1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
63 * Print a warning if need_resched is set for the given duration (if
64 * LATENCY_WARN is enabled).
66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
69 __read_mostly int sysctl_resched_latency_warn_ms = 100;
70 __read_mostly int sysctl_resched_latency_warn_once = 1;
71 #endif /* CONFIG_SCHED_DEBUG */
74 * Number of tasks to iterate in a single balance run.
75 * Limited because this is done with IRQs disabled.
77 const_debug unsigned int sysctl_sched_nr_migrate = 32;
80 * period over which we measure -rt task CPU usage in us.
83 unsigned int sysctl_sched_rt_period = 1000000;
85 __read_mostly int scheduler_running;
87 #ifdef CONFIG_SCHED_CORE
89 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
91 /* kernel prio, less is more */
92 static inline int __task_prio(struct task_struct *p)
94 if (p->sched_class == &stop_sched_class) /* trumps deadline */
97 if (rt_prio(p->prio)) /* includes deadline */
98 return p->prio; /* [-1, 99] */
100 if (p->sched_class == &idle_sched_class)
101 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
103 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
113 /* real prio, less is less */
114 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
117 int pa = __task_prio(a), pb = __task_prio(b);
125 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
126 return !dl_time_before(a->dl.deadline, b->dl.deadline);
128 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
129 return cfs_prio_less(a, b, in_fi);
134 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
136 if (a->core_cookie < b->core_cookie)
139 if (a->core_cookie > b->core_cookie)
142 /* flip prio, so high prio is leftmost */
143 if (prio_less(b, a, task_rq(a)->core->core_forceidle))
149 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
151 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
153 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
156 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
158 const struct task_struct *p = __node_2_sc(node);
159 unsigned long cookie = (unsigned long)key;
161 if (cookie < p->core_cookie)
164 if (cookie > p->core_cookie)
170 static void sched_core_enqueue(struct rq *rq, struct task_struct *p)
172 rq->core->core_task_seq++;
177 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
180 static void sched_core_dequeue(struct rq *rq, struct task_struct *p)
182 rq->core->core_task_seq++;
187 rb_erase(&p->core_node, &rq->core_tree);
191 * Find left-most (aka, highest priority) task matching @cookie.
193 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
195 struct rb_node *node;
197 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
199 * The idle task always matches any cookie!
202 return idle_sched_class.pick_task(rq);
204 return __node_2_sc(node);
207 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
209 struct rb_node *node = &p->core_node;
211 node = rb_next(node);
215 p = container_of(node, struct task_struct, core_node);
216 if (p->core_cookie != cookie)
223 * Magic required such that:
225 * raw_spin_rq_lock(rq);
227 * raw_spin_rq_unlock(rq);
229 * ends up locking and unlocking the _same_ lock, and all CPUs
230 * always agree on what rq has what lock.
232 * XXX entirely possible to selectively enable cores, don't bother for now.
235 static DEFINE_MUTEX(sched_core_mutex);
236 static atomic_t sched_core_count;
237 static struct cpumask sched_core_mask;
239 static void __sched_core_flip(bool enabled)
246 * Toggle the online cores, one by one.
248 cpumask_copy(&sched_core_mask, cpu_online_mask);
249 for_each_cpu(cpu, &sched_core_mask) {
250 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
254 for_each_cpu(t, smt_mask) {
255 /* supports up to SMT8 */
256 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
259 for_each_cpu(t, smt_mask)
260 cpu_rq(t)->core_enabled = enabled;
262 for_each_cpu(t, smt_mask)
263 raw_spin_unlock(&cpu_rq(t)->__lock);
266 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
270 * Toggle the offline CPUs.
272 cpumask_copy(&sched_core_mask, cpu_possible_mask);
273 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
275 for_each_cpu(cpu, &sched_core_mask)
276 cpu_rq(cpu)->core_enabled = enabled;
281 static void sched_core_assert_empty(void)
285 for_each_possible_cpu(cpu)
286 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
289 static void __sched_core_enable(void)
291 static_branch_enable(&__sched_core_enabled);
293 * Ensure all previous instances of raw_spin_rq_*lock() have finished
294 * and future ones will observe !sched_core_disabled().
297 __sched_core_flip(true);
298 sched_core_assert_empty();
301 static void __sched_core_disable(void)
303 sched_core_assert_empty();
304 __sched_core_flip(false);
305 static_branch_disable(&__sched_core_enabled);
308 void sched_core_get(void)
310 if (atomic_inc_not_zero(&sched_core_count))
313 mutex_lock(&sched_core_mutex);
314 if (!atomic_read(&sched_core_count))
315 __sched_core_enable();
317 smp_mb__before_atomic();
318 atomic_inc(&sched_core_count);
319 mutex_unlock(&sched_core_mutex);
322 static void __sched_core_put(struct work_struct *work)
324 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
325 __sched_core_disable();
326 mutex_unlock(&sched_core_mutex);
330 void sched_core_put(void)
332 static DECLARE_WORK(_work, __sched_core_put);
335 * "There can be only one"
337 * Either this is the last one, or we don't actually need to do any
338 * 'work'. If it is the last *again*, we rely on
339 * WORK_STRUCT_PENDING_BIT.
341 if (!atomic_add_unless(&sched_core_count, -1, 1))
342 schedule_work(&_work);
345 #else /* !CONFIG_SCHED_CORE */
347 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
348 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
350 #endif /* CONFIG_SCHED_CORE */
353 * part of the period that we allow rt tasks to run in us.
356 int sysctl_sched_rt_runtime = 950000;
360 * Serialization rules:
366 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
369 * rq2->lock where: rq1 < rq2
373 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
374 * local CPU's rq->lock, it optionally removes the task from the runqueue and
375 * always looks at the local rq data structures to find the most eligible task
378 * Task enqueue is also under rq->lock, possibly taken from another CPU.
379 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
380 * the local CPU to avoid bouncing the runqueue state around [ see
381 * ttwu_queue_wakelist() ]
383 * Task wakeup, specifically wakeups that involve migration, are horribly
384 * complicated to avoid having to take two rq->locks.
388 * System-calls and anything external will use task_rq_lock() which acquires
389 * both p->pi_lock and rq->lock. As a consequence the state they change is
390 * stable while holding either lock:
392 * - sched_setaffinity()/
393 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
394 * - set_user_nice(): p->se.load, p->*prio
395 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
396 * p->se.load, p->rt_priority,
397 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
398 * - sched_setnuma(): p->numa_preferred_nid
399 * - sched_move_task()/
400 * cpu_cgroup_fork(): p->sched_task_group
401 * - uclamp_update_active() p->uclamp*
403 * p->state <- TASK_*:
405 * is changed locklessly using set_current_state(), __set_current_state() or
406 * set_special_state(), see their respective comments, or by
407 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
410 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
412 * is set by activate_task() and cleared by deactivate_task(), under
413 * rq->lock. Non-zero indicates the task is runnable, the special
414 * ON_RQ_MIGRATING state is used for migration without holding both
415 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
417 * p->on_cpu <- { 0, 1 }:
419 * is set by prepare_task() and cleared by finish_task() such that it will be
420 * set before p is scheduled-in and cleared after p is scheduled-out, both
421 * under rq->lock. Non-zero indicates the task is running on its CPU.
423 * [ The astute reader will observe that it is possible for two tasks on one
424 * CPU to have ->on_cpu = 1 at the same time. ]
426 * task_cpu(p): is changed by set_task_cpu(), the rules are:
428 * - Don't call set_task_cpu() on a blocked task:
430 * We don't care what CPU we're not running on, this simplifies hotplug,
431 * the CPU assignment of blocked tasks isn't required to be valid.
433 * - for try_to_wake_up(), called under p->pi_lock:
435 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
437 * - for migration called under rq->lock:
438 * [ see task_on_rq_migrating() in task_rq_lock() ]
440 * o move_queued_task()
443 * - for migration called under double_rq_lock():
445 * o __migrate_swap_task()
446 * o push_rt_task() / pull_rt_task()
447 * o push_dl_task() / pull_dl_task()
448 * o dl_task_offline_migration()
452 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
454 raw_spinlock_t *lock;
456 /* Matches synchronize_rcu() in __sched_core_enable() */
458 if (sched_core_disabled()) {
459 raw_spin_lock_nested(&rq->__lock, subclass);
460 /* preempt_count *MUST* be > 1 */
461 preempt_enable_no_resched();
466 lock = __rq_lockp(rq);
467 raw_spin_lock_nested(lock, subclass);
468 if (likely(lock == __rq_lockp(rq))) {
469 /* preempt_count *MUST* be > 1 */
470 preempt_enable_no_resched();
473 raw_spin_unlock(lock);
477 bool raw_spin_rq_trylock(struct rq *rq)
479 raw_spinlock_t *lock;
482 /* Matches synchronize_rcu() in __sched_core_enable() */
484 if (sched_core_disabled()) {
485 ret = raw_spin_trylock(&rq->__lock);
491 lock = __rq_lockp(rq);
492 ret = raw_spin_trylock(lock);
493 if (!ret || (likely(lock == __rq_lockp(rq)))) {
497 raw_spin_unlock(lock);
501 void raw_spin_rq_unlock(struct rq *rq)
503 raw_spin_unlock(rq_lockp(rq));
508 * double_rq_lock - safely lock two runqueues
510 void double_rq_lock(struct rq *rq1, struct rq *rq2)
512 lockdep_assert_irqs_disabled();
514 if (rq_order_less(rq2, rq1))
517 raw_spin_rq_lock(rq1);
518 if (__rq_lockp(rq1) == __rq_lockp(rq2))
521 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
526 * __task_rq_lock - lock the rq @p resides on.
528 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
533 lockdep_assert_held(&p->pi_lock);
537 raw_spin_rq_lock(rq);
538 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
542 raw_spin_rq_unlock(rq);
544 while (unlikely(task_on_rq_migrating(p)))
550 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
552 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
553 __acquires(p->pi_lock)
559 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
561 raw_spin_rq_lock(rq);
563 * move_queued_task() task_rq_lock()
566 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
567 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
568 * [S] ->cpu = new_cpu [L] task_rq()
572 * If we observe the old CPU in task_rq_lock(), the acquire of
573 * the old rq->lock will fully serialize against the stores.
575 * If we observe the new CPU in task_rq_lock(), the address
576 * dependency headed by '[L] rq = task_rq()' and the acquire
577 * will pair with the WMB to ensure we then also see migrating.
579 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
583 raw_spin_rq_unlock(rq);
584 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
586 while (unlikely(task_on_rq_migrating(p)))
592 * RQ-clock updating methods:
595 static void update_rq_clock_task(struct rq *rq, s64 delta)
598 * In theory, the compile should just see 0 here, and optimize out the call
599 * to sched_rt_avg_update. But I don't trust it...
601 s64 __maybe_unused steal = 0, irq_delta = 0;
603 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
604 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
607 * Since irq_time is only updated on {soft,}irq_exit, we might run into
608 * this case when a previous update_rq_clock() happened inside a
611 * When this happens, we stop ->clock_task and only update the
612 * prev_irq_time stamp to account for the part that fit, so that a next
613 * update will consume the rest. This ensures ->clock_task is
616 * It does however cause some slight miss-attribution of {soft,}irq
617 * time, a more accurate solution would be to update the irq_time using
618 * the current rq->clock timestamp, except that would require using
621 if (irq_delta > delta)
624 rq->prev_irq_time += irq_delta;
627 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
628 if (static_key_false((¶virt_steal_rq_enabled))) {
629 steal = paravirt_steal_clock(cpu_of(rq));
630 steal -= rq->prev_steal_time_rq;
632 if (unlikely(steal > delta))
635 rq->prev_steal_time_rq += steal;
640 rq->clock_task += delta;
642 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
643 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
644 update_irq_load_avg(rq, irq_delta + steal);
646 update_rq_clock_pelt(rq, delta);
649 void update_rq_clock(struct rq *rq)
653 lockdep_assert_rq_held(rq);
655 if (rq->clock_update_flags & RQCF_ACT_SKIP)
658 #ifdef CONFIG_SCHED_DEBUG
659 if (sched_feat(WARN_DOUBLE_CLOCK))
660 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
661 rq->clock_update_flags |= RQCF_UPDATED;
664 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
668 update_rq_clock_task(rq, delta);
671 #ifdef CONFIG_SCHED_HRTICK
673 * Use HR-timers to deliver accurate preemption points.
676 static void hrtick_clear(struct rq *rq)
678 if (hrtimer_active(&rq->hrtick_timer))
679 hrtimer_cancel(&rq->hrtick_timer);
683 * High-resolution timer tick.
684 * Runs from hardirq context with interrupts disabled.
686 static enum hrtimer_restart hrtick(struct hrtimer *timer)
688 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
691 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
695 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
698 return HRTIMER_NORESTART;
703 static void __hrtick_restart(struct rq *rq)
705 struct hrtimer *timer = &rq->hrtick_timer;
706 ktime_t time = rq->hrtick_time;
708 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
712 * called from hardirq (IPI) context
714 static void __hrtick_start(void *arg)
720 __hrtick_restart(rq);
725 * Called to set the hrtick timer state.
727 * called with rq->lock held and irqs disabled
729 void hrtick_start(struct rq *rq, u64 delay)
731 struct hrtimer *timer = &rq->hrtick_timer;
735 * Don't schedule slices shorter than 10000ns, that just
736 * doesn't make sense and can cause timer DoS.
738 delta = max_t(s64, delay, 10000LL);
739 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
742 __hrtick_restart(rq);
744 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
749 * Called to set the hrtick timer state.
751 * called with rq->lock held and irqs disabled
753 void hrtick_start(struct rq *rq, u64 delay)
756 * Don't schedule slices shorter than 10000ns, that just
757 * doesn't make sense. Rely on vruntime for fairness.
759 delay = max_t(u64, delay, 10000LL);
760 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
761 HRTIMER_MODE_REL_PINNED_HARD);
764 #endif /* CONFIG_SMP */
766 static void hrtick_rq_init(struct rq *rq)
769 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
771 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
772 rq->hrtick_timer.function = hrtick;
774 #else /* CONFIG_SCHED_HRTICK */
775 static inline void hrtick_clear(struct rq *rq)
779 static inline void hrtick_rq_init(struct rq *rq)
782 #endif /* CONFIG_SCHED_HRTICK */
785 * cmpxchg based fetch_or, macro so it works for different integer types
787 #define fetch_or(ptr, mask) \
789 typeof(ptr) _ptr = (ptr); \
790 typeof(mask) _mask = (mask); \
791 typeof(*_ptr) _old, _val = *_ptr; \
794 _old = cmpxchg(_ptr, _val, _val | _mask); \
802 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
804 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
805 * this avoids any races wrt polling state changes and thereby avoids
808 static bool set_nr_and_not_polling(struct task_struct *p)
810 struct thread_info *ti = task_thread_info(p);
811 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
815 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
817 * If this returns true, then the idle task promises to call
818 * sched_ttwu_pending() and reschedule soon.
820 static bool set_nr_if_polling(struct task_struct *p)
822 struct thread_info *ti = task_thread_info(p);
823 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
826 if (!(val & _TIF_POLLING_NRFLAG))
828 if (val & _TIF_NEED_RESCHED)
830 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
839 static bool set_nr_and_not_polling(struct task_struct *p)
841 set_tsk_need_resched(p);
846 static bool set_nr_if_polling(struct task_struct *p)
853 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
855 struct wake_q_node *node = &task->wake_q;
858 * Atomically grab the task, if ->wake_q is !nil already it means
859 * it's already queued (either by us or someone else) and will get the
860 * wakeup due to that.
862 * In order to ensure that a pending wakeup will observe our pending
863 * state, even in the failed case, an explicit smp_mb() must be used.
865 smp_mb__before_atomic();
866 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
870 * The head is context local, there can be no concurrency.
873 head->lastp = &node->next;
878 * wake_q_add() - queue a wakeup for 'later' waking.
879 * @head: the wake_q_head to add @task to
880 * @task: the task to queue for 'later' wakeup
882 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
883 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
886 * This function must be used as-if it were wake_up_process(); IOW the task
887 * must be ready to be woken at this location.
889 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
891 if (__wake_q_add(head, task))
892 get_task_struct(task);
896 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
897 * @head: the wake_q_head to add @task to
898 * @task: the task to queue for 'later' wakeup
900 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
901 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
904 * This function must be used as-if it were wake_up_process(); IOW the task
905 * must be ready to be woken at this location.
907 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
908 * that already hold reference to @task can call the 'safe' version and trust
909 * wake_q to do the right thing depending whether or not the @task is already
912 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
914 if (!__wake_q_add(head, task))
915 put_task_struct(task);
918 void wake_up_q(struct wake_q_head *head)
920 struct wake_q_node *node = head->first;
922 while (node != WAKE_Q_TAIL) {
923 struct task_struct *task;
925 task = container_of(node, struct task_struct, wake_q);
926 /* Task can safely be re-inserted now: */
928 task->wake_q.next = NULL;
931 * wake_up_process() executes a full barrier, which pairs with
932 * the queueing in wake_q_add() so as not to miss wakeups.
934 wake_up_process(task);
935 put_task_struct(task);
940 * resched_curr - mark rq's current task 'to be rescheduled now'.
942 * On UP this means the setting of the need_resched flag, on SMP it
943 * might also involve a cross-CPU call to trigger the scheduler on
946 void resched_curr(struct rq *rq)
948 struct task_struct *curr = rq->curr;
951 lockdep_assert_rq_held(rq);
953 if (test_tsk_need_resched(curr))
958 if (cpu == smp_processor_id()) {
959 set_tsk_need_resched(curr);
960 set_preempt_need_resched();
964 if (set_nr_and_not_polling(curr))
965 smp_send_reschedule(cpu);
967 trace_sched_wake_idle_without_ipi(cpu);
970 void resched_cpu(int cpu)
972 struct rq *rq = cpu_rq(cpu);
975 raw_spin_rq_lock_irqsave(rq, flags);
976 if (cpu_online(cpu) || cpu == smp_processor_id())
978 raw_spin_rq_unlock_irqrestore(rq, flags);
982 #ifdef CONFIG_NO_HZ_COMMON
984 * In the semi idle case, use the nearest busy CPU for migrating timers
985 * from an idle CPU. This is good for power-savings.
987 * We don't do similar optimization for completely idle system, as
988 * selecting an idle CPU will add more delays to the timers than intended
989 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
991 int get_nohz_timer_target(void)
993 int i, cpu = smp_processor_id(), default_cpu = -1;
994 struct sched_domain *sd;
996 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1003 for_each_domain(cpu, sd) {
1004 for_each_cpu_and(i, sched_domain_span(sd),
1005 housekeeping_cpumask(HK_FLAG_TIMER)) {
1016 if (default_cpu == -1)
1017 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1025 * When add_timer_on() enqueues a timer into the timer wheel of an
1026 * idle CPU then this timer might expire before the next timer event
1027 * which is scheduled to wake up that CPU. In case of a completely
1028 * idle system the next event might even be infinite time into the
1029 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1030 * leaves the inner idle loop so the newly added timer is taken into
1031 * account when the CPU goes back to idle and evaluates the timer
1032 * wheel for the next timer event.
1034 static void wake_up_idle_cpu(int cpu)
1036 struct rq *rq = cpu_rq(cpu);
1038 if (cpu == smp_processor_id())
1041 if (set_nr_and_not_polling(rq->idle))
1042 smp_send_reschedule(cpu);
1044 trace_sched_wake_idle_without_ipi(cpu);
1047 static bool wake_up_full_nohz_cpu(int cpu)
1050 * We just need the target to call irq_exit() and re-evaluate
1051 * the next tick. The nohz full kick at least implies that.
1052 * If needed we can still optimize that later with an
1055 if (cpu_is_offline(cpu))
1056 return true; /* Don't try to wake offline CPUs. */
1057 if (tick_nohz_full_cpu(cpu)) {
1058 if (cpu != smp_processor_id() ||
1059 tick_nohz_tick_stopped())
1060 tick_nohz_full_kick_cpu(cpu);
1068 * Wake up the specified CPU. If the CPU is going offline, it is the
1069 * caller's responsibility to deal with the lost wakeup, for example,
1070 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1072 void wake_up_nohz_cpu(int cpu)
1074 if (!wake_up_full_nohz_cpu(cpu))
1075 wake_up_idle_cpu(cpu);
1078 static void nohz_csd_func(void *info)
1080 struct rq *rq = info;
1081 int cpu = cpu_of(rq);
1085 * Release the rq::nohz_csd.
1087 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1088 WARN_ON(!(flags & NOHZ_KICK_MASK));
1090 rq->idle_balance = idle_cpu(cpu);
1091 if (rq->idle_balance && !need_resched()) {
1092 rq->nohz_idle_balance = flags;
1093 raise_softirq_irqoff(SCHED_SOFTIRQ);
1097 #endif /* CONFIG_NO_HZ_COMMON */
1099 #ifdef CONFIG_NO_HZ_FULL
1100 bool sched_can_stop_tick(struct rq *rq)
1102 int fifo_nr_running;
1104 /* Deadline tasks, even if single, need the tick */
1105 if (rq->dl.dl_nr_running)
1109 * If there are more than one RR tasks, we need the tick to affect the
1110 * actual RR behaviour.
1112 if (rq->rt.rr_nr_running) {
1113 if (rq->rt.rr_nr_running == 1)
1120 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1121 * forced preemption between FIFO tasks.
1123 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1124 if (fifo_nr_running)
1128 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1129 * if there's more than one we need the tick for involuntary
1132 if (rq->nr_running > 1)
1137 #endif /* CONFIG_NO_HZ_FULL */
1138 #endif /* CONFIG_SMP */
1140 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1141 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1143 * Iterate task_group tree rooted at *from, calling @down when first entering a
1144 * node and @up when leaving it for the final time.
1146 * Caller must hold rcu_lock or sufficient equivalent.
1148 int walk_tg_tree_from(struct task_group *from,
1149 tg_visitor down, tg_visitor up, void *data)
1151 struct task_group *parent, *child;
1157 ret = (*down)(parent, data);
1160 list_for_each_entry_rcu(child, &parent->children, siblings) {
1167 ret = (*up)(parent, data);
1168 if (ret || parent == from)
1172 parent = parent->parent;
1179 int tg_nop(struct task_group *tg, void *data)
1185 static void set_load_weight(struct task_struct *p, bool update_load)
1187 int prio = p->static_prio - MAX_RT_PRIO;
1188 struct load_weight *load = &p->se.load;
1191 * SCHED_IDLE tasks get minimal weight:
1193 if (task_has_idle_policy(p)) {
1194 load->weight = scale_load(WEIGHT_IDLEPRIO);
1195 load->inv_weight = WMULT_IDLEPRIO;
1200 * SCHED_OTHER tasks have to update their load when changing their
1203 if (update_load && p->sched_class == &fair_sched_class) {
1204 reweight_task(p, prio);
1206 load->weight = scale_load(sched_prio_to_weight[prio]);
1207 load->inv_weight = sched_prio_to_wmult[prio];
1211 #ifdef CONFIG_UCLAMP_TASK
1213 * Serializes updates of utilization clamp values
1215 * The (slow-path) user-space triggers utilization clamp value updates which
1216 * can require updates on (fast-path) scheduler's data structures used to
1217 * support enqueue/dequeue operations.
1218 * While the per-CPU rq lock protects fast-path update operations, user-space
1219 * requests are serialized using a mutex to reduce the risk of conflicting
1220 * updates or API abuses.
1222 static DEFINE_MUTEX(uclamp_mutex);
1224 /* Max allowed minimum utilization */
1225 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1227 /* Max allowed maximum utilization */
1228 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1231 * By default RT tasks run at the maximum performance point/capacity of the
1232 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1233 * SCHED_CAPACITY_SCALE.
1235 * This knob allows admins to change the default behavior when uclamp is being
1236 * used. In battery powered devices, particularly, running at the maximum
1237 * capacity and frequency will increase energy consumption and shorten the
1240 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1242 * This knob will not override the system default sched_util_clamp_min defined
1245 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1247 /* All clamps are required to be less or equal than these values */
1248 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1251 * This static key is used to reduce the uclamp overhead in the fast path. It
1252 * primarily disables the call to uclamp_rq_{inc, dec}() in
1253 * enqueue/dequeue_task().
1255 * This allows users to continue to enable uclamp in their kernel config with
1256 * minimum uclamp overhead in the fast path.
1258 * As soon as userspace modifies any of the uclamp knobs, the static key is
1259 * enabled, since we have an actual users that make use of uclamp
1262 * The knobs that would enable this static key are:
1264 * * A task modifying its uclamp value with sched_setattr().
1265 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1266 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1268 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1270 /* Integer rounded range for each bucket */
1271 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1273 #define for_each_clamp_id(clamp_id) \
1274 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1276 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1278 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1281 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1283 if (clamp_id == UCLAMP_MIN)
1285 return SCHED_CAPACITY_SCALE;
1288 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1289 unsigned int value, bool user_defined)
1291 uc_se->value = value;
1292 uc_se->bucket_id = uclamp_bucket_id(value);
1293 uc_se->user_defined = user_defined;
1296 static inline unsigned int
1297 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1298 unsigned int clamp_value)
1301 * Avoid blocked utilization pushing up the frequency when we go
1302 * idle (which drops the max-clamp) by retaining the last known
1305 if (clamp_id == UCLAMP_MAX) {
1306 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1310 return uclamp_none(UCLAMP_MIN);
1313 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1314 unsigned int clamp_value)
1316 /* Reset max-clamp retention only on idle exit */
1317 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1320 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1324 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1325 unsigned int clamp_value)
1327 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1328 int bucket_id = UCLAMP_BUCKETS - 1;
1331 * Since both min and max clamps are max aggregated, find the
1332 * top most bucket with tasks in.
1334 for ( ; bucket_id >= 0; bucket_id--) {
1335 if (!bucket[bucket_id].tasks)
1337 return bucket[bucket_id].value;
1340 /* No tasks -- default clamp values */
1341 return uclamp_idle_value(rq, clamp_id, clamp_value);
1344 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1346 unsigned int default_util_min;
1347 struct uclamp_se *uc_se;
1349 lockdep_assert_held(&p->pi_lock);
1351 uc_se = &p->uclamp_req[UCLAMP_MIN];
1353 /* Only sync if user didn't override the default */
1354 if (uc_se->user_defined)
1357 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1358 uclamp_se_set(uc_se, default_util_min, false);
1361 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1369 /* Protect updates to p->uclamp_* */
1370 rq = task_rq_lock(p, &rf);
1371 __uclamp_update_util_min_rt_default(p);
1372 task_rq_unlock(rq, p, &rf);
1375 static void uclamp_sync_util_min_rt_default(void)
1377 struct task_struct *g, *p;
1380 * copy_process() sysctl_uclamp
1381 * uclamp_min_rt = X;
1382 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1383 * // link thread smp_mb__after_spinlock()
1384 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1385 * sched_post_fork() for_each_process_thread()
1386 * __uclamp_sync_rt() __uclamp_sync_rt()
1388 * Ensures that either sched_post_fork() will observe the new
1389 * uclamp_min_rt or for_each_process_thread() will observe the new
1392 read_lock(&tasklist_lock);
1393 smp_mb__after_spinlock();
1394 read_unlock(&tasklist_lock);
1397 for_each_process_thread(g, p)
1398 uclamp_update_util_min_rt_default(p);
1402 static inline struct uclamp_se
1403 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1405 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1406 #ifdef CONFIG_UCLAMP_TASK_GROUP
1407 struct uclamp_se uc_max;
1410 * Tasks in autogroups or root task group will be
1411 * restricted by system defaults.
1413 if (task_group_is_autogroup(task_group(p)))
1415 if (task_group(p) == &root_task_group)
1418 uc_max = task_group(p)->uclamp[clamp_id];
1419 if (uc_req.value > uc_max.value || !uc_req.user_defined)
1427 * The effective clamp bucket index of a task depends on, by increasing
1429 * - the task specific clamp value, when explicitly requested from userspace
1430 * - the task group effective clamp value, for tasks not either in the root
1431 * group or in an autogroup
1432 * - the system default clamp value, defined by the sysadmin
1434 static inline struct uclamp_se
1435 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1437 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1438 struct uclamp_se uc_max = uclamp_default[clamp_id];
1440 /* System default restrictions always apply */
1441 if (unlikely(uc_req.value > uc_max.value))
1447 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1449 struct uclamp_se uc_eff;
1451 /* Task currently refcounted: use back-annotated (effective) value */
1452 if (p->uclamp[clamp_id].active)
1453 return (unsigned long)p->uclamp[clamp_id].value;
1455 uc_eff = uclamp_eff_get(p, clamp_id);
1457 return (unsigned long)uc_eff.value;
1461 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1462 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1463 * updates the rq's clamp value if required.
