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>
16 #include <linux/blkdev.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(pelt_thermal_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
44 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
46 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
48 #ifdef CONFIG_SCHED_DEBUG
50 * Debugging: various feature bits
52 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
53 * sysctl_sched_features, defined in sched.h, to allow constants propagation
54 * at compile time and compiler optimization based on features default.
56 #define SCHED_FEAT(name, enabled) \
57 (1UL << __SCHED_FEAT_##name) * enabled |
58 const_debug unsigned int sysctl_sched_features =
64 * Print a warning if need_resched is set for the given duration (if
65 * LATENCY_WARN is enabled).
67 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
70 __read_mostly int sysctl_resched_latency_warn_ms = 100;
71 __read_mostly int sysctl_resched_latency_warn_once = 1;
72 #endif /* CONFIG_SCHED_DEBUG */
75 * Number of tasks to iterate in a single balance run.
76 * Limited because this is done with IRQs disabled.
78 #ifdef CONFIG_PREEMPT_RT
79 const_debug unsigned int sysctl_sched_nr_migrate = 8;
81 const_debug unsigned int sysctl_sched_nr_migrate = 32;
85 * period over which we measure -rt task CPU usage in us.
88 unsigned int sysctl_sched_rt_period = 1000000;
90 __read_mostly int scheduler_running;
92 #ifdef CONFIG_SCHED_CORE
94 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
96 /* kernel prio, less is more */
97 static inline int __task_prio(struct task_struct *p)
99 if (p->sched_class == &stop_sched_class) /* trumps deadline */
102 if (rt_prio(p->prio)) /* includes deadline */
103 return p->prio; /* [-1, 99] */
105 if (p->sched_class == &idle_sched_class)
106 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
108 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
118 /* real prio, less is less */
119 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
122 int pa = __task_prio(a), pb = __task_prio(b);
130 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
131 return !dl_time_before(a->dl.deadline, b->dl.deadline);
133 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
134 return cfs_prio_less(a, b, in_fi);
139 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
141 if (a->core_cookie < b->core_cookie)
144 if (a->core_cookie > b->core_cookie)
147 /* flip prio, so high prio is leftmost */
148 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
154 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
156 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
158 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
161 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
163 const struct task_struct *p = __node_2_sc(node);
164 unsigned long cookie = (unsigned long)key;
166 if (cookie < p->core_cookie)
169 if (cookie > p->core_cookie)
175 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
177 rq->core->core_task_seq++;
182 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
185 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
187 rq->core->core_task_seq++;
189 if (sched_core_enqueued(p)) {
190 rb_erase(&p->core_node, &rq->core_tree);
191 RB_CLEAR_NODE(&p->core_node);
195 * Migrating the last task off the cpu, with the cpu in forced idle
196 * state. Reschedule to create an accounting edge for forced idle,
197 * and re-examine whether the core is still in forced idle state.
199 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
200 rq->core->core_forceidle_count && rq->curr == rq->idle)
205 * Find left-most (aka, highest priority) task matching @cookie.
207 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
209 struct rb_node *node;
211 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
213 * The idle task always matches any cookie!
216 return idle_sched_class.pick_task(rq);
218 return __node_2_sc(node);
221 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
223 struct rb_node *node = &p->core_node;
225 node = rb_next(node);
229 p = container_of(node, struct task_struct, core_node);
230 if (p->core_cookie != cookie)
237 * Magic required such that:
239 * raw_spin_rq_lock(rq);
241 * raw_spin_rq_unlock(rq);
243 * ends up locking and unlocking the _same_ lock, and all CPUs
244 * always agree on what rq has what lock.
246 * XXX entirely possible to selectively enable cores, don't bother for now.
249 static DEFINE_MUTEX(sched_core_mutex);
250 static atomic_t sched_core_count;
251 static struct cpumask sched_core_mask;
253 static void sched_core_lock(int cpu, unsigned long *flags)
255 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
258 local_irq_save(*flags);
259 for_each_cpu(t, smt_mask)
260 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
263 static void sched_core_unlock(int cpu, unsigned long *flags)
265 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
268 for_each_cpu(t, smt_mask)
269 raw_spin_unlock(&cpu_rq(t)->__lock);
270 local_irq_restore(*flags);
273 static void __sched_core_flip(bool enabled)
281 * Toggle the online cores, one by one.
283 cpumask_copy(&sched_core_mask, cpu_online_mask);
284 for_each_cpu(cpu, &sched_core_mask) {
285 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
287 sched_core_lock(cpu, &flags);
289 for_each_cpu(t, smt_mask)
290 cpu_rq(t)->core_enabled = enabled;
292 cpu_rq(cpu)->core->core_forceidle_start = 0;
294 sched_core_unlock(cpu, &flags);
296 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
300 * Toggle the offline CPUs.
302 cpumask_copy(&sched_core_mask, cpu_possible_mask);
303 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
305 for_each_cpu(cpu, &sched_core_mask)
306 cpu_rq(cpu)->core_enabled = enabled;
311 static void sched_core_assert_empty(void)
315 for_each_possible_cpu(cpu)
316 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
319 static void __sched_core_enable(void)
321 static_branch_enable(&__sched_core_enabled);
323 * Ensure all previous instances of raw_spin_rq_*lock() have finished
324 * and future ones will observe !sched_core_disabled().
327 __sched_core_flip(true);
328 sched_core_assert_empty();
331 static void __sched_core_disable(void)
333 sched_core_assert_empty();
334 __sched_core_flip(false);
335 static_branch_disable(&__sched_core_enabled);
338 void sched_core_get(void)
340 if (atomic_inc_not_zero(&sched_core_count))
343 mutex_lock(&sched_core_mutex);
344 if (!atomic_read(&sched_core_count))
345 __sched_core_enable();
347 smp_mb__before_atomic();
348 atomic_inc(&sched_core_count);
349 mutex_unlock(&sched_core_mutex);
352 static void __sched_core_put(struct work_struct *work)
354 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
355 __sched_core_disable();
356 mutex_unlock(&sched_core_mutex);
360 void sched_core_put(void)
362 static DECLARE_WORK(_work, __sched_core_put);
365 * "There can be only one"
367 * Either this is the last one, or we don't actually need to do any
368 * 'work'. If it is the last *again*, we rely on
369 * WORK_STRUCT_PENDING_BIT.
371 if (!atomic_add_unless(&sched_core_count, -1, 1))
372 schedule_work(&_work);
375 #else /* !CONFIG_SCHED_CORE */
377 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
379 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
381 #endif /* CONFIG_SCHED_CORE */
384 * part of the period that we allow rt tasks to run in us.
387 int sysctl_sched_rt_runtime = 950000;
391 * Serialization rules:
397 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
400 * rq2->lock where: rq1 < rq2
404 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
405 * local CPU's rq->lock, it optionally removes the task from the runqueue and
406 * always looks at the local rq data structures to find the most eligible task
409 * Task enqueue is also under rq->lock, possibly taken from another CPU.
410 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
411 * the local CPU to avoid bouncing the runqueue state around [ see
412 * ttwu_queue_wakelist() ]
414 * Task wakeup, specifically wakeups that involve migration, are horribly
415 * complicated to avoid having to take two rq->locks.
419 * System-calls and anything external will use task_rq_lock() which acquires
420 * both p->pi_lock and rq->lock. As a consequence the state they change is
421 * stable while holding either lock:
423 * - sched_setaffinity()/
424 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
425 * - set_user_nice(): p->se.load, p->*prio
426 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
427 * p->se.load, p->rt_priority,
428 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
429 * - sched_setnuma(): p->numa_preferred_nid
430 * - sched_move_task()/
431 * cpu_cgroup_fork(): p->sched_task_group
432 * - uclamp_update_active() p->uclamp*
434 * p->state <- TASK_*:
436 * is changed locklessly using set_current_state(), __set_current_state() or
437 * set_special_state(), see their respective comments, or by
438 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
441 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
443 * is set by activate_task() and cleared by deactivate_task(), under
444 * rq->lock. Non-zero indicates the task is runnable, the special
445 * ON_RQ_MIGRATING state is used for migration without holding both
446 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
448 * p->on_cpu <- { 0, 1 }:
450 * is set by prepare_task() and cleared by finish_task() such that it will be
451 * set before p is scheduled-in and cleared after p is scheduled-out, both
452 * under rq->lock. Non-zero indicates the task is running on its CPU.
454 * [ The astute reader will observe that it is possible for two tasks on one
455 * CPU to have ->on_cpu = 1 at the same time. ]
457 * task_cpu(p): is changed by set_task_cpu(), the rules are:
459 * - Don't call set_task_cpu() on a blocked task:
461 * We don't care what CPU we're not running on, this simplifies hotplug,
462 * the CPU assignment of blocked tasks isn't required to be valid.
464 * - for try_to_wake_up(), called under p->pi_lock:
466 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
468 * - for migration called under rq->lock:
469 * [ see task_on_rq_migrating() in task_rq_lock() ]
471 * o move_queued_task()
474 * - for migration called under double_rq_lock():
476 * o __migrate_swap_task()
477 * o push_rt_task() / pull_rt_task()
478 * o push_dl_task() / pull_dl_task()
479 * o dl_task_offline_migration()
483 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
485 raw_spinlock_t *lock;
487 /* Matches synchronize_rcu() in __sched_core_enable() */
489 if (sched_core_disabled()) {
490 raw_spin_lock_nested(&rq->__lock, subclass);
491 /* preempt_count *MUST* be > 1 */
492 preempt_enable_no_resched();
497 lock = __rq_lockp(rq);
498 raw_spin_lock_nested(lock, subclass);
499 if (likely(lock == __rq_lockp(rq))) {
500 /* preempt_count *MUST* be > 1 */
501 preempt_enable_no_resched();
504 raw_spin_unlock(lock);
508 bool raw_spin_rq_trylock(struct rq *rq)
510 raw_spinlock_t *lock;
513 /* Matches synchronize_rcu() in __sched_core_enable() */
515 if (sched_core_disabled()) {
516 ret = raw_spin_trylock(&rq->__lock);
522 lock = __rq_lockp(rq);
523 ret = raw_spin_trylock(lock);
524 if (!ret || (likely(lock == __rq_lockp(rq)))) {
528 raw_spin_unlock(lock);
532 void raw_spin_rq_unlock(struct rq *rq)
534 raw_spin_unlock(rq_lockp(rq));
539 * double_rq_lock - safely lock two runqueues
541 void double_rq_lock(struct rq *rq1, struct rq *rq2)
543 lockdep_assert_irqs_disabled();
545 if (rq_order_less(rq2, rq1))
548 raw_spin_rq_lock(rq1);
549 if (__rq_lockp(rq1) == __rq_lockp(rq2))
552 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
557 * __task_rq_lock - lock the rq @p resides on.
559 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
564 lockdep_assert_held(&p->pi_lock);
568 raw_spin_rq_lock(rq);
569 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
573 raw_spin_rq_unlock(rq);
575 while (unlikely(task_on_rq_migrating(p)))
581 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
583 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
584 __acquires(p->pi_lock)
590 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
592 raw_spin_rq_lock(rq);
594 * move_queued_task() task_rq_lock()
597 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
598 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
599 * [S] ->cpu = new_cpu [L] task_rq()
603 * If we observe the old CPU in task_rq_lock(), the acquire of
604 * the old rq->lock will fully serialize against the stores.
606 * If we observe the new CPU in task_rq_lock(), the address
607 * dependency headed by '[L] rq = task_rq()' and the acquire
608 * will pair with the WMB to ensure we then also see migrating.
610 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
614 raw_spin_rq_unlock(rq);
615 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
617 while (unlikely(task_on_rq_migrating(p)))
623 * RQ-clock updating methods:
626 static void update_rq_clock_task(struct rq *rq, s64 delta)
629 * In theory, the compile should just see 0 here, and optimize out the call
630 * to sched_rt_avg_update. But I don't trust it...
632 s64 __maybe_unused steal = 0, irq_delta = 0;
634 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
635 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
638 * Since irq_time is only updated on {soft,}irq_exit, we might run into
639 * this case when a previous update_rq_clock() happened inside a
642 * When this happens, we stop ->clock_task and only update the
643 * prev_irq_time stamp to account for the part that fit, so that a next
644 * update will consume the rest. This ensures ->clock_task is
647 * It does however cause some slight miss-attribution of {soft,}irq
648 * time, a more accurate solution would be to update the irq_time using
649 * the current rq->clock timestamp, except that would require using
652 if (irq_delta > delta)
655 rq->prev_irq_time += irq_delta;
658 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
659 if (static_key_false((¶virt_steal_rq_enabled))) {
660 steal = paravirt_steal_clock(cpu_of(rq));
661 steal -= rq->prev_steal_time_rq;
663 if (unlikely(steal > delta))
666 rq->prev_steal_time_rq += steal;
671 rq->clock_task += delta;
673 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
674 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
675 update_irq_load_avg(rq, irq_delta + steal);
677 update_rq_clock_pelt(rq, delta);
680 void update_rq_clock(struct rq *rq)
684 lockdep_assert_rq_held(rq);
686 if (rq->clock_update_flags & RQCF_ACT_SKIP)
689 #ifdef CONFIG_SCHED_DEBUG
690 if (sched_feat(WARN_DOUBLE_CLOCK))
691 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
692 rq->clock_update_flags |= RQCF_UPDATED;
695 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
699 update_rq_clock_task(rq, delta);
702 #ifdef CONFIG_SCHED_HRTICK
704 * Use HR-timers to deliver accurate preemption points.
707 static void hrtick_clear(struct rq *rq)
709 if (hrtimer_active(&rq->hrtick_timer))
710 hrtimer_cancel(&rq->hrtick_timer);
714 * High-resolution timer tick.
715 * Runs from hardirq context with interrupts disabled.
717 static enum hrtimer_restart hrtick(struct hrtimer *timer)
719 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
722 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
726 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
729 return HRTIMER_NORESTART;
734 static void __hrtick_restart(struct rq *rq)
736 struct hrtimer *timer = &rq->hrtick_timer;
737 ktime_t time = rq->hrtick_time;
739 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
743 * called from hardirq (IPI) context
745 static void __hrtick_start(void *arg)
751 __hrtick_restart(rq);
756 * Called to set the hrtick timer state.
758 * called with rq->lock held and irqs disabled
760 void hrtick_start(struct rq *rq, u64 delay)
762 struct hrtimer *timer = &rq->hrtick_timer;
766 * Don't schedule slices shorter than 10000ns, that just
767 * doesn't make sense and can cause timer DoS.
769 delta = max_t(s64, delay, 10000LL);
770 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
773 __hrtick_restart(rq);
775 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
780 * Called to set the hrtick timer state.
782 * called with rq->lock held and irqs disabled
784 void hrtick_start(struct rq *rq, u64 delay)
787 * Don't schedule slices shorter than 10000ns, that just
788 * doesn't make sense. Rely on vruntime for fairness.
790 delay = max_t(u64, delay, 10000LL);
791 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
792 HRTIMER_MODE_REL_PINNED_HARD);
795 #endif /* CONFIG_SMP */
797 static void hrtick_rq_init(struct rq *rq)
800 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
802 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
803 rq->hrtick_timer.function = hrtick;
805 #else /* CONFIG_SCHED_HRTICK */
806 static inline void hrtick_clear(struct rq *rq)
810 static inline void hrtick_rq_init(struct rq *rq)
813 #endif /* CONFIG_SCHED_HRTICK */
816 * cmpxchg based fetch_or, macro so it works for different integer types
818 #define fetch_or(ptr, mask) \
820 typeof(ptr) _ptr = (ptr); \
821 typeof(mask) _mask = (mask); \
822 typeof(*_ptr) _old, _val = *_ptr; \
825 _old = cmpxchg(_ptr, _val, _val | _mask); \
833 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
835 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
836 * this avoids any races wrt polling state changes and thereby avoids
839 static bool set_nr_and_not_polling(struct task_struct *p)
841 struct thread_info *ti = task_thread_info(p);
842 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
846 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
848 * If this returns true, then the idle task promises to call
849 * sched_ttwu_pending() and reschedule soon.
851 static bool set_nr_if_polling(struct task_struct *p)
853 struct thread_info *ti = task_thread_info(p);
854 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
857 if (!(val & _TIF_POLLING_NRFLAG))
859 if (val & _TIF_NEED_RESCHED)
861 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
870 static bool set_nr_and_not_polling(struct task_struct *p)
872 set_tsk_need_resched(p);
877 static bool set_nr_if_polling(struct task_struct *p)
884 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
886 struct wake_q_node *node = &task->wake_q;
889 * Atomically grab the task, if ->wake_q is !nil already it means
890 * it's already queued (either by us or someone else) and will get the
891 * wakeup due to that.
893 * In order to ensure that a pending wakeup will observe our pending
894 * state, even in the failed case, an explicit smp_mb() must be used.
896 smp_mb__before_atomic();
897 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
901 * The head is context local, there can be no concurrency.
904 head->lastp = &node->next;
909 * wake_q_add() - queue a wakeup for 'later' waking.
910 * @head: the wake_q_head to add @task to
911 * @task: the task to queue for 'later' wakeup
913 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
914 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
917 * This function must be used as-if it were wake_up_process(); IOW the task
918 * must be ready to be woken at this location.
920 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
922 if (__wake_q_add(head, task))
923 get_task_struct(task);
927 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
928 * @head: the wake_q_head to add @task to
929 * @task: the task to queue for 'later' wakeup
931 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
932 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
935 * This function must be used as-if it were wake_up_process(); IOW the task
936 * must be ready to be woken at this location.
938 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
939 * that already hold reference to @task can call the 'safe' version and trust
940 * wake_q to do the right thing depending whether or not the @task is already
943 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
945 if (!__wake_q_add(head, task))
946 put_task_struct(task);
949 void wake_up_q(struct wake_q_head *head)
951 struct wake_q_node *node = head->first;
953 while (node != WAKE_Q_TAIL) {
954 struct task_struct *task;
956 task = container_of(node, struct task_struct, wake_q);
957 /* Task can safely be re-inserted now: */
959 task->wake_q.next = NULL;
962 * wake_up_process() executes a full barrier, which pairs with
963 * the queueing in wake_q_add() so as not to miss wakeups.
965 wake_up_process(task);
966 put_task_struct(task);
971 * resched_curr - mark rq's current task 'to be rescheduled now'.
973 * On UP this means the setting of the need_resched flag, on SMP it
974 * might also involve a cross-CPU call to trigger the scheduler on
977 void resched_curr(struct rq *rq)
979 struct task_struct *curr = rq->curr;
982 lockdep_assert_rq_held(rq);
984 if (test_tsk_need_resched(curr))
989 if (cpu == smp_processor_id()) {
990 set_tsk_need_resched(curr);
991 set_preempt_need_resched();
995 if (set_nr_and_not_polling(curr))
996 smp_send_reschedule(cpu);
998 trace_sched_wake_idle_without_ipi(cpu);
1001 void resched_cpu(int cpu)
1003 struct rq *rq = cpu_rq(cpu);
1004 unsigned long flags;
1006 raw_spin_rq_lock_irqsave(rq, flags);
1007 if (cpu_online(cpu) || cpu == smp_processor_id())
1009 raw_spin_rq_unlock_irqrestore(rq, flags);
1013 #ifdef CONFIG_NO_HZ_COMMON
1015 * In the semi idle case, use the nearest busy CPU for migrating timers
1016 * from an idle CPU. This is good for power-savings.
1018 * We don't do similar optimization for completely idle system, as
1019 * selecting an idle CPU will add more delays to the timers than intended
1020 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1022 int get_nohz_timer_target(void)
1024 int i, cpu = smp_processor_id(), default_cpu = -1;
1025 struct sched_domain *sd;
1026 const struct cpumask *hk_mask;
1028 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1034 hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1037 for_each_domain(cpu, sd) {
1038 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1049 if (default_cpu == -1)
1050 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1058 * When add_timer_on() enqueues a timer into the timer wheel of an
1059 * idle CPU then this timer might expire before the next timer event
1060 * which is scheduled to wake up that CPU. In case of a completely
1061 * idle system the next event might even be infinite time into the
1062 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1063 * leaves the inner idle loop so the newly added timer is taken into
1064 * account when the CPU goes back to idle and evaluates the timer
1065 * wheel for the next timer event.
1067 static void wake_up_idle_cpu(int cpu)
1069 struct rq *rq = cpu_rq(cpu);
1071 if (cpu == smp_processor_id())
1074 if (set_nr_and_not_polling(rq->idle))
1075 smp_send_reschedule(cpu);
1077 trace_sched_wake_idle_without_ipi(cpu);
1080 static bool wake_up_full_nohz_cpu(int cpu)
1083 * We just need the target to call irq_exit() and re-evaluate
1084 * the next tick. The nohz full kick at least implies that.
1085 * If needed we can still optimize that later with an
1088 if (cpu_is_offline(cpu))
1089 return true; /* Don't try to wake offline CPUs. */
1090 if (tick_nohz_full_cpu(cpu)) {
1091 if (cpu != smp_processor_id() ||
1092 tick_nohz_tick_stopped())
1093 tick_nohz_full_kick_cpu(cpu);
1101 * Wake up the specified CPU. If the CPU is going offline, it is the
1102 * caller's responsibility to deal with the lost wakeup, for example,
1103 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1105 void wake_up_nohz_cpu(int cpu)
1107 if (!wake_up_full_nohz_cpu(cpu))
1108 wake_up_idle_cpu(cpu);
1111 static void nohz_csd_func(void *info)
1113 struct rq *rq = info;
1114 int cpu = cpu_of(rq);
1118 * Release the rq::nohz_csd.
1120 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1121 WARN_ON(!(flags & NOHZ_KICK_MASK));
1123 rq->idle_balance = idle_cpu(cpu);
1124 if (rq->idle_balance && !need_resched()) {
1125 rq->nohz_idle_balance = flags;
1126 raise_softirq_irqoff(SCHED_SOFTIRQ);
1130 #endif /* CONFIG_NO_HZ_COMMON */
1132 #ifdef CONFIG_NO_HZ_FULL
1133 bool sched_can_stop_tick(struct rq *rq)
1135 int fifo_nr_running;
1137 /* Deadline tasks, even if single, need the tick */
1138 if (rq->dl.dl_nr_running)
1142 * If there are more than one RR tasks, we need the tick to affect the
1143 * actual RR behaviour.
1145 if (rq->rt.rr_nr_running) {
1146 if (rq->rt.rr_nr_running == 1)
1153 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1154 * forced preemption between FIFO tasks.
1156 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1157 if (fifo_nr_running)
1161 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1162 * if there's more than one we need the tick for involuntary
1165 if (rq->nr_running > 1)
1170 #endif /* CONFIG_NO_HZ_FULL */
1171 #endif /* CONFIG_SMP */
1173 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1174 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1176 * Iterate task_group tree rooted at *from, calling @down when first entering a
1177 * node and @up when leaving it for the final time.
1179 * Caller must hold rcu_lock or sufficient equivalent.
1181 int walk_tg_tree_from(struct task_group *from,
1182 tg_visitor down, tg_visitor up, void *data)
1184 struct task_group *parent, *child;
1190 ret = (*down)(parent, data);
1193 list_for_each_entry_rcu(child, &parent->children, siblings) {
1200 ret = (*up)(parent, data);
1201 if (ret || parent == from)
1205 parent = parent->parent;
1212 int tg_nop(struct task_group *tg, void *data)
1218 static void set_load_weight(struct task_struct *p, bool update_load)
1220 int prio = p->static_prio - MAX_RT_PRIO;
1221 struct load_weight *load = &p->se.load;
1224 * SCHED_IDLE tasks get minimal weight:
1226 if (task_has_idle_policy(p)) {
1227 load->weight = scale_load(WEIGHT_IDLEPRIO);
1228 load->inv_weight = WMULT_IDLEPRIO;
1233 * SCHED_OTHER tasks have to update their load when changing their
1236 if (update_load && p->sched_class == &fair_sched_class) {
1237 reweight_task(p, prio);
1239 load->weight = scale_load(sched_prio_to_weight[prio]);
1240 load->inv_weight = sched_prio_to_wmult[prio];
1244 #ifdef CONFIG_UCLAMP_TASK
1246 * Serializes updates of utilization clamp values
1248 * The (slow-path) user-space triggers utilization clamp value updates which
1249 * can require updates on (fast-path) scheduler's data structures used to
1250 * support enqueue/dequeue operations.
1251 * While the per-CPU rq lock protects fast-path update operations, user-space
1252 * requests are serialized using a mutex to reduce the risk of conflicting
1253 * updates or API abuses.
1255 static DEFINE_MUTEX(uclamp_mutex);
1257 /* Max allowed minimum utilization */
1258 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1260 /* Max allowed maximum utilization */
1261 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1264 * By default RT tasks run at the maximum performance point/capacity of the
1265 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1266 * SCHED_CAPACITY_SCALE.
1268 * This knob allows admins to change the default behavior when uclamp is being
1269 * used. In battery powered devices, particularly, running at the maximum
1270 * capacity and frequency will increase energy consumption and shorten the
1273 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1275 * This knob will not override the system default sched_util_clamp_min defined
1278 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1280 /* All clamps are required to be less or equal than these values */
1281 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1284 * This static key is used to reduce the uclamp overhead in the fast path. It
1285 * primarily disables the call to uclamp_rq_{inc, dec}() in
1286 * enqueue/dequeue_task().
1288 * This allows users to continue to enable uclamp in their kernel config with
1289 * minimum uclamp overhead in the fast path.
1291 * As soon as userspace modifies any of the uclamp knobs, the static key is
1292 * enabled, since we have an actual users that make use of uclamp
1295 * The knobs that would enable this static key are:
1297 * * A task modifying its uclamp value with sched_setattr().
1298 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1299 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1301 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1303 /* Integer rounded range for each bucket */
1304 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1306 #define for_each_clamp_id(clamp_id) \
1307 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1309 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1311 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1314 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1316 if (clamp_id == UCLAMP_MIN)
1318 return SCHED_CAPACITY_SCALE;
1321 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1322 unsigned int value, bool user_defined)
1324 uc_se->value = value;
1325 uc_se->bucket_id = uclamp_bucket_id(value);
1326 uc_se->user_defined = user_defined;
1329 static inline unsigned int
1330 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1331 unsigned int clamp_value)
1334 * Avoid blocked utilization pushing up the frequency when we go
1335 * idle (which drops the max-clamp) by retaining the last known
1338 if (clamp_id == UCLAMP_MAX) {
1339 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1343 return uclamp_none(UCLAMP_MIN);
1346 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1347 unsigned int clamp_value)
1349 /* Reset max-clamp retention only on idle exit */
1350 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1353 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1357 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1358 unsigned int clamp_value)
1360 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1361 int bucket_id = UCLAMP_BUCKETS - 1;
1364 * Since both min and max clamps are max aggregated, find the
1365 * top most bucket with tasks in.
1367 for ( ; bucket_id >= 0; bucket_id--) {
1368 if (!bucket[bucket_id].tasks)
1370 return bucket[bucket_id].value;
1373 /* No tasks -- default clamp values */
1374 return uclamp_idle_value(rq, clamp_id, clamp_value);
1377 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1379 unsigned int default_util_min;
1380 struct uclamp_se *uc_se;
1382 lockdep_assert_held(&p->pi_lock);
1384 uc_se = &p->uclamp_req[UCLAMP_MIN];
1386 /* Only sync if user didn't override the default */
1387 if (uc_se->user_defined)
1390 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1391 uclamp_se_set(uc_se, default_util_min, false);
1394 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1402 /* Protect updates to p->uclamp_* */
1403 rq = task_rq_lock(p, &rf);
1404 __uclamp_update_util_min_rt_default(p);
1405 task_rq_unlock(rq, p, &rf);
1408 static void uclamp_sync_util_min_rt_default(void)
1410 struct task_struct *g, *p;
1413 * copy_process() sysctl_uclamp
1414 * uclamp_min_rt = X;
1415 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1416 * // link thread smp_mb__after_spinlock()
1417 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1418 * sched_post_fork() for_each_process_thread()
1419 * __uclamp_sync_rt() __uclamp_sync_rt()
1421 * Ensures that either sched_post_fork() will observe the new
1422 * uclamp_min_rt or for_each_process_thread() will observe the new
1425 read_lock(&tasklist_lock);
1426 smp_mb__after_spinlock();
1427 read_unlock(&tasklist_lock);
1430 for_each_process_thread(g, p)
1431 uclamp_update_util_min_rt_default(p);
1435 static inline struct uclamp_se
1436 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1438 /* Copy by value as we could modify it */
1439 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1440 #ifdef CONFIG_UCLAMP_TASK_GROUP
1441 unsigned int tg_min, tg_max, value;
1444 * Tasks in autogroups or root task group will be
1445 * restricted by system defaults.
1447 if (task_group_is_autogroup(task_group(p)))
1449 if (task_group(p) == &root_task_group)
1452 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1453 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1454 value = uc_req.value;
1455 value = clamp(value, tg_min, tg_max);
1456 uclamp_se_set(&uc_req, value, false);
1463 * The effective clamp bucket index of a task depends on, by increasing
1465 * - the task specific clamp value, when explicitly requested from userspace
1466 * - the task group effective clamp value, for tasks not either in the root
1467 * group or in an autogroup
1468 * - the system default clamp value, defined by the sysadmin
1470 static inline struct uclamp_se
1471 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1473 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1474 struct uclamp_se uc_max = uclamp_default[clamp_id];
1476 /* System default restrictions always apply */
1477 if (unlikely(uc_req.value > uc_max.value))
1483 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1485 struct uclamp_se uc_eff;
1487 /* Task currently refcounted: use back-annotated (effective) value */
1488 if (p->uclamp[clamp_id].active)
1489 return (unsigned long)p->uclamp[clamp_id].value;
1491 uc_eff = uclamp_eff_get(p, clamp_id);
1493 return (unsigned long)uc_eff.value;
1497 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1498 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1499 * updates the rq's clamp value if required.
