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(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
63 * Print a warning if need_resched is set for the given duration (if
64 * LATENCY_WARN is enabled).
66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
69 __read_mostly int sysctl_resched_latency_warn_ms = 100;
70 __read_mostly int sysctl_resched_latency_warn_once = 1;
71 #endif /* CONFIG_SCHED_DEBUG */
74 * Number of tasks to iterate in a single balance run.
75 * Limited because this is done with IRQs disabled.
77 #ifdef CONFIG_PREEMPT_RT
78 const_debug unsigned int sysctl_sched_nr_migrate = 8;
80 const_debug unsigned int sysctl_sched_nr_migrate = 32;
84 * period over which we measure -rt task CPU usage in us.
87 unsigned int sysctl_sched_rt_period = 1000000;
89 __read_mostly int scheduler_running;
91 #ifdef CONFIG_SCHED_CORE
93 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
95 /* kernel prio, less is more */
96 static inline int __task_prio(struct task_struct *p)
98 if (p->sched_class == &stop_sched_class) /* trumps deadline */
101 if (rt_prio(p->prio)) /* includes deadline */
102 return p->prio; /* [-1, 99] */
104 if (p->sched_class == &idle_sched_class)
105 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
107 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
117 /* real prio, less is less */
118 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
121 int pa = __task_prio(a), pb = __task_prio(b);
129 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
130 return !dl_time_before(a->dl.deadline, b->dl.deadline);
132 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
133 return cfs_prio_less(a, b, in_fi);
138 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
140 if (a->core_cookie < b->core_cookie)
143 if (a->core_cookie > b->core_cookie)
146 /* flip prio, so high prio is leftmost */
147 if (prio_less(b, a, task_rq(a)->core->core_forceidle))
153 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
155 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
157 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
160 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
162 const struct task_struct *p = __node_2_sc(node);
163 unsigned long cookie = (unsigned long)key;
165 if (cookie < p->core_cookie)
168 if (cookie > p->core_cookie)
174 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
176 rq->core->core_task_seq++;
181 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
184 void sched_core_dequeue(struct rq *rq, struct task_struct *p)
186 rq->core->core_task_seq++;
188 if (!sched_core_enqueued(p))
191 rb_erase(&p->core_node, &rq->core_tree);
192 RB_CLEAR_NODE(&p->core_node);
196 * Find left-most (aka, highest priority) task matching @cookie.
198 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
200 struct rb_node *node;
202 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
204 * The idle task always matches any cookie!
207 return idle_sched_class.pick_task(rq);
209 return __node_2_sc(node);
212 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
214 struct rb_node *node = &p->core_node;
216 node = rb_next(node);
220 p = container_of(node, struct task_struct, core_node);
221 if (p->core_cookie != cookie)
228 * Magic required such that:
230 * raw_spin_rq_lock(rq);
232 * raw_spin_rq_unlock(rq);
234 * ends up locking and unlocking the _same_ lock, and all CPUs
235 * always agree on what rq has what lock.
237 * XXX entirely possible to selectively enable cores, don't bother for now.
240 static DEFINE_MUTEX(sched_core_mutex);
241 static atomic_t sched_core_count;
242 static struct cpumask sched_core_mask;
244 static void sched_core_lock(int cpu, unsigned long *flags)
246 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
249 local_irq_save(*flags);
250 for_each_cpu(t, smt_mask)
251 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
254 static void sched_core_unlock(int cpu, unsigned long *flags)
256 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
259 for_each_cpu(t, smt_mask)
260 raw_spin_unlock(&cpu_rq(t)->__lock);
261 local_irq_restore(*flags);
264 static void __sched_core_flip(bool enabled)
272 * Toggle the online cores, one by one.
274 cpumask_copy(&sched_core_mask, cpu_online_mask);
275 for_each_cpu(cpu, &sched_core_mask) {
276 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
278 sched_core_lock(cpu, &flags);
280 for_each_cpu(t, smt_mask)
281 cpu_rq(t)->core_enabled = enabled;
283 sched_core_unlock(cpu, &flags);
285 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
289 * Toggle the offline CPUs.
291 cpumask_copy(&sched_core_mask, cpu_possible_mask);
292 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
294 for_each_cpu(cpu, &sched_core_mask)
295 cpu_rq(cpu)->core_enabled = enabled;
300 static void sched_core_assert_empty(void)
304 for_each_possible_cpu(cpu)
305 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
308 static void __sched_core_enable(void)
310 static_branch_enable(&__sched_core_enabled);
312 * Ensure all previous instances of raw_spin_rq_*lock() have finished
313 * and future ones will observe !sched_core_disabled().
316 __sched_core_flip(true);
317 sched_core_assert_empty();
320 static void __sched_core_disable(void)
322 sched_core_assert_empty();
323 __sched_core_flip(false);
324 static_branch_disable(&__sched_core_enabled);
327 void sched_core_get(void)
329 if (atomic_inc_not_zero(&sched_core_count))
332 mutex_lock(&sched_core_mutex);
333 if (!atomic_read(&sched_core_count))
334 __sched_core_enable();
336 smp_mb__before_atomic();
337 atomic_inc(&sched_core_count);
338 mutex_unlock(&sched_core_mutex);
341 static void __sched_core_put(struct work_struct *work)
343 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
344 __sched_core_disable();
345 mutex_unlock(&sched_core_mutex);
349 void sched_core_put(void)
351 static DECLARE_WORK(_work, __sched_core_put);
354 * "There can be only one"
356 * Either this is the last one, or we don't actually need to do any
357 * 'work'. If it is the last *again*, we rely on
358 * WORK_STRUCT_PENDING_BIT.
360 if (!atomic_add_unless(&sched_core_count, -1, 1))
361 schedule_work(&_work);
364 #else /* !CONFIG_SCHED_CORE */
366 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
367 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
369 #endif /* CONFIG_SCHED_CORE */
372 * part of the period that we allow rt tasks to run in us.
375 int sysctl_sched_rt_runtime = 950000;
379 * Serialization rules:
385 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
388 * rq2->lock where: rq1 < rq2
392 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
393 * local CPU's rq->lock, it optionally removes the task from the runqueue and
394 * always looks at the local rq data structures to find the most eligible task
397 * Task enqueue is also under rq->lock, possibly taken from another CPU.
398 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
399 * the local CPU to avoid bouncing the runqueue state around [ see
400 * ttwu_queue_wakelist() ]
402 * Task wakeup, specifically wakeups that involve migration, are horribly
403 * complicated to avoid having to take two rq->locks.
407 * System-calls and anything external will use task_rq_lock() which acquires
408 * both p->pi_lock and rq->lock. As a consequence the state they change is
409 * stable while holding either lock:
411 * - sched_setaffinity()/
412 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
413 * - set_user_nice(): p->se.load, p->*prio
414 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
415 * p->se.load, p->rt_priority,
416 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
417 * - sched_setnuma(): p->numa_preferred_nid
418 * - sched_move_task()/
419 * cpu_cgroup_fork(): p->sched_task_group
420 * - uclamp_update_active() p->uclamp*
422 * p->state <- TASK_*:
424 * is changed locklessly using set_current_state(), __set_current_state() or
425 * set_special_state(), see their respective comments, or by
426 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
429 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
431 * is set by activate_task() and cleared by deactivate_task(), under
432 * rq->lock. Non-zero indicates the task is runnable, the special
433 * ON_RQ_MIGRATING state is used for migration without holding both
434 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
436 * p->on_cpu <- { 0, 1 }:
438 * is set by prepare_task() and cleared by finish_task() such that it will be
439 * set before p is scheduled-in and cleared after p is scheduled-out, both
440 * under rq->lock. Non-zero indicates the task is running on its CPU.
442 * [ The astute reader will observe that it is possible for two tasks on one
443 * CPU to have ->on_cpu = 1 at the same time. ]
445 * task_cpu(p): is changed by set_task_cpu(), the rules are:
447 * - Don't call set_task_cpu() on a blocked task:
449 * We don't care what CPU we're not running on, this simplifies hotplug,
450 * the CPU assignment of blocked tasks isn't required to be valid.
452 * - for try_to_wake_up(), called under p->pi_lock:
454 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
456 * - for migration called under rq->lock:
457 * [ see task_on_rq_migrating() in task_rq_lock() ]
459 * o move_queued_task()
462 * - for migration called under double_rq_lock():
464 * o __migrate_swap_task()
465 * o push_rt_task() / pull_rt_task()
466 * o push_dl_task() / pull_dl_task()
467 * o dl_task_offline_migration()
471 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
473 raw_spinlock_t *lock;
475 /* Matches synchronize_rcu() in __sched_core_enable() */
477 if (sched_core_disabled()) {
478 raw_spin_lock_nested(&rq->__lock, subclass);
479 /* preempt_count *MUST* be > 1 */
480 preempt_enable_no_resched();
485 lock = __rq_lockp(rq);
486 raw_spin_lock_nested(lock, subclass);
487 if (likely(lock == __rq_lockp(rq))) {
488 /* preempt_count *MUST* be > 1 */
489 preempt_enable_no_resched();
492 raw_spin_unlock(lock);
496 bool raw_spin_rq_trylock(struct rq *rq)
498 raw_spinlock_t *lock;
501 /* Matches synchronize_rcu() in __sched_core_enable() */
503 if (sched_core_disabled()) {
504 ret = raw_spin_trylock(&rq->__lock);
510 lock = __rq_lockp(rq);
511 ret = raw_spin_trylock(lock);
512 if (!ret || (likely(lock == __rq_lockp(rq)))) {
516 raw_spin_unlock(lock);
520 void raw_spin_rq_unlock(struct rq *rq)
522 raw_spin_unlock(rq_lockp(rq));
527 * double_rq_lock - safely lock two runqueues
529 void double_rq_lock(struct rq *rq1, struct rq *rq2)
531 lockdep_assert_irqs_disabled();
533 if (rq_order_less(rq2, rq1))
536 raw_spin_rq_lock(rq1);
537 if (__rq_lockp(rq1) == __rq_lockp(rq2))
540 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
545 * __task_rq_lock - lock the rq @p resides on.
547 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
552 lockdep_assert_held(&p->pi_lock);
556 raw_spin_rq_lock(rq);
557 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
561 raw_spin_rq_unlock(rq);
563 while (unlikely(task_on_rq_migrating(p)))
569 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
571 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
572 __acquires(p->pi_lock)
578 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
580 raw_spin_rq_lock(rq);
582 * move_queued_task() task_rq_lock()
585 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
586 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
587 * [S] ->cpu = new_cpu [L] task_rq()
591 * If we observe the old CPU in task_rq_lock(), the acquire of
592 * the old rq->lock will fully serialize against the stores.
594 * If we observe the new CPU in task_rq_lock(), the address
595 * dependency headed by '[L] rq = task_rq()' and the acquire
596 * will pair with the WMB to ensure we then also see migrating.
598 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
602 raw_spin_rq_unlock(rq);
603 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
605 while (unlikely(task_on_rq_migrating(p)))
611 * RQ-clock updating methods:
614 static void update_rq_clock_task(struct rq *rq, s64 delta)
617 * In theory, the compile should just see 0 here, and optimize out the call
618 * to sched_rt_avg_update. But I don't trust it...
620 s64 __maybe_unused steal = 0, irq_delta = 0;
622 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
623 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
626 * Since irq_time is only updated on {soft,}irq_exit, we might run into
627 * this case when a previous update_rq_clock() happened inside a
630 * When this happens, we stop ->clock_task and only update the
631 * prev_irq_time stamp to account for the part that fit, so that a next
632 * update will consume the rest. This ensures ->clock_task is
635 * It does however cause some slight miss-attribution of {soft,}irq
636 * time, a more accurate solution would be to update the irq_time using
637 * the current rq->clock timestamp, except that would require using
640 if (irq_delta > delta)
643 rq->prev_irq_time += irq_delta;
646 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
647 if (static_key_false((¶virt_steal_rq_enabled))) {
648 steal = paravirt_steal_clock(cpu_of(rq));
649 steal -= rq->prev_steal_time_rq;
651 if (unlikely(steal > delta))
654 rq->prev_steal_time_rq += steal;
659 rq->clock_task += delta;
661 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
662 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
663 update_irq_load_avg(rq, irq_delta + steal);
665 update_rq_clock_pelt(rq, delta);
668 void update_rq_clock(struct rq *rq)
672 lockdep_assert_rq_held(rq);
674 if (rq->clock_update_flags & RQCF_ACT_SKIP)
677 #ifdef CONFIG_SCHED_DEBUG
678 if (sched_feat(WARN_DOUBLE_CLOCK))
679 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
680 rq->clock_update_flags |= RQCF_UPDATED;
683 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
687 update_rq_clock_task(rq, delta);
690 #ifdef CONFIG_SCHED_HRTICK
692 * Use HR-timers to deliver accurate preemption points.
695 static void hrtick_clear(struct rq *rq)
697 if (hrtimer_active(&rq->hrtick_timer))
698 hrtimer_cancel(&rq->hrtick_timer);
702 * High-resolution timer tick.
703 * Runs from hardirq context with interrupts disabled.
705 static enum hrtimer_restart hrtick(struct hrtimer *timer)
707 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
710 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
714 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
717 return HRTIMER_NORESTART;
722 static void __hrtick_restart(struct rq *rq)
724 struct hrtimer *timer = &rq->hrtick_timer;
725 ktime_t time = rq->hrtick_time;
727 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
731 * called from hardirq (IPI) context
733 static void __hrtick_start(void *arg)
739 __hrtick_restart(rq);
744 * Called to set the hrtick timer state.
746 * called with rq->lock held and irqs disabled
748 void hrtick_start(struct rq *rq, u64 delay)
750 struct hrtimer *timer = &rq->hrtick_timer;
754 * Don't schedule slices shorter than 10000ns, that just
755 * doesn't make sense and can cause timer DoS.
757 delta = max_t(s64, delay, 10000LL);
758 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
761 __hrtick_restart(rq);
763 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
768 * Called to set the hrtick timer state.
770 * called with rq->lock held and irqs disabled
772 void hrtick_start(struct rq *rq, u64 delay)
775 * Don't schedule slices shorter than 10000ns, that just
776 * doesn't make sense. Rely on vruntime for fairness.
778 delay = max_t(u64, delay, 10000LL);
779 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
780 HRTIMER_MODE_REL_PINNED_HARD);
783 #endif /* CONFIG_SMP */
785 static void hrtick_rq_init(struct rq *rq)
788 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
790 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
791 rq->hrtick_timer.function = hrtick;
793 #else /* CONFIG_SCHED_HRTICK */
794 static inline void hrtick_clear(struct rq *rq)
798 static inline void hrtick_rq_init(struct rq *rq)
801 #endif /* CONFIG_SCHED_HRTICK */
804 * cmpxchg based fetch_or, macro so it works for different integer types
806 #define fetch_or(ptr, mask) \
808 typeof(ptr) _ptr = (ptr); \
809 typeof(mask) _mask = (mask); \
810 typeof(*_ptr) _old, _val = *_ptr; \
813 _old = cmpxchg(_ptr, _val, _val | _mask); \
821 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
823 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
824 * this avoids any races wrt polling state changes and thereby avoids
827 static bool set_nr_and_not_polling(struct task_struct *p)
829 struct thread_info *ti = task_thread_info(p);
830 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
834 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
836 * If this returns true, then the idle task promises to call
837 * sched_ttwu_pending() and reschedule soon.
839 static bool set_nr_if_polling(struct task_struct *p)
841 struct thread_info *ti = task_thread_info(p);
842 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
845 if (!(val & _TIF_POLLING_NRFLAG))
847 if (val & _TIF_NEED_RESCHED)
849 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
858 static bool set_nr_and_not_polling(struct task_struct *p)
860 set_tsk_need_resched(p);
865 static bool set_nr_if_polling(struct task_struct *p)
872 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
874 struct wake_q_node *node = &task->wake_q;
877 * Atomically grab the task, if ->wake_q is !nil already it means
878 * it's already queued (either by us or someone else) and will get the
879 * wakeup due to that.
881 * In order to ensure that a pending wakeup will observe our pending
882 * state, even in the failed case, an explicit smp_mb() must be used.
884 smp_mb__before_atomic();
885 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
889 * The head is context local, there can be no concurrency.
892 head->lastp = &node->next;
897 * wake_q_add() - queue a wakeup for 'later' waking.
898 * @head: the wake_q_head to add @task to
899 * @task: the task to queue for 'later' wakeup
901 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
902 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
905 * This function must be used as-if it were wake_up_process(); IOW the task
906 * must be ready to be woken at this location.
908 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
910 if (__wake_q_add(head, task))
911 get_task_struct(task);
915 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
916 * @head: the wake_q_head to add @task to
917 * @task: the task to queue for 'later' wakeup
919 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
920 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
923 * This function must be used as-if it were wake_up_process(); IOW the task
924 * must be ready to be woken at this location.
926 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
927 * that already hold reference to @task can call the 'safe' version and trust
928 * wake_q to do the right thing depending whether or not the @task is already
931 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
933 if (!__wake_q_add(head, task))
934 put_task_struct(task);
937 void wake_up_q(struct wake_q_head *head)
939 struct wake_q_node *node = head->first;
941 while (node != WAKE_Q_TAIL) {
942 struct task_struct *task;
944 task = container_of(node, struct task_struct, wake_q);
945 /* Task can safely be re-inserted now: */
947 task->wake_q.next = NULL;
950 * wake_up_process() executes a full barrier, which pairs with
951 * the queueing in wake_q_add() so as not to miss wakeups.
953 wake_up_process(task);
954 put_task_struct(task);
959 * resched_curr - mark rq's current task 'to be rescheduled now'.
961 * On UP this means the setting of the need_resched flag, on SMP it
962 * might also involve a cross-CPU call to trigger the scheduler on
965 void resched_curr(struct rq *rq)
967 struct task_struct *curr = rq->curr;
970 lockdep_assert_rq_held(rq);
972 if (test_tsk_need_resched(curr))
977 if (cpu == smp_processor_id()) {
978 set_tsk_need_resched(curr);
979 set_preempt_need_resched();
983 if (set_nr_and_not_polling(curr))
984 smp_send_reschedule(cpu);
986 trace_sched_wake_idle_without_ipi(cpu);
989 void resched_cpu(int cpu)
991 struct rq *rq = cpu_rq(cpu);
994 raw_spin_rq_lock_irqsave(rq, flags);
995 if (cpu_online(cpu) || cpu == smp_processor_id())
997 raw_spin_rq_unlock_irqrestore(rq, flags);
1001 #ifdef CONFIG_NO_HZ_COMMON
1003 * In the semi idle case, use the nearest busy CPU for migrating timers
1004 * from an idle CPU. This is good for power-savings.
1006 * We don't do similar optimization for completely idle system, as
1007 * selecting an idle CPU will add more delays to the timers than intended
1008 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1010 int get_nohz_timer_target(void)
1012 int i, cpu = smp_processor_id(), default_cpu = -1;
1013 struct sched_domain *sd;
1014 const struct cpumask *hk_mask;
1016 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1022 hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1025 for_each_domain(cpu, sd) {
1026 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1037 if (default_cpu == -1)
1038 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1046 * When add_timer_on() enqueues a timer into the timer wheel of an
1047 * idle CPU then this timer might expire before the next timer event
1048 * which is scheduled to wake up that CPU. In case of a completely
1049 * idle system the next event might even be infinite time into the
1050 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1051 * leaves the inner idle loop so the newly added timer is taken into
1052 * account when the CPU goes back to idle and evaluates the timer
1053 * wheel for the next timer event.
1055 static void wake_up_idle_cpu(int cpu)
1057 struct rq *rq = cpu_rq(cpu);
1059 if (cpu == smp_processor_id())
1062 if (set_nr_and_not_polling(rq->idle))
1063 smp_send_reschedule(cpu);
1065 trace_sched_wake_idle_without_ipi(cpu);
1068 static bool wake_up_full_nohz_cpu(int cpu)
1071 * We just need the target to call irq_exit() and re-evaluate
1072 * the next tick. The nohz full kick at least implies that.
1073 * If needed we can still optimize that later with an
1076 if (cpu_is_offline(cpu))
1077 return true; /* Don't try to wake offline CPUs. */
1078 if (tick_nohz_full_cpu(cpu)) {
1079 if (cpu != smp_processor_id() ||
1080 tick_nohz_tick_stopped())
1081 tick_nohz_full_kick_cpu(cpu);
1089 * Wake up the specified CPU. If the CPU is going offline, it is the
1090 * caller's responsibility to deal with the lost wakeup, for example,
1091 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1093 void wake_up_nohz_cpu(int cpu)
1095 if (!wake_up_full_nohz_cpu(cpu))
1096 wake_up_idle_cpu(cpu);
1099 static void nohz_csd_func(void *info)
1101 struct rq *rq = info;
1102 int cpu = cpu_of(rq);
1106 * Release the rq::nohz_csd.
1108 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1109 WARN_ON(!(flags & NOHZ_KICK_MASK));
1111 rq->idle_balance = idle_cpu(cpu);
1112 if (rq->idle_balance && !need_resched()) {
1113 rq->nohz_idle_balance = flags;
1114 raise_softirq_irqoff(SCHED_SOFTIRQ);
1118 #endif /* CONFIG_NO_HZ_COMMON */
1120 #ifdef CONFIG_NO_HZ_FULL
1121 bool sched_can_stop_tick(struct rq *rq)
1123 int fifo_nr_running;
1125 /* Deadline tasks, even if single, need the tick */
1126 if (rq->dl.dl_nr_running)
1130 * If there are more than one RR tasks, we need the tick to affect the
1131 * actual RR behaviour.
1133 if (rq->rt.rr_nr_running) {
1134 if (rq->rt.rr_nr_running == 1)
1141 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1142 * forced preemption between FIFO tasks.
1144 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1145 if (fifo_nr_running)
1149 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1150 * if there's more than one we need the tick for involuntary
1153 if (rq->nr_running > 1)
1158 #endif /* CONFIG_NO_HZ_FULL */
1159 #endif /* CONFIG_SMP */
1161 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1162 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1164 * Iterate task_group tree rooted at *from, calling @down when first entering a
1165 * node and @up when leaving it for the final time.
1167 * Caller must hold rcu_lock or sufficient equivalent.
1169 int walk_tg_tree_from(struct task_group *from,
1170 tg_visitor down, tg_visitor up, void *data)
1172 struct task_group *parent, *child;
1178 ret = (*down)(parent, data);
1181 list_for_each_entry_rcu(child, &parent->children, siblings) {
1188 ret = (*up)(parent, data);
1189 if (ret || parent == from)
1193 parent = parent->parent;
1200 int tg_nop(struct task_group *tg, void *data)
1206 static void set_load_weight(struct task_struct *p, bool update_load)
1208 int prio = p->static_prio - MAX_RT_PRIO;
1209 struct load_weight *load = &p->se.load;
1212 * SCHED_IDLE tasks get minimal weight:
1214 if (task_has_idle_policy(p)) {
1215 load->weight = scale_load(WEIGHT_IDLEPRIO);
1216 load->inv_weight = WMULT_IDLEPRIO;
1221 * SCHED_OTHER tasks have to update their load when changing their
1224 if (update_load && p->sched_class == &fair_sched_class) {
1225 reweight_task(p, prio);
1227 load->weight = scale_load(sched_prio_to_weight[prio]);
1228 load->inv_weight = sched_prio_to_wmult[prio];
1232 #ifdef CONFIG_UCLAMP_TASK
1234 * Serializes updates of utilization clamp values
1236 * The (slow-path) user-space triggers utilization clamp value updates which
1237 * can require updates on (fast-path) scheduler's data structures used to
1238 * support enqueue/dequeue operations.
1239 * While the per-CPU rq lock protects fast-path update operations, user-space
1240 * requests are serialized using a mutex to reduce the risk of conflicting
1241 * updates or API abuses.
1243 static DEFINE_MUTEX(uclamp_mutex);
1245 /* Max allowed minimum utilization */
1246 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1248 /* Max allowed maximum utilization */
1249 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1252 * By default RT tasks run at the maximum performance point/capacity of the
1253 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1254 * SCHED_CAPACITY_SCALE.
1256 * This knob allows admins to change the default behavior when uclamp is being
1257 * used. In battery powered devices, particularly, running at the maximum
1258 * capacity and frequency will increase energy consumption and shorten the
1261 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1263 * This knob will not override the system default sched_util_clamp_min defined
1266 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1268 /* All clamps are required to be less or equal than these values */
1269 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1272 * This static key is used to reduce the uclamp overhead in the fast path. It
1273 * primarily disables the call to uclamp_rq_{inc, dec}() in
1274 * enqueue/dequeue_task().
1276 * This allows users to continue to enable uclamp in their kernel config with
1277 * minimum uclamp overhead in the fast path.
1279 * As soon as userspace modifies any of the uclamp knobs, the static key is
1280 * enabled, since we have an actual users that make use of uclamp
1283 * The knobs that would enable this static key are:
1285 * * A task modifying its uclamp value with sched_setattr().
1286 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1287 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1289 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1291 /* Integer rounded range for each bucket */
1292 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1294 #define for_each_clamp_id(clamp_id) \
1295 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1297 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1299 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1302 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1304 if (clamp_id == UCLAMP_MIN)
1306 return SCHED_CAPACITY_SCALE;
1309 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1310 unsigned int value, bool user_defined)
1312 uc_se->value = value;
1313 uc_se->bucket_id = uclamp_bucket_id(value);
1314 uc_se->user_defined = user_defined;
1317 static inline unsigned int
1318 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1319 unsigned int clamp_value)
1322 * Avoid blocked utilization pushing up the frequency when we go
1323 * idle (which drops the max-clamp) by retaining the last known
1326 if (clamp_id == UCLAMP_MAX) {
1327 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1331 return uclamp_none(UCLAMP_MIN);
1334 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1335 unsigned int clamp_value)
1337 /* Reset max-clamp retention only on idle exit */
1338 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1341 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1345 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1346 unsigned int clamp_value)
1348 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1349 int bucket_id = UCLAMP_BUCKETS - 1;
1352 * Since both min and max clamps are max aggregated, find the
1353 * top most bucket with tasks in.
1355 for ( ; bucket_id >= 0; bucket_id--) {
1356 if (!bucket[bucket_id].tasks)
1358 return bucket[bucket_id].value;
1361 /* No tasks -- default clamp values */
1362 return uclamp_idle_value(rq, clamp_id, clamp_value);
1365 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1367 unsigned int default_util_min;
1368 struct uclamp_se *uc_se;
1370 lockdep_assert_held(&p->pi_lock);
1372 uc_se = &p->uclamp_req[UCLAMP_MIN];
1374 /* Only sync if user didn't override the default */
1375 if (uc_se->user_defined)
1378 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1379 uclamp_se_set(uc_se, default_util_min, false);
1382 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1390 /* Protect updates to p->uclamp_* */
1391 rq = task_rq_lock(p, &rf);
1392 __uclamp_update_util_min_rt_default(p);
1393 task_rq_unlock(rq, p, &rf);
1396 static void uclamp_sync_util_min_rt_default(void)
1398 struct task_struct *g, *p;
1401 * copy_process() sysctl_uclamp
1402 * uclamp_min_rt = X;
1403 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1404 * // link thread smp_mb__after_spinlock()
1405 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1406 * sched_post_fork() for_each_process_thread()
1407 * __uclamp_sync_rt() __uclamp_sync_rt()
1409 * Ensures that either sched_post_fork() will observe the new
1410 * uclamp_min_rt or for_each_process_thread() will observe the new
1413 read_lock(&tasklist_lock);
1414 smp_mb__after_spinlock();
1415 read_unlock(&tasklist_lock);
1418 for_each_process_thread(g, p)
1419 uclamp_update_util_min_rt_default(p);
1423 static inline struct uclamp_se
1424 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1426 /* Copy by value as we could modify it */
1427 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1428 #ifdef CONFIG_UCLAMP_TASK_GROUP
1429 unsigned int tg_min, tg_max, value;
1432 * Tasks in autogroups or root task group will be
1433 * restricted by system defaults.
1435 if (task_group_is_autogroup(task_group(p)))
1437 if (task_group(p) == &root_task_group)
1440 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1441 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1442 value = uc_req.value;
1443 value = clamp(value, tg_min, tg_max);
1444 uclamp_se_set(&uc_req, value, false);
1451 * The effective clamp bucket index of a task depends on, by increasing
1453 * - the task specific clamp value, when explicitly requested from userspace
1454 * - the task group effective clamp value, for tasks not either in the root
1455 * group or in an autogroup
1456 * - the system default clamp value, defined by the sysadmin
1458 static inline struct uclamp_se
1459 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1461 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1462 struct uclamp_se uc_max = uclamp_default[clamp_id];
1464 /* System default restrictions always apply */
1465 if (unlikely(uc_req.value > uc_max.value))
1471 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1473 struct uclamp_se uc_eff;
1475 /* Task currently refcounted: use back-annotated (effective) value */
1476 if (p->uclamp[clamp_id].active)
1477 return (unsigned long)p->uclamp[clamp_id].value;
1479 uc_eff = uclamp_eff_get(p, clamp_id);
1481 return (unsigned long)uc_eff.value;
1485 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1486 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1487 * updates the rq's clamp value if required.
1489 * Tasks can have a task-specific value requested from user-space, track
1490 * within each bucket the maximum value for tasks refcounted in it.
1491 * This "local max aggregation" allows to track the exact "requested" value
1492 * for each bucket when all its RUNNABLE tasks require the same clamp.