1465 * Tasks can have a task-specific value requested from user-space, track
1466 * within each bucket the maximum value for tasks refcounted in it.
1467 * This "local max aggregation" allows to track the exact "requested" value
1468 * for each bucket when all its RUNNABLE tasks require the same clamp.
1470 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1471 enum uclamp_id clamp_id)
1473 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1474 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1475 struct uclamp_bucket *bucket;
1477 lockdep_assert_rq_held(rq);
1479 /* Update task effective clamp */
1480 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1482 bucket = &uc_rq->bucket[uc_se->bucket_id];
1484 uc_se->active = true;
1486 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1489 * Local max aggregation: rq buckets always track the max
1490 * "requested" clamp value of its RUNNABLE tasks.
1492 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1493 bucket->value = uc_se->value;
1495 if (uc_se->value > READ_ONCE(uc_rq->value))
1496 WRITE_ONCE(uc_rq->value, uc_se->value);
1500 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1501 * is released. If this is the last task reference counting the rq's max
1502 * active clamp value, then the rq's clamp value is updated.
1504 * Both refcounted tasks and rq's cached clamp values are expected to be
1505 * always valid. If it's detected they are not, as defensive programming,
1506 * enforce the expected state and warn.
1508 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1509 enum uclamp_id clamp_id)
1511 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1512 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1513 struct uclamp_bucket *bucket;
1514 unsigned int bkt_clamp;
1515 unsigned int rq_clamp;
1517 lockdep_assert_rq_held(rq);
1520 * If sched_uclamp_used was enabled after task @p was enqueued,
1521 * we could end up with unbalanced call to uclamp_rq_dec_id().
1523 * In this case the uc_se->active flag should be false since no uclamp
1524 * accounting was performed at enqueue time and we can just return
1527 * Need to be careful of the following enqueue/dequeue ordering
1531 * // sched_uclamp_used gets enabled
1534 * // Must not decrement bucket->tasks here
1537 * where we could end up with stale data in uc_se and
1538 * bucket[uc_se->bucket_id].
1540 * The following check here eliminates the possibility of such race.
1542 if (unlikely(!uc_se->active))
1545 bucket = &uc_rq->bucket[uc_se->bucket_id];
1547 SCHED_WARN_ON(!bucket->tasks);
1548 if (likely(bucket->tasks))
1551 uc_se->active = false;
1554 * Keep "local max aggregation" simple and accept to (possibly)
1555 * overboost some RUNNABLE tasks in the same bucket.
1556 * The rq clamp bucket value is reset to its base value whenever
1557 * there are no more RUNNABLE tasks refcounting it.
1559 if (likely(bucket->tasks))
1562 rq_clamp = READ_ONCE(uc_rq->value);
1564 * Defensive programming: this should never happen. If it happens,
1565 * e.g. due to future modification, warn and fixup the expected value.
1567 SCHED_WARN_ON(bucket->value > rq_clamp);
1568 if (bucket->value >= rq_clamp) {
1569 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1570 WRITE_ONCE(uc_rq->value, bkt_clamp);
1574 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1576 enum uclamp_id clamp_id;
1579 * Avoid any overhead until uclamp is actually used by the userspace.
1581 * The condition is constructed such that a NOP is generated when
1582 * sched_uclamp_used is disabled.
1584 if (!static_branch_unlikely(&sched_uclamp_used))
1587 if (unlikely(!p->sched_class->uclamp_enabled))
1590 for_each_clamp_id(clamp_id)
1591 uclamp_rq_inc_id(rq, p, clamp_id);
1593 /* Reset clamp idle holding when there is one RUNNABLE task */
1594 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1595 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1598 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1600 enum uclamp_id clamp_id;
1603 * Avoid any overhead until uclamp is actually used by the userspace.
1605 * The condition is constructed such that a NOP is generated when
1606 * sched_uclamp_used is disabled.
1608 if (!static_branch_unlikely(&sched_uclamp_used))
1611 if (unlikely(!p->sched_class->uclamp_enabled))
1614 for_each_clamp_id(clamp_id)
1615 uclamp_rq_dec_id(rq, p, clamp_id);
1619 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1625 * Lock the task and the rq where the task is (or was) queued.
1627 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1628 * price to pay to safely serialize util_{min,max} updates with
1629 * enqueues, dequeues and migration operations.
1630 * This is the same locking schema used by __set_cpus_allowed_ptr().
1632 rq = task_rq_lock(p, &rf);
1635 * Setting the clamp bucket is serialized by task_rq_lock().
1636 * If the task is not yet RUNNABLE and its task_struct is not
1637 * affecting a valid clamp bucket, the next time it's enqueued,
1638 * it will already see the updated clamp bucket value.
1640 if (p->uclamp[clamp_id].active) {
1641 uclamp_rq_dec_id(rq, p, clamp_id);
1642 uclamp_rq_inc_id(rq, p, clamp_id);
1645 task_rq_unlock(rq, p, &rf);
1648 #ifdef CONFIG_UCLAMP_TASK_GROUP
1650 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1651 unsigned int clamps)
1653 enum uclamp_id clamp_id;
1654 struct css_task_iter it;
1655 struct task_struct *p;
1657 css_task_iter_start(css, 0, &it);
1658 while ((p = css_task_iter_next(&it))) {
1659 for_each_clamp_id(clamp_id) {
1660 if ((0x1 << clamp_id) & clamps)
1661 uclamp_update_active(p, clamp_id);
1664 css_task_iter_end(&it);
1667 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1668 static void uclamp_update_root_tg(void)
1670 struct task_group *tg = &root_task_group;
1672 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1673 sysctl_sched_uclamp_util_min, false);
1674 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1675 sysctl_sched_uclamp_util_max, false);
1678 cpu_util_update_eff(&root_task_group.css);
1682 static void uclamp_update_root_tg(void) { }
1685 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1686 void *buffer, size_t *lenp, loff_t *ppos)
1688 bool update_root_tg = false;
1689 int old_min, old_max, old_min_rt;
1692 mutex_lock(&uclamp_mutex);
1693 old_min = sysctl_sched_uclamp_util_min;
1694 old_max = sysctl_sched_uclamp_util_max;
1695 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1697 result = proc_dointvec(table, write, buffer, lenp, ppos);
1703 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1704 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1705 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1711 if (old_min != sysctl_sched_uclamp_util_min) {
1712 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1713 sysctl_sched_uclamp_util_min, false);
1714 update_root_tg = true;
1716 if (old_max != sysctl_sched_uclamp_util_max) {
1717 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1718 sysctl_sched_uclamp_util_max, false);
1719 update_root_tg = true;
1722 if (update_root_tg) {
1723 static_branch_enable(&sched_uclamp_used);
1724 uclamp_update_root_tg();
1727 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1728 static_branch_enable(&sched_uclamp_used);
1729 uclamp_sync_util_min_rt_default();
1733 * We update all RUNNABLE tasks only when task groups are in use.
1734 * Otherwise, keep it simple and do just a lazy update at each next
1735 * task enqueue time.
1741 sysctl_sched_uclamp_util_min = old_min;
1742 sysctl_sched_uclamp_util_max = old_max;
1743 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1745 mutex_unlock(&uclamp_mutex);
1750 static int uclamp_validate(struct task_struct *p,
1751 const struct sched_attr *attr)
1753 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1754 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1756 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1757 util_min = attr->sched_util_min;
1759 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1763 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1764 util_max = attr->sched_util_max;
1766 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1770 if (util_min != -1 && util_max != -1 && util_min > util_max)
1774 * We have valid uclamp attributes; make sure uclamp is enabled.
1776 * We need to do that here, because enabling static branches is a
1777 * blocking operation which obviously cannot be done while holding
1780 static_branch_enable(&sched_uclamp_used);
1785 static bool uclamp_reset(const struct sched_attr *attr,
1786 enum uclamp_id clamp_id,
1787 struct uclamp_se *uc_se)
1789 /* Reset on sched class change for a non user-defined clamp value. */
1790 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1791 !uc_se->user_defined)
1794 /* Reset on sched_util_{min,max} == -1. */
1795 if (clamp_id == UCLAMP_MIN &&
1796 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1797 attr->sched_util_min == -1) {
1801 if (clamp_id == UCLAMP_MAX &&
1802 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1803 attr->sched_util_max == -1) {
1810 static void __setscheduler_uclamp(struct task_struct *p,
1811 const struct sched_attr *attr)
1813 enum uclamp_id clamp_id;
1815 for_each_clamp_id(clamp_id) {
1816 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1819 if (!uclamp_reset(attr, clamp_id, uc_se))
1823 * RT by default have a 100% boost value that could be modified
1826 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1827 value = sysctl_sched_uclamp_util_min_rt_default;
1829 value = uclamp_none(clamp_id);
1831 uclamp_se_set(uc_se, value, false);
1835 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1838 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1839 attr->sched_util_min != -1) {
1840 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1841 attr->sched_util_min, true);
1844 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1845 attr->sched_util_max != -1) {
1846 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1847 attr->sched_util_max, true);
1851 static void uclamp_fork(struct task_struct *p)
1853 enum uclamp_id clamp_id;
1856 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1857 * as the task is still at its early fork stages.
1859 for_each_clamp_id(clamp_id)
1860 p->uclamp[clamp_id].active = false;
1862 if (likely(!p->sched_reset_on_fork))
1865 for_each_clamp_id(clamp_id) {
1866 uclamp_se_set(&p->uclamp_req[clamp_id],
1867 uclamp_none(clamp_id), false);
1871 static void uclamp_post_fork(struct task_struct *p)
1873 uclamp_update_util_min_rt_default(p);
1876 static void __init init_uclamp_rq(struct rq *rq)
1878 enum uclamp_id clamp_id;
1879 struct uclamp_rq *uc_rq = rq->uclamp;
1881 for_each_clamp_id(clamp_id) {
1882 uc_rq[clamp_id] = (struct uclamp_rq) {
1883 .value = uclamp_none(clamp_id)
1887 rq->uclamp_flags = 0;
1890 static void __init init_uclamp(void)
1892 struct uclamp_se uc_max = {};
1893 enum uclamp_id clamp_id;
1896 for_each_possible_cpu(cpu)
1897 init_uclamp_rq(cpu_rq(cpu));
1899 for_each_clamp_id(clamp_id) {
1900 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1901 uclamp_none(clamp_id), false);
1904 /* System defaults allow max clamp values for both indexes */
1905 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1906 for_each_clamp_id(clamp_id) {
1907 uclamp_default[clamp_id] = uc_max;
1908 #ifdef CONFIG_UCLAMP_TASK_GROUP
1909 root_task_group.uclamp_req[clamp_id] = uc_max;
1910 root_task_group.uclamp[clamp_id] = uc_max;
1915 #else /* CONFIG_UCLAMP_TASK */
1916 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1917 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1918 static inline int uclamp_validate(struct task_struct *p,
1919 const struct sched_attr *attr)
1923 static void __setscheduler_uclamp(struct task_struct *p,
1924 const struct sched_attr *attr) { }
1925 static inline void uclamp_fork(struct task_struct *p) { }
1926 static inline void uclamp_post_fork(struct task_struct *p) { }
1927 static inline void init_uclamp(void) { }
1928 #endif /* CONFIG_UCLAMP_TASK */
1930 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1932 if (!(flags & ENQUEUE_NOCLOCK))
1933 update_rq_clock(rq);
1935 if (!(flags & ENQUEUE_RESTORE)) {
1936 sched_info_enqueue(rq, p);
1937 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1940 uclamp_rq_inc(rq, p);
1941 p->sched_class->enqueue_task(rq, p, flags);
1943 if (sched_core_enabled(rq))
1944 sched_core_enqueue(rq, p);
1947 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1949 if (sched_core_enabled(rq))
1950 sched_core_dequeue(rq, p);
1952 if (!(flags & DEQUEUE_NOCLOCK))
1953 update_rq_clock(rq);
1955 if (!(flags & DEQUEUE_SAVE)) {
1956 sched_info_dequeue(rq, p);
1957 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1960 uclamp_rq_dec(rq, p);
1961 p->sched_class->dequeue_task(rq, p, flags);
1964 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1966 enqueue_task(rq, p, flags);
1968 p->on_rq = TASK_ON_RQ_QUEUED;
1971 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1973 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1975 dequeue_task(rq, p, flags);
1979 * __normal_prio - return the priority that is based on the static prio
1981 static inline int __normal_prio(struct task_struct *p)
1983 return p->static_prio;
1987 * Calculate the expected normal priority: i.e. priority
1988 * without taking RT-inheritance into account. Might be
1989 * boosted by interactivity modifiers. Changes upon fork,
1990 * setprio syscalls, and whenever the interactivity
1991 * estimator recalculates.
1993 static inline int normal_prio(struct task_struct *p)
1997 if (task_has_dl_policy(p))
1998 prio = MAX_DL_PRIO-1;
1999 else if (task_has_rt_policy(p))
2000 prio = MAX_RT_PRIO-1 - p->rt_priority;
2002 prio = __normal_prio(p);
2007 * Calculate the current priority, i.e. the priority
2008 * taken into account by the scheduler. This value might
2009 * be boosted by RT tasks, or might be boosted by
2010 * interactivity modifiers. Will be RT if the task got
2011 * RT-boosted. If not then it returns p->normal_prio.
2013 static int effective_prio(struct task_struct *p)
2015 p->normal_prio = normal_prio(p);
2017 * If we are RT tasks or we were boosted to RT priority,
2018 * keep the priority unchanged. Otherwise, update priority
2019 * to the normal priority:
2021 if (!rt_prio(p->prio))
2022 return p->normal_prio;
2027 * task_curr - is this task currently executing on a CPU?
2028 * @p: the task in question.
2030 * Return: 1 if the task is currently executing. 0 otherwise.
2032 inline int task_curr(const struct task_struct *p)
2034 return cpu_curr(task_cpu(p)) == p;
2038 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2039 * use the balance_callback list if you want balancing.
2041 * this means any call to check_class_changed() must be followed by a call to
2042 * balance_callback().
2044 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2045 const struct sched_class *prev_class,
2048 if (prev_class != p->sched_class) {
2049 if (prev_class->switched_from)
2050 prev_class->switched_from(rq, p);
2052 p->sched_class->switched_to(rq, p);
2053 } else if (oldprio != p->prio || dl_task(p))
2054 p->sched_class->prio_changed(rq, p, oldprio);
2057 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2059 if (p->sched_class == rq->curr->sched_class)
2060 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2061 else if (p->sched_class > rq->curr->sched_class)
2065 * A queue event has occurred, and we're going to schedule. In
2066 * this case, we can save a useless back to back clock update.
2068 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2069 rq_clock_skip_update(rq);
2075 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2077 static int __set_cpus_allowed_ptr(struct task_struct *p,
2078 const struct cpumask *new_mask,
2081 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2083 if (likely(!p->migration_disabled))
2086 if (p->cpus_ptr != &p->cpus_mask)
2090 * Violates locking rules! see comment in __do_set_cpus_allowed().
2092 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2095 void migrate_disable(void)
2097 struct task_struct *p = current;
2099 if (p->migration_disabled) {
2100 p->migration_disabled++;
2105 this_rq()->nr_pinned++;
2106 p->migration_disabled = 1;
2109 EXPORT_SYMBOL_GPL(migrate_disable);
2111 void migrate_enable(void)
2113 struct task_struct *p = current;
2115 if (p->migration_disabled > 1) {
2116 p->migration_disabled--;
2121 * Ensure stop_task runs either before or after this, and that
2122 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2125 if (p->cpus_ptr != &p->cpus_mask)
2126 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2128 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2129 * regular cpus_mask, otherwise things that race (eg.
2130 * select_fallback_rq) get confused.
2133 p->migration_disabled = 0;
2134 this_rq()->nr_pinned--;
2137 EXPORT_SYMBOL_GPL(migrate_enable);
2139 static inline bool rq_has_pinned_tasks(struct rq *rq)
2141 return rq->nr_pinned;
2145 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2146 * __set_cpus_allowed_ptr() and select_fallback_rq().
2148 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2150 /* When not in the task's cpumask, no point in looking further. */
2151 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2154 /* migrate_disabled() must be allowed to finish. */
2155 if (is_migration_disabled(p))
2156 return cpu_online(cpu);
2158 /* Non kernel threads are not allowed during either online or offline. */
2159 if (!(p->flags & PF_KTHREAD))
2160 return cpu_active(cpu);
2162 /* KTHREAD_IS_PER_CPU is always allowed. */
2163 if (kthread_is_per_cpu(p))
2164 return cpu_online(cpu);
2166 /* Regular kernel threads don't get to stay during offline. */
2170 /* But are allowed during online. */
2171 return cpu_online(cpu);
2175 * This is how migration works:
2177 * 1) we invoke migration_cpu_stop() on the target CPU using
2179 * 2) stopper starts to run (implicitly forcing the migrated thread
2181 * 3) it checks whether the migrated task is still in the wrong runqueue.
2182 * 4) if it's in the wrong runqueue then the migration thread removes
2183 * it and puts it into the right queue.
2184 * 5) stopper completes and stop_one_cpu() returns and the migration
2189 * move_queued_task - move a queued task to new rq.
2191 * Returns (locked) new rq. Old rq's lock is released.
2193 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2194 struct task_struct *p, int new_cpu)
2196 lockdep_assert_rq_held(rq);
2198 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2199 set_task_cpu(p, new_cpu);
2202 rq = cpu_rq(new_cpu);
2205 BUG_ON(task_cpu(p) != new_cpu);
2206 activate_task(rq, p, 0);
2207 check_preempt_curr(rq, p, 0);
2212 struct migration_arg {
2213 struct task_struct *task;
2215 struct set_affinity_pending *pending;
2219 * @refs: number of wait_for_completion()
2220 * @stop_pending: is @stop_work in use
2222 struct set_affinity_pending {
2224 unsigned int stop_pending;
2225 struct completion done;
2226 struct cpu_stop_work stop_work;
2227 struct migration_arg arg;
2231 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2232 * this because either it can't run here any more (set_cpus_allowed()
2233 * away from this CPU, or CPU going down), or because we're
2234 * attempting to rebalance this task on exec (sched_exec).
2236 * So we race with normal scheduler movements, but that's OK, as long
2237 * as the task is no longer on this CPU.
2239 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2240 struct task_struct *p, int dest_cpu)
2242 /* Affinity changed (again). */
2243 if (!is_cpu_allowed(p, dest_cpu))
2246 update_rq_clock(rq);
2247 rq = move_queued_task(rq, rf, p, dest_cpu);
2253 * migration_cpu_stop - this will be executed by a highprio stopper thread
2254 * and performs thread migration by bumping thread off CPU then
2255 * 'pushing' onto another runqueue.
2257 static int migration_cpu_stop(void *data)
2259 struct migration_arg *arg = data;
2260 struct set_affinity_pending *pending = arg->pending;
2261 struct task_struct *p = arg->task;
2262 int dest_cpu = arg->dest_cpu;
2263 struct rq *rq = this_rq();
2264 bool complete = false;
2268 * The original target CPU might have gone down and we might
2269 * be on another CPU but it doesn't matter.
2271 local_irq_save(rf.flags);
2273 * We need to explicitly wake pending tasks before running
2274 * __migrate_task() such that we will not miss enforcing cpus_ptr
2275 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2277 flush_smp_call_function_from_idle();
2279 raw_spin_lock(&p->pi_lock);
2283 * If we were passed a pending, then ->stop_pending was set, thus
2284 * p->migration_pending must have remained stable.
2286 WARN_ON_ONCE(pending && pending != p->migration_pending);
2289 * If task_rq(p) != rq, it cannot be migrated here, because we're
2290 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2291 * we're holding p->pi_lock.
2293 if (task_rq(p) == rq) {
2294 if (is_migration_disabled(p))
2298 p->migration_pending = NULL;
2303 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2306 dest_cpu = cpumask_any_distribute(&p->cpus_mask);
2309 if (task_on_rq_queued(p))
2310 rq = __migrate_task(rq, &rf, p, dest_cpu);
2312 p->wake_cpu = dest_cpu;
2315 * XXX __migrate_task() can fail, at which point we might end
2316 * up running on a dodgy CPU, AFAICT this can only happen
2317 * during CPU hotplug, at which point we'll get pushed out
2318 * anyway, so it's probably not a big deal.
2321 } else if (pending) {
2323 * This happens when we get migrated between migrate_enable()'s
2324 * preempt_enable() and scheduling the stopper task. At that
2325 * point we're a regular task again and not current anymore.
2327 * A !PREEMPT kernel has a giant hole here, which makes it far
2332 * The task moved before the stopper got to run. We're holding
2333 * ->pi_lock, so the allowed mask is stable - if it got
2334 * somewhere allowed, we're done.
2336 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2337 p->migration_pending = NULL;
2343 * When migrate_enable() hits a rq mis-match we can't reliably
2344 * determine is_migration_disabled() and so have to chase after
2347 WARN_ON_ONCE(!pending->stop_pending);
2348 task_rq_unlock(rq, p, &rf);
2349 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2350 &pending->arg, &pending->stop_work);
2355 pending->stop_pending = false;
2356 task_rq_unlock(rq, p, &rf);
2359 complete_all(&pending->done);
2364 int push_cpu_stop(void *arg)
2366 struct rq *lowest_rq = NULL, *rq = this_rq();
2367 struct task_struct *p = arg;
2369 raw_spin_lock_irq(&p->pi_lock);
2370 raw_spin_rq_lock(rq);
2372 if (task_rq(p) != rq)
2375 if (is_migration_disabled(p)) {
2376 p->migration_flags |= MDF_PUSH;
2380 p->migration_flags &= ~MDF_PUSH;
2382 if (p->sched_class->find_lock_rq)
2383 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2388 // XXX validate p is still the highest prio task
2389 if (task_rq(p) == rq) {
2390 deactivate_task(rq, p, 0);
2391 set_task_cpu(p, lowest_rq->cpu);
2392 activate_task(lowest_rq, p, 0);
2393 resched_curr(lowest_rq);
2396 double_unlock_balance(rq, lowest_rq);
2399 rq->push_busy = false;
2400 raw_spin_rq_unlock(rq);
2401 raw_spin_unlock_irq(&p->pi_lock);
2408 * sched_class::set_cpus_allowed must do the below, but is not required to
2409 * actually call this function.
2411 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2413 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2414 p->cpus_ptr = new_mask;
2418 cpumask_copy(&p->cpus_mask, new_mask);
2419 p->nr_cpus_allowed = cpumask_weight(new_mask);
2423 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2425 struct rq *rq = task_rq(p);
2426 bool queued, running;
2429 * This here violates the locking rules for affinity, since we're only
2430 * supposed to change these variables while holding both rq->lock and
2433 * HOWEVER, it magically works, because ttwu() is the only code that
2434 * accesses these variables under p->pi_lock and only does so after
2435 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2436 * before finish_task().
2438 * XXX do further audits, this smells like something putrid.
2440 if (flags & SCA_MIGRATE_DISABLE)
2441 SCHED_WARN_ON(!p->on_cpu);
2443 lockdep_assert_held(&p->pi_lock);
2445 queued = task_on_rq_queued(p);
2446 running = task_current(rq, p);
2450 * Because __kthread_bind() calls this on blocked tasks without
2453 lockdep_assert_rq_held(rq);
2454 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2457 put_prev_task(rq, p);
2459 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2462 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2464 set_next_task(rq, p);
2467 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2469 __do_set_cpus_allowed(p, new_mask, 0);
2473 * This function is wildly self concurrent; here be dragons.
2476 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2477 * designated task is enqueued on an allowed CPU. If that task is currently
2478 * running, we have to kick it out using the CPU stopper.
2480 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2483 * Initial conditions: P0->cpus_mask = [0, 1]
2487 * migrate_disable();
2489 * set_cpus_allowed_ptr(P0, [1]);
2491 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2492 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2493 * This means we need the following scheme:
2497 * migrate_disable();
2499 * set_cpus_allowed_ptr(P0, [1]);
2503 * __set_cpus_allowed_ptr();
2504 * <wakes local stopper>
2505 * `--> <woken on migration completion>
2507 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2508 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2509 * task p are serialized by p->pi_lock, which we can leverage: the one that
2510 * should come into effect at the end of the Migrate-Disable region is the last
2511 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2512 * but we still need to properly signal those waiting tasks at the appropriate
2515 * This is implemented using struct set_affinity_pending. The first
2516 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2517 * setup an instance of that struct and install it on the targeted task_struct.
2518 * Any and all further callers will reuse that instance. Those then wait for
2519 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2520 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2523 * (1) In the cases covered above. There is one more where the completion is
2524 * signaled within affine_move_task() itself: when a subsequent affinity request
2525 * occurs after the stopper bailed out due to the targeted task still being
2526 * Migrate-Disable. Consider:
2528 * Initial conditions: P0->cpus_mask = [0, 1]
2532 * migrate_disable();
2534 * set_cpus_allowed_ptr(P0, [1]);
2537 * migration_cpu_stop()
2538 * is_migration_disabled()
2540 * set_cpus_allowed_ptr(P0, [0, 1]);
2541 * <signal completion>
2544 * Note that the above is safe vs a concurrent migrate_enable(), as any
2545 * pending affinity completion is preceded by an uninstallation of
2546 * p->migration_pending done with p->pi_lock held.
2548 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2549 int dest_cpu, unsigned int flags)
2551 struct set_affinity_pending my_pending = { }, *pending = NULL;
2552 bool stop_pending, complete = false;
2554 /* Can the task run on the task's current CPU? If so, we're done */
2555 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2556 struct task_struct *push_task = NULL;
2558 if ((flags & SCA_MIGRATE_ENABLE) &&
2559 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2560 rq->push_busy = true;
2561 push_task = get_task_struct(p);
2565 * If there are pending waiters, but no pending stop_work,
2566 * then complete now.
2568 pending = p->migration_pending;
2569 if (pending && !pending->stop_pending) {
2570 p->migration_pending = NULL;
2574 task_rq_unlock(rq, p, rf);
2577 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2582 complete_all(&pending->done);
2587 if (!(flags & SCA_MIGRATE_ENABLE)) {
2588 /* serialized by p->pi_lock */
2589 if (!p->migration_pending) {
2590 /* Install the request */
2591 refcount_set(&my_pending.refs, 1);
2592 init_completion(&my_pending.done);
2593 my_pending.arg = (struct migration_arg) {
2595 .dest_cpu = -1, /* any */
2596 .pending = &my_pending,
2599 p->migration_pending = &my_pending;
2601 pending = p->migration_pending;
2602 refcount_inc(&pending->refs);
2605 pending = p->migration_pending;
2607 * - !MIGRATE_ENABLE:
2608 * we'll have installed a pending if there wasn't one already.
2611 * we're here because the current CPU isn't matching anymore,
2612 * the only way that can happen is because of a concurrent
2613 * set_cpus_allowed_ptr() call, which should then still be
2614 * pending completion.
2616 * Either way, we really should have a @pending here.
2618 if (WARN_ON_ONCE(!pending)) {
2619 task_rq_unlock(rq, p, rf);
2623 if (task_running(rq, p) || p->state == TASK_WAKING) {
2625 * MIGRATE_ENABLE gets here because 'p == current', but for
2626 * anything else we cannot do is_migration_disabled(), punt
2627 * and have the stopper function handle it all race-free.
2629 stop_pending = pending->stop_pending;
2631 pending->stop_pending = true;
2633 if (flags & SCA_MIGRATE_ENABLE)
2634 p->migration_flags &= ~MDF_PUSH;
2636 task_rq_unlock(rq, p, rf);
2638 if (!stop_pending) {
2639 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2640 &pending->arg, &pending->stop_work);
2643 if (flags & SCA_MIGRATE_ENABLE)
2647 if (!is_migration_disabled(p)) {
2648 if (task_on_rq_queued(p))
2649 rq = move_queued_task(rq, rf, p, dest_cpu);
2651 if (!pending->stop_pending) {
2652 p->migration_pending = NULL;
2656 task_rq_unlock(rq, p, rf);
2659 complete_all(&pending->done);
2662 wait_for_completion(&pending->done);
2664 if (refcount_dec_and_test(&pending->refs))
2665 wake_up_var(&pending->refs); /* No UaF, just an address */
2668 * Block the original owner of &pending until all subsequent callers
2669 * have seen the completion and decremented the refcount
2671 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2674 WARN_ON_ONCE(my_pending.stop_pending);
2680 * Change a given task's CPU affinity. Migrate the thread to a
2681 * proper CPU and schedule it away if the CPU it's executing on
2682 * is removed from the allowed bitmask.
