);
337 static void __sched_core_put(struct work_struct *work)
339 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
340 __sched_core_disable();
341 mutex_unlock(&sched_core_mutex);
345 void sched_core_put(void)
347 static DECLARE_WORK(_work, __sched_core_put);
350 * "There can be only one"
352 * Either this is the last one, or we don't actually need to do any
353 * 'work'. If it is the last *again*, we rely on
354 * WORK_STRUCT_PENDING_BIT.
356 if (!atomic_add_unless(&sched_core_count, -1, 1))
357 schedule_work(&_work);
360 #else /* !CONFIG_SCHED_CORE */
362 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
363 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
365 #endif /* CONFIG_SCHED_CORE */
368 * part of the period that we allow rt tasks to run in us.
371 int sysctl_sched_rt_runtime = 950000;
375 * Serialization rules:
381 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
384 * rq2->lock where: rq1 < rq2
388 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
389 * local CPU's rq->lock, it optionally removes the task from the runqueue and
390 * always looks at the local rq data structures to find the most eligible task
393 * Task enqueue is also under rq->lock, possibly taken from another CPU.
394 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
395 * the local CPU to avoid bouncing the runqueue state around [ see
396 * ttwu_queue_wakelist() ]
398 * Task wakeup, specifically wakeups that involve migration, are horribly
399 * complicated to avoid having to take two rq->locks.
403 * System-calls and anything external will use task_rq_lock() which acquires
404 * both p->pi_lock and rq->lock. As a consequence the state they change is
405 * stable while holding either lock:
407 * - sched_setaffinity()/
408 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
409 * - set_user_nice(): p->se.load, p->*prio
410 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
411 * p->se.load, p->rt_priority,
412 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
413 * - sched_setnuma(): p->numa_preferred_nid
414 * - sched_move_task()/
415 * cpu_cgroup_fork(): p->sched_task_group
416 * - uclamp_update_active() p->uclamp*
418 * p->state <- TASK_*:
420 * is changed locklessly using set_current_state(), __set_current_state() or
421 * set_special_state(), see their respective comments, or by
422 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
425 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
427 * is set by activate_task() and cleared by deactivate_task(), under
428 * rq->lock. Non-zero indicates the task is runnable, the special
429 * ON_RQ_MIGRATING state is used for migration without holding both
430 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
432 * p->on_cpu <- { 0, 1 }:
434 * is set by prepare_task() and cleared by finish_task() such that it will be
435 * set before p is scheduled-in and cleared after p is scheduled-out, both
436 * under rq->lock. Non-zero indicates the task is running on its CPU.
438 * [ The astute reader will observe that it is possible for two tasks on one
439 * CPU to have ->on_cpu = 1 at the same time. ]
441 * task_cpu(p): is changed by set_task_cpu(), the rules are:
443 * - Don't call set_task_cpu() on a blocked task:
445 * We don't care what CPU we're not running on, this simplifies hotplug,
446 * the CPU assignment of blocked tasks isn't required to be valid.
448 * - for try_to_wake_up(), called under p->pi_lock:
450 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
452 * - for migration called under rq->lock:
453 * [ see task_on_rq_migrating() in task_rq_lock() ]
455 * o move_queued_task()
458 * - for migration called under double_rq_lock():
460 * o __migrate_swap_task()
461 * o push_rt_task() / pull_rt_task()
462 * o push_dl_task() / pull_dl_task()
463 * o dl_task_offline_migration()
467 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
469 raw_spinlock_t *lock;
471 /* Matches synchronize_rcu() in __sched_core_enable() */
473 if (sched_core_disabled()) {
474 raw_spin_lock_nested(&rq->__lock, subclass);
475 /* preempt_count *MUST* be > 1 */
476 preempt_enable_no_resched();
481 lock = __rq_lockp(rq);
482 raw_spin_lock_nested(lock, subclass);
483 if (likely(lock == __rq_lockp(rq))) {
484 /* preempt_count *MUST* be > 1 */
485 preempt_enable_no_resched();
488 raw_spin_unlock(lock);
492 bool raw_spin_rq_trylock(struct rq *rq)
494 raw_spinlock_t *lock;
497 /* Matches synchronize_rcu() in __sched_core_enable() */
499 if (sched_core_disabled()) {
500 ret = raw_spin_trylock(&rq->__lock);
506 lock = __rq_lockp(rq);
507 ret = raw_spin_trylock(lock);
508 if (!ret || (likely(lock == __rq_lockp(rq)))) {
512 raw_spin_unlock(lock);
516 void raw_spin_rq_unlock(struct rq *rq)
518 raw_spin_unlock(rq_lockp(rq));
523 * double_rq_lock - safely lock two runqueues
525 void double_rq_lock(struct rq *rq1, struct rq *rq2)
527 lockdep_assert_irqs_disabled();
529 if (rq_order_less(rq2, rq1))
532 raw_spin_rq_lock(rq1);
533 if (__rq_lockp(rq1) == __rq_lockp(rq2))
536 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
541 * __task_rq_lock - lock the rq @p resides on.
543 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
548 lockdep_assert_held(&p->pi_lock);
552 raw_spin_rq_lock(rq);
553 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
557 raw_spin_rq_unlock(rq);
559 while (unlikely(task_on_rq_migrating(p)))
565 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
567 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
568 __acquires(p->pi_lock)
574 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
576 raw_spin_rq_lock(rq);
578 * move_queued_task() task_rq_lock()
581 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
582 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
583 * [S] ->cpu = new_cpu [L] task_rq()
587 * If we observe the old CPU in task_rq_lock(), the acquire of
588 * the old rq->lock will fully serialize against the stores.
590 * If we observe the new CPU in task_rq_lock(), the address
591 * dependency headed by '[L] rq = task_rq()' and the acquire
592 * will pair with the WMB to ensure we then also see migrating.
594 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
598 raw_spin_rq_unlock(rq);
599 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
601 while (unlikely(task_on_rq_migrating(p)))
607 * RQ-clock updating methods:
610 static void update_rq_clock_task(struct rq *rq, s64 delta)
613 * In theory, the compile should just see 0 here, and optimize out the call
614 * to sched_rt_avg_update. But I don't trust it...
616 s64 __maybe_unused steal = 0, irq_delta = 0;
618 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
619 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
622 * Since irq_time is only updated on {soft,}irq_exit, we might run into
623 * this case when a previous update_rq_clock() happened inside a
626 * When this happens, we stop ->clock_task and only update the
627 * prev_irq_time stamp to account for the part that fit, so that a next
628 * update will consume the rest. This ensures ->clock_task is
631 * It does however cause some slight miss-attribution of {soft,}irq
632 * time, a more accurate solution would be to update the irq_time using
633 * the current rq->clock timestamp, except that would require using
636 if (irq_delta > delta)
639 rq->prev_irq_time += irq_delta;
642 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
643 if (static_key_false((¶virt_steal_rq_enabled))) {
644 steal = paravirt_steal_clock(cpu_of(rq));
645 steal -= rq->prev_steal_time_rq;
647 if (unlikely(steal > delta))
650 rq->prev_steal_time_rq += steal;
655 rq->clock_task += delta;
657 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
658 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
659 update_irq_load_avg(rq, irq_delta + steal);
661 update_rq_clock_pelt(rq, delta);
664 void update_rq_clock(struct rq *rq)
668 lockdep_assert_rq_held(rq);
670 if (rq->clock_update_flags & RQCF_ACT_SKIP)
673 #ifdef CONFIG_SCHED_DEBUG
674 if (sched_feat(WARN_DOUBLE_CLOCK))
675 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
676 rq->clock_update_flags |= RQCF_UPDATED;
679 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
683 update_rq_clock_task(rq, delta);
686 #ifdef CONFIG_SCHED_HRTICK
688 * Use HR-timers to deliver accurate preemption points.
691 static void hrtick_clear(struct rq *rq)
693 if (hrtimer_active(&rq->hrtick_timer))
694 hrtimer_cancel(&rq->hrtick_timer);
698 * High-resolution timer tick.
699 * Runs from hardirq context with interrupts disabled.
701 static enum hrtimer_restart hrtick(struct hrtimer *timer)
703 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
706 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
710 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
713 return HRTIMER_NORESTART;
718 static void __hrtick_restart(struct rq *rq)
720 struct hrtimer *timer = &rq->hrtick_timer;
721 ktime_t time = rq->hrtick_time;
723 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
727 * called from hardirq (IPI) context
729 static void __hrtick_start(void *arg)
735 __hrtick_restart(rq);
740 * Called to set the hrtick timer state.
742 * called with rq->lock held and irqs disabled
744 void hrtick_start(struct rq *rq, u64 delay)
746 struct hrtimer *timer = &rq->hrtick_timer;
750 * Don't schedule slices shorter than 10000ns, that just
751 * doesn't make sense and can cause timer DoS.
753 delta = max_t(s64, delay, 10000LL);
754 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
757 __hrtick_restart(rq);
759 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
764 * Called to set the hrtick timer state.
766 * called with rq->lock held and irqs disabled
768 void hrtick_start(struct rq *rq, u64 delay)
771 * Don't schedule slices shorter than 10000ns, that just
772 * doesn't make sense. Rely on vruntime for fairness.
774 delay = max_t(u64, delay, 10000LL);
775 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
776 HRTIMER_MODE_REL_PINNED_HARD);
779 #endif /* CONFIG_SMP */
781 static void hrtick_rq_init(struct rq *rq)
784 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
786 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
787 rq->hrtick_timer.function = hrtick;
789 #else /* CONFIG_SCHED_HRTICK */
790 static inline void hrtick_clear(struct rq *rq)
794 static inline void hrtick_rq_init(struct rq *rq)
797 #endif /* CONFIG_SCHED_HRTICK */
800 * cmpxchg based fetch_or, macro so it works for different integer types
802 #define fetch_or(ptr, mask) \
804 typeof(ptr) _ptr = (ptr); \
805 typeof(mask) _mask = (mask); \
806 typeof(*_ptr) _old, _val = *_ptr; \
809 _old = cmpxchg(_ptr, _val, _val | _mask); \
817 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
819 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
820 * this avoids any races wrt polling state changes and thereby avoids
823 static bool set_nr_and_not_polling(struct task_struct *p)
825 struct thread_info *ti = task_thread_info(p);
826 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
830 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
832 * If this returns true, then the idle task promises to call
833 * sched_ttwu_pending() and reschedule soon.
835 static bool set_nr_if_polling(struct task_struct *p)
837 struct thread_info *ti = task_thread_info(p);
838 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
841 if (!(val & _TIF_POLLING_NRFLAG))
843 if (val & _TIF_NEED_RESCHED)
845 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
854 static bool set_nr_and_not_polling(struct task_struct *p)
856 set_tsk_need_resched(p);
861 static bool set_nr_if_polling(struct task_struct *p)
868 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
870 struct wake_q_node *node = &task->wake_q;
873 * Atomically grab the task, if ->wake_q is !nil already it means
874 * it's already queued (either by us or someone else) and will get the
875 * wakeup due to that.
877 * In order to ensure that a pending wakeup will observe our pending
878 * state, even in the failed case, an explicit smp_mb() must be used.
880 smp_mb__before_atomic();
881 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
885 * The head is context local, there can be no concurrency.
888 head->lastp = &node->next;
893 * wake_q_add() - queue a wakeup for 'later' waking.
894 * @head: the wake_q_head to add @task to
895 * @task: the task to queue for 'later' wakeup
897 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
898 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
901 * This function must be used as-if it were wake_up_process(); IOW the task
902 * must be ready to be woken at this location.
904 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
906 if (__wake_q_add(head, task))
907 get_task_struct(task);
911 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
912 * @head: the wake_q_head to add @task to
913 * @task: the task to queue for 'later' wakeup
915 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
916 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
919 * This function must be used as-if it were wake_up_process(); IOW the task
920 * must be ready to be woken at this location.
922 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
923 * that already hold reference to @task can call the 'safe' version and trust
924 * wake_q to do the right thing depending whether or not the @task is already
927 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
929 if (!__wake_q_add(head, task))
930 put_task_struct(task);
933 void wake_up_q(struct wake_q_head *head)
935 struct wake_q_node *node = head->first;
937 while (node != WAKE_Q_TAIL) {
938 struct task_struct *task;
940 task = container_of(node, struct task_struct, wake_q);
941 /* Task can safely be re-inserted now: */
943 task->wake_q.next = NULL;
946 * wake_up_process() executes a full barrier, which pairs with
947 * the queueing in wake_q_add() so as not to miss wakeups.
949 wake_up_process(task);
950 put_task_struct(task);
955 * resched_curr - mark rq's current task 'to be rescheduled now'.
957 * On UP this means the setting of the need_resched flag, on SMP it
958 * might also involve a cross-CPU call to trigger the scheduler on
961 void resched_curr(struct rq *rq)
963 struct task_struct *curr = rq->curr;
966 lockdep_assert_rq_held(rq);
968 if (test_tsk_need_resched(curr))
973 if (cpu == smp_processor_id()) {
974 set_tsk_need_resched(curr);
975 set_preempt_need_resched();
979 if (set_nr_and_not_polling(curr))
980 smp_send_reschedule(cpu);
982 trace_sched_wake_idle_without_ipi(cpu);
985 void resched_cpu(int cpu)
987 struct rq *rq = cpu_rq(cpu);
990 raw_spin_rq_lock_irqsave(rq, flags);
991 if (cpu_online(cpu) || cpu == smp_processor_id())
993 raw_spin_rq_unlock_irqrestore(rq, flags);
997 #ifdef CONFIG_NO_HZ_COMMON
999 * In the semi idle case, use the nearest busy CPU for migrating timers
1000 * from an idle CPU. This is good for power-savings.
1002 * We don't do similar optimization for completely idle system, as
1003 * selecting an idle CPU will add more delays to the timers than intended
1004 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1006 int get_nohz_timer_target(void)
1008 int i, cpu = smp_processor_id(), default_cpu = -1;
1009 struct sched_domain *sd;
1010 const struct cpumask *hk_mask;
1012 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1018 hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1021 for_each_domain(cpu, sd) {
1022 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1033 if (default_cpu == -1)
1034 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1042 * When add_timer_on() enqueues a timer into the timer wheel of an
1043 * idle CPU then this timer might expire before the next timer event
1044 * which is scheduled to wake up that CPU. In case of a completely
1045 * idle system the next event might even be infinite time into the
1046 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1047 * leaves the inner idle loop so the newly added timer is taken into
1048 * account when the CPU goes back to idle and evaluates the timer
1049 * wheel for the next timer event.
1051 static void wake_up_idle_cpu(int cpu)
1053 struct rq *rq = cpu_rq(cpu);
1055 if (cpu == smp_processor_id())
1058 if (set_nr_and_not_polling(rq->idle))
1059 smp_send_reschedule(cpu);
1061 trace_sched_wake_idle_without_ipi(cpu);
1064 static bool wake_up_full_nohz_cpu(int cpu)
1067 * We just need the target to call irq_exit() and re-evaluate
1068 * the next tick. The nohz full kick at least implies that.
1069 * If needed we can still optimize that later with an
1072 if (cpu_is_offline(cpu))
1073 return true; /* Don't try to wake offline CPUs. */
1074 if (tick_nohz_full_cpu(cpu)) {
1075 if (cpu != smp_processor_id() ||
1076 tick_nohz_tick_stopped())
1077 tick_nohz_full_kick_cpu(cpu);
1085 * Wake up the specified CPU. If the CPU is going offline, it is the
1086 * caller's responsibility to deal with the lost wakeup, for example,
1087 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1089 void wake_up_nohz_cpu(int cpu)
1091 if (!wake_up_full_nohz_cpu(cpu))
1092 wake_up_idle_cpu(cpu);
1095 static void nohz_csd_func(void *info)
1097 struct rq *rq = info;
1098 int cpu = cpu_of(rq);
1102 * Release the rq::nohz_csd.
1104 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1105 WARN_ON(!(flags & NOHZ_KICK_MASK));
1107 rq->idle_balance = idle_cpu(cpu);
1108 if (rq->idle_balance && !need_resched()) {
1109 rq->nohz_idle_balance = flags;
1110 raise_softirq_irqoff(SCHED_SOFTIRQ);
1114 #endif /* CONFIG_NO_HZ_COMMON */
1116 #ifdef CONFIG_NO_HZ_FULL
1117 bool sched_can_stop_tick(struct rq *rq)
1119 int fifo_nr_running;
1121 /* Deadline tasks, even if single, need the tick */
1122 if (rq->dl.dl_nr_running)
1126 * If there are more than one RR tasks, we need the tick to affect the
1127 * actual RR behaviour.
1129 if (rq->rt.rr_nr_running) {
1130 if (rq->rt.rr_nr_running == 1)
1137 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1138 * forced preemption between FIFO tasks.
1140 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1141 if (fifo_nr_running)
1145 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1146 * if there's more than one we need the tick for involuntary
1149 if (rq->nr_running > 1)
1154 #endif /* CONFIG_NO_HZ_FULL */
1155 #endif /* CONFIG_SMP */
1157 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1158 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1160 * Iterate task_group tree rooted at *from, calling @down when first entering a
1161 * node and @up when leaving it for the final time.
1163 * Caller must hold rcu_lock or sufficient equivalent.
1165 int walk_tg_tree_from(struct task_group *from,
1166 tg_visitor down, tg_visitor up, void *data)
1168 struct task_group *parent, *child;
1174 ret = (*down)(parent, data);
1177 list_for_each_entry_rcu(child, &parent->children, siblings) {
1184 ret = (*up)(parent, data);
1185 if (ret || parent == from)
1189 parent = parent->parent;
1196 int tg_nop(struct task_group *tg, void *data)
1202 static void set_load_weight(struct task_struct *p, bool update_load)
1204 int prio = p->static_prio - MAX_RT_PRIO;
1205 struct load_weight *load = &p->se.load;
1208 * SCHED_IDLE tasks get minimal weight:
1210 if (task_has_idle_policy(p)) {
1211 load->weight = scale_load(WEIGHT_IDLEPRIO);
1212 load->inv_weight = WMULT_IDLEPRIO;
1217 * SCHED_OTHER tasks have to update their load when changing their
1220 if (update_load && p->sched_class == &fair_sched_class) {
1221 reweight_task(p, prio);
1223 load->weight = scale_load(sched_prio_to_weight[prio]);
1224 load->inv_weight = sched_prio_to_wmult[prio];
1228 #ifdef CONFIG_UCLAMP_TASK
1230 * Serializes updates of utilization clamp values
1232 * The (slow-path) user-space triggers utilization clamp value updates which
1233 * can require updates on (fast-path) scheduler's data structures used to
1234 * support enqueue/dequeue operations.
1235 * While the per-CPU rq lock protects fast-path update operations, user-space
1236 * requests are serialized using a mutex to reduce the risk of conflicting
1237 * updates or API abuses.
1239 static DEFINE_MUTEX(uclamp_mutex);
1241 /* Max allowed minimum utilization */
1242 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1244 /* Max allowed maximum utilization */
1245 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1248 * By default RT tasks run at the maximum performance point/capacity of the
1249 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1250 * SCHED_CAPACITY_SCALE.
1252 * This knob allows admins to change the default behavior when uclamp is being
1253 * used. In battery powered devices, particularly, running at the maximum
1254 * capacity and frequency will increase energy consumption and shorten the
1257 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1259 * This knob will not override the system default sched_util_clamp_min defined
1262 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1264 /* All clamps are required to be less or equal than these values */
1265 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1268 * This static key is used to reduce the uclamp overhead in the fast path. It
1269 * primarily disables the call to uclamp_rq_{inc, dec}() in
1270 * enqueue/dequeue_task().
1272 * This allows users to continue to enable uclamp in their kernel config with
1273 * minimum uclamp overhead in the fast path.
1275 * As soon as userspace modifies any of the uclamp knobs, the static key is
1276 * enabled, since we have an actual users that make use of uclamp
1279 * The knobs that would enable this static key are:
1281 * * A task modifying its uclamp value with sched_setattr().
1282 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1283 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1285 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1287 /* Integer rounded range for each bucket */
1288 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1290 #define for_each_clamp_id(clamp_id) \
1291 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1293 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1295 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1298 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1300 if (clamp_id == UCLAMP_MIN)
1302 return SCHED_CAPACITY_SCALE;
1305 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1306 unsigned int value, bool user_defined)
1308 uc_se->value = value;
1309 uc_se->bucket_id = uclamp_bucket_id(value);
1310 uc_se->user_defined = user_defined;
1313 static inline unsigned int
1314 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1315 unsigned int clamp_value)
1318 * Avoid blocked utilization pushing up the frequency when we go
1319 * idle (which drops the max-clamp) by retaining the last known
1322 if (clamp_id == UCLAMP_MAX) {
1323 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1327 return uclamp_none(UCLAMP_MIN);
1330 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1331 unsigned int clamp_value)
1333 /* Reset max-clamp retention only on idle exit */
1334 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1337 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1341 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1342 unsigned int clamp_value)
1344 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1345 int bucket_id = UCLAMP_BUCKETS - 1;
1348 * Since both min and max clamps are max aggregated, find the
1349 * top most bucket with tasks in.
1351 for ( ; bucket_id >= 0; bucket_id--) {
1352 if (!bucket[bucket_id].tasks)
1354 return bucket[bucket_id].value;
1357 /* No tasks -- default clamp values */
1358 return uclamp_idle_value(rq, clamp_id, clamp_value);
1361 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1363 unsigned int default_util_min;
1364 struct uclamp_se *uc_se;
1366 lockdep_assert_held(&p->pi_lock);
1368 uc_se = &p->uclamp_req[UCLAMP_MIN];
1370 /* Only sync if user didn't override the default */
1371 if (uc_se->user_defined)
1374 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1375 uclamp_se_set(uc_se, default_util_min, false);
1378 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1386 /* Protect updates to p->uclamp_* */
1387 rq = task_rq_lock(p, &rf);
1388 __uclamp_update_util_min_rt_default(p);
1389 task_rq_unlock(rq, p, &rf);
1392 static void uclamp_sync_util_min_rt_default(void)
1394 struct task_struct *g, *p;
1397 * copy_process() sysctl_uclamp
1398 * uclamp_min_rt = X;
1399 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1400 * // link thread smp_mb__after_spinlock()
1401 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1402 * sched_post_fork() for_each_process_thread()
1403 * __uclamp_sync_rt() __uclamp_sync_rt()
1405 * Ensures that either sched_post_fork() will observe the new
1406 * uclamp_min_rt or for_each_process_thread() will observe the new
1409 read_lock(&tasklist_lock);
1410 smp_mb__after_spinlock();
1411 read_unlock(&tasklist_lock);
1414 for_each_process_thread(g, p)
1415 uclamp_update_util_min_rt_default(p);
1419 static inline struct uclamp_se
1420 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1422 /* Copy by value as we could modify it */
1423 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1424 #ifdef CONFIG_UCLAMP_TASK_GROUP
1425 unsigned int tg_min, tg_max, value;
1428 * Tasks in autogroups or root task group will be
1429 * restricted by system defaults.
1431 if (task_group_is_autogroup(task_group(p)))
1433 if (task_group(p) == &root_task_group)
1436 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1437 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1438 value = uc_req.value;
1439 value = clamp(value, tg_min, tg_max);
1440 uclamp_se_set(&uc_req, value, false);
1447 * The effective clamp bucket index of a task depends on, by increasing
1449 * - the task specific clamp value, when explicitly requested from userspace
1450 * - the task group effective clamp value, for tasks not either in the root
1451 * group or in an autogroup
1452 * - the system default clamp value, defined by the sysadmin
1454 static inline struct uclamp_se
1455 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1457 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1458 struct uclamp_se uc_max = uclamp_default[clamp_id];
1460 /* System default restrictions always apply */
1461 if (unlikely(uc_req.value > uc_max.value))
1467 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1469 struct uclamp_se uc_eff;
1471 /* Task currently refcounted: use back-annotated (effective) value */
1472 if (p->uclamp[clamp_id].active)
1473 return (unsigned long)p->uclamp[clamp_id].value;
1475 uc_eff = uclamp_eff_get(p, clamp_id);
1477 return (unsigned long)uc_eff.value;
1481 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1482 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1483 * updates the rq's clamp value if required.
1485 * Tasks can have a task-specific value requested from user-space, track
1486 * within each bucket the maximum value for tasks refcounted in it.
1487 * This "local max aggregation" allows to track the exact "requested" value
1488 * for each bucket when all its RUNNABLE tasks require the same clamp.
1490 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1491 enum uclamp_id clamp_id)
1493 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1494 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1495 struct uclamp_bucket *bucket;
1497 lockdep_assert_rq_held(rq);
1499 /* Update task effective clamp */
1500 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1502 bucket = &uc_rq->bucket[uc_se->bucket_id];
1504 uc_se->active = true;
1506 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1509 * Local max aggregation: rq buckets always track the max
1510 * "requested" clamp value of its RUNNABLE tasks.
1512 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1513 bucket->value = uc_se->value;
1515 if (uc_se->value > READ_ONCE(uc_rq->value))
1516 WRITE_ONCE(uc_rq->value, uc_se->value);
1520 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1521 * is released. If this is the last task reference counting the rq's max
1522 * active clamp value, then the rq's clamp value is updated.
1524 * Both refcounted tasks and rq's cached clamp values are expected to be
1525 * always valid. If it's detected they are not, as defensive programming,
1526 * enforce the expected state and warn.
1528 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1529 enum uclamp_id clamp_id)
1531 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1532 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1533 struct uclamp_bucket *bucket;
1534 unsigned int bkt_clamp;
1535 unsigned int rq_clamp;
1537 lockdep_assert_rq_held(rq);
1540 * If sched_uclamp_used was enabled after task @p was enqueued,
1541 * we could end up with unbalanced call to uclamp_rq_dec_id().
1543 * In this case the uc_se->active flag should be false since no uclamp
1544 * accounting was performed at enqueue time and we can just return
1547 * Need to be careful of the following enqueue/dequeue ordering
1551 * // sched_uclamp_used gets enabled
1554 * // Must not decrement bucket->tasks here
1557 * where we could end up with stale data in uc_se and
1558 * bucket[uc_se->bucket_id].
1560 * The following check here eliminates the possibility of such race.
1562 if (unlikely(!uc_se->active))
1565 bucket = &uc_rq->bucket[uc_se->bucket_id];
1567 SCHED_WARN_ON(!bucket->tasks);
1568 if (likely(bucket->tasks))
1571 uc_se->active = false;
1574 * Keep "local max aggregation" simple and accept to (possibly)
1575 * overboost some RUNNABLE tasks in the same bucket.
1576 * The rq clamp bucket value is reset to its base value whenever
1577 * there are no more RUNNABLE tasks refcounting it.
1579 if (likely(bucket->tasks))
1582 rq_clamp = READ_ONCE(uc_rq->value);
1584 * Defensive programming: this should never happen. If it happens,
1585 * e.g. due to future modification, warn and fixup the expected value.
1587 SCHED_WARN_ON(bucket->value > rq_clamp);
1588 if (bucket->value >= rq_clamp) {
1589 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1590 WRITE_ONCE(uc_rq->value, bkt_clamp);
1594 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1596 enum uclamp_id clamp_id;
1599 * Avoid any overhead until uclamp is actually used by the userspace.
1601 * The condition is constructed such that a NOP is generated when
1602 * sched_uclamp_used is disabled.
1604 if (!static_branch_unlikely(&sched_uclamp_used))
1607 if (unlikely(!p->sched_class->uclamp_enabled))
1610 for_each_clamp_id(clamp_id)
1611 uclamp_rq_inc_id(rq, p, clamp_id);
1613 /* Reset clamp idle holding when there is one RUNNABLE task */
1614 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1615 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1618 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1620 enum uclamp_id clamp_id;
1623 * Avoid any overhead until uclamp is actually used by the userspace.
1625 * The condition is constructed such that a NOP is generated when
1626 * sched_uclamp_used is disabled.
1628 if (!static_branch_unlikely(&sched_uclamp_used))
1631 if (unlikely(!p->sched_class->uclamp_enabled))
1634 for_each_clamp_id(clamp_id)
1635 uclamp_rq_dec_id(rq, p, clamp_id);
1638 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1639 enum uclamp_id clamp_id)
1641 if (!p->uclamp[clamp_id].active)
1644 uclamp_rq_dec_id(rq, p, clamp_id);
1645 uclamp_rq_inc_id(rq, p, clamp_id);
1648 * Make sure to clear the idle flag if we've transiently reached 0
1649 * active tasks on rq.
1651 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1652 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1656 uclamp_update_active(struct task_struct *p)
1658 enum uclamp_id clamp_id;
1663 * Lock the task and the rq where the task is (or was) queued.
