1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
63 * Print a warning if need_resched is set for the given duration (if
64 * LATENCY_WARN is enabled).
66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
69 __read_mostly int sysctl_resched_latency_warn_ms = 100;
70 __read_mostly int sysctl_resched_latency_warn_once = 1;
71 #endif /* CONFIG_SCHED_DEBUG */
74 * Number of tasks to iterate in a single balance run.
75 * Limited because this is done with IRQs disabled.
77 const_debug unsigned int sysctl_sched_nr_migrate = 32;
80 * period over which we measure -rt task CPU usage in us.
83 unsigned int sysctl_sched_rt_period = 1000000;
85 __read_mostly int scheduler_running;
87 #ifdef CONFIG_SCHED_CORE
89 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
91 /* kernel prio, less is more */
92 static inline int __task_prio(struct task_struct *p)
94 if (p->sched_class == &stop_sched_class) /* trumps deadline */
97 if (rt_prio(p->prio)) /* includes deadline */
98 return p->prio; /* [-1, 99] */
100 if (p->sched_class == &idle_sched_class)
101 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
103 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
113 /* real prio, less is less */
114 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
117 int pa = __task_prio(a), pb = __task_prio(b);
125 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
126 return !dl_time_before(a->dl.deadline, b->dl.deadline);
128 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
129 return cfs_prio_less(a, b, in_fi);
134 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
136 if (a->core_cookie < b->core_cookie)
139 if (a->core_cookie > b->core_cookie)
142 /* flip prio, so high prio is leftmost */
143 if (prio_less(b, a, task_rq(a)->core->core_forceidle))
149 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
151 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
153 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
156 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
158 const struct task_struct *p = __node_2_sc(node);
159 unsigned long cookie = (unsigned long)key;
161 if (cookie < p->core_cookie)
164 if (cookie > p->core_cookie)
170 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
172 rq->core->core_task_seq++;
177 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
180 void sched_core_dequeue(struct rq *rq, struct task_struct *p)
182 rq->core->core_task_seq++;
184 if (!sched_core_enqueued(p))
187 rb_erase(&p->core_node, &rq->core_tree);
188 RB_CLEAR_NODE(&p->core_node);
192 * Find left-most (aka, highest priority) task matching @cookie.
194 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
196 struct rb_node *node;
198 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
200 * The idle task always matches any cookie!
203 return idle_sched_class.pick_task(rq);
205 return __node_2_sc(node);
208 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
210 struct rb_node *node = &p->core_node;
212 node = rb_next(node);
216 p = container_of(node, struct task_struct, core_node);
217 if (p->core_cookie != cookie)
224 * Magic required such that:
226 * raw_spin_rq_lock(rq);
228 * raw_spin_rq_unlock(rq);
230 * ends up locking and unlocking the _same_ lock, and all CPUs
231 * always agree on what rq has what lock.
233 * XXX entirely possible to selectively enable cores, don't bother for now.
236 static DEFINE_MUTEX(sched_core_mutex);
237 static atomic_t sched_core_count;
238 static struct cpumask sched_core_mask;
240 static void __sched_core_flip(bool enabled)
247 * Toggle the online cores, one by one.
249 cpumask_copy(&sched_core_mask, cpu_online_mask);
250 for_each_cpu(cpu, &sched_core_mask) {
251 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
255 for_each_cpu(t, smt_mask) {
256 /* supports up to SMT8 */
257 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
260 for_each_cpu(t, smt_mask)
261 cpu_rq(t)->core_enabled = enabled;
263 for_each_cpu(t, smt_mask)
264 raw_spin_unlock(&cpu_rq(t)->__lock);
267 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
271 * Toggle the offline CPUs.
273 cpumask_copy(&sched_core_mask, cpu_possible_mask);
274 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
276 for_each_cpu(cpu, &sched_core_mask)
277 cpu_rq(cpu)->core_enabled = enabled;
282 static void sched_core_assert_empty(void)
286 for_each_possible_cpu(cpu)
287 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
290 static void __sched_core_enable(void)
292 static_branch_enable(&__sched_core_enabled);
294 * Ensure all previous instances of raw_spin_rq_*lock() have finished
295 * and future ones will observe !sched_core_disabled().
298 __sched_core_flip(true);
299 sched_core_assert_empty();
302 static void __sched_core_disable(void)
304 sched_core_assert_empty();
305 __sched_core_flip(false);
306 static_branch_disable(&__sched_core_enabled);
309 void sched_core_get(void)
311 if (atomic_inc_not_zero(&sched_core_count))
314 mutex_lock(&sched_core_mutex);
315 if (!atomic_read(&sched_core_count))
316 __sched_core_enable();
318 smp_mb__before_atomic();
319 atomic_inc(&sched_core_count);
320 mutex_unlock(&sched_core_mutex);
323 static void __sched_core_put(struct work_struct *work)
325 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
326 __sched_core_disable();
327 mutex_unlock(&sched_core_mutex);
331 void sched_core_put(void)
333 static DECLARE_WORK(_work, __sched_core_put);
336 * "There can be only one"
338 * Either this is the last one, or we don't actually need to do any
339 * 'work'. If it is the last *again*, we rely on
340 * WORK_STRUCT_PENDING_BIT.
342 if (!atomic_add_unless(&sched_core_count, -1, 1))
343 schedule_work(&_work);
346 #else /* !CONFIG_SCHED_CORE */
348 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
349 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
351 #endif /* CONFIG_SCHED_CORE */
354 * part of the period that we allow rt tasks to run in us.
357 int sysctl_sched_rt_runtime = 950000;
361 * Serialization rules:
367 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
370 * rq2->lock where: rq1 < rq2
374 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
375 * local CPU's rq->lock, it optionally removes the task from the runqueue and
376 * always looks at the local rq data structures to find the most eligible task
379 * Task enqueue is also under rq->lock, possibly taken from another CPU.
380 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
381 * the local CPU to avoid bouncing the runqueue state around [ see
382 * ttwu_queue_wakelist() ]
384 * Task wakeup, specifically wakeups that involve migration, are horribly
385 * complicated to avoid having to take two rq->locks.
389 * System-calls and anything external will use task_rq_lock() which acquires
390 * both p->pi_lock and rq->lock. As a consequence the state they change is
391 * stable while holding either lock:
393 * - sched_setaffinity()/
394 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
395 * - set_user_nice(): p->se.load, p->*prio
396 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
397 * p->se.load, p->rt_priority,
398 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
399 * - sched_setnuma(): p->numa_preferred_nid
400 * - sched_move_task()/
401 * cpu_cgroup_fork(): p->sched_task_group
402 * - uclamp_update_active() p->uclamp*
404 * p->state <- TASK_*:
406 * is changed locklessly using set_current_state(), __set_current_state() or
407 * set_special_state(), see their respective comments, or by
408 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
411 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
413 * is set by activate_task() and cleared by deactivate_task(), under
414 * rq->lock. Non-zero indicates the task is runnable, the special
415 * ON_RQ_MIGRATING state is used for migration without holding both
416 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
418 * p->on_cpu <- { 0, 1 }:
420 * is set by prepare_task() and cleared by finish_task() such that it will be
421 * set before p is scheduled-in and cleared after p is scheduled-out, both
422 * under rq->lock. Non-zero indicates the task is running on its CPU.
424 * [ The astute reader will observe that it is possible for two tasks on one
425 * CPU to have ->on_cpu = 1 at the same time. ]
427 * task_cpu(p): is changed by set_task_cpu(), the rules are:
429 * - Don't call set_task_cpu() on a blocked task:
431 * We don't care what CPU we're not running on, this simplifies hotplug,
432 * the CPU assignment of blocked tasks isn't required to be valid.
434 * - for try_to_wake_up(), called under p->pi_lock:
436 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
438 * - for migration called under rq->lock:
439 * [ see task_on_rq_migrating() in task_rq_lock() ]
441 * o move_queued_task()
444 * - for migration called under double_rq_lock():
446 * o __migrate_swap_task()
447 * o push_rt_task() / pull_rt_task()
448 * o push_dl_task() / pull_dl_task()
449 * o dl_task_offline_migration()
453 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
455 raw_spinlock_t *lock;
457 /* Matches synchronize_rcu() in __sched_core_enable() */
459 if (sched_core_disabled()) {
460 raw_spin_lock_nested(&rq->__lock, subclass);
461 /* preempt_count *MUST* be > 1 */
462 preempt_enable_no_resched();
467 lock = __rq_lockp(rq);
468 raw_spin_lock_nested(lock, subclass);
469 if (likely(lock == __rq_lockp(rq))) {
470 /* preempt_count *MUST* be > 1 */
471 preempt_enable_no_resched();
474 raw_spin_unlock(lock);
478 bool raw_spin_rq_trylock(struct rq *rq)
480 raw_spinlock_t *lock;
483 /* Matches synchronize_rcu() in __sched_core_enable() */
485 if (sched_core_disabled()) {
486 ret = raw_spin_trylock(&rq->__lock);
492 lock = __rq_lockp(rq);
493 ret = raw_spin_trylock(lock);
494 if (!ret || (likely(lock == __rq_lockp(rq)))) {
498 raw_spin_unlock(lock);
502 void raw_spin_rq_unlock(struct rq *rq)
504 raw_spin_unlock(rq_lockp(rq));
509 * double_rq_lock - safely lock two runqueues
511 void double_rq_lock(struct rq *rq1, struct rq *rq2)
513 lockdep_assert_irqs_disabled();
515 if (rq_order_less(rq2, rq1))
518 raw_spin_rq_lock(rq1);
519 if (__rq_lockp(rq1) == __rq_lockp(rq2))
522 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
527 * __task_rq_lock - lock the rq @p resides on.
529 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
534 lockdep_assert_held(&p->pi_lock);
538 raw_spin_rq_lock(rq);
539 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
543 raw_spin_rq_unlock(rq);
545 while (unlikely(task_on_rq_migrating(p)))
551 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
553 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
554 __acquires(p->pi_lock)
560 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
562 raw_spin_rq_lock(rq);
564 * move_queued_task() task_rq_lock()
567 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
568 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
569 * [S] ->cpu = new_cpu [L] task_rq()
573 * If we observe the old CPU in task_rq_lock(), the acquire of
574 * the old rq->lock will fully serialize against the stores.
576 * If we observe the new CPU in task_rq_lock(), the address
577 * dependency headed by '[L] rq = task_rq()' and the acquire
578 * will pair with the WMB to ensure we then also see migrating.
580 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
584 raw_spin_rq_unlock(rq);
585 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
587 while (unlikely(task_on_rq_migrating(p)))
593 * RQ-clock updating methods:
596 static void update_rq_clock_task(struct rq *rq, s64 delta)
599 * In theory, the compile should just see 0 here, and optimize out the call
600 * to sched_rt_avg_update. But I don't trust it...
602 s64 __maybe_unused steal = 0, irq_delta = 0;
604 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
605 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
608 * Since irq_time is only updated on {soft,}irq_exit, we might run into
609 * this case when a previous update_rq_clock() happened inside a
612 * When this happens, we stop ->clock_task and only update the
613 * prev_irq_time stamp to account for the part that fit, so that a next
614 * update will consume the rest. This ensures ->clock_task is
617 * It does however cause some slight miss-attribution of {soft,}irq
618 * time, a more accurate solution would be to update the irq_time using
619 * the current rq->clock timestamp, except that would require using
622 if (irq_delta > delta)
625 rq->prev_irq_time += irq_delta;
628 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
629 if (static_key_false((¶virt_steal_rq_enabled))) {
630 steal = paravirt_steal_clock(cpu_of(rq));
631 steal -= rq->prev_steal_time_rq;
633 if (unlikely(steal > delta))
636 rq->prev_steal_time_rq += steal;
641 rq->clock_task += delta;
643 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
644 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
645 update_irq_load_avg(rq, irq_delta + steal);
647 update_rq_clock_pelt(rq, delta);
650 void update_rq_clock(struct rq *rq)
654 lockdep_assert_rq_held(rq);
656 if (rq->clock_update_flags & RQCF_ACT_SKIP)
659 #ifdef CONFIG_SCHED_DEBUG
660 if (sched_feat(WARN_DOUBLE_CLOCK))
661 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
662 rq->clock_update_flags |= RQCF_UPDATED;
665 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
669 update_rq_clock_task(rq, delta);
672 #ifdef CONFIG_SCHED_HRTICK
674 * Use HR-timers to deliver accurate preemption points.
677 static void hrtick_clear(struct rq *rq)
679 if (hrtimer_active(&rq->hrtick_timer))
680 hrtimer_cancel(&rq->hrtick_timer);
684 * High-resolution timer tick.
685 * Runs from hardirq context with interrupts disabled.
687 static enum hrtimer_restart hrtick(struct hrtimer *timer)
689 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
692 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
696 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
699 return HRTIMER_NORESTART;
704 static void __hrtick_restart(struct rq *rq)
706 struct hrtimer *timer = &rq->hrtick_timer;
707 ktime_t time = rq->hrtick_time;
709 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
713 * called from hardirq (IPI) context
715 static void __hrtick_start(void *arg)
721 __hrtick_restart(rq);
726 * Called to set the hrtick timer state.