1501 * Tasks can have a task-specific value requested from user-space, track
1502 * within each bucket the maximum value for tasks refcounted in it.
1503 * This "local max aggregation" allows to track the exact "requested" value
1504 * for each bucket when all its RUNNABLE tasks require the same clamp.
1506 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1507 enum uclamp_id clamp_id)
1509 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1510 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1511 struct uclamp_bucket *bucket;
1513 lockdep_assert_rq_held(rq);
1515 /* Update task effective clamp */
1516 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1518 bucket = &uc_rq->bucket[uc_se->bucket_id];
1520 uc_se->active = true;
1522 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1525 * Local max aggregation: rq buckets always track the max
1526 * "requested" clamp value of its RUNNABLE tasks.
1528 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1529 bucket->value = uc_se->value;
1531 if (uc_se->value > READ_ONCE(uc_rq->value))
1532 WRITE_ONCE(uc_rq->value, uc_se->value);
1536 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1537 * is released. If this is the last task reference counting the rq's max
1538 * active clamp value, then the rq's clamp value is updated.
1540 * Both refcounted tasks and rq's cached clamp values are expected to be
1541 * always valid. If it's detected they are not, as defensive programming,
1542 * enforce the expected state and warn.
1544 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1545 enum uclamp_id clamp_id)
1547 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1548 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1549 struct uclamp_bucket *bucket;
1550 unsigned int bkt_clamp;
1551 unsigned int rq_clamp;
1553 lockdep_assert_rq_held(rq);
1556 * If sched_uclamp_used was enabled after task @p was enqueued,
1557 * we could end up with unbalanced call to uclamp_rq_dec_id().
1559 * In this case the uc_se->active flag should be false since no uclamp
1560 * accounting was performed at enqueue time and we can just return
1563 * Need to be careful of the following enqueue/dequeue ordering
1567 * // sched_uclamp_used gets enabled
1570 * // Must not decrement bucket->tasks here
1573 * where we could end up with stale data in uc_se and
1574 * bucket[uc_se->bucket_id].
1576 * The following check here eliminates the possibility of such race.
1578 if (unlikely(!uc_se->active))
1581 bucket = &uc_rq->bucket[uc_se->bucket_id];
1583 SCHED_WARN_ON(!bucket->tasks);
1584 if (likely(bucket->tasks))
1587 uc_se->active = false;
1590 * Keep "local max aggregation" simple and accept to (possibly)
1591 * overboost some RUNNABLE tasks in the same bucket.
1592 * The rq clamp bucket value is reset to its base value whenever
1593 * there are no more RUNNABLE tasks refcounting it.
1595 if (likely(bucket->tasks))
1598 rq_clamp = READ_ONCE(uc_rq->value);
1600 * Defensive programming: this should never happen. If it happens,
1601 * e.g. due to future modification, warn and fixup the expected value.
1603 SCHED_WARN_ON(bucket->value > rq_clamp);
1604 if (bucket->value >= rq_clamp) {
1605 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1606 WRITE_ONCE(uc_rq->value, bkt_clamp);
1610 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1612 enum uclamp_id clamp_id;
1615 * Avoid any overhead until uclamp is actually used by the userspace.
1617 * The condition is constructed such that a NOP is generated when
1618 * sched_uclamp_used is disabled.
1620 if (!static_branch_unlikely(&sched_uclamp_used))
1623 if (unlikely(!p->sched_class->uclamp_enabled))
1626 for_each_clamp_id(clamp_id)
1627 uclamp_rq_inc_id(rq, p, clamp_id);
1629 /* Reset clamp idle holding when there is one RUNNABLE task */
1630 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1631 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1634 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1636 enum uclamp_id clamp_id;
1639 * Avoid any overhead until uclamp is actually used by the userspace.
1641 * The condition is constructed such that a NOP is generated when
1642 * sched_uclamp_used is disabled.
1644 if (!static_branch_unlikely(&sched_uclamp_used))
1647 if (unlikely(!p->sched_class->uclamp_enabled))
1650 for_each_clamp_id(clamp_id)
1651 uclamp_rq_dec_id(rq, p, clamp_id);
1654 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1655 enum uclamp_id clamp_id)
1657 if (!p->uclamp[clamp_id].active)
1660 uclamp_rq_dec_id(rq, p, clamp_id);
1661 uclamp_rq_inc_id(rq, p, clamp_id);
1664 * Make sure to clear the idle flag if we've transiently reached 0
1665 * active tasks on rq.
1667 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1668 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1672 uclamp_update_active(struct task_struct *p)
1674 enum uclamp_id clamp_id;
1679 * Lock the task and the rq where the task is (or was) queued.
1681 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1682 * price to pay to safely serialize util_{min,max} updates with
1683 * enqueues, dequeues and migration operations.
1684 * This is the same locking schema used by __set_cpus_allowed_ptr().
1686 rq = task_rq_lock(p, &rf);
1689 * Setting the clamp bucket is serialized by task_rq_lock().
1690 * If the task is not yet RUNNABLE and its task_struct is not
1691 * affecting a valid clamp bucket, the next time it's enqueued,
1692 * it will already see the updated clamp bucket value.
1694 for_each_clamp_id(clamp_id)
1695 uclamp_rq_reinc_id(rq, p, clamp_id);
1697 task_rq_unlock(rq, p, &rf);
1700 #ifdef CONFIG_UCLAMP_TASK_GROUP
1702 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1704 struct css_task_iter it;
1705 struct task_struct *p;
1707 css_task_iter_start(css, 0, &it);
1708 while ((p = css_task_iter_next(&it)))
1709 uclamp_update_active(p);
1710 css_task_iter_end(&it);
1713 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1714 static void uclamp_update_root_tg(void)
1716 struct task_group *tg = &root_task_group;
1718 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1719 sysctl_sched_uclamp_util_min, false);
1720 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1721 sysctl_sched_uclamp_util_max, false);
1724 cpu_util_update_eff(&root_task_group.css);
1728 static void uclamp_update_root_tg(void) { }
1731 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1732 void *buffer, size_t *lenp, loff_t *ppos)
1734 bool update_root_tg = false;
1735 int old_min, old_max, old_min_rt;
1738 mutex_lock(&uclamp_mutex);
1739 old_min = sysctl_sched_uclamp_util_min;
1740 old_max = sysctl_sched_uclamp_util_max;
1741 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1743 result = proc_dointvec(table, write, buffer, lenp, ppos);
1749 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1750 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1751 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1757 if (old_min != sysctl_sched_uclamp_util_min) {
1758 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1759 sysctl_sched_uclamp_util_min, false);
1760 update_root_tg = true;
1762 if (old_max != sysctl_sched_uclamp_util_max) {
1763 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1764 sysctl_sched_uclamp_util_max, false);
1765 update_root_tg = true;
1768 if (update_root_tg) {
1769 static_branch_enable(&sched_uclamp_used);
1770 uclamp_update_root_tg();
1773 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1774 static_branch_enable(&sched_uclamp_used);
1775 uclamp_sync_util_min_rt_default();
1779 * We update all RUNNABLE tasks only when task groups are in use.
1780 * Otherwise, keep it simple and do just a lazy update at each next
1781 * task enqueue time.
1787 sysctl_sched_uclamp_util_min = old_min;
1788 sysctl_sched_uclamp_util_max = old_max;
1789 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1791 mutex_unlock(&uclamp_mutex);
1796 static int uclamp_validate(struct task_struct *p,
1797 const struct sched_attr *attr)
1799 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1800 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1802 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1803 util_min = attr->sched_util_min;
1805 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1809 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1810 util_max = attr->sched_util_max;
1812 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1816 if (util_min != -1 && util_max != -1 && util_min > util_max)
1820 * We have valid uclamp attributes; make sure uclamp is enabled.
1822 * We need to do that here, because enabling static branches is a
1823 * blocking operation which obviously cannot be done while holding
1826 static_branch_enable(&sched_uclamp_used);
1831 static bool uclamp_reset(const struct sched_attr *attr,
1832 enum uclamp_id clamp_id,
1833 struct uclamp_se *uc_se)
1835 /* Reset on sched class change for a non user-defined clamp value. */
1836 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1837 !uc_se->user_defined)
1840 /* Reset on sched_util_{min,max} == -1. */
1841 if (clamp_id == UCLAMP_MIN &&
1842 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1843 attr->sched_util_min == -1) {
1847 if (clamp_id == UCLAMP_MAX &&
1848 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1849 attr->sched_util_max == -1) {
1856 static void __setscheduler_uclamp(struct task_struct *p,
1857 const struct sched_attr *attr)
1859 enum uclamp_id clamp_id;
1861 for_each_clamp_id(clamp_id) {
1862 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1865 if (!uclamp_reset(attr, clamp_id, uc_se))
1869 * RT by default have a 100% boost value that could be modified
1872 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1873 value = sysctl_sched_uclamp_util_min_rt_default;
1875 value = uclamp_none(clamp_id);
1877 uclamp_se_set(uc_se, value, false);
1881 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1884 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1885 attr->sched_util_min != -1) {
1886 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1887 attr->sched_util_min, true);
1890 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1891 attr->sched_util_max != -1) {
1892 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1893 attr->sched_util_max, true);
1897 static void uclamp_fork(struct task_struct *p)
1899 enum uclamp_id clamp_id;
1902 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1903 * as the task is still at its early fork stages.
1905 for_each_clamp_id(clamp_id)
1906 p->uclamp[clamp_id].active = false;
1908 if (likely(!p->sched_reset_on_fork))
1911 for_each_clamp_id(clamp_id) {
1912 uclamp_se_set(&p->uclamp_req[clamp_id],
1913 uclamp_none(clamp_id), false);
1917 static void uclamp_post_fork(struct task_struct *p)
1919 uclamp_update_util_min_rt_default(p);
1922 static void __init init_uclamp_rq(struct rq *rq)
1924 enum uclamp_id clamp_id;
1925 struct uclamp_rq *uc_rq = rq->uclamp;
1927 for_each_clamp_id(clamp_id) {
1928 uc_rq[clamp_id] = (struct uclamp_rq) {
1929 .value = uclamp_none(clamp_id)
1933 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1936 static void __init init_uclamp(void)
1938 struct uclamp_se uc_max = {};
1939 enum uclamp_id clamp_id;
1942 for_each_possible_cpu(cpu)
1943 init_uclamp_rq(cpu_rq(cpu));
1945 for_each_clamp_id(clamp_id) {
1946 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1947 uclamp_none(clamp_id), false);
1950 /* System defaults allow max clamp values for both indexes */
1951 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1952 for_each_clamp_id(clamp_id) {
1953 uclamp_default[clamp_id] = uc_max;
1954 #ifdef CONFIG_UCLAMP_TASK_GROUP
1955 root_task_group.uclamp_req[clamp_id] = uc_max;
1956 root_task_group.uclamp[clamp_id] = uc_max;
1961 #else /* CONFIG_UCLAMP_TASK */
1962 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1963 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1964 static inline int uclamp_validate(struct task_struct *p,
1965 const struct sched_attr *attr)
1969 static void __setscheduler_uclamp(struct task_struct *p,
1970 const struct sched_attr *attr) { }
1971 static inline void uclamp_fork(struct task_struct *p) { }
1972 static inline void uclamp_post_fork(struct task_struct *p) { }
1973 static inline void init_uclamp(void) { }
1974 #endif /* CONFIG_UCLAMP_TASK */
1976 bool sched_task_on_rq(struct task_struct *p)
1978 return task_on_rq_queued(p);
1981 unsigned long get_wchan(struct task_struct *p)
1983 unsigned long ip = 0;
1986 if (!p || p == current)
1989 /* Only get wchan if task is blocked and we can keep it that way. */
1990 raw_spin_lock_irq(&p->pi_lock);
1991 state = READ_ONCE(p->__state);
1992 smp_rmb(); /* see try_to_wake_up() */
1993 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
1994 ip = __get_wchan(p);
1995 raw_spin_unlock_irq(&p->pi_lock);
2000 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2002 if (!(flags & ENQUEUE_NOCLOCK))
2003 update_rq_clock(rq);
2005 if (!(flags & ENQUEUE_RESTORE)) {
2006 sched_info_enqueue(rq, p);
2007 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2010 uclamp_rq_inc(rq, p);
2011 p->sched_class->enqueue_task(rq, p, flags);
2013 if (sched_core_enabled(rq))
2014 sched_core_enqueue(rq, p);
2017 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2019 if (sched_core_enabled(rq))
2020 sched_core_dequeue(rq, p, flags);
2022 if (!(flags & DEQUEUE_NOCLOCK))
2023 update_rq_clock(rq);
2025 if (!(flags & DEQUEUE_SAVE)) {
2026 sched_info_dequeue(rq, p);
2027 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2030 uclamp_rq_dec(rq, p);
2031 p->sched_class->dequeue_task(rq, p, flags);
2034 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2036 enqueue_task(rq, p, flags);
2038 p->on_rq = TASK_ON_RQ_QUEUED;
2041 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2043 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2045 dequeue_task(rq, p, flags);
2048 static inline int __normal_prio(int policy, int rt_prio, int nice)
2052 if (dl_policy(policy))
2053 prio = MAX_DL_PRIO - 1;
2054 else if (rt_policy(policy))
2055 prio = MAX_RT_PRIO - 1 - rt_prio;
2057 prio = NICE_TO_PRIO(nice);
2063 * Calculate the expected normal priority: i.e. priority
2064 * without taking RT-inheritance into account. Might be
2065 * boosted by interactivity modifiers. Changes upon fork,
2066 * setprio syscalls, and whenever the interactivity
2067 * estimator recalculates.
2069 static inline int normal_prio(struct task_struct *p)
2071 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2075 * Calculate the current priority, i.e. the priority
2076 * taken into account by the scheduler. This value might
2077 * be boosted by RT tasks, or might be boosted by
2078 * interactivity modifiers. Will be RT if the task got
2079 * RT-boosted. If not then it returns p->normal_prio.
2081 static int effective_prio(struct task_struct *p)
2083 p->normal_prio = normal_prio(p);
2085 * If we are RT tasks or we were boosted to RT priority,
2086 * keep the priority unchanged. Otherwise, update priority
2087 * to the normal priority:
2089 if (!rt_prio(p->prio))
2090 return p->normal_prio;
2095 * task_curr - is this task currently executing on a CPU?
2096 * @p: the task in question.
2098 * Return: 1 if the task is currently executing. 0 otherwise.
2100 inline int task_curr(const struct task_struct *p)
2102 return cpu_curr(task_cpu(p)) == p;
2106 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2107 * use the balance_callback list if you want balancing.
2109 * this means any call to check_class_changed() must be followed by a call to
2110 * balance_callback().
2112 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2113 const struct sched_class *prev_class,
2116 if (prev_class != p->sched_class) {
2117 if (prev_class->switched_from)
2118 prev_class->switched_from(rq, p);
2120 p->sched_class->switched_to(rq, p);
2121 } else if (oldprio != p->prio || dl_task(p))
2122 p->sched_class->prio_changed(rq, p, oldprio);
2125 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2127 if (p->sched_class == rq->curr->sched_class)
2128 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2129 else if (p->sched_class > rq->curr->sched_class)
2133 * A queue event has occurred, and we're going to schedule. In
2134 * this case, we can save a useless back to back clock update.
2136 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2137 rq_clock_skip_update(rq);
2143 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2145 static int __set_cpus_allowed_ptr(struct task_struct *p,
2146 const struct cpumask *new_mask,
2149 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2151 if (likely(!p->migration_disabled))
2154 if (p->cpus_ptr != &p->cpus_mask)
2158 * Violates locking rules! see comment in __do_set_cpus_allowed().
2160 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2163 void migrate_disable(void)
2165 struct task_struct *p = current;
2167 if (p->migration_disabled) {
2168 p->migration_disabled++;
2173 this_rq()->nr_pinned++;
2174 p->migration_disabled = 1;
2177 EXPORT_SYMBOL_GPL(migrate_disable);
2179 void migrate_enable(void)
2181 struct task_struct *p = current;
2183 if (p->migration_disabled > 1) {
2184 p->migration_disabled--;
2188 if (WARN_ON_ONCE(!p->migration_disabled))
2192 * Ensure stop_task runs either before or after this, and that
2193 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2196 if (p->cpus_ptr != &p->cpus_mask)
2197 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2199 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2200 * regular cpus_mask, otherwise things that race (eg.
2201 * select_fallback_rq) get confused.
2204 p->migration_disabled = 0;
2205 this_rq()->nr_pinned--;
2208 EXPORT_SYMBOL_GPL(migrate_enable);
2210 static inline bool rq_has_pinned_tasks(struct rq *rq)
2212 return rq->nr_pinned;
2216 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2217 * __set_cpus_allowed_ptr() and select_fallback_rq().
2219 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2221 /* When not in the task's cpumask, no point in looking further. */
2222 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2225 /* migrate_disabled() must be allowed to finish. */
2226 if (is_migration_disabled(p))
2227 return cpu_online(cpu);
2229 /* Non kernel threads are not allowed during either online or offline. */
2230 if (!(p->flags & PF_KTHREAD))
2231 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2233 /* KTHREAD_IS_PER_CPU is always allowed. */
2234 if (kthread_is_per_cpu(p))
2235 return cpu_online(cpu);
2237 /* Regular kernel threads don't get to stay during offline. */
2241 /* But are allowed during online. */
2242 return cpu_online(cpu);
2246 * This is how migration works:
2248 * 1) we invoke migration_cpu_stop() on the target CPU using
2250 * 2) stopper starts to run (implicitly forcing the migrated thread
2252 * 3) it checks whether the migrated task is still in the wrong runqueue.
2253 * 4) if it's in the wrong runqueue then the migration thread removes
2254 * it and puts it into the right queue.
2255 * 5) stopper completes and stop_one_cpu() returns and the migration
2260 * move_queued_task - move a queued task to new rq.
2262 * Returns (locked) new rq. Old rq's lock is released.
2264 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2265 struct task_struct *p, int new_cpu)
2267 lockdep_assert_rq_held(rq);
2269 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2270 set_task_cpu(p, new_cpu);
2273 rq = cpu_rq(new_cpu);
2276 BUG_ON(task_cpu(p) != new_cpu);
2277 activate_task(rq, p, 0);
2278 check_preempt_curr(rq, p, 0);
2283 struct migration_arg {
2284 struct task_struct *task;
2286 struct set_affinity_pending *pending;
2290 * @refs: number of wait_for_completion()
2291 * @stop_pending: is @stop_work in use
2293 struct set_affinity_pending {
2295 unsigned int stop_pending;
2296 struct completion done;
2297 struct cpu_stop_work stop_work;
2298 struct migration_arg arg;
2302 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2303 * this because either it can't run here any more (set_cpus_allowed()
2304 * away from this CPU, or CPU going down), or because we're
2305 * attempting to rebalance this task on exec (sched_exec).
2307 * So we race with normal scheduler movements, but that's OK, as long
2308 * as the task is no longer on this CPU.
2310 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2311 struct task_struct *p, int dest_cpu)
2313 /* Affinity changed (again). */
2314 if (!is_cpu_allowed(p, dest_cpu))
2317 update_rq_clock(rq);
2318 rq = move_queued_task(rq, rf, p, dest_cpu);
2324 * migration_cpu_stop - this will be executed by a highprio stopper thread
2325 * and performs thread migration by bumping thread off CPU then
2326 * 'pushing' onto another runqueue.
2328 static int migration_cpu_stop(void *data)
2330 struct migration_arg *arg = data;
2331 struct set_affinity_pending *pending = arg->pending;
2332 struct task_struct *p = arg->task;
2333 struct rq *rq = this_rq();
2334 bool complete = false;
2338 * The original target CPU might have gone down and we might
2339 * be on another CPU but it doesn't matter.
2341 local_irq_save(rf.flags);
2343 * We need to explicitly wake pending tasks before running
2344 * __migrate_task() such that we will not miss enforcing cpus_ptr
2345 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2347 flush_smp_call_function_from_idle();
2349 raw_spin_lock(&p->pi_lock);
2353 * If we were passed a pending, then ->stop_pending was set, thus
2354 * p->migration_pending must have remained stable.
2356 WARN_ON_ONCE(pending && pending != p->migration_pending);
2359 * If task_rq(p) != rq, it cannot be migrated here, because we're
2360 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2361 * we're holding p->pi_lock.
2363 if (task_rq(p) == rq) {
2364 if (is_migration_disabled(p))
2368 p->migration_pending = NULL;
2371 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2375 if (task_on_rq_queued(p))
2376 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2378 p->wake_cpu = arg->dest_cpu;
2381 * XXX __migrate_task() can fail, at which point we might end
2382 * up running on a dodgy CPU, AFAICT this can only happen
2383 * during CPU hotplug, at which point we'll get pushed out
2384 * anyway, so it's probably not a big deal.
2387 } else if (pending) {
2389 * This happens when we get migrated between migrate_enable()'s
2390 * preempt_enable() and scheduling the stopper task. At that
2391 * point we're a regular task again and not current anymore.
2393 * A !PREEMPT kernel has a giant hole here, which makes it far
2398 * The task moved before the stopper got to run. We're holding
2399 * ->pi_lock, so the allowed mask is stable - if it got
2400 * somewhere allowed, we're done.
2402 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2403 p->migration_pending = NULL;
2409 * When migrate_enable() hits a rq mis-match we can't reliably
2410 * determine is_migration_disabled() and so have to chase after
2413 WARN_ON_ONCE(!pending->stop_pending);
2414 task_rq_unlock(rq, p, &rf);
2415 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2416 &pending->arg, &pending->stop_work);
2421 pending->stop_pending = false;
2422 task_rq_unlock(rq, p, &rf);
2425 complete_all(&pending->done);
2430 int push_cpu_stop(void *arg)
2432 struct rq *lowest_rq = NULL, *rq = this_rq();
2433 struct task_struct *p = arg;
2435 raw_spin_lock_irq(&p->pi_lock);
2436 raw_spin_rq_lock(rq);
2438 if (task_rq(p) != rq)
2441 if (is_migration_disabled(p)) {
2442 p->migration_flags |= MDF_PUSH;
2446 p->migration_flags &= ~MDF_PUSH;
2448 if (p->sched_class->find_lock_rq)
2449 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2454 // XXX validate p is still the highest prio task
2455 if (task_rq(p) == rq) {
2456 deactivate_task(rq, p, 0);
2457 set_task_cpu(p, lowest_rq->cpu);
2458 activate_task(lowest_rq, p, 0);
2459 resched_curr(lowest_rq);
2462 double_unlock_balance(rq, lowest_rq);
2465 rq->push_busy = false;
2466 raw_spin_rq_unlock(rq);
2467 raw_spin_unlock_irq(&p->pi_lock);
2474 * sched_class::set_cpus_allowed must do the below, but is not required to
2475 * actually call this function.
2477 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2479 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2480 p->cpus_ptr = new_mask;
2484 cpumask_copy(&p->cpus_mask, new_mask);
2485 p->nr_cpus_allowed = cpumask_weight(new_mask);
2489 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2491 struct rq *rq = task_rq(p);
2492 bool queued, running;
2495 * This here violates the locking rules for affinity, since we're only
2496 * supposed to change these variables while holding both rq->lock and
2499 * HOWEVER, it magically works, because ttwu() is the only code that
2500 * accesses these variables under p->pi_lock and only does so after
2501 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2502 * before finish_task().
2504 * XXX do further audits, this smells like something putrid.
2506 if (flags & SCA_MIGRATE_DISABLE)
2507 SCHED_WARN_ON(!p->on_cpu);
2509 lockdep_assert_held(&p->pi_lock);
2511 queued = task_on_rq_queued(p);
2512 running = task_current(rq, p);
2516 * Because __kthread_bind() calls this on blocked tasks without
2519 lockdep_assert_rq_held(rq);
2520 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2523 put_prev_task(rq, p);
2525 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2528 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2530 set_next_task(rq, p);
2533 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2535 __do_set_cpus_allowed(p, new_mask, 0);
2538 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2541 if (!src->user_cpus_ptr)
2544 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2545 if (!dst->user_cpus_ptr)
2548 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2552 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2554 struct cpumask *user_mask = NULL;
2556 swap(p->user_cpus_ptr, user_mask);
2561 void release_user_cpus_ptr(struct task_struct *p)
2563 kfree(clear_user_cpus_ptr(p));
2567 * This function is wildly self concurrent; here be dragons.
2570 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2571 * designated task is enqueued on an allowed CPU. If that task is currently
2572 * running, we have to kick it out using the CPU stopper.
2574 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2577 * Initial conditions: P0->cpus_mask = [0, 1]
2581 * migrate_disable();
2583 * set_cpus_allowed_ptr(P0, [1]);
2585 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2586 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2587 * This means we need the following scheme:
2591 * migrate_disable();
2593 * set_cpus_allowed_ptr(P0, [1]);
2597 * __set_cpus_allowed_ptr();
2598 * <wakes local stopper>
2599 * `--> <woken on migration completion>
2601 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2602 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2603 * task p are serialized by p->pi_lock, which we can leverage: the one that
2604 * should come into effect at the end of the Migrate-Disable region is the last
2605 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2606 * but we still need to properly signal those waiting tasks at the appropriate
2609 * This is implemented using struct set_affinity_pending. The first
2610 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2611 * setup an instance of that struct and install it on the targeted task_struct.
2612 * Any and all further callers will reuse that instance. Those then wait for
2613 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2614 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2617 * (1) In the cases covered above. There is one more where the completion is
2618 * signaled within affine_move_task() itself: when a subsequent affinity request
2619 * occurs after the stopper bailed out due to the targeted task still being
2620 * Migrate-Disable. Consider:
2622 * Initial conditions: P0->cpus_mask = [0, 1]
2626 * migrate_disable();
2628 * set_cpus_allowed_ptr(P0, [1]);
2631 * migration_cpu_stop()
2632 * is_migration_disabled()
2634 * set_cpus_allowed_ptr(P0, [0, 1]);
2635 * <signal completion>
2638 * Note that the above is safe vs a concurrent migrate_enable(), as any
2639 * pending affinity completion is preceded by an uninstallation of
2640 * p->migration_pending done with p->pi_lock held.
2642 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2643 int dest_cpu, unsigned int flags)
2645 struct set_affinity_pending my_pending = { }, *pending = NULL;
2646 bool stop_pending, complete = false;
2648 /* Can the task run on the task's current CPU? If so, we're done */
2649 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2650 struct task_struct *push_task = NULL;
2652 if ((flags & SCA_MIGRATE_ENABLE) &&
2653 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2654 rq->push_busy = true;
2655 push_task = get_task_struct(p);
2659 * If there are pending waiters, but no pending stop_work,
2660 * then complete now.
2662 pending = p->migration_pending;
2663 if (pending && !pending->stop_pending) {
2664 p->migration_pending = NULL;
2668 task_rq_unlock(rq, p, rf);
2671 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2676 complete_all(&pending->done);
2681 if (!(flags & SCA_MIGRATE_ENABLE)) {
2682 /* serialized by p->pi_lock */
2683 if (!p->migration_pending) {
2684 /* Install the request */
2685 refcount_set(&my_pending.refs, 1);
2686 init_completion(&my_pending.done);
2687 my_pending.arg = (struct migration_arg) {
2689 .dest_cpu = dest_cpu,
2690 .pending = &my_pending,
2693 p->migration_pending = &my_pending;
2695 pending = p->migration_pending;
2696 refcount_inc(&pending->refs);
2698 * Affinity has changed, but we've already installed a
2699 * pending. migration_cpu_stop() *must* see this, else
2700 * we risk a completion of the pending despite having a
2701 * task on a disallowed CPU.
2703 * Serialized by p->pi_lock, so this is safe.
2705 pending->arg.dest_cpu = dest_cpu;
2708 pending = p->migration_pending;
2710 * - !MIGRATE_ENABLE:
2711 * we'll have installed a pending if there wasn't one already.
2714 * we're here because the current CPU isn't matching anymore,
2715 * the only way that can happen is because of a concurrent
2716 * set_cpus_allowed_ptr() call, which should then still be
2717 * pending completion.
2719 * Either way, we really should have a @pending here.
2721 if (WARN_ON_ONCE(!pending)) {
2722 task_rq_unlock(rq, p, rf);
2726 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2728 * MIGRATE_ENABLE gets here because 'p == current', but for
2729 * anything else we cannot do is_migration_disabled(), punt
2730 * and have the stopper function handle it all race-free.
2732 stop_pending = pending->stop_pending;
2734 pending->stop_pending = true;
2736 if (flags & SCA_MIGRATE_ENABLE)
2737 p->migration_flags &= ~MDF_PUSH;
2739 task_rq_unlock(rq, p, rf);
2741 if (!stop_pending) {
2742 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2743 &pending->arg, &pending->stop_work);
2746 if (flags & SCA_MIGRATE_ENABLE)
2750 if (!is_migration_disabled(p)) {
2751 if (task_on_rq_queued(p))
2752 rq = move_queued_task(rq, rf, p, dest_cpu);
2754 if (!pending->stop_pending) {
2755 p->migration_pending = NULL;
2759 task_rq_unlock(rq, p, rf);
2762 complete_all(&pending->done);
2765 wait_for_completion(&pending->done);
2767 if (refcount_dec_and_test(&pending->refs))
2768 wake_up_var(&pending->refs); /* No UaF, just an address */
2771 * Block the original owner of &pending until all subsequent callers
2772 * have seen the completion and decremented the refcount
2774 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2777 WARN_ON_ONCE(my_pending.stop_pending);
2783 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2785 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2786 const struct cpumask *new_mask,
2789 struct rq_flags *rf)
2790 __releases(rq->lock)
2791 __releases(p->pi_lock)
2793 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2794 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2795 bool kthread = p->flags & PF_KTHREAD;
2796 struct cpumask *user_mask = NULL;
2797 unsigned int dest_cpu;
2800 update_rq_clock(rq);
2802 if (kthread || is_migration_disabled(p)) {
2804 * Kernel threads are allowed on online && !active CPUs,
2805 * however, during cpu-hot-unplug, even these might get pushed
2806 * away if not KTHREAD_IS_PER_CPU.
