1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../io_uring/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
44 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
46 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
48 #ifdef CONFIG_SCHED_DEBUG
50 * Debugging: various feature bits
52 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
53 * sysctl_sched_features, defined in sched.h, to allow constants propagation
54 * at compile time and compiler optimization based on features default.
56 #define SCHED_FEAT(name, enabled) \
57 (1UL << __SCHED_FEAT_##name) * enabled |
58 const_debug unsigned int sysctl_sched_features =
64 * Print a warning if need_resched is set for the given duration (if
65 * LATENCY_WARN is enabled).
67 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
70 __read_mostly int sysctl_resched_latency_warn_ms = 100;
71 __read_mostly int sysctl_resched_latency_warn_once = 1;
72 #endif /* CONFIG_SCHED_DEBUG */
75 * Number of tasks to iterate in a single balance run.
76 * Limited because this is done with IRQs disabled.
78 #ifdef CONFIG_PREEMPT_RT
79 const_debug unsigned int sysctl_sched_nr_migrate = 8;
81 const_debug unsigned int sysctl_sched_nr_migrate = 32;
85 * period over which we measure -rt task CPU usage in us.
88 unsigned int sysctl_sched_rt_period = 1000000;
90 __read_mostly int scheduler_running;
92 #ifdef CONFIG_SCHED_CORE
94 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
96 /* kernel prio, less is more */
97 static inline int __task_prio(struct task_struct *p)
99 if (p->sched_class == &stop_sched_class) /* trumps deadline */
102 if (rt_prio(p->prio)) /* includes deadline */
103 return p->prio; /* [-1, 99] */
105 if (p->sched_class == &idle_sched_class)
106 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
108 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
118 /* real prio, less is less */
119 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
122 int pa = __task_prio(a), pb = __task_prio(b);
130 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
131 return !dl_time_before(a->dl.deadline, b->dl.deadline);
133 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
134 return cfs_prio_less(a, b, in_fi);
139 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
141 if (a->core_cookie < b->core_cookie)
144 if (a->core_cookie > b->core_cookie)
147 /* flip prio, so high prio is leftmost */
148 if (prio_less(b, a, task_rq(a)->core->core_forceidle))
154 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
156 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
158 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
161 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
163 const struct task_struct *p = __node_2_sc(node);
164 unsigned long cookie = (unsigned long)key;
166 if (cookie < p->core_cookie)
169 if (cookie > p->core_cookie)
175 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
177 rq->core->core_task_seq++;
182 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
185 void sched_core_dequeue(struct rq *rq, struct task_struct *p)
187 rq->core->core_task_seq++;
189 if (!sched_core_enqueued(p))
192 rb_erase(&p->core_node, &rq->core_tree);
193 RB_CLEAR_NODE(&p->core_node);
197 * Find left-most (aka, highest priority) task matching @cookie.
199 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
201 struct rb_node *node;
203 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
205 * The idle task always matches any cookie!
208 return idle_sched_class.pick_task(rq);
210 return __node_2_sc(node);
213 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
215 struct rb_node *node = &p->core_node;
217 node = rb_next(node);
221 p = container_of(node, struct task_struct, core_node);
222 if (p->core_cookie != cookie)
229 * Magic required such that:
231 * raw_spin_rq_lock(rq);
233 * raw_spin_rq_unlock(rq);
235 * ends up locking and unlocking the _same_ lock, and all CPUs
236 * always agree on what rq has what lock.
238 * XXX entirely possible to selectively enable cores, don't bother for now.
241 static DEFINE_MUTEX(sched_core_mutex);
242 static atomic_t sched_core_count;
243 static struct cpumask sched_core_mask;
245 static void sched_core_lock(int cpu, unsigned long *flags)
247 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
250 local_irq_save(*flags);
251 for_each_cpu(t, smt_mask)
252 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
255 static void sched_core_unlock(int cpu, unsigned long *flags)
257 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
260 for_each_cpu(t, smt_mask)
261 raw_spin_unlock(&cpu_rq(t)->__lock);
262 local_irq_restore(*flags);
265 static void __sched_core_flip(bool enabled)
273 * Toggle the online cores, one by one.
275 cpumask_copy(&sched_core_mask, cpu_online_mask);
276 for_each_cpu(cpu, &sched_core_mask) {
277 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
279 sched_core_lock(cpu, &flags);
281 for_each_cpu(t, smt_mask)
282 cpu_rq(t)->core_enabled = enabled;
284 sched_core_unlock(cpu, &flags);
286 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
290 * Toggle the offline CPUs.
292 cpumask_copy(&sched_core_mask, cpu_possible_mask);
293 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
295 for_each_cpu(cpu, &sched_core_mask)
296 cpu_rq(cpu)->core_enabled = enabled;
301 static void sched_core_assert_empty(void)
305 for_each_possible_cpu(cpu)
306 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
309 static void __sched_core_enable(void)
311 static_branch_enable(&__sched_core_enabled);
313 * Ensure all previous instances of raw_spin_rq_*lock() have finished
314 * and future ones will observe !sched_core_disabled().
317 __sched_core_flip(true);
318 sched_core_assert_empty();
321 static void __sched_core_disable(void)
323 sched_core_assert_empty();
324 __sched_core_flip(false);
325 static_branch_disable(&__sched_core_enabled);
328 void sched_core_get(void)
330 if (atomic_inc_not_zero(&sched_core_count))
333 mutex_lock(&sched_core_mutex);
334 if (!atomic_read(&sched_core_count))
335 __sched_core_enable();
337 smp_mb__before_atomic();
338 atomic_inc(&sched_core_count);
339 mutex_unlock(&sched_core_mutex);
342 static void __sched_core_put(struct work_struct *work)
344 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
345 __sched_core_disable();
346 mutex_unlock(&sched_core_mutex);
350 void sched_core_put(void)
352 static DECLARE_WORK(_work, __sched_core_put);
355 * "There can be only one"
357 * Either this is the last one, or we don't actually need to do any
358 * 'work'. If it is the last *again*, we rely on
359 * WORK_STRUCT_PENDING_BIT.
361 if (!atomic_add_unless(&sched_core_count, -1, 1))
362 schedule_work(&_work);
365 #else /* !CONFIG_SCHED_CORE */
367 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
368 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
370 #endif /* CONFIG_SCHED_CORE */
373 * part of the period that we allow rt tasks to run in us.
376 int sysctl_sched_rt_runtime = 950000;
380 * Serialization rules:
386 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
389 * rq2->lock where: rq1 < rq2
393 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
394 * local CPU's rq->lock, it optionally removes the task from the runqueue and
395 * always looks at the local rq data structures to find the most eligible task
398 * Task enqueue is also under rq->lock, possibly taken from another CPU.
399 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
400 * the local CPU to avoid bouncing the runqueue state around [ see
401 * ttwu_queue_wakelist() ]
403 * Task wakeup, specifically wakeups that involve migration, are horribly
404 * complicated to avoid having to take two rq->locks.
408 * System-calls and anything external will use task_rq_lock() which acquires
409 * both p->pi_lock and rq->lock. As a consequence the state they change is
410 * stable while holding either lock:
412 * - sched_setaffinity()/
413 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
414 * - set_user_nice(): p->se.load, p->*prio
415 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
416 * p->se.load, p->rt_priority,
417 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
418 * - sched_setnuma(): p->numa_preferred_nid
419 * - sched_move_task()/
420 * cpu_cgroup_fork(): p->sched_task_group
421 * - uclamp_update_active() p->uclamp*
423 * p->state <- TASK_*:
425 * is changed locklessly using set_current_state(), __set_current_state() or
426 * set_special_state(), see their respective comments, or by
427 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
430 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
432 * is set by activate_task() and cleared by deactivate_task(), under
433 * rq->lock. Non-zero indicates the task is runnable, the special
434 * ON_RQ_MIGRATING state is used for migration without holding both
435 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
437 * p->on_cpu <- { 0, 1 }:
439 * is set by prepare_task() and cleared by finish_task() such that it will be
440 * set before p is scheduled-in and cleared after p is scheduled-out, both
441 * under rq->lock. Non-zero indicates the task is running on its CPU.
443 * [ The astute reader will observe that it is possible for two tasks on one
444 * CPU to have ->on_cpu = 1 at the same time. ]
446 * task_cpu(p): is changed by set_task_cpu(), the rules are:
448 * - Don't call set_task_cpu() on a blocked task:
450 * We don't care what CPU we're not running on, this simplifies hotplug,
451 * the CPU assignment of blocked tasks isn't required to be valid.
453 * - for try_to_wake_up(), called under p->pi_lock:
455 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
457 * - for migration called under rq->lock:
458 * [ see task_on_rq_migrating() in task_rq_lock() ]
460 * o move_queued_task()
463 * - for migration called under double_rq_lock():
465 * o __migrate_swap_task()
466 * o push_rt_task() / pull_rt_task()
467 * o push_dl_task() / pull_dl_task()
468 * o dl_task_offline_migration()
472 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
474 raw_spinlock_t *lock;
476 /* Matches synchronize_rcu() in __sched_core_enable() */
478 if (sched_core_disabled()) {
479 raw_spin_lock_nested(&rq->__lock, subclass);
480 /* preempt_count *MUST* be > 1 */
481 preempt_enable_no_resched();
486 lock = __rq_lockp(rq);
487 raw_spin_lock_nested(lock, subclass);
488 if (likely(lock == __rq_lockp(rq))) {
489 /* preempt_count *MUST* be > 1 */
490 preempt_enable_no_resched();
493 raw_spin_unlock(lock);
497 bool raw_spin_rq_trylock(struct rq *rq)
499 raw_spinlock_t *lock;
502 /* Matches synchronize_rcu() in __sched_core_enable() */
504 if (sched_core_disabled()) {
505 ret = raw_spin_trylock(&rq->__lock);
511 lock = __rq_lockp(rq);
512 ret = raw_spin_trylock(lock);
513 if (!ret || (likely(lock == __rq_lockp(rq)))) {
517 raw_spin_unlock(lock);
521 void raw_spin_rq_unlock(struct rq *rq)
523 raw_spin_unlock(rq_lockp(rq));
528 * double_rq_lock - safely lock two runqueues
530 void double_rq_lock(struct rq *rq1, struct rq *rq2)
532 lockdep_assert_irqs_disabled();
534 if (rq_order_less(rq2, rq1))
537 raw_spin_rq_lock(rq1);
538 if (__rq_lockp(rq1) != __rq_lockp(rq2))
539 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
541 double_rq_clock_clear_update(rq1, rq2);
546 * __task_rq_lock - lock the rq @p resides on.
548 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
553 lockdep_assert_held(&p->pi_lock);
557 raw_spin_rq_lock(rq);
558 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
562 raw_spin_rq_unlock(rq);
564 while (unlikely(task_on_rq_migrating(p)))
570 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
572 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
573 __acquires(p->pi_lock)
579 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
581 raw_spin_rq_lock(rq);
583 * move_queued_task() task_rq_lock()
586 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
587 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
588 * [S] ->cpu = new_cpu [L] task_rq()
592 * If we observe the old CPU in task_rq_lock(), the acquire of
593 * the old rq->lock will fully serialize against the stores.
595 * If we observe the new CPU in task_rq_lock(), the address
596 * dependency headed by '[L] rq = task_rq()' and the acquire
597 * will pair with the WMB to ensure we then also see migrating.
599 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
603 raw_spin_rq_unlock(rq);
604 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
606 while (unlikely(task_on_rq_migrating(p)))
612 * RQ-clock updating methods:
615 static void update_rq_clock_task(struct rq *rq, s64 delta)
618 * In theory, the compile should just see 0 here, and optimize out the call
619 * to sched_rt_avg_update. But I don't trust it...
621 s64 __maybe_unused steal = 0, irq_delta = 0;
623 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
624 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
627 * Since irq_time is only updated on {soft,}irq_exit, we might run into
628 * this case when a previous update_rq_clock() happened inside a
631 * When this happens, we stop ->clock_task and only update the
632 * prev_irq_time stamp to account for the part that fit, so that a next
633 * update will consume the rest. This ensures ->clock_task is
636 * It does however cause some slight miss-attribution of {soft,}irq
637 * time, a more accurate solution would be to update the irq_time using
638 * the current rq->clock timestamp, except that would require using
641 if (irq_delta > delta)
644 rq->prev_irq_time += irq_delta;
647 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
648 if (static_key_false((¶virt_steal_rq_enabled))) {
649 steal = paravirt_steal_clock(cpu_of(rq));
650 steal -= rq->prev_steal_time_rq;
652 if (unlikely(steal > delta))
655 rq->prev_steal_time_rq += steal;
660 rq->clock_task += delta;
662 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
663 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
664 update_irq_load_avg(rq, irq_delta + steal);
666 update_rq_clock_pelt(rq, delta);
669 void update_rq_clock(struct rq *rq)
673 lockdep_assert_rq_held(rq);
675 if (rq->clock_update_flags & RQCF_ACT_SKIP)
678 #ifdef CONFIG_SCHED_DEBUG
679 if (sched_feat(WARN_DOUBLE_CLOCK))
680 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
681 rq->clock_update_flags |= RQCF_UPDATED;
684 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
688 update_rq_clock_task(rq, delta);
691 #ifdef CONFIG_SCHED_HRTICK
693 * Use HR-timers to deliver accurate preemption points.
696 static void hrtick_clear(struct rq *rq)
698 if (hrtimer_active(&rq->hrtick_timer))
699 hrtimer_cancel(&rq->hrtick_timer);
703 * High-resolution timer tick.
704 * Runs from hardirq context with interrupts disabled.
706 static enum hrtimer_restart hrtick(struct hrtimer *timer)
708 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
711 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
715 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
718 return HRTIMER_NORESTART;
723 static void __hrtick_restart(struct rq *rq)
725 struct hrtimer *timer = &rq->hrtick_timer;
726 ktime_t time = rq->hrtick_time;
728 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
732 * called from hardirq (IPI) context
734 static void __hrtick_start(void *arg)
740 __hrtick_restart(rq);
745 * Called to set the hrtick timer state.
747 * called with rq->lock held and irqs disabled
749 void hrtick_start(struct rq *rq, u64 delay)
751 struct hrtimer *timer = &rq->hrtick_timer;
755 * Don't schedule slices shorter than 10000ns, that just
756 * doesn't make sense and can cause timer DoS.
758 delta = max_t(s64, delay, 10000LL);
759 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
762 __hrtick_restart(rq);
764 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
769 * Called to set the hrtick timer state.
771 * called with rq->lock held and irqs disabled
773 void hrtick_start(struct rq *rq, u64 delay)
776 * Don't schedule slices shorter than 10000ns, that just
777 * doesn't make sense. Rely on vruntime for fairness.
779 delay = max_t(u64, delay, 10000LL);
780 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
781 HRTIMER_MODE_REL_PINNED_HARD);
784 #endif /* CONFIG_SMP */
786 static void hrtick_rq_init(struct rq *rq)
789 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
791 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
792 rq->hrtick_timer.function = hrtick;
794 #else /* CONFIG_SCHED_HRTICK */
795 static inline void hrtick_clear(struct rq *rq)
799 static inline void hrtick_rq_init(struct rq *rq)
802 #endif /* CONFIG_SCHED_HRTICK */
805 * cmpxchg based fetch_or, macro so it works for different integer types
807 #define fetch_or(ptr, mask) \
809 typeof(ptr) _ptr = (ptr); \
810 typeof(mask) _mask = (mask); \
811 typeof(*_ptr) _old, _val = *_ptr; \
814 _old = cmpxchg(_ptr, _val, _val | _mask); \
822 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
824 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
825 * this avoids any races wrt polling state changes and thereby avoids
828 static bool set_nr_and_not_polling(struct task_struct *p)
830 struct thread_info *ti = task_thread_info(p);
831 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
835 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
837 * If this returns true, then the idle task promises to call
838 * sched_ttwu_pending() and reschedule soon.
840 static bool set_nr_if_polling(struct task_struct *p)
842 struct thread_info *ti = task_thread_info(p);
843 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
846 if (!(val & _TIF_POLLING_NRFLAG))
848 if (val & _TIF_NEED_RESCHED)
850 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
859 static bool set_nr_and_not_polling(struct task_struct *p)
861 set_tsk_need_resched(p);
866 static bool set_nr_if_polling(struct task_struct *p)
873 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
875 struct wake_q_node *node = &task->wake_q;
878 * Atomically grab the task, if ->wake_q is !nil already it means
879 * it's already queued (either by us or someone else) and will get the
880 * wakeup due to that.
882 * In order to ensure that a pending wakeup will observe our pending
883 * state, even in the failed case, an explicit smp_mb() must be used.
885 smp_mb__before_atomic();
886 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
890 * The head is context local, there can be no concurrency.
893 head->lastp = &node->next;
898 * wake_q_add() - queue a wakeup for 'later' waking.
899 * @head: the wake_q_head to add @task to
900 * @task: the task to queue for 'later' wakeup
902 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
903 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
906 * This function must be used as-if it were wake_up_process(); IOW the task
907 * must be ready to be woken at this location.
909 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
911 if (__wake_q_add(head, task))
912 get_task_struct(task);
916 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
917 * @head: the wake_q_head to add @task to
918 * @task: the task to queue for 'later' wakeup
920 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
921 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
924 * This function must be used as-if it were wake_up_process(); IOW the task
925 * must be ready to be woken at this location.
927 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
928 * that already hold reference to @task can call the 'safe' version and trust
929 * wake_q to do the right thing depending whether or not the @task is already
932 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
934 if (!__wake_q_add(head, task))
935 put_task_struct(task);
938 void wake_up_q(struct wake_q_head *head)
940 struct wake_q_node *node = head->first;
942 while (node != WAKE_Q_TAIL) {
943 struct task_struct *task;
945 task = container_of(node, struct task_struct, wake_q);
946 /* Task can safely be re-inserted now: */
948 task->wake_q.next = NULL;
951 * wake_up_process() executes a full barrier, which pairs with
952 * the queueing in wake_q_add() so as not to miss wakeups.
954 wake_up_process(task);
955 put_task_struct(task);
960 * resched_curr - mark rq's current task 'to be rescheduled now'.
962 * On UP this means the setting of the need_resched flag, on SMP it
963 * might also involve a cross-CPU call to trigger the scheduler on
966 void resched_curr(struct rq *rq)
968 struct task_struct *curr = rq->curr;
971 lockdep_assert_rq_held(rq);
973 if (test_tsk_need_resched(curr))
978 if (cpu == smp_processor_id()) {
979 set_tsk_need_resched(curr);
980 set_preempt_need_resched();
984 if (set_nr_and_not_polling(curr))
985 smp_send_reschedule(cpu);
987 trace_sched_wake_idle_without_ipi(cpu);
990 #ifdef CONFIG_PREEMPT_LAZY
992 static int tsk_is_polling(struct task_struct *p)
994 #ifdef TIF_POLLING_NRFLAG
995 return test_tsk_thread_flag(p, TIF_POLLING_NRFLAG);
1001 void resched_curr_lazy(struct rq *rq)
1003 struct task_struct *curr = rq->curr;
1006 if (!sched_feat(PREEMPT_LAZY)) {
1011 if (test_tsk_need_resched(curr))
1014 if (test_tsk_need_resched_lazy(curr))
1017 set_tsk_need_resched_lazy(curr);
1020 if (cpu == smp_processor_id())
1023 /* NEED_RESCHED_LAZY must be visible before we test polling */
1025 if (!tsk_is_polling(curr))
1026 smp_send_reschedule(cpu);
1030 void resched_cpu(int cpu)
1032 struct rq *rq = cpu_rq(cpu);
1033 unsigned long flags;
1035 raw_spin_rq_lock_irqsave(rq, flags);
1036 if (cpu_online(cpu) || cpu == smp_processor_id())
1038 raw_spin_rq_unlock_irqrestore(rq, flags);
1042 #ifdef CONFIG_NO_HZ_COMMON
1044 * In the semi idle case, use the nearest busy CPU for migrating timers
1045 * from an idle CPU. This is good for power-savings.
1047 * We don't do similar optimization for completely idle system, as
1048 * selecting an idle CPU will add more delays to the timers than intended
1049 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1051 int get_nohz_timer_target(void)
1053 int i, cpu = smp_processor_id(), default_cpu = -1;
1054 struct sched_domain *sd;
1055 const struct cpumask *hk_mask;
1057 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1063 hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1066 for_each_domain(cpu, sd) {
1067 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1078 if (default_cpu == -1)
1079 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1087 * When add_timer_on() enqueues a timer into the timer wheel of an
1088 * idle CPU then this timer might expire before the next timer event
1089 * which is scheduled to wake up that CPU. In case of a completely
1090 * idle system the next event might even be infinite time into the
1091 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1092 * leaves the inner idle loop so the newly added timer is taken into
1093 * account when the CPU goes back to idle and evaluates the timer
1094 * wheel for the next timer event.
1096 static void wake_up_idle_cpu(int cpu)
1098 struct rq *rq = cpu_rq(cpu);
1100 if (cpu == smp_processor_id())
1103 if (set_nr_and_not_polling(rq->idle))
1104 smp_send_reschedule(cpu);
1106 trace_sched_wake_idle_without_ipi(cpu);
1109 static bool wake_up_full_nohz_cpu(int cpu)
1112 * We just need the target to call irq_exit() and re-evaluate
1113 * the next tick. The nohz full kick at least implies that.
1114 * If needed we can still optimize that later with an
1117 if (cpu_is_offline(cpu))
1118 return true; /* Don't try to wake offline CPUs. */
1119 if (tick_nohz_full_cpu(cpu)) {
1120 if (cpu != smp_processor_id() ||
1121 tick_nohz_tick_stopped())
1122 tick_nohz_full_kick_cpu(cpu);
1130 * Wake up the specified CPU. If the CPU is going offline, it is the
1131 * caller's responsibility to deal with the lost wakeup, for example,
1132 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1134 void wake_up_nohz_cpu(int cpu)
1136 if (!wake_up_full_nohz_cpu(cpu))
1137 wake_up_idle_cpu(cpu);
1140 static void nohz_csd_func(void *info)
1142 struct rq *rq = info;
1143 int cpu = cpu_of(rq);
1147 * Release the rq::nohz_csd.
1149 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1150 WARN_ON(!(flags & NOHZ_KICK_MASK));
1152 rq->idle_balance = idle_cpu(cpu);
1153 if (rq->idle_balance && !need_resched()) {
1154 rq->nohz_idle_balance = flags;
1155 raise_softirq_irqoff(SCHED_SOFTIRQ);
1159 #endif /* CONFIG_NO_HZ_COMMON */
1161 #ifdef CONFIG_NO_HZ_FULL
1162 bool sched_can_stop_tick(struct rq *rq)
1164 int fifo_nr_running;
1166 /* Deadline tasks, even if single, need the tick */
1167 if (rq->dl.dl_nr_running)
1171 * If there are more than one RR tasks, we need the tick to affect the
1172 * actual RR behaviour.
1174 if (rq->rt.rr_nr_running) {
1175 if (rq->rt.rr_nr_running == 1)
1182 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1183 * forced preemption between FIFO tasks.
1185 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1186 if (fifo_nr_running)
1190 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1191 * if there's more than one we need the tick for involuntary
1194 if (rq->nr_running > 1)
1199 #endif /* CONFIG_NO_HZ_FULL */
1200 #endif /* CONFIG_SMP */
1202 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1203 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1205 * Iterate task_group tree rooted at *from, calling @down when first entering a
1206 * node and @up when leaving it for the final time.
1208 * Caller must hold rcu_lock or sufficient equivalent.
1210 int walk_tg_tree_from(struct task_group *from,
1211 tg_visitor down, tg_visitor up, void *data)
1213 struct task_group *parent, *child;
1219 ret = (*down)(parent, data);
1222 list_for_each_entry_rcu(child, &parent->children, siblings) {
1229 ret = (*up)(parent, data);
1230 if (ret || parent == from)
1234 parent = parent->parent;
1241 int tg_nop(struct task_group *tg, void *data)
1247 static void set_load_weight(struct task_struct *p, bool update_load)
1249 int prio = p->static_prio - MAX_RT_PRIO;
1250 struct load_weight *load = &p->se.load;
1253 * SCHED_IDLE tasks get minimal weight:
1255 if (task_has_idle_policy(p)) {
1256 load->weight = scale_load(WEIGHT_IDLEPRIO);
1257 load->inv_weight = WMULT_IDLEPRIO;
1262 * SCHED_OTHER tasks have to update their load when changing their
1265 if (update_load && p->sched_class == &fair_sched_class) {
1266 reweight_task(p, prio);
1268 load->weight = scale_load(sched_prio_to_weight[prio]);
1269 load->inv_weight = sched_prio_to_wmult[prio];
1273 #ifdef CONFIG_UCLAMP_TASK
1275 * Serializes updates of utilization clamp values
1277 * The (slow-path) user-space triggers utilization clamp value updates which
1278 * can require updates on (fast-path) scheduler's data structures used to
1279 * support enqueue/dequeue operations.
1280 * While the per-CPU rq lock protects fast-path update operations, user-space
1281 * requests are serialized using a mutex to reduce the risk of conflicting
1282 * updates or API abuses.
1284 static DEFINE_MUTEX(uclamp_mutex);
1286 /* Max allowed minimum utilization */
1287 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1289 /* Max allowed maximum utilization */
1290 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1293 * By default RT tasks run at the maximum performance point/capacity of the
1294 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1295 * SCHED_CAPACITY_SCALE.
1297 * This knob allows admins to change the default behavior when uclamp is being
1298 * used. In battery powered devices, particularly, running at the maximum
1299 * capacity and frequency will increase energy consumption and shorten the
1302 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1304 * This knob will not override the system default sched_util_clamp_min defined
1307 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1309 /* All clamps are required to be less or equal than these values */
1310 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1313 * This static key is used to reduce the uclamp overhead in the fast path. It
1314 * primarily disables the call to uclamp_rq_{inc, dec}() in
1315 * enqueue/dequeue_task().
1317 * This allows users to continue to enable uclamp in their kernel config with
1318 * minimum uclamp overhead in the fast path.
1320 * As soon as userspace modifies any of the uclamp knobs, the static key is
1321 * enabled, since we have an actual users that make use of uclamp
1324 * The knobs that would enable this static key are:
1326 * * A task modifying its uclamp value with sched_setattr().
1327 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1328 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1330 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1332 /* Integer rounded range for each bucket */
1333 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1335 #define for_each_clamp_id(clamp_id) \
1336 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1338 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1340 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1343 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1345 if (clamp_id == UCLAMP_MIN)
1347 return SCHED_CAPACITY_SCALE;
1350 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1351 unsigned int value, bool user_defined)
1353 uc_se->value = value;
1354 uc_se->bucket_id = uclamp_bucket_id(value);
1355 uc_se->user_defined = user_defined;
1358 static inline unsigned int
1359 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1360 unsigned int clamp_value)
1363 * Avoid blocked utilization pushing up the frequency when we go
1364 * idle (which drops the max-clamp) by retaining the last known
1367 if (clamp_id == UCLAMP_MAX) {
1368 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1372 return uclamp_none(UCLAMP_MIN);
1375 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1376 unsigned int clamp_value)
1378 /* Reset max-clamp retention only on idle exit */
1379 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1382 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1386 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1387 unsigned int clamp_value)
1389 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1390 int bucket_id = UCLAMP_BUCKETS - 1;
1393 * Since both min and max clamps are max aggregated, find the
1394 * top most bucket with tasks in.
1396 for ( ; bucket_id >= 0; bucket_id--) {
1397 if (!bucket[bucket_id].tasks)
1399 return bucket[bucket_id].value;
1402 /* No tasks -- default clamp values */
1403 return uclamp_idle_value(rq, clamp_id, clamp_value);
1406 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1408 unsigned int default_util_min;
1409 struct uclamp_se *uc_se;
1411 lockdep_assert_held(&p->pi_lock);
1413 uc_se = &p->uclamp_req[UCLAMP_MIN];
1415 /* Only sync if user didn't override the default */
1416 if (uc_se->user_defined)
1419 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1420 uclamp_se_set(uc_se, default_util_min, false);
1423 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1431 /* Protect updates to p->uclamp_* */
1432 rq = task_rq_lock(p, &rf);
1433 __uclamp_update_util_min_rt_default(p);
1434 task_rq_unlock(rq, p, &rf);
1437 static void uclamp_sync_util_min_rt_default(void)
1439 struct task_struct *g, *p;
1442 * copy_process() sysctl_uclamp
1443 * uclamp_min_rt = X;
1444 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1445 * // link thread smp_mb__after_spinlock()
1446 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1447 * sched_post_fork() for_each_process_thread()
1448 * __uclamp_sync_rt() __uclamp_sync_rt()
1450 * Ensures that either sched_post_fork() will observe the new
1451 * uclamp_min_rt or for_each_process_thread() will observe the new
1454 read_lock(&tasklist_lock);
1455 smp_mb__after_spinlock();
1456 read_unlock(&tasklist_lock);
1459 for_each_process_thread(g, p)
1460 uclamp_update_util_min_rt_default(p);
1464 static inline struct uclamp_se
1465 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1467 /* Copy by value as we could modify it */
1468 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1469 #ifdef CONFIG_UCLAMP_TASK_GROUP
1470 unsigned int tg_min, tg_max, value;
1473 * Tasks in autogroups or root task group will be
1474 * restricted by system defaults.
1476 if (task_group_is_autogroup(task_group(p)))
1478 if (task_group(p) == &root_task_group)
1481 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1482 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1483 value = uc_req.value;
1484 value = clamp(value, tg_min, tg_max);
1485 uclamp_se_set(&uc_req, value, false);
1492 * The effective clamp bucket index of a task depends on, by increasing
1494 * - the task specific clamp value, when explicitly requested from userspace
1495 * - the task group effective clamp value, for tasks not either in the root
1496 * group or in an autogroup
1497 * - the system default clamp value, defined by the sysadmin
1499 static inline struct uclamp_se
1500 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1502 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1503 struct uclamp_se uc_max = uclamp_default[clamp_id];
1505 /* System default restrictions always apply */
1506 if (unlikely(uc_req.value > uc_max.value))
1512 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1514 struct uclamp_se uc_eff;
1516 /* Task currently refcounted: use back-annotated (effective) value */
1517 if (p->uclamp[clamp_id].active)
1518 return (unsigned long)p->uclamp[clamp_id].value;
1520 uc_eff = uclamp_eff_get(p, clamp_id);
1522 return (unsigned long)uc_eff.value;
1526 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1527 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1528 * updates the rq's clamp value if required.