1665 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1666 * price to pay to safely serialize util_{min,max} updates with
1667 * enqueues, dequeues and migration operations.
1668 * This is the same locking schema used by __set_cpus_allowed_ptr().
1670 rq = task_rq_lock(p, &rf);
1673 * Setting the clamp bucket is serialized by task_rq_lock().
1674 * If the task is not yet RUNNABLE and its task_struct is not
1675 * affecting a valid clamp bucket, the next time it's enqueued,
1676 * it will already see the updated clamp bucket value.
1678 for_each_clamp_id(clamp_id)
1679 uclamp_rq_reinc_id(rq, p, clamp_id);
1681 task_rq_unlock(rq, p, &rf);
1684 #ifdef CONFIG_UCLAMP_TASK_GROUP
1686 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1688 struct css_task_iter it;
1689 struct task_struct *p;
1691 css_task_iter_start(css, 0, &it);
1692 while ((p = css_task_iter_next(&it)))
1693 uclamp_update_active(p);
1694 css_task_iter_end(&it);
1697 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1698 static void uclamp_update_root_tg(void)
1700 struct task_group *tg = &root_task_group;
1702 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1703 sysctl_sched_uclamp_util_min, false);
1704 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1705 sysctl_sched_uclamp_util_max, false);
1708 cpu_util_update_eff(&root_task_group.css);
1712 static void uclamp_update_root_tg(void) { }
1715 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1716 void *buffer, size_t *lenp, loff_t *ppos)
1718 bool update_root_tg = false;
1719 int old_min, old_max, old_min_rt;
1722 mutex_lock(&uclamp_mutex);
1723 old_min = sysctl_sched_uclamp_util_min;
1724 old_max = sysctl_sched_uclamp_util_max;
1725 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1727 result = proc_dointvec(table, write, buffer, lenp, ppos);
1733 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1734 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1735 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1741 if (old_min != sysctl_sched_uclamp_util_min) {
1742 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1743 sysctl_sched_uclamp_util_min, false);
1744 update_root_tg = true;
1746 if (old_max != sysctl_sched_uclamp_util_max) {
1747 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1748 sysctl_sched_uclamp_util_max, false);
1749 update_root_tg = true;
1752 if (update_root_tg) {
1753 static_branch_enable(&sched_uclamp_used);
1754 uclamp_update_root_tg();
1757 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1758 static_branch_enable(&sched_uclamp_used);
1759 uclamp_sync_util_min_rt_default();
1763 * We update all RUNNABLE tasks only when task groups are in use.
1764 * Otherwise, keep it simple and do just a lazy update at each next
1765 * task enqueue time.
1771 sysctl_sched_uclamp_util_min = old_min;
1772 sysctl_sched_uclamp_util_max = old_max;
1773 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1775 mutex_unlock(&uclamp_mutex);
1780 static int uclamp_validate(struct task_struct *p,
1781 const struct sched_attr *attr)
1783 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1784 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1786 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1787 util_min = attr->sched_util_min;
1789 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1793 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1794 util_max = attr->sched_util_max;
1796 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1800 if (util_min != -1 && util_max != -1 && util_min > util_max)
1804 * We have valid uclamp attributes; make sure uclamp is enabled.
1806 * We need to do that here, because enabling static branches is a
1807 * blocking operation which obviously cannot be done while holding
1810 static_branch_enable(&sched_uclamp_used);
1815 static bool uclamp_reset(const struct sched_attr *attr,
1816 enum uclamp_id clamp_id,
1817 struct uclamp_se *uc_se)
1819 /* Reset on sched class change for a non user-defined clamp value. */
1820 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1821 !uc_se->user_defined)
1824 /* Reset on sched_util_{min,max} == -1. */
1825 if (clamp_id == UCLAMP_MIN &&
1826 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1827 attr->sched_util_min == -1) {
1831 if (clamp_id == UCLAMP_MAX &&
1832 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1833 attr->sched_util_max == -1) {
1840 static void __setscheduler_uclamp(struct task_struct *p,
1841 const struct sched_attr *attr)
1843 enum uclamp_id clamp_id;
1845 for_each_clamp_id(clamp_id) {
1846 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1849 if (!uclamp_reset(attr, clamp_id, uc_se))
1853 * RT by default have a 100% boost value that could be modified
1856 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1857 value = sysctl_sched_uclamp_util_min_rt_default;
1859 value = uclamp_none(clamp_id);
1861 uclamp_se_set(uc_se, value, false);
1865 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1868 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1869 attr->sched_util_min != -1) {
1870 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1871 attr->sched_util_min, true);
1874 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1875 attr->sched_util_max != -1) {
1876 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1877 attr->sched_util_max, true);
1881 static void uclamp_fork(struct task_struct *p)
1883 enum uclamp_id clamp_id;
1886 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1887 * as the task is still at its early fork stages.
1889 for_each_clamp_id(clamp_id)
1890 p->uclamp[clamp_id].active = false;
1892 if (likely(!p->sched_reset_on_fork))
1895 for_each_clamp_id(clamp_id) {
1896 uclamp_se_set(&p->uclamp_req[clamp_id],
1897 uclamp_none(clamp_id), false);
1901 static void uclamp_post_fork(struct task_struct *p)
1903 uclamp_update_util_min_rt_default(p);
1906 static void __init init_uclamp_rq(struct rq *rq)
1908 enum uclamp_id clamp_id;
1909 struct uclamp_rq *uc_rq = rq->uclamp;
1911 for_each_clamp_id(clamp_id) {
1912 uc_rq[clamp_id] = (struct uclamp_rq) {
1913 .value = uclamp_none(clamp_id)
1917 rq->uclamp_flags = 0;
1920 static void __init init_uclamp(void)
1922 struct uclamp_se uc_max = {};
1923 enum uclamp_id clamp_id;
1926 for_each_possible_cpu(cpu)
1927 init_uclamp_rq(cpu_rq(cpu));
1929 for_each_clamp_id(clamp_id) {
1930 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1931 uclamp_none(clamp_id), false);
1934 /* System defaults allow max clamp values for both indexes */
1935 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1936 for_each_clamp_id(clamp_id) {
1937 uclamp_default[clamp_id] = uc_max;
1938 #ifdef CONFIG_UCLAMP_TASK_GROUP
1939 root_task_group.uclamp_req[clamp_id] = uc_max;
1940 root_task_group.uclamp[clamp_id] = uc_max;
1945 #else /* CONFIG_UCLAMP_TASK */
1946 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1947 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1948 static inline int uclamp_validate(struct task_struct *p,
1949 const struct sched_attr *attr)
1953 static void __setscheduler_uclamp(struct task_struct *p,
1954 const struct sched_attr *attr) { }
1955 static inline void uclamp_fork(struct task_struct *p) { }
1956 static inline void uclamp_post_fork(struct task_struct *p) { }
1957 static inline void init_uclamp(void) { }
1958 #endif /* CONFIG_UCLAMP_TASK */
1960 bool sched_task_on_rq(struct task_struct *p)
1962 return task_on_rq_queued(p);
1965 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1967 if (!(flags & ENQUEUE_NOCLOCK))
1968 update_rq_clock(rq);
1970 if (!(flags & ENQUEUE_RESTORE)) {
1971 sched_info_enqueue(rq, p);
1972 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1975 uclamp_rq_inc(rq, p);
1976 p->sched_class->enqueue_task(rq, p, flags);
1978 if (sched_core_enabled(rq))
1979 sched_core_enqueue(rq, p);
1982 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1984 if (sched_core_enabled(rq))
1985 sched_core_dequeue(rq, p);
1987 if (!(flags & DEQUEUE_NOCLOCK))
1988 update_rq_clock(rq);
1990 if (!(flags & DEQUEUE_SAVE)) {
1991 sched_info_dequeue(rq, p);
1992 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1995 uclamp_rq_dec(rq, p);
1996 p->sched_class->dequeue_task(rq, p, flags);
1999 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2001 enqueue_task(rq, p, flags);
2003 p->on_rq = TASK_ON_RQ_QUEUED;
2006 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2008 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2010 dequeue_task(rq, p, flags);
2013 static inline int __normal_prio(int policy, int rt_prio, int nice)
2017 if (dl_policy(policy))
2018 prio = MAX_DL_PRIO - 1;
2019 else if (rt_policy(policy))
2020 prio = MAX_RT_PRIO - 1 - rt_prio;
2022 prio = NICE_TO_PRIO(nice);
2028 * Calculate the expected normal priority: i.e. priority
2029 * without taking RT-inheritance into account. Might be
2030 * boosted by interactivity modifiers. Changes upon fork,
2031 * setprio syscalls, and whenever the interactivity
2032 * estimator recalculates.
2034 static inline int normal_prio(struct task_struct *p)
2036 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2040 * Calculate the current priority, i.e. the priority
2041 * taken into account by the scheduler. This value might
2042 * be boosted by RT tasks, or might be boosted by
2043 * interactivity modifiers. Will be RT if the task got
2044 * RT-boosted. If not then it returns p->normal_prio.
2046 static int effective_prio(struct task_struct *p)
2048 p->normal_prio = normal_prio(p);
2050 * If we are RT tasks or we were boosted to RT priority,
2051 * keep the priority unchanged. Otherwise, update priority
2052 * to the normal priority:
2054 if (!rt_prio(p->prio))
2055 return p->normal_prio;
2060 * task_curr - is this task currently executing on a CPU?
2061 * @p: the task in question.
2063 * Return: 1 if the task is currently executing. 0 otherwise.
2065 inline int task_curr(const struct task_struct *p)
2067 return cpu_curr(task_cpu(p)) == p;
2071 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2072 * use the balance_callback list if you want balancing.
2074 * this means any call to check_class_changed() must be followed by a call to
2075 * balance_callback().
2077 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2078 const struct sched_class *prev_class,
2081 if (prev_class != p->sched_class) {
2082 if (prev_class->switched_from)
2083 prev_class->switched_from(rq, p);
2085 p->sched_class->switched_to(rq, p);
2086 } else if (oldprio != p->prio || dl_task(p))
2087 p->sched_class->prio_changed(rq, p, oldprio);
2090 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2092 if (p->sched_class == rq->curr->sched_class)
2093 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2094 else if (p->sched_class > rq->curr->sched_class)
2098 * A queue event has occurred, and we're going to schedule. In
2099 * this case, we can save a useless back to back clock update.
2101 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2102 rq_clock_skip_update(rq);
2108 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2110 static int __set_cpus_allowed_ptr(struct task_struct *p,
2111 const struct cpumask *new_mask,
2114 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2116 if (likely(!p->migration_disabled))
2119 if (p->cpus_ptr != &p->cpus_mask)
2123 * Violates locking rules! see comment in __do_set_cpus_allowed().
2125 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2128 void migrate_disable(void)
2130 struct task_struct *p = current;
2132 if (p->migration_disabled) {
2133 p->migration_disabled++;
2138 this_rq()->nr_pinned++;
2139 p->migration_disabled = 1;
2142 EXPORT_SYMBOL_GPL(migrate_disable);
2144 void migrate_enable(void)
2146 struct task_struct *p = current;
2148 if (p->migration_disabled > 1) {
2149 p->migration_disabled--;
2154 * Ensure stop_task runs either before or after this, and that
2155 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2158 if (p->cpus_ptr != &p->cpus_mask)
2159 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2161 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2162 * regular cpus_mask, otherwise things that race (eg.
2163 * select_fallback_rq) get confused.
2166 p->migration_disabled = 0;
2167 this_rq()->nr_pinned--;
2170 EXPORT_SYMBOL_GPL(migrate_enable);
2172 static inline bool rq_has_pinned_tasks(struct rq *rq)
2174 return rq->nr_pinned;
2178 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2179 * __set_cpus_allowed_ptr() and select_fallback_rq().
2181 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2183 /* When not in the task's cpumask, no point in looking further. */
2184 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2187 /* migrate_disabled() must be allowed to finish. */
2188 if (is_migration_disabled(p))
2189 return cpu_online(cpu);
2191 /* Non kernel threads are not allowed during either online or offline. */
2192 if (!(p->flags & PF_KTHREAD))
2193 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2195 /* KTHREAD_IS_PER_CPU is always allowed. */
2196 if (kthread_is_per_cpu(p))
2197 return cpu_online(cpu);
2199 /* Regular kernel threads don't get to stay during offline. */
2203 /* But are allowed during online. */
2204 return cpu_online(cpu);
2208 * This is how migration works:
2210 * 1) we invoke migration_cpu_stop() on the target CPU using
2212 * 2) stopper starts to run (implicitly forcing the migrated thread
2214 * 3) it checks whether the migrated task is still in the wrong runqueue.
2215 * 4) if it's in the wrong runqueue then the migration thread removes
2216 * it and puts it into the right queue.
2217 * 5) stopper completes and stop_one_cpu() returns and the migration
2222 * move_queued_task - move a queued task to new rq.
2224 * Returns (locked) new rq. Old rq's lock is released.
2226 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2227 struct task_struct *p, int new_cpu)
2229 lockdep_assert_rq_held(rq);
2231 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2232 set_task_cpu(p, new_cpu);
2235 rq = cpu_rq(new_cpu);
2238 BUG_ON(task_cpu(p) != new_cpu);
2239 activate_task(rq, p, 0);
2240 check_preempt_curr(rq, p, 0);
2245 struct migration_arg {
2246 struct task_struct *task;
2248 struct set_affinity_pending *pending;
2252 * @refs: number of wait_for_completion()
2253 * @stop_pending: is @stop_work in use
2255 struct set_affinity_pending {
2257 unsigned int stop_pending;
2258 struct completion done;
2259 struct cpu_stop_work stop_work;
2260 struct migration_arg arg;
2264 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2265 * this because either it can't run here any more (set_cpus_allowed()
2266 * away from this CPU, or CPU going down), or because we're
2267 * attempting to rebalance this task on exec (sched_exec).
2269 * So we race with normal scheduler movements, but that's OK, as long
2270 * as the task is no longer on this CPU.
2272 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2273 struct task_struct *p, int dest_cpu)
2275 /* Affinity changed (again). */
2276 if (!is_cpu_allowed(p, dest_cpu))
2279 update_rq_clock(rq);
2280 rq = move_queued_task(rq, rf, p, dest_cpu);
2286 * migration_cpu_stop - this will be executed by a highprio stopper thread
2287 * and performs thread migration by bumping thread off CPU then
2288 * 'pushing' onto another runqueue.
2290 static int migration_cpu_stop(void *data)
2292 struct migration_arg *arg = data;
2293 struct set_affinity_pending *pending = arg->pending;
2294 struct task_struct *p = arg->task;
2295 struct rq *rq = this_rq();
2296 bool complete = false;
2300 * The original target CPU might have gone down and we might
2301 * be on another CPU but it doesn't matter.
2303 local_irq_save(rf.flags);
2305 * We need to explicitly wake pending tasks before running
2306 * __migrate_task() such that we will not miss enforcing cpus_ptr
2307 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2309 flush_smp_call_function_from_idle();
2311 raw_spin_lock(&p->pi_lock);
2315 * If we were passed a pending, then ->stop_pending was set, thus
2316 * p->migration_pending must have remained stable.
2318 WARN_ON_ONCE(pending && pending != p->migration_pending);
2321 * If task_rq(p) != rq, it cannot be migrated here, because we're
2322 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2323 * we're holding p->pi_lock.
2325 if (task_rq(p) == rq) {
2326 if (is_migration_disabled(p))
2330 p->migration_pending = NULL;
2333 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2337 if (task_on_rq_queued(p))
2338 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2340 p->wake_cpu = arg->dest_cpu;
2343 * XXX __migrate_task() can fail, at which point we might end
2344 * up running on a dodgy CPU, AFAICT this can only happen
2345 * during CPU hotplug, at which point we'll get pushed out
2346 * anyway, so it's probably not a big deal.
2349 } else if (pending) {
2351 * This happens when we get migrated between migrate_enable()'s
2352 * preempt_enable() and scheduling the stopper task. At that
2353 * point we're a regular task again and not current anymore.
2355 * A !PREEMPT kernel has a giant hole here, which makes it far
2360 * The task moved before the stopper got to run. We're holding
2361 * ->pi_lock, so the allowed mask is stable - if it got
2362 * somewhere allowed, we're done.
2364 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2365 p->migration_pending = NULL;
2371 * When migrate_enable() hits a rq mis-match we can't reliably
2372 * determine is_migration_disabled() and so have to chase after
2375 WARN_ON_ONCE(!pending->stop_pending);
2376 task_rq_unlock(rq, p, &rf);
2377 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2378 &pending->arg, &pending->stop_work);
2383 pending->stop_pending = false;
2384 task_rq_unlock(rq, p, &rf);
2387 complete_all(&pending->done);
2392 int push_cpu_stop(void *arg)
2394 struct rq *lowest_rq = NULL, *rq = this_rq();
2395 struct task_struct *p = arg;
2397 raw_spin_lock_irq(&p->pi_lock);
2398 raw_spin_rq_lock(rq);
2400 if (task_rq(p) != rq)
2403 if (is_migration_disabled(p)) {
2404 p->migration_flags |= MDF_PUSH;
2408 p->migration_flags &= ~MDF_PUSH;
2410 if (p->sched_class->find_lock_rq)
2411 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2416 // XXX validate p is still the highest prio task
2417 if (task_rq(p) == rq) {
2418 deactivate_task(rq, p, 0);
2419 set_task_cpu(p, lowest_rq->cpu);
2420 activate_task(lowest_rq, p, 0);
2421 resched_curr(lowest_rq);
2424 double_unlock_balance(rq, lowest_rq);
2427 rq->push_busy = false;
2428 raw_spin_rq_unlock(rq);
2429 raw_spin_unlock_irq(&p->pi_lock);
2436 * sched_class::set_cpus_allowed must do the below, but is not required to
2437 * actually call this function.
2439 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2441 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2442 p->cpus_ptr = new_mask;
2446 cpumask_copy(&p->cpus_mask, new_mask);
2447 p->nr_cpus_allowed = cpumask_weight(new_mask);
2451 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2453 struct rq *rq = task_rq(p);
2454 bool queued, running;
2457 * This here violates the locking rules for affinity, since we're only
2458 * supposed to change these variables while holding both rq->lock and
2461 * HOWEVER, it magically works, because ttwu() is the only code that
2462 * accesses these variables under p->pi_lock and only does so after
2463 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2464 * before finish_task().
2466 * XXX do further audits, this smells like something putrid.
2468 if (flags & SCA_MIGRATE_DISABLE)
2469 SCHED_WARN_ON(!p->on_cpu);
2471 lockdep_assert_held(&p->pi_lock);
2473 queued = task_on_rq_queued(p);
2474 running = task_current(rq, p);
2478 * Because __kthread_bind() calls this on blocked tasks without
2481 lockdep_assert_rq_held(rq);
2482 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2485 put_prev_task(rq, p);
2487 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2490 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2492 set_next_task(rq, p);
2495 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2497 __do_set_cpus_allowed(p, new_mask, 0);
2500 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2503 if (!src->user_cpus_ptr)
2506 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2507 if (!dst->user_cpus_ptr)
2510 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2514 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2516 struct cpumask *user_mask = NULL;
2518 swap(p->user_cpus_ptr, user_mask);
2523 void release_user_cpus_ptr(struct task_struct *p)
2525 kfree(clear_user_cpus_ptr(p));
2529 * This function is wildly self concurrent; here be dragons.
2532 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2533 * designated task is enqueued on an allowed CPU. If that task is currently
2534 * running, we have to kick it out using the CPU stopper.
2536 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2539 * Initial conditions: P0->cpus_mask = [0, 1]
2543 * migrate_disable();
2545 * set_cpus_allowed_ptr(P0, [1]);
2547 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2548 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2549 * This means we need the following scheme:
2553 * migrate_disable();
2555 * set_cpus_allowed_ptr(P0, [1]);
2559 * __set_cpus_allowed_ptr();
2560 * <wakes local stopper>
2561 * `--> <woken on migration completion>
2563 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2564 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2565 * task p are serialized by p->pi_lock, which we can leverage: the one that
2566 * should come into effect at the end of the Migrate-Disable region is the last
2567 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2568 * but we still need to properly signal those waiting tasks at the appropriate
2571 * This is implemented using struct set_affinity_pending. The first
2572 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2573 * setup an instance of that struct and install it on the targeted task_struct.
2574 * Any and all further callers will reuse that instance. Those then wait for
2575 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2576 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2579 * (1) In the cases covered above. There is one more where the completion is
2580 * signaled within affine_move_task() itself: when a subsequent affinity request
2581 * occurs after the stopper bailed out due to the targeted task still being
2582 * Migrate-Disable. Consider:
2584 * Initial conditions: P0->cpus_mask = [0, 1]
2588 * migrate_disable();
2590 * set_cpus_allowed_ptr(P0, [1]);
2593 * migration_cpu_stop()
2594 * is_migration_disabled()
2596 * set_cpus_allowed_ptr(P0, [0, 1]);
2597 * <signal completion>
2600 * Note that the above is safe vs a concurrent migrate_enable(), as any
2601 * pending affinity completion is preceded by an uninstallation of
2602 * p->migration_pending done with p->pi_lock held.
2604 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2605 int dest_cpu, unsigned int flags)
2607 struct set_affinity_pending my_pending = { }, *pending = NULL;
2608 bool stop_pending, complete = false;
2610 /* Can the task run on the task's current CPU? If so, we're done */
2611 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2612 struct task_struct *push_task = NULL;
2614 if ((flags & SCA_MIGRATE_ENABLE) &&
2615 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2616 rq->push_busy = true;
2617 push_task = get_task_struct(p);
2621 * If there are pending waiters, but no pending stop_work,
2622 * then complete now.
2624 pending = p->migration_pending;
2625 if (pending && !pending->stop_pending) {
2626 p->migration_pending = NULL;
2630 task_rq_unlock(rq, p, rf);
2633 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2638 complete_all(&pending->done);
2643 if (!(flags & SCA_MIGRATE_ENABLE)) {
2644 /* serialized by p->pi_lock */
2645 if (!p->migration_pending) {
2646 /* Install the request */
2647 refcount_set(&my_pending.refs, 1);
2648 init_completion(&my_pending.done);
2649 my_pending.arg = (struct migration_arg) {
2651 .dest_cpu = dest_cpu,
2652 .pending = &my_pending,
2655 p->migration_pending = &my_pending;
2657 pending = p->migration_pending;
2658 refcount_inc(&pending->refs);
2660 * Affinity has changed, but we've already installed a
2661 * pending. migration_cpu_stop() *must* see this, else
2662 * we risk a completion of the pending despite having a
2663 * task on a disallowed CPU.
2665 * Serialized by p->pi_lock, so this is safe.
2667 pending->arg.dest_cpu = dest_cpu;
2670 pending = p->migration_pending;
2672 * - !MIGRATE_ENABLE:
2673 * we'll have installed a pending if there wasn't one already.
2676 * we're here because the current CPU isn't matching anymore,
2677 * the only way that can happen is because of a concurrent
2678 * set_cpus_allowed_ptr() call, which should then still be
2679 * pending completion.
2681 * Either way, we really should have a @pending here.
2683 if (WARN_ON_ONCE(!pending)) {
2684 task_rq_unlock(rq, p, rf);
2688 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2690 * MIGRATE_ENABLE gets here because 'p == current', but for
2691 * anything else we cannot do is_migration_disabled(), punt
2692 * and have the stopper function handle it all race-free.
2694 stop_pending = pending->stop_pending;
2696 pending->stop_pending = true;
2698 if (flags & SCA_MIGRATE_ENABLE)
2699 p->migration_flags &= ~MDF_PUSH;
2701 task_rq_unlock(rq, p, rf);
2703 if (!stop_pending) {
2704 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2705 &pending->arg, &pending->stop_work);
2708 if (flags & SCA_MIGRATE_ENABLE)
2712 if (!is_migration_disabled(p)) {
2713 if (task_on_rq_queued(p))
2714 rq = move_queued_task(rq, rf, p, dest_cpu);
2716 if (!pending->stop_pending) {
2717 p->migration_pending = NULL;
2721 task_rq_unlock(rq, p, rf);
2724 complete_all(&pending->done);
2727 wait_for_completion(&pending->done);
2729 if (refcount_dec_and_test(&pending->refs))
2730 wake_up_var(&pending->refs); /* No UaF, just an address */
2733 * Block the original owner of &pending until all subsequent callers
2734 * have seen the completion and decremented the refcount
2736 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2739 WARN_ON_ONCE(my_pending.stop_pending);
2745 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2747 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2748 const struct cpumask *new_mask,
2751 struct rq_flags *rf)
2752 __releases(rq->lock)
2753 __releases(p->pi_lock)
2755 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2756 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2757 bool kthread = p->flags & PF_KTHREAD;
2758 struct cpumask *user_mask = NULL;
2759 unsigned int dest_cpu;
2762 update_rq_clock(rq);
2764 if (kthread || is_migration_disabled(p)) {
2766 * Kernel threads are allowed on online && !active CPUs,
2767 * however, during cpu-hot-unplug, even these might get pushed
2768 * away if not KTHREAD_IS_PER_CPU.
2770 * Specifically, migration_disabled() tasks must not fail the
2771 * cpumask_any_and_distribute() pick below, esp. so on
2772 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2773 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2775 cpu_valid_mask = cpu_online_mask;
2778 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2784 * Must re-check here, to close a race against __kthread_bind(),
2785 * sched_setaffinity() is not guaranteed to observe the flag.
2787 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2792 if (!(flags & SCA_MIGRATE_ENABLE)) {
2793 if (cpumask_equal(&p->cpus_mask, new_mask))
2796 if (WARN_ON_ONCE(p == current &&
2797 is_migration_disabled(p) &&
2798 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2805 * Picking a ~random cpu helps in cases where we are changing affinity
2806 * for groups of tasks (ie. cpuset), so that load balancing is not
2807 * immediately required to distribute the tasks within their new mask.
2809 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2810 if (dest_cpu >= nr_cpu_ids) {
2815 __do_set_cpus_allowed(p, new_mask, flags);
2817 if (flags & SCA_USER)
2818 user_mask = clear_user_cpus_ptr(p);
2820 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2827 task_rq_unlock(rq, p, rf);
2833 * Change a given task's CPU affinity. Migrate the thread to a
2834 * proper CPU and schedule it away if the CPU it's executing on
2835 * is removed from the allowed bitmask.
2837 * NOTE: the caller must have a valid reference to the task, the
2838 * task must not exit() & deallocate itself prematurely. The
2839 * call is not atomic; no spinlocks may be held.
2841 static int __set_cpus_allowed_ptr(struct task_struct *p,
2842 const struct cpumask *new_mask, u32 flags)
2847 rq = task_rq_lock(p, &rf);
2848 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2851 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2853 return __set_cpus_allowed_ptr(p, new_mask, 0);
2855 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2858 * Change a given task's CPU affinity to the intersection of its current
2859 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2860 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2861 * If the resulting mask is empty, leave the affinity unchanged and return
2864 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2865 struct cpumask *new_mask,
2866 const struct cpumask *subset_mask)
2868 struct cpumask *user_mask = NULL;
2873 if (!p->user_cpus_ptr) {
2874 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2879 rq = task_rq_lock(p, &rf);
2882 * Forcefully restricting the affinity of a deadline task is
2883 * likely to cause problems, so fail and noisily override the
2886 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2891 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2897 * We're about to butcher the task affinity, so keep track of what
2898 * the user asked for in case we're able to restore it later on.
2901 cpumask_copy(user_mask, p->cpus_ptr);
2902 p->user_cpus_ptr = user_mask;
2905 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2908 task_rq_unlock(rq, p, &rf);
2914 * Restrict the CPU affinity of task @p so that it is a subset of
2915 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2916 * old affinity mask. If the resulting mask is empty, we warn and walk
2917 * up the cpuset hierarchy until we find a suitable mask.
2919 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2921 cpumask_var_t new_mask;
2922 const struct cpumask *override_mask = task_cpu_possible_mask(p);
2924 alloc_cpumask_var(&new_mask, GFP_KERNEL);
2927 * __migrate_task() can fail silently in the face of concurrent
2928 * offlining of the chosen destination CPU, so take the hotplug
2929 * lock to ensure that the migration succeeds.
2932 if (!cpumask_available(new_mask))
2935 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2939 * We failed to find a valid subset of the affinity mask for the
2940 * task, so override it based on its cpuset hierarchy.