2808 * Specifically, migration_disabled() tasks must not fail the
2809 * cpumask_any_and_distribute() pick below, esp. so on
2810 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2811 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2813 cpu_valid_mask = cpu_online_mask;
2816 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2822 * Must re-check here, to close a race against __kthread_bind(),
2823 * sched_setaffinity() is not guaranteed to observe the flag.
2825 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2830 if (!(flags & SCA_MIGRATE_ENABLE)) {
2831 if (cpumask_equal(&p->cpus_mask, new_mask))
2834 if (WARN_ON_ONCE(p == current &&
2835 is_migration_disabled(p) &&
2836 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2843 * Picking a ~random cpu helps in cases where we are changing affinity
2844 * for groups of tasks (ie. cpuset), so that load balancing is not
2845 * immediately required to distribute the tasks within their new mask.
2847 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2848 if (dest_cpu >= nr_cpu_ids) {
2853 __do_set_cpus_allowed(p, new_mask, flags);
2855 if (flags & SCA_USER)
2856 user_mask = clear_user_cpus_ptr(p);
2858 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2865 task_rq_unlock(rq, p, rf);
2871 * Change a given task's CPU affinity. Migrate the thread to a
2872 * proper CPU and schedule it away if the CPU it's executing on
2873 * is removed from the allowed bitmask.
2875 * NOTE: the caller must have a valid reference to the task, the
2876 * task must not exit() & deallocate itself prematurely. The
2877 * call is not atomic; no spinlocks may be held.
2879 static int __set_cpus_allowed_ptr(struct task_struct *p,
2880 const struct cpumask *new_mask, u32 flags)
2885 rq = task_rq_lock(p, &rf);
2886 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2889 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2891 return __set_cpus_allowed_ptr(p, new_mask, 0);
2893 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2896 * Change a given task's CPU affinity to the intersection of its current
2897 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2898 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2899 * If the resulting mask is empty, leave the affinity unchanged and return
2902 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2903 struct cpumask *new_mask,
2904 const struct cpumask *subset_mask)
2906 struct cpumask *user_mask = NULL;
2911 if (!p->user_cpus_ptr) {
2912 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2917 rq = task_rq_lock(p, &rf);
2920 * Forcefully restricting the affinity of a deadline task is
2921 * likely to cause problems, so fail and noisily override the
2924 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2929 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2935 * We're about to butcher the task affinity, so keep track of what
2936 * the user asked for in case we're able to restore it later on.
2939 cpumask_copy(user_mask, p->cpus_ptr);
2940 p->user_cpus_ptr = user_mask;
2943 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2946 task_rq_unlock(rq, p, &rf);
2952 * Restrict the CPU affinity of task @p so that it is a subset of
2953 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2954 * old affinity mask. If the resulting mask is empty, we warn and walk
2955 * up the cpuset hierarchy until we find a suitable mask.
2957 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2959 cpumask_var_t new_mask;
2960 const struct cpumask *override_mask = task_cpu_possible_mask(p);
2962 alloc_cpumask_var(&new_mask, GFP_KERNEL);
2965 * __migrate_task() can fail silently in the face of concurrent
2966 * offlining of the chosen destination CPU, so take the hotplug
2967 * lock to ensure that the migration succeeds.
2970 if (!cpumask_available(new_mask))
2973 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2977 * We failed to find a valid subset of the affinity mask for the
2978 * task, so override it based on its cpuset hierarchy.
2980 cpuset_cpus_allowed(p, new_mask);
2981 override_mask = new_mask;
2984 if (printk_ratelimit()) {
2985 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2986 task_pid_nr(p), p->comm,
2987 cpumask_pr_args(override_mask));
2990 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2993 free_cpumask_var(new_mask);
2997 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3000 * Restore the affinity of a task @p which was previously restricted by a
3001 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3002 * @p->user_cpus_ptr.
3004 * It is the caller's responsibility to serialise this with any calls to
3005 * force_compatible_cpus_allowed_ptr(@p).
3007 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3009 struct cpumask *user_mask = p->user_cpus_ptr;
3010 unsigned long flags;
3013 * Try to restore the old affinity mask. If this fails, then
3014 * we free the mask explicitly to avoid it being inherited across
3015 * a subsequent fork().
3017 if (!user_mask || !__sched_setaffinity(p, user_mask))
3020 raw_spin_lock_irqsave(&p->pi_lock, flags);
3021 user_mask = clear_user_cpus_ptr(p);
3022 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3027 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3029 #ifdef CONFIG_SCHED_DEBUG
3030 unsigned int state = READ_ONCE(p->__state);
3033 * We should never call set_task_cpu() on a blocked task,
3034 * ttwu() will sort out the placement.
3036 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3039 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3040 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3041 * time relying on p->on_rq.
3043 WARN_ON_ONCE(state == TASK_RUNNING &&
3044 p->sched_class == &fair_sched_class &&
3045 (p->on_rq && !task_on_rq_migrating(p)));
3047 #ifdef CONFIG_LOCKDEP
3049 * The caller should hold either p->pi_lock or rq->lock, when changing
3050 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3052 * sched_move_task() holds both and thus holding either pins the cgroup,
3055 * Furthermore, all task_rq users should acquire both locks, see
3058 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3059 lockdep_is_held(__rq_lockp(task_rq(p)))));
3062 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3064 WARN_ON_ONCE(!cpu_online(new_cpu));
3066 WARN_ON_ONCE(is_migration_disabled(p));
3069 trace_sched_migrate_task(p, new_cpu);
3071 if (task_cpu(p) != new_cpu) {
3072 if (p->sched_class->migrate_task_rq)
3073 p->sched_class->migrate_task_rq(p, new_cpu);
3074 p->se.nr_migrations++;
3076 perf_event_task_migrate(p);
3079 __set_task_cpu(p, new_cpu);
3082 #ifdef CONFIG_NUMA_BALANCING
3083 static void __migrate_swap_task(struct task_struct *p, int cpu)
3085 if (task_on_rq_queued(p)) {
3086 struct rq *src_rq, *dst_rq;
3087 struct rq_flags srf, drf;
3089 src_rq = task_rq(p);
3090 dst_rq = cpu_rq(cpu);
3092 rq_pin_lock(src_rq, &srf);
3093 rq_pin_lock(dst_rq, &drf);
3095 deactivate_task(src_rq, p, 0);
3096 set_task_cpu(p, cpu);
3097 activate_task(dst_rq, p, 0);
3098 check_preempt_curr(dst_rq, p, 0);
3100 rq_unpin_lock(dst_rq, &drf);
3101 rq_unpin_lock(src_rq, &srf);
3105 * Task isn't running anymore; make it appear like we migrated
3106 * it before it went to sleep. This means on wakeup we make the
3107 * previous CPU our target instead of where it really is.
3113 struct migration_swap_arg {
3114 struct task_struct *src_task, *dst_task;
3115 int src_cpu, dst_cpu;
3118 static int migrate_swap_stop(void *data)
3120 struct migration_swap_arg *arg = data;
3121 struct rq *src_rq, *dst_rq;
3124 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3127 src_rq = cpu_rq(arg->src_cpu);
3128 dst_rq = cpu_rq(arg->dst_cpu);
3130 double_raw_lock(&arg->src_task->pi_lock,
3131 &arg->dst_task->pi_lock);
3132 double_rq_lock(src_rq, dst_rq);
3134 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3137 if (task_cpu(arg->src_task) != arg->src_cpu)
3140 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3143 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3146 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3147 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3152 double_rq_unlock(src_rq, dst_rq);
3153 raw_spin_unlock(&arg->dst_task->pi_lock);
3154 raw_spin_unlock(&arg->src_task->pi_lock);
3160 * Cross migrate two tasks
3162 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3163 int target_cpu, int curr_cpu)
3165 struct migration_swap_arg arg;
3168 arg = (struct migration_swap_arg){
3170 .src_cpu = curr_cpu,
3172 .dst_cpu = target_cpu,
3175 if (arg.src_cpu == arg.dst_cpu)
3179 * These three tests are all lockless; this is OK since all of them
3180 * will be re-checked with proper locks held further down the line.
3182 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3185 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3188 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3191 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3192 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3197 #endif /* CONFIG_NUMA_BALANCING */
3200 * wait_task_inactive - wait for a thread to unschedule.
3202 * If @match_state is nonzero, it's the @p->state value just checked and
3203 * not expected to change. If it changes, i.e. @p might have woken up,
3204 * then return zero. When we succeed in waiting for @p to be off its CPU,
3205 * we return a positive number (its total switch count). If a second call
3206 * a short while later returns the same number, the caller can be sure that
3207 * @p has remained unscheduled the whole time.
3209 * The caller must ensure that the task *will* unschedule sometime soon,
3210 * else this function might spin for a *long* time. This function can't
3211 * be called with interrupts off, or it may introduce deadlock with
3212 * smp_call_function() if an IPI is sent by the same process we are
3213 * waiting to become inactive.
3215 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3217 int running, queued;
3224 * We do the initial early heuristics without holding
3225 * any task-queue locks at all. We'll only try to get
3226 * the runqueue lock when things look like they will
3232 * If the task is actively running on another CPU
3233 * still, just relax and busy-wait without holding
3236 * NOTE! Since we don't hold any locks, it's not
3237 * even sure that "rq" stays as the right runqueue!
3238 * But we don't care, since "task_running()" will
3239 * return false if the runqueue has changed and p
3240 * is actually now running somewhere else!
3242 while (task_running(rq, p)) {
3243 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3249 * Ok, time to look more closely! We need the rq
3250 * lock now, to be *sure*. If we're wrong, we'll
3251 * just go back and repeat.
3253 rq = task_rq_lock(p, &rf);
3254 trace_sched_wait_task(p);
3255 running = task_running(rq, p);
3256 queued = task_on_rq_queued(p);
3258 if (!match_state || READ_ONCE(p->__state) == match_state)
3259 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3260 task_rq_unlock(rq, p, &rf);
3263 * If it changed from the expected state, bail out now.
3265 if (unlikely(!ncsw))
3269 * Was it really running after all now that we
3270 * checked with the proper locks actually held?
3272 * Oops. Go back and try again..
3274 if (unlikely(running)) {
3280 * It's not enough that it's not actively running,
3281 * it must be off the runqueue _entirely_, and not
3284 * So if it was still runnable (but just not actively
3285 * running right now), it's preempted, and we should
3286 * yield - it could be a while.
3288 if (unlikely(queued)) {
3289 ktime_t to = NSEC_PER_SEC / HZ;
3291 set_current_state(TASK_UNINTERRUPTIBLE);
3292 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3297 * Ahh, all good. It wasn't running, and it wasn't
3298 * runnable, which means that it will never become
3299 * running in the future either. We're all done!
3308 * kick_process - kick a running thread to enter/exit the kernel
3309 * @p: the to-be-kicked thread
3311 * Cause a process which is running on another CPU to enter
3312 * kernel-mode, without any delay. (to get signals handled.)
3314 * NOTE: this function doesn't have to take the runqueue lock,
3315 * because all it wants to ensure is that the remote task enters
3316 * the kernel. If the IPI races and the task has been migrated
3317 * to another CPU then no harm is done and the purpose has been
3320 void kick_process(struct task_struct *p)
3326 if ((cpu != smp_processor_id()) && task_curr(p))
3327 smp_send_reschedule(cpu);
3330 EXPORT_SYMBOL_GPL(kick_process);
3333 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3335 * A few notes on cpu_active vs cpu_online:
3337 * - cpu_active must be a subset of cpu_online
3339 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3340 * see __set_cpus_allowed_ptr(). At this point the newly online
3341 * CPU isn't yet part of the sched domains, and balancing will not
3344 * - on CPU-down we clear cpu_active() to mask the sched domains and
3345 * avoid the load balancer to place new tasks on the to be removed
3346 * CPU. Existing tasks will remain running there and will be taken
3349 * This means that fallback selection must not select !active CPUs.
3350 * And can assume that any active CPU must be online. Conversely
3351 * select_task_rq() below may allow selection of !active CPUs in order
3352 * to satisfy the above rules.
3354 static int select_fallback_rq(int cpu, struct task_struct *p)
3356 int nid = cpu_to_node(cpu);
3357 const struct cpumask *nodemask = NULL;
3358 enum { cpuset, possible, fail } state = cpuset;
3362 * If the node that the CPU is on has been offlined, cpu_to_node()
3363 * will return -1. There is no CPU on the node, and we should
3364 * select the CPU on the other node.
3367 nodemask = cpumask_of_node(nid);
3369 /* Look for allowed, online CPU in same node. */
3370 for_each_cpu(dest_cpu, nodemask) {
3371 if (is_cpu_allowed(p, dest_cpu))
3377 /* Any allowed, online CPU? */
3378 for_each_cpu(dest_cpu, p->cpus_ptr) {
3379 if (!is_cpu_allowed(p, dest_cpu))
3385 /* No more Mr. Nice Guy. */
3388 if (cpuset_cpus_allowed_fallback(p)) {
3395 * XXX When called from select_task_rq() we only
3396 * hold p->pi_lock and again violate locking order.
3398 * More yuck to audit.
3400 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3410 if (state != cpuset) {
3412 * Don't tell them about moving exiting tasks or
3413 * kernel threads (both mm NULL), since they never
3416 if (p->mm && printk_ratelimit()) {
3417 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3418 task_pid_nr(p), p->comm, cpu);
3426 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3429 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3431 lockdep_assert_held(&p->pi_lock);
3433 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3434 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3436 cpu = cpumask_any(p->cpus_ptr);
3439 * In order not to call set_task_cpu() on a blocking task we need
3440 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3443 * Since this is common to all placement strategies, this lives here.
3445 * [ this allows ->select_task() to simply return task_cpu(p) and
3446 * not worry about this generic constraint ]
3448 if (unlikely(!is_cpu_allowed(p, cpu)))
3449 cpu = select_fallback_rq(task_cpu(p), p);
3454 void sched_set_stop_task(int cpu, struct task_struct *stop)
3456 static struct lock_class_key stop_pi_lock;
3457 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3458 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3462 * Make it appear like a SCHED_FIFO task, its something
3463 * userspace knows about and won't get confused about.
3465 * Also, it will make PI more or less work without too
3466 * much confusion -- but then, stop work should not
3467 * rely on PI working anyway.
3469 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3471 stop->sched_class = &stop_sched_class;
3474 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3475 * adjust the effective priority of a task. As a result,
3476 * rt_mutex_setprio() can trigger (RT) balancing operations,
3477 * which can then trigger wakeups of the stop thread to push
3478 * around the current task.
3480 * The stop task itself will never be part of the PI-chain, it
3481 * never blocks, therefore that ->pi_lock recursion is safe.
3482 * Tell lockdep about this by placing the stop->pi_lock in its
3485 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3488 cpu_rq(cpu)->stop = stop;
3492 * Reset it back to a normal scheduling class so that
3493 * it can die in pieces.
3495 old_stop->sched_class = &rt_sched_class;
3499 #else /* CONFIG_SMP */
3501 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3502 const struct cpumask *new_mask,
3505 return set_cpus_allowed_ptr(p, new_mask);
3508 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3510 static inline bool rq_has_pinned_tasks(struct rq *rq)
3515 #endif /* !CONFIG_SMP */
3518 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3522 if (!schedstat_enabled())
3528 if (cpu == rq->cpu) {
3529 __schedstat_inc(rq->ttwu_local);
3530 __schedstat_inc(p->stats.nr_wakeups_local);
3532 struct sched_domain *sd;
3534 __schedstat_inc(p->stats.nr_wakeups_remote);
3536 for_each_domain(rq->cpu, sd) {
3537 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3538 __schedstat_inc(sd->ttwu_wake_remote);
3545 if (wake_flags & WF_MIGRATED)
3546 __schedstat_inc(p->stats.nr_wakeups_migrate);
3547 #endif /* CONFIG_SMP */
3549 __schedstat_inc(rq->ttwu_count);
3550 __schedstat_inc(p->stats.nr_wakeups);
3552 if (wake_flags & WF_SYNC)
3553 __schedstat_inc(p->stats.nr_wakeups_sync);
3557 * Mark the task runnable and perform wakeup-preemption.
3559 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3560 struct rq_flags *rf)
3562 check_preempt_curr(rq, p, wake_flags);
3563 WRITE_ONCE(p->__state, TASK_RUNNING);
3564 trace_sched_wakeup(p);
3567 if (p->sched_class->task_woken) {
3569 * Our task @p is fully woken up and running; so it's safe to
3570 * drop the rq->lock, hereafter rq is only used for statistics.
3572 rq_unpin_lock(rq, rf);
3573 p->sched_class->task_woken(rq, p);
3574 rq_repin_lock(rq, rf);
3577 if (rq->idle_stamp) {
3578 u64 delta = rq_clock(rq) - rq->idle_stamp;
3579 u64 max = 2*rq->max_idle_balance_cost;
3581 update_avg(&rq->avg_idle, delta);
3583 if (rq->avg_idle > max)
3586 rq->wake_stamp = jiffies;
3587 rq->wake_avg_idle = rq->avg_idle / 2;
3595 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3596 struct rq_flags *rf)
3598 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3600 lockdep_assert_rq_held(rq);
3602 if (p->sched_contributes_to_load)
3603 rq->nr_uninterruptible--;
3606 if (wake_flags & WF_MIGRATED)
3607 en_flags |= ENQUEUE_MIGRATED;
3611 delayacct_blkio_end(p);
3612 atomic_dec(&task_rq(p)->nr_iowait);
3615 activate_task(rq, p, en_flags);
3616 ttwu_do_wakeup(rq, p, wake_flags, rf);
3620 * Consider @p being inside a wait loop:
3623 * set_current_state(TASK_UNINTERRUPTIBLE);
3630 * __set_current_state(TASK_RUNNING);
3632 * between set_current_state() and schedule(). In this case @p is still
3633 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3636 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3637 * then schedule() must still happen and p->state can be changed to
3638 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3639 * need to do a full wakeup with enqueue.
3641 * Returns: %true when the wakeup is done,
3644 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3650 rq = __task_rq_lock(p, &rf);
3651 if (task_on_rq_queued(p)) {
3652 /* check_preempt_curr() may use rq clock */
3653 update_rq_clock(rq);
3654 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3657 __task_rq_unlock(rq, &rf);
3663 void sched_ttwu_pending(void *arg)
3665 struct llist_node *llist = arg;
3666 struct rq *rq = this_rq();
3667 struct task_struct *p, *t;
3674 * rq::ttwu_pending racy indication of out-standing wakeups.
3675 * Races such that false-negatives are possible, since they
3676 * are shorter lived that false-positives would be.
3678 WRITE_ONCE(rq->ttwu_pending, 0);
3680 rq_lock_irqsave(rq, &rf);
3681 update_rq_clock(rq);
3683 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3684 if (WARN_ON_ONCE(p->on_cpu))
3685 smp_cond_load_acquire(&p->on_cpu, !VAL);
3687 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3688 set_task_cpu(p, cpu_of(rq));
3690 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3693 rq_unlock_irqrestore(rq, &rf);
3696 void send_call_function_single_ipi(int cpu)
3698 struct rq *rq = cpu_rq(cpu);
3700 if (!set_nr_if_polling(rq->idle))
3701 arch_send_call_function_single_ipi(cpu);
3703 trace_sched_wake_idle_without_ipi(cpu);
3707 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3708 * necessary. The wakee CPU on receipt of the IPI will queue the task
3709 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3710 * of the wakeup instead of the waker.
3712 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3714 struct rq *rq = cpu_rq(cpu);
3716 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3718 WRITE_ONCE(rq->ttwu_pending, 1);
3719 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3722 void wake_up_if_idle(int cpu)
3724 struct rq *rq = cpu_rq(cpu);
3729 if (!is_idle_task(rcu_dereference(rq->curr)))
3732 rq_lock_irqsave(rq, &rf);
3733 if (is_idle_task(rq->curr))
3735 /* Else CPU is not idle, do nothing here: */
3736 rq_unlock_irqrestore(rq, &rf);
3742 bool cpus_share_cache(int this_cpu, int that_cpu)
3744 if (this_cpu == that_cpu)
3747 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3750 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3753 * Do not complicate things with the async wake_list while the CPU is
3756 if (!cpu_active(cpu))
3760 * If the CPU does not share cache, then queue the task on the
3761 * remote rqs wakelist to avoid accessing remote data.
3763 if (!cpus_share_cache(smp_processor_id(), cpu))
3767 * If the task is descheduling and the only running task on the
3768 * CPU then use the wakelist to offload the task activation to
3769 * the soon-to-be-idle CPU as the current CPU is likely busy.
3770 * nr_running is checked to avoid unnecessary task stacking.
3772 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3778 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3780 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3781 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3784 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3785 __ttwu_queue_wakelist(p, cpu, wake_flags);
3792 #else /* !CONFIG_SMP */
3794 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3799 #endif /* CONFIG_SMP */
3801 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3803 struct rq *rq = cpu_rq(cpu);
3806 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3810 update_rq_clock(rq);
3811 ttwu_do_activate(rq, p, wake_flags, &rf);
3816 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3818 * The caller holds p::pi_lock if p != current or has preemption
3819 * disabled when p == current.
3821 * The rules of PREEMPT_RT saved_state:
3823 * The related locking code always holds p::pi_lock when updating
3824 * p::saved_state, which means the code is fully serialized in both cases.
3826 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3827 * bits set. This allows to distinguish all wakeup scenarios.
3829 static __always_inline
3830 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3832 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3833 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3834 state != TASK_RTLOCK_WAIT);
3837 if (READ_ONCE(p->__state) & state) {
3842 #ifdef CONFIG_PREEMPT_RT
3844 * Saved state preserves the task state across blocking on
3845 * an RT lock. If the state matches, set p::saved_state to
3846 * TASK_RUNNING, but do not wake the task because it waits
3847 * for a lock wakeup. Also indicate success because from
3848 * the regular waker's point of view this has succeeded.
3850 * After acquiring the lock the task will restore p::__state
3851 * from p::saved_state which ensures that the regular
3852 * wakeup is not lost. The restore will also set
3853 * p::saved_state to TASK_RUNNING so any further tests will
3854 * not result in false positives vs. @success
3856 if (p->saved_state & state) {
3857 p->saved_state = TASK_RUNNING;
3865 * Notes on Program-Order guarantees on SMP systems.
3869 * The basic program-order guarantee on SMP systems is that when a task [t]
3870 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3871 * execution on its new CPU [c1].
3873 * For migration (of runnable tasks) this is provided by the following means:
3875 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3876 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3877 * rq(c1)->lock (if not at the same time, then in that order).
3878 * C) LOCK of the rq(c1)->lock scheduling in task
3880 * Release/acquire chaining guarantees that B happens after A and C after B.
3881 * Note: the CPU doing B need not be c0 or c1
3890 * UNLOCK rq(0)->lock
3892 * LOCK rq(0)->lock // orders against CPU0
3894 * UNLOCK rq(0)->lock
3898 * UNLOCK rq(1)->lock
3900 * LOCK rq(1)->lock // orders against CPU2
3903 * UNLOCK rq(1)->lock
3906 * BLOCKING -- aka. SLEEP + WAKEUP
3908 * For blocking we (obviously) need to provide the same guarantee as for
3909 * migration. However the means are completely different as there is no lock
3910 * chain to provide order. Instead we do:
3912 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3913 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3917 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3919 * LOCK rq(0)->lock LOCK X->pi_lock
3922 * smp_store_release(X->on_cpu, 0);
3924 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3930 * X->state = RUNNING
3931 * UNLOCK rq(2)->lock
3933 * LOCK rq(2)->lock // orders against CPU1
3936 * UNLOCK rq(2)->lock
3939 * UNLOCK rq(0)->lock
3942 * However, for wakeups there is a second guarantee we must provide, namely we
3943 * must ensure that CONDITION=1 done by the caller can not be reordered with
3944 * accesses to the task state; see try_to_wake_up() and set_current_state().
3948 * try_to_wake_up - wake up a thread
3949 * @p: the thread to be awakened
3950 * @state: the mask of task states that can be woken
3951 * @wake_flags: wake modifier flags (WF_*)
3953 * Conceptually does:
3955 * If (@state & @p->state) @p->state = TASK_RUNNING.
3957 * If the task was not queued/runnable, also place it back on a runqueue.
3959 * This function is atomic against schedule() which would dequeue the task.
3961 * It issues a full memory barrier before accessing @p->state, see the comment
3962 * with set_current_state().
3964 * Uses p->pi_lock to serialize against concurrent wake-ups.
3966 * Relies on p->pi_lock stabilizing:
3969 * - p->sched_task_group
3970 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3972 * Tries really hard to only take one task_rq(p)->lock for performance.
3973 * Takes rq->lock in:
3974 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3975 * - ttwu_queue() -- new rq, for enqueue of the task;
3976 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3978 * As a consequence we race really badly with just about everything. See the
3979 * many memory barriers and their comments for details.
3981 * Return: %true if @p->state changes (an actual wakeup was done),
3985 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3987 unsigned long flags;
3988 int cpu, success = 0;
3993 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3994 * == smp_processor_id()'. Together this means we can special
3995 * case the whole 'p->on_rq && ttwu_runnable()' case below
3996 * without taking any locks.
3999 * - we rely on Program-Order guarantees for all the ordering,
4000 * - we're serialized against set_special_state() by virtue of
4001 * it disabling IRQs (this allows not taking ->pi_lock).
4003 if (!ttwu_state_match(p, state, &success))
4006 trace_sched_waking(p);
4007 WRITE_ONCE(p->__state, TASK_RUNNING);
4008 trace_sched_wakeup(p);
4013 * If we are going to wake up a thread waiting for CONDITION we
4014 * need to ensure that CONDITION=1 done by the caller can not be
4015 * reordered with p->state check below. This pairs with smp_store_mb()
4016 * in set_current_state() that the waiting thread does.
4018 raw_spin_lock_irqsave(&p->pi_lock, flags);
4019 smp_mb__after_spinlock();
4020 if (!ttwu_state_match(p, state, &success))
4023 trace_sched_waking(p);
4026 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4027 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4028 * in smp_cond_load_acquire() below.
4030 * sched_ttwu_pending() try_to_wake_up()
4031 * STORE p->on_rq = 1 LOAD p->state
4034 * __schedule() (switch to task 'p')
4035 * LOCK rq->lock smp_rmb();
4036 * smp_mb__after_spinlock();
4040 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4042 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4043 * __schedule(). See the comment for smp_mb__after_spinlock().
4045 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4048 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4053 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4054 * possible to, falsely, observe p->on_cpu == 0.
4056 * One must be running (->on_cpu == 1) in order to remove oneself
4057 * from the runqueue.
4059 * __schedule() (switch to task 'p') try_to_wake_up()
4060 * STORE p->on_cpu = 1 LOAD p->on_rq
4063 * __schedule() (put 'p' to sleep)
4064 * LOCK rq->lock smp_rmb();
4065 * smp_mb__after_spinlock();
4066 * STORE p->on_rq = 0 LOAD p->on_cpu
4068 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4069 * __schedule(). See the comment for smp_mb__after_spinlock().
4071 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4072 * schedule()'s deactivate_task() has 'happened' and p will no longer
4073 * care about it's own p->state. See the comment in __schedule().
4075 smp_acquire__after_ctrl_dep();
4078 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4079 * == 0), which means we need to do an enqueue, change p->state to
4080 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4081 * enqueue, such as ttwu_queue_wakelist().
4083 WRITE_ONCE(p->__state, TASK_WAKING);
4086 * If the owning (remote) CPU is still in the middle of schedule() with
4087 * this task as prev, considering queueing p on the remote CPUs wake_list
4088 * which potentially sends an IPI instead of spinning on p->on_cpu to
4089 * let the waker make forward progress. This is safe because IRQs are
4090 * disabled and the IPI will deliver after on_cpu is cleared.
4092 * Ensure we load task_cpu(p) after p->on_cpu:
4094 * set_task_cpu(p, cpu);
4095 * STORE p->cpu = @cpu
4096 * __schedule() (switch to task 'p')
4098 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4099 * STORE p->on_cpu = 1 LOAD p->cpu
4101 * to ensure we observe the correct CPU on which the task is currently
4104 if (smp_load_acquire(&p->on_cpu) &&
4105 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4109 * If the owning (remote) CPU is still in the middle of schedule() with
4110 * this task as prev, wait until it's done referencing the task.
4112 * Pairs with the smp_store_release() in finish_task().
4114 * This ensures that tasks getting woken will be fully ordered against
4115 * their previous state and preserve Program Order.
4117 smp_cond_load_acquire(&p->on_cpu, !VAL);
4119 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4120 if (task_cpu(p) != cpu) {
4122 delayacct_blkio_end(p);
4123 atomic_dec(&task_rq(p)->nr_iowait);
4126 wake_flags |= WF_MIGRATED;
4127 psi_ttwu_dequeue(p);
4128 set_task_cpu(p, cpu);
4132 #endif /* CONFIG_SMP */
4134 ttwu_queue(p, cpu, wake_flags);
4136 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4139 ttwu_stat(p, task_cpu(p), wake_flags);
4146 * task_call_func - Invoke a function on task in fixed state
4147 * @p: Process for which the function is to be invoked, can be @current.
4148 * @func: Function to invoke.
4149 * @arg: Argument to function.
4151 * Fix the task in it's current state by avoiding wakeups and or rq operations
4152 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4153 * to work out what the state is, if required. Given that @func can be invoked
4154 * with a runqueue lock held, it had better be quite lightweight.