728 * called with rq->lock held and irqs disabled
730 void hrtick_start(struct rq *rq, u64 delay)
732 struct hrtimer *timer = &rq->hrtick_timer;
736 * Don't schedule slices shorter than 10000ns, that just
737 * doesn't make sense and can cause timer DoS.
739 delta = max_t(s64, delay, 10000LL);
740 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
743 __hrtick_restart(rq);
745 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
750 * Called to set the hrtick timer state.
752 * called with rq->lock held and irqs disabled
754 void hrtick_start(struct rq *rq, u64 delay)
757 * Don't schedule slices shorter than 10000ns, that just
758 * doesn't make sense. Rely on vruntime for fairness.
760 delay = max_t(u64, delay, 10000LL);
761 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
762 HRTIMER_MODE_REL_PINNED_HARD);
765 #endif /* CONFIG_SMP */
767 static void hrtick_rq_init(struct rq *rq)
770 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
772 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
773 rq->hrtick_timer.function = hrtick;
775 #else /* CONFIG_SCHED_HRTICK */
776 static inline void hrtick_clear(struct rq *rq)
780 static inline void hrtick_rq_init(struct rq *rq)
783 #endif /* CONFIG_SCHED_HRTICK */
786 * cmpxchg based fetch_or, macro so it works for different integer types
788 #define fetch_or(ptr, mask) \
790 typeof(ptr) _ptr = (ptr); \
791 typeof(mask) _mask = (mask); \
792 typeof(*_ptr) _old, _val = *_ptr; \
795 _old = cmpxchg(_ptr, _val, _val | _mask); \
803 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
805 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
806 * this avoids any races wrt polling state changes and thereby avoids
809 static bool set_nr_and_not_polling(struct task_struct *p)
811 struct thread_info *ti = task_thread_info(p);
812 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
816 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
818 * If this returns true, then the idle task promises to call
819 * sched_ttwu_pending() and reschedule soon.
821 static bool set_nr_if_polling(struct task_struct *p)
823 struct thread_info *ti = task_thread_info(p);
824 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
827 if (!(val & _TIF_POLLING_NRFLAG))
829 if (val & _TIF_NEED_RESCHED)
831 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
840 static bool set_nr_and_not_polling(struct task_struct *p)
842 set_tsk_need_resched(p);
847 static bool set_nr_if_polling(struct task_struct *p)
854 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
856 struct wake_q_node *node = &task->wake_q;
859 * Atomically grab the task, if ->wake_q is !nil already it means
860 * it's already queued (either by us or someone else) and will get the
861 * wakeup due to that.
863 * In order to ensure that a pending wakeup will observe our pending
864 * state, even in the failed case, an explicit smp_mb() must be used.
866 smp_mb__before_atomic();
867 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
871 * The head is context local, there can be no concurrency.
874 head->lastp = &node->next;
879 * wake_q_add() - queue a wakeup for 'later' waking.
880 * @head: the wake_q_head to add @task to
881 * @task: the task to queue for 'later' wakeup
883 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
884 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
887 * This function must be used as-if it were wake_up_process(); IOW the task
888 * must be ready to be woken at this location.
890 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
892 if (__wake_q_add(head, task))
893 get_task_struct(task);
897 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
898 * @head: the wake_q_head to add @task to
899 * @task: the task to queue for 'later' wakeup
901 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
902 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
905 * This function must be used as-if it were wake_up_process(); IOW the task
906 * must be ready to be woken at this location.
908 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
909 * that already hold reference to @task can call the 'safe' version and trust
910 * wake_q to do the right thing depending whether or not the @task is already
913 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
915 if (!__wake_q_add(head, task))
916 put_task_struct(task);
919 void wake_up_q(struct wake_q_head *head)
921 struct wake_q_node *node = head->first;
923 while (node != WAKE_Q_TAIL) {
924 struct task_struct *task;
926 task = container_of(node, struct task_struct, wake_q);
927 /* Task can safely be re-inserted now: */
929 task->wake_q.next = NULL;
932 * wake_up_process() executes a full barrier, which pairs with
933 * the queueing in wake_q_add() so as not to miss wakeups.
935 wake_up_process(task);
936 put_task_struct(task);
941 * resched_curr - mark rq's current task 'to be rescheduled now'.
943 * On UP this means the setting of the need_resched flag, on SMP it
944 * might also involve a cross-CPU call to trigger the scheduler on
947 void resched_curr(struct rq *rq)
949 struct task_struct *curr = rq->curr;
952 lockdep_assert_rq_held(rq);
954 if (test_tsk_need_resched(curr))
959 if (cpu == smp_processor_id()) {
960 set_tsk_need_resched(curr);
961 set_preempt_need_resched();
965 if (set_nr_and_not_polling(curr))
966 smp_send_reschedule(cpu);
968 trace_sched_wake_idle_without_ipi(cpu);
971 void resched_cpu(int cpu)
973 struct rq *rq = cpu_rq(cpu);
976 raw_spin_rq_lock_irqsave(rq, flags);
977 if (cpu_online(cpu) || cpu == smp_processor_id())
979 raw_spin_rq_unlock_irqrestore(rq, flags);
983 #ifdef CONFIG_NO_HZ_COMMON
985 * In the semi idle case, use the nearest busy CPU for migrating timers
986 * from an idle CPU. This is good for power-savings.
988 * We don't do similar optimization for completely idle system, as
989 * selecting an idle CPU will add more delays to the timers than intended
990 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
992 int get_nohz_timer_target(void)
994 int i, cpu = smp_processor_id(), default_cpu = -1;
995 struct sched_domain *sd;
997 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1004 for_each_domain(cpu, sd) {
1005 for_each_cpu_and(i, sched_domain_span(sd),
1006 housekeeping_cpumask(HK_FLAG_TIMER)) {
1017 if (default_cpu == -1)
1018 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1026 * When add_timer_on() enqueues a timer into the timer wheel of an
1027 * idle CPU then this timer might expire before the next timer event
1028 * which is scheduled to wake up that CPU. In case of a completely
1029 * idle system the next event might even be infinite time into the
1030 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1031 * leaves the inner idle loop so the newly added timer is taken into
1032 * account when the CPU goes back to idle and evaluates the timer
1033 * wheel for the next timer event.
1035 static void wake_up_idle_cpu(int cpu)
1037 struct rq *rq = cpu_rq(cpu);
1039 if (cpu == smp_processor_id())
1042 if (set_nr_and_not_polling(rq->idle))
1043 smp_send_reschedule(cpu);
1045 trace_sched_wake_idle_without_ipi(cpu);
1048 static bool wake_up_full_nohz_cpu(int cpu)
1051 * We just need the target to call irq_exit() and re-evaluate
1052 * the next tick. The nohz full kick at least implies that.
1053 * If needed we can still optimize that later with an
1056 if (cpu_is_offline(cpu))
1057 return true; /* Don't try to wake offline CPUs. */
1058 if (tick_nohz_full_cpu(cpu)) {
1059 if (cpu != smp_processor_id() ||
1060 tick_nohz_tick_stopped())
1061 tick_nohz_full_kick_cpu(cpu);
1069 * Wake up the specified CPU. If the CPU is going offline, it is the
1070 * caller's responsibility to deal with the lost wakeup, for example,
1071 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1073 void wake_up_nohz_cpu(int cpu)
1075 if (!wake_up_full_nohz_cpu(cpu))
1076 wake_up_idle_cpu(cpu);
1079 static void nohz_csd_func(void *info)
1081 struct rq *rq = info;
1082 int cpu = cpu_of(rq);
1086 * Release the rq::nohz_csd.
1088 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1089 WARN_ON(!(flags & NOHZ_KICK_MASK));
1091 rq->idle_balance = idle_cpu(cpu);
1092 if (rq->idle_balance && !need_resched()) {
1093 rq->nohz_idle_balance = flags;
1094 raise_softirq_irqoff(SCHED_SOFTIRQ);
1098 #endif /* CONFIG_NO_HZ_COMMON */
1100 #ifdef CONFIG_NO_HZ_FULL
1101 bool sched_can_stop_tick(struct rq *rq)
1103 int fifo_nr_running;
1105 /* Deadline tasks, even if single, need the tick */
1106 if (rq->dl.dl_nr_running)
1110 * If there are more than one RR tasks, we need the tick to affect the
1111 * actual RR behaviour.
1113 if (rq->rt.rr_nr_running) {
1114 if (rq->rt.rr_nr_running == 1)
1121 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1122 * forced preemption between FIFO tasks.
1124 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1125 if (fifo_nr_running)
1129 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1130 * if there's more than one we need the tick for involuntary
1133 if (rq->nr_running > 1)
1138 #endif /* CONFIG_NO_HZ_FULL */
1139 #endif /* CONFIG_SMP */
1141 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1142 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1144 * Iterate task_group tree rooted at *from, calling @down when first entering a
1145 * node and @up when leaving it for the final time.
1147 * Caller must hold rcu_lock or sufficient equivalent.
1149 int walk_tg_tree_from(struct task_group *from,
1150 tg_visitor down, tg_visitor up, void *data)
1152 struct task_group *parent, *child;
1158 ret = (*down)(parent, data);
1161 list_for_each_entry_rcu(child, &parent->children, siblings) {
1168 ret = (*up)(parent, data);
1169 if (ret || parent == from)
1173 parent = parent->parent;
1180 int tg_nop(struct task_group *tg, void *data)
1186 static void set_load_weight(struct task_struct *p, bool update_load)
1188 int prio = p->static_prio - MAX_RT_PRIO;
1189 struct load_weight *load = &p->se.load;
1192 * SCHED_IDLE tasks get minimal weight:
1194 if (task_has_idle_policy(p)) {
1195 load->weight = scale_load(WEIGHT_IDLEPRIO);
1196 load->inv_weight = WMULT_IDLEPRIO;
1201 * SCHED_OTHER tasks have to update their load when changing their
1204 if (update_load && p->sched_class == &fair_sched_class) {
1205 reweight_task(p, prio);
1207 load->weight = scale_load(sched_prio_to_weight[prio]);
1208 load->inv_weight = sched_prio_to_wmult[prio];
1212 #ifdef CONFIG_UCLAMP_TASK
1214 * Serializes updates of utilization clamp values
1216 * The (slow-path) user-space triggers utilization clamp value updates which
1217 * can require updates on (fast-path) scheduler's data structures used to
1218 * support enqueue/dequeue operations.
1219 * While the per-CPU rq lock protects fast-path update operations, user-space
1220 * requests are serialized using a mutex to reduce the risk of conflicting
1221 * updates or API abuses.
1223 static DEFINE_MUTEX(uclamp_mutex);
1225 /* Max allowed minimum utilization */
1226 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1228 /* Max allowed maximum utilization */
1229 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1232 * By default RT tasks run at the maximum performance point/capacity of the
1233 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1234 * SCHED_CAPACITY_SCALE.
1236 * This knob allows admins to change the default behavior when uclamp is being
1237 * used. In battery powered devices, particularly, running at the maximum
1238 * capacity and frequency will increase energy consumption and shorten the
1241 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1243 * This knob will not override the system default sched_util_clamp_min defined
1246 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1248 /* All clamps are required to be less or equal than these values */
1249 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1252 * This static key is used to reduce the uclamp overhead in the fast path. It
1253 * primarily disables the call to uclamp_rq_{inc, dec}() in
1254 * enqueue/dequeue_task().
1256 * This allows users to continue to enable uclamp in their kernel config with
1257 * minimum uclamp overhead in the fast path.
1259 * As soon as userspace modifies any of the uclamp knobs, the static key is
1260 * enabled, since we have an actual users that make use of uclamp
1263 * The knobs that would enable this static key are:
1265 * * A task modifying its uclamp value with sched_setattr().
1266 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1267 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1269 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1271 /* Integer rounded range for each bucket */
1272 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1274 #define for_each_clamp_id(clamp_id) \
1275 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1277 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1279 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1282 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1284 if (clamp_id == UCLAMP_MIN)
1286 return SCHED_CAPACITY_SCALE;
1289 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1290 unsigned int value, bool user_defined)
1292 uc_se->value = value;
1293 uc_se->bucket_id = uclamp_bucket_id(value);
1294 uc_se->user_defined = user_defined;
1297 static inline unsigned int
1298 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1299 unsigned int clamp_value)
1302 * Avoid blocked utilization pushing up the frequency when we go
1303 * idle (which drops the max-clamp) by retaining the last known
1306 if (clamp_id == UCLAMP_MAX) {
1307 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1311 return uclamp_none(UCLAMP_MIN);
1314 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1315 unsigned int clamp_value)
1317 /* Reset max-clamp retention only on idle exit */
1318 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1321 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1325 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1326 unsigned int clamp_value)
1328 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1329 int bucket_id = UCLAMP_BUCKETS - 1;
1332 * Since both min and max clamps are max aggregated, find the
1333 * top most bucket with tasks in.