2684 * NOTE: the caller must have a valid reference to the task, the
2685 * task must not exit() & deallocate itself prematurely. The
2686 * call is not atomic; no spinlocks may be held.
2688 static int __set_cpus_allowed_ptr(struct task_struct *p,
2689 const struct cpumask *new_mask,
2692 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2693 unsigned int dest_cpu;
2698 rq = task_rq_lock(p, &rf);
2699 update_rq_clock(rq);
2701 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2703 * Kernel threads are allowed on online && !active CPUs,
2704 * however, during cpu-hot-unplug, even these might get pushed
2705 * away if not KTHREAD_IS_PER_CPU.
2707 * Specifically, migration_disabled() tasks must not fail the
2708 * cpumask_any_and_distribute() pick below, esp. so on
2709 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2710 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2712 cpu_valid_mask = cpu_online_mask;
2716 * Must re-check here, to close a race against __kthread_bind(),
2717 * sched_setaffinity() is not guaranteed to observe the flag.
2719 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2724 if (!(flags & SCA_MIGRATE_ENABLE)) {
2725 if (cpumask_equal(&p->cpus_mask, new_mask))
2728 if (WARN_ON_ONCE(p == current &&
2729 is_migration_disabled(p) &&
2730 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2737 * Picking a ~random cpu helps in cases where we are changing affinity
2738 * for groups of tasks (ie. cpuset), so that load balancing is not
2739 * immediately required to distribute the tasks within their new mask.
2741 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2742 if (dest_cpu >= nr_cpu_ids) {
2747 __do_set_cpus_allowed(p, new_mask, flags);
2749 return affine_move_task(rq, p, &rf, dest_cpu, flags);
2752 task_rq_unlock(rq, p, &rf);
2757 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2759 return __set_cpus_allowed_ptr(p, new_mask, 0);
2761 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2763 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2765 #ifdef CONFIG_SCHED_DEBUG
2767 * We should never call set_task_cpu() on a blocked task,
2768 * ttwu() will sort out the placement.
2770 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2774 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2775 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2776 * time relying on p->on_rq.
2778 WARN_ON_ONCE(p->state == TASK_RUNNING &&
2779 p->sched_class == &fair_sched_class &&
2780 (p->on_rq && !task_on_rq_migrating(p)));
2782 #ifdef CONFIG_LOCKDEP
2784 * The caller should hold either p->pi_lock or rq->lock, when changing
2785 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2787 * sched_move_task() holds both and thus holding either pins the cgroup,
2790 * Furthermore, all task_rq users should acquire both locks, see
2793 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2794 lockdep_is_held(__rq_lockp(task_rq(p)))));
2797 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2799 WARN_ON_ONCE(!cpu_online(new_cpu));
2801 WARN_ON_ONCE(is_migration_disabled(p));
2804 trace_sched_migrate_task(p, new_cpu);
2806 if (task_cpu(p) != new_cpu) {
2807 if (p->sched_class->migrate_task_rq)
2808 p->sched_class->migrate_task_rq(p, new_cpu);
2809 p->se.nr_migrations++;
2811 perf_event_task_migrate(p);
2814 __set_task_cpu(p, new_cpu);
2817 #ifdef CONFIG_NUMA_BALANCING
2818 static void __migrate_swap_task(struct task_struct *p, int cpu)
2820 if (task_on_rq_queued(p)) {
2821 struct rq *src_rq, *dst_rq;
2822 struct rq_flags srf, drf;
2824 src_rq = task_rq(p);
2825 dst_rq = cpu_rq(cpu);
2827 rq_pin_lock(src_rq, &srf);
2828 rq_pin_lock(dst_rq, &drf);
2830 deactivate_task(src_rq, p, 0);
2831 set_task_cpu(p, cpu);
2832 activate_task(dst_rq, p, 0);
2833 check_preempt_curr(dst_rq, p, 0);
2835 rq_unpin_lock(dst_rq, &drf);
2836 rq_unpin_lock(src_rq, &srf);
2840 * Task isn't running anymore; make it appear like we migrated
2841 * it before it went to sleep. This means on wakeup we make the
2842 * previous CPU our target instead of where it really is.
2848 struct migration_swap_arg {
2849 struct task_struct *src_task, *dst_task;
2850 int src_cpu, dst_cpu;
2853 static int migrate_swap_stop(void *data)
2855 struct migration_swap_arg *arg = data;
2856 struct rq *src_rq, *dst_rq;
2859 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2862 src_rq = cpu_rq(arg->src_cpu);
2863 dst_rq = cpu_rq(arg->dst_cpu);
2865 double_raw_lock(&arg->src_task->pi_lock,
2866 &arg->dst_task->pi_lock);
2867 double_rq_lock(src_rq, dst_rq);
2869 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2872 if (task_cpu(arg->src_task) != arg->src_cpu)
2875 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2878 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2881 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2882 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2887 double_rq_unlock(src_rq, dst_rq);
2888 raw_spin_unlock(&arg->dst_task->pi_lock);
2889 raw_spin_unlock(&arg->src_task->pi_lock);
2895 * Cross migrate two tasks
2897 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2898 int target_cpu, int curr_cpu)
2900 struct migration_swap_arg arg;
2903 arg = (struct migration_swap_arg){
2905 .src_cpu = curr_cpu,
2907 .dst_cpu = target_cpu,
2910 if (arg.src_cpu == arg.dst_cpu)
2914 * These three tests are all lockless; this is OK since all of them
2915 * will be re-checked with proper locks held further down the line.
2917 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2920 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2923 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2926 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2927 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2932 #endif /* CONFIG_NUMA_BALANCING */
2935 * wait_task_inactive - wait for a thread to unschedule.
2937 * If @match_state is nonzero, it's the @p->state value just checked and
2938 * not expected to change. If it changes, i.e. @p might have woken up,
2939 * then return zero. When we succeed in waiting for @p to be off its CPU,
2940 * we return a positive number (its total switch count). If a second call
2941 * a short while later returns the same number, the caller can be sure that
2942 * @p has remained unscheduled the whole time.
2944 * The caller must ensure that the task *will* unschedule sometime soon,
2945 * else this function might spin for a *long* time. This function can't
2946 * be called with interrupts off, or it may introduce deadlock with
2947 * smp_call_function() if an IPI is sent by the same process we are
2948 * waiting to become inactive.
2950 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2952 int running, queued;
2959 * We do the initial early heuristics without holding
2960 * any task-queue locks at all. We'll only try to get
2961 * the runqueue lock when things look like they will
2967 * If the task is actively running on another CPU
2968 * still, just relax and busy-wait without holding
2971 * NOTE! Since we don't hold any locks, it's not
2972 * even sure that "rq" stays as the right runqueue!
2973 * But we don't care, since "task_running()" will
2974 * return false if the runqueue has changed and p
2975 * is actually now running somewhere else!
2977 while (task_running(rq, p)) {
2978 if (match_state && unlikely(p->state != match_state))
2984 * Ok, time to look more closely! We need the rq
2985 * lock now, to be *sure*. If we're wrong, we'll
2986 * just go back and repeat.
2988 rq = task_rq_lock(p, &rf);
2989 trace_sched_wait_task(p);
2990 running = task_running(rq, p);
2991 queued = task_on_rq_queued(p);
2993 if (!match_state || p->state == match_state)
2994 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2995 task_rq_unlock(rq, p, &rf);
2998 * If it changed from the expected state, bail out now.
3000 if (unlikely(!ncsw))
3004 * Was it really running after all now that we
3005 * checked with the proper locks actually held?
3007 * Oops. Go back and try again..
3009 if (unlikely(running)) {
3015 * It's not enough that it's not actively running,
3016 * it must be off the runqueue _entirely_, and not
3019 * So if it was still runnable (but just not actively
3020 * running right now), it's preempted, and we should
3021 * yield - it could be a while.
3023 if (unlikely(queued)) {
3024 ktime_t to = NSEC_PER_SEC / HZ;
3026 set_current_state(TASK_UNINTERRUPTIBLE);
3027 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
3032 * Ahh, all good. It wasn't running, and it wasn't
3033 * runnable, which means that it will never become
3034 * running in the future either. We're all done!
3043 * kick_process - kick a running thread to enter/exit the kernel
3044 * @p: the to-be-kicked thread
3046 * Cause a process which is running on another CPU to enter
3047 * kernel-mode, without any delay. (to get signals handled.)
3049 * NOTE: this function doesn't have to take the runqueue lock,
3050 * because all it wants to ensure is that the remote task enters
3051 * the kernel. If the IPI races and the task has been migrated
3052 * to another CPU then no harm is done and the purpose has been
3055 void kick_process(struct task_struct *p)
3061 if ((cpu != smp_processor_id()) && task_curr(p))
3062 smp_send_reschedule(cpu);
3065 EXPORT_SYMBOL_GPL(kick_process);
3068 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3070 * A few notes on cpu_active vs cpu_online:
3072 * - cpu_active must be a subset of cpu_online
3074 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3075 * see __set_cpus_allowed_ptr(). At this point the newly online
3076 * CPU isn't yet part of the sched domains, and balancing will not
3079 * - on CPU-down we clear cpu_active() to mask the sched domains and
3080 * avoid the load balancer to place new tasks on the to be removed
3081 * CPU. Existing tasks will remain running there and will be taken
3084 * This means that fallback selection must not select !active CPUs.
3085 * And can assume that any active CPU must be online. Conversely
3086 * select_task_rq() below may allow selection of !active CPUs in order
3087 * to satisfy the above rules.
3089 static int select_fallback_rq(int cpu, struct task_struct *p)
3091 int nid = cpu_to_node(cpu);
3092 const struct cpumask *nodemask = NULL;
3093 enum { cpuset, possible, fail } state = cpuset;
3097 * If the node that the CPU is on has been offlined, cpu_to_node()
3098 * will return -1. There is no CPU on the node, and we should
3099 * select the CPU on the other node.
3102 nodemask = cpumask_of_node(nid);
3104 /* Look for allowed, online CPU in same node. */
3105 for_each_cpu(dest_cpu, nodemask) {
3106 if (!cpu_active(dest_cpu))
3108 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
3114 /* Any allowed, online CPU? */
3115 for_each_cpu(dest_cpu, p->cpus_ptr) {
3116 if (!is_cpu_allowed(p, dest_cpu))
3122 /* No more Mr. Nice Guy. */
3125 if (IS_ENABLED(CONFIG_CPUSETS)) {
3126 cpuset_cpus_allowed_fallback(p);
3133 * XXX When called from select_task_rq() we only
3134 * hold p->pi_lock and again violate locking order.
3136 * More yuck to audit.
3138 do_set_cpus_allowed(p, cpu_possible_mask);
3149 if (state != cpuset) {
3151 * Don't tell them about moving exiting tasks or
3152 * kernel threads (both mm NULL), since they never
3155 if (p->mm && printk_ratelimit()) {
3156 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3157 task_pid_nr(p), p->comm, cpu);
3165 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3168 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3170 lockdep_assert_held(&p->pi_lock);
3172 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3173 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3175 cpu = cpumask_any(p->cpus_ptr);
3178 * In order not to call set_task_cpu() on a blocking task we need
3179 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3182 * Since this is common to all placement strategies, this lives here.
3184 * [ this allows ->select_task() to simply return task_cpu(p) and
3185 * not worry about this generic constraint ]
3187 if (unlikely(!is_cpu_allowed(p, cpu)))
3188 cpu = select_fallback_rq(task_cpu(p), p);
3193 void sched_set_stop_task(int cpu, struct task_struct *stop)
3195 static struct lock_class_key stop_pi_lock;
3196 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3197 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3201 * Make it appear like a SCHED_FIFO task, its something
3202 * userspace knows about and won't get confused about.
3204 * Also, it will make PI more or less work without too
3205 * much confusion -- but then, stop work should not
3206 * rely on PI working anyway.
3208 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3210 stop->sched_class = &stop_sched_class;
3213 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3214 * adjust the effective priority of a task. As a result,
3215 * rt_mutex_setprio() can trigger (RT) balancing operations,
3216 * which can then trigger wakeups of the stop thread to push
3217 * around the current task.
3219 * The stop task itself will never be part of the PI-chain, it
3220 * never blocks, therefore that ->pi_lock recursion is safe.
3221 * Tell lockdep about this by placing the stop->pi_lock in its
3224 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3227 cpu_rq(cpu)->stop = stop;
3231 * Reset it back to a normal scheduling class so that
3232 * it can die in pieces.
3234 old_stop->sched_class = &rt_sched_class;
3238 #else /* CONFIG_SMP */
3240 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3241 const struct cpumask *new_mask,
3244 return set_cpus_allowed_ptr(p, new_mask);
3247 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3249 static inline bool rq_has_pinned_tasks(struct rq *rq)
3254 #endif /* !CONFIG_SMP */
3257 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3261 if (!schedstat_enabled())
3267 if (cpu == rq->cpu) {
3268 __schedstat_inc(rq->ttwu_local);
3269 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3271 struct sched_domain *sd;
3273 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3275 for_each_domain(rq->cpu, sd) {
3276 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3277 __schedstat_inc(sd->ttwu_wake_remote);
3284 if (wake_flags & WF_MIGRATED)
3285 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3286 #endif /* CONFIG_SMP */
3288 __schedstat_inc(rq->ttwu_count);
3289 __schedstat_inc(p->se.statistics.nr_wakeups);
3291 if (wake_flags & WF_SYNC)
3292 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3296 * Mark the task runnable and perform wakeup-preemption.
3298 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3299 struct rq_flags *rf)
3301 check_preempt_curr(rq, p, wake_flags);
3302 p->state = TASK_RUNNING;
3303 trace_sched_wakeup(p);
3306 if (p->sched_class->task_woken) {
3308 * Our task @p is fully woken up and running; so it's safe to
3309 * drop the rq->lock, hereafter rq is only used for statistics.
3311 rq_unpin_lock(rq, rf);
3312 p->sched_class->task_woken(rq, p);
3313 rq_repin_lock(rq, rf);
3316 if (rq->idle_stamp) {
3317 u64 delta = rq_clock(rq) - rq->idle_stamp;
3318 u64 max = 2*rq->max_idle_balance_cost;
3320 update_avg(&rq->avg_idle, delta);
3322 if (rq->avg_idle > max)
3331 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3332 struct rq_flags *rf)
3334 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3336 lockdep_assert_rq_held(rq);
3338 if (p->sched_contributes_to_load)
3339 rq->nr_uninterruptible--;
3342 if (wake_flags & WF_MIGRATED)
3343 en_flags |= ENQUEUE_MIGRATED;
3347 delayacct_blkio_end(p);
3348 atomic_dec(&task_rq(p)->nr_iowait);
3351 activate_task(rq, p, en_flags);
3352 ttwu_do_wakeup(rq, p, wake_flags, rf);
3356 * Consider @p being inside a wait loop:
3359 * set_current_state(TASK_UNINTERRUPTIBLE);
3366 * __set_current_state(TASK_RUNNING);
3368 * between set_current_state() and schedule(). In this case @p is still
3369 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3372 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3373 * then schedule() must still happen and p->state can be changed to
3374 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3375 * need to do a full wakeup with enqueue.
3377 * Returns: %true when the wakeup is done,
3380 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3386 rq = __task_rq_lock(p, &rf);
3387 if (task_on_rq_queued(p)) {
3388 /* check_preempt_curr() may use rq clock */
3389 update_rq_clock(rq);
3390 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3393 __task_rq_unlock(rq, &rf);
3399 void sched_ttwu_pending(void *arg)
3401 struct llist_node *llist = arg;
3402 struct rq *rq = this_rq();
3403 struct task_struct *p, *t;
3410 * rq::ttwu_pending racy indication of out-standing wakeups.
3411 * Races such that false-negatives are possible, since they
3412 * are shorter lived that false-positives would be.
3414 WRITE_ONCE(rq->ttwu_pending, 0);
3416 rq_lock_irqsave(rq, &rf);
3417 update_rq_clock(rq);
3419 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3420 if (WARN_ON_ONCE(p->on_cpu))
3421 smp_cond_load_acquire(&p->on_cpu, !VAL);
3423 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3424 set_task_cpu(p, cpu_of(rq));
3426 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3429 rq_unlock_irqrestore(rq, &rf);
3432 void send_call_function_single_ipi(int cpu)
3434 struct rq *rq = cpu_rq(cpu);
3436 if (!set_nr_if_polling(rq->idle))
3437 arch_send_call_function_single_ipi(cpu);
3439 trace_sched_wake_idle_without_ipi(cpu);
3443 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3444 * necessary. The wakee CPU on receipt of the IPI will queue the task
3445 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3446 * of the wakeup instead of the waker.
3448 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3450 struct rq *rq = cpu_rq(cpu);
3452 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3454 WRITE_ONCE(rq->ttwu_pending, 1);
3455 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3458 void wake_up_if_idle(int cpu)
3460 struct rq *rq = cpu_rq(cpu);
3465 if (!is_idle_task(rcu_dereference(rq->curr)))
3468 if (set_nr_if_polling(rq->idle)) {
3469 trace_sched_wake_idle_without_ipi(cpu);
3471 rq_lock_irqsave(rq, &rf);
3472 if (is_idle_task(rq->curr))
3473 smp_send_reschedule(cpu);
3474 /* Else CPU is not idle, do nothing here: */
3475 rq_unlock_irqrestore(rq, &rf);
3482 bool cpus_share_cache(int this_cpu, int that_cpu)
3484 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3487 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3490 * Do not complicate things with the async wake_list while the CPU is
3493 if (!cpu_active(cpu))
3497 * If the CPU does not share cache, then queue the task on the
3498 * remote rqs wakelist to avoid accessing remote data.
3500 if (!cpus_share_cache(smp_processor_id(), cpu))
3504 * If the task is descheduling and the only running task on the
3505 * CPU then use the wakelist to offload the task activation to
3506 * the soon-to-be-idle CPU as the current CPU is likely busy.
3507 * nr_running is checked to avoid unnecessary task stacking.
3509 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3515 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3517 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3518 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3521 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3522 __ttwu_queue_wakelist(p, cpu, wake_flags);
3529 #else /* !CONFIG_SMP */
3531 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3536 #endif /* CONFIG_SMP */
3538 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3540 struct rq *rq = cpu_rq(cpu);
3543 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3547 update_rq_clock(rq);
3548 ttwu_do_activate(rq, p, wake_flags, &rf);
3553 * Notes on Program-Order guarantees on SMP systems.
3557 * The basic program-order guarantee on SMP systems is that when a task [t]
3558 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3559 * execution on its new CPU [c1].
3561 * For migration (of runnable tasks) this is provided by the following means:
3563 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3564 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3565 * rq(c1)->lock (if not at the same time, then in that order).
3566 * C) LOCK of the rq(c1)->lock scheduling in task
3568 * Release/acquire chaining guarantees that B happens after A and C after B.
3569 * Note: the CPU doing B need not be c0 or c1
3578 * UNLOCK rq(0)->lock
3580 * LOCK rq(0)->lock // orders against CPU0
3582 * UNLOCK rq(0)->lock
3586 * UNLOCK rq(1)->lock
3588 * LOCK rq(1)->lock // orders against CPU2
3591 * UNLOCK rq(1)->lock
3594 * BLOCKING -- aka. SLEEP + WAKEUP
3596 * For blocking we (obviously) need to provide the same guarantee as for
3597 * migration. However the means are completely different as there is no lock
3598 * chain to provide order. Instead we do:
3600 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3601 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3605 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3607 * LOCK rq(0)->lock LOCK X->pi_lock
3610 * smp_store_release(X->on_cpu, 0);
3612 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3618 * X->state = RUNNING
3619 * UNLOCK rq(2)->lock
3621 * LOCK rq(2)->lock // orders against CPU1
3624 * UNLOCK rq(2)->lock
3627 * UNLOCK rq(0)->lock
3630 * However, for wakeups there is a second guarantee we must provide, namely we
3631 * must ensure that CONDITION=1 done by the caller can not be reordered with
3632 * accesses to the task state; see try_to_wake_up() and set_current_state().
3636 * try_to_wake_up - wake up a thread
3637 * @p: the thread to be awakened
3638 * @state: the mask of task states that can be woken
3639 * @wake_flags: wake modifier flags (WF_*)
3641 * Conceptually does:
3643 * If (@state & @p->state) @p->state = TASK_RUNNING.
3645 * If the task was not queued/runnable, also place it back on a runqueue.
3647 * This function is atomic against schedule() which would dequeue the task.
3649 * It issues a full memory barrier before accessing @p->state, see the comment
3650 * with set_current_state().
3652 * Uses p->pi_lock to serialize against concurrent wake-ups.
3654 * Relies on p->pi_lock stabilizing:
3657 * - p->sched_task_group
3658 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3660 * Tries really hard to only take one task_rq(p)->lock for performance.
3661 * Takes rq->lock in:
3662 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3663 * - ttwu_queue() -- new rq, for enqueue of the task;
3664 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3666 * As a consequence we race really badly with just about everything. See the
3667 * many memory barriers and their comments for details.
3669 * Return: %true if @p->state changes (an actual wakeup was done),
3673 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3675 unsigned long flags;
3676 int cpu, success = 0;
3681 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3682 * == smp_processor_id()'. Together this means we can special
3683 * case the whole 'p->on_rq && ttwu_runnable()' case below
3684 * without taking any locks.
3687 * - we rely on Program-Order guarantees for all the ordering,
3688 * - we're serialized against set_special_state() by virtue of
3689 * it disabling IRQs (this allows not taking ->pi_lock).
3691 if (!(p->state & state))
3695 trace_sched_waking(p);
3696 p->state = TASK_RUNNING;
3697 trace_sched_wakeup(p);
3702 * If we are going to wake up a thread waiting for CONDITION we
3703 * need to ensure that CONDITION=1 done by the caller can not be
3704 * reordered with p->state check below. This pairs with smp_store_mb()
3705 * in set_current_state() that the waiting thread does.
3707 raw_spin_lock_irqsave(&p->pi_lock, flags);
3708 smp_mb__after_spinlock();
3709 if (!(p->state & state))
3712 trace_sched_waking(p);
3714 /* We're going to change ->state: */
3718 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3719 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3720 * in smp_cond_load_acquire() below.
3722 * sched_ttwu_pending() try_to_wake_up()
3723 * STORE p->on_rq = 1 LOAD p->state
3726 * __schedule() (switch to task 'p')
3727 * LOCK rq->lock smp_rmb();
3728 * smp_mb__after_spinlock();
3732 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3734 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3735 * __schedule(). See the comment for smp_mb__after_spinlock().
3737 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3740 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3745 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3746 * possible to, falsely, observe p->on_cpu == 0.
3748 * One must be running (->on_cpu == 1) in order to remove oneself
3749 * from the runqueue.
3751 * __schedule() (switch to task 'p') try_to_wake_up()
3752 * STORE p->on_cpu = 1 LOAD p->on_rq
3755 * __schedule() (put 'p' to sleep)
3756 * LOCK rq->lock smp_rmb();
3757 * smp_mb__after_spinlock();
3758 * STORE p->on_rq = 0 LOAD p->on_cpu
3760 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3761 * __schedule(). See the comment for smp_mb__after_spinlock().
3763 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3764 * schedule()'s deactivate_task() has 'happened' and p will no longer
3765 * care about it's own p->state. See the comment in __schedule().
3767 smp_acquire__after_ctrl_dep();
3770 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3771 * == 0), which means we need to do an enqueue, change p->state to
3772 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3773 * enqueue, such as ttwu_queue_wakelist().
3775 p->state = TASK_WAKING;
3778 * If the owning (remote) CPU is still in the middle of schedule() with
3779 * this task as prev, considering queueing p on the remote CPUs wake_list
3780 * which potentially sends an IPI instead of spinning on p->on_cpu to
3781 * let the waker make forward progress. This is safe because IRQs are
3782 * disabled and the IPI will deliver after on_cpu is cleared.
3784 * Ensure we load task_cpu(p) after p->on_cpu:
3786 * set_task_cpu(p, cpu);
3787 * STORE p->cpu = @cpu
3788 * __schedule() (switch to task 'p')
3790 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3791 * STORE p->on_cpu = 1 LOAD p->cpu
3793 * to ensure we observe the correct CPU on which the task is currently
3796 if (smp_load_acquire(&p->on_cpu) &&
3797 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3801 * If the owning (remote) CPU is still in the middle of schedule() with
3802 * this task as prev, wait until it's done referencing the task.
3804 * Pairs with the smp_store_release() in finish_task().
3806 * This ensures that tasks getting woken will be fully ordered against
3807 * their previous state and preserve Program Order.
3809 smp_cond_load_acquire(&p->on_cpu, !VAL);
3811 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3812 if (task_cpu(p) != cpu) {
3814 delayacct_blkio_end(p);
3815 atomic_dec(&task_rq(p)->nr_iowait);
3818 wake_flags |= WF_MIGRATED;
3819 psi_ttwu_dequeue(p);
3820 set_task_cpu(p, cpu);
3824 #endif /* CONFIG_SMP */
3826 ttwu_queue(p, cpu, wake_flags);
3828 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3831 ttwu_stat(p, task_cpu(p), wake_flags);
3838 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3839 * @p: Process for which the function is to be invoked, can be @current.
3840 * @func: Function to invoke.
3841 * @arg: Argument to function.
3843 * If the specified task can be quickly locked into a definite state
3844 * (either sleeping or on a given runqueue), arrange to keep it in that
3845 * state while invoking @func(@arg). This function can use ->on_rq and
3846 * task_curr() to work out what the state is, if required. Given that
3847 * @func can be invoked with a runqueue lock held, it had better be quite
3851 * @false if the task slipped out from under the locks.
3852 * @true if the task was locked onto a runqueue or is sleeping.
3853 * However, @func can override this by returning @false.
3855 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3861 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3863 rq = __task_rq_lock(p, &rf);
3864 if (task_rq(p) == rq)
3873 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3878 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
3883 * wake_up_process - Wake up a specific process
3884 * @p: The process to be woken up.
3886 * Attempt to wake up the nominated process and move it to the set of runnable
3889 * Return: 1 if the process was woken up, 0 if it was already running.
3891 * This function executes a full memory barrier before accessing the task state.
3893 int wake_up_process(struct task_struct *p)
3895 return try_to_wake_up(p, TASK_NORMAL, 0);
3897 EXPORT_SYMBOL(wake_up_process);
3899 int wake_up_state(struct task_struct *p, unsigned int state)
3901 return try_to_wake_up(p, state, 0);
3905 * Perform scheduler related setup for a newly forked process p.
3906 * p is forked by current.
3908 * __sched_fork() is basic setup used by init_idle() too:
3910 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3915 p->se.exec_start = 0;
3916 p->se.sum_exec_runtime = 0;
3917 p->se.prev_sum_exec_runtime = 0;
3918 p->se.nr_migrations = 0;
3920 INIT_LIST_HEAD(&p->se.group_node);
3922 #ifdef CONFIG_FAIR_GROUP_SCHED
3923 p->se.cfs_rq = NULL;
3926 #ifdef CONFIG_SCHEDSTATS
3927 /* Even if schedstat is disabled, there should not be garbage */
3928 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3931 RB_CLEAR_NODE(&p->dl.rb_node);
3932 init_dl_task_timer(&p->dl);
3933 init_dl_inactive_task_timer(&p->dl);
3934 __dl_clear_params(p);
3936 INIT_LIST_HEAD(&p->rt.run_list);
3938 p->rt.time_slice = sched_rr_timeslice;
3942 #ifdef CONFIG_PREEMPT_NOTIFIERS
3943 INIT_HLIST_HEAD(&p->preempt_notifiers);
3946 #ifdef CONFIG_COMPACTION
3947 p->capture_control = NULL;
3949 init_numa_balancing(clone_flags, p);
3951 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3952 p->migration_pending = NULL;
3956 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3958 #ifdef CONFIG_NUMA_BALANCING
3960 void set_numabalancing_state(bool enabled)
3963 static_branch_enable(&sched_numa_balancing);
3965 static_branch_disable(&sched_numa_balancing);
3968 #ifdef CONFIG_PROC_SYSCTL
3969 int sysctl_numa_balancing(struct ctl_table *table, int write,
3970 void *buffer, size_t *lenp, loff_t *ppos)
3974 int state = static_branch_likely(&sched_numa_balancing);
3976 if (write && !capable(CAP_SYS_ADMIN))
3981 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3985 set_numabalancing_state(state);
3991 #ifdef CONFIG_SCHEDSTATS
3993 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3994 static bool __initdata __sched_schedstats = false;
3996 static void set_schedstats(bool enabled)
3999 static_branch_enable(&sched_schedstats);
4001 static_branch_disable(&sched_schedstats);
4004 void force_schedstat_enabled(void)
4006 if (!schedstat_enabled()) {
4007 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4008 static_branch_enable(&sched_schedstats);
4012 static int __init setup_schedstats(char *str)
4019 * This code is called before jump labels have been set up, so we can't
4020 * change the static branch directly just yet. Instead set a temporary
4021 * variable so init_schedstats() can do it later.