1494 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1495 enum uclamp_id clamp_id)
1497 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1498 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1499 struct uclamp_bucket *bucket;
1501 lockdep_assert_rq_held(rq);
1503 /* Update task effective clamp */
1504 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1506 bucket = &uc_rq->bucket[uc_se->bucket_id];
1508 uc_se->active = true;
1510 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1513 * Local max aggregation: rq buckets always track the max
1514 * "requested" clamp value of its RUNNABLE tasks.
1516 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1517 bucket->value = uc_se->value;
1519 if (uc_se->value > READ_ONCE(uc_rq->value))
1520 WRITE_ONCE(uc_rq->value, uc_se->value);
1524 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1525 * is released. If this is the last task reference counting the rq's max
1526 * active clamp value, then the rq's clamp value is updated.
1528 * Both refcounted tasks and rq's cached clamp values are expected to be
1529 * always valid. If it's detected they are not, as defensive programming,
1530 * enforce the expected state and warn.
1532 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1533 enum uclamp_id clamp_id)
1535 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1536 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1537 struct uclamp_bucket *bucket;
1538 unsigned int bkt_clamp;
1539 unsigned int rq_clamp;
1541 lockdep_assert_rq_held(rq);
1544 * If sched_uclamp_used was enabled after task @p was enqueued,
1545 * we could end up with unbalanced call to uclamp_rq_dec_id().
1547 * In this case the uc_se->active flag should be false since no uclamp
1548 * accounting was performed at enqueue time and we can just return
1551 * Need to be careful of the following enqueue/dequeue ordering
1555 * // sched_uclamp_used gets enabled
1558 * // Must not decrement bucket->tasks here
1561 * where we could end up with stale data in uc_se and
1562 * bucket[uc_se->bucket_id].
1564 * The following check here eliminates the possibility of such race.
1566 if (unlikely(!uc_se->active))
1569 bucket = &uc_rq->bucket[uc_se->bucket_id];
1571 SCHED_WARN_ON(!bucket->tasks);
1572 if (likely(bucket->tasks))
1575 uc_se->active = false;
1578 * Keep "local max aggregation" simple and accept to (possibly)
1579 * overboost some RUNNABLE tasks in the same bucket.
1580 * The rq clamp bucket value is reset to its base value whenever
1581 * there are no more RUNNABLE tasks refcounting it.
1583 if (likely(bucket->tasks))
1586 rq_clamp = READ_ONCE(uc_rq->value);
1588 * Defensive programming: this should never happen. If it happens,
1589 * e.g. due to future modification, warn and fixup the expected value.
1591 SCHED_WARN_ON(bucket->value > rq_clamp);
1592 if (bucket->value >= rq_clamp) {
1593 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1594 WRITE_ONCE(uc_rq->value, bkt_clamp);
1598 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1600 enum uclamp_id clamp_id;
1603 * Avoid any overhead until uclamp is actually used by the userspace.
1605 * The condition is constructed such that a NOP is generated when
1606 * sched_uclamp_used is disabled.
1608 if (!static_branch_unlikely(&sched_uclamp_used))
1611 if (unlikely(!p->sched_class->uclamp_enabled))
1614 for_each_clamp_id(clamp_id)
1615 uclamp_rq_inc_id(rq, p, clamp_id);
1617 /* Reset clamp idle holding when there is one RUNNABLE task */
1618 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1619 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1622 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1624 enum uclamp_id clamp_id;
1627 * Avoid any overhead until uclamp is actually used by the userspace.
1629 * The condition is constructed such that a NOP is generated when
1630 * sched_uclamp_used is disabled.
1632 if (!static_branch_unlikely(&sched_uclamp_used))
1635 if (unlikely(!p->sched_class->uclamp_enabled))
1638 for_each_clamp_id(clamp_id)
1639 uclamp_rq_dec_id(rq, p, clamp_id);
1642 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1643 enum uclamp_id clamp_id)
1645 if (!p->uclamp[clamp_id].active)
1648 uclamp_rq_dec_id(rq, p, clamp_id);
1649 uclamp_rq_inc_id(rq, p, clamp_id);
1652 * Make sure to clear the idle flag if we've transiently reached 0
1653 * active tasks on rq.
1655 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1656 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1660 uclamp_update_active(struct task_struct *p)
1662 enum uclamp_id clamp_id;
1667 * Lock the task and the rq where the task is (or was) queued.
1669 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1670 * price to pay to safely serialize util_{min,max} updates with
1671 * enqueues, dequeues and migration operations.
1672 * This is the same locking schema used by __set_cpus_allowed_ptr().
1674 rq = task_rq_lock(p, &rf);
1677 * Setting the clamp bucket is serialized by task_rq_lock().
1678 * If the task is not yet RUNNABLE and its task_struct is not
1679 * affecting a valid clamp bucket, the next time it's enqueued,
1680 * it will already see the updated clamp bucket value.
1682 for_each_clamp_id(clamp_id)
1683 uclamp_rq_reinc_id(rq, p, clamp_id);
1685 task_rq_unlock(rq, p, &rf);
1688 #ifdef CONFIG_UCLAMP_TASK_GROUP
1690 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1692 struct css_task_iter it;
1693 struct task_struct *p;
1695 css_task_iter_start(css, 0, &it);
1696 while ((p = css_task_iter_next(&it)))
1697 uclamp_update_active(p);
1698 css_task_iter_end(&it);
1701 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1702 static void uclamp_update_root_tg(void)
1704 struct task_group *tg = &root_task_group;
1706 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1707 sysctl_sched_uclamp_util_min, false);
1708 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1709 sysctl_sched_uclamp_util_max, false);
1712 cpu_util_update_eff(&root_task_group.css);
1716 static void uclamp_update_root_tg(void) { }
1719 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1720 void *buffer, size_t *lenp, loff_t *ppos)
1722 bool update_root_tg = false;
1723 int old_min, old_max, old_min_rt;
1726 mutex_lock(&uclamp_mutex);
1727 old_min = sysctl_sched_uclamp_util_min;
1728 old_max = sysctl_sched_uclamp_util_max;
1729 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1731 result = proc_dointvec(table, write, buffer, lenp, ppos);
1737 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1738 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1739 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1745 if (old_min != sysctl_sched_uclamp_util_min) {
1746 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1747 sysctl_sched_uclamp_util_min, false);
1748 update_root_tg = true;
1750 if (old_max != sysctl_sched_uclamp_util_max) {
1751 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1752 sysctl_sched_uclamp_util_max, false);
1753 update_root_tg = true;
1756 if (update_root_tg) {
1757 static_branch_enable(&sched_uclamp_used);
1758 uclamp_update_root_tg();
1761 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1762 static_branch_enable(&sched_uclamp_used);
1763 uclamp_sync_util_min_rt_default();
1767 * We update all RUNNABLE tasks only when task groups are in use.
1768 * Otherwise, keep it simple and do just a lazy update at each next
1769 * task enqueue time.
1775 sysctl_sched_uclamp_util_min = old_min;
1776 sysctl_sched_uclamp_util_max = old_max;
1777 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1779 mutex_unlock(&uclamp_mutex);
1784 static int uclamp_validate(struct task_struct *p,
1785 const struct sched_attr *attr)
1787 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1788 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1790 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1791 util_min = attr->sched_util_min;
1793 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1797 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1798 util_max = attr->sched_util_max;
1800 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1804 if (util_min != -1 && util_max != -1 && util_min > util_max)
1808 * We have valid uclamp attributes; make sure uclamp is enabled.
1810 * We need to do that here, because enabling static branches is a
1811 * blocking operation which obviously cannot be done while holding
1814 static_branch_enable(&sched_uclamp_used);
1819 static bool uclamp_reset(const struct sched_attr *attr,
1820 enum uclamp_id clamp_id,
1821 struct uclamp_se *uc_se)
1823 /* Reset on sched class change for a non user-defined clamp value. */
1824 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1825 !uc_se->user_defined)
1828 /* Reset on sched_util_{min,max} == -1. */
1829 if (clamp_id == UCLAMP_MIN &&
1830 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1831 attr->sched_util_min == -1) {
1835 if (clamp_id == UCLAMP_MAX &&
1836 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1837 attr->sched_util_max == -1) {
1844 static void __setscheduler_uclamp(struct task_struct *p,
1845 const struct sched_attr *attr)
1847 enum uclamp_id clamp_id;
1849 for_each_clamp_id(clamp_id) {
1850 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1853 if (!uclamp_reset(attr, clamp_id, uc_se))
1857 * RT by default have a 100% boost value that could be modified
1860 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1861 value = sysctl_sched_uclamp_util_min_rt_default;
1863 value = uclamp_none(clamp_id);
1865 uclamp_se_set(uc_se, value, false);
1869 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1872 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1873 attr->sched_util_min != -1) {
1874 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1875 attr->sched_util_min, true);
1878 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1879 attr->sched_util_max != -1) {
1880 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1881 attr->sched_util_max, true);
1885 static void uclamp_fork(struct task_struct *p)
1887 enum uclamp_id clamp_id;
1890 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1891 * as the task is still at its early fork stages.
1893 for_each_clamp_id(clamp_id)
1894 p->uclamp[clamp_id].active = false;
1896 if (likely(!p->sched_reset_on_fork))
1899 for_each_clamp_id(clamp_id) {
1900 uclamp_se_set(&p->uclamp_req[clamp_id],
1901 uclamp_none(clamp_id), false);
1905 static void uclamp_post_fork(struct task_struct *p)
1907 uclamp_update_util_min_rt_default(p);
1910 static void __init init_uclamp_rq(struct rq *rq)
1912 enum uclamp_id clamp_id;
1913 struct uclamp_rq *uc_rq = rq->uclamp;
1915 for_each_clamp_id(clamp_id) {
1916 uc_rq[clamp_id] = (struct uclamp_rq) {
1917 .value = uclamp_none(clamp_id)
1921 rq->uclamp_flags = 0;
1924 static void __init init_uclamp(void)
1926 struct uclamp_se uc_max = {};
1927 enum uclamp_id clamp_id;
1930 for_each_possible_cpu(cpu)
1931 init_uclamp_rq(cpu_rq(cpu));
1933 for_each_clamp_id(clamp_id) {
1934 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1935 uclamp_none(clamp_id), false);
1938 /* System defaults allow max clamp values for both indexes */
1939 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1940 for_each_clamp_id(clamp_id) {
1941 uclamp_default[clamp_id] = uc_max;
1942 #ifdef CONFIG_UCLAMP_TASK_GROUP
1943 root_task_group.uclamp_req[clamp_id] = uc_max;
1944 root_task_group.uclamp[clamp_id] = uc_max;
1949 #else /* CONFIG_UCLAMP_TASK */
1950 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1951 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1952 static inline int uclamp_validate(struct task_struct *p,
1953 const struct sched_attr *attr)
1957 static void __setscheduler_uclamp(struct task_struct *p,
1958 const struct sched_attr *attr) { }
1959 static inline void uclamp_fork(struct task_struct *p) { }
1960 static inline void uclamp_post_fork(struct task_struct *p) { }
1961 static inline void init_uclamp(void) { }
1962 #endif /* CONFIG_UCLAMP_TASK */
1964 bool sched_task_on_rq(struct task_struct *p)
1966 return task_on_rq_queued(p);
1969 unsigned long get_wchan(struct task_struct *p)
1971 unsigned long ip = 0;
1974 if (!p || p == current)
1977 /* Only get wchan if task is blocked and we can keep it that way. */
1978 raw_spin_lock_irq(&p->pi_lock);
1979 state = READ_ONCE(p->__state);
1980 smp_rmb(); /* see try_to_wake_up() */
1981 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
1982 ip = __get_wchan(p);
1983 raw_spin_unlock_irq(&p->pi_lock);
1988 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1990 if (!(flags & ENQUEUE_NOCLOCK))
1991 update_rq_clock(rq);
1993 if (!(flags & ENQUEUE_RESTORE)) {
1994 sched_info_enqueue(rq, p);
1995 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1998 uclamp_rq_inc(rq, p);
1999 p->sched_class->enqueue_task(rq, p, flags);
2001 if (sched_core_enabled(rq))
2002 sched_core_enqueue(rq, p);
2005 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2007 if (sched_core_enabled(rq))
2008 sched_core_dequeue(rq, p);
2010 if (!(flags & DEQUEUE_NOCLOCK))
2011 update_rq_clock(rq);
2013 if (!(flags & DEQUEUE_SAVE)) {
2014 sched_info_dequeue(rq, p);
2015 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2018 uclamp_rq_dec(rq, p);
2019 p->sched_class->dequeue_task(rq, p, flags);
2022 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2024 enqueue_task(rq, p, flags);
2026 p->on_rq = TASK_ON_RQ_QUEUED;
2029 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2031 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2033 dequeue_task(rq, p, flags);
2036 static inline int __normal_prio(int policy, int rt_prio, int nice)
2040 if (dl_policy(policy))
2041 prio = MAX_DL_PRIO - 1;
2042 else if (rt_policy(policy))
2043 prio = MAX_RT_PRIO - 1 - rt_prio;
2045 prio = NICE_TO_PRIO(nice);
2051 * Calculate the expected normal priority: i.e. priority
2052 * without taking RT-inheritance into account. Might be
2053 * boosted by interactivity modifiers. Changes upon fork,
2054 * setprio syscalls, and whenever the interactivity
2055 * estimator recalculates.
2057 static inline int normal_prio(struct task_struct *p)
2059 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2063 * Calculate the current priority, i.e. the priority
2064 * taken into account by the scheduler. This value might
2065 * be boosted by RT tasks, or might be boosted by
2066 * interactivity modifiers. Will be RT if the task got
2067 * RT-boosted. If not then it returns p->normal_prio.
2069 static int effective_prio(struct task_struct *p)
2071 p->normal_prio = normal_prio(p);
2073 * If we are RT tasks or we were boosted to RT priority,
2074 * keep the priority unchanged. Otherwise, update priority
2075 * to the normal priority:
2077 if (!rt_prio(p->prio))
2078 return p->normal_prio;
2083 * task_curr - is this task currently executing on a CPU?
2084 * @p: the task in question.
2086 * Return: 1 if the task is currently executing. 0 otherwise.
2088 inline int task_curr(const struct task_struct *p)
2090 return cpu_curr(task_cpu(p)) == p;
2094 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2095 * use the balance_callback list if you want balancing.
2097 * this means any call to check_class_changed() must be followed by a call to
2098 * balance_callback().
2100 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2101 const struct sched_class *prev_class,
2104 if (prev_class != p->sched_class) {
2105 if (prev_class->switched_from)
2106 prev_class->switched_from(rq, p);
2108 p->sched_class->switched_to(rq, p);
2109 } else if (oldprio != p->prio || dl_task(p))
2110 p->sched_class->prio_changed(rq, p, oldprio);
2113 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2115 if (p->sched_class == rq->curr->sched_class)
2116 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2117 else if (p->sched_class > rq->curr->sched_class)
2121 * A queue event has occurred, and we're going to schedule. In
2122 * this case, we can save a useless back to back clock update.
2124 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2125 rq_clock_skip_update(rq);
2131 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2133 static int __set_cpus_allowed_ptr(struct task_struct *p,
2134 const struct cpumask *new_mask,
2137 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2139 if (likely(!p->migration_disabled))
2142 if (p->cpus_ptr != &p->cpus_mask)
2146 * Violates locking rules! see comment in __do_set_cpus_allowed().
2148 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2151 void migrate_disable(void)
2153 struct task_struct *p = current;
2155 if (p->migration_disabled) {
2156 p->migration_disabled++;
2161 this_rq()->nr_pinned++;
2162 p->migration_disabled = 1;
2165 EXPORT_SYMBOL_GPL(migrate_disable);
2167 void migrate_enable(void)
2169 struct task_struct *p = current;
2171 if (p->migration_disabled > 1) {
2172 p->migration_disabled--;
2177 * Ensure stop_task runs either before or after this, and that
2178 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2181 if (p->cpus_ptr != &p->cpus_mask)
2182 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2184 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2185 * regular cpus_mask, otherwise things that race (eg.
2186 * select_fallback_rq) get confused.
2189 p->migration_disabled = 0;
2190 this_rq()->nr_pinned--;
2193 EXPORT_SYMBOL_GPL(migrate_enable);
2195 static inline bool rq_has_pinned_tasks(struct rq *rq)
2197 return rq->nr_pinned;
2201 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2202 * __set_cpus_allowed_ptr() and select_fallback_rq().
2204 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2206 /* When not in the task's cpumask, no point in looking further. */
2207 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2210 /* migrate_disabled() must be allowed to finish. */
2211 if (is_migration_disabled(p))
2212 return cpu_online(cpu);
2214 /* Non kernel threads are not allowed during either online or offline. */
2215 if (!(p->flags & PF_KTHREAD))
2216 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2218 /* KTHREAD_IS_PER_CPU is always allowed. */
2219 if (kthread_is_per_cpu(p))
2220 return cpu_online(cpu);
2222 /* Regular kernel threads don't get to stay during offline. */
2226 /* But are allowed during online. */
2227 return cpu_online(cpu);
2231 * This is how migration works:
2233 * 1) we invoke migration_cpu_stop() on the target CPU using
2235 * 2) stopper starts to run (implicitly forcing the migrated thread
2237 * 3) it checks whether the migrated task is still in the wrong runqueue.
2238 * 4) if it's in the wrong runqueue then the migration thread removes
2239 * it and puts it into the right queue.
2240 * 5) stopper completes and stop_one_cpu() returns and the migration
2245 * move_queued_task - move a queued task to new rq.
2247 * Returns (locked) new rq. Old rq's lock is released.
2249 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2250 struct task_struct *p, int new_cpu)
2252 lockdep_assert_rq_held(rq);
2254 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2255 set_task_cpu(p, new_cpu);
2258 rq = cpu_rq(new_cpu);
2261 BUG_ON(task_cpu(p) != new_cpu);
2262 activate_task(rq, p, 0);
2263 check_preempt_curr(rq, p, 0);
2268 struct migration_arg {
2269 struct task_struct *task;
2271 struct set_affinity_pending *pending;
2275 * @refs: number of wait_for_completion()
2276 * @stop_pending: is @stop_work in use
2278 struct set_affinity_pending {
2280 unsigned int stop_pending;
2281 struct completion done;
2282 struct cpu_stop_work stop_work;
2283 struct migration_arg arg;
2287 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2288 * this because either it can't run here any more (set_cpus_allowed()
2289 * away from this CPU, or CPU going down), or because we're
2290 * attempting to rebalance this task on exec (sched_exec).
2292 * So we race with normal scheduler movements, but that's OK, as long
2293 * as the task is no longer on this CPU.
2295 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2296 struct task_struct *p, int dest_cpu)
2298 /* Affinity changed (again). */
2299 if (!is_cpu_allowed(p, dest_cpu))
2302 update_rq_clock(rq);
2303 rq = move_queued_task(rq, rf, p, dest_cpu);
2309 * migration_cpu_stop - this will be executed by a highprio stopper thread
2310 * and performs thread migration by bumping thread off CPU then
2311 * 'pushing' onto another runqueue.
2313 static int migration_cpu_stop(void *data)
2315 struct migration_arg *arg = data;
2316 struct set_affinity_pending *pending = arg->pending;
2317 struct task_struct *p = arg->task;
2318 struct rq *rq = this_rq();
2319 bool complete = false;
2323 * The original target CPU might have gone down and we might
2324 * be on another CPU but it doesn't matter.
2326 local_irq_save(rf.flags);
2328 * We need to explicitly wake pending tasks before running
2329 * __migrate_task() such that we will not miss enforcing cpus_ptr
2330 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2332 flush_smp_call_function_from_idle();
2334 raw_spin_lock(&p->pi_lock);
2338 * If we were passed a pending, then ->stop_pending was set, thus
2339 * p->migration_pending must have remained stable.
2341 WARN_ON_ONCE(pending && pending != p->migration_pending);
2344 * If task_rq(p) != rq, it cannot be migrated here, because we're
2345 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2346 * we're holding p->pi_lock.
2348 if (task_rq(p) == rq) {
2349 if (is_migration_disabled(p))
2353 p->migration_pending = NULL;
2356 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2360 if (task_on_rq_queued(p))
2361 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2363 p->wake_cpu = arg->dest_cpu;
2366 * XXX __migrate_task() can fail, at which point we might end
2367 * up running on a dodgy CPU, AFAICT this can only happen
2368 * during CPU hotplug, at which point we'll get pushed out
2369 * anyway, so it's probably not a big deal.
2372 } else if (pending) {
2374 * This happens when we get migrated between migrate_enable()'s
2375 * preempt_enable() and scheduling the stopper task. At that
2376 * point we're a regular task again and not current anymore.
2378 * A !PREEMPT kernel has a giant hole here, which makes it far
2383 * The task moved before the stopper got to run. We're holding
2384 * ->pi_lock, so the allowed mask is stable - if it got
2385 * somewhere allowed, we're done.
2387 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2388 p->migration_pending = NULL;
2394 * When migrate_enable() hits a rq mis-match we can't reliably
2395 * determine is_migration_disabled() and so have to chase after
2398 WARN_ON_ONCE(!pending->stop_pending);
2399 task_rq_unlock(rq, p, &rf);
2400 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2401 &pending->arg, &pending->stop_work);
2406 pending->stop_pending = false;
2407 task_rq_unlock(rq, p, &rf);
2410 complete_all(&pending->done);
2415 int push_cpu_stop(void *arg)
2417 struct rq *lowest_rq = NULL, *rq = this_rq();
2418 struct task_struct *p = arg;
2420 raw_spin_lock_irq(&p->pi_lock);
2421 raw_spin_rq_lock(rq);
2423 if (task_rq(p) != rq)
2426 if (is_migration_disabled(p)) {
2427 p->migration_flags |= MDF_PUSH;
2431 p->migration_flags &= ~MDF_PUSH;
2433 if (p->sched_class->find_lock_rq)
2434 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2439 // XXX validate p is still the highest prio task
2440 if (task_rq(p) == rq) {
2441 deactivate_task(rq, p, 0);
2442 set_task_cpu(p, lowest_rq->cpu);
2443 activate_task(lowest_rq, p, 0);
2444 resched_curr(lowest_rq);
2447 double_unlock_balance(rq, lowest_rq);
2450 rq->push_busy = false;
2451 raw_spin_rq_unlock(rq);
2452 raw_spin_unlock_irq(&p->pi_lock);
2459 * sched_class::set_cpus_allowed must do the below, but is not required to
2460 * actually call this function.
2462 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2464 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2465 p->cpus_ptr = new_mask;
2469 cpumask_copy(&p->cpus_mask, new_mask);
2470 p->nr_cpus_allowed = cpumask_weight(new_mask);
2474 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2476 struct rq *rq = task_rq(p);
2477 bool queued, running;
2480 * This here violates the locking rules for affinity, since we're only
2481 * supposed to change these variables while holding both rq->lock and
2484 * HOWEVER, it magically works, because ttwu() is the only code that
2485 * accesses these variables under p->pi_lock and only does so after
2486 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2487 * before finish_task().
2489 * XXX do further audits, this smells like something putrid.
2491 if (flags & SCA_MIGRATE_DISABLE)
2492 SCHED_WARN_ON(!p->on_cpu);
2494 lockdep_assert_held(&p->pi_lock);
2496 queued = task_on_rq_queued(p);
2497 running = task_current(rq, p);
2501 * Because __kthread_bind() calls this on blocked tasks without
2504 lockdep_assert_rq_held(rq);
2505 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2508 put_prev_task(rq, p);
2510 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2513 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2515 set_next_task(rq, p);
2518 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2520 __do_set_cpus_allowed(p, new_mask, 0);
2523 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2526 if (!src->user_cpus_ptr)
2529 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2530 if (!dst->user_cpus_ptr)
2533 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2537 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2539 struct cpumask *user_mask = NULL;
2541 swap(p->user_cpus_ptr, user_mask);
2546 void release_user_cpus_ptr(struct task_struct *p)
2548 kfree(clear_user_cpus_ptr(p));
2552 * This function is wildly self concurrent; here be dragons.
2555 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2556 * designated task is enqueued on an allowed CPU. If that task is currently
2557 * running, we have to kick it out using the CPU stopper.
2559 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2562 * Initial conditions: P0->cpus_mask = [0, 1]
2566 * migrate_disable();
2568 * set_cpus_allowed_ptr(P0, [1]);
2570 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2571 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2572 * This means we need the following scheme:
2576 * migrate_disable();
2578 * set_cpus_allowed_ptr(P0, [1]);
2582 * __set_cpus_allowed_ptr();
2583 * <wakes local stopper>
2584 * `--> <woken on migration completion>
2586 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2587 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2588 * task p are serialized by p->pi_lock, which we can leverage: the one that
2589 * should come into effect at the end of the Migrate-Disable region is the last
2590 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2591 * but we still need to properly signal those waiting tasks at the appropriate
2594 * This is implemented using struct set_affinity_pending. The first
2595 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2596 * setup an instance of that struct and install it on the targeted task_struct.
2597 * Any and all further callers will reuse that instance. Those then wait for
2598 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2599 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2602 * (1) In the cases covered above. There is one more where the completion is
2603 * signaled within affine_move_task() itself: when a subsequent affinity request
2604 * occurs after the stopper bailed out due to the targeted task still being
2605 * Migrate-Disable. Consider:
2607 * Initial conditions: P0->cpus_mask = [0, 1]
2611 * migrate_disable();
2613 * set_cpus_allowed_ptr(P0, [1]);
2616 * migration_cpu_stop()
2617 * is_migration_disabled()
2619 * set_cpus_allowed_ptr(P0, [0, 1]);
2620 * <signal completion>
2623 * Note that the above is safe vs a concurrent migrate_enable(), as any
2624 * pending affinity completion is preceded by an uninstallation of
2625 * p->migration_pending done with p->pi_lock held.
2627 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2628 int dest_cpu, unsigned int flags)
2630 struct set_affinity_pending my_pending = { }, *pending = NULL;
2631 bool stop_pending, complete = false;
2633 /* Can the task run on the task's current CPU? If so, we're done */
2634 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2635 struct task_struct *push_task = NULL;
2637 if ((flags & SCA_MIGRATE_ENABLE) &&
2638 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2639 rq->push_busy = true;
2640 push_task = get_task_struct(p);
2644 * If there are pending waiters, but no pending stop_work,
2645 * then complete now.
2647 pending = p->migration_pending;
2648 if (pending && !pending->stop_pending) {
2649 p->migration_pending = NULL;
2653 task_rq_unlock(rq, p, rf);
2656 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2661 complete_all(&pending->done);
2666 if (!(flags & SCA_MIGRATE_ENABLE)) {
2667 /* serialized by p->pi_lock */
2668 if (!p->migration_pending) {
2669 /* Install the request */
2670 refcount_set(&my_pending.refs, 1);
2671 init_completion(&my_pending.done);
2672 my_pending.arg = (struct migration_arg) {
2674 .dest_cpu = dest_cpu,
2675 .pending = &my_pending,
2678 p->migration_pending = &my_pending;
2680 pending = p->migration_pending;
2681 refcount_inc(&pending->refs);
2683 * Affinity has changed, but we've already installed a
2684 * pending. migration_cpu_stop() *must* see this, else
2685 * we risk a completion of the pending despite having a
2686 * task on a disallowed CPU.
2688 * Serialized by p->pi_lock, so this is safe.
2690 pending->arg.dest_cpu = dest_cpu;
2693 pending = p->migration_pending;
2695 * - !MIGRATE_ENABLE:
2696 * we'll have installed a pending if there wasn't one already.
2699 * we're here because the current CPU isn't matching anymore,
2700 * the only way that can happen is because of a concurrent
2701 * set_cpus_allowed_ptr() call, which should then still be
2702 * pending completion.
2704 * Either way, we really should have a @pending here.
2706 if (WARN_ON_ONCE(!pending)) {
2707 task_rq_unlock(rq, p, rf);
2711 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2713 * MIGRATE_ENABLE gets here because 'p == current', but for
2714 * anything else we cannot do is_migration_disabled(), punt
2715 * and have the stopper function handle it all race-free.
2717 stop_pending = pending->stop_pending;
2719 pending->stop_pending = true;
2721 if (flags & SCA_MIGRATE_ENABLE)
2722 p->migration_flags &= ~MDF_PUSH;
2724 task_rq_unlock(rq, p, rf);
2726 if (!stop_pending) {
2727 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2728 &pending->arg, &pending->stop_work);
2731 if (flags & SCA_MIGRATE_ENABLE)
2735 if (!is_migration_disabled(p)) {
2736 if (task_on_rq_queued(p))
2737 rq = move_queued_task(rq, rf, p, dest_cpu);
2739 if (!pending->stop_pending) {
2740 p->migration_pending = NULL;
2744 task_rq_unlock(rq, p, rf);
2747 complete_all(&pending->done);
2750 wait_for_completion(&pending->done);
2752 if (refcount_dec_and_test(&pending->refs))
2753 wake_up_var(&pending->refs); /* No UaF, just an address */
2756 * Block the original owner of &pending until all subsequent callers
2757 * have seen the completion and decremented the refcount
2759 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2762 WARN_ON_ONCE(my_pending.stop_pending);
2768 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2770 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2771 const struct cpumask *new_mask,
2774 struct rq_flags *rf)
2775 __releases(rq->lock)
2776 __releases(p->pi_lock)
2778 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2779 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2780 bool kthread = p->flags & PF_KTHREAD;
2781 struct cpumask *user_mask = NULL;
2782 unsigned int dest_cpu;
2785 update_rq_clock(rq);
2787 if (kthread || is_migration_disabled(p)) {
2789 * Kernel threads are allowed on online && !active CPUs,
2790 * however, during cpu-hot-unplug, even these might get pushed
2791 * away if not KTHREAD_IS_PER_CPU.