1335 for ( ; bucket_id >= 0; bucket_id--) {
1336 if (!bucket[bucket_id].tasks)
1338 return bucket[bucket_id].value;
1341 /* No tasks -- default clamp values */
1342 return uclamp_idle_value(rq, clamp_id, clamp_value);
1345 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1347 unsigned int default_util_min;
1348 struct uclamp_se *uc_se;
1350 lockdep_assert_held(&p->pi_lock);
1352 uc_se = &p->uclamp_req[UCLAMP_MIN];
1354 /* Only sync if user didn't override the default */
1355 if (uc_se->user_defined)
1358 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1359 uclamp_se_set(uc_se, default_util_min, false);
1362 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1370 /* Protect updates to p->uclamp_* */
1371 rq = task_rq_lock(p, &rf);
1372 __uclamp_update_util_min_rt_default(p);
1373 task_rq_unlock(rq, p, &rf);
1376 static void uclamp_sync_util_min_rt_default(void)
1378 struct task_struct *g, *p;
1381 * copy_process() sysctl_uclamp
1382 * uclamp_min_rt = X;
1383 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1384 * // link thread smp_mb__after_spinlock()
1385 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1386 * sched_post_fork() for_each_process_thread()
1387 * __uclamp_sync_rt() __uclamp_sync_rt()
1389 * Ensures that either sched_post_fork() will observe the new
1390 * uclamp_min_rt or for_each_process_thread() will observe the new
1393 read_lock(&tasklist_lock);
1394 smp_mb__after_spinlock();
1395 read_unlock(&tasklist_lock);
1398 for_each_process_thread(g, p)
1399 uclamp_update_util_min_rt_default(p);
1403 static inline struct uclamp_se
1404 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1406 /* Copy by value as we could modify it */
1407 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1408 #ifdef CONFIG_UCLAMP_TASK_GROUP
1409 unsigned int tg_min, tg_max, value;
1412 * Tasks in autogroups or root task group will be
1413 * restricted by system defaults.
1415 if (task_group_is_autogroup(task_group(p)))
1417 if (task_group(p) == &root_task_group)
1420 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1421 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1422 value = uc_req.value;
1423 value = clamp(value, tg_min, tg_max);
1424 uclamp_se_set(&uc_req, value, false);
1431 * The effective clamp bucket index of a task depends on, by increasing
1433 * - the task specific clamp value, when explicitly requested from userspace
1434 * - the task group effective clamp value, for tasks not either in the root
1435 * group or in an autogroup
1436 * - the system default clamp value, defined by the sysadmin
1438 static inline struct uclamp_se
1439 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1441 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1442 struct uclamp_se uc_max = uclamp_default[clamp_id];
1444 /* System default restrictions always apply */
1445 if (unlikely(uc_req.value > uc_max.value))
1451 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1453 struct uclamp_se uc_eff;
1455 /* Task currently refcounted: use back-annotated (effective) value */
1456 if (p->uclamp[clamp_id].active)
1457 return (unsigned long)p->uclamp[clamp_id].value;
1459 uc_eff = uclamp_eff_get(p, clamp_id);
1461 return (unsigned long)uc_eff.value;
1465 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1466 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1467 * updates the rq's clamp value if required.
1469 * Tasks can have a task-specific value requested from user-space, track
1470 * within each bucket the maximum value for tasks refcounted in it.
1471 * This "local max aggregation" allows to track the exact "requested" value
1472 * for each bucket when all its RUNNABLE tasks require the same clamp.
1474 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1475 enum uclamp_id clamp_id)
1477 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1478 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1479 struct uclamp_bucket *bucket;
1481 lockdep_assert_rq_held(rq);
1483 /* Update task effective clamp */
1484 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1486 bucket = &uc_rq->bucket[uc_se->bucket_id];
1488 uc_se->active = true;
1490 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1493 * Local max aggregation: rq buckets always track the max
1494 * "requested" clamp value of its RUNNABLE tasks.
1496 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1497 bucket->value = uc_se->value;
1499 if (uc_se->value > READ_ONCE(uc_rq->value))
1500 WRITE_ONCE(uc_rq->value, uc_se->value);
1504 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1505 * is released. If this is the last task reference counting the rq's max
1506 * active clamp value, then the rq's clamp value is updated.
1508 * Both refcounted tasks and rq's cached clamp values are expected to be
1509 * always valid. If it's detected they are not, as defensive programming,
1510 * enforce the expected state and warn.
1512 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1513 enum uclamp_id clamp_id)
1515 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1516 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1517 struct uclamp_bucket *bucket;
1518 unsigned int bkt_clamp;
1519 unsigned int rq_clamp;
1521 lockdep_assert_rq_held(rq);
1524 * If sched_uclamp_used was enabled after task @p was enqueued,
1525 * we could end up with unbalanced call to uclamp_rq_dec_id().
1527 * In this case the uc_se->active flag should be false since no uclamp
1528 * accounting was performed at enqueue time and we can just return
1531 * Need to be careful of the following enqueue/dequeue ordering
1535 * // sched_uclamp_used gets enabled
1538 * // Must not decrement bucket->tasks here
1541 * where we could end up with stale data in uc_se and
1542 * bucket[uc_se->bucket_id].
1544 * The following check here eliminates the possibility of such race.
1546 if (unlikely(!uc_se->active))
1549 bucket = &uc_rq->bucket[uc_se->bucket_id];
1551 SCHED_WARN_ON(!bucket->tasks);
1552 if (likely(bucket->tasks))
1555 uc_se->active = false;
1558 * Keep "local max aggregation" simple and accept to (possibly)
1559 * overboost some RUNNABLE tasks in the same bucket.
1560 * The rq clamp bucket value is reset to its base value whenever
1561 * there are no more RUNNABLE tasks refcounting it.
1563 if (likely(bucket->tasks))
1566 rq_clamp = READ_ONCE(uc_rq->value);
1568 * Defensive programming: this should never happen. If it happens,
1569 * e.g. due to future modification, warn and fixup the expected value.
1571 SCHED_WARN_ON(bucket->value > rq_clamp);
1572 if (bucket->value >= rq_clamp) {
1573 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1574 WRITE_ONCE(uc_rq->value, bkt_clamp);
1578 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1580 enum uclamp_id clamp_id;
1583 * Avoid any overhead until uclamp is actually used by the userspace.
1585 * The condition is constructed such that a NOP is generated when
1586 * sched_uclamp_used is disabled.
1588 if (!static_branch_unlikely(&sched_uclamp_used))
1591 if (unlikely(!p->sched_class->uclamp_enabled))
1594 for_each_clamp_id(clamp_id)
1595 uclamp_rq_inc_id(rq, p, clamp_id);
1597 /* Reset clamp idle holding when there is one RUNNABLE task */
1598 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1599 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1602 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1604 enum uclamp_id clamp_id;
1607 * Avoid any overhead until uclamp is actually used by the userspace.
1609 * The condition is constructed such that a NOP is generated when
1610 * sched_uclamp_used is disabled.
1612 if (!static_branch_unlikely(&sched_uclamp_used))
1615 if (unlikely(!p->sched_class->uclamp_enabled))
1618 for_each_clamp_id(clamp_id)
1619 uclamp_rq_dec_id(rq, p, clamp_id);
1623 uclamp_update_active(struct task_struct *p)
1625 enum uclamp_id clamp_id;
1630 * Lock the task and the rq where the task is (or was) queued.
1632 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1633 * price to pay to safely serialize util_{min,max} updates with
1634 * enqueues, dequeues and migration operations.
1635 * This is the same locking schema used by __set_cpus_allowed_ptr().
1637 rq = task_rq_lock(p, &rf);
1640 * Setting the clamp bucket is serialized by task_rq_lock().
1641 * If the task is not yet RUNNABLE and its task_struct is not
1642 * affecting a valid clamp bucket, the next time it's enqueued,
1643 * it will already see the updated clamp bucket value.
1645 for_each_clamp_id(clamp_id) {
1646 if (p->uclamp[clamp_id].active) {
1647 uclamp_rq_dec_id(rq, p, clamp_id);
1648 uclamp_rq_inc_id(rq, p, clamp_id);
1652 task_rq_unlock(rq, p, &rf);
1655 #ifdef CONFIG_UCLAMP_TASK_GROUP
1657 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1659 struct css_task_iter it;
1660 struct task_struct *p;
1662 css_task_iter_start(css, 0, &it);
1663 while ((p = css_task_iter_next(&it)))
1664 uclamp_update_active(p);
1665 css_task_iter_end(&it);
1668 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1669 static void uclamp_update_root_tg(void)
1671 struct task_group *tg = &root_task_group;
1673 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1674 sysctl_sched_uclamp_util_min, false);
1675 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1676 sysctl_sched_uclamp_util_max, false);
1679 cpu_util_update_eff(&root_task_group.css);
1683 static void uclamp_update_root_tg(void) { }
1686 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1687 void *buffer, size_t *lenp, loff_t *ppos)
1689 bool update_root_tg = false;
1690 int old_min, old_max, old_min_rt;
1693 mutex_lock(&uclamp_mutex);
1694 old_min = sysctl_sched_uclamp_util_min;
1695 old_max = sysctl_sched_uclamp_util_max;
1696 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1698 result = proc_dointvec(table, write, buffer, lenp, ppos);
1704 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1705 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1706 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1712 if (old_min != sysctl_sched_uclamp_util_min) {
1713 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1714 sysctl_sched_uclamp_util_min, false);
1715 update_root_tg = true;
1717 if (old_max != sysctl_sched_uclamp_util_max) {
1718 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1719 sysctl_sched_uclamp_util_max, false);
1720 update_root_tg = true;
1723 if (update_root_tg) {
1724 static_branch_enable(&sched_uclamp_used);
1725 uclamp_update_root_tg();
1728 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1729 static_branch_enable(&sched_uclamp_used);
1730 uclamp_sync_util_min_rt_default();
1734 * We update all RUNNABLE tasks only when task groups are in use.
1735 * Otherwise, keep it simple and do just a lazy update at each next
1736 * task enqueue time.
1742 sysctl_sched_uclamp_util_min = old_min;
1743 sysctl_sched_uclamp_util_max = old_max;
1744 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1746 mutex_unlock(&uclamp_mutex);
1751 static int uclamp_validate(struct task_struct *p,
1752 const struct sched_attr *attr)
1754 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1755 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1757 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1758 util_min = attr->sched_util_min;
1760 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1764 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1765 util_max = attr->sched_util_max;
1767 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1771 if (util_min != -1 && util_max != -1 && util_min > util_max)
1775 * We have valid uclamp attributes; make sure uclamp is enabled.
1777 * We need to do that here, because enabling static branches is a
1778 * blocking operation which obviously cannot be done while holding
1781 static_branch_enable(&sched_uclamp_used);
1786 static bool uclamp_reset(const struct sched_attr *attr,
1787 enum uclamp_id clamp_id,
1788 struct uclamp_se *uc_se)
1790 /* Reset on sched class change for a non user-defined clamp value. */
1791 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1792 !uc_se->user_defined)
1795 /* Reset on sched_util_{min,max} == -1. */
1796 if (clamp_id == UCLAMP_MIN &&
1797 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1798 attr->sched_util_min == -1) {
1802 if (clamp_id == UCLAMP_MAX &&
1803 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1804 attr->sched_util_max == -1) {
1811 static void __setscheduler_uclamp(struct task_struct *p,
1812 const struct sched_attr *attr)
1814 enum uclamp_id clamp_id;
1816 for_each_clamp_id(clamp_id) {
1817 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1820 if (!uclamp_reset(attr, clamp_id, uc_se))
1824 * RT by default have a 100% boost value that could be modified
1827 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1828 value = sysctl_sched_uclamp_util_min_rt_default;
1830 value = uclamp_none(clamp_id);
1832 uclamp_se_set(uc_se, value, false);
1836 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1839 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1840 attr->sched_util_min != -1) {
1841 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1842 attr->sched_util_min, true);
1845 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1846 attr->sched_util_max != -1) {
1847 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1848 attr->sched_util_max, true);
1852 static void uclamp_fork(struct task_struct *p)
1854 enum uclamp_id clamp_id;
1857 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1858 * as the task is still at its early fork stages.