1530 * Tasks can have a task-specific value requested from user-space, track
1531 * within each bucket the maximum value for tasks refcounted in it.
1532 * This "local max aggregation" allows to track the exact "requested" value
1533 * for each bucket when all its RUNNABLE tasks require the same clamp.
1535 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1536 enum uclamp_id clamp_id)
1538 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1539 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1540 struct uclamp_bucket *bucket;
1542 lockdep_assert_rq_held(rq);
1544 /* Update task effective clamp */
1545 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1547 bucket = &uc_rq->bucket[uc_se->bucket_id];
1549 uc_se->active = true;
1551 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1554 * Local max aggregation: rq buckets always track the max
1555 * "requested" clamp value of its RUNNABLE tasks.
1557 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1558 bucket->value = uc_se->value;
1560 if (uc_se->value > READ_ONCE(uc_rq->value))
1561 WRITE_ONCE(uc_rq->value, uc_se->value);
1565 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1566 * is released. If this is the last task reference counting the rq's max
1567 * active clamp value, then the rq's clamp value is updated.
1569 * Both refcounted tasks and rq's cached clamp values are expected to be
1570 * always valid. If it's detected they are not, as defensive programming,
1571 * enforce the expected state and warn.
1573 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1574 enum uclamp_id clamp_id)
1576 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1577 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1578 struct uclamp_bucket *bucket;
1579 unsigned int bkt_clamp;
1580 unsigned int rq_clamp;
1582 lockdep_assert_rq_held(rq);
1585 * If sched_uclamp_used was enabled after task @p was enqueued,
1586 * we could end up with unbalanced call to uclamp_rq_dec_id().
1588 * In this case the uc_se->active flag should be false since no uclamp
1589 * accounting was performed at enqueue time and we can just return
1592 * Need to be careful of the following enqueue/dequeue ordering
1596 * // sched_uclamp_used gets enabled
1599 * // Must not decrement bucket->tasks here
1602 * where we could end up with stale data in uc_se and
1603 * bucket[uc_se->bucket_id].
1605 * The following check here eliminates the possibility of such race.
1607 if (unlikely(!uc_se->active))
1610 bucket = &uc_rq->bucket[uc_se->bucket_id];
1612 SCHED_WARN_ON(!bucket->tasks);
1613 if (likely(bucket->tasks))
1616 uc_se->active = false;
1619 * Keep "local max aggregation" simple and accept to (possibly)
1620 * overboost some RUNNABLE tasks in the same bucket.
1621 * The rq clamp bucket value is reset to its base value whenever
1622 * there are no more RUNNABLE tasks refcounting it.
1624 if (likely(bucket->tasks))
1627 rq_clamp = READ_ONCE(uc_rq->value);
1629 * Defensive programming: this should never happen. If it happens,
1630 * e.g. due to future modification, warn and fixup the expected value.
1632 SCHED_WARN_ON(bucket->value > rq_clamp);
1633 if (bucket->value >= rq_clamp) {
1634 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1635 WRITE_ONCE(uc_rq->value, bkt_clamp);
1639 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1641 enum uclamp_id clamp_id;
1644 * Avoid any overhead until uclamp is actually used by the userspace.
1646 * The condition is constructed such that a NOP is generated when
1647 * sched_uclamp_used is disabled.
1649 if (!static_branch_unlikely(&sched_uclamp_used))
1652 if (unlikely(!p->sched_class->uclamp_enabled))
1655 for_each_clamp_id(clamp_id)
1656 uclamp_rq_inc_id(rq, p, clamp_id);
1658 /* Reset clamp idle holding when there is one RUNNABLE task */
1659 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1660 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1663 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1665 enum uclamp_id clamp_id;
1668 * Avoid any overhead until uclamp is actually used by the userspace.
1670 * The condition is constructed such that a NOP is generated when
1671 * sched_uclamp_used is disabled.
1673 if (!static_branch_unlikely(&sched_uclamp_used))
1676 if (unlikely(!p->sched_class->uclamp_enabled))
1679 for_each_clamp_id(clamp_id)
1680 uclamp_rq_dec_id(rq, p, clamp_id);
1683 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1684 enum uclamp_id clamp_id)
1686 if (!p->uclamp[clamp_id].active)
1689 uclamp_rq_dec_id(rq, p, clamp_id);
1690 uclamp_rq_inc_id(rq, p, clamp_id);
1693 * Make sure to clear the idle flag if we've transiently reached 0
1694 * active tasks on rq.
1696 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1697 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1701 uclamp_update_active(struct task_struct *p)
1703 enum uclamp_id clamp_id;
1708 * Lock the task and the rq where the task is (or was) queued.
1710 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1711 * price to pay to safely serialize util_{min,max} updates with
1712 * enqueues, dequeues and migration operations.
1713 * This is the same locking schema used by __set_cpus_allowed_ptr().
1715 rq = task_rq_lock(p, &rf);
1718 * Setting the clamp bucket is serialized by task_rq_lock().
1719 * If the task is not yet RUNNABLE and its task_struct is not
1720 * affecting a valid clamp bucket, the next time it's enqueued,
1721 * it will already see the updated clamp bucket value.
1723 for_each_clamp_id(clamp_id)
1724 uclamp_rq_reinc_id(rq, p, clamp_id);
1726 task_rq_unlock(rq, p, &rf);
1729 #ifdef CONFIG_UCLAMP_TASK_GROUP
1731 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1733 struct css_task_iter it;
1734 struct task_struct *p;
1736 css_task_iter_start(css, 0, &it);
1737 while ((p = css_task_iter_next(&it)))
1738 uclamp_update_active(p);
1739 css_task_iter_end(&it);
1742 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1743 static void uclamp_update_root_tg(void)
1745 struct task_group *tg = &root_task_group;
1747 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1748 sysctl_sched_uclamp_util_min, false);
1749 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1750 sysctl_sched_uclamp_util_max, false);
1753 cpu_util_update_eff(&root_task_group.css);
1757 static void uclamp_update_root_tg(void) { }
1760 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1761 void *buffer, size_t *lenp, loff_t *ppos)
1763 bool update_root_tg = false;
1764 int old_min, old_max, old_min_rt;
1767 mutex_lock(&uclamp_mutex);
1768 old_min = sysctl_sched_uclamp_util_min;
1769 old_max = sysctl_sched_uclamp_util_max;
1770 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1772 result = proc_dointvec(table, write, buffer, lenp, ppos);
1778 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1779 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1780 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1786 if (old_min != sysctl_sched_uclamp_util_min) {
1787 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1788 sysctl_sched_uclamp_util_min, false);
1789 update_root_tg = true;
1791 if (old_max != sysctl_sched_uclamp_util_max) {
1792 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1793 sysctl_sched_uclamp_util_max, false);
1794 update_root_tg = true;
1797 if (update_root_tg) {
1798 static_branch_enable(&sched_uclamp_used);
1799 uclamp_update_root_tg();
1802 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1803 static_branch_enable(&sched_uclamp_used);
1804 uclamp_sync_util_min_rt_default();
1808 * We update all RUNNABLE tasks only when task groups are in use.
1809 * Otherwise, keep it simple and do just a lazy update at each next
1810 * task enqueue time.
1816 sysctl_sched_uclamp_util_min = old_min;
1817 sysctl_sched_uclamp_util_max = old_max;
1818 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1820 mutex_unlock(&uclamp_mutex);
1825 static int uclamp_validate(struct task_struct *p,
1826 const struct sched_attr *attr)
1828 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1829 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1831 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1832 util_min = attr->sched_util_min;
1834 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1838 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1839 util_max = attr->sched_util_max;
1841 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1845 if (util_min != -1 && util_max != -1 && util_min > util_max)
1849 * We have valid uclamp attributes; make sure uclamp is enabled.
1851 * We need to do that here, because enabling static branches is a
1852 * blocking operation which obviously cannot be done while holding
1855 static_branch_enable(&sched_uclamp_used);
1860 static bool uclamp_reset(const struct sched_attr *attr,
1861 enum uclamp_id clamp_id,
1862 struct uclamp_se *uc_se)
1864 /* Reset on sched class change for a non user-defined clamp value. */
1865 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1866 !uc_se->user_defined)
1869 /* Reset on sched_util_{min,max} == -1. */
1870 if (clamp_id == UCLAMP_MIN &&
1871 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1872 attr->sched_util_min == -1) {
1876 if (clamp_id == UCLAMP_MAX &&
1877 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1878 attr->sched_util_max == -1) {
1885 static void __setscheduler_uclamp(struct task_struct *p,
1886 const struct sched_attr *attr)
1888 enum uclamp_id clamp_id;
1890 for_each_clamp_id(clamp_id) {
1891 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1894 if (!uclamp_reset(attr, clamp_id, uc_se))
1898 * RT by default have a 100% boost value that could be modified
1901 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1902 value = sysctl_sched_uclamp_util_min_rt_default;
1904 value = uclamp_none(clamp_id);
1906 uclamp_se_set(uc_se, value, false);
1910 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1913 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1914 attr->sched_util_min != -1) {
1915 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1916 attr->sched_util_min, true);
1919 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1920 attr->sched_util_max != -1) {
1921 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1922 attr->sched_util_max, true);
1926 static void uclamp_fork(struct task_struct *p)
1928 enum uclamp_id clamp_id;
1931 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1932 * as the task is still at its early fork stages.
1934 for_each_clamp_id(clamp_id)
1935 p->uclamp[clamp_id].active = false;
1937 if (likely(!p->sched_reset_on_fork))
1940 for_each_clamp_id(clamp_id) {
1941 uclamp_se_set(&p->uclamp_req[clamp_id],
1942 uclamp_none(clamp_id), false);
1946 static void uclamp_post_fork(struct task_struct *p)
1948 uclamp_update_util_min_rt_default(p);
1951 static void __init init_uclamp_rq(struct rq *rq)
1953 enum uclamp_id clamp_id;
1954 struct uclamp_rq *uc_rq = rq->uclamp;
1956 for_each_clamp_id(clamp_id) {
1957 uc_rq[clamp_id] = (struct uclamp_rq) {
1958 .value = uclamp_none(clamp_id)
1962 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1965 static void __init init_uclamp(void)
1967 struct uclamp_se uc_max = {};
1968 enum uclamp_id clamp_id;
1971 for_each_possible_cpu(cpu)
1972 init_uclamp_rq(cpu_rq(cpu));
1974 for_each_clamp_id(clamp_id) {
1975 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1976 uclamp_none(clamp_id), false);
1979 /* System defaults allow max clamp values for both indexes */
1980 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1981 for_each_clamp_id(clamp_id) {
1982 uclamp_default[clamp_id] = uc_max;
1983 #ifdef CONFIG_UCLAMP_TASK_GROUP
1984 root_task_group.uclamp_req[clamp_id] = uc_max;
1985 root_task_group.uclamp[clamp_id] = uc_max;
1990 #else /* CONFIG_UCLAMP_TASK */
1991 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1992 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1993 static inline int uclamp_validate(struct task_struct *p,
1994 const struct sched_attr *attr)
1998 static void __setscheduler_uclamp(struct task_struct *p,
1999 const struct sched_attr *attr) { }
2000 static inline void uclamp_fork(struct task_struct *p) { }
2001 static inline void uclamp_post_fork(struct task_struct *p) { }
2002 static inline void init_uclamp(void) { }
2003 #endif /* CONFIG_UCLAMP_TASK */
2005 bool sched_task_on_rq(struct task_struct *p)
2007 return task_on_rq_queued(p);
2010 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2012 if (!(flags & ENQUEUE_NOCLOCK))
2013 update_rq_clock(rq);
2015 if (!(flags & ENQUEUE_RESTORE)) {
2016 sched_info_enqueue(rq, p);
2017 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2020 uclamp_rq_inc(rq, p);
2021 p->sched_class->enqueue_task(rq, p, flags);
2023 if (sched_core_enabled(rq))
2024 sched_core_enqueue(rq, p);
2027 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2029 if (sched_core_enabled(rq))
2030 sched_core_dequeue(rq, p);
2032 if (!(flags & DEQUEUE_NOCLOCK))
2033 update_rq_clock(rq);
2035 if (!(flags & DEQUEUE_SAVE)) {
2036 sched_info_dequeue(rq, p);
2037 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2040 uclamp_rq_dec(rq, p);
2041 p->sched_class->dequeue_task(rq, p, flags);
2044 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2046 enqueue_task(rq, p, flags);
2048 p->on_rq = TASK_ON_RQ_QUEUED;
2051 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2053 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2055 dequeue_task(rq, p, flags);
2058 static inline int __normal_prio(int policy, int rt_prio, int nice)
2062 if (dl_policy(policy))
2063 prio = MAX_DL_PRIO - 1;
2064 else if (rt_policy(policy))
2065 prio = MAX_RT_PRIO - 1 - rt_prio;
2067 prio = NICE_TO_PRIO(nice);
2073 * Calculate the expected normal priority: i.e. priority
2074 * without taking RT-inheritance into account. Might be
2075 * boosted by interactivity modifiers. Changes upon fork,
2076 * setprio syscalls, and whenever the interactivity
2077 * estimator recalculates.
2079 static inline int normal_prio(struct task_struct *p)
2081 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2085 * Calculate the current priority, i.e. the priority
2086 * taken into account by the scheduler. This value might
2087 * be boosted by RT tasks, or might be boosted by
2088 * interactivity modifiers. Will be RT if the task got
2089 * RT-boosted. If not then it returns p->normal_prio.
2091 static int effective_prio(struct task_struct *p)
2093 p->normal_prio = normal_prio(p);
2095 * If we are RT tasks or we were boosted to RT priority,
2096 * keep the priority unchanged. Otherwise, update priority
2097 * to the normal priority:
2099 if (!rt_prio(p->prio))
2100 return p->normal_prio;
2105 * task_curr - is this task currently executing on a CPU?
2106 * @p: the task in question.
2108 * Return: 1 if the task is currently executing. 0 otherwise.
2110 inline int task_curr(const struct task_struct *p)
2112 return cpu_curr(task_cpu(p)) == p;
2116 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2117 * use the balance_callback list if you want balancing.
2119 * this means any call to check_class_changed() must be followed by a call to
2120 * balance_callback().
2122 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2123 const struct sched_class *prev_class,
2126 if (prev_class != p->sched_class) {
2127 if (prev_class->switched_from)
2128 prev_class->switched_from(rq, p);
2130 p->sched_class->switched_to(rq, p);
2131 } else if (oldprio != p->prio || dl_task(p))
2132 p->sched_class->prio_changed(rq, p, oldprio);
2135 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2137 if (p->sched_class == rq->curr->sched_class)
2138 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2139 else if (p->sched_class > rq->curr->sched_class)
2143 * A queue event has occurred, and we're going to schedule. In
2144 * this case, we can save a useless back to back clock update.
2146 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2147 rq_clock_skip_update(rq);
2153 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2155 static int __set_cpus_allowed_ptr(struct task_struct *p,
2156 const struct cpumask *new_mask,
2159 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2161 if (likely(!p->migration_disabled))
2164 if (p->cpus_ptr != &p->cpus_mask)
2168 * Violates locking rules! see comment in __do_set_cpus_allowed().
2170 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2173 void migrate_disable(void)
2175 struct task_struct *p = current;
2177 if (p->migration_disabled) {
2178 p->migration_disabled++;
2183 this_rq()->nr_pinned++;
2184 p->migration_disabled = 1;
2185 preempt_lazy_disable();
2188 EXPORT_SYMBOL_GPL(migrate_disable);
2190 void migrate_enable(void)
2192 struct task_struct *p = current;
2194 if (p->migration_disabled > 1) {
2195 p->migration_disabled--;
2197 } else if (WARN_ON_ONCE(p->migration_disabled == 0)) {
2202 * Ensure stop_task runs either before or after this, and that
2203 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2206 if (p->cpus_ptr != &p->cpus_mask)
2207 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2209 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2210 * regular cpus_mask, otherwise things that race (eg.
2211 * select_fallback_rq) get confused.
2214 p->migration_disabled = 0;
2215 this_rq()->nr_pinned--;
2216 preempt_lazy_enable();
2219 EXPORT_SYMBOL_GPL(migrate_enable);
2221 static inline bool rq_has_pinned_tasks(struct rq *rq)
2223 return rq->nr_pinned;
2227 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2228 * __set_cpus_allowed_ptr() and select_fallback_rq().
2230 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2232 /* When not in the task's cpumask, no point in looking further. */
2233 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2236 /* migrate_disabled() must be allowed to finish. */
2237 if (is_migration_disabled(p))
2238 return cpu_online(cpu);
2240 /* Non kernel threads are not allowed during either online or offline. */
2241 if (!(p->flags & PF_KTHREAD))
2242 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2244 /* KTHREAD_IS_PER_CPU is always allowed. */
2245 if (kthread_is_per_cpu(p))
2246 return cpu_online(cpu);
2248 /* Regular kernel threads don't get to stay during offline. */
2252 /* But are allowed during online. */
2253 return cpu_online(cpu);
2257 * This is how migration works:
2259 * 1) we invoke migration_cpu_stop() on the target CPU using
2261 * 2) stopper starts to run (implicitly forcing the migrated thread
2263 * 3) it checks whether the migrated task is still in the wrong runqueue.
2264 * 4) if it's in the wrong runqueue then the migration thread removes
2265 * it and puts it into the right queue.
2266 * 5) stopper completes and stop_one_cpu() returns and the migration
2271 * move_queued_task - move a queued task to new rq.
2273 * Returns (locked) new rq. Old rq's lock is released.
2275 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2276 struct task_struct *p, int new_cpu)
2278 lockdep_assert_rq_held(rq);
2280 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2281 set_task_cpu(p, new_cpu);
2284 rq = cpu_rq(new_cpu);
2287 BUG_ON(task_cpu(p) != new_cpu);
2288 activate_task(rq, p, 0);
2289 check_preempt_curr(rq, p, 0);
2294 struct migration_arg {
2295 struct task_struct *task;
2297 struct set_affinity_pending *pending;
2301 * @refs: number of wait_for_completion()
2302 * @stop_pending: is @stop_work in use
2304 struct set_affinity_pending {
2306 unsigned int stop_pending;
2307 struct completion done;
2308 struct cpu_stop_work stop_work;
2309 struct migration_arg arg;
2313 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2314 * this because either it can't run here any more (set_cpus_allowed()
2315 * away from this CPU, or CPU going down), or because we're
2316 * attempting to rebalance this task on exec (sched_exec).
2318 * So we race with normal scheduler movements, but that's OK, as long
2319 * as the task is no longer on this CPU.
2321 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2322 struct task_struct *p, int dest_cpu)
2324 /* Affinity changed (again). */
2325 if (!is_cpu_allowed(p, dest_cpu))
2328 update_rq_clock(rq);
2329 rq = move_queued_task(rq, rf, p, dest_cpu);
2335 * migration_cpu_stop - this will be executed by a highprio stopper thread
2336 * and performs thread migration by bumping thread off CPU then
2337 * 'pushing' onto another runqueue.
2339 static int migration_cpu_stop(void *data)
2341 struct migration_arg *arg = data;
2342 struct set_affinity_pending *pending = arg->pending;
2343 struct task_struct *p = arg->task;
2344 struct rq *rq = this_rq();
2345 bool complete = false;
2349 * The original target CPU might have gone down and we might
2350 * be on another CPU but it doesn't matter.
2352 local_irq_save(rf.flags);
2354 * We need to explicitly wake pending tasks before running
2355 * __migrate_task() such that we will not miss enforcing cpus_ptr
2356 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2358 flush_smp_call_function_from_idle();
2360 raw_spin_lock(&p->pi_lock);
2364 * If we were passed a pending, then ->stop_pending was set, thus
2365 * p->migration_pending must have remained stable.
2367 WARN_ON_ONCE(pending && pending != p->migration_pending);
2370 * If task_rq(p) != rq, it cannot be migrated here, because we're
2371 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2372 * we're holding p->pi_lock.
2374 if (task_rq(p) == rq) {
2375 if (is_migration_disabled(p))
2379 p->migration_pending = NULL;
2382 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2386 if (task_on_rq_queued(p))
2387 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2389 p->wake_cpu = arg->dest_cpu;
2392 * XXX __migrate_task() can fail, at which point we might end
2393 * up running on a dodgy CPU, AFAICT this can only happen
2394 * during CPU hotplug, at which point we'll get pushed out
2395 * anyway, so it's probably not a big deal.
2398 } else if (pending) {
2400 * This happens when we get migrated between migrate_enable()'s
2401 * preempt_enable() and scheduling the stopper task. At that
2402 * point we're a regular task again and not current anymore.
2404 * A !PREEMPT kernel has a giant hole here, which makes it far
2409 * The task moved before the stopper got to run. We're holding
2410 * ->pi_lock, so the allowed mask is stable - if it got
2411 * somewhere allowed, we're done.
2413 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2414 p->migration_pending = NULL;
2420 * When migrate_enable() hits a rq mis-match we can't reliably
2421 * determine is_migration_disabled() and so have to chase after
2424 WARN_ON_ONCE(!pending->stop_pending);
2425 task_rq_unlock(rq, p, &rf);
2426 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2427 &pending->arg, &pending->stop_work);
2432 pending->stop_pending = false;
2433 task_rq_unlock(rq, p, &rf);
2436 complete_all(&pending->done);
2441 int push_cpu_stop(void *arg)
2443 struct rq *lowest_rq = NULL, *rq = this_rq();
2444 struct task_struct *p = arg;
2446 raw_spin_lock_irq(&p->pi_lock);
2447 raw_spin_rq_lock(rq);
2449 if (task_rq(p) != rq)
2452 if (is_migration_disabled(p)) {
2453 p->migration_flags |= MDF_PUSH;
2457 p->migration_flags &= ~MDF_PUSH;
2459 if (p->sched_class->find_lock_rq)
2460 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2465 // XXX validate p is still the highest prio task
2466 if (task_rq(p) == rq) {
2467 deactivate_task(rq, p, 0);
2468 set_task_cpu(p, lowest_rq->cpu);
2469 activate_task(lowest_rq, p, 0);
2470 resched_curr(lowest_rq);
2473 double_unlock_balance(rq, lowest_rq);
2476 rq->push_busy = false;
2477 raw_spin_rq_unlock(rq);
2478 raw_spin_unlock_irq(&p->pi_lock);
2485 * sched_class::set_cpus_allowed must do the below, but is not required to
2486 * actually call this function.
2488 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2490 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2491 p->cpus_ptr = new_mask;
2495 cpumask_copy(&p->cpus_mask, new_mask);
2496 p->nr_cpus_allowed = cpumask_weight(new_mask);
2500 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2502 struct rq *rq = task_rq(p);
2503 bool queued, running;
2506 * This here violates the locking rules for affinity, since we're only
2507 * supposed to change these variables while holding both rq->lock and
2510 * HOWEVER, it magically works, because ttwu() is the only code that
2511 * accesses these variables under p->pi_lock and only does so after
2512 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2513 * before finish_task().
2515 * XXX do further audits, this smells like something putrid.
2517 if (flags & SCA_MIGRATE_DISABLE)
2518 SCHED_WARN_ON(!p->on_cpu);
2520 lockdep_assert_held(&p->pi_lock);
2522 queued = task_on_rq_queued(p);
2523 running = task_current(rq, p);
2527 * Because __kthread_bind() calls this on blocked tasks without
2530 lockdep_assert_rq_held(rq);
2531 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2534 put_prev_task(rq, p);
2536 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2539 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2541 set_next_task(rq, p);
2544 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2546 __do_set_cpus_allowed(p, new_mask, 0);
2549 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2552 cpumask_t *user_mask;
2553 unsigned long flags;
2556 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2557 * may differ by now due to racing.
2559 dst->user_cpus_ptr = NULL;
2562 * This check is racy and losing the race is a valid situation.
2563 * It is not worth the extra overhead of taking the pi_lock on
2566 if (data_race(!src->user_cpus_ptr))
2569 user_mask = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2574 * Use pi_lock to protect content of user_cpus_ptr
2576 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2577 * do_set_cpus_allowed().
2579 raw_spin_lock_irqsave(&src->pi_lock, flags);
2580 if (src->user_cpus_ptr) {
2581 swap(dst->user_cpus_ptr, user_mask);
2582 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2584 raw_spin_unlock_irqrestore(&src->pi_lock, flags);
2586 if (unlikely(user_mask))
2592 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2594 struct cpumask *user_mask = NULL;
2596 swap(p->user_cpus_ptr, user_mask);
2601 void release_user_cpus_ptr(struct task_struct *p)
2603 kfree(clear_user_cpus_ptr(p));
2607 * This function is wildly self concurrent; here be dragons.
2610 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2611 * designated task is enqueued on an allowed CPU. If that task is currently
2612 * running, we have to kick it out using the CPU stopper.
2614 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2617 * Initial conditions: P0->cpus_mask = [0, 1]
2621 * migrate_disable();
2623 * set_cpus_allowed_ptr(P0, [1]);
2625 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2626 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2627 * This means we need the following scheme:
2631 * migrate_disable();
2633 * set_cpus_allowed_ptr(P0, [1]);
2637 * __set_cpus_allowed_ptr();
2638 * <wakes local stopper>
2639 * `--> <woken on migration completion>
2641 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2642 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2643 * task p are serialized by p->pi_lock, which we can leverage: the one that
2644 * should come into effect at the end of the Migrate-Disable region is the last
2645 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2646 * but we still need to properly signal those waiting tasks at the appropriate
2649 * This is implemented using struct set_affinity_pending. The first
2650 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2651 * setup an instance of that struct and install it on the targeted task_struct.
2652 * Any and all further callers will reuse that instance. Those then wait for
2653 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2654 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2657 * (1) In the cases covered above. There is one more where the completion is
2658 * signaled within affine_move_task() itself: when a subsequent affinity request
2659 * occurs after the stopper bailed out due to the targeted task still being
2660 * Migrate-Disable. Consider:
2662 * Initial conditions: P0->cpus_mask = [0, 1]
2666 * migrate_disable();
2668 * set_cpus_allowed_ptr(P0, [1]);
2671 * migration_cpu_stop()
2672 * is_migration_disabled()
2674 * set_cpus_allowed_ptr(P0, [0, 1]);
2675 * <signal completion>
2678 * Note that the above is safe vs a concurrent migrate_enable(), as any
2679 * pending affinity completion is preceded by an uninstallation of
2680 * p->migration_pending done with p->pi_lock held.
2682 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2683 int dest_cpu, unsigned int flags)
2685 struct set_affinity_pending my_pending = { }, *pending = NULL;
2686 bool stop_pending, complete = false;
2688 /* Can the task run on the task's current CPU? If so, we're done */
2689 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2690 struct task_struct *push_task = NULL;
2692 if ((flags & SCA_MIGRATE_ENABLE) &&
2693 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2694 rq->push_busy = true;
2695 push_task = get_task_struct(p);
2699 * If there are pending waiters, but no pending stop_work,
2700 * then complete now.
2702 pending = p->migration_pending;
2703 if (pending && !pending->stop_pending) {
2704 p->migration_pending = NULL;
2708 task_rq_unlock(rq, p, rf);
2711 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2716 complete_all(&pending->done);
2721 if (!(flags & SCA_MIGRATE_ENABLE)) {
2722 /* serialized by p->pi_lock */
2723 if (!p->migration_pending) {
2724 /* Install the request */
2725 refcount_set(&my_pending.refs, 1);
2726 init_completion(&my_pending.done);
2727 my_pending.arg = (struct migration_arg) {
2729 .dest_cpu = dest_cpu,
2730 .pending = &my_pending,
2733 p->migration_pending = &my_pending;
2735 pending = p->migration_pending;
2736 refcount_inc(&pending->refs);
2738 * Affinity has changed, but we've already installed a
2739 * pending. migration_cpu_stop() *must* see this, else
2740 * we risk a completion of the pending despite having a
2741 * task on a disallowed CPU.
2743 * Serialized by p->pi_lock, so this is safe.
2745 pending->arg.dest_cpu = dest_cpu;
2748 pending = p->migration_pending;
2750 * - !MIGRATE_ENABLE:
2751 * we'll have installed a pending if there wasn't one already.
2754 * we're here because the current CPU isn't matching anymore,
2755 * the only way that can happen is because of a concurrent
2756 * set_cpus_allowed_ptr() call, which should then still be
2757 * pending completion.
2759 * Either way, we really should have a @pending here.
2761 if (WARN_ON_ONCE(!pending)) {
2762 task_rq_unlock(rq, p, rf);
2766 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2768 * MIGRATE_ENABLE gets here because 'p == current', but for
2769 * anything else we cannot do is_migration_disabled(), punt
2770 * and have the stopper function handle it all race-free.
2772 stop_pending = pending->stop_pending;
2774 pending->stop_pending = true;
2776 if (flags & SCA_MIGRATE_ENABLE)
2777 p->migration_flags &= ~MDF_PUSH;
2779 task_rq_unlock(rq, p, rf);
2781 if (!stop_pending) {
2782 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2783 &pending->arg, &pending->stop_work);
2786 if (flags & SCA_MIGRATE_ENABLE)
2790 if (!is_migration_disabled(p)) {
2791 if (task_on_rq_queued(p))
2792 rq = move_queued_task(rq, rf, p, dest_cpu);
2794 if (!pending->stop_pending) {
2795 p->migration_pending = NULL;
2799 task_rq_unlock(rq, p, rf);
2802 complete_all(&pending->done);
2805 wait_for_completion(&pending->done);
2807 if (refcount_dec_and_test(&pending->refs))
2808 wake_up_var(&pending->refs); /* No UaF, just an address */
2811 * Block the original owner of &pending until all subsequent callers
2812 * have seen the completion and decremented the refcount
2814 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2817 WARN_ON_ONCE(my_pending.stop_pending);
2823 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2825 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2826 const struct cpumask *new_mask,
2829 struct rq_flags *rf)
2830 __releases(rq->lock)
2831 __releases(p->pi_lock)
2833 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2834 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2835 bool kthread = p->flags & PF_KTHREAD;
2836 struct cpumask *user_mask = NULL;
2837 unsigned int dest_cpu;
2840 update_rq_clock(rq);
2842 if (kthread || is_migration_disabled(p)) {
2844 * Kernel threads are allowed on online && !active CPUs,
2845 * however, during cpu-hot-unplug, even these might get pushed
2846 * away if not KTHREAD_IS_PER_CPU.