2942 cpuset_cpus_allowed(p, new_mask);
2943 override_mask = new_mask;
2946 if (printk_ratelimit()) {
2947 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2948 task_pid_nr(p), p->comm,
2949 cpumask_pr_args(override_mask));
2952 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2955 free_cpumask_var(new_mask);
2959 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
2962 * Restore the affinity of a task @p which was previously restricted by a
2963 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
2964 * @p->user_cpus_ptr.
2966 * It is the caller's responsibility to serialise this with any calls to
2967 * force_compatible_cpus_allowed_ptr(@p).
2969 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
2971 struct cpumask *user_mask = p->user_cpus_ptr;
2972 unsigned long flags;
2975 * Try to restore the old affinity mask. If this fails, then
2976 * we free the mask explicitly to avoid it being inherited across
2977 * a subsequent fork().
2979 if (!user_mask || !__sched_setaffinity(p, user_mask))
2982 raw_spin_lock_irqsave(&p->pi_lock, flags);
2983 user_mask = clear_user_cpus_ptr(p);
2984 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2989 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2991 #ifdef CONFIG_SCHED_DEBUG
2992 unsigned int state = READ_ONCE(p->__state);
2995 * We should never call set_task_cpu() on a blocked task,
2996 * ttwu() will sort out the placement.
2998 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3001 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3002 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3003 * time relying on p->on_rq.
3005 WARN_ON_ONCE(state == TASK_RUNNING &&
3006 p->sched_class == &fair_sched_class &&
3007 (p->on_rq && !task_on_rq_migrating(p)));
3009 #ifdef CONFIG_LOCKDEP
3011 * The caller should hold either p->pi_lock or rq->lock, when changing
3012 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3014 * sched_move_task() holds both and thus holding either pins the cgroup,
3017 * Furthermore, all task_rq users should acquire both locks, see
3020 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3021 lockdep_is_held(__rq_lockp(task_rq(p)))));
3024 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3026 WARN_ON_ONCE(!cpu_online(new_cpu));
3028 WARN_ON_ONCE(is_migration_disabled(p));
3031 trace_sched_migrate_task(p, new_cpu);
3033 if (task_cpu(p) != new_cpu) {
3034 if (p->sched_class->migrate_task_rq)
3035 p->sched_class->migrate_task_rq(p, new_cpu);
3036 p->se.nr_migrations++;
3038 perf_event_task_migrate(p);
3041 __set_task_cpu(p, new_cpu);
3044 #ifdef CONFIG_NUMA_BALANCING
3045 static void __migrate_swap_task(struct task_struct *p, int cpu)
3047 if (task_on_rq_queued(p)) {
3048 struct rq *src_rq, *dst_rq;
3049 struct rq_flags srf, drf;
3051 src_rq = task_rq(p);
3052 dst_rq = cpu_rq(cpu);
3054 rq_pin_lock(src_rq, &srf);
3055 rq_pin_lock(dst_rq, &drf);
3057 deactivate_task(src_rq, p, 0);
3058 set_task_cpu(p, cpu);
3059 activate_task(dst_rq, p, 0);
3060 check_preempt_curr(dst_rq, p, 0);
3062 rq_unpin_lock(dst_rq, &drf);
3063 rq_unpin_lock(src_rq, &srf);
3067 * Task isn't running anymore; make it appear like we migrated
3068 * it before it went to sleep. This means on wakeup we make the
3069 * previous CPU our target instead of where it really is.
3075 struct migration_swap_arg {
3076 struct task_struct *src_task, *dst_task;
3077 int src_cpu, dst_cpu;
3080 static int migrate_swap_stop(void *data)
3082 struct migration_swap_arg *arg = data;
3083 struct rq *src_rq, *dst_rq;
3086 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3089 src_rq = cpu_rq(arg->src_cpu);
3090 dst_rq = cpu_rq(arg->dst_cpu);
3092 double_raw_lock(&arg->src_task->pi_lock,
3093 &arg->dst_task->pi_lock);
3094 double_rq_lock(src_rq, dst_rq);
3096 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3099 if (task_cpu(arg->src_task) != arg->src_cpu)
3102 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3105 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3108 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3109 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3114 double_rq_unlock(src_rq, dst_rq);
3115 raw_spin_unlock(&arg->dst_task->pi_lock);
3116 raw_spin_unlock(&arg->src_task->pi_lock);
3122 * Cross migrate two tasks
3124 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3125 int target_cpu, int curr_cpu)
3127 struct migration_swap_arg arg;
3130 arg = (struct migration_swap_arg){
3132 .src_cpu = curr_cpu,
3134 .dst_cpu = target_cpu,
3137 if (arg.src_cpu == arg.dst_cpu)
3141 * These three tests are all lockless; this is OK since all of them
3142 * will be re-checked with proper locks held further down the line.
3144 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3147 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3150 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3153 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3154 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3159 #endif /* CONFIG_NUMA_BALANCING */
3162 * wait_task_inactive - wait for a thread to unschedule.
3164 * If @match_state is nonzero, it's the @p->state value just checked and
3165 * not expected to change. If it changes, i.e. @p might have woken up,
3166 * then return zero. When we succeed in waiting for @p to be off its CPU,
3167 * we return a positive number (its total switch count). If a second call
3168 * a short while later returns the same number, the caller can be sure that
3169 * @p has remained unscheduled the whole time.
3171 * The caller must ensure that the task *will* unschedule sometime soon,
3172 * else this function might spin for a *long* time. This function can't
3173 * be called with interrupts off, or it may introduce deadlock with
3174 * smp_call_function() if an IPI is sent by the same process we are
3175 * waiting to become inactive.
3177 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3179 int running, queued;
3186 * We do the initial early heuristics without holding
3187 * any task-queue locks at all. We'll only try to get
3188 * the runqueue lock when things look like they will
3194 * If the task is actively running on another CPU
3195 * still, just relax and busy-wait without holding
3198 * NOTE! Since we don't hold any locks, it's not
3199 * even sure that "rq" stays as the right runqueue!
3200 * But we don't care, since "task_running()" will
3201 * return false if the runqueue has changed and p
3202 * is actually now running somewhere else!
3204 while (task_running(rq, p)) {
3205 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3211 * Ok, time to look more closely! We need the rq
3212 * lock now, to be *sure*. If we're wrong, we'll
3213 * just go back and repeat.
3215 rq = task_rq_lock(p, &rf);
3216 trace_sched_wait_task(p);
3217 running = task_running(rq, p);
3218 queued = task_on_rq_queued(p);
3220 if (!match_state || READ_ONCE(p->__state) == match_state)
3221 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3222 task_rq_unlock(rq, p, &rf);
3225 * If it changed from the expected state, bail out now.
3227 if (unlikely(!ncsw))
3231 * Was it really running after all now that we
3232 * checked with the proper locks actually held?
3234 * Oops. Go back and try again..
3236 if (unlikely(running)) {
3242 * It's not enough that it's not actively running,
3243 * it must be off the runqueue _entirely_, and not
3246 * So if it was still runnable (but just not actively
3247 * running right now), it's preempted, and we should
3248 * yield - it could be a while.
3250 if (unlikely(queued)) {
3251 ktime_t to = NSEC_PER_SEC / HZ;
3253 set_current_state(TASK_UNINTERRUPTIBLE);
3254 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
3259 * Ahh, all good. It wasn't running, and it wasn't
3260 * runnable, which means that it will never become
3261 * running in the future either. We're all done!
3270 * kick_process - kick a running thread to enter/exit the kernel
3271 * @p: the to-be-kicked thread
3273 * Cause a process which is running on another CPU to enter
3274 * kernel-mode, without any delay. (to get signals handled.)
3276 * NOTE: this function doesn't have to take the runqueue lock,
3277 * because all it wants to ensure is that the remote task enters
3278 * the kernel. If the IPI races and the task has been migrated
3279 * to another CPU then no harm is done and the purpose has been
3282 void kick_process(struct task_struct *p)
3288 if ((cpu != smp_processor_id()) && task_curr(p))
3289 smp_send_reschedule(cpu);
3292 EXPORT_SYMBOL_GPL(kick_process);
3295 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3297 * A few notes on cpu_active vs cpu_online:
3299 * - cpu_active must be a subset of cpu_online
3301 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3302 * see __set_cpus_allowed_ptr(). At this point the newly online
3303 * CPU isn't yet part of the sched domains, and balancing will not
3306 * - on CPU-down we clear cpu_active() to mask the sched domains and
3307 * avoid the load balancer to place new tasks on the to be removed
3308 * CPU. Existing tasks will remain running there and will be taken
3311 * This means that fallback selection must not select !active CPUs.
3312 * And can assume that any active CPU must be online. Conversely
3313 * select_task_rq() below may allow selection of !active CPUs in order
3314 * to satisfy the above rules.
3316 static int select_fallback_rq(int cpu, struct task_struct *p)
3318 int nid = cpu_to_node(cpu);
3319 const struct cpumask *nodemask = NULL;
3320 enum { cpuset, possible, fail } state = cpuset;
3324 * If the node that the CPU is on has been offlined, cpu_to_node()
3325 * will return -1. There is no CPU on the node, and we should
3326 * select the CPU on the other node.
3329 nodemask = cpumask_of_node(nid);
3331 /* Look for allowed, online CPU in same node. */
3332 for_each_cpu(dest_cpu, nodemask) {
3333 if (is_cpu_allowed(p, dest_cpu))
3339 /* Any allowed, online CPU? */
3340 for_each_cpu(dest_cpu, p->cpus_ptr) {
3341 if (!is_cpu_allowed(p, dest_cpu))
3347 /* No more Mr. Nice Guy. */
3350 if (cpuset_cpus_allowed_fallback(p)) {
3357 * XXX When called from select_task_rq() we only
3358 * hold p->pi_lock and again violate locking order.
3360 * More yuck to audit.
3362 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3372 if (state != cpuset) {
3374 * Don't tell them about moving exiting tasks or
3375 * kernel threads (both mm NULL), since they never
3378 if (p->mm && printk_ratelimit()) {
3379 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3380 task_pid_nr(p), p->comm, cpu);
3388 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3391 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3393 lockdep_assert_held(&p->pi_lock);
3395 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3396 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3398 cpu = cpumask_any(p->cpus_ptr);
3401 * In order not to call set_task_cpu() on a blocking task we need
3402 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3405 * Since this is common to all placement strategies, this lives here.
3407 * [ this allows ->select_task() to simply return task_cpu(p) and
3408 * not worry about this generic constraint ]
3410 if (unlikely(!is_cpu_allowed(p, cpu)))
3411 cpu = select_fallback_rq(task_cpu(p), p);
3416 void sched_set_stop_task(int cpu, struct task_struct *stop)
3418 static struct lock_class_key stop_pi_lock;
3419 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3420 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3424 * Make it appear like a SCHED_FIFO task, its something
3425 * userspace knows about and won't get confused about.
3427 * Also, it will make PI more or less work without too
3428 * much confusion -- but then, stop work should not
3429 * rely on PI working anyway.
3431 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3433 stop->sched_class = &stop_sched_class;
3436 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3437 * adjust the effective priority of a task. As a result,
3438 * rt_mutex_setprio() can trigger (RT) balancing operations,
3439 * which can then trigger wakeups of the stop thread to push
3440 * around the current task.
3442 * The stop task itself will never be part of the PI-chain, it
3443 * never blocks, therefore that ->pi_lock recursion is safe.
3444 * Tell lockdep about this by placing the stop->pi_lock in its
3447 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3450 cpu_rq(cpu)->stop = stop;
3454 * Reset it back to a normal scheduling class so that
3455 * it can die in pieces.
3457 old_stop->sched_class = &rt_sched_class;
3461 #else /* CONFIG_SMP */
3463 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3464 const struct cpumask *new_mask,
3467 return set_cpus_allowed_ptr(p, new_mask);
3470 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3472 static inline bool rq_has_pinned_tasks(struct rq *rq)
3477 #endif /* !CONFIG_SMP */
3480 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3484 if (!schedstat_enabled())
3490 if (cpu == rq->cpu) {
3491 __schedstat_inc(rq->ttwu_local);
3492 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3494 struct sched_domain *sd;
3496 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3498 for_each_domain(rq->cpu, sd) {
3499 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3500 __schedstat_inc(sd->ttwu_wake_remote);
3507 if (wake_flags & WF_MIGRATED)
3508 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3509 #endif /* CONFIG_SMP */
3511 __schedstat_inc(rq->ttwu_count);
3512 __schedstat_inc(p->se.statistics.nr_wakeups);
3514 if (wake_flags & WF_SYNC)
3515 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3519 * Mark the task runnable and perform wakeup-preemption.
3521 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3522 struct rq_flags *rf)
3524 check_preempt_curr(rq, p, wake_flags);
3525 WRITE_ONCE(p->__state, TASK_RUNNING);
3526 trace_sched_wakeup(p);
3529 if (p->sched_class->task_woken) {
3531 * Our task @p is fully woken up and running; so it's safe to
3532 * drop the rq->lock, hereafter rq is only used for statistics.
3534 rq_unpin_lock(rq, rf);
3535 p->sched_class->task_woken(rq, p);
3536 rq_repin_lock(rq, rf);
3539 if (rq->idle_stamp) {
3540 u64 delta = rq_clock(rq) - rq->idle_stamp;
3541 u64 max = 2*rq->max_idle_balance_cost;
3543 update_avg(&rq->avg_idle, delta);
3545 if (rq->avg_idle > max)
3548 rq->wake_stamp = jiffies;
3549 rq->wake_avg_idle = rq->avg_idle / 2;
3557 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3558 struct rq_flags *rf)
3560 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3562 lockdep_assert_rq_held(rq);
3564 if (p->sched_contributes_to_load)
3565 rq->nr_uninterruptible--;
3568 if (wake_flags & WF_MIGRATED)
3569 en_flags |= ENQUEUE_MIGRATED;
3573 delayacct_blkio_end(p);
3574 atomic_dec(&task_rq(p)->nr_iowait);
3577 activate_task(rq, p, en_flags);
3578 ttwu_do_wakeup(rq, p, wake_flags, rf);
3582 * Consider @p being inside a wait loop:
3585 * set_current_state(TASK_UNINTERRUPTIBLE);
3592 * __set_current_state(TASK_RUNNING);
3594 * between set_current_state() and schedule(). In this case @p is still
3595 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3598 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3599 * then schedule() must still happen and p->state can be changed to
3600 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3601 * need to do a full wakeup with enqueue.
3603 * Returns: %true when the wakeup is done,
3606 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3612 rq = __task_rq_lock(p, &rf);
3613 if (task_on_rq_queued(p)) {
3614 /* check_preempt_curr() may use rq clock */
3615 update_rq_clock(rq);
3616 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3619 __task_rq_unlock(rq, &rf);
3625 void sched_ttwu_pending(void *arg)
3627 struct llist_node *llist = arg;
3628 struct rq *rq = this_rq();
3629 struct task_struct *p, *t;
3636 * rq::ttwu_pending racy indication of out-standing wakeups.
3637 * Races such that false-negatives are possible, since they
3638 * are shorter lived that false-positives would be.
3640 WRITE_ONCE(rq->ttwu_pending, 0);
3642 rq_lock_irqsave(rq, &rf);
3643 update_rq_clock(rq);
3645 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3646 if (WARN_ON_ONCE(p->on_cpu))
3647 smp_cond_load_acquire(&p->on_cpu, !VAL);
3649 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3650 set_task_cpu(p, cpu_of(rq));
3652 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3655 rq_unlock_irqrestore(rq, &rf);
3658 void send_call_function_single_ipi(int cpu)
3660 struct rq *rq = cpu_rq(cpu);
3662 if (!set_nr_if_polling(rq->idle))
3663 arch_send_call_function_single_ipi(cpu);
3665 trace_sched_wake_idle_without_ipi(cpu);
3669 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3670 * necessary. The wakee CPU on receipt of the IPI will queue the task
3671 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3672 * of the wakeup instead of the waker.
3674 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3676 struct rq *rq = cpu_rq(cpu);
3678 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3680 WRITE_ONCE(rq->ttwu_pending, 1);
3681 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3684 void wake_up_if_idle(int cpu)
3686 struct rq *rq = cpu_rq(cpu);
3691 if (!is_idle_task(rcu_dereference(rq->curr)))
3694 if (set_nr_if_polling(rq->idle)) {
3695 trace_sched_wake_idle_without_ipi(cpu);
3697 rq_lock_irqsave(rq, &rf);
3698 if (is_idle_task(rq->curr))
3699 smp_send_reschedule(cpu);
3700 /* Else CPU is not idle, do nothing here: */
3701 rq_unlock_irqrestore(rq, &rf);
3708 bool cpus_share_cache(int this_cpu, int that_cpu)
3710 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3713 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3716 * Do not complicate things with the async wake_list while the CPU is
3719 if (!cpu_active(cpu))
3723 * If the CPU does not share cache, then queue the task on the
3724 * remote rqs wakelist to avoid accessing remote data.
3726 if (!cpus_share_cache(smp_processor_id(), cpu))
3730 * If the task is descheduling and the only running task on the
3731 * CPU then use the wakelist to offload the task activation to
3732 * the soon-to-be-idle CPU as the current CPU is likely busy.
3733 * nr_running is checked to avoid unnecessary task stacking.
3735 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3741 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3743 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3744 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3747 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3748 __ttwu_queue_wakelist(p, cpu, wake_flags);
3755 #else /* !CONFIG_SMP */
3757 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3762 #endif /* CONFIG_SMP */
3764 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3766 struct rq *rq = cpu_rq(cpu);
3769 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3773 update_rq_clock(rq);
3774 ttwu_do_activate(rq, p, wake_flags, &rf);
3779 * Notes on Program-Order guarantees on SMP systems.
3783 * The basic program-order guarantee on SMP systems is that when a task [t]
3784 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3785 * execution on its new CPU [c1].
3787 * For migration (of runnable tasks) this is provided by the following means:
3789 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3790 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3791 * rq(c1)->lock (if not at the same time, then in that order).
3792 * C) LOCK of the rq(c1)->lock scheduling in task
3794 * Release/acquire chaining guarantees that B happens after A and C after B.
3795 * Note: the CPU doing B need not be c0 or c1
3804 * UNLOCK rq(0)->lock
3806 * LOCK rq(0)->lock // orders against CPU0
3808 * UNLOCK rq(0)->lock
3812 * UNLOCK rq(1)->lock
3814 * LOCK rq(1)->lock // orders against CPU2
3817 * UNLOCK rq(1)->lock
3820 * BLOCKING -- aka. SLEEP + WAKEUP
3822 * For blocking we (obviously) need to provide the same guarantee as for
3823 * migration. However the means are completely different as there is no lock
3824 * chain to provide order. Instead we do:
3826 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3827 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3831 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3833 * LOCK rq(0)->lock LOCK X->pi_lock
3836 * smp_store_release(X->on_cpu, 0);
3838 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3844 * X->state = RUNNING
3845 * UNLOCK rq(2)->lock
3847 * LOCK rq(2)->lock // orders against CPU1
3850 * UNLOCK rq(2)->lock
3853 * UNLOCK rq(0)->lock
3856 * However, for wakeups there is a second guarantee we must provide, namely we
3857 * must ensure that CONDITION=1 done by the caller can not be reordered with
3858 * accesses to the task state; see try_to_wake_up() and set_current_state().
3862 * try_to_wake_up - wake up a thread
3863 * @p: the thread to be awakened
3864 * @state: the mask of task states that can be woken
3865 * @wake_flags: wake modifier flags (WF_*)
3867 * Conceptually does:
3869 * If (@state & @p->state) @p->state = TASK_RUNNING.
3871 * If the task was not queued/runnable, also place it back on a runqueue.
3873 * This function is atomic against schedule() which would dequeue the task.
3875 * It issues a full memory barrier before accessing @p->state, see the comment
3876 * with set_current_state().
3878 * Uses p->pi_lock to serialize against concurrent wake-ups.
3880 * Relies on p->pi_lock stabilizing:
3883 * - p->sched_task_group
3884 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3886 * Tries really hard to only take one task_rq(p)->lock for performance.
3887 * Takes rq->lock in:
3888 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3889 * - ttwu_queue() -- new rq, for enqueue of the task;
3890 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3892 * As a consequence we race really badly with just about everything. See the
3893 * many memory barriers and their comments for details.
3895 * Return: %true if @p->state changes (an actual wakeup was done),
3899 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3901 unsigned long flags;
3902 int cpu, success = 0;
3907 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3908 * == smp_processor_id()'. Together this means we can special
3909 * case the whole 'p->on_rq && ttwu_runnable()' case below
3910 * without taking any locks.
3913 * - we rely on Program-Order guarantees for all the ordering,
3914 * - we're serialized against set_special_state() by virtue of
3915 * it disabling IRQs (this allows not taking ->pi_lock).
3917 if (!(READ_ONCE(p->__state) & state))
3921 trace_sched_waking(p);
3922 WRITE_ONCE(p->__state, TASK_RUNNING);
3923 trace_sched_wakeup(p);
3928 * If we are going to wake up a thread waiting for CONDITION we
3929 * need to ensure that CONDITION=1 done by the caller can not be
3930 * reordered with p->state check below. This pairs with smp_store_mb()
3931 * in set_current_state() that the waiting thread does.
3933 raw_spin_lock_irqsave(&p->pi_lock, flags);
3934 smp_mb__after_spinlock();
3935 if (!(READ_ONCE(p->__state) & state))
3938 trace_sched_waking(p);
3940 /* We're going to change ->state: */
3944 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3945 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3946 * in smp_cond_load_acquire() below.
3948 * sched_ttwu_pending() try_to_wake_up()
3949 * STORE p->on_rq = 1 LOAD p->state
3952 * __schedule() (switch to task 'p')
3953 * LOCK rq->lock smp_rmb();
3954 * smp_mb__after_spinlock();
3958 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3960 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3961 * __schedule(). See the comment for smp_mb__after_spinlock().
3963 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3966 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3971 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3972 * possible to, falsely, observe p->on_cpu == 0.
3974 * One must be running (->on_cpu == 1) in order to remove oneself
3975 * from the runqueue.
3977 * __schedule() (switch to task 'p') try_to_wake_up()
3978 * STORE p->on_cpu = 1 LOAD p->on_rq
3981 * __schedule() (put 'p' to sleep)
3982 * LOCK rq->lock smp_rmb();
3983 * smp_mb__after_spinlock();
3984 * STORE p->on_rq = 0 LOAD p->on_cpu
3986 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3987 * __schedule(). See the comment for smp_mb__after_spinlock().
3989 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3990 * schedule()'s deactivate_task() has 'happened' and p will no longer
3991 * care about it's own p->state. See the comment in __schedule().
3993 smp_acquire__after_ctrl_dep();
3996 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3997 * == 0), which means we need to do an enqueue, change p->state to
3998 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3999 * enqueue, such as ttwu_queue_wakelist().
4001 WRITE_ONCE(p->__state, TASK_WAKING);
4004 * If the owning (remote) CPU is still in the middle of schedule() with
4005 * this task as prev, considering queueing p on the remote CPUs wake_list
4006 * which potentially sends an IPI instead of spinning on p->on_cpu to
4007 * let the waker make forward progress. This is safe because IRQs are
4008 * disabled and the IPI will deliver after on_cpu is cleared.
4010 * Ensure we load task_cpu(p) after p->on_cpu:
4012 * set_task_cpu(p, cpu);
4013 * STORE p->cpu = @cpu
4014 * __schedule() (switch to task 'p')
4016 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4017 * STORE p->on_cpu = 1 LOAD p->cpu
4019 * to ensure we observe the correct CPU on which the task is currently
4022 if (smp_load_acquire(&p->on_cpu) &&
4023 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4027 * If the owning (remote) CPU is still in the middle of schedule() with
4028 * this task as prev, wait until it's done referencing the task.
4030 * Pairs with the smp_store_release() in finish_task().
4032 * This ensures that tasks getting woken will be fully ordered against
4033 * their previous state and preserve Program Order.
4035 smp_cond_load_acquire(&p->on_cpu, !VAL);
4037 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4038 if (task_cpu(p) != cpu) {
4040 delayacct_blkio_end(p);
4041 atomic_dec(&task_rq(p)->nr_iowait);
4044 wake_flags |= WF_MIGRATED;
4045 psi_ttwu_dequeue(p);
4046 set_task_cpu(p, cpu);
4050 #endif /* CONFIG_SMP */
4052 ttwu_queue(p, cpu, wake_flags);
4054 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4057 ttwu_stat(p, task_cpu(p), wake_flags);
4064 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
4065 * @p: Process for which the function is to be invoked, can be @current.
4066 * @func: Function to invoke.
4067 * @arg: Argument to function.
4069 * If the specified task can be quickly locked into a definite state
4070 * (either sleeping or on a given runqueue), arrange to keep it in that
4071 * state while invoking @func(@arg). This function can use ->on_rq and
4072 * task_curr() to work out what the state is, if required. Given that
4073 * @func can be invoked with a runqueue lock held, it had better be quite
4077 * @false if the task slipped out from under the locks.
4078 * @true if the task was locked onto a runqueue or is sleeping.
4079 * However, @func can override this by returning @false.
4081 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
4087 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4089 rq = __task_rq_lock(p, &rf);
4090 if (task_rq(p) == rq)
4094 switch (READ_ONCE(p->__state)) {
4099 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
4104 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4109 * wake_up_process - Wake up a specific process
4110 * @p: The process to be woken up.
4112 * Attempt to wake up the nominated process and move it to the set of runnable
4115 * Return: 1 if the process was woken up, 0 if it was already running.
4117 * This function executes a full memory barrier before accessing the task state.
4119 int wake_up_process(struct task_struct *p)
4121 return try_to_wake_up(p, TASK_NORMAL, 0);
4123 EXPORT_SYMBOL(wake_up_process);
4125 int wake_up_state(struct task_struct *p, unsigned int state)
4127 return try_to_wake_up(p, state, 0);
4131 * Perform scheduler related setup for a newly forked process p.
4132 * p is forked by current.
4134 * __sched_fork() is basic setup used by init_idle() too:
4136 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4141 p->se.exec_start = 0;
4142 p->se.sum_exec_runtime = 0;
4143 p->se.prev_sum_exec_runtime = 0;
4144 p->se.nr_migrations = 0;
4146 INIT_LIST_HEAD(&p->se.group_node);
4148 #ifdef CONFIG_FAIR_GROUP_SCHED
4149 p->se.cfs_rq = NULL;
4152 #ifdef CONFIG_SCHEDSTATS
4153 /* Even if schedstat is disabled, there should not be garbage */
4154 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
4157 RB_CLEAR_NODE(&p->dl.rb_node);
4158 init_dl_task_timer(&p->dl);
4159 init_dl_inactive_task_timer(&p->dl);
4160 __dl_clear_params(p);
4162 INIT_LIST_HEAD(&p->rt.run_list);
4164 p->rt.time_slice = sched_rr_timeslice;
4168 #ifdef CONFIG_PREEMPT_NOTIFIERS
4169 INIT_HLIST_HEAD(&p->preempt_notifiers);
4172 #ifdef CONFIG_COMPACTION
4173 p->capture_control = NULL;
4175 init_numa_balancing(clone_flags, p);
4177 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4178 p->migration_pending = NULL;
4182 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4184 #ifdef CONFIG_NUMA_BALANCING
4186 void set_numabalancing_state(bool enabled)
4189 static_branch_enable(&sched_numa_balancing);
4191 static_branch_disable(&sched_numa_balancing);
4194 #ifdef CONFIG_PROC_SYSCTL
4195 int sysctl_numa_balancing(struct ctl_table *table, int write,
4196 void *buffer, size_t *lenp, loff_t *ppos)
4200 int state = static_branch_likely(&sched_numa_balancing);
4202 if (write && !capable(CAP_SYS_ADMIN))
4207 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4211 set_numabalancing_state(state);
4217 #ifdef CONFIG_SCHEDSTATS
4219 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4221 static void set_schedstats(bool enabled)
4224 static_branch_enable(&sched_schedstats);
4226 static_branch_disable(&sched_schedstats);
4229 void force_schedstat_enabled(void)
4231 if (!schedstat_enabled()) {
4232 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4233 static_branch_enable(&sched_schedstats);
4237 static int __init setup_schedstats(char *str)
4243 if (!strcmp(str, "enable")) {
4244 set_schedstats(true);
4246 } else if (!strcmp(str, "disable")) {
4247 set_schedstats(false);
4252 pr_warn("Unable to parse schedstats=\n");
4256 __setup("schedstats=", setup_schedstats);
4258 #ifdef CONFIG_PROC_SYSCTL
4259 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4260 size_t *lenp, loff_t *ppos)
4264 int state = static_branch_likely(&sched_schedstats);
4266 if (write && !capable(CAP_SYS_ADMIN))
4271 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4275 set_schedstats(state);
4278 #endif /* CONFIG_PROC_SYSCTL */
4279 #endif /* CONFIG_SCHEDSTATS */
4282 * fork()/clone()-time setup:
4284 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4286 unsigned long flags;
4288 __sched_fork(clone_flags, p);
4290 * We mark the process as NEW here. This guarantees that
4291 * nobody will actually run it, and a signal or other external
4292 * event cannot wake it up and insert it on the runqueue either.