1860 for_each_clamp_id(clamp_id)
1861 p->uclamp[clamp_id].active = false;
1863 if (likely(!p->sched_reset_on_fork))
1866 for_each_clamp_id(clamp_id) {
1867 uclamp_se_set(&p->uclamp_req[clamp_id],
1868 uclamp_none(clamp_id), false);
1872 static void uclamp_post_fork(struct task_struct *p)
1874 uclamp_update_util_min_rt_default(p);
1877 static void __init init_uclamp_rq(struct rq *rq)
1879 enum uclamp_id clamp_id;
1880 struct uclamp_rq *uc_rq = rq->uclamp;
1882 for_each_clamp_id(clamp_id) {
1883 uc_rq[clamp_id] = (struct uclamp_rq) {
1884 .value = uclamp_none(clamp_id)
1888 rq->uclamp_flags = 0;
1891 static void __init init_uclamp(void)
1893 struct uclamp_se uc_max = {};
1894 enum uclamp_id clamp_id;
1897 for_each_possible_cpu(cpu)
1898 init_uclamp_rq(cpu_rq(cpu));
1900 for_each_clamp_id(clamp_id) {
1901 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1902 uclamp_none(clamp_id), false);
1905 /* System defaults allow max clamp values for both indexes */
1906 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1907 for_each_clamp_id(clamp_id) {
1908 uclamp_default[clamp_id] = uc_max;
1909 #ifdef CONFIG_UCLAMP_TASK_GROUP
1910 root_task_group.uclamp_req[clamp_id] = uc_max;
1911 root_task_group.uclamp[clamp_id] = uc_max;
1916 #else /* CONFIG_UCLAMP_TASK */
1917 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1918 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1919 static inline int uclamp_validate(struct task_struct *p,
1920 const struct sched_attr *attr)
1924 static void __setscheduler_uclamp(struct task_struct *p,
1925 const struct sched_attr *attr) { }
1926 static inline void uclamp_fork(struct task_struct *p) { }
1927 static inline void uclamp_post_fork(struct task_struct *p) { }
1928 static inline void init_uclamp(void) { }
1929 #endif /* CONFIG_UCLAMP_TASK */
1931 bool sched_task_on_rq(struct task_struct *p)
1933 return task_on_rq_queued(p);
1936 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1938 if (!(flags & ENQUEUE_NOCLOCK))
1939 update_rq_clock(rq);
1941 if (!(flags & ENQUEUE_RESTORE)) {
1942 sched_info_enqueue(rq, p);
1943 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1946 uclamp_rq_inc(rq, p);
1947 p->sched_class->enqueue_task(rq, p, flags);
1949 if (sched_core_enabled(rq))
1950 sched_core_enqueue(rq, p);
1953 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1955 if (sched_core_enabled(rq))
1956 sched_core_dequeue(rq, p);
1958 if (!(flags & DEQUEUE_NOCLOCK))
1959 update_rq_clock(rq);
1961 if (!(flags & DEQUEUE_SAVE)) {
1962 sched_info_dequeue(rq, p);
1963 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1966 uclamp_rq_dec(rq, p);
1967 p->sched_class->dequeue_task(rq, p, flags);
1970 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1972 enqueue_task(rq, p, flags);
1974 p->on_rq = TASK_ON_RQ_QUEUED;
1977 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1979 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1981 dequeue_task(rq, p, flags);
1984 static inline int __normal_prio(int policy, int rt_prio, int nice)
1988 if (dl_policy(policy))
1989 prio = MAX_DL_PRIO - 1;
1990 else if (rt_policy(policy))
1991 prio = MAX_RT_PRIO - 1 - rt_prio;
1993 prio = NICE_TO_PRIO(nice);
1999 * Calculate the expected normal priority: i.e. priority
2000 * without taking RT-inheritance into account. Might be
2001 * boosted by interactivity modifiers. Changes upon fork,
2002 * setprio syscalls, and whenever the interactivity
2003 * estimator recalculates.
2005 static inline int normal_prio(struct task_struct *p)
2007 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2011 * Calculate the current priority, i.e. the priority
2012 * taken into account by the scheduler. This value might
2013 * be boosted by RT tasks, or might be boosted by
2014 * interactivity modifiers. Will be RT if the task got
2015 * RT-boosted. If not then it returns p->normal_prio.
2017 static int effective_prio(struct task_struct *p)
2019 p->normal_prio = normal_prio(p);
2021 * If we are RT tasks or we were boosted to RT priority,
2022 * keep the priority unchanged. Otherwise, update priority
2023 * to the normal priority:
2025 if (!rt_prio(p->prio))
2026 return p->normal_prio;
2031 * task_curr - is this task currently executing on a CPU?
2032 * @p: the task in question.
2034 * Return: 1 if the task is currently executing. 0 otherwise.
2036 inline int task_curr(const struct task_struct *p)
2038 return cpu_curr(task_cpu(p)) == p;
2042 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2043 * use the balance_callback list if you want balancing.
2045 * this means any call to check_class_changed() must be followed by a call to
2046 * balance_callback().
2048 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2049 const struct sched_class *prev_class,
2052 if (prev_class != p->sched_class) {
2053 if (prev_class->switched_from)
2054 prev_class->switched_from(rq, p);
2056 p->sched_class->switched_to(rq, p);
2057 } else if (oldprio != p->prio || dl_task(p))
2058 p->sched_class->prio_changed(rq, p, oldprio);
2061 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2063 if (p->sched_class == rq->curr->sched_class)
2064 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2065 else if (p->sched_class > rq->curr->sched_class)
2069 * A queue event has occurred, and we're going to schedule. In
2070 * this case, we can save a useless back to back clock update.
2072 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2073 rq_clock_skip_update(rq);
2079 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2081 static int __set_cpus_allowed_ptr(struct task_struct *p,
2082 const struct cpumask *new_mask,
2085 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2087 if (likely(!p->migration_disabled))
2090 if (p->cpus_ptr != &p->cpus_mask)
2094 * Violates locking rules! see comment in __do_set_cpus_allowed().
2096 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2099 void migrate_disable(void)
2101 struct task_struct *p = current;
2103 if (p->migration_disabled) {
2104 p->migration_disabled++;
2109 this_rq()->nr_pinned++;
2110 p->migration_disabled = 1;
2113 EXPORT_SYMBOL_GPL(migrate_disable);
2115 void migrate_enable(void)
2117 struct task_struct *p = current;
2119 if (p->migration_disabled > 1) {
2120 p->migration_disabled--;
2125 * Ensure stop_task runs either before or after this, and that
2126 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2129 if (p->cpus_ptr != &p->cpus_mask)
2130 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2132 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2133 * regular cpus_mask, otherwise things that race (eg.
2134 * select_fallback_rq) get confused.
2137 p->migration_disabled = 0;
2138 this_rq()->nr_pinned--;
2141 EXPORT_SYMBOL_GPL(migrate_enable);
2143 static inline bool rq_has_pinned_tasks(struct rq *rq)
2145 return rq->nr_pinned;
2149 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2150 * __set_cpus_allowed_ptr() and select_fallback_rq().
2152 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2154 /* When not in the task's cpumask, no point in looking further. */
2155 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2158 /* migrate_disabled() must be allowed to finish. */
2159 if (is_migration_disabled(p))
2160 return cpu_online(cpu);
2162 /* Non kernel threads are not allowed during either online or offline. */
2163 if (!(p->flags & PF_KTHREAD))
2164 return cpu_active(cpu);
2166 /* KTHREAD_IS_PER_CPU is always allowed. */
2167 if (kthread_is_per_cpu(p))
2168 return cpu_online(cpu);
2170 /* Regular kernel threads don't get to stay during offline. */
2174 /* But are allowed during online. */
2175 return cpu_online(cpu);
2179 * This is how migration works:
2181 * 1) we invoke migration_cpu_stop() on the target CPU using
2183 * 2) stopper starts to run (implicitly forcing the migrated thread
2185 * 3) it checks whether the migrated task is still in the wrong runqueue.
2186 * 4) if it's in the wrong runqueue then the migration thread removes
2187 * it and puts it into the right queue.
2188 * 5) stopper completes and stop_one_cpu() returns and the migration
2193 * move_queued_task - move a queued task to new rq.
2195 * Returns (locked) new rq. Old rq's lock is released.
2197 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2198 struct task_struct *p, int new_cpu)
2200 lockdep_assert_rq_held(rq);
2202 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2203 set_task_cpu(p, new_cpu);
2206 rq = cpu_rq(new_cpu);
2209 BUG_ON(task_cpu(p) != new_cpu);
2210 activate_task(rq, p, 0);
2211 check_preempt_curr(rq, p, 0);
2216 struct migration_arg {
2217 struct task_struct *task;
2219 struct set_affinity_pending *pending;
2223 * @refs: number of wait_for_completion()
2224 * @stop_pending: is @stop_work in use
2226 struct set_affinity_pending {
2228 unsigned int stop_pending;
2229 struct completion done;
2230 struct cpu_stop_work stop_work;
2231 struct migration_arg arg;
2235 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2236 * this because either it can't run here any more (set_cpus_allowed()
2237 * away from this CPU, or CPU going down), or because we're
2238 * attempting to rebalance this task on exec (sched_exec).
2240 * So we race with normal scheduler movements, but that's OK, as long
2241 * as the task is no longer on this CPU.
2243 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2244 struct task_struct *p, int dest_cpu)
2246 /* Affinity changed (again). */
2247 if (!is_cpu_allowed(p, dest_cpu))
2250 update_rq_clock(rq);
2251 rq = move_queued_task(rq, rf, p, dest_cpu);
2257 * migration_cpu_stop - this will be executed by a highprio stopper thread
2258 * and performs thread migration by bumping thread off CPU then
2259 * 'pushing' onto another runqueue.
2261 static int migration_cpu_stop(void *data)
2263 struct migration_arg *arg = data;
2264 struct set_affinity_pending *pending = arg->pending;
2265 struct task_struct *p = arg->task;
2266 struct rq *rq = this_rq();
2267 bool complete = false;
2271 * The original target CPU might have gone down and we might
2272 * be on another CPU but it doesn't matter.
2274 local_irq_save(rf.flags);
2276 * We need to explicitly wake pending tasks before running
2277 * __migrate_task() such that we will not miss enforcing cpus_ptr
2278 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2280 flush_smp_call_function_from_idle();
2282 raw_spin_lock(&p->pi_lock);
2286 * If we were passed a pending, then ->stop_pending was set, thus
2287 * p->migration_pending must have remained stable.
2289 WARN_ON_ONCE(pending && pending != p->migration_pending);
2292 * If task_rq(p) != rq, it cannot be migrated here, because we're
2293 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2294 * we're holding p->pi_lock.
2296 if (task_rq(p) == rq) {
2297 if (is_migration_disabled(p))
2301 p->migration_pending = NULL;
2304 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2308 if (task_on_rq_queued(p))
2309 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2311 p->wake_cpu = arg->dest_cpu;
2314 * XXX __migrate_task() can fail, at which point we might end
2315 * up running on a dodgy CPU, AFAICT this can only happen
2316 * during CPU hotplug, at which point we'll get pushed out
2317 * anyway, so it's probably not a big deal.
2320 } else if (pending) {
2322 * This happens when we get migrated between migrate_enable()'s
2323 * preempt_enable() and scheduling the stopper task. At that
2324 * point we're a regular task again and not current anymore.
2326 * A !PREEMPT kernel has a giant hole here, which makes it far
2331 * The task moved before the stopper got to run. We're holding
2332 * ->pi_lock, so the allowed mask is stable - if it got
2333 * somewhere allowed, we're done.
2335 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2336 p->migration_pending = NULL;
2342 * When migrate_enable() hits a rq mis-match we can't reliably
2343 * determine is_migration_disabled() and so have to chase after
2346 WARN_ON_ONCE(!pending->stop_pending);
2347 task_rq_unlock(rq, p, &rf);
2348 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2349 &pending->arg, &pending->stop_work);
2354 pending->stop_pending = false;
2355 task_rq_unlock(rq, p, &rf);
2358 complete_all(&pending->done);
2363 int push_cpu_stop(void *arg)
2365 struct rq *lowest_rq = NULL, *rq = this_rq();
2366 struct task_struct *p = arg;
2368 raw_spin_lock_irq(&p->pi_lock);
2369 raw_spin_rq_lock(rq);
2371 if (task_rq(p) != rq)
2374 if (is_migration_disabled(p)) {
2375 p->migration_flags |= MDF_PUSH;
2379 p->migration_flags &= ~MDF_PUSH;
2381 if (p->sched_class->find_lock_rq)
2382 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2387 // XXX validate p is still the highest prio task
2388 if (task_rq(p) == rq) {
2389 deactivate_task(rq, p, 0);
2390 set_task_cpu(p, lowest_rq->cpu);
2391 activate_task(lowest_rq, p, 0);
2392 resched_curr(lowest_rq);
2395 double_unlock_balance(rq, lowest_rq);
2398 rq->push_busy = false;
2399 raw_spin_rq_unlock(rq);
2400 raw_spin_unlock_irq(&p->pi_lock);
2407 * sched_class::set_cpus_allowed must do the below, but is not required to
2408 * actually call this function.
2410 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2412 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2413 p->cpus_ptr = new_mask;
2417 cpumask_copy(&p->cpus_mask, new_mask);
2418 p->nr_cpus_allowed = cpumask_weight(new_mask);
2422 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2424 struct rq *rq = task_rq(p);
2425 bool queued, running;
2428 * This here violates the locking rules for affinity, since we're only
2429 * supposed to change these variables while holding both rq->lock and
2432 * HOWEVER, it magically works, because ttwu() is the only code that
2433 * accesses these variables under p->pi_lock and only does so after
2434 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2435 * before finish_task().
2437 * XXX do further audits, this smells like something putrid.