2848 * Specifically, migration_disabled() tasks must not fail the
2849 * cpumask_any_and_distribute() pick below, esp. so on
2850 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2851 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2853 cpu_valid_mask = cpu_online_mask;
2856 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2862 * Must re-check here, to close a race against __kthread_bind(),
2863 * sched_setaffinity() is not guaranteed to observe the flag.
2865 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2870 if (!(flags & SCA_MIGRATE_ENABLE)) {
2871 if (cpumask_equal(&p->cpus_mask, new_mask))
2874 if (WARN_ON_ONCE(p == current &&
2875 is_migration_disabled(p) &&
2876 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2883 * Picking a ~random cpu helps in cases where we are changing affinity
2884 * for groups of tasks (ie. cpuset), so that load balancing is not
2885 * immediately required to distribute the tasks within their new mask.
2887 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2888 if (dest_cpu >= nr_cpu_ids) {
2893 __do_set_cpus_allowed(p, new_mask, flags);
2895 if (flags & SCA_USER)
2896 user_mask = clear_user_cpus_ptr(p);
2898 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2905 task_rq_unlock(rq, p, rf);
2911 * Change a given task's CPU affinity. Migrate the thread to a
2912 * proper CPU and schedule it away if the CPU it's executing on
2913 * is removed from the allowed bitmask.
2915 * NOTE: the caller must have a valid reference to the task, the
2916 * task must not exit() & deallocate itself prematurely. The
2917 * call is not atomic; no spinlocks may be held.
2919 static int __set_cpus_allowed_ptr(struct task_struct *p,
2920 const struct cpumask *new_mask, u32 flags)
2925 rq = task_rq_lock(p, &rf);
2926 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2929 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2931 return __set_cpus_allowed_ptr(p, new_mask, 0);
2933 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2936 * Change a given task's CPU affinity to the intersection of its current
2937 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2938 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2939 * If the resulting mask is empty, leave the affinity unchanged and return
2942 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2943 struct cpumask *new_mask,
2944 const struct cpumask *subset_mask)
2946 struct cpumask *user_mask = NULL;
2951 if (!p->user_cpus_ptr) {
2952 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2957 rq = task_rq_lock(p, &rf);
2960 * Forcefully restricting the affinity of a deadline task is
2961 * likely to cause problems, so fail and noisily override the
2964 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2969 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2975 * We're about to butcher the task affinity, so keep track of what
2976 * the user asked for in case we're able to restore it later on.
2979 cpumask_copy(user_mask, p->cpus_ptr);
2980 p->user_cpus_ptr = user_mask;
2983 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2986 task_rq_unlock(rq, p, &rf);
2992 * Restrict the CPU affinity of task @p so that it is a subset of
2993 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2994 * old affinity mask. If the resulting mask is empty, we warn and walk
2995 * up the cpuset hierarchy until we find a suitable mask.
2997 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2999 cpumask_var_t new_mask;
3000 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3002 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3005 * __migrate_task() can fail silently in the face of concurrent
3006 * offlining of the chosen destination CPU, so take the hotplug
3007 * lock to ensure that the migration succeeds.
3010 if (!cpumask_available(new_mask))
3013 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3017 * We failed to find a valid subset of the affinity mask for the
3018 * task, so override it based on its cpuset hierarchy.
3020 cpuset_cpus_allowed(p, new_mask);
3021 override_mask = new_mask;
3024 if (printk_ratelimit()) {
3025 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3026 task_pid_nr(p), p->comm,
3027 cpumask_pr_args(override_mask));
3030 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3033 free_cpumask_var(new_mask);
3037 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3040 * Restore the affinity of a task @p which was previously restricted by a
3041 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3042 * @p->user_cpus_ptr.
3044 * It is the caller's responsibility to serialise this with any calls to
3045 * force_compatible_cpus_allowed_ptr(@p).
3047 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3049 struct cpumask *user_mask = p->user_cpus_ptr;
3050 unsigned long flags;
3053 * Try to restore the old affinity mask. If this fails, then
3054 * we free the mask explicitly to avoid it being inherited across
3055 * a subsequent fork().
3057 if (!user_mask || !__sched_setaffinity(p, user_mask))
3060 raw_spin_lock_irqsave(&p->pi_lock, flags);
3061 user_mask = clear_user_cpus_ptr(p);
3062 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3067 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3069 #ifdef CONFIG_SCHED_DEBUG
3070 unsigned int state = READ_ONCE(p->__state);
3073 * We should never call set_task_cpu() on a blocked task,
3074 * ttwu() will sort out the placement.
3076 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3079 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3080 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3081 * time relying on p->on_rq.
3083 WARN_ON_ONCE(state == TASK_RUNNING &&
3084 p->sched_class == &fair_sched_class &&
3085 (p->on_rq && !task_on_rq_migrating(p)));
3087 #ifdef CONFIG_LOCKDEP
3089 * The caller should hold either p->pi_lock or rq->lock, when changing
3090 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3092 * sched_move_task() holds both and thus holding either pins the cgroup,
3095 * Furthermore, all task_rq users should acquire both locks, see
3098 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3099 lockdep_is_held(__rq_lockp(task_rq(p)))));
3102 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3104 WARN_ON_ONCE(!cpu_online(new_cpu));
3106 WARN_ON_ONCE(is_migration_disabled(p));
3109 trace_sched_migrate_task(p, new_cpu);
3111 if (task_cpu(p) != new_cpu) {
3112 if (p->sched_class->migrate_task_rq)
3113 p->sched_class->migrate_task_rq(p, new_cpu);
3114 p->se.nr_migrations++;
3116 perf_event_task_migrate(p);
3119 __set_task_cpu(p, new_cpu);
3122 #ifdef CONFIG_NUMA_BALANCING
3123 static void __migrate_swap_task(struct task_struct *p, int cpu)
3125 if (task_on_rq_queued(p)) {
3126 struct rq *src_rq, *dst_rq;
3127 struct rq_flags srf, drf;
3129 src_rq = task_rq(p);
3130 dst_rq = cpu_rq(cpu);
3132 rq_pin_lock(src_rq, &srf);
3133 rq_pin_lock(dst_rq, &drf);
3135 deactivate_task(src_rq, p, 0);
3136 set_task_cpu(p, cpu);
3137 activate_task(dst_rq, p, 0);
3138 check_preempt_curr(dst_rq, p, 0);
3140 rq_unpin_lock(dst_rq, &drf);
3141 rq_unpin_lock(src_rq, &srf);
3145 * Task isn't running anymore; make it appear like we migrated
3146 * it before it went to sleep. This means on wakeup we make the
3147 * previous CPU our target instead of where it really is.
3153 struct migration_swap_arg {
3154 struct task_struct *src_task, *dst_task;
3155 int src_cpu, dst_cpu;
3158 static int migrate_swap_stop(void *data)
3160 struct migration_swap_arg *arg = data;
3161 struct rq *src_rq, *dst_rq;
3164 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3167 src_rq = cpu_rq(arg->src_cpu);
3168 dst_rq = cpu_rq(arg->dst_cpu);
3170 double_raw_lock(&arg->src_task->pi_lock,
3171 &arg->dst_task->pi_lock);
3172 double_rq_lock(src_rq, dst_rq);
3174 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3177 if (task_cpu(arg->src_task) != arg->src_cpu)
3180 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3183 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3186 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3187 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3192 double_rq_unlock(src_rq, dst_rq);
3193 raw_spin_unlock(&arg->dst_task->pi_lock);
3194 raw_spin_unlock(&arg->src_task->pi_lock);
3200 * Cross migrate two tasks
3202 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3203 int target_cpu, int curr_cpu)
3205 struct migration_swap_arg arg;
3208 arg = (struct migration_swap_arg){
3210 .src_cpu = curr_cpu,
3212 .dst_cpu = target_cpu,
3215 if (arg.src_cpu == arg.dst_cpu)
3219 * These three tests are all lockless; this is OK since all of them
3220 * will be re-checked with proper locks held further down the line.
3222 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3225 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3228 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3231 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3232 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3237 #endif /* CONFIG_NUMA_BALANCING */
3240 * wait_task_inactive - wait for a thread to unschedule.
3242 * If @match_state is nonzero, it's the @p->state value just checked and
3243 * not expected to change. If it changes, i.e. @p might have woken up,
3244 * then return zero. When we succeed in waiting for @p to be off its CPU,
3245 * we return a positive number (its total switch count). If a second call
3246 * a short while later returns the same number, the caller can be sure that
3247 * @p has remained unscheduled the whole time.
3249 * The caller must ensure that the task *will* unschedule sometime soon,
3250 * else this function might spin for a *long* time. This function can't
3251 * be called with interrupts off, or it may introduce deadlock with
3252 * smp_call_function() if an IPI is sent by the same process we are
3253 * waiting to become inactive.
3255 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3257 int running, queued;
3264 * We do the initial early heuristics without holding
3265 * any task-queue locks at all. We'll only try to get
3266 * the runqueue lock when things look like they will
3272 * If the task is actively running on another CPU
3273 * still, just relax and busy-wait without holding
3276 * NOTE! Since we don't hold any locks, it's not
3277 * even sure that "rq" stays as the right runqueue!
3278 * But we don't care, since "task_running()" will
3279 * return false if the runqueue has changed and p
3280 * is actually now running somewhere else!
3282 while (task_running(rq, p)) {
3283 if (match_state && !task_match_state_lock(p, match_state))
3289 * Ok, time to look more closely! We need the rq
3290 * lock now, to be *sure*. If we're wrong, we'll
3291 * just go back and repeat.
3293 rq = task_rq_lock(p, &rf);
3294 trace_sched_wait_task(p);
3295 running = task_running(rq, p);
3296 queued = task_on_rq_queued(p);
3298 if (!match_state || task_match_state_or_saved(p, match_state))
3299 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3300 task_rq_unlock(rq, p, &rf);
3303 * If it changed from the expected state, bail out now.
3305 if (unlikely(!ncsw))
3309 * Was it really running after all now that we
3310 * checked with the proper locks actually held?
3312 * Oops. Go back and try again..
3314 if (unlikely(running)) {
3320 * It's not enough that it's not actively running,
3321 * it must be off the runqueue _entirely_, and not
3324 * So if it was still runnable (but just not actively
3325 * running right now), it's preempted, and we should
3326 * yield - it could be a while.
3328 if (unlikely(queued)) {
3329 ktime_t to = NSEC_PER_SEC / HZ;
3331 set_current_state(TASK_UNINTERRUPTIBLE);
3332 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3337 * Ahh, all good. It wasn't running, and it wasn't
3338 * runnable, which means that it will never become
3339 * running in the future either. We're all done!
3348 * kick_process - kick a running thread to enter/exit the kernel
3349 * @p: the to-be-kicked thread
3351 * Cause a process which is running on another CPU to enter
3352 * kernel-mode, without any delay. (to get signals handled.)
3354 * NOTE: this function doesn't have to take the runqueue lock,
3355 * because all it wants to ensure is that the remote task enters
3356 * the kernel. If the IPI races and the task has been migrated
3357 * to another CPU then no harm is done and the purpose has been
3360 void kick_process(struct task_struct *p)
3366 if ((cpu != smp_processor_id()) && task_curr(p))
3367 smp_send_reschedule(cpu);
3370 EXPORT_SYMBOL_GPL(kick_process);
3373 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3375 * A few notes on cpu_active vs cpu_online:
3377 * - cpu_active must be a subset of cpu_online
3379 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3380 * see __set_cpus_allowed_ptr(). At this point the newly online
3381 * CPU isn't yet part of the sched domains, and balancing will not
3384 * - on CPU-down we clear cpu_active() to mask the sched domains and
3385 * avoid the load balancer to place new tasks on the to be removed
3386 * CPU. Existing tasks will remain running there and will be taken
3389 * This means that fallback selection must not select !active CPUs.
3390 * And can assume that any active CPU must be online. Conversely
3391 * select_task_rq() below may allow selection of !active CPUs in order
3392 * to satisfy the above rules.
3394 static int select_fallback_rq(int cpu, struct task_struct *p)
3396 int nid = cpu_to_node(cpu);
3397 const struct cpumask *nodemask = NULL;
3398 enum { cpuset, possible, fail } state = cpuset;
3402 * If the node that the CPU is on has been offlined, cpu_to_node()
3403 * will return -1. There is no CPU on the node, and we should
3404 * select the CPU on the other node.
3407 nodemask = cpumask_of_node(nid);
3409 /* Look for allowed, online CPU in same node. */
3410 for_each_cpu(dest_cpu, nodemask) {
3411 if (is_cpu_allowed(p, dest_cpu))
3417 /* Any allowed, online CPU? */
3418 for_each_cpu(dest_cpu, p->cpus_ptr) {
3419 if (!is_cpu_allowed(p, dest_cpu))
3425 /* No more Mr. Nice Guy. */
3428 if (cpuset_cpus_allowed_fallback(p)) {
3435 * XXX When called from select_task_rq() we only
3436 * hold p->pi_lock and again violate locking order.
3438 * More yuck to audit.
3440 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3450 if (state != cpuset) {
3452 * Don't tell them about moving exiting tasks or
3453 * kernel threads (both mm NULL), since they never
3456 if (p->mm && printk_ratelimit()) {
3457 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3458 task_pid_nr(p), p->comm, cpu);
3466 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3469 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3471 lockdep_assert_held(&p->pi_lock);
3473 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3474 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3476 cpu = cpumask_any(p->cpus_ptr);
3479 * In order not to call set_task_cpu() on a blocking task we need
3480 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3483 * Since this is common to all placement strategies, this lives here.
3485 * [ this allows ->select_task() to simply return task_cpu(p) and
3486 * not worry about this generic constraint ]
3488 if (unlikely(!is_cpu_allowed(p, cpu)))
3489 cpu = select_fallback_rq(task_cpu(p), p);
3494 void sched_set_stop_task(int cpu, struct task_struct *stop)
3496 static struct lock_class_key stop_pi_lock;
3497 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3498 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3502 * Make it appear like a SCHED_FIFO task, its something
3503 * userspace knows about and won't get confused about.
3505 * Also, it will make PI more or less work without too
3506 * much confusion -- but then, stop work should not
3507 * rely on PI working anyway.
3509 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3511 stop->sched_class = &stop_sched_class;
3514 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3515 * adjust the effective priority of a task. As a result,
3516 * rt_mutex_setprio() can trigger (RT) balancing operations,
3517 * which can then trigger wakeups of the stop thread to push
3518 * around the current task.
3520 * The stop task itself will never be part of the PI-chain, it
3521 * never blocks, therefore that ->pi_lock recursion is safe.
3522 * Tell lockdep about this by placing the stop->pi_lock in its
3525 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3528 cpu_rq(cpu)->stop = stop;
3532 * Reset it back to a normal scheduling class so that
3533 * it can die in pieces.
3535 old_stop->sched_class = &rt_sched_class;
3539 #else /* CONFIG_SMP */
3541 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3542 const struct cpumask *new_mask,
3545 return set_cpus_allowed_ptr(p, new_mask);
3548 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3550 static inline bool rq_has_pinned_tasks(struct rq *rq)
3555 #endif /* !CONFIG_SMP */
3558 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3562 if (!schedstat_enabled())
3568 if (cpu == rq->cpu) {
3569 __schedstat_inc(rq->ttwu_local);
3570 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3572 struct sched_domain *sd;
3574 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3576 for_each_domain(rq->cpu, sd) {
3577 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3578 __schedstat_inc(sd->ttwu_wake_remote);
3585 if (wake_flags & WF_MIGRATED)
3586 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3587 #endif /* CONFIG_SMP */
3589 __schedstat_inc(rq->ttwu_count);
3590 __schedstat_inc(p->se.statistics.nr_wakeups);
3592 if (wake_flags & WF_SYNC)
3593 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3597 * Mark the task runnable and perform wakeup-preemption.
3599 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3600 struct rq_flags *rf)
3602 check_preempt_curr(rq, p, wake_flags);
3603 WRITE_ONCE(p->__state, TASK_RUNNING);
3604 trace_sched_wakeup(p);
3607 if (p->sched_class->task_woken) {
3609 * Our task @p is fully woken up and running; so it's safe to
3610 * drop the rq->lock, hereafter rq is only used for statistics.
3612 rq_unpin_lock(rq, rf);
3613 p->sched_class->task_woken(rq, p);
3614 rq_repin_lock(rq, rf);
3617 if (rq->idle_stamp) {
3618 u64 delta = rq_clock(rq) - rq->idle_stamp;
3619 u64 max = 2*rq->max_idle_balance_cost;
3621 update_avg(&rq->avg_idle, delta);
3623 if (rq->avg_idle > max)
3626 rq->wake_stamp = jiffies;
3627 rq->wake_avg_idle = rq->avg_idle / 2;
3635 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3636 struct rq_flags *rf)
3638 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3640 lockdep_assert_rq_held(rq);
3642 if (p->sched_contributes_to_load)
3643 rq->nr_uninterruptible--;
3646 if (wake_flags & WF_MIGRATED)
3647 en_flags |= ENQUEUE_MIGRATED;
3651 delayacct_blkio_end(p);
3652 atomic_dec(&task_rq(p)->nr_iowait);
3655 activate_task(rq, p, en_flags);
3656 ttwu_do_wakeup(rq, p, wake_flags, rf);
3660 * Consider @p being inside a wait loop:
3663 * set_current_state(TASK_UNINTERRUPTIBLE);
3670 * __set_current_state(TASK_RUNNING);
3672 * between set_current_state() and schedule(). In this case @p is still
3673 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3676 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3677 * then schedule() must still happen and p->state can be changed to
3678 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3679 * need to do a full wakeup with enqueue.
3681 * Returns: %true when the wakeup is done,
3684 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3690 rq = __task_rq_lock(p, &rf);
3691 if (task_on_rq_queued(p)) {
3692 /* check_preempt_curr() may use rq clock */
3693 update_rq_clock(rq);
3694 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3697 __task_rq_unlock(rq, &rf);
3703 void sched_ttwu_pending(void *arg)
3705 struct llist_node *llist = arg;
3706 struct rq *rq = this_rq();
3707 struct task_struct *p, *t;
3714 * rq::ttwu_pending racy indication of out-standing wakeups.
3715 * Races such that false-negatives are possible, since they
3716 * are shorter lived that false-positives would be.
3718 WRITE_ONCE(rq->ttwu_pending, 0);
3720 rq_lock_irqsave(rq, &rf);
3721 update_rq_clock(rq);
3723 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3724 if (WARN_ON_ONCE(p->on_cpu))
3725 smp_cond_load_acquire(&p->on_cpu, !VAL);
3727 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3728 set_task_cpu(p, cpu_of(rq));
3730 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3733 rq_unlock_irqrestore(rq, &rf);
3736 void send_call_function_single_ipi(int cpu)
3738 struct rq *rq = cpu_rq(cpu);
3740 if (!set_nr_if_polling(rq->idle))
3741 arch_send_call_function_single_ipi(cpu);
3743 trace_sched_wake_idle_without_ipi(cpu);
3747 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3748 * necessary. The wakee CPU on receipt of the IPI will queue the task
3749 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3750 * of the wakeup instead of the waker.
3752 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3754 struct rq *rq = cpu_rq(cpu);
3756 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3758 WRITE_ONCE(rq->ttwu_pending, 1);
3759 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3762 void wake_up_if_idle(int cpu)
3764 struct rq *rq = cpu_rq(cpu);
3769 if (!is_idle_task(rcu_dereference(rq->curr)))
3772 if (set_nr_if_polling(rq->idle)) {
3773 trace_sched_wake_idle_without_ipi(cpu);
3775 rq_lock_irqsave(rq, &rf);
3776 if (is_idle_task(rq->curr))
3777 smp_send_reschedule(cpu);
3778 /* Else CPU is not idle, do nothing here: */
3779 rq_unlock_irqrestore(rq, &rf);
3786 bool cpus_share_cache(int this_cpu, int that_cpu)
3788 if (this_cpu == that_cpu)
3791 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3794 static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
3797 * Do not complicate things with the async wake_list while the CPU is
3800 if (!cpu_active(cpu))
3803 /* Ensure the task will still be allowed to run on the CPU. */
3804 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3808 * If the CPU does not share cache, then queue the task on the
3809 * remote rqs wakelist to avoid accessing remote data.
3811 if (!cpus_share_cache(smp_processor_id(), cpu))
3814 if (cpu == smp_processor_id())
3818 * If the wakee cpu is idle, or the task is descheduling and the
3819 * only running task on the CPU, then use the wakelist to offload
3820 * the task activation to the idle (or soon-to-be-idle) CPU as
3821 * the current CPU is likely busy. nr_running is checked to
3822 * avoid unnecessary task stacking.
3824 * Note that we can only get here with (wakee) p->on_rq=0,
3825 * p->on_cpu can be whatever, we've done the dequeue, so
3826 * the wakee has been accounted out of ->nr_running.
3828 if (!cpu_rq(cpu)->nr_running)
3834 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3836 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
3837 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3838 __ttwu_queue_wakelist(p, cpu, wake_flags);
3845 #else /* !CONFIG_SMP */
3847 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3852 #endif /* CONFIG_SMP */
3854 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3856 struct rq *rq = cpu_rq(cpu);
3859 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3863 update_rq_clock(rq);
3864 ttwu_do_activate(rq, p, wake_flags, &rf);
3869 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3871 * The caller holds p::pi_lock if p != current or has preemption
3872 * disabled when p == current.
3874 * The rules of PREEMPT_RT saved_state:
3876 * The related locking code always holds p::pi_lock when updating
3877 * p::saved_state, which means the code is fully serialized in both cases.
3879 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3880 * bits set. This allows to distinguish all wakeup scenarios.
3882 static __always_inline
3883 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3885 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3886 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3887 state != TASK_RTLOCK_WAIT);
3890 if (READ_ONCE(p->__state) & state) {
3895 #ifdef CONFIG_PREEMPT_RT
3897 * Saved state preserves the task state across blocking on
3898 * an RT lock. If the state matches, set p::saved_state to
3899 * TASK_RUNNING, but do not wake the task because it waits
3900 * for a lock wakeup. Also indicate success because from
3901 * the regular waker's point of view this has succeeded.
3903 * After acquiring the lock the task will restore p::__state
3904 * from p::saved_state which ensures that the regular
3905 * wakeup is not lost. The restore will also set
3906 * p::saved_state to TASK_RUNNING so any further tests will
3907 * not result in false positives vs. @success
3909 if (p->saved_state & state) {
3910 p->saved_state = TASK_RUNNING;
3918 * Notes on Program-Order guarantees on SMP systems.
3922 * The basic program-order guarantee on SMP systems is that when a task [t]
3923 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3924 * execution on its new CPU [c1].
3926 * For migration (of runnable tasks) this is provided by the following means:
3928 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3929 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3930 * rq(c1)->lock (if not at the same time, then in that order).
3931 * C) LOCK of the rq(c1)->lock scheduling in task
3933 * Release/acquire chaining guarantees that B happens after A and C after B.
3934 * Note: the CPU doing B need not be c0 or c1
3943 * UNLOCK rq(0)->lock
3945 * LOCK rq(0)->lock // orders against CPU0
3947 * UNLOCK rq(0)->lock
3951 * UNLOCK rq(1)->lock
3953 * LOCK rq(1)->lock // orders against CPU2
3956 * UNLOCK rq(1)->lock
3959 * BLOCKING -- aka. SLEEP + WAKEUP
3961 * For blocking we (obviously) need to provide the same guarantee as for
3962 * migration. However the means are completely different as there is no lock
3963 * chain to provide order. Instead we do:
3965 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3966 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3970 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3972 * LOCK rq(0)->lock LOCK X->pi_lock
3975 * smp_store_release(X->on_cpu, 0);
3977 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3983 * X->state = RUNNING
3984 * UNLOCK rq(2)->lock
3986 * LOCK rq(2)->lock // orders against CPU1
3989 * UNLOCK rq(2)->lock
3992 * UNLOCK rq(0)->lock
3995 * However, for wakeups there is a second guarantee we must provide, namely we
3996 * must ensure that CONDITION=1 done by the caller can not be reordered with
3997 * accesses to the task state; see try_to_wake_up() and set_current_state().
4001 * try_to_wake_up - wake up a thread
4002 * @p: the thread to be awakened
4003 * @state: the mask of task states that can be woken
4004 * @wake_flags: wake modifier flags (WF_*)
4006 * Conceptually does:
4008 * If (@state & @p->state) @p->state = TASK_RUNNING.
4010 * If the task was not queued/runnable, also place it back on a runqueue.
4012 * This function is atomic against schedule() which would dequeue the task.
4014 * It issues a full memory barrier before accessing @p->state, see the comment
4015 * with set_current_state().
4017 * Uses p->pi_lock to serialize against concurrent wake-ups.
4019 * Relies on p->pi_lock stabilizing:
4022 * - p->sched_task_group
4023 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4025 * Tries really hard to only take one task_rq(p)->lock for performance.
4026 * Takes rq->lock in:
4027 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4028 * - ttwu_queue() -- new rq, for enqueue of the task;
4029 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4031 * As a consequence we race really badly with just about everything. See the
4032 * many memory barriers and their comments for details.
4034 * Return: %true if @p->state changes (an actual wakeup was done),
4038 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
4040 unsigned long flags;
4041 int cpu, success = 0;
4046 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4047 * == smp_processor_id()'. Together this means we can special
4048 * case the whole 'p->on_rq && ttwu_runnable()' case below
4049 * without taking any locks.
4052 * - we rely on Program-Order guarantees for all the ordering,
4053 * - we're serialized against set_special_state() by virtue of
4054 * it disabling IRQs (this allows not taking ->pi_lock).
4056 if (!ttwu_state_match(p, state, &success))
4059 trace_sched_waking(p);
4060 WRITE_ONCE(p->__state, TASK_RUNNING);
4061 trace_sched_wakeup(p);
4066 * If we are going to wake up a thread waiting for CONDITION we
4067 * need to ensure that CONDITION=1 done by the caller can not be
4068 * reordered with p->state check below. This pairs with smp_store_mb()
4069 * in set_current_state() that the waiting thread does.
4071 raw_spin_lock_irqsave(&p->pi_lock, flags);
4072 smp_mb__after_spinlock();
4073 if (!ttwu_state_match(p, state, &success))
4076 trace_sched_waking(p);
4079 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4080 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4081 * in smp_cond_load_acquire() below.
4083 * sched_ttwu_pending() try_to_wake_up()
4084 * STORE p->on_rq = 1 LOAD p->state
4087 * __schedule() (switch to task 'p')
4088 * LOCK rq->lock smp_rmb();
4089 * smp_mb__after_spinlock();
4093 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4095 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4096 * __schedule(). See the comment for smp_mb__after_spinlock().
4098 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4101 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4106 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4107 * possible to, falsely, observe p->on_cpu == 0.
4109 * One must be running (->on_cpu == 1) in order to remove oneself
4110 * from the runqueue.
4112 * __schedule() (switch to task 'p') try_to_wake_up()
4113 * STORE p->on_cpu = 1 LOAD p->on_rq
4116 * __schedule() (put 'p' to sleep)
4117 * LOCK rq->lock smp_rmb();
4118 * smp_mb__after_spinlock();
4119 * STORE p->on_rq = 0 LOAD p->on_cpu
4121 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4122 * __schedule(). See the comment for smp_mb__after_spinlock().
4124 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4125 * schedule()'s deactivate_task() has 'happened' and p will no longer
4126 * care about it's own p->state. See the comment in __schedule().
4128 smp_acquire__after_ctrl_dep();
4131 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4132 * == 0), which means we need to do an enqueue, change p->state to
4133 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4134 * enqueue, such as ttwu_queue_wakelist().
4136 WRITE_ONCE(p->__state, TASK_WAKING);
4139 * If the owning (remote) CPU is still in the middle of schedule() with
4140 * this task as prev, considering queueing p on the remote CPUs wake_list
4141 * which potentially sends an IPI instead of spinning on p->on_cpu to
4142 * let the waker make forward progress. This is safe because IRQs are
4143 * disabled and the IPI will deliver after on_cpu is cleared.
4145 * Ensure we load task_cpu(p) after p->on_cpu:
4147 * set_task_cpu(p, cpu);
4148 * STORE p->cpu = @cpu
4149 * __schedule() (switch to task 'p')
4151 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4152 * STORE p->on_cpu = 1 LOAD p->cpu
4154 * to ensure we observe the correct CPU on which the task is currently
4157 if (smp_load_acquire(&p->on_cpu) &&
4158 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
4162 * If the owning (remote) CPU is still in the middle of schedule() with
4163 * this task as prev, wait until it's done referencing the task.
4165 * Pairs with the smp_store_release() in finish_task().
4167 * This ensures that tasks getting woken will be fully ordered against
4168 * their previous state and preserve Program Order.
4170 smp_cond_load_acquire(&p->on_cpu, !VAL);
4172 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4173 if (task_cpu(p) != cpu) {
4175 delayacct_blkio_end(p);
4176 atomic_dec(&task_rq(p)->nr_iowait);
4179 wake_flags |= WF_MIGRATED;
4180 psi_ttwu_dequeue(p);
4181 set_task_cpu(p, cpu);
4185 #endif /* CONFIG_SMP */
4187 ttwu_queue(p, cpu, wake_flags);
4189 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4192 ttwu_stat(p, task_cpu(p), wake_flags);
4199 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
4200 * @p: Process for which the function is to be invoked, can be @current.
4201 * @func: Function to invoke.
4202 * @arg: Argument to function.