4294 p->__state = TASK_NEW;
4297 * Make sure we do not leak PI boosting priority to the child.
4299 p->prio = current->normal_prio;
4304 * Revert to default priority/policy on fork if requested.
4306 if (unlikely(p->sched_reset_on_fork)) {
4307 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4308 p->policy = SCHED_NORMAL;
4309 p->static_prio = NICE_TO_PRIO(0);
4311 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4312 p->static_prio = NICE_TO_PRIO(0);
4314 p->prio = p->normal_prio = p->static_prio;
4315 set_load_weight(p, false);
4318 * We don't need the reset flag anymore after the fork. It has
4319 * fulfilled its duty:
4321 p->sched_reset_on_fork = 0;
4324 if (dl_prio(p->prio))
4326 else if (rt_prio(p->prio))
4327 p->sched_class = &rt_sched_class;
4329 p->sched_class = &fair_sched_class;
4331 init_entity_runnable_average(&p->se);
4334 * The child is not yet in the pid-hash so no cgroup attach races,
4335 * and the cgroup is pinned to this child due to cgroup_fork()
4336 * is ran before sched_fork().
4338 * Silence PROVE_RCU.
4340 raw_spin_lock_irqsave(&p->pi_lock, flags);
4343 * We're setting the CPU for the first time, we don't migrate,
4344 * so use __set_task_cpu().
4346 __set_task_cpu(p, smp_processor_id());
4347 if (p->sched_class->task_fork)
4348 p->sched_class->task_fork(p);
4349 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4351 #ifdef CONFIG_SCHED_INFO
4352 if (likely(sched_info_on()))
4353 memset(&p->sched_info, 0, sizeof(p->sched_info));
4355 #if defined(CONFIG_SMP)
4358 init_task_preempt_count(p);
4360 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4361 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4366 void sched_post_fork(struct task_struct *p)
4368 uclamp_post_fork(p);
4371 unsigned long to_ratio(u64 period, u64 runtime)
4373 if (runtime == RUNTIME_INF)
4377 * Doing this here saves a lot of checks in all
4378 * the calling paths, and returning zero seems
4379 * safe for them anyway.
4384 return div64_u64(runtime << BW_SHIFT, period);
4388 * wake_up_new_task - wake up a newly created task for the first time.
4390 * This function will do some initial scheduler statistics housekeeping
4391 * that must be done for every newly created context, then puts the task
4392 * on the runqueue and wakes it.
4394 void wake_up_new_task(struct task_struct *p)
4399 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4400 WRITE_ONCE(p->__state, TASK_RUNNING);
4403 * Fork balancing, do it here and not earlier because:
4404 * - cpus_ptr can change in the fork path
4405 * - any previously selected CPU might disappear through hotplug
4407 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4408 * as we're not fully set-up yet.
4410 p->recent_used_cpu = task_cpu(p);
4412 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4414 rq = __task_rq_lock(p, &rf);
4415 update_rq_clock(rq);
4416 post_init_entity_util_avg(p);
4418 activate_task(rq, p, ENQUEUE_NOCLOCK);
4419 trace_sched_wakeup_new(p);
4420 check_preempt_curr(rq, p, WF_FORK);
4422 if (p->sched_class->task_woken) {
4424 * Nothing relies on rq->lock after this, so it's fine to
4427 rq_unpin_lock(rq, &rf);
4428 p->sched_class->task_woken(rq, p);
4429 rq_repin_lock(rq, &rf);
4432 task_rq_unlock(rq, p, &rf);
4435 #ifdef CONFIG_PREEMPT_NOTIFIERS
4437 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4439 void preempt_notifier_inc(void)
4441 static_branch_inc(&preempt_notifier_key);
4443 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4445 void preempt_notifier_dec(void)
4447 static_branch_dec(&preempt_notifier_key);
4449 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4452 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4453 * @notifier: notifier struct to register
4455 void preempt_notifier_register(struct preempt_notifier *notifier)
4457 if (!static_branch_unlikely(&preempt_notifier_key))
4458 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4460 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4462 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4465 * preempt_notifier_unregister - no longer interested in preemption notifications
4466 * @notifier: notifier struct to unregister
4468 * This is *not* safe to call from within a preemption notifier.
4470 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4472 hlist_del(¬ifier->link);
4474 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4476 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4478 struct preempt_notifier *notifier;
4480 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4481 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4484 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4486 if (static_branch_unlikely(&preempt_notifier_key))
4487 __fire_sched_in_preempt_notifiers(curr);
4491 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4492 struct task_struct *next)
4494 struct preempt_notifier *notifier;
4496 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4497 notifier->ops->sched_out(notifier, next);
4500 static __always_inline void
4501 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4502 struct task_struct *next)
4504 if (static_branch_unlikely(&preempt_notifier_key))
4505 __fire_sched_out_preempt_notifiers(curr, next);
4508 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4510 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4515 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4516 struct task_struct *next)
4520 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4522 static inline void prepare_task(struct task_struct *next)
4526 * Claim the task as running, we do this before switching to it
4527 * such that any running task will have this set.
4529 * See the ttwu() WF_ON_CPU case and its ordering comment.
4531 WRITE_ONCE(next->on_cpu, 1);
4535 static inline void finish_task(struct task_struct *prev)
4539 * This must be the very last reference to @prev from this CPU. After
4540 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4541 * must ensure this doesn't happen until the switch is completely
4544 * In particular, the load of prev->state in finish_task_switch() must
4545 * happen before this.
4547 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4549 smp_store_release(&prev->on_cpu, 0);
4555 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4557 void (*func)(struct rq *rq);
4558 struct callback_head *next;
4560 lockdep_assert_rq_held(rq);
4563 func = (void (*)(struct rq *))head->func;
4572 static void balance_push(struct rq *rq);
4574 struct callback_head balance_push_callback = {
4576 .func = (void (*)(struct callback_head *))balance_push,
4579 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4581 struct callback_head *head = rq->balance_callback;
4583 lockdep_assert_rq_held(rq);
4585 rq->balance_callback = NULL;
4590 static void __balance_callbacks(struct rq *rq)
4592 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4595 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4597 unsigned long flags;
4599 if (unlikely(head)) {
4600 raw_spin_rq_lock_irqsave(rq, flags);
4601 do_balance_callbacks(rq, head);
4602 raw_spin_rq_unlock_irqrestore(rq, flags);
4608 static inline void __balance_callbacks(struct rq *rq)
4612 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4617 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4624 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4627 * Since the runqueue lock will be released by the next
4628 * task (which is an invalid locking op but in the case
4629 * of the scheduler it's an obvious special-case), so we
4630 * do an early lockdep release here:
4632 rq_unpin_lock(rq, rf);
4633 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4634 #ifdef CONFIG_DEBUG_SPINLOCK
4635 /* this is a valid case when another task releases the spinlock */
4636 rq_lockp(rq)->owner = next;
4640 static inline void finish_lock_switch(struct rq *rq)
4643 * If we are tracking spinlock dependencies then we have to
4644 * fix up the runqueue lock - which gets 'carried over' from
4645 * prev into current:
4647 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4648 __balance_callbacks(rq);
4649 raw_spin_rq_unlock_irq(rq);
4653 * NOP if the arch has not defined these:
4656 #ifndef prepare_arch_switch
4657 # define prepare_arch_switch(next) do { } while (0)
4660 #ifndef finish_arch_post_lock_switch
4661 # define finish_arch_post_lock_switch() do { } while (0)
4664 static inline void kmap_local_sched_out(void)
4666 #ifdef CONFIG_KMAP_LOCAL
4667 if (unlikely(current->kmap_ctrl.idx))
4668 __kmap_local_sched_out();
4672 static inline void kmap_local_sched_in(void)
4674 #ifdef CONFIG_KMAP_LOCAL
4675 if (unlikely(current->kmap_ctrl.idx))
4676 __kmap_local_sched_in();
4681 * prepare_task_switch - prepare to switch tasks
4682 * @rq: the runqueue preparing to switch
4683 * @prev: the current task that is being switched out
4684 * @next: the task we are going to switch to.
4686 * This is called with the rq lock held and interrupts off. It must
4687 * be paired with a subsequent finish_task_switch after the context
4690 * prepare_task_switch sets up locking and calls architecture specific
4694 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4695 struct task_struct *next)
4697 kcov_prepare_switch(prev);
4698 sched_info_switch(rq, prev, next);
4699 perf_event_task_sched_out(prev, next);
4701 fire_sched_out_preempt_notifiers(prev, next);
4702 kmap_local_sched_out();
4704 prepare_arch_switch(next);
4708 * finish_task_switch - clean up after a task-switch
4709 * @prev: the thread we just switched away from.
4711 * finish_task_switch must be called after the context switch, paired
4712 * with a prepare_task_switch call before the context switch.
4713 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4714 * and do any other architecture-specific cleanup actions.
4716 * Note that we may have delayed dropping an mm in context_switch(). If
4717 * so, we finish that here outside of the runqueue lock. (Doing it
4718 * with the lock held can cause deadlocks; see schedule() for
4721 * The context switch have flipped the stack from under us and restored the
4722 * local variables which were saved when this task called schedule() in the
4723 * past. prev == current is still correct but we need to recalculate this_rq
4724 * because prev may have moved to another CPU.
4726 static struct rq *finish_task_switch(struct task_struct *prev)
4727 __releases(rq->lock)
4729 struct rq *rq = this_rq();
4730 struct mm_struct *mm = rq->prev_mm;
4734 * The previous task will have left us with a preempt_count of 2
4735 * because it left us after:
4738 * preempt_disable(); // 1
4740 * raw_spin_lock_irq(&rq->lock) // 2
4742 * Also, see FORK_PREEMPT_COUNT.
4744 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4745 "corrupted preempt_count: %s/%d/0x%x\n",
4746 current->comm, current->pid, preempt_count()))
4747 preempt_count_set(FORK_PREEMPT_COUNT);
4752 * A task struct has one reference for the use as "current".
4753 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4754 * schedule one last time. The schedule call will never return, and
4755 * the scheduled task must drop that reference.
4757 * We must observe prev->state before clearing prev->on_cpu (in
4758 * finish_task), otherwise a concurrent wakeup can get prev
4759 * running on another CPU and we could rave with its RUNNING -> DEAD
4760 * transition, resulting in a double drop.
4762 prev_state = READ_ONCE(prev->__state);
4763 vtime_task_switch(prev);
4764 perf_event_task_sched_in(prev, current);
4766 tick_nohz_task_switch();
4767 finish_lock_switch(rq);
4768 finish_arch_post_lock_switch();
4769 kcov_finish_switch(current);
4771 * kmap_local_sched_out() is invoked with rq::lock held and
4772 * interrupts disabled. There is no requirement for that, but the
4773 * sched out code does not have an interrupt enabled section.
4774 * Restoring the maps on sched in does not require interrupts being
4777 kmap_local_sched_in();
4779 fire_sched_in_preempt_notifiers(current);
4781 * When switching through a kernel thread, the loop in
4782 * membarrier_{private,global}_expedited() may have observed that
4783 * kernel thread and not issued an IPI. It is therefore possible to
4784 * schedule between user->kernel->user threads without passing though
4785 * switch_mm(). Membarrier requires a barrier after storing to
4786 * rq->curr, before returning to userspace, so provide them here:
4788 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4789 * provided by mmdrop(),
4790 * - a sync_core for SYNC_CORE.
4793 membarrier_mm_sync_core_before_usermode(mm);
4796 if (unlikely(prev_state == TASK_DEAD)) {
4797 if (prev->sched_class->task_dead)
4798 prev->sched_class->task_dead(prev);
4801 * Remove function-return probe instances associated with this
4802 * task and put them back on the free list.
4804 kprobe_flush_task(prev);
4806 /* Task is done with its stack. */
4807 put_task_stack(prev);
4809 put_task_struct_rcu_user(prev);
4816 * schedule_tail - first thing a freshly forked thread must call.
4817 * @prev: the thread we just switched away from.
4819 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4820 __releases(rq->lock)
4823 * New tasks start with FORK_PREEMPT_COUNT, see there and
4824 * finish_task_switch() for details.
4826 * finish_task_switch() will drop rq->lock() and lower preempt_count
4827 * and the preempt_enable() will end up enabling preemption (on
4828 * PREEMPT_COUNT kernels).
4831 finish_task_switch(prev);
4834 if (current->set_child_tid)
4835 put_user(task_pid_vnr(current), current->set_child_tid);
4837 calculate_sigpending();
4841 * context_switch - switch to the new MM and the new thread's register state.
4843 static __always_inline struct rq *
4844 context_switch(struct rq *rq, struct task_struct *prev,
4845 struct task_struct *next, struct rq_flags *rf)
4847 prepare_task_switch(rq, prev, next);
4850 * For paravirt, this is coupled with an exit in switch_to to
4851 * combine the page table reload and the switch backend into
4854 arch_start_context_switch(prev);
4857 * kernel -> kernel lazy + transfer active
4858 * user -> kernel lazy + mmgrab() active
4860 * kernel -> user switch + mmdrop() active
4861 * user -> user switch
4863 if (!next->mm) { // to kernel
4864 enter_lazy_tlb(prev->active_mm, next);
4866 next->active_mm = prev->active_mm;
4867 if (prev->mm) // from user
4868 mmgrab(prev->active_mm);
4870 prev->active_mm = NULL;
4872 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4874 * sys_membarrier() requires an smp_mb() between setting
4875 * rq->curr / membarrier_switch_mm() and returning to userspace.
4877 * The below provides this either through switch_mm(), or in
4878 * case 'prev->active_mm == next->mm' through
4879 * finish_task_switch()'s mmdrop().
4881 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4883 if (!prev->mm) { // from kernel
4884 /* will mmdrop() in finish_task_switch(). */
4885 rq->prev_mm = prev->active_mm;
4886 prev->active_mm = NULL;
4890 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4892 prepare_lock_switch(rq, next, rf);
4894 /* Here we just switch the register state and the stack. */
4895 switch_to(prev, next, prev);
4898 return finish_task_switch(prev);
4902 * nr_running and nr_context_switches:
4904 * externally visible scheduler statistics: current number of runnable
4905 * threads, total number of context switches performed since bootup.
4907 unsigned int nr_running(void)
4909 unsigned int i, sum = 0;
4911 for_each_online_cpu(i)
4912 sum += cpu_rq(i)->nr_running;
4918 * Check if only the current task is running on the CPU.
4920 * Caution: this function does not check that the caller has disabled
4921 * preemption, thus the result might have a time-of-check-to-time-of-use
4922 * race. The caller is responsible to use it correctly, for example:
4924 * - from a non-preemptible section (of course)
4926 * - from a thread that is bound to a single CPU
4928 * - in a loop with very short iterations (e.g. a polling loop)
4930 bool single_task_running(void)
4932 return raw_rq()->nr_running == 1;
4934 EXPORT_SYMBOL(single_task_running);
4936 unsigned long long nr_context_switches(void)
4939 unsigned long long sum = 0;
4941 for_each_possible_cpu(i)
4942 sum += cpu_rq(i)->nr_switches;
4948 * Consumers of these two interfaces, like for example the cpuidle menu
4949 * governor, are using nonsensical data. Preferring shallow idle state selection
4950 * for a CPU that has IO-wait which might not even end up running the task when
4951 * it does become runnable.
4954 unsigned int nr_iowait_cpu(int cpu)
4956 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4960 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4962 * The idea behind IO-wait account is to account the idle time that we could
4963 * have spend running if it were not for IO. That is, if we were to improve the
4964 * storage performance, we'd have a proportional reduction in IO-wait time.
4966 * This all works nicely on UP, where, when a task blocks on IO, we account
4967 * idle time as IO-wait, because if the storage were faster, it could've been
4968 * running and we'd not be idle.
4970 * This has been extended to SMP, by doing the same for each CPU. This however
4973 * Imagine for instance the case where two tasks block on one CPU, only the one
4974 * CPU will have IO-wait accounted, while the other has regular idle. Even
4975 * though, if the storage were faster, both could've ran at the same time,
4976 * utilising both CPUs.
4978 * This means, that when looking globally, the current IO-wait accounting on
4979 * SMP is a lower bound, by reason of under accounting.
4981 * Worse, since the numbers are provided per CPU, they are sometimes
4982 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4983 * associated with any one particular CPU, it can wake to another CPU than it
4984 * blocked on. This means the per CPU IO-wait number is meaningless.
4986 * Task CPU affinities can make all that even more 'interesting'.
4989 unsigned int nr_iowait(void)
4991 unsigned int i, sum = 0;
4993 for_each_possible_cpu(i)
4994 sum += nr_iowait_cpu(i);
5002 * sched_exec - execve() is a valuable balancing opportunity, because at
5003 * this point the task has the smallest effective memory and cache footprint.
5005 void sched_exec(void)
5007 struct task_struct *p = current;
5008 unsigned long flags;
5011 raw_spin_lock_irqsave(&p->pi_lock, flags);
5012 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5013 if (dest_cpu == smp_processor_id())
5016 if (likely(cpu_active(dest_cpu))) {
5017 struct migration_arg arg = { p, dest_cpu };
5019 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5020 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5024 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5029 DEFINE_PER_CPU(struct kernel_stat, kstat);
5030 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5032 EXPORT_PER_CPU_SYMBOL(kstat);
5033 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5036 * The function fair_sched_class.update_curr accesses the struct curr
5037 * and its field curr->exec_start; when called from task_sched_runtime(),
5038 * we observe a high rate of cache misses in practice.
5039 * Prefetching this data results in improved performance.
5041 static inline void prefetch_curr_exec_start(struct task_struct *p)
5043 #ifdef CONFIG_FAIR_GROUP_SCHED
5044 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5046 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5049 prefetch(&curr->exec_start);
5053 * Return accounted runtime for the task.
5054 * In case the task is currently running, return the runtime plus current's
5055 * pending runtime that have not been accounted yet.
5057 unsigned long long task_sched_runtime(struct task_struct *p)
5063 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5065 * 64-bit doesn't need locks to atomically read a 64-bit value.
5066 * So we have a optimization chance when the task's delta_exec is 0.
5067 * Reading ->on_cpu is racy, but this is ok.
5069 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5070 * If we race with it entering CPU, unaccounted time is 0. This is
5071 * indistinguishable from the read occurring a few cycles earlier.
5072 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5073 * been accounted, so we're correct here as well.
5075 if (!p->on_cpu || !task_on_rq_queued(p))
5076 return p->se.sum_exec_runtime;
5079 rq = task_rq_lock(p, &rf);
5081 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5082 * project cycles that may never be accounted to this
5083 * thread, breaking clock_gettime().
5085 if (task_current(rq, p) && task_on_rq_queued(p)) {
5086 prefetch_curr_exec_start(p);
5087 update_rq_clock(rq);
5088 p->sched_class->update_curr(rq);
5090 ns = p->se.sum_exec_runtime;
5091 task_rq_unlock(rq, p, &rf);
5096 #ifdef CONFIG_SCHED_DEBUG
5097 static u64 cpu_resched_latency(struct rq *rq)
5099 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5100 u64 resched_latency, now = rq_clock(rq);
5101 static bool warned_once;
5103 if (sysctl_resched_latency_warn_once && warned_once)
5106 if (!need_resched() || !latency_warn_ms)
5109 if (system_state == SYSTEM_BOOTING)
5112 if (!rq->last_seen_need_resched_ns) {
5113 rq->last_seen_need_resched_ns = now;
5114 rq->ticks_without_resched = 0;
5118 rq->ticks_without_resched++;
5119 resched_latency = now - rq->last_seen_need_resched_ns;
5120 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5125 return resched_latency;
5128 static int __init setup_resched_latency_warn_ms(char *str)
5132 if ((kstrtol(str, 0, &val))) {
5133 pr_warn("Unable to set resched_latency_warn_ms\n");
5137 sysctl_resched_latency_warn_ms = val;
5140 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5142 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5143 #endif /* CONFIG_SCHED_DEBUG */
5146 * This function gets called by the timer code, with HZ frequency.
5147 * We call it with interrupts disabled.
5149 void scheduler_tick(void)
5151 int cpu = smp_processor_id();
5152 struct rq *rq = cpu_rq(cpu);
5153 struct task_struct *curr = rq->curr;
5155 unsigned long thermal_pressure;
5156 u64 resched_latency;
5158 arch_scale_freq_tick();
5163 update_rq_clock(rq);
5164 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5165 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5166 curr->sched_class->task_tick(rq, curr, 0);
5167 if (sched_feat(LATENCY_WARN))
5168 resched_latency = cpu_resched_latency(rq);
5169 calc_global_load_tick(rq);
5173 if (sched_feat(LATENCY_WARN) && resched_latency)
5174 resched_latency_warn(cpu, resched_latency);
5176 perf_event_task_tick();
5179 rq->idle_balance = idle_cpu(cpu);
5180 trigger_load_balance(rq);
5184 #ifdef CONFIG_NO_HZ_FULL
5189 struct delayed_work work;
5191 /* Values for ->state, see diagram below. */
5192 #define TICK_SCHED_REMOTE_OFFLINE 0
5193 #define TICK_SCHED_REMOTE_OFFLINING 1
5194 #define TICK_SCHED_REMOTE_RUNNING 2
5197 * State diagram for ->state:
5200 * TICK_SCHED_REMOTE_OFFLINE
5203 * | | sched_tick_remote()
5206 * +--TICK_SCHED_REMOTE_OFFLINING
5209 * sched_tick_start() | | sched_tick_stop()
5212 * TICK_SCHED_REMOTE_RUNNING
5215 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5216 * and sched_tick_start() are happy to leave the state in RUNNING.
5219 static struct tick_work __percpu *tick_work_cpu;
5221 static void sched_tick_remote(struct work_struct *work)
5223 struct delayed_work *dwork = to_delayed_work(work);
5224 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5225 int cpu = twork->cpu;
5226 struct rq *rq = cpu_rq(cpu);
5227 struct task_struct *curr;
5233 * Handle the tick only if it appears the remote CPU is running in full
5234 * dynticks mode. The check is racy by nature, but missing a tick or
5235 * having one too much is no big deal because the scheduler tick updates
5236 * statistics and checks timeslices in a time-independent way, regardless
5237 * of when exactly it is running.
5239 if (!tick_nohz_tick_stopped_cpu(cpu))
5242 rq_lock_irq(rq, &rf);
5244 if (cpu_is_offline(cpu))
5247 update_rq_clock(rq);
5249 if (!is_idle_task(curr)) {
5251 * Make sure the next tick runs within a reasonable
5254 delta = rq_clock_task(rq) - curr->se.exec_start;
5255 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5257 curr->sched_class->task_tick(rq, curr, 0);
5259 calc_load_nohz_remote(rq);
5261 rq_unlock_irq(rq, &rf);
5265 * Run the remote tick once per second (1Hz). This arbitrary
5266 * frequency is large enough to avoid overload but short enough
5267 * to keep scheduler internal stats reasonably up to date. But
5268 * first update state to reflect hotplug activity if required.
5270 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5271 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5272 if (os == TICK_SCHED_REMOTE_RUNNING)
5273 queue_delayed_work(system_unbound_wq, dwork, HZ);
5276 static void sched_tick_start(int cpu)
5279 struct tick_work *twork;
5281 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5284 WARN_ON_ONCE(!tick_work_cpu);
5286 twork = per_cpu_ptr(tick_work_cpu, cpu);
5287 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5288 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5289 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5291 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5292 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5296 #ifdef CONFIG_HOTPLUG_CPU
5297 static void sched_tick_stop(int cpu)
5299 struct tick_work *twork;
5302 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5305 WARN_ON_ONCE(!tick_work_cpu);
5307 twork = per_cpu_ptr(tick_work_cpu, cpu);
5308 /* There cannot be competing actions, but don't rely on stop-machine. */
5309 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5310 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5311 /* Don't cancel, as this would mess up the state machine. */
5313 #endif /* CONFIG_HOTPLUG_CPU */
5315 int __init sched_tick_offload_init(void)
5317 tick_work_cpu = alloc_percpu(struct tick_work);
5318 BUG_ON(!tick_work_cpu);
5322 #else /* !CONFIG_NO_HZ_FULL */
5323 static inline void sched_tick_start(int cpu) { }
5324 static inline void sched_tick_stop(int cpu) { }
5327 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5328 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5330 * If the value passed in is equal to the current preempt count
5331 * then we just disabled preemption. Start timing the latency.
5333 static inline void preempt_latency_start(int val)
5335 if (preempt_count() == val) {
5336 unsigned long ip = get_lock_parent_ip();
5337 #ifdef CONFIG_DEBUG_PREEMPT
5338 current->preempt_disable_ip = ip;
5340 trace_preempt_off(CALLER_ADDR0, ip);
5344 void preempt_count_add(int val)
5346 #ifdef CONFIG_DEBUG_PREEMPT
5350 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5353 __preempt_count_add(val);
5354 #ifdef CONFIG_DEBUG_PREEMPT
5356 * Spinlock count overflowing soon?
5358 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5361 preempt_latency_start(val);
5363 EXPORT_SYMBOL(preempt_count_add);
5364 NOKPROBE_SYMBOL(preempt_count_add);
5367 * If the value passed in equals to the current preempt count
5368 * then we just enabled preemption. Stop timing the latency.
5370 static inline void preempt_latency_stop(int val)
5372 if (preempt_count() == val)
5373 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5376 void preempt_count_sub(int val)
5378 #ifdef CONFIG_DEBUG_PREEMPT
5382 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5385 * Is the spinlock portion underflowing?
5387 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5388 !(preempt_count() & PREEMPT_MASK)))
5392 preempt_latency_stop(val);
5393 __preempt_count_sub(val);
5395 EXPORT_SYMBOL(preempt_count_sub);
5396 NOKPROBE_SYMBOL(preempt_count_sub);
5399 static inline void preempt_latency_start(int val) { }
5400 static inline void preempt_latency_stop(int val) { }
5403 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5405 #ifdef CONFIG_DEBUG_PREEMPT
5406 return p->preempt_disable_ip;
5413 * Print scheduling while atomic bug:
5415 static noinline void __schedule_bug(struct task_struct *prev)
5417 /* Save this before calling printk(), since that will clobber it */
5418 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5420 if (oops_in_progress)
5423 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5424 prev->comm, prev->pid, preempt_count());
5426 debug_show_held_locks(prev);
5428 if (irqs_disabled())
5429 print_irqtrace_events(prev);
5430 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5431 && in_atomic_preempt_off()) {
5432 pr_err("Preemption disabled at:");
5433 print_ip_sym(KERN_ERR, preempt_disable_ip);
5436 panic("scheduling while atomic\n");
5439 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5443 * Various schedule()-time debugging checks and statistics:
5445 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5447 #ifdef CONFIG_SCHED_STACK_END_CHECK
5448 if (task_stack_end_corrupted(prev))
5449 panic("corrupted stack end detected inside scheduler\n");
5451 if (task_scs_end_corrupted(prev))
5452 panic("corrupted shadow stack detected inside scheduler\n");
5455 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5456 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5457 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5458 prev->comm, prev->pid, prev->non_block_count);
5460 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5464 if (unlikely(in_atomic_preempt_off())) {
5465 __schedule_bug(prev);
5466 preempt_count_set(PREEMPT_DISABLED);
5469 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5471 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5473 schedstat_inc(this_rq()->sched_count);
5476 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5477 struct rq_flags *rf)
5480 const struct sched_class *class;
5482 * We must do the balancing pass before put_prev_task(), such
5483 * that when we release the rq->lock the task is in the same
5484 * state as before we took rq->lock.
5486 * We can terminate the balance pass as soon as we know there is
5487 * a runnable task of @class priority or higher.
5489 for_class_range(class, prev->sched_class, &idle_sched_class) {
5490 if (class->balance(rq, prev, rf))
5495 put_prev_task(rq, prev);
5499 * Pick up the highest-prio task:
5501 static inline struct task_struct *
5502 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5504 const struct sched_class *class;
5505 struct task_struct *p;
5508 * Optimization: we know that if all tasks are in the fair class we can
5509 * call that function directly, but only if the @prev task wasn't of a
5510 * higher scheduling class, because otherwise those lose the
5511 * opportunity to pull in more work from other CPUs.