2439 if (flags & SCA_MIGRATE_DISABLE)
2440 SCHED_WARN_ON(!p->on_cpu);
2442 lockdep_assert_held(&p->pi_lock);
2444 queued = task_on_rq_queued(p);
2445 running = task_current(rq, p);
2449 * Because __kthread_bind() calls this on blocked tasks without
2452 lockdep_assert_rq_held(rq);
2453 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2456 put_prev_task(rq, p);
2458 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2461 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2463 set_next_task(rq, p);
2466 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2468 __do_set_cpus_allowed(p, new_mask, 0);
2472 * This function is wildly self concurrent; here be dragons.
2475 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2476 * designated task is enqueued on an allowed CPU. If that task is currently
2477 * running, we have to kick it out using the CPU stopper.
2479 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2482 * Initial conditions: P0->cpus_mask = [0, 1]
2486 * migrate_disable();
2488 * set_cpus_allowed_ptr(P0, [1]);
2490 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2491 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2492 * This means we need the following scheme:
2496 * migrate_disable();
2498 * set_cpus_allowed_ptr(P0, [1]);
2502 * __set_cpus_allowed_ptr();
2503 * <wakes local stopper>
2504 * `--> <woken on migration completion>
2506 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2507 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2508 * task p are serialized by p->pi_lock, which we can leverage: the one that
2509 * should come into effect at the end of the Migrate-Disable region is the last
2510 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2511 * but we still need to properly signal those waiting tasks at the appropriate
2514 * This is implemented using struct set_affinity_pending. The first
2515 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2516 * setup an instance of that struct and install it on the targeted task_struct.
2517 * Any and all further callers will reuse that instance. Those then wait for
2518 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2519 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2522 * (1) In the cases covered above. There is one more where the completion is
2523 * signaled within affine_move_task() itself: when a subsequent affinity request
2524 * occurs after the stopper bailed out due to the targeted task still being
2525 * Migrate-Disable. Consider:
2527 * Initial conditions: P0->cpus_mask = [0, 1]
2531 * migrate_disable();
2533 * set_cpus_allowed_ptr(P0, [1]);
2536 * migration_cpu_stop()
2537 * is_migration_disabled()
2539 * set_cpus_allowed_ptr(P0, [0, 1]);
2540 * <signal completion>
2543 * Note that the above is safe vs a concurrent migrate_enable(), as any
2544 * pending affinity completion is preceded by an uninstallation of
2545 * p->migration_pending done with p->pi_lock held.
2547 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2548 int dest_cpu, unsigned int flags)
2550 struct set_affinity_pending my_pending = { }, *pending = NULL;
2551 bool stop_pending, complete = false;
2553 /* Can the task run on the task's current CPU? If so, we're done */
2554 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2555 struct task_struct *push_task = NULL;
2557 if ((flags & SCA_MIGRATE_ENABLE) &&
2558 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2559 rq->push_busy = true;
2560 push_task = get_task_struct(p);
2564 * If there are pending waiters, but no pending stop_work,
2565 * then complete now.
2567 pending = p->migration_pending;
2568 if (pending && !pending->stop_pending) {
2569 p->migration_pending = NULL;
2573 task_rq_unlock(rq, p, rf);
2576 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2581 complete_all(&pending->done);
2586 if (!(flags & SCA_MIGRATE_ENABLE)) {
2587 /* serialized by p->pi_lock */
2588 if (!p->migration_pending) {
2589 /* Install the request */
2590 refcount_set(&my_pending.refs, 1);
2591 init_completion(&my_pending.done);
2592 my_pending.arg = (struct migration_arg) {
2594 .dest_cpu = dest_cpu,
2595 .pending = &my_pending,
2598 p->migration_pending = &my_pending;
2600 pending = p->migration_pending;
2601 refcount_inc(&pending->refs);
2603 * Affinity has changed, but we've already installed a
2604 * pending. migration_cpu_stop() *must* see this, else
2605 * we risk a completion of the pending despite having a
2606 * task on a disallowed CPU.
2608 * Serialized by p->pi_lock, so this is safe.
2610 pending->arg.dest_cpu = dest_cpu;
2613 pending = p->migration_pending;
2615 * - !MIGRATE_ENABLE:
2616 * we'll have installed a pending if there wasn't one already.
2619 * we're here because the current CPU isn't matching anymore,
2620 * the only way that can happen is because of a concurrent
2621 * set_cpus_allowed_ptr() call, which should then still be
2622 * pending completion.
2624 * Either way, we really should have a @pending here.
2626 if (WARN_ON_ONCE(!pending)) {
2627 task_rq_unlock(rq, p, rf);
2631 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2633 * MIGRATE_ENABLE gets here because 'p == current', but for
2634 * anything else we cannot do is_migration_disabled(), punt
2635 * and have the stopper function handle it all race-free.
2637 stop_pending = pending->stop_pending;
2639 pending->stop_pending = true;
2641 if (flags & SCA_MIGRATE_ENABLE)
2642 p->migration_flags &= ~MDF_PUSH;
2644 task_rq_unlock(rq, p, rf);
2646 if (!stop_pending) {
2647 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2648 &pending->arg, &pending->stop_work);
2651 if (flags & SCA_MIGRATE_ENABLE)
2655 if (!is_migration_disabled(p)) {
2656 if (task_on_rq_queued(p))
2657 rq = move_queued_task(rq, rf, p, dest_cpu);
2659 if (!pending->stop_pending) {
2660 p->migration_pending = NULL;
2664 task_rq_unlock(rq, p, rf);
2667 complete_all(&pending->done);
2670 wait_for_completion(&pending->done);
2672 if (refcount_dec_and_test(&pending->refs))
2673 wake_up_var(&pending->refs); /* No UaF, just an address */
2676 * Block the original owner of &pending until all subsequent callers
2677 * have seen the completion and decremented the refcount
2679 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2682 WARN_ON_ONCE(my_pending.stop_pending);
2688 * Change a given task's CPU affinity. Migrate the thread to a
2689 * proper CPU and schedule it away if the CPU it's executing on
2690 * is removed from the allowed bitmask.
2692 * NOTE: the caller must have a valid reference to the task, the
2693 * task must not exit() & deallocate itself prematurely. The
2694 * call is not atomic; no spinlocks may be held.
2696 static int __set_cpus_allowed_ptr(struct task_struct *p,
2697 const struct cpumask *new_mask,
2700 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2701 unsigned int dest_cpu;
2706 rq = task_rq_lock(p, &rf);
2707 update_rq_clock(rq);
2709 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2711 * Kernel threads are allowed on online && !active CPUs,
2712 * however, during cpu-hot-unplug, even these might get pushed
2713 * away if not KTHREAD_IS_PER_CPU.
2715 * Specifically, migration_disabled() tasks must not fail the
2716 * cpumask_any_and_distribute() pick below, esp. so on
2717 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2718 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2720 cpu_valid_mask = cpu_online_mask;
2724 * Must re-check here, to close a race against __kthread_bind(),
2725 * sched_setaffinity() is not guaranteed to observe the flag.
2727 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2732 if (!(flags & SCA_MIGRATE_ENABLE)) {
2733 if (cpumask_equal(&p->cpus_mask, new_mask))
2736 if (WARN_ON_ONCE(p == current &&
2737 is_migration_disabled(p) &&
2738 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2745 * Picking a ~random cpu helps in cases where we are changing affinity
2746 * for groups of tasks (ie. cpuset), so that load balancing is not
2747 * immediately required to distribute the tasks within their new mask.
2749 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2750 if (dest_cpu >= nr_cpu_ids) {
2755 __do_set_cpus_allowed(p, new_mask, flags);
2757 return affine_move_task(rq, p, &rf, dest_cpu, flags);
2760 task_rq_unlock(rq, p, &rf);
2765 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2767 return __set_cpus_allowed_ptr(p, new_mask, 0);
2769 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2771 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2773 #ifdef CONFIG_SCHED_DEBUG
2774 unsigned int state = READ_ONCE(p->__state);
2777 * We should never call set_task_cpu() on a blocked task,
2778 * ttwu() will sort out the placement.
2780 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
2783 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2784 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2785 * time relying on p->on_rq.
2787 WARN_ON_ONCE(state == TASK_RUNNING &&
2788 p->sched_class == &fair_sched_class &&
2789 (p->on_rq && !task_on_rq_migrating(p)));
2791 #ifdef CONFIG_LOCKDEP
2793 * The caller should hold either p->pi_lock or rq->lock, when changing
2794 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2796 * sched_move_task() holds both and thus holding either pins the cgroup,
2799 * Furthermore, all task_rq users should acquire both locks, see
2802 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2803 lockdep_is_held(__rq_lockp(task_rq(p)))));
2806 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2808 WARN_ON_ONCE(!cpu_online(new_cpu));
2810 WARN_ON_ONCE(is_migration_disabled(p));
2813 trace_sched_migrate_task(p, new_cpu);
2815 if (task_cpu(p) != new_cpu) {
2816 if (p->sched_class->migrate_task_rq)
2817 p->sched_class->migrate_task_rq(p, new_cpu);
2818 p->se.nr_migrations++;
2820 perf_event_task_migrate(p);
2823 __set_task_cpu(p, new_cpu);
2826 #ifdef CONFIG_NUMA_BALANCING
2827 static void __migrate_swap_task(struct task_struct *p, int cpu)
2829 if (task_on_rq_queued(p)) {
2830 struct rq *src_rq, *dst_rq;
2831 struct rq_flags srf, drf;
2833 src_rq = task_rq(p);
2834 dst_rq = cpu_rq(cpu);
2836 rq_pin_lock(src_rq, &srf);
2837 rq_pin_lock(dst_rq, &drf);
2839 deactivate_task(src_rq, p, 0);
2840 set_task_cpu(p, cpu);
2841 activate_task(dst_rq, p, 0);
2842 check_preempt_curr(dst_rq, p, 0);
2844 rq_unpin_lock(dst_rq, &drf);
2845 rq_unpin_lock(src_rq, &srf);
2849 * Task isn't running anymore; make it appear like we migrated
2850 * it before it went to sleep. This means on wakeup we make the
2851 * previous CPU our target instead of where it really is.
2857 struct migration_swap_arg {
2858 struct task_struct *src_task, *dst_task;
2859 int src_cpu, dst_cpu;
2862 static int migrate_swap_stop(void *data)
2864 struct migration_swap_arg *arg = data;
2865 struct rq *src_rq, *dst_rq;
2868 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2871 src_rq = cpu_rq(arg->src_cpu);
2872 dst_rq = cpu_rq(arg->dst_cpu);
2874 double_raw_lock(&arg->src_task->pi_lock,
2875 &arg->dst_task->pi_lock);
2876 double_rq_lock(src_rq, dst_rq);
2878 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2881 if (task_cpu(arg->src_task) != arg->src_cpu)
2884 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2887 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2890 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2891 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2896 double_rq_unlock(src_rq, dst_rq);
2897 raw_spin_unlock(&arg->dst_task->pi_lock);
2898 raw_spin_unlock(&arg->src_task->pi_lock);
2904 * Cross migrate two tasks
2906 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2907 int target_cpu, int curr_cpu)
2909 struct migration_swap_arg arg;
2912 arg = (struct migration_swap_arg){
2914 .src_cpu = curr_cpu,
2916 .dst_cpu = target_cpu,
2919 if (arg.src_cpu == arg.dst_cpu)
2923 * These three tests are all lockless; this is OK since all of them
2924 * will be re-checked with proper locks held further down the line.
2926 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2929 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2932 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2935 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2936 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2941 #endif /* CONFIG_NUMA_BALANCING */
2944 * wait_task_inactive - wait for a thread to unschedule.
2946 * If @match_state is nonzero, it's the @p->state value just checked and
2947 * not expected to change. If it changes, i.e. @p might have woken up,
2948 * then return zero. When we succeed in waiting for @p to be off its CPU,
2949 * we return a positive number (its total switch count). If a second call
2950 * a short while later returns the same number, the caller can be sure that
2951 * @p has remained unscheduled the whole time.
2953 * The caller must ensure that the task *will* unschedule sometime soon,
2954 * else this function might spin for a *long* time. This function can't
2955 * be called with interrupts off, or it may introduce deadlock with
2956 * smp_call_function() if an IPI is sent by the same process we are
2957 * waiting to become inactive.
2959 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2961 int running, queued;
2968 * We do the initial early heuristics without holding
2969 * any task-queue locks at all. We'll only try to get
2970 * the runqueue lock when things look like they will
2976 * If the task is actively running on another CPU
2977 * still, just relax and busy-wait without holding
2980 * NOTE! Since we don't hold any locks, it's not
2981 * even sure that "rq" stays as the right runqueue!
2982 * But we don't care, since "task_running()" will
2983 * return false if the runqueue has changed and p
2984 * is actually now running somewhere else!
2986 while (task_running(rq, p)) {
2987 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
2993 * Ok, time to look more closely! We need the rq
2994 * lock now, to be *sure*. If we're wrong, we'll
2995 * just go back and repeat.
2997 rq = task_rq_lock(p, &rf);
2998 trace_sched_wait_task(p);
2999 running = task_running(rq, p);
3000 queued = task_on_rq_queued(p);
3002 if (!match_state || READ_ONCE(p->__state) == match_state)
3003 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3004 task_rq_unlock(rq, p, &rf);
3007 * If it changed from the expected state, bail out now.