4204 * If the specified task can be quickly locked into a definite state
4205 * (either sleeping or on a given runqueue), arrange to keep it in that
4206 * state while invoking @func(@arg). This function can use ->on_rq and
4207 * task_curr() to work out what the state is, if required. Given that
4208 * @func can be invoked with a runqueue lock held, it had better be quite
4212 * @false if the task slipped out from under the locks.
4213 * @true if the task was locked onto a runqueue or is sleeping.
4214 * However, @func can override this by returning @false.
4216 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
4222 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4224 rq = __task_rq_lock(p, &rf);
4225 if (task_rq(p) == rq)
4229 switch (READ_ONCE(p->__state)) {
4234 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
4239 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4244 * wake_up_process - Wake up a specific process
4245 * @p: The process to be woken up.
4247 * Attempt to wake up the nominated process and move it to the set of runnable
4250 * Return: 1 if the process was woken up, 0 if it was already running.
4252 * This function executes a full memory barrier before accessing the task state.
4254 int wake_up_process(struct task_struct *p)
4256 return try_to_wake_up(p, TASK_NORMAL, 0);
4258 EXPORT_SYMBOL(wake_up_process);
4260 int wake_up_state(struct task_struct *p, unsigned int state)
4262 return try_to_wake_up(p, state, 0);
4266 * Perform scheduler related setup for a newly forked process p.
4267 * p is forked by current.
4269 * __sched_fork() is basic setup used by init_idle() too:
4271 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4276 p->se.exec_start = 0;
4277 p->se.sum_exec_runtime = 0;
4278 p->se.prev_sum_exec_runtime = 0;
4279 p->se.nr_migrations = 0;
4281 INIT_LIST_HEAD(&p->se.group_node);
4283 #ifdef CONFIG_FAIR_GROUP_SCHED
4284 p->se.cfs_rq = NULL;
4287 #ifdef CONFIG_SCHEDSTATS
4288 /* Even if schedstat is disabled, there should not be garbage */
4289 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
4292 RB_CLEAR_NODE(&p->dl.rb_node);
4293 init_dl_task_timer(&p->dl);
4294 init_dl_inactive_task_timer(&p->dl);
4295 __dl_clear_params(p);
4297 INIT_LIST_HEAD(&p->rt.run_list);
4299 p->rt.time_slice = sched_rr_timeslice;
4303 #ifdef CONFIG_PREEMPT_NOTIFIERS
4304 INIT_HLIST_HEAD(&p->preempt_notifiers);
4307 #ifdef CONFIG_COMPACTION
4308 p->capture_control = NULL;
4310 init_numa_balancing(clone_flags, p);
4312 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4313 p->migration_pending = NULL;
4317 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4319 #ifdef CONFIG_NUMA_BALANCING
4321 void set_numabalancing_state(bool enabled)
4324 static_branch_enable(&sched_numa_balancing);
4326 static_branch_disable(&sched_numa_balancing);
4329 #ifdef CONFIG_PROC_SYSCTL
4330 int sysctl_numa_balancing(struct ctl_table *table, int write,
4331 void *buffer, size_t *lenp, loff_t *ppos)
4335 int state = static_branch_likely(&sched_numa_balancing);
4337 if (write && !capable(CAP_SYS_ADMIN))
4342 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4346 set_numabalancing_state(state);
4352 #ifdef CONFIG_SCHEDSTATS
4354 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4356 static void set_schedstats(bool enabled)
4359 static_branch_enable(&sched_schedstats);
4361 static_branch_disable(&sched_schedstats);
4364 void force_schedstat_enabled(void)
4366 if (!schedstat_enabled()) {
4367 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4368 static_branch_enable(&sched_schedstats);
4372 static int __init setup_schedstats(char *str)
4378 if (!strcmp(str, "enable")) {
4379 set_schedstats(true);
4381 } else if (!strcmp(str, "disable")) {
4382 set_schedstats(false);
4387 pr_warn("Unable to parse schedstats=\n");
4391 __setup("schedstats=", setup_schedstats);
4393 #ifdef CONFIG_PROC_SYSCTL
4394 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4395 size_t *lenp, loff_t *ppos)
4399 int state = static_branch_likely(&sched_schedstats);
4401 if (write && !capable(CAP_SYS_ADMIN))
4406 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4410 set_schedstats(state);
4413 #endif /* CONFIG_PROC_SYSCTL */
4414 #endif /* CONFIG_SCHEDSTATS */
4417 * fork()/clone()-time setup:
4419 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4421 __sched_fork(clone_flags, p);
4423 * We mark the process as NEW here. This guarantees that
4424 * nobody will actually run it, and a signal or other external
4425 * event cannot wake it up and insert it on the runqueue either.
4427 p->__state = TASK_NEW;
4430 * Make sure we do not leak PI boosting priority to the child.
4432 p->prio = current->normal_prio;
4437 * Revert to default priority/policy on fork if requested.
4439 if (unlikely(p->sched_reset_on_fork)) {
4440 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4441 p->policy = SCHED_NORMAL;
4442 p->static_prio = NICE_TO_PRIO(0);
4444 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4445 p->static_prio = NICE_TO_PRIO(0);
4447 p->prio = p->normal_prio = p->static_prio;
4448 set_load_weight(p, false);
4451 * We don't need the reset flag anymore after the fork. It has
4452 * fulfilled its duty:
4454 p->sched_reset_on_fork = 0;
4457 if (dl_prio(p->prio))
4459 else if (rt_prio(p->prio))
4460 p->sched_class = &rt_sched_class;
4462 p->sched_class = &fair_sched_class;
4464 init_entity_runnable_average(&p->se);
4467 #ifdef CONFIG_SCHED_INFO
4468 if (likely(sched_info_on()))
4469 memset(&p->sched_info, 0, sizeof(p->sched_info));
4471 #if defined(CONFIG_SMP)
4474 init_task_preempt_count(p);
4475 #ifdef CONFIG_HAVE_PREEMPT_LAZY
4476 task_thread_info(p)->preempt_lazy_count = 0;
4479 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4480 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4485 void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4487 unsigned long flags;
4490 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4491 * required yet, but lockdep gets upset if rules are violated.
4493 raw_spin_lock_irqsave(&p->pi_lock, flags);
4494 #ifdef CONFIG_CGROUP_SCHED
4496 struct task_group *tg;
4497 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4498 struct task_group, css);
4499 tg = autogroup_task_group(p, tg);
4500 p->sched_task_group = tg;
4505 * We're setting the CPU for the first time, we don't migrate,
4506 * so use __set_task_cpu().
4508 __set_task_cpu(p, smp_processor_id());
4509 if (p->sched_class->task_fork)
4510 p->sched_class->task_fork(p);
4511 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4514 void sched_post_fork(struct task_struct *p)
4516 uclamp_post_fork(p);
4519 unsigned long to_ratio(u64 period, u64 runtime)
4521 if (runtime == RUNTIME_INF)
4525 * Doing this here saves a lot of checks in all
4526 * the calling paths, and returning zero seems
4527 * safe for them anyway.
4532 return div64_u64(runtime << BW_SHIFT, period);
4536 * wake_up_new_task - wake up a newly created task for the first time.
4538 * This function will do some initial scheduler statistics housekeeping
4539 * that must be done for every newly created context, then puts the task
4540 * on the runqueue and wakes it.
4542 void wake_up_new_task(struct task_struct *p)
4547 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4548 WRITE_ONCE(p->__state, TASK_RUNNING);
4551 * Fork balancing, do it here and not earlier because:
4552 * - cpus_ptr can change in the fork path
4553 * - any previously selected CPU might disappear through hotplug
4555 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4556 * as we're not fully set-up yet.
4558 p->recent_used_cpu = task_cpu(p);
4560 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4562 rq = __task_rq_lock(p, &rf);
4563 update_rq_clock(rq);
4564 post_init_entity_util_avg(p);
4566 activate_task(rq, p, ENQUEUE_NOCLOCK);
4567 trace_sched_wakeup_new(p);
4568 check_preempt_curr(rq, p, WF_FORK);
4570 if (p->sched_class->task_woken) {
4572 * Nothing relies on rq->lock after this, so it's fine to
4575 rq_unpin_lock(rq, &rf);
4576 p->sched_class->task_woken(rq, p);
4577 rq_repin_lock(rq, &rf);
4580 task_rq_unlock(rq, p, &rf);
4583 #ifdef CONFIG_PREEMPT_NOTIFIERS
4585 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4587 void preempt_notifier_inc(void)
4589 static_branch_inc(&preempt_notifier_key);
4591 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4593 void preempt_notifier_dec(void)
4595 static_branch_dec(&preempt_notifier_key);
4597 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4600 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4601 * @notifier: notifier struct to register
4603 void preempt_notifier_register(struct preempt_notifier *notifier)
4605 if (!static_branch_unlikely(&preempt_notifier_key))
4606 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4608 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4610 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4613 * preempt_notifier_unregister - no longer interested in preemption notifications
4614 * @notifier: notifier struct to unregister
4616 * This is *not* safe to call from within a preemption notifier.
4618 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4620 hlist_del(¬ifier->link);
4622 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4624 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4626 struct preempt_notifier *notifier;
4628 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4629 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4632 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4634 if (static_branch_unlikely(&preempt_notifier_key))
4635 __fire_sched_in_preempt_notifiers(curr);
4639 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4640 struct task_struct *next)
4642 struct preempt_notifier *notifier;
4644 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4645 notifier->ops->sched_out(notifier, next);
4648 static __always_inline void
4649 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4650 struct task_struct *next)
4652 if (static_branch_unlikely(&preempt_notifier_key))
4653 __fire_sched_out_preempt_notifiers(curr, next);
4656 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4658 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4663 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4664 struct task_struct *next)
4668 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4670 static inline void prepare_task(struct task_struct *next)
4674 * Claim the task as running, we do this before switching to it
4675 * such that any running task will have this set.
4677 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4678 * its ordering comment.
4680 WRITE_ONCE(next->on_cpu, 1);
4684 static inline void finish_task(struct task_struct *prev)
4688 * This must be the very last reference to @prev from this CPU. After
4689 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4690 * must ensure this doesn't happen until the switch is completely
4693 * In particular, the load of prev->state in finish_task_switch() must
4694 * happen before this.
4696 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4698 smp_store_release(&prev->on_cpu, 0);
4704 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4706 void (*func)(struct rq *rq);
4707 struct callback_head *next;
4709 lockdep_assert_rq_held(rq);
4712 func = (void (*)(struct rq *))head->func;
4721 static void balance_push(struct rq *rq);
4724 * balance_push_callback is a right abuse of the callback interface and plays
4725 * by significantly different rules.
4727 * Where the normal balance_callback's purpose is to be ran in the same context
4728 * that queued it (only later, when it's safe to drop rq->lock again),
4729 * balance_push_callback is specifically targeted at __schedule().
4731 * This abuse is tolerated because it places all the unlikely/odd cases behind
4732 * a single test, namely: rq->balance_callback == NULL.
4734 struct callback_head balance_push_callback = {
4736 .func = (void (*)(struct callback_head *))balance_push,
4739 static inline struct callback_head *
4740 __splice_balance_callbacks(struct rq *rq, bool split)
4742 struct callback_head *head = rq->balance_callback;
4747 lockdep_assert_rq_held(rq);
4749 * Must not take balance_push_callback off the list when
4750 * splice_balance_callbacks() and balance_callbacks() are not
4751 * in the same rq->lock section.
4753 * In that case it would be possible for __schedule() to interleave
4754 * and observe the list empty.
4756 if (split && head == &balance_push_callback)
4759 rq->balance_callback = NULL;
4764 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4766 return __splice_balance_callbacks(rq, true);
4769 static void __balance_callbacks(struct rq *rq)
4771 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
4774 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4776 unsigned long flags;
4778 if (unlikely(head)) {
4779 raw_spin_rq_lock_irqsave(rq, flags);
4780 do_balance_callbacks(rq, head);
4781 raw_spin_rq_unlock_irqrestore(rq, flags);
4787 static inline void __balance_callbacks(struct rq *rq)
4791 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4796 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4803 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4806 * Since the runqueue lock will be released by the next
4807 * task (which is an invalid locking op but in the case
4808 * of the scheduler it's an obvious special-case), so we
4809 * do an early lockdep release here:
4811 rq_unpin_lock(rq, rf);
4812 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4813 #ifdef CONFIG_DEBUG_SPINLOCK
4814 /* this is a valid case when another task releases the spinlock */
4815 rq_lockp(rq)->owner = next;
4819 static inline void finish_lock_switch(struct rq *rq)
4822 * If we are tracking spinlock dependencies then we have to
4823 * fix up the runqueue lock - which gets 'carried over' from
4824 * prev into current:
4826 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4827 __balance_callbacks(rq);
4828 raw_spin_rq_unlock_irq(rq);
4832 * NOP if the arch has not defined these:
4835 #ifndef prepare_arch_switch
4836 # define prepare_arch_switch(next) do { } while (0)
4839 #ifndef finish_arch_post_lock_switch
4840 # define finish_arch_post_lock_switch() do { } while (0)
4843 static inline void kmap_local_sched_out(void)
4845 #ifdef CONFIG_KMAP_LOCAL
4846 if (unlikely(current->kmap_ctrl.idx))
4847 __kmap_local_sched_out();
4851 static inline void kmap_local_sched_in(void)
4853 #ifdef CONFIG_KMAP_LOCAL
4854 if (unlikely(current->kmap_ctrl.idx))
4855 __kmap_local_sched_in();
4860 * prepare_task_switch - prepare to switch tasks
4861 * @rq: the runqueue preparing to switch
4862 * @prev: the current task that is being switched out
4863 * @next: the task we are going to switch to.
4865 * This is called with the rq lock held and interrupts off. It must
4866 * be paired with a subsequent finish_task_switch after the context
4869 * prepare_task_switch sets up locking and calls architecture specific
4873 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4874 struct task_struct *next)
4876 kcov_prepare_switch(prev);
4877 sched_info_switch(rq, prev, next);
4878 perf_event_task_sched_out(prev, next);
4880 fire_sched_out_preempt_notifiers(prev, next);
4881 kmap_local_sched_out();
4883 prepare_arch_switch(next);
4887 * finish_task_switch - clean up after a task-switch
4888 * @prev: the thread we just switched away from.
4890 * finish_task_switch must be called after the context switch, paired
4891 * with a prepare_task_switch call before the context switch.
4892 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4893 * and do any other architecture-specific cleanup actions.
4895 * Note that we may have delayed dropping an mm in context_switch(). If
4896 * so, we finish that here outside of the runqueue lock. (Doing it
4897 * with the lock held can cause deadlocks; see schedule() for
4900 * The context switch have flipped the stack from under us and restored the
4901 * local variables which were saved when this task called schedule() in the
4902 * past. prev == current is still correct but we need to recalculate this_rq
4903 * because prev may have moved to another CPU.
4905 static struct rq *finish_task_switch(struct task_struct *prev)
4906 __releases(rq->lock)
4908 struct rq *rq = this_rq();
4909 struct mm_struct *mm = rq->prev_mm;
4913 * The previous task will have left us with a preempt_count of 2
4914 * because it left us after:
4917 * preempt_disable(); // 1
4919 * raw_spin_lock_irq(&rq->lock) // 2
4921 * Also, see FORK_PREEMPT_COUNT.
4923 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4924 "corrupted preempt_count: %s/%d/0x%x\n",
4925 current->comm, current->pid, preempt_count()))
4926 preempt_count_set(FORK_PREEMPT_COUNT);
4931 * A task struct has one reference for the use as "current".
4932 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4933 * schedule one last time. The schedule call will never return, and
4934 * the scheduled task must drop that reference.
4936 * We must observe prev->state before clearing prev->on_cpu (in
4937 * finish_task), otherwise a concurrent wakeup can get prev
4938 * running on another CPU and we could rave with its RUNNING -> DEAD
4939 * transition, resulting in a double drop.
4941 prev_state = READ_ONCE(prev->__state);
4942 vtime_task_switch(prev);
4943 perf_event_task_sched_in(prev, current);
4945 tick_nohz_task_switch();
4946 finish_lock_switch(rq);
4947 finish_arch_post_lock_switch();
4948 kcov_finish_switch(current);
4950 * kmap_local_sched_out() is invoked with rq::lock held and
4951 * interrupts disabled. There is no requirement for that, but the
4952 * sched out code does not have an interrupt enabled section.
4953 * Restoring the maps on sched in does not require interrupts being
4956 kmap_local_sched_in();
4958 fire_sched_in_preempt_notifiers(current);
4960 * When switching through a kernel thread, the loop in
4961 * membarrier_{private,global}_expedited() may have observed that
4962 * kernel thread and not issued an IPI. It is therefore possible to
4963 * schedule between user->kernel->user threads without passing though
4964 * switch_mm(). Membarrier requires a barrier after storing to
4965 * rq->curr, before returning to userspace, so provide them here:
4967 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4968 * provided by mmdrop(),
4969 * - a sync_core for SYNC_CORE.
4972 membarrier_mm_sync_core_before_usermode(mm);
4975 if (unlikely(prev_state == TASK_DEAD)) {
4976 if (prev->sched_class->task_dead)
4977 prev->sched_class->task_dead(prev);
4980 * Release VMAP'ed task stack immediate for reuse. On RT
4981 * enabled kernels this is delayed for latency reasons.
4983 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
4984 put_task_stack(prev);
4986 put_task_struct_rcu_user(prev);
4993 * schedule_tail - first thing a freshly forked thread must call.
4994 * @prev: the thread we just switched away from.
4996 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4997 __releases(rq->lock)
5000 * New tasks start with FORK_PREEMPT_COUNT, see there and
5001 * finish_task_switch() for details.
5003 * finish_task_switch() will drop rq->lock() and lower preempt_count
5004 * and the preempt_enable() will end up enabling preemption (on
5005 * PREEMPT_COUNT kernels).
5008 finish_task_switch(prev);
5011 if (current->set_child_tid)
5012 put_user(task_pid_vnr(current), current->set_child_tid);
5014 calculate_sigpending();
5018 * context_switch - switch to the new MM and the new thread's register state.
5020 static __always_inline struct rq *
5021 context_switch(struct rq *rq, struct task_struct *prev,
5022 struct task_struct *next, struct rq_flags *rf)
5024 prepare_task_switch(rq, prev, next);
5027 * For paravirt, this is coupled with an exit in switch_to to
5028 * combine the page table reload and the switch backend into
5031 arch_start_context_switch(prev);
5034 * kernel -> kernel lazy + transfer active
5035 * user -> kernel lazy + mmgrab() active
5037 * kernel -> user switch + mmdrop() active
5038 * user -> user switch
5040 if (!next->mm) { // to kernel
5041 enter_lazy_tlb(prev->active_mm, next);
5043 next->active_mm = prev->active_mm;
5044 if (prev->mm) // from user
5045 mmgrab(prev->active_mm);
5047 prev->active_mm = NULL;
5049 membarrier_switch_mm(rq, prev->active_mm, next->mm);
5051 * sys_membarrier() requires an smp_mb() between setting
5052 * rq->curr / membarrier_switch_mm() and returning to userspace.
5054 * The below provides this either through switch_mm(), or in
5055 * case 'prev->active_mm == next->mm' through
5056 * finish_task_switch()'s mmdrop().
5058 switch_mm_irqs_off(prev->active_mm, next->mm, next);
5060 if (!prev->mm) { // from kernel
5061 /* will mmdrop() in finish_task_switch(). */
5062 rq->prev_mm = prev->active_mm;
5063 prev->active_mm = NULL;
5067 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5069 prepare_lock_switch(rq, next, rf);
5071 /* Here we just switch the register state and the stack. */
5072 switch_to(prev, next, prev);
5075 return finish_task_switch(prev);
5079 * nr_running and nr_context_switches:
5081 * externally visible scheduler statistics: current number of runnable
5082 * threads, total number of context switches performed since bootup.
5084 unsigned int nr_running(void)
5086 unsigned int i, sum = 0;
5088 for_each_online_cpu(i)
5089 sum += cpu_rq(i)->nr_running;
5095 * Check if only the current task is running on the CPU.
5097 * Caution: this function does not check that the caller has disabled
5098 * preemption, thus the result might have a time-of-check-to-time-of-use
5099 * race. The caller is responsible to use it correctly, for example:
5101 * - from a non-preemptible section (of course)
5103 * - from a thread that is bound to a single CPU
5105 * - in a loop with very short iterations (e.g. a polling loop)
5107 bool single_task_running(void)
5109 return raw_rq()->nr_running == 1;
5111 EXPORT_SYMBOL(single_task_running);
5113 unsigned long long nr_context_switches(void)
5116 unsigned long long sum = 0;
5118 for_each_possible_cpu(i)
5119 sum += cpu_rq(i)->nr_switches;
5125 * Consumers of these two interfaces, like for example the cpuidle menu
5126 * governor, are using nonsensical data. Preferring shallow idle state selection
5127 * for a CPU that has IO-wait which might not even end up running the task when
5128 * it does become runnable.
5131 unsigned int nr_iowait_cpu(int cpu)
5133 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5137 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5139 * The idea behind IO-wait account is to account the idle time that we could
5140 * have spend running if it were not for IO. That is, if we were to improve the
5141 * storage performance, we'd have a proportional reduction in IO-wait time.
5143 * This all works nicely on UP, where, when a task blocks on IO, we account
5144 * idle time as IO-wait, because if the storage were faster, it could've been
5145 * running and we'd not be idle.
5147 * This has been extended to SMP, by doing the same for each CPU. This however
5150 * Imagine for instance the case where two tasks block on one CPU, only the one
5151 * CPU will have IO-wait accounted, while the other has regular idle. Even
5152 * though, if the storage were faster, both could've ran at the same time,
5153 * utilising both CPUs.
5155 * This means, that when looking globally, the current IO-wait accounting on
5156 * SMP is a lower bound, by reason of under accounting.
5158 * Worse, since the numbers are provided per CPU, they are sometimes
5159 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5160 * associated with any one particular CPU, it can wake to another CPU than it
5161 * blocked on. This means the per CPU IO-wait number is meaningless.
5163 * Task CPU affinities can make all that even more 'interesting'.
5166 unsigned int nr_iowait(void)
5168 unsigned int i, sum = 0;
5170 for_each_possible_cpu(i)
5171 sum += nr_iowait_cpu(i);
5179 * sched_exec - execve() is a valuable balancing opportunity, because at
5180 * this point the task has the smallest effective memory and cache footprint.
5182 void sched_exec(void)
5184 struct task_struct *p = current;
5185 unsigned long flags;
5188 raw_spin_lock_irqsave(&p->pi_lock, flags);
5189 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5190 if (dest_cpu == smp_processor_id())
5193 if (likely(cpu_active(dest_cpu))) {
5194 struct migration_arg arg = { p, dest_cpu };
5196 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5197 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5201 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5206 DEFINE_PER_CPU(struct kernel_stat, kstat);
5207 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5209 EXPORT_PER_CPU_SYMBOL(kstat);
5210 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5213 * The function fair_sched_class.update_curr accesses the struct curr
5214 * and its field curr->exec_start; when called from task_sched_runtime(),
5215 * we observe a high rate of cache misses in practice.
5216 * Prefetching this data results in improved performance.
5218 static inline void prefetch_curr_exec_start(struct task_struct *p)
5220 #ifdef CONFIG_FAIR_GROUP_SCHED
5221 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5223 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5226 prefetch(&curr->exec_start);
5230 * Return accounted runtime for the task.
5231 * In case the task is currently running, return the runtime plus current's
5232 * pending runtime that have not been accounted yet.
5234 unsigned long long task_sched_runtime(struct task_struct *p)
5240 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5242 * 64-bit doesn't need locks to atomically read a 64-bit value.
5243 * So we have a optimization chance when the task's delta_exec is 0.
5244 * Reading ->on_cpu is racy, but this is ok.
5246 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5247 * If we race with it entering CPU, unaccounted time is 0. This is
5248 * indistinguishable from the read occurring a few cycles earlier.
5249 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5250 * been accounted, so we're correct here as well.
5252 if (!p->on_cpu || !task_on_rq_queued(p))
5253 return p->se.sum_exec_runtime;
5256 rq = task_rq_lock(p, &rf);
5258 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5259 * project cycles that may never be accounted to this
5260 * thread, breaking clock_gettime().
5262 if (task_current(rq, p) && task_on_rq_queued(p)) {
5263 prefetch_curr_exec_start(p);
5264 update_rq_clock(rq);
5265 p->sched_class->update_curr(rq);
5267 ns = p->se.sum_exec_runtime;
5268 task_rq_unlock(rq, p, &rf);
5273 #ifdef CONFIG_SCHED_DEBUG
5274 static u64 cpu_resched_latency(struct rq *rq)
5276 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5277 u64 resched_latency, now = rq_clock(rq);
5278 static bool warned_once;
5280 if (sysctl_resched_latency_warn_once && warned_once)
5283 if (!need_resched() || !latency_warn_ms)
5286 if (system_state == SYSTEM_BOOTING)
5289 if (!rq->last_seen_need_resched_ns) {
5290 rq->last_seen_need_resched_ns = now;
5291 rq->ticks_without_resched = 0;
5295 rq->ticks_without_resched++;
5296 resched_latency = now - rq->last_seen_need_resched_ns;
5297 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5302 return resched_latency;
5305 static int __init setup_resched_latency_warn_ms(char *str)
5309 if ((kstrtol(str, 0, &val))) {
5310 pr_warn("Unable to set resched_latency_warn_ms\n");
5314 sysctl_resched_latency_warn_ms = val;
5317 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5319 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5320 #endif /* CONFIG_SCHED_DEBUG */
5323 * This function gets called by the timer code, with HZ frequency.
5324 * We call it with interrupts disabled.
5326 void scheduler_tick(void)
5328 int cpu = smp_processor_id();
5329 struct rq *rq = cpu_rq(cpu);
5330 struct task_struct *curr = rq->curr;
5332 unsigned long thermal_pressure;
5333 u64 resched_latency;
5335 arch_scale_freq_tick();
5340 update_rq_clock(rq);
5341 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5342 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5343 curr->sched_class->task_tick(rq, curr, 0);
5344 if (sched_feat(LATENCY_WARN))
5345 resched_latency = cpu_resched_latency(rq);
5346 calc_global_load_tick(rq);
5350 if (sched_feat(LATENCY_WARN) && resched_latency)
5351 resched_latency_warn(cpu, resched_latency);
5353 perf_event_task_tick();
5356 rq->idle_balance = idle_cpu(cpu);
5357 trigger_load_balance(rq);
5361 #ifdef CONFIG_NO_HZ_FULL
5366 struct delayed_work work;
5368 /* Values for ->state, see diagram below. */
5369 #define TICK_SCHED_REMOTE_OFFLINE 0
5370 #define TICK_SCHED_REMOTE_OFFLINING 1
5371 #define TICK_SCHED_REMOTE_RUNNING 2
5374 * State diagram for ->state:
5377 * TICK_SCHED_REMOTE_OFFLINE
5380 * | | sched_tick_remote()
5383 * +--TICK_SCHED_REMOTE_OFFLINING
5386 * sched_tick_start() | | sched_tick_stop()
5389 * TICK_SCHED_REMOTE_RUNNING
5392 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5393 * and sched_tick_start() are happy to leave the state in RUNNING.
5396 static struct tick_work __percpu *tick_work_cpu;
5398 static void sched_tick_remote(struct work_struct *work)
5400 struct delayed_work *dwork = to_delayed_work(work);
5401 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5402 int cpu = twork->cpu;
5403 struct rq *rq = cpu_rq(cpu);
5404 struct task_struct *curr;
5410 * Handle the tick only if it appears the remote CPU is running in full
5411 * dynticks mode. The check is racy by nature, but missing a tick or
5412 * having one too much is no big deal because the scheduler tick updates
5413 * statistics and checks timeslices in a time-independent way, regardless
5414 * of when exactly it is running.
5416 if (!tick_nohz_tick_stopped_cpu(cpu))
5419 rq_lock_irq(rq, &rf);
5421 if (cpu_is_offline(cpu))
5424 update_rq_clock(rq);
5426 if (!is_idle_task(curr)) {
5428 * Make sure the next tick runs within a reasonable
5431 delta = rq_clock_task(rq) - curr->se.exec_start;
5432 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5434 curr->sched_class->task_tick(rq, curr, 0);
5436 calc_load_nohz_remote(rq);
5438 rq_unlock_irq(rq, &rf);
5442 * Run the remote tick once per second (1Hz). This arbitrary
5443 * frequency is large enough to avoid overload but short enough
5444 * to keep scheduler internal stats reasonably up to date. But
5445 * first update state to reflect hotplug activity if required.
5447 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5448 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5449 if (os == TICK_SCHED_REMOTE_RUNNING)
5450 queue_delayed_work(system_unbound_wq, dwork, HZ);
5453 static void sched_tick_start(int cpu)
5456 struct tick_work *twork;
5458 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5461 WARN_ON_ONCE(!tick_work_cpu);
5463 twork = per_cpu_ptr(tick_work_cpu, cpu);
5464 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5465 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5466 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5468 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5469 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5473 #ifdef CONFIG_HOTPLUG_CPU
5474 static void sched_tick_stop(int cpu)
5476 struct tick_work *twork;
5479 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5482 WARN_ON_ONCE(!tick_work_cpu);
5484 twork = per_cpu_ptr(tick_work_cpu, cpu);
5485 /* There cannot be competing actions, but don't rely on stop-machine. */
5486 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5487 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5488 /* Don't cancel, as this would mess up the state machine. */
5490 #endif /* CONFIG_HOTPLUG_CPU */
5492 int __init sched_tick_offload_init(void)
5494 tick_work_cpu = alloc_percpu(struct tick_work);
5495 BUG_ON(!tick_work_cpu);
5499 #else /* !CONFIG_NO_HZ_FULL */
5500 static inline void sched_tick_start(int cpu) { }
5501 static inline void sched_tick_stop(int cpu) { }
5504 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5505 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5507 * If the value passed in is equal to the current preempt count
5508 * then we just disabled preemption. Start timing the latency.