4157 * Whatever @func returns
4159 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4161 struct rq *rq = NULL;
4166 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4168 state = READ_ONCE(p->__state);
4171 * Ensure we load p->on_rq after p->__state, otherwise it would be
4172 * possible to, falsely, observe p->on_rq == 0.
4174 * See try_to_wake_up() for a longer comment.
4179 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4180 * the task is blocked. Make sure to check @state since ttwu() can drop
4181 * locks at the end, see ttwu_queue_wakelist().
4183 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4184 rq = __task_rq_lock(p, &rf);
4187 * At this point the task is pinned; either:
4188 * - blocked and we're holding off wakeups (pi->lock)
4189 * - woken, and we're holding off enqueue (rq->lock)
4190 * - queued, and we're holding off schedule (rq->lock)
4191 * - running, and we're holding off de-schedule (rq->lock)
4193 * The called function (@func) can use: task_curr(), p->on_rq and
4194 * p->__state to differentiate between these states.
4201 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4206 * wake_up_process - Wake up a specific process
4207 * @p: The process to be woken up.
4209 * Attempt to wake up the nominated process and move it to the set of runnable
4212 * Return: 1 if the process was woken up, 0 if it was already running.
4214 * This function executes a full memory barrier before accessing the task state.
4216 int wake_up_process(struct task_struct *p)
4218 return try_to_wake_up(p, TASK_NORMAL, 0);
4220 EXPORT_SYMBOL(wake_up_process);
4222 int wake_up_state(struct task_struct *p, unsigned int state)
4224 return try_to_wake_up(p, state, 0);
4228 * Perform scheduler related setup for a newly forked process p.
4229 * p is forked by current.
4231 * __sched_fork() is basic setup used by init_idle() too:
4233 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4238 p->se.exec_start = 0;
4239 p->se.sum_exec_runtime = 0;
4240 p->se.prev_sum_exec_runtime = 0;
4241 p->se.nr_migrations = 0;
4243 INIT_LIST_HEAD(&p->se.group_node);
4245 #ifdef CONFIG_FAIR_GROUP_SCHED
4246 p->se.cfs_rq = NULL;
4249 #ifdef CONFIG_SCHEDSTATS
4250 /* Even if schedstat is disabled, there should not be garbage */
4251 memset(&p->stats, 0, sizeof(p->stats));
4254 RB_CLEAR_NODE(&p->dl.rb_node);
4255 init_dl_task_timer(&p->dl);
4256 init_dl_inactive_task_timer(&p->dl);
4257 __dl_clear_params(p);
4259 INIT_LIST_HEAD(&p->rt.run_list);
4261 p->rt.time_slice = sched_rr_timeslice;
4265 #ifdef CONFIG_PREEMPT_NOTIFIERS
4266 INIT_HLIST_HEAD(&p->preempt_notifiers);
4269 #ifdef CONFIG_COMPACTION
4270 p->capture_control = NULL;
4272 init_numa_balancing(clone_flags, p);
4274 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4275 p->migration_pending = NULL;
4279 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4281 #ifdef CONFIG_NUMA_BALANCING
4283 void set_numabalancing_state(bool enabled)
4286 static_branch_enable(&sched_numa_balancing);
4288 static_branch_disable(&sched_numa_balancing);
4291 #ifdef CONFIG_PROC_SYSCTL
4292 int sysctl_numa_balancing(struct ctl_table *table, int write,
4293 void *buffer, size_t *lenp, loff_t *ppos)
4297 int state = static_branch_likely(&sched_numa_balancing);
4299 if (write && !capable(CAP_SYS_ADMIN))
4304 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4308 set_numabalancing_state(state);
4314 #ifdef CONFIG_SCHEDSTATS
4316 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4318 static void set_schedstats(bool enabled)
4321 static_branch_enable(&sched_schedstats);
4323 static_branch_disable(&sched_schedstats);
4326 void force_schedstat_enabled(void)
4328 if (!schedstat_enabled()) {
4329 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4330 static_branch_enable(&sched_schedstats);
4334 static int __init setup_schedstats(char *str)
4340 if (!strcmp(str, "enable")) {
4341 set_schedstats(true);
4343 } else if (!strcmp(str, "disable")) {
4344 set_schedstats(false);
4349 pr_warn("Unable to parse schedstats=\n");
4353 __setup("schedstats=", setup_schedstats);
4355 #ifdef CONFIG_PROC_SYSCTL
4356 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4357 size_t *lenp, loff_t *ppos)
4361 int state = static_branch_likely(&sched_schedstats);
4363 if (write && !capable(CAP_SYS_ADMIN))
4368 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4372 set_schedstats(state);
4375 #endif /* CONFIG_PROC_SYSCTL */
4376 #endif /* CONFIG_SCHEDSTATS */
4379 * fork()/clone()-time setup:
4381 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4383 __sched_fork(clone_flags, p);
4385 * We mark the process as NEW here. This guarantees that
4386 * nobody will actually run it, and a signal or other external
4387 * event cannot wake it up and insert it on the runqueue either.
4389 p->__state = TASK_NEW;
4392 * Make sure we do not leak PI boosting priority to the child.
4394 p->prio = current->normal_prio;
4399 * Revert to default priority/policy on fork if requested.
4401 if (unlikely(p->sched_reset_on_fork)) {
4402 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4403 p->policy = SCHED_NORMAL;
4404 p->static_prio = NICE_TO_PRIO(0);
4406 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4407 p->static_prio = NICE_TO_PRIO(0);
4409 p->prio = p->normal_prio = p->static_prio;
4410 set_load_weight(p, false);
4413 * We don't need the reset flag anymore after the fork. It has
4414 * fulfilled its duty:
4416 p->sched_reset_on_fork = 0;
4419 if (dl_prio(p->prio))
4421 else if (rt_prio(p->prio))
4422 p->sched_class = &rt_sched_class;
4424 p->sched_class = &fair_sched_class;
4426 init_entity_runnable_average(&p->se);
4429 #ifdef CONFIG_SCHED_INFO
4430 if (likely(sched_info_on()))
4431 memset(&p->sched_info, 0, sizeof(p->sched_info));
4433 #if defined(CONFIG_SMP)
4436 init_task_preempt_count(p);
4438 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4439 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4444 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4446 unsigned long flags;
4449 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4450 * required yet, but lockdep gets upset if rules are violated.
4452 raw_spin_lock_irqsave(&p->pi_lock, flags);
4453 #ifdef CONFIG_CGROUP_SCHED
4455 struct task_group *tg;
4456 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4457 struct task_group, css);
4458 tg = autogroup_task_group(p, tg);
4459 p->sched_task_group = tg;
4464 * We're setting the CPU for the first time, we don't migrate,
4465 * so use __set_task_cpu().
4467 __set_task_cpu(p, smp_processor_id());
4468 if (p->sched_class->task_fork)
4469 p->sched_class->task_fork(p);
4470 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4473 void sched_post_fork(struct task_struct *p)
4475 uclamp_post_fork(p);
4478 unsigned long to_ratio(u64 period, u64 runtime)
4480 if (runtime == RUNTIME_INF)
4484 * Doing this here saves a lot of checks in all
4485 * the calling paths, and returning zero seems
4486 * safe for them anyway.
4491 return div64_u64(runtime << BW_SHIFT, period);
4495 * wake_up_new_task - wake up a newly created task for the first time.
4497 * This function will do some initial scheduler statistics housekeeping
4498 * that must be done for every newly created context, then puts the task
4499 * on the runqueue and wakes it.
4501 void wake_up_new_task(struct task_struct *p)
4506 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4507 WRITE_ONCE(p->__state, TASK_RUNNING);
4510 * Fork balancing, do it here and not earlier because:
4511 * - cpus_ptr can change in the fork path
4512 * - any previously selected CPU might disappear through hotplug
4514 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4515 * as we're not fully set-up yet.
4517 p->recent_used_cpu = task_cpu(p);
4519 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4521 rq = __task_rq_lock(p, &rf);
4522 update_rq_clock(rq);
4523 post_init_entity_util_avg(p);
4525 activate_task(rq, p, ENQUEUE_NOCLOCK);
4526 trace_sched_wakeup_new(p);
4527 check_preempt_curr(rq, p, WF_FORK);
4529 if (p->sched_class->task_woken) {
4531 * Nothing relies on rq->lock after this, so it's fine to
4534 rq_unpin_lock(rq, &rf);
4535 p->sched_class->task_woken(rq, p);
4536 rq_repin_lock(rq, &rf);
4539 task_rq_unlock(rq, p, &rf);
4542 #ifdef CONFIG_PREEMPT_NOTIFIERS
4544 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4546 void preempt_notifier_inc(void)
4548 static_branch_inc(&preempt_notifier_key);
4550 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4552 void preempt_notifier_dec(void)
4554 static_branch_dec(&preempt_notifier_key);
4556 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4559 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4560 * @notifier: notifier struct to register
4562 void preempt_notifier_register(struct preempt_notifier *notifier)
4564 if (!static_branch_unlikely(&preempt_notifier_key))
4565 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4567 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4569 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4572 * preempt_notifier_unregister - no longer interested in preemption notifications
4573 * @notifier: notifier struct to unregister
4575 * This is *not* safe to call from within a preemption notifier.
4577 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4579 hlist_del(¬ifier->link);
4581 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4583 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4585 struct preempt_notifier *notifier;
4587 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4588 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4591 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4593 if (static_branch_unlikely(&preempt_notifier_key))
4594 __fire_sched_in_preempt_notifiers(curr);
4598 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4599 struct task_struct *next)
4601 struct preempt_notifier *notifier;
4603 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4604 notifier->ops->sched_out(notifier, next);
4607 static __always_inline void
4608 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4609 struct task_struct *next)
4611 if (static_branch_unlikely(&preempt_notifier_key))
4612 __fire_sched_out_preempt_notifiers(curr, next);
4615 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4617 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4622 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4623 struct task_struct *next)
4627 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4629 static inline void prepare_task(struct task_struct *next)
4633 * Claim the task as running, we do this before switching to it
4634 * such that any running task will have this set.
4636 * See the ttwu() WF_ON_CPU case and its ordering comment.
4638 WRITE_ONCE(next->on_cpu, 1);
4642 static inline void finish_task(struct task_struct *prev)
4646 * This must be the very last reference to @prev from this CPU. After
4647 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4648 * must ensure this doesn't happen until the switch is completely
4651 * In particular, the load of prev->state in finish_task_switch() must
4652 * happen before this.
4654 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4656 smp_store_release(&prev->on_cpu, 0);
4662 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4664 void (*func)(struct rq *rq);
4665 struct callback_head *next;
4667 lockdep_assert_rq_held(rq);
4670 func = (void (*)(struct rq *))head->func;
4679 static void balance_push(struct rq *rq);
4681 struct callback_head balance_push_callback = {
4683 .func = (void (*)(struct callback_head *))balance_push,
4686 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4688 struct callback_head *head = rq->balance_callback;
4690 lockdep_assert_rq_held(rq);
4692 rq->balance_callback = NULL;
4697 static void __balance_callbacks(struct rq *rq)
4699 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4702 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4704 unsigned long flags;
4706 if (unlikely(head)) {
4707 raw_spin_rq_lock_irqsave(rq, flags);
4708 do_balance_callbacks(rq, head);
4709 raw_spin_rq_unlock_irqrestore(rq, flags);
4715 static inline void __balance_callbacks(struct rq *rq)
4719 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4724 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4731 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4734 * Since the runqueue lock will be released by the next
4735 * task (which is an invalid locking op but in the case
4736 * of the scheduler it's an obvious special-case), so we
4737 * do an early lockdep release here:
4739 rq_unpin_lock(rq, rf);
4740 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4741 #ifdef CONFIG_DEBUG_SPINLOCK
4742 /* this is a valid case when another task releases the spinlock */
4743 rq_lockp(rq)->owner = next;
4747 static inline void finish_lock_switch(struct rq *rq)
4750 * If we are tracking spinlock dependencies then we have to
4751 * fix up the runqueue lock - which gets 'carried over' from
4752 * prev into current:
4754 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4755 __balance_callbacks(rq);
4756 raw_spin_rq_unlock_irq(rq);
4760 * NOP if the arch has not defined these:
4763 #ifndef prepare_arch_switch
4764 # define prepare_arch_switch(next) do { } while (0)
4767 #ifndef finish_arch_post_lock_switch
4768 # define finish_arch_post_lock_switch() do { } while (0)
4771 static inline void kmap_local_sched_out(void)
4773 #ifdef CONFIG_KMAP_LOCAL
4774 if (unlikely(current->kmap_ctrl.idx))
4775 __kmap_local_sched_out();
4779 static inline void kmap_local_sched_in(void)
4781 #ifdef CONFIG_KMAP_LOCAL
4782 if (unlikely(current->kmap_ctrl.idx))
4783 __kmap_local_sched_in();
4788 * prepare_task_switch - prepare to switch tasks
4789 * @rq: the runqueue preparing to switch
4790 * @prev: the current task that is being switched out
4791 * @next: the task we are going to switch to.
4793 * This is called with the rq lock held and interrupts off. It must
4794 * be paired with a subsequent finish_task_switch after the context
4797 * prepare_task_switch sets up locking and calls architecture specific
4801 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4802 struct task_struct *next)
4804 kcov_prepare_switch(prev);
4805 sched_info_switch(rq, prev, next);
4806 perf_event_task_sched_out(prev, next);
4808 fire_sched_out_preempt_notifiers(prev, next);
4809 kmap_local_sched_out();
4811 prepare_arch_switch(next);
4815 * finish_task_switch - clean up after a task-switch
4816 * @prev: the thread we just switched away from.
4818 * finish_task_switch must be called after the context switch, paired
4819 * with a prepare_task_switch call before the context switch.
4820 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4821 * and do any other architecture-specific cleanup actions.
4823 * Note that we may have delayed dropping an mm in context_switch(). If
4824 * so, we finish that here outside of the runqueue lock. (Doing it
4825 * with the lock held can cause deadlocks; see schedule() for
4828 * The context switch have flipped the stack from under us and restored the
4829 * local variables which were saved when this task called schedule() in the
4830 * past. prev == current is still correct but we need to recalculate this_rq
4831 * because prev may have moved to another CPU.
4833 static struct rq *finish_task_switch(struct task_struct *prev)
4834 __releases(rq->lock)
4836 struct rq *rq = this_rq();
4837 struct mm_struct *mm = rq->prev_mm;
4841 * The previous task will have left us with a preempt_count of 2
4842 * because it left us after:
4845 * preempt_disable(); // 1
4847 * raw_spin_lock_irq(&rq->lock) // 2
4849 * Also, see FORK_PREEMPT_COUNT.
4851 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4852 "corrupted preempt_count: %s/%d/0x%x\n",
4853 current->comm, current->pid, preempt_count()))
4854 preempt_count_set(FORK_PREEMPT_COUNT);
4859 * A task struct has one reference for the use as "current".
4860 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4861 * schedule one last time. The schedule call will never return, and
4862 * the scheduled task must drop that reference.
4864 * We must observe prev->state before clearing prev->on_cpu (in
4865 * finish_task), otherwise a concurrent wakeup can get prev
4866 * running on another CPU and we could rave with its RUNNING -> DEAD
4867 * transition, resulting in a double drop.
4869 prev_state = READ_ONCE(prev->__state);
4870 vtime_task_switch(prev);
4871 perf_event_task_sched_in(prev, current);
4873 tick_nohz_task_switch();
4874 finish_lock_switch(rq);
4875 finish_arch_post_lock_switch();
4876 kcov_finish_switch(current);
4878 * kmap_local_sched_out() is invoked with rq::lock held and
4879 * interrupts disabled. There is no requirement for that, but the
4880 * sched out code does not have an interrupt enabled section.
4881 * Restoring the maps on sched in does not require interrupts being
4884 kmap_local_sched_in();
4886 fire_sched_in_preempt_notifiers(current);
4888 * When switching through a kernel thread, the loop in
4889 * membarrier_{private,global}_expedited() may have observed that
4890 * kernel thread and not issued an IPI. It is therefore possible to
4891 * schedule between user->kernel->user threads without passing though
4892 * switch_mm(). Membarrier requires a barrier after storing to
4893 * rq->curr, before returning to userspace, so provide them here:
4895 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4896 * provided by mmdrop(),
4897 * - a sync_core for SYNC_CORE.
4900 membarrier_mm_sync_core_before_usermode(mm);
4903 if (unlikely(prev_state == TASK_DEAD)) {
4904 if (prev->sched_class->task_dead)
4905 prev->sched_class->task_dead(prev);
4907 /* Task is done with its stack. */
4908 put_task_stack(prev);
4910 put_task_struct_rcu_user(prev);
4917 * schedule_tail - first thing a freshly forked thread must call.
4918 * @prev: the thread we just switched away from.
4920 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4921 __releases(rq->lock)
4924 * New tasks start with FORK_PREEMPT_COUNT, see there and
4925 * finish_task_switch() for details.
4927 * finish_task_switch() will drop rq->lock() and lower preempt_count
4928 * and the preempt_enable() will end up enabling preemption (on
4929 * PREEMPT_COUNT kernels).
4932 finish_task_switch(prev);
4935 if (current->set_child_tid)
4936 put_user(task_pid_vnr(current), current->set_child_tid);
4938 calculate_sigpending();
4942 * context_switch - switch to the new MM and the new thread's register state.
4944 static __always_inline struct rq *
4945 context_switch(struct rq *rq, struct task_struct *prev,
4946 struct task_struct *next, struct rq_flags *rf)
4948 prepare_task_switch(rq, prev, next);
4951 * For paravirt, this is coupled with an exit in switch_to to
4952 * combine the page table reload and the switch backend into
4955 arch_start_context_switch(prev);
4958 * kernel -> kernel lazy + transfer active
4959 * user -> kernel lazy + mmgrab() active
4961 * kernel -> user switch + mmdrop() active
4962 * user -> user switch
4964 if (!next->mm) { // to kernel
4965 enter_lazy_tlb(prev->active_mm, next);
4967 next->active_mm = prev->active_mm;
4968 if (prev->mm) // from user
4969 mmgrab(prev->active_mm);
4971 prev->active_mm = NULL;
4973 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4975 * sys_membarrier() requires an smp_mb() between setting
4976 * rq->curr / membarrier_switch_mm() and returning to userspace.
4978 * The below provides this either through switch_mm(), or in
4979 * case 'prev->active_mm == next->mm' through
4980 * finish_task_switch()'s mmdrop().
4982 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4984 if (!prev->mm) { // from kernel
4985 /* will mmdrop() in finish_task_switch(). */
4986 rq->prev_mm = prev->active_mm;
4987 prev->active_mm = NULL;
4991 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4993 prepare_lock_switch(rq, next, rf);
4995 /* Here we just switch the register state and the stack. */
4996 switch_to(prev, next, prev);
4999 return finish_task_switch(prev);
5003 * nr_running and nr_context_switches:
5005 * externally visible scheduler statistics: current number of runnable
5006 * threads, total number of context switches performed since bootup.
5008 unsigned int nr_running(void)
5010 unsigned int i, sum = 0;
5012 for_each_online_cpu(i)
5013 sum += cpu_rq(i)->nr_running;
5019 * Check if only the current task is running on the CPU.
5021 * Caution: this function does not check that the caller has disabled
5022 * preemption, thus the result might have a time-of-check-to-time-of-use
5023 * race. The caller is responsible to use it correctly, for example:
5025 * - from a non-preemptible section (of course)
5027 * - from a thread that is bound to a single CPU
5029 * - in a loop with very short iterations (e.g. a polling loop)
5031 bool single_task_running(void)
5033 return raw_rq()->nr_running == 1;
5035 EXPORT_SYMBOL(single_task_running);
5037 unsigned long long nr_context_switches(void)
5040 unsigned long long sum = 0;
5042 for_each_possible_cpu(i)
5043 sum += cpu_rq(i)->nr_switches;
5049 * Consumers of these two interfaces, like for example the cpuidle menu
5050 * governor, are using nonsensical data. Preferring shallow idle state selection
5051 * for a CPU that has IO-wait which might not even end up running the task when
5052 * it does become runnable.
5055 unsigned int nr_iowait_cpu(int cpu)
5057 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5061 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5063 * The idea behind IO-wait account is to account the idle time that we could
5064 * have spend running if it were not for IO. That is, if we were to improve the
5065 * storage performance, we'd have a proportional reduction in IO-wait time.
5067 * This all works nicely on UP, where, when a task blocks on IO, we account
5068 * idle time as IO-wait, because if the storage were faster, it could've been
5069 * running and we'd not be idle.
5071 * This has been extended to SMP, by doing the same for each CPU. This however
5074 * Imagine for instance the case where two tasks block on one CPU, only the one
5075 * CPU will have IO-wait accounted, while the other has regular idle. Even
5076 * though, if the storage were faster, both could've ran at the same time,
5077 * utilising both CPUs.
5079 * This means, that when looking globally, the current IO-wait accounting on
5080 * SMP is a lower bound, by reason of under accounting.
5082 * Worse, since the numbers are provided per CPU, they are sometimes
5083 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5084 * associated with any one particular CPU, it can wake to another CPU than it
5085 * blocked on. This means the per CPU IO-wait number is meaningless.
5087 * Task CPU affinities can make all that even more 'interesting'.
5090 unsigned int nr_iowait(void)
5092 unsigned int i, sum = 0;
5094 for_each_possible_cpu(i)
5095 sum += nr_iowait_cpu(i);
5103 * sched_exec - execve() is a valuable balancing opportunity, because at
5104 * this point the task has the smallest effective memory and cache footprint.
5106 void sched_exec(void)
5108 struct task_struct *p = current;
5109 unsigned long flags;
5112 raw_spin_lock_irqsave(&p->pi_lock, flags);
5113 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5114 if (dest_cpu == smp_processor_id())
5117 if (likely(cpu_active(dest_cpu))) {
5118 struct migration_arg arg = { p, dest_cpu };
5120 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5121 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5125 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5130 DEFINE_PER_CPU(struct kernel_stat, kstat);
5131 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5133 EXPORT_PER_CPU_SYMBOL(kstat);
5134 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5137 * The function fair_sched_class.update_curr accesses the struct curr
5138 * and its field curr->exec_start; when called from task_sched_runtime(),
5139 * we observe a high rate of cache misses in practice.
5140 * Prefetching this data results in improved performance.
5142 static inline void prefetch_curr_exec_start(struct task_struct *p)
5144 #ifdef CONFIG_FAIR_GROUP_SCHED
5145 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5147 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5150 prefetch(&curr->exec_start);
5154 * Return accounted runtime for the task.
5155 * In case the task is currently running, return the runtime plus current's
5156 * pending runtime that have not been accounted yet.
5158 unsigned long long task_sched_runtime(struct task_struct *p)
5164 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5166 * 64-bit doesn't need locks to atomically read a 64-bit value.
5167 * So we have a optimization chance when the task's delta_exec is 0.
5168 * Reading ->on_cpu is racy, but this is ok.
5170 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5171 * If we race with it entering CPU, unaccounted time is 0. This is
5172 * indistinguishable from the read occurring a few cycles earlier.
5173 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5174 * been accounted, so we're correct here as well.
5176 if (!p->on_cpu || !task_on_rq_queued(p))
5177 return p->se.sum_exec_runtime;
5180 rq = task_rq_lock(p, &rf);
5182 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5183 * project cycles that may never be accounted to this
5184 * thread, breaking clock_gettime().
5186 if (task_current(rq, p) && task_on_rq_queued(p)) {
5187 prefetch_curr_exec_start(p);
5188 update_rq_clock(rq);
5189 p->sched_class->update_curr(rq);
5191 ns = p->se.sum_exec_runtime;
5192 task_rq_unlock(rq, p, &rf);
5197 #ifdef CONFIG_SCHED_DEBUG
5198 static u64 cpu_resched_latency(struct rq *rq)
5200 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5201 u64 resched_latency, now = rq_clock(rq);
5202 static bool warned_once;
5204 if (sysctl_resched_latency_warn_once && warned_once)
5207 if (!need_resched() || !latency_warn_ms)
5210 if (system_state == SYSTEM_BOOTING)
5213 if (!rq->last_seen_need_resched_ns) {
5214 rq->last_seen_need_resched_ns = now;
5215 rq->ticks_without_resched = 0;
5219 rq->ticks_without_resched++;
5220 resched_latency = now - rq->last_seen_need_resched_ns;
5221 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5226 return resched_latency;
5229 static int __init setup_resched_latency_warn_ms(char *str)
5233 if ((kstrtol(str, 0, &val))) {
5234 pr_warn("Unable to set resched_latency_warn_ms\n");
5238 sysctl_resched_latency_warn_ms = val;
5241 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5243 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5244 #endif /* CONFIG_SCHED_DEBUG */
5247 * This function gets called by the timer code, with HZ frequency.
5248 * We call it with interrupts disabled.
5250 void scheduler_tick(void)
5252 int cpu = smp_processor_id();
5253 struct rq *rq = cpu_rq(cpu);
5254 struct task_struct *curr = rq->curr;
5256 unsigned long thermal_pressure;
5257 u64 resched_latency;
5259 arch_scale_freq_tick();
5264 update_rq_clock(rq);
5265 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5266 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5267 curr->sched_class->task_tick(rq, curr, 0);
5268 if (sched_feat(LATENCY_WARN))
5269 resched_latency = cpu_resched_latency(rq);
5270 calc_global_load_tick(rq);
5271 sched_core_tick(rq);
5275 if (sched_feat(LATENCY_WARN) && resched_latency)
5276 resched_latency_warn(cpu, resched_latency);
5278 perf_event_task_tick();
5281 rq->idle_balance = idle_cpu(cpu);
5282 trigger_load_balance(rq);
5286 #ifdef CONFIG_NO_HZ_FULL
5291 struct delayed_work work;
5293 /* Values for ->state, see diagram below. */
5294 #define TICK_SCHED_REMOTE_OFFLINE 0
5295 #define TICK_SCHED_REMOTE_OFFLINING 1
5296 #define TICK_SCHED_REMOTE_RUNNING 2
5299 * State diagram for ->state:
5302 * TICK_SCHED_REMOTE_OFFLINE
5305 * | | sched_tick_remote()
5308 * +--TICK_SCHED_REMOTE_OFFLINING
5311 * sched_tick_start() | | sched_tick_stop()
5314 * TICK_SCHED_REMOTE_RUNNING
5317 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5318 * and sched_tick_start() are happy to leave the state in RUNNING.
5321 static struct tick_work __percpu *tick_work_cpu;
5323 static void sched_tick_remote(struct work_struct *work)
5325 struct delayed_work *dwork = to_delayed_work(work);
5326 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5327 int cpu = twork->cpu;
5328 struct rq *rq = cpu_rq(cpu);
5329 struct task_struct *curr;
5335 * Handle the tick only if it appears the remote CPU is running in full
5336 * dynticks mode. The check is racy by nature, but missing a tick or
5337 * having one too much is no big deal because the scheduler tick updates
5338 * statistics and checks timeslices in a time-independent way, regardless
5339 * of when exactly it is running.
5341 if (!tick_nohz_tick_stopped_cpu(cpu))
5344 rq_lock_irq(rq, &rf);
5346 if (cpu_is_offline(cpu))
5349 update_rq_clock(rq);
5351 if (!is_idle_task(curr)) {
5353 * Make sure the next tick runs within a reasonable
5356 delta = rq_clock_task(rq) - curr->se.exec_start;
5357 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5359 curr->sched_class->task_tick(rq, curr, 0);
5361 calc_load_nohz_remote(rq);
5363 rq_unlock_irq(rq, &rf);
5367 * Run the remote tick once per second (1Hz). This arbitrary
5368 * frequency is large enough to avoid overload but short enough
5369 * to keep scheduler internal stats reasonably up to date. But
5370 * first update state to reflect hotplug activity if required.
5372 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5373 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5374 if (os == TICK_SCHED_REMOTE_RUNNING)
5375 queue_delayed_work(system_unbound_wq, dwork, HZ);
5378 static void sched_tick_start(int cpu)
5381 struct tick_work *twork;
5383 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5386 WARN_ON_ONCE(!tick_work_cpu);
5388 twork = per_cpu_ptr(tick_work_cpu, cpu);
5389 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5390 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5391 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5393 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5394 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5398 #ifdef CONFIG_HOTPLUG_CPU
5399 static void sched_tick_stop(int cpu)
5401 struct tick_work *twork;
5404 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5407 WARN_ON_ONCE(!tick_work_cpu);
5409 twork = per_cpu_ptr(tick_work_cpu, cpu);
5410 /* There cannot be competing actions, but don't rely on stop-machine. */
5411 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5412 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5413 /* Don't cancel, as this would mess up the state machine. */
5415 #endif /* CONFIG_HOTPLUG_CPU */
5417 int __init sched_tick_offload_init(void)
5419 tick_work_cpu = alloc_percpu(struct tick_work);
5420 BUG_ON(!tick_work_cpu);
5424 #else /* !CONFIG_NO_HZ_FULL */
5425 static inline void sched_tick_start(int cpu) { }
5426 static inline void sched_tick_stop(int cpu) { }
5429 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5430 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5432 * If the value passed in is equal to the current preempt count
5433 * then we just disabled preemption. Start timing the latency.
5435 static inline void preempt_latency_start(int val)
5437 if (preempt_count() == val) {
5438 unsigned long ip = get_lock_parent_ip();
5439 #ifdef CONFIG_DEBUG_PREEMPT
5440 current->preempt_disable_ip = ip;
5442 trace_preempt_off(CALLER_ADDR0, ip);
5446 void preempt_count_add(int val)
5448 #ifdef CONFIG_DEBUG_PREEMPT
5452 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5455 __preempt_count_add(val);
5456 #ifdef CONFIG_DEBUG_PREEMPT
5458 * Spinlock count overflowing soon?