3009 if (unlikely(!ncsw))
3013 * Was it really running after all now that we
3014 * checked with the proper locks actually held?
3016 * Oops. Go back and try again..
3018 if (unlikely(running)) {
3024 * It's not enough that it's not actively running,
3025 * it must be off the runqueue _entirely_, and not
3028 * So if it was still runnable (but just not actively
3029 * running right now), it's preempted, and we should
3030 * yield - it could be a while.
3032 if (unlikely(queued)) {
3033 ktime_t to = NSEC_PER_SEC / HZ;
3035 set_current_state(TASK_UNINTERRUPTIBLE);
3036 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
3041 * Ahh, all good. It wasn't running, and it wasn't
3042 * runnable, which means that it will never become
3043 * running in the future either. We're all done!
3052 * kick_process - kick a running thread to enter/exit the kernel
3053 * @p: the to-be-kicked thread
3055 * Cause a process which is running on another CPU to enter
3056 * kernel-mode, without any delay. (to get signals handled.)
3058 * NOTE: this function doesn't have to take the runqueue lock,
3059 * because all it wants to ensure is that the remote task enters
3060 * the kernel. If the IPI races and the task has been migrated
3061 * to another CPU then no harm is done and the purpose has been
3064 void kick_process(struct task_struct *p)
3070 if ((cpu != smp_processor_id()) && task_curr(p))
3071 smp_send_reschedule(cpu);
3074 EXPORT_SYMBOL_GPL(kick_process);
3077 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3079 * A few notes on cpu_active vs cpu_online:
3081 * - cpu_active must be a subset of cpu_online
3083 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3084 * see __set_cpus_allowed_ptr(). At this point the newly online
3085 * CPU isn't yet part of the sched domains, and balancing will not
3088 * - on CPU-down we clear cpu_active() to mask the sched domains and
3089 * avoid the load balancer to place new tasks on the to be removed
3090 * CPU. Existing tasks will remain running there and will be taken
3093 * This means that fallback selection must not select !active CPUs.
3094 * And can assume that any active CPU must be online. Conversely
3095 * select_task_rq() below may allow selection of !active CPUs in order
3096 * to satisfy the above rules.
3098 static int select_fallback_rq(int cpu, struct task_struct *p)
3100 int nid = cpu_to_node(cpu);
3101 const struct cpumask *nodemask = NULL;
3102 enum { cpuset, possible, fail } state = cpuset;
3106 * If the node that the CPU is on has been offlined, cpu_to_node()
3107 * will return -1. There is no CPU on the node, and we should
3108 * select the CPU on the other node.
3111 nodemask = cpumask_of_node(nid);
3113 /* Look for allowed, online CPU in same node. */
3114 for_each_cpu(dest_cpu, nodemask) {
3115 if (!cpu_active(dest_cpu))
3117 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
3123 /* Any allowed, online CPU? */
3124 for_each_cpu(dest_cpu, p->cpus_ptr) {
3125 if (!is_cpu_allowed(p, dest_cpu))
3131 /* No more Mr. Nice Guy. */
3134 if (IS_ENABLED(CONFIG_CPUSETS)) {
3135 cpuset_cpus_allowed_fallback(p);
3142 * XXX When called from select_task_rq() we only
3143 * hold p->pi_lock and again violate locking order.
3145 * More yuck to audit.
3147 do_set_cpus_allowed(p, cpu_possible_mask);
3158 if (state != cpuset) {
3160 * Don't tell them about moving exiting tasks or
3161 * kernel threads (both mm NULL), since they never
3164 if (p->mm && printk_ratelimit()) {
3165 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3166 task_pid_nr(p), p->comm, cpu);
3174 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3177 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3179 lockdep_assert_held(&p->pi_lock);
3181 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3182 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3184 cpu = cpumask_any(p->cpus_ptr);
3187 * In order not to call set_task_cpu() on a blocking task we need
3188 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3191 * Since this is common to all placement strategies, this lives here.
3193 * [ this allows ->select_task() to simply return task_cpu(p) and
3194 * not worry about this generic constraint ]
3196 if (unlikely(!is_cpu_allowed(p, cpu)))
3197 cpu = select_fallback_rq(task_cpu(p), p);
3202 void sched_set_stop_task(int cpu, struct task_struct *stop)
3204 static struct lock_class_key stop_pi_lock;
3205 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3206 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3210 * Make it appear like a SCHED_FIFO task, its something
3211 * userspace knows about and won't get confused about.
3213 * Also, it will make PI more or less work without too
3214 * much confusion -- but then, stop work should not
3215 * rely on PI working anyway.
3217 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3219 stop->sched_class = &stop_sched_class;
3222 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3223 * adjust the effective priority of a task. As a result,
3224 * rt_mutex_setprio() can trigger (RT) balancing operations,
3225 * which can then trigger wakeups of the stop thread to push
3226 * around the current task.
3228 * The stop task itself will never be part of the PI-chain, it
3229 * never blocks, therefore that ->pi_lock recursion is safe.
3230 * Tell lockdep about this by placing the stop->pi_lock in its
3233 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3236 cpu_rq(cpu)->stop = stop;
3240 * Reset it back to a normal scheduling class so that
3241 * it can die in pieces.
3243 old_stop->sched_class = &rt_sched_class;
3247 #else /* CONFIG_SMP */
3249 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3250 const struct cpumask *new_mask,
3253 return set_cpus_allowed_ptr(p, new_mask);
3256 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3258 static inline bool rq_has_pinned_tasks(struct rq *rq)
3263 #endif /* !CONFIG_SMP */
3266 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3270 if (!schedstat_enabled())
3276 if (cpu == rq->cpu) {
3277 __schedstat_inc(rq->ttwu_local);
3278 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3280 struct sched_domain *sd;
3282 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3284 for_each_domain(rq->cpu, sd) {
3285 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3286 __schedstat_inc(sd->ttwu_wake_remote);
3293 if (wake_flags & WF_MIGRATED)
3294 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3295 #endif /* CONFIG_SMP */
3297 __schedstat_inc(rq->ttwu_count);
3298 __schedstat_inc(p->se.statistics.nr_wakeups);
3300 if (wake_flags & WF_SYNC)
3301 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3305 * Mark the task runnable and perform wakeup-preemption.
3307 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3308 struct rq_flags *rf)
3310 check_preempt_curr(rq, p, wake_flags);
3311 WRITE_ONCE(p->__state, TASK_RUNNING);
3312 trace_sched_wakeup(p);
3315 if (p->sched_class->task_woken) {
3317 * Our task @p is fully woken up and running; so it's safe to
3318 * drop the rq->lock, hereafter rq is only used for statistics.
3320 rq_unpin_lock(rq, rf);
3321 p->sched_class->task_woken(rq, p);
3322 rq_repin_lock(rq, rf);
3325 if (rq->idle_stamp) {
3326 u64 delta = rq_clock(rq) - rq->idle_stamp;
3327 u64 max = 2*rq->max_idle_balance_cost;
3329 update_avg(&rq->avg_idle, delta);
3331 if (rq->avg_idle > max)
3334 rq->wake_stamp = jiffies;
3335 rq->wake_avg_idle = rq->avg_idle / 2;
3343 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3344 struct rq_flags *rf)
3346 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3348 lockdep_assert_rq_held(rq);
3350 if (p->sched_contributes_to_load)
3351 rq->nr_uninterruptible--;
3354 if (wake_flags & WF_MIGRATED)
3355 en_flags |= ENQUEUE_MIGRATED;
3359 delayacct_blkio_end(p);
3360 atomic_dec(&task_rq(p)->nr_iowait);
3363 activate_task(rq, p, en_flags);
3364 ttwu_do_wakeup(rq, p, wake_flags, rf);
3368 * Consider @p being inside a wait loop:
3371 * set_current_state(TASK_UNINTERRUPTIBLE);
3378 * __set_current_state(TASK_RUNNING);
3380 * between set_current_state() and schedule(). In this case @p is still
3381 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3384 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3385 * then schedule() must still happen and p->state can be changed to
3386 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3387 * need to do a full wakeup with enqueue.
3389 * Returns: %true when the wakeup is done,
3392 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3398 rq = __task_rq_lock(p, &rf);
3399 if (task_on_rq_queued(p)) {
3400 /* check_preempt_curr() may use rq clock */
3401 update_rq_clock(rq);
3402 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3405 __task_rq_unlock(rq, &rf);
3411 void sched_ttwu_pending(void *arg)
3413 struct llist_node *llist = arg;
3414 struct rq *rq = this_rq();
3415 struct task_struct *p, *t;
3422 * rq::ttwu_pending racy indication of out-standing wakeups.
3423 * Races such that false-negatives are possible, since they
3424 * are shorter lived that false-positives would be.
3426 WRITE_ONCE(rq->ttwu_pending, 0);
3428 rq_lock_irqsave(rq, &rf);
3429 update_rq_clock(rq);
3431 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3432 if (WARN_ON_ONCE(p->on_cpu))
3433 smp_cond_load_acquire(&p->on_cpu, !VAL);
3435 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3436 set_task_cpu(p, cpu_of(rq));
3438 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3441 rq_unlock_irqrestore(rq, &rf);
3444 void send_call_function_single_ipi(int cpu)
3446 struct rq *rq = cpu_rq(cpu);
3448 if (!set_nr_if_polling(rq->idle))
3449 arch_send_call_function_single_ipi(cpu);
3451 trace_sched_wake_idle_without_ipi(cpu);
3455 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3456 * necessary. The wakee CPU on receipt of the IPI will queue the task
3457 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3458 * of the wakeup instead of the waker.
3460 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3462 struct rq *rq = cpu_rq(cpu);
3464 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3466 WRITE_ONCE(rq->ttwu_pending, 1);
3467 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3470 void wake_up_if_idle(int cpu)
3472 struct rq *rq = cpu_rq(cpu);
3477 if (!is_idle_task(rcu_dereference(rq->curr)))
3480 if (set_nr_if_polling(rq->idle)) {
3481 trace_sched_wake_idle_without_ipi(cpu);
3483 rq_lock_irqsave(rq, &rf);
3484 if (is_idle_task(rq->curr))
3485 smp_send_reschedule(cpu);
3486 /* Else CPU is not idle, do nothing here: */
3487 rq_unlock_irqrestore(rq, &rf);
3494 bool cpus_share_cache(int this_cpu, int that_cpu)
3496 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3499 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3502 * Do not complicate things with the async wake_list while the CPU is
3505 if (!cpu_active(cpu))
3509 * If the CPU does not share cache, then queue the task on the
3510 * remote rqs wakelist to avoid accessing remote data.
3512 if (!cpus_share_cache(smp_processor_id(), cpu))
3516 * If the task is descheduling and the only running task on the
3517 * CPU then use the wakelist to offload the task activation to
3518 * the soon-to-be-idle CPU as the current CPU is likely busy.
3519 * nr_running is checked to avoid unnecessary task stacking.
3521 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3527 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3529 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3530 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3533 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3534 __ttwu_queue_wakelist(p, cpu, wake_flags);
3541 #else /* !CONFIG_SMP */
3543 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3548 #endif /* CONFIG_SMP */
3550 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3552 struct rq *rq = cpu_rq(cpu);
3555 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3559 update_rq_clock(rq);
3560 ttwu_do_activate(rq, p, wake_flags, &rf);
3565 * Notes on Program-Order guarantees on SMP systems.
3569 * The basic program-order guarantee on SMP systems is that when a task [t]
3570 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3571 * execution on its new CPU [c1].
3573 * For migration (of runnable tasks) this is provided by the following means:
3575 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3576 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3577 * rq(c1)->lock (if not at the same time, then in that order).
3578 * C) LOCK of the rq(c1)->lock scheduling in task
3580 * Release/acquire chaining guarantees that B happens after A and C after B.
3581 * Note: the CPU doing B need not be c0 or c1
3590 * UNLOCK rq(0)->lock
3592 * LOCK rq(0)->lock // orders against CPU0
3594 * UNLOCK rq(0)->lock
3598 * UNLOCK rq(1)->lock
3600 * LOCK rq(1)->lock // orders against CPU2
3603 * UNLOCK rq(1)->lock
3606 * BLOCKING -- aka. SLEEP + WAKEUP
3608 * For blocking we (obviously) need to provide the same guarantee as for
3609 * migration. However the means are completely different as there is no lock
3610 * chain to provide order. Instead we do:
3612 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3613 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3617 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3619 * LOCK rq(0)->lock LOCK X->pi_lock
3622 * smp_store_release(X->on_cpu, 0);
3624 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3630 * X->state = RUNNING
3631 * UNLOCK rq(2)->lock
3633 * LOCK rq(2)->lock // orders against CPU1
3636 * UNLOCK rq(2)->lock
3639 * UNLOCK rq(0)->lock
3642 * However, for wakeups there is a second guarantee we must provide, namely we
3643 * must ensure that CONDITION=1 done by the caller can not be reordered with
3644 * accesses to the task state; see try_to_wake_up() and set_current_state().