4023 if (!strcmp(str, "enable")) {
4024 __sched_schedstats = true;
4026 } else if (!strcmp(str, "disable")) {
4027 __sched_schedstats = false;
4032 pr_warn("Unable to parse schedstats=\n");
4036 __setup("schedstats=", setup_schedstats);
4038 static void __init init_schedstats(void)
4040 set_schedstats(__sched_schedstats);
4043 #ifdef CONFIG_PROC_SYSCTL
4044 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4045 size_t *lenp, loff_t *ppos)
4049 int state = static_branch_likely(&sched_schedstats);
4051 if (write && !capable(CAP_SYS_ADMIN))
4056 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4060 set_schedstats(state);
4063 #endif /* CONFIG_PROC_SYSCTL */
4064 #else /* !CONFIG_SCHEDSTATS */
4065 static inline void init_schedstats(void) {}
4066 #endif /* CONFIG_SCHEDSTATS */
4069 * fork()/clone()-time setup:
4071 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4073 unsigned long flags;
4075 __sched_fork(clone_flags, p);
4077 * We mark the process as NEW here. This guarantees that
4078 * nobody will actually run it, and a signal or other external
4079 * event cannot wake it up and insert it on the runqueue either.
4081 p->state = TASK_NEW;
4084 * Make sure we do not leak PI boosting priority to the child.
4086 p->prio = current->normal_prio;
4091 * Revert to default priority/policy on fork if requested.
4093 if (unlikely(p->sched_reset_on_fork)) {
4094 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4095 p->policy = SCHED_NORMAL;
4096 p->static_prio = NICE_TO_PRIO(0);
4098 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4099 p->static_prio = NICE_TO_PRIO(0);
4101 p->prio = p->normal_prio = __normal_prio(p);
4102 set_load_weight(p, false);
4105 * We don't need the reset flag anymore after the fork. It has
4106 * fulfilled its duty:
4108 p->sched_reset_on_fork = 0;
4111 if (dl_prio(p->prio))
4113 else if (rt_prio(p->prio))
4114 p->sched_class = &rt_sched_class;
4116 p->sched_class = &fair_sched_class;
4118 init_entity_runnable_average(&p->se);
4121 * The child is not yet in the pid-hash so no cgroup attach races,
4122 * and the cgroup is pinned to this child due to cgroup_fork()
4123 * is ran before sched_fork().
4125 * Silence PROVE_RCU.
4127 raw_spin_lock_irqsave(&p->pi_lock, flags);
4130 * We're setting the CPU for the first time, we don't migrate,
4131 * so use __set_task_cpu().
4133 __set_task_cpu(p, smp_processor_id());
4134 if (p->sched_class->task_fork)
4135 p->sched_class->task_fork(p);
4136 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4138 #ifdef CONFIG_SCHED_INFO
4139 if (likely(sched_info_on()))
4140 memset(&p->sched_info, 0, sizeof(p->sched_info));
4142 #if defined(CONFIG_SMP)
4145 init_task_preempt_count(p);
4147 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4148 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4153 void sched_post_fork(struct task_struct *p)
4155 uclamp_post_fork(p);
4158 unsigned long to_ratio(u64 period, u64 runtime)
4160 if (runtime == RUNTIME_INF)
4164 * Doing this here saves a lot of checks in all
4165 * the calling paths, and returning zero seems
4166 * safe for them anyway.
4171 return div64_u64(runtime << BW_SHIFT, period);
4175 * wake_up_new_task - wake up a newly created task for the first time.
4177 * This function will do some initial scheduler statistics housekeeping
4178 * that must be done for every newly created context, then puts the task
4179 * on the runqueue and wakes it.
4181 void wake_up_new_task(struct task_struct *p)
4186 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4187 p->state = TASK_RUNNING;
4190 * Fork balancing, do it here and not earlier because:
4191 * - cpus_ptr can change in the fork path
4192 * - any previously selected CPU might disappear through hotplug
4194 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4195 * as we're not fully set-up yet.
4197 p->recent_used_cpu = task_cpu(p);
4199 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4201 rq = __task_rq_lock(p, &rf);
4202 update_rq_clock(rq);
4203 post_init_entity_util_avg(p);
4205 activate_task(rq, p, ENQUEUE_NOCLOCK);
4206 trace_sched_wakeup_new(p);
4207 check_preempt_curr(rq, p, WF_FORK);
4209 if (p->sched_class->task_woken) {
4211 * Nothing relies on rq->lock after this, so it's fine to
4214 rq_unpin_lock(rq, &rf);
4215 p->sched_class->task_woken(rq, p);
4216 rq_repin_lock(rq, &rf);
4219 task_rq_unlock(rq, p, &rf);
4222 #ifdef CONFIG_PREEMPT_NOTIFIERS
4224 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4226 void preempt_notifier_inc(void)
4228 static_branch_inc(&preempt_notifier_key);
4230 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4232 void preempt_notifier_dec(void)
4234 static_branch_dec(&preempt_notifier_key);
4236 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4239 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4240 * @notifier: notifier struct to register
4242 void preempt_notifier_register(struct preempt_notifier *notifier)
4244 if (!static_branch_unlikely(&preempt_notifier_key))
4245 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4247 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4249 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4252 * preempt_notifier_unregister - no longer interested in preemption notifications
4253 * @notifier: notifier struct to unregister
4255 * This is *not* safe to call from within a preemption notifier.
4257 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4259 hlist_del(¬ifier->link);
4261 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4263 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4265 struct preempt_notifier *notifier;
4267 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4268 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4271 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4273 if (static_branch_unlikely(&preempt_notifier_key))
4274 __fire_sched_in_preempt_notifiers(curr);
4278 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4279 struct task_struct *next)
4281 struct preempt_notifier *notifier;
4283 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4284 notifier->ops->sched_out(notifier, next);
4287 static __always_inline void
4288 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4289 struct task_struct *next)
4291 if (static_branch_unlikely(&preempt_notifier_key))
4292 __fire_sched_out_preempt_notifiers(curr, next);
4295 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4297 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4302 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4303 struct task_struct *next)
4307 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4309 static inline void prepare_task(struct task_struct *next)
4313 * Claim the task as running, we do this before switching to it
4314 * such that any running task will have this set.
4316 * See the ttwu() WF_ON_CPU case and its ordering comment.
4318 WRITE_ONCE(next->on_cpu, 1);
4322 static inline void finish_task(struct task_struct *prev)
4326 * This must be the very last reference to @prev from this CPU. After
4327 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4328 * must ensure this doesn't happen until the switch is completely
4331 * In particular, the load of prev->state in finish_task_switch() must
4332 * happen before this.
4334 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4336 smp_store_release(&prev->on_cpu, 0);
4342 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4344 void (*func)(struct rq *rq);
4345 struct callback_head *next;
4347 lockdep_assert_rq_held(rq);
4350 func = (void (*)(struct rq *))head->func;
4359 static void balance_push(struct rq *rq);
4361 struct callback_head balance_push_callback = {
4363 .func = (void (*)(struct callback_head *))balance_push,
4366 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4368 struct callback_head *head = rq->balance_callback;
4370 lockdep_assert_rq_held(rq);
4372 rq->balance_callback = NULL;
4377 static void __balance_callbacks(struct rq *rq)
4379 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4382 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4384 unsigned long flags;
4386 if (unlikely(head)) {
4387 raw_spin_rq_lock_irqsave(rq, flags);
4388 do_balance_callbacks(rq, head);
4389 raw_spin_rq_unlock_irqrestore(rq, flags);
4395 static inline void __balance_callbacks(struct rq *rq)
4399 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4404 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4411 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4414 * Since the runqueue lock will be released by the next
4415 * task (which is an invalid locking op but in the case
4416 * of the scheduler it's an obvious special-case), so we
4417 * do an early lockdep release here:
4419 rq_unpin_lock(rq, rf);
4420 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4421 #ifdef CONFIG_DEBUG_SPINLOCK
4422 /* this is a valid case when another task releases the spinlock */
4423 rq_lockp(rq)->owner = next;
4427 static inline void finish_lock_switch(struct rq *rq)
4430 * If we are tracking spinlock dependencies then we have to
4431 * fix up the runqueue lock - which gets 'carried over' from
4432 * prev into current:
4434 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4435 __balance_callbacks(rq);
4436 raw_spin_rq_unlock_irq(rq);
4440 * NOP if the arch has not defined these:
4443 #ifndef prepare_arch_switch
4444 # define prepare_arch_switch(next) do { } while (0)
4447 #ifndef finish_arch_post_lock_switch
4448 # define finish_arch_post_lock_switch() do { } while (0)
4451 static inline void kmap_local_sched_out(void)
4453 #ifdef CONFIG_KMAP_LOCAL
4454 if (unlikely(current->kmap_ctrl.idx))
4455 __kmap_local_sched_out();
4459 static inline void kmap_local_sched_in(void)
4461 #ifdef CONFIG_KMAP_LOCAL
4462 if (unlikely(current->kmap_ctrl.idx))
4463 __kmap_local_sched_in();
4468 * prepare_task_switch - prepare to switch tasks
4469 * @rq: the runqueue preparing to switch
4470 * @prev: the current task that is being switched out
4471 * @next: the task we are going to switch to.
4473 * This is called with the rq lock held and interrupts off. It must
4474 * be paired with a subsequent finish_task_switch after the context
4477 * prepare_task_switch sets up locking and calls architecture specific
4481 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4482 struct task_struct *next)
4484 kcov_prepare_switch(prev);
4485 sched_info_switch(rq, prev, next);
4486 perf_event_task_sched_out(prev, next);
4488 fire_sched_out_preempt_notifiers(prev, next);
4489 kmap_local_sched_out();
4491 prepare_arch_switch(next);
4495 * finish_task_switch - clean up after a task-switch
4496 * @prev: the thread we just switched away from.
4498 * finish_task_switch must be called after the context switch, paired
4499 * with a prepare_task_switch call before the context switch.
4500 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4501 * and do any other architecture-specific cleanup actions.
4503 * Note that we may have delayed dropping an mm in context_switch(). If
4504 * so, we finish that here outside of the runqueue lock. (Doing it
4505 * with the lock held can cause deadlocks; see schedule() for
4508 * The context switch have flipped the stack from under us and restored the
4509 * local variables which were saved when this task called schedule() in the
4510 * past. prev == current is still correct but we need to recalculate this_rq
4511 * because prev may have moved to another CPU.
4513 static struct rq *finish_task_switch(struct task_struct *prev)
4514 __releases(rq->lock)
4516 struct rq *rq = this_rq();
4517 struct mm_struct *mm = rq->prev_mm;
4521 * The previous task will have left us with a preempt_count of 2
4522 * because it left us after:
4525 * preempt_disable(); // 1
4527 * raw_spin_lock_irq(&rq->lock) // 2
4529 * Also, see FORK_PREEMPT_COUNT.
4531 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4532 "corrupted preempt_count: %s/%d/0x%x\n",
4533 current->comm, current->pid, preempt_count()))
4534 preempt_count_set(FORK_PREEMPT_COUNT);
4539 * A task struct has one reference for the use as "current".
4540 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4541 * schedule one last time. The schedule call will never return, and
4542 * the scheduled task must drop that reference.
4544 * We must observe prev->state before clearing prev->on_cpu (in
4545 * finish_task), otherwise a concurrent wakeup can get prev
4546 * running on another CPU and we could rave with its RUNNING -> DEAD
4547 * transition, resulting in a double drop.
4549 prev_state = prev->state;
4550 vtime_task_switch(prev);
4551 perf_event_task_sched_in(prev, current);
4553 finish_lock_switch(rq);
4554 finish_arch_post_lock_switch();
4555 kcov_finish_switch(current);
4557 * kmap_local_sched_out() is invoked with rq::lock held and
4558 * interrupts disabled. There is no requirement for that, but the
4559 * sched out code does not have an interrupt enabled section.
4560 * Restoring the maps on sched in does not require interrupts being
4563 kmap_local_sched_in();
4565 fire_sched_in_preempt_notifiers(current);
4567 * When switching through a kernel thread, the loop in
4568 * membarrier_{private,global}_expedited() may have observed that
4569 * kernel thread and not issued an IPI. It is therefore possible to
4570 * schedule between user->kernel->user threads without passing though
4571 * switch_mm(). Membarrier requires a barrier after storing to
4572 * rq->curr, before returning to userspace, so provide them here:
4574 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4575 * provided by mmdrop(),
4576 * - a sync_core for SYNC_CORE.
4579 membarrier_mm_sync_core_before_usermode(mm);
4582 if (unlikely(prev_state == TASK_DEAD)) {
4583 if (prev->sched_class->task_dead)
4584 prev->sched_class->task_dead(prev);
4587 * Remove function-return probe instances associated with this
4588 * task and put them back on the free list.
4590 kprobe_flush_task(prev);
4592 /* Task is done with its stack. */
4593 put_task_stack(prev);
4595 put_task_struct_rcu_user(prev);
4598 tick_nohz_task_switch();
4603 * schedule_tail - first thing a freshly forked thread must call.
4604 * @prev: the thread we just switched away from.
4606 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4607 __releases(rq->lock)
4610 * New tasks start with FORK_PREEMPT_COUNT, see there and
4611 * finish_task_switch() for details.
4613 * finish_task_switch() will drop rq->lock() and lower preempt_count
4614 * and the preempt_enable() will end up enabling preemption (on
4615 * PREEMPT_COUNT kernels).
4618 finish_task_switch(prev);
4621 if (current->set_child_tid)
4622 put_user(task_pid_vnr(current), current->set_child_tid);
4624 calculate_sigpending();
4628 * context_switch - switch to the new MM and the new thread's register state.
4630 static __always_inline struct rq *
4631 context_switch(struct rq *rq, struct task_struct *prev,
4632 struct task_struct *next, struct rq_flags *rf)
4634 prepare_task_switch(rq, prev, next);
4637 * For paravirt, this is coupled with an exit in switch_to to
4638 * combine the page table reload and the switch backend into
4641 arch_start_context_switch(prev);
4644 * kernel -> kernel lazy + transfer active
4645 * user -> kernel lazy + mmgrab() active
4647 * kernel -> user switch + mmdrop() active
4648 * user -> user switch
4650 if (!next->mm) { // to kernel
4651 enter_lazy_tlb(prev->active_mm, next);
4653 next->active_mm = prev->active_mm;
4654 if (prev->mm) // from user
4655 mmgrab(prev->active_mm);
4657 prev->active_mm = NULL;
4659 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4661 * sys_membarrier() requires an smp_mb() between setting
4662 * rq->curr / membarrier_switch_mm() and returning to userspace.
4664 * The below provides this either through switch_mm(), or in
4665 * case 'prev->active_mm == next->mm' through
4666 * finish_task_switch()'s mmdrop().
4668 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4670 if (!prev->mm) { // from kernel
4671 /* will mmdrop() in finish_task_switch(). */
4672 rq->prev_mm = prev->active_mm;
4673 prev->active_mm = NULL;
4677 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4679 prepare_lock_switch(rq, next, rf);
4681 /* Here we just switch the register state and the stack. */
4682 switch_to(prev, next, prev);
4685 return finish_task_switch(prev);
4689 * nr_running and nr_context_switches:
4691 * externally visible scheduler statistics: current number of runnable
4692 * threads, total number of context switches performed since bootup.
4694 unsigned long nr_running(void)
4696 unsigned long i, sum = 0;
4698 for_each_online_cpu(i)
4699 sum += cpu_rq(i)->nr_running;
4705 * Check if only the current task is running on the CPU.
4707 * Caution: this function does not check that the caller has disabled
4708 * preemption, thus the result might have a time-of-check-to-time-of-use
4709 * race. The caller is responsible to use it correctly, for example:
4711 * - from a non-preemptible section (of course)
4713 * - from a thread that is bound to a single CPU
4715 * - in a loop with very short iterations (e.g. a polling loop)
4717 bool single_task_running(void)
4719 return raw_rq()->nr_running == 1;
4721 EXPORT_SYMBOL(single_task_running);
4723 unsigned long long nr_context_switches(void)
4726 unsigned long long sum = 0;
4728 for_each_possible_cpu(i)
4729 sum += cpu_rq(i)->nr_switches;
4735 * Consumers of these two interfaces, like for example the cpuidle menu
4736 * governor, are using nonsensical data. Preferring shallow idle state selection
4737 * for a CPU that has IO-wait which might not even end up running the task when
4738 * it does become runnable.
4741 unsigned long nr_iowait_cpu(int cpu)
4743 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4747 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4749 * The idea behind IO-wait account is to account the idle time that we could
4750 * have spend running if it were not for IO. That is, if we were to improve the
4751 * storage performance, we'd have a proportional reduction in IO-wait time.
4753 * This all works nicely on UP, where, when a task blocks on IO, we account
4754 * idle time as IO-wait, because if the storage were faster, it could've been
4755 * running and we'd not be idle.
4757 * This has been extended to SMP, by doing the same for each CPU. This however
4760 * Imagine for instance the case where two tasks block on one CPU, only the one
4761 * CPU will have IO-wait accounted, while the other has regular idle. Even
4762 * though, if the storage were faster, both could've ran at the same time,
4763 * utilising both CPUs.
4765 * This means, that when looking globally, the current IO-wait accounting on
4766 * SMP is a lower bound, by reason of under accounting.
4768 * Worse, since the numbers are provided per CPU, they are sometimes
4769 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4770 * associated with any one particular CPU, it can wake to another CPU than it
4771 * blocked on. This means the per CPU IO-wait number is meaningless.
4773 * Task CPU affinities can make all that even more 'interesting'.
4776 unsigned long nr_iowait(void)
4778 unsigned long i, sum = 0;
4780 for_each_possible_cpu(i)
4781 sum += nr_iowait_cpu(i);
4789 * sched_exec - execve() is a valuable balancing opportunity, because at
4790 * this point the task has the smallest effective memory and cache footprint.
4792 void sched_exec(void)
4794 struct task_struct *p = current;
4795 unsigned long flags;
4798 raw_spin_lock_irqsave(&p->pi_lock, flags);
4799 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4800 if (dest_cpu == smp_processor_id())
4803 if (likely(cpu_active(dest_cpu))) {
4804 struct migration_arg arg = { p, dest_cpu };
4806 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4807 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4811 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4816 DEFINE_PER_CPU(struct kernel_stat, kstat);
4817 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4819 EXPORT_PER_CPU_SYMBOL(kstat);
4820 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4823 * The function fair_sched_class.update_curr accesses the struct curr
4824 * and its field curr->exec_start; when called from task_sched_runtime(),
4825 * we observe a high rate of cache misses in practice.
4826 * Prefetching this data results in improved performance.
4828 static inline void prefetch_curr_exec_start(struct task_struct *p)
4830 #ifdef CONFIG_FAIR_GROUP_SCHED
4831 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4833 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4836 prefetch(&curr->exec_start);
4840 * Return accounted runtime for the task.
4841 * In case the task is currently running, return the runtime plus current's
4842 * pending runtime that have not been accounted yet.
4844 unsigned long long task_sched_runtime(struct task_struct *p)
4850 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4852 * 64-bit doesn't need locks to atomically read a 64-bit value.
4853 * So we have a optimization chance when the task's delta_exec is 0.
4854 * Reading ->on_cpu is racy, but this is ok.
4856 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4857 * If we race with it entering CPU, unaccounted time is 0. This is
4858 * indistinguishable from the read occurring a few cycles earlier.
4859 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4860 * been accounted, so we're correct here as well.
4862 if (!p->on_cpu || !task_on_rq_queued(p))
4863 return p->se.sum_exec_runtime;
4866 rq = task_rq_lock(p, &rf);
4868 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4869 * project cycles that may never be accounted to this
4870 * thread, breaking clock_gettime().
4872 if (task_current(rq, p) && task_on_rq_queued(p)) {
4873 prefetch_curr_exec_start(p);
4874 update_rq_clock(rq);
4875 p->sched_class->update_curr(rq);
4877 ns = p->se.sum_exec_runtime;
4878 task_rq_unlock(rq, p, &rf);
4883 #ifdef CONFIG_SCHED_DEBUG
4884 static u64 cpu_resched_latency(struct rq *rq)
4886 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
4887 u64 resched_latency, now = rq_clock(rq);
4888 static bool warned_once;
4890 if (sysctl_resched_latency_warn_once && warned_once)
4893 if (!need_resched() || !latency_warn_ms)
4896 if (system_state == SYSTEM_BOOTING)
4899 if (!rq->last_seen_need_resched_ns) {
4900 rq->last_seen_need_resched_ns = now;
4901 rq->ticks_without_resched = 0;
4905 rq->ticks_without_resched++;
4906 resched_latency = now - rq->last_seen_need_resched_ns;
4907 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
4912 return resched_latency;
4915 static int __init setup_resched_latency_warn_ms(char *str)
4919 if ((kstrtol(str, 0, &val))) {
4920 pr_warn("Unable to set resched_latency_warn_ms\n");
4924 sysctl_resched_latency_warn_ms = val;
4927 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
4929 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
4930 #endif /* CONFIG_SCHED_DEBUG */
4933 * This function gets called by the timer code, with HZ frequency.
4934 * We call it with interrupts disabled.
4936 void scheduler_tick(void)
4938 int cpu = smp_processor_id();
4939 struct rq *rq = cpu_rq(cpu);
4940 struct task_struct *curr = rq->curr;
4942 unsigned long thermal_pressure;
4943 u64 resched_latency;
4945 arch_scale_freq_tick();
4950 update_rq_clock(rq);
4951 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4952 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4953 curr->sched_class->task_tick(rq, curr, 0);
4954 if (sched_feat(LATENCY_WARN))
4955 resched_latency = cpu_resched_latency(rq);
4956 calc_global_load_tick(rq);
4960 if (sched_feat(LATENCY_WARN) && resched_latency)
4961 resched_latency_warn(cpu, resched_latency);
4963 perf_event_task_tick();
4966 rq->idle_balance = idle_cpu(cpu);
4967 trigger_load_balance(rq);
4971 #ifdef CONFIG_NO_HZ_FULL
4976 struct delayed_work work;
4978 /* Values for ->state, see diagram below. */
4979 #define TICK_SCHED_REMOTE_OFFLINE 0
4980 #define TICK_SCHED_REMOTE_OFFLINING 1
4981 #define TICK_SCHED_REMOTE_RUNNING 2
4984 * State diagram for ->state:
4987 * TICK_SCHED_REMOTE_OFFLINE
4990 * | | sched_tick_remote()
4993 * +--TICK_SCHED_REMOTE_OFFLINING
4996 * sched_tick_start() | | sched_tick_stop()
4999 * TICK_SCHED_REMOTE_RUNNING
5002 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5003 * and sched_tick_start() are happy to leave the state in RUNNING.
5006 static struct tick_work __percpu *tick_work_cpu;
5008 static void sched_tick_remote(struct work_struct *work)
5010 struct delayed_work *dwork = to_delayed_work(work);
5011 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5012 int cpu = twork->cpu;
5013 struct rq *rq = cpu_rq(cpu);
5014 struct task_struct *curr;
5020 * Handle the tick only if it appears the remote CPU is running in full
5021 * dynticks mode. The check is racy by nature, but missing a tick or
5022 * having one too much is no big deal because the scheduler tick updates
5023 * statistics and checks timeslices in a time-independent way, regardless
5024 * of when exactly it is running.
5026 if (!tick_nohz_tick_stopped_cpu(cpu))
5029 rq_lock_irq(rq, &rf);
5031 if (cpu_is_offline(cpu))
5034 update_rq_clock(rq);
5036 if (!is_idle_task(curr)) {
5038 * Make sure the next tick runs within a reasonable
5041 delta = rq_clock_task(rq) - curr->se.exec_start;
5042 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5044 curr->sched_class->task_tick(rq, curr, 0);
5046 calc_load_nohz_remote(rq);
5048 rq_unlock_irq(rq, &rf);
5052 * Run the remote tick once per second (1Hz). This arbitrary
5053 * frequency is large enough to avoid overload but short enough
5054 * to keep scheduler internal stats reasonably up to date. But
5055 * first update state to reflect hotplug activity if required.
5057 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5058 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5059 if (os == TICK_SCHED_REMOTE_RUNNING)
5060 queue_delayed_work(system_unbound_wq, dwork, HZ);
5063 static void sched_tick_start(int cpu)
5066 struct tick_work *twork;
5068 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5071 WARN_ON_ONCE(!tick_work_cpu);
5073 twork = per_cpu_ptr(tick_work_cpu, cpu);
5074 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5075 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5076 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5078 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5079 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5083 #ifdef CONFIG_HOTPLUG_CPU
5084 static void sched_tick_stop(int cpu)
5086 struct tick_work *twork;
5089 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5092 WARN_ON_ONCE(!tick_work_cpu);
5094 twork = per_cpu_ptr(tick_work_cpu, cpu);
5095 /* There cannot be competing actions, but don't rely on stop-machine. */
5096 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5097 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5098 /* Don't cancel, as this would mess up the state machine. */
5100 #endif /* CONFIG_HOTPLUG_CPU */
5102 int __init sched_tick_offload_init(void)
5104 tick_work_cpu = alloc_percpu(struct tick_work);
5105 BUG_ON(!tick_work_cpu);
5109 #else /* !CONFIG_NO_HZ_FULL */
5110 static inline void sched_tick_start(int cpu) { }
5111 static inline void sched_tick_stop(int cpu) { }
5114 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5115 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5117 * If the value passed in is equal to the current preempt count
5118 * then we just disabled preemption. Start timing the latency.
5120 static inline void preempt_latency_start(int val)
5122 if (preempt_count() == val) {
5123 unsigned long ip = get_lock_parent_ip();
5124 #ifdef CONFIG_DEBUG_PREEMPT
5125 current->preempt_disable_ip = ip;
5127 trace_preempt_off(CALLER_ADDR0, ip);
5131 void preempt_count_add(int val)
5133 #ifdef CONFIG_DEBUG_PREEMPT
5137 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5140 __preempt_count_add(val);
5141 #ifdef CONFIG_DEBUG_PREEMPT
5143 * Spinlock count overflowing soon?
5145 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5148 preempt_latency_start(val);
5150 EXPORT_SYMBOL(preempt_count_add);
5151 NOKPROBE_SYMBOL(preempt_count_add);
5154 * If the value passed in equals to the current preempt count
5155 * then we just enabled preemption. Stop timing the latency.
5157 static inline void preempt_latency_stop(int val)
5159 if (preempt_count() == val)
5160 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5163 void preempt_count_sub(int val)
5165 #ifdef CONFIG_DEBUG_PREEMPT
5169 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5172 * Is the spinlock portion underflowing?