5460 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5463 preempt_latency_start(val);
5465 EXPORT_SYMBOL(preempt_count_add);
5466 NOKPROBE_SYMBOL(preempt_count_add);
5469 * If the value passed in equals to the current preempt count
5470 * then we just enabled preemption. Stop timing the latency.
5472 static inline void preempt_latency_stop(int val)
5474 if (preempt_count() == val)
5475 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5478 void preempt_count_sub(int val)
5480 #ifdef CONFIG_DEBUG_PREEMPT
5484 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5487 * Is the spinlock portion underflowing?
5489 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5490 !(preempt_count() & PREEMPT_MASK)))
5494 preempt_latency_stop(val);
5495 __preempt_count_sub(val);
5497 EXPORT_SYMBOL(preempt_count_sub);
5498 NOKPROBE_SYMBOL(preempt_count_sub);
5501 static inline void preempt_latency_start(int val) { }
5502 static inline void preempt_latency_stop(int val) { }
5505 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5507 #ifdef CONFIG_DEBUG_PREEMPT
5508 return p->preempt_disable_ip;
5515 * Print scheduling while atomic bug:
5517 static noinline void __schedule_bug(struct task_struct *prev)
5519 /* Save this before calling printk(), since that will clobber it */
5520 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5522 if (oops_in_progress)
5525 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5526 prev->comm, prev->pid, preempt_count());
5528 debug_show_held_locks(prev);
5530 if (irqs_disabled())
5531 print_irqtrace_events(prev);
5532 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5533 && in_atomic_preempt_off()) {
5534 pr_err("Preemption disabled at:");
5535 print_ip_sym(KERN_ERR, preempt_disable_ip);
5538 panic("scheduling while atomic\n");
5541 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5545 * Various schedule()-time debugging checks and statistics:
5547 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5549 #ifdef CONFIG_SCHED_STACK_END_CHECK
5550 if (task_stack_end_corrupted(prev))
5551 panic("corrupted stack end detected inside scheduler\n");
5553 if (task_scs_end_corrupted(prev))
5554 panic("corrupted shadow stack detected inside scheduler\n");
5557 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5558 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5559 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5560 prev->comm, prev->pid, prev->non_block_count);
5562 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5566 if (unlikely(in_atomic_preempt_off())) {
5567 __schedule_bug(prev);
5568 preempt_count_set(PREEMPT_DISABLED);
5571 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5573 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5575 schedstat_inc(this_rq()->sched_count);
5578 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5579 struct rq_flags *rf)
5582 const struct sched_class *class;
5584 * We must do the balancing pass before put_prev_task(), such
5585 * that when we release the rq->lock the task is in the same
5586 * state as before we took rq->lock.
5588 * We can terminate the balance pass as soon as we know there is
5589 * a runnable task of @class priority or higher.
5591 for_class_range(class, prev->sched_class, &idle_sched_class) {
5592 if (class->balance(rq, prev, rf))
5597 put_prev_task(rq, prev);
5601 * Pick up the highest-prio task:
5603 static inline struct task_struct *
5604 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5606 const struct sched_class *class;
5607 struct task_struct *p;
5610 * Optimization: we know that if all tasks are in the fair class we can
5611 * call that function directly, but only if the @prev task wasn't of a
5612 * higher scheduling class, because otherwise those lose the
5613 * opportunity to pull in more work from other CPUs.
5615 if (likely(prev->sched_class <= &fair_sched_class &&
5616 rq->nr_running == rq->cfs.h_nr_running)) {
5618 p = pick_next_task_fair(rq, prev, rf);
5619 if (unlikely(p == RETRY_TASK))
5622 /* Assume the next prioritized class is idle_sched_class */
5624 put_prev_task(rq, prev);
5625 p = pick_next_task_idle(rq);
5632 put_prev_task_balance(rq, prev, rf);
5634 for_each_class(class) {
5635 p = class->pick_next_task(rq);
5640 BUG(); /* The idle class should always have a runnable task. */
5643 #ifdef CONFIG_SCHED_CORE
5644 static inline bool is_task_rq_idle(struct task_struct *t)
5646 return (task_rq(t)->idle == t);
5649 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5651 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5654 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5656 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5659 return a->core_cookie == b->core_cookie;
5662 static inline struct task_struct *pick_task(struct rq *rq)
5664 const struct sched_class *class;
5665 struct task_struct *p;
5667 for_each_class(class) {
5668 p = class->pick_task(rq);
5673 BUG(); /* The idle class should always have a runnable task. */
5676 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5678 static void queue_core_balance(struct rq *rq);
5680 static struct task_struct *
5681 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5683 struct task_struct *next, *p, *max = NULL;
5684 const struct cpumask *smt_mask;
5685 bool fi_before = false;
5686 bool core_clock_updated = (rq == rq->core);
5687 unsigned long cookie;
5688 int i, cpu, occ = 0;
5692 if (!sched_core_enabled(rq))
5693 return __pick_next_task(rq, prev, rf);
5697 /* Stopper task is switching into idle, no need core-wide selection. */
5698 if (cpu_is_offline(cpu)) {
5700 * Reset core_pick so that we don't enter the fastpath when
5701 * coming online. core_pick would already be migrated to
5702 * another cpu during offline.
5704 rq->core_pick = NULL;
5705 return __pick_next_task(rq, prev, rf);
5709 * If there were no {en,de}queues since we picked (IOW, the task
5710 * pointers are all still valid), and we haven't scheduled the last
5711 * pick yet, do so now.
5713 * rq->core_pick can be NULL if no selection was made for a CPU because
5714 * it was either offline or went offline during a sibling's core-wide
5715 * selection. In this case, do a core-wide selection.
5717 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5718 rq->core->core_pick_seq != rq->core_sched_seq &&
5720 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5722 next = rq->core_pick;
5724 put_prev_task(rq, prev);
5725 set_next_task(rq, next);
5728 rq->core_pick = NULL;
5732 put_prev_task_balance(rq, prev, rf);
5734 smt_mask = cpu_smt_mask(cpu);
5735 need_sync = !!rq->core->core_cookie;
5738 rq->core->core_cookie = 0UL;
5739 if (rq->core->core_forceidle_count) {
5740 if (!core_clock_updated) {
5741 update_rq_clock(rq->core);
5742 core_clock_updated = true;
5744 sched_core_account_forceidle(rq);
5745 /* reset after accounting force idle */
5746 rq->core->core_forceidle_start = 0;
5747 rq->core->core_forceidle_count = 0;
5748 rq->core->core_forceidle_occupation = 0;
5754 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5756 * @task_seq guards the task state ({en,de}queues)
5757 * @pick_seq is the @task_seq we did a selection on
5758 * @sched_seq is the @pick_seq we scheduled
5760 * However, preemptions can cause multiple picks on the same task set.
5761 * 'Fix' this by also increasing @task_seq for every pick.
5763 rq->core->core_task_seq++;
5766 * Optimize for common case where this CPU has no cookies
5767 * and there are no cookied tasks running on siblings.
5770 next = pick_task(rq);
5771 if (!next->core_cookie) {
5772 rq->core_pick = NULL;
5774 * For robustness, update the min_vruntime_fi for
5775 * unconstrained picks as well.
5777 WARN_ON_ONCE(fi_before);
5778 task_vruntime_update(rq, next, false);
5784 * For each thread: do the regular task pick and find the max prio task
5787 * Tie-break prio towards the current CPU
5789 for_each_cpu_wrap(i, smt_mask, cpu) {
5793 * Current cpu always has its clock updated on entrance to
5794 * pick_next_task(). If the current cpu is not the core,
5795 * the core may also have been updated above.
5797 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5798 update_rq_clock(rq_i);
5800 p = rq_i->core_pick = pick_task(rq_i);
5801 if (!max || prio_less(max, p, fi_before))
5805 cookie = rq->core->core_cookie = max->core_cookie;
5808 * For each thread: try and find a runnable task that matches @max or
5811 for_each_cpu(i, smt_mask) {
5813 p = rq_i->core_pick;
5815 if (!cookie_equals(p, cookie)) {
5818 p = sched_core_find(rq_i, cookie);
5820 p = idle_sched_class.pick_task(rq_i);
5823 rq_i->core_pick = p;
5825 if (p == rq_i->idle) {
5826 if (rq_i->nr_running) {
5827 rq->core->core_forceidle_count++;
5829 rq->core->core_forceidle_seq++;
5836 if (schedstat_enabled() && rq->core->core_forceidle_count) {
5837 rq->core->core_forceidle_start = rq_clock(rq->core);
5838 rq->core->core_forceidle_occupation = occ;
5841 rq->core->core_pick_seq = rq->core->core_task_seq;
5842 next = rq->core_pick;
5843 rq->core_sched_seq = rq->core->core_pick_seq;
5845 /* Something should have been selected for current CPU */
5846 WARN_ON_ONCE(!next);
5849 * Reschedule siblings
5851 * NOTE: L1TF -- at this point we're no longer running the old task and
5852 * sending an IPI (below) ensures the sibling will no longer be running
5853 * their task. This ensures there is no inter-sibling overlap between
5854 * non-matching user state.
5856 for_each_cpu(i, smt_mask) {
5860 * An online sibling might have gone offline before a task
5861 * could be picked for it, or it might be offline but later
5862 * happen to come online, but its too late and nothing was
5863 * picked for it. That's Ok - it will pick tasks for itself,
5866 if (!rq_i->core_pick)
5870 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5871 * fi_before fi update?
5877 if (!(fi_before && rq->core->core_forceidle_count))
5878 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
5880 rq_i->core_pick->core_occupation = occ;
5883 rq_i->core_pick = NULL;
5887 /* Did we break L1TF mitigation requirements? */
5888 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5890 if (rq_i->curr == rq_i->core_pick) {
5891 rq_i->core_pick = NULL;
5899 set_next_task(rq, next);
5901 if (rq->core->core_forceidle_count && next == rq->idle)
5902 queue_core_balance(rq);
5907 static bool try_steal_cookie(int this, int that)
5909 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5910 struct task_struct *p;
5911 unsigned long cookie;
5912 bool success = false;
5914 local_irq_disable();
5915 double_rq_lock(dst, src);
5917 cookie = dst->core->core_cookie;
5921 if (dst->curr != dst->idle)
5924 p = sched_core_find(src, cookie);
5929 if (p == src->core_pick || p == src->curr)
5932 if (!is_cpu_allowed(p, this))
5935 if (p->core_occupation > dst->idle->core_occupation)
5938 deactivate_task(src, p, 0);
5939 set_task_cpu(p, this);
5940 activate_task(dst, p, 0);
5948 p = sched_core_next(p, cookie);
5952 double_rq_unlock(dst, src);
5958 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5962 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5969 if (try_steal_cookie(cpu, i))
5976 static void sched_core_balance(struct rq *rq)
5978 struct sched_domain *sd;
5979 int cpu = cpu_of(rq);
5983 raw_spin_rq_unlock_irq(rq);
5984 for_each_domain(cpu, sd) {
5988 if (steal_cookie_task(cpu, sd))
5991 raw_spin_rq_lock_irq(rq);
5996 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5998 static void queue_core_balance(struct rq *rq)
6000 if (!sched_core_enabled(rq))
6003 if (!rq->core->core_cookie)
6006 if (!rq->nr_running) /* not forced idle */
6009 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6012 static void sched_core_cpu_starting(unsigned int cpu)
6014 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6015 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6016 unsigned long flags;
6019 sched_core_lock(cpu, &flags);
6021 WARN_ON_ONCE(rq->core != rq);
6023 /* if we're the first, we'll be our own leader */
6024 if (cpumask_weight(smt_mask) == 1)
6027 /* find the leader */
6028 for_each_cpu(t, smt_mask) {
6032 if (rq->core == rq) {
6038 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6041 /* install and validate core_rq */
6042 for_each_cpu(t, smt_mask) {
6048 WARN_ON_ONCE(rq->core != core_rq);
6052 sched_core_unlock(cpu, &flags);
6055 static void sched_core_cpu_deactivate(unsigned int cpu)
6057 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6058 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6059 unsigned long flags;
6062 sched_core_lock(cpu, &flags);
6064 /* if we're the last man standing, nothing to do */
6065 if (cpumask_weight(smt_mask) == 1) {
6066 WARN_ON_ONCE(rq->core != rq);
6070 /* if we're not the leader, nothing to do */
6074 /* find a new leader */
6075 for_each_cpu(t, smt_mask) {
6078 core_rq = cpu_rq(t);
6082 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6085 /* copy the shared state to the new leader */
6086 core_rq->core_task_seq = rq->core_task_seq;
6087 core_rq->core_pick_seq = rq->core_pick_seq;
6088 core_rq->core_cookie = rq->core_cookie;
6089 core_rq->core_forceidle_count = rq->core_forceidle_count;
6090 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6091 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6094 * Accounting edge for forced idle is handled in pick_next_task().
6095 * Don't need another one here, since the hotplug thread shouldn't
6098 core_rq->core_forceidle_start = 0;
6100 /* install new leader */
6101 for_each_cpu(t, smt_mask) {
6107 sched_core_unlock(cpu, &flags);
6110 static inline void sched_core_cpu_dying(unsigned int cpu)
6112 struct rq *rq = cpu_rq(cpu);
6118 #else /* !CONFIG_SCHED_CORE */
6120 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6121 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6122 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6124 static struct task_struct *
6125 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6127 return __pick_next_task(rq, prev, rf);
6130 #endif /* CONFIG_SCHED_CORE */
6133 * Constants for the sched_mode argument of __schedule().
6135 * The mode argument allows RT enabled kernels to differentiate a
6136 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6137 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6138 * optimize the AND operation out and just check for zero.
6141 #define SM_PREEMPT 0x1
6142 #define SM_RTLOCK_WAIT 0x2
6144 #ifndef CONFIG_PREEMPT_RT
6145 # define SM_MASK_PREEMPT (~0U)
6147 # define SM_MASK_PREEMPT SM_PREEMPT
6151 * __schedule() is the main scheduler function.
6153 * The main means of driving the scheduler and thus entering this function are:
6155 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6157 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6158 * paths. For example, see arch/x86/entry_64.S.
6160 * To drive preemption between tasks, the scheduler sets the flag in timer
6161 * interrupt handler scheduler_tick().
6163 * 3. Wakeups don't really cause entry into schedule(). They add a
6164 * task to the run-queue and that's it.
6166 * Now, if the new task added to the run-queue preempts the current
6167 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6168 * called on the nearest possible occasion:
6170 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6172 * - in syscall or exception context, at the next outmost
6173 * preempt_enable(). (this might be as soon as the wake_up()'s
6176 * - in IRQ context, return from interrupt-handler to
6177 * preemptible context
6179 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6182 * - cond_resched() call
6183 * - explicit schedule() call
6184 * - return from syscall or exception to user-space
6185 * - return from interrupt-handler to user-space
6187 * WARNING: must be called with preemption disabled!
6189 static void __sched notrace __schedule(unsigned int sched_mode)
6191 struct task_struct *prev, *next;
6192 unsigned long *switch_count;
6193 unsigned long prev_state;
6198 cpu = smp_processor_id();
6202 schedule_debug(prev, !!sched_mode);
6204 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6207 local_irq_disable();
6208 rcu_note_context_switch(!!sched_mode);
6211 * Make sure that signal_pending_state()->signal_pending() below
6212 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6213 * done by the caller to avoid the race with signal_wake_up():
6215 * __set_current_state(@state) signal_wake_up()
6216 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6217 * wake_up_state(p, state)
6218 * LOCK rq->lock LOCK p->pi_state
6219 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6220 * if (signal_pending_state()) if (p->state & @state)
6222 * Also, the membarrier system call requires a full memory barrier
6223 * after coming from user-space, before storing to rq->curr.
6226 smp_mb__after_spinlock();
6228 /* Promote REQ to ACT */
6229 rq->clock_update_flags <<= 1;
6230 update_rq_clock(rq);
6232 switch_count = &prev->nivcsw;
6235 * We must load prev->state once (task_struct::state is volatile), such
6238 * - we form a control dependency vs deactivate_task() below.
6239 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
6241 prev_state = READ_ONCE(prev->__state);
6242 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6243 if (signal_pending_state(prev_state, prev)) {
6244 WRITE_ONCE(prev->__state, TASK_RUNNING);
6246 prev->sched_contributes_to_load =
6247 (prev_state & TASK_UNINTERRUPTIBLE) &&
6248 !(prev_state & TASK_NOLOAD) &&
6249 !(prev->flags & PF_FROZEN);
6251 if (prev->sched_contributes_to_load)
6252 rq->nr_uninterruptible++;
6255 * __schedule() ttwu()
6256 * prev_state = prev->state; if (p->on_rq && ...)
6257 * if (prev_state) goto out;
6258 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6259 * p->state = TASK_WAKING
6261 * Where __schedule() and ttwu() have matching control dependencies.
6263 * After this, schedule() must not care about p->state any more.
6265 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6267 if (prev->in_iowait) {
6268 atomic_inc(&rq->nr_iowait);
6269 delayacct_blkio_start();
6272 switch_count = &prev->nvcsw;
6275 next = pick_next_task(rq, prev, &rf);
6276 clear_tsk_need_resched(prev);
6277 clear_preempt_need_resched();
6278 #ifdef CONFIG_SCHED_DEBUG
6279 rq->last_seen_need_resched_ns = 0;
6282 if (likely(prev != next)) {
6285 * RCU users of rcu_dereference(rq->curr) may not see
6286 * changes to task_struct made by pick_next_task().
6288 RCU_INIT_POINTER(rq->curr, next);
6290 * The membarrier system call requires each architecture
6291 * to have a full memory barrier after updating
6292 * rq->curr, before returning to user-space.
6294 * Here are the schemes providing that barrier on the
6295 * various architectures:
6296 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6297 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6298 * - finish_lock_switch() for weakly-ordered
6299 * architectures where spin_unlock is a full barrier,
6300 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6301 * is a RELEASE barrier),
6305 migrate_disable_switch(rq, prev);
6306 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6308 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next);
6310 /* Also unlocks the rq: */
6311 rq = context_switch(rq, prev, next, &rf);
6313 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6315 rq_unpin_lock(rq, &rf);
6316 __balance_callbacks(rq);
6317 raw_spin_rq_unlock_irq(rq);
6321 void __noreturn do_task_dead(void)
6323 /* Causes final put_task_struct in finish_task_switch(): */
6324 set_special_state(TASK_DEAD);
6326 /* Tell freezer to ignore us: */
6327 current->flags |= PF_NOFREEZE;
6329 __schedule(SM_NONE);
6332 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6337 static inline void sched_submit_work(struct task_struct *tsk)
6339 unsigned int task_flags;
6341 if (task_is_running(tsk))
6344 task_flags = tsk->flags;
6346 * If a worker goes to sleep, notify and ask workqueue whether it
6347 * wants to wake up a task to maintain concurrency.
6349 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6350 if (task_flags & PF_WQ_WORKER)
6351 wq_worker_sleeping(tsk);
6353 io_wq_worker_sleeping(tsk);
6356 if (tsk_is_pi_blocked(tsk))
6360 * If we are going to sleep and we have plugged IO queued,
6361 * make sure to submit it to avoid deadlocks.
6363 if (blk_needs_flush_plug(tsk))
6364 blk_flush_plug(tsk->plug, true);
6367 static void sched_update_worker(struct task_struct *tsk)
6369 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6370 if (tsk->flags & PF_WQ_WORKER)
6371 wq_worker_running(tsk);
6373 io_wq_worker_running(tsk);
6377 asmlinkage __visible void __sched schedule(void)
6379 struct task_struct *tsk = current;
6381 sched_submit_work(tsk);
6384 __schedule(SM_NONE);
6385 sched_preempt_enable_no_resched();
6386 } while (need_resched());
6387 sched_update_worker(tsk);
6389 EXPORT_SYMBOL(schedule);
6392 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6393 * state (have scheduled out non-voluntarily) by making sure that all
6394 * tasks have either left the run queue or have gone into user space.
6395 * As idle tasks do not do either, they must not ever be preempted
6396 * (schedule out non-voluntarily).
6398 * schedule_idle() is similar to schedule_preempt_disable() except that it
6399 * never enables preemption because it does not call sched_submit_work().
6401 void __sched schedule_idle(void)
6404 * As this skips calling sched_submit_work(), which the idle task does
6405 * regardless because that function is a nop when the task is in a
6406 * TASK_RUNNING state, make sure this isn't used someplace that the
6407 * current task can be in any other state. Note, idle is always in the
6408 * TASK_RUNNING state.
6410 WARN_ON_ONCE(current->__state);
6412 __schedule(SM_NONE);
6413 } while (need_resched());
6416 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6417 asmlinkage __visible void __sched schedule_user(void)
6420 * If we come here after a random call to set_need_resched(),
6421 * or we have been woken up remotely but the IPI has not yet arrived,
6422 * we haven't yet exited the RCU idle mode. Do it here manually until
6423 * we find a better solution.
6425 * NB: There are buggy callers of this function. Ideally we
6426 * should warn if prev_state != CONTEXT_USER, but that will trigger
6427 * too frequently to make sense yet.
6429 enum ctx_state prev_state = exception_enter();
6431 exception_exit(prev_state);
6436 * schedule_preempt_disabled - called with preemption disabled
6438 * Returns with preemption disabled. Note: preempt_count must be 1
6440 void __sched schedule_preempt_disabled(void)
6442 sched_preempt_enable_no_resched();
6447 #ifdef CONFIG_PREEMPT_RT
6448 void __sched notrace schedule_rtlock(void)
6452 __schedule(SM_RTLOCK_WAIT);
6453 sched_preempt_enable_no_resched();
6454 } while (need_resched());
6456 NOKPROBE_SYMBOL(schedule_rtlock);
6459 static void __sched notrace preempt_schedule_common(void)
6463 * Because the function tracer can trace preempt_count_sub()
6464 * and it also uses preempt_enable/disable_notrace(), if
6465 * NEED_RESCHED is set, the preempt_enable_notrace() called
6466 * by the function tracer will call this function again and
6467 * cause infinite recursion.
6469 * Preemption must be disabled here before the function
6470 * tracer can trace. Break up preempt_disable() into two
6471 * calls. One to disable preemption without fear of being
6472 * traced. The other to still record the preemption latency,
6473 * which can also be traced by the function tracer.
6475 preempt_disable_notrace();
6476 preempt_latency_start(1);
6477 __schedule(SM_PREEMPT);
6478 preempt_latency_stop(1);
6479 preempt_enable_no_resched_notrace();
6482 * Check again in case we missed a preemption opportunity
6483 * between schedule and now.
6485 } while (need_resched());
6488 #ifdef CONFIG_PREEMPTION
6490 * This is the entry point to schedule() from in-kernel preemption
6491 * off of preempt_enable.
6493 asmlinkage __visible void __sched notrace preempt_schedule(void)
6496 * If there is a non-zero preempt_count or interrupts are disabled,
6497 * we do not want to preempt the current task. Just return..
6499 if (likely(!preemptible()))
6502 preempt_schedule_common();
6504 NOKPROBE_SYMBOL(preempt_schedule);
6505 EXPORT_SYMBOL(preempt_schedule);
6507 #ifdef CONFIG_PREEMPT_DYNAMIC
6508 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6509 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6514 * preempt_schedule_notrace - preempt_schedule called by tracing
6516 * The tracing infrastructure uses preempt_enable_notrace to prevent
6517 * recursion and tracing preempt enabling caused by the tracing
6518 * infrastructure itself. But as tracing can happen in areas coming
6519 * from userspace or just about to enter userspace, a preempt enable
6520 * can occur before user_exit() is called. This will cause the scheduler
6521 * to be called when the system is still in usermode.
6523 * To prevent this, the preempt_enable_notrace will use this function
6524 * instead of preempt_schedule() to exit user context if needed before
6525 * calling the scheduler.
6527 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6529 enum ctx_state prev_ctx;
6531 if (likely(!preemptible()))
6536 * Because the function tracer can trace preempt_count_sub()
6537 * and it also uses preempt_enable/disable_notrace(), if
6538 * NEED_RESCHED is set, the preempt_enable_notrace() called
6539 * by the function tracer will call this function again and
6540 * cause infinite recursion.
6542 * Preemption must be disabled here before the function
6543 * tracer can trace. Break up preempt_disable() into two
6544 * calls. One to disable preemption without fear of being
6545 * traced. The other to still record the preemption latency,
6546 * which can also be traced by the function tracer.
6548 preempt_disable_notrace();
6549 preempt_latency_start(1);
6551 * Needs preempt disabled in case user_exit() is traced
6552 * and the tracer calls preempt_enable_notrace() causing
6553 * an infinite recursion.
6555 prev_ctx = exception_enter();
6556 __schedule(SM_PREEMPT);
6557 exception_exit(prev_ctx);
6559 preempt_latency_stop(1);
6560 preempt_enable_no_resched_notrace();
6561 } while (need_resched());
6563 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6565 #ifdef CONFIG_PREEMPT_DYNAMIC
6566 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6567 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6570 #endif /* CONFIG_PREEMPTION */
6572 #ifdef CONFIG_PREEMPT_DYNAMIC
6574 #include <linux/entry-common.h>
6579 * SC:preempt_schedule
6580 * SC:preempt_schedule_notrace
6581 * SC:irqentry_exit_cond_resched
6585 * cond_resched <- __cond_resched
6586 * might_resched <- RET0
6587 * preempt_schedule <- NOP
6588 * preempt_schedule_notrace <- NOP
6589 * irqentry_exit_cond_resched <- NOP
6592 * cond_resched <- __cond_resched
6593 * might_resched <- __cond_resched
6594 * preempt_schedule <- NOP
6595 * preempt_schedule_notrace <- NOP
6596 * irqentry_exit_cond_resched <- NOP
6599 * cond_resched <- RET0
6600 * might_resched <- RET0
6601 * preempt_schedule <- preempt_schedule
6602 * preempt_schedule_notrace <- preempt_schedule_notrace
6603 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6607 preempt_dynamic_undefined = -1,
6608 preempt_dynamic_none,
6609 preempt_dynamic_voluntary,
6610 preempt_dynamic_full,
6613 int preempt_dynamic_mode = preempt_dynamic_undefined;
6615 int sched_dynamic_mode(const char *str)
6617 if (!strcmp(str, "none"))
6618 return preempt_dynamic_none;
6620 if (!strcmp(str, "voluntary"))
6621 return preempt_dynamic_voluntary;
6623 if (!strcmp(str, "full"))
6624 return preempt_dynamic_full;
6629 void sched_dynamic_update(int mode)
6632 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6633 * the ZERO state, which is invalid.
6635 static_call_update(cond_resched, __cond_resched);
6636 static_call_update(might_resched, __cond_resched);
6637 static_call_update(preempt_schedule, __preempt_schedule_func);
6638 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6639 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6642 case preempt_dynamic_none:
6643 static_call_update(cond_resched, __cond_resched);
6644 static_call_update(might_resched, (void *)&__static_call_return0);
6645 static_call_update(preempt_schedule, NULL);
6646 static_call_update(preempt_schedule_notrace, NULL);
6647 static_call_update(irqentry_exit_cond_resched, NULL);
6648 pr_info("Dynamic Preempt: none\n");
6651 case preempt_dynamic_voluntary:
6652 static_call_update(cond_resched, __cond_resched);
6653 static_call_update(might_resched, __cond_resched);
6654 static_call_update(preempt_schedule, NULL);
6655 static_call_update(preempt_schedule_notrace, NULL);
6656 static_call_update(irqentry_exit_cond_resched, NULL);
6657 pr_info("Dynamic Preempt: voluntary\n");
6660 case preempt_dynamic_full:
6661 static_call_update(cond_resched, (void *)&__static_call_return0);
6662 static_call_update(might_resched, (void *)&__static_call_return0);
6663 static_call_update(preempt_schedule, __preempt_schedule_func);
6664 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6665 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6666 pr_info("Dynamic Preempt: full\n");
6670 preempt_dynamic_mode = mode;
6673 static int __init setup_preempt_mode(char *str)
6675 int mode = sched_dynamic_mode(str);
6677 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6681 sched_dynamic_update(mode);
6684 __setup("preempt=", setup_preempt_mode);
6686 static void __init preempt_dynamic_init(void)
6688 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
6689 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
6690 sched_dynamic_update(preempt_dynamic_none);
6691 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
6692 sched_dynamic_update(preempt_dynamic_voluntary);
6694 /* Default static call setting, nothing to do */
6695 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
6696 preempt_dynamic_mode = preempt_dynamic_full;
6697 pr_info("Dynamic Preempt: full\n");
6702 #else /* !CONFIG_PREEMPT_DYNAMIC */
6704 static inline void preempt_dynamic_init(void) { }
6706 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
6709 * This is the entry point to schedule() from kernel preemption
6710 * off of irq context.
6711 * Note, that this is called and return with irqs disabled. This will
6712 * protect us against recursive calling from irq.
6714 asmlinkage __visible void __sched preempt_schedule_irq(void)
6716 enum ctx_state prev_state;
6718 /* Catch callers which need to be fixed */
6719 BUG_ON(preempt_count() || !irqs_disabled());
6721 prev_state = exception_enter();
6726 __schedule(SM_PREEMPT);
6727 local_irq_disable();
6728 sched_preempt_enable_no_resched();
6729 } while (need_resched());
6731 exception_exit(prev_state);
6734 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6737 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6738 return try_to_wake_up(curr->private, mode, wake_flags);
6740 EXPORT_SYMBOL(default_wake_function);
6742 static void __setscheduler_prio(struct task_struct *p, int prio)
6745 p->sched_class = &dl_sched_class;
6746 else if (rt_prio(prio))
6747 p->sched_class = &rt_sched_class;
6749 p->sched_class = &fair_sched_class;
6754 #ifdef CONFIG_RT_MUTEXES
6756 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6759 prio = min(prio, pi_task->prio);
6764 static inline int rt_effective_prio(struct task_struct *p, int prio)
6766 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6768 return __rt_effective_prio(pi_task, prio);
6772 * rt_mutex_setprio - set the current priority of a task
6774 * @pi_task: donor task
6776 * This function changes the 'effective' priority of a task. It does
6777 * not touch ->normal_prio like __setscheduler().