3648 * try_to_wake_up - wake up a thread
3649 * @p: the thread to be awakened
3650 * @state: the mask of task states that can be woken
3651 * @wake_flags: wake modifier flags (WF_*)
3653 * Conceptually does:
3655 * If (@state & @p->state) @p->state = TASK_RUNNING.
3657 * If the task was not queued/runnable, also place it back on a runqueue.
3659 * This function is atomic against schedule() which would dequeue the task.
3661 * It issues a full memory barrier before accessing @p->state, see the comment
3662 * with set_current_state().
3664 * Uses p->pi_lock to serialize against concurrent wake-ups.
3666 * Relies on p->pi_lock stabilizing:
3669 * - p->sched_task_group
3670 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3672 * Tries really hard to only take one task_rq(p)->lock for performance.
3673 * Takes rq->lock in:
3674 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3675 * - ttwu_queue() -- new rq, for enqueue of the task;
3676 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3678 * As a consequence we race really badly with just about everything. See the
3679 * many memory barriers and their comments for details.
3681 * Return: %true if @p->state changes (an actual wakeup was done),
3685 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3687 unsigned long flags;
3688 int cpu, success = 0;
3693 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3694 * == smp_processor_id()'. Together this means we can special
3695 * case the whole 'p->on_rq && ttwu_runnable()' case below
3696 * without taking any locks.
3699 * - we rely on Program-Order guarantees for all the ordering,
3700 * - we're serialized against set_special_state() by virtue of
3701 * it disabling IRQs (this allows not taking ->pi_lock).
3703 if (!(READ_ONCE(p->__state) & state))
3707 trace_sched_waking(p);
3708 WRITE_ONCE(p->__state, TASK_RUNNING);
3709 trace_sched_wakeup(p);
3714 * If we are going to wake up a thread waiting for CONDITION we
3715 * need to ensure that CONDITION=1 done by the caller can not be
3716 * reordered with p->state check below. This pairs with smp_store_mb()
3717 * in set_current_state() that the waiting thread does.
3719 raw_spin_lock_irqsave(&p->pi_lock, flags);
3720 smp_mb__after_spinlock();
3721 if (!(READ_ONCE(p->__state) & state))
3724 trace_sched_waking(p);
3726 /* We're going to change ->state: */
3730 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3731 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3732 * in smp_cond_load_acquire() below.
3734 * sched_ttwu_pending() try_to_wake_up()
3735 * STORE p->on_rq = 1 LOAD p->state
3738 * __schedule() (switch to task 'p')
3739 * LOCK rq->lock smp_rmb();
3740 * smp_mb__after_spinlock();
3744 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3746 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3747 * __schedule(). See the comment for smp_mb__after_spinlock().
3749 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3752 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3757 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3758 * possible to, falsely, observe p->on_cpu == 0.
3760 * One must be running (->on_cpu == 1) in order to remove oneself
3761 * from the runqueue.
3763 * __schedule() (switch to task 'p') try_to_wake_up()
3764 * STORE p->on_cpu = 1 LOAD p->on_rq
3767 * __schedule() (put 'p' to sleep)
3768 * LOCK rq->lock smp_rmb();
3769 * smp_mb__after_spinlock();
3770 * STORE p->on_rq = 0 LOAD p->on_cpu
3772 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3773 * __schedule(). See the comment for smp_mb__after_spinlock().
3775 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3776 * schedule()'s deactivate_task() has 'happened' and p will no longer
3777 * care about it's own p->state. See the comment in __schedule().
3779 smp_acquire__after_ctrl_dep();
3782 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3783 * == 0), which means we need to do an enqueue, change p->state to
3784 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3785 * enqueue, such as ttwu_queue_wakelist().
3787 WRITE_ONCE(p->__state, TASK_WAKING);
3790 * If the owning (remote) CPU is still in the middle of schedule() with
3791 * this task as prev, considering queueing p on the remote CPUs wake_list
3792 * which potentially sends an IPI instead of spinning on p->on_cpu to
3793 * let the waker make forward progress. This is safe because IRQs are
3794 * disabled and the IPI will deliver after on_cpu is cleared.
3796 * Ensure we load task_cpu(p) after p->on_cpu:
3798 * set_task_cpu(p, cpu);
3799 * STORE p->cpu = @cpu
3800 * __schedule() (switch to task 'p')
3802 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3803 * STORE p->on_cpu = 1 LOAD p->cpu
3805 * to ensure we observe the correct CPU on which the task is currently
3808 if (smp_load_acquire(&p->on_cpu) &&
3809 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3813 * If the owning (remote) CPU is still in the middle of schedule() with
3814 * this task as prev, wait until it's done referencing the task.
3816 * Pairs with the smp_store_release() in finish_task().
3818 * This ensures that tasks getting woken will be fully ordered against
3819 * their previous state and preserve Program Order.
3821 smp_cond_load_acquire(&p->on_cpu, !VAL);
3823 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3824 if (task_cpu(p) != cpu) {
3826 delayacct_blkio_end(p);
3827 atomic_dec(&task_rq(p)->nr_iowait);
3830 wake_flags |= WF_MIGRATED;
3831 psi_ttwu_dequeue(p);
3832 set_task_cpu(p, cpu);
3836 #endif /* CONFIG_SMP */
3838 ttwu_queue(p, cpu, wake_flags);
3840 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3843 ttwu_stat(p, task_cpu(p), wake_flags);
3850 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3851 * @p: Process for which the function is to be invoked, can be @current.
3852 * @func: Function to invoke.
3853 * @arg: Argument to function.
3855 * If the specified task can be quickly locked into a definite state
3856 * (either sleeping or on a given runqueue), arrange to keep it in that
3857 * state while invoking @func(@arg). This function can use ->on_rq and
3858 * task_curr() to work out what the state is, if required. Given that
3859 * @func can be invoked with a runqueue lock held, it had better be quite
3863 * @false if the task slipped out from under the locks.
3864 * @true if the task was locked onto a runqueue or is sleeping.
3865 * However, @func can override this by returning @false.
3867 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3873 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3875 rq = __task_rq_lock(p, &rf);
3876 if (task_rq(p) == rq)
3880 switch (READ_ONCE(p->__state)) {
3885 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3890 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
3895 * wake_up_process - Wake up a specific process
3896 * @p: The process to be woken up.
3898 * Attempt to wake up the nominated process and move it to the set of runnable
3901 * Return: 1 if the process was woken up, 0 if it was already running.
3903 * This function executes a full memory barrier before accessing the task state.
3905 int wake_up_process(struct task_struct *p)
3907 return try_to_wake_up(p, TASK_NORMAL, 0);
3909 EXPORT_SYMBOL(wake_up_process);
3911 int wake_up_state(struct task_struct *p, unsigned int state)
3913 return try_to_wake_up(p, state, 0);
3917 * Perform scheduler related setup for a newly forked process p.
3918 * p is forked by current.
3920 * __sched_fork() is basic setup used by init_idle() too:
3922 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3927 p->se.exec_start = 0;
3928 p->se.sum_exec_runtime = 0;
3929 p->se.prev_sum_exec_runtime = 0;
3930 p->se.nr_migrations = 0;
3932 INIT_LIST_HEAD(&p->se.group_node);
3934 #ifdef CONFIG_FAIR_GROUP_SCHED
3935 p->se.cfs_rq = NULL;
3938 #ifdef CONFIG_SCHEDSTATS
3939 /* Even if schedstat is disabled, there should not be garbage */
3940 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3943 RB_CLEAR_NODE(&p->dl.rb_node);
3944 init_dl_task_timer(&p->dl);
3945 init_dl_inactive_task_timer(&p->dl);
3946 __dl_clear_params(p);
3948 INIT_LIST_HEAD(&p->rt.run_list);
3950 p->rt.time_slice = sched_rr_timeslice;
3954 #ifdef CONFIG_PREEMPT_NOTIFIERS
3955 INIT_HLIST_HEAD(&p->preempt_notifiers);
3958 #ifdef CONFIG_COMPACTION
3959 p->capture_control = NULL;
3961 init_numa_balancing(clone_flags, p);
3963 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3964 p->migration_pending = NULL;
3968 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3970 #ifdef CONFIG_NUMA_BALANCING
3972 void set_numabalancing_state(bool enabled)
3975 static_branch_enable(&sched_numa_balancing);
3977 static_branch_disable(&sched_numa_balancing);
3980 #ifdef CONFIG_PROC_SYSCTL
3981 int sysctl_numa_balancing(struct ctl_table *table, int write,
3982 void *buffer, size_t *lenp, loff_t *ppos)
3986 int state = static_branch_likely(&sched_numa_balancing);
3988 if (write && !capable(CAP_SYS_ADMIN))
3993 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3997 set_numabalancing_state(state);
4003 #ifdef CONFIG_SCHEDSTATS
4005 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4007 static void set_schedstats(bool enabled)
4010 static_branch_enable(&sched_schedstats);
4012 static_branch_disable(&sched_schedstats);
4015 void force_schedstat_enabled(void)
4017 if (!schedstat_enabled()) {
4018 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4019 static_branch_enable(&sched_schedstats);
4023 static int __init setup_schedstats(char *str)
4029 if (!strcmp(str, "enable")) {
4030 set_schedstats(true);
4032 } else if (!strcmp(str, "disable")) {
4033 set_schedstats(false);
4038 pr_warn("Unable to parse schedstats=\n");
4042 __setup("schedstats=", setup_schedstats);
4044 #ifdef CONFIG_PROC_SYSCTL
4045 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4046 size_t *lenp, loff_t *ppos)
4050 int state = static_branch_likely(&sched_schedstats);
4052 if (write && !capable(CAP_SYS_ADMIN))
4057 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4061 set_schedstats(state);
4064 #endif /* CONFIG_PROC_SYSCTL */
4065 #endif /* CONFIG_SCHEDSTATS */
4068 * fork()/clone()-time setup:
4070 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4072 unsigned long flags;
4074 __sched_fork(clone_flags, p);
4076 * We mark the process as NEW here. This guarantees that
4077 * nobody will actually run it, and a signal or other external
4078 * event cannot wake it up and insert it on the runqueue either.
4080 p->__state = TASK_NEW;
4083 * Make sure we do not leak PI boosting priority to the child.
4085 p->prio = current->normal_prio;
4090 * Revert to default priority/policy on fork if requested.
4092 if (unlikely(p->sched_reset_on_fork)) {
4093 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4094 p->policy = SCHED_NORMAL;
4095 p->static_prio = NICE_TO_PRIO(0);
4097 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4098 p->static_prio = NICE_TO_PRIO(0);
4100 p->prio = p->normal_prio = p->static_prio;
4101 set_load_weight(p, false);
4104 * We don't need the reset flag anymore after the fork. It has
4105 * fulfilled its duty:
4107 p->sched_reset_on_fork = 0;
4110 if (dl_prio(p->prio))
4112 else if (rt_prio(p->prio))
4113 p->sched_class = &rt_sched_class;
4115 p->sched_class = &fair_sched_class;
4117 init_entity_runnable_average(&p->se);
4120 * The child is not yet in the pid-hash so no cgroup attach races,
4121 * and the cgroup is pinned to this child due to cgroup_fork()
4122 * is ran before sched_fork().
4124 * Silence PROVE_RCU.
4126 raw_spin_lock_irqsave(&p->pi_lock, flags);
4129 * We're setting the CPU for the first time, we don't migrate,
4130 * so use __set_task_cpu().
4132 __set_task_cpu(p, smp_processor_id());
4133 if (p->sched_class->task_fork)
4134 p->sched_class->task_fork(p);
4135 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4137 #ifdef CONFIG_SCHED_INFO
4138 if (likely(sched_info_on()))
4139 memset(&p->sched_info, 0, sizeof(p->sched_info));
4141 #if defined(CONFIG_SMP)
4144 init_task_preempt_count(p);
4146 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4147 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4152 void sched_post_fork(struct task_struct *p)
4154 uclamp_post_fork(p);
4157 unsigned long to_ratio(u64 period, u64 runtime)
4159 if (runtime == RUNTIME_INF)
4163 * Doing this here saves a lot of checks in all
4164 * the calling paths, and returning zero seems
4165 * safe for them anyway.
4170 return div64_u64(runtime << BW_SHIFT, period);
4174 * wake_up_new_task - wake up a newly created task for the first time.
4176 * This function will do some initial scheduler statistics housekeeping
4177 * that must be done for every newly created context, then puts the task
4178 * on the runqueue and wakes it.
4180 void wake_up_new_task(struct task_struct *p)
4185 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4186 WRITE_ONCE(p->__state, TASK_RUNNING);
4189 * Fork balancing, do it here and not earlier because:
4190 * - cpus_ptr can change in the fork path
4191 * - any previously selected CPU might disappear through hotplug
4193 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4194 * as we're not fully set-up yet.