5510 static inline void preempt_latency_start(int val)
5512 if (preempt_count() == val) {
5513 unsigned long ip = get_lock_parent_ip();
5514 #ifdef CONFIG_DEBUG_PREEMPT
5515 current->preempt_disable_ip = ip;
5517 trace_preempt_off(CALLER_ADDR0, ip);
5521 void preempt_count_add(int val)
5523 #ifdef CONFIG_DEBUG_PREEMPT
5527 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5530 __preempt_count_add(val);
5531 #ifdef CONFIG_DEBUG_PREEMPT
5533 * Spinlock count overflowing soon?
5535 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5538 preempt_latency_start(val);
5540 EXPORT_SYMBOL(preempt_count_add);
5541 NOKPROBE_SYMBOL(preempt_count_add);
5544 * If the value passed in equals to the current preempt count
5545 * then we just enabled preemption. Stop timing the latency.
5547 static inline void preempt_latency_stop(int val)
5549 if (preempt_count() == val)
5550 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5553 void preempt_count_sub(int val)
5555 #ifdef CONFIG_DEBUG_PREEMPT
5559 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5562 * Is the spinlock portion underflowing?
5564 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5565 !(preempt_count() & PREEMPT_MASK)))
5569 preempt_latency_stop(val);
5570 __preempt_count_sub(val);
5572 EXPORT_SYMBOL(preempt_count_sub);
5573 NOKPROBE_SYMBOL(preempt_count_sub);
5576 static inline void preempt_latency_start(int val) { }
5577 static inline void preempt_latency_stop(int val) { }
5580 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5582 #ifdef CONFIG_DEBUG_PREEMPT
5583 return p->preempt_disable_ip;
5590 * Print scheduling while atomic bug:
5592 static noinline void __schedule_bug(struct task_struct *prev)
5594 /* Save this before calling printk(), since that will clobber it */
5595 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5597 if (oops_in_progress)
5600 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5601 prev->comm, prev->pid, preempt_count());
5603 debug_show_held_locks(prev);
5605 if (irqs_disabled())
5606 print_irqtrace_events(prev);
5607 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5608 && in_atomic_preempt_off()) {
5609 pr_err("Preemption disabled at:");
5610 print_ip_sym(KERN_ERR, preempt_disable_ip);
5612 check_panic_on_warn("scheduling while atomic");
5615 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5619 * Various schedule()-time debugging checks and statistics:
5621 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5623 #ifdef CONFIG_SCHED_STACK_END_CHECK
5624 if (task_stack_end_corrupted(prev))
5625 panic("corrupted stack end detected inside scheduler\n");
5627 if (task_scs_end_corrupted(prev))
5628 panic("corrupted shadow stack detected inside scheduler\n");
5631 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5632 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5633 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5634 prev->comm, prev->pid, prev->non_block_count);
5636 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5640 if (unlikely(in_atomic_preempt_off())) {
5641 __schedule_bug(prev);
5642 preempt_count_set(PREEMPT_DISABLED);
5645 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5647 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5649 schedstat_inc(this_rq()->sched_count);
5652 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5653 struct rq_flags *rf)
5656 const struct sched_class *class;
5658 * We must do the balancing pass before put_prev_task(), such
5659 * that when we release the rq->lock the task is in the same
5660 * state as before we took rq->lock.
5662 * We can terminate the balance pass as soon as we know there is
5663 * a runnable task of @class priority or higher.
5665 for_class_range(class, prev->sched_class, &idle_sched_class) {
5666 if (class->balance(rq, prev, rf))
5671 put_prev_task(rq, prev);
5675 * Pick up the highest-prio task:
5677 static inline struct task_struct *
5678 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5680 const struct sched_class *class;
5681 struct task_struct *p;
5684 * Optimization: we know that if all tasks are in the fair class we can
5685 * call that function directly, but only if the @prev task wasn't of a
5686 * higher scheduling class, because otherwise those lose the
5687 * opportunity to pull in more work from other CPUs.
5689 if (likely(prev->sched_class <= &fair_sched_class &&
5690 rq->nr_running == rq->cfs.h_nr_running)) {
5692 p = pick_next_task_fair(rq, prev, rf);
5693 if (unlikely(p == RETRY_TASK))
5696 /* Assume the next prioritized class is idle_sched_class */
5698 put_prev_task(rq, prev);
5699 p = pick_next_task_idle(rq);
5706 put_prev_task_balance(rq, prev, rf);
5708 for_each_class(class) {
5709 p = class->pick_next_task(rq);
5714 /* The idle class should always have a runnable task: */
5718 #ifdef CONFIG_SCHED_CORE
5719 static inline bool is_task_rq_idle(struct task_struct *t)
5721 return (task_rq(t)->idle == t);
5724 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5726 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5729 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5731 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5734 return a->core_cookie == b->core_cookie;
5737 // XXX fairness/fwd progress conditions
5740 * - NULL if there is no runnable task for this class.
5741 * - the highest priority task for this runqueue if it matches
5742 * rq->core->core_cookie or its priority is greater than max.
5743 * - Else returns idle_task.
5745 static struct task_struct *
5746 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5748 struct task_struct *class_pick, *cookie_pick;
5749 unsigned long cookie = rq->core->core_cookie;
5751 class_pick = class->pick_task(rq);
5757 * If class_pick is tagged, return it only if it has
5758 * higher priority than max.
5760 if (max && class_pick->core_cookie &&
5761 prio_less(class_pick, max, in_fi))
5762 return idle_sched_class.pick_task(rq);
5768 * If class_pick is idle or matches cookie, return early.
5770 if (cookie_equals(class_pick, cookie))
5773 cookie_pick = sched_core_find(rq, cookie);
5776 * If class > max && class > cookie, it is the highest priority task on
5777 * the core (so far) and it must be selected, otherwise we must go with
5778 * the cookie pick in order to satisfy the constraint.
5780 if (prio_less(cookie_pick, class_pick, in_fi) &&
5781 (!max || prio_less(max, class_pick, in_fi)))
5787 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5789 static struct task_struct *
5790 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5792 struct task_struct *next, *max = NULL;
5793 const struct sched_class *class;
5794 const struct cpumask *smt_mask;
5795 bool fi_before = false;
5796 int i, j, cpu, occ = 0;
5799 if (!sched_core_enabled(rq))
5800 return __pick_next_task(rq, prev, rf);
5804 /* Stopper task is switching into idle, no need core-wide selection. */
5805 if (cpu_is_offline(cpu)) {
5807 * Reset core_pick so that we don't enter the fastpath when
5808 * coming online. core_pick would already be migrated to
5809 * another cpu during offline.
5811 rq->core_pick = NULL;
5812 return __pick_next_task(rq, prev, rf);
5816 * If there were no {en,de}queues since we picked (IOW, the task
5817 * pointers are all still valid), and we haven't scheduled the last
5818 * pick yet, do so now.
5820 * rq->core_pick can be NULL if no selection was made for a CPU because
5821 * it was either offline or went offline during a sibling's core-wide
5822 * selection. In this case, do a core-wide selection.
5824 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5825 rq->core->core_pick_seq != rq->core_sched_seq &&
5827 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5829 next = rq->core_pick;
5831 put_prev_task(rq, prev);
5832 set_next_task(rq, next);
5835 rq->core_pick = NULL;
5839 put_prev_task_balance(rq, prev, rf);
5841 smt_mask = cpu_smt_mask(cpu);
5842 need_sync = !!rq->core->core_cookie;
5845 rq->core->core_cookie = 0UL;
5846 if (rq->core->core_forceidle) {
5849 rq->core->core_forceidle = false;
5853 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5855 * @task_seq guards the task state ({en,de}queues)
5856 * @pick_seq is the @task_seq we did a selection on
5857 * @sched_seq is the @pick_seq we scheduled
5859 * However, preemptions can cause multiple picks on the same task set.
5860 * 'Fix' this by also increasing @task_seq for every pick.
5862 rq->core->core_task_seq++;
5865 * Optimize for common case where this CPU has no cookies
5866 * and there are no cookied tasks running on siblings.
5869 for_each_class(class) {
5870 next = class->pick_task(rq);
5875 if (!next->core_cookie) {
5876 rq->core_pick = NULL;
5878 * For robustness, update the min_vruntime_fi for
5879 * unconstrained picks as well.
5881 WARN_ON_ONCE(fi_before);
5882 task_vruntime_update(rq, next, false);
5887 for_each_cpu(i, smt_mask) {
5888 struct rq *rq_i = cpu_rq(i);
5890 rq_i->core_pick = NULL;
5893 update_rq_clock(rq_i);
5897 * Try and select tasks for each sibling in descending sched_class
5900 for_each_class(class) {
5902 for_each_cpu_wrap(i, smt_mask, cpu) {
5903 struct rq *rq_i = cpu_rq(i);
5904 struct task_struct *p;
5906 if (rq_i->core_pick)
5910 * If this sibling doesn't yet have a suitable task to
5911 * run; ask for the most eligible task, given the
5912 * highest priority task already selected for this
5915 p = pick_task(rq_i, class, max, fi_before);
5919 if (!is_task_rq_idle(p))
5922 rq_i->core_pick = p;
5923 if (rq_i->idle == p && rq_i->nr_running) {
5924 rq->core->core_forceidle = true;
5926 rq->core->core_forceidle_seq++;
5930 * If this new candidate is of higher priority than the
5931 * previous; and they're incompatible; we need to wipe
5932 * the slate and start over. pick_task makes sure that
5933 * p's priority is more than max if it doesn't match
5936 * NOTE: this is a linear max-filter and is thus bounded
5937 * in execution time.
5939 if (!max || !cookie_match(max, p)) {
5940 struct task_struct *old_max = max;
5942 rq->core->core_cookie = p->core_cookie;
5946 rq->core->core_forceidle = false;
5947 for_each_cpu(j, smt_mask) {
5951 cpu_rq(j)->core_pick = NULL;
5960 rq->core->core_pick_seq = rq->core->core_task_seq;
5961 next = rq->core_pick;
5962 rq->core_sched_seq = rq->core->core_pick_seq;
5964 /* Something should have been selected for current CPU */
5965 WARN_ON_ONCE(!next);
5968 * Reschedule siblings
5970 * NOTE: L1TF -- at this point we're no longer running the old task and
5971 * sending an IPI (below) ensures the sibling will no longer be running
5972 * their task. This ensures there is no inter-sibling overlap between
5973 * non-matching user state.
5975 for_each_cpu(i, smt_mask) {
5976 struct rq *rq_i = cpu_rq(i);
5979 * An online sibling might have gone offline before a task
5980 * could be picked for it, or it might be offline but later
5981 * happen to come online, but its too late and nothing was
5982 * picked for it. That's Ok - it will pick tasks for itself,
5985 if (!rq_i->core_pick)
5989 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5990 * fi_before fi update?
5996 if (!(fi_before && rq->core->core_forceidle))
5997 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5999 rq_i->core_pick->core_occupation = occ;
6002 rq_i->core_pick = NULL;
6006 /* Did we break L1TF mitigation requirements? */
6007 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6009 if (rq_i->curr == rq_i->core_pick) {
6010 rq_i->core_pick = NULL;
6018 set_next_task(rq, next);
6022 static bool try_steal_cookie(int this, int that)
6024 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6025 struct task_struct *p;
6026 unsigned long cookie;
6027 bool success = false;
6029 local_irq_disable();
6030 double_rq_lock(dst, src);
6032 cookie = dst->core->core_cookie;
6036 if (dst->curr != dst->idle)
6039 p = sched_core_find(src, cookie);
6044 if (p == src->core_pick || p == src->curr)
6047 if (!is_cpu_allowed(p, this))
6050 if (p->core_occupation > dst->idle->core_occupation)
6053 deactivate_task(src, p, 0);
6054 set_task_cpu(p, this);
6055 activate_task(dst, p, 0);
6063 p = sched_core_next(p, cookie);
6067 double_rq_unlock(dst, src);
6073 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6077 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6084 if (try_steal_cookie(cpu, i))
6091 static void sched_core_balance(struct rq *rq)
6093 struct sched_domain *sd;
6094 int cpu = cpu_of(rq);
6098 raw_spin_rq_unlock_irq(rq);
6099 for_each_domain(cpu, sd) {
6103 if (steal_cookie_task(cpu, sd))
6106 raw_spin_rq_lock_irq(rq);
6111 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
6113 void queue_core_balance(struct rq *rq)
6115 if (!sched_core_enabled(rq))
6118 if (!rq->core->core_cookie)
6121 if (!rq->nr_running) /* not forced idle */
6124 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6127 static void sched_core_cpu_starting(unsigned int cpu)
6129 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6130 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6131 unsigned long flags;
6134 sched_core_lock(cpu, &flags);
6136 WARN_ON_ONCE(rq->core != rq);
6138 /* if we're the first, we'll be our own leader */
6139 if (cpumask_weight(smt_mask) == 1)
6142 /* find the leader */
6143 for_each_cpu(t, smt_mask) {
6147 if (rq->core == rq) {
6153 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6156 /* install and validate core_rq */
6157 for_each_cpu(t, smt_mask) {
6163 WARN_ON_ONCE(rq->core != core_rq);
6167 sched_core_unlock(cpu, &flags);
6170 static void sched_core_cpu_deactivate(unsigned int cpu)
6172 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6173 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6174 unsigned long flags;
6177 sched_core_lock(cpu, &flags);
6179 /* if we're the last man standing, nothing to do */
6180 if (cpumask_weight(smt_mask) == 1) {
6181 WARN_ON_ONCE(rq->core != rq);
6185 /* if we're not the leader, nothing to do */
6189 /* find a new leader */
6190 for_each_cpu(t, smt_mask) {
6193 core_rq = cpu_rq(t);
6197 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6200 /* copy the shared state to the new leader */
6201 core_rq->core_task_seq = rq->core_task_seq;
6202 core_rq->core_pick_seq = rq->core_pick_seq;
6203 core_rq->core_cookie = rq->core_cookie;
6204 core_rq->core_forceidle = rq->core_forceidle;
6205 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6207 /* install new leader */
6208 for_each_cpu(t, smt_mask) {
6214 sched_core_unlock(cpu, &flags);
6217 static inline void sched_core_cpu_dying(unsigned int cpu)
6219 struct rq *rq = cpu_rq(cpu);
6225 #else /* !CONFIG_SCHED_CORE */
6227 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6228 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6229 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6231 static struct task_struct *
6232 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6234 return __pick_next_task(rq, prev, rf);
6237 #endif /* CONFIG_SCHED_CORE */
6240 * Constants for the sched_mode argument of __schedule().
6242 * The mode argument allows RT enabled kernels to differentiate a
6243 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6244 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6245 * optimize the AND operation out and just check for zero.
6248 #define SM_PREEMPT 0x1
6249 #define SM_RTLOCK_WAIT 0x2
6251 #ifndef CONFIG_PREEMPT_RT
6252 # define SM_MASK_PREEMPT (~0U)
6254 # define SM_MASK_PREEMPT SM_PREEMPT
6258 * __schedule() is the main scheduler function.
6260 * The main means of driving the scheduler and thus entering this function are:
6262 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6264 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6265 * paths. For example, see arch/x86/entry_64.S.
6267 * To drive preemption between tasks, the scheduler sets the flag in timer
6268 * interrupt handler scheduler_tick().
6270 * 3. Wakeups don't really cause entry into schedule(). They add a
6271 * task to the run-queue and that's it.
6273 * Now, if the new task added to the run-queue preempts the current
6274 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6275 * called on the nearest possible occasion:
6277 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6279 * - in syscall or exception context, at the next outmost
6280 * preempt_enable(). (this might be as soon as the wake_up()'s
6283 * - in IRQ context, return from interrupt-handler to
6284 * preemptible context
6286 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6289 * - cond_resched() call
6290 * - explicit schedule() call
6291 * - return from syscall or exception to user-space
6292 * - return from interrupt-handler to user-space
6294 * WARNING: must be called with preemption disabled!
6296 static void __sched notrace __schedule(unsigned int sched_mode)
6298 struct task_struct *prev, *next;
6299 unsigned long *switch_count;
6300 unsigned long prev_state;
6305 cpu = smp_processor_id();
6309 schedule_debug(prev, !!sched_mode);
6311 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6314 local_irq_disable();
6315 rcu_note_context_switch(!!sched_mode);
6318 * Make sure that signal_pending_state()->signal_pending() below
6319 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6320 * done by the caller to avoid the race with signal_wake_up():
6322 * __set_current_state(@state) signal_wake_up()
6323 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6324 * wake_up_state(p, state)
6325 * LOCK rq->lock LOCK p->pi_state
6326 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6327 * if (signal_pending_state()) if (p->state & @state)
6329 * Also, the membarrier system call requires a full memory barrier
6330 * after coming from user-space, before storing to rq->curr.
6333 smp_mb__after_spinlock();
6335 /* Promote REQ to ACT */
6336 rq->clock_update_flags <<= 1;
6337 update_rq_clock(rq);
6339 switch_count = &prev->nivcsw;
6342 * We must load prev->state once (task_struct::state is volatile), such
6345 * - we form a control dependency vs deactivate_task() below.
6346 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
6348 prev_state = READ_ONCE(prev->__state);
6349 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6350 if (signal_pending_state(prev_state, prev)) {
6351 WRITE_ONCE(prev->__state, TASK_RUNNING);
6353 prev->sched_contributes_to_load =
6354 (prev_state & TASK_UNINTERRUPTIBLE) &&
6355 !(prev_state & TASK_NOLOAD) &&
6356 !(prev->flags & PF_FROZEN);
6358 if (prev->sched_contributes_to_load)
6359 rq->nr_uninterruptible++;
6362 * __schedule() ttwu()
6363 * prev_state = prev->state; if (p->on_rq && ...)
6364 * if (prev_state) goto out;
6365 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6366 * p->state = TASK_WAKING
6368 * Where __schedule() and ttwu() have matching control dependencies.
6370 * After this, schedule() must not care about p->state any more.
6372 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6374 if (prev->in_iowait) {
6375 atomic_inc(&rq->nr_iowait);
6376 delayacct_blkio_start();
6379 switch_count = &prev->nvcsw;
6382 next = pick_next_task(rq, prev, &rf);
6383 clear_tsk_need_resched(prev);
6384 clear_tsk_need_resched_lazy(prev);
6385 clear_preempt_need_resched();
6386 #ifdef CONFIG_SCHED_DEBUG
6387 rq->last_seen_need_resched_ns = 0;
6390 if (likely(prev != next)) {
6393 * RCU users of rcu_dereference(rq->curr) may not see
6394 * changes to task_struct made by pick_next_task().
6396 RCU_INIT_POINTER(rq->curr, next);
6398 * The membarrier system call requires each architecture
6399 * to have a full memory barrier after updating
6400 * rq->curr, before returning to user-space.
6402 * Here are the schemes providing that barrier on the
6403 * various architectures:
6404 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6405 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6406 * - finish_lock_switch() for weakly-ordered
6407 * architectures where spin_unlock is a full barrier,
6408 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6409 * is a RELEASE barrier),
6413 migrate_disable_switch(rq, prev);
6414 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6416 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next);
6418 /* Also unlocks the rq: */
6419 rq = context_switch(rq, prev, next, &rf);
6421 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6423 rq_unpin_lock(rq, &rf);
6424 __balance_callbacks(rq);
6425 raw_spin_rq_unlock_irq(rq);
6429 void __noreturn do_task_dead(void)
6431 /* Causes final put_task_struct in finish_task_switch(): */
6432 set_special_state(TASK_DEAD);
6434 /* Tell freezer to ignore us: */
6435 current->flags |= PF_NOFREEZE;
6437 __schedule(SM_NONE);
6440 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6445 static inline void sched_submit_work(struct task_struct *tsk)
6447 unsigned int task_flags;
6449 if (task_is_running(tsk))
6452 task_flags = tsk->flags;
6454 * If a worker went to sleep, notify and ask workqueue whether
6455 * it wants to wake up a task to maintain concurrency.
6456 * As this function is called inside the schedule() context,
6457 * we disable preemption to avoid it calling schedule() again
6458 * in the possible wakeup of a kworker and because wq_worker_sleeping()
6461 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6463 if (task_flags & PF_WQ_WORKER)
6464 wq_worker_sleeping(tsk);
6466 io_wq_worker_sleeping(tsk);
6467 preempt_enable_no_resched();
6471 * spinlock and rwlock must not flush block requests. This will
6472 * deadlock if the callback attempts to acquire a lock which is
6475 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
6478 * If we are going to sleep and we have plugged IO queued,
6479 * make sure to submit it to avoid deadlocks.
6481 if (blk_needs_flush_plug(tsk))
6482 blk_schedule_flush_plug(tsk);
6485 static void sched_update_worker(struct task_struct *tsk)
6487 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6488 if (tsk->flags & PF_WQ_WORKER)
6489 wq_worker_running(tsk);
6491 io_wq_worker_running(tsk);
6495 asmlinkage __visible void __sched schedule(void)
6497 struct task_struct *tsk = current;
6499 sched_submit_work(tsk);
6502 __schedule(SM_NONE);
6503 sched_preempt_enable_no_resched();
6504 } while (need_resched());
6505 sched_update_worker(tsk);
6507 EXPORT_SYMBOL(schedule);
6510 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6511 * state (have scheduled out non-voluntarily) by making sure that all
6512 * tasks have either left the run queue or have gone into user space.
6513 * As idle tasks do not do either, they must not ever be preempted
6514 * (schedule out non-voluntarily).
6516 * schedule_idle() is similar to schedule_preempt_disable() except that it
6517 * never enables preemption because it does not call sched_submit_work().
6519 void __sched schedule_idle(void)
6522 * As this skips calling sched_submit_work(), which the idle task does
6523 * regardless because that function is a nop when the task is in a
6524 * TASK_RUNNING state, make sure this isn't used someplace that the
6525 * current task can be in any other state. Note, idle is always in the
6526 * TASK_RUNNING state.
6528 WARN_ON_ONCE(current->__state);
6530 __schedule(SM_NONE);
6531 } while (need_resched());
6534 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6535 asmlinkage __visible void __sched schedule_user(void)
6538 * If we come here after a random call to set_need_resched(),
6539 * or we have been woken up remotely but the IPI has not yet arrived,
6540 * we haven't yet exited the RCU idle mode. Do it here manually until
6541 * we find a better solution.
6543 * NB: There are buggy callers of this function. Ideally we
6544 * should warn if prev_state != CONTEXT_USER, but that will trigger
6545 * too frequently to make sense yet.
6547 enum ctx_state prev_state = exception_enter();
6549 exception_exit(prev_state);
6554 * schedule_preempt_disabled - called with preemption disabled
6556 * Returns with preemption disabled. Note: preempt_count must be 1
6558 void __sched schedule_preempt_disabled(void)
6560 sched_preempt_enable_no_resched();
6565 #ifdef CONFIG_PREEMPT_RT
6566 void __sched notrace schedule_rtlock(void)
6570 __schedule(SM_RTLOCK_WAIT);
6571 sched_preempt_enable_no_resched();
6572 } while (need_resched());
6574 NOKPROBE_SYMBOL(schedule_rtlock);
6577 static void __sched notrace preempt_schedule_common(void)
6581 * Because the function tracer can trace preempt_count_sub()
6582 * and it also uses preempt_enable/disable_notrace(), if
6583 * NEED_RESCHED is set, the preempt_enable_notrace() called
6584 * by the function tracer will call this function again and
6585 * cause infinite recursion.
6587 * Preemption must be disabled here before the function
6588 * tracer can trace. Break up preempt_disable() into two
6589 * calls. One to disable preemption without fear of being
6590 * traced. The other to still record the preemption latency,
6591 * which can also be traced by the function tracer.
6593 preempt_disable_notrace();
6594 preempt_latency_start(1);
6595 __schedule(SM_PREEMPT);
6596 preempt_latency_stop(1);
6597 preempt_enable_no_resched_notrace();
6600 * Check again in case we missed a preemption opportunity
6601 * between schedule and now.
6603 } while (need_resched());
6606 #ifdef CONFIG_PREEMPT_LAZY
6608 * If TIF_NEED_RESCHED is then we allow to be scheduled away since this is
6609 * set by a RT task. Oterwise we try to avoid beeing scheduled out as long as
6610 * preempt_lazy_count counter >0.
6612 static __always_inline int preemptible_lazy(void)
6614 if (test_thread_flag(TIF_NEED_RESCHED))
6616 if (current_thread_info()->preempt_lazy_count)
6623 static inline int preemptible_lazy(void)
6630 #ifdef CONFIG_PREEMPTION
6632 * This is the entry point to schedule() from in-kernel preemption
6633 * off of preempt_enable.
6635 asmlinkage __visible void __sched notrace preempt_schedule(void)
6638 * If there is a non-zero preempt_count or interrupts are disabled,
6639 * we do not want to preempt the current task. Just return..
6641 if (likely(!preemptible()))
6643 if (!preemptible_lazy())
6645 preempt_schedule_common();
6647 NOKPROBE_SYMBOL(preempt_schedule);
6648 EXPORT_SYMBOL(preempt_schedule);
6650 #ifdef CONFIG_PREEMPT_DYNAMIC
6651 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6652 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6657 * preempt_schedule_notrace - preempt_schedule called by tracing
6659 * The tracing infrastructure uses preempt_enable_notrace to prevent
6660 * recursion and tracing preempt enabling caused by the tracing
6661 * infrastructure itself. But as tracing can happen in areas coming
6662 * from userspace or just about to enter userspace, a preempt enable
6663 * can occur before user_exit() is called. This will cause the scheduler
6664 * to be called when the system is still in usermode.
6666 * To prevent this, the preempt_enable_notrace will use this function
6667 * instead of preempt_schedule() to exit user context if needed before
6668 * calling the scheduler.
6670 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6672 enum ctx_state prev_ctx;
6674 if (likely(!preemptible()))
6677 if (!preemptible_lazy())
6682 * Because the function tracer can trace preempt_count_sub()
6683 * and it also uses preempt_enable/disable_notrace(), if
6684 * NEED_RESCHED is set, the preempt_enable_notrace() called
6685 * by the function tracer will call this function again and
6686 * cause infinite recursion.
6688 * Preemption must be disabled here before the function
6689 * tracer can trace. Break up preempt_disable() into two
6690 * calls. One to disable preemption without fear of being
6691 * traced. The other to still record the preemption latency,
6692 * which can also be traced by the function tracer.
6694 preempt_disable_notrace();
6695 preempt_latency_start(1);
6697 * Needs preempt disabled in case user_exit() is traced
6698 * and the tracer calls preempt_enable_notrace() causing
6699 * an infinite recursion.
6701 prev_ctx = exception_enter();
6702 __schedule(SM_PREEMPT);
6703 exception_exit(prev_ctx);
6705 preempt_latency_stop(1);
6706 preempt_enable_no_resched_notrace();
6707 } while (need_resched());
6709 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6711 #ifdef CONFIG_PREEMPT_DYNAMIC
6712 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6713 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6716 #endif /* CONFIG_PREEMPTION */
6718 #ifdef CONFIG_PREEMPT_DYNAMIC
6720 #include <linux/entry-common.h>
6725 * SC:preempt_schedule
6726 * SC:preempt_schedule_notrace
6727 * SC:irqentry_exit_cond_resched
6731 * cond_resched <- __cond_resched
6732 * might_resched <- RET0
6733 * preempt_schedule <- NOP
6734 * preempt_schedule_notrace <- NOP
6735 * irqentry_exit_cond_resched <- NOP
6738 * cond_resched <- __cond_resched
6739 * might_resched <- __cond_resched
6740 * preempt_schedule <- NOP
6741 * preempt_schedule_notrace <- NOP
6742 * irqentry_exit_cond_resched <- NOP
6745 * cond_resched <- RET0
6746 * might_resched <- RET0
6747 * preempt_schedule <- preempt_schedule
6748 * preempt_schedule_notrace <- preempt_schedule_notrace
6749 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6753 preempt_dynamic_none = 0,
6754 preempt_dynamic_voluntary,
6755 preempt_dynamic_full,
6758 int preempt_dynamic_mode = preempt_dynamic_full;
6760 int sched_dynamic_mode(const char *str)
6762 if (!strcmp(str, "none"))
6763 return preempt_dynamic_none;
6765 if (!strcmp(str, "voluntary"))
6766 return preempt_dynamic_voluntary;
6768 if (!strcmp(str, "full"))
6769 return preempt_dynamic_full;
6774 void sched_dynamic_update(int mode)
6777 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6778 * the ZERO state, which is invalid.
6780 static_call_update(cond_resched, __cond_resched);
6781 static_call_update(might_resched, __cond_resched);
6782 static_call_update(preempt_schedule, __preempt_schedule_func);
6783 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6784 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6787 case preempt_dynamic_none:
6788 static_call_update(cond_resched, __cond_resched);
6789 static_call_update(might_resched, (void *)&__static_call_return0);
6790 static_call_update(preempt_schedule, NULL);
6791 static_call_update(preempt_schedule_notrace, NULL);
6792 static_call_update(irqentry_exit_cond_resched, NULL);
6793 pr_info("Dynamic Preempt: none\n");
6796 case preempt_dynamic_voluntary:
6797 static_call_update(cond_resched, __cond_resched);
6798 static_call_update(might_resched, __cond_resched);
6799 static_call_update(preempt_schedule, NULL);
6800 static_call_update(preempt_schedule_notrace, NULL);
6801 static_call_update(irqentry_exit_cond_resched, NULL);
6802 pr_info("Dynamic Preempt: voluntary\n");
6805 case preempt_dynamic_full:
6806 static_call_update(cond_resched, (void *)&__static_call_return0);
6807 static_call_update(might_resched, (void *)&__static_call_return0);
6808 static_call_update(preempt_schedule, __preempt_schedule_func);
6809 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6810 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6811 pr_info("Dynamic Preempt: full\n");
6815 preempt_dynamic_mode = mode;
6818 static int __init setup_preempt_mode(char *str)
6820 int mode = sched_dynamic_mode(str);
6822 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6826 sched_dynamic_update(mode);
6829 __setup("preempt=", setup_preempt_mode);
6831 #endif /* CONFIG_PREEMPT_DYNAMIC */
6834 * This is the entry point to schedule() from kernel preemption
6835 * off of irq context.
6836 * Note, that this is called and return with irqs disabled. This will
6837 * protect us against recursive calling from irq.