6779 * Used by the rt_mutex code to implement priority inheritance
6780 * logic. Call site only calls if the priority of the task changed.
6782 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6784 int prio, oldprio, queued, running, queue_flag =
6785 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6786 const struct sched_class *prev_class;
6790 /* XXX used to be waiter->prio, not waiter->task->prio */
6791 prio = __rt_effective_prio(pi_task, p->normal_prio);
6794 * If nothing changed; bail early.
6796 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6799 rq = __task_rq_lock(p, &rf);
6800 update_rq_clock(rq);
6802 * Set under pi_lock && rq->lock, such that the value can be used under
6805 * Note that there is loads of tricky to make this pointer cache work
6806 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6807 * ensure a task is de-boosted (pi_task is set to NULL) before the
6808 * task is allowed to run again (and can exit). This ensures the pointer
6809 * points to a blocked task -- which guarantees the task is present.
6811 p->pi_top_task = pi_task;
6814 * For FIFO/RR we only need to set prio, if that matches we're done.
6816 if (prio == p->prio && !dl_prio(prio))
6820 * Idle task boosting is a nono in general. There is one
6821 * exception, when PREEMPT_RT and NOHZ is active:
6823 * The idle task calls get_next_timer_interrupt() and holds
6824 * the timer wheel base->lock on the CPU and another CPU wants
6825 * to access the timer (probably to cancel it). We can safely
6826 * ignore the boosting request, as the idle CPU runs this code
6827 * with interrupts disabled and will complete the lock
6828 * protected section without being interrupted. So there is no
6829 * real need to boost.
6831 if (unlikely(p == rq->idle)) {
6832 WARN_ON(p != rq->curr);
6833 WARN_ON(p->pi_blocked_on);
6837 trace_sched_pi_setprio(p, pi_task);
6840 if (oldprio == prio)
6841 queue_flag &= ~DEQUEUE_MOVE;
6843 prev_class = p->sched_class;
6844 queued = task_on_rq_queued(p);
6845 running = task_current(rq, p);
6847 dequeue_task(rq, p, queue_flag);
6849 put_prev_task(rq, p);
6852 * Boosting condition are:
6853 * 1. -rt task is running and holds mutex A
6854 * --> -dl task blocks on mutex A
6856 * 2. -dl task is running and holds mutex A
6857 * --> -dl task blocks on mutex A and could preempt the
6860 if (dl_prio(prio)) {
6861 if (!dl_prio(p->normal_prio) ||
6862 (pi_task && dl_prio(pi_task->prio) &&
6863 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6864 p->dl.pi_se = pi_task->dl.pi_se;
6865 queue_flag |= ENQUEUE_REPLENISH;
6867 p->dl.pi_se = &p->dl;
6869 } else if (rt_prio(prio)) {
6870 if (dl_prio(oldprio))
6871 p->dl.pi_se = &p->dl;
6873 queue_flag |= ENQUEUE_HEAD;
6875 if (dl_prio(oldprio))
6876 p->dl.pi_se = &p->dl;
6877 if (rt_prio(oldprio))
6881 __setscheduler_prio(p, prio);
6884 enqueue_task(rq, p, queue_flag);
6886 set_next_task(rq, p);
6888 check_class_changed(rq, p, prev_class, oldprio);
6890 /* Avoid rq from going away on us: */
6893 rq_unpin_lock(rq, &rf);
6894 __balance_callbacks(rq);
6895 raw_spin_rq_unlock(rq);
6900 static inline int rt_effective_prio(struct task_struct *p, int prio)
6906 void set_user_nice(struct task_struct *p, long nice)
6908 bool queued, running;
6913 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6916 * We have to be careful, if called from sys_setpriority(),
6917 * the task might be in the middle of scheduling on another CPU.
6919 rq = task_rq_lock(p, &rf);
6920 update_rq_clock(rq);
6923 * The RT priorities are set via sched_setscheduler(), but we still
6924 * allow the 'normal' nice value to be set - but as expected
6925 * it won't have any effect on scheduling until the task is
6926 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6928 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6929 p->static_prio = NICE_TO_PRIO(nice);
6932 queued = task_on_rq_queued(p);
6933 running = task_current(rq, p);
6935 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6937 put_prev_task(rq, p);
6939 p->static_prio = NICE_TO_PRIO(nice);
6940 set_load_weight(p, true);
6942 p->prio = effective_prio(p);
6945 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6947 set_next_task(rq, p);
6950 * If the task increased its priority or is running and
6951 * lowered its priority, then reschedule its CPU:
6953 p->sched_class->prio_changed(rq, p, old_prio);
6956 task_rq_unlock(rq, p, &rf);
6958 EXPORT_SYMBOL(set_user_nice);
6961 * can_nice - check if a task can reduce its nice value
6965 int can_nice(const struct task_struct *p, const int nice)
6967 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6968 int nice_rlim = nice_to_rlimit(nice);
6970 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6971 capable(CAP_SYS_NICE));
6974 #ifdef __ARCH_WANT_SYS_NICE
6977 * sys_nice - change the priority of the current process.
6978 * @increment: priority increment
6980 * sys_setpriority is a more generic, but much slower function that
6981 * does similar things.
6983 SYSCALL_DEFINE1(nice, int, increment)
6988 * Setpriority might change our priority at the same moment.
6989 * We don't have to worry. Conceptually one call occurs first
6990 * and we have a single winner.
6992 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6993 nice = task_nice(current) + increment;
6995 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6996 if (increment < 0 && !can_nice(current, nice))
6999 retval = security_task_setnice(current, nice);
7003 set_user_nice(current, nice);
7010 * task_prio - return the priority value of a given task.
7011 * @p: the task in question.
7013 * Return: The priority value as seen by users in /proc.
7015 * sched policy return value kernel prio user prio/nice
7017 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7018 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7019 * deadline -101 -1 0
7021 int task_prio(const struct task_struct *p)
7023 return p->prio - MAX_RT_PRIO;
7027 * idle_cpu - is a given CPU idle currently?
7028 * @cpu: the processor in question.
7030 * Return: 1 if the CPU is currently idle. 0 otherwise.
7032 int idle_cpu(int cpu)
7034 struct rq *rq = cpu_rq(cpu);
7036 if (rq->curr != rq->idle)
7043 if (rq->ttwu_pending)
7051 * available_idle_cpu - is a given CPU idle for enqueuing work.
7052 * @cpu: the CPU in question.
7054 * Return: 1 if the CPU is currently idle. 0 otherwise.
7056 int available_idle_cpu(int cpu)
7061 if (vcpu_is_preempted(cpu))
7068 * idle_task - return the idle task for a given CPU.
7069 * @cpu: the processor in question.
7071 * Return: The idle task for the CPU @cpu.
7073 struct task_struct *idle_task(int cpu)
7075 return cpu_rq(cpu)->idle;
7080 * This function computes an effective utilization for the given CPU, to be
7081 * used for frequency selection given the linear relation: f = u * f_max.
7083 * The scheduler tracks the following metrics:
7085 * cpu_util_{cfs,rt,dl,irq}()
7088 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7089 * synchronized windows and are thus directly comparable.
7091 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7092 * which excludes things like IRQ and steal-time. These latter are then accrued
7093 * in the irq utilization.
7095 * The DL bandwidth number otoh is not a measured metric but a value computed
7096 * based on the task model parameters and gives the minimal utilization
7097 * required to meet deadlines.
7099 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7100 unsigned long max, enum cpu_util_type type,
7101 struct task_struct *p)
7103 unsigned long dl_util, util, irq;
7104 struct rq *rq = cpu_rq(cpu);
7106 if (!uclamp_is_used() &&
7107 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7112 * Early check to see if IRQ/steal time saturates the CPU, can be
7113 * because of inaccuracies in how we track these -- see
7114 * update_irq_load_avg().
7116 irq = cpu_util_irq(rq);
7117 if (unlikely(irq >= max))
7121 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7122 * CFS tasks and we use the same metric to track the effective
7123 * utilization (PELT windows are synchronized) we can directly add them
7124 * to obtain the CPU's actual utilization.
7126 * CFS and RT utilization can be boosted or capped, depending on
7127 * utilization clamp constraints requested by currently RUNNABLE
7129 * When there are no CFS RUNNABLE tasks, clamps are released and
7130 * frequency will be gracefully reduced with the utilization decay.
7132 util = util_cfs + cpu_util_rt(rq);
7133 if (type == FREQUENCY_UTIL)
7134 util = uclamp_rq_util_with(rq, util, p);
7136 dl_util = cpu_util_dl(rq);
7139 * For frequency selection we do not make cpu_util_dl() a permanent part
7140 * of this sum because we want to use cpu_bw_dl() later on, but we need
7141 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7142 * that we select f_max when there is no idle time.
7144 * NOTE: numerical errors or stop class might cause us to not quite hit
7145 * saturation when we should -- something for later.
7147 if (util + dl_util >= max)
7151 * OTOH, for energy computation we need the estimated running time, so
7152 * include util_dl and ignore dl_bw.
7154 if (type == ENERGY_UTIL)
7158 * There is still idle time; further improve the number by using the
7159 * irq metric. Because IRQ/steal time is hidden from the task clock we
7160 * need to scale the task numbers:
7163 * U' = irq + --------- * U
7166 util = scale_irq_capacity(util, irq, max);
7170 * Bandwidth required by DEADLINE must always be granted while, for
7171 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7172 * to gracefully reduce the frequency when no tasks show up for longer
7175 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7176 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7177 * an interface. So, we only do the latter for now.
7179 if (type == FREQUENCY_UTIL)
7180 util += cpu_bw_dl(rq);
7182 return min(max, util);
7185 unsigned long sched_cpu_util(int cpu, unsigned long max)
7187 return effective_cpu_util(cpu, cpu_util_cfs(cpu), max,
7190 #endif /* CONFIG_SMP */
7193 * find_process_by_pid - find a process with a matching PID value.
7194 * @pid: the pid in question.
7196 * The task of @pid, if found. %NULL otherwise.
7198 static struct task_struct *find_process_by_pid(pid_t pid)
7200 return pid ? find_task_by_vpid(pid) : current;
7204 * sched_setparam() passes in -1 for its policy, to let the functions
7205 * it calls know not to change it.
7207 #define SETPARAM_POLICY -1
7209 static void __setscheduler_params(struct task_struct *p,
7210 const struct sched_attr *attr)
7212 int policy = attr->sched_policy;
7214 if (policy == SETPARAM_POLICY)
7219 if (dl_policy(policy))
7220 __setparam_dl(p, attr);
7221 else if (fair_policy(policy))
7222 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7225 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7226 * !rt_policy. Always setting this ensures that things like
7227 * getparam()/getattr() don't report silly values for !rt tasks.
7229 p->rt_priority = attr->sched_priority;
7230 p->normal_prio = normal_prio(p);
7231 set_load_weight(p, true);
7235 * Check the target process has a UID that matches the current process's:
7237 static bool check_same_owner(struct task_struct *p)
7239 const struct cred *cred = current_cred(), *pcred;
7243 pcred = __task_cred(p);
7244 match = (uid_eq(cred->euid, pcred->euid) ||
7245 uid_eq(cred->euid, pcred->uid));
7250 static int __sched_setscheduler(struct task_struct *p,
7251 const struct sched_attr *attr,
7254 int oldpolicy = -1, policy = attr->sched_policy;
7255 int retval, oldprio, newprio, queued, running;
7256 const struct sched_class *prev_class;
7257 struct callback_head *head;
7260 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7263 /* The pi code expects interrupts enabled */
7264 BUG_ON(pi && in_interrupt());
7266 /* Double check policy once rq lock held: */
7268 reset_on_fork = p->sched_reset_on_fork;
7269 policy = oldpolicy = p->policy;
7271 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7273 if (!valid_policy(policy))
7277 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7281 * Valid priorities for SCHED_FIFO and SCHED_RR are
7282 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7283 * SCHED_BATCH and SCHED_IDLE is 0.
7285 if (attr->sched_priority > MAX_RT_PRIO-1)
7287 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7288 (rt_policy(policy) != (attr->sched_priority != 0)))
7292 * Allow unprivileged RT tasks to decrease priority:
7294 if (user && !capable(CAP_SYS_NICE)) {
7295 if (fair_policy(policy)) {
7296 if (attr->sched_nice < task_nice(p) &&
7297 !can_nice(p, attr->sched_nice))
7301 if (rt_policy(policy)) {
7302 unsigned long rlim_rtprio =
7303 task_rlimit(p, RLIMIT_RTPRIO);
7305 /* Can't set/change the rt policy: */
7306 if (policy != p->policy && !rlim_rtprio)
7309 /* Can't increase priority: */
7310 if (attr->sched_priority > p->rt_priority &&
7311 attr->sched_priority > rlim_rtprio)
7316 * Can't set/change SCHED_DEADLINE policy at all for now
7317 * (safest behavior); in the future we would like to allow
7318 * unprivileged DL tasks to increase their relative deadline
7319 * or reduce their runtime (both ways reducing utilization)
7321 if (dl_policy(policy))
7325 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7326 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7328 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7329 if (!can_nice(p, task_nice(p)))
7333 /* Can't change other user's priorities: */
7334 if (!check_same_owner(p))
7337 /* Normal users shall not reset the sched_reset_on_fork flag: */
7338 if (p->sched_reset_on_fork && !reset_on_fork)
7343 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7346 retval = security_task_setscheduler(p);
7351 /* Update task specific "requested" clamps */
7352 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7353 retval = uclamp_validate(p, attr);
7362 * Make sure no PI-waiters arrive (or leave) while we are
7363 * changing the priority of the task:
7365 * To be able to change p->policy safely, the appropriate
7366 * runqueue lock must be held.
7368 rq = task_rq_lock(p, &rf);
7369 update_rq_clock(rq);
7372 * Changing the policy of the stop threads its a very bad idea:
7374 if (p == rq->stop) {
7380 * If not changing anything there's no need to proceed further,
7381 * but store a possible modification of reset_on_fork.
7383 if (unlikely(policy == p->policy)) {
7384 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7386 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7388 if (dl_policy(policy) && dl_param_changed(p, attr))
7390 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7393 p->sched_reset_on_fork = reset_on_fork;
7400 #ifdef CONFIG_RT_GROUP_SCHED
7402 * Do not allow realtime tasks into groups that have no runtime
7405 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7406 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7407 !task_group_is_autogroup(task_group(p))) {
7413 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7414 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7415 cpumask_t *span = rq->rd->span;
7418 * Don't allow tasks with an affinity mask smaller than
7419 * the entire root_domain to become SCHED_DEADLINE. We
7420 * will also fail if there's no bandwidth available.
7422 if (!cpumask_subset(span, p->cpus_ptr) ||
7423 rq->rd->dl_bw.bw == 0) {
7431 /* Re-check policy now with rq lock held: */
7432 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7433 policy = oldpolicy = -1;
7434 task_rq_unlock(rq, p, &rf);
7436 cpuset_read_unlock();
7441 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7442 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7445 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7450 p->sched_reset_on_fork = reset_on_fork;
7453 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7456 * Take priority boosted tasks into account. If the new
7457 * effective priority is unchanged, we just store the new
7458 * normal parameters and do not touch the scheduler class and
7459 * the runqueue. This will be done when the task deboost
7462 newprio = rt_effective_prio(p, newprio);
7463 if (newprio == oldprio)
7464 queue_flags &= ~DEQUEUE_MOVE;
7467 queued = task_on_rq_queued(p);
7468 running = task_current(rq, p);
7470 dequeue_task(rq, p, queue_flags);
7472 put_prev_task(rq, p);
7474 prev_class = p->sched_class;
7476 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7477 __setscheduler_params(p, attr);
7478 __setscheduler_prio(p, newprio);
7480 __setscheduler_uclamp(p, attr);
7484 * We enqueue to tail when the priority of a task is
7485 * increased (user space view).
7487 if (oldprio < p->prio)
7488 queue_flags |= ENQUEUE_HEAD;
7490 enqueue_task(rq, p, queue_flags);
7493 set_next_task(rq, p);
7495 check_class_changed(rq, p, prev_class, oldprio);
7497 /* Avoid rq from going away on us: */
7499 head = splice_balance_callbacks(rq);
7500 task_rq_unlock(rq, p, &rf);
7503 cpuset_read_unlock();
7504 rt_mutex_adjust_pi(p);
7507 /* Run balance callbacks after we've adjusted the PI chain: */
7508 balance_callbacks(rq, head);
7514 task_rq_unlock(rq, p, &rf);
7516 cpuset_read_unlock();
7520 static int _sched_setscheduler(struct task_struct *p, int policy,
7521 const struct sched_param *param, bool check)
7523 struct sched_attr attr = {
7524 .sched_policy = policy,
7525 .sched_priority = param->sched_priority,
7526 .sched_nice = PRIO_TO_NICE(p->static_prio),
7529 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7530 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7531 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7532 policy &= ~SCHED_RESET_ON_FORK;
7533 attr.sched_policy = policy;
7536 return __sched_setscheduler(p, &attr, check, true);
7539 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7540 * @p: the task in question.
7541 * @policy: new policy.
7542 * @param: structure containing the new RT priority.
7544 * Use sched_set_fifo(), read its comment.
7546 * Return: 0 on success. An error code otherwise.
7548 * NOTE that the task may be already dead.
7550 int sched_setscheduler(struct task_struct *p, int policy,
7551 const struct sched_param *param)
7553 return _sched_setscheduler(p, policy, param, true);
7556 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7558 return __sched_setscheduler(p, attr, true, true);
7561 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7563 return __sched_setscheduler(p, attr, false, true);
7565 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7568 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7569 * @p: the task in question.
7570 * @policy: new policy.
7571 * @param: structure containing the new RT priority.
7573 * Just like sched_setscheduler, only don't bother checking if the
7574 * current context has permission. For example, this is needed in
7575 * stop_machine(): we create temporary high priority worker threads,
7576 * but our caller might not have that capability.
7578 * Return: 0 on success. An error code otherwise.
7580 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7581 const struct sched_param *param)
7583 return _sched_setscheduler(p, policy, param, false);
7587 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7588 * incapable of resource management, which is the one thing an OS really should
7591 * This is of course the reason it is limited to privileged users only.
7593 * Worse still; it is fundamentally impossible to compose static priority
7594 * workloads. You cannot take two correctly working static prio workloads
7595 * and smash them together and still expect them to work.
7597 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7601 * The administrator _MUST_ configure the system, the kernel simply doesn't
7602 * know enough information to make a sensible choice.
7604 void sched_set_fifo(struct task_struct *p)
7606 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7607 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7609 EXPORT_SYMBOL_GPL(sched_set_fifo);
7612 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7614 void sched_set_fifo_low(struct task_struct *p)
7616 struct sched_param sp = { .sched_priority = 1 };
7617 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7619 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7621 void sched_set_normal(struct task_struct *p, int nice)
7623 struct sched_attr attr = {
7624 .sched_policy = SCHED_NORMAL,
7627 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7629 EXPORT_SYMBOL_GPL(sched_set_normal);
7632 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7634 struct sched_param lparam;
7635 struct task_struct *p;
7638 if (!param || pid < 0)
7640 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7645 p = find_process_by_pid(pid);
7651 retval = sched_setscheduler(p, policy, &lparam);
7659 * Mimics kernel/events/core.c perf_copy_attr().
7661 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7666 /* Zero the full structure, so that a short copy will be nice: */
7667 memset(attr, 0, sizeof(*attr));
7669 ret = get_user(size, &uattr->size);
7673 /* ABI compatibility quirk: */
7675 size = SCHED_ATTR_SIZE_VER0;
7676 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7679 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7686 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7687 size < SCHED_ATTR_SIZE_VER1)
7691 * XXX: Do we want to be lenient like existing syscalls; or do we want
7692 * to be strict and return an error on out-of-bounds values?
7694 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7699 put_user(sizeof(*attr), &uattr->size);
7703 static void get_params(struct task_struct *p, struct sched_attr *attr)
7705 if (task_has_dl_policy(p))
7706 __getparam_dl(p, attr);
7707 else if (task_has_rt_policy(p))
7708 attr->sched_priority = p->rt_priority;
7710 attr->sched_nice = task_nice(p);
7714 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7715 * @pid: the pid in question.
7716 * @policy: new policy.
7717 * @param: structure containing the new RT priority.
7719 * Return: 0 on success. An error code otherwise.
7721 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7726 return do_sched_setscheduler(pid, policy, param);
7730 * sys_sched_setparam - set/change the RT priority of a thread
7731 * @pid: the pid in question.
7732 * @param: structure containing the new RT priority.
7734 * Return: 0 on success. An error code otherwise.
7736 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7738 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7742 * sys_sched_setattr - same as above, but with extended sched_attr
7743 * @pid: the pid in question.
7744 * @uattr: structure containing the extended parameters.
7745 * @flags: for future extension.
7747 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7748 unsigned int, flags)
7750 struct sched_attr attr;
7751 struct task_struct *p;
7754 if (!uattr || pid < 0 || flags)
7757 retval = sched_copy_attr(uattr, &attr);
7761 if ((int)attr.sched_policy < 0)
7763 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7764 attr.sched_policy = SETPARAM_POLICY;
7768 p = find_process_by_pid(pid);
7774 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7775 get_params(p, &attr);
7776 retval = sched_setattr(p, &attr);
7784 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7785 * @pid: the pid in question.
7787 * Return: On success, the policy of the thread. Otherwise, a negative error
7790 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7792 struct task_struct *p;
7800 p = find_process_by_pid(pid);
7802 retval = security_task_getscheduler(p);
7805 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7812 * sys_sched_getparam - get the RT priority of a thread
7813 * @pid: the pid in question.
7814 * @param: structure containing the RT priority.
7816 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7819 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7821 struct sched_param lp = { .sched_priority = 0 };
7822 struct task_struct *p;
7825 if (!param || pid < 0)
7829 p = find_process_by_pid(pid);
7834 retval = security_task_getscheduler(p);
7838 if (task_has_rt_policy(p))
7839 lp.sched_priority = p->rt_priority;
7843 * This one might sleep, we cannot do it with a spinlock held ...
7845 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7855 * Copy the kernel size attribute structure (which might be larger
7856 * than what user-space knows about) to user-space.
7858 * Note that all cases are valid: user-space buffer can be larger or
7859 * smaller than the kernel-space buffer. The usual case is that both
7860 * have the same size.
7863 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7864 struct sched_attr *kattr,
7867 unsigned int ksize = sizeof(*kattr);
7869 if (!access_ok(uattr, usize))
7873 * sched_getattr() ABI forwards and backwards compatibility:
7875 * If usize == ksize then we just copy everything to user-space and all is good.
7877 * If usize < ksize then we only copy as much as user-space has space for,
7878 * this keeps ABI compatibility as well. We skip the rest.
7880 * If usize > ksize then user-space is using a newer version of the ABI,
7881 * which part the kernel doesn't know about. Just ignore it - tooling can
7882 * detect the kernel's knowledge of attributes from the attr->size value
7883 * which is set to ksize in this case.
7885 kattr->size = min(usize, ksize);
7887 if (copy_to_user(uattr, kattr, kattr->size))
7894 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7895 * @pid: the pid in question.
7896 * @uattr: structure containing the extended parameters.
7897 * @usize: sizeof(attr) for fwd/bwd comp.
7898 * @flags: for future extension.
7900 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7901 unsigned int, usize, unsigned int, flags)
7903 struct sched_attr kattr = { };
7904 struct task_struct *p;
7907 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7908 usize < SCHED_ATTR_SIZE_VER0 || flags)
7912 p = find_process_by_pid(pid);
7917 retval = security_task_getscheduler(p);
7921 kattr.sched_policy = p->policy;
7922 if (p->sched_reset_on_fork)
7923 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7924 get_params(p, &kattr);
7925 kattr.sched_flags &= SCHED_FLAG_ALL;
7927 #ifdef CONFIG_UCLAMP_TASK
7929 * This could race with another potential updater, but this is fine
7930 * because it'll correctly read the old or the new value. We don't need
7931 * to guarantee who wins the race as long as it doesn't return garbage.
7933 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7934 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7939 return sched_attr_copy_to_user(uattr, &kattr, usize);
7947 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7952 * If the task isn't a deadline task or admission control is
7953 * disabled then we don't care about affinity changes.
7955 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7959 * Since bandwidth control happens on root_domain basis,
7960 * if admission test is enabled, we only admit -deadline
7961 * tasks allowed to run on all the CPUs in the task's
7965 if (!cpumask_subset(task_rq(p)->rd->span, mask))
7973 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
7976 cpumask_var_t cpus_allowed, new_mask;
7978 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
7981 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7983 goto out_free_cpus_allowed;
7986 cpuset_cpus_allowed(p, cpus_allowed);
7987 cpumask_and(new_mask, mask, cpus_allowed);
7989 retval = dl_task_check_affinity(p, new_mask);
7991 goto out_free_new_mask;
7993 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
7995 goto out_free_new_mask;
7997 cpuset_cpus_allowed(p, cpus_allowed);
7998 if (!cpumask_subset(new_mask, cpus_allowed)) {
8000 * We must have raced with a concurrent cpuset update.
8001 * Just reset the cpumask to the cpuset's cpus_allowed.
8003 cpumask_copy(new_mask, cpus_allowed);
8008 free_cpumask_var(new_mask);
8009 out_free_cpus_allowed:
8010 free_cpumask_var(cpus_allowed);
8014 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8016 struct task_struct *p;
8021 p = find_process_by_pid(pid);
8027 /* Prevent p going away */
8031 if (p->flags & PF_NO_SETAFFINITY) {
8036 if (!check_same_owner(p)) {
8038 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8046 retval = security_task_setscheduler(p);
8050 retval = __sched_setaffinity(p, in_mask);
8056 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8057 struct cpumask *new_mask)
8059 if (len < cpumask_size())
8060 cpumask_clear(new_mask);
8061 else if (len > cpumask_size())
8062 len = cpumask_size();
8064 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8068 * sys_sched_setaffinity - set the CPU affinity of a process
8069 * @pid: pid of the process
8070 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8071 * @user_mask_ptr: user-space pointer to the new CPU mask
8073 * Return: 0 on success. An error code otherwise.
8075 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8076 unsigned long __user *, user_mask_ptr)
8078 cpumask_var_t new_mask;
8081 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8084 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8086 retval = sched_setaffinity(pid, new_mask);
8087 free_cpumask_var(new_mask);
8091 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8093 struct task_struct *p;
8094 unsigned long flags;
8100 p = find_process_by_pid(pid);
8104 retval = security_task_getscheduler(p);
8108 raw_spin_lock_irqsave(&p->pi_lock, flags);
8109 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8110 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8119 * sys_sched_getaffinity - get the CPU affinity of a process
8120 * @pid: pid of the process
8121 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8122 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8124 * Return: size of CPU mask copied to user_mask_ptr on success. An
8125 * error code otherwise.
8127 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8128 unsigned long __user *, user_mask_ptr)
8133 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8135 if (len & (sizeof(unsigned long)-1))
8138 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8141 ret = sched_getaffinity(pid, mask);
8143 unsigned int retlen = min(len, cpumask_size());
8145 if (copy_to_user(user_mask_ptr, mask, retlen))
8150 free_cpumask_var(mask);
8155 static void do_sched_yield(void)
8160 rq = this_rq_lock_irq(&rf);
8162 schedstat_inc(rq->yld_count);
8163 current->sched_class->yield_task(rq);
8166 rq_unlock_irq(rq, &rf);
8167 sched_preempt_enable_no_resched();
8173 * sys_sched_yield - yield the current processor to other threads.
8175 * This function yields the current CPU to other tasks. If there are no
8176 * other threads running on this CPU then this function will return.
8180 SYSCALL_DEFINE0(sched_yield)
8186 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8187 int __sched __cond_resched(void)
8189 if (should_resched(0)) {
8190 preempt_schedule_common();
8194 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8195 * whether the current CPU is in an RCU read-side critical section,
8196 * so the tick can report quiescent states even for CPUs looping
8197 * in kernel context. In contrast, in non-preemptible kernels,
8198 * RCU readers leave no in-memory hints, which means that CPU-bound
8199 * processes executing in kernel context might never report an
8200 * RCU quiescent state. Therefore, the following code causes
8201 * cond_resched() to report a quiescent state, but only when RCU
8202 * is in urgent need of one.
8204 #ifndef CONFIG_PREEMPT_RCU
8209 EXPORT_SYMBOL(__cond_resched);
8212 #ifdef CONFIG_PREEMPT_DYNAMIC
8213 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8214 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8216 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8217 EXPORT_STATIC_CALL_TRAMP(might_resched);
8221 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8222 * call schedule, and on return reacquire the lock.
8224 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8225 * operations here to prevent schedule() from being called twice (once via
8226 * spin_unlock(), once by hand).
8228 int __cond_resched_lock(spinlock_t *lock)
8230 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8233 lockdep_assert_held(lock);
8235 if (spin_needbreak(lock) || resched) {
8237 if (!_cond_resched())
8244 EXPORT_SYMBOL(__cond_resched_lock);
8246 int __cond_resched_rwlock_read(rwlock_t *lock)
8248 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8251 lockdep_assert_held_read(lock);
8253 if (rwlock_needbreak(lock) || resched) {
8255 if (!_cond_resched())
8262 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8264 int __cond_resched_rwlock_write(rwlock_t *lock)
8266 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8269 lockdep_assert_held_write(lock);
8271 if (rwlock_needbreak(lock) || resched) {
8273 if (!_cond_resched())
8280 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8283 * yield - yield the current processor to other threads.
8285 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8287 * The scheduler is at all times free to pick the calling task as the most
8288 * eligible task to run, if removing the yield() call from your code breaks
8289 * it, it's already broken.