4196 p->recent_used_cpu = task_cpu(p);
4198 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4200 rq = __task_rq_lock(p, &rf);
4201 update_rq_clock(rq);
4202 post_init_entity_util_avg(p);
4204 activate_task(rq, p, ENQUEUE_NOCLOCK);
4205 trace_sched_wakeup_new(p);
4206 check_preempt_curr(rq, p, WF_FORK);
4208 if (p->sched_class->task_woken) {
4210 * Nothing relies on rq->lock after this, so it's fine to
4213 rq_unpin_lock(rq, &rf);
4214 p->sched_class->task_woken(rq, p);
4215 rq_repin_lock(rq, &rf);
4218 task_rq_unlock(rq, p, &rf);
4221 #ifdef CONFIG_PREEMPT_NOTIFIERS
4223 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4225 void preempt_notifier_inc(void)
4227 static_branch_inc(&preempt_notifier_key);
4229 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4231 void preempt_notifier_dec(void)
4233 static_branch_dec(&preempt_notifier_key);
4235 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4238 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4239 * @notifier: notifier struct to register
4241 void preempt_notifier_register(struct preempt_notifier *notifier)
4243 if (!static_branch_unlikely(&preempt_notifier_key))
4244 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4246 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4248 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4251 * preempt_notifier_unregister - no longer interested in preemption notifications
4252 * @notifier: notifier struct to unregister
4254 * This is *not* safe to call from within a preemption notifier.
4256 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4258 hlist_del(¬ifier->link);
4260 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4262 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4264 struct preempt_notifier *notifier;
4266 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4267 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4270 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4272 if (static_branch_unlikely(&preempt_notifier_key))
4273 __fire_sched_in_preempt_notifiers(curr);
4277 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4278 struct task_struct *next)
4280 struct preempt_notifier *notifier;
4282 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4283 notifier->ops->sched_out(notifier, next);
4286 static __always_inline void
4287 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4288 struct task_struct *next)
4290 if (static_branch_unlikely(&preempt_notifier_key))
4291 __fire_sched_out_preempt_notifiers(curr, next);
4294 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4296 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4301 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4302 struct task_struct *next)
4306 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4308 static inline void prepare_task(struct task_struct *next)
4312 * Claim the task as running, we do this before switching to it
4313 * such that any running task will have this set.
4315 * See the ttwu() WF_ON_CPU case and its ordering comment.
4317 WRITE_ONCE(next->on_cpu, 1);
4321 static inline void finish_task(struct task_struct *prev)
4325 * This must be the very last reference to @prev from this CPU. After
4326 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4327 * must ensure this doesn't happen until the switch is completely
4330 * In particular, the load of prev->state in finish_task_switch() must
4331 * happen before this.
4333 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4335 smp_store_release(&prev->on_cpu, 0);
4341 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4343 void (*func)(struct rq *rq);
4344 struct callback_head *next;
4346 lockdep_assert_rq_held(rq);
4349 func = (void (*)(struct rq *))head->func;
4358 static void balance_push(struct rq *rq);
4360 struct callback_head balance_push_callback = {
4362 .func = (void (*)(struct callback_head *))balance_push,
4365 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4367 struct callback_head *head = rq->balance_callback;
4369 lockdep_assert_rq_held(rq);
4371 rq->balance_callback = NULL;
4376 static void __balance_callbacks(struct rq *rq)
4378 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4381 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4383 unsigned long flags;
4385 if (unlikely(head)) {
4386 raw_spin_rq_lock_irqsave(rq, flags);
4387 do_balance_callbacks(rq, head);
4388 raw_spin_rq_unlock_irqrestore(rq, flags);
4394 static inline void __balance_callbacks(struct rq *rq)
4398 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4403 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4410 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4413 * Since the runqueue lock will be released by the next
4414 * task (which is an invalid locking op but in the case
4415 * of the scheduler it's an obvious special-case), so we
4416 * do an early lockdep release here:
4418 rq_unpin_lock(rq, rf);
4419 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4420 #ifdef CONFIG_DEBUG_SPINLOCK
4421 /* this is a valid case when another task releases the spinlock */
4422 rq_lockp(rq)->owner = next;
4426 static inline void finish_lock_switch(struct rq *rq)
4429 * If we are tracking spinlock dependencies then we have to
4430 * fix up the runqueue lock - which gets 'carried over' from
4431 * prev into current:
4433 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4434 __balance_callbacks(rq);
4435 raw_spin_rq_unlock_irq(rq);
4439 * NOP if the arch has not defined these:
4442 #ifndef prepare_arch_switch
4443 # define prepare_arch_switch(next) do { } while (0)
4446 #ifndef finish_arch_post_lock_switch
4447 # define finish_arch_post_lock_switch() do { } while (0)
4450 static inline void kmap_local_sched_out(void)
4452 #ifdef CONFIG_KMAP_LOCAL
4453 if (unlikely(current->kmap_ctrl.idx))
4454 __kmap_local_sched_out();
4458 static inline void kmap_local_sched_in(void)
4460 #ifdef CONFIG_KMAP_LOCAL
4461 if (unlikely(current->kmap_ctrl.idx))
4462 __kmap_local_sched_in();
4467 * prepare_task_switch - prepare to switch tasks
4468 * @rq: the runqueue preparing to switch
4469 * @prev: the current task that is being switched out
4470 * @next: the task we are going to switch to.
4472 * This is called with the rq lock held and interrupts off. It must
4473 * be paired with a subsequent finish_task_switch after the context
4476 * prepare_task_switch sets up locking and calls architecture specific
4480 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4481 struct task_struct *next)
4483 kcov_prepare_switch(prev);
4484 sched_info_switch(rq, prev, next);
4485 perf_event_task_sched_out(prev, next);
4487 fire_sched_out_preempt_notifiers(prev, next);
4488 kmap_local_sched_out();
4490 prepare_arch_switch(next);
4494 * finish_task_switch - clean up after a task-switch
4495 * @prev: the thread we just switched away from.
4497 * finish_task_switch must be called after the context switch, paired
4498 * with a prepare_task_switch call before the context switch.
4499 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4500 * and do any other architecture-specific cleanup actions.
4502 * Note that we may have delayed dropping an mm in context_switch(). If
4503 * so, we finish that here outside of the runqueue lock. (Doing it
4504 * with the lock held can cause deadlocks; see schedule() for
4507 * The context switch have flipped the stack from under us and restored the
4508 * local variables which were saved when this task called schedule() in the
4509 * past. prev == current is still correct but we need to recalculate this_rq
4510 * because prev may have moved to another CPU.
4512 static struct rq *finish_task_switch(struct task_struct *prev)
4513 __releases(rq->lock)
4515 struct rq *rq = this_rq();
4516 struct mm_struct *mm = rq->prev_mm;
4520 * The previous task will have left us with a preempt_count of 2
4521 * because it left us after:
4524 * preempt_disable(); // 1
4526 * raw_spin_lock_irq(&rq->lock) // 2
4528 * Also, see FORK_PREEMPT_COUNT.
4530 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4531 "corrupted preempt_count: %s/%d/0x%x\n",
4532 current->comm, current->pid, preempt_count()))
4533 preempt_count_set(FORK_PREEMPT_COUNT);
4538 * A task struct has one reference for the use as "current".
4539 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4540 * schedule one last time. The schedule call will never return, and
4541 * the scheduled task must drop that reference.
4543 * We must observe prev->state before clearing prev->on_cpu (in
4544 * finish_task), otherwise a concurrent wakeup can get prev
4545 * running on another CPU and we could rave with its RUNNING -> DEAD
4546 * transition, resulting in a double drop.
4548 prev_state = READ_ONCE(prev->__state);
4549 vtime_task_switch(prev);
4550 perf_event_task_sched_in(prev, current);
4552 tick_nohz_task_switch();
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);
4602 * schedule_tail - first thing a freshly forked thread must call.
4603 * @prev: the thread we just switched away from.
4605 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4606 __releases(rq->lock)
4609 * New tasks start with FORK_PREEMPT_COUNT, see there and
4610 * finish_task_switch() for details.
4612 * finish_task_switch() will drop rq->lock() and lower preempt_count
4613 * and the preempt_enable() will end up enabling preemption (on
4614 * PREEMPT_COUNT kernels).
4617 finish_task_switch(prev);
4620 if (current->set_child_tid)
4621 put_user(task_pid_vnr(current), current->set_child_tid);
4623 calculate_sigpending();
4627 * context_switch - switch to the new MM and the new thread's register state.
4629 static __always_inline struct rq *
4630 context_switch(struct rq *rq, struct task_struct *prev,
4631 struct task_struct *next, struct rq_flags *rf)
4633 prepare_task_switch(rq, prev, next);
4636 * For paravirt, this is coupled with an exit in switch_to to
4637 * combine the page table reload and the switch backend into
4640 arch_start_context_switch(prev);
4643 * kernel -> kernel lazy + transfer active
4644 * user -> kernel lazy + mmgrab() active
4646 * kernel -> user switch + mmdrop() active
4647 * user -> user switch
4649 if (!next->mm) { // to kernel
4650 enter_lazy_tlb(prev->active_mm, next);
4652 next->active_mm = prev->active_mm;
4653 if (prev->mm) // from user
4654 mmgrab(prev->active_mm);
4656 prev->active_mm = NULL;
4658 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4660 * sys_membarrier() requires an smp_mb() between setting
4661 * rq->curr / membarrier_switch_mm() and returning to userspace.
4663 * The below provides this either through switch_mm(), or in
4664 * case 'prev->active_mm == next->mm' through
4665 * finish_task_switch()'s mmdrop().
4667 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4669 if (!prev->mm) { // from kernel
4670 /* will mmdrop() in finish_task_switch(). */
4671 rq->prev_mm = prev->active_mm;
4672 prev->active_mm = NULL;
4676 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4678 prepare_lock_switch(rq, next, rf);
4680 /* Here we just switch the register state and the stack. */
4681 switch_to(prev, next, prev);
4684 return finish_task_switch(prev);
4688 * nr_running and nr_context_switches:
4690 * externally visible scheduler statistics: current number of runnable
4691 * threads, total number of context switches performed since bootup.
4693 unsigned int nr_running(void)
4695 unsigned int i, sum = 0;
4697 for_each_online_cpu(i)
4698 sum += cpu_rq(i)->nr_running;
4704 * Check if only the current task is running on the CPU.
4706 * Caution: this function does not check that the caller has disabled
4707 * preemption, thus the result might have a time-of-check-to-time-of-use
4708 * race. The caller is responsible to use it correctly, for example:
4710 * - from a non-preemptible section (of course)
4712 * - from a thread that is bound to a single CPU
4714 * - in a loop with very short iterations (e.g. a polling loop)
4716 bool single_task_running(void)
4718 return raw_rq()->nr_running == 1;
4720 EXPORT_SYMBOL(single_task_running);
4722 unsigned long long nr_context_switches(void)
4725 unsigned long long sum = 0;
4727 for_each_possible_cpu(i)
4728 sum += cpu_rq(i)->nr_switches;
4734 * Consumers of these two interfaces, like for example the cpuidle menu
4735 * governor, are using nonsensical data. Preferring shallow idle state selection
4736 * for a CPU that has IO-wait which might not even end up running the task when
4737 * it does become runnable.
4740 unsigned int nr_iowait_cpu(int cpu)
4742 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4746 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4748 * The idea behind IO-wait account is to account the idle time that we could
4749 * have spend running if it were not for IO. That is, if we were to improve the
4750 * storage performance, we'd have a proportional reduction in IO-wait time.
4752 * This all works nicely on UP, where, when a task blocks on IO, we account
4753 * idle time as IO-wait, because if the storage were faster, it could've been
4754 * running and we'd not be idle.
4756 * This has been extended to SMP, by doing the same for each CPU. This however
4759 * Imagine for instance the case where two tasks block on one CPU, only the one
4760 * CPU will have IO-wait accounted, while the other has regular idle. Even
4761 * though, if the storage were faster, both could've ran at the same time,
4762 * utilising both CPUs.
4764 * This means, that when looking globally, the current IO-wait accounting on
4765 * SMP is a lower bound, by reason of under accounting.
4767 * Worse, since the numbers are provided per CPU, they are sometimes
4768 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4769 * associated with any one particular CPU, it can wake to another CPU than it
4770 * blocked on. This means the per CPU IO-wait number is meaningless.
4772 * Task CPU affinities can make all that even more 'interesting'.
4775 unsigned int nr_iowait(void)
4777 unsigned int i, sum = 0;
4779 for_each_possible_cpu(i)
4780 sum += nr_iowait_cpu(i);
4788 * sched_exec - execve() is a valuable balancing opportunity, because at
4789 * this point the task has the smallest effective memory and cache footprint.
4791 void sched_exec(void)
4793 struct task_struct *p = current;
4794 unsigned long flags;
4797 raw_spin_lock_irqsave(&p->pi_lock, flags);
4798 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4799 if (dest_cpu == smp_processor_id())
4802 if (likely(cpu_active(dest_cpu))) {
4803 struct migration_arg arg = { p, dest_cpu };
4805 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4806 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4810 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4815 DEFINE_PER_CPU(struct kernel_stat, kstat);
4816 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4818 EXPORT_PER