5174 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5175 !(preempt_count() & PREEMPT_MASK)))
5179 preempt_latency_stop(val);
5180 __preempt_count_sub(val);
5182 EXPORT_SYMBOL(preempt_count_sub);
5183 NOKPROBE_SYMBOL(preempt_count_sub);
5186 static inline void preempt_latency_start(int val) { }
5187 static inline void preempt_latency_stop(int val) { }
5190 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5192 #ifdef CONFIG_DEBUG_PREEMPT
5193 return p->preempt_disable_ip;
5200 * Print scheduling while atomic bug:
5202 static noinline void __schedule_bug(struct task_struct *prev)
5204 /* Save this before calling printk(), since that will clobber it */
5205 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5207 if (oops_in_progress)
5210 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5211 prev->comm, prev->pid, preempt_count());
5213 debug_show_held_locks(prev);
5215 if (irqs_disabled())
5216 print_irqtrace_events(prev);
5217 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5218 && in_atomic_preempt_off()) {
5219 pr_err("Preemption disabled at:");
5220 print_ip_sym(KERN_ERR, preempt_disable_ip);
5223 panic("scheduling while atomic\n");
5226 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5230 * Various schedule()-time debugging checks and statistics:
5232 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5234 #ifdef CONFIG_SCHED_STACK_END_CHECK
5235 if (task_stack_end_corrupted(prev))
5236 panic("corrupted stack end detected inside scheduler\n");
5238 if (task_scs_end_corrupted(prev))
5239 panic("corrupted shadow stack detected inside scheduler\n");
5242 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5243 if (!preempt && prev->state && prev->non_block_count) {
5244 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5245 prev->comm, prev->pid, prev->non_block_count);
5247 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5251 if (unlikely(in_atomic_preempt_off())) {
5252 __schedule_bug(prev);
5253 preempt_count_set(PREEMPT_DISABLED);
5256 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5258 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5260 schedstat_inc(this_rq()->sched_count);
5263 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5264 struct rq_flags *rf)
5267 const struct sched_class *class;
5269 * We must do the balancing pass before put_prev_task(), such
5270 * that when we release the rq->lock the task is in the same
5271 * state as before we took rq->lock.
5273 * We can terminate the balance pass as soon as we know there is
5274 * a runnable task of @class priority or higher.
5276 for_class_range(class, prev->sched_class, &idle_sched_class) {
5277 if (class->balance(rq, prev, rf))
5282 put_prev_task(rq, prev);
5286 * Pick up the highest-prio task:
5288 static inline struct task_struct *
5289 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5291 const struct sched_class *class;
5292 struct task_struct *p;
5295 * Optimization: we know that if all tasks are in the fair class we can
5296 * call that function directly, but only if the @prev task wasn't of a
5297 * higher scheduling class, because otherwise those lose the
5298 * opportunity to pull in more work from other CPUs.
5300 if (likely(prev->sched_class <= &fair_sched_class &&
5301 rq->nr_running == rq->cfs.h_nr_running)) {
5303 p = pick_next_task_fair(rq, prev, rf);
5304 if (unlikely(p == RETRY_TASK))
5307 /* Assumes fair_sched_class->next == idle_sched_class */
5309 put_prev_task(rq, prev);
5310 p = pick_next_task_idle(rq);
5317 put_prev_task_balance(rq, prev, rf);
5319 for_each_class(class) {
5320 p = class->pick_next_task(rq);
5325 /* The idle class should always have a runnable task: */
5329 #ifdef CONFIG_SCHED_CORE
5330 static inline bool is_task_rq_idle(struct task_struct *t)
5332 return (task_rq(t)->idle == t);
5335 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5337 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5340 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5342 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5345 return a->core_cookie == b->core_cookie;
5348 // XXX fairness/fwd progress conditions
5351 * - NULL if there is no runnable task for this class.
5352 * - the highest priority task for this runqueue if it matches
5353 * rq->core->core_cookie or its priority is greater than max.
5354 * - Else returns idle_task.
5356 static struct task_struct *
5357 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5359 struct task_struct *class_pick, *cookie_pick;
5360 unsigned long cookie = rq->core->core_cookie;
5362 class_pick = class->pick_task(rq);
5368 * If class_pick is tagged, return it only if it has
5369 * higher priority than max.
5371 if (max && class_pick->core_cookie &&
5372 prio_less(class_pick, max, in_fi))
5373 return idle_sched_class.pick_task(rq);
5379 * If class_pick is idle or matches cookie, return early.
5381 if (cookie_equals(class_pick, cookie))
5384 cookie_pick = sched_core_find(rq, cookie);
5387 * If class > max && class > cookie, it is the highest priority task on
5388 * the core (so far) and it must be selected, otherwise we must go with
5389 * the cookie pick in order to satisfy the constraint.
5391 if (prio_less(cookie_pick, class_pick, in_fi) &&
5392 (!max || prio_less(max, class_pick, in_fi)))
5398 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5400 static struct task_struct *
5401 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5403 struct task_struct *next, *max = NULL;
5404 const struct sched_class *class;
5405 const struct cpumask *smt_mask;
5406 bool fi_before = false;
5407 int i, j, cpu, occ = 0;
5410 if (!sched_core_enabled(rq))
5411 return __pick_next_task(rq, prev, rf);
5415 /* Stopper task is switching into idle, no need core-wide selection. */
5416 if (cpu_is_offline(cpu)) {
5418 * Reset core_pick so that we don't enter the fastpath when
5419 * coming online. core_pick would already be migrated to
5420 * another cpu during offline.
5422 rq->core_pick = NULL;
5423 return __pick_next_task(rq, prev, rf);
5427 * If there were no {en,de}queues since we picked (IOW, the task
5428 * pointers are all still valid), and we haven't scheduled the last
5429 * pick yet, do so now.
5431 * rq->core_pick can be NULL if no selection was made for a CPU because
5432 * it was either offline or went offline during a sibling's core-wide
5433 * selection. In this case, do a core-wide selection.
5435 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5436 rq->core->core_pick_seq != rq->core_sched_seq &&
5438 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5440 next = rq->core_pick;
5442 put_prev_task(rq, prev);
5443 set_next_task(rq, next);
5446 rq->core_pick = NULL;
5450 put_prev_task_balance(rq, prev, rf);
5452 smt_mask = cpu_smt_mask(cpu);
5453 need_sync = !!rq->core->core_cookie;
5456 rq->core->core_cookie = 0UL;
5457 if (rq->core->core_forceidle) {
5460 rq->core->core_forceidle = false;
5464 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5466 * @task_seq guards the task state ({en,de}queues)
5467 * @pick_seq is the @task_seq we did a selection on
5468 * @sched_seq is the @pick_seq we scheduled
5470 * However, preemptions can cause multiple picks on the same task set.
5471 * 'Fix' this by also increasing @task_seq for every pick.
5473 rq->core->core_task_seq++;
5476 * Optimize for common case where this CPU has no cookies
5477 * and there are no cookied tasks running on siblings.
5480 for_each_class(class) {
5481 next = class->pick_task(rq);
5486 if (!next->core_cookie) {
5487 rq->core_pick = NULL;
5489 * For robustness, update the min_vruntime_fi for
5490 * unconstrained picks as well.
5492 WARN_ON_ONCE(fi_before);
5493 task_vruntime_update(rq, next, false);
5498 for_each_cpu(i, smt_mask) {
5499 struct rq *rq_i = cpu_rq(i);
5501 rq_i->core_pick = NULL;
5504 update_rq_clock(rq_i);
5508 * Try and select tasks for each sibling in decending sched_class
5511 for_each_class(class) {
5513 for_each_cpu_wrap(i, smt_mask, cpu) {
5514 struct rq *rq_i = cpu_rq(i);
5515 struct task_struct *p;
5517 if (rq_i->core_pick)
5521 * If this sibling doesn't yet have a suitable task to
5522 * run; ask for the most elegible task, given the
5523 * highest priority task already selected for this
5526 p = pick_task(rq_i, class, max, fi_before);
5530 if (!is_task_rq_idle(p))
5533 rq_i->core_pick = p;
5534 if (rq_i->idle == p && rq_i->nr_running) {
5535 rq->core->core_forceidle = true;
5537 rq->core->core_forceidle_seq++;
5541 * If this new candidate is of higher priority than the
5542 * previous; and they're incompatible; we need to wipe
5543 * the slate and start over. pick_task makes sure that
5544 * p's priority is more than max if it doesn't match
5547 * NOTE: this is a linear max-filter and is thus bounded
5548 * in execution time.
5550 if (!max || !cookie_match(max, p)) {
5551 struct task_struct *old_max = max;
5553 rq->core->core_cookie = p->core_cookie;
5557 rq->core->core_forceidle = false;
5558 for_each_cpu(j, smt_mask) {
5562 cpu_rq(j)->core_pick = NULL;
5571 rq->core->core_pick_seq = rq->core->core_task_seq;
5572 next = rq->core_pick;
5573 rq->core_sched_seq = rq->core->core_pick_seq;
5575 /* Something should have been selected for current CPU */
5576 WARN_ON_ONCE(!next);
5579 * Reschedule siblings
5581 * NOTE: L1TF -- at this point we're no longer running the old task and
5582 * sending an IPI (below) ensures the sibling will no longer be running
5583 * their task. This ensures there is no inter-sibling overlap between
5584 * non-matching user state.
5586 for_each_cpu(i, smt_mask) {
5587 struct rq *rq_i = cpu_rq(i);
5590 * An online sibling might have gone offline before a task
5591 * could be picked for it, or it might be offline but later
5592 * happen to come online, but its too late and nothing was
5593 * picked for it. That's Ok - it will pick tasks for itself,
5596 if (!rq_i->core_pick)
5600 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5601 * fi_before fi update?
5607 if (!(fi_before && rq->core->core_forceidle))
5608 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5610 rq_i->core_pick->core_occupation = occ;
5613 rq_i->core_pick = NULL;
5617 /* Did we break L1TF mitigation requirements? */
5618 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5620 if (rq_i->curr == rq_i->core_pick) {
5621 rq_i->core_pick = NULL;
5629 set_next_task(rq, next);
5633 static bool try_steal_cookie(int this, int that)
5635 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5636 struct task_struct *p;
5637 unsigned long cookie;
5638 bool success = false;
5640 local_irq_disable();
5641 double_rq_lock(dst, src);
5643 cookie = dst->core->core_cookie;
5647 if (dst->curr != dst->idle)
5650 p = sched_core_find(src, cookie);
5655 if (p == src->core_pick || p == src->curr)
5658 if (!cpumask_test_cpu(this, &p->cpus_mask))
5661 if (p->core_occupation > dst->idle->core_occupation)
5664 p->on_rq = TASK_ON_RQ_MIGRATING;
5665 deactivate_task(src, p, 0);
5666 set_task_cpu(p, this);
5667 activate_task(dst, p, 0);
5668 p->on_rq = TASK_ON_RQ_QUEUED;
5676 p = sched_core_next(p, cookie);
5680 double_rq_unlock(dst, src);
5686 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5690 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5697 if (try_steal_cookie(cpu, i))
5704 static void sched_core_balance(struct rq *rq)
5706 struct sched_domain *sd;
5707 int cpu = cpu_of(rq);
5711 raw_spin_rq_unlock_irq(rq);
5712 for_each_domain(cpu, sd) {
5716 if (steal_cookie_task(cpu, sd))
5719 raw_spin_rq_lock_irq(rq);
5724 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5726 void queue_core_balance(struct rq *rq)
5728 if (!sched_core_enabled(rq))
5731 if (!rq->core->core_cookie)
5734 if (!rq->nr_running) /* not forced idle */
5737 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5740 static inline void sched_core_cpu_starting(unsigned int cpu)
5742 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5743 struct rq *rq, *core_rq = NULL;
5746 core_rq = cpu_rq(cpu)->core;
5749 for_each_cpu(i, smt_mask) {
5751 if (rq->core && rq->core == rq)
5756 core_rq = cpu_rq(cpu);
5758 for_each_cpu(i, smt_mask) {
5761 WARN_ON_ONCE(rq->core && rq->core != core_rq);
5766 #else /* !CONFIG_SCHED_CORE */
5768 static inline void sched_core_cpu_starting(unsigned int cpu) {}
5770 static struct task_struct *
5771 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5773 return __pick_next_task(rq, prev, rf);
5776 #endif /* CONFIG_SCHED_CORE */
5779 * __schedule() is the main scheduler function.
5781 * The main means of driving the scheduler and thus entering this function are:
5783 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
5785 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
5786 * paths. For example, see arch/x86/entry_64.S.
5788 * To drive preemption between tasks, the scheduler sets the flag in timer
5789 * interrupt handler scheduler_tick().
5791 * 3. Wakeups don't really cause entry into schedule(). They add a
5792 * task to the run-queue and that's it.
5794 * Now, if the new task added to the run-queue preempts the current
5795 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
5796 * called on the nearest possible occasion:
5798 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
5800 * - in syscall or exception context, at the next outmost
5801 * preempt_enable(). (this might be as soon as the wake_up()'s
5804 * - in IRQ context, return from interrupt-handler to
5805 * preemptible context
5807 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
5810 * - cond_resched() call
5811 * - explicit schedule() call
5812 * - return from syscall or exception to user-space
5813 * - return from interrupt-handler to user-space
5815 * WARNING: must be called with preemption disabled!
5817 static void __sched notrace __schedule(bool preempt)
5819 struct task_struct *prev, *next;
5820 unsigned long *switch_count;
5821 unsigned long prev_state;
5826 cpu = smp_processor_id();
5830 schedule_debug(prev, preempt);
5832 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
5835 local_irq_disable();
5836 rcu_note_context_switch(preempt);
5839 * Make sure that signal_pending_state()->signal_pending() below
5840 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
5841 * done by the caller to avoid the race with signal_wake_up():
5843 * __set_current_state(@state) signal_wake_up()
5844 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
5845 * wake_up_state(p, state)
5846 * LOCK rq->lock LOCK p->pi_state
5847 * smp_mb__after_spinlock() smp_mb__after_spinlock()
5848 * if (signal_pending_state()) if (p->state & @state)
5850 * Also, the membarrier system call requires a full memory barrier
5851 * after coming from user-space, before storing to rq->curr.
5854 smp_mb__after_spinlock();
5856 /* Promote REQ to ACT */
5857 rq->clock_update_flags <<= 1;
5858 update_rq_clock(rq);
5860 switch_count = &prev->nivcsw;
5863 * We must load prev->state once (task_struct::state is volatile), such
5866 * - we form a control dependency vs deactivate_task() below.
5867 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5869 prev_state = prev->state;
5870 if (!preempt && prev_state) {
5871 if (signal_pending_state(prev_state, prev)) {
5872 prev->state = TASK_RUNNING;
5874 prev->sched_contributes_to_load =
5875 (prev_state & TASK_UNINTERRUPTIBLE) &&
5876 !(prev_state & TASK_NOLOAD) &&
5877 !(prev->flags & PF_FROZEN);
5879 if (prev->sched_contributes_to_load)
5880 rq->nr_uninterruptible++;
5883 * __schedule() ttwu()
5884 * prev_state = prev->state; if (p->on_rq && ...)
5885 * if (prev_state) goto out;
5886 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5887 * p->state = TASK_WAKING
5889 * Where __schedule() and ttwu() have matching control dependencies.
5891 * After this, schedule() must not care about p->state any more.
5893 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
5895 if (prev->in_iowait) {
5896 atomic_inc(&rq->nr_iowait);
5897 delayacct_blkio_start();
5900 switch_count = &prev->nvcsw;
5903 next = pick_next_task(rq, prev, &rf);
5904 clear_tsk_need_resched(prev);
5905 clear_preempt_need_resched();
5906 #ifdef CONFIG_SCHED_DEBUG
5907 rq->last_seen_need_resched_ns = 0;
5910 if (likely(prev != next)) {
5913 * RCU users of rcu_dereference(rq->curr) may not see
5914 * changes to task_struct made by pick_next_task().
5916 RCU_INIT_POINTER(rq->curr, next);
5918 * The membarrier system call requires each architecture
5919 * to have a full memory barrier after updating
5920 * rq->curr, before returning to user-space.
5922 * Here are the schemes providing that barrier on the
5923 * various architectures:
5924 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5925 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5926 * - finish_lock_switch() for weakly-ordered
5927 * architectures where spin_unlock is a full barrier,
5928 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5929 * is a RELEASE barrier),
5933 migrate_disable_switch(rq, prev);
5934 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5936 trace_sched_switch(preempt, prev, next);
5938 /* Also unlocks the rq: */
5939 rq = context_switch(rq, prev, next, &rf);
5941 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5943 rq_unpin_lock(rq, &rf);
5944 __balance_callbacks(rq);
5945 raw_spin_rq_unlock_irq(rq);
5949 void __noreturn do_task_dead(void)
5951 /* Causes final put_task_struct in finish_task_switch(): */
5952 set_special_state(TASK_DEAD);
5954 /* Tell freezer to ignore us: */
5955 current->flags |= PF_NOFREEZE;
5960 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5965 static inline void sched_submit_work(struct task_struct *tsk)
5967 unsigned int task_flags;
5972 task_flags = tsk->flags;
5974 * If a worker went to sleep, notify and ask workqueue whether
5975 * it wants to wake up a task to maintain concurrency.
5976 * As this function is called inside the schedule() context,
5977 * we disable preemption to avoid it calling schedule() again
5978 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5981 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5983 if (task_flags & PF_WQ_WORKER)
5984 wq_worker_sleeping(tsk);
5986 io_wq_worker_sleeping(tsk);
5987 preempt_enable_no_resched();
5990 if (tsk_is_pi_blocked(tsk))
5994 * If we are going to sleep and we have plugged IO queued,
5995 * make sure to submit it to avoid deadlocks.
5997 if (blk_needs_flush_plug(tsk))
5998 blk_schedule_flush_plug(tsk);
6001 static void sched_update_worker(struct task_struct *tsk)
6003 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6004 if (tsk->flags & PF_WQ_WORKER)
6005 wq_worker_running(tsk);
6007 io_wq_worker_running(tsk);
6011 asmlinkage __visible void __sched schedule(void)
6013 struct task_struct *tsk = current;
6015 sched_submit_work(tsk);
6019 sched_preempt_enable_no_resched();
6020 } while (need_resched());
6021 sched_update_worker(tsk);
6023 EXPORT_SYMBOL(schedule);
6026 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6027 * state (have scheduled out non-voluntarily) by making sure that all
6028 * tasks have either left the run queue or have gone into user space.
6029 * As idle tasks do not do either, they must not ever be preempted
6030 * (schedule out non-voluntarily).
6032 * schedule_idle() is similar to schedule_preempt_disable() except that it
6033 * never enables preemption because it does not call sched_submit_work().
6035 void __sched schedule_idle(void)
6038 * As this skips calling sched_submit_work(), which the idle task does
6039 * regardless because that function is a nop when the task is in a
6040 * TASK_RUNNING state, make sure this isn't used someplace that the
6041 * current task can be in any other state. Note, idle is always in the
6042 * TASK_RUNNING state.
6044 WARN_ON_ONCE(current->state);
6047 } while (need_resched());
6050 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6051 asmlinkage __visible void __sched schedule_user(void)
6054 * If we come here after a random call to set_need_resched(),
6055 * or we have been woken up remotely but the IPI has not yet arrived,
6056 * we haven't yet exited the RCU idle mode. Do it here manually until
6057 * we find a better solution.
6059 * NB: There are buggy callers of this function. Ideally we
6060 * should warn if prev_state != CONTEXT_USER, but that will trigger
6061 * too frequently to make sense yet.
6063 enum ctx_state prev_state = exception_enter();
6065 exception_exit(prev_state);
6070 * schedule_preempt_disabled - called with preemption disabled
6072 * Returns with preemption disabled. Note: preempt_count must be 1
6074 void __sched schedule_preempt_disabled(void)
6076 sched_preempt_enable_no_resched();
6081 static void __sched notrace preempt_schedule_common(void)
6085 * Because the function tracer can trace preempt_count_sub()
6086 * and it also uses preempt_enable/disable_notrace(), if
6087 * NEED_RESCHED is set, the preempt_enable_notrace() called
6088 * by the function tracer will call this function again and
6089 * cause infinite recursion.
6091 * Preemption must be disabled here before the function
6092 * tracer can trace. Break up preempt_disable() into two
6093 * calls. One to disable preemption without fear of being
6094 * traced. The other to still record the preemption latency,
6095 * which can also be traced by the function tracer.
6097 preempt_disable_notrace();
6098 preempt_latency_start(1);
6100 preempt_latency_stop(1);
6101 preempt_enable_no_resched_notrace();
6104 * Check again in case we missed a preemption opportunity
6105 * between schedule and now.
6107 } while (need_resched());
6110 #ifdef CONFIG_PREEMPTION
6112 * This is the entry point to schedule() from in-kernel preemption
6113 * off of preempt_enable.
6115 asmlinkage __visible void __sched notrace preempt_schedule(void)
6118 * If there is a non-zero preempt_count or interrupts are disabled,
6119 * we do not want to preempt the current task. Just return..
6121 if (likely(!preemptible()))
6124 preempt_schedule_common();
6126 NOKPROBE_SYMBOL(preempt_schedule);
6127 EXPORT_SYMBOL(preempt_schedule);
6129 #ifdef CONFIG_PREEMPT_DYNAMIC
6130 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6131 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6136 * preempt_schedule_notrace - preempt_schedule called by tracing
6138 * The tracing infrastructure uses preempt_enable_notrace to prevent
6139 * recursion and tracing preempt enabling caused by the tracing
6140 * infrastructure itself. But as tracing can happen in areas coming
6141 * from userspace or just about to enter userspace, a preempt enable
6142 * can occur before user_exit() is called. This will cause the scheduler
6143 * to be called when the system is still in usermode.
6145 * To prevent this, the preempt_enable_notrace will use this function
6146 * instead of preempt_schedule() to exit user context if needed before
6147 * calling the scheduler.
6149 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6151 enum ctx_state prev_ctx;
6153 if (likely(!preemptible()))
6158 * Because the function tracer can trace preempt_count_sub()
6159 * and it also uses preempt_enable/disable_notrace(), if
6160 * NEED_RESCHED is set, the preempt_enable_notrace() called
6161 * by the function tracer will call this function again and
6162 * cause infinite recursion.
6164 * Preemption must be disabled here before the function
6165 * tracer can trace. Break up preempt_disable() into two
6166 * calls. One to disable preemption without fear of being
6167 * traced. The other to still record the preemption latency,
6168 * which can also be traced by the function tracer.
6170 preempt_disable_notrace();
6171 preempt_latency_start(1);
6173 * Needs preempt disabled in case user_exit() is traced
6174 * and the tracer calls preempt_enable_notrace() causing
6175 * an infinite recursion.
6177 prev_ctx = exception_enter();
6179 exception_exit(prev_ctx);
6181 preempt_latency_stop(1);
6182 preempt_enable_no_resched_notrace();
6183 } while (need_resched());
6185 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6187 #ifdef CONFIG_PREEMPT_DYNAMIC
6188 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6189 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6192 #endif /* CONFIG_PREEMPTION */
6194 #ifdef CONFIG_PREEMPT_DYNAMIC
6196 #include <linux/entry-common.h>
6201 * SC:preempt_schedule
6202 * SC:preempt_schedule_notrace
6203 * SC:irqentry_exit_cond_resched
6207 * cond_resched <- __cond_resched
6208 * might_resched <- RET0
6209 * preempt_schedule <- NOP
6210 * preempt_schedule_notrace <- NOP
6211 * irqentry_exit_cond_resched <- NOP
6214 * cond_resched <- __cond_resched
6215 * might_resched <- __cond_resched
6216 * preempt_schedule <- NOP
6217 * preempt_schedule_notrace <- NOP
6218 * irqentry_exit_cond_resched <- NOP
6221 * cond_resched <- RET0
6222 * might_resched <- RET0
6223 * preempt_schedule <- preempt_schedule
6224 * preempt_schedule_notrace <- preempt_schedule_notrace
6225 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6229 preempt_dynamic_none = 0,
6230 preempt_dynamic_voluntary,
6231 preempt_dynamic_full,
6234 int preempt_dynamic_mode = preempt_dynamic_full;
6236 int sched_dynamic_mode(const char *str)
6238 if (!strcmp(str, "none"))
6239 return preempt_dynamic_none;
6241 if (!strcmp(str, "voluntary"))
6242 return preempt_dynamic_voluntary;
6244 if (!strcmp(str, "full"))
6245 return preempt_dynamic_full;
6250 void sched_dynamic_update(int mode)
6253 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6254 * the ZERO state, which is invalid.
6256 static_call_update(cond_resched, __cond_resched);
6257 static_call_update(might_resched, __cond_resched);
6258 static_call_update(preempt_schedule, __preempt_schedule_func);
6259 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6260 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6263 case preempt_dynamic_none:
6264 static_call_update(cond_resched, __cond_resched);
6265 static_call_update(might_resched, (void *)&__static_call_return0);
6266 static_call_update(preempt_schedule, NULL);
6267 static_call_update(preempt_schedule_notrace, NULL);
6268 static_call_update(irqentry_exit_cond_resched, NULL);
6269 pr_info("Dynamic Preempt: none\n");
6272 case preempt_dynamic_voluntary:
6273 static_call_update(cond_resched, __cond_resched);
6274 static_call_update(might_resched, __cond_resched);
6275 static_call_update(preempt_schedule, NULL);
6276 static_call_update(preempt_schedule_notrace, NULL);
6277 static_call_update(irqentry_exit_cond_resched, NULL);
6278 pr_info("Dynamic Preempt: voluntary\n");
6281 case preempt_dynamic_full:
6282 static_call_update(cond_resched, (void *)&__static_call_return0);
6283 static_call_update(might_resched, (void *)&__static_call_return0);
6284 static_call_update(preempt_schedule, __preempt_schedule_func);
6285 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6286 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6287 pr_info("Dynamic Preempt: full\n");
6291 preempt_dynamic_mode = mode;
6294 static int __init setup_preempt_mode(char *str)
6296 int mode = sched_dynamic_mode(str);
6298 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6302 sched_dynamic_update(mode);
6305 __setup("preempt=", setup_preempt_mode);
6307 #endif /* CONFIG_PREEMPT_DYNAMIC */
6310 * This is the entry point to schedule() from kernel preemption
6311 * off of irq context.
6312 * Note, that this is called and return with irqs disabled. This will
6313 * protect us against recursive calling from irq.
6315 asmlinkage __visible void __sched preempt_schedule_irq(void)
6317 enum ctx_state prev_state;
6319 /* Catch callers which need to be fixed */
6320 BUG_ON(preempt_count() || !irqs_disabled());
6322 prev_state = exception_enter();
6328 local_irq_disable();
6329 sched_preempt_enable_no_resched();
6330 } while (need_resched());
6332 exception_exit(prev_state);
6335 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6338 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6339 return try_to_wake_up(curr->private, mode, wake_flags);
6341 EXPORT_SYMBOL(default_wake_function);
6343 #ifdef CONFIG_RT_MUTEXES
6345 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6348 prio = min(prio, pi_task->prio);
6353 static inline int rt_effective_prio(struct task_struct *p, int prio)
6355 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6357 return __rt_effective_prio(pi_task, prio);
6361 * rt_mutex_setprio - set the current priority of a task
6363 * @pi_task: donor task
6365 * This function changes the 'effective' priority of a task. It does
6366 * not touch ->normal_prio like __setscheduler().
6368 * Used by the rt_mutex code to implement priority inheritance
6369 * logic. Call site only calls if the priority of the task changed.
6371 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6373 int prio, oldprio, queued, running, queue_flag =
6374 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6375 const struct sched_class *prev_class;
6379 /* XXX used to be waiter->prio, not waiter->task->prio */
6380 prio = __rt_effective_prio(pi_task, p->normal_prio);
6383 * If nothing changed; bail early.
6385 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6388 rq = __task_rq_lock(p, &rf);
6389 update_rq_clock(rq);
6391 * Set under pi_lock && rq->lock, such that the value can be used under
6394 * Note that there is loads of tricky to make this pointer cache work
6395 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6396 * ensure a task is de-boosted (pi_task is set to NULL) before the
6397 * task is allowed to run again (and can exit). This ensures the pointer
6398 * points to a blocked task -- which guarantees the task is present.
6400 p->pi_top_task = pi_task;
6403 * For FIFO/RR we only need to set prio, if that matches we're done.
6405 if (prio == p->prio && !dl_prio(prio))
6409 * Idle task boosting is a nono in general. There is one
6410 * exception, when PREEMPT_RT and NOHZ is active:
6412 * The idle task calls get_next_timer_interrupt() and holds
6413 * the timer wheel base->lock on the CPU and another CPU wants
6414 * to access the timer (probably to cancel it). We can safely
6415 * ignore the boosting request, as the idle CPU runs this code
6416 * with interrupts disabled and will complete the lock
6417 * protected section without being interrupted. So there is no
6418 * real need to boost.