6839 asmlinkage __visible void __sched preempt_schedule_irq(void)
6841 enum ctx_state prev_state;
6843 /* Catch callers which need to be fixed */
6844 BUG_ON(preempt_count() || !irqs_disabled());
6846 prev_state = exception_enter();
6851 __schedule(SM_PREEMPT);
6852 local_irq_disable();
6853 sched_preempt_enable_no_resched();
6854 } while (need_resched());
6856 exception_exit(prev_state);
6859 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6862 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6863 return try_to_wake_up(curr->private, mode, wake_flags);
6865 EXPORT_SYMBOL(default_wake_function);
6867 static void __setscheduler_prio(struct task_struct *p, int prio)
6870 p->sched_class = &dl_sched_class;
6871 else if (rt_prio(prio))
6872 p->sched_class = &rt_sched_class;
6874 p->sched_class = &fair_sched_class;
6879 #ifdef CONFIG_RT_MUTEXES
6881 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6884 prio = min(prio, pi_task->prio);
6889 static inline int rt_effective_prio(struct task_struct *p, int prio)
6891 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6893 return __rt_effective_prio(pi_task, prio);
6897 * rt_mutex_setprio - set the current priority of a task
6899 * @pi_task: donor task
6901 * This function changes the 'effective' priority of a task. It does
6902 * not touch ->normal_prio like __setscheduler().
6904 * Used by the rt_mutex code to implement priority inheritance
6905 * logic. Call site only calls if the priority of the task changed.
6907 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6909 int prio, oldprio, queued, running, queue_flag =
6910 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6911 const struct sched_class *prev_class;
6915 /* XXX used to be waiter->prio, not waiter->task->prio */
6916 prio = __rt_effective_prio(pi_task, p->normal_prio);
6919 * If nothing changed; bail early.
6921 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6924 rq = __task_rq_lock(p, &rf);
6925 update_rq_clock(rq);
6927 * Set under pi_lock && rq->lock, such that the value can be used under
6930 * Note that there is loads of tricky to make this pointer cache work
6931 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6932 * ensure a task is de-boosted (pi_task is set to NULL) before the
6933 * task is allowed to run again (and can exit). This ensures the pointer
6934 * points to a blocked task -- which guarantees the task is present.
6936 p->pi_top_task = pi_task;
6939 * For FIFO/RR we only need to set prio, if that matches we're done.
6941 if (prio == p->prio && !dl_prio(prio))
6945 * Idle task boosting is a nono in general. There is one
6946 * exception, when PREEMPT_RT and NOHZ is active:
6948 * The idle task calls get_next_timer_interrupt() and holds
6949 * the timer wheel base->lock on the CPU and another CPU wants
6950 * to access the timer (probably to cancel it). We can safely
6951 * ignore the boosting request, as the idle CPU runs this code
6952 * with interrupts disabled and will complete the lock
6953 * protected section without being interrupted. So there is no
6954 * real need to boost.
6956 if (unlikely(p == rq->idle)) {
6957 WARN_ON(p != rq->curr);
6958 WARN_ON(p->pi_blocked_on);
6962 trace_sched_pi_setprio(p, pi_task);
6965 if (oldprio == prio)
6966 queue_flag &= ~DEQUEUE_MOVE;
6968 prev_class = p->sched_class;
6969 queued = task_on_rq_queued(p);
6970 running = task_current(rq, p);
6972 dequeue_task(rq, p, queue_flag);
6974 put_prev_task(rq, p);
6977 * Boosting condition are:
6978 * 1. -rt task is running and holds mutex A
6979 * --> -dl task blocks on mutex A
6981 * 2. -dl task is running and holds mutex A
6982 * --> -dl task blocks on mutex A and could preempt the
6985 if (dl_prio(prio)) {
6986 if (!dl_prio(p->normal_prio) ||
6987 (pi_task && dl_prio(pi_task->prio) &&
6988 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6989 p->dl.pi_se = pi_task->dl.pi_se;
6990 queue_flag |= ENQUEUE_REPLENISH;
6992 p->dl.pi_se = &p->dl;
6994 } else if (rt_prio(prio)) {
6995 if (dl_prio(oldprio))
6996 p->dl.pi_se = &p->dl;
6998 queue_flag |= ENQUEUE_HEAD;
7000 if (dl_prio(oldprio))
7001 p->dl.pi_se = &p->dl;
7002 if (rt_prio(oldprio))
7006 __setscheduler_prio(p, prio);
7009 enqueue_task(rq, p, queue_flag);
7011 set_next_task(rq, p);
7013 check_class_changed(rq, p, prev_class, oldprio);
7015 /* Avoid rq from going away on us: */
7018 rq_unpin_lock(rq, &rf);
7019 __balance_callbacks(rq);
7020 raw_spin_rq_unlock(rq);
7025 static inline int rt_effective_prio(struct task_struct *p, int prio)
7031 void set_user_nice(struct task_struct *p, long nice)
7033 bool queued, running;
7038 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
7041 * We have to be careful, if called from sys_setpriority(),
7042 * the task might be in the middle of scheduling on another CPU.
7044 rq = task_rq_lock(p, &rf);
7045 update_rq_clock(rq);
7048 * The RT priorities are set via sched_setscheduler(), but we still
7049 * allow the 'normal' nice value to be set - but as expected
7050 * it won't have any effect on scheduling until the task is
7051 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
7053 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
7054 p->static_prio = NICE_TO_PRIO(nice);
7057 queued = task_on_rq_queued(p);
7058 running = task_current(rq, p);
7060 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
7062 put_prev_task(rq, p);
7064 p->static_prio = NICE_TO_PRIO(nice);
7065 set_load_weight(p, true);
7067 p->prio = effective_prio(p);
7070 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7072 set_next_task(rq, p);
7075 * If the task increased its priority or is running and
7076 * lowered its priority, then reschedule its CPU:
7078 p->sched_class->prio_changed(rq, p, old_prio);
7081 task_rq_unlock(rq, p, &rf);
7083 EXPORT_SYMBOL(set_user_nice);
7086 * can_nice - check if a task can reduce its nice value
7090 int can_nice(const struct task_struct *p, const int nice)
7092 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7093 int nice_rlim = nice_to_rlimit(nice);
7095 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
7096 capable(CAP_SYS_NICE));
7099 #ifdef __ARCH_WANT_SYS_NICE
7102 * sys_nice - change the priority of the current process.
7103 * @increment: priority increment
7105 * sys_setpriority is a more generic, but much slower function that
7106 * does similar things.
7108 SYSCALL_DEFINE1(nice, int, increment)
7113 * Setpriority might change our priority at the same moment.
7114 * We don't have to worry. Conceptually one call occurs first
7115 * and we have a single winner.
7117 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
7118 nice = task_nice(current) + increment;
7120 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
7121 if (increment < 0 && !can_nice(current, nice))
7124 retval = security_task_setnice(current, nice);
7128 set_user_nice(current, nice);
7135 * task_prio - return the priority value of a given task.
7136 * @p: the task in question.
7138 * Return: The priority value as seen by users in /proc.
7140 * sched policy return value kernel prio user prio/nice
7142 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7143 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7144 * deadline -101 -1 0
7146 int task_prio(const struct task_struct *p)
7148 return p->prio - MAX_RT_PRIO;
7152 * idle_cpu - is a given CPU idle currently?
7153 * @cpu: the processor in question.
7155 * Return: 1 if the CPU is currently idle. 0 otherwise.
7157 int idle_cpu(int cpu)
7159 struct rq *rq = cpu_rq(cpu);
7161 if (rq->curr != rq->idle)
7168 if (rq->ttwu_pending)
7176 * available_idle_cpu - is a given CPU idle for enqueuing work.
7177 * @cpu: the CPU in question.
7179 * Return: 1 if the CPU is currently idle. 0 otherwise.
7181 int available_idle_cpu(int cpu)
7186 if (vcpu_is_preempted(cpu))
7193 * idle_task - return the idle task for a given CPU.
7194 * @cpu: the processor in question.
7196 * Return: The idle task for the CPU @cpu.
7198 struct task_struct *idle_task(int cpu)
7200 return cpu_rq(cpu)->idle;
7205 * This function computes an effective utilization for the given CPU, to be
7206 * used for frequency selection given the linear relation: f = u * f_max.
7208 * The scheduler tracks the following metrics:
7210 * cpu_util_{cfs,rt,dl,irq}()
7213 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7214 * synchronized windows and are thus directly comparable.
7216 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7217 * which excludes things like IRQ and steal-time. These latter are then accrued
7218 * in the irq utilization.
7220 * The DL bandwidth number otoh is not a measured metric but a value computed
7221 * based on the task model parameters and gives the minimal utilization
7222 * required to meet deadlines.
7224 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7225 unsigned long max, enum cpu_util_type type,
7226 struct task_struct *p)
7228 unsigned long dl_util, util, irq;
7229 struct rq *rq = cpu_rq(cpu);
7231 if (!uclamp_is_used() &&
7232 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7237 * Early check to see if IRQ/steal time saturates the CPU, can be
7238 * because of inaccuracies in how we track these -- see
7239 * update_irq_load_avg().
7241 irq = cpu_util_irq(rq);
7242 if (unlikely(irq >= max))
7246 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7247 * CFS tasks and we use the same metric to track the effective
7248 * utilization (PELT windows are synchronized) we can directly add them
7249 * to obtain the CPU's actual utilization.
7251 * CFS and RT utilization can be boosted or capped, depending on
7252 * utilization clamp constraints requested by currently RUNNABLE
7254 * When there are no CFS RUNNABLE tasks, clamps are released and
7255 * frequency will be gracefully reduced with the utilization decay.
7257 util = util_cfs + cpu_util_rt(rq);
7258 if (type == FREQUENCY_UTIL)
7259 util = uclamp_rq_util_with(rq, util, p);
7261 dl_util = cpu_util_dl(rq);
7264 * For frequency selection we do not make cpu_util_dl() a permanent part
7265 * of this sum because we want to use cpu_bw_dl() later on, but we need
7266 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7267 * that we select f_max when there is no idle time.
7269 * NOTE: numerical errors or stop class might cause us to not quite hit
7270 * saturation when we should -- something for later.
7272 if (util + dl_util >= max)
7276 * OTOH, for energy computation we need the estimated running time, so
7277 * include util_dl and ignore dl_bw.
7279 if (type == ENERGY_UTIL)
7283 * There is still idle time; further improve the number by using the
7284 * irq metric. Because IRQ/steal time is hidden from the task clock we
7285 * need to scale the task numbers:
7288 * U' = irq + --------- * U
7291 util = scale_irq_capacity(util, irq, max);
7295 * Bandwidth required by DEADLINE must always be granted while, for
7296 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7297 * to gracefully reduce the frequency when no tasks show up for longer
7300 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7301 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7302 * an interface. So, we only do the latter for now.
7304 if (type == FREQUENCY_UTIL)
7305 util += cpu_bw_dl(rq);
7307 return min(max, util);
7310 unsigned long sched_cpu_util(int cpu, unsigned long max)
7312 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
7315 #endif /* CONFIG_SMP */
7318 * find_process_by_pid - find a process with a matching PID value.
7319 * @pid: the pid in question.
7321 * The task of @pid, if found. %NULL otherwise.
7323 static struct task_struct *find_process_by_pid(pid_t pid)
7325 return pid ? find_task_by_vpid(pid) : current;
7329 * sched_setparam() passes in -1 for its policy, to let the functions
7330 * it calls know not to change it.
7332 #define SETPARAM_POLICY -1
7334 static void __setscheduler_params(struct task_struct *p,
7335 const struct sched_attr *attr)
7337 int policy = attr->sched_policy;
7339 if (policy == SETPARAM_POLICY)
7344 if (dl_policy(policy))
7345 __setparam_dl(p, attr);
7346 else if (fair_policy(policy))
7347 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7350 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7351 * !rt_policy. Always setting this ensures that things like
7352 * getparam()/getattr() don't report silly values for !rt tasks.
7354 p->rt_priority = attr->sched_priority;
7355 p->normal_prio = normal_prio(p);
7356 set_load_weight(p, true);
7360 * Check the target process has a UID that matches the current process's:
7362 static bool check_same_owner(struct task_struct *p)
7364 const struct cred *cred = current_cred(), *pcred;
7368 pcred = __task_cred(p);
7369 match = (uid_eq(cred->euid, pcred->euid) ||
7370 uid_eq(cred->euid, pcred->uid));
7375 static int __sched_setscheduler(struct task_struct *p,
7376 const struct sched_attr *attr,
7379 int oldpolicy = -1, policy = attr->sched_policy;
7380 int retval, oldprio, newprio, queued, running;
7381 const struct sched_class *prev_class;
7382 struct callback_head *head;
7385 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7388 /* The pi code expects interrupts enabled */
7389 BUG_ON(pi && in_interrupt());
7391 /* Double check policy once rq lock held: */
7393 reset_on_fork = p->sched_reset_on_fork;
7394 policy = oldpolicy = p->policy;
7396 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7398 if (!valid_policy(policy))
7402 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7406 * Valid priorities for SCHED_FIFO and SCHED_RR are
7407 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7408 * SCHED_BATCH and SCHED_IDLE is 0.
7410 if (attr->sched_priority > MAX_RT_PRIO-1)
7412 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7413 (rt_policy(policy) != (attr->sched_priority != 0)))
7417 * Allow unprivileged RT tasks to decrease priority:
7419 if (user && !capable(CAP_SYS_NICE)) {
7420 if (fair_policy(policy)) {
7421 if (attr->sched_nice < task_nice(p) &&
7422 !can_nice(p, attr->sched_nice))
7426 if (rt_policy(policy)) {
7427 unsigned long rlim_rtprio =
7428 task_rlimit(p, RLIMIT_RTPRIO);
7430 /* Can't set/change the rt policy: */
7431 if (policy != p->policy && !rlim_rtprio)
7434 /* Can't increase priority: */
7435 if (attr->sched_priority > p->rt_priority &&
7436 attr->sched_priority > rlim_rtprio)
7441 * Can't set/change SCHED_DEADLINE policy at all for now
7442 * (safest behavior); in the future we would like to allow
7443 * unprivileged DL tasks to increase their relative deadline
7444 * or reduce their runtime (both ways reducing utilization)
7446 if (dl_policy(policy))
7450 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7451 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7453 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7454 if (!can_nice(p, task_nice(p)))
7458 /* Can't change other user's priorities: */
7459 if (!check_same_owner(p))
7462 /* Normal users shall not reset the sched_reset_on_fork flag: */
7463 if (p->sched_reset_on_fork && !reset_on_fork)
7468 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7471 retval = security_task_setscheduler(p);
7476 /* Update task specific "requested" clamps */
7477 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7478 retval = uclamp_validate(p, attr);
7487 * Make sure no PI-waiters arrive (or leave) while we are
7488 * changing the priority of the task:
7490 * To be able to change p->policy safely, the appropriate
7491 * runqueue lock must be held.
7493 rq = task_rq_lock(p, &rf);
7494 update_rq_clock(rq);
7497 * Changing the policy of the stop threads its a very bad idea:
7499 if (p == rq->stop) {
7505 * If not changing anything there's no need to proceed further,
7506 * but store a possible modification of reset_on_fork.
7508 if (unlikely(policy == p->policy)) {
7509 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7511 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7513 if (dl_policy(policy) && dl_param_changed(p, attr))
7515 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7518 p->sched_reset_on_fork = reset_on_fork;
7525 #ifdef CONFIG_RT_GROUP_SCHED
7527 * Do not allow realtime tasks into groups that have no runtime
7530 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7531 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7532 !task_group_is_autogroup(task_group(p))) {
7538 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7539 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7540 cpumask_t *span = rq->rd->span;
7543 * Don't allow tasks with an affinity mask smaller than
7544 * the entire root_domain to become SCHED_DEADLINE. We
7545 * will also fail if there's no bandwidth available.
7547 if (!cpumask_subset(span, p->cpus_ptr) ||
7548 rq->rd->dl_bw.bw == 0) {
7556 /* Re-check policy now with rq lock held: */
7557 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7558 policy = oldpolicy = -1;
7559 task_rq_unlock(rq, p, &rf);
7561 cpuset_read_unlock();
7566 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7567 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7570 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7575 p->sched_reset_on_fork = reset_on_fork;
7578 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7581 * Take priority boosted tasks into account. If the new
7582 * effective priority is unchanged, we just store the new
7583 * normal parameters and do not touch the scheduler class and
7584 * the runqueue. This will be done when the task deboost
7587 newprio = rt_effective_prio(p, newprio);
7588 if (newprio == oldprio)
7589 queue_flags &= ~DEQUEUE_MOVE;
7592 queued = task_on_rq_queued(p);
7593 running = task_current(rq, p);
7595 dequeue_task(rq, p, queue_flags);
7597 put_prev_task(rq, p);
7599 prev_class = p->sched_class;
7601 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7602 __setscheduler_params(p, attr);
7603 __setscheduler_prio(p, newprio);
7605 __setscheduler_uclamp(p, attr);
7609 * We enqueue to tail when the priority of a task is
7610 * increased (user space view).
7612 if (oldprio < p->prio)
7613 queue_flags |= ENQUEUE_HEAD;
7615 enqueue_task(rq, p, queue_flags);
7618 set_next_task(rq, p);
7620 check_class_changed(rq, p, prev_class, oldprio);
7622 /* Avoid rq from going away on us: */
7624 head = splice_balance_callbacks(rq);
7625 task_rq_unlock(rq, p, &rf);
7628 cpuset_read_unlock();
7629 rt_mutex_adjust_pi(p);
7632 /* Run balance callbacks after we've adjusted the PI chain: */
7633 balance_callbacks(rq, head);
7639 task_rq_unlock(rq, p, &rf);
7641 cpuset_read_unlock();
7645 static int _sched_setscheduler(struct task_struct *p, int policy,
7646 const struct sched_param *param, bool check)
7648 struct sched_attr attr = {
7649 .sched_policy = policy,
7650 .sched_priority = param->sched_priority,
7651 .sched_nice = PRIO_TO_NICE(p->static_prio),
7654 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7655 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7656 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7657 policy &= ~SCHED_RESET_ON_FORK;
7658 attr.sched_policy = policy;
7661 return __sched_setscheduler(p, &attr, check, true);
7664 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7665 * @p: the task in question.
7666 * @policy: new policy.
7667 * @param: structure containing the new RT priority.
7669 * Use sched_set_fifo(), read its comment.
7671 * Return: 0 on success. An error code otherwise.
7673 * NOTE that the task may be already dead.
7675 int sched_setscheduler(struct task_struct *p, int policy,
7676 const struct sched_param *param)
7678 return _sched_setscheduler(p, policy, param, true);
7680 EXPORT_SYMBOL_GPL(sched_setscheduler);
7682 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7684 return __sched_setscheduler(p, attr, true, true);
7686 EXPORT_SYMBOL_GPL(sched_setattr);
7688 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7690 return __sched_setscheduler(p, attr, false, true);
7692 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7695 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7696 * @p: the task in question.
7697 * @policy: new policy.
7698 * @param: structure containing the new RT priority.
7700 * Just like sched_setscheduler, only don't bother checking if the
7701 * current context has permission. For example, this is needed in
7702 * stop_machine(): we create temporary high priority worker threads,
7703 * but our caller might not have that capability.
7705 * Return: 0 on success. An error code otherwise.
7707 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7708 const struct sched_param *param)
7710 return _sched_setscheduler(p, policy, param, false);
7712 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
7715 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7716 * incapable of resource management, which is the one thing an OS really should
7719 * This is of course the reason it is limited to privileged users only.
7721 * Worse still; it is fundamentally impossible to compose static priority
7722 * workloads. You cannot take two correctly working static prio workloads
7723 * and smash them together and still expect them to work.
7725 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7729 * The administrator _MUST_ configure the system, the kernel simply doesn't
7730 * know enough information to make a sensible choice.
7732 void sched_set_fifo(struct task_struct *p)
7734 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7735 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7737 EXPORT_SYMBOL_GPL(sched_set_fifo);
7740 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7742 void sched_set_fifo_low(struct task_struct *p)
7744 struct sched_param sp = { .sched_priority = 1 };
7745 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7747 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7749 void sched_set_normal(struct task_struct *p, int nice)
7751 struct sched_attr attr = {
7752 .sched_policy = SCHED_NORMAL,
7755 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7757 EXPORT_SYMBOL_GPL(sched_set_normal);
7760 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7762 struct sched_param lparam;
7763 struct task_struct *p;
7766 if (!param || pid < 0)
7768 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7773 p = find_process_by_pid(pid);
7779 retval = sched_setscheduler(p, policy, &lparam);
7787 * Mimics kernel/events/core.c perf_copy_attr().
7789 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7794 /* Zero the full structure, so that a short copy will be nice: */
7795 memset(attr, 0, sizeof(*attr));
7797 ret = get_user(size, &uattr->size);
7801 /* ABI compatibility quirk: */
7803 size = SCHED_ATTR_SIZE_VER0;
7804 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7807 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7814 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7815 size < SCHED_ATTR_SIZE_VER1)
7819 * XXX: Do we want to be lenient like existing syscalls; or do we want
7820 * to be strict and return an error on out-of-bounds values?
7822 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7827 put_user(sizeof(*attr), &uattr->size);
7831 static void get_params(struct task_struct *p, struct sched_attr *attr)
7833 if (task_has_dl_policy(p))
7834 __getparam_dl(p, attr);
7835 else if (task_has_rt_policy(p))
7836 attr->sched_priority = p->rt_priority;
7838 attr->sched_nice = task_nice(p);
7842 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7843 * @pid: the pid in question.
7844 * @policy: new policy.
7845 * @param: structure containing the new RT priority.
7847 * Return: 0 on success. An error code otherwise.
7849 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7854 return do_sched_setscheduler(pid, policy, param);
7858 * sys_sched_setparam - set/change the RT priority of a thread
7859 * @pid: the pid in question.
7860 * @param: structure containing the new RT priority.
7862 * Return: 0 on success. An error code otherwise.
7864 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7866 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7870 * sys_sched_setattr - same as above, but with extended sched_attr
7871 * @pid: the pid in question.
7872 * @uattr: structure containing the extended parameters.
7873 * @flags: for future extension.
7875 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7876 unsigned int, flags)
7878 struct sched_attr attr;
7879 struct task_struct *p;
7882 if (!uattr || pid < 0 || flags)
7885 retval = sched_copy_attr(uattr, &attr);
7889 if ((int)attr.sched_policy < 0)
7891 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7892 attr.sched_policy = SETPARAM_POLICY;
7896 p = find_process_by_pid(pid);
7902 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7903 get_params(p, &attr);
7904 retval = sched_setattr(p, &attr);
7912 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7913 * @pid: the pid in question.
7915 * Return: On success, the policy of the thread. Otherwise, a negative error
7918 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7920 struct task_struct *p;
7928 p = find_process_by_pid(pid);
7930 retval = security_task_getscheduler(p);
7933 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7940 * sys_sched_getparam - get the RT priority of a thread
7941 * @pid: the pid in question.
7942 * @param: structure containing the RT priority.
7944 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7947 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7949 struct sched_param lp = { .sched_priority = 0 };
7950 struct task_struct *p;
7953 if (!param || pid < 0)
7957 p = find_process_by_pid(pid);
7962 retval = security_task_getscheduler(p);
7966 if (task_has_rt_policy(p))
7967 lp.sched_priority = p->rt_priority;
7971 * This one might sleep, we cannot do it with a spinlock held ...
7973 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7983 * Copy the kernel size attribute structure (which might be larger
7984 * than what user-space knows about) to user-space.
7986 * Note that all cases are valid: user-space buffer can be larger or
7987 * smaller than the kernel-space buffer. The usual case is that both
7988 * have the same size.
7991 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7992 struct sched_attr *kattr,
7995 unsigned int ksize = sizeof(*kattr);
7997 if (!access_ok(uattr, usize))
8001 * sched_getattr() ABI forwards and backwards compatibility:
8003 * If usize == ksize then we just copy everything to user-space and all is good.
8005 * If usize < ksize then we only copy as much as user-space has space for,
8006 * this keeps ABI compatibility as well. We skip the rest.
8008 * If usize > ksize then user-space is using a newer version of the ABI,
8009 * which part the kernel doesn't know about. Just ignore it - tooling can
8010 * detect the kernel's knowledge of attributes from the attr->size value
8011 * which is set to ksize in this case.
8013 kattr->size = min(usize, ksize);
8015 if (copy_to_user(uattr, kattr, kattr->size))
8022 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
8023 * @pid: the pid in question.
8024 * @uattr: structure containing the extended parameters.
8025 * @usize: sizeof(attr) for fwd/bwd comp.
8026 * @flags: for future extension.
8028 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
8029 unsigned int, usize, unsigned int, flags)
8031 struct sched_attr kattr = { };
8032 struct task_struct *p;
8035 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8036 usize < SCHED_ATTR_SIZE_VER0 || flags)
8040 p = find_process_by_pid(pid);
8045 retval = security_task_getscheduler(p);
8049 kattr.sched_policy = p->policy;
8050 if (p->sched_reset_on_fork)
8051 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
8052 get_params(p, &kattr);
8053 kattr.sched_flags &= SCHED_FLAG_ALL;
8055 #ifdef CONFIG_UCLAMP_TASK
8057 * This could race with another potential updater, but this is fine
8058 * because it'll correctly read the old or the new value. We don't need
8059 * to guarantee who wins the race as long as it doesn't return garbage.
8061 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8062 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
8067 return sched_attr_copy_to_user(uattr, &kattr, usize);
8075 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8080 * If the task isn't a deadline task or admission control is
8081 * disabled then we don't care about affinity changes.
8083 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8087 * Since bandwidth control happens on root_domain basis,
8088 * if admission test is enabled, we only admit -deadline
8089 * tasks allowed to run on all the CPUs in the task's
8093 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8101 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
8104 cpumask_var_t cpus_allowed, new_mask;
8106 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8109 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8111 goto out_free_cpus_allowed;
8114 cpuset_cpus_allowed(p, cpus_allowed);
8115 cpumask_and(new_mask, mask, cpus_allowed);
8117 retval = dl_task_check_affinity(p, new_mask);
8119 goto out_free_new_mask;
8121 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
8123 goto out_free_new_mask;
8125 cpuset_cpus_allowed(p, cpus_allowed);
8126 if (!cpumask_subset(new_mask, cpus_allowed)) {
8128 * We must have raced with a concurrent cpuset update.
8129 * Just reset the cpumask to the cpuset's cpus_allowed.
8131 cpumask_copy(new_mask, cpus_allowed);
8136 free_cpumask_var(new_mask);
8137 out_free_cpus_allowed:
8138 free_cpumask_var(cpus_allowed);
8142 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8144 struct task_struct *p;
8149 p = find_process_by_pid(pid);
8155 /* Prevent p going away */
8159 if (p->flags & PF_NO_SETAFFINITY) {
8164 if (!check_same_owner(p)) {
8166 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8174 retval = security_task_setscheduler(p);
8178 retval = __sched_setaffinity(p, in_mask);
8184 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8185 struct cpumask *new_mask)
8187 if (len < cpumask_size())
8188 cpumask_clear(new_mask);
8189 else if (len > cpumask_size())
8190 len = cpumask_size();
8192 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8196 * sys_sched_setaffinity - set the CPU affinity of a process
8197 * @pid: pid of the process
8198 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8199 * @user_mask_ptr: user-space pointer to the new CPU mask
8201 * Return: 0 on success. An error code otherwise.
8203 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8204 unsigned long __user *, user_mask_ptr)
8206 cpumask_var_t new_mask;
8209 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8212 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8214 retval = sched_setaffinity(pid, new_mask);
8215 free_cpumask_var(new_mask);
8219 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8221 struct task_struct *p;
8222 unsigned long flags;
8228 p = find_process_by_pid(pid);
8232 retval = security_task_getscheduler(p);
8236 raw_spin_lock_irqsave(&p->pi_lock, flags);
8237 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8238 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8247 * sys_sched_getaffinity - get the CPU affinity of a process
8248 * @pid: pid of the process
8249 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8250 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8252 * Return: size of CPU mask copied to user_mask_ptr on success. An
8253 * error code otherwise.
8255 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8256 unsigned long __user *, user_mask_ptr)
8261 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8263 if (len & (sizeof(unsigned long)-1))
8266 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8269 ret = sched_getaffinity(pid, mask);
8271 unsigned int retlen = min(len, cpumask_size());
8273 if (copy_to_user(user_mask_ptr, mask, retlen))
8278 free_cpumask_var(mask);
8283 static void do_sched_yield(void)
8288 rq = this_rq_lock_irq(&rf);
8290 schedstat_inc(rq->yld_count);
8291 current->sched_class->yield_task(rq);
8294 rq_unlock_irq(rq, &rf);
8295 sched_preempt_enable_no_resched();
8301 * sys_sched_yield - yield the current processor to other threads.
8303 * This function yields the current CPU to other tasks. If there are no
8304 * other threads running on this CPU then this function will return.
8308 SYSCALL_DEFINE0(sched_yield)
8314 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8315 int __sched __cond_resched(void)
8317 if (should_resched(0)) {
8318 preempt_schedule_common();
8322 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8323 * whether the current CPU is in an RCU read-side critical section,
8324 * so the tick can report quiescent states even for CPUs looping
8325 * in kernel context. In contrast, in non-preemptible kernels,
8326 * RCU readers leave no in-memory hints, which means that CPU-bound
8327 * processes executing in kernel context might never report an
8328 * RCU quiescent state. Therefore, the following code causes
8329 * cond_resched() to report a quiescent state, but only when RCU
8330 * is in urgent need of one.
8332 #ifndef CONFIG_PREEMPT_RCU
8337 EXPORT_SYMBOL(__cond_resched);
8340 #ifdef CONFIG_PREEMPT_DYNAMIC
8341 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8342 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8344 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8345 EXPORT_STATIC_CALL_TRAMP(might_resched);
8349 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8350 * call schedule, and on return reacquire the lock.
8352 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8353 * operations here to prevent schedule() from being called twice (once via
8354 * spin_unlock(), once by hand).