5513 if (likely(prev->sched_class <= &fair_sched_class &&
5514 rq->nr_running == rq->cfs.h_nr_running)) {
5516 p = pick_next_task_fair(rq, prev, rf);
5517 if (unlikely(p == RETRY_TASK))
5520 /* Assume the next prioritized class is idle_sched_class */
5522 put_prev_task(rq, prev);
5523 p = pick_next_task_idle(rq);
5530 put_prev_task_balance(rq, prev, rf);
5532 for_each_class(class) {
5533 p = class->pick_next_task(rq);
5538 /* The idle class should always have a runnable task: */
5542 #ifdef CONFIG_SCHED_CORE
5543 static inline bool is_task_rq_idle(struct task_struct *t)
5545 return (task_rq(t)->idle == t);
5548 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5550 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5553 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5555 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5558 return a->core_cookie == b->core_cookie;
5561 // XXX fairness/fwd progress conditions
5564 * - NULL if there is no runnable task for this class.
5565 * - the highest priority task for this runqueue if it matches
5566 * rq->core->core_cookie or its priority is greater than max.
5567 * - Else returns idle_task.
5569 static struct task_struct *
5570 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5572 struct task_struct *class_pick, *cookie_pick;
5573 unsigned long cookie = rq->core->core_cookie;
5575 class_pick = class->pick_task(rq);
5581 * If class_pick is tagged, return it only if it has
5582 * higher priority than max.
5584 if (max && class_pick->core_cookie &&
5585 prio_less(class_pick, max, in_fi))
5586 return idle_sched_class.pick_task(rq);
5592 * If class_pick is idle or matches cookie, return early.
5594 if (cookie_equals(class_pick, cookie))
5597 cookie_pick = sched_core_find(rq, cookie);
5600 * If class > max && class > cookie, it is the highest priority task on
5601 * the core (so far) and it must be selected, otherwise we must go with
5602 * the cookie pick in order to satisfy the constraint.
5604 if (prio_less(cookie_pick, class_pick, in_fi) &&
5605 (!max || prio_less(max, class_pick, in_fi)))
5611 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5613 static struct task_struct *
5614 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5616 struct task_struct *next, *max = NULL;
5617 const struct sched_class *class;
5618 const struct cpumask *smt_mask;
5619 bool fi_before = false;
5620 int i, j, cpu, occ = 0;
5623 if (!sched_core_enabled(rq))
5624 return __pick_next_task(rq, prev, rf);
5628 /* Stopper task is switching into idle, no need core-wide selection. */
5629 if (cpu_is_offline(cpu)) {
5631 * Reset core_pick so that we don't enter the fastpath when
5632 * coming online. core_pick would already be migrated to
5633 * another cpu during offline.
5635 rq->core_pick = NULL;
5636 return __pick_next_task(rq, prev, rf);
5640 * If there were no {en,de}queues since we picked (IOW, the task
5641 * pointers are all still valid), and we haven't scheduled the last
5642 * pick yet, do so now.
5644 * rq->core_pick can be NULL if no selection was made for a CPU because
5645 * it was either offline or went offline during a sibling's core-wide
5646 * selection. In this case, do a core-wide selection.
5648 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5649 rq->core->core_pick_seq != rq->core_sched_seq &&
5651 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5653 next = rq->core_pick;
5655 put_prev_task(rq, prev);
5656 set_next_task(rq, next);
5659 rq->core_pick = NULL;
5663 put_prev_task_balance(rq, prev, rf);
5665 smt_mask = cpu_smt_mask(cpu);
5666 need_sync = !!rq->core->core_cookie;
5669 rq->core->core_cookie = 0UL;
5670 if (rq->core->core_forceidle) {
5673 rq->core->core_forceidle = false;
5677 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5679 * @task_seq guards the task state ({en,de}queues)
5680 * @pick_seq is the @task_seq we did a selection on
5681 * @sched_seq is the @pick_seq we scheduled
5683 * However, preemptions can cause multiple picks on the same task set.
5684 * 'Fix' this by also increasing @task_seq for every pick.
5686 rq->core->core_task_seq++;
5689 * Optimize for common case where this CPU has no cookies
5690 * and there are no cookied tasks running on siblings.
5693 for_each_class(class) {
5694 next = class->pick_task(rq);
5699 if (!next->core_cookie) {
5700 rq->core_pick = NULL;
5702 * For robustness, update the min_vruntime_fi for
5703 * unconstrained picks as well.
5705 WARN_ON_ONCE(fi_before);
5706 task_vruntime_update(rq, next, false);
5711 for_each_cpu(i, smt_mask) {
5712 struct rq *rq_i = cpu_rq(i);
5714 rq_i->core_pick = NULL;
5717 update_rq_clock(rq_i);
5721 * Try and select tasks for each sibling in descending sched_class
5724 for_each_class(class) {
5726 for_each_cpu_wrap(i, smt_mask, cpu) {
5727 struct rq *rq_i = cpu_rq(i);
5728 struct task_struct *p;
5730 if (rq_i->core_pick)
5734 * If this sibling doesn't yet have a suitable task to
5735 * run; ask for the most eligible task, given the
5736 * highest priority task already selected for this
5739 p = pick_task(rq_i, class, max, fi_before);
5743 if (!is_task_rq_idle(p))
5746 rq_i->core_pick = p;
5747 if (rq_i->idle == p && rq_i->nr_running) {
5748 rq->core->core_forceidle = true;
5750 rq->core->core_forceidle_seq++;
5754 * If this new candidate is of higher priority than the
5755 * previous; and they're incompatible; we need to wipe
5756 * the slate and start over. pick_task makes sure that
5757 * p's priority is more than max if it doesn't match
5760 * NOTE: this is a linear max-filter and is thus bounded
5761 * in execution time.
5763 if (!max || !cookie_match(max, p)) {
5764 struct task_struct *old_max = max;
5766 rq->core->core_cookie = p->core_cookie;
5770 rq->core->core_forceidle = false;
5771 for_each_cpu(j, smt_mask) {
5775 cpu_rq(j)->core_pick = NULL;
5784 rq->core->core_pick_seq = rq->core->core_task_seq;
5785 next = rq->core_pick;
5786 rq->core_sched_seq = rq->core->core_pick_seq;
5788 /* Something should have been selected for current CPU */
5789 WARN_ON_ONCE(!next);
5792 * Reschedule siblings
5794 * NOTE: L1TF -- at this point we're no longer running the old task and
5795 * sending an IPI (below) ensures the sibling will no longer be running
5796 * their task. This ensures there is no inter-sibling overlap between
5797 * non-matching user state.
5799 for_each_cpu(i, smt_mask) {
5800 struct rq *rq_i = cpu_rq(i);
5803 * An online sibling might have gone offline before a task
5804 * could be picked for it, or it might be offline but later
5805 * happen to come online, but its too late and nothing was
5806 * picked for it. That's Ok - it will pick tasks for itself,
5809 if (!rq_i->core_pick)
5813 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5814 * fi_before fi update?
5820 if (!(fi_before && rq->core->core_forceidle))
5821 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5823 rq_i->core_pick->core_occupation = occ;
5826 rq_i->core_pick = NULL;
5830 /* Did we break L1TF mitigation requirements? */
5831 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5833 if (rq_i->curr == rq_i->core_pick) {
5834 rq_i->core_pick = NULL;
5842 set_next_task(rq, next);
5846 static bool try_steal_cookie(int this, int that)
5848 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5849 struct task_struct *p;
5850 unsigned long cookie;
5851 bool success = false;
5853 local_irq_disable();
5854 double_rq_lock(dst, src);
5856 cookie = dst->core->core_cookie;
5860 if (dst->curr != dst->idle)
5863 p = sched_core_find(src, cookie);
5868 if (p == src->core_pick || p == src->curr)
5871 if (!cpumask_test_cpu(this, &p->cpus_mask))
5874 if (p->core_occupation > dst->idle->core_occupation)
5877 deactivate_task(src, p, 0);
5878 set_task_cpu(p, this);
5879 activate_task(dst, p, 0);
5887 p = sched_core_next(p, cookie);
5891 double_rq_unlock(dst, src);
5897 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5901 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5908 if (try_steal_cookie(cpu, i))
5915 static void sched_core_balance(struct rq *rq)
5917 struct sched_domain *sd;
5918 int cpu = cpu_of(rq);
5922 raw_spin_rq_unlock_irq(rq);
5923 for_each_domain(cpu, sd) {
5927 if (steal_cookie_task(cpu, sd))
5930 raw_spin_rq_lock_irq(rq);
5935 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5937 void queue_core_balance(struct rq *rq)
5939 if (!sched_core_enabled(rq))
5942 if (!rq->core->core_cookie)
5945 if (!rq->nr_running) /* not forced idle */
5948 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5951 static void sched_core_cpu_starting(unsigned int cpu)
5953 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5954 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
5955 unsigned long flags;
5958 sched_core_lock(cpu, &flags);
5960 WARN_ON_ONCE(rq->core != rq);
5962 /* if we're the first, we'll be our own leader */
5963 if (cpumask_weight(smt_mask) == 1)
5966 /* find the leader */
5967 for_each_cpu(t, smt_mask) {
5971 if (rq->core == rq) {
5977 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
5980 /* install and validate core_rq */
5981 for_each_cpu(t, smt_mask) {
5987 WARN_ON_ONCE(rq->core != core_rq);
5991 sched_core_unlock(cpu, &flags);
5994 static void sched_core_cpu_deactivate(unsigned int cpu)
5996 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5997 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
5998 unsigned long flags;
6001 sched_core_lock(cpu, &flags);
6003 /* if we're the last man standing, nothing to do */
6004 if (cpumask_weight(smt_mask) == 1) {
6005 WARN_ON_ONCE(rq->core != rq);
6009 /* if we're not the leader, nothing to do */
6013 /* find a new leader */
6014 for_each_cpu(t, smt_mask) {
6017 core_rq = cpu_rq(t);
6021 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6024 /* copy the shared state to the new leader */
6025 core_rq->core_task_seq = rq->core_task_seq;
6026 core_rq->core_pick_seq = rq->core_pick_seq;
6027 core_rq->core_cookie = rq->core_cookie;
6028 core_rq->core_forceidle = rq->core_forceidle;
6029 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6031 /* install new leader */
6032 for_each_cpu(t, smt_mask) {
6038 sched_core_unlock(cpu, &flags);
6041 static inline void sched_core_cpu_dying(unsigned int cpu)
6043 struct rq *rq = cpu_rq(cpu);
6049 #else /* !CONFIG_SCHED_CORE */
6051 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6052 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6053 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6055 static struct task_struct *
6056 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6058 return __pick_next_task(rq, prev, rf);
6061 #endif /* CONFIG_SCHED_CORE */
6064 * __schedule() is the main scheduler function.
6066 * The main means of driving the scheduler and thus entering this function are:
6068 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6070 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6071 * paths. For example, see arch/x86/entry_64.S.
6073 * To drive preemption between tasks, the scheduler sets the flag in timer
6074 * interrupt handler scheduler_tick().
6076 * 3. Wakeups don't really cause entry into schedule(). They add a
6077 * task to the run-queue and that's it.
6079 * Now, if the new task added to the run-queue preempts the current
6080 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6081 * called on the nearest possible occasion:
6083 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6085 * - in syscall or exception context, at the next outmost
6086 * preempt_enable(). (this might be as soon as the wake_up()'s
6089 * - in IRQ context, return from interrupt-handler to
6090 * preemptible context
6092 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6095 * - cond_resched() call
6096 * - explicit schedule() call
6097 * - return from syscall or exception to user-space
6098 * - return from interrupt-handler to user-space
6100 * WARNING: must be called with preemption disabled!
6102 static void __sched notrace __schedule(bool preempt)
6104 struct task_struct *prev, *next;
6105 unsigned long *switch_count;
6106 unsigned long prev_state;
6111 cpu = smp_processor_id();
6115 schedule_debug(prev, preempt);
6117 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6120 local_irq_disable();
6121 rcu_note_context_switch(preempt);
6124 * Make sure that signal_pending_state()->signal_pending() below
6125 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6126 * done by the caller to avoid the race with signal_wake_up():
6128 * __set_current_state(@state) signal_wake_up()
6129 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6130 * wake_up_state(p, state)
6131 * LOCK rq->lock LOCK p->pi_state
6132 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6133 * if (signal_pending_state()) if (p->state & @state)
6135 * Also, the membarrier system call requires a full memory barrier
6136 * after coming from user-space, before storing to rq->curr.
6139 smp_mb__after_spinlock();
6141 /* Promote REQ to ACT */
6142 rq->clock_update_flags <<= 1;
6143 update_rq_clock(rq);
6145 switch_count = &prev->nivcsw;
6148 * We must load prev->state once (task_struct::state is volatile), such
6151 * - we form a control dependency vs deactivate_task() below.
6152 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
6154 prev_state = READ_ONCE(prev->__state);
6155 if (!preempt && prev_state) {
6156 if (signal_pending_state(prev_state, prev)) {
6157 WRITE_ONCE(prev->__state, TASK_RUNNING);
6159 prev->sched_contributes_to_load =
6160 (prev_state & TASK_UNINTERRUPTIBLE) &&
6161 !(prev_state & TASK_NOLOAD) &&
6162 !(prev->flags & PF_FROZEN);
6164 if (prev->sched_contributes_to_load)
6165 rq->nr_uninterruptible++;
6168 * __schedule() ttwu()
6169 * prev_state = prev->state; if (p->on_rq && ...)
6170 * if (prev_state) goto out;
6171 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6172 * p->state = TASK_WAKING
6174 * Where __schedule() and ttwu() have matching control dependencies.
6176 * After this, schedule() must not care about p->state any more.
6178 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6180 if (prev->in_iowait) {
6181 atomic_inc(&rq->nr_iowait);
6182 delayacct_blkio_start();
6185 switch_count = &prev->nvcsw;
6188 next = pick_next_task(rq, prev, &rf);
6189 clear_tsk_need_resched(prev);
6190 clear_preempt_need_resched();
6191 #ifdef CONFIG_SCHED_DEBUG
6192 rq->last_seen_need_resched_ns = 0;
6195 if (likely(prev != next)) {
6198 * RCU users of rcu_dereference(rq->curr) may not see
6199 * changes to task_struct made by pick_next_task().
6201 RCU_INIT_POINTER(rq->curr, next);
6203 * The membarrier system call requires each architecture
6204 * to have a full memory barrier after updating
6205 * rq->curr, before returning to user-space.
6207 * Here are the schemes providing that barrier on the
6208 * various architectures:
6209 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6210 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6211 * - finish_lock_switch() for weakly-ordered
6212 * architectures where spin_unlock is a full barrier,
6213 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6214 * is a RELEASE barrier),
6218 migrate_disable_switch(rq, prev);
6219 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6221 trace_sched_switch(preempt, prev, next);
6223 /* Also unlocks the rq: */
6224 rq = context_switch(rq, prev, next, &rf);
6226 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6228 rq_unpin_lock(rq, &rf);
6229 __balance_callbacks(rq);
6230 raw_spin_rq_unlock_irq(rq);
6234 void __noreturn do_task_dead(void)
6236 /* Causes final put_task_struct in finish_task_switch(): */
6237 set_special_state(TASK_DEAD);
6239 /* Tell freezer to ignore us: */
6240 current->flags |= PF_NOFREEZE;
6245 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6250 static inline void sched_submit_work(struct task_struct *tsk)
6252 unsigned int task_flags;
6254 if (task_is_running(tsk))
6257 task_flags = tsk->flags;
6259 * If a worker went to sleep, notify and ask workqueue whether
6260 * it wants to wake up a task to maintain concurrency.
6261 * As this function is called inside the schedule() context,
6262 * we disable preemption to avoid it calling schedule() again
6263 * in the possible wakeup of a kworker and because wq_worker_sleeping()
6266 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6268 if (task_flags & PF_WQ_WORKER)
6269 wq_worker_sleeping(tsk);
6271 io_wq_worker_sleeping(tsk);
6272 preempt_enable_no_resched();
6275 if (tsk_is_pi_blocked(tsk))
6279 * If we are going to sleep and we have plugged IO queued,
6280 * make sure to submit it to avoid deadlocks.
6282 if (blk_needs_flush_plug(tsk))
6283 blk_schedule_flush_plug(tsk);
6286 static void sched_update_worker(struct task_struct *tsk)
6288 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6289 if (tsk->flags & PF_WQ_WORKER)
6290 wq_worker_running(tsk);
6292 io_wq_worker_running(tsk);
6296 asmlinkage __visible void __sched schedule(void)
6298 struct task_struct *tsk = current;
6300 sched_submit_work(tsk);
6304 sched_preempt_enable_no_resched();
6305 } while (need_resched());
6306 sched_update_worker(tsk);
6308 EXPORT_SYMBOL(schedule);
6311 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6312 * state (have scheduled out non-voluntarily) by making sure that all
6313 * tasks have either left the run queue or have gone into user space.
6314 * As idle tasks do not do either, they must not ever be preempted
6315 * (schedule out non-voluntarily).
6317 * schedule_idle() is similar to schedule_preempt_disable() except that it
6318 * never enables preemption because it does not call sched_submit_work().
6320 void __sched schedule_idle(void)
6323 * As this skips calling sched_submit_work(), which the idle task does
6324 * regardless because that function is a nop when the task is in a
6325 * TASK_RUNNING state, make sure this isn't used someplace that the
6326 * current task can be in any other state. Note, idle is always in the
6327 * TASK_RUNNING state.
6329 WARN_ON_ONCE(current->__state);
6332 } while (need_resched());
6335 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6336 asmlinkage __visible void __sched schedule_user(void)
6339 * If we come here after a random call to set_need_resched(),
6340 * or we have been woken up remotely but the IPI has not yet arrived,
6341 * we haven't yet exited the RCU idle mode. Do it here manually until
6342 * we find a better solution.
6344 * NB: There are buggy callers of this function. Ideally we
6345 * should warn if prev_state != CONTEXT_USER, but that will trigger
6346 * too frequently to make sense yet.
6348 enum ctx_state prev_state = exception_enter();
6350 exception_exit(prev_state);
6355 * schedule_preempt_disabled - called with preemption disabled
6357 * Returns with preemption disabled. Note: preempt_count must be 1
6359 void __sched schedule_preempt_disabled(void)
6361 sched_preempt_enable_no_resched();
6366 static void __sched notrace preempt_schedule_common(void)
6370 * Because the function tracer can trace preempt_count_sub()
6371 * and it also uses preempt_enable/disable_notrace(), if
6372 * NEED_RESCHED is set, the preempt_enable_notrace() called
6373 * by the function tracer will call this function again and
6374 * cause infinite recursion.
6376 * Preemption must be disabled here before the function
6377 * tracer can trace. Break up preempt_disable() into two
6378 * calls. One to disable preemption without fear of being
6379 * traced. The other to still record the preemption latency,
6380 * which can also be traced by the function tracer.
6382 preempt_disable_notrace();
6383 preempt_latency_start(1);
6385 preempt_latency_stop(1);
6386 preempt_enable_no_resched_notrace();
6389 * Check again in case we missed a preemption opportunity
6390 * between schedule and now.
6392 } while (need_resched());
6395 #ifdef CONFIG_PREEMPTION
6397 * This is the entry point to schedule() from in-kernel preemption
6398 * off of preempt_enable.
6400 asmlinkage __visible void __sched notrace preempt_schedule(void)
6403 * If there is a non-zero preempt_count or interrupts are disabled,
6404 * we do not want to preempt the current task. Just return..
6406 if (likely(!preemptible()))
6409 preempt_schedule_common();
6411 NOKPROBE_SYMBOL(preempt_schedule);
6412 EXPORT_SYMBOL(preempt_schedule);
6414 #ifdef CONFIG_PREEMPT_DYNAMIC
6415 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6416 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6421 * preempt_schedule_notrace - preempt_schedule called by tracing
6423 * The tracing infrastructure uses preempt_enable_notrace to prevent
6424 * recursion and tracing preempt enabling caused by the tracing
6425 * infrastructure itself. But as tracing can happen in areas coming
6426 * from userspace or just about to enter userspace, a preempt enable
6427 * can occur before user_exit() is called. This will cause the scheduler
6428 * to be called when the system is still in usermode.
6430 * To prevent this, the preempt_enable_notrace will use this function
6431 * instead of preempt_schedule() to exit user context if needed before
6432 * calling the scheduler.
6434 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6436 enum ctx_state prev_ctx;
6438 if (likely(!preemptible()))
6443 * Because the function tracer can trace preempt_count_sub()
6444 * and it also uses preempt_enable/disable_notrace(), if
6445 * NEED_RESCHED is set, the preempt_enable_notrace() called
6446 * by the function tracer will call this function again and
6447 * cause infinite recursion.
6449 * Preemption must be disabled here before the function
6450 * tracer can trace. Break up preempt_disable() into two
6451 * calls. One to disable preemption without fear of being
6452 * traced. The other to still record the preemption latency,
6453 * which can also be traced by the function tracer.
6455 preempt_disable_notrace();
6456 preempt_latency_start(1);
6458 * Needs preempt disabled in case user_exit() is traced
6459 * and the tracer calls preempt_enable_notrace() causing
6460 * an infinite recursion.
6462 prev_ctx = exception_enter();
6464 exception_exit(prev_ctx);
6466 preempt_latency_stop(1);
6467 preempt_enable_no_resched_notrace();
6468 } while (need_resched());
6470 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6472 #ifdef CONFIG_PREEMPT_DYNAMIC
6473 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6474 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6477 #endif /* CONFIG_PREEMPTION */
6479 #ifdef CONFIG_PREEMPT_DYNAMIC
6481 #include <linux/entry-common.h>
6486 * SC:preempt_schedule
6487 * SC:preempt_schedule_notrace
6488 * SC:irqentry_exit_cond_resched
6492 * cond_resched <- __cond_resched
6493 * might_resched <- RET0
6494 * preempt_schedule <- NOP
6495 * preempt_schedule_notrace <- NOP
6496 * irqentry_exit_cond_resched <- NOP
6499 * cond_resched <- __cond_resched
6500 * might_resched <- __cond_resched
6501 * preempt_schedule <- NOP
6502 * preempt_schedule_notrace <- NOP
6503 * irqentry_exit_cond_resched <- NOP
6506 * cond_resched <- RET0
6507 * might_resched <- RET0
6508 * preempt_schedule <- preempt_schedule
6509 * preempt_schedule_notrace <- preempt_schedule_notrace
6510 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6514 preempt_dynamic_none = 0,
6515 preempt_dynamic_voluntary,
6516 preempt_dynamic_full,
6519 int preempt_dynamic_mode = preempt_dynamic_full;
6521 int sched_dynamic_mode(const char *str)
6523 if (!strcmp(str, "none"))
6524 return preempt_dynamic_none;
6526 if (!strcmp(str, "voluntary"))
6527 return preempt_dynamic_voluntary;
6529 if (!strcmp(str, "full"))
6530 return preempt_dynamic_full;
6535 void sched_dynamic_update(int mode)
6538 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6539 * the ZERO state, which is invalid.
6541 static_call_update(cond_resched, __cond_resched);
6542 static_call_update(might_resched, __cond_resched);
6543 static_call_update(preempt_schedule, __preempt_schedule_func);
6544 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6545 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6548 case preempt_dynamic_none:
6549 static_call_update(cond_resched, __cond_resched);
6550 static_call_update(might_resched, (void *)&__static_call_return0);
6551 static_call_update(preempt_schedule, NULL);
6552 static_call_update(preempt_schedule_notrace, NULL);
6553 static_call_update(irqentry_exit_cond_resched, NULL);
6554 pr_info("Dynamic Preempt: none\n");
6557 case preempt_dynamic_voluntary:
6558 static_call_update(cond_resched, __cond_resched);
6559 static_call_update(might_resched, __cond_resched);
6560 static_call_update(preempt_schedule, NULL);
6561 static_call_update(preempt_schedule_notrace, NULL);
6562 static_call_update(irqentry_exit_cond_resched, NULL);
6563 pr_info("Dynamic Preempt: voluntary\n");
6566 case preempt_dynamic_full:
6567 static_call_update(cond_resched, (void *)&__static_call_return0);
6568 static_call_update(might_resched, (void *)&__static_call_return0);
6569 static_call_update(preempt_schedule, __preempt_schedule_func);
6570 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6571 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6572 pr_info("Dynamic Preempt: full\n");
6576 preempt_dynamic_mode = mode;
6579 static int __init setup_preempt_mode(char *str)
6581 int mode = sched_dynamic_mode(str);
6583 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6587 sched_dynamic_update(mode);
6590 __setup("preempt=", setup_preempt_mode);
6592 #endif /* CONFIG_PREEMPT_DYNAMIC */
6595 * This is the entry point to schedule() from kernel preemption
6596 * off of irq context.
6597 * Note, that this is called and return with irqs disabled. This will
6598 * protect us against recursive calling from irq.
6600 asmlinkage __visible void __sched preempt_schedule_irq(void)
6602 enum ctx_state prev_state;
6604 /* Catch callers which need to be fixed */
6605 BUG_ON(preempt_count() || !irqs_disabled());
6607 prev_state = exception_enter();
6613 local_irq_disable();
6614 sched_preempt_enable_no_resched();
6615 } while (need_resched());
6617 exception_exit(prev_state);
6620 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6623 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6624 return try_to_wake_up(curr->private, mode, wake_flags);
6626 EXPORT_SYMBOL(default_wake_function);
6628 static void __setscheduler_prio(struct task_struct *p, int prio)
6631 p->sched_class = &dl_sched_class;
6632 else if (rt_prio(prio))
6633 p->sched_class = &rt_sched_class;
6635 p->sched_class = &fair_sched_class;
6640 #ifdef CONFIG_RT_MUTEXES
6642 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6645 prio = min(prio, pi_task->prio);
6650 static inline int rt_effective_prio(struct task_struct *p, int prio)
6652 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6654 return __rt_effective_prio(pi_task, prio);
6658 * rt_mutex_setprio - set the current priority of a task
6660 * @pi_task: donor task
6662 * This function changes the 'effective' priority of a task. It does
6663 * not touch ->normal_prio like __setscheduler().
6665 * Used by the rt_mutex code to implement priority inheritance
6666 * logic. Call site only calls if the priority of the task changed.
6668 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6670 int prio, oldprio, queued, running, queue_flag =
6671 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6672 const struct sched_class *prev_class;
6676 /* XXX used to be waiter->prio, not waiter->task->prio */
6677 prio = __rt_effective_prio(pi_task, p->normal_prio);
6680 * If nothing changed; bail early.
6682 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6685 rq = __task_rq_lock(p, &rf);
6686 update_rq_clock(rq);
6688 * Set under pi_lock && rq->lock, such that the value can be used under
6691 * Note that there is loads of tricky to make this pointer cache work
6692 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6693 * ensure a task is de-boosted (pi_task is set to NULL) before the
6694 * task is allowed to run again (and can exit). This ensures the pointer
6695 * points to a blocked task -- which guarantees the task is present.
6697 p->pi_top_task = pi_task;
6700 * For FIFO/RR we only need to set prio, if that matches we're done.
6702 if (prio == p->prio && !dl_prio(prio))
6706 * Idle task boosting is a nono in general. There is one
6707 * exception, when PREEMPT_RT and NOHZ is active:
6709 * The idle task calls get_next_timer_interrupt() and holds
6710 * the timer wheel base->lock on the CPU and another CPU wants
6711 * to access the timer (probably to cancel it). We can safely
6712 * ignore the boosting request, as the idle CPU runs this code
6713 * with interrupts disabled and will complete the lock
6714 * protected section without being interrupted. So there is no
6715 * real need to boost.