2793 * Specifically, migration_disabled() tasks must not fail the
2794 * cpumask_any_and_distribute() pick below, esp. so on
2795 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2796 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2798 cpu_valid_mask = cpu_online_mask;
2801 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2807 * Must re-check here, to close a race against __kthread_bind(),
2808 * sched_setaffinity() is not guaranteed to observe the flag.
2810 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2815 if (!(flags & SCA_MIGRATE_ENABLE)) {
2816 if (cpumask_equal(&p->cpus_mask, new_mask))
2819 if (WARN_ON_ONCE(p == current &&
2820 is_migration_disabled(p) &&
2821 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2828 * Picking a ~random cpu helps in cases where we are changing affinity
2829 * for groups of tasks (ie. cpuset), so that load balancing is not
2830 * immediately required to distribute the tasks within their new mask.
2832 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2833 if (dest_cpu >= nr_cpu_ids) {
2838 __do_set_cpus_allowed(p, new_mask, flags);
2840 if (flags & SCA_USER)
2841 user_mask = clear_user_cpus_ptr(p);
2843 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2850 task_rq_unlock(rq, p, rf);
2856 * Change a given task's CPU affinity. Migrate the thread to a
2857 * proper CPU and schedule it away if the CPU it's executing on
2858 * is removed from the allowed bitmask.
2860 * NOTE: the caller must have a valid reference to the task, the
2861 * task must not exit() & deallocate itself prematurely. The
2862 * call is not atomic; no spinlocks may be held.
2864 static int __set_cpus_allowed_ptr(struct task_struct *p,
2865 const struct cpumask *new_mask, u32 flags)
2870 rq = task_rq_lock(p, &rf);
2871 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2874 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2876 return __set_cpus_allowed_ptr(p, new_mask, 0);
2878 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2881 * Change a given task's CPU affinity to the intersection of its current
2882 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2883 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2884 * If the resulting mask is empty, leave the affinity unchanged and return
2887 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2888 struct cpumask *new_mask,
2889 const struct cpumask *subset_mask)
2891 struct cpumask *user_mask = NULL;
2896 if (!p->user_cpus_ptr) {
2897 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2902 rq = task_rq_lock(p, &rf);
2905 * Forcefully restricting the affinity of a deadline task is
2906 * likely to cause problems, so fail and noisily override the
2909 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2914 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2920 * We're about to butcher the task affinity, so keep track of what
2921 * the user asked for in case we're able to restore it later on.
2924 cpumask_copy(user_mask, p->cpus_ptr);
2925 p->user_cpus_ptr = user_mask;
2928 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2931 task_rq_unlock(rq, p, &rf);
2937 * Restrict the CPU affinity of task @p so that it is a subset of
2938 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2939 * old affinity mask. If the resulting mask is empty, we warn and walk
2940 * up the cpuset hierarchy until we find a suitable mask.
2942 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2944 cpumask_var_t new_mask;
2945 const struct cpumask *override_mask = task_cpu_possible_mask(p);
2947 alloc_cpumask_var(&new_mask, GFP_KERNEL);
2950 * __migrate_task() can fail silently in the face of concurrent
2951 * offlining of the chosen destination CPU, so take the hotplug
2952 * lock to ensure that the migration succeeds.
2955 if (!cpumask_available(new_mask))
2958 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2962 * We failed to find a valid subset of the affinity mask for the
2963 * task, so override it based on its cpuset hierarchy.
2965 cpuset_cpus_allowed(p, new_mask);
2966 override_mask = new_mask;
2969 if (printk_ratelimit()) {
2970 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2971 task_pid_nr(p), p->comm,
2972 cpumask_pr_args(override_mask));
2975 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2978 free_cpumask_var(new_mask);
2982 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
2985 * Restore the affinity of a task @p which was previously restricted by a
2986 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
2987 * @p->user_cpus_ptr.
2989 * It is the caller's responsibility to serialise this with any calls to
2990 * force_compatible_cpus_allowed_ptr(@p).
2992 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
2994 struct cpumask *user_mask = p->user_cpus_ptr;
2995 unsigned long flags;
2998 * Try to restore the old affinity mask. If this fails, then
2999 * we free the mask explicitly to avoid it being inherited across
3000 * a subsequent fork().
3002 if (!user_mask || !__sched_setaffinity(p, user_mask))
3005 raw_spin_lock_irqsave(&p->pi_lock, flags);
3006 user_mask = clear_user_cpus_ptr(p);
3007 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3012 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3014 #ifdef CONFIG_SCHED_DEBUG
3015 unsigned int state = READ_ONCE(p->__state);
3018 * We should never call set_task_cpu() on a blocked task,
3019 * ttwu() will sort out the placement.
3021 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3024 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3025 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3026 * time relying on p->on_rq.
3028 WARN_ON_ONCE(state == TASK_RUNNING &&
3029 p->sched_class == &fair_sched_class &&
3030 (p->on_rq && !task_on_rq_migrating(p)));
3032 #ifdef CONFIG_LOCKDEP
3034 * The caller should hold either p->pi_lock or rq->lock, when changing
3035 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3037 * sched_move_task() holds both and thus holding either pins the cgroup,
3040 * Furthermore, all task_rq users should acquire both locks, see
3043 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3044 lockdep_is_held(__rq_lockp(task_rq(p)))));
3047 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3049 WARN_ON_ONCE(!cpu_online(new_cpu));
3051 WARN_ON_ONCE(is_migration_disabled(p));
3054 trace_sched_migrate_task(p, new_cpu);
3056 if (task_cpu(p) != new_cpu) {
3057 if (p->sched_class->migrate_task_rq)
3058 p->sched_class->migrate_task_rq(p, new_cpu);
3059 p->se.nr_migrations++;
3061 perf_event_task_migrate(p);
3064 __set_task_cpu(p, new_cpu);
3067 #ifdef CONFIG_NUMA_BALANCING
3068 static void __migrate_swap_task(struct task_struct *p, int cpu)
3070 if (task_on_rq_queued(p)) {
3071 struct rq *src_rq, *dst_rq;
3072 struct rq_flags srf, drf;
3074 src_rq = task_rq(p);
3075 dst_rq = cpu_rq(cpu);
3077 rq_pin_lock(src_rq, &srf);
3078 rq_pin_lock(dst_rq, &drf);
3080 deactivate_task(src_rq, p, 0);
3081 set_task_cpu(p, cpu);
3082 activate_task(dst_rq, p, 0);
3083 check_preempt_curr(dst_rq, p, 0);
3085 rq_unpin_lock(dst_rq, &drf);
3086 rq_unpin_lock(src_rq, &srf);
3090 * Task isn't running anymore; make it appear like we migrated
3091 * it before it went to sleep. This means on wakeup we make the
3092 * previous CPU our target instead of where it really is.
3098 struct migration_swap_arg {
3099 struct task_struct *src_task, *dst_task;
3100 int src_cpu, dst_cpu;
3103 static int migrate_swap_stop(void *data)
3105 struct migration_swap_arg *arg = data;
3106 struct rq *src_rq, *dst_rq;
3109 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3112 src_rq = cpu_rq(arg->src_cpu);
3113 dst_rq = cpu_rq(arg->dst_cpu);
3115 double_raw_lock(&arg->src_task->pi_lock,
3116 &arg->dst_task->pi_lock);
3117 double_rq_lock(src_rq, dst_rq);
3119 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3122 if (task_cpu(arg->src_task) != arg->src_cpu)
3125 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3128 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3131 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3132 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3137 double_rq_unlock(src_rq, dst_rq);
3138 raw_spin_unlock(&arg->dst_task->pi_lock);
3139 raw_spin_unlock(&arg->src_task->pi_lock);
3145 * Cross migrate two tasks
3147 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3148 int target_cpu, int curr_cpu)
3150 struct migration_swap_arg arg;
3153 arg = (struct migration_swap_arg){
3155 .src_cpu = curr_cpu,
3157 .dst_cpu = target_cpu,
3160 if (arg.src_cpu == arg.dst_cpu)
3164 * These three tests are all lockless; this is OK since all of them
3165 * will be re-checked with proper locks held further down the line.
3167 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3170 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3173 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3176 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3177 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3182 #endif /* CONFIG_NUMA_BALANCING */
3185 * wait_task_inactive - wait for a thread to unschedule.
3187 * If @match_state is nonzero, it's the @p->state value just checked and
3188 * not expected to change. If it changes, i.e. @p might have woken up,
3189 * then return zero. When we succeed in waiting for @p to be off its CPU,
3190 * we return a positive number (its total switch count). If a second call
3191 * a short while later returns the same number, the caller can be sure that
3192 * @p has remained unscheduled the whole time.
3194 * The caller must ensure that the task *will* unschedule sometime soon,
3195 * else this function might spin for a *long* time. This function can't
3196 * be called with interrupts off, or it may introduce deadlock with
3197 * smp_call_function() if an IPI is sent by the same process we are
3198 * waiting to become inactive.
3200 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3202 int running, queued;
3209 * We do the initial early heuristics without holding
3210 * any task-queue locks at all. We'll only try to get
3211 * the runqueue lock when things look like they will
3217 * If the task is actively running on another CPU
3218 * still, just relax and busy-wait without holding
3221 * NOTE! Since we don't hold any locks, it's not
3222 * even sure that "rq" stays as the right runqueue!
3223 * But we don't care, since "task_running()" will
3224 * return false if the runqueue has changed and p
3225 * is actually now running somewhere else!
3227 while (task_running(rq, p)) {
3228 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3234 * Ok, time to look more closely! We need the rq
3235 * lock now, to be *sure*. If we're wrong, we'll
3236 * just go back and repeat.
3238 rq = task_rq_lock(p, &rf);
3239 trace_sched_wait_task(p);
3240 running = task_running(rq, p);
3241 queued = task_on_rq_queued(p);
3243 if (!match_state || READ_ONCE(p->__state) == match_state)
3244 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3245 task_rq_unlock(rq, p, &rf);
3248 * If it changed from the expected state, bail out now.
3250 if (unlikely(!ncsw))
3254 * Was it really running after all now that we
3255 * checked with the proper locks actually held?
3257 * Oops. Go back and try again..
3259 if (unlikely(running)) {
3265 * It's not enough that it's not actively running,
3266 * it must be off the runqueue _entirely_, and not
3269 * So if it was still runnable (but just not actively
3270 * running right now), it's preempted, and we should
3271 * yield - it could be a while.
3273 if (unlikely(queued)) {
3274 ktime_t to = NSEC_PER_SEC / HZ;
3276 set_current_state(TASK_UNINTERRUPTIBLE);
3277 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3282 * Ahh, all good. It wasn't running, and it wasn't
3283 * runnable, which means that it will never become
3284 * running in the future either. We're all done!
3293 * kick_process - kick a running thread to enter/exit the kernel
3294 * @p: the to-be-kicked thread
3296 * Cause a process which is running on another CPU to enter
3297 * kernel-mode, without any delay. (to get signals handled.)
3299 * NOTE: this function doesn't have to take the runqueue lock,
3300 * because all it wants to ensure is that the remote task enters
3301 * the kernel. If the IPI races and the task has been migrated
3302 * to another CPU then no harm is done and the purpose has been
3305 void kick_process(struct task_struct *p)
3311 if ((cpu != smp_processor_id()) && task_curr(p))
3312 smp_send_reschedule(cpu);
3315 EXPORT_SYMBOL_GPL(kick_process);
3318 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3320 * A few notes on cpu_active vs cpu_online:
3322 * - cpu_active must be a subset of cpu_online
3324 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3325 * see __set_cpus_allowed_ptr(). At this point the newly online
3326 * CPU isn't yet part of the sched domains, and balancing will not
3329 * - on CPU-down we clear cpu_active() to mask the sched domains and
3330 * avoid the load balancer to place new tasks on the to be removed
3331 * CPU. Existing tasks will remain running there and will be taken
3334 * This means that fallback selection must not select !active CPUs.
3335 * And can assume that any active CPU must be online. Conversely
3336 * select_task_rq() below may allow selection of !active CPUs in order
3337 * to satisfy the above rules.
3339 static int select_fallback_rq(int cpu, struct task_struct *p)
3341 int nid = cpu_to_node(cpu);
3342 const struct cpumask *nodemask = NULL;
3343 enum { cpuset, possible, fail } state = cpuset;
3347 * If the node that the CPU is on has been offlined, cpu_to_node()
3348 * will return -1. There is no CPU on the node, and we should
3349 * select the CPU on the other node.
3352 nodemask = cpumask_of_node(nid);
3354 /* Look for allowed, online CPU in same node. */
3355 for_each_cpu(dest_cpu, nodemask) {
3356 if (is_cpu_allowed(p, dest_cpu))
3362 /* Any allowed, online CPU? */
3363 for_each_cpu(dest_cpu, p->cpus_ptr) {
3364 if (!is_cpu_allowed(p, dest_cpu))
3370 /* No more Mr. Nice Guy. */
3373 if (cpuset_cpus_allowed_fallback(p)) {
3380 * XXX When called from select_task_rq() we only
3381 * hold p->pi_lock and again violate locking order.
3383 * More yuck to audit.
3385 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3395 if (state != cpuset) {
3397 * Don't tell them about moving exiting tasks or
3398 * kernel threads (both mm NULL), since they never
3401 if (p->mm && printk_ratelimit()) {
3402 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3403 task_pid_nr(p), p->comm, cpu);
3411 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3414 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3416 lockdep_assert_held(&p->pi_lock);
3418 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3419 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3421 cpu = cpumask_any(p->cpus_ptr);
3424 * In order not to call set_task_cpu() on a blocking task we need
3425 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3428 * Since this is common to all placement strategies, this lives here.
3430 * [ this allows ->select_task() to simply return task_cpu(p) and
3431 * not worry about this generic constraint ]
3433 if (unlikely(!is_cpu_allowed(p, cpu)))
3434 cpu = select_fallback_rq(task_cpu(p), p);
3439 void sched_set_stop_task(int cpu, struct task_struct *stop)
3441 static struct lock_class_key stop_pi_lock;
3442 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3443 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3447 * Make it appear like a SCHED_FIFO task, its something
3448 * userspace knows about and won't get confused about.
3450 * Also, it will make PI more or less work without too
3451 * much confusion -- but then, stop work should not
3452 * rely on PI working anyway.
3454 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3456 stop->sched_class = &stop_sched_class;
3459 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3460 * adjust the effective priority of a task. As a result,
3461 * rt_mutex_setprio() can trigger (RT) balancing operations,
3462 * which can then trigger wakeups of the stop thread to push
3463 * around the current task.
3465 * The stop task itself will never be part of the PI-chain, it
3466 * never blocks, therefore that ->pi_lock recursion is safe.
3467 * Tell lockdep about this by placing the stop->pi_lock in its
3470 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3473 cpu_rq(cpu)->stop = stop;
3477 * Reset it back to a normal scheduling class so that
3478 * it can die in pieces.
3480 old_stop->sched_class = &rt_sched_class;
3484 #else /* CONFIG_SMP */
3486 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3487 const struct cpumask *new_mask,
3490 return set_cpus_allowed_ptr(p, new_mask);
3493 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3495 static inline bool rq_has_pinned_tasks(struct rq *rq)
3500 #endif /* !CONFIG_SMP */
3503 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3507 if (!schedstat_enabled())
3513 if (cpu == rq->cpu) {
3514 __schedstat_inc(rq->ttwu_local);
3515 __schedstat_inc(p->stats.nr_wakeups_local);
3517 struct sched_domain *sd;
3519 __schedstat_inc(p->stats.nr_wakeups_remote);
3521 for_each_domain(rq->cpu, sd) {
3522 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3523 __schedstat_inc(sd->ttwu_wake_remote);
3530 if (wake_flags & WF_MIGRATED)
3531 __schedstat_inc(p->stats.nr_wakeups_migrate);
3532 #endif /* CONFIG_SMP */
3534 __schedstat_inc(rq->ttwu_count);
3535 __schedstat_inc(p->stats.nr_wakeups);
3537 if (wake_flags & WF_SYNC)
3538 __schedstat_inc(p->stats.nr_wakeups_sync);
3542 * Mark the task runnable and perform wakeup-preemption.
3544 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3545 struct rq_flags *rf)
3547 check_preempt_curr(rq, p, wake_flags);
3548 WRITE_ONCE(p->__state, TASK_RUNNING);
3549 trace_sched_wakeup(p);
3552 if (p->sched_class->task_woken) {
3554 * Our task @p is fully woken up and running; so it's safe to
3555 * drop the rq->lock, hereafter rq is only used for statistics.
3557 rq_unpin_lock(rq, rf);
3558 p->sched_class->task_woken(rq, p);
3559 rq_repin_lock(rq, rf);
3562 if (rq->idle_stamp) {
3563 u64 delta = rq_clock(rq) - rq->idle_stamp;
3564 u64 max = 2*rq->max_idle_balance_cost;
3566 update_avg(&rq->avg_idle, delta);
3568 if (rq->avg_idle > max)
3571 rq->wake_stamp = jiffies;
3572 rq->wake_avg_idle = rq->avg_idle / 2;
3580 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3581 struct rq_flags *rf)
3583 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3585 lockdep_assert_rq_held(rq);
3587 if (p->sched_contributes_to_load)
3588 rq->nr_uninterruptible--;
3591 if (wake_flags & WF_MIGRATED)
3592 en_flags |= ENQUEUE_MIGRATED;
3596 delayacct_blkio_end(p);
3597 atomic_dec(&task_rq(p)->nr_iowait);
3600 activate_task(rq, p, en_flags);
3601 ttwu_do_wakeup(rq, p, wake_flags, rf);
3605 * Consider @p being inside a wait loop:
3608 * set_current_state(TASK_UNINTERRUPTIBLE);
3615 * __set_current_state(TASK_RUNNING);
3617 * between set_current_state() and schedule(). In this case @p is still
3618 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3621 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3622 * then schedule() must still happen and p->state can be changed to
3623 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3624 * need to do a full wakeup with enqueue.
3626 * Returns: %true when the wakeup is done,
3629 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3635 rq = __task_rq_lock(p, &rf);
3636 if (task_on_rq_queued(p)) {
3637 /* check_preempt_curr() may use rq clock */
3638 update_rq_clock(rq);
3639 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3642 __task_rq_unlock(rq, &rf);
3648 void sched_ttwu_pending(void *arg)
3650 struct llist_node *llist = arg;
3651 struct rq *rq = this_rq();
3652 struct task_struct *p, *t;
3659 * rq::ttwu_pending racy indication of out-standing wakeups.
3660 * Races such that false-negatives are possible, since they
3661 * are shorter lived that false-positives would be.
3663 WRITE_ONCE(rq->ttwu_pending, 0);
3665 rq_lock_irqsave(rq, &rf);
3666 update_rq_clock(rq);
3668 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3669 if (WARN_ON_ONCE(p->on_cpu))
3670 smp_cond_load_acquire(&p->on_cpu, !VAL);
3672 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3673 set_task_cpu(p, cpu_of(rq));
3675 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3678 rq_unlock_irqrestore(rq, &rf);
3681 void send_call_function_single_ipi(int cpu)
3683 struct rq *rq = cpu_rq(cpu);
3685 if (!set_nr_if_polling(rq->idle))
3686 arch_send_call_function_single_ipi(cpu);
3688 trace_sched_wake_idle_without_ipi(cpu);
3692 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3693 * necessary. The wakee CPU on receipt of the IPI will queue the task
3694 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3695 * of the wakeup instead of the waker.
3697 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3699 struct rq *rq = cpu_rq(cpu);
3701 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3703 WRITE_ONCE(rq->ttwu_pending, 1);
3704 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3707 void wake_up_if_idle(int cpu)
3709 struct rq *rq = cpu_rq(cpu);
3714 if (!is_idle_task(rcu_dereference(rq->curr)))
3717 rq_lock_irqsave(rq, &rf);
3718 if (is_idle_task(rq->curr))
3720 /* Else CPU is not idle, do nothing here: */
3721 rq_unlock_irqrestore(rq, &rf);
3727 bool cpus_share_cache(int this_cpu, int that_cpu)
3729 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3732 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3735 * Do not complicate things with the async wake_list while the CPU is
3738 if (!cpu_active(cpu))
3742 * If the CPU does not share cache, then queue the task on the
3743 * remote rqs wakelist to avoid accessing remote data.
3745 if (!cpus_share_cache(smp_processor_id(), cpu))
3749 * If the task is descheduling and the only running task on the
3750 * CPU then use the wakelist to offload the task activation to
3751 * the soon-to-be-idle CPU as the current CPU is likely busy.
3752 * nr_running is checked to avoid unnecessary task stacking.
3754 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3760 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3762 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3763 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3766 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3767 __ttwu_queue_wakelist(p, cpu, wake_flags);
3774 #else /* !CONFIG_SMP */
3776 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3781 #endif /* CONFIG_SMP */
3783 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3785 struct rq *rq = cpu_rq(cpu);
3788 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3792 update_rq_clock(rq);
3793 ttwu_do_activate(rq, p, wake_flags, &rf);
3798 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3800 * The caller holds p::pi_lock if p != current or has preemption
3801 * disabled when p == current.
3803 * The rules of PREEMPT_RT saved_state:
3805 * The related locking code always holds p::pi_lock when updating
3806 * p::saved_state, which means the code is fully serialized in both cases.
3808 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3809 * bits set. This allows to distinguish all wakeup scenarios.
3811 static __always_inline
3812 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3814 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3815 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3816 state != TASK_RTLOCK_WAIT);
3819 if (READ_ONCE(p->__state) & state) {
3824 #ifdef CONFIG_PREEMPT_RT
3826 * Saved state preserves the task state across blocking on
3827 * an RT lock. If the state matches, set p::saved_state to
3828 * TASK_RUNNING, but do not wake the task because it waits
3829 * for a lock wakeup. Also indicate success because from
3830 * the regular waker's point of view this has succeeded.
3832 * After acquiring the lock the task will restore p::__state
3833 * from p::saved_state which ensures that the regular
3834 * wakeup is not lost. The restore will also set
3835 * p::saved_state to TASK_RUNNING so any further tests will
3836 * not result in false positives vs. @success
3838 if (p->saved_state & state) {
3839 p->saved_state = TASK_RUNNING;
3847 * Notes on Program-Order guarantees on SMP systems.
3851 * The basic program-order guarantee on SMP systems is that when a task [t]
3852 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3853 * execution on its new CPU [c1].
3855 * For migration (of runnable tasks) this is provided by the following means:
3857 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3858 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3859 * rq(c1)->lock (if not at the same time, then in that order).
3860 * C) LOCK of the rq(c1)->lock scheduling in task
3862 * Release/acquire chaining guarantees that B happens after A and C after B.
3863 * Note: the CPU doing B need not be c0 or c1
3872 * UNLOCK rq(0)->lock
3874 * LOCK rq(0)->lock // orders against CPU0
3876 * UNLOCK rq(0)->lock
3880 * UNLOCK rq(1)->lock
3882 * LOCK rq(1)->lock // orders against CPU2
3885 * UNLOCK rq(1)->lock
3888 * BLOCKING -- aka. SLEEP + WAKEUP
3890 * For blocking we (obviously) need to provide the same guarantee as for
3891 * migration. However the means are completely different as there is no lock
3892 * chain to provide order. Instead we do:
3894 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3895 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3899 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3901 * LOCK rq(0)->lock LOCK X->pi_lock
3904 * smp_store_release(X->on_cpu, 0);
3906 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3912 * X->state = RUNNING
3913 * UNLOCK rq(2)->lock
3915 * LOCK rq(2)->lock // orders against CPU1
3918 * UNLOCK rq(2)->lock
3921 * UNLOCK rq(0)->lock
3924 * However, for wakeups there is a second guarantee we must provide, namely we
3925 * must ensure that CONDITION=1 done by the caller can not be reordered with
3926 * accesses to the task state; see try_to_wake_up() and set_current_state().
3930 * try_to_wake_up - wake up a thread
3931 * @p: the thread to be awakened
3932 * @state: the mask of task states that can be woken
3933 * @wake_flags: wake modifier flags (WF_*)
3935 * Conceptually does:
3937 * If (@state & @p->state) @p->state = TASK_RUNNING.
3939 * If the task was not queued/runnable, also place it back on a runqueue.
3941 * This function is atomic against schedule() which would dequeue the task.
3943 * It issues a full memory barrier before accessing @p->state, see the comment
3944 * with set_current_state().
3946 * Uses p->pi_lock to serialize against concurrent wake-ups.
3948 * Relies on p->pi_lock stabilizing:
3951 * - p->sched_task_group
3952 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3954 * Tries really hard to only take one task_rq(p)->lock for performance.
3955 * Takes rq->lock in:
3956 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3957 * - ttwu_queue() -- new rq, for enqueue of the task;
3958 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3960 * As a consequence we race really badly with just about everything. See the
3961 * many memory barriers and their comments for details.
3963 * Return: %true if @p->state changes (an actual wakeup was done),
3967 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3969 unsigned long flags;
3970 int cpu, success = 0;
3975 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3976 * == smp_processor_id()'. Together this means we can special
3977 * case the whole 'p->on_rq && ttwu_runnable()' case below
3978 * without taking any locks.
3981 * - we rely on Program-Order guarantees for all the ordering,
3982 * - we're serialized against set_special_state() by virtue of
3983 * it disabling IRQs (this allows not taking ->pi_lock).
3985 if (!ttwu_state_match(p, state, &success))
3988 trace_sched_waking(p);
3989 WRITE_ONCE(p->__state, TASK_RUNNING);
3990 trace_sched_wakeup(p);
3995 * If we are going to wake up a thread waiting for CONDITION we
3996 * need to ensure that CONDITION=1 done by the caller can not be
3997 * reordered with p->state check below. This pairs with smp_store_mb()
3998 * in set_current_state() that the waiting thread does.
4000 raw_spin_lock_irqsave(&p->pi_lock, flags);
4001 smp_mb__after_spinlock();
4002 if (!ttwu_state_match(p, state, &success))
4005 trace_sched_waking(p);
4008 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4009 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4010 * in smp_cond_load_acquire() below.
4012 * sched_ttwu_pending() try_to_wake_up()
4013 * STORE p->on_rq = 1 LOAD p->state
4016 * __schedule() (switch to task 'p')
4017 * LOCK rq->lock smp_rmb();
4018 * smp_mb__after_spinlock();
4022 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4024 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4025 * __schedule(). See the comment for smp_mb__after_spinlock().
4027 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4030 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4035 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4036 * possible to, falsely, observe p->on_cpu == 0.
4038 * One must be running (->on_cpu == 1) in order to remove oneself
4039 * from the runqueue.
4041 * __schedule() (switch to task 'p') try_to_wake_up()
4042 * STORE p->on_cpu = 1 LOAD p->on_rq
4045 * __schedule() (put 'p' to sleep)
4046 * LOCK rq->lock smp_rmb();
4047 * smp_mb__after_spinlock();
4048 * STORE p->on_rq = 0 LOAD p->on_cpu
4050 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4051 * __schedule(). See the comment for smp_mb__after_spinlock().
4053 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4054 * schedule()'s deactivate_task() has 'happened' and p will no longer
4055 * care about it's own p->state. See the comment in __schedule().
4057 smp_acquire__after_ctrl_dep();
4060 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4061 * == 0), which means we need to do an enqueue, change p->state to
4062 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4063 * enqueue, such as ttwu_queue_wakelist().
4065 WRITE_ONCE(p->__state, TASK_WAKING);
4068 * If the owning (remote) CPU is still in the middle of schedule() with
4069 * this task as prev, considering queueing p on the remote CPUs wake_list
4070 * which potentially sends an IPI instead of spinning on p->on_cpu to
4071 * let the waker make forward progress. This is safe because IRQs are
4072 * disabled and the IPI will deliver after on_cpu is cleared.
4074 * Ensure we load task_cpu(p) after p->on_cpu:
4076 * set_task_cpu(p, cpu);
4077 * STORE p->cpu = @cpu
4078 * __schedule() (switch to task 'p')
4080 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4081 * STORE p->on_cpu = 1 LOAD p->cpu
4083 * to ensure we observe the correct CPU on which the task is currently
4086 if (smp_load_acquire(&p->on_cpu) &&
4087 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4091 * If the owning (remote) CPU is still in the middle of schedule() with
4092 * this task as prev, wait until it's done referencing the task.
4094 * Pairs with the smp_store_release() in finish_task().
4096 * This ensures that tasks getting woken will be fully ordered against
4097 * their previous state and preserve Program Order.
4099 smp_cond_load_acquire(&p->on_cpu, !VAL);
4101 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4102 if (task_cpu(p) != cpu) {
4104 delayacct_blkio_end(p);
4105 atomic_dec(&task_rq(p)->nr_iowait);
4108 wake_flags |= WF_MIGRATED;
4109 psi_ttwu_dequeue(p);
4110 set_task_cpu(p, cpu);
4114 #endif /* CONFIG_SMP */
4116 ttwu_queue(p, cpu, wake_flags);
4118 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4121 ttwu_stat(p, task_cpu(p), wake_flags);
4128 * task_call_func - Invoke a function on task in fixed state
4129 * @p: Process for which the function is to be invoked, can be @current.
4130 * @func: Function to invoke.
4131 * @arg: Argument to function.