8356 int __cond_resched_lock(spinlock_t *lock)
8358 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8361 lockdep_assert_held(lock);
8363 if (spin_needbreak(lock) || resched) {
8365 if (!_cond_resched())
8372 EXPORT_SYMBOL(__cond_resched_lock);
8374 int __cond_resched_rwlock_read(rwlock_t *lock)
8376 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8379 lockdep_assert_held_read(lock);
8381 if (rwlock_needbreak(lock) || resched) {
8383 if (!_cond_resched())
8390 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8392 int __cond_resched_rwlock_write(rwlock_t *lock)
8394 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8397 lockdep_assert_held_write(lock);
8399 if (rwlock_needbreak(lock) || resched) {
8401 if (!_cond_resched())
8408 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8411 * yield - yield the current processor to other threads.
8413 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8415 * The scheduler is at all times free to pick the calling task as the most
8416 * eligible task to run, if removing the yield() call from your code breaks
8417 * it, it's already broken.
8419 * Typical broken usage is:
8424 * where one assumes that yield() will let 'the other' process run that will
8425 * make event true. If the current task is a SCHED_FIFO task that will never
8426 * happen. Never use yield() as a progress guarantee!!
8428 * If you want to use yield() to wait for something, use wait_event().
8429 * If you want to use yield() to be 'nice' for others, use cond_resched().
8430 * If you still want to use yield(), do not!
8432 void __sched yield(void)
8434 set_current_state(TASK_RUNNING);
8437 EXPORT_SYMBOL(yield);
8440 * yield_to - yield the current processor to another thread in
8441 * your thread group, or accelerate that thread toward the
8442 * processor it's on.
8444 * @preempt: whether task preemption is allowed or not
8446 * It's the caller's job to ensure that the target task struct
8447 * can't go away on us before we can do any checks.
8450 * true (>0) if we indeed boosted the target task.
8451 * false (0) if we failed to boost the target.
8452 * -ESRCH if there's no task to yield to.
8454 int __sched yield_to(struct task_struct *p, bool preempt)
8456 struct task_struct *curr = current;
8457 struct rq *rq, *p_rq;
8458 unsigned long flags;
8461 local_irq_save(flags);
8467 * If we're the only runnable task on the rq and target rq also
8468 * has only one task, there's absolutely no point in yielding.
8470 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8475 double_rq_lock(rq, p_rq);
8476 if (task_rq(p) != p_rq) {
8477 double_rq_unlock(rq, p_rq);
8481 if (!curr->sched_class->yield_to_task)
8484 if (curr->sched_class != p->sched_class)
8487 if (task_running(p_rq, p) || !task_is_running(p))
8490 yielded = curr->sched_class->yield_to_task(rq, p);
8492 schedstat_inc(rq->yld_count);
8494 * Make p's CPU reschedule; pick_next_entity takes care of
8497 if (preempt && rq != p_rq)
8502 double_rq_unlock(rq, p_rq);
8504 local_irq_restore(flags);
8511 EXPORT_SYMBOL_GPL(yield_to);
8513 int io_schedule_prepare(void)
8515 int old_iowait = current->in_iowait;
8517 current->in_iowait = 1;
8518 blk_schedule_flush_plug(current);
8523 void io_schedule_finish(int token)
8525 current->in_iowait = token;
8529 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8530 * that process accounting knows that this is a task in IO wait state.
8532 long __sched io_schedule_timeout(long timeout)
8537 token = io_schedule_prepare();
8538 ret = schedule_timeout(timeout);
8539 io_schedule_finish(token);
8543 EXPORT_SYMBOL(io_schedule_timeout);
8545 void __sched io_schedule(void)
8549 token = io_schedule_prepare();
8551 io_schedule_finish(token);
8553 EXPORT_SYMBOL(io_schedule);
8556 * sys_sched_get_priority_max - return maximum RT priority.
8557 * @policy: scheduling class.
8559 * Return: On success, this syscall returns the maximum
8560 * rt_priority that can be used by a given scheduling class.
8561 * On failure, a negative error code is returned.
8563 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8570 ret = MAX_RT_PRIO-1;
8572 case SCHED_DEADLINE:
8583 * sys_sched_get_priority_min - return minimum RT priority.
8584 * @policy: scheduling class.
8586 * Return: On success, this syscall returns the minimum
8587 * rt_priority that can be used by a given scheduling class.
8588 * On failure, a negative error code is returned.
8590 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8599 case SCHED_DEADLINE:
8608 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8610 struct task_struct *p;
8611 unsigned int time_slice;
8621 p = find_process_by_pid(pid);
8625 retval = security_task_getscheduler(p);
8629 rq = task_rq_lock(p, &rf);
8631 if (p->sched_class->get_rr_interval)
8632 time_slice = p->sched_class->get_rr_interval(rq, p);
8633 task_rq_unlock(rq, p, &rf);
8636 jiffies_to_timespec64(time_slice, t);
8645 * sys_sched_rr_get_interval - return the default timeslice of a process.
8646 * @pid: pid of the process.
8647 * @interval: userspace pointer to the timeslice value.
8649 * this syscall writes the default timeslice value of a given process
8650 * into the user-space timespec buffer. A value of '0' means infinity.
8652 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8655 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8656 struct __kernel_timespec __user *, interval)
8658 struct timespec64 t;
8659 int retval = sched_rr_get_interval(pid, &t);
8662 retval = put_timespec64(&t, interval);
8667 #ifdef CONFIG_COMPAT_32BIT_TIME
8668 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8669 struct old_timespec32 __user *, interval)
8671 struct timespec64 t;
8672 int retval = sched_rr_get_interval(pid, &t);
8675 retval = put_old_timespec32(&t, interval);
8680 void sched_show_task(struct task_struct *p)
8682 unsigned long free = 0;
8685 if (!try_get_task_stack(p))
8688 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8690 if (task_is_running(p))
8691 pr_cont(" running task ");
8692 #ifdef CONFIG_DEBUG_STACK_USAGE
8693 free = stack_not_used(p);
8698 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8700 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8701 free, task_pid_nr(p), ppid,
8702 (unsigned long)task_thread_info(p)->flags);
8704 print_worker_info(KERN_INFO, p);
8705 print_stop_info(KERN_INFO, p);
8706 show_stack(p, NULL, KERN_INFO);
8709 EXPORT_SYMBOL_GPL(sched_show_task);
8712 state_filter_match(unsigned long state_filter, struct task_struct *p)
8714 unsigned int state = READ_ONCE(p->__state);
8716 /* no filter, everything matches */
8720 /* filter, but doesn't match */
8721 if (!(state & state_filter))
8725 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8728 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8735 void show_state_filter(unsigned int state_filter)
8737 struct task_struct *g, *p;
8740 for_each_process_thread(g, p) {
8742 * reset the NMI-timeout, listing all files on a slow
8743 * console might take a lot of time:
8744 * Also, reset softlockup watchdogs on all CPUs, because
8745 * another CPU might be blocked waiting for us to process
8748 touch_nmi_watchdog();
8749 touch_all_softlockup_watchdogs();
8750 if (state_filter_match(state_filter, p))
8754 #ifdef CONFIG_SCHED_DEBUG
8756 sysrq_sched_debug_show();
8760 * Only show locks if all tasks are dumped:
8763 debug_show_all_locks();
8767 * init_idle - set up an idle thread for a given CPU
8768 * @idle: task in question
8769 * @cpu: CPU the idle task belongs to
8771 * NOTE: this function does not set the idle thread's NEED_RESCHED
8772 * flag, to make booting more robust.
8774 void __init init_idle(struct task_struct *idle, int cpu)
8776 struct rq *rq = cpu_rq(cpu);
8777 unsigned long flags;
8779 __sched_fork(0, idle);
8782 * The idle task doesn't need the kthread struct to function, but it
8783 * is dressed up as a per-CPU kthread and thus needs to play the part
8784 * if we want to avoid special-casing it in code that deals with per-CPU
8787 set_kthread_struct(idle);
8789 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8790 raw_spin_rq_lock(rq);
8792 idle->__state = TASK_RUNNING;
8793 idle->se.exec_start = sched_clock();
8795 * PF_KTHREAD should already be set at this point; regardless, make it
8796 * look like a proper per-CPU kthread.
8798 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8799 kthread_set_per_cpu(idle, cpu);
8803 * It's possible that init_idle() gets called multiple times on a task,
8804 * in that case do_set_cpus_allowed() will not do the right thing.
8806 * And since this is boot we can forgo the serialization.
8808 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8811 * We're having a chicken and egg problem, even though we are
8812 * holding rq->lock, the CPU isn't yet set to this CPU so the
8813 * lockdep check in task_group() will fail.
8815 * Similar case to sched_fork(). / Alternatively we could
8816 * use task_rq_lock() here and obtain the other rq->lock.
8821 __set_task_cpu(idle, cpu);
8825 rcu_assign_pointer(rq->curr, idle);
8826 idle->on_rq = TASK_ON_RQ_QUEUED;
8830 raw_spin_rq_unlock(rq);
8831 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8833 /* Set the preempt count _outside_ the spinlocks! */
8834 init_idle_preempt_count(idle, cpu);
8835 #ifdef CONFIG_HAVE_PREEMPT_LAZY
8836 task_thread_info(idle)->preempt_lazy_count = 0;
8839 * The idle tasks have their own, simple scheduling class:
8841 idle->sched_class = &idle_sched_class;
8842 ftrace_graph_init_idle_task(idle, cpu);
8843 vtime_init_idle(idle, cpu);
8845 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8851 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8852 const struct cpumask *trial)
8856 if (!cpumask_weight(cur))
8859 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8864 int task_can_attach(struct task_struct *p,
8865 const struct cpumask *cs_effective_cpus)
8870 * Kthreads which disallow setaffinity shouldn't be moved
8871 * to a new cpuset; we don't want to change their CPU
8872 * affinity and isolating such threads by their set of
8873 * allowed nodes is unnecessary. Thus, cpusets are not
8874 * applicable for such threads. This prevents checking for
8875 * success of set_cpus_allowed_ptr() on all attached tasks
8876 * before cpus_mask may be changed.
8878 if (p->flags & PF_NO_SETAFFINITY) {
8883 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8884 cs_effective_cpus)) {
8885 int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
8887 if (unlikely(cpu >= nr_cpu_ids))
8889 ret = dl_cpu_busy(cpu, p);
8896 bool sched_smp_initialized __read_mostly;
8898 #ifdef CONFIG_NUMA_BALANCING
8899 /* Migrate current task p to target_cpu */
8900 int migrate_task_to(struct task_struct *p, int target_cpu)
8902 struct migration_arg arg = { p, target_cpu };
8903 int curr_cpu = task_cpu(p);
8905 if (curr_cpu == target_cpu)
8908 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8911 /* TODO: This is not properly updating schedstats */
8913 trace_sched_move_numa(p, curr_cpu, target_cpu);
8914 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8918 * Requeue a task on a given node and accurately track the number of NUMA
8919 * tasks on the runqueues
8921 void sched_setnuma(struct task_struct *p, int nid)
8923 bool queued, running;
8927 rq = task_rq_lock(p, &rf);
8928 queued = task_on_rq_queued(p);
8929 running = task_current(rq, p);
8932 dequeue_task(rq, p, DEQUEUE_SAVE);
8934 put_prev_task(rq, p);
8936 p->numa_preferred_nid = nid;
8939 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8941 set_next_task(rq, p);
8942 task_rq_unlock(rq, p, &rf);
8944 #endif /* CONFIG_NUMA_BALANCING */
8946 #ifdef CONFIG_HOTPLUG_CPU
8948 * Ensure that the idle task is using init_mm right before its CPU goes
8951 void idle_task_exit(void)
8953 struct mm_struct *mm = current->active_mm;
8955 BUG_ON(cpu_online(smp_processor_id()));
8956 BUG_ON(current != this_rq()->idle);
8958 if (mm != &init_mm) {
8959 switch_mm(mm, &init_mm, current);
8960 finish_arch_post_lock_switch();
8963 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8966 static int __balance_push_cpu_stop(void *arg)
8968 struct task_struct *p = arg;
8969 struct rq *rq = this_rq();
8973 raw_spin_lock_irq(&p->pi_lock);
8976 update_rq_clock(rq);
8978 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8979 cpu = select_fallback_rq(rq->cpu, p);
8980 rq = __migrate_task(rq, &rf, p, cpu);
8984 raw_spin_unlock_irq(&p->pi_lock);
8991 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8994 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8996 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8997 * effective when the hotplug motion is down.
8999 static void balance_push(struct rq *rq)
9001 struct task_struct *push_task = rq->curr;
9003 lockdep_assert_rq_held(rq);
9006 * Ensure the thing is persistent until balance_push_set(.on = false);
9008 rq->balance_callback = &balance_push_callback;
9011 * Only active while going offline and when invoked on the outgoing
9014 if (!cpu_dying(rq->cpu) || rq != this_rq())
9018 * Both the cpu-hotplug and stop task are in this case and are
9019 * required to complete the hotplug process.
9021 if (kthread_is_per_cpu(push_task) ||
9022 is_migration_disabled(push_task)) {
9025 * If this is the idle task on the outgoing CPU try to wake
9026 * up the hotplug control thread which might wait for the
9027 * last task to vanish. The rcuwait_active() check is
9028 * accurate here because the waiter is pinned on this CPU
9029 * and can't obviously be running in parallel.
9031 * On RT kernels this also has to check whether there are
9032 * pinned and scheduled out tasks on the runqueue. They
9033 * need to leave the migrate disabled section first.
9035 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9036 rcuwait_active(&rq->hotplug_wait)) {
9037 raw_spin_rq_unlock(rq);
9038 rcuwait_wake_up(&rq->hotplug_wait);
9039 raw_spin_rq_lock(rq);
9044 get_task_struct(push_task);
9046 * Temporarily drop rq->lock such that we can wake-up the stop task.
9047 * Both preemption and IRQs are still disabled.
9049 raw_spin_rq_unlock(rq);
9050 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9051 this_cpu_ptr(&push_work));
9053 * At this point need_resched() is true and we'll take the loop in
9054 * schedule(). The next pick is obviously going to be the stop task
9055 * which kthread_is_per_cpu() and will push this task away.
9057 raw_spin_rq_lock(rq);
9060 static void balance_push_set(int cpu, bool on)
9062 struct rq *rq = cpu_rq(cpu);
9065 rq_lock_irqsave(rq, &rf);
9067 WARN_ON_ONCE(rq->balance_callback);
9068 rq->balance_callback = &balance_push_callback;
9069 } else if (rq->balance_callback == &balance_push_callback) {
9070 rq->balance_callback = NULL;
9072 rq_unlock_irqrestore(rq, &rf);
9076 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9077 * inactive. All tasks which are not per CPU kernel threads are either
9078 * pushed off this CPU now via balance_push() or placed on a different CPU
9079 * during wakeup. Wait until the CPU is quiescent.
9081 static void balance_hotplug_wait(void)
9083 struct rq *rq = this_rq();
9085 rcuwait_wait_event(&rq->hotplug_wait,
9086 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
9087 TASK_UNINTERRUPTIBLE);
9092 static inline void balance_push(struct rq *rq)
9096 static inline void balance_push_set(int cpu, bool on)
9100 static inline void balance_hotplug_wait(void)
9104 #endif /* CONFIG_HOTPLUG_CPU */
9106 void set_rq_online(struct rq *rq)
9109 const struct sched_class *class;
9111 cpumask_set_cpu(rq->cpu, rq->rd->online);
9114 for_each_class(class) {
9115 if (class->rq_online)
9116 class->rq_online(rq);
9121 void set_rq_offline(struct rq *rq)
9124 const struct sched_class *class;
9126 for_each_class(class) {
9127 if (class->rq_offline)
9128 class->rq_offline(rq);
9131 cpumask_clear_cpu(rq->cpu, rq->rd->online);
9137 * used to mark begin/end of suspend/resume:
9139 static int num_cpus_frozen;
9142 * Update cpusets according to cpu_active mask. If cpusets are
9143 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9144 * around partition_sched_domains().
9146 * If we come here as part of a suspend/resume, don't touch cpusets because we
9147 * want to restore it back to its original state upon resume anyway.
9149 static void cpuset_cpu_active(void)
9151 if (cpuhp_tasks_frozen) {
9153 * num_cpus_frozen tracks how many CPUs are involved in suspend
9154 * resume sequence. As long as this is not the last online
9155 * operation in the resume sequence, just build a single sched
9156 * domain, ignoring cpusets.
9158 partition_sched_domains(1, NULL, NULL);
9159 if (--num_cpus_frozen)
9162 * This is the last CPU online operation. So fall through and
9163 * restore the original sched domains by considering the
9164 * cpuset configurations.
9166 cpuset_force_rebuild();
9168 cpuset_update_active_cpus();
9171 static int cpuset_cpu_inactive(unsigned int cpu)
9173 if (!cpuhp_tasks_frozen) {
9174 int ret = dl_cpu_busy(cpu, NULL);
9178 cpuset_update_active_cpus();
9181 partition_sched_domains(1, NULL, NULL);
9186 int sched_cpu_activate(unsigned int cpu)
9188 struct rq *rq = cpu_rq(cpu);
9192 * Clear the balance_push callback and prepare to schedule
9195 balance_push_set(cpu, false);
9197 #ifdef CONFIG_SCHED_SMT
9199 * When going up, increment the number of cores with SMT present.
9201 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9202 static_branch_inc_cpuslocked(&sched_smt_present);
9204 set_cpu_active(cpu, true);
9206 if (sched_smp_initialized) {
9207 sched_domains_numa_masks_set(cpu);
9208 cpuset_cpu_active();
9212 * Put the rq online, if not already. This happens:
9214 * 1) In the early boot process, because we build the real domains
9215 * after all CPUs have been brought up.
9217 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9220 rq_lock_irqsave(rq, &rf);
9222 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9225 rq_unlock_irqrestore(rq, &rf);
9230 int sched_cpu_deactivate(unsigned int cpu)
9232 struct rq *rq = cpu_rq(cpu);
9237 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9238 * load balancing when not active
9240 nohz_balance_exit_idle(rq);
9242 set_cpu_active(cpu, false);
9245 * From this point forward, this CPU will refuse to run any task that
9246 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9247 * push those tasks away until this gets cleared, see
9248 * sched_cpu_dying().
9250 balance_push_set(cpu, true);
9253 * We've cleared cpu_active_mask / set balance_push, wait for all
9254 * preempt-disabled and RCU users of this state to go away such that
9255 * all new such users will observe it.
9257 * Specifically, we rely on ttwu to no longer target this CPU, see
9258 * ttwu_queue_cond() and is_cpu_allowed().
9260 * Do sync before park smpboot threads to take care the rcu boost case.
9264 rq_lock_irqsave(rq, &rf);
9266 update_rq_clock(rq);
9267 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9270 rq_unlock_irqrestore(rq, &rf);
9272 #ifdef CONFIG_SCHED_SMT
9274 * When going down, decrement the number of cores with SMT present.
9276 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9277 static_branch_dec_cpuslocked(&sched_smt_present);
9279 sched_core_cpu_deactivate(cpu);
9282 if (!sched_smp_initialized)
9285 ret = cpuset_cpu_inactive(cpu);
9287 balance_push_set(cpu, false);
9288 set_cpu_active(cpu, true);
9291 sched_domains_numa_masks_clear(cpu);
9295 static void sched_rq_cpu_starting(unsigned int cpu)
9297 struct rq *rq = cpu_rq(cpu);
9299 rq->calc_load_update = calc_load_update;
9300 update_max_interval();
9303 int sched_cpu_starting(unsigned int cpu)
9305 sched_core_cpu_starting(cpu);
9306 sched_rq_cpu_starting(cpu);
9307 sched_tick_start(cpu);
9311 #ifdef CONFIG_HOTPLUG_CPU
9314 * Invoked immediately before the stopper thread is invoked to bring the
9315 * CPU down completely. At this point all per CPU kthreads except the
9316 * hotplug thread (current) and the stopper thread (inactive) have been
9317 * either parked or have been unbound from the outgoing CPU. Ensure that
9318 * any of those which might be on the way out are gone.
9320 * If after this point a bound task is being woken on this CPU then the
9321 * responsible hotplug callback has failed to do it's job.
9322 * sched_cpu_dying() will catch it with the appropriate fireworks.
9324 int sched_cpu_wait_empty(unsigned int cpu)
9326 balance_hotplug_wait();
9331 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9332 * might have. Called from the CPU stopper task after ensuring that the
9333 * stopper is the last running task on the CPU, so nr_active count is
9334 * stable. We need to take the teardown thread which is calling this into
9335 * account, so we hand in adjust = 1 to the load calculation.
9337 * Also see the comment "Global load-average calculations".
9339 static void calc_load_migrate(struct rq *rq)
9341 long delta = calc_load_fold_active(rq, 1);
9344 atomic_long_add(delta, &calc_load_tasks);
9347 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9349 struct task_struct *g, *p;
9350 int cpu = cpu_of(rq);
9352 lockdep_assert_rq_held(rq);
9354 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9355 for_each_process_thread(g, p) {
9356 if (task_cpu(p) != cpu)
9359 if (!task_on_rq_queued(p))
9362 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9366 int sched_cpu_dying(unsigned int cpu)
9368 struct rq *rq = cpu_rq(cpu);
9371 /* Handle pending wakeups and then migrate everything off */
9372 sched_tick_stop(cpu);
9374 rq_lock_irqsave(rq, &rf);
9375 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9376 WARN(true, "Dying CPU not properly vacated!");
9377 dump_rq_tasks(rq, KERN_WARNING);
9379 rq_unlock_irqrestore(rq, &rf);
9381 calc_load_migrate(rq);
9382 update_max_interval();
9384 sched_core_cpu_dying(cpu);
9389 void __init sched_init_smp(void)
9394 * There's no userspace yet to cause hotplug operations; hence all the
9395 * CPU masks are stable and all blatant races in the below code cannot
9398 mutex_lock(&sched_domains_mutex);
9399 sched_init_domains(cpu_active_mask);
9400 mutex_unlock(&sched_domains_mutex);
9402 /* Move init over to a non-isolated CPU */
9403 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9405 current->flags &= ~PF_NO_SETAFFINITY;
9406 sched_init_granularity();
9408 init_sched_rt_class();
9409 init_sched_dl_class();
9411 sched_smp_initialized = true;
9414 static int __init migration_init(void)
9416 sched_cpu_starting(smp_processor_id());
9419 early_initcall(migration_init);
9422 void __init sched_init_smp(void)
9424 sched_init_granularity();
9426 #endif /* CONFIG_SMP */
9428 int in_sched_functions(unsigned long addr)
9430 return in_lock_functions(addr) ||
9431 (addr >= (unsigned long)__sched_text_start
9432 && addr < (unsigned long)__sched_text_end);
9435 #ifdef CONFIG_CGROUP_SCHED
9437 * Default task group.
9438 * Every task in system belongs to this group at bootup.
9440 struct task_group root_task_group;
9441 LIST_HEAD(task_groups);
9443 /* Cacheline aligned slab cache for task_group */
9444 static struct kmem_cache *task_group_cache __read_mostly;
9447 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9448 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9450 void __init sched_init(void)
9452 unsigned long ptr = 0;
9455 /* Make sure the linker didn't screw up */
9456 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9457 &fair_sched_class + 1 != &rt_sched_class ||
9458 &rt_sched_class + 1 != &dl_sched_class);
9460 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9465 #ifdef CONFIG_FAIR_GROUP_SCHED
9466 ptr += 2 * nr_cpu_ids * sizeof(void **);
9468 #ifdef CONFIG_RT_GROUP_SCHED
9469 ptr += 2 * nr_cpu_ids * sizeof(void **);
9472 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9474 #ifdef CONFIG_FAIR_GROUP_SCHED
9475 root_task_group.se = (struct sched_entity **)ptr;
9476 ptr += nr_cpu_ids * sizeof(void **);
9478 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9479 ptr += nr_cpu_ids * sizeof(void **);
9481 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9482 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9483 #endif /* CONFIG_FAIR_GROUP_SCHED */
9484 #ifdef CONFIG_RT_GROUP_SCHED
9485 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9486 ptr += nr_cpu_ids * sizeof(void **);
9488 root_task_group.rt_rq = (struct rt_rq **)ptr;
9489 ptr += nr_cpu_ids * sizeof(void **);
9491 #endif /* CONFIG_RT_GROUP_SCHED */
9493 #ifdef CONFIG_CPUMASK_OFFSTACK
9494 for_each_possible_cpu(i) {
9495 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9496 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9497 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9498 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9500 #endif /* CONFIG_CPUMASK_OFFSTACK */
9502 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9503 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9506 init_defrootdomain();
9509 #ifdef CONFIG_RT_GROUP_SCHED
9510 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9511 global_rt_period(), global_rt_runtime());
9512 #endif /* CONFIG_RT_GROUP_SCHED */
9514 #ifdef CONFIG_CGROUP_SCHED
9515 task_group_cache = KMEM_CACHE(task_group, 0);
9517 list_add(&root_task_group.list, &task_groups);
9518 INIT_LIST_HEAD(&root_task_group.children);
9519 INIT_LIST_HEAD(&root_task_group.siblings);
9520 autogroup_init(&init_task);
9521 #endif /* CONFIG_CGROUP_SCHED */
9523 for_each_possible_cpu(i) {
9527 raw_spin_lock_init(&rq->__lock);
9529 rq->calc_load_active = 0;
9530 rq->calc_load_update = jiffies + LOAD_FREQ;
9531 init_cfs_rq(&rq->cfs);
9532 init_rt_rq(&rq->rt);
9533 init_dl_rq(&rq->dl);
9534 #ifdef CONFIG_FAIR_GROUP_SCHED
9535 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9536 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9538 * How much CPU bandwidth does root_task_group get?
9540 * In case of task-groups formed thr' the cgroup filesystem, it
9541 * gets 100% of the CPU resources in the system. This overall
9542 * system CPU resource is divided among the tasks of
9543 * root_task_group and its child task-groups in a fair manner,
9544 * based on each entity's (task or task-group's) weight
9545 * (se->load.weight).
9547 * In other words, if root_task_group has 10 tasks of weight
9548 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9549 * then A0's share of the CPU resource is:
9551 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9553 * We achieve this by letting root_task_group's tasks sit
9554 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9556 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9557 #endif /* CONFIG_FAIR_GROUP_SCHED */
9559 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9560 #ifdef CONFIG_RT_GROUP_SCHED
9561 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9566 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9567 rq->balance_callback = &balance_push_callback;
9568 rq->active_balance = 0;
9569 rq->next_balance = jiffies;
9574 rq->avg_idle = 2*sysctl_sched_migration_cost;
9575 rq->wake_stamp = jiffies;
9576 rq->wake_avg_idle = rq->avg_idle;
9577 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9579 INIT_LIST_HEAD(&rq->cfs_tasks);
9581 rq_attach_root(rq, &def_root_domain);
9582 #ifdef CONFIG_NO_HZ_COMMON
9583 rq->last_blocked_load_update_tick = jiffies;
9584 atomic_set(&rq->nohz_flags, 0);
9586 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9588 #ifdef CONFIG_HOTPLUG_CPU
9589 rcuwait_init(&rq->hotplug_wait);
9591 #endif /* CONFIG_SMP */
9593 atomic_set(&rq->nr_iowait, 0);
9595 #ifdef CONFIG_SCHED_CORE
9597 rq->core_pick = NULL;
9598 rq->core_enabled = 0;
9599 rq->core_tree = RB_ROOT;
9600 rq->core_forceidle = false;
9602 rq->core_cookie = 0UL;
9606 set_load_weight(&init_task, false);
9609 * The boot idle thread does lazy MMU switching as well:
9612 enter_lazy_tlb(&init_mm, current);
9615 * Make us the idle thread. Technically, schedule() should not be
9616 * called from this thread, however somewhere below it might be,
9617 * but because we are the idle thread, we just pick up running again
9618 * when this runqueue becomes "idle".
9620 init_idle(current, smp_processor_id());
9622 calc_load_update = jiffies + LOAD_FREQ;
9625 idle_thread_set_boot_cpu();
9626 balance_push_set(smp_processor_id(), false);
9628 init_sched_fair_class();
9634 scheduler_running = 1;
9637 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9639 void __might_sleep(const char *file, int line)
9641 unsigned int state = get_current_state();
9643 * Blocking primitives will set (and therefore destroy) current->state,
9644 * since we will exit with TASK_RUNNING make sure we enter with it,
9645 * otherwise we will destroy state.
9647 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9648 "do not call blocking ops when !TASK_RUNNING; "
9649 "state=%x set at [<%p>] %pS\n", state,
9650 (void *)current->task_state_change,
9651 (void *)current->task_state_change);
9653 __might_resched(file, line, 0);
9655 EXPORT_SYMBOL(__might_sleep);
9657 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9659 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9662 if (preempt_count() == preempt_offset)
9665 pr_err("Preemption disabled at:");
9666 print_ip_sym(KERN_ERR, ip);
9669 static inline bool resched_offsets_ok(unsigned int offsets)
9671 unsigned int nested = preempt_count();
9673 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9675 return nested == offsets;
9678 void __might_resched(const char *file, int line, unsigned int offsets)
9680 /* Ratelimiting timestamp: */
9681 static unsigned long prev_jiffy;
9683 unsigned long preempt_disable_ip;
9685 /* WARN_ON_ONCE() by default, no rate limit required: */
9688 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9689 !is_idle_task(current) && !current->non_block_count) ||
9690 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9694 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9696 prev_jiffy = jiffies;
9698 /* Save this before calling printk(), since that will clobber it: */
9699 preempt_disable_ip = get_preempt_disable_ip(current);
9701 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9703 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9704 in_atomic(), irqs_disabled(), current->non_block_count,
9705 current->pid, current->comm);
9706 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9707 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9709 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9710 pr_err("RCU nest depth: %d, expected: %u\n",
9711 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9714 if (task_stack_end_corrupted(current))
9715 pr_emerg("Thread overran stack, or stack corrupted\n");
9717 debug_show_held_locks(current);
9718 if (irqs_disabled())
9719 print_irqtrace_events(current);
9721 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9722 preempt_disable_ip);
9725 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9727 EXPORT_SYMBOL(__might_resched);
9729 void __cant_sleep(const char *file, int line, int preempt_offset)
9731 static unsigned long prev_jiffy;
9733 if (irqs_disabled())
9736 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9739 if (preempt_count() > preempt_offset)
9742 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9744 prev_jiffy = jiffies;
9746 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9747 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9748 in_atomic(), irqs_disabled(),
9749 current->pid, current->comm);
9751 debug_show_held_locks(current);
9753 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9755 EXPORT_SYMBOL_GPL(__cant_sleep);
9758 void __cant_migrate(const char *file, int line)
9760 static unsigned long prev_jiffy;
9762 if (irqs_disabled())
9765 if (is_migration_disabled(current))
9768 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9771 if (preempt_count() > 0)
9774 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9776 prev_jiffy = jiffies;
9778 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9779 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9780 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9781 current->pid, current->comm);
9783 debug_show_held_locks(current);
9785 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9787 EXPORT_SYMBOL_GPL(__cant_migrate);
9791 #ifdef CONFIG_MAGIC_SYSRQ
9792 void normalize_rt_tasks(void)
9794 struct task_struct *g, *p;
9795 struct sched_attr attr = {
9796 .sched_policy = SCHED_NORMAL,
9799 read_lock(&tasklist_lock);
9800 for_each_process_thread(g, p) {
9802 * Only normalize user tasks:
9804 if (p->flags & PF_KTHREAD)
9807 p->se.exec_start = 0;
9808 schedstat_set(p->se.statistics.wait_start, 0);
9809 schedstat_set(p->se.statistics.sleep_start, 0);
9810 schedstat_set(p->se.statistics.block_start, 0);
9812 if (!dl_task(p) && !rt_task(p)) {
9814 * Renice negative nice level userspace
9817 if (task_nice(p) < 0)
9818 set_user_nice(p, 0);
9822 __sched_setscheduler(p, &attr, false, false);
9824 read_unlock(&tasklist_lock);
9827 #endif /* CONFIG_MAGIC_SYSRQ */
9829 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9831 * These functions are only useful for the IA64 MCA handling, or kdb.