6717 if (unlikely(p == rq->idle)) {
6718 WARN_ON(p != rq->curr);
6719 WARN_ON(p->pi_blocked_on);
6723 trace_sched_pi_setprio(p, pi_task);
6726 if (oldprio == prio)
6727 queue_flag &= ~DEQUEUE_MOVE;
6729 prev_class = p->sched_class;
6730 queued = task_on_rq_queued(p);
6731 running = task_current(rq, p);
6733 dequeue_task(rq, p, queue_flag);
6735 put_prev_task(rq, p);
6738 * Boosting condition are:
6739 * 1. -rt task is running and holds mutex A
6740 * --> -dl task blocks on mutex A
6742 * 2. -dl task is running and holds mutex A
6743 * --> -dl task blocks on mutex A and could preempt the
6746 if (dl_prio(prio)) {
6747 if (!dl_prio(p->normal_prio) ||
6748 (pi_task && dl_prio(pi_task->prio) &&
6749 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6750 p->dl.pi_se = pi_task->dl.pi_se;
6751 queue_flag |= ENQUEUE_REPLENISH;
6753 p->dl.pi_se = &p->dl;
6755 } else if (rt_prio(prio)) {
6756 if (dl_prio(oldprio))
6757 p->dl.pi_se = &p->dl;
6759 queue_flag |= ENQUEUE_HEAD;
6761 if (dl_prio(oldprio))
6762 p->dl.pi_se = &p->dl;
6763 if (rt_prio(oldprio))
6767 __setscheduler_prio(p, prio);
6770 enqueue_task(rq, p, queue_flag);
6772 set_next_task(rq, p);
6774 check_class_changed(rq, p, prev_class, oldprio);
6776 /* Avoid rq from going away on us: */
6779 rq_unpin_lock(rq, &rf);
6780 __balance_callbacks(rq);
6781 raw_spin_rq_unlock(rq);
6786 static inline int rt_effective_prio(struct task_struct *p, int prio)
6792 void set_user_nice(struct task_struct *p, long nice)
6794 bool queued, running;
6799 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6802 * We have to be careful, if called from sys_setpriority(),
6803 * the task might be in the middle of scheduling on another CPU.
6805 rq = task_rq_lock(p, &rf);
6806 update_rq_clock(rq);
6809 * The RT priorities are set via sched_setscheduler(), but we still
6810 * allow the 'normal' nice value to be set - but as expected
6811 * it won't have any effect on scheduling until the task is
6812 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6814 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6815 p->static_prio = NICE_TO_PRIO(nice);
6818 queued = task_on_rq_queued(p);
6819 running = task_current(rq, p);
6821 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6823 put_prev_task(rq, p);
6825 p->static_prio = NICE_TO_PRIO(nice);
6826 set_load_weight(p, true);
6828 p->prio = effective_prio(p);
6831 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6833 set_next_task(rq, p);
6836 * If the task increased its priority or is running and
6837 * lowered its priority, then reschedule its CPU:
6839 p->sched_class->prio_changed(rq, p, old_prio);
6842 task_rq_unlock(rq, p, &rf);
6844 EXPORT_SYMBOL(set_user_nice);
6847 * can_nice - check if a task can reduce its nice value
6851 int can_nice(const struct task_struct *p, const int nice)
6853 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6854 int nice_rlim = nice_to_rlimit(nice);
6856 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6857 capable(CAP_SYS_NICE));
6860 #ifdef __ARCH_WANT_SYS_NICE
6863 * sys_nice - change the priority of the current process.
6864 * @increment: priority increment
6866 * sys_setpriority is a more generic, but much slower function that
6867 * does similar things.
6869 SYSCALL_DEFINE1(nice, int, increment)
6874 * Setpriority might change our priority at the same moment.
6875 * We don't have to worry. Conceptually one call occurs first
6876 * and we have a single winner.
6878 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6879 nice = task_nice(current) + increment;
6881 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6882 if (increment < 0 && !can_nice(current, nice))
6885 retval = security_task_setnice(current, nice);
6889 set_user_nice(current, nice);
6896 * task_prio - return the priority value of a given task.
6897 * @p: the task in question.
6899 * Return: The priority value as seen by users in /proc.
6901 * sched policy return value kernel prio user prio/nice
6903 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6904 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6905 * deadline -101 -1 0
6907 int task_prio(const struct task_struct *p)
6909 return p->prio - MAX_RT_PRIO;
6913 * idle_cpu - is a given CPU idle currently?
6914 * @cpu: the processor in question.
6916 * Return: 1 if the CPU is currently idle. 0 otherwise.
6918 int idle_cpu(int cpu)
6920 struct rq *rq = cpu_rq(cpu);
6922 if (rq->curr != rq->idle)
6929 if (rq->ttwu_pending)
6937 * available_idle_cpu - is a given CPU idle for enqueuing work.
6938 * @cpu: the CPU in question.
6940 * Return: 1 if the CPU is currently idle. 0 otherwise.
6942 int available_idle_cpu(int cpu)
6947 if (vcpu_is_preempted(cpu))
6954 * idle_task - return the idle task for a given CPU.
6955 * @cpu: the processor in question.
6957 * Return: The idle task for the CPU @cpu.
6959 struct task_struct *idle_task(int cpu)
6961 return cpu_rq(cpu)->idle;
6966 * This function computes an effective utilization for the given CPU, to be
6967 * used for frequency selection given the linear relation: f = u * f_max.
6969 * The scheduler tracks the following metrics:
6971 * cpu_util_{cfs,rt,dl,irq}()
6974 * Where the cfs,rt and dl util numbers are tracked with the same metric and
6975 * synchronized windows and are thus directly comparable.
6977 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
6978 * which excludes things like IRQ and steal-time. These latter are then accrued
6979 * in the irq utilization.
6981 * The DL bandwidth number otoh is not a measured metric but a value computed
6982 * based on the task model parameters and gives the minimal utilization
6983 * required to meet deadlines.
6985 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
6986 unsigned long max, enum cpu_util_type type,
6987 struct task_struct *p)
6989 unsigned long dl_util, util, irq;
6990 struct rq *rq = cpu_rq(cpu);
6992 if (!uclamp_is_used() &&
6993 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
6998 * Early check to see if IRQ/steal time saturates the CPU, can be
6999 * because of inaccuracies in how we track these -- see
7000 * update_irq_load_avg().
7002 irq = cpu_util_irq(rq);
7003 if (unlikely(irq >= max))
7007 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7008 * CFS tasks and we use the same metric to track the effective
7009 * utilization (PELT windows are synchronized) we can directly add them
7010 * to obtain the CPU's actual utilization.
7012 * CFS and RT utilization can be boosted or capped, depending on
7013 * utilization clamp constraints requested by currently RUNNABLE
7015 * When there are no CFS RUNNABLE tasks, clamps are released and
7016 * frequency will be gracefully reduced with the utilization decay.
7018 util = util_cfs + cpu_util_rt(rq);
7019 if (type == FREQUENCY_UTIL)
7020 util = uclamp_rq_util_with(rq, util, p);
7022 dl_util = cpu_util_dl(rq);
7025 * For frequency selection we do not make cpu_util_dl() a permanent part
7026 * of this sum because we want to use cpu_bw_dl() later on, but we need
7027 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7028 * that we select f_max when there is no idle time.
7030 * NOTE: numerical errors or stop class might cause us to not quite hit
7031 * saturation when we should -- something for later.
7033 if (util + dl_util >= max)
7037 * OTOH, for energy computation we need the estimated running time, so
7038 * include util_dl and ignore dl_bw.
7040 if (type == ENERGY_UTIL)
7044 * There is still idle time; further improve the number by using the
7045 * irq metric. Because IRQ/steal time is hidden from the task clock we
7046 * need to scale the task numbers:
7049 * U' = irq + --------- * U
7052 util = scale_irq_capacity(util, irq, max);
7056 * Bandwidth required by DEADLINE must always be granted while, for
7057 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7058 * to gracefully reduce the frequency when no tasks show up for longer
7061 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7062 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7063 * an interface. So, we only do the latter for now.
7065 if (type == FREQUENCY_UTIL)
7066 util += cpu_bw_dl(rq);
7068 return min(max, util);
7071 unsigned long sched_cpu_util(int cpu, unsigned long max)
7073 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
7076 #endif /* CONFIG_SMP */
7079 * find_process_by_pid - find a process with a matching PID value.
7080 * @pid: the pid in question.
7082 * The task of @pid, if found. %NULL otherwise.
7084 static struct task_struct *find_process_by_pid(pid_t pid)
7086 return pid ? find_task_by_vpid(pid) : current;
7090 * sched_setparam() passes in -1 for its policy, to let the functions
7091 * it calls know not to change it.
7093 #define SETPARAM_POLICY -1
7095 static void __setscheduler_params(struct task_struct *p,
7096 const struct sched_attr *attr)
7098 int policy = attr->sched_policy;
7100 if (policy == SETPARAM_POLICY)
7105 if (dl_policy(policy))
7106 __setparam_dl(p, attr);
7107 else if (fair_policy(policy))
7108 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7111 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7112 * !rt_policy. Always setting this ensures that things like
7113 * getparam()/getattr() don't report silly values for !rt tasks.
7115 p->rt_priority = attr->sched_priority;
7116 p->normal_prio = normal_prio(p);
7117 set_load_weight(p, true);
7121 * Check the target process has a UID that matches the current process's:
7123 static bool check_same_owner(struct task_struct *p)
7125 const struct cred *cred = current_cred(), *pcred;
7129 pcred = __task_cred(p);
7130 match = (uid_eq(cred->euid, pcred->euid) ||
7131 uid_eq(cred->euid, pcred->uid));
7136 static int __sched_setscheduler(struct task_struct *p,
7137 const struct sched_attr *attr,
7140 int oldpolicy = -1, policy = attr->sched_policy;
7141 int retval, oldprio, newprio, queued, running;
7142 const struct sched_class *prev_class;
7143 struct callback_head *head;
7146 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7149 /* The pi code expects interrupts enabled */
7150 BUG_ON(pi && in_interrupt());
7152 /* Double check policy once rq lock held: */
7154 reset_on_fork = p->sched_reset_on_fork;
7155 policy = oldpolicy = p->policy;
7157 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7159 if (!valid_policy(policy))
7163 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7167 * Valid priorities for SCHED_FIFO and SCHED_RR are
7168 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7169 * SCHED_BATCH and SCHED_IDLE is 0.
7171 if (attr->sched_priority > MAX_RT_PRIO-1)
7173 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7174 (rt_policy(policy) != (attr->sched_priority != 0)))
7178 * Allow unprivileged RT tasks to decrease priority:
7180 if (user && !capable(CAP_SYS_NICE)) {
7181 if (fair_policy(policy)) {
7182 if (attr->sched_nice < task_nice(p) &&
7183 !can_nice(p, attr->sched_nice))
7187 if (rt_policy(policy)) {
7188 unsigned long rlim_rtprio =
7189 task_rlimit(p, RLIMIT_RTPRIO);
7191 /* Can't set/change the rt policy: */
7192 if (policy != p->policy && !rlim_rtprio)
7195 /* Can't increase priority: */
7196 if (attr->sched_priority > p->rt_priority &&
7197 attr->sched_priority > rlim_rtprio)
7202 * Can't set/change SCHED_DEADLINE policy at all for now
7203 * (safest behavior); in the future we would like to allow
7204 * unprivileged DL tasks to increase their relative deadline
7205 * or reduce their runtime (both ways reducing utilization)
7207 if (dl_policy(policy))
7211 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7212 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7214 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7215 if (!can_nice(p, task_nice(p)))
7219 /* Can't change other user's priorities: */
7220 if (!check_same_owner(p))
7223 /* Normal users shall not reset the sched_reset_on_fork flag: */
7224 if (p->sched_reset_on_fork && !reset_on_fork)
7229 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7232 retval = security_task_setscheduler(p);
7237 /* Update task specific "requested" clamps */
7238 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7239 retval = uclamp_validate(p, attr);
7248 * Make sure no PI-waiters arrive (or leave) while we are
7249 * changing the priority of the task:
7251 * To be able to change p->policy safely, the appropriate
7252 * runqueue lock must be held.
7254 rq = task_rq_lock(p, &rf);
7255 update_rq_clock(rq);
7258 * Changing the policy of the stop threads its a very bad idea:
7260 if (p == rq->stop) {
7266 * If not changing anything there's no need to proceed further,
7267 * but store a possible modification of reset_on_fork.
7269 if (unlikely(policy == p->policy)) {
7270 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7272 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7274 if (dl_policy(policy) && dl_param_changed(p, attr))
7276 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7279 p->sched_reset_on_fork = reset_on_fork;
7286 #ifdef CONFIG_RT_GROUP_SCHED
7288 * Do not allow realtime tasks into groups that have no runtime
7291 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7292 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7293 !task_group_is_autogroup(task_group(p))) {
7299 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7300 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7301 cpumask_t *span = rq->rd->span;
7304 * Don't allow tasks with an affinity mask smaller than
7305 * the entire root_domain to become SCHED_DEADLINE. We
7306 * will also fail if there's no bandwidth available.
7308 if (!cpumask_subset(span, p->cpus_ptr) ||
7309 rq->rd->dl_bw.bw == 0) {
7317 /* Re-check policy now with rq lock held: */
7318 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7319 policy = oldpolicy = -1;
7320 task_rq_unlock(rq, p, &rf);
7322 cpuset_read_unlock();
7327 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7328 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7331 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7336 p->sched_reset_on_fork = reset_on_fork;
7339 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7342 * Take priority boosted tasks into account. If the new
7343 * effective priority is unchanged, we just store the new
7344 * normal parameters and do not touch the scheduler class and
7345 * the runqueue. This will be done when the task deboost
7348 newprio = rt_effective_prio(p, newprio);
7349 if (newprio == oldprio)
7350 queue_flags &= ~DEQUEUE_MOVE;
7353 queued = task_on_rq_queued(p);
7354 running = task_current(rq, p);
7356 dequeue_task(rq, p, queue_flags);
7358 put_prev_task(rq, p);
7360 prev_class = p->sched_class;
7362 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7363 __setscheduler_params(p, attr);
7364 __setscheduler_prio(p, newprio);
7366 __setscheduler_uclamp(p, attr);
7370 * We enqueue to tail when the priority of a task is
7371 * increased (user space view).
7373 if (oldprio < p->prio)
7374 queue_flags |= ENQUEUE_HEAD;
7376 enqueue_task(rq, p, queue_flags);
7379 set_next_task(rq, p);
7381 check_class_changed(rq, p, prev_class, oldprio);
7383 /* Avoid rq from going away on us: */
7385 head = splice_balance_callbacks(rq);
7386 task_rq_unlock(rq, p, &rf);
7389 cpuset_read_unlock();
7390 rt_mutex_adjust_pi(p);
7393 /* Run balance callbacks after we've adjusted the PI chain: */
7394 balance_callbacks(rq, head);
7400 task_rq_unlock(rq, p, &rf);
7402 cpuset_read_unlock();
7406 static int _sched_setscheduler(struct task_struct *p, int policy,
7407 const struct sched_param *param, bool check)
7409 struct sched_attr attr = {
7410 .sched_policy = policy,
7411 .sched_priority = param->sched_priority,
7412 .sched_nice = PRIO_TO_NICE(p->static_prio),
7415 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7416 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7417 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7418 policy &= ~SCHED_RESET_ON_FORK;
7419 attr.sched_policy = policy;
7422 return __sched_setscheduler(p, &attr, check, true);
7425 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7426 * @p: the task in question.
7427 * @policy: new policy.
7428 * @param: structure containing the new RT priority.
7430 * Use sched_set_fifo(), read its comment.
7432 * Return: 0 on success. An error code otherwise.
7434 * NOTE that the task may be already dead.
7436 int sched_setscheduler(struct task_struct *p, int policy,
7437 const struct sched_param *param)
7439 return _sched_setscheduler(p, policy, param, true);
7442 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7444 return __sched_setscheduler(p, attr, true, true);
7447 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7449 return __sched_setscheduler(p, attr, false, true);
7451 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7454 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7455 * @p: the task in question.
7456 * @policy: new policy.
7457 * @param: structure containing the new RT priority.
7459 * Just like sched_setscheduler, only don't bother checking if the
7460 * current context has permission. For example, this is needed in
7461 * stop_machine(): we create temporary high priority worker threads,
7462 * but our caller might not have that capability.
7464 * Return: 0 on success. An error code otherwise.
7466 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7467 const struct sched_param *param)
7469 return _sched_setscheduler(p, policy, param, false);
7473 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7474 * incapable of resource management, which is the one thing an OS really should
7477 * This is of course the reason it is limited to privileged users only.
7479 * Worse still; it is fundamentally impossible to compose static priority
7480 * workloads. You cannot take two correctly working static prio workloads
7481 * and smash them together and still expect them to work.
7483 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7487 * The administrator _MUST_ configure the system, the kernel simply doesn't
7488 * know enough information to make a sensible choice.
7490 void sched_set_fifo(struct task_struct *p)
7492 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7493 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7495 EXPORT_SYMBOL_GPL(sched_set_fifo);
7498 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7500 void sched_set_fifo_low(struct task_struct *p)
7502 struct sched_param sp = { .sched_priority = 1 };
7503 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7505 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7507 void sched_set_normal(struct task_struct *p, int nice)
7509 struct sched_attr attr = {
7510 .sched_policy = SCHED_NORMAL,
7513 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7515 EXPORT_SYMBOL_GPL(sched_set_normal);
7518 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7520 struct sched_param lparam;
7521 struct task_struct *p;
7524 if (!param || pid < 0)
7526 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7531 p = find_process_by_pid(pid);
7537 retval = sched_setscheduler(p, policy, &lparam);
7545 * Mimics kernel/events/core.c perf_copy_attr().
7547 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7552 /* Zero the full structure, so that a short copy will be nice: */
7553 memset(attr, 0, sizeof(*attr));
7555 ret = get_user(size, &uattr->size);
7559 /* ABI compatibility quirk: */
7561 size = SCHED_ATTR_SIZE_VER0;
7562 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7565 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7572 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7573 size < SCHED_ATTR_SIZE_VER1)
7577 * XXX: Do we want to be lenient like existing syscalls; or do we want
7578 * to be strict and return an error on out-of-bounds values?
7580 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7585 put_user(sizeof(*attr), &uattr->size);
7589 static void get_params(struct task_struct *p, struct sched_attr *attr)
7591 if (task_has_dl_policy(p))
7592 __getparam_dl(p, attr);
7593 else if (task_has_rt_policy(p))
7594 attr->sched_priority = p->rt_priority;
7596 attr->sched_nice = task_nice(p);
7600 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7601 * @pid: the pid in question.
7602 * @policy: new policy.
7603 * @param: structure containing the new RT priority.
7605 * Return: 0 on success. An error code otherwise.
7607 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7612 return do_sched_setscheduler(pid, policy, param);
7616 * sys_sched_setparam - set/change the RT priority of a thread
7617 * @pid: the pid in question.
7618 * @param: structure containing the new RT priority.
7620 * Return: 0 on success. An error code otherwise.
7622 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7624 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7628 * sys_sched_setattr - same as above, but with extended sched_attr
7629 * @pid: the pid in question.
7630 * @uattr: structure containing the extended parameters.
7631 * @flags: for future extension.
7633 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7634 unsigned int, flags)
7636 struct sched_attr attr;
7637 struct task_struct *p;
7640 if (!uattr || pid < 0 || flags)
7643 retval = sched_copy_attr(uattr, &attr);
7647 if ((int)attr.sched_policy < 0)
7649 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7650 attr.sched_policy = SETPARAM_POLICY;
7654 p = find_process_by_pid(pid);
7660 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7661 get_params(p, &attr);
7662 retval = sched_setattr(p, &attr);
7670 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7671 * @pid: the pid in question.
7673 * Return: On success, the policy of the thread. Otherwise, a negative error
7676 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7678 struct task_struct *p;
7686 p = find_process_by_pid(pid);
7688 retval = security_task_getscheduler(p);
7691 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7698 * sys_sched_getparam - get the RT priority of a thread
7699 * @pid: the pid in question.
7700 * @param: structure containing the RT priority.
7702 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7705 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7707 struct sched_param lp = { .sched_priority = 0 };
7708 struct task_struct *p;
7711 if (!param || pid < 0)
7715 p = find_process_by_pid(pid);
7720 retval = security_task_getscheduler(p);
7724 if (task_has_rt_policy(p))
7725 lp.sched_priority = p->rt_priority;
7729 * This one might sleep, we cannot do it with a spinlock held ...
7731 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7741 * Copy the kernel size attribute structure (which might be larger
7742 * than what user-space knows about) to user-space.
7744 * Note that all cases are valid: user-space buffer can be larger or
7745 * smaller than the kernel-space buffer. The usual case is that both
7746 * have the same size.
7749 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7750 struct sched_attr *kattr,
7753 unsigned int ksize = sizeof(*kattr);
7755 if (!access_ok(uattr, usize))
7759 * sched_getattr() ABI forwards and backwards compatibility:
7761 * If usize == ksize then we just copy everything to user-space and all is good.
7763 * If usize < ksize then we only copy as much as user-space has space for,
7764 * this keeps ABI compatibility as well. We skip the rest.
7766 * If usize > ksize then user-space is using a newer version of the ABI,
7767 * which part the kernel doesn't know about. Just ignore it - tooling can
7768 * detect the kernel's knowledge of attributes from the attr->size value
7769 * which is set to ksize in this case.
7771 kattr->size = min(usize, ksize);
7773 if (copy_to_user(uattr, kattr, kattr->size))
7780 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7781 * @pid: the pid in question.
7782 * @uattr: structure containing the extended parameters.
7783 * @usize: sizeof(attr) for fwd/bwd comp.
7784 * @flags: for future extension.
7786 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7787 unsigned int, usize, unsigned int, flags)
7789 struct sched_attr kattr = { };
7790 struct task_struct *p;
7793 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7794 usize < SCHED_ATTR_SIZE_VER0 || flags)
7798 p = find_process_by_pid(pid);
7803 retval = security_task_getscheduler(p);
7807 kattr.sched_policy = p->policy;
7808 if (p->sched_reset_on_fork)
7809 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7810 get_params(p, &kattr);
7811 kattr.sched_flags &= SCHED_FLAG_ALL;
7813 #ifdef CONFIG_UCLAMP_TASK
7815 * This could race with another potential updater, but this is fine
7816 * because it'll correctly read the old or the new value. We don't need
7817 * to guarantee who wins the race as long as it doesn't return garbage.
7819 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7820 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7825 return sched_attr_copy_to_user(uattr, &kattr, usize);
7833 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7838 * If the task isn't a deadline task or admission control is
7839 * disabled then we don't care about affinity changes.
7841 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7845 * Since bandwidth control happens on root_domain basis,
7846 * if admission test is enabled, we only admit -deadline
7847 * tasks allowed to run on all the CPUs in the task's
7851 if (!cpumask_subset(task_rq(p)->rd->span, mask))
7859 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
7862 cpumask_var_t cpus_allowed, new_mask;
7864 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
7867 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7869 goto out_free_cpus_allowed;
7872 cpuset_cpus_allowed(p, cpus_allowed);
7873 cpumask_and(new_mask, mask, cpus_allowed);
7875 retval = dl_task_check_affinity(p, new_mask);
7877 goto out_free_new_mask;
7879 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
7881 goto out_free_new_mask;
7883 cpuset_cpus_allowed(p, cpus_allowed);
7884 if (!cpumask_subset(new_mask, cpus_allowed)) {
7886 * We must have raced with a concurrent cpuset update.
7887 * Just reset the cpumask to the cpuset's cpus_allowed.
7889 cpumask_copy(new_mask, cpus_allowed);
7894 free_cpumask_var(new_mask);
7895 out_free_cpus_allowed:
7896 free_cpumask_var(cpus_allowed);
7900 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7902 struct task_struct *p;
7907 p = find_process_by_pid(pid);
7913 /* Prevent p going away */
7917 if (p->flags & PF_NO_SETAFFINITY) {
7922 if (!check_same_owner(p)) {
7924 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7932 retval = security_task_setscheduler(p);
7936 retval = __sched_setaffinity(p, in_mask);
7942 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7943 struct cpumask *new_mask)
7945 if (len < cpumask_size())
7946 cpumask_clear(new_mask);
7947 else if (len > cpumask_size())
7948 len = cpumask_size();
7950 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
7954 * sys_sched_setaffinity - set the CPU affinity of a process
7955 * @pid: pid of the process
7956 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7957 * @user_mask_ptr: user-space pointer to the new CPU mask
7959 * Return: 0 on success. An error code otherwise.
7961 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
7962 unsigned long __user *, user_mask_ptr)
7964 cpumask_var_t new_mask;
7967 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
7970 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
7972 retval = sched_setaffinity(pid, new_mask);
7973 free_cpumask_var(new_mask);
7977 long sched_getaffinity(pid_t pid, struct cpumask *mask)
7979 struct task_struct *p;
7980 unsigned long flags;
7986 p = find_process_by_pid(pid);
7990 retval = security_task_getscheduler(p);
7994 raw_spin_lock_irqsave(&p->pi_lock, flags);
7995 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
7996 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8005 * sys_sched_getaffinity - get the CPU affinity of a process
8006 * @pid: pid of the process
8007 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8008 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8010 * Return: size of CPU mask copied to user_mask_ptr on success. An
8011 * error code otherwise.
8013 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8014 unsigned long __user *, user_mask_ptr)
8019 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8021 if (len & (sizeof(unsigned long)-1))
8024 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8027 ret = sched_getaffinity(pid, mask);
8029 unsigned int retlen = min(len, cpumask_size());
8031 if (copy_to_user(user_mask_ptr, mask, retlen))
8036 free_cpumask_var(mask);
8041 static void do_sched_yield(void)
8046 rq = this_rq_lock_irq(&rf);
8048 schedstat_inc(rq->yld_count);
8049 current->sched_class->yield_task(rq);
8052 rq_unlock_irq(rq, &rf);
8053 sched_preempt_enable_no_resched();
8059 * sys_sched_yield - yield the current processor to other threads.
8061 * This function yields the current CPU to other tasks. If there are no
8062 * other threads running on this CPU then this function will return.
8066 SYSCALL_DEFINE0(sched_yield)
8072 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8073 int __sched __cond_resched(void)
8075 if (should_resched(0)) {
8076 preempt_schedule_common();
8080 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8081 * whether the current CPU is in an RCU read-side critical section,
8082 * so the tick can report quiescent states even for CPUs looping
8083 * in kernel context. In contrast, in non-preemptible kernels,
8084 * RCU readers leave no in-memory hints, which means that CPU-bound
8085 * processes executing in kernel context might never report an
8086 * RCU quiescent state. Therefore, the following code causes
8087 * cond_resched() to report a quiescent state, but only when RCU
8088 * is in urgent need of one.
8090 #ifndef CONFIG_PREEMPT_RCU
8095 EXPORT_SYMBOL(__cond_resched);
8098 #ifdef CONFIG_PREEMPT_DYNAMIC
8099 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8100 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8102 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8103 EXPORT_STATIC_CALL_TRAMP(might_resched);
8107 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8108 * call schedule, and on return reacquire the lock.
8110 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8111 * operations here to prevent schedule() from being called twice (once via
8112 * spin_unlock(), once by hand).
8114 int __cond_resched_lock(spinlock_t *lock)
8116 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8119 lockdep_assert_held(lock);
8121 if (spin_needbreak(lock) || resched) {
8124 preempt_schedule_common();
8132 EXPORT_SYMBOL(__cond_resched_lock);
8134 int __cond_resched_rwlock_read(rwlock_t *lock)
8136 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8139 lockdep_assert_held_read(lock);
8141 if (rwlock_needbreak(lock) || resched) {
8144 preempt_schedule_common();
8152 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8154 int __cond_resched_rwlock_write(rwlock_t *lock)
8156 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8159 lockdep_assert_held_write(lock);
8161 if (rwlock_needbreak(lock) || resched) {
8164 preempt_schedule_common();
8172 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8175 * yield - yield the current processor to other threads.
8177 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8179 * The scheduler is at all times free to pick the calling task as the most
8180 * eligible task to run, if removing the yield() call from your code breaks
8181 * it, it's already broken.
8183 * Typical broken usage is:
8188 * where one assumes that yield() will let 'the other' process run that will
8189 * make event true. If the current task is a SCHED_FIFO task that will never
8190 * happen. Never use yield() as a progress guarantee!!
8192 * If you want to use yield() to wait for something, use wait_event().
8193 * If you want to use yield() to be 'nice' for others, use cond_resched().
8194 * If you still want to use yield(), do not!
8196 void __sched yield(void)
8198 set_current_state(TASK_RUNNING);
8201 EXPORT_SYMBOL(yield);
8204 * yield_to - yield the current processor to another thread in
8205 * your thread group, or accelerate that thread toward the
8206 * processor it's on.
8208 * @preempt: whether task preemption is allowed or not
8210 * It's the caller's job to ensure that the target task struct
8211 * can't go away on us before we can do any checks.
8214 * true (>0) if we indeed boosted the target task.
8215 * false (0) if we failed to boost the target.
8216 * -ESRCH if there's no task to yield to.
8218 int __sched yield_to(struct task_struct *p, bool preempt)
8220 struct task_struct *curr = current;
8221 struct rq *rq, *p_rq;
8222 unsigned long flags;
8225 local_irq_save(flags);
8231 * If we're the only runnable task on the rq and target rq also
8232 * has only one task, there's absolutely no point in yielding.