8291 * Typical broken usage is:
8296 * where one assumes that yield() will let 'the other' process run that will
8297 * make event true. If the current task is a SCHED_FIFO task that will never
8298 * happen. Never use yield() as a progress guarantee!!
8300 * If you want to use yield() to wait for something, use wait_event().
8301 * If you want to use yield() to be 'nice' for others, use cond_resched().
8302 * If you still want to use yield(), do not!
8304 void __sched yield(void)
8306 set_current_state(TASK_RUNNING);
8309 EXPORT_SYMBOL(yield);
8312 * yield_to - yield the current processor to another thread in
8313 * your thread group, or accelerate that thread toward the
8314 * processor it's on.
8316 * @preempt: whether task preemption is allowed or not
8318 * It's the caller's job to ensure that the target task struct
8319 * can't go away on us before we can do any checks.
8322 * true (>0) if we indeed boosted the target task.
8323 * false (0) if we failed to boost the target.
8324 * -ESRCH if there's no task to yield to.
8326 int __sched yield_to(struct task_struct *p, bool preempt)
8328 struct task_struct *curr = current;
8329 struct rq *rq, *p_rq;
8330 unsigned long flags;
8333 local_irq_save(flags);
8339 * If we're the only runnable task on the rq and target rq also
8340 * has only one task, there's absolutely no point in yielding.
8342 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8347 double_rq_lock(rq, p_rq);
8348 if (task_rq(p) != p_rq) {
8349 double_rq_unlock(rq, p_rq);
8353 if (!curr->sched_class->yield_to_task)
8356 if (curr->sched_class != p->sched_class)
8359 if (task_running(p_rq, p) || !task_is_running(p))
8362 yielded = curr->sched_class->yield_to_task(rq, p);
8364 schedstat_inc(rq->yld_count);
8366 * Make p's CPU reschedule; pick_next_entity takes care of
8369 if (preempt && rq != p_rq)
8374 double_rq_unlock(rq, p_rq);
8376 local_irq_restore(flags);
8383 EXPORT_SYMBOL_GPL(yield_to);
8385 int io_schedule_prepare(void)
8387 int old_iowait = current->in_iowait;
8389 current->in_iowait = 1;
8391 blk_flush_plug(current->plug, true);
8396 void io_schedule_finish(int token)
8398 current->in_iowait = token;
8402 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8403 * that process accounting knows that this is a task in IO wait state.
8405 long __sched io_schedule_timeout(long timeout)
8410 token = io_schedule_prepare();
8411 ret = schedule_timeout(timeout);
8412 io_schedule_finish(token);
8416 EXPORT_SYMBOL(io_schedule_timeout);
8418 void __sched io_schedule(void)
8422 token = io_schedule_prepare();
8424 io_schedule_finish(token);
8426 EXPORT_SYMBOL(io_schedule);
8429 * sys_sched_get_priority_max - return maximum RT priority.
8430 * @policy: scheduling class.
8432 * Return: On success, this syscall returns the maximum
8433 * rt_priority that can be used by a given scheduling class.
8434 * On failure, a negative error code is returned.
8436 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8443 ret = MAX_RT_PRIO-1;
8445 case SCHED_DEADLINE:
8456 * sys_sched_get_priority_min - return minimum RT priority.
8457 * @policy: scheduling class.
8459 * Return: On success, this syscall returns the minimum
8460 * rt_priority that can be used by a given scheduling class.
8461 * On failure, a negative error code is returned.
8463 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8472 case SCHED_DEADLINE:
8481 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8483 struct task_struct *p;
8484 unsigned int time_slice;
8494 p = find_process_by_pid(pid);
8498 retval = security_task_getscheduler(p);
8502 rq = task_rq_lock(p, &rf);
8504 if (p->sched_class->get_rr_interval)
8505 time_slice = p->sched_class->get_rr_interval(rq, p);
8506 task_rq_unlock(rq, p, &rf);
8509 jiffies_to_timespec64(time_slice, t);
8518 * sys_sched_rr_get_interval - return the default timeslice of a process.
8519 * @pid: pid of the process.
8520 * @interval: userspace pointer to the timeslice value.
8522 * this syscall writes the default timeslice value of a given process
8523 * into the user-space timespec buffer. A value of '0' means infinity.
8525 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8528 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8529 struct __kernel_timespec __user *, interval)
8531 struct timespec64 t;
8532 int retval = sched_rr_get_interval(pid, &t);
8535 retval = put_timespec64(&t, interval);
8540 #ifdef CONFIG_COMPAT_32BIT_TIME
8541 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8542 struct old_timespec32 __user *, interval)
8544 struct timespec64 t;
8545 int retval = sched_rr_get_interval(pid, &t);
8548 retval = put_old_timespec32(&t, interval);
8553 void sched_show_task(struct task_struct *p)
8555 unsigned long free = 0;
8558 if (!try_get_task_stack(p))
8561 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8563 if (task_is_running(p))
8564 pr_cont(" running task ");
8565 #ifdef CONFIG_DEBUG_STACK_USAGE
8566 free = stack_not_used(p);
8571 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8573 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8574 free, task_pid_nr(p), ppid,
8575 read_task_thread_flags(p));
8577 print_worker_info(KERN_INFO, p);
8578 print_stop_info(KERN_INFO, p);
8579 show_stack(p, NULL, KERN_INFO);
8582 EXPORT_SYMBOL_GPL(sched_show_task);
8585 state_filter_match(unsigned long state_filter, struct task_struct *p)
8587 unsigned int state = READ_ONCE(p->__state);
8589 /* no filter, everything matches */
8593 /* filter, but doesn't match */
8594 if (!(state & state_filter))
8598 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8601 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8608 void show_state_filter(unsigned int state_filter)
8610 struct task_struct *g, *p;
8613 for_each_process_thread(g, p) {
8615 * reset the NMI-timeout, listing all files on a slow
8616 * console might take a lot of time:
8617 * Also, reset softlockup watchdogs on all CPUs, because
8618 * another CPU might be blocked waiting for us to process
8621 touch_nmi_watchdog();
8622 touch_all_softlockup_watchdogs();
8623 if (state_filter_match(state_filter, p))
8627 #ifdef CONFIG_SCHED_DEBUG
8629 sysrq_sched_debug_show();
8633 * Only show locks if all tasks are dumped:
8636 debug_show_all_locks();
8640 * init_idle - set up an idle thread for a given CPU
8641 * @idle: task in question
8642 * @cpu: CPU the idle task belongs to
8644 * NOTE: this function does not set the idle thread's NEED_RESCHED
8645 * flag, to make booting more robust.
8647 void __init init_idle(struct task_struct *idle, int cpu)
8649 struct rq *rq = cpu_rq(cpu);
8650 unsigned long flags;
8652 __sched_fork(0, idle);
8654 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8655 raw_spin_rq_lock(rq);
8657 idle->__state = TASK_RUNNING;
8658 idle->se.exec_start = sched_clock();
8660 * PF_KTHREAD should already be set at this point; regardless, make it
8661 * look like a proper per-CPU kthread.
8663 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8664 kthread_set_per_cpu(idle, cpu);
8668 * It's possible that init_idle() gets called multiple times on a task,
8669 * in that case do_set_cpus_allowed() will not do the right thing.
8671 * And since this is boot we can forgo the serialization.
8673 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8676 * We're having a chicken and egg problem, even though we are
8677 * holding rq->lock, the CPU isn't yet set to this CPU so the
8678 * lockdep check in task_group() will fail.
8680 * Similar case to sched_fork(). / Alternatively we could
8681 * use task_rq_lock() here and obtain the other rq->lock.
8686 __set_task_cpu(idle, cpu);
8690 rcu_assign_pointer(rq->curr, idle);
8691 idle->on_rq = TASK_ON_RQ_QUEUED;
8695 raw_spin_rq_unlock(rq);
8696 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8698 /* Set the preempt count _outside_ the spinlocks! */
8699 init_idle_preempt_count(idle, cpu);
8702 * The idle tasks have their own, simple scheduling class:
8704 idle->sched_class = &idle_sched_class;
8705 ftrace_graph_init_idle_task(idle, cpu);
8706 vtime_init_idle(idle, cpu);
8708 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8714 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8715 const struct cpumask *trial)
8719 if (!cpumask_weight(cur))
8722 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8727 int task_can_attach(struct task_struct *p,
8728 const struct cpumask *cs_cpus_allowed)
8733 * Kthreads which disallow setaffinity shouldn't be moved
8734 * to a new cpuset; we don't want to change their CPU
8735 * affinity and isolating such threads by their set of
8736 * allowed nodes is unnecessary. Thus, cpusets are not
8737 * applicable for such threads. This prevents checking for
8738 * success of set_cpus_allowed_ptr() on all attached tasks
8739 * before cpus_mask may be changed.
8741 if (p->flags & PF_NO_SETAFFINITY) {
8746 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8748 ret = dl_task_can_attach(p, cs_cpus_allowed);
8754 bool sched_smp_initialized __read_mostly;
8756 #ifdef CONFIG_NUMA_BALANCING
8757 /* Migrate current task p to target_cpu */
8758 int migrate_task_to(struct task_struct *p, int target_cpu)
8760 struct migration_arg arg = { p, target_cpu };
8761 int curr_cpu = task_cpu(p);
8763 if (curr_cpu == target_cpu)
8766 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8769 /* TODO: This is not properly updating schedstats */
8771 trace_sched_move_numa(p, curr_cpu, target_cpu);
8772 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8776 * Requeue a task on a given node and accurately track the number of NUMA
8777 * tasks on the runqueues
8779 void sched_setnuma(struct task_struct *p, int nid)
8781 bool queued, running;
8785 rq = task_rq_lock(p, &rf);
8786 queued = task_on_rq_queued(p);
8787 running = task_current(rq, p);
8790 dequeue_task(rq, p, DEQUEUE_SAVE);
8792 put_prev_task(rq, p);
8794 p->numa_preferred_nid = nid;
8797 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8799 set_next_task(rq, p);
8800 task_rq_unlock(rq, p, &rf);
8802 #endif /* CONFIG_NUMA_BALANCING */
8804 #ifdef CONFIG_HOTPLUG_CPU
8806 * Ensure that the idle task is using init_mm right before its CPU goes
8809 void idle_task_exit(void)
8811 struct mm_struct *mm = current->active_mm;
8813 BUG_ON(cpu_online(smp_processor_id()));
8814 BUG_ON(current != this_rq()->idle);
8816 if (mm != &init_mm) {
8817 switch_mm(mm, &init_mm, current);
8818 finish_arch_post_lock_switch();
8821 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8824 static int __balance_push_cpu_stop(void *arg)
8826 struct task_struct *p = arg;
8827 struct rq *rq = this_rq();
8831 raw_spin_lock_irq(&p->pi_lock);
8834 update_rq_clock(rq);
8836 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8837 cpu = select_fallback_rq(rq->cpu, p);
8838 rq = __migrate_task(rq, &rf, p, cpu);
8842 raw_spin_unlock_irq(&p->pi_lock);
8849 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8852 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8854 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8855 * effective when the hotplug motion is down.
8857 static void balance_push(struct rq *rq)
8859 struct task_struct *push_task = rq->curr;
8861 lockdep_assert_rq_held(rq);
8864 * Ensure the thing is persistent until balance_push_set(.on = false);
8866 rq->balance_callback = &balance_push_callback;
8869 * Only active while going offline and when invoked on the outgoing
8872 if (!cpu_dying(rq->cpu) || rq != this_rq())
8876 * Both the cpu-hotplug and stop task are in this case and are
8877 * required to complete the hotplug process.
8879 if (kthread_is_per_cpu(push_task) ||
8880 is_migration_disabled(push_task)) {
8883 * If this is the idle task on the outgoing CPU try to wake
8884 * up the hotplug control thread which might wait for the
8885 * last task to vanish. The rcuwait_active() check is
8886 * accurate here because the waiter is pinned on this CPU
8887 * and can't obviously be running in parallel.
8889 * On RT kernels this also has to check whether there are
8890 * pinned and scheduled out tasks on the runqueue. They
8891 * need to leave the migrate disabled section first.
8893 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8894 rcuwait_active(&rq->hotplug_wait)) {
8895 raw_spin_rq_unlock(rq);
8896 rcuwait_wake_up(&rq->hotplug_wait);
8897 raw_spin_rq_lock(rq);
8902 get_task_struct(push_task);
8904 * Temporarily drop rq->lock such that we can wake-up the stop task.
8905 * Both preemption and IRQs are still disabled.
8907 raw_spin_rq_unlock(rq);
8908 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8909 this_cpu_ptr(&push_work));
8911 * At this point need_resched() is true and we'll take the loop in
8912 * schedule(). The next pick is obviously going to be the stop task
8913 * which kthread_is_per_cpu() and will push this task away.
8915 raw_spin_rq_lock(rq);
8918 static void balance_push_set(int cpu, bool on)
8920 struct rq *rq = cpu_rq(cpu);
8923 rq_lock_irqsave(rq, &rf);
8925 WARN_ON_ONCE(rq->balance_callback);
8926 rq->balance_callback = &balance_push_callback;
8927 } else if (rq->balance_callback == &balance_push_callback) {
8928 rq->balance_callback = NULL;
8930 rq_unlock_irqrestore(rq, &rf);
8934 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8935 * inactive. All tasks which are not per CPU kernel threads are either
8936 * pushed off this CPU now via balance_push() or placed on a different CPU
8937 * during wakeup. Wait until the CPU is quiescent.
8939 static void balance_hotplug_wait(void)
8941 struct rq *rq = this_rq();
8943 rcuwait_wait_event(&rq->hotplug_wait,
8944 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8945 TASK_UNINTERRUPTIBLE);
8950 static inline void balance_push(struct rq *rq)
8954 static inline void balance_push_set(int cpu, bool on)
8958 static inline void balance_hotplug_wait(void)
8962 #endif /* CONFIG_HOTPLUG_CPU */
8964 void set_rq_online(struct rq *rq)
8967 const struct sched_class *class;
8969 cpumask_set_cpu(rq->cpu, rq->rd->online);
8972 for_each_class(class) {
8973 if (class->rq_online)
8974 class->rq_online(rq);
8979 void set_rq_offline(struct rq *rq)
8982 const struct sched_class *class;
8984 for_each_class(class) {
8985 if (class->rq_offline)
8986 class->rq_offline(rq);
8989 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8995 * used to mark begin/end of suspend/resume:
8997 static int num_cpus_frozen;
9000 * Update cpusets according to cpu_active mask. If cpusets are
9001 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9002 * around partition_sched_domains().
9004 * If we come here as part of a suspend/resume, don't touch cpusets because we
9005 * want to restore it back to its original state upon resume anyway.
9007 static void cpuset_cpu_active(void)
9009 if (cpuhp_tasks_frozen) {
9011 * num_cpus_frozen tracks how many CPUs are involved in suspend
9012 * resume sequence. As long as this is not the last online
9013 * operation in the resume sequence, just build a single sched
9014 * domain, ignoring cpusets.
9016 partition_sched_domains(1, NULL, NULL);
9017 if (--num_cpus_frozen)
9020 * This is the last CPU online operation. So fall through and
9021 * restore the original sched domains by considering the
9022 * cpuset configurations.
9024 cpuset_force_rebuild();
9026 cpuset_update_active_cpus();
9029 static int cpuset_cpu_inactive(unsigned int cpu)
9031 if (!cpuhp_tasks_frozen) {
9032 if (dl_cpu_busy(cpu))
9034 cpuset_update_active_cpus();
9037 partition_sched_domains(1, NULL, NULL);
9042 int sched_cpu_activate(unsigned int cpu)
9044 struct rq *rq = cpu_rq(cpu);
9048 * Clear the balance_push callback and prepare to schedule
9051 balance_push_set(cpu, false);
9053 #ifdef CONFIG_SCHED_SMT
9055 * When going up, increment the number of cores with SMT present.
9057 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9058 static_branch_inc_cpuslocked(&sched_smt_present);
9060 set_cpu_active(cpu, true);
9062 if (sched_smp_initialized) {
9063 sched_domains_numa_masks_set(cpu);
9064 cpuset_cpu_active();
9068 * Put the rq online, if not already. This happens:
9070 * 1) In the early boot process, because we build the real domains
9071 * after all CPUs have been brought up.
9073 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9076 rq_lock_irqsave(rq, &rf);
9078 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9081 rq_unlock_irqrestore(rq, &rf);
9086 int sched_cpu_deactivate(unsigned int cpu)
9088 struct rq *rq = cpu_rq(cpu);
9093 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9094 * load balancing when not active
9096 nohz_balance_exit_idle(rq);
9098 set_cpu_active(cpu, false);
9101 * From this point forward, this CPU will refuse to run any task that
9102 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9103 * push those tasks away until this gets cleared, see
9104 * sched_cpu_dying().
9106 balance_push_set(cpu, true);
9109 * We've cleared cpu_active_mask / set balance_push, wait for all
9110 * preempt-disabled and RCU users of this state to go away such that
9111 * all new such users will observe it.
9113 * Specifically, we rely on ttwu to no longer target this CPU, see
9114 * ttwu_queue_cond() and is_cpu_allowed().
9116 * Do sync before park smpboot threads to take care the rcu boost case.
9120 rq_lock_irqsave(rq, &rf);
9122 update_rq_clock(rq);
9123 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9126 rq_unlock_irqrestore(rq, &rf);
9128 #ifdef CONFIG_SCHED_SMT
9130 * When going down, decrement the number of cores with SMT present.
9132 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9133 static_branch_dec_cpuslocked(&sched_smt_present);
9135 sched_core_cpu_deactivate(cpu);
9138 if (!sched_smp_initialized)
9141 ret = cpuset_cpu_inactive(cpu);
9143 balance_push_set(cpu, false);
9144 set_cpu_active(cpu, true);
9147 sched_domains_numa_masks_clear(cpu);
9151 static void sched_rq_cpu_starting(unsigned int cpu)
9153 struct rq *rq = cpu_rq(cpu);
9155 rq->calc_load_update = calc_load_update;
9156 update_max_interval();
9159 int sched_cpu_starting(unsigned int cpu)
9161 sched_core_cpu_starting(cpu);
9162 sched_rq_cpu_starting(cpu);
9163 sched_tick_start(cpu);
9167 #ifdef CONFIG_HOTPLUG_CPU
9170 * Invoked immediately before the stopper thread is invoked to bring the
9171 * CPU down completely. At this point all per CPU kthreads except the
9172 * hotplug thread (current) and the stopper thread (inactive) have been
9173 * either parked or have been unbound from the outgoing CPU. Ensure that
9174 * any of those which might be on the way out are gone.
9176 * If after this point a bound task is being woken on this CPU then the
9177 * responsible hotplug callback has failed to do it's job.
9178 * sched_cpu_dying() will catch it with the appropriate fireworks.
9180 int sched_cpu_wait_empty(unsigned int cpu)
9182 balance_hotplug_wait();
9187 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9188 * might have. Called from the CPU stopper task after ensuring that the
9189 * stopper is the last running task on the CPU, so nr_active count is
9190 * stable. We need to take the teardown thread which is calling this into
9191 * account, so we hand in adjust = 1 to the load calculation.
9193 * Also see the comment "Global load-average calculations".
9195 static void calc_load_migrate(struct rq *rq)
9197 long delta = calc_load_fold_active(rq, 1);
9200 atomic_long_add(delta, &calc_load_tasks);
9203 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9205 struct task_struct *g, *p;
9206 int cpu = cpu_of(rq);
9208 lockdep_assert_rq_held(rq);
9210 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9211 for_each_process_thread(g, p) {
9212 if (task_cpu(p) != cpu)
9215 if (!task_on_rq_queued(p))
9218 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9222 int sched_cpu_dying(unsigned int cpu)
9224 struct rq *rq = cpu_rq(cpu);
9227 /* Handle pending wakeups and then migrate everything off */
9228 sched_tick_stop(cpu);
9230 rq_lock_irqsave(rq, &rf);
9231 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9232 WARN(true, "Dying CPU not properly vacated!");
9233 dump_rq_tasks(rq, KERN_WARNING);
9235 rq_unlock_irqrestore(rq, &rf);
9237 calc_load_migrate(rq);
9238 update_max_interval();
9240 sched_core_cpu_dying(cpu);
9245 void __init sched_init_smp(void)
9250 * There's no userspace yet to cause hotplug operations; hence all the
9251 * CPU masks are stable and all blatant races in the below code cannot
9254 mutex_lock(&sched_domains_mutex);
9255 sched_init_domains(cpu_active_mask);
9256 mutex_unlock(&sched_domains_mutex);
9258 /* Move init over to a non-isolated CPU */
9259 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9261 current->flags &= ~PF_NO_SETAFFINITY;
9262 sched_init_granularity();
9264 init_sched_rt_class();
9265 init_sched_dl_class();
9267 sched_smp_initialized = true;
9270 static int __init migration_init(void)
9272 sched_cpu_starting(smp_processor_id());
9275 early_initcall(migration_init);
9278 void __init sched_init_smp(void)
9280 sched_init_granularity();
9282 #endif /* CONFIG_SMP */
9284 int in_sched_functions(unsigned long addr)
9286 return in_lock_functions(addr) ||
9287 (addr >= (unsigned long)__sched_text_start
9288 && addr < (unsigned long)__sched_text_end);
9291 #ifdef CONFIG_CGROUP_SCHED
9293 * Default task group.
9294 * Every task in system belongs to this group at bootup.
9296 struct task_group root_task_group;
9297 LIST_HEAD(task_groups);
9299 /* Cacheline aligned slab cache for task_group */
9300 static struct kmem_cache *task_group_cache __read_mostly;
9303 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9304 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9306 void __init sched_init(void)
9308 unsigned long ptr = 0;
9311 /* Make sure the linker didn't screw up */
9312 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9313 &fair_sched_class + 1 != &rt_sched_class ||
9314 &rt_sched_class + 1 != &dl_sched_class);
9316 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9321 #ifdef CONFIG_FAIR_GROUP_SCHED
9322 ptr += 2 * nr_cpu_ids * sizeof(void **);
9324 #ifdef CONFIG_RT_GROUP_SCHED
9325 ptr += 2 * nr_cpu_ids * sizeof(void **);
9328 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9330 #ifdef CONFIG_FAIR_GROUP_SCHED
9331 root_task_group.se = (struct sched_entity **)ptr;
9332 ptr += nr_cpu_ids * sizeof(void **);
9334 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9335 ptr += nr_cpu_ids * sizeof(void **);
9337 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9338 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9339 #endif /* CONFIG_FAIR_GROUP_SCHED */
9340 #ifdef CONFIG_RT_GROUP_SCHED
9341 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9342 ptr += nr_cpu_ids * sizeof(void **);
9344 root_task_group.rt_rq = (struct rt_rq **)ptr;
9345 ptr += nr_cpu_ids * sizeof(void **);
9347 #endif /* CONFIG_RT_GROUP_SCHED */
9349 #ifdef CONFIG_CPUMASK_OFFSTACK
9350 for_each_possible_cpu(i) {
9351 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9352 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9353 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9354 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9356 #endif /* CONFIG_CPUMASK_OFFSTACK */
9358 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9359 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9362 init_defrootdomain();
9365 #ifdef CONFIG_RT_GROUP_SCHED
9366 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9367 global_rt_period(), global_rt_runtime());
9368 #endif /* CONFIG_RT_GROUP_SCHED */
9370 #ifdef CONFIG_CGROUP_SCHED
9371 task_group_cache = KMEM_CACHE(task_group, 0);
9373 list_add(&root_task_group.list, &task_groups);
9374 INIT_LIST_HEAD(&root_task_group.children);
9375 INIT_LIST_HEAD(&root_task_group.siblings);
9376 autogroup_init(&init_task);
9377 #endif /* CONFIG_CGROUP_SCHED */
9379 for_each_possible_cpu(i) {
9383 raw_spin_lock_init(&rq->__lock);
9385 rq->calc_load_active = 0;
9386 rq->calc_load_update = jiffies + LOAD_FREQ;
9387 init_cfs_rq(&rq->cfs);
9388 init_rt_rq(&rq->rt);
9389 init_dl_rq(&rq->dl);
9390 #ifdef CONFIG_FAIR_GROUP_SCHED
9391 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9392 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9394 * How much CPU bandwidth does root_task_group get?
9396 * In case of task-groups formed thr' the cgroup filesystem, it
9397 * gets 100% of the CPU resources in the system. This overall
9398 * system CPU resource is divided among the tasks of
9399 * root_task_group and its child task-groups in a fair manner,
9400 * based on each entity's (task or task-group's) weight
9401 * (se->load.weight).
9403 * In other words, if root_task_group has 10 tasks of weight
9404 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9405 * then A0's share of the CPU resource is:
9407 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9409 * We achieve this by letting root_task_group's tasks sit
9410 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9412 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9413 #endif /* CONFIG_FAIR_GROUP_SCHED */
9415 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9416 #ifdef CONFIG_RT_GROUP_SCHED
9417 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9422 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9423 rq->balance_callback = &balance_push_callback;
9424 rq->active_balance = 0;
9425 rq->next_balance = jiffies;
9430 rq->avg_idle = 2*sysctl_sched_migration_cost;
9431 rq->wake_stamp = jiffies;
9432 rq->wake_avg_idle = rq->avg_idle;
9433 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9435 INIT_LIST_HEAD(&rq->cfs_tasks);
9437 rq_attach_root(rq, &def_root_domain);
9438 #ifdef CONFIG_NO_HZ_COMMON
9439 rq->last_blocked_load_update_tick = jiffies;
9440 atomic_set(&rq->nohz_flags, 0);
9442 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9444 #ifdef CONFIG_HOTPLUG_CPU
9445 rcuwait_init(&rq->hotplug_wait);
9447 #endif /* CONFIG_SMP */
9449 atomic_set(&rq->nr_iowait, 0);
9451 #ifdef CONFIG_SCHED_CORE
9453 rq->core_pick = NULL;
9454 rq->core_enabled = 0;
9455 rq->core_tree = RB_ROOT;
9456 rq->core_forceidle_count = 0;
9457 rq->core_forceidle_occupation = 0;
9458 rq->core_forceidle_start = 0;
9460 rq->core_cookie = 0UL;
9464 set_load_weight(&init_task, false);
9467 * The boot idle thread does lazy MMU switching as well:
9470 enter_lazy_tlb(&init_mm, current);
9473 * The idle task doesn't need the kthread struct to function, but it
9474 * is dressed up as a per-CPU kthread and thus needs to play the part
9475 * if we want to avoid special-casing it in code that deals with per-CPU
9478 WARN_ON(!set_kthread_struct(current));
9481 * Make us the idle thread. Technically, schedule() should not be
9482 * called from this thread, however somewhere below it might be,
9483 * but because we are the idle thread, we just pick up running again
9484 * when this runqueue becomes "idle".
9486 init_idle(current, smp_processor_id());
9488 calc_load_update = jiffies + LOAD_FREQ;
9491 idle_thread_set_boot_cpu();
9492 balance_push_set(smp_processor_id(), false);
9494 init_sched_fair_class();
9500 preempt_dynamic_init();
9502 scheduler_running = 1;
9505 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9507 void __might_sleep(const char *file, int line)
9509 unsigned int state = get_current_state();
9511 * Blocking primitives will set (and therefore destroy) current->state,
9512 * since we will exit with TASK_RUNNING make sure we enter with it,
9513 * otherwise we will destroy state.
9515 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9516 "do not call blocking ops when !TASK_RUNNING; "
9517 "state=%x set at [<%p>] %pS\n", state,
9518 (void *)current->task_state_change,
9519 (void *)current->task_state_change);
9521 __might_resched(file, line, 0);
9523 EXPORT_SYMBOL(__might_sleep);
9525 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9527 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9530 if (preempt_count() == preempt_offset)
9533 pr_err("Preemption disabled at:");
9534 print_ip_sym(KERN_ERR, ip);
9537 static inline bool resched_offsets_ok(unsigned int offsets)
9539 unsigned int nested = preempt_count();
9541 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9543 return nested == offsets;
9546 void __might_resched(const char *file, int line, unsigned int offsets)
9548 /* Ratelimiting timestamp: */
9549 static unsigned long prev_jiffy;
9551 unsigned long preempt_disable_ip;
9553 /* WARN_ON_ONCE() by default, no rate limit required: */
9556 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9557 !is_idle_task(current) && !current->non_block_count) ||
9558 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9562 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9564 prev_jiffy = jiffies;
9566 /* Save this before calling printk(), since that will clobber it: */
9567 preempt_disable_ip = get_preempt_disable_ip(current);
9569 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9571 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9572 in_atomic(), irqs_disabled(), current->non_block_count,
9573 current->pid, current->comm);
9574 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9575 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9577 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9578 pr_err("RCU nest depth: %d, expected: %u\n",
9579 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9582 if (task_stack_end_corrupted(current))
9583 pr_emerg("Thread overran stack, or stack corrupted\n");
9585 debug_show_held_locks(current);
9586 if (irqs_disabled())
9587 print_irqtrace_events(current);
9589 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9590 preempt_disable_ip);
9593 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9595 EXPORT_SYMBOL(__might_resched);
9597 void __cant_sleep(const char *file, int line, int preempt_offset)
9599 static unsigned long prev_jiffy;
9601 if (irqs_disabled())
9604 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9607 if (preempt_count() > preempt_offset)
9610 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9612 prev_jiffy = jiffies;
9614 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9615 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9616 in_atomic(), irqs_disabled(),
9617 current->pid, current->comm);
9619 debug_show_held_locks(current);
9621 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9623 EXPORT_SYMBOL_GPL(__cant_sleep);
9626 void __cant_migrate(const char *file, int line)
9628 static unsigned long prev_jiffy;
9630 if (irqs_disabled())
9633 if (is_migration_disabled(current))
9636 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9639 if (preempt_count() > 0)
9642 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9644 prev_jiffy = jiffies;
9646 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9647 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9648 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9649 current->pid, current->comm);
9651 debug_show_held_locks(current);
9653 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9655 EXPORT_SYMBOL_GPL(__cant_migrate);
9659 #ifdef CONFIG_MAGIC_SYSRQ
9660 void normalize_rt_tasks(void)
9662 struct task_struct *g, *p;
9663 struct sched_attr attr = {
9664 .sched_policy = SCHED_NORMAL,
9667 read_lock(&tasklist_lock);
9668 for_each_process_thread(g, p) {
9670 * Only normalize user tasks:
9672 if (p->flags & PF_KTHREAD)
9675 p->se.exec_start = 0;
9676 schedstat_set(p->stats.wait_start, 0);
9677 schedstat_set(p->stats.sleep_start, 0);
9678 schedstat_set(p->stats.block_start, 0);
9680 if (!dl_task(p) && !rt_task(p)) {
9682 * Renice negative nice level userspace
9685 if (task_nice(p) < 0)
9686 set_user_nice(p, 0);
9690 __sched_setscheduler(p, &attr, false, false);
9692 read_unlock(&tasklist_lock);
9695 #endif /* CONFIG_MAGIC_SYSRQ */
9697 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9699 * These functions are only useful for the IA64 MCA handling, or kdb.