4133 * Fix the task in it's current state by avoiding wakeups and or rq operations
4134 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4135 * to work out what the state is, if required. Given that @func can be invoked
4136 * with a runqueue lock held, it had better be quite lightweight.
4139 * Whatever @func returns
4141 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4143 struct rq *rq = NULL;
4148 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4150 state = READ_ONCE(p->__state);
4153 * Ensure we load p->on_rq after p->__state, otherwise it would be
4154 * possible to, falsely, observe p->on_rq == 0.
4156 * See try_to_wake_up() for a longer comment.
4161 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4162 * the task is blocked. Make sure to check @state since ttwu() can drop
4163 * locks at the end, see ttwu_queue_wakelist().
4165 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4166 rq = __task_rq_lock(p, &rf);
4169 * At this point the task is pinned; either:
4170 * - blocked and we're holding off wakeups (pi->lock)
4171 * - woken, and we're holding off enqueue (rq->lock)
4172 * - queued, and we're holding off schedule (rq->lock)
4173 * - running, and we're holding off de-schedule (rq->lock)
4175 * The called function (@func) can use: task_curr(), p->on_rq and
4176 * p->__state to differentiate between these states.
4183 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4188 * wake_up_process - Wake up a specific process
4189 * @p: The process to be woken up.
4191 * Attempt to wake up the nominated process and move it to the set of runnable
4194 * Return: 1 if the process was woken up, 0 if it was already running.
4196 * This function executes a full memory barrier before accessing the task state.
4198 int wake_up_process(struct task_struct *p)
4200 return try_to_wake_up(p, TASK_NORMAL, 0);
4202 EXPORT_SYMBOL(wake_up_process);
4204 int wake_up_state(struct task_struct *p, unsigned int state)
4206 return try_to_wake_up(p, state, 0);
4210 * Perform scheduler related setup for a newly forked process p.
4211 * p is forked by current.
4213 * __sched_fork() is basic setup used by init_idle() too:
4215 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4220 p->se.exec_start = 0;
4221 p->se.sum_exec_runtime = 0;
4222 p->se.prev_sum_exec_runtime = 0;
4223 p->se.nr_migrations = 0;
4225 INIT_LIST_HEAD(&p->se.group_node);
4227 #ifdef CONFIG_FAIR_GROUP_SCHED
4228 p->se.cfs_rq = NULL;
4231 #ifdef CONFIG_SCHEDSTATS
4232 /* Even if schedstat is disabled, there should not be garbage */
4233 memset(&p->stats, 0, sizeof(p->stats));
4236 RB_CLEAR_NODE(&p->dl.rb_node);
4237 init_dl_task_timer(&p->dl);
4238 init_dl_inactive_task_timer(&p->dl);
4239 __dl_clear_params(p);
4241 INIT_LIST_HEAD(&p->rt.run_list);
4243 p->rt.time_slice = sched_rr_timeslice;
4247 #ifdef CONFIG_PREEMPT_NOTIFIERS
4248 INIT_HLIST_HEAD(&p->preempt_notifiers);
4251 #ifdef CONFIG_COMPACTION
4252 p->capture_control = NULL;
4254 init_numa_balancing(clone_flags, p);
4256 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4257 p->migration_pending = NULL;
4261 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4263 #ifdef CONFIG_NUMA_BALANCING
4265 void set_numabalancing_state(bool enabled)
4268 static_branch_enable(&sched_numa_balancing);
4270 static_branch_disable(&sched_numa_balancing);
4273 #ifdef CONFIG_PROC_SYSCTL
4274 int sysctl_numa_balancing(struct ctl_table *table, int write,
4275 void *buffer, size_t *lenp, loff_t *ppos)
4279 int state = static_branch_likely(&sched_numa_balancing);
4281 if (write && !capable(CAP_SYS_ADMIN))
4286 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4290 set_numabalancing_state(state);
4296 #ifdef CONFIG_SCHEDSTATS
4298 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4300 static void set_schedstats(bool enabled)
4303 static_branch_enable(&sched_schedstats);
4305 static_branch_disable(&sched_schedstats);
4308 void force_schedstat_enabled(void)
4310 if (!schedstat_enabled()) {
4311 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4312 static_branch_enable(&sched_schedstats);
4316 static int __init setup_schedstats(char *str)
4322 if (!strcmp(str, "enable")) {
4323 set_schedstats(true);
4325 } else if (!strcmp(str, "disable")) {
4326 set_schedstats(false);
4331 pr_warn("Unable to parse schedstats=\n");
4335 __setup("schedstats=", setup_schedstats);
4337 #ifdef CONFIG_PROC_SYSCTL
4338 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4339 size_t *lenp, loff_t *ppos)
4343 int state = static_branch_likely(&sched_schedstats);
4345 if (write && !capable(CAP_SYS_ADMIN))
4350 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4354 set_schedstats(state);
4357 #endif /* CONFIG_PROC_SYSCTL */
4358 #endif /* CONFIG_SCHEDSTATS */
4361 * fork()/clone()-time setup:
4363 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4365 __sched_fork(clone_flags, p);
4367 * We mark the process as NEW here. This guarantees that
4368 * nobody will actually run it, and a signal or other external
4369 * event cannot wake it up and insert it on the runqueue either.
4371 p->__state = TASK_NEW;
4374 * Make sure we do not leak PI boosting priority to the child.
4376 p->prio = current->normal_prio;
4381 * Revert to default priority/policy on fork if requested.
4383 if (unlikely(p->sched_reset_on_fork)) {
4384 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4385 p->policy = SCHED_NORMAL;
4386 p->static_prio = NICE_TO_PRIO(0);
4388 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4389 p->static_prio = NICE_TO_PRIO(0);
4391 p->prio = p->normal_prio = p->static_prio;
4392 set_load_weight(p, false);
4395 * We don't need the reset flag anymore after the fork. It has
4396 * fulfilled its duty:
4398 p->sched_reset_on_fork = 0;
4401 if (dl_prio(p->prio))
4403 else if (rt_prio(p->prio))
4404 p->sched_class = &rt_sched_class;
4406 p->sched_class = &fair_sched_class;
4408 init_entity_runnable_average(&p->se);
4410 #ifdef CONFIG_SCHED_INFO
4411 if (likely(sched_info_on()))
4412 memset(&p->sched_info, 0, sizeof(p->sched_info));
4414 #if defined(CONFIG_SMP)
4417 init_task_preempt_count(p);
4419 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4420 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4425 void sched_post_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4427 unsigned long flags;
4428 #ifdef CONFIG_CGROUP_SCHED
4429 struct task_group *tg;
4432 raw_spin_lock_irqsave(&p->pi_lock, flags);
4433 #ifdef CONFIG_CGROUP_SCHED
4434 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4435 struct task_group, css);
4436 p->sched_task_group = autogroup_task_group(p, tg);
4440 * We're setting the CPU for the first time, we don't migrate,
4441 * so use __set_task_cpu().
4443 __set_task_cpu(p, smp_processor_id());
4444 if (p->sched_class->task_fork)
4445 p->sched_class->task_fork(p);
4446 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4448 uclamp_post_fork(p);
4451 unsigned long to_ratio(u64 period, u64 runtime)
4453 if (runtime == RUNTIME_INF)
4457 * Doing this here saves a lot of checks in all
4458 * the calling paths, and returning zero seems
4459 * safe for them anyway.
4464 return div64_u64(runtime << BW_SHIFT, period);
4468 * wake_up_new_task - wake up a newly created task for the first time.
4470 * This function will do some initial scheduler statistics housekeeping
4471 * that must be done for every newly created context, then puts the task
4472 * on the runqueue and wakes it.
4474 void wake_up_new_task(struct task_struct *p)
4479 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4480 WRITE_ONCE(p->__state, TASK_RUNNING);
4483 * Fork balancing, do it here and not earlier because:
4484 * - cpus_ptr can change in the fork path
4485 * - any previously selected CPU might disappear through hotplug
4487 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4488 * as we're not fully set-up yet.
4490 p->recent_used_cpu = task_cpu(p);
4492 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4494 rq = __task_rq_lock(p, &rf);
4495 update_rq_clock(rq);
4496 post_init_entity_util_avg(p);
4498 activate_task(rq, p, ENQUEUE_NOCLOCK);
4499 trace_sched_wakeup_new(p);
4500 check_preempt_curr(rq, p, WF_FORK);
4502 if (p->sched_class->task_woken) {
4504 * Nothing relies on rq->lock after this, so it's fine to
4507 rq_unpin_lock(rq, &rf);
4508 p->sched_class->task_woken(rq, p);
4509 rq_repin_lock(rq, &rf);
4512 task_rq_unlock(rq, p, &rf);
4515 #ifdef CONFIG_PREEMPT_NOTIFIERS
4517 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4519 void preempt_notifier_inc(void)
4521 static_branch_inc(&preempt_notifier_key);
4523 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4525 void preempt_notifier_dec(void)
4527 static_branch_dec(&preempt_notifier_key);
4529 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4532 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4533 * @notifier: notifier struct to register
4535 void preempt_notifier_register(struct preempt_notifier *notifier)
4537 if (!static_branch_unlikely(&preempt_notifier_key))
4538 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4540 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4542 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4545 * preempt_notifier_unregister - no longer interested in preemption notifications
4546 * @notifier: notifier struct to unregister
4548 * This is *not* safe to call from within a preemption notifier.
4550 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4552 hlist_del(¬ifier->link);
4554 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4556 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4558 struct preempt_notifier *notifier;
4560 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4561 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4564 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4566 if (static_branch_unlikely(&preempt_notifier_key))
4567 __fire_sched_in_preempt_notifiers(curr);
4571 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4572 struct task_struct *next)
4574 struct preempt_notifier *notifier;
4576 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4577 notifier->ops->sched_out(notifier, next);
4580 static __always_inline void
4581 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4582 struct task_struct *next)
4584 if (static_branch_unlikely(&preempt_notifier_key))
4585 __fire_sched_out_preempt_notifiers(curr, next);
4588 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4590 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4595 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4596 struct task_struct *next)
4600 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4602 static inline void prepare_task(struct task_struct *next)
4606 * Claim the task as running, we do this before switching to it
4607 * such that any running task will have this set.
4609 * See the ttwu() WF_ON_CPU case and its ordering comment.
4611 WRITE_ONCE(next->on_cpu, 1);
4615 static inline void finish_task(struct task_struct *prev)
4619 * This must be the very last reference to @prev from this CPU. After
4620 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4621 * must ensure this doesn't happen until the switch is completely
4624 * In particular, the load of prev->state in finish_task_switch() must
4625 * happen before this.
4627 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4629 smp_store_release(&prev->on_cpu, 0);
4635 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4637 void (*func)(struct rq *rq);
4638 struct callback_head *next;
4640 lockdep_assert_rq_held(rq);
4643 func = (void (*)(struct rq *))head->func;
4652 static void balance_push(struct rq *rq);
4654 struct callback_head balance_push_callback = {
4656 .func = (void (*)(struct callback_head *))balance_push,
4659 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4661 struct callback_head *head = rq->balance_callback;
4663 lockdep_assert_rq_held(rq);
4665 rq->balance_callback = NULL;
4670 static void __balance_callbacks(struct rq *rq)
4672 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4675 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4677 unsigned long flags;
4679 if (unlikely(head)) {
4680 raw_spin_rq_lock_irqsave(rq, flags);
4681 do_balance_callbacks(rq, head);
4682 raw_spin_rq_unlock_irqrestore(rq, flags);
4688 static inline void __balance_callbacks(struct rq *rq)
4692 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4697 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4704 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4707 * Since the runqueue lock will be released by the next
4708 * task (which is an invalid locking op but in the case
4709 * of the scheduler it's an obvious special-case), so we
4710 * do an early lockdep release here:
4712 rq_unpin_lock(rq, rf);
4713 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4714 #ifdef CONFIG_DEBUG_SPINLOCK
4715 /* this is a valid case when another task releases the spinlock */
4716 rq_lockp(rq)->owner = next;
4720 static inline void finish_lock_switch(struct rq *rq)
4723 * If we are tracking spinlock dependencies then we have to
4724 * fix up the runqueue lock - which gets 'carried over' from
4725 * prev into current:
4727 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4728 __balance_callbacks(rq);
4729 raw_spin_rq_unlock_irq(rq);
4733 * NOP if the arch has not defined these:
4736 #ifndef prepare_arch_switch
4737 # define prepare_arch_switch(next) do { } while (0)
4740 #ifndef finish_arch_post_lock_switch
4741 # define finish_arch_post_lock_switch() do { } while (0)
4744 static inline void kmap_local_sched_out(void)
4746 #ifdef CONFIG_KMAP_LOCAL
4747 if (unlikely(current->kmap_ctrl.idx))
4748 __kmap_local_sched_out();
4752 static inline void kmap_local_sched_in(void)
4754 #ifdef CONFIG_KMAP_LOCAL
4755 if (unlikely(current->kmap_ctrl.idx))
4756 __kmap_local_sched_in();
4761 * prepare_task_switch - prepare to switch tasks
4762 * @rq: the runqueue preparing to switch
4763 * @prev: the current task that is being switched out
4764 * @next: the task we are going to switch to.
4766 * This is called with the rq lock held and interrupts off. It must
4767 * be paired with a subsequent finish_task_switch after the context
4770 * prepare_task_switch sets up locking and calls architecture specific
4774 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4775 struct task_struct *next)
4777 kcov_prepare_switch(prev);
4778 sched_info_switch(rq, prev, next);
4779 perf_event_task_sched_out(prev, next);
4781 fire_sched_out_preempt_notifiers(prev, next);
4782 kmap_local_sched_out();
4784 prepare_arch_switch(next);
4788 * finish_task_switch - clean up after a task-switch
4789 * @prev: the thread we just switched away from.
4791 * finish_task_switch must be called after the context switch, paired
4792 * with a prepare_task_switch call before the context switch.
4793 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4794 * and do any other architecture-specific cleanup actions.
4796 * Note that we may have delayed dropping an mm in context_switch(). If
4797 * so, we finish that here outside of the runqueue lock. (Doing it
4798 * with the lock held can cause deadlocks; see schedule() for
4801 * The context switch have flipped the stack from under us and restored the
4802 * local variables which were saved when this task called schedule() in the
4803 * past. prev == current is still correct but we need to recalculate this_rq
4804 * because prev may have moved to another CPU.
4806 static struct rq *finish_task_switch(struct task_struct *prev)
4807 __releases(rq->lock)
4809 struct rq *rq = this_rq();
4810 struct mm_struct *mm = rq->prev_mm;
4814 * The previous task will have left us with a preempt_count of 2
4815 * because it left us after:
4818 * preempt_disable(); // 1
4820 * raw_spin_lock_irq(&rq->lock) // 2
4822 * Also, see FORK_PREEMPT_COUNT.
4824 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4825 "corrupted preempt_count: %s/%d/0x%x\n",
4826 current->comm, current->pid, preempt_count()))
4827 preempt_count_set(FORK_PREEMPT_COUNT);
4832 * A task struct has one reference for the use as "current".
4833 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4834 * schedule one last time. The schedule call will never return, and
4835 * the scheduled task must drop that reference.
4837 * We must observe prev->state before clearing prev->on_cpu (in
4838 * finish_task), otherwise a concurrent wakeup can get prev
4839 * running on another CPU and we could rave with its RUNNING -> DEAD
4840 * transition, resulting in a double drop.
4842 prev_state = READ_ONCE(prev->__state);
4843 vtime_task_switch(prev);
4844 perf_event_task_sched_in(prev, current);
4846 tick_nohz_task_switch();
4847 finish_lock_switch(rq);
4848 finish_arch_post_lock_switch();
4849 kcov_finish_switch(current);
4851 * kmap_local_sched_out() is invoked with rq::lock held and
4852 * interrupts disabled. There is no requirement for that, but the
4853 * sched out code does not have an interrupt enabled section.
4854 * Restoring the maps on sched in does not require interrupts being
4857 kmap_local_sched_in();
4859 fire_sched_in_preempt_notifiers(current);
4861 * When switching through a kernel thread, the loop in
4862 * membarrier_{private,global}_expedited() may have observed that
4863 * kernel thread and not issued an IPI. It is therefore possible to
4864 * schedule between user->kernel->user threads without passing though
4865 * switch_mm(). Membarrier requires a barrier after storing to
4866 * rq->curr, before returning to userspace, so provide them here:
4868 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4869 * provided by mmdrop(),
4870 * - a sync_core for SYNC_CORE.
4873 membarrier_mm_sync_core_before_usermode(mm);
4876 if (unlikely(prev_state == TASK_DEAD)) {
4877 if (prev->sched_class->task_dead)
4878 prev->sched_class->task_dead(prev);
4880 /* Task is done with its stack. */
4881 put_task_stack(prev);
4883 put_task_struct_rcu_user(prev);
4890 * schedule_tail - first thing a freshly forked thread must call.
4891 * @prev: the thread we just switched away from.
4893 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4894 __releases(rq->lock)
4897 * New tasks start with FORK_PREEMPT_COUNT, see there and
4898 * finish_task_switch() for details.
4900 * finish_task_switch() will drop rq->lock() and lower preempt_count
4901 * and the preempt_enable() will end up enabling preemption (on
4902 * PREEMPT_COUNT kernels).
4905 finish_task_switch(prev);
4908 if (current->set_child_tid)
4909 put_user(task_pid_vnr(current), current->set_child_tid);
4911 calculate_sigpending();
4915 * context_switch - switch to the new MM and the new thread's register state.
4917 static __always_inline struct rq *
4918 context_switch(struct rq *rq, struct task_struct *prev,
4919 struct task_struct *next, struct rq_flags *rf)
4921 prepare_task_switch(rq, prev, next);
4924 * For paravirt, this is coupled with an exit in switch_to to
4925 * combine the page table reload and the switch backend into
4928 arch_start_context_switch(prev);
4931 * kernel -> kernel lazy + transfer active
4932 * user -> kernel lazy + mmgrab() active
4934 * kernel -> user switch + mmdrop() active
4935 * user -> user switch
4937 if (!next->mm) { // to kernel
4938 enter_lazy_tlb(prev->active_mm, next);
4940 next->active_mm = prev->active_mm;
4941 if (prev->mm) // from user
4942 mmgrab(prev->active_mm);
4944 prev->active_mm = NULL;
4946 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4948 * sys_membarrier() requires an smp_mb() between setting
4949 * rq->curr / membarrier_switch_mm() and returning to userspace.
4951 * The below provides this either through switch_mm(), or in
4952 * case 'prev->active_mm == next->mm' through
4953 * finish_task_switch()'s mmdrop().
4955 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4957 if (!prev->mm) { // from kernel
4958 /* will mmdrop() in finish_task_switch(). */
4959 rq->prev_mm = prev->active_mm;
4960 prev->active_mm = NULL;
4964 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4966 prepare_lock_switch(rq, next, rf);
4968 /* Here we just switch the register state and the stack. */
4969 switch_to(prev, next, prev);
4972 return finish_task_switch(prev);
4976 * nr_running and nr_context_switches:
4978 * externally visible scheduler statistics: current number of runnable
4979 * threads, total number of context switches performed since bootup.
4981 unsigned int nr_running(void)
4983 unsigned int i, sum = 0;
4985 for_each_online_cpu(i)
4986 sum += cpu_rq(i)->nr_running;
4992 * Check if only the current task is running on the CPU.
4994 * Caution: this function does not check that the caller has disabled
4995 * preemption, thus the result might have a time-of-check-to-time-of-use
4996 * race. The caller is responsible to use it correctly, for example:
4998 * - from a non-preemptible section (of course)
5000 * - from a thread that is bound to a single CPU
5002 * - in a loop with very short iterations (e.g. a polling loop)
5004 bool single_task_running(void)
5006 return raw_rq()->nr_running == 1;
5008 EXPORT_SYMBOL(single_task_running);
5010 unsigned long long nr_context_switches(void)
5013 unsigned long long sum = 0;
5015 for_each_possible_cpu(i)
5016 sum += cpu_rq(i)->nr_switches;
5022 * Consumers of these two interfaces, like for example the cpuidle menu
5023 * governor, are using nonsensical data. Preferring shallow idle state selection
5024 * for a CPU that has IO-wait which might not even end up running the task when
5025 * it does become runnable.
5028 unsigned int nr_iowait_cpu(int cpu)
5030 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5034 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5036 * The idea behind IO-wait account is to account the idle time that we could
5037 * have spend running if it were not for IO. That is, if we were to improve the
5038 * storage performance, we'd have a proportional reduction in IO-wait time.
5040 * This all works nicely on UP, where, when a task blocks on IO, we account
5041 * idle time as IO-wait, because if the storage were faster, it could've been
5042 * running and we'd not be idle.
5044 * This has been extended to SMP, by doing the same for each CPU. This however
5047 * Imagine for instance the case where two tasks block on one CPU, only the one
5048 * CPU will have IO-wait accounted, while the other has regular idle. Even
5049 * though, if the storage were faster, both could've ran at the same time,
5050 * utilising both CPUs.
5052 * This means, that when looking globally, the current IO-wait accounting on
5053 * SMP is a lower bound, by reason of under accounting.
5055 * Worse, since the numbers are provided per CPU, they are sometimes
5056 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5057 * associated with any one particular CPU, it can wake to another CPU than it
5058 * blocked on. This means the per CPU IO-wait number is meaningless.
5060 * Task CPU affinities can make all that even more 'interesting'.
5063 unsigned int nr_iowait(void)
5065 unsigned int i, sum = 0;
5067 for_each_possible_cpu(i)
5068 sum += nr_iowait_cpu(i);
5076 * sched_exec - execve() is a valuable balancing opportunity, because at
5077 * this point the task has the smallest effective memory and cache footprint.
5079 void sched_exec(void)
5081 struct task_struct *p = current;
5082 unsigned long flags;
5085 raw_spin_lock_irqsave(&p->pi_lock, flags);
5086 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5087 if (dest_cpu == smp_processor_id())
5090 if (likely(cpu_active(dest_cpu))) {
5091 struct migration_arg arg = { p, dest_cpu };
5093 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5094 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5098 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5103 DEFINE_PER_CPU(struct kernel_stat, kstat);
5104 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5106 EXPORT_PER_CPU_SYMBOL(kstat);
5107 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5110 * The function fair_sched_class.update_curr accesses the struct curr
5111 * and its field curr->exec_start; when called from task_sched_runtime(),
5112 * we observe a high rate of cache misses in practice.
5113 * Prefetching this data results in improved performance.
5115 static inline void prefetch_curr_exec_start(struct task_struct *p)
5117 #ifdef CONFIG_FAIR_GROUP_SCHED
5118 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5120 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5123 prefetch(&curr->exec_start);
5127 * Return accounted runtime for the task.
5128 * In case the task is currently running, return the runtime plus current's
5129 * pending runtime that have not been accounted yet.
5131 unsigned long long task_sched_runtime(struct task_struct *p)
5137 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5139 * 64-bit doesn't need locks to atomically read a 64-bit value.
5140 * So we have a optimization chance when the task's delta_exec is 0.
5141 * Reading ->on_cpu is racy, but this is ok.
5143 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5144 * If we race with it entering CPU, unaccounted time is 0. This is
5145 * indistinguishable from the read occurring a few cycles earlier.
5146 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5147 * been accounted, so we're correct here as well.
5149 if (!p->on_cpu || !task_on_rq_queued(p))
5150 return p->se.sum_exec_runtime;
5153 rq = task_rq_lock(p, &rf);
5155 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5156 * project cycles that may never be accounted to this
5157 * thread, breaking clock_gettime().
5159 if (task_current(rq, p) && task_on_rq_queued(p)) {
5160 prefetch_curr_exec_start(p);
5161 update_rq_clock(rq);
5162 p->sched_class->update_curr(rq);
5164 ns = p->se.sum_exec_runtime;
5165 task_rq_unlock(rq, p, &rf);
5170 #ifdef CONFIG_SCHED_DEBUG
5171 static u64 cpu_resched_latency(struct rq *rq)
5173 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5174 u64 resched_latency, now = rq_clock(rq);
5175 static bool warned_once;
5177 if (sysctl_resched_latency_warn_once && warned_once)
5180 if (!need_resched() || !latency_warn_ms)
5183 if (system_state == SYSTEM_BOOTING)
5186 if (!rq->last_seen_need_resched_ns) {
5187 rq->last_seen_need_resched_ns = now;
5188 rq->ticks_without_resched = 0;
5192 rq->ticks_without_resched++;
5193 resched_latency = now - rq->last_seen_need_resched_ns;
5194 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5199 return resched_latency;
5202 static int __init setup_resched_latency_warn_ms(char *str)
5206 if ((kstrtol(str, 0, &val))) {
5207 pr_warn("Unable to set resched_latency_warn_ms\n");
5211 sysctl_resched_latency_warn_ms = val;
5214 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5216 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5217 #endif /* CONFIG_SCHED_DEBUG */
5220 * This function gets called by the timer code, with HZ frequency.
5221 * We call it with interrupts disabled.
5223 void scheduler_tick(void)
5225 int cpu = smp_processor_id();
5226 struct rq *rq = cpu_rq(cpu);
5227 struct task_struct *curr = rq->curr;
5229 unsigned long thermal_pressure;
5230 u64 resched_latency;
5232 arch_scale_freq_tick();
5237 update_rq_clock(rq);
5238 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5239 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5240 curr->sched_class->task_tick(rq, curr, 0);
5241 if (sched_feat(LATENCY_WARN))
5242 resched_latency = cpu_resched_latency(rq);
5243 calc_global_load_tick(rq);
5247 if (sched_feat(LATENCY_WARN) && resched_latency)
5248 resched_latency_warn(cpu, resched_latency);
5250 perf_event_task_tick();
5253 rq->idle_balance = idle_cpu(cpu);
5254 trigger_load_balance(rq);
5258 #ifdef CONFIG_NO_HZ_FULL
5263 struct delayed_work work;
5265 /* Values for ->state, see diagram below. */
5266 #define TICK_SCHED_REMOTE_OFFLINE 0
5267 #define TICK_SCHED_REMOTE_OFFLINING 1
5268 #define TICK_SCHED_REMOTE_RUNNING 2
5271 * State diagram for ->state:
5274 * TICK_SCHED_REMOTE_OFFLINE
5277 * | | sched_tick_remote()
5280 * +--TICK_SCHED_REMOTE_OFFLINING
5283 * sched_tick_start() | | sched_tick_stop()
5286 * TICK_SCHED_REMOTE_RUNNING
5289 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5290 * and sched_tick_start() are happy to leave the state in RUNNING.
5293 static struct tick_work __percpu *tick_work_cpu;
5295 static void sched_tick_remote(struct work_struct *work)
5297 struct delayed_work *dwork = to_delayed_work(work);
5298 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5299 int cpu = twork->cpu;
5300 struct rq *rq = cpu_rq(cpu);
5301 struct task_struct *curr;
5307 * Handle the tick only if it appears the remote CPU is running in full
5308 * dynticks mode. The check is racy by nature, but missing a tick or
5309 * having one too much is no big deal because the scheduler tick updates
5310 * statistics and checks timeslices in a time-independent way, regardless
5311 * of when exactly it is running.
5313 if (!tick_nohz_tick_stopped_cpu(cpu))
5316 rq_lock_irq(rq, &rf);
5318 if (cpu_is_offline(cpu))
5321 update_rq_clock(rq);
5323 if (!is_idle_task(curr)) {
5325 * Make sure the next tick runs within a reasonable
5328 delta = rq_clock_task(rq) - curr->se.exec_start;
5329 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5331 curr->sched_class->task_tick(rq, curr, 0);
5333 calc_load_nohz_remote(rq);
5335 rq_unlock_irq(rq, &rf);
5339 * Run the remote tick once per second (1Hz). This arbitrary
5340 * frequency is large enough to avoid overload but short enough
5341 * to keep scheduler internal stats reasonably up to date. But
5342 * first update state to reflect hotplug activity if required.
5344 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5345 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5346 if (os == TICK_SCHED_REMOTE_RUNNING)
5347 queue_delayed_work(system_unbound_wq, dwork, HZ);
5350 static void sched_tick_start(int cpu)
5353 struct tick_work *twork;
5355 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5358 WARN_ON_ONCE(!tick_work_cpu);
5360 twork = per_cpu_ptr(tick_work_cpu, cpu);
5361 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5362 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5363 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5365 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5366 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5370 #ifdef CONFIG_HOTPLUG_CPU
5371 static void sched_tick_stop(int cpu)
5373 struct tick_work *twork;
5376 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5379 WARN_ON_ONCE(!tick_work_cpu);
5381 twork = per_cpu_ptr(tick_work_cpu, cpu);
5382 /* There cannot be competing actions, but don't rely on stop-machine. */
5383 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5384 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5385 /* Don't cancel, as this would mess up the state machine. */
5387 #endif /* CONFIG_HOTPLUG_CPU */
5389 int __init sched_tick_offload_init(void)
5391 tick_work_cpu = alloc_percpu(struct tick_work);
5392 BUG_ON(!tick_work_cpu);
5396 #else /* !CONFIG_NO_HZ_FULL */
5397 static inline void sched_tick_start(int cpu) { }
5398 static inline void sched_tick_stop(int cpu) { }
5401 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5402 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5404 * If the value passed in is equal to the current preempt count
5405 * then we just disabled preemption. Start timing the latency.