6420 if (unlikely(p == rq->idle)) {
6421 WARN_ON(p != rq->curr);
6422 WARN_ON(p->pi_blocked_on);
6426 trace_sched_pi_setprio(p, pi_task);
6429 if (oldprio == prio)
6430 queue_flag &= ~DEQUEUE_MOVE;
6432 prev_class = p->sched_class;
6433 queued = task_on_rq_queued(p);
6434 running = task_current(rq, p);
6436 dequeue_task(rq, p, queue_flag);
6438 put_prev_task(rq, p);
6441 * Boosting condition are:
6442 * 1. -rt task is running and holds mutex A
6443 * --> -dl task blocks on mutex A
6445 * 2. -dl task is running and holds mutex A
6446 * --> -dl task blocks on mutex A and could preempt the
6449 if (dl_prio(prio)) {
6450 if (!dl_prio(p->normal_prio) ||
6451 (pi_task && dl_prio(pi_task->prio) &&
6452 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6453 p->dl.pi_se = pi_task->dl.pi_se;
6454 queue_flag |= ENQUEUE_REPLENISH;
6456 p->dl.pi_se = &p->dl;
6458 p->sched_class = &dl_sched_class;
6459 } else if (rt_prio(prio)) {
6460 if (dl_prio(oldprio))
6461 p->dl.pi_se = &p->dl;
6463 queue_flag |= ENQUEUE_HEAD;
6464 p->sched_class = &rt_sched_class;
6466 if (dl_prio(oldprio))
6467 p->dl.pi_se = &p->dl;
6468 if (rt_prio(oldprio))
6470 p->sched_class = &fair_sched_class;
6476 enqueue_task(rq, p, queue_flag);
6478 set_next_task(rq, p);
6480 check_class_changed(rq, p, prev_class, oldprio);
6482 /* Avoid rq from going away on us: */
6485 rq_unpin_lock(rq, &rf);
6486 __balance_callbacks(rq);
6487 raw_spin_rq_unlock(rq);
6492 static inline int rt_effective_prio(struct task_struct *p, int prio)
6498 void set_user_nice(struct task_struct *p, long nice)
6500 bool queued, running;
6505 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6508 * We have to be careful, if called from sys_setpriority(),
6509 * the task might be in the middle of scheduling on another CPU.
6511 rq = task_rq_lock(p, &rf);
6512 update_rq_clock(rq);
6515 * The RT priorities are set via sched_setscheduler(), but we still
6516 * allow the 'normal' nice value to be set - but as expected
6517 * it won't have any effect on scheduling until the task is
6518 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6520 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6521 p->static_prio = NICE_TO_PRIO(nice);
6524 queued = task_on_rq_queued(p);
6525 running = task_current(rq, p);
6527 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6529 put_prev_task(rq, p);
6531 p->static_prio = NICE_TO_PRIO(nice);
6532 set_load_weight(p, true);
6534 p->prio = effective_prio(p);
6537 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6539 set_next_task(rq, p);
6542 * If the task increased its priority or is running and
6543 * lowered its priority, then reschedule its CPU:
6545 p->sched_class->prio_changed(rq, p, old_prio);
6548 task_rq_unlock(rq, p, &rf);
6550 EXPORT_SYMBOL(set_user_nice);
6553 * can_nice - check if a task can reduce its nice value
6557 int can_nice(const struct task_struct *p, const int nice)
6559 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6560 int nice_rlim = nice_to_rlimit(nice);
6562 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6563 capable(CAP_SYS_NICE));
6566 #ifdef __ARCH_WANT_SYS_NICE
6569 * sys_nice - change the priority of the current process.
6570 * @increment: priority increment
6572 * sys_setpriority is a more generic, but much slower function that
6573 * does similar things.
6575 SYSCALL_DEFINE1(nice, int, increment)
6580 * Setpriority might change our priority at the same moment.
6581 * We don't have to worry. Conceptually one call occurs first
6582 * and we have a single winner.
6584 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6585 nice = task_nice(current) + increment;
6587 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6588 if (increment < 0 && !can_nice(current, nice))
6591 retval = security_task_setnice(current, nice);
6595 set_user_nice(current, nice);
6602 * task_prio - return the priority value of a given task.
6603 * @p: the task in question.
6605 * Return: The priority value as seen by users in /proc.
6607 * sched policy return value kernel prio user prio/nice
6609 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6610 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6611 * deadline -101 -1 0
6613 int task_prio(const struct task_struct *p)
6615 return p->prio - MAX_RT_PRIO;
6619 * idle_cpu - is a given CPU idle currently?
6620 * @cpu: the processor in question.
6622 * Return: 1 if the CPU is currently idle. 0 otherwise.
6624 int idle_cpu(int cpu)
6626 struct rq *rq = cpu_rq(cpu);
6628 if (rq->curr != rq->idle)
6635 if (rq->ttwu_pending)
6643 * available_idle_cpu - is a given CPU idle for enqueuing work.
6644 * @cpu: the CPU in question.
6646 * Return: 1 if the CPU is currently idle. 0 otherwise.
6648 int available_idle_cpu(int cpu)
6653 if (vcpu_is_preempted(cpu))
6660 * idle_task - return the idle task for a given CPU.
6661 * @cpu: the processor in question.
6663 * Return: The idle task for the CPU @cpu.
6665 struct task_struct *idle_task(int cpu)
6667 return cpu_rq(cpu)->idle;
6672 * This function computes an effective utilization for the given CPU, to be
6673 * used for frequency selection given the linear relation: f = u * f_max.
6675 * The scheduler tracks the following metrics:
6677 * cpu_util_{cfs,rt,dl,irq}()
6680 * Where the cfs,rt and dl util numbers are tracked with the same metric and
6681 * synchronized windows and are thus directly comparable.
6683 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
6684 * which excludes things like IRQ and steal-time. These latter are then accrued
6685 * in the irq utilization.
6687 * The DL bandwidth number otoh is not a measured metric but a value computed
6688 * based on the task model parameters and gives the minimal utilization
6689 * required to meet deadlines.
6691 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
6692 unsigned long max, enum cpu_util_type type,
6693 struct task_struct *p)
6695 unsigned long dl_util, util, irq;
6696 struct rq *rq = cpu_rq(cpu);
6698 if (!uclamp_is_used() &&
6699 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
6704 * Early check to see if IRQ/steal time saturates the CPU, can be
6705 * because of inaccuracies in how we track these -- see
6706 * update_irq_load_avg().
6708 irq = cpu_util_irq(rq);
6709 if (unlikely(irq >= max))
6713 * Because the time spend on RT/DL tasks is visible as 'lost' time to
6714 * CFS tasks and we use the same metric to track the effective
6715 * utilization (PELT windows are synchronized) we can directly add them
6716 * to obtain the CPU's actual utilization.
6718 * CFS and RT utilization can be boosted or capped, depending on
6719 * utilization clamp constraints requested by currently RUNNABLE
6721 * When there are no CFS RUNNABLE tasks, clamps are released and
6722 * frequency will be gracefully reduced with the utilization decay.
6724 util = util_cfs + cpu_util_rt(rq);
6725 if (type == FREQUENCY_UTIL)
6726 util = uclamp_rq_util_with(rq, util, p);
6728 dl_util = cpu_util_dl(rq);
6731 * For frequency selection we do not make cpu_util_dl() a permanent part
6732 * of this sum because we want to use cpu_bw_dl() later on, but we need
6733 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
6734 * that we select f_max when there is no idle time.
6736 * NOTE: numerical errors or stop class might cause us to not quite hit
6737 * saturation when we should -- something for later.
6739 if (util + dl_util >= max)
6743 * OTOH, for energy computation we need the estimated running time, so
6744 * include util_dl and ignore dl_bw.
6746 if (type == ENERGY_UTIL)
6750 * There is still idle time; further improve the number by using the
6751 * irq metric. Because IRQ/steal time is hidden from the task clock we
6752 * need to scale the task numbers:
6755 * U' = irq + --------- * U
6758 util = scale_irq_capacity(util, irq, max);
6762 * Bandwidth required by DEADLINE must always be granted while, for
6763 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
6764 * to gracefully reduce the frequency when no tasks show up for longer
6767 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
6768 * bw_dl as requested freq. However, cpufreq is not yet ready for such
6769 * an interface. So, we only do the latter for now.
6771 if (type == FREQUENCY_UTIL)
6772 util += cpu_bw_dl(rq);
6774 return min(max, util);
6777 unsigned long sched_cpu_util(int cpu, unsigned long max)
6779 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
6782 #endif /* CONFIG_SMP */
6785 * find_process_by_pid - find a process with a matching PID value.
6786 * @pid: the pid in question.
6788 * The task of @pid, if found. %NULL otherwise.
6790 static struct task_struct *find_process_by_pid(pid_t pid)
6792 return pid ? find_task_by_vpid(pid) : current;
6796 * sched_setparam() passes in -1 for its policy, to let the functions
6797 * it calls know not to change it.
6799 #define SETPARAM_POLICY -1
6801 static void __setscheduler_params(struct task_struct *p,
6802 const struct sched_attr *attr)
6804 int policy = attr->sched_policy;
6806 if (policy == SETPARAM_POLICY)
6811 if (dl_policy(policy))
6812 __setparam_dl(p, attr);
6813 else if (fair_policy(policy))
6814 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6817 * __sched_setscheduler() ensures attr->sched_priority == 0 when
6818 * !rt_policy. Always setting this ensures that things like
6819 * getparam()/getattr() don't report silly values for !rt tasks.
6821 p->rt_priority = attr->sched_priority;
6822 p->normal_prio = normal_prio(p);
6823 set_load_weight(p, true);
6826 /* Actually do priority change: must hold pi & rq lock. */
6827 static void __setscheduler(struct rq *rq, struct task_struct *p,
6828 const struct sched_attr *attr, bool keep_boost)
6831 * If params can't change scheduling class changes aren't allowed
6834 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
6837 __setscheduler_params(p, attr);
6840 * Keep a potential priority boosting if called from
6841 * sched_setscheduler().
6843 p->prio = normal_prio(p);
6845 p->prio = rt_effective_prio(p, p->prio);
6847 if (dl_prio(p->prio))
6848 p->sched_class = &dl_sched_class;
6849 else if (rt_prio(p->prio))
6850 p->sched_class = &rt_sched_class;
6852 p->sched_class = &fair_sched_class;
6856 * Check the target process has a UID that matches the current process's:
6858 static bool check_same_owner(struct task_struct *p)
6860 const struct cred *cred = current_cred(), *pcred;
6864 pcred = __task_cred(p);
6865 match = (uid_eq(cred->euid, pcred->euid) ||
6866 uid_eq(cred->euid, pcred->uid));
6871 static int __sched_setscheduler(struct task_struct *p,
6872 const struct sched_attr *attr,
6875 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
6876 MAX_RT_PRIO - 1 - attr->sched_priority;
6877 int retval, oldprio, oldpolicy = -1, queued, running;
6878 int new_effective_prio, policy = attr->sched_policy;
6879 const struct sched_class *prev_class;
6880 struct callback_head *head;
6883 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6886 /* The pi code expects interrupts enabled */
6887 BUG_ON(pi && in_interrupt());
6889 /* Double check policy once rq lock held: */
6891 reset_on_fork = p->sched_reset_on_fork;
6892 policy = oldpolicy = p->policy;
6894 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
6896 if (!valid_policy(policy))
6900 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
6904 * Valid priorities for SCHED_FIFO and SCHED_RR are
6905 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6906 * SCHED_BATCH and SCHED_IDLE is 0.
6908 if (attr->sched_priority > MAX_RT_PRIO-1)
6910 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
6911 (rt_policy(policy) != (attr->sched_priority != 0)))
6915 * Allow unprivileged RT tasks to decrease priority:
6917 if (user && !capable(CAP_SYS_NICE)) {
6918 if (fair_policy(policy)) {
6919 if (attr->sched_nice < task_nice(p) &&
6920 !can_nice(p, attr->sched_nice))
6924 if (rt_policy(policy)) {
6925 unsigned long rlim_rtprio =
6926 task_rlimit(p, RLIMIT_RTPRIO);
6928 /* Can't set/change the rt policy: */
6929 if (policy != p->policy && !rlim_rtprio)
6932 /* Can't increase priority: */
6933 if (attr->sched_priority > p->rt_priority &&
6934 attr->sched_priority > rlim_rtprio)
6939 * Can't set/change SCHED_DEADLINE policy at all for now
6940 * (safest behavior); in the future we would like to allow
6941 * unprivileged DL tasks to increase their relative deadline
6942 * or reduce their runtime (both ways reducing utilization)
6944 if (dl_policy(policy))
6948 * Treat SCHED_IDLE as nice 20. Only allow a switch to
6949 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6951 if (task_has_idle_policy(p) && !idle_policy(policy)) {
6952 if (!can_nice(p, task_nice(p)))
6956 /* Can't change other user's priorities: */
6957 if (!check_same_owner(p))
6960 /* Normal users shall not reset the sched_reset_on_fork flag: */
6961 if (p->sched_reset_on_fork && !reset_on_fork)
6966 if (attr->sched_flags & SCHED_FLAG_SUGOV)
6969 retval = security_task_setscheduler(p);
6974 /* Update task specific "requested" clamps */
6975 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
6976 retval = uclamp_validate(p, attr);
6985 * Make sure no PI-waiters arrive (or leave) while we are
6986 * changing the priority of the task:
6988 * To be able to change p->policy safely, the appropriate
6989 * runqueue lock must be held.
6991 rq = task_rq_lock(p, &rf);
6992 update_rq_clock(rq);
6995 * Changing the policy of the stop threads its a very bad idea:
6997 if (p == rq->stop) {
7003 * If not changing anything there's no need to proceed further,
7004 * but store a possible modification of reset_on_fork.
7006 if (unlikely(policy == p->policy)) {
7007 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7009 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7011 if (dl_policy(policy) && dl_param_changed(p, attr))
7013 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7016 p->sched_reset_on_fork = reset_on_fork;
7023 #ifdef CONFIG_RT_GROUP_SCHED
7025 * Do not allow realtime tasks into groups that have no runtime
7028 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7029 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7030 !task_group_is_autogroup(task_group(p))) {
7036 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7037 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7038 cpumask_t *span = rq->rd->span;
7041 * Don't allow tasks with an affinity mask smaller than
7042 * the entire root_domain to become SCHED_DEADLINE. We
7043 * will also fail if there's no bandwidth available.
7045 if (!cpumask_subset(span, p->cpus_ptr) ||
7046 rq->rd->dl_bw.bw == 0) {
7054 /* Re-check policy now with rq lock held: */
7055 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7056 policy = oldpolicy = -1;
7057 task_rq_unlock(rq, p, &rf);
7059 cpuset_read_unlock();
7064 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7065 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7068 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7073 p->sched_reset_on_fork = reset_on_fork;
7078 * Take priority boosted tasks into account. If the new
7079 * effective priority is unchanged, we just store the new
7080 * normal parameters and do not touch the scheduler class and
7081 * the runqueue. This will be done when the task deboost
7084 new_effective_prio = rt_effective_prio(p, newprio);
7085 if (new_effective_prio == oldprio)
7086 queue_flags &= ~DEQUEUE_MOVE;
7089 queued = task_on_rq_queued(p);
7090 running = task_current(rq, p);
7092 dequeue_task(rq, p, queue_flags);
7094 put_prev_task(rq, p);
7096 prev_class = p->sched_class;
7098 __setscheduler(rq, p, attr, pi);
7099 __setscheduler_uclamp(p, attr);
7103 * We enqueue to tail when the priority of a task is
7104 * increased (user space view).
7106 if (oldprio < p->prio)
7107 queue_flags |= ENQUEUE_HEAD;
7109 enqueue_task(rq, p, queue_flags);
7112 set_next_task(rq, p);
7114 check_class_changed(rq, p, prev_class, oldprio);
7116 /* Avoid rq from going away on us: */
7118 head = splice_balance_callbacks(rq);
7119 task_rq_unlock(rq, p, &rf);
7122 cpuset_read_unlock();
7123 rt_mutex_adjust_pi(p);
7126 /* Run balance callbacks after we've adjusted the PI chain: */
7127 balance_callbacks(rq, head);
7133 task_rq_unlock(rq, p, &rf);
7135 cpuset_read_unlock();
7139 static int _sched_setscheduler(struct task_struct *p, int policy,
7140 const struct sched_param *param, bool check)
7142 struct sched_attr attr = {
7143 .sched_policy = policy,
7144 .sched_priority = param->sched_priority,
7145 .sched_nice = PRIO_TO_NICE(p->static_prio),
7148 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7149 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7150 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7151 policy &= ~SCHED_RESET_ON_FORK;
7152 attr.sched_policy = policy;
7155 return __sched_setscheduler(p, &attr, check, true);
7158 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7159 * @p: the task in question.
7160 * @policy: new policy.
7161 * @param: structure containing the new RT priority.
7163 * Use sched_set_fifo(), read its comment.
7165 * Return: 0 on success. An error code otherwise.
7167 * NOTE that the task may be already dead.
7169 int sched_setscheduler(struct task_struct *p, int policy,
7170 const struct sched_param *param)
7172 return _sched_setscheduler(p, policy, param, true);
7175 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7177 return __sched_setscheduler(p, attr, true, true);
7180 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7182 return __sched_setscheduler(p, attr, false, true);
7184 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7187 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7188 * @p: the task in question.
7189 * @policy: new policy.
7190 * @param: structure containing the new RT priority.
7192 * Just like sched_setscheduler, only don't bother checking if the
7193 * current context has permission. For example, this is needed in
7194 * stop_machine(): we create temporary high priority worker threads,
7195 * but our caller might not have that capability.
7197 * Return: 0 on success. An error code otherwise.
7199 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7200 const struct sched_param *param)
7202 return _sched_setscheduler(p, policy, param, false);
7206 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7207 * incapable of resource management, which is the one thing an OS really should
7210 * This is of course the reason it is limited to privileged users only.
7212 * Worse still; it is fundamentally impossible to compose static priority
7213 * workloads. You cannot take two correctly working static prio workloads
7214 * and smash them together and still expect them to work.
7216 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7220 * The administrator _MUST_ configure the system, the kernel simply doesn't
7221 * know enough information to make a sensible choice.
7223 void sched_set_fifo(struct task_struct *p)
7225 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7226 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7228 EXPORT_SYMBOL_GPL(sched_set_fifo);
7231 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7233 void sched_set_fifo_low(struct task_struct *p)
7235 struct sched_param sp = { .sched_priority = 1 };
7236 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7238 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7240 void sched_set_normal(struct task_struct *p, int nice)
7242 struct sched_attr attr = {
7243 .sched_policy = SCHED_NORMAL,
7246 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7248 EXPORT_SYMBOL_GPL(sched_set_normal);
7251 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7253 struct sched_param lparam;
7254 struct task_struct *p;
7257 if (!param || pid < 0)
7259 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7264 p = find_process_by_pid(pid);
7270 retval = sched_setscheduler(p, policy, &lparam);
7278 * Mimics kernel/events/core.c perf_copy_attr().
7280 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7285 /* Zero the full structure, so that a short copy will be nice: */
7286 memset(attr, 0, sizeof(*attr));
7288 ret = get_user(size, &uattr->size);
7292 /* ABI compatibility quirk: */
7294 size = SCHED_ATTR_SIZE_VER0;
7295 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7298 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7305 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7306 size < SCHED_ATTR_SIZE_VER1)
7310 * XXX: Do we want to be lenient like existing syscalls; or do we want
7311 * to be strict and return an error on out-of-bounds values?
7313 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7318 put_user(sizeof(*attr), &uattr->size);
7323 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7324 * @pid: the pid in question.
7325 * @policy: new policy.
7326 * @param: structure containing the new RT priority.
7328 * Return: 0 on success. An error code otherwise.
7330 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7335 return do_sched_setscheduler(pid, policy, param);
7339 * sys_sched_setparam - set/change the RT priority of a thread
7340 * @pid: the pid in question.
7341 * @param: structure containing the new RT priority.
7343 * Return: 0 on success. An error code otherwise.
7345 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7347 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7351 * sys_sched_setattr - same as above, but with extended sched_attr
7352 * @pid: the pid in question.
7353 * @uattr: structure containing the extended parameters.
7354 * @flags: for future extension.
7356 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7357 unsigned int, flags)
7359 struct sched_attr attr;
7360 struct task_struct *p;
7363 if (!uattr || pid < 0 || flags)
7366 retval = sched_copy_attr(uattr, &attr);
7370 if ((int)attr.sched_policy < 0)
7372 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7373 attr.sched_policy = SETPARAM_POLICY;
7377 p = find_process_by_pid(pid);
7383 retval = sched_setattr(p, &attr);
7391 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7392 * @pid: the pid in question.
7394 * Return: On success, the policy of the thread. Otherwise, a negative error
7397 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7399 struct task_struct *p;
7407 p = find_process_by_pid(pid);
7409 retval = security_task_getscheduler(p);
7412 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7419 * sys_sched_getparam - get the RT priority of a thread
7420 * @pid: the pid in question.
7421 * @param: structure containing the RT priority.
7423 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7426 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7428 struct sched_param lp = { .sched_priority = 0 };
7429 struct task_struct *p;
7432 if (!param || pid < 0)
7436 p = find_process_by_pid(pid);
7441 retval = security_task_getscheduler(p);
7445 if (task_has_rt_policy(p))
7446 lp.sched_priority = p->rt_priority;
7450 * This one might sleep, we cannot do it with a spinlock held ...
7452 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7462 * Copy the kernel size attribute structure (which might be larger
7463 * than what user-space knows about) to user-space.
7465 * Note that all cases are valid: user-space buffer can be larger or
7466 * smaller than the kernel-space buffer. The usual case is that both
7467 * have the same size.
7470 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7471 struct sched_attr *kattr,
7474 unsigned int ksize = sizeof(*kattr);
7476 if (!access_ok(uattr, usize))
7480 * sched_getattr() ABI forwards and backwards compatibility:
7482 * If usize == ksize then we just copy everything to user-space and all is good.
7484 * If usize < ksize then we only copy as much as user-space has space for,
7485 * this keeps ABI compatibility as well. We skip the rest.
7487 * If usize > ksize then user-space is using a newer version of the ABI,
7488 * which part the kernel doesn't know about. Just ignore it - tooling can
7489 * detect the kernel's knowledge of attributes from the attr->size value
7490 * which is set to ksize in this case.
7492 kattr->size = min(usize, ksize);
7494 if (copy_to_user(uattr, kattr, kattr->size))
7501 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7502 * @pid: the pid in question.
7503 * @uattr: structure containing the extended parameters.
7504 * @usize: sizeof(attr) for fwd/bwd comp.
7505 * @flags: for future extension.
7507 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7508 unsigned int, usize, unsigned int, flags)
7510 struct sched_attr kattr = { };
7511 struct task_struct *p;
7514 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7515 usize < SCHED_ATTR_SIZE_VER0 || flags)
7519 p = find_process_by_pid(pid);
7524 retval = security_task_getscheduler(p);
7528 kattr.sched_policy = p->policy;
7529 if (p->sched_reset_on_fork)
7530 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7531 if (task_has_dl_policy(p))
7532 __getparam_dl(p, &kattr);
7533 else if (task_has_rt_policy(p))
7534 kattr.sched_priority = p->rt_priority;
7536 kattr.sched_nice = task_nice(p);
7538 #ifdef CONFIG_UCLAMP_TASK
7540 * This could race with another potential updater, but this is fine
7541 * because it'll correctly read the old or the new value. We don't need
7542 * to guarantee who wins the race as long as it doesn't return garbage.
7544 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7545 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7550 return sched_attr_copy_to_user(uattr, &kattr, usize);
7557 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7559 cpumask_var_t cpus_allowed, new_mask;
7560 struct task_struct *p;
7565 p = find_process_by_pid(pid);
7571 /* Prevent p going away */
7575 if (p->flags & PF_NO_SETAFFINITY) {
7579 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
7583 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7585 goto out_free_cpus_allowed;
7588 if (!check_same_owner(p)) {
7590 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7592 goto out_free_new_mask;
7597 retval = security_task_setscheduler(p);
7599 goto out_free_new_mask;
7602 cpuset_cpus_allowed(p, cpus_allowed);
7603 cpumask_and(new_mask, in_mask, cpus_allowed);
7606 * Since bandwidth control happens on root_domain basis,
7607 * if admission test is enabled, we only admit -deadline
7608 * tasks allowed to run on all the CPUs in the task's
7612 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
7614 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
7617 goto out_free_new_mask;
7623 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
7626 cpuset_cpus_allowed(p, cpus_allowed);
7627 if (!cpumask_subset(new_mask, cpus_allowed)) {
7629 * We must have raced with a concurrent cpuset
7630 * update. Just reset the cpus_allowed to the
7631 * cpuset's cpus_allowed
7633 cpumask_copy(new_mask, cpus_allowed);
7638 free_cpumask_var(new_mask);
7639 out_free_cpus_allowed:
7640 free_cpumask_var(cpus_allowed);
7646 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7647 struct cpumask *new_mask)
7649 if (len < cpumask_size())
7650 cpumask_clear(new_mask);
7651 else if (len > cpumask_size())
7652 len = cpumask_size();
7654 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
7658 * sys_sched_setaffinity - set the CPU affinity of a process
7659 * @pid: pid of the process
7660 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7661 * @user_mask_ptr: user-space pointer to the new CPU mask
7663 * Return: 0 on success. An error code otherwise.
7665 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
7666 unsigned long __user *, user_mask_ptr)
7668 cpumask_var_t new_mask;
7671 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
7674 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
7676 retval = sched_setaffinity(pid, new_mask);
7677 free_cpumask_var(new_mask);
7681 long sched_getaffinity(pid_t pid, struct cpumask *mask)
7683 struct task_struct *p;
7684 unsigned long flags;
7690 p = find_process_by_pid(pid);
7694 retval = security_task_getscheduler(p);
7698 raw_spin_lock_irqsave(&p->pi_lock, flags);
7699 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
7700 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
7709 * sys_sched_getaffinity - get the CPU affinity of a process
7710 * @pid: pid of the process
7711 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7712 * @user_mask_ptr: user-space pointer to hold the current CPU mask
7714 * Return: size of CPU mask copied to user_mask_ptr on success. An
7715 * error code otherwise.
7717 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
7718 unsigned long __user *, user_mask_ptr)
7723 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
7725 if (len & (sizeof(unsigned long)-1))
7728 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
7731 ret = sched_getaffinity(pid, mask);
7733 unsigned int retlen = min(len, cpumask_size());
7735 if (copy_to_user(user_mask_ptr, mask, retlen))
7740 free_cpumask_var(mask);
7745 static void do_sched_yield(void)
7750 rq = this_rq_lock_irq(&rf);
7752 schedstat_inc(rq->yld_count);
7753 current->sched_class->yield_task(rq);
7756 rq_unlock_irq(rq, &rf);
7757 sched_preempt_enable_no_resched();
7763 * sys_sched_yield - yield the current processor to other threads.
7765 * This function yields the current CPU to other tasks. If there are no
7766 * other threads running on this CPU then this function will return.
7770 SYSCALL_DEFINE0(sched_yield)
7776 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
7777 int __sched __cond_resched(void)
7779 if (should_resched(0)) {
7780 preempt_schedule_common();
7783 #ifndef CONFIG_PREEMPT_RCU
7788 EXPORT_SYMBOL(__cond_resched);
7791 #ifdef CONFIG_PREEMPT_DYNAMIC
7792 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
7793 EXPORT_STATIC_CALL_TRAMP(cond_resched);
7795 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
7796 EXPORT_STATIC_CALL_TRAMP(might_resched);
7800 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7801 * call schedule, and on return reacquire the lock.
7803 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7804 * operations here to prevent schedule() from being called twice (once via
7805 * spin_unlock(), once by hand).
7807 int __cond_resched_lock(spinlock_t *lock)
7809 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7812 lockdep_assert_held(lock);
7814 if (spin_needbreak(lock) || resched) {
7817 preempt_schedule_common();
7825 EXPORT_SYMBOL(__cond_resched_lock);
7827 int __cond_resched_rwlock_read(rwlock_t *lock)
7829 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7832 lockdep_assert_held_read(lock);
7834 if (rwlock_needbreak(lock) || resched) {
7837 preempt_schedule_common();
7845 EXPORT_SYMBOL(__cond_resched_rwlock_read);
7847 int __cond_resched_rwlock_write(rwlock_t *lock)
7849 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7852 lockdep_assert_held_write(lock);
7854 if (rwlock_needbreak(lock) || resched) {
7857 preempt_schedule_common();
7865 EXPORT_SYMBOL(__cond_resched_rwlock_write);
7868 * yield - yield the current processor to other threads.