9701 * They can only be called when the whole system has been
9702 * stopped - every CPU needs to be quiescent, and no scheduling
9703 * activity can take place. Using them for anything else would
9704 * be a serious bug, and as a result, they aren't even visible
9705 * under any other configuration.
9709 * curr_task - return the current task for a given CPU.
9710 * @cpu: the processor in question.
9712 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9714 * Return: The current task for @cpu.
9716 struct task_struct *curr_task(int cpu)
9718 return cpu_curr(cpu);
9721 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9725 * ia64_set_curr_task - set the current task for a given CPU.
9726 * @cpu: the processor in question.
9727 * @p: the task pointer to set.
9729 * Description: This function must only be used when non-maskable interrupts
9730 * are serviced on a separate stack. It allows the architecture to switch the
9731 * notion of the current task on a CPU in a non-blocking manner. This function
9732 * must be called with all CPU's synchronized, and interrupts disabled, the
9733 * and caller must save the original value of the current task (see
9734 * curr_task() above) and restore that value before reenabling interrupts and
9735 * re-starting the system.
9737 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9739 void ia64_set_curr_task(int cpu, struct task_struct *p)
9746 #ifdef CONFIG_CGROUP_SCHED
9747 /* task_group_lock serializes the addition/removal of task groups */
9748 static DEFINE_SPINLOCK(task_group_lock);
9750 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9751 struct task_group *parent)
9753 #ifdef CONFIG_UCLAMP_TASK_GROUP
9754 enum uclamp_id clamp_id;
9756 for_each_clamp_id(clamp_id) {
9757 uclamp_se_set(&tg->uclamp_req[clamp_id],
9758 uclamp_none(clamp_id), false);
9759 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9764 static void sched_free_group(struct task_group *tg)
9766 free_fair_sched_group(tg);
9767 free_rt_sched_group(tg);
9769 kmem_cache_free(task_group_cache, tg);
9772 static void sched_free_group_rcu(struct rcu_head *rcu)
9774 sched_free_group(container_of(rcu, struct task_group, rcu));
9777 static void sched_unregister_group(struct task_group *tg)
9779 unregister_fair_sched_group(tg);
9780 unregister_rt_sched_group(tg);
9782 * We have to wait for yet another RCU grace period to expire, as
9783 * print_cfs_stats() might run concurrently.
9785 call_rcu(&tg->rcu, sched_free_group_rcu);
9788 /* allocate runqueue etc for a new task group */
9789 struct task_group *sched_create_group(struct task_group *parent)
9791 struct task_group *tg;
9793 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9795 return ERR_PTR(-ENOMEM);
9797 if (!alloc_fair_sched_group(tg, parent))
9800 if (!alloc_rt_sched_group(tg, parent))
9803 alloc_uclamp_sched_group(tg, parent);
9808 sched_free_group(tg);
9809 return ERR_PTR(-ENOMEM);
9812 void sched_online_group(struct task_group *tg, struct task_group *parent)
9814 unsigned long flags;
9816 spin_lock_irqsave(&task_group_lock, flags);
9817 list_add_rcu(&tg->list, &task_groups);
9819 /* Root should already exist: */
9822 tg->parent = parent;
9823 INIT_LIST_HEAD(&tg->children);
9824 list_add_rcu(&tg->siblings, &parent->children);
9825 spin_unlock_irqrestore(&task_group_lock, flags);
9827 online_fair_sched_group(tg);
9830 /* rcu callback to free various structures associated with a task group */
9831 static void sched_unregister_group_rcu(struct rcu_head *rhp)
9833 /* Now it should be safe to free those cfs_rqs: */
9834 sched_unregister_group(container_of(rhp, struct task_group, rcu));
9837 void sched_destroy_group(struct task_group *tg)
9839 /* Wait for possible concurrent references to cfs_rqs complete: */
9840 call_rcu(&tg->rcu, sched_unregister_group_rcu);
9843 void sched_release_group(struct task_group *tg)
9845 unsigned long flags;
9848 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
9849 * sched_cfs_period_timer()).
9851 * For this to be effective, we have to wait for all pending users of
9852 * this task group to leave their RCU critical section to ensure no new
9853 * user will see our dying task group any more. Specifically ensure
9854 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
9856 * We therefore defer calling unregister_fair_sched_group() to
9857 * sched_unregister_group() which is guarantied to get called only after the
9858 * current RCU grace period has expired.
9860 spin_lock_irqsave(&task_group_lock, flags);
9861 list_del_rcu(&tg->list);
9862 list_del_rcu(&tg->siblings);
9863 spin_unlock_irqrestore(&task_group_lock, flags);
9866 static void sched_change_group(struct task_struct *tsk, int type)
9868 struct task_group *tg;
9871 * All callers are synchronized by task_rq_lock(); we do not use RCU
9872 * which is pointless here. Thus, we pass "true" to task_css_check()
9873 * to prevent lockdep warnings.
9875 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9876 struct task_group, css);
9877 tg = autogroup_task_group(tsk, tg);
9878 tsk->sched_task_group = tg;
9880 #ifdef CONFIG_FAIR_GROUP_SCHED
9881 if (tsk->sched_class->task_change_group)
9882 tsk->sched_class->task_change_group(tsk, type);
9885 set_task_rq(tsk, task_cpu(tsk));
9889 * Change task's runqueue when it moves between groups.
9891 * The caller of this function should have put the task in its new group by
9892 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9895 void sched_move_task(struct task_struct *tsk)
9897 int queued, running, queue_flags =
9898 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9902 rq = task_rq_lock(tsk, &rf);
9903 update_rq_clock(rq);
9905 running = task_current(rq, tsk);
9906 queued = task_on_rq_queued(tsk);
9909 dequeue_task(rq, tsk, queue_flags);
9911 put_prev_task(rq, tsk);
9913 sched_change_group(tsk, TASK_MOVE_GROUP);
9916 enqueue_task(rq, tsk, queue_flags);
9918 set_next_task(rq, tsk);
9920 * After changing group, the running task may have joined a
9921 * throttled one but it's still the running task. Trigger a
9922 * resched to make sure that task can still run.
9927 task_rq_unlock(rq, tsk, &rf);
9930 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9932 return css ? container_of(css, struct task_group, css) : NULL;
9935 static struct cgroup_subsys_state *
9936 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9938 struct task_group *parent = css_tg(parent_css);
9939 struct task_group *tg;
9942 /* This is early initialization for the top cgroup */
9943 return &root_task_group.css;
9946 tg = sched_create_group(parent);
9948 return ERR_PTR(-ENOMEM);
9953 /* Expose task group only after completing cgroup initialization */
9954 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9956 struct task_group *tg = css_tg(css);
9957 struct task_group *parent = css_tg(css->parent);
9960 sched_online_group(tg, parent);
9962 #ifdef CONFIG_UCLAMP_TASK_GROUP
9963 /* Propagate the effective uclamp value for the new group */
9964 mutex_lock(&uclamp_mutex);
9966 cpu_util_update_eff(css);
9968 mutex_unlock(&uclamp_mutex);
9974 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9976 struct task_group *tg = css_tg(css);
9978 sched_release_group(tg);
9981 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9983 struct task_group *tg = css_tg(css);
9986 * Relies on the RCU grace period between css_released() and this.
9988 sched_unregister_group(tg);
9992 * This is called before wake_up_new_task(), therefore we really only
9993 * have to set its group bits, all the other stuff does not apply.
9995 static void cpu_cgroup_fork(struct task_struct *task)
10000 rq = task_rq_lock(task, &rf);
10002 update_rq_clock(rq);
10003 sched_change_group(task, TASK_SET_GROUP);
10005 task_rq_unlock(rq, task, &rf);
10008 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10010 struct task_struct *task;
10011 struct cgroup_subsys_state *css;
10014 cgroup_taskset_for_each(task, css, tset) {
10015 #ifdef CONFIG_RT_GROUP_SCHED
10016 if (!sched_rt_can_attach(css_tg(css), task))
10020 * Serialize against wake_up_new_task() such that if it's
10021 * running, we're sure to observe its full state.
10023 raw_spin_lock_irq(&task->pi_lock);
10025 * Avoid calling sched_move_task() before wake_up_new_task()
10026 * has happened. This would lead to problems with PELT, due to
10027 * move wanting to detach+attach while we're not attached yet.
10029 if (READ_ONCE(task->__state) == TASK_NEW)
10031 raw_spin_unlock_irq(&task->pi_lock);
10039 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10041 struct task_struct *task;
10042 struct cgroup_subsys_state *css;
10044 cgroup_taskset_for_each(task, css, tset)
10045 sched_move_task(task);
10048 #ifdef CONFIG_UCLAMP_TASK_GROUP
10049 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10051 struct cgroup_subsys_state *top_css = css;
10052 struct uclamp_se *uc_parent = NULL;
10053 struct uclamp_se *uc_se = NULL;
10054 unsigned int eff[UCLAMP_CNT];
10055 enum uclamp_id clamp_id;
10056 unsigned int clamps;
10058 lockdep_assert_held(&uclamp_mutex);
10059 SCHED_WARN_ON(!rcu_read_lock_held());
10061 css_for_each_descendant_pre(css, top_css) {
10062 uc_parent = css_tg(css)->parent
10063 ? css_tg(css)->parent->uclamp : NULL;
10065 for_each_clamp_id(clamp_id) {
10066 /* Assume effective clamps matches requested clamps */
10067 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10068 /* Cap effective clamps with parent's effective clamps */
10070 eff[clamp_id] > uc_parent[clamp_id].value) {
10071 eff[clamp_id] = uc_parent[clamp_id].value;
10074 /* Ensure protection is always capped by limit */
10075 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10077 /* Propagate most restrictive effective clamps */
10079 uc_se = css_tg(css)->uclamp;
10080 for_each_clamp_id(clamp_id) {
10081 if (eff[clamp_id] == uc_se[clamp_id].value)
10083 uc_se[clamp_id].value = eff[clamp_id];
10084 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10085 clamps |= (0x1 << clamp_id);
10088 css = css_rightmost_descendant(css);
10092 /* Immediately update descendants RUNNABLE tasks */
10093 uclamp_update_active_tasks(css);
10098 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10099 * C expression. Since there is no way to convert a macro argument (N) into a
10100 * character constant, use two levels of macros.
10102 #define _POW10(exp) ((unsigned int)1e##exp)
10103 #define POW10(exp) _POW10(exp)
10105 struct uclamp_request {
10106 #define UCLAMP_PERCENT_SHIFT 2
10107 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10113 static inline struct uclamp_request
10114 capacity_from_percent(char *buf)
10116 struct uclamp_request req = {
10117 .percent = UCLAMP_PERCENT_SCALE,
10118 .util = SCHED_CAPACITY_SCALE,
10123 if (strcmp(buf, "max")) {
10124 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10128 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10133 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10134 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10140 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10141 size_t nbytes, loff_t off,
10142 enum uclamp_id clamp_id)
10144 struct uclamp_request req;
10145 struct task_group *tg;
10147 req = capacity_from_percent(buf);
10151 static_branch_enable(&sched_uclamp_used);
10153 mutex_lock(&uclamp_mutex);
10156 tg = css_tg(of_css(of));
10157 if (tg->uclamp_req[clamp_id].value != req.util)
10158 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10161 * Because of not recoverable conversion rounding we keep track of the
10162 * exact requested value
10164 tg->uclamp_pct[clamp_id] = req.percent;
10166 /* Update effective clamps to track the most restrictive value */
10167 cpu_util_update_eff(of_css(of));
10170 mutex_unlock(&uclamp_mutex);
10175 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10176 char *buf, size_t nbytes,
10179 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10182 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10183 char *buf, size_t nbytes,
10186 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10189 static inline void cpu_uclamp_print(struct seq_file *sf,
10190 enum uclamp_id clamp_id)
10192 struct task_group *tg;
10198 tg = css_tg(seq_css(sf));
10199 util_clamp = tg->uclamp_req[clamp_id].value;
10202 if (util_clamp == SCHED_CAPACITY_SCALE) {
10203 seq_puts(sf, "max\n");
10207 percent = tg->uclamp_pct[clamp_id];
10208 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10209 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10212 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10214 cpu_uclamp_print(sf, UCLAMP_MIN);
10218 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10220 cpu_uclamp_print(sf, UCLAMP_MAX);
10223 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10225 #ifdef CONFIG_FAIR_GROUP_SCHED
10226 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10227 struct cftype *cftype, u64 shareval)
10229 if (shareval > scale_load_down(ULONG_MAX))
10230 shareval = MAX_SHARES;
10231 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10234 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10235 struct cftype *cft)
10237 struct task_group *tg = css_tg(css);
10239 return (u64) scale_load_down(tg->shares);
10242 #ifdef CONFIG_CFS_BANDWIDTH
10243 static DEFINE_MUTEX(cfs_constraints_mutex);
10245 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10246 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10247 /* More than 203 days if BW_SHIFT equals 20. */
10248 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10250 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10252 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10255 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10256 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10258 if (tg == &root_task_group)
10262 * Ensure we have at some amount of bandwidth every period. This is
10263 * to prevent reaching a state of large arrears when throttled via
10264 * entity_tick() resulting in prolonged exit starvation.
10266 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10270 * Likewise, bound things on the other side by preventing insane quota
10271 * periods. This also allows us to normalize in computing quota
10274 if (period > max_cfs_quota_period)
10278 * Bound quota to defend quota against overflow during bandwidth shift.
10280 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10283 if (quota != RUNTIME_INF && (burst > quota ||
10284 burst + quota > max_cfs_runtime))
10288 * Prevent race between setting of cfs_rq->runtime_enabled and
10289 * unthrottle_offline_cfs_rqs().
10292 mutex_lock(&cfs_constraints_mutex);
10293 ret = __cfs_schedulable(tg, period, quota);
10297 runtime_enabled = quota != RUNTIME_INF;
10298 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10300 * If we need to toggle cfs_bandwidth_used, off->on must occur
10301 * before making related changes, and on->off must occur afterwards
10303 if (runtime_enabled && !runtime_was_enabled)
10304 cfs_bandwidth_usage_inc();
10305 raw_spin_lock_irq(&cfs_b->lock);
10306 cfs_b->period = ns_to_ktime(period);
10307 cfs_b->quota = quota;
10308 cfs_b->burst = burst;
10310 __refill_cfs_bandwidth_runtime(cfs_b);
10312 /* Restart the period timer (if active) to handle new period expiry: */
10313 if (runtime_enabled)
10314 start_cfs_bandwidth(cfs_b);
10316 raw_spin_unlock_irq(&cfs_b->lock);
10318 for_each_online_cpu(i) {
10319 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10320 struct rq *rq = cfs_rq->rq;
10321 struct rq_flags rf;
10323 rq_lock_irq(rq, &rf);
10324 cfs_rq->runtime_enabled = runtime_enabled;
10325 cfs_rq->runtime_remaining = 0;
10327 if (cfs_rq->throttled)
10328 unthrottle_cfs_rq(cfs_rq);
10329 rq_unlock_irq(rq, &rf);
10331 if (runtime_was_enabled && !runtime_enabled)
10332 cfs_bandwidth_usage_dec();
10334 mutex_unlock(&cfs_constraints_mutex);
10335 cpus_read_unlock();
10340 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10342 u64 quota, period, burst;
10344 period = ktime_to_ns(tg->cfs_bandwidth.period);
10345 burst = tg->cfs_bandwidth.burst;
10346 if (cfs_quota_us < 0)
10347 quota = RUNTIME_INF;
10348 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10349 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10353 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10356 static long tg_get_cfs_quota(struct task_group *tg)
10360 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10363 quota_us = tg->cfs_bandwidth.quota;
10364 do_div(quota_us, NSEC_PER_USEC);
10369 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10371 u64 quota, period, burst;
10373 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10376 period = (u64)cfs_period_us * NSEC_PER_USEC;
10377 quota = tg->cfs_bandwidth.quota;
10378 burst = tg->cfs_bandwidth.burst;
10380 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10383 static long tg_get_cfs_period(struct task_group *tg)
10387 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10388 do_div(cfs_period_us, NSEC_PER_USEC);
10390 return cfs_period_us;
10393 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10395 u64 quota, period, burst;
10397 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10400 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10401 period = ktime_to_ns(tg->cfs_bandwidth.period);
10402 quota = tg->cfs_bandwidth.quota;
10404 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10407 static long tg_get_cfs_burst(struct task_group *tg)
10411 burst_us = tg->cfs_bandwidth.burst;
10412 do_div(burst_us, NSEC_PER_USEC);
10417 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10418 struct cftype *cft)
10420 return tg_get_cfs_quota(css_tg(css));
10423 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10424 struct cftype *cftype, s64 cfs_quota_us)
10426 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10429 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10430 struct cftype *cft)
10432 return tg_get_cfs_period(css_tg(css));
10435 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10436 struct cftype *cftype, u64 cfs_period_us)
10438 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10441 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10442 struct cftype *cft)
10444 return tg_get_cfs_burst(css_tg(css));
10447 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10448 struct cftype *cftype, u64 cfs_burst_us)
10450 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10453 struct cfs_schedulable_data {
10454 struct task_group *tg;
10459 * normalize group quota/period to be quota/max_period
10460 * note: units are usecs
10462 static u64 normalize_cfs_quota(struct task_group *tg,
10463 struct cfs_schedulable_data *d)
10468 period = d->period;
10471 period = tg_get_cfs_period(tg);
10472 quota = tg_get_cfs_quota(tg);
10475 /* note: these should typically be equivalent */
10476 if (quota == RUNTIME_INF || quota == -1)
10477 return RUNTIME_INF;
10479 return to_ratio(period, quota);
10482 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10484 struct cfs_schedulable_data *d = data;
10485 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10486 s64 quota = 0, parent_quota = -1;
10489 quota = RUNTIME_INF;
10491 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10493 quota = normalize_cfs_quota(tg, d);
10494 parent_quota = parent_b->hierarchical_quota;
10497 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10498 * always take the min. On cgroup1, only inherit when no
10501 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10502 quota = min(quota, parent_quota);
10504 if (quota == RUNTIME_INF)
10505 quota = parent_quota;
10506 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10510 cfs_b->hierarchical_quota = quota;
10515 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10518 struct cfs_schedulable_data data = {
10524 if (quota != RUNTIME_INF) {
10525 do_div(data.period, NSEC_PER_USEC);
10526 do_div(data.quota, NSEC_PER_USEC);
10530 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10536 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10538 struct task_group *tg = css_tg(seq_css(sf));
10539 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10541 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10542 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10543 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10545 if (schedstat_enabled() && tg != &root_task_group) {
10546 struct sched_statistics *stats;
10550 for_each_possible_cpu(i) {
10551 stats = __schedstats_from_se(tg->se[i]);
10552 ws += schedstat_val(stats->wait_sum);
10555 seq_printf(sf, "wait_sum %llu\n", ws);
10558 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10559 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10563 #endif /* CONFIG_CFS_BANDWIDTH */
10564 #endif /* CONFIG_FAIR_GROUP_SCHED */
10566 #ifdef CONFIG_RT_GROUP_SCHED
10567 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10568 struct cftype *cft, s64 val)
10570 return sched_group_set_rt_runtime(css_tg(css), val);
10573 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10574 struct cftype *cft)
10576 return sched_group_rt_runtime(css_tg(css));
10579 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10580 struct cftype *cftype, u64 rt_period_us)
10582 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10585 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10586 struct cftype *cft)
10588 return sched_group_rt_period(css_tg(css));
10590 #endif /* CONFIG_RT_GROUP_SCHED */
10592 #ifdef CONFIG_FAIR_GROUP_SCHED
10593 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10594 struct cftype *cft)
10596 return css_tg(css)->idle;
10599 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10600 struct cftype *cft, s64 idle)
10602 return sched_group_set_idle(css_tg(css), idle);
10606 static struct cftype cpu_legacy_files[] = {
10607 #ifdef CONFIG_FAIR_GROUP_SCHED
10610 .read_u64 = cpu_shares_read_u64,
10611 .write_u64 = cpu_shares_write_u64,
10615 .read_s64 = cpu_idle_read_s64,
10616 .write_s64 = cpu_idle_write_s64,
10619 #ifdef CONFIG_CFS_BANDWIDTH
10621 .name = "cfs_quota_us",
10622 .read_s64 = cpu_cfs_quota_read_s64,
10623 .write_s64 = cpu_cfs_quota_write_s64,
10626 .name = "cfs_period_us",
10627 .read_u64 = cpu_cfs_period_read_u64,
10628 .write_u64 = cpu_cfs_period_write_u64,
10631 .name = "cfs_burst_us",
10632 .read_u64 = cpu_cfs_burst_read_u64,
10633 .write_u64 = cpu_cfs_burst_write_u64,
10637 .seq_show = cpu_cfs_stat_show,
10640 #ifdef CONFIG_RT_GROUP_SCHED
10642 .name = "rt_runtime_us",
10643 .read_s64 = cpu_rt_runtime_read,
10644 .write_s64 = cpu_rt_runtime_write,
10647 .name = "rt_period_us",
10648 .read_u64 = cpu_rt_period_read_uint,
10649 .write_u64 = cpu_rt_period_write_uint,
10652 #ifdef CONFIG_UCLAMP_TASK_GROUP
10654 .name = "uclamp.min",
10655 .flags = CFTYPE_NOT_ON_ROOT,
10656 .seq_show = cpu_uclamp_min_show,
10657 .write = cpu_uclamp_min_write,
10660 .name = "uclamp.max",
10661 .flags = CFTYPE_NOT_ON_ROOT,
10662 .seq_show = cpu_uclamp_max_show,
10663 .write = cpu_uclamp_max_write,
10666 { } /* Terminate */
10669 static int cpu_extra_stat_show(struct seq_file *sf,
10670 struct cgroup_subsys_state *css)
10672 #ifdef CONFIG_CFS_BANDWIDTH
10674 struct task_group *tg = css_tg(css);
10675 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10676 u64 throttled_usec, burst_usec;
10678 throttled_usec = cfs_b->throttled_time;
10679 do_div(throttled_usec, NSEC_PER_USEC);
10680 burst_usec = cfs_b->burst_time;
10681 do_div(burst_usec, NSEC_PER_USEC);
10683 seq_printf(sf, "nr_periods %d\n"
10684 "nr_throttled %d\n"
10685 "throttled_usec %llu\n"
10687 "burst_usec %llu\n",
10688 cfs_b->nr_periods, cfs_b->nr_throttled,
10689 throttled_usec, cfs_b->nr_burst, burst_usec);
10695 #ifdef CONFIG_FAIR_GROUP_SCHED
10696 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10697 struct cftype *cft)
10699 struct task_group *tg = css_tg(css);
10700 u64 weight = scale_load_down(tg->shares);
10702 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10705 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10706 struct cftype *cft, u64 weight)
10709 * cgroup weight knobs should use the common MIN, DFL and MAX
10710 * values which are 1, 100 and 10000 respectively. While it loses
10711 * a bit of range on both ends, it maps pretty well onto the shares
10712 * value used by scheduler and the round-trip conversions preserve
10713 * the original value over the entire range.
10715 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10718 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10720 return sched_group_set_shares(css_tg(css), scale_load(weight));
10723 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10724 struct cftype *cft)
10726 unsigned long weight = scale_load_down(css_tg(css)->shares);
10727 int last_delta = INT_MAX;
10730 /* find the closest nice value to the current weight */
10731 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10732 delta = abs(sched_prio_to_weight[prio] - weight);
10733 if (delta >= last_delta)
10735 last_delta = delta;
10738 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10741 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10742 struct cftype *cft, s64 nice)
10744 unsigned long weight;
10747 if (nice < MIN_NICE || nice > MAX_NICE)
10750 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10751 idx = array_index_nospec(idx, 40);
10752 weight = sched_prio_to_weight[idx];
10754 return sched_group_set_shares(css_tg(css), scale_load(weight));
10758 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10759 long period, long quota)
10762 seq_puts(sf, "max");
10764 seq_printf(sf, "%ld", quota);
10766 seq_printf(sf, " %ld\n", period);
10769 /* caller should put the current value in *@periodp before calling */
10770 static int __maybe_unused cpu_period_quota_parse(char *buf,
10771 u64 *periodp, u64 *quotap)
10773 char tok[21]; /* U64_MAX */
10775 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10778 *periodp *= NSEC_PER_USEC;
10780 if (sscanf(tok, "%llu", quotap))
10781 *quotap *= NSEC_PER_USEC;
10782 else if (!strcmp(tok, "max"))
10783 *quotap = RUNTIME_INF;
10790 #ifdef CONFIG_CFS_BANDWIDTH
10791 static int cpu_max_show(struct seq_file *sf, void *v)
10793 struct task_group *tg = css_tg(seq_css(sf));
10795 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10799 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10800 char *buf, size_t nbytes, loff_t off)
10802 struct task_group *tg = css_tg(of_css(of));
10803 u64 period = tg_get_cfs_period(tg);
10804 u64 burst = tg_get_cfs_burst(tg);
10808 ret = cpu_period_quota_parse(buf, &period, "a);
10810 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10811 return ret ?: nbytes;
10815 static struct cftype cpu_files[] = {
10816 #ifdef CONFIG_FAIR_GROUP_SCHED
10819 .flags = CFTYPE_NOT_ON_ROOT,
10820 .read_u64 = cpu_weight_read_u64,
10821 .write_u64 = cpu_weight_write_u64,
10824 .name = "weight.nice",
10825 .flags = CFTYPE_NOT_ON_ROOT,
10826 .read_s64 = cpu_weight_nice_read_s64,
10827 .write_s64 = cpu_weight_nice_write_s64,
10831 .flags = CFTYPE_NOT_ON_ROOT,
10832 .read_s64 = cpu_idle_read_s64,
10833 .write_s64 = cpu_idle_write_s64,
10836 #ifdef CONFIG_CFS_BANDWIDTH
10839 .flags = CFTYPE_NOT_ON_ROOT,
10840 .seq_show = cpu_max_show,
10841 .write = cpu_max_write,
10844 .name = "max.burst",
10845 .flags = CFTYPE_NOT_ON_ROOT,
10846 .read_u64 = cpu_cfs_burst_read_u64,
10847 .write_u64 = cpu_cfs_burst_write_u64,
10850 #ifdef CONFIG_UCLAMP_TASK_GROUP
10852 .name = "uclamp.min",
10853 .flags = CFTYPE_NOT_ON_ROOT,
10854 .seq_show = cpu_uclamp_min_show,
10855 .write = cpu_uclamp_min_write,
10858 .name = "uclamp.max",
10859 .flags = CFTYPE_NOT_ON_ROOT,
10860 .seq_show = cpu_uclamp_max_show,
10861 .write = cpu_uclamp_max_write,
10864 { } /* terminate */
10867 struct cgroup_subsys cpu_cgrp_subsys = {
10868 .css_alloc = cpu_cgroup_css_alloc,
10869 .css_online = cpu_cgroup_css_online,
10870 .css_released = cpu_cgroup_css_released,
10871 .css_free = cpu_cgroup_css_free,
10872 .css_extra_stat_show = cpu_extra_stat_show,
10873 .fork = cpu_cgroup_fork,
10874 .can_attach = cpu_cgroup_can_attach,
10875 .attach = cpu_cgroup_attach,
10876 .legacy_cftypes = cpu_legacy_files,
10877 .dfl_cftypes = cpu_files,
10878 .early_init = true,
10882 #endif /* CONFIG_CGROUP_SCHED */
10884 void dump_cpu_task(int cpu)
10886 pr_info("Task dump for CPU %d:\n", cpu);
10887 sched_show_task(cpu_curr(cpu));
10891 * Nice levels are multiplicative, with a gentle 10% change for every
10892 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10893 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10894 * that remained on nice 0.
10896 * The "10% effect" is relative and cumulative: from _any_ nice level,
10897 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10898 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10899 * If a task goes up by ~10% and another task goes down by ~10% then
10900 * the relative distance between them is ~25%.)
10902 const int sched_prio_to_weight[40] = {
10903 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10904 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10905 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10906 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10907 /* 0 */ 1024, 820, 655, 526, 423,
10908 /* 5 */ 335, 272, 215, 172, 137,
10909 /* 10 */ 110, 87, 70, 56, 45,
10910 /* 15 */ 36, 29, 23, 18, 15,
10914 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10916 * In cases where the weight does not change often, we can use the
10917 * precalculated inverse to speed up arithmetics by turning divisions
10918 * into multiplications:
10920 const u32 sched_prio_to_wmult[40] = {
10921 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10922 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10923 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10924 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10925 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10926 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10927 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10928 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10931 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10933 trace_sched_update_nr_running_tp(rq, count);