5407 static inline void preempt_latency_start(int val)
5409 if (preempt_count() == val) {
5410 unsigned long ip = get_lock_parent_ip();
5411 #ifdef CONFIG_DEBUG_PREEMPT
5412 current->preempt_disable_ip = ip;
5414 trace_preempt_off(CALLER_ADDR0, ip);
5418 void preempt_count_add(int val)
5420 #ifdef CONFIG_DEBUG_PREEMPT
5424 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5427 __preempt_count_add(val);
5428 #ifdef CONFIG_DEBUG_PREEMPT
5430 * Spinlock count overflowing soon?
5432 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5435 preempt_latency_start(val);
5437 EXPORT_SYMBOL(preempt_count_add);
5438 NOKPROBE_SYMBOL(preempt_count_add);
5441 * If the value passed in equals to the current preempt count
5442 * then we just enabled preemption. Stop timing the latency.
5444 static inline void preempt_latency_stop(int val)
5446 if (preempt_count() == val)
5447 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5450 void preempt_count_sub(int val)
5452 #ifdef CONFIG_DEBUG_PREEMPT
5456 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5459 * Is the spinlock portion underflowing?
5461 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5462 !(preempt_count() & PREEMPT_MASK)))
5466 preempt_latency_stop(val);
5467 __preempt_count_sub(val);
5469 EXPORT_SYMBOL(preempt_count_sub);
5470 NOKPROBE_SYMBOL(preempt_count_sub);
5473 static inline void preempt_latency_start(int val) { }
5474 static inline void preempt_latency_stop(int val) { }
5477 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5479 #ifdef CONFIG_DEBUG_PREEMPT
5480 return p->preempt_disable_ip;
5487 * Print scheduling while atomic bug:
5489 static noinline void __schedule_bug(struct task_struct *prev)
5491 /* Save this before calling printk(), since that will clobber it */
5492 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5494 if (oops_in_progress)
5497 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5498 prev->comm, prev->pid, preempt_count());
5500 debug_show_held_locks(prev);
5502 if (irqs_disabled())
5503 print_irqtrace_events(prev);
5504 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5505 && in_atomic_preempt_off()) {
5506 pr_err("Preemption disabled at:");
5507 print_ip_sym(KERN_ERR, preempt_disable_ip);
5510 panic("scheduling while atomic\n");
5513 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5517 * Various schedule()-time debugging checks and statistics:
5519 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5521 #ifdef CONFIG_SCHED_STACK_END_CHECK
5522 if (task_stack_end_corrupted(prev))
5523 panic("corrupted stack end detected inside scheduler\n");
5525 if (task_scs_end_corrupted(prev))
5526 panic("corrupted shadow stack detected inside scheduler\n");
5529 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5530 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5531 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5532 prev->comm, prev->pid, prev->non_block_count);
5534 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5538 if (unlikely(in_atomic_preempt_off())) {
5539 __schedule_bug(prev);
5540 preempt_count_set(PREEMPT_DISABLED);
5543 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5545 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5547 schedstat_inc(this_rq()->sched_count);
5550 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5551 struct rq_flags *rf)
5554 const struct sched_class *class;
5556 * We must do the balancing pass before put_prev_task(), such
5557 * that when we release the rq->lock the task is in the same
5558 * state as before we took rq->lock.
5560 * We can terminate the balance pass as soon as we know there is
5561 * a runnable task of @class priority or higher.
5563 for_class_range(class, prev->sched_class, &idle_sched_class) {
5564 if (class->balance(rq, prev, rf))
5569 put_prev_task(rq, prev);
5573 * Pick up the highest-prio task:
5575 static inline struct task_struct *
5576 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5578 const struct sched_class *class;
5579 struct task_struct *p;
5582 * Optimization: we know that if all tasks are in the fair class we can
5583 * call that function directly, but only if the @prev task wasn't of a
5584 * higher scheduling class, because otherwise those lose the
5585 * opportunity to pull in more work from other CPUs.
5587 if (likely(prev->sched_class <= &fair_sched_class &&
5588 rq->nr_running == rq->cfs.h_nr_running)) {
5590 p = pick_next_task_fair(rq, prev, rf);
5591 if (unlikely(p == RETRY_TASK))
5594 /* Assume the next prioritized class is idle_sched_class */
5596 put_prev_task(rq, prev);
5597 p = pick_next_task_idle(rq);
5604 put_prev_task_balance(rq, prev, rf);
5606 for_each_class(class) {
5607 p = class->pick_next_task(rq);
5612 BUG(); /* The idle class should always have a runnable task. */
5615 #ifdef CONFIG_SCHED_CORE
5616 static inline bool is_task_rq_idle(struct task_struct *t)
5618 return (task_rq(t)->idle == t);
5621 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5623 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5626 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5628 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5631 return a->core_cookie == b->core_cookie;
5634 static inline struct task_struct *pick_task(struct rq *rq)
5636 const struct sched_class *class;
5637 struct task_struct *p;
5639 for_each_class(class) {
5640 p = class->pick_task(rq);
5645 BUG(); /* The idle class should always have a runnable task. */
5648 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5650 static struct task_struct *
5651 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5653 struct task_struct *next, *p, *max = NULL;
5654 const struct cpumask *smt_mask;
5655 bool fi_before = false;
5656 unsigned long cookie;
5657 int i, cpu, occ = 0;
5661 if (!sched_core_enabled(rq))
5662 return __pick_next_task(rq, prev, rf);
5666 /* Stopper task is switching into idle, no need core-wide selection. */
5667 if (cpu_is_offline(cpu)) {
5669 * Reset core_pick so that we don't enter the fastpath when
5670 * coming online. core_pick would already be migrated to
5671 * another cpu during offline.
5673 rq->core_pick = NULL;
5674 return __pick_next_task(rq, prev, rf);
5678 * If there were no {en,de}queues since we picked (IOW, the task
5679 * pointers are all still valid), and we haven't scheduled the last
5680 * pick yet, do so now.
5682 * rq->core_pick can be NULL if no selection was made for a CPU because
5683 * it was either offline or went offline during a sibling's core-wide
5684 * selection. In this case, do a core-wide selection.
5686 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5687 rq->core->core_pick_seq != rq->core_sched_seq &&
5689 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5691 next = rq->core_pick;
5693 put_prev_task(rq, prev);
5694 set_next_task(rq, next);
5697 rq->core_pick = NULL;
5701 put_prev_task_balance(rq, prev, rf);
5703 smt_mask = cpu_smt_mask(cpu);
5704 need_sync = !!rq->core->core_cookie;
5707 rq->core->core_cookie = 0UL;
5708 if (rq->core->core_forceidle) {
5711 rq->core->core_forceidle = false;
5715 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5717 * @task_seq guards the task state ({en,de}queues)
5718 * @pick_seq is the @task_seq we did a selection on
5719 * @sched_seq is the @pick_seq we scheduled
5721 * However, preemptions can cause multiple picks on the same task set.
5722 * 'Fix' this by also increasing @task_seq for every pick.
5724 rq->core->core_task_seq++;
5727 * Optimize for common case where this CPU has no cookies
5728 * and there are no cookied tasks running on siblings.
5731 next = pick_task(rq);
5732 if (!next->core_cookie) {
5733 rq->core_pick = NULL;
5735 * For robustness, update the min_vruntime_fi for
5736 * unconstrained picks as well.
5738 WARN_ON_ONCE(fi_before);
5739 task_vruntime_update(rq, next, false);
5745 * For each thread: do the regular task pick and find the max prio task
5748 * Tie-break prio towards the current CPU
5750 for_each_cpu_wrap(i, smt_mask, cpu) {
5754 update_rq_clock(rq_i);
5756 p = rq_i->core_pick = pick_task(rq_i);
5757 if (!max || prio_less(max, p, fi_before))
5761 cookie = rq->core->core_cookie = max->core_cookie;
5764 * For each thread: try and find a runnable task that matches @max or
5767 for_each_cpu(i, smt_mask) {
5769 p = rq_i->core_pick;
5771 if (!cookie_equals(p, cookie)) {
5774 p = sched_core_find(rq_i, cookie);
5776 p = idle_sched_class.pick_task(rq_i);
5779 rq_i->core_pick = p;
5781 if (p == rq_i->idle) {
5782 if (rq_i->nr_running) {
5783 rq->core->core_forceidle = true;
5785 rq->core->core_forceidle_seq++;
5792 rq->core->core_pick_seq = rq->core->core_task_seq;
5793 next = rq->core_pick;
5794 rq->core_sched_seq = rq->core->core_pick_seq;
5796 /* Something should have been selected for current CPU */
5797 WARN_ON_ONCE(!next);
5800 * Reschedule siblings
5802 * NOTE: L1TF -- at this point we're no longer running the old task and
5803 * sending an IPI (below) ensures the sibling will no longer be running
5804 * their task. This ensures there is no inter-sibling overlap between
5805 * non-matching user state.
5807 for_each_cpu(i, smt_mask) {
5811 * An online sibling might have gone offline before a task
5812 * could be picked for it, or it might be offline but later
5813 * happen to come online, but its too late and nothing was
5814 * picked for it. That's Ok - it will pick tasks for itself,
5817 if (!rq_i->core_pick)
5821 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5822 * fi_before fi update?
5828 if (!(fi_before && rq->core->core_forceidle))
5829 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5831 rq_i->core_pick->core_occupation = occ;
5834 rq_i->core_pick = NULL;
5838 /* Did we break L1TF mitigation requirements? */
5839 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5841 if (rq_i->curr == rq_i->core_pick) {
5842 rq_i->core_pick = NULL;
5850 set_next_task(rq, next);
5854 static bool try_steal_cookie(int this, int that)
5856 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5857 struct task_struct *p;
5858 unsigned long cookie;
5859 bool success = false;
5861 local_irq_disable();
5862 double_rq_lock(dst, src);
5864 cookie = dst->core->core_cookie;
5868 if (dst->curr != dst->idle)
5871 p = sched_core_find(src, cookie);
5876 if (p == src->core_pick || p == src->curr)
5879 if (!cpumask_test_cpu(this, &p->cpus_mask))
5882 if (p->core_occupation > dst->idle->core_occupation)
5885 deactivate_task(src, p, 0);
5886 set_task_cpu(p, this);
5887 activate_task(dst, p, 0);
5895 p = sched_core_next(p, cookie);
5899 double_rq_unlock(dst, src);
5905 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5909 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5916 if (try_steal_cookie(cpu, i))
5923 static void sched_core_balance(struct rq *rq)
5925 struct sched_domain *sd;
5926 int cpu = cpu_of(rq);
5930 raw_spin_rq_unlock_irq(rq);
5931 for_each_domain(cpu, sd) {
5935 if (steal_cookie_task(cpu, sd))
5938 raw_spin_rq_lock_irq(rq);
5943 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5945 void queue_core_balance(struct rq *rq)
5947 if (!sched_core_enabled(rq))
5950 if (!rq->core->core_cookie)
5953 if (!rq->nr_running) /* not forced idle */
5956 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5959 static void sched_core_cpu_starting(unsigned int cpu)
5961 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5962 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
5963 unsigned long flags;
5966 sched_core_lock(cpu, &flags);
5968 WARN_ON_ONCE(rq->core != rq);
5970 /* if we're the first, we'll be our own leader */
5971 if (cpumask_weight(smt_mask) == 1)
5974 /* find the leader */
5975 for_each_cpu(t, smt_mask) {
5979 if (rq->core == rq) {
5985 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
5988 /* install and validate core_rq */
5989 for_each_cpu(t, smt_mask) {
5995 WARN_ON_ONCE(rq->core != core_rq);
5999 sched_core_unlock(cpu, &flags);
6002 static void sched_core_cpu_deactivate(unsigned int cpu)
6004 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6005 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6006 unsigned long flags;
6009 sched_core_lock(cpu, &flags);
6011 /* if we're the last man standing, nothing to do */
6012 if (cpumask_weight(smt_mask) == 1) {
6013 WARN_ON_ONCE(rq->core != rq);
6017 /* if we're not the leader, nothing to do */
6021 /* find a new leader */
6022 for_each_cpu(t, smt_mask) {
6025 core_rq = cpu_rq(t);
6029 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6032 /* copy the shared state to the new leader */
6033 core_rq->core_task_seq = rq->core_task_seq;
6034 core_rq->core_pick_seq = rq->core_pick_seq;
6035 core_rq->core_cookie = rq->core_cookie;
6036 core_rq->core_forceidle = rq->core_forceidle;
6037 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6039 /* install new leader */
6040 for_each_cpu(t, smt_mask) {
6046 sched_core_unlock(cpu, &flags);
6049 static inline void sched_core_cpu_dying(unsigned int cpu)
6051 struct rq *rq = cpu_rq(cpu);
6057 #else /* !CONFIG_SCHED_CORE */
6059 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6060 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6061 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6063 static struct task_struct *
6064 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6066 return __pick_next_task(rq, prev, rf);
6069 #endif /* CONFIG_SCHED_CORE */
6072 * Constants for the sched_mode argument of __schedule().
6074 * The mode argument allows RT enabled kernels to differentiate a
6075 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6076 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6077 * optimize the AND operation out and just check for zero.
6080 #define SM_PREEMPT 0x1
6081 #define SM_RTLOCK_WAIT 0x2
6083 #ifndef CONFIG_PREEMPT_RT
6084 # define SM_MASK_PREEMPT (~0U)
6086 # define SM_MASK_PREEMPT SM_PREEMPT
6090 * __schedule() is the main scheduler function.
6092 * The main means of driving the scheduler and thus entering this function are:
6094 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6096 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6097 * paths. For example, see arch/x86/entry_64.S.
6099 * To drive preemption between tasks, the scheduler sets the flag in timer
6100 * interrupt handler scheduler_tick().
6102 * 3. Wakeups don't really cause entry into schedule(). They add a
6103 * task to the run-queue and that's it.
6105 * Now, if the new task added to the run-queue preempts the current
6106 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6107 * called on the nearest possible occasion:
6109 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6111 * - in syscall or exception context, at the next outmost
6112 * preempt_enable(). (this might be as soon as the wake_up()'s
6115 * - in IRQ context, return from interrupt-handler to
6116 * preemptible context
6118 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6121 * - cond_resched() call
6122 * - explicit schedule() call
6123 * - return from syscall or exception to user-space
6124 * - return from interrupt-handler to user-space
6126 * WARNING: must be called with preemption disabled!
6128 static void __sched notrace __schedule(unsigned int sched_mode)
6130 struct task_struct *prev, *next;
6131 unsigned long *switch_count;
6132 unsigned long prev_state;
6137 cpu = smp_processor_id();
6141 schedule_debug(prev, !!sched_mode);
6143 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6146 local_irq_disable();
6147 rcu_note_context_switch(!!sched_mode);
6150 * Make sure that signal_pending_state()->signal_pending() below
6151 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6152 * done by the caller to avoid the race with signal_wake_up():
6154 * __set_current_state(@state) signal_wake_up()
6155 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6156 * wake_up_state(p, state)
6157 * LOCK rq->lock LOCK p->pi_state
6158 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6159 * if (signal_pending_state()) if (p->state & @state)
6161 * Also, the membarrier system call requires a full memory barrier
6162 * after coming from user-space, before storing to rq->curr.
6165 smp_mb__after_spinlock();
6167 /* Promote REQ to ACT */
6168 rq->clock_update_flags <<= 1;
6169 update_rq_clock(rq);
6171 switch_count = &prev->nivcsw;
6174 * We must load prev->state once (task_struct::state is volatile), such
6177 * - we form a control dependency vs deactivate_task() below.
6178 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
6180 prev_state = READ_ONCE(prev->__state);
6181 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6182 if (signal_pending_state(prev_state, prev)) {
6183 WRITE_ONCE(prev->__state, TASK_RUNNING);
6185 prev->sched_contributes_to_load =
6186 (prev_state & TASK_UNINTERRUPTIBLE) &&
6187 !(prev_state & TASK_NOLOAD) &&
6188 !(prev->flags & PF_FROZEN);
6190 if (prev->sched_contributes_to_load)
6191 rq->nr_uninterruptible++;
6194 * __schedule() ttwu()
6195 * prev_state = prev->state; if (p->on_rq && ...)
6196 * if (prev_state) goto out;
6197 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6198 * p->state = TASK_WAKING
6200 * Where __schedule() and ttwu() have matching control dependencies.
6202 * After this, schedule() must not care about p->state any more.
6204 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6206 if (prev->in_iowait) {
6207 atomic_inc(&rq->nr_iowait);
6208 delayacct_blkio_start();
6211 switch_count = &prev->nvcsw;
6214 next = pick_next_task(rq, prev, &rf);
6215 clear_tsk_need_resched(prev);
6216 clear_preempt_need_resched();
6217 #ifdef CONFIG_SCHED_DEBUG
6218 rq->last_seen_need_resched_ns = 0;
6221 if (likely(prev != next)) {
6224 * RCU users of rcu_dereference(rq->curr) may not see
6225 * changes to task_struct made by pick_next_task().
6227 RCU_INIT_POINTER(rq->curr, next);
6229 * The membarrier system call requires each architecture
6230 * to have a full memory barrier after updating
6231 * rq->curr, before returning to user-space.
6233 * Here are the schemes providing that barrier on the
6234 * various architectures:
6235 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6236 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6237 * - finish_lock_switch() for weakly-ordered
6238 * architectures where spin_unlock is a full barrier,
6239 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6240 * is a RELEASE barrier),
6244 migrate_disable_switch(rq, prev);
6245 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6247 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next);
6249 /* Also unlocks the rq: */
6250 rq = context_switch(rq, prev, next, &rf);
6252 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6254 rq_unpin_lock(rq, &rf);
6255 __balance_callbacks(rq);
6256 raw_spin_rq_unlock_irq(rq);
6260 void __noreturn do_task_dead(void)
6262 /* Causes final put_task_struct in finish_task_switch(): */
6263 set_special_state(TASK_DEAD);
6265 /* Tell freezer to ignore us: */
6266 current->flags |= PF_NOFREEZE;
6268 __schedule(SM_NONE);
6271 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6276 static inline void sched_submit_work(struct task_struct *tsk)
6278 unsigned int task_flags;
6280 if (task_is_running(tsk))
6283 task_flags = tsk->flags;
6285 * If a worker goes to sleep, notify and ask workqueue whether it
6286 * wants to wake up a task to maintain concurrency.
6288 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6289 if (task_flags & PF_WQ_WORKER)
6290 wq_worker_sleeping(tsk);
6292 io_wq_worker_sleeping(tsk);
6295 if (tsk_is_pi_blocked(tsk))
6299 * If we are going to sleep and we have plugged IO queued,
6300 * make sure to submit it to avoid deadlocks.
6302 if (blk_needs_flush_plug(tsk))
6303 blk_flush_plug(tsk->plug, true);
6306 static void sched_update_worker(struct task_struct *tsk)
6308 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6309 if (tsk->flags & PF_WQ_WORKER)
6310 wq_worker_running(tsk);
6312 io_wq_worker_running(tsk);
6316 asmlinkage __visible void __sched schedule(void)
6318 struct task_struct *tsk = current;
6320 sched_submit_work(tsk);
6323 __schedule(SM_NONE);
6324 sched_preempt_enable_no_resched();
6325 } while (need_resched());
6326 sched_update_worker(tsk);
6328 EXPORT_SYMBOL(schedule);
6331 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6332 * state (have scheduled out non-voluntarily) by making sure that all
6333 * tasks have either left the run queue or have gone into user space.
6334 * As idle tasks do not do either, they must not ever be preempted
6335 * (schedule out non-voluntarily).
6337 * schedule_idle() is similar to schedule_preempt_disable() except that it
6338 * never enables preemption because it does not call sched_submit_work().
6340 void __sched schedule_idle(void)
6343 * As this skips calling sched_submit_work(), which the idle task does
6344 * regardless because that function is a nop when the task is in a
6345 * TASK_RUNNING state, make sure this isn't used someplace that the
6346 * current task can be in any other state. Note, idle is always in the
6347 * TASK_RUNNING state.
6349 WARN_ON_ONCE(current->__state);
6351 __schedule(SM_NONE);
6352 } while (need_resched());
6355 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6356 asmlinkage __visible void __sched schedule_user(void)
6359 * If we come here after a random call to set_need_resched(),
6360 * or we have been woken up remotely but the IPI has not yet arrived,
6361 * we haven't yet exited the RCU idle mode. Do it here manually until
6362 * we find a better solution.
6364 * NB: There are buggy callers of this function. Ideally we
6365 * should warn if prev_state != CONTEXT_USER, but that will trigger
6366 * too frequently to make sense yet.
6368 enum ctx_state prev_state = exception_enter();
6370 exception_exit(prev_state);
6375 * schedule_preempt_disabled - called with preemption disabled
6377 * Returns with preemption disabled. Note: preempt_count must be 1
6379 void __sched schedule_preempt_disabled(void)
6381 sched_preempt_enable_no_resched();
6386 #ifdef CONFIG_PREEMPT_RT
6387 void __sched notrace schedule_rtlock(void)
6391 __schedule(SM_RTLOCK_WAIT);
6392 sched_preempt_enable_no_resched();
6393 } while (need_resched());
6395 NOKPROBE_SYMBOL(schedule_rtlock);
6398 static void __sched notrace preempt_schedule_common(void)
6402 * Because the function tracer can trace preempt_count_sub()
6403 * and it also uses preempt_enable/disable_notrace(), if
6404 * NEED_RESCHED is set, the preempt_enable_notrace() called
6405 * by the function tracer will call this function again and
6406 * cause infinite recursion.
6408 * Preemption must be disabled here before the function
6409 * tracer can trace. Break up preempt_disable() into two
6410 * calls. One to disable preemption without fear of being
6411 * traced. The other to still record the preemption latency,
6412 * which can also be traced by the function tracer.
6414 preempt_disable_notrace();
6415 preempt_latency_start(1);
6416 __schedule(SM_PREEMPT);
6417 preempt_latency_stop(1);
6418 preempt_enable_no_resched_notrace();
6421 * Check again in case we missed a preemption opportunity
6422 * between schedule and now.
6424 } while (need_resched());
6427 #ifdef CONFIG_PREEMPTION
6429 * This is the entry point to schedule() from in-kernel preemption
6430 * off of preempt_enable.
6432 asmlinkage __visible void __sched notrace preempt_schedule(void)
6435 * If there is a non-zero preempt_count or interrupts are disabled,
6436 * we do not want to preempt the current task. Just return..
6438 if (likely(!preemptible()))
6441 preempt_schedule_common();
6443 NOKPROBE_SYMBOL(preempt_schedule);
6444 EXPORT_SYMBOL(preempt_schedule);
6446 #ifdef CONFIG_PREEMPT_DYNAMIC
6447 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6448 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6453 * preempt_schedule_notrace - preempt_schedule called by tracing
6455 * The tracing infrastructure uses preempt_enable_notrace to prevent
6456 * recursion and tracing preempt enabling caused by the tracing
6457 * infrastructure itself. But as tracing can happen in areas coming
6458 * from userspace or just about to enter userspace, a preempt enable
6459 * can occur before user_exit() is called. This will cause the scheduler
6460 * to be called when the system is still in usermode.
6462 * To prevent this, the preempt_enable_notrace will use this function
6463 * instead of preempt_schedule() to exit user context if needed before
6464 * calling the scheduler.
6466 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6468 enum ctx_state prev_ctx;
6470 if (likely(!preemptible()))
6475 * Because the function tracer can trace preempt_count_sub()
6476 * and it also uses preempt_enable/disable_notrace(), if
6477 * NEED_RESCHED is set, the preempt_enable_notrace() called
6478 * by the function tracer will call this function again and
6479 * cause infinite recursion.
6481 * Preemption must be disabled here before the function
6482 * tracer can trace. Break up preempt_disable() into two
6483 * calls. One to disable preemption without fear of being
6484 * traced. The other to still record the preemption latency,
6485 * which can also be traced by the function tracer.
6487 preempt_disable_notrace();
6488 preempt_latency_start(1);
6490 * Needs preempt disabled in case user_exit() is traced
6491 * and the tracer calls preempt_enable_notrace() causing
6492 * an infinite recursion.
6494 prev_ctx = exception_enter();
6495 __schedule(SM_PREEMPT);
6496 exception_exit(prev_ctx);
6498 preempt_latency_stop(1);
6499 preempt_enable_no_resched_notrace();
6500 } while (need_resched());
6502 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6504 #ifdef CONFIG_PREEMPT_DYNAMIC
6505 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6506 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6509 #endif /* CONFIG_PREEMPTION */
6511 #ifdef CONFIG_PREEMPT_DYNAMIC
6513 #include <linux/entry-common.h>
6518 * SC:preempt_schedule
6519 * SC:preempt_schedule_notrace
6520 * SC:irqentry_exit_cond_resched
6524 * cond_resched <- __cond_resched
6525 * might_resched <- RET0
6526 * preempt_schedule <- NOP
6527 * preempt_schedule_notrace <- NOP
6528 * irqentry_exit_cond_resched <- NOP
6531 * cond_resched <- __cond_resched
6532 * might_resched <- __cond_resched
6533 * preempt_schedule <- NOP
6534 * preempt_schedule_notrace <- NOP
6535 * irqentry_exit_cond_resched <- NOP
6538 * cond_resched <- RET0
6539 * might_resched <- RET0
6540 * preempt_schedule <- preempt_schedule
6541 * preempt_schedule_notrace <- preempt_schedule_notrace
6542 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6546 preempt_dynamic_undefined = -1,
6547 preempt_dynamic_none,
6548 preempt_dynamic_voluntary,
6549 preempt_dynamic_full,
6552 int preempt_dynamic_mode = preempt_dynamic_undefined;
6554 int sched_dynamic_mode(const char *str)
6556 if (!strcmp(str, "none"))
6557 return preempt_dynamic_none;
6559 if (!strcmp(str, "voluntary"))
6560 return preempt_dynamic_voluntary;
6562 if (!strcmp(str, "full"))
6563 return preempt_dynamic_full;
6568 void sched_dynamic_update(int mode)
6571 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6572 * the ZERO state, which is invalid.
6574 static_call_update(cond_resched, __cond_resched);
6575 static_call_update(might_resched, __cond_resched);
6576 static_call_update(preempt_schedule, __preempt_schedule_func);
6577 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6578 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6581 case preempt_dynamic_none:
6582 static_call_update(cond_resched, __cond_resched);
6583 static_call_update(might_resched, (void *)&__static_call_return0);
6584 static_call_update(preempt_schedule, NULL);
6585 static_call_update(preempt_schedule_notrace, NULL);
6586 static_call_update(irqentry_exit_cond_resched, NULL);
6587 pr_info("Dynamic Preempt: none\n");
6590 case preempt_dynamic_voluntary:
6591 static_call_update(cond_resched, __cond_resched);
6592 static_call_update(might_resched, __cond_resched);
6593 static_call_update(preempt_schedule, NULL);
6594 static_call_update(preempt_schedule_notrace, NULL);
6595 static_call_update(irqentry_exit_cond_resched, NULL);
6596 pr_info("Dynamic Preempt: voluntary\n");
6599 case preempt_dynamic_full:
6600 static_call_update(cond_resched, (void *)&__static_call_return0);
6601 static_call_update(might_resched, (void *)&__static_call_return0);
6602 static_call_update(preempt_schedule, __preempt_schedule_func);
6603 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6604 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6605 pr_info("Dynamic Preempt: full\n");
6609 preempt_dynamic_mode = mode;
6612 static int __init setup_preempt_mode(char *str)
6614 int mode = sched_dynamic_mode(str);
6616 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6620 sched_dynamic_update(mode);
6623 __setup("preempt=", setup_preempt_mode);
6625 static void __init preempt_dynamic_init(void)
6627 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
6628 if (IS_ENABLED(CONFIG_PREEMPT_NONE_BEHAVIOUR)) {
6629 sched_dynamic_update(preempt_dynamic_none);
6630 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY_BEHAVIOUR)) {
6631 sched_dynamic_update(preempt_dynamic_voluntary);
6633 /* Default static call setting, nothing to do */
6634 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_BEHAVIOUR));
6635 preempt_dynamic_mode = preempt_dynamic_full;
6636 pr_info("Dynamic Preempt: full\n");
6641 #else /* !CONFIG_PREEMPT_DYNAMIC */
6643 static inline void preempt_dynamic_init(void) { }
6645 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
6648 * This is the entry point to schedule() from kernel preemption
6649 * off of irq context.
6650 * Note, that this is called and return with irqs disabled. This will
6651 * protect us against recursive calling from irq.
6653 asmlinkage __visible void __sched preempt_schedule_irq(void)
6655 enum ctx_state prev_state;
6657 /* Catch callers which need to be fixed */
6658 BUG_ON(preempt_count() || !irqs_disabled());
6660 prev_state = exception_enter();
6665 __schedule(SM_PREEMPT);
6666 local_irq_disable();
6667 sched_preempt_enable_no_resched();
6668 } while (need_resched());
6670 exception_exit(prev_state);
6673 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6676 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6677 return try_to_wake_up(curr->private, mode, wake_flags);
6679 EXPORT_SYMBOL(default_wake_function);
6681 static void __setscheduler_prio(struct task_struct *p, int prio)
6684 p->sched_class = &dl_sched_class;
6685 else if (rt_prio(prio))
6686 p->sched_class = &rt_sched_class;
6688 p->sched_class = &fair_sched_class;
6693 #ifdef CONFIG_RT_MUTEXES
6695 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6698 prio = min(prio, pi_task->prio);
6703 static inline int rt_effective_prio(struct task_struct *p, int prio)
6705 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6707 return __rt_effective_prio(pi_task, prio);
6711 * rt_mutex_setprio - set the current priority of a task
6713 * @pi_task: donor task
6715 * This function changes the 'effective' priority of a task. It does
6716 * not touch ->normal_prio like __setscheduler().