7870 * Do not ever use this function, there's a 99% chance you're doing it wrong.
7872 * The scheduler is at all times free to pick the calling task as the most
7873 * eligible task to run, if removing the yield() call from your code breaks
7874 * it, it's already broken.
7876 * Typical broken usage is:
7881 * where one assumes that yield() will let 'the other' process run that will
7882 * make event true. If the current task is a SCHED_FIFO task that will never
7883 * happen. Never use yield() as a progress guarantee!!
7885 * If you want to use yield() to wait for something, use wait_event().
7886 * If you want to use yield() to be 'nice' for others, use cond_resched().
7887 * If you still want to use yield(), do not!
7889 void __sched yield(void)
7891 set_current_state(TASK_RUNNING);
7894 EXPORT_SYMBOL(yield);
7897 * yield_to - yield the current processor to another thread in
7898 * your thread group, or accelerate that thread toward the
7899 * processor it's on.
7901 * @preempt: whether task preemption is allowed or not
7903 * It's the caller's job to ensure that the target task struct
7904 * can't go away on us before we can do any checks.
7907 * true (>0) if we indeed boosted the target task.
7908 * false (0) if we failed to boost the target.
7909 * -ESRCH if there's no task to yield to.
7911 int __sched yield_to(struct task_struct *p, bool preempt)
7913 struct task_struct *curr = current;
7914 struct rq *rq, *p_rq;
7915 unsigned long flags;
7918 local_irq_save(flags);
7924 * If we're the only runnable task on the rq and target rq also
7925 * has only one task, there's absolutely no point in yielding.
7927 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
7932 double_rq_lock(rq, p_rq);
7933 if (task_rq(p) != p_rq) {
7934 double_rq_unlock(rq, p_rq);
7938 if (!curr->sched_class->yield_to_task)
7941 if (curr->sched_class != p->sched_class)
7944 if (task_running(p_rq, p) || p->state)
7947 yielded = curr->sched_class->yield_to_task(rq, p);
7949 schedstat_inc(rq->yld_count);
7951 * Make p's CPU reschedule; pick_next_entity takes care of
7954 if (preempt && rq != p_rq)
7959 double_rq_unlock(rq, p_rq);
7961 local_irq_restore(flags);
7968 EXPORT_SYMBOL_GPL(yield_to);
7970 int io_schedule_prepare(void)
7972 int old_iowait = current->in_iowait;
7974 current->in_iowait = 1;
7975 blk_schedule_flush_plug(current);
7980 void io_schedule_finish(int token)
7982 current->in_iowait = token;
7986 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7987 * that process accounting knows that this is a task in IO wait state.
7989 long __sched io_schedule_timeout(long timeout)
7994 token = io_schedule_prepare();
7995 ret = schedule_timeout(timeout);
7996 io_schedule_finish(token);
8000 EXPORT_SYMBOL(io_schedule_timeout);
8002 void __sched io_schedule(void)
8006 token = io_schedule_prepare();
8008 io_schedule_finish(token);
8010 EXPORT_SYMBOL(io_schedule);
8013 * sys_sched_get_priority_max - return maximum RT priority.
8014 * @policy: scheduling class.
8016 * Return: On success, this syscall returns the maximum
8017 * rt_priority that can be used by a given scheduling class.
8018 * On failure, a negative error code is returned.
8020 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8027 ret = MAX_RT_PRIO-1;
8029 case SCHED_DEADLINE:
8040 * sys_sched_get_priority_min - return minimum RT priority.
8041 * @policy: scheduling class.
8043 * Return: On success, this syscall returns the minimum
8044 * rt_priority that can be used by a given scheduling class.
8045 * On failure, a negative error code is returned.
8047 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8056 case SCHED_DEADLINE:
8065 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8067 struct task_struct *p;
8068 unsigned int time_slice;
8078 p = find_process_by_pid(pid);
8082 retval = security_task_getscheduler(p);
8086 rq = task_rq_lock(p, &rf);
8088 if (p->sched_class->get_rr_interval)
8089 time_slice = p->sched_class->get_rr_interval(rq, p);
8090 task_rq_unlock(rq, p, &rf);
8093 jiffies_to_timespec64(time_slice, t);
8102 * sys_sched_rr_get_interval - return the default timeslice of a process.
8103 * @pid: pid of the process.
8104 * @interval: userspace pointer to the timeslice value.
8106 * this syscall writes the default timeslice value of a given process
8107 * into the user-space timespec buffer. A value of '0' means infinity.
8109 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8112 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8113 struct __kernel_timespec __user *, interval)
8115 struct timespec64 t;
8116 int retval = sched_rr_get_interval(pid, &t);
8119 retval = put_timespec64(&t, interval);
8124 #ifdef CONFIG_COMPAT_32BIT_TIME
8125 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8126 struct old_timespec32 __user *, interval)
8128 struct timespec64 t;
8129 int retval = sched_rr_get_interval(pid, &t);
8132 retval = put_old_timespec32(&t, interval);
8137 void sched_show_task(struct task_struct *p)
8139 unsigned long free = 0;
8142 if (!try_get_task_stack(p))
8145 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8147 if (p->state == TASK_RUNNING)
8148 pr_cont(" running task ");
8149 #ifdef CONFIG_DEBUG_STACK_USAGE
8150 free = stack_not_used(p);
8155 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8157 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8158 free, task_pid_nr(p), ppid,
8159 (unsigned long)task_thread_info(p)->flags);
8161 print_worker_info(KERN_INFO, p);
8162 print_stop_info(KERN_INFO, p);
8163 show_stack(p, NULL, KERN_INFO);
8166 EXPORT_SYMBOL_GPL(sched_show_task);
8169 state_filter_match(unsigned long state_filter, struct task_struct *p)
8171 /* no filter, everything matches */
8175 /* filter, but doesn't match */
8176 if (!(p->state & state_filter))
8180 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8183 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
8190 void show_state_filter(unsigned long state_filter)
8192 struct task_struct *g, *p;
8195 for_each_process_thread(g, p) {
8197 * reset the NMI-timeout, listing all files on a slow
8198 * console might take a lot of time:
8199 * Also, reset softlockup watchdogs on all CPUs, because
8200 * another CPU might be blocked waiting for us to process
8203 touch_nmi_watchdog();
8204 touch_all_softlockup_watchdogs();
8205 if (state_filter_match(state_filter, p))
8209 #ifdef CONFIG_SCHED_DEBUG
8211 sysrq_sched_debug_show();
8215 * Only show locks if all tasks are dumped:
8218 debug_show_all_locks();
8222 * init_idle - set up an idle thread for a given CPU
8223 * @idle: task in question
8224 * @cpu: CPU the idle task belongs to
8226 * NOTE: this function does not set the idle thread's NEED_RESCHED
8227 * flag, to make booting more robust.
8229 void init_idle(struct task_struct *idle, int cpu)
8231 struct rq *rq = cpu_rq(cpu);
8232 unsigned long flags;
8234 __sched_fork(0, idle);
8236 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8237 raw_spin_rq_lock(rq);
8239 idle->state = TASK_RUNNING;
8240 idle->se.exec_start = sched_clock();
8241 idle->flags |= PF_IDLE;
8243 scs_task_reset(idle);
8244 kasan_unpoison_task_stack(idle);
8248 * It's possible that init_idle() gets called multiple times on a task,
8249 * in that case do_set_cpus_allowed() will not do the right thing.
8251 * And since this is boot we can forgo the serialization.
8253 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8256 * We're having a chicken and egg problem, even though we are
8257 * holding rq->lock, the CPU isn't yet set to this CPU so the
8258 * lockdep check in task_group() will fail.
8260 * Similar case to sched_fork(). / Alternatively we could
8261 * use task_rq_lock() here and obtain the other rq->lock.
8266 __set_task_cpu(idle, cpu);
8270 rcu_assign_pointer(rq->curr, idle);
8271 idle->on_rq = TASK_ON_RQ_QUEUED;
8275 raw_spin_rq_unlock(rq);
8276 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8278 /* Set the preempt count _outside_ the spinlocks! */
8279 init_idle_preempt_count(idle, cpu);
8282 * The idle tasks have their own, simple scheduling class:
8284 idle->sched_class = &idle_sched_class;
8285 ftrace_graph_init_idle_task(idle, cpu);
8286 vtime_init_idle(idle, cpu);
8288 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8294 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8295 const struct cpumask *trial)
8299 if (!cpumask_weight(cur))
8302 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8307 int task_can_attach(struct task_struct *p,
8308 const struct cpumask *cs_cpus_allowed)
8313 * Kthreads which disallow setaffinity shouldn't be moved
8314 * to a new cpuset; we don't want to change their CPU
8315 * affinity and isolating such threads by their set of
8316 * allowed nodes is unnecessary. Thus, cpusets are not
8317 * applicable for such threads. This prevents checking for
8318 * success of set_cpus_allowed_ptr() on all attached tasks
8319 * before cpus_mask may be changed.
8321 if (p->flags & PF_NO_SETAFFINITY) {
8326 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8328 ret = dl_task_can_attach(p, cs_cpus_allowed);
8334 bool sched_smp_initialized __read_mostly;
8336 #ifdef CONFIG_NUMA_BALANCING
8337 /* Migrate current task p to target_cpu */
8338 int migrate_task_to(struct task_struct *p, int target_cpu)
8340 struct migration_arg arg = { p, target_cpu };
8341 int curr_cpu = task_cpu(p);
8343 if (curr_cpu == target_cpu)
8346 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8349 /* TODO: This is not properly updating schedstats */
8351 trace_sched_move_numa(p, curr_cpu, target_cpu);
8352 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8356 * Requeue a task on a given node and accurately track the number of NUMA
8357 * tasks on the runqueues
8359 void sched_setnuma(struct task_struct *p, int nid)
8361 bool queued, running;
8365 rq = task_rq_lock(p, &rf);
8366 queued = task_on_rq_queued(p);
8367 running = task_current(rq, p);
8370 dequeue_task(rq, p, DEQUEUE_SAVE);
8372 put_prev_task(rq, p);
8374 p->numa_preferred_nid = nid;
8377 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8379 set_next_task(rq, p);
8380 task_rq_unlock(rq, p, &rf);
8382 #endif /* CONFIG_NUMA_BALANCING */
8384 #ifdef CONFIG_HOTPLUG_CPU
8386 * Ensure that the idle task is using init_mm right before its CPU goes
8389 void idle_task_exit(void)
8391 struct mm_struct *mm = current->active_mm;
8393 BUG_ON(cpu_online(smp_processor_id()));
8394 BUG_ON(current != this_rq()->idle);
8396 if (mm != &init_mm) {
8397 switch_mm(mm, &init_mm, current);
8398 finish_arch_post_lock_switch();
8401 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8404 static int __balance_push_cpu_stop(void *arg)
8406 struct task_struct *p = arg;
8407 struct rq *rq = this_rq();
8411 raw_spin_lock_irq(&p->pi_lock);
8414 update_rq_clock(rq);
8416 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8417 cpu = select_fallback_rq(rq->cpu, p);
8418 rq = __migrate_task(rq, &rf, p, cpu);
8422 raw_spin_unlock_irq(&p->pi_lock);
8429 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8432 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8434 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8435 * effective when the hotplug motion is down.
8437 static void balance_push(struct rq *rq)
8439 struct task_struct *push_task = rq->curr;
8441 lockdep_assert_rq_held(rq);
8442 SCHED_WARN_ON(rq->cpu != smp_processor_id());
8445 * Ensure the thing is persistent until balance_push_set(.on = false);
8447 rq->balance_callback = &balance_push_callback;
8450 * Only active while going offline.
8452 if (!cpu_dying(rq->cpu))
8456 * Both the cpu-hotplug and stop task are in this case and are
8457 * required to complete the hotplug process.
8459 * XXX: the idle task does not match kthread_is_per_cpu() due to
8460 * histerical raisins.
8462 if (rq->idle == push_task ||
8463 kthread_is_per_cpu(push_task) ||
8464 is_migration_disabled(push_task)) {
8467 * If this is the idle task on the outgoing CPU try to wake
8468 * up the hotplug control thread which might wait for the
8469 * last task to vanish. The rcuwait_active() check is
8470 * accurate here because the waiter is pinned on this CPU
8471 * and can't obviously be running in parallel.
8473 * On RT kernels this also has to check whether there are
8474 * pinned and scheduled out tasks on the runqueue. They
8475 * need to leave the migrate disabled section first.
8477 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8478 rcuwait_active(&rq->hotplug_wait)) {
8479 raw_spin_rq_unlock(rq);
8480 rcuwait_wake_up(&rq->hotplug_wait);
8481 raw_spin_rq_lock(rq);
8486 get_task_struct(push_task);
8488 * Temporarily drop rq->lock such that we can wake-up the stop task.
8489 * Both preemption and IRQs are still disabled.
8491 raw_spin_rq_unlock(rq);
8492 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8493 this_cpu_ptr(&push_work));
8495 * At this point need_resched() is true and we'll take the loop in
8496 * schedule(). The next pick is obviously going to be the stop task
8497 * which kthread_is_per_cpu() and will push this task away.
8499 raw_spin_rq_lock(rq);
8502 static void balance_push_set(int cpu, bool on)
8504 struct rq *rq = cpu_rq(cpu);
8507 rq_lock_irqsave(rq, &rf);
8509 WARN_ON_ONCE(rq->balance_callback);
8510 rq->balance_callback = &balance_push_callback;
8511 } else if (rq->balance_callback == &balance_push_callback) {
8512 rq->balance_callback = NULL;
8514 rq_unlock_irqrestore(rq, &rf);
8518 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8519 * inactive. All tasks which are not per CPU kernel threads are either
8520 * pushed off this CPU now via balance_push() or placed on a different CPU
8521 * during wakeup. Wait until the CPU is quiescent.
8523 static void balance_hotplug_wait(void)
8525 struct rq *rq = this_rq();
8527 rcuwait_wait_event(&rq->hotplug_wait,
8528 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8529 TASK_UNINTERRUPTIBLE);
8534 static inline void balance_push(struct rq *rq)
8538 static inline void balance_push_set(int cpu, bool on)
8542 static inline void balance_hotplug_wait(void)
8546 #endif /* CONFIG_HOTPLUG_CPU */
8548 void set_rq_online(struct rq *rq)
8551 const struct sched_class *class;
8553 cpumask_set_cpu(rq->cpu, rq->rd->online);
8556 for_each_class(class) {
8557 if (class->rq_online)
8558 class->rq_online(rq);
8563 void set_rq_offline(struct rq *rq)
8566 const struct sched_class *class;
8568 for_each_class(class) {
8569 if (class->rq_offline)
8570 class->rq_offline(rq);
8573 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8579 * used to mark begin/end of suspend/resume:
8581 static int num_cpus_frozen;
8584 * Update cpusets according to cpu_active mask. If cpusets are
8585 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8586 * around partition_sched_domains().
8588 * If we come here as part of a suspend/resume, don't touch cpusets because we
8589 * want to restore it back to its original state upon resume anyway.
8591 static void cpuset_cpu_active(void)
8593 if (cpuhp_tasks_frozen) {
8595 * num_cpus_frozen tracks how many CPUs are involved in suspend
8596 * resume sequence. As long as this is not the last online
8597 * operation in the resume sequence, just build a single sched
8598 * domain, ignoring cpusets.
8600 partition_sched_domains(1, NULL, NULL);
8601 if (--num_cpus_frozen)
8604 * This is the last CPU online operation. So fall through and
8605 * restore the original sched domains by considering the
8606 * cpuset configurations.
8608 cpuset_force_rebuild();
8610 cpuset_update_active_cpus();
8613 static int cpuset_cpu_inactive(unsigned int cpu)
8615 if (!cpuhp_tasks_frozen) {
8616 if (dl_cpu_busy(cpu))
8618 cpuset_update_active_cpus();
8621 partition_sched_domains(1, NULL, NULL);
8626 int sched_cpu_activate(unsigned int cpu)
8628 struct rq *rq = cpu_rq(cpu);
8632 * Clear the balance_push callback and prepare to schedule
8635 balance_push_set(cpu, false);
8637 #ifdef CONFIG_SCHED_SMT
8639 * When going up, increment the number of cores with SMT present.
8641 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8642 static_branch_inc_cpuslocked(&sched_smt_present);
8644 set_cpu_active(cpu, true);
8646 if (sched_smp_initialized) {
8647 sched_domains_numa_masks_set(cpu);
8648 cpuset_cpu_active();
8652 * Put the rq online, if not already. This happens:
8654 * 1) In the early boot process, because we build the real domains
8655 * after all CPUs have been brought up.
8657 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
8660 rq_lock_irqsave(rq, &rf);
8662 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8665 rq_unlock_irqrestore(rq, &rf);
8670 int sched_cpu_deactivate(unsigned int cpu)
8672 struct rq *rq = cpu_rq(cpu);
8677 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
8678 * load balancing when not active
8680 nohz_balance_exit_idle(rq);
8682 set_cpu_active(cpu, false);
8685 * From this point forward, this CPU will refuse to run any task that
8686 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
8687 * push those tasks away until this gets cleared, see
8688 * sched_cpu_dying().
8690 balance_push_set(cpu, true);
8693 * We've cleared cpu_active_mask / set balance_push, wait for all
8694 * preempt-disabled and RCU users of this state to go away such that
8695 * all new such users will observe it.
8697 * Specifically, we rely on ttwu to no longer target this CPU, see
8698 * ttwu_queue_cond() and is_cpu_allowed().
8700 * Do sync before park smpboot threads to take care the rcu boost case.
8704 rq_lock_irqsave(rq, &rf);
8706 update_rq_clock(rq);
8707 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8710 rq_unlock_irqrestore(rq, &rf);
8712 #ifdef CONFIG_SCHED_SMT
8714 * When going down, decrement the number of cores with SMT present.
8716 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8717 static_branch_dec_cpuslocked(&sched_smt_present);
8720 if (!sched_smp_initialized)
8723 ret = cpuset_cpu_inactive(cpu);
8725 balance_push_set(cpu, false);
8726 set_cpu_active(cpu, true);
8729 sched_domains_numa_masks_clear(cpu);
8733 static void sched_rq_cpu_starting(unsigned int cpu)
8735 struct rq *rq = cpu_rq(cpu);
8737 rq->calc_load_update = calc_load_update;
8738 update_max_interval();
8741 int sched_cpu_starting(unsigned int cpu)
8743 sched_core_cpu_starting(cpu);
8744 sched_rq_cpu_starting(cpu);
8745 sched_tick_start(cpu);
8749 #ifdef CONFIG_HOTPLUG_CPU
8752 * Invoked immediately before the stopper thread is invoked to bring the
8753 * CPU down completely. At this point all per CPU kthreads except the
8754 * hotplug thread (current) and the stopper thread (inactive) have been
8755 * either parked or have been unbound from the outgoing CPU. Ensure that
8756 * any of those which might be on the way out are gone.
8758 * If after this point a bound task is being woken on this CPU then the
8759 * responsible hotplug callback has failed to do it's job.
8760 * sched_cpu_dying() will catch it with the appropriate fireworks.
8762 int sched_cpu_wait_empty(unsigned int cpu)
8764 balance_hotplug_wait();
8769 * Since this CPU is going 'away' for a while, fold any nr_active delta we
8770 * might have. Called from the CPU stopper task after ensuring that the
8771 * stopper is the last running task on the CPU, so nr_active count is
8772 * stable. We need to take the teardown thread which is calling this into
8773 * account, so we hand in adjust = 1 to the load calculation.
8775 * Also see the comment "Global load-average calculations".
8777 static void calc_load_migrate(struct rq *rq)
8779 long delta = calc_load_fold_active(rq, 1);
8782 atomic_long_add(delta, &calc_load_tasks);
8785 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
8787 struct task_struct *g, *p;
8788 int cpu = cpu_of(rq);
8790 lockdep_assert_rq_held(rq);
8792 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
8793 for_each_process_thread(g, p) {
8794 if (task_cpu(p) != cpu)
8797 if (!task_on_rq_queued(p))
8800 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
8804 int sched_cpu_dying(unsigned int cpu)
8806 struct rq *rq = cpu_rq(cpu);
8809 /* Handle pending wakeups and then migrate everything off */
8810 sched_tick_stop(cpu);
8812 rq_lock_irqsave(rq, &rf);
8813 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
8814 WARN(true, "Dying CPU not properly vacated!");
8815 dump_rq_tasks(rq, KERN_WARNING);
8817 rq_unlock_irqrestore(rq, &rf);
8819 calc_load_migrate(rq);
8820 update_max_interval();
8826 void __init sched_init_smp(void)
8831 * There's no userspace yet to cause hotplug operations; hence all the
8832 * CPU masks are stable and all blatant races in the below code cannot
8835 mutex_lock(&sched_domains_mutex);
8836 sched_init_domains(cpu_active_mask);
8837 mutex_unlock(&sched_domains_mutex);
8839 /* Move init over to a non-isolated CPU */
8840 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
8842 sched_init_granularity();
8844 init_sched_rt_class();
8845 init_sched_dl_class();
8847 sched_smp_initialized = true;
8850 static int __init migration_init(void)
8852 sched_cpu_starting(smp_processor_id());
8855 early_initcall(migration_init);
8858 void __init sched_init_smp(void)
8860 sched_init_granularity();
8862 #endif /* CONFIG_SMP */
8864 int in_sched_functions(unsigned long addr)
8866 return in_lock_functions(addr) ||
8867 (addr >= (unsigned long)__sched_text_start
8868 && addr < (unsigned long)__sched_text_end);
8871 #ifdef CONFIG_CGROUP_SCHED
8873 * Default task group.
8874 * Every task in system belongs to this group at bootup.
8876 struct task_group root_task_group;
8877 LIST_HEAD(task_groups);
8879 /* Cacheline aligned slab cache for task_group */
8880 static struct kmem_cache *task_group_cache __read_mostly;
8883 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
8884 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
8886 void __init sched_init(void)
8888 unsigned long ptr = 0;
8891 /* Make sure the linker didn't screw up */
8892 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
8893 &fair_sched_class + 1 != &rt_sched_class ||
8894 &rt_sched_class + 1 != &dl_sched_class);
8896 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
8901 #ifdef CONFIG_FAIR_GROUP_SCHED
8902 ptr += 2 * nr_cpu_ids * sizeof(void **);
8904 #ifdef CONFIG_RT_GROUP_SCHED
8905 ptr += 2 * nr_cpu_ids * sizeof(void **);
8908 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
8910 #ifdef CONFIG_FAIR_GROUP_SCHED
8911 root_task_group.se = (struct sched_entity **)ptr;
8912 ptr += nr_cpu_ids * sizeof(void **);
8914 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8915 ptr += nr_cpu_ids * sizeof(void **);
8917 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
8918 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8919 #endif /* CONFIG_FAIR_GROUP_SCHED */
8920 #ifdef CONFIG_RT_GROUP_SCHED
8921 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8922 ptr += nr_cpu_ids * sizeof(void **);
8924 root_task_group.rt_rq = (struct rt_rq **)ptr;
8925 ptr += nr_cpu_ids * sizeof(void **);
8927 #endif /* CONFIG_RT_GROUP_SCHED */
8929 #ifdef CONFIG_CPUMASK_OFFSTACK
8930 for_each_possible_cpu(i) {
8931 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
8932 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8933 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
8934 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8936 #endif /* CONFIG_CPUMASK_OFFSTACK */
8938 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
8939 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
8942 init_defrootdomain();
8945 #ifdef CONFIG_RT_GROUP_SCHED
8946 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8947 global_rt_period(), global_rt_runtime());
8948 #endif /* CONFIG_RT_GROUP_SCHED */
8950 #ifdef CONFIG_CGROUP_SCHED
8951 task_group_cache = KMEM_CACHE(task_group, 0);
8953 list_add(&root_task_group.list, &task_groups);
8954 INIT_LIST_HEAD(&root_task_group.children);
8955 INIT_LIST_HEAD(&root_task_group.siblings);
8956 autogroup_init(&init_task);
8957 #endif /* CONFIG_CGROUP_SCHED */
8959 for_each_possible_cpu(i) {
8963 raw_spin_lock_init(&rq->__lock);
8965 rq->calc_load_active = 0;
8966 rq->calc_load_update = jiffies + LOAD_FREQ;
8967 init_cfs_rq(&rq->cfs);
8968 init_rt_rq(&rq->rt);
8969 init_dl_rq(&rq->dl);
8970 #ifdef CONFIG_FAIR_GROUP_SCHED
8971 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8972 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
8974 * How much CPU bandwidth does root_task_group get?
8976 * In case of task-groups formed thr' the cgroup filesystem, it
8977 * gets 100% of the CPU resources in the system. This overall
8978 * system CPU resource is divided among the tasks of
8979 * root_task_group and its child task-groups in a fair manner,
8980 * based on each entity's (task or task-group's) weight
8981 * (se->load.weight).
8983 * In other words, if root_task_group has 10 tasks of weight
8984 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8985 * then A0's share of the CPU resource is:
8987 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8989 * We achieve this by letting root_task_group's tasks sit
8990 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8992 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8993 #endif /* CONFIG_FAIR_GROUP_SCHED */
8995 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8996 #ifdef CONFIG_RT_GROUP_SCHED
8997 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9002 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9003 rq->balance_callback = &balance_push_callback;
9004 rq->active_balance = 0;
9005 rq->next_balance = jiffies;
9010 rq->avg_idle = 2*sysctl_sched_migration_cost;
9011 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9013 INIT_LIST_HEAD(&rq->cfs_tasks);
9015 rq_attach_root(rq, &def_root_domain);
9016 #ifdef CONFIG_NO_HZ_COMMON
9017 rq->last_blocked_load_update_tick = jiffies;
9018 atomic_set(&rq->nohz_flags, 0);
9020 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9022 #ifdef CONFIG_HOTPLUG_CPU
9023 rcuwait_init(&rq->hotplug_wait);
9025 #endif /* CONFIG_SMP */
9027 atomic_set(&rq->nr_iowait, 0);
9029 #ifdef CONFIG_SCHED_CORE
9031 rq->core_pick = NULL;
9032 rq->core_enabled = 0;
9033 rq->core_tree = RB_ROOT;
9034 rq->core_forceidle = false;
9036 rq->core_cookie = 0UL;
9040 set_load_weight(&init_task, false);
9043 * The boot idle thread does lazy MMU switching as well:
9046 enter_lazy_tlb(&init_mm, current);
9049 * Make us the idle thread. Technically, schedule() should not be
9050 * called from this thread, however somewhere below it might be,
9051 * but because we are the idle thread, we just pick up running again
9052 * when this runqueue becomes "idle".
9054 init_idle(current, smp_processor_id());
9056 calc_load_update = jiffies + LOAD_FREQ;
9059 idle_thread_set_boot_cpu();
9060 balance_push_set(smp_processor_id(), false);
9062 init_sched_fair_class();
9070 scheduler_running = 1;
9073 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9074 static inline int preempt_count_equals(int preempt_offset)
9076 int nested = preempt_count() + rcu_preempt_depth();
9078 return (nested == preempt_offset);
9081 void __might_sleep(const char *file, int line, int preempt_offset)
9084 * Blocking primitives will set (and therefore destroy) current->state,
9085 * since we will exit with TASK_RUNNING make sure we enter with it,
9086 * otherwise we will destroy state.