9833 * They can only be called when the whole system has been
9834 * stopped - every CPU needs to be quiescent, and no scheduling
9835 * activity can take place. Using them for anything else would
9836 * be a serious bug, and as a result, they aren't even visible
9837 * under any other configuration.
9841 * curr_task - return the current task for a given CPU.
9842 * @cpu: the processor in question.
9844 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9846 * Return: The current task for @cpu.
9848 struct task_struct *curr_task(int cpu)
9850 return cpu_curr(cpu);
9853 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9857 * ia64_set_curr_task - set the current task for a given CPU.
9858 * @cpu: the processor in question.
9859 * @p: the task pointer to set.
9861 * Description: This function must only be used when non-maskable interrupts
9862 * are serviced on a separate stack. It allows the architecture to switch the
9863 * notion of the current task on a CPU in a non-blocking manner. This function
9864 * must be called with all CPU's synchronized, and interrupts disabled, the
9865 * and caller must save the original value of the current task (see
9866 * curr_task() above) and restore that value before reenabling interrupts and
9867 * re-starting the system.
9869 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9871 void ia64_set_curr_task(int cpu, struct task_struct *p)
9878 #ifdef CONFIG_CGROUP_SCHED
9879 /* task_group_lock serializes the addition/removal of task groups */
9880 static DEFINE_SPINLOCK(task_group_lock);
9882 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9883 struct task_group *parent)
9885 #ifdef CONFIG_UCLAMP_TASK_GROUP
9886 enum uclamp_id clamp_id;
9888 for_each_clamp_id(clamp_id) {
9889 uclamp_se_set(&tg->uclamp_req[clamp_id],
9890 uclamp_none(clamp_id), false);
9891 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9896 static void sched_free_group(struct task_group *tg)
9898 free_fair_sched_group(tg);
9899 free_rt_sched_group(tg);
9901 kmem_cache_free(task_group_cache, tg);
9904 static void sched_free_group_rcu(struct rcu_head *rcu)
9906 sched_free_group(container_of(rcu, struct task_group, rcu));
9909 static void sched_unregister_group(struct task_group *tg)
9911 unregister_fair_sched_group(tg);
9912 unregister_rt_sched_group(tg);
9914 * We have to wait for yet another RCU grace period to expire, as
9915 * print_cfs_stats() might run concurrently.
9917 call_rcu(&tg->rcu, sched_free_group_rcu);
9920 /* allocate runqueue etc for a new task group */
9921 struct task_group *sched_create_group(struct task_group *parent)
9923 struct task_group *tg;
9925 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9927 return ERR_PTR(-ENOMEM);
9929 if (!alloc_fair_sched_group(tg, parent))
9932 if (!alloc_rt_sched_group(tg, parent))
9935 alloc_uclamp_sched_group(tg, parent);
9940 sched_free_group(tg);
9941 return ERR_PTR(-ENOMEM);
9944 void sched_online_group(struct task_group *tg, struct task_group *parent)
9946 unsigned long flags;
9948 spin_lock_irqsave(&task_group_lock, flags);
9949 list_add_rcu(&tg->list, &task_groups);
9951 /* Root should already exist: */
9954 tg->parent = parent;
9955 INIT_LIST_HEAD(&tg->children);
9956 list_add_rcu(&tg->siblings, &parent->children);
9957 spin_unlock_irqrestore(&task_group_lock, flags);
9959 online_fair_sched_group(tg);
9962 /* rcu callback to free various structures associated with a task group */
9963 static void sched_unregister_group_rcu(struct rcu_head *rhp)
9965 /* Now it should be safe to free those cfs_rqs: */
9966 sched_unregister_group(container_of(rhp, struct task_group, rcu));
9969 void sched_destroy_group(struct task_group *tg)
9971 /* Wait for possible concurrent references to cfs_rqs complete: */
9972 call_rcu(&tg->rcu, sched_unregister_group_rcu);
9975 void sched_release_group(struct task_group *tg)
9977 unsigned long flags;
9980 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
9981 * sched_cfs_period_timer()).
9983 * For this to be effective, we have to wait for all pending users of
9984 * this task group to leave their RCU critical section to ensure no new
9985 * user will see our dying task group any more. Specifically ensure
9986 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
9988 * We therefore defer calling unregister_fair_sched_group() to
9989 * sched_unregister_group() which is guarantied to get called only after the
9990 * current RCU grace period has expired.
9992 spin_lock_irqsave(&task_group_lock, flags);
9993 list_del_rcu(&tg->list);
9994 list_del_rcu(&tg->siblings);
9995 spin_unlock_irqrestore(&task_group_lock, flags);
9998 static void sched_change_group(struct task_struct *tsk, int type)
10000 struct task_group *tg;
10003 * All callers are synchronized by task_rq_lock(); we do not use RCU
10004 * which is pointless here. Thus, we pass "true" to task_css_check()
10005 * to prevent lockdep warnings.
10007 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10008 struct task_group, css);
10009 tg = autogroup_task_group(tsk, tg);
10010 tsk->sched_task_group = tg;
10012 #ifdef CONFIG_FAIR_GROUP_SCHED
10013 if (tsk->sched_class->task_change_group)
10014 tsk->sched_class->task_change_group(tsk, type);
10017 set_task_rq(tsk, task_cpu(tsk));
10021 * Change task's runqueue when it moves between groups.
10023 * The caller of this function should have put the task in its new group by
10024 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10027 void sched_move_task(struct task_struct *tsk)
10029 int queued, running, queue_flags =
10030 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
10031 struct rq_flags rf;
10034 rq = task_rq_lock(tsk, &rf);
10035 update_rq_clock(rq);
10037 running = task_current(rq, tsk);
10038 queued = task_on_rq_queued(tsk);
10041 dequeue_task(rq, tsk, queue_flags);
10043 put_prev_task(rq, tsk);
10045 sched_change_group(tsk, TASK_MOVE_GROUP);
10048 enqueue_task(rq, tsk, queue_flags);
10050 set_next_task(rq, tsk);
10052 * After changing group, the running task may have joined a
10053 * throttled one but it's still the running task. Trigger a
10054 * resched to make sure that task can still run.
10059 task_rq_unlock(rq, tsk, &rf);
10062 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
10064 return css ? container_of(css, struct task_group, css) : NULL;
10067 static struct cgroup_subsys_state *
10068 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
10070 struct task_group *parent = css_tg(parent_css);
10071 struct task_group *tg;
10074 /* This is early initialization for the top cgroup */
10075 return &root_task_group.css;
10078 tg = sched_create_group(parent);
10080 return ERR_PTR(-ENOMEM);
10085 /* Expose task group only after completing cgroup initialization */
10086 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10088 struct task_group *tg = css_tg(css);
10089 struct task_group *parent = css_tg(css->parent);
10092 sched_online_group(tg, parent);
10094 #ifdef CONFIG_UCLAMP_TASK_GROUP
10095 /* Propagate the effective uclamp value for the new group */
10096 mutex_lock(&uclamp_mutex);
10098 cpu_util_update_eff(css);
10100 mutex_unlock(&uclamp_mutex);
10106 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
10108 struct task_group *tg = css_tg(css);
10110 sched_release_group(tg);
10113 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
10115 struct task_group *tg = css_tg(css);
10118 * Relies on the RCU grace period between css_released() and this.
10120 sched_unregister_group(tg);
10124 * This is called before wake_up_new_task(), therefore we really only
10125 * have to set its group bits, all the other stuff does not apply.
10127 static void cpu_cgroup_fork(struct task_struct *task)
10129 struct rq_flags rf;
10132 rq = task_rq_lock(task, &rf);
10134 update_rq_clock(rq);
10135 sched_change_group(task, TASK_SET_GROUP);
10137 task_rq_unlock(rq, task, &rf);
10140 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10142 struct task_struct *task;
10143 struct cgroup_subsys_state *css;
10146 cgroup_taskset_for_each(task, css, tset) {
10147 #ifdef CONFIG_RT_GROUP_SCHED
10148 if (!sched_rt_can_attach(css_tg(css), task))
10152 * Serialize against wake_up_new_task() such that if it's
10153 * running, we're sure to observe its full state.
10155 raw_spin_lock_irq(&task->pi_lock);
10157 * Avoid calling sched_move_task() before wake_up_new_task()
10158 * has happened. This would lead to problems with PELT, due to
10159 * move wanting to detach+attach while we're not attached yet.
10161 if (READ_ONCE(task->__state) == TASK_NEW)
10163 raw_spin_unlock_irq(&task->pi_lock);
10171 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10173 struct task_struct *task;
10174 struct cgroup_subsys_state *css;
10176 cgroup_taskset_for_each(task, css, tset)
10177 sched_move_task(task);
10180 #ifdef CONFIG_UCLAMP_TASK_GROUP
10181 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10183 struct cgroup_subsys_state *top_css = css;
10184 struct uclamp_se *uc_parent = NULL;
10185 struct uclamp_se *uc_se = NULL;
10186 unsigned int eff[UCLAMP_CNT];
10187 enum uclamp_id clamp_id;
10188 unsigned int clamps;
10190 lockdep_assert_held(&uclamp_mutex);
10191 SCHED_WARN_ON(!rcu_read_lock_held());
10193 css_for_each_descendant_pre(css, top_css) {
10194 uc_parent = css_tg(css)->parent
10195 ? css_tg(css)->parent->uclamp : NULL;
10197 for_each_clamp_id(clamp_id) {
10198 /* Assume effective clamps matches requested clamps */
10199 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10200 /* Cap effective clamps with parent's effective clamps */
10202 eff[clamp_id] > uc_parent[clamp_id].value) {
10203 eff[clamp_id] = uc_parent[clamp_id].value;
10206 /* Ensure protection is always capped by limit */
10207 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10209 /* Propagate most restrictive effective clamps */
10211 uc_se = css_tg(css)->uclamp;
10212 for_each_clamp_id(clamp_id) {
10213 if (eff[clamp_id] == uc_se[clamp_id].value)
10215 uc_se[clamp_id].value = eff[clamp_id];
10216 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10217 clamps |= (0x1 << clamp_id);
10220 css = css_rightmost_descendant(css);
10224 /* Immediately update descendants RUNNABLE tasks */
10225 uclamp_update_active_tasks(css);
10230 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10231 * C expression. Since there is no way to convert a macro argument (N) into a
10232 * character constant, use two levels of macros.
10234 #define _POW10(exp) ((unsigned int)1e##exp)
10235 #define POW10(exp) _POW10(exp)
10237 struct uclamp_request {
10238 #define UCLAMP_PERCENT_SHIFT 2
10239 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10245 static inline struct uclamp_request
10246 capacity_from_percent(char *buf)
10248 struct uclamp_request req = {
10249 .percent = UCLAMP_PERCENT_SCALE,
10250 .util = SCHED_CAPACITY_SCALE,
10255 if (strcmp(buf, "max")) {
10256 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10260 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10265 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10266 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10272 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10273 size_t nbytes, loff_t off,
10274 enum uclamp_id clamp_id)
10276 struct uclamp_request req;
10277 struct task_group *tg;
10279 req = capacity_from_percent(buf);
10283 static_branch_enable(&sched_uclamp_used);
10285 mutex_lock(&uclamp_mutex);
10288 tg = css_tg(of_css(of));
10289 if (tg->uclamp_req[clamp_id].value != req.util)
10290 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10293 * Because of not recoverable conversion rounding we keep track of the
10294 * exact requested value
10296 tg->uclamp_pct[clamp_id] = req.percent;
10298 /* Update effective clamps to track the most restrictive value */
10299 cpu_util_update_eff(of_css(of));
10302 mutex_unlock(&uclamp_mutex);
10307 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10308 char *buf, size_t nbytes,
10311 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10314 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10315 char *buf, size_t nbytes,
10318 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10321 static inline void cpu_uclamp_print(struct seq_file *sf,
10322 enum uclamp_id clamp_id)
10324 struct task_group *tg;
10330 tg = css_tg(seq_css(sf));
10331 util_clamp = tg->uclamp_req[clamp_id].value;
10334 if (util_clamp == SCHED_CAPACITY_SCALE) {
10335 seq_puts(sf, "max\n");
10339 percent = tg->uclamp_pct[clamp_id];
10340 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10341 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10344 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10346 cpu_uclamp_print(sf, UCLAMP_MIN);
10350 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10352 cpu_uclamp_print(sf, UCLAMP_MAX);
10355 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10357 #ifdef CONFIG_FAIR_GROUP_SCHED
10358 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10359 struct cftype *cftype, u64 shareval)
10361 if (shareval > scale_load_down(ULONG_MAX))
10362 shareval = MAX_SHARES;
10363 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10366 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10367 struct cftype *cft)
10369 struct task_group *tg = css_tg(css);
10371 return (u64) scale_load_down(tg->shares);
10374 #ifdef CONFIG_CFS_BANDWIDTH
10375 static DEFINE_MUTEX(cfs_constraints_mutex);
10377 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10378 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10379 /* More than 203 days if BW_SHIFT equals 20. */
10380 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10382 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10384 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10387 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10388 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10390 if (tg == &root_task_group)
10394 * Ensure we have at some amount of bandwidth every period. This is
10395 * to prevent reaching a state of large arrears when throttled via
10396 * entity_tick() resulting in prolonged exit starvation.
10398 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10402 * Likewise, bound things on the other side by preventing insane quota
10403 * periods. This also allows us to normalize in computing quota
10406 if (period > max_cfs_quota_period)
10410 * Bound quota to defend quota against overflow during bandwidth shift.
10412 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10415 if (quota != RUNTIME_INF && (burst > quota ||
10416 burst + quota > max_cfs_runtime))
10420 * Prevent race between setting of cfs_rq->runtime_enabled and
10421 * unthrottle_offline_cfs_rqs().
10424 mutex_lock(&cfs_constraints_mutex);
10425 ret = __cfs_schedulable(tg, period, quota);
10429 runtime_enabled = quota != RUNTIME_INF;
10430 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10432 * If we need to toggle cfs_bandwidth_used, off->on must occur
10433 * before making related changes, and on->off must occur afterwards
10435 if (runtime_enabled && !runtime_was_enabled)
10436 cfs_bandwidth_usage_inc();
10437 raw_spin_lock_irq(&cfs_b->lock);
10438 cfs_b->period = ns_to_ktime(period);
10439 cfs_b->quota = quota;
10440 cfs_b->burst = burst;
10442 __refill_cfs_bandwidth_runtime(cfs_b);
10444 /* Restart the period timer (if active) to handle new period expiry: */
10445 if (runtime_enabled)
10446 start_cfs_bandwidth(cfs_b);
10448 raw_spin_unlock_irq(&cfs_b->lock);
10450 for_each_online_cpu(i) {
10451 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10452 struct rq *rq = cfs_rq->rq;
10453 struct rq_flags rf;
10455 rq_lock_irq(rq, &rf);
10456 cfs_rq->runtime_enabled = runtime_enabled;
10457 cfs_rq->runtime_remaining = 0;
10459 if (cfs_rq->throttled)
10460 unthrottle_cfs_rq(cfs_rq);
10461 rq_unlock_irq(rq, &rf);
10463 if (runtime_was_enabled && !runtime_enabled)
10464 cfs_bandwidth_usage_dec();
10466 mutex_unlock(&cfs_constraints_mutex);
10467 cpus_read_unlock();
10472 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10474 u64 quota, period, burst;
10476 period = ktime_to_ns(tg->cfs_bandwidth.period);
10477 burst = tg->cfs_bandwidth.burst;
10478 if (cfs_quota_us < 0)
10479 quota = RUNTIME_INF;
10480 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10481 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10485 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10488 static long tg_get_cfs_quota(struct task_group *tg)
10492 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10495 quota_us = tg->cfs_bandwidth.quota;
10496 do_div(quota_us, NSEC_PER_USEC);
10501 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10503 u64 quota, period, burst;
10505 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10508 period = (u64)cfs_period_us * NSEC_PER_USEC;
10509 quota = tg->cfs_bandwidth.quota;
10510 burst = tg->cfs_bandwidth.burst;
10512 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10515 static long tg_get_cfs_period(struct task_group *tg)
10519 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10520 do_div(cfs_period_us, NSEC_PER_USEC);
10522 return cfs_period_us;
10525 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10527 u64 quota, period, burst;
10529 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10532 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10533 period = ktime_to_ns(tg->cfs_bandwidth.period);
10534 quota = tg->cfs_bandwidth.quota;
10536 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10539 static long tg_get_cfs_burst(struct task_group *tg)
10543 burst_us = tg->cfs_bandwidth.burst;
10544 do_div(burst_us, NSEC_PER_USEC);
10549 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10550 struct cftype *cft)
10552 return tg_get_cfs_quota(css_tg(css));
10555 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10556 struct cftype *cftype, s64 cfs_quota_us)
10558 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10561 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10562 struct cftype *cft)
10564 return tg_get_cfs_period(css_tg(css));
10567 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10568 struct cftype *cftype, u64 cfs_period_us)
10570 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10573 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10574 struct cftype *cft)
10576 return tg_get_cfs_burst(css_tg(css));
10579 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10580 struct cftype *cftype, u64 cfs_burst_us)
10582 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10585 struct cfs_schedulable_data {
10586 struct task_group *tg;
10591 * normalize group quota/period to be quota/max_period
10592 * note: units are usecs
10594 static u64 normalize_cfs_quota(struct task_group *tg,
10595 struct cfs_schedulable_data *d)
10600 period = d->period;
10603 period = tg_get_cfs_period(tg);
10604 quota = tg_get_cfs_quota(tg);
10607 /* note: these should typically be equivalent */
10608 if (quota == RUNTIME_INF || quota == -1)
10609 return RUNTIME_INF;
10611 return to_ratio(period, quota);
10614 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10616 struct cfs_schedulable_data *d = data;
10617 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10618 s64 quota = 0, parent_quota = -1;
10621 quota = RUNTIME_INF;
10623 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10625 quota = normalize_cfs_quota(tg, d);
10626 parent_quota = parent_b->hierarchical_quota;
10629 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10630 * always take the min. On cgroup1, only inherit when no
10633 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10634 quota = min(quota, parent_quota);
10636 if (quota == RUNTIME_INF)
10637 quota = parent_quota;
10638 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10642 cfs_b->hierarchical_quota = quota;
10647 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10650 struct cfs_schedulable_data data = {
10656 if (quota != RUNTIME_INF) {
10657 do_div(data.period, NSEC_PER_USEC);
10658 do_div(data.quota, NSEC_PER_USEC);
10662 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10668 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10670 struct task_group *tg = css_tg(seq_css(sf));
10671 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10673 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10674 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10675 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10677 if (schedstat_enabled() && tg != &root_task_group) {
10681 for_each_possible_cpu(i)
10682 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10684 seq_printf(sf, "wait_sum %llu\n", ws);
10689 #endif /* CONFIG_CFS_BANDWIDTH */
10690 #endif /* CONFIG_FAIR_GROUP_SCHED */
10692 #ifdef CONFIG_RT_GROUP_SCHED
10693 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10694 struct cftype *cft, s64 val)
10696 return sched_group_set_rt_runtime(css_tg(css), val);
10699 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10700 struct cftype *cft)
10702 return sched_group_rt_runtime(css_tg(css));
10705 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10706 struct cftype *cftype, u64 rt_period_us)
10708 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10711 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10712 struct cftype *cft)
10714 return sched_group_rt_period(css_tg(css));
10716 #endif /* CONFIG_RT_GROUP_SCHED */
10718 #ifdef CONFIG_FAIR_GROUP_SCHED
10719 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10720 struct cftype *cft)
10722 return css_tg(css)->idle;
10725 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10726 struct cftype *cft, s64 idle)
10728 return sched_group_set_idle(css_tg(css), idle);
10732 static struct cftype cpu_legacy_files[] = {
10733 #ifdef CONFIG_FAIR_GROUP_SCHED
10736 .read_u64 = cpu_shares_read_u64,
10737 .write_u64 = cpu_shares_write_u64,
10741 .read_s64 = cpu_idle_read_s64,
10742 .write_s64 = cpu_idle_write_s64,
10745 #ifdef CONFIG_CFS_BANDWIDTH
10747 .name = "cfs_quota_us",
10748 .read_s64 = cpu_cfs_quota_read_s64,
10749 .write_s64 = cpu_cfs_quota_write_s64,
10752 .name = "cfs_period_us",
10753 .read_u64 = cpu_cfs_period_read_u64,
10754 .write_u64 = cpu_cfs_period_write_u64,
10757 .name = "cfs_burst_us",
10758 .read_u64 = cpu_cfs_burst_read_u64,
10759 .write_u64 = cpu_cfs_burst_write_u64,
10763 .seq_show = cpu_cfs_stat_show,
10766 #ifdef CONFIG_RT_GROUP_SCHED
10768 .name = "rt_runtime_us",
10769 .read_s64 = cpu_rt_runtime_read,
10770 .write_s64 = cpu_rt_runtime_write,
10773 .name = "rt_period_us",
10774 .read_u64 = cpu_rt_period_read_uint,
10775 .write_u64 = cpu_rt_period_write_uint,
10778 #ifdef CONFIG_UCLAMP_TASK_GROUP
10780 .name = "uclamp.min",
10781 .flags = CFTYPE_NOT_ON_ROOT,
10782 .seq_show = cpu_uclamp_min_show,
10783 .write = cpu_uclamp_min_write,
10786 .name = "uclamp.max",
10787 .flags = CFTYPE_NOT_ON_ROOT,
10788 .seq_show = cpu_uclamp_max_show,
10789 .write = cpu_uclamp_max_write,
10792 { } /* Terminate */
10795 static int cpu_extra_stat_show(struct seq_file *sf,
10796 struct cgroup_subsys_state *css)
10798 #ifdef CONFIG_CFS_BANDWIDTH
10800 struct task_group *tg = css_tg(css);
10801 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10802 u64 throttled_usec;
10804 throttled_usec = cfs_b->throttled_time;
10805 do_div(throttled_usec, NSEC_PER_USEC);
10807 seq_printf(sf, "nr_periods %d\n"
10808 "nr_throttled %d\n"
10809 "throttled_usec %llu\n",
10810 cfs_b->nr_periods, cfs_b->nr_throttled,
10817 #ifdef CONFIG_FAIR_GROUP_SCHED
10818 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10819 struct cftype *cft)
10821 struct task_group *tg = css_tg(css);
10822 u64 weight = scale_load_down(tg->shares);
10824 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10827 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10828 struct cftype *cft, u64 weight)
10831 * cgroup weight knobs should use the common MIN, DFL and MAX
10832 * values which are 1, 100 and 10000 respectively. While it loses
10833 * a bit of range on both ends, it maps pretty well onto the shares
10834 * value used by scheduler and the round-trip conversions preserve
10835 * the original value over the entire range.
10837 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10840 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10842 return sched_group_set_shares(css_tg(css), scale_load(weight));
10845 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10846 struct cftype *cft)
10848 unsigned long weight = scale_load_down(css_tg(css)->shares);
10849 int last_delta = INT_MAX;
10852 /* find the closest nice value to the current weight */
10853 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10854 delta = abs(sched_prio_to_weight[prio] - weight);
10855 if (delta >= last_delta)
10857 last_delta = delta;
10860 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10863 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10864 struct cftype *cft, s64 nice)
10866 unsigned long weight;
10869 if (nice < MIN_NICE || nice > MAX_NICE)
10872 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10873 idx = array_index_nospec(idx, 40);
10874 weight = sched_prio_to_weight[idx];
10876 return sched_group_set_shares(css_tg(css), scale_load(weight));
10880 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10881 long period, long quota)
10884 seq_puts(sf, "max");
10886 seq_printf(sf, "%ld", quota);
10888 seq_printf(sf, " %ld\n", period);
10891 /* caller should put the current value in *@periodp before calling */
10892 static int __maybe_unused cpu_period_quota_parse(char *buf,
10893 u64 *periodp, u64 *quotap)
10895 char tok[21]; /* U64_MAX */
10897 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10900 *periodp *= NSEC_PER_USEC;
10902 if (sscanf(tok, "%llu", quotap))
10903 *quotap *= NSEC_PER_USEC;
10904 else if (!strcmp(tok, "max"))
10905 *quotap = RUNTIME_INF;
10912 #ifdef CONFIG_CFS_BANDWIDTH
10913 static int cpu_max_show(struct seq_file *sf, void *v)
10915 struct task_group *tg = css_tg(seq_css(sf));
10917 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10921 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10922 char *buf, size_t nbytes, loff_t off)
10924 struct task_group *tg = css_tg(of_css(of));
10925 u64 period = tg_get_cfs_period(tg);
10926 u64 burst = tg_get_cfs_burst(tg);
10930 ret = cpu_period_quota_parse(buf, &period, "a);
10932 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10933 return ret ?: nbytes;
10937 static struct cftype cpu_files[] = {
10938 #ifdef CONFIG_FAIR_GROUP_SCHED
10941 .flags = CFTYPE_NOT_ON_ROOT,
10942 .read_u64 = cpu_weight_read_u64,
10943 .write_u64 = cpu_weight_write_u64,
10946 .name = "weight.nice",
10947 .flags = CFTYPE_NOT_ON_ROOT,
10948 .read_s64 = cpu_weight_nice_read_s64,
10949 .write_s64 = cpu_weight_nice_write_s64,
10953 .flags = CFTYPE_NOT_ON_ROOT,
10954 .read_s64 = cpu_idle_read_s64,
10955 .write_s64 = cpu_idle_write_s64,
10958 #ifdef CONFIG_CFS_BANDWIDTH
10961 .flags = CFTYPE_NOT_ON_ROOT,
10962 .seq_show = cpu_max_show,
10963 .write = cpu_max_write,
10966 .name = "max.burst",
10967 .flags = CFTYPE_NOT_ON_ROOT,
10968 .read_u64 = cpu_cfs_burst_read_u64,
10969 .write_u64 = cpu_cfs_burst_write_u64,
10972 #ifdef CONFIG_UCLAMP_TASK_GROUP
10974 .name = "uclamp.min",
10975 .flags = CFTYPE_NOT_ON_ROOT,
10976 .seq_show = cpu_uclamp_min_show,
10977 .write = cpu_uclamp_min_write,
10980 .name = "uclamp.max",
10981 .flags = CFTYPE_NOT_ON_ROOT,
10982 .seq_show = cpu_uclamp_max_show,
10983 .write = cpu_uclamp_max_write,
10986 { } /* terminate */
10989 struct cgroup_subsys cpu_cgrp_subsys = {
10990 .css_alloc = cpu_cgroup_css_alloc,
10991 .css_online = cpu_cgroup_css_online,
10992 .css_released = cpu_cgroup_css_released,
10993 .css_free = cpu_cgroup_css_free,
10994 .css_extra_stat_show = cpu_extra_stat_show,
10995 .fork = cpu_cgroup_fork,
10996 .can_attach = cpu_cgroup_can_attach,
10997 .attach = cpu_cgroup_attach,
10998 .legacy_cftypes = cpu_legacy_files,
10999 .dfl_cftypes = cpu_files,
11000 .early_init = true,
11004 #endif /* CONFIG_CGROUP_SCHED */
11006 void dump_cpu_task(int cpu)
11008 pr_info("Task dump for CPU %d:\n", cpu);
11009 sched_show_task(cpu_curr(cpu));
11013 * Nice levels are multiplicative, with a gentle 10% change for every
11014 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11015 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11016 * that remained on nice 0.
11018 * The "10% effect" is relative and cumulative: from _any_ nice level,
11019 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11020 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11021 * If a task goes up by ~10% and another task goes down by ~10% then
11022 * the relative distance between them is ~25%.)
11024 const int sched_prio_to_weight[40] = {
11025 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11026 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11027 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11028 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11029 /* 0 */ 1024, 820, 655, 526, 423,
11030 /* 5 */ 335, 272, 215, 172, 137,
11031 /* 10 */ 110, 87, 70, 56, 45,
11032 /* 15 */ 36, 29, 23, 18, 15,
11036 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11038 * In cases where the weight does not change often, we can use the
11039 * precalculated inverse to speed up arithmetics by turning divisions
11040 * into multiplications:
11042 const u32 sched_prio_to_wmult[40] = {
11043 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11044 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11045 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11046 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11047 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11048 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11049 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11050 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11053 void call_trace_sched_update_nr_running(struct rq *rq, int count)
11055 trace_sched_update_nr_running_tp(rq, count);