6718 * Used by the rt_mutex code to implement priority inheritance
6719 * logic. Call site only calls if the priority of the task changed.
6721 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6723 int prio, oldprio, queued, running, queue_flag =
6724 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6725 const struct sched_class *prev_class;
6729 /* XXX used to be waiter->prio, not waiter->task->prio */
6730 prio = __rt_effective_prio(pi_task, p->normal_prio);
6733 * If nothing changed; bail early.
6735 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6738 rq = __task_rq_lock(p, &rf);
6739 update_rq_clock(rq);
6741 * Set under pi_lock && rq->lock, such that the value can be used under
6744 * Note that there is loads of tricky to make this pointer cache work
6745 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6746 * ensure a task is de-boosted (pi_task is set to NULL) before the
6747 * task is allowed to run again (and can exit). This ensures the pointer
6748 * points to a blocked task -- which guarantees the task is present.
6750 p->pi_top_task = pi_task;
6753 * For FIFO/RR we only need to set prio, if that matches we're done.
6755 if (prio == p->prio && !dl_prio(prio))
6759 * Idle task boosting is a nono in general. There is one
6760 * exception, when PREEMPT_RT and NOHZ is active:
6762 * The idle task calls get_next_timer_interrupt() and holds
6763 * the timer wheel base->lock on the CPU and another CPU wants
6764 * to access the timer (probably to cancel it). We can safely
6765 * ignore the boosting request, as the idle CPU runs this code
6766 * with interrupts disabled and will complete the lock
6767 * protected section without being interrupted. So there is no
6768 * real need to boost.
6770 if (unlikely(p == rq->idle)) {
6771 WARN_ON(p != rq->curr);
6772 WARN_ON(p->pi_blocked_on);
6776 trace_sched_pi_setprio(p, pi_task);
6779 if (oldprio == prio)
6780 queue_flag &= ~DEQUEUE_MOVE;
6782 prev_class = p->sched_class;
6783 queued = task_on_rq_queued(p);
6784 running = task_current(rq, p);
6786 dequeue_task(rq, p, queue_flag);
6788 put_prev_task(rq, p);
6791 * Boosting condition are:
6792 * 1. -rt task is running and holds mutex A
6793 * --> -dl task blocks on mutex A
6795 * 2. -dl task is running and holds mutex A
6796 * --> -dl task blocks on mutex A and could preempt the
6799 if (dl_prio(prio)) {
6800 if (!dl_prio(p->normal_prio) ||
6801 (pi_task && dl_prio(pi_task->prio) &&
6802 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6803 p->dl.pi_se = pi_task->dl.pi_se;
6804 queue_flag |= ENQUEUE_REPLENISH;
6806 p->dl.pi_se = &p->dl;
6808 } else if (rt_prio(prio)) {
6809 if (dl_prio(oldprio))
6810 p->dl.pi_se = &p->dl;
6812 queue_flag |= ENQUEUE_HEAD;
6814 if (dl_prio(oldprio))
6815 p->dl.pi_se = &p->dl;
6816 if (rt_prio(oldprio))
6820 __setscheduler_prio(p, prio);
6823 enqueue_task(rq, p, queue_flag);
6825 set_next_task(rq, p);
6827 check_class_changed(rq, p, prev_class, oldprio);
6829 /* Avoid rq from going away on us: */
6832 rq_unpin_lock(rq, &rf);
6833 __balance_callbacks(rq);
6834 raw_spin_rq_unlock(rq);
6839 static inline int rt_effective_prio(struct task_struct *p, int prio)
6845 void set_user_nice(struct task_struct *p, long nice)
6847 bool queued, running;
6852 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6855 * We have to be careful, if called from sys_setpriority(),
6856 * the task might be in the middle of scheduling on another CPU.
6858 rq = task_rq_lock(p, &rf);
6859 update_rq_clock(rq);
6862 * The RT priorities are set via sched_setscheduler(), but we still
6863 * allow the 'normal' nice value to be set - but as expected
6864 * it won't have any effect on scheduling until the task is
6865 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6867 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6868 p->static_prio = NICE_TO_PRIO(nice);
6871 queued = task_on_rq_queued(p);
6872 running = task_current(rq, p);
6874 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6876 put_prev_task(rq, p);
6878 p->static_prio = NICE_TO_PRIO(nice);
6879 set_load_weight(p, true);
6881 p->prio = effective_prio(p);
6884 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6886 set_next_task(rq, p);
6889 * If the task increased its priority or is running and
6890 * lowered its priority, then reschedule its CPU:
6892 p->sched_class->prio_changed(rq, p, old_prio);
6895 task_rq_unlock(rq, p, &rf);
6897 EXPORT_SYMBOL(set_user_nice);
6900 * can_nice - check if a task can reduce its nice value
6904 int can_nice(const struct task_struct *p, const int nice)
6906 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6907 int nice_rlim = nice_to_rlimit(nice);
6909 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6910 capable(CAP_SYS_NICE));
6913 #ifdef __ARCH_WANT_SYS_NICE
6916 * sys_nice - change the priority of the current process.
6917 * @increment: priority increment
6919 * sys_setpriority is a more generic, but much slower function that
6920 * does similar things.
6922 SYSCALL_DEFINE1(nice, int, increment)
6927 * Setpriority might change our priority at the same moment.
6928 * We don't have to worry. Conceptually one call occurs first
6929 * and we have a single winner.
6931 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6932 nice = task_nice(current) + increment;
6934 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6935 if (increment < 0 && !can_nice(current, nice))
6938 retval = security_task_setnice(current, nice);
6942 set_user_nice(current, nice);
6949 * task_prio - return the priority value of a given task.
6950 * @p: the task in question.
6952 * Return: The priority value as seen by users in /proc.
6954 * sched policy return value kernel prio user prio/nice
6956 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6957 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6958 * deadline -101 -1 0
6960 int task_prio(const struct task_struct *p)
6962 return p->prio - MAX_RT_PRIO;
6966 * idle_cpu - is a given CPU idle currently?
6967 * @cpu: the processor in question.
6969 * Return: 1 if the CPU is currently idle. 0 otherwise.
6971 int idle_cpu(int cpu)
6973 struct rq *rq = cpu_rq(cpu);
6975 if (rq->curr != rq->idle)
6982 if (rq->ttwu_pending)
6990 * available_idle_cpu - is a given CPU idle for enqueuing work.
6991 * @cpu: the CPU in question.
6993 * Return: 1 if the CPU is currently idle. 0 otherwise.
6995 int available_idle_cpu(int cpu)
7000 if (vcpu_is_preempted(cpu))
7007 * idle_task - return the idle task for a given CPU.
7008 * @cpu: the processor in question.
7010 * Return: The idle task for the CPU @cpu.
7012 struct task_struct *idle_task(int cpu)
7014 return cpu_rq(cpu)->idle;
7019 * This function computes an effective utilization for the given CPU, to be
7020 * used for frequency selection given the linear relation: f = u * f_max.
7022 * The scheduler tracks the following metrics:
7024 * cpu_util_{cfs,rt,dl,irq}()
7027 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7028 * synchronized windows and are thus directly comparable.
7030 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7031 * which excludes things like IRQ and steal-time. These latter are then accrued
7032 * in the irq utilization.
7034 * The DL bandwidth number otoh is not a measured metric but a value computed
7035 * based on the task model parameters and gives the minimal utilization
7036 * required to meet deadlines.
7038 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7039 unsigned long max, enum cpu_util_type type,
7040 struct task_struct *p)
7042 unsigned long dl_util, util, irq;
7043 struct rq *rq = cpu_rq(cpu);
7045 if (!uclamp_is_used() &&
7046 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7051 * Early check to see if IRQ/steal time saturates the CPU, can be
7052 * because of inaccuracies in how we track these -- see
7053 * update_irq_load_avg().
7055 irq = cpu_util_irq(rq);
7056 if (unlikely(irq >= max))
7060 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7061 * CFS tasks and we use the same metric to track the effective
7062 * utilization (PELT windows are synchronized) we can directly add them
7063 * to obtain the CPU's actual utilization.
7065 * CFS and RT utilization can be boosted or capped, depending on
7066 * utilization clamp constraints requested by currently RUNNABLE
7068 * When there are no CFS RUNNABLE tasks, clamps are released and
7069 * frequency will be gracefully reduced with the utilization decay.
7071 util = util_cfs + cpu_util_rt(rq);
7072 if (type == FREQUENCY_UTIL)
7073 util = uclamp_rq_util_with(rq, util, p);
7075 dl_util = cpu_util_dl(rq);
7078 * For frequency selection we do not make cpu_util_dl() a permanent part
7079 * of this sum because we want to use cpu_bw_dl() later on, but we need
7080 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7081 * that we select f_max when there is no idle time.
7083 * NOTE: numerical errors or stop class might cause us to not quite hit
7084 * saturation when we should -- something for later.
7086 if (util + dl_util >= max)
7090 * OTOH, for energy computation we need the estimated running time, so
7091 * include util_dl and ignore dl_bw.
7093 if (type == ENERGY_UTIL)
7097 * There is still idle time; further improve the number by using the
7098 * irq metric. Because IRQ/steal time is hidden from the task clock we
7099 * need to scale the task numbers:
7102 * U' = irq + --------- * U
7105 util = scale_irq_capacity(util, irq, max);
7109 * Bandwidth required by DEADLINE must always be granted while, for
7110 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7111 * to gracefully reduce the frequency when no tasks show up for longer
7114 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7115 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7116 * an interface. So, we only do the latter for now.
7118 if (type == FREQUENCY_UTIL)
7119 util += cpu_bw_dl(rq);
7121 return min(max, util);
7124 unsigned long sched_cpu_util(int cpu, unsigned long max)
7126 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
7129 #endif /* CONFIG_SMP */
7132 * find_process_by_pid - find a process with a matching PID value.
7133 * @pid: the pid in question.
7135 * The task of @pid, if found. %NULL otherwise.
7137 static struct task_struct *find_process_by_pid(pid_t pid)
7139 return pid ? find_task_by_vpid(pid) : current;
7143 * sched_setparam() passes in -1 for its policy, to let the functions
7144 * it calls know not to change it.
7146 #define SETPARAM_POLICY -1
7148 static void __setscheduler_params(struct task_struct *p,
7149 const struct sched_attr *attr)
7151 int policy = attr->sched_policy;
7153 if (policy == SETPARAM_POLICY)
7158 if (dl_policy(policy))
7159 __setparam_dl(p, attr);
7160 else if (fair_policy(policy))
7161 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7164 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7165 * !rt_policy. Always setting this ensures that things like
7166 * getparam()/getattr() don't report silly values for !rt tasks.
7168 p->rt_priority = attr->sched_priority;
7169 p->normal_prio = normal_prio(p);
7170 set_load_weight(p, true);
7174 * Check the target process has a UID that matches the current process's:
7176 static bool check_same_owner(struct task_struct *p)
7178 const struct cred *cred = current_cred(), *pcred;
7182 pcred = __task_cred(p);
7183 match = (uid_eq(cred->euid, pcred->euid) ||
7184 uid_eq(cred->euid, pcred->uid));
7189 static int __sched_setscheduler(struct task_struct *p,
7190 const struct sched_attr *attr,
7193 int oldpolicy = -1, policy = attr->sched_policy;
7194 int retval, oldprio, newprio, queued, running;
7195 const struct sched_class *prev_class;
7196 struct callback_head *head;
7199 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7202 /* The pi code expects interrupts enabled */
7203 BUG_ON(pi && in_interrupt());
7205 /* Double check policy once rq lock held: */
7207 reset_on_fork = p->sched_reset_on_fork;
7208 policy = oldpolicy = p->policy;
7210 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7212 if (!valid_policy(policy))
7216 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7220 * Valid priorities for SCHED_FIFO and SCHED_RR are
7221 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7222 * SCHED_BATCH and SCHED_IDLE is 0.
7224 if (attr->sched_priority > MAX_RT_PRIO-1)
7226 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7227 (rt_policy(policy) != (attr->sched_priority != 0)))
7231 * Allow unprivileged RT tasks to decrease priority:
7233 if (user && !capable(CAP_SYS_NICE)) {
7234 if (fair_policy(policy)) {
7235 if (attr->sched_nice < task_nice(p) &&
7236 !can_nice(p, attr->sched_nice))
7240 if (rt_policy(policy)) {
7241 unsigned long rlim_rtprio =
7242 task_rlimit(p, RLIMIT_RTPRIO);
7244 /* Can't set/change the rt policy: */
7245 if (policy != p->policy && !rlim_rtprio)
7248 /* Can't increase priority: */
7249 if (attr->sched_priority > p->rt_priority &&
7250 attr->sched_priority > rlim_rtprio)
7255 * Can't set/change SCHED_DEADLINE policy at all for now
7256 * (safest behavior); in the future we would like to allow
7257 * unprivileged DL tasks to increase their relative deadline
7258 * or reduce their runtime (both ways reducing utilization)
7260 if (dl_policy(policy))
7264 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7265 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7267 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7268 if (!can_nice(p, task_nice(p)))
7272 /* Can't change other user's priorities: */
7273 if (!check_same_owner(p))
7276 /* Normal users shall not reset the sched_reset_on_fork flag: */
7277 if (p->sched_reset_on_fork && !reset_on_fork)
7282 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7285 retval = security_task_setscheduler(p);
7290 /* Update task specific "requested" clamps */
7291 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7292 retval = uclamp_validate(p, attr);
7301 * Make sure no PI-waiters arrive (or leave) while we are
7302 * changing the priority of the task:
7304 * To be able to change p->policy safely, the appropriate
7305 * runqueue lock must be held.
7307 rq = task_rq_lock(p, &rf);
7308 update_rq_clock(rq);
7311 * Changing the policy of the stop threads its a very bad idea:
7313 if (p == rq->stop) {
7319 * If not changing anything there's no need to proceed further,
7320 * but store a possible modification of reset_on_fork.
7322 if (unlikely(policy == p->policy)) {
7323 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7325 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7327 if (dl_policy(policy) && dl_param_changed(p, attr))
7329 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7332 p->sched_reset_on_fork = reset_on_fork;
7339 #ifdef CONFIG_RT_GROUP_SCHED
7341 * Do not allow realtime tasks into groups that have no runtime
7344 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7345 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7346 !task_group_is_autogroup(task_group(p))) {
7352 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7353 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7354 cpumask_t *span = rq->rd->span;
7357 * Don't allow tasks with an affinity mask smaller than
7358 * the entire root_domain to become SCHED_DEADLINE. We
7359 * will also fail if there's no bandwidth available.
7361 if (!cpumask_subset(span, p->cpus_ptr) ||
7362 rq->rd->dl_bw.bw == 0) {
7370 /* Re-check policy now with rq lock held: */
7371 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7372 policy = oldpolicy = -1;
7373 task_rq_unlock(rq, p, &rf);
7375 cpuset_read_unlock();
7380 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7381 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7384 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7389 p->sched_reset_on_fork = reset_on_fork;
7392 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7395 * Take priority boosted tasks into account. If the new
7396 * effective priority is unchanged, we just store the new
7397 * normal parameters and do not touch the scheduler class and
7398 * the runqueue. This will be done when the task deboost
7401 newprio = rt_effective_prio(p, newprio);
7402 if (newprio == oldprio)
7403 queue_flags &= ~DEQUEUE_MOVE;
7406 queued = task_on_rq_queued(p);
7407 running = task_current(rq, p);
7409 dequeue_task(rq, p, queue_flags);
7411 put_prev_task(rq, p);
7413 prev_class = p->sched_class;
7415 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7416 __setscheduler_params(p, attr);
7417 __setscheduler_prio(p, newprio);
7419 __setscheduler_uclamp(p, attr);
7423 * We enqueue to tail when the priority of a task is
7424 * increased (user space view).
7426 if (oldprio < p->prio)
7427 queue_flags |= ENQUEUE_HEAD;
7429 enqueue_task(rq, p, queue_flags);
7432 set_next_task(rq, p);
7434 check_class_changed(rq, p, prev_class, oldprio);
7436 /* Avoid rq from going away on us: */
7438 head = splice_balance_callbacks(rq);
7439 task_rq_unlock(rq, p, &rf);
7442 cpuset_read_unlock();
7443 rt_mutex_adjust_pi(p);
7446 /* Run balance callbacks after we've adjusted the PI chain: */
7447 balance_callbacks(rq, head);
7453 task_rq_unlock(rq, p, &rf);
7455 cpuset_read_unlock();
7459 static int _sched_setscheduler(struct task_struct *p, int policy,
7460 const struct sched_param *param, bool check)
7462 struct sched_attr attr = {
7463 .sched_policy = policy,
7464 .sched_priority = param->sched_priority,
7465 .sched_nice = PRIO_TO_NICE(p->static_prio),
7468 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7469 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7470 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7471 policy &= ~SCHED_RESET_ON_FORK;
7472 attr.sched_policy = policy;
7475 return __sched_setscheduler(p, &attr, check, true);
7478 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7479 * @p: the task in question.
7480 * @policy: new policy.
7481 * @param: structure containing the new RT priority.
7483 * Use sched_set_fifo(), read its comment.
7485 * Return: 0 on success. An error code otherwise.
7487 * NOTE that the task may be already dead.
7489 int sched_setscheduler(struct task_struct *p, int policy,
7490 const struct sched_param *param)
7492 return _sched_setscheduler(p, policy, param, true);
7495 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7497 return __sched_setscheduler(p, attr, true, true);
7500 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7502 return __sched_setscheduler(p, attr, false, true);
7504 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7507 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7508 * @p: the task in question.
7509 * @policy: new policy.
7510 * @param: structure containing the new RT priority.
7512 * Just like sched_setscheduler, only don't bother checking if the
7513 * current context has permission. For example, this is needed in
7514 * stop_machine(): we create temporary high priority worker threads,
7515 * but our caller might not have that capability.
7517 * Return: 0 on success. An error code otherwise.
7519 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7520 const struct sched_param *param)
7522 return _sched_setscheduler(p, policy, param, false);
7526 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7527 * incapable of resource management, which is the one thing an OS really should
7530 * This is of course the reason it is limited to privileged users only.
7532 * Worse still; it is fundamentally impossible to compose static priority
7533 * workloads. You cannot take two correctly working static prio workloads
7534 * and smash them together and still expect them to work.
7536 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7540 * The administrator _MUST_ configure the system, the kernel simply doesn't
7541 * know enough information to make a sensible choice.
7543 void sched_set_fifo(struct task_struct *p)
7545 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7546 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7548 EXPORT_SYMBOL_GPL(sched_set_fifo);
7551 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7553 void sched_set_fifo_low(struct task_struct *p)
7555 struct sched_param sp = { .sched_priority = 1 };
7556 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7558 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7560 void sched_set_normal(struct task_struct *p, int nice)
7562 struct sched_attr attr = {
7563 .sched_policy = SCHED_NORMAL,
7566 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7568 EXPORT_SYMBOL_GPL(sched_set_normal);
7571 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7573 struct sched_param lparam;
7574 struct task_struct *p;
7577 if (!param || pid < 0)
7579 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7584 p = find_process_by_pid(pid);
7590 retval = sched_setscheduler(p, policy, &lparam);
7598 * Mimics kernel/events/core.c perf_copy_attr().
7600 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7605 /* Zero the full structure, so that a short copy will be nice: */
7606 memset(attr, 0, sizeof(*attr));
7608 ret = get_user(size, &uattr->size);
7612 /* ABI compatibility quirk: */
7614 size = SCHED_ATTR_SIZE_VER0;
7615 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7618 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7625 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7626 size < SCHED_ATTR_SIZE_VER1)
7630 * XXX: Do we want to be lenient like existing syscalls; or do we want
7631 * to be strict and return an error on out-of-bounds values?
7633 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7638 put_user(sizeof(*attr), &uattr->size);
7642 static void get_params(struct task_struct *p, struct sched_attr *attr)
7644 if (task_has_dl_policy(p))
7645 __getparam_dl(p, attr);
7646 else if (task_has_rt_policy(p))
7647 attr->sched_priority = p->rt_priority;
7649 attr->sched_nice = task_nice(p);
7653 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7654 * @pid: the pid in question.
7655 * @policy: new policy.
7656 * @param: structure containing the new RT priority.
7658 * Return: 0 on success. An error code otherwise.
7660 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7665 return do_sched_setscheduler(pid, policy, param);
7669 * sys_sched_setparam - set/change the RT priority of a thread
7670 * @pid: the pid in question.
7671 * @param: structure containing the new RT priority.
7673 * Return: 0 on success. An error code otherwise.
7675 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7677 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7681 * sys_sched_setattr - same as above, but with extended sched_attr
7682 * @pid: the pid in question.
7683 * @uattr: structure containing the extended parameters.
7684 * @flags: for future extension.
7686 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7687 unsigned int, flags)
7689 struct sched_attr attr;
7690 struct task_struct *p;
7693 if (!uattr || pid < 0 || flags)
7696 retval = sched_copy_attr(uattr, &attr);
7700 if ((int)attr.sched_policy < 0)
7702 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7703 attr.sched_policy = SETPARAM_POLICY;
7707 p = find_process_by_pid(pid);
7713 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7714 get_params(p, &attr);
7715 retval = sched_setattr(p, &attr);
7723 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7724 * @pid: the pid in question.
7726 * Return: On success, the policy of the thread. Otherwise, a negative error
7729 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7731 struct task_struct *p;
7739 p = find_process_by_pid(pid);
7741 retval = security_task_getscheduler(p);
7744 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7751 * sys_sched_getparam - get the RT priority of a thread
7752 * @pid: the pid in question.
7753 * @param: structure containing the RT priority.
7755 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7758 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7760 struct sched_param lp = { .sched_priority = 0 };
7761 struct task_struct *p;
7764 if (!param || pid < 0)
7768 p = find_process_by_pid(pid);
7773 retval = security_task_getscheduler(p);
7777 if (task_has_rt_policy(p))
7778 lp.sched_priority = p->rt_priority;
7782 * This one might sleep, we cannot do it with a spinlock held ...
7784 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7794 * Copy the kernel size attribute structure (which might be larger
7795 * than what user-space knows about) to user-space.
7797 * Note that all cases are valid: user-space buffer can be larger or
7798 * smaller than the kernel-space buffer. The usual case is that both
7799 * have the same size.
7802 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7803 struct sched_attr *kattr,
7806 unsigned int ksize = sizeof(*kattr);
7808 if (!access_ok(uattr, usize))
7812 * sched_getattr() ABI forwards and backwards compatibility:
7814 * If usize == ksize then we just copy everything to user-space and all is good.
7816 * If usize < ksize then we only copy as much as user-space has space for,
7817 * this keeps ABI compatibility as well. We skip the rest.
7819 * If usize > ksize then user-space is using a newer version of the ABI,
7820 * which part the kernel doesn't know about. Just ignore it - tooling can
7821 * detect the kernel's knowledge of attributes from the attr->size value
7822 * which is set to ksize in this case.
7824 kattr->size = min(usize, ksize);
7826 if (copy_to_user(uattr, kattr, kattr->size))
7833 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7834 * @pid: the pid in question.
7835 * @uattr: structure containing the extended parameters.
7836 * @usize: sizeof(attr) for fwd/bwd comp.
7837 * @flags: for future extension.
7839 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7840 unsigned int, usize, unsigned int, flags)
7842 struct sched_attr kattr = { };
7843 struct task_struct *p;
7846 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7847 usize < SCHED_ATTR_SIZE_VER0 || flags)
7851 p = find_process_by_pid(pid);
7856 retval = security_task_getscheduler(p);
7860 kattr.sched_policy = p->policy;
7861 if (p->sched_reset_on_fork)
7862 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7863 get_params(p, &kattr);
7864 kattr.sched_flags &= SCHED_FLAG_ALL;
7866 #ifdef CONFIG_UCLAMP_TASK
7868 * This could race with another potential updater, but this is fine
7869 * because it'll correctly read the old or the new value. We don't need
7870 * to guarantee who wins the race as long as it doesn't return garbage.
7872 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7873 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7878 return sched_attr_copy_to_user(uattr, &kattr, usize);
7886 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7891 * If the task isn't a deadline task or admission control is
7892 * disabled then we don't care about affinity changes.
7894 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7898 * Since bandwidth control happens on root_domain basis,
7899 * if admission test is enabled, we only admit -deadline
7900 * tasks allowed to run on all the CPUs in the task's
7904 if (!cpumask_subset(task_rq(p)->rd->span, mask))
7912 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
7915 cpumask_var_t cpus_allowed, new_mask;
7917 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
7920 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7922 goto out_free_cpus_allowed;
7925 cpuset_cpus_allowed(p, cpus_allowed);
7926 cpumask_and(new_mask, mask, cpus_allowed);
7928 retval = dl_task_check_affinity(p, new_mask);
7930 goto out_free_new_mask;
7932 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
7934 goto out_free_new_mask;
7936 cpuset_cpus_allowed(p, cpus_allowed);
7937 if (!cpumask_subset(new_mask, cpus_allowed)) {
7939 * We must have raced with a concurrent cpuset update.
7940 * Just reset the cpumask to the cpuset's cpus_allowed.
7942 cpumask_copy(new_mask, cpus_allowed);
7947 free_cpumask_var(new_mask);
7948 out_free_cpus_allowed:
7949 free_cpumask_var(cpus_allowed);
7953 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7955 struct task_struct *p;
7960 p = find_process_by_pid(pid);
7966 /* Prevent p going away */
7970 if (p->flags & PF_NO_SETAFFINITY) {
7975 if (!check_same_owner(p)) {
7977 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7985 retval = security_task_setscheduler(p);
7989 retval = __sched_setaffinity(p, in_mask);
7995 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7996 struct cpumask *new_mask)
7998 if (len < cpumask_size())
7999 cpumask_clear(new_mask);
8000 else if (len > cpumask_size())
8001 len = cpumask_size();
8003 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8007 * sys_sched_setaffinity - set the CPU affinity of a process
8008 * @pid: pid of the process
8009 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8010 * @user_mask_ptr: user-space pointer to the new CPU mask
8012 * Return: 0 on success. An error code otherwise.
8014 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8015 unsigned long __user *, user_mask_ptr)
8017 cpumask_var_t new_mask;
8020 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8023 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8025 retval = sched_setaffinity(pid, new_mask);
8026 free_cpumask_var(new_mask);
8030 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8032 struct task_struct *p;
8033 unsigned long flags;
8039 p = find_process_by_pid(pid);
8043 retval = security_task_getscheduler(p);
8047 raw_spin_lock_irqsave(&p->pi_lock, flags);
8048 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8049 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8058 * sys_sched_getaffinity - get the CPU affinity of a process
8059 * @pid: pid of the process
8060 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8061 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8063 * Return: size of CPU mask copied to user_mask_ptr on success. An
8064 * error code otherwise.
8066 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8067 unsigned long __user *, user_mask_ptr)
8072 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8074 if (len & (sizeof(unsigned long)-1))
8077 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8080 ret = sched_getaffinity(pid, mask);
8082 unsigned int retlen = min(len, cpumask_size());
8084 if (copy_to_user(user_mask_ptr, mask, retlen))
8089 free_cpumask_var(mask);
8094 static void do_sched_yield(void)
8099 rq = this_rq_lock_irq(&rf);
8101 schedstat_inc(rq->yld_count);
8102 current->sched_class->yield_task(rq);
8105 rq_unlock_irq(rq, &rf);
8106 sched_preempt_enable_no_resched();
8112 * sys_sched_yield - yield the current processor to other threads.
8114 * This function yields the current CPU to other tasks. If there are no
8115 * other threads running on this CPU then this function will return.
8119 SYSCALL_DEFINE0(sched_yield)
8125 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8126 int __sched __cond_resched(void)
8128 if (should_resched(0)) {
8129 preempt_schedule_common();
8133 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8134 * whether the current CPU is in an RCU read-side critical section,
8135 * so the tick can report quiescent states even for CPUs looping
8136 * in kernel context. In contrast, in non-preemptible kernels,
8137 * RCU readers leave no in-memory hints, which means that CPU-bound
8138 * processes executing in kernel context might never report an
8139 * RCU quiescent state. Therefore, the following code causes
8140 * cond_resched() to report a quiescent state, but only when RCU
8141 * is in urgent need of one.
8143 #ifndef CONFIG_PREEMPT_RCU
8148 EXPORT_SYMBOL(__cond_resched);
8151 #ifdef CONFIG_PREEMPT_DYNAMIC
8152 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8153 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8155 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8156 EXPORT_STATIC_CALL_TRAMP(might_resched);
8160 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8161 * call schedule, and on return reacquire the lock.
8163 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8164 * operations here to prevent schedule() from being called twice (once via
8165 * spin_unlock(), once by hand).
8167 int __cond_resched_lock(spinlock_t *lock)
8169 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8172 lockdep_assert_held(lock);
8174 if (spin_needbreak(lock) || resched) {
8177 preempt_schedule_common();
8185 EXPORT_SYMBOL(__cond_resched_lock);
8187 int __cond_resched_rwlock_read(rwlock_t *lock)
8189 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8192 lockdep_assert_held_read(lock);
8194 if (rwlock_needbreak(lock) || resched) {
8197 preempt_schedule_common();
8205 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8207 int __cond_resched_rwlock_write(rwlock_t *lock)
8209 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8212 lockdep_assert_held_write(lock);
8214 if (rwlock_needbreak(lock) || resched) {
8217 preempt_schedule_common();
8225 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8228 * yield - yield the current processor to other threads.