9088 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
9089 "do not call blocking ops when !TASK_RUNNING; "
9090 "state=%lx set at [<%p>] %pS\n",
9092 (void *)current->task_state_change,
9093 (void *)current->task_state_change);
9095 ___might_sleep(file, line, preempt_offset);
9097 EXPORT_SYMBOL(__might_sleep);
9099 void ___might_sleep(const char *file, int line, int preempt_offset)
9101 /* Ratelimiting timestamp: */
9102 static unsigned long prev_jiffy;
9104 unsigned long preempt_disable_ip;
9106 /* WARN_ON_ONCE() by default, no rate limit required: */
9109 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
9110 !is_idle_task(current) && !current->non_block_count) ||
9111 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9115 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9117 prev_jiffy = jiffies;
9119 /* Save this before calling printk(), since that will clobber it: */
9120 preempt_disable_ip = get_preempt_disable_ip(current);
9123 "BUG: sleeping function called from invalid context at %s:%d\n",
9126 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9127 in_atomic(), irqs_disabled(), current->non_block_count,
9128 current->pid, current->comm);
9130 if (task_stack_end_corrupted(current))
9131 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
9133 debug_show_held_locks(current);
9134 if (irqs_disabled())
9135 print_irqtrace_events(current);
9136 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
9137 && !preempt_count_equals(preempt_offset)) {
9138 pr_err("Preemption disabled at:");
9139 print_ip_sym(KERN_ERR, preempt_disable_ip);
9142 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9144 EXPORT_SYMBOL(___might_sleep);
9146 void __cant_sleep(const char *file, int line, int preempt_offset)
9148 static unsigned long prev_jiffy;
9150 if (irqs_disabled())
9153 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9156 if (preempt_count() > preempt_offset)
9159 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9161 prev_jiffy = jiffies;
9163 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9164 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9165 in_atomic(), irqs_disabled(),
9166 current->pid, current->comm);
9168 debug_show_held_locks(current);
9170 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9172 EXPORT_SYMBOL_GPL(__cant_sleep);
9175 void __cant_migrate(const char *file, int line)
9177 static unsigned long prev_jiffy;
9179 if (irqs_disabled())
9182 if (is_migration_disabled(current))
9185 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9188 if (preempt_count() > 0)
9191 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9193 prev_jiffy = jiffies;
9195 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9196 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9197 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9198 current->pid, current->comm);
9200 debug_show_held_locks(current);
9202 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9204 EXPORT_SYMBOL_GPL(__cant_migrate);
9208 #ifdef CONFIG_MAGIC_SYSRQ
9209 void normalize_rt_tasks(void)
9211 struct task_struct *g, *p;
9212 struct sched_attr attr = {
9213 .sched_policy = SCHED_NORMAL,
9216 read_lock(&tasklist_lock);
9217 for_each_process_thread(g, p) {
9219 * Only normalize user tasks:
9221 if (p->flags & PF_KTHREAD)
9224 p->se.exec_start = 0;
9225 schedstat_set(p->se.statistics.wait_start, 0);
9226 schedstat_set(p->se.statistics.sleep_start, 0);
9227 schedstat_set(p->se.statistics.block_start, 0);
9229 if (!dl_task(p) && !rt_task(p)) {
9231 * Renice negative nice level userspace
9234 if (task_nice(p) < 0)
9235 set_user_nice(p, 0);
9239 __sched_setscheduler(p, &attr, false, false);
9241 read_unlock(&tasklist_lock);
9244 #endif /* CONFIG_MAGIC_SYSRQ */
9246 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9248 * These functions are only useful for the IA64 MCA handling, or kdb.
9250 * They can only be called when the whole system has been
9251 * stopped - every CPU needs to be quiescent, and no scheduling
9252 * activity can take place. Using them for anything else would
9253 * be a serious bug, and as a result, they aren't even visible
9254 * under any other configuration.
9258 * curr_task - return the current task for a given CPU.
9259 * @cpu: the processor in question.
9261 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9263 * Return: The current task for @cpu.
9265 struct task_struct *curr_task(int cpu)
9267 return cpu_curr(cpu);
9270 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9274 * ia64_set_curr_task - set the current task for a given CPU.
9275 * @cpu: the processor in question.
9276 * @p: the task pointer to set.
9278 * Description: This function must only be used when non-maskable interrupts
9279 * are serviced on a separate stack. It allows the architecture to switch the
9280 * notion of the current task on a CPU in a non-blocking manner. This function
9281 * must be called with all CPU's synchronized, and interrupts disabled, the
9282 * and caller must save the original value of the current task (see
9283 * curr_task() above) and restore that value before reenabling interrupts and
9284 * re-starting the system.
9286 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9288 void ia64_set_curr_task(int cpu, struct task_struct *p)
9295 #ifdef CONFIG_CGROUP_SCHED
9296 /* task_group_lock serializes the addition/removal of task groups */
9297 static DEFINE_SPINLOCK(task_group_lock);
9299 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9300 struct task_group *parent)
9302 #ifdef CONFIG_UCLAMP_TASK_GROUP
9303 enum uclamp_id clamp_id;
9305 for_each_clamp_id(clamp_id) {
9306 uclamp_se_set(&tg->uclamp_req[clamp_id],
9307 uclamp_none(clamp_id), false);
9308 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9313 static void sched_free_group(struct task_group *tg)
9315 free_fair_sched_group(tg);
9316 free_rt_sched_group(tg);
9318 kmem_cache_free(task_group_cache, tg);
9321 /* allocate runqueue etc for a new task group */
9322 struct task_group *sched_create_group(struct task_group *parent)
9324 struct task_group *tg;
9326 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9328 return ERR_PTR(-ENOMEM);
9330 if (!alloc_fair_sched_group(tg, parent))
9333 if (!alloc_rt_sched_group(tg, parent))
9336 alloc_uclamp_sched_group(tg, parent);
9341 sched_free_group(tg);
9342 return ERR_PTR(-ENOMEM);
9345 void sched_online_group(struct task_group *tg, struct task_group *parent)
9347 unsigned long flags;
9349 spin_lock_irqsave(&task_group_lock, flags);
9350 list_add_rcu(&tg->list, &task_groups);
9352 /* Root should already exist: */
9355 tg->parent = parent;
9356 INIT_LIST_HEAD(&tg->children);
9357 list_add_rcu(&tg->siblings, &parent->children);
9358 spin_unlock_irqrestore(&task_group_lock, flags);
9360 online_fair_sched_group(tg);
9363 /* rcu callback to free various structures associated with a task group */
9364 static void sched_free_group_rcu(struct rcu_head *rhp)
9366 /* Now it should be safe to free those cfs_rqs: */
9367 sched_free_group(container_of(rhp, struct task_group, rcu));
9370 void sched_destroy_group(struct task_group *tg)
9372 /* Wait for possible concurrent references to cfs_rqs complete: */
9373 call_rcu(&tg->rcu, sched_free_group_rcu);
9376 void sched_offline_group(struct task_group *tg)
9378 unsigned long flags;
9380 /* End participation in shares distribution: */
9381 unregister_fair_sched_group(tg);
9383 spin_lock_irqsave(&task_group_lock, flags);
9384 list_del_rcu(&tg->list);
9385 list_del_rcu(&tg->siblings);
9386 spin_unlock_irqrestore(&task_group_lock, flags);
9389 static void sched_change_group(struct task_struct *tsk, int type)
9391 struct task_group *tg;
9394 * All callers are synchronized by task_rq_lock(); we do not use RCU
9395 * which is pointless here. Thus, we pass "true" to task_css_check()
9396 * to prevent lockdep warnings.
9398 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9399 struct task_group, css);
9400 tg = autogroup_task_group(tsk, tg);
9401 tsk->sched_task_group = tg;
9403 #ifdef CONFIG_FAIR_GROUP_SCHED
9404 if (tsk->sched_class->task_change_group)
9405 tsk->sched_class->task_change_group(tsk, type);
9408 set_task_rq(tsk, task_cpu(tsk));
9412 * Change task's runqueue when it moves between groups.
9414 * The caller of this function should have put the task in its new group by
9415 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9418 void sched_move_task(struct task_struct *tsk)
9420 int queued, running, queue_flags =
9421 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9425 rq = task_rq_lock(tsk, &rf);
9426 update_rq_clock(rq);
9428 running = task_current(rq, tsk);
9429 queued = task_on_rq_queued(tsk);
9432 dequeue_task(rq, tsk, queue_flags);
9434 put_prev_task(rq, tsk);
9436 sched_change_group(tsk, TASK_MOVE_GROUP);
9439 enqueue_task(rq, tsk, queue_flags);
9441 set_next_task(rq, tsk);
9443 * After changing group, the running task may have joined a
9444 * throttled one but it's still the running task. Trigger a
9445 * resched to make sure that task can still run.
9450 task_rq_unlock(rq, tsk, &rf);
9453 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9455 return css ? container_of(css, struct task_group, css) : NULL;
9458 static struct cgroup_subsys_state *
9459 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9461 struct task_group *parent = css_tg(parent_css);
9462 struct task_group *tg;
9465 /* This is early initialization for the top cgroup */
9466 return &root_task_group.css;
9469 tg = sched_create_group(parent);
9471 return ERR_PTR(-ENOMEM);
9476 /* Expose task group only after completing cgroup initialization */
9477 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9479 struct task_group *tg = css_tg(css);
9480 struct task_group *parent = css_tg(css->parent);
9483 sched_online_group(tg, parent);
9485 #ifdef CONFIG_UCLAMP_TASK_GROUP
9486 /* Propagate the effective uclamp value for the new group */
9487 cpu_util_update_eff(css);
9493 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9495 struct task_group *tg = css_tg(css);
9497 sched_offline_group(tg);
9500 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9502 struct task_group *tg = css_tg(css);
9505 * Relies on the RCU grace period between css_released() and this.
9507 sched_free_group(tg);
9511 * This is called before wake_up_new_task(), therefore we really only
9512 * have to set its group bits, all the other stuff does not apply.
9514 static void cpu_cgroup_fork(struct task_struct *task)
9519 rq = task_rq_lock(task, &rf);
9521 update_rq_clock(rq);
9522 sched_change_group(task, TASK_SET_GROUP);
9524 task_rq_unlock(rq, task, &rf);
9527 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9529 struct task_struct *task;
9530 struct cgroup_subsys_state *css;
9533 cgroup_taskset_for_each(task, css, tset) {
9534 #ifdef CONFIG_RT_GROUP_SCHED
9535 if (!sched_rt_can_attach(css_tg(css), task))
9539 * Serialize against wake_up_new_task() such that if it's
9540 * running, we're sure to observe its full state.
9542 raw_spin_lock_irq(&task->pi_lock);
9544 * Avoid calling sched_move_task() before wake_up_new_task()
9545 * has happened. This would lead to problems with PELT, due to
9546 * move wanting to detach+attach while we're not attached yet.
9548 if (task->state == TASK_NEW)
9550 raw_spin_unlock_irq(&task->pi_lock);
9558 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9560 struct task_struct *task;
9561 struct cgroup_subsys_state *css;
9563 cgroup_taskset_for_each(task, css, tset)
9564 sched_move_task(task);
9567 #ifdef CONFIG_UCLAMP_TASK_GROUP
9568 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9570 struct cgroup_subsys_state *top_css = css;
9571 struct uclamp_se *uc_parent = NULL;
9572 struct uclamp_se *uc_se = NULL;
9573 unsigned int eff[UCLAMP_CNT];
9574 enum uclamp_id clamp_id;
9575 unsigned int clamps;
9577 css_for_each_descendant_pre(css, top_css) {
9578 uc_parent = css_tg(css)->parent
9579 ? css_tg(css)->parent->uclamp : NULL;
9581 for_each_clamp_id(clamp_id) {
9582 /* Assume effective clamps matches requested clamps */
9583 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
9584 /* Cap effective clamps with parent's effective clamps */
9586 eff[clamp_id] > uc_parent[clamp_id].value) {
9587 eff[clamp_id] = uc_parent[clamp_id].value;
9590 /* Ensure protection is always capped by limit */
9591 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
9593 /* Propagate most restrictive effective clamps */
9595 uc_se = css_tg(css)->uclamp;
9596 for_each_clamp_id(clamp_id) {
9597 if (eff[clamp_id] == uc_se[clamp_id].value)
9599 uc_se[clamp_id].value = eff[clamp_id];
9600 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
9601 clamps |= (0x1 << clamp_id);
9604 css = css_rightmost_descendant(css);
9608 /* Immediately update descendants RUNNABLE tasks */
9609 uclamp_update_active_tasks(css, clamps);
9614 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
9615 * C expression. Since there is no way to convert a macro argument (N) into a
9616 * character constant, use two levels of macros.
9618 #define _POW10(exp) ((unsigned int)1e##exp)
9619 #define POW10(exp) _POW10(exp)
9621 struct uclamp_request {
9622 #define UCLAMP_PERCENT_SHIFT 2
9623 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
9629 static inline struct uclamp_request
9630 capacity_from_percent(char *buf)
9632 struct uclamp_request req = {
9633 .percent = UCLAMP_PERCENT_SCALE,
9634 .util = SCHED_CAPACITY_SCALE,
9639 if (strcmp(buf, "max")) {
9640 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
9644 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
9649 req.util = req.percent << SCHED_CAPACITY_SHIFT;
9650 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
9656 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
9657 size_t nbytes, loff_t off,
9658 enum uclamp_id clamp_id)
9660 struct uclamp_request req;
9661 struct task_group *tg;
9663 req = capacity_from_percent(buf);
9667 static_branch_enable(&sched_uclamp_used);
9669 mutex_lock(&uclamp_mutex);
9672 tg = css_tg(of_css(of));
9673 if (tg->uclamp_req[clamp_id].value != req.util)
9674 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
9677 * Because of not recoverable conversion rounding we keep track of the
9678 * exact requested value
9680 tg->uclamp_pct[clamp_id] = req.percent;
9682 /* Update effective clamps to track the most restrictive value */
9683 cpu_util_update_eff(of_css(of));
9686 mutex_unlock(&uclamp_mutex);
9691 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
9692 char *buf, size_t nbytes,
9695 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
9698 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
9699 char *buf, size_t nbytes,
9702 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
9705 static inline void cpu_uclamp_print(struct seq_file *sf,
9706 enum uclamp_id clamp_id)
9708 struct task_group *tg;
9714 tg = css_tg(seq_css(sf));
9715 util_clamp = tg->uclamp_req[clamp_id].value;
9718 if (util_clamp == SCHED_CAPACITY_SCALE) {
9719 seq_puts(sf, "max\n");
9723 percent = tg->uclamp_pct[clamp_id];
9724 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
9725 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
9728 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
9730 cpu_uclamp_print(sf, UCLAMP_MIN);
9734 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
9736 cpu_uclamp_print(sf, UCLAMP_MAX);
9739 #endif /* CONFIG_UCLAMP_TASK_GROUP */
9741 #ifdef CONFIG_FAIR_GROUP_SCHED
9742 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
9743 struct cftype *cftype, u64 shareval)
9745 if (shareval > scale_load_down(ULONG_MAX))
9746 shareval = MAX_SHARES;
9747 return sched_group_set_shares(css_tg(css), scale_load(shareval));
9750 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
9753 struct task_group *tg = css_tg(css);
9755 return (u64) scale_load_down(tg->shares);
9758 #ifdef CONFIG_CFS_BANDWIDTH
9759 static DEFINE_MUTEX(cfs_constraints_mutex);
9761 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9762 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9763 /* More than 203 days if BW_SHIFT equals 20. */
9764 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
9766 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9768 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9770 int i, ret = 0, runtime_enabled, runtime_was_enabled;
9771 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9773 if (tg == &root_task_group)
9777 * Ensure we have at some amount of bandwidth every period. This is
9778 * to prevent reaching a state of large arrears when throttled via
9779 * entity_tick() resulting in prolonged exit starvation.
9781 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9785 * Likewise, bound things on the other side by preventing insane quota
9786 * periods. This also allows us to normalize in computing quota
9789 if (period > max_cfs_quota_period)
9793 * Bound quota to defend quota against overflow during bandwidth shift.
9795 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
9799 * Prevent race between setting of cfs_rq->runtime_enabled and
9800 * unthrottle_offline_cfs_rqs().
9803 mutex_lock(&cfs_constraints_mutex);
9804 ret = __cfs_schedulable(tg, period, quota);
9808 runtime_enabled = quota != RUNTIME_INF;
9809 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
9811 * If we need to toggle cfs_bandwidth_used, off->on must occur
9812 * before making related changes, and on->off must occur afterwards
9814 if (runtime_enabled && !runtime_was_enabled)
9815 cfs_bandwidth_usage_inc();
9816 raw_spin_lock_irq(&cfs_b->lock);
9817 cfs_b->period = ns_to_ktime(period);
9818 cfs_b->quota = quota;
9820 __refill_cfs_bandwidth_runtime(cfs_b);
9822 /* Restart the period timer (if active) to handle new period expiry: */
9823 if (runtime_enabled)
9824 start_cfs_bandwidth(cfs_b);
9826 raw_spin_unlock_irq(&cfs_b->lock);
9828 for_each_online_cpu(i) {
9829 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9830 struct rq *rq = cfs_rq->rq;
9833 rq_lock_irq(rq, &rf);
9834 cfs_rq->runtime_enabled = runtime_enabled;
9835 cfs_rq->runtime_remaining = 0;
9837 if (cfs_rq->throttled)
9838 unthrottle_cfs_rq(cfs_rq);
9839 rq_unlock_irq(rq, &rf);
9841 if (runtime_was_enabled && !runtime_enabled)
9842 cfs_bandwidth_usage_dec();
9844 mutex_unlock(&cfs_constraints_mutex);
9850 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9854 period = ktime_to_ns(tg->cfs_bandwidth.period);
9855 if (cfs_quota_us < 0)
9856 quota = RUNTIME_INF;
9857 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
9858 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9862 return tg_set_cfs_bandwidth(tg, period, quota);
9865 static long tg_get_cfs_quota(struct task_group *tg)
9869 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
9872 quota_us = tg->cfs_bandwidth.quota;
9873 do_div(quota_us, NSEC_PER_USEC);
9878 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9882 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
9885 period = (u64)cfs_period_us * NSEC_PER_USEC;
9886 quota = tg->cfs_bandwidth.quota;
9888 return tg_set_cfs_bandwidth(tg, period, quota);
9891 static long tg_get_cfs_period(struct task_group *tg)
9895 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
9896 do_div(cfs_period_us, NSEC_PER_USEC);
9898 return cfs_period_us;
9901 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
9904 return tg_get_cfs_quota(css_tg(css));
9907 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
9908 struct cftype *cftype, s64 cfs_quota_us)
9910 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
9913 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
9916 return tg_get_cfs_period(css_tg(css));
9919 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
9920 struct cftype *cftype, u64 cfs_period_us)
9922 return tg_set_cfs_period(css_tg(css), cfs_period_us);
9925 struct cfs_schedulable_data {
9926 struct task_group *tg;
9931 * normalize group quota/period to be quota/max_period
9932 * note: units are usecs
9934 static u64 normalize_cfs_quota(struct task_group *tg,
9935 struct cfs_schedulable_data *d)
9943 period = tg_get_cfs_period(tg);
9944 quota = tg_get_cfs_quota(tg);
9947 /* note: these should typically be equivalent */
9948 if (quota == RUNTIME_INF || quota == -1)
9951 return to_ratio(period, quota);
9954 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9956 struct cfs_schedulable_data *d = data;
9957 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9958 s64 quota = 0, parent_quota = -1;
9961 quota = RUNTIME_INF;
9963 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
9965 quota = normalize_cfs_quota(tg, d);
9966 parent_quota = parent_b->hierarchical_quota;
9969 * Ensure max(child_quota) <= parent_quota. On cgroup2,
9970 * always take the min. On cgroup1, only inherit when no
9973 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
9974 quota = min(quota, parent_quota);
9976 if (quota == RUNTIME_INF)
9977 quota = parent_quota;
9978 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9982 cfs_b->hierarchical_quota = quota;
9987 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9990 struct cfs_schedulable_data data = {
9996 if (quota != RUNTIME_INF) {
9997 do_div(data.period, NSEC_PER_USEC);
9998 do_div(data.quota, NSEC_PER_USEC);
10002 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10008 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10010 struct task_group *tg = css_tg(seq_css(sf));
10011 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10013 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10014 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10015 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10017 if (schedstat_enabled() && tg != &root_task_group) {
10021 for_each_possible_cpu(i)
10022 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10024 seq_printf(sf, "wait_sum %llu\n", ws);
10029 #endif /* CONFIG_CFS_BANDWIDTH */
10030 #endif /* CONFIG_FAIR_GROUP_SCHED */
10032 #ifdef CONFIG_RT_GROUP_SCHED
10033 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10034 struct cftype *cft, s64 val)
10036 return sched_group_set_rt_runtime(css_tg(css), val);
10039 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10040 struct cftype *cft)
10042 return sched_group_rt_runtime(css_tg(css));
10045 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10046 struct cftype *cftype, u64 rt_period_us)
10048 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10051 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10052 struct cftype *cft)
10054 return sched_group_rt_period(css_tg(css));
10056 #endif /* CONFIG_RT_GROUP_SCHED */
10058 static struct cftype cpu_legacy_files[] = {
10059 #ifdef CONFIG_FAIR_GROUP_SCHED
10062 .read_u64 = cpu_shares_read_u64,
10063 .write_u64 = cpu_shares_write_u64,
10066 #ifdef CONFIG_CFS_BANDWIDTH
10068 .name = "cfs_quota_us",
10069 .read_s64 = cpu_cfs_quota_read_s64,
10070 .write_s64 = cpu_cfs_quota_write_s64,
10073 .name = "cfs_period_us",
10074 .read_u64 = cpu_cfs_period_read_u64,
10075 .write_u64 = cpu_cfs_period_write_u64,
10079 .seq_show = cpu_cfs_stat_show,
10082 #ifdef CONFIG_RT_GROUP_SCHED
10084 .name = "rt_runtime_us",
10085 .read_s64 = cpu_rt_runtime_read,
10086 .write_s64 = cpu_rt_runtime_write,
10089 .name = "rt_period_us",
10090 .read_u64 = cpu_rt_period_read_uint,
10091 .write_u64 = cpu_rt_period_write_uint,
10094 #ifdef CONFIG_UCLAMP_TASK_GROUP
10096 .name = "uclamp.min",
10097 .flags = CFTYPE_NOT_ON_ROOT,
10098 .seq_show = cpu_uclamp_min_show,
10099 .write = cpu_uclamp_min_write,
10102 .name = "uclamp.max",
10103 .flags = CFTYPE_NOT_ON_ROOT,
10104 .seq_show = cpu_uclamp_max_show,
10105 .write = cpu_uclamp_max_write,
10108 { } /* Terminate */
10111 static int cpu_extra_stat_show(struct seq_file *sf,
10112 struct cgroup_subsys_state *css)
10114 #ifdef CONFIG_CFS_BANDWIDTH
10116 struct task_group *tg = css_tg(css);
10117 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10118 u64 throttled_usec;
10120 throttled_usec = cfs_b->throttled_time;
10121 do_div(throttled_usec, NSEC_PER_USEC);
10123 seq_printf(sf, "nr_periods %d\n"
10124 "nr_throttled %d\n"
10125 "throttled_usec %llu\n",
10126 cfs_b->nr_periods, cfs_b->nr_throttled,
10133 #ifdef CONFIG_FAIR_GROUP_SCHED
10134 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10135 struct cftype *cft)
10137 struct task_group *tg = css_tg(css);
10138 u64 weight = scale_load_down(tg->shares);
10140 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10143 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10144 struct cftype *cft, u64 weight)
10147 * cgroup weight knobs should use the common MIN, DFL and MAX
10148 * values which are 1, 100 and 10000 respectively. While it loses
10149 * a bit of range on both ends, it maps pretty well onto the shares
10150 * value used by scheduler and the round-trip conversions preserve
10151 * the original value over the entire range.
10153 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10156 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10158 return sched_group_set_shares(css_tg(css), scale_load(weight));
10161 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10162 struct cftype *cft)
10164 unsigned long weight = scale_load_down(css_tg(css)->shares);
10165 int last_delta = INT_MAX;
10168 /* find the closest nice value to the current weight */
10169 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10170 delta = abs(sched_prio_to_weight[prio] - weight);
10171 if (delta >= last_delta)
10173 last_delta = delta;
10176 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10179 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10180 struct cftype *cft, s64 nice)
10182 unsigned long weight;
10185 if (nice < MIN_NICE || nice > MAX_NICE)
10188 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10189 idx = array_index_nospec(idx, 40);
10190 weight = sched_prio_to_weight[idx];
10192 return sched_group_set_shares(css_tg(css), scale_load(weight));
10196 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10197 long period, long quota)
10200 seq_puts(sf, "max");
10202 seq_printf(sf, "%ld", quota);
10204 seq_printf(sf, " %ld\n", period);
10207 /* caller should put the current value in *@periodp before calling */
10208 static int __maybe_unused cpu_period_quota_parse(char *buf,
10209 u64 *periodp, u64 *quotap)
10211 char tok[21]; /* U64_MAX */
10213 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10216 *periodp *= NSEC_PER_USEC;
10218 if (sscanf(tok, "%llu", quotap))
10219 *quotap *= NSEC_PER_USEC;
10220 else if (!strcmp(tok, "max"))
10221 *quotap = RUNTIME_INF;
10228 #ifdef CONFIG_CFS_BANDWIDTH
10229 static int cpu_max_show(struct seq_file *sf, void *v)
10231 struct task_group *tg = css_tg(seq_css(sf));
10233 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10237 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10238 char *buf, size_t nbytes, loff_t off)
10240 struct task_group *tg = css_tg(of_css(of));
10241 u64 period = tg_get_cfs_period(tg);
10245 ret = cpu_period_quota_parse(buf, &period, "a);
10247 ret = tg_set_cfs_bandwidth(tg, period, quota);
10248 return ret ?: nbytes;
10252 static struct cftype cpu_files[] = {
10253 #ifdef CONFIG_FAIR_GROUP_SCHED
10256 .flags = CFTYPE_NOT_ON_ROOT,
10257 .read_u64 = cpu_weight_read_u64,
10258 .write_u64 = cpu_weight_write_u64,
10261 .name = "weight.nice",
10262 .flags = CFTYPE_NOT_ON_ROOT,
10263 .read_s64 = cpu_weight_nice_read_s64,
10264 .write_s64 = cpu_weight_nice_write_s64,
10267 #ifdef CONFIG_CFS_BANDWIDTH
10270 .flags = CFTYPE_NOT_ON_ROOT,
10271 .seq_show = cpu_max_show,
10272 .write = cpu_max_write,
10275 #ifdef CONFIG_UCLAMP_TASK_GROUP
10277 .name = "uclamp.min",
10278 .flags = CFTYPE_NOT_ON_ROOT,
10279 .seq_show = cpu_uclamp_min_show,
10280 .write = cpu_uclamp_min_write,
10283 .name = "uclamp.max",
10284 .flags = CFTYPE_NOT_ON_ROOT,
10285 .seq_show = cpu_uclamp_max_show,
10286 .write = cpu_uclamp_max_write,
10289 { } /* terminate */
10292 struct cgroup_subsys cpu_cgrp_subsys = {
10293 .css_alloc = cpu_cgroup_css_alloc,
10294 .css_online = cpu_cgroup_css_online,
10295 .css_released = cpu_cgroup_css_released,
10296 .css_free = cpu_cgroup_css_free,
10297 .css_extra_stat_show = cpu_extra_stat_show,
10298 .fork = cpu_cgroup_fork,
10299 .can_attach = cpu_cgroup_can_attach,
10300 .attach = cpu_cgroup_attach,
10301 .legacy_cftypes = cpu_legacy_files,
10302 .dfl_cftypes = cpu_files,
10303 .early_init = true,
10307 #endif /* CONFIG_CGROUP_SCHED */
10309 void dump_cpu_task(int cpu)
10311 pr_info("Task dump for CPU %d:\n", cpu);
10312 sched_show_task(cpu_curr(cpu));
10316 * Nice levels are multiplicative, with a gentle 10% change for every
10317 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10318 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10319 * that remained on nice 0.
10321 * The "10% effect" is relative and cumulative: from _any_ nice level,
10322 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10323 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10324 * If a task goes up by ~10% and another task goes down by ~10% then
10325 * the relative distance between them is ~25%.)
10327 const int sched_prio_to_weight[40] = {
10328 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10329 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10330 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10331 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10332 /* 0 */ 1024, 820, 655, 526, 423,
10333 /* 5 */ 335, 272, 215, 172, 137,
10334 /* 10 */ 110, 87, 70, 56, 45,
10335 /* 15 */ 36, 29, 23, 18, 15,
10339 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10341 * In cases where the weight does not change often, we can use the
10342 * precalculated inverse to speed up arithmetics by turning divisions
10343 * into multiplications:
10345 const u32 sched_prio_to_wmult[40] = {
10346 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10347 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10348 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10349 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10350 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10351 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10352 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10353 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10356 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10358 trace_sched_update_nr_running_tp(rq, count);