8234 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8239 double_rq_lock(rq, p_rq);
8240 if (task_rq(p) != p_rq) {
8241 double_rq_unlock(rq, p_rq);
8245 if (!curr->sched_class->yield_to_task)
8248 if (curr->sched_class != p->sched_class)
8251 if (task_running(p_rq, p) || !task_is_running(p))
8254 yielded = curr->sched_class->yield_to_task(rq, p);
8256 schedstat_inc(rq->yld_count);
8258 * Make p's CPU reschedule; pick_next_entity takes care of
8261 if (preempt && rq != p_rq)
8266 double_rq_unlock(rq, p_rq);
8268 local_irq_restore(flags);
8275 EXPORT_SYMBOL_GPL(yield_to);
8277 int io_schedule_prepare(void)
8279 int old_iowait = current->in_iowait;
8281 current->in_iowait = 1;
8282 blk_schedule_flush_plug(current);
8287 void io_schedule_finish(int token)
8289 current->in_iowait = token;
8293 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8294 * that process accounting knows that this is a task in IO wait state.
8296 long __sched io_schedule_timeout(long timeout)
8301 token = io_schedule_prepare();
8302 ret = schedule_timeout(timeout);
8303 io_schedule_finish(token);
8307 EXPORT_SYMBOL(io_schedule_timeout);
8309 void __sched io_schedule(void)
8313 token = io_schedule_prepare();
8315 io_schedule_finish(token);
8317 EXPORT_SYMBOL(io_schedule);
8320 * sys_sched_get_priority_max - return maximum RT priority.
8321 * @policy: scheduling class.
8323 * Return: On success, this syscall returns the maximum
8324 * rt_priority that can be used by a given scheduling class.
8325 * On failure, a negative error code is returned.
8327 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8334 ret = MAX_RT_PRIO-1;
8336 case SCHED_DEADLINE:
8347 * sys_sched_get_priority_min - return minimum RT priority.
8348 * @policy: scheduling class.
8350 * Return: On success, this syscall returns the minimum
8351 * rt_priority that can be used by a given scheduling class.
8352 * On failure, a negative error code is returned.
8354 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8363 case SCHED_DEADLINE:
8372 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8374 struct task_struct *p;
8375 unsigned int time_slice;
8385 p = find_process_by_pid(pid);
8389 retval = security_task_getscheduler(p);
8393 rq = task_rq_lock(p, &rf);
8395 if (p->sched_class->get_rr_interval)
8396 time_slice = p->sched_class->get_rr_interval(rq, p);
8397 task_rq_unlock(rq, p, &rf);
8400 jiffies_to_timespec64(time_slice, t);
8409 * sys_sched_rr_get_interval - return the default timeslice of a process.
8410 * @pid: pid of the process.
8411 * @interval: userspace pointer to the timeslice value.
8413 * this syscall writes the default timeslice value of a given process
8414 * into the user-space timespec buffer. A value of '0' means infinity.
8416 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8419 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8420 struct __kernel_timespec __user *, interval)
8422 struct timespec64 t;
8423 int retval = sched_rr_get_interval(pid, &t);
8426 retval = put_timespec64(&t, interval);
8431 #ifdef CONFIG_COMPAT_32BIT_TIME
8432 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8433 struct old_timespec32 __user *, interval)
8435 struct timespec64 t;
8436 int retval = sched_rr_get_interval(pid, &t);
8439 retval = put_old_timespec32(&t, interval);
8444 void sched_show_task(struct task_struct *p)
8446 unsigned long free = 0;
8449 if (!try_get_task_stack(p))
8452 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8454 if (task_is_running(p))
8455 pr_cont(" running task ");
8456 #ifdef CONFIG_DEBUG_STACK_USAGE
8457 free = stack_not_used(p);
8462 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8464 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8465 free, task_pid_nr(p), ppid,
8466 (unsigned long)task_thread_info(p)->flags);
8468 print_worker_info(KERN_INFO, p);
8469 print_stop_info(KERN_INFO, p);
8470 show_stack(p, NULL, KERN_INFO);
8473 EXPORT_SYMBOL_GPL(sched_show_task);
8476 state_filter_match(unsigned long state_filter, struct task_struct *p)
8478 unsigned int state = READ_ONCE(p->__state);
8480 /* no filter, everything matches */
8484 /* filter, but doesn't match */
8485 if (!(state & state_filter))
8489 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8492 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8499 void show_state_filter(unsigned int state_filter)
8501 struct task_struct *g, *p;
8504 for_each_process_thread(g, p) {
8506 * reset the NMI-timeout, listing all files on a slow
8507 * console might take a lot of time:
8508 * Also, reset softlockup watchdogs on all CPUs, because
8509 * another CPU might be blocked waiting for us to process
8512 touch_nmi_watchdog();
8513 touch_all_softlockup_watchdogs();
8514 if (state_filter_match(state_filter, p))
8518 #ifdef CONFIG_SCHED_DEBUG
8520 sysrq_sched_debug_show();
8524 * Only show locks if all tasks are dumped:
8527 debug_show_all_locks();
8531 * init_idle - set up an idle thread for a given CPU
8532 * @idle: task in question
8533 * @cpu: CPU the idle task belongs to
8535 * NOTE: this function does not set the idle thread's NEED_RESCHED
8536 * flag, to make booting more robust.
8538 void __init init_idle(struct task_struct *idle, int cpu)
8540 struct rq *rq = cpu_rq(cpu);
8541 unsigned long flags;
8543 __sched_fork(0, idle);
8546 * The idle task doesn't need the kthread struct to function, but it
8547 * is dressed up as a per-CPU kthread and thus needs to play the part
8548 * if we want to avoid special-casing it in code that deals with per-CPU
8551 set_kthread_struct(idle);
8553 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8554 raw_spin_rq_lock(rq);
8556 idle->__state = TASK_RUNNING;
8557 idle->se.exec_start = sched_clock();
8559 * PF_KTHREAD should already be set at this point; regardless, make it
8560 * look like a proper per-CPU kthread.
8562 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8563 kthread_set_per_cpu(idle, cpu);
8565 scs_task_reset(idle);
8566 kasan_unpoison_task_stack(idle);
8570 * It's possible that init_idle() gets called multiple times on a task,
8571 * in that case do_set_cpus_allowed() will not do the right thing.
8573 * And since this is boot we can forgo the serialization.
8575 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8578 * We're having a chicken and egg problem, even though we are
8579 * holding rq->lock, the CPU isn't yet set to this CPU so the
8580 * lockdep check in task_group() will fail.
8582 * Similar case to sched_fork(). / Alternatively we could
8583 * use task_rq_lock() here and obtain the other rq->lock.
8588 __set_task_cpu(idle, cpu);
8592 rcu_assign_pointer(rq->curr, idle);
8593 idle->on_rq = TASK_ON_RQ_QUEUED;
8597 raw_spin_rq_unlock(rq);
8598 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8600 /* Set the preempt count _outside_ the spinlocks! */
8601 init_idle_preempt_count(idle, cpu);
8604 * The idle tasks have their own, simple scheduling class:
8606 idle->sched_class = &idle_sched_class;
8607 ftrace_graph_init_idle_task(idle, cpu);
8608 vtime_init_idle(idle, cpu);
8610 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8616 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8617 const struct cpumask *trial)
8621 if (!cpumask_weight(cur))
8624 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8629 int task_can_attach(struct task_struct *p,
8630 const struct cpumask *cs_cpus_allowed)
8635 * Kthreads which disallow setaffinity shouldn't be moved
8636 * to a new cpuset; we don't want to change their CPU
8637 * affinity and isolating such threads by their set of
8638 * allowed nodes is unnecessary. Thus, cpusets are not
8639 * applicable for such threads. This prevents checking for
8640 * success of set_cpus_allowed_ptr() on all attached tasks
8641 * before cpus_mask may be changed.
8643 if (p->flags & PF_NO_SETAFFINITY) {
8648 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8650 ret = dl_task_can_attach(p, cs_cpus_allowed);
8656 bool sched_smp_initialized __read_mostly;
8658 #ifdef CONFIG_NUMA_BALANCING
8659 /* Migrate current task p to target_cpu */
8660 int migrate_task_to(struct task_struct *p, int target_cpu)
8662 struct migration_arg arg = { p, target_cpu };
8663 int curr_cpu = task_cpu(p);
8665 if (curr_cpu == target_cpu)
8668 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8671 /* TODO: This is not properly updating schedstats */
8673 trace_sched_move_numa(p, curr_cpu, target_cpu);
8674 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8678 * Requeue a task on a given node and accurately track the number of NUMA
8679 * tasks on the runqueues
8681 void sched_setnuma(struct task_struct *p, int nid)
8683 bool queued, running;
8687 rq = task_rq_lock(p, &rf);
8688 queued = task_on_rq_queued(p);
8689 running = task_current(rq, p);
8692 dequeue_task(rq, p, DEQUEUE_SAVE);
8694 put_prev_task(rq, p);
8696 p->numa_preferred_nid = nid;
8699 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8701 set_next_task(rq, p);
8702 task_rq_unlock(rq, p, &rf);
8704 #endif /* CONFIG_NUMA_BALANCING */
8706 #ifdef CONFIG_HOTPLUG_CPU
8708 * Ensure that the idle task is using init_mm right before its CPU goes
8711 void idle_task_exit(void)
8713 struct mm_struct *mm = current->active_mm;
8715 BUG_ON(cpu_online(smp_processor_id()));
8716 BUG_ON(current != this_rq()->idle);
8718 if (mm != &init_mm) {
8719 switch_mm(mm, &init_mm, current);
8720 finish_arch_post_lock_switch();
8723 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8726 static int __balance_push_cpu_stop(void *arg)
8728 struct task_struct *p = arg;
8729 struct rq *rq = this_rq();
8733 raw_spin_lock_irq(&p->pi_lock);
8736 update_rq_clock(rq);
8738 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8739 cpu = select_fallback_rq(rq->cpu, p);
8740 rq = __migrate_task(rq, &rf, p, cpu);
8744 raw_spin_unlock_irq(&p->pi_lock);
8751 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8754 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8756 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8757 * effective when the hotplug motion is down.
8759 static void balance_push(struct rq *rq)
8761 struct task_struct *push_task = rq->curr;
8763 lockdep_assert_rq_held(rq);
8764 SCHED_WARN_ON(rq->cpu != smp_processor_id());
8767 * Ensure the thing is persistent until balance_push_set(.on = false);
8769 rq->balance_callback = &balance_push_callback;
8772 * Only active while going offline.
8774 if (!cpu_dying(rq->cpu))
8778 * Both the cpu-hotplug and stop task are in this case and are
8779 * required to complete the hotplug process.
8781 if (kthread_is_per_cpu(push_task) ||
8782 is_migration_disabled(push_task)) {
8785 * If this is the idle task on the outgoing CPU try to wake
8786 * up the hotplug control thread which might wait for the
8787 * last task to vanish. The rcuwait_active() check is
8788 * accurate here because the waiter is pinned on this CPU
8789 * and can't obviously be running in parallel.
8791 * On RT kernels this also has to check whether there are
8792 * pinned and scheduled out tasks on the runqueue. They
8793 * need to leave the migrate disabled section first.
8795 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8796 rcuwait_active(&rq->hotplug_wait)) {
8797 raw_spin_rq_unlock(rq);
8798 rcuwait_wake_up(&rq->hotplug_wait);
8799 raw_spin_rq_lock(rq);
8804 get_task_struct(push_task);
8806 * Temporarily drop rq->lock such that we can wake-up the stop task.
8807 * Both preemption and IRQs are still disabled.
8809 raw_spin_rq_unlock(rq);
8810 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8811 this_cpu_ptr(&push_work));
8813 * At this point need_resched() is true and we'll take the loop in
8814 * schedule(). The next pick is obviously going to be the stop task
8815 * which kthread_is_per_cpu() and will push this task away.
8817 raw_spin_rq_lock(rq);
8820 static void balance_push_set(int cpu, bool on)
8822 struct rq *rq = cpu_rq(cpu);
8825 rq_lock_irqsave(rq, &rf);
8827 WARN_ON_ONCE(rq->balance_callback);
8828 rq->balance_callback = &balance_push_callback;
8829 } else if (rq->balance_callback == &balance_push_callback) {
8830 rq->balance_callback = NULL;
8832 rq_unlock_irqrestore(rq, &rf);
8836 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8837 * inactive. All tasks which are not per CPU kernel threads are either
8838 * pushed off this CPU now via balance_push() or placed on a different CPU
8839 * during wakeup. Wait until the CPU is quiescent.
8841 static void balance_hotplug_wait(void)
8843 struct rq *rq = this_rq();
8845 rcuwait_wait_event(&rq->hotplug_wait,
8846 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8847 TASK_UNINTERRUPTIBLE);
8852 static inline void balance_push(struct rq *rq)
8856 static inline void balance_push_set(int cpu, bool on)
8860 static inline void balance_hotplug_wait(void)
8864 #endif /* CONFIG_HOTPLUG_CPU */
8866 void set_rq_online(struct rq *rq)
8869 const struct sched_class *class;
8871 cpumask_set_cpu(rq->cpu, rq->rd->online);
8874 for_each_class(class) {
8875 if (class->rq_online)
8876 class->rq_online(rq);
8881 void set_rq_offline(struct rq *rq)
8884 const struct sched_class *class;
8886 for_each_class(class) {
8887 if (class->rq_offline)
8888 class->rq_offline(rq);
8891 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8897 * used to mark begin/end of suspend/resume:
8899 static int num_cpus_frozen;
8902 * Update cpusets according to cpu_active mask. If cpusets are
8903 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8904 * around partition_sched_domains().
8906 * If we come here as part of a suspend/resume, don't touch cpusets because we
8907 * want to restore it back to its original state upon resume anyway.
8909 static void cpuset_cpu_active(void)
8911 if (cpuhp_tasks_frozen) {
8913 * num_cpus_frozen tracks how many CPUs are involved in suspend
8914 * resume sequence. As long as this is not the last online
8915 * operation in the resume sequence, just build a single sched
8916 * domain, ignoring cpusets.
8918 partition_sched_domains(1, NULL, NULL);
8919 if (--num_cpus_frozen)
8922 * This is the last CPU online operation. So fall through and
8923 * restore the original sched domains by considering the
8924 * cpuset configurations.
8926 cpuset_force_rebuild();
8928 cpuset_update_active_cpus();
8931 static int cpuset_cpu_inactive(unsigned int cpu)
8933 if (!cpuhp_tasks_frozen) {
8934 if (dl_cpu_busy(cpu))
8936 cpuset_update_active_cpus();
8939 partition_sched_domains(1, NULL, NULL);
8944 int sched_cpu_activate(unsigned int cpu)
8946 struct rq *rq = cpu_rq(cpu);
8950 * Clear the balance_push callback and prepare to schedule
8953 balance_push_set(cpu, false);
8955 #ifdef CONFIG_SCHED_SMT
8957 * When going up, increment the number of cores with SMT present.
8959 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8960 static_branch_inc_cpuslocked(&sched_smt_present);
8962 set_cpu_active(cpu, true);
8964 if (sched_smp_initialized) {
8965 sched_domains_numa_masks_set(cpu);
8966 cpuset_cpu_active();
8970 * Put the rq online, if not already. This happens:
8972 * 1) In the early boot process, because we build the real domains
8973 * after all CPUs have been brought up.
8975 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
8978 rq_lock_irqsave(rq, &rf);
8980 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8983 rq_unlock_irqrestore(rq, &rf);
8988 int sched_cpu_deactivate(unsigned int cpu)
8990 struct rq *rq = cpu_rq(cpu);
8995 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
8996 * load balancing when not active
8998 nohz_balance_exit_idle(rq);
9000 set_cpu_active(cpu, false);
9003 * From this point forward, this CPU will refuse to run any task that
9004 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9005 * push those tasks away until this gets cleared, see
9006 * sched_cpu_dying().
9008 balance_push_set(cpu, true);
9011 * We've cleared cpu_active_mask / set balance_push, wait for all
9012 * preempt-disabled and RCU users of this state to go away such that
9013 * all new such users will observe it.
9015 * Specifically, we rely on ttwu to no longer target this CPU, see
9016 * ttwu_queue_cond() and is_cpu_allowed().
9018 * Do sync before park smpboot threads to take care the rcu boost case.
9022 rq_lock_irqsave(rq, &rf);
9024 update_rq_clock(rq);
9025 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9028 rq_unlock_irqrestore(rq, &rf);
9030 #ifdef CONFIG_SCHED_SMT
9032 * When going down, decrement the number of cores with SMT present.
9034 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9035 static_branch_dec_cpuslocked(&sched_smt_present);
9037 sched_core_cpu_deactivate(cpu);
9040 if (!sched_smp_initialized)
9043 ret = cpuset_cpu_inactive(cpu);
9045 balance_push_set(cpu, false);
9046 set_cpu_active(cpu, true);
9049 sched_domains_numa_masks_clear(cpu);
9053 static void sched_rq_cpu_starting(unsigned int cpu)
9055 struct rq *rq = cpu_rq(cpu);
9057 rq->calc_load_update = calc_load_update;
9058 update_max_interval();
9061 int sched_cpu_starting(unsigned int cpu)
9063 sched_core_cpu_starting(cpu);
9064 sched_rq_cpu_starting(cpu);
9065 sched_tick_start(cpu);
9069 #ifdef CONFIG_HOTPLUG_CPU
9072 * Invoked immediately before the stopper thread is invoked to bring the
9073 * CPU down completely. At this point all per CPU kthreads except the
9074 * hotplug thread (current) and the stopper thread (inactive) have been
9075 * either parked or have been unbound from the outgoing CPU. Ensure that
9076 * any of those which might be on the way out are gone.
9078 * If after this point a bound task is being woken on this CPU then the
9079 * responsible hotplug callback has failed to do it's job.
9080 * sched_cpu_dying() will catch it with the appropriate fireworks.
9082 int sched_cpu_wait_empty(unsigned int cpu)
9084 balance_hotplug_wait();
9089 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9090 * might have. Called from the CPU stopper task after ensuring that the
9091 * stopper is the last running task on the CPU, so nr_active count is
9092 * stable. We need to take the teardown thread which is calling this into
9093 * account, so we hand in adjust = 1 to the load calculation.
9095 * Also see the comment "Global load-average calculations".
9097 static void calc_load_migrate(struct rq *rq)
9099 long delta = calc_load_fold_active(rq, 1);
9102 atomic_long_add(delta, &calc_load_tasks);
9105 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9107 struct task_struct *g, *p;
9108 int cpu = cpu_of(rq);
9110 lockdep_assert_rq_held(rq);
9112 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9113 for_each_process_thread(g, p) {
9114 if (task_cpu(p) != cpu)
9117 if (!task_on_rq_queued(p))
9120 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9124 int sched_cpu_dying(unsigned int cpu)
9126 struct rq *rq = cpu_rq(cpu);
9129 /* Handle pending wakeups and then migrate everything off */
9130 sched_tick_stop(cpu);
9132 rq_lock_irqsave(rq, &rf);
9133 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9134 WARN(true, "Dying CPU not properly vacated!");
9135 dump_rq_tasks(rq, KERN_WARNING);
9137 rq_unlock_irqrestore(rq, &rf);
9139 calc_load_migrate(rq);
9140 update_max_interval();
9142 sched_core_cpu_dying(cpu);
9147 void __init sched_init_smp(void)
9152 * There's no userspace yet to cause hotplug operations; hence all the
9153 * CPU masks are stable and all blatant races in the below code cannot
9156 mutex_lock(&sched_domains_mutex);
9157 sched_init_domains(cpu_active_mask);
9158 mutex_unlock(&sched_domains_mutex);
9160 /* Move init over to a non-isolated CPU */
9161 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9163 current->flags &= ~PF_NO_SETAFFINITY;
9164 sched_init_granularity();
9166 init_sched_rt_class();
9167 init_sched_dl_class();
9169 sched_smp_initialized = true;
9172 static int __init migration_init(void)
9174 sched_cpu_starting(smp_processor_id());
9177 early_initcall(migration_init);
9180 void __init sched_init_smp(void)
9182 sched_init_granularity();
9184 #endif /* CONFIG_SMP */
9186 int in_sched_functions(unsigned long addr)
9188 return in_lock_functions(addr) ||
9189 (addr >= (unsigned long)__sched_text_start
9190 && addr < (unsigned long)__sched_text_end);
9193 #ifdef CONFIG_CGROUP_SCHED
9195 * Default task group.
9196 * Every task in system belongs to this group at bootup.
9198 struct task_group root_task_group;
9199 LIST_HEAD(task_groups);
9201 /* Cacheline aligned slab cache for task_group */
9202 static struct kmem_cache *task_group_cache __read_mostly;
9205 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9206 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9208 void __init sched_init(void)
9210 unsigned long ptr = 0;
9213 /* Make sure the linker didn't screw up */
9214 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9215 &fair_sched_class + 1 != &rt_sched_class ||
9216 &rt_sched_class + 1 != &dl_sched_class);
9218 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9223 #ifdef CONFIG_FAIR_GROUP_SCHED
9224 ptr += 2 * nr_cpu_ids * sizeof(void **);
9226 #ifdef CONFIG_RT_GROUP_SCHED
9227 ptr += 2 * nr_cpu_ids * sizeof(void **);
9230 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9232 #ifdef CONFIG_FAIR_GROUP_SCHED
9233 root_task_group.se = (struct sched_entity **)ptr;
9234 ptr += nr_cpu_ids * sizeof(void **);
9236 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9237 ptr += nr_cpu_ids * sizeof(void **);
9239 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9240 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9241 #endif /* CONFIG_FAIR_GROUP_SCHED */
9242 #ifdef CONFIG_RT_GROUP_SCHED
9243 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9244 ptr += nr_cpu_ids * sizeof(void **);
9246 root_task_group.rt_rq = (struct rt_rq **)ptr;
9247 ptr += nr_cpu_ids * sizeof(void **);
9249 #endif /* CONFIG_RT_GROUP_SCHED */
9251 #ifdef CONFIG_CPUMASK_OFFSTACK
9252 for_each_possible_cpu(i) {
9253 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9254 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9255 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9256 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9258 #endif /* CONFIG_CPUMASK_OFFSTACK */
9260 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9261 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9264 init_defrootdomain();
9267 #ifdef CONFIG_RT_GROUP_SCHED
9268 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9269 global_rt_period(), global_rt_runtime());
9270 #endif /* CONFIG_RT_GROUP_SCHED */
9272 #ifdef CONFIG_CGROUP_SCHED
9273 task_group_cache = KMEM_CACHE(task_group, 0);
9275 list_add(&root_task_group.list, &task_groups);
9276 INIT_LIST_HEAD(&root_task_group.children);
9277 INIT_LIST_HEAD(&root_task_group.siblings);
9278 autogroup_init(&init_task);
9279 #endif /* CONFIG_CGROUP_SCHED */
9281 for_each_possible_cpu(i) {
9285 raw_spin_lock_init(&rq->__lock);
9287 rq->calc_load_active = 0;
9288 rq->calc_load_update = jiffies + LOAD_FREQ;
9289 init_cfs_rq(&rq->cfs);
9290 init_rt_rq(&rq->rt);
9291 init_dl_rq(&rq->dl);
9292 #ifdef CONFIG_FAIR_GROUP_SCHED
9293 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9294 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9296 * How much CPU bandwidth does root_task_group get?
9298 * In case of task-groups formed thr' the cgroup filesystem, it
9299 * gets 100% of the CPU resources in the system. This overall
9300 * system CPU resource is divided among the tasks of
9301 * root_task_group and its child task-groups in a fair manner,
9302 * based on each entity's (task or task-group's) weight
9303 * (se->load.weight).
9305 * In other words, if root_task_group has 10 tasks of weight
9306 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9307 * then A0's share of the CPU resource is:
9309 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9311 * We achieve this by letting root_task_group's tasks sit
9312 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9314 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9315 #endif /* CONFIG_FAIR_GROUP_SCHED */
9317 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9318 #ifdef CONFIG_RT_GROUP_SCHED
9319 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9324 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9325 rq->balance_callback = &balance_push_callback;
9326 rq->active_balance = 0;
9327 rq->next_balance = jiffies;
9332 rq->avg_idle = 2*sysctl_sched_migration_cost;
9333 rq->wake_stamp = jiffies;
9334 rq->wake_avg_idle = rq->avg_idle;
9335 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9337 INIT_LIST_HEAD(&rq->cfs_tasks);
9339 rq_attach_root(rq, &def_root_domain);
9340 #ifdef CONFIG_NO_HZ_COMMON
9341 rq->last_blocked_load_update_tick = jiffies;
9342 atomic_set(&rq->nohz_flags, 0);
9344 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9346 #ifdef CONFIG_HOTPLUG_CPU
9347 rcuwait_init(&rq->hotplug_wait);
9349 #endif /* CONFIG_SMP */
9351 atomic_set(&rq->nr_iowait, 0);
9353 #ifdef CONFIG_SCHED_CORE
9355 rq->core_pick = NULL;
9356 rq->core_enabled = 0;
9357 rq->core_tree = RB_ROOT;
9358 rq->core_forceidle = false;
9360 rq->core_cookie = 0UL;
9364 set_load_weight(&init_task, false);
9367 * The boot idle thread does lazy MMU switching as well:
9370 enter_lazy_tlb(&init_mm, current);
9373 * Make us the idle thread. Technically, schedule() should not be
9374 * called from this thread, however somewhere below it might be,
9375 * but because we are the idle thread, we just pick up running again
9376 * when this runqueue becomes "idle".
9378 init_idle(current, smp_processor_id());
9380 calc_load_update = jiffies + LOAD_FREQ;
9383 idle_thread_set_boot_cpu();
9384 balance_push_set(smp_processor_id(), false);
9386 init_sched_fair_class();
9392 scheduler_running = 1;
9395 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9396 static inline int preempt_count_equals(int preempt_offset)
9398 int nested = preempt_count() + rcu_preempt_depth();
9400 return (nested == preempt_offset);
9403 void __might_sleep(const char *file, int line, int preempt_offset)
9405 unsigned int state = get_current_state();
9407 * Blocking primitives will set (and therefore destroy) current->state,
9408 * since we will exit with TASK_RUNNING make sure we enter with it,
9409 * otherwise we will destroy state.
9411 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9412 "do not call blocking ops when !TASK_RUNNING; "
9413 "state=%x set at [<%p>] %pS\n", state,
9414 (void *)current->task_state_change,
9415 (void *)current->task_state_change);
9417 ___might_sleep(file, line, preempt_offset);
9419 EXPORT_SYMBOL(__might_sleep);
9421 void ___might_sleep(const char *file, int line, int preempt_offset)
9423 /* Ratelimiting timestamp: */
9424 static unsigned long prev_jiffy;
9426 unsigned long preempt_disable_ip;
9428 /* WARN_ON_ONCE() by default, no rate limit required: */
9431 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
9432 !is_idle_task(current) && !current->non_block_count) ||
9433 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9437 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9439 prev_jiffy = jiffies;
9441 /* Save this before calling printk(), since that will clobber it: */
9442 preempt_disable_ip = get_preempt_disable_ip(current);
9445 "BUG: sleeping function called from invalid context at %s:%d\n",
9448 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9449 in_atomic(), irqs_disabled(), current->non_block_count,
9450 current->pid, current->comm);
9452 if (task_stack_end_corrupted(current))
9453 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
9455 debug_show_held_locks(current);
9456 if (irqs_disabled())
9457 print_irqtrace_events(current);
9458 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
9459 && !preempt_count_equals(preempt_offset)) {
9460 pr_err("Preemption disabled at:");
9461 print_ip_sym(KERN_ERR, preempt_disable_ip);
9464 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9466 EXPORT_SYMBOL(___might_sleep);
9468 void __cant_sleep(const char *file, int line, int preempt_offset)
9470 static unsigned long prev_jiffy;
9472 if (irqs_disabled())
9475 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9478 if (preempt_count() > preempt_offset)
9481 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9483 prev_jiffy = jiffies;
9485 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9486 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9487 in_atomic(), irqs_disabled(),
9488 current->pid, current->comm);
9490 debug_show_held_locks(current);
9492 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9494 EXPORT_SYMBOL_GPL(__cant_sleep);
9497 void __cant_migrate(const char *file, int line)
9499 static unsigned long prev_jiffy;
9501 if (irqs_disabled())
9504 if (is_migration_disabled(current))
9507 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9510 if (preempt_count() > 0)
9513 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9515 prev_jiffy = jiffies;
9517 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9518 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9519 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9520 current->pid, current->comm);
9522 debug_show_held_locks(current);
9524 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9526 EXPORT_SYMBOL_GPL(__cant_migrate);
9530 #ifdef CONFIG_MAGIC_SYSRQ
9531 void normalize_rt_tasks(void)
9533 struct task_struct *g, *p;
9534 struct sched_attr attr = {
9535 .sched_policy = SCHED_NORMAL,
9538 read_lock(&tasklist_lock);
9539 for_each_process_thread(g, p) {
9541 * Only normalize user tasks:
9543 if (p->flags & PF_KTHREAD)
9546 p->se.exec_start = 0;
9547 schedstat_set(p->se.statistics.wait_start, 0);
9548 schedstat_set(p->se.statistics.sleep_start, 0);
9549 schedstat_set(p->se.statistics.block_start, 0);
9551 if (!dl_task(p) && !rt_task(p)) {
9553 * Renice negative nice level userspace
9556 if (task_nice(p) < 0)
9557 set_user_nice(p, 0);
9561 __sched_setscheduler(p, &attr, false, false);
9563 read_unlock(&tasklist_lock);
9566 #endif /* CONFIG_MAGIC_SYSRQ */
9568 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9570 * These functions are only useful for the IA64 MCA handling, or kdb.