8230 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8232 * The scheduler is at all times free to pick the calling task as the most
8233 * eligible task to run, if removing the yield() call from your code breaks
8234 * it, it's already broken.
8236 * Typical broken usage is:
8241 * where one assumes that yield() will let 'the other' process run that will
8242 * make event true. If the current task is a SCHED_FIFO task that will never
8243 * happen. Never use yield() as a progress guarantee!!
8245 * If you want to use yield() to wait for something, use wait_event().
8246 * If you want to use yield() to be 'nice' for others, use cond_resched().
8247 * If you still want to use yield(), do not!
8249 void __sched yield(void)
8251 set_current_state(TASK_RUNNING);
8254 EXPORT_SYMBOL(yield);
8257 * yield_to - yield the current processor to another thread in
8258 * your thread group, or accelerate that thread toward the
8259 * processor it's on.
8261 * @preempt: whether task preemption is allowed or not
8263 * It's the caller's job to ensure that the target task struct
8264 * can't go away on us before we can do any checks.
8267 * true (>0) if we indeed boosted the target task.
8268 * false (0) if we failed to boost the target.
8269 * -ESRCH if there's no task to yield to.
8271 int __sched yield_to(struct task_struct *p, bool preempt)
8273 struct task_struct *curr = current;
8274 struct rq *rq, *p_rq;
8275 unsigned long flags;
8278 local_irq_save(flags);
8284 * If we're the only runnable task on the rq and target rq also
8285 * has only one task, there's absolutely no point in yielding.
8287 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8292 double_rq_lock(rq, p_rq);
8293 if (task_rq(p) != p_rq) {
8294 double_rq_unlock(rq, p_rq);
8298 if (!curr->sched_class->yield_to_task)
8301 if (curr->sched_class != p->sched_class)
8304 if (task_running(p_rq, p) || !task_is_running(p))
8307 yielded = curr->sched_class->yield_to_task(rq, p);
8309 schedstat_inc(rq->yld_count);
8311 * Make p's CPU reschedule; pick_next_entity takes care of
8314 if (preempt && rq != p_rq)
8319 double_rq_unlock(rq, p_rq);
8321 local_irq_restore(flags);
8328 EXPORT_SYMBOL_GPL(yield_to);
8330 int io_schedule_prepare(void)
8332 int old_iowait = current->in_iowait;
8334 current->in_iowait = 1;
8336 blk_flush_plug(current->plug, true);
8341 void io_schedule_finish(int token)
8343 current->in_iowait = token;
8347 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8348 * that process accounting knows that this is a task in IO wait state.
8350 long __sched io_schedule_timeout(long timeout)
8355 token = io_schedule_prepare();
8356 ret = schedule_timeout(timeout);
8357 io_schedule_finish(token);
8361 EXPORT_SYMBOL(io_schedule_timeout);
8363 void __sched io_schedule(void)
8367 token = io_schedule_prepare();
8369 io_schedule_finish(token);
8371 EXPORT_SYMBOL(io_schedule);
8374 * sys_sched_get_priority_max - return maximum RT priority.
8375 * @policy: scheduling class.
8377 * Return: On success, this syscall returns the maximum
8378 * rt_priority that can be used by a given scheduling class.
8379 * On failure, a negative error code is returned.
8381 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8388 ret = MAX_RT_PRIO-1;
8390 case SCHED_DEADLINE:
8401 * sys_sched_get_priority_min - return minimum RT priority.
8402 * @policy: scheduling class.
8404 * Return: On success, this syscall returns the minimum
8405 * rt_priority that can be used by a given scheduling class.
8406 * On failure, a negative error code is returned.
8408 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8417 case SCHED_DEADLINE:
8426 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8428 struct task_struct *p;
8429 unsigned int time_slice;
8439 p = find_process_by_pid(pid);
8443 retval = security_task_getscheduler(p);
8447 rq = task_rq_lock(p, &rf);
8449 if (p->sched_class->get_rr_interval)
8450 time_slice = p->sched_class->get_rr_interval(rq, p);
8451 task_rq_unlock(rq, p, &rf);
8454 jiffies_to_timespec64(time_slice, t);
8463 * sys_sched_rr_get_interval - return the default timeslice of a process.
8464 * @pid: pid of the process.
8465 * @interval: userspace pointer to the timeslice value.
8467 * this syscall writes the default timeslice value of a given process
8468 * into the user-space timespec buffer. A value of '0' means infinity.
8470 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8473 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8474 struct __kernel_timespec __user *, interval)
8476 struct timespec64 t;
8477 int retval = sched_rr_get_interval(pid, &t);
8480 retval = put_timespec64(&t, interval);
8485 #ifdef CONFIG_COMPAT_32BIT_TIME
8486 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8487 struct old_timespec32 __user *, interval)
8489 struct timespec64 t;
8490 int retval = sched_rr_get_interval(pid, &t);
8493 retval = put_old_timespec32(&t, interval);
8498 void sched_show_task(struct task_struct *p)
8500 unsigned long free = 0;
8503 if (!try_get_task_stack(p))
8506 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8508 if (task_is_running(p))
8509 pr_cont(" running task ");
8510 #ifdef CONFIG_DEBUG_STACK_USAGE
8511 free = stack_not_used(p);
8516 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8518 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8519 free, task_pid_nr(p), ppid,
8520 (unsigned long)task_thread_info(p)->flags);
8522 print_worker_info(KERN_INFO, p);
8523 print_stop_info(KERN_INFO, p);
8524 show_stack(p, NULL, KERN_INFO);
8527 EXPORT_SYMBOL_GPL(sched_show_task);
8530 state_filter_match(unsigned long state_filter, struct task_struct *p)
8532 unsigned int state = READ_ONCE(p->__state);
8534 /* no filter, everything matches */
8538 /* filter, but doesn't match */
8539 if (!(state & state_filter))
8543 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8546 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8553 void show_state_filter(unsigned int state_filter)
8555 struct task_struct *g, *p;
8558 for_each_process_thread(g, p) {
8560 * reset the NMI-timeout, listing all files on a slow
8561 * console might take a lot of time:
8562 * Also, reset softlockup watchdogs on all CPUs, because
8563 * another CPU might be blocked waiting for us to process
8566 touch_nmi_watchdog();
8567 touch_all_softlockup_watchdogs();
8568 if (state_filter_match(state_filter, p))
8572 #ifdef CONFIG_SCHED_DEBUG
8574 sysrq_sched_debug_show();
8578 * Only show locks if all tasks are dumped:
8581 debug_show_all_locks();
8585 * init_idle - set up an idle thread for a given CPU
8586 * @idle: task in question
8587 * @cpu: CPU the idle task belongs to
8589 * NOTE: this function does not set the idle thread's NEED_RESCHED
8590 * flag, to make booting more robust.
8592 void __init init_idle(struct task_struct *idle, int cpu)
8594 struct rq *rq = cpu_rq(cpu);
8595 unsigned long flags;
8597 __sched_fork(0, idle);
8600 * The idle task doesn't need the kthread struct to function, but it
8601 * is dressed up as a per-CPU kthread and thus needs to play the part
8602 * if we want to avoid special-casing it in code that deals with per-CPU
8605 set_kthread_struct(idle);
8607 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8608 raw_spin_rq_lock(rq);
8610 idle->__state = TASK_RUNNING;
8611 idle->se.exec_start = sched_clock();
8613 * PF_KTHREAD should already be set at this point; regardless, make it
8614 * look like a proper per-CPU kthread.
8616 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8617 kthread_set_per_cpu(idle, cpu);
8619 scs_task_reset(idle);
8620 kasan_unpoison_task_stack(idle);
8624 * It's possible that init_idle() gets called multiple times on a task,
8625 * in that case do_set_cpus_allowed() will not do the right thing.
8627 * And since this is boot we can forgo the serialization.
8629 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8632 * We're having a chicken and egg problem, even though we are
8633 * holding rq->lock, the CPU isn't yet set to this CPU so the
8634 * lockdep check in task_group() will fail.
8636 * Similar case to sched_fork(). / Alternatively we could
8637 * use task_rq_lock() here and obtain the other rq->lock.
8642 __set_task_cpu(idle, cpu);
8646 rcu_assign_pointer(rq->curr, idle);
8647 idle->on_rq = TASK_ON_RQ_QUEUED;
8651 raw_spin_rq_unlock(rq);
8652 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8654 /* Set the preempt count _outside_ the spinlocks! */
8655 init_idle_preempt_count(idle, cpu);
8658 * The idle tasks have their own, simple scheduling class:
8660 idle->sched_class = &idle_sched_class;
8661 ftrace_graph_init_idle_task(idle, cpu);
8662 vtime_init_idle(idle, cpu);
8664 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8670 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8671 const struct cpumask *trial)
8675 if (!cpumask_weight(cur))
8678 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8683 int task_can_attach(struct task_struct *p,
8684 const struct cpumask *cs_cpus_allowed)
8689 * Kthreads which disallow setaffinity shouldn't be moved
8690 * to a new cpuset; we don't want to change their CPU
8691 * affinity and isolating such threads by their set of
8692 * allowed nodes is unnecessary. Thus, cpusets are not
8693 * applicable for such threads. This prevents checking for
8694 * success of set_cpus_allowed_ptr() on all attached tasks
8695 * before cpus_mask may be changed.
8697 if (p->flags & PF_NO_SETAFFINITY) {
8702 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8704 ret = dl_task_can_attach(p, cs_cpus_allowed);
8710 bool sched_smp_initialized __read_mostly;
8712 #ifdef CONFIG_NUMA_BALANCING
8713 /* Migrate current task p to target_cpu */
8714 int migrate_task_to(struct task_struct *p, int target_cpu)
8716 struct migration_arg arg = { p, target_cpu };
8717 int curr_cpu = task_cpu(p);
8719 if (curr_cpu == target_cpu)
8722 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8725 /* TODO: This is not properly updating schedstats */
8727 trace_sched_move_numa(p, curr_cpu, target_cpu);
8728 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8732 * Requeue a task on a given node and accurately track the number of NUMA
8733 * tasks on the runqueues
8735 void sched_setnuma(struct task_struct *p, int nid)
8737 bool queued, running;
8741 rq = task_rq_lock(p, &rf);
8742 queued = task_on_rq_queued(p);
8743 running = task_current(rq, p);
8746 dequeue_task(rq, p, DEQUEUE_SAVE);
8748 put_prev_task(rq, p);
8750 p->numa_preferred_nid = nid;
8753 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8755 set_next_task(rq, p);
8756 task_rq_unlock(rq, p, &rf);
8758 #endif /* CONFIG_NUMA_BALANCING */
8760 #ifdef CONFIG_HOTPLUG_CPU
8762 * Ensure that the idle task is using init_mm right before its CPU goes
8765 void idle_task_exit(void)
8767 struct mm_struct *mm = current->active_mm;
8769 BUG_ON(cpu_online(smp_processor_id()));
8770 BUG_ON(current != this_rq()->idle);
8772 if (mm != &init_mm) {
8773 switch_mm(mm, &init_mm, current);
8774 finish_arch_post_lock_switch();
8777 scs_task_reset(current);
8778 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8781 static int __balance_push_cpu_stop(void *arg)
8783 struct task_struct *p = arg;
8784 struct rq *rq = this_rq();
8788 raw_spin_lock_irq(&p->pi_lock);
8791 update_rq_clock(rq);
8793 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8794 cpu = select_fallback_rq(rq->cpu, p);
8795 rq = __migrate_task(rq, &rf, p, cpu);
8799 raw_spin_unlock_irq(&p->pi_lock);
8806 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8809 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8811 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8812 * effective when the hotplug motion is down.
8814 static void balance_push(struct rq *rq)
8816 struct task_struct *push_task = rq->curr;
8818 lockdep_assert_rq_held(rq);
8821 * Ensure the thing is persistent until balance_push_set(.on = false);
8823 rq->balance_callback = &balance_push_callback;
8826 * Only active while going offline and when invoked on the outgoing
8829 if (!cpu_dying(rq->cpu) || rq != this_rq())
8833 * Both the cpu-hotplug and stop task are in this case and are
8834 * required to complete the hotplug process.
8836 if (kthread_is_per_cpu(push_task) ||
8837 is_migration_disabled(push_task)) {
8840 * If this is the idle task on the outgoing CPU try to wake
8841 * up the hotplug control thread which might wait for the
8842 * last task to vanish. The rcuwait_active() check is
8843 * accurate here because the waiter is pinned on this CPU
8844 * and can't obviously be running in parallel.
8846 * On RT kernels this also has to check whether there are
8847 * pinned and scheduled out tasks on the runqueue. They
8848 * need to leave the migrate disabled section first.
8850 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8851 rcuwait_active(&rq->hotplug_wait)) {
8852 raw_spin_rq_unlock(rq);
8853 rcuwait_wake_up(&rq->hotplug_wait);
8854 raw_spin_rq_lock(rq);
8859 get_task_struct(push_task);
8861 * Temporarily drop rq->lock such that we can wake-up the stop task.
8862 * Both preemption and IRQs are still disabled.
8864 raw_spin_rq_unlock(rq);
8865 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8866 this_cpu_ptr(&push_work));
8868 * At this point need_resched() is true and we'll take the loop in
8869 * schedule(). The next pick is obviously going to be the stop task
8870 * which kthread_is_per_cpu() and will push this task away.
8872 raw_spin_rq_lock(rq);
8875 static void balance_push_set(int cpu, bool on)
8877 struct rq *rq = cpu_rq(cpu);
8880 rq_lock_irqsave(rq, &rf);
8882 WARN_ON_ONCE(rq->balance_callback);
8883 rq->balance_callback = &balance_push_callback;
8884 } else if (rq->balance_callback == &balance_push_callback) {
8885 rq->balance_callback = NULL;
8887 rq_unlock_irqrestore(rq, &rf);
8891 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8892 * inactive. All tasks which are not per CPU kernel threads are either
8893 * pushed off this CPU now via balance_push() or placed on a different CPU
8894 * during wakeup. Wait until the CPU is quiescent.
8896 static void balance_hotplug_wait(void)
8898 struct rq *rq = this_rq();
8900 rcuwait_wait_event(&rq->hotplug_wait,
8901 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8902 TASK_UNINTERRUPTIBLE);
8907 static inline void balance_push(struct rq *rq)
8911 static inline void balance_push_set(int cpu, bool on)
8915 static inline void balance_hotplug_wait(void)
8919 #endif /* CONFIG_HOTPLUG_CPU */
8921 void set_rq_online(struct rq *rq)
8924 const struct sched_class *class;
8926 cpumask_set_cpu(rq->cpu, rq->rd->online);
8929 for_each_class(class) {
8930 if (class->rq_online)
8931 class->rq_online(rq);
8936 void set_rq_offline(struct rq *rq)
8939 const struct sched_class *class;
8941 for_each_class(class) {
8942 if (class->rq_offline)
8943 class->rq_offline(rq);
8946 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8952 * used to mark begin/end of suspend/resume:
8954 static int num_cpus_frozen;
8957 * Update cpusets according to cpu_active mask. If cpusets are
8958 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8959 * around partition_sched_domains().
8961 * If we come here as part of a suspend/resume, don't touch cpusets because we
8962 * want to restore it back to its original state upon resume anyway.
8964 static void cpuset_cpu_active(void)
8966 if (cpuhp_tasks_frozen) {
8968 * num_cpus_frozen tracks how many CPUs are involved in suspend
8969 * resume sequence. As long as this is not the last online
8970 * operation in the resume sequence, just build a single sched
8971 * domain, ignoring cpusets.
8973 partition_sched_domains(1, NULL, NULL);
8974 if (--num_cpus_frozen)
8977 * This is the last CPU online operation. So fall through and
8978 * restore the original sched domains by considering the
8979 * cpuset configurations.
8981 cpuset_force_rebuild();
8983 cpuset_update_active_cpus();
8986 static int cpuset_cpu_inactive(unsigned int cpu)
8988 if (!cpuhp_tasks_frozen) {
8989 if (dl_cpu_busy(cpu))
8991 cpuset_update_active_cpus();
8994 partition_sched_domains(1, NULL, NULL);
8999 int sched_cpu_activate(unsigned int cpu)
9001 struct rq *rq = cpu_rq(cpu);
9005 * Clear the balance_push callback and prepare to schedule
9008 balance_push_set(cpu, false);
9010 #ifdef CONFIG_SCHED_SMT
9012 * When going up, increment the number of cores with SMT present.
9014 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9015 static_branch_inc_cpuslocked(&sched_smt_present);
9017 set_cpu_active(cpu, true);
9019 if (sched_smp_initialized) {
9020 sched_domains_numa_masks_set(cpu);
9021 cpuset_cpu_active();
9025 * Put the rq online, if not already. This happens:
9027 * 1) In the early boot process, because we build the real domains
9028 * after all CPUs have been brought up.
9030 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9033 rq_lock_irqsave(rq, &rf);
9035 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9038 rq_unlock_irqrestore(rq, &rf);
9043 int sched_cpu_deactivate(unsigned int cpu)
9045 struct rq *rq = cpu_rq(cpu);
9050 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9051 * load balancing when not active
9053 nohz_balance_exit_idle(rq);
9055 set_cpu_active(cpu, false);
9058 * From this point forward, this CPU will refuse to run any task that
9059 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9060 * push those tasks away until this gets cleared, see
9061 * sched_cpu_dying().
9063 balance_push_set(cpu, true);
9066 * We've cleared cpu_active_mask / set balance_push, wait for all
9067 * preempt-disabled and RCU users of this state to go away such that
9068 * all new such users will observe it.
9070 * Specifically, we rely on ttwu to no longer target this CPU, see
9071 * ttwu_queue_cond() and is_cpu_allowed().
9073 * Do sync before park smpboot threads to take care the rcu boost case.
9077 rq_lock_irqsave(rq, &rf);
9079 update_rq_clock(rq);
9080 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9083 rq_unlock_irqrestore(rq, &rf);
9085 #ifdef CONFIG_SCHED_SMT
9087 * When going down, decrement the number of cores with SMT present.
9089 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9090 static_branch_dec_cpuslocked(&sched_smt_present);
9092 sched_core_cpu_deactivate(cpu);
9095 if (!sched_smp_initialized)
9098 ret = cpuset_cpu_inactive(cpu);
9100 balance_push_set(cpu, false);
9101 set_cpu_active(cpu, true);
9104 sched_domains_numa_masks_clear(cpu);
9108 static void sched_rq_cpu_starting(unsigned int cpu)
9110 struct rq *rq = cpu_rq(cpu);
9112 rq->calc_load_update = calc_load_update;
9113 update_max_interval();
9116 int sched_cpu_starting(unsigned int cpu)
9118 sched_core_cpu_starting(cpu);
9119 sched_rq_cpu_starting(cpu);
9120 sched_tick_start(cpu);
9124 #ifdef CONFIG_HOTPLUG_CPU
9127 * Invoked immediately before the stopper thread is invoked to bring the
9128 * CPU down completely. At this point all per CPU kthreads except the
9129 * hotplug thread (current) and the stopper thread (inactive) have been
9130 * either parked or have been unbound from the outgoing CPU. Ensure that
9131 * any of those which might be on the way out are gone.
9133 * If after this point a bound task is being woken on this CPU then the
9134 * responsible hotplug callback has failed to do it's job.
9135 * sched_cpu_dying() will catch it with the appropriate fireworks.
9137 int sched_cpu_wait_empty(unsigned int cpu)
9139 balance_hotplug_wait();
9144 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9145 * might have. Called from the CPU stopper task after ensuring that the
9146 * stopper is the last running task on the CPU, so nr_active count is
9147 * stable. We need to take the teardown thread which is calling this into
9148 * account, so we hand in adjust = 1 to the load calculation.
9150 * Also see the comment "Global load-average calculations".
9152 static void calc_load_migrate(struct rq *rq)
9154 long delta = calc_load_fold_active(rq, 1);
9157 atomic_long_add(delta, &calc_load_tasks);
9160 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9162 struct task_struct *g, *p;
9163 int cpu = cpu_of(rq);
9165 lockdep_assert_rq_held(rq);
9167 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9168 for_each_process_thread(g, p) {
9169 if (task_cpu(p) != cpu)
9172 if (!task_on_rq_queued(p))
9175 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9179 int sched_cpu_dying(unsigned int cpu)
9181 struct rq *rq = cpu_rq(cpu);
9184 /* Handle pending wakeups and then migrate everything off */
9185 sched_tick_stop(cpu);
9187 rq_lock_irqsave(rq, &rf);
9188 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9189 WARN(true, "Dying CPU not properly vacated!");
9190 dump_rq_tasks(rq, KERN_WARNING);
9192 rq_unlock_irqrestore(rq, &rf);
9194 calc_load_migrate(rq);
9195 update_max_interval();
9197 sched_core_cpu_dying(cpu);
9202 void __init sched_init_smp(void)
9207 * There's no userspace yet to cause hotplug operations; hence all the
9208 * CPU masks are stable and all blatant races in the below code cannot
9211 mutex_lock(&sched_domains_mutex);
9212 sched_init_domains(cpu_active_mask);
9213 mutex_unlock(&sched_domains_mutex);
9215 /* Move init over to a non-isolated CPU */
9216 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9218 current->flags &= ~PF_NO_SETAFFINITY;
9219 sched_init_granularity();
9221 init_sched_rt_class();
9222 init_sched_dl_class();
9224 sched_smp_initialized = true;
9227 static int __init migration_init(void)
9229 sched_cpu_starting(smp_processor_id());
9232 early_initcall(migration_init);
9235 void __init sched_init_smp(void)
9237 sched_init_granularity();
9239 #endif /* CONFIG_SMP */
9241 int in_sched_functions(unsigned long addr)
9243 return in_lock_functions(addr) ||
9244 (addr >= (unsigned long)__sched_text_start
9245 && addr < (unsigned long)__sched_text_end);
9248 #ifdef CONFIG_CGROUP_SCHED
9250 * Default task group.
9251 * Every task in system belongs to this group at bootup.
9253 struct task_group root_task_group;
9254 LIST_HEAD(task_groups);
9256 /* Cacheline aligned slab cache for task_group */
9257 static struct kmem_cache *task_group_cache __read_mostly;
9260 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9261 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9263 void __init sched_init(void)
9265 unsigned long ptr = 0;
9268 /* Make sure the linker didn't screw up */
9269 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9270 &fair_sched_class + 1 != &rt_sched_class ||
9271 &rt_sched_class + 1 != &dl_sched_class);
9273 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9278 #ifdef CONFIG_FAIR_GROUP_SCHED
9279 ptr += 2 * nr_cpu_ids * sizeof(void **);
9281 #ifdef CONFIG_RT_GROUP_SCHED
9282 ptr += 2 * nr_cpu_ids * sizeof(void **);
9285 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9287 #ifdef CONFIG_FAIR_GROUP_SCHED
9288 root_task_group.se = (struct sched_entity **)ptr;
9289 ptr += nr_cpu_ids * sizeof(void **);
9291 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9292 ptr += nr_cpu_ids * sizeof(void **);
9294 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9295 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9296 #endif /* CONFIG_FAIR_GROUP_SCHED */
9297 #ifdef CONFIG_RT_GROUP_SCHED
9298 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9299 ptr += nr_cpu_ids * sizeof(void **);
9301 root_task_group.rt_rq = (struct rt_rq **)ptr;
9302 ptr += nr_cpu_ids * sizeof(void **);
9304 #endif /* CONFIG_RT_GROUP_SCHED */
9306 #ifdef CONFIG_CPUMASK_OFFSTACK
9307 for_each_possible_cpu(i) {
9308 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9309 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9310 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9311 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9313 #endif /* CONFIG_CPUMASK_OFFSTACK */
9315 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9316 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9319 init_defrootdomain();
9322 #ifdef CONFIG_RT_GROUP_SCHED
9323 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9324 global_rt_period(), global_rt_runtime());
9325 #endif /* CONFIG_RT_GROUP_SCHED */
9327 #ifdef CONFIG_CGROUP_SCHED
9328 task_group_cache = KMEM_CACHE(task_group, 0);
9330 list_add(&root_task_group.list, &task_groups);
9331 INIT_LIST_HEAD(&root_task_group.children);
9332 INIT_LIST_HEAD(&root_task_group.siblings);
9333 autogroup_init(&init_task);
9334 #endif /* CONFIG_CGROUP_SCHED */
9336 for_each_possible_cpu(i) {
9340 raw_spin_lock_init(&rq->__lock);
9342 rq->calc_load_active = 0;
9343 rq->calc_load_update = jiffies + LOAD_FREQ;
9344 init_cfs_rq(&rq->cfs);
9345 init_rt_rq(&rq->rt);
9346 init_dl_rq(&rq->dl);
9347 #ifdef CONFIG_FAIR_GROUP_SCHED
9348 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9349 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9351 * How much CPU bandwidth does root_task_group get?
9353 * In case of task-groups formed thr' the cgroup filesystem, it
9354 * gets 100% of the CPU resources in the system. This overall
9355 * system CPU resource is divided among the tasks of
9356 * root_task_group and its child task-groups in a fair manner,
9357 * based on each entity's (task or task-group's) weight
9358 * (se->load.weight).
9360 * In other words, if root_task_group has 10 tasks of weight
9361 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9362 * then A0's share of the CPU resource is:
9364 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9366 * We achieve this by letting root_task_group's tasks sit
9367 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9369 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9370 #endif /* CONFIG_FAIR_GROUP_SCHED */
9372 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9373 #ifdef CONFIG_RT_GROUP_SCHED
9374 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9379 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9380 rq->balance_callback = &balance_push_callback;
9381 rq->active_balance = 0;
9382 rq->next_balance = jiffies;
9387 rq->avg_idle = 2*sysctl_sched_migration_cost;
9388 rq->wake_stamp = jiffies;
9389 rq->wake_avg_idle = rq->avg_idle;
9390 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9392 INIT_LIST_HEAD(&rq->cfs_tasks);
9394 rq_attach_root(rq, &def_root_domain);
9395 #ifdef CONFIG_NO_HZ_COMMON
9396 rq->last_blocked_load_update_tick = jiffies;
9397 atomic_set(&rq->nohz_flags, 0);
9399 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9401 #ifdef CONFIG_HOTPLUG_CPU
9402 rcuwait_init(&rq->hotplug_wait);
9404 #endif /* CONFIG_SMP */
9406 atomic_set(&rq->nr_iowait, 0);
9408 #ifdef CONFIG_SCHED_CORE
9410 rq->core_pick = NULL;
9411 rq->core_enabled = 0;
9412 rq->core_tree = RB_ROOT;
9413 rq->core_forceidle = false;
9415 rq->core_cookie = 0UL;
9419 set_load_weight(&init_task, false);
9422 * The boot idle thread does lazy MMU switching as well:
9425 enter_lazy_tlb(&init_mm, current);
9428 * Make us the idle thread. Technically, schedule() should not be
9429 * called from this thread, however somewhere below it might be,
9430 * but because we are the idle thread, we just pick up running again
9431 * when this runqueue becomes "idle".
9433 init_idle(current, smp_processor_id());
9435 calc_load_update = jiffies + LOAD_FREQ;
9438 idle_thread_set_boot_cpu();
9439 balance_push_set(smp_processor_id(), false);
9441 init_sched_fair_class();
9447 preempt_dynamic_init();
9449 scheduler_running = 1;
9452 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9454 void __might_sleep(const char *file, int line)
9456 unsigned int state = get_current_state();
9458 * Blocking primitives will set (and therefore destroy) current->state,
9459 * since we will exit with TASK_RUNNING make sure we enter with it,
9460 * otherwise we will destroy state.
9462 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9463 "do not call blocking ops when !TASK_RUNNING; "
9464 "state=%x set at [<%p>] %pS\n", state,
9465 (void *)current->task_state_change,
9466 (void *)current->task_state_change);
9468 __might_resched(file, line, 0);
9470 EXPORT_SYMBOL(__might_sleep);
9472 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9474 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9477 if (preempt_count() == preempt_offset)
9480 pr_err("Preemption disabled at:");
9481 print_ip_sym(KERN_ERR, ip);
9484 static inline bool resched_offsets_ok(unsigned int offsets)
9486 unsigned int nested = preempt_count();
9488 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9490 return nested == offsets;
9493 void __might_resched(const char *file, int line, unsigned int offsets)
9495 /* Ratelimiting timestamp: */
9496 static unsigned long prev_jiffy;
9498 unsigned long preempt_disable_ip;
9500 /* WARN_ON_ONCE() by default, no rate limit required: */
9503 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9504 !is_idle_task(current) && !current->non_block_count) ||
9505 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9509 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9511 prev_jiffy = jiffies;
9513 /* Save this before calling printk(), since that will clobber it: */
9514 preempt_disable_ip = get_preempt_disable_ip(current);
9516 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9518 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9519 in_atomic(), irqs_disabled(), current->non_block_count,
9520 current->pid, current->comm);
9521 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9522 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9524 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9525 pr_err("RCU nest depth: %d, expected: %u\n",
9526 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9529 if (task_stack_end_corrupted(current))
9530 pr_emerg("Thread overran stack, or stack corrupted\n");
9532 debug_show_held_locks(current);
9533 if (irqs_disabled())
9534 print_irqtrace_events(current);
9536 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9537 preempt_disable_ip);
9540 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9542 EXPORT_SYMBOL(__might_resched);
9544 void __cant_sleep(const char *file, int line, int preempt_offset)
9546 static unsigned long prev_jiffy;
9548 if (irqs_disabled())
9551 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9554 if (preempt_count() > preempt_offset)
9557 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9559 prev_jiffy = jiffies;
9561 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9562 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9563 in_atomic(), irqs_disabled(),
9564 current->pid, current->comm);
9566 debug_show_held_locks(current);
9568 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9570 EXPORT_SYMBOL_GPL(__cant_sleep);
9573 void __cant_migrate(const char *file, int line)
9575 static unsigned long prev_jiffy;
9577 if (irqs_disabled())
9580 if (is_migration_disabled(current))
9583 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9586 if (preempt_count() > 0)
9589 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9591 prev_jiffy = jiffies;
9593 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9594 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9595 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9596 current->pid, current->comm);
9598 debug_show_held_locks(current);
9600 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9602 EXPORT_SYMBOL_GPL(__cant_migrate);
9606 #ifdef CONFIG_MAGIC_SYSRQ
9607 void normalize_rt_tasks(void)
9609 struct task_struct *g, *p;
9610 struct sched_attr attr = {
9611 .sched_policy = SCHED_NORMAL,
9614 read_lock(&tasklist_lock);
9615 for_each_process_thread(g, p) {
9617 * Only normalize user tasks:
9619 if (p->flags & PF_KTHREAD)
9622 p->se.exec_start = 0;
9623 schedstat_set(p->stats.wait_start, 0);
9624 schedstat_set(p->stats.sleep_start, 0);
9625 schedstat_set(p->stats.block_start, 0);
9627 if (!dl_task(p) && !rt_task(p)) {
9629 * Renice negative nice level userspace
9632 if (task_nice(p) < 0)
9633 set_user_nice(p, 0);
9637 __sched_setscheduler(p, &attr, false, false);
9639 read_unlock(&tasklist_lock);
9642 #endif /* CONFIG_MAGIC_SYSRQ */
9644 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9646 * These functions are only useful for the IA64 MCA handling, or kdb.
9648 * They can only be called when the whole system has been
9649 * stopped - every CPU needs to be quiescent, and no scheduling
9650 * activity can take place. Using them for anything else would
9651 * be a serious bug, and as a result, they aren't even visible
9652 * under any other configuration.
9656 * curr_task - return the current task for a given CPU.
9657 * @cpu: the processor in question.
9659 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9661 * Return: The current task for @cpu.
9663 struct task_struct *curr_task(int cpu)
9665 return cpu_curr(cpu);
9668 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9672 * ia64_set_curr_task - set the current task for a given CPU.
9673 * @cpu: the processor in question.
9674 * @p: the task pointer to set.
9676 * Description: This function must only be used when non-maskable interrupts
9677 * are serviced on a separate stack. It allows the architecture to switch the
9678 * notion of the current task on a CPU in a non-blocking manner. This function
9679 * must be called with all CPU's synchronized, and interrupts disabled, the
9680 * and caller must save the original value of the current task (see
9681 * curr_task() above) and restore that value before reenabling interrupts and
9682 * re-starting the system.
9684 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9686 void ia64_set_curr_task(int cpu, struct task_struct *p)
9693 #ifdef CONFIG_CGROUP_SCHED
9694 /* task_group_lock serializes the addition/removal of task groups */
9695 static DEFINE_SPINLOCK(task_group_lock);
9697 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9698 struct task_group *parent)
9700 #ifdef CONFIG_UCLAMP_TASK_GROUP
9701 enum uclamp_id clamp_id;
9703 for_each_clamp_id(clamp_id) {
9704 uclamp_se_set(&tg->uclamp_req[clamp_id],
9705 uclamp_none(clamp_id), false);
9706 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9711 static void sched_free_group(struct task_group *tg)
9713 free_fair_sched_group(tg);
9714 free_rt_sched_group(tg);
9716 kmem_cache_free(task_group_cache, tg);
9719 /* allocate runqueue etc for a new task group */
9720 struct task_group *sched_create_group(struct task_group *parent)
9722 struct task_group *tg;
9724 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9726 return ERR_PTR(-ENOMEM);
9728 if (!alloc_fair_sched_group(tg, parent))
9731 if (!alloc_rt_sched_group(tg, parent))
9734 alloc_uclamp_sched_group(tg, parent);
9739 sched_free_group(tg);
9740 return ERR_PTR(-ENOMEM);
9743 void sched_online_group(struct task_group *tg, struct task_group *parent)
9745 unsigned long flags;
9747 spin_lock_irqsave(&task_group_lock, flags);
9748 list_add_rcu(&tg->list, &task_groups);
9750 /* Root should already exist: */
9753 tg->parent = parent;
9754 INIT_LIST_HEAD(&tg->children);
9755 list_add_rcu(&tg->siblings, &parent->children);
9756 spin_unlock_irqrestore(&task_group_lock, flags);
9758 online_fair_sched_group(tg);
9761 /* rcu callback to free various structures associated with a task group */
9762 static void sched_free_group_rcu(struct rcu_head *rhp)
9764 /* Now it should be safe to free those cfs_rqs: */
9765 sched_free_group(container_of(rhp, struct task_group, rcu));
9768 void sched_destroy_group(struct task_group *tg)
9770 /* Wait for possible concurrent references to cfs_rqs complete: */
9771 call_rcu(&tg->rcu, sched_free_group_rcu);
9774 void sched_offline_group(struct task_group *tg)
9776 unsigned long flags;
9778 /* End participation in shares distribution: */
9779 unregister_fair_sched_group(tg);
9781 spin_lock_irqsave(&task_group_lock, flags);
9782 list_del_rcu(&tg->list);
9783 list_del_rcu(&tg->siblings);
9784 spin_unlock_irqrestore(&task_group_lock, flags);
9787 static void sched_change_group(struct task_struct *tsk, int type)
9789 struct task_group *tg;
9792 * All callers are synchronized by task_rq_lock(); we do not use RCU
9793 * which is pointless here. Thus, we pass "true" to task_css_check()
9794 * to prevent lockdep warnings.
9796 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9797 struct task_group, css);
9798 tg = autogroup_task_group(tsk, tg);
9799 tsk->sched_task_group = tg;
9801 #ifdef CONFIG_FAIR_GROUP_SCHED
9802 if (tsk->sched_class->task_change_group)
9803 tsk->sched_class->task_change_group(tsk, type);
9806 set_task_rq(tsk, task_cpu(tsk));
9810 * Change task's runqueue when it moves between groups.
9812 * The caller of this function should have put the task in its new group by
9813 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9816 void sched_move_task(struct task_struct *tsk)
9818 int queued, running, queue_flags =
9819 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9823 rq = task_rq_lock(tsk, &rf);
9824 update_rq_clock(rq);
9826 running = task_current(rq, tsk);
9827 queued = task_on_rq_queued(tsk);
9830 dequeue_task(rq, tsk, queue_flags);
9832 put_prev_task(rq, tsk);
9834 sched_change_group(tsk, TASK_MOVE_GROUP);
9837 enqueue_task(rq, tsk, queue_flags);
9839 set_next_task(rq, tsk);
9841 * After changing group, the running task may have joined a
9842 * throttled one but it's still the running task. Trigger a
9843 * resched to make sure that task can still run.
9848 task_rq_unlock(rq, tsk, &rf);
9851 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9853 return css ? container_of(css, struct task_group, css) : NULL;
9856 static struct cgroup_subsys_state *
9857 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9859 struct task_group *parent = css_tg(parent_css);
9860 struct task_group *tg;
9863 /* This is early initialization for the top cgroup */
9864 return &root_task_group.css;
9867 tg = sched_create_group(parent);
9869 return ERR_PTR(-ENOMEM);
9874 /* Expose task group only after completing cgroup initialization */
9875 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9877 struct task_group *tg = css_tg(css);
9878 struct task_group *parent = css_tg(css->parent);
9881 sched_online_group(tg, parent);
9883 #ifdef CONFIG_UCLAMP_TASK_GROUP
9884 /* Propagate the effective uclamp value for the new group */
9885 mutex_lock(&uclamp_mutex);
9887 cpu_util_update_eff(css);
9889 mutex_unlock(&uclamp_mutex);
9895 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9897 struct task_group *tg = css_tg(css);
9899 sched_offline_group(tg);
9902 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9904 struct task_group *tg = css_tg(css);
9907 * Relies on the RCU grace period between css_released() and this.
9909 sched_free_group(tg);
9913 * This is called before wake_up_new_task(), therefore we really only
9914 * have to set its group bits, all the other stuff does not apply.
9916 static void cpu_cgroup_fork(struct task_struct *task)
9921 rq = task_rq_lock(task, &rf);
9923 update_rq_clock(rq);
9924 sched_change_group(task, TASK_SET_GROUP);
9926 task_rq_unlock(rq, task, &rf);
9929 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9931 struct task_struct *task;
9932 struct cgroup_subsys_state *css;
9935 cgroup_taskset_for_each(task, css, tset) {
9936 #ifdef CONFIG_RT_GROUP_SCHED
9937 if (!sched_rt_can_attach(css_tg(css), task))
9941 * Serialize against wake_up_new_task() such that if it's
9942 * running, we're sure to observe its full state.
9944 raw_spin_lock_irq(&task->pi_lock);
9946 * Avoid calling sched_move_task() before wake_up_new_task()
9947 * has happened. This would lead to problems with PELT, due to
9948 * move wanting to detach+attach while we're not attached yet.
9950 if (READ_ONCE(task->__state) == TASK_NEW)
9952 raw_spin_unlock_irq(&task->pi_lock);
9960 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9962 struct task_struct *task;
9963 struct cgroup_subsys_state *css;
9965 cgroup_taskset_for_each(task, css, tset)
9966 sched_move_task(task);
9969 #ifdef CONFIG_UCLAMP_TASK_GROUP
9970 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9972 struct cgroup_subsys_state *top_css = css;
9973 struct uclamp_se *uc_parent = NULL;
9974 struct uclamp_se *uc_se = NULL;
9975 unsigned int eff[UCLAMP_CNT];
9976 enum uclamp_id clamp_id;
9977 unsigned int clamps;
9979 lockdep_assert_held(&uclamp_mutex);
9980 SCHED_WARN_ON(!rcu_read_lock_held());
9982 css_for_each_descendant_pre(css, top_css) {
9983 uc_parent = css_tg(css)->parent
9984 ? css_tg(css)->parent->uclamp : NULL;
9986 for_each_clamp_id(clamp_id) {
9987 /* Assume effective clamps matches requested clamps */
9988 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
9989 /* Cap effective clamps with parent's effective clamps */
9991 eff[clamp_id] > uc_parent[clamp_id].value) {
9992 eff[clamp_id] = uc_parent[clamp_id].value;
9995 /* Ensure protection is always capped by limit */
9996 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
9998 /* Propagate most restrictive effective clamps */
10000 uc_se = css_tg(css)->uclamp;
10001 for_each_clamp_id(clamp_id) {
10002 if (eff[clamp_id] == uc_se[clamp_id].value)
10004 uc_se[clamp_id].value = eff[clamp_id];
10005 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10006 clamps |= (0x1 << clamp_id);
10009 css = css_rightmost_descendant(css);
10013 /* Immediately update descendants RUNNABLE tasks */
10014 uclamp_update_active_tasks(css);
10019 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10020 * C expression. Since there is no way to convert a macro argument (N) into a
10021 * character constant, use two levels of macros.
10023 #define _POW10(exp) ((unsigned int)1e##exp)
10024 #define POW10(exp) _POW10(exp)
10026 struct uclamp_request {
10027 #define UCLAMP_PERCENT_SHIFT 2
10028 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10034 static inline struct uclamp_request
10035 capacity_from_percent(char *buf)
10037 struct uclamp_request req = {
10038 .percent = UCLAMP_PERCENT_SCALE,
10039 .util = SCHED_CAPACITY_SCALE,
10044 if (strcmp(buf, "max")) {
10045 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10049 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10054 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10055 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10061 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10062 size_t nbytes, loff_t off,
10063 enum uclamp_id clamp_id)
10065 struct uclamp_request req;
10066 struct task_group *tg;
10068 req = capacity_from_percent(buf);
10072 static_branch_enable(&sched_uclamp_used);
10074 mutex_lock(&uclamp_mutex);
10077 tg = css_tg(of_css(of));
10078 if (tg->uclamp_req[clamp_id].value != req.util)
10079 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10082 * Because of not recoverable conversion rounding we keep track of the
10083 * exact requested value
10085 tg->uclamp_pct[clamp_id] = req.percent;
10087 /* Update effective clamps to track the most restrictive value */
10088 cpu_util_update_eff(of_css(of));
10091 mutex_unlock(&uclamp_mutex);
10096 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10097 char *buf, size_t nbytes,
10100 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10103 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10104 char *buf, size_t nbytes,
10107 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10110 static inline void cpu_uclamp_print(struct seq_file *sf,
10111 enum uclamp_id clamp_id)
10113 struct task_group *tg;
10119 tg = css_tg(seq_css(sf));
10120 util_clamp = tg->uclamp_req[clamp_id].value;
10123 if (util_clamp == SCHED_CAPACITY_SCALE) {
10124 seq_puts(sf, "max\n");
10128 percent = tg->uclamp_pct[clamp_id];
10129 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10130 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10133 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10135 cpu_uclamp_print(sf, UCLAMP_MIN);
10139 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10141 cpu_uclamp_print(sf, UCLAMP_MAX);
10144 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10146 #ifdef CONFIG_FAIR_GROUP_SCHED
10147 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10148 struct cftype *cftype, u64 shareval)
10150 if (shareval > scale_load_down(ULONG_MAX))
10151 shareval = MAX_SHARES;
10152 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10155 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10156 struct cftype *cft)
10158 struct task_group *tg = css_tg(css);
10160 return (u64) scale_load_down(tg->shares);
10163 #ifdef CONFIG_CFS_BANDWIDTH
10164 static DEFINE_MUTEX(cfs_constraints_mutex);
10166 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10167 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10168 /* More than 203 days if BW_SHIFT equals 20. */
10169 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10171 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10173 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10176 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10177 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10179 if (tg == &root_task_group)
10183 * Ensure we have at some amount of bandwidth every period. This is
10184 * to prevent reaching a state of large arrears when throttled via
10185 * entity_tick() resulting in prolonged exit starvation.
10187 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10191 * Likewise, bound things on the other side by preventing insane quota
10192 * periods. This also allows us to normalize in computing quota
10195 if (period > max_cfs_quota_period)
10199 * Bound quota to defend quota against overflow during bandwidth shift.
10201 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10204 if (quota != RUNTIME_INF && (burst > quota ||
10205 burst + quota > max_cfs_runtime))
10209 * Prevent race between setting of cfs_rq->runtime_enabled and
10210 * unthrottle_offline_cfs_rqs().
10213 mutex_lock(&cfs_constraints_mutex);
10214 ret = __cfs_schedulable(tg, period, quota);
10218 runtime_enabled = quota != RUNTIME_INF;
10219 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10221 * If we need to toggle cfs_bandwidth_used, off->on must occur
10222 * before making related changes, and on->off must occur afterwards
10224 if (runtime_enabled && !runtime_was_enabled)
10225 cfs_bandwidth_usage_inc();
10226 raw_spin_lock_irq(&cfs_b->lock);
10227 cfs_b->period = ns_to_ktime(period);
10228 cfs_b->quota = quota;
10229 cfs_b->burst = burst;
10231 __refill_cfs_bandwidth_runtime(cfs_b);
10233 /* Restart the period timer (if active) to handle new period expiry: */
10234 if (runtime_enabled)
10235 start_cfs_bandwidth(cfs_b);
10237 raw_spin_unlock_irq(&cfs_b->lock);
10239 for_each_online_cpu(i) {
10240 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10241 struct rq *rq = cfs_rq->rq;
10242 struct rq_flags rf;
10244 rq_lock_irq(rq, &rf);
10245 cfs_rq->runtime_enabled = runtime_enabled;
10246 cfs_rq->runtime_remaining = 0;
10248 if (cfs_rq->throttled)
10249 unthrottle_cfs_rq(cfs_rq);
10250 rq_unlock_irq(rq, &rf);
10252 if (runtime_was_enabled && !runtime_enabled)
10253 cfs_bandwidth_usage_dec();
10255 mutex_unlock(&cfs_constraints_mutex);
10256 cpus_read_unlock();
10261 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10263 u64 quota, period, burst;
10265 period = ktime_to_ns(tg->cfs_bandwidth.period);
10266 burst = tg->cfs_bandwidth.burst;
10267 if (cfs_quota_us < 0)
10268 quota = RUNTIME_INF;
10269 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10270 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10274 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10277 static long tg_get_cfs_quota(struct task_group *tg)
10281 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10284 quota_us = tg->cfs_bandwidth.quota;
10285 do_div(quota_us, NSEC_PER_USEC);
10290 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10292 u64 quota, period, burst;
10294 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10297 period = (u64)cfs_period_us * NSEC_PER_USEC;
10298 quota = tg->cfs_bandwidth.quota;
10299 burst = tg->cfs_bandwidth.burst;
10301 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10304 static long tg_get_cfs_period(struct task_group *tg)
10308 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10309 do_div(cfs_period_us, NSEC_PER_USEC);
10311 return cfs_period_us;
10314 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10316 u64 quota, period, burst;
10318 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10321 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10322 period = ktime_to_ns(tg->cfs_bandwidth.period);
10323 quota = tg->cfs_bandwidth.quota;
10325 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10328 static long tg_get_cfs_burst(struct task_group *tg)
10332 burst_us = tg->cfs_bandwidth.burst;
10333 do_div(burst_us, NSEC_PER_USEC);
10338 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10339 struct cftype *cft)
10341 return tg_get_cfs_quota(css_tg(css));
10344 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10345 struct cftype *cftype, s64 cfs_quota_us)
10347 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10350 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10351 struct cftype *cft)
10353 return tg_get_cfs_period(css_tg(css));
10356 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10357 struct cftype *cftype, u64 cfs_period_us)
10359 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10362 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10363 struct cftype *cft)
10365 return tg_get_cfs_burst(css_tg(css));
10368 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10369 struct cftype *cftype, u64 cfs_burst_us)
10371 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10374 struct cfs_schedulable_data {
10375 struct task_group *tg;
10380 * normalize group quota/period to be quota/max_period
10381 * note: units are usecs
10383 static u64 normalize_cfs_quota(struct task_group *tg,
10384 struct cfs_schedulable_data *d)
10389 period = d->period;
10392 period = tg_get_cfs_period(tg);
10393 quota = tg_get_cfs_quota(tg);
10396 /* note: these should typically be equivalent */
10397 if (quota == RUNTIME_INF || quota == -1)
10398 return RUNTIME_INF;
10400 return to_ratio(period, quota);
10403 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10405 struct cfs_schedulable_data *d = data;
10406 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10407 s64 quota = 0, parent_quota = -1;
10410 quota = RUNTIME_INF;
10412 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10414 quota = normalize_cfs_quota(tg, d);
10415 parent_quota = parent_b->hierarchical_quota;
10418 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10419 * always take the min. On cgroup1, only inherit when no
10422 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10423 quota = min(quota, parent_quota);
10425 if (quota == RUNTIME_INF)
10426 quota = parent_quota;
10427 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10431 cfs_b->hierarchical_quota = quota;
10436 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10439 struct cfs_schedulable_data data = {
10445 if (quota != RUNTIME_INF) {
10446 do_div(data.period, NSEC_PER_USEC);
10447 do_div(data.quota, NSEC_PER_USEC);
10451 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10457 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10459 struct task_group *tg = css_tg(seq_css(sf));
10460 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10462 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10463 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10464 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10466 if (schedstat_enabled() && tg != &root_task_group) {
10467 struct sched_statistics *stats;
10471 for_each_possible_cpu(i) {
10472 stats = __schedstats_from_se(tg->se[i]);
10473 ws += schedstat_val(stats->wait_sum);
10476 seq_printf(sf, "wait_sum %llu\n", ws);
10479 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10480 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10484 #endif /* CONFIG_CFS_BANDWIDTH */
10485 #endif /* CONFIG_FAIR_GROUP_SCHED */
10487 #ifdef CONFIG_RT_GROUP_SCHED
10488 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10489 struct cftype *cft, s64 val)
10491 return sched_group_set_rt_runtime(css_tg(css), val);
10494 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10495 struct cftype *cft)
10497 return sched_group_rt_runtime(css_tg(css));
10500 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10501 struct cftype *cftype, u64 rt_period_us)
10503 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10506 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10507 struct cftype *cft)
10509 return sched_group_rt_period(css_tg(css));
10511 #endif /* CONFIG_RT_GROUP_SCHED */
10513 #ifdef CONFIG_FAIR_GROUP_SCHED
10514 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10515 struct cftype *cft)
10517 return css_tg(css)->idle;
10520 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10521 struct cftype *cft, s64 idle)
10523 return sched_group_set_idle(css_tg(css), idle);
10527 static struct cftype cpu_legacy_files[] = {
10528 #ifdef CONFIG_FAIR_GROUP_SCHED
10531 .read_u64 = cpu_shares_read_u64,
10532 .write_u64 = cpu_shares_write_u64,
10536 .read_s64 = cpu_idle_read_s64,
10537 .write_s64 = cpu_idle_write_s64,
10540 #ifdef CONFIG_CFS_BANDWIDTH
10542 .name = "cfs_quota_us",
10543 .read_s64 = cpu_cfs_quota_read_s64,
10544 .write_s64 = cpu_cfs_quota_write_s64,
10547 .name = "cfs_period_us",
10548 .read_u64 = cpu_cfs_period_read_u64,
10549 .write_u64 = cpu_cfs_period_write_u64,
10552 .name = "cfs_burst_us",
10553 .read_u64 = cpu_cfs_burst_read_u64,
10554 .write_u64 = cpu_cfs_burst_write_u64,
10558 .seq_show = cpu_cfs_stat_show,
10561 #ifdef CONFIG_RT_GROUP_SCHED
10563 .name = "rt_runtime_us",
10564 .read_s64 = cpu_rt_runtime_read,
10565 .write_s64 = cpu_rt_runtime_write,
10568 .name = "rt_period_us",
10569 .read_u64 = cpu_rt_period_read_uint,
10570 .write_u64 = cpu_rt_period_write_uint,
10573 #ifdef CONFIG_UCLAMP_TASK_GROUP
10575 .name = "uclamp.min",
10576 .flags = CFTYPE_NOT_ON_ROOT,
10577 .seq_show = cpu_uclamp_min_show,
10578 .write = cpu_uclamp_min_write,
10581 .name = "uclamp.max",
10582 .flags = CFTYPE_NOT_ON_ROOT,
10583 .seq_show = cpu_uclamp_max_show,
10584 .write = cpu_uclamp_max_write,
10587 { } /* Terminate */
10590 static int cpu_extra_stat_show(struct seq_file *sf,
10591 struct cgroup_subsys_state *css)
10593 #ifdef CONFIG_CFS_BANDWIDTH
10595 struct task_group *tg = css_tg(css);
10596 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10597 u64 throttled_usec, burst_usec;
10599 throttled_usec = cfs_b->throttled_time;
10600 do_div(throttled_usec, NSEC_PER_USEC);
10601 burst_usec = cfs_b->burst_time;
10602 do_div(burst_usec, NSEC_PER_USEC);
10604 seq_printf(sf, "nr_periods %d\n"
10605 "nr_throttled %d\n"
10606 "throttled_usec %llu\n"
10608 "burst_usec %llu\n",
10609 cfs_b->nr_periods, cfs_b->nr_throttled,
10610 throttled_usec, cfs_b->nr_burst, burst_usec);
10616 #ifdef CONFIG_FAIR_GROUP_SCHED
10617 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10618 struct cftype *cft)
10620 struct task_group *tg = css_tg(css);
10621 u64 weight = scale_load_down(tg->shares);
10623 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10626 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10627 struct cftype *cft, u64 weight)
10630 * cgroup weight knobs should use the common MIN, DFL and MAX
10631 * values which are 1, 100 and 10000 respectively. While it loses
10632 * a bit of range on both ends, it maps pretty well onto the shares
10633 * value used by scheduler and the round-trip conversions preserve
10634 * the original value over the entire range.
10636 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10639 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10641 return sched_group_set_shares(css_tg(css), scale_load(weight));
10644 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10645 struct cftype *cft)
10647 unsigned long weight = scale_load_down(css_tg(css)->shares);
10648 int last_delta = INT_MAX;
10651 /* find the closest nice value to the current weight */
10652 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10653 delta = abs(sched_prio_to_weight[prio] - weight);
10654 if (delta >= last_delta)
10656 last_delta = delta;
10659 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10662 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10663 struct cftype *cft, s64 nice)
10665 unsigned long weight;
10668 if (nice < MIN_NICE || nice > MAX_NICE)
10671 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10672 idx = array_index_nospec(idx, 40);
10673 weight = sched_prio_to_weight[idx];
10675 return sched_group_set_shares(css_tg(css), scale_load(weight));
10679 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10680 long period, long quota)
10683 seq_puts(sf, "max");
10685 seq_printf(sf, "%ld", quota);
10687 seq_printf(sf, " %ld\n", period);
10690 /* caller should put the current value in *@periodp before calling */
10691 static int __maybe_unused cpu_period_quota_parse(char *buf,
10692 u64 *periodp, u64 *quotap)
10694 char tok[21]; /* U64_MAX */
10696 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10699 *periodp *= NSEC_PER_USEC;
10701 if (sscanf(tok, "%llu", quotap))
10702 *quotap *= NSEC_PER_USEC;
10703 else if (!strcmp(tok, "max"))
10704 *quotap = RUNTIME_INF;
10711 #ifdef CONFIG_CFS_BANDWIDTH
10712 static int cpu_max_show(struct seq_file *sf, void *v)
10714 struct task_group *tg = css_tg(seq_css(sf));
10716 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10720 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10721 char *buf, size_t nbytes, loff_t off)
10723 struct task_group *tg = css_tg(of_css(of));
10724 u64 period = tg_get_cfs_period(tg);
10725 u64 burst = tg_get_cfs_burst(tg);
10729 ret = cpu_period_quota_parse(buf, &period, "a);
10731 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10732 return ret ?: nbytes;
10736 static struct cftype cpu_files[] = {
10737 #ifdef CONFIG_FAIR_GROUP_SCHED
10740 .flags = CFTYPE_NOT_ON_ROOT,
10741 .read_u64 = cpu_weight_read_u64,
10742 .write_u64 = cpu_weight_write_u64,
10745 .name = "weight.nice",
10746 .flags = CFTYPE_NOT_ON_ROOT,
10747 .read_s64 = cpu_weight_nice_read_s64,
10748 .write_s64 = cpu_weight_nice_write_s64,
10752 .flags = CFTYPE_NOT_ON_ROOT,
10753 .read_s64 = cpu_idle_read_s64,
10754 .write_s64 = cpu_idle_write_s64,
10757 #ifdef CONFIG_CFS_BANDWIDTH
10760 .flags = CFTYPE_NOT_ON_ROOT,
10761 .seq_show = cpu_max_show,
10762 .write = cpu_max_write,
10765 .name = "max.burst",
10766 .flags = CFTYPE_NOT_ON_ROOT,
10767 .read_u64 = cpu_cfs_burst_read_u64,
10768 .write_u64 = cpu_cfs_burst_write_u64,
10771 #ifdef CONFIG_UCLAMP_TASK_GROUP
10773 .name = "uclamp.min",
10774 .flags = CFTYPE_NOT_ON_ROOT,
10775 .seq_show = cpu_uclamp_min_show,
10776 .write = cpu_uclamp_min_write,
10779 .name = "uclamp.max",
10780 .flags = CFTYPE_NOT_ON_ROOT,
10781 .seq_show = cpu_uclamp_max_show,
10782 .write = cpu_uclamp_max_write,
10785 { } /* terminate */
10788 struct cgroup_subsys cpu_cgrp_subsys = {
10789 .css_alloc = cpu_cgroup_css_alloc,
10790 .css_online = cpu_cgroup_css_online,
10791 .css_released = cpu_cgroup_css_released,
10792 .css_free = cpu_cgroup_css_free,
10793 .css_extra_stat_show = cpu_extra_stat_show,
10794 .fork = cpu_cgroup_fork,
10795 .can_attach = cpu_cgroup_can_attach,
10796 .attach = cpu_cgroup_attach,
10797 .legacy_cftypes = cpu_legacy_files,
10798 .dfl_cftypes = cpu_files,
10799 .early_init = true,
10803 #endif /* CONFIG_CGROUP_SCHED */
10805 void dump_cpu_task(int cpu)
10807 pr_info("Task dump for CPU %d:\n", cpu);
10808 sched_show_task(cpu_curr(cpu));
10812 * Nice levels are multiplicative, with a gentle 10% change for every
10813 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10814 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10815 * that remained on nice 0.
10817 * The "10% effect" is relative and cumulative: from _any_ nice level,
10818 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10819 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10820 * If a task goes up by ~10% and another task goes down by ~10% then
10821 * the relative distance between them is ~25%.)
10823 const int sched_prio_to_weight[40] = {
10824 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10825 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10826 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10827 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10828 /* 0 */ 1024, 820, 655, 526, 423,
10829 /* 5 */ 335, 272, 215, 172, 137,
10830 /* 10 */ 110, 87, 70, 56, 45,
10831 /* 15 */ 36, 29, 23, 18, 15,
10835 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10837 * In cases where the weight does not change often, we can use the
10838 * precalculated inverse to speed up arithmetics by turning divisions
10839 * into multiplications:
10841 const u32 sched_prio_to_wmult[40] = {
10842 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10843 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10844 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10845 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10846 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10847 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10848 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10849 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10852 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10854 trace_sched_update_nr_running_tp(rq, count);