9572 * They can only be called when the whole system has been
9573 * stopped - every CPU needs to be quiescent, and no scheduling
9574 * activity can take place. Using them for anything else would
9575 * be a serious bug, and as a result, they aren't even visible
9576 * under any other configuration.
9580 * curr_task - return the current task for a given CPU.
9581 * @cpu: the processor in question.
9583 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9585 * Return: The current task for @cpu.
9587 struct task_struct *curr_task(int cpu)
9589 return cpu_curr(cpu);
9592 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9596 * ia64_set_curr_task - set the current task for a given CPU.
9597 * @cpu: the processor in question.
9598 * @p: the task pointer to set.
9600 * Description: This function must only be used when non-maskable interrupts
9601 * are serviced on a separate stack. It allows the architecture to switch the
9602 * notion of the current task on a CPU in a non-blocking manner. This function
9603 * must be called with all CPU's synchronized, and interrupts disabled, the
9604 * and caller must save the original value of the current task (see
9605 * curr_task() above) and restore that value before reenabling interrupts and
9606 * re-starting the system.
9608 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9610 void ia64_set_curr_task(int cpu, struct task_struct *p)
9617 #ifdef CONFIG_CGROUP_SCHED
9618 /* task_group_lock serializes the addition/removal of task groups */
9619 static DEFINE_SPINLOCK(task_group_lock);
9621 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9622 struct task_group *parent)
9624 #ifdef CONFIG_UCLAMP_TASK_GROUP
9625 enum uclamp_id clamp_id;
9627 for_each_clamp_id(clamp_id) {
9628 uclamp_se_set(&tg->uclamp_req[clamp_id],
9629 uclamp_none(clamp_id), false);
9630 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9635 static void sched_free_group(struct task_group *tg)
9637 free_fair_sched_group(tg);
9638 free_rt_sched_group(tg);
9640 kmem_cache_free(task_group_cache, tg);
9643 /* allocate runqueue etc for a new task group */
9644 struct task_group *sched_create_group(struct task_group *parent)
9646 struct task_group *tg;
9648 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9650 return ERR_PTR(-ENOMEM);
9652 if (!alloc_fair_sched_group(tg, parent))
9655 if (!alloc_rt_sched_group(tg, parent))
9658 alloc_uclamp_sched_group(tg, parent);
9663 sched_free_group(tg);
9664 return ERR_PTR(-ENOMEM);
9667 void sched_online_group(struct task_group *tg, struct task_group *parent)
9669 unsigned long flags;
9671 spin_lock_irqsave(&task_group_lock, flags);
9672 list_add_rcu(&tg->list, &task_groups);
9674 /* Root should already exist: */
9677 tg->parent = parent;
9678 INIT_LIST_HEAD(&tg->children);
9679 list_add_rcu(&tg->siblings, &parent->children);
9680 spin_unlock_irqrestore(&task_group_lock, flags);
9682 online_fair_sched_group(tg);
9685 /* rcu callback to free various structures associated with a task group */
9686 static void sched_free_group_rcu(struct rcu_head *rhp)
9688 /* Now it should be safe to free those cfs_rqs: */
9689 sched_free_group(container_of(rhp, struct task_group, rcu));
9692 void sched_destroy_group(struct task_group *tg)
9694 /* Wait for possible concurrent references to cfs_rqs complete: */
9695 call_rcu(&tg->rcu, sched_free_group_rcu);
9698 void sched_offline_group(struct task_group *tg)
9700 unsigned long flags;
9702 /* End participation in shares distribution: */
9703 unregister_fair_sched_group(tg);
9705 spin_lock_irqsave(&task_group_lock, flags);
9706 list_del_rcu(&tg->list);
9707 list_del_rcu(&tg->siblings);
9708 spin_unlock_irqrestore(&task_group_lock, flags);
9711 static void sched_change_group(struct task_struct *tsk, int type)
9713 struct task_group *tg;
9716 * All callers are synchronized by task_rq_lock(); we do not use RCU
9717 * which is pointless here. Thus, we pass "true" to task_css_check()
9718 * to prevent lockdep warnings.
9720 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9721 struct task_group, css);
9722 tg = autogroup_task_group(tsk, tg);
9723 tsk->sched_task_group = tg;
9725 #ifdef CONFIG_FAIR_GROUP_SCHED
9726 if (tsk->sched_class->task_change_group)
9727 tsk->sched_class->task_change_group(tsk, type);
9730 set_task_rq(tsk, task_cpu(tsk));
9734 * Change task's runqueue when it moves between groups.
9736 * The caller of this function should have put the task in its new group by
9737 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9740 void sched_move_task(struct task_struct *tsk)
9742 int queued, running, queue_flags =
9743 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9747 rq = task_rq_lock(tsk, &rf);
9748 update_rq_clock(rq);
9750 running = task_current(rq, tsk);
9751 queued = task_on_rq_queued(tsk);
9754 dequeue_task(rq, tsk, queue_flags);
9756 put_prev_task(rq, tsk);
9758 sched_change_group(tsk, TASK_MOVE_GROUP);
9761 enqueue_task(rq, tsk, queue_flags);
9763 set_next_task(rq, tsk);
9765 * After changing group, the running task may have joined a
9766 * throttled one but it's still the running task. Trigger a
9767 * resched to make sure that task can still run.
9772 task_rq_unlock(rq, tsk, &rf);
9775 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9777 return css ? container_of(css, struct task_group, css) : NULL;
9780 static struct cgroup_subsys_state *
9781 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9783 struct task_group *parent = css_tg(parent_css);
9784 struct task_group *tg;
9787 /* This is early initialization for the top cgroup */
9788 return &root_task_group.css;
9791 tg = sched_create_group(parent);
9793 return ERR_PTR(-ENOMEM);
9798 /* Expose task group only after completing cgroup initialization */
9799 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9801 struct task_group *tg = css_tg(css);
9802 struct task_group *parent = css_tg(css->parent);
9805 sched_online_group(tg, parent);
9807 #ifdef CONFIG_UCLAMP_TASK_GROUP
9808 /* Propagate the effective uclamp value for the new group */
9809 mutex_lock(&uclamp_mutex);
9811 cpu_util_update_eff(css);
9813 mutex_unlock(&uclamp_mutex);
9819 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9821 struct task_group *tg = css_tg(css);
9823 sched_offline_group(tg);
9826 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9828 struct task_group *tg = css_tg(css);
9831 * Relies on the RCU grace period between css_released() and this.
9833 sched_free_group(tg);
9837 * This is called before wake_up_new_task(), therefore we really only
9838 * have to set its group bits, all the other stuff does not apply.
9840 static void cpu_cgroup_fork(struct task_struct *task)
9845 rq = task_rq_lock(task, &rf);
9847 update_rq_clock(rq);
9848 sched_change_group(task, TASK_SET_GROUP);
9850 task_rq_unlock(rq, task, &rf);
9853 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9855 struct task_struct *task;
9856 struct cgroup_subsys_state *css;
9859 cgroup_taskset_for_each(task, css, tset) {
9860 #ifdef CONFIG_RT_GROUP_SCHED
9861 if (!sched_rt_can_attach(css_tg(css), task))
9865 * Serialize against wake_up_new_task() such that if it's
9866 * running, we're sure to observe its full state.
9868 raw_spin_lock_irq(&task->pi_lock);
9870 * Avoid calling sched_move_task() before wake_up_new_task()
9871 * has happened. This would lead to problems with PELT, due to
9872 * move wanting to detach+attach while we're not attached yet.
9874 if (READ_ONCE(task->__state) == TASK_NEW)
9876 raw_spin_unlock_irq(&task->pi_lock);
9884 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9886 struct task_struct *task;
9887 struct cgroup_subsys_state *css;
9889 cgroup_taskset_for_each(task, css, tset)
9890 sched_move_task(task);
9893 #ifdef CONFIG_UCLAMP_TASK_GROUP
9894 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9896 struct cgroup_subsys_state *top_css = css;
9897 struct uclamp_se *uc_parent = NULL;
9898 struct uclamp_se *uc_se = NULL;
9899 unsigned int eff[UCLAMP_CNT];
9900 enum uclamp_id clamp_id;
9901 unsigned int clamps;
9903 lockdep_assert_held(&uclamp_mutex);
9904 SCHED_WARN_ON(!rcu_read_lock_held());
9906 css_for_each_descendant_pre(css, top_css) {
9907 uc_parent = css_tg(css)->parent
9908 ? css_tg(css)->parent->uclamp : NULL;
9910 for_each_clamp_id(clamp_id) {
9911 /* Assume effective clamps matches requested clamps */
9912 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
9913 /* Cap effective clamps with parent's effective clamps */
9915 eff[clamp_id] > uc_parent[clamp_id].value) {
9916 eff[clamp_id] = uc_parent[clamp_id].value;
9919 /* Ensure protection is always capped by limit */
9920 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
9922 /* Propagate most restrictive effective clamps */
9924 uc_se = css_tg(css)->uclamp;
9925 for_each_clamp_id(clamp_id) {
9926 if (eff[clamp_id] == uc_se[clamp_id].value)
9928 uc_se[clamp_id].value = eff[clamp_id];
9929 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
9930 clamps |= (0x1 << clamp_id);
9933 css = css_rightmost_descendant(css);
9937 /* Immediately update descendants RUNNABLE tasks */
9938 uclamp_update_active_tasks(css);
9943 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
9944 * C expression. Since there is no way to convert a macro argument (N) into a
9945 * character constant, use two levels of macros.
9947 #define _POW10(exp) ((unsigned int)1e##exp)
9948 #define POW10(exp) _POW10(exp)
9950 struct uclamp_request {
9951 #define UCLAMP_PERCENT_SHIFT 2
9952 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
9958 static inline struct uclamp_request
9959 capacity_from_percent(char *buf)
9961 struct uclamp_request req = {
9962 .percent = UCLAMP_PERCENT_SCALE,
9963 .util = SCHED_CAPACITY_SCALE,
9968 if (strcmp(buf, "max")) {
9969 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
9973 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
9978 req.util = req.percent << SCHED_CAPACITY_SHIFT;
9979 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
9985 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
9986 size_t nbytes, loff_t off,
9987 enum uclamp_id clamp_id)
9989 struct uclamp_request req;
9990 struct task_group *tg;
9992 req = capacity_from_percent(buf);
9996 static_branch_enable(&sched_uclamp_used);
9998 mutex_lock(&uclamp_mutex);
10001 tg = css_tg(of_css(of));
10002 if (tg->uclamp_req[clamp_id].value != req.util)
10003 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10006 * Because of not recoverable conversion rounding we keep track of the
10007 * exact requested value
10009 tg->uclamp_pct[clamp_id] = req.percent;
10011 /* Update effective clamps to track the most restrictive value */
10012 cpu_util_update_eff(of_css(of));
10015 mutex_unlock(&uclamp_mutex);
10020 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10021 char *buf, size_t nbytes,
10024 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10027 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10028 char *buf, size_t nbytes,
10031 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10034 static inline void cpu_uclamp_print(struct seq_file *sf,
10035 enum uclamp_id clamp_id)
10037 struct task_group *tg;
10043 tg = css_tg(seq_css(sf));
10044 util_clamp = tg->uclamp_req[clamp_id].value;
10047 if (util_clamp == SCHED_CAPACITY_SCALE) {
10048 seq_puts(sf, "max\n");
10052 percent = tg->uclamp_pct[clamp_id];
10053 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10054 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10057 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10059 cpu_uclamp_print(sf, UCLAMP_MIN);
10063 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10065 cpu_uclamp_print(sf, UCLAMP_MAX);
10068 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10070 #ifdef CONFIG_FAIR_GROUP_SCHED
10071 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10072 struct cftype *cftype, u64 shareval)
10074 if (shareval > scale_load_down(ULONG_MAX))
10075 shareval = MAX_SHARES;
10076 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10079 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10080 struct cftype *cft)
10082 struct task_group *tg = css_tg(css);
10084 return (u64) scale_load_down(tg->shares);
10087 #ifdef CONFIG_CFS_BANDWIDTH
10088 static DEFINE_MUTEX(cfs_constraints_mutex);
10090 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10091 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10092 /* More than 203 days if BW_SHIFT equals 20. */
10093 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10095 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10097 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10100 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10101 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10103 if (tg == &root_task_group)
10107 * Ensure we have at some amount of bandwidth every period. This is
10108 * to prevent reaching a state of large arrears when throttled via
10109 * entity_tick() resulting in prolonged exit starvation.
10111 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10115 * Likewise, bound things on the other side by preventing insane quota
10116 * periods. This also allows us to normalize in computing quota
10119 if (period > max_cfs_quota_period)
10123 * Bound quota to defend quota against overflow during bandwidth shift.
10125 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10128 if (quota != RUNTIME_INF && (burst > quota ||
10129 burst + quota > max_cfs_runtime))
10133 * Prevent race between setting of cfs_rq->runtime_enabled and
10134 * unthrottle_offline_cfs_rqs().
10137 mutex_lock(&cfs_constraints_mutex);
10138 ret = __cfs_schedulable(tg, period, quota);
10142 runtime_enabled = quota != RUNTIME_INF;
10143 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10145 * If we need to toggle cfs_bandwidth_used, off->on must occur
10146 * before making related changes, and on->off must occur afterwards
10148 if (runtime_enabled && !runtime_was_enabled)
10149 cfs_bandwidth_usage_inc();
10150 raw_spin_lock_irq(&cfs_b->lock);
10151 cfs_b->period = ns_to_ktime(period);
10152 cfs_b->quota = quota;
10153 cfs_b->burst = burst;
10155 __refill_cfs_bandwidth_runtime(cfs_b);
10157 /* Restart the period timer (if active) to handle new period expiry: */
10158 if (runtime_enabled)
10159 start_cfs_bandwidth(cfs_b);
10161 raw_spin_unlock_irq(&cfs_b->lock);
10163 for_each_online_cpu(i) {
10164 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10165 struct rq *rq = cfs_rq->rq;
10166 struct rq_flags rf;
10168 rq_lock_irq(rq, &rf);
10169 cfs_rq->runtime_enabled = runtime_enabled;
10170 cfs_rq->runtime_remaining = 0;
10172 if (cfs_rq->throttled)
10173 unthrottle_cfs_rq(cfs_rq);
10174 rq_unlock_irq(rq, &rf);
10176 if (runtime_was_enabled && !runtime_enabled)
10177 cfs_bandwidth_usage_dec();
10179 mutex_unlock(&cfs_constraints_mutex);
10180 cpus_read_unlock();
10185 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10187 u64 quota, period, burst;
10189 period = ktime_to_ns(tg->cfs_bandwidth.period);
10190 burst = tg->cfs_bandwidth.burst;
10191 if (cfs_quota_us < 0)
10192 quota = RUNTIME_INF;
10193 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10194 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10198 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10201 static long tg_get_cfs_quota(struct task_group *tg)
10205 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10208 quota_us = tg->cfs_bandwidth.quota;
10209 do_div(quota_us, NSEC_PER_USEC);
10214 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10216 u64 quota, period, burst;
10218 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10221 period = (u64)cfs_period_us * NSEC_PER_USEC;
10222 quota = tg->cfs_bandwidth.quota;
10223 burst = tg->cfs_bandwidth.burst;
10225 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10228 static long tg_get_cfs_period(struct task_group *tg)
10232 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10233 do_div(cfs_period_us, NSEC_PER_USEC);
10235 return cfs_period_us;
10238 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10240 u64 quota, period, burst;
10242 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10245 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10246 period = ktime_to_ns(tg->cfs_bandwidth.period);
10247 quota = tg->cfs_bandwidth.quota;
10249 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10252 static long tg_get_cfs_burst(struct task_group *tg)
10256 burst_us = tg->cfs_bandwidth.burst;
10257 do_div(burst_us, NSEC_PER_USEC);
10262 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10263 struct cftype *cft)
10265 return tg_get_cfs_quota(css_tg(css));
10268 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10269 struct cftype *cftype, s64 cfs_quota_us)
10271 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10274 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10275 struct cftype *cft)
10277 return tg_get_cfs_period(css_tg(css));
10280 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10281 struct cftype *cftype, u64 cfs_period_us)
10283 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10286 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10287 struct cftype *cft)
10289 return tg_get_cfs_burst(css_tg(css));
10292 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10293 struct cftype *cftype, u64 cfs_burst_us)
10295 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10298 struct cfs_schedulable_data {
10299 struct task_group *tg;
10304 * normalize group quota/period to be quota/max_period
10305 * note: units are usecs
10307 static u64 normalize_cfs_quota(struct task_group *tg,
10308 struct cfs_schedulable_data *d)
10313 period = d->period;
10316 period = tg_get_cfs_period(tg);
10317 quota = tg_get_cfs_quota(tg);
10320 /* note: these should typically be equivalent */
10321 if (quota == RUNTIME_INF || quota == -1)
10322 return RUNTIME_INF;
10324 return to_ratio(period, quota);
10327 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10329 struct cfs_schedulable_data *d = data;
10330 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10331 s64 quota = 0, parent_quota = -1;
10334 quota = RUNTIME_INF;
10336 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10338 quota = normalize_cfs_quota(tg, d);
10339 parent_quota = parent_b->hierarchical_quota;
10342 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10343 * always take the min. On cgroup1, only inherit when no
10346 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10347 quota = min(quota, parent_quota);
10349 if (quota == RUNTIME_INF)
10350 quota = parent_quota;
10351 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10355 cfs_b->hierarchical_quota = quota;
10360 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10363 struct cfs_schedulable_data data = {
10369 if (quota != RUNTIME_INF) {
10370 do_div(data.period, NSEC_PER_USEC);
10371 do_div(data.quota, NSEC_PER_USEC);
10375 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10381 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10383 struct task_group *tg = css_tg(seq_css(sf));
10384 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10386 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10387 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10388 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10390 if (schedstat_enabled() && tg != &root_task_group) {
10394 for_each_possible_cpu(i)
10395 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10397 seq_printf(sf, "wait_sum %llu\n", ws);
10402 #endif /* CONFIG_CFS_BANDWIDTH */
10403 #endif /* CONFIG_FAIR_GROUP_SCHED */
10405 #ifdef CONFIG_RT_GROUP_SCHED
10406 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10407 struct cftype *cft, s64 val)
10409 return sched_group_set_rt_runtime(css_tg(css), val);
10412 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10413 struct cftype *cft)
10415 return sched_group_rt_runtime(css_tg(css));
10418 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10419 struct cftype *cftype, u64 rt_period_us)
10421 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10424 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10425 struct cftype *cft)
10427 return sched_group_rt_period(css_tg(css));
10429 #endif /* CONFIG_RT_GROUP_SCHED */
10431 #ifdef CONFIG_FAIR_GROUP_SCHED
10432 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10433 struct cftype *cft)
10435 return css_tg(css)->idle;
10438 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10439 struct cftype *cft, s64 idle)
10441 return sched_group_set_idle(css_tg(css), idle);
10445 static struct cftype cpu_legacy_files[] = {
10446 #ifdef CONFIG_FAIR_GROUP_SCHED
10449 .read_u64 = cpu_shares_read_u64,
10450 .write_u64 = cpu_shares_write_u64,
10454 .read_s64 = cpu_idle_read_s64,
10455 .write_s64 = cpu_idle_write_s64,
10458 #ifdef CONFIG_CFS_BANDWIDTH
10460 .name = "cfs_quota_us",
10461 .read_s64 = cpu_cfs_quota_read_s64,
10462 .write_s64 = cpu_cfs_quota_write_s64,
10465 .name = "cfs_period_us",
10466 .read_u64 = cpu_cfs_period_read_u64,
10467 .write_u64 = cpu_cfs_period_write_u64,
10470 .name = "cfs_burst_us",
10471 .read_u64 = cpu_cfs_burst_read_u64,
10472 .write_u64 = cpu_cfs_burst_write_u64,
10476 .seq_show = cpu_cfs_stat_show,
10479 #ifdef CONFIG_RT_GROUP_SCHED
10481 .name = "rt_runtime_us",
10482 .read_s64 = cpu_rt_runtime_read,
10483 .write_s64 = cpu_rt_runtime_write,
10486 .name = "rt_period_us",
10487 .read_u64 = cpu_rt_period_read_uint,
10488 .write_u64 = cpu_rt_period_write_uint,
10491 #ifdef CONFIG_UCLAMP_TASK_GROUP
10493 .name = "uclamp.min",
10494 .flags = CFTYPE_NOT_ON_ROOT,
10495 .seq_show = cpu_uclamp_min_show,
10496 .write = cpu_uclamp_min_write,
10499 .name = "uclamp.max",
10500 .flags = CFTYPE_NOT_ON_ROOT,
10501 .seq_show = cpu_uclamp_max_show,
10502 .write = cpu_uclamp_max_write,
10505 { } /* Terminate */
10508 static int cpu_extra_stat_show(struct seq_file *sf,
10509 struct cgroup_subsys_state *css)
10511 #ifdef CONFIG_CFS_BANDWIDTH
10513 struct task_group *tg = css_tg(css);
10514 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10515 u64 throttled_usec;
10517 throttled_usec = cfs_b->throttled_time;
10518 do_div(throttled_usec, NSEC_PER_USEC);
10520 seq_printf(sf, "nr_periods %d\n"
10521 "nr_throttled %d\n"
10522 "throttled_usec %llu\n",
10523 cfs_b->nr_periods, cfs_b->nr_throttled,
10530 #ifdef CONFIG_FAIR_GROUP_SCHED
10531 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10532 struct cftype *cft)
10534 struct task_group *tg = css_tg(css);
10535 u64 weight = scale_load_down(tg->shares);
10537 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10540 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10541 struct cftype *cft, u64 weight)
10544 * cgroup weight knobs should use the common MIN, DFL and MAX
10545 * values which are 1, 100 and 10000 respectively. While it loses
10546 * a bit of range on both ends, it maps pretty well onto the shares
10547 * value used by scheduler and the round-trip conversions preserve
10548 * the original value over the entire range.
10550 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10553 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10555 return sched_group_set_shares(css_tg(css), scale_load(weight));
10558 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10559 struct cftype *cft)
10561 unsigned long weight = scale_load_down(css_tg(css)->shares);
10562 int last_delta = INT_MAX;
10565 /* find the closest nice value to the current weight */
10566 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10567 delta = abs(sched_prio_to_weight[prio] - weight);
10568 if (delta >= last_delta)
10570 last_delta = delta;
10573 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10576 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10577 struct cftype *cft, s64 nice)
10579 unsigned long weight;
10582 if (nice < MIN_NICE || nice > MAX_NICE)
10585 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10586 idx = array_index_nospec(idx, 40);
10587 weight = sched_prio_to_weight[idx];
10589 return sched_group_set_shares(css_tg(css), scale_load(weight));
10593 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10594 long period, long quota)
10597 seq_puts(sf, "max");
10599 seq_printf(sf, "%ld", quota);
10601 seq_printf(sf, " %ld\n", period);
10604 /* caller should put the current value in *@periodp before calling */
10605 static int __maybe_unused cpu_period_quota_parse(char *buf,
10606 u64 *periodp, u64 *quotap)
10608 char tok[21]; /* U64_MAX */
10610 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10613 *periodp *= NSEC_PER_USEC;
10615 if (sscanf(tok, "%llu", quotap))
10616 *quotap *= NSEC_PER_USEC;
10617 else if (!strcmp(tok, "max"))
10618 *quotap = RUNTIME_INF;
10625 #ifdef CONFIG_CFS_BANDWIDTH
10626 static int cpu_max_show(struct seq_file *sf, void *v)
10628 struct task_group *tg = css_tg(seq_css(sf));
10630 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10634 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10635 char *buf, size_t nbytes, loff_t off)
10637 struct task_group *tg = css_tg(of_css(of));
10638 u64 period = tg_get_cfs_period(tg);
10639 u64 burst = tg_get_cfs_burst(tg);
10643 ret = cpu_period_quota_parse(buf, &period, "a);
10645 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10646 return ret ?: nbytes;
10650 static struct cftype cpu_files[] = {
10651 #ifdef CONFIG_FAIR_GROUP_SCHED
10654 .flags = CFTYPE_NOT_ON_ROOT,
10655 .read_u64 = cpu_weight_read_u64,
10656 .write_u64 = cpu_weight_write_u64,
10659 .name = "weight.nice",
10660 .flags = CFTYPE_NOT_ON_ROOT,
10661 .read_s64 = cpu_weight_nice_read_s64,
10662 .write_s64 = cpu_weight_nice_write_s64,
10666 .flags = CFTYPE_NOT_ON_ROOT,
10667 .read_s64 = cpu_idle_read_s64,
10668 .write_s64 = cpu_idle_write_s64,
10671 #ifdef CONFIG_CFS_BANDWIDTH
10674 .flags = CFTYPE_NOT_ON_ROOT,
10675 .seq_show = cpu_max_show,
10676 .write = cpu_max_write,
10679 .name = "max.burst",
10680 .flags = CFTYPE_NOT_ON_ROOT,
10681 .read_u64 = cpu_cfs_burst_read_u64,
10682 .write_u64 = cpu_cfs_burst_write_u64,
10685 #ifdef CONFIG_UCLAMP_TASK_GROUP
10687 .name = "uclamp.min",
10688 .flags = CFTYPE_NOT_ON_ROOT,
10689 .seq_show = cpu_uclamp_min_show,
10690 .write = cpu_uclamp_min_write,
10693 .name = "uclamp.max",
10694 .flags = CFTYPE_NOT_ON_ROOT,
10695 .seq_show = cpu_uclamp_max_show,
10696 .write = cpu_uclamp_max_write,
10699 { } /* terminate */
10702 struct cgroup_subsys cpu_cgrp_subsys = {
10703 .css_alloc = cpu_cgroup_css_alloc,
10704 .css_online = cpu_cgroup_css_online,
10705 .css_released = cpu_cgroup_css_released,
10706 .css_free = cpu_cgroup_css_free,
10707 .css_extra_stat_show = cpu_extra_stat_show,
10708 .fork = cpu_cgroup_fork,
10709 .can_attach = cpu_cgroup_can_attach,
10710 .attach = cpu_cgroup_attach,
10711 .legacy_cftypes = cpu_legacy_files,
10712 .dfl_cftypes = cpu_files,
10713 .early_init = true,
10717 #endif /* CONFIG_CGROUP_SCHED */
10719 void dump_cpu_task(int cpu)
10721 pr_info("Task dump for CPU %d:\n", cpu);
10722 sched_show_task(cpu_curr(cpu));
10726 * Nice levels are multiplicative, with a gentle 10% change for every
10727 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10728 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10729 * that remained on nice 0.
10731 * The "10% effect" is relative and cumulative: from _any_ nice level,
10732 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10733 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10734 * If a task goes up by ~10% and another task goes down by ~10% then
10735 * the relative distance between them is ~25%.)
10737 const int sched_prio_to_weight[40] = {
10738 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10739 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10740 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10741 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10742 /* 0 */ 1024, 820, 655, 526, 423,
10743 /* 5 */ 335, 272, 215, 172, 137,
10744 /* 10 */ 110, 87, 70, 56, 45,
10745 /* 15 */ 36, 29, 23, 18, 15,
10749 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10751 * In cases where the weight does not change often, we can use the
10752 * precalculated inverse to speed up arithmetics by turning divisions
10753 * into multiplications:
10755 const u32 sched_prio_to_wmult[40] = {
10756 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10757 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10758 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10759 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10760 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10761 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10762 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10763 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10766 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10768 trace_sched_update_nr_running_tp(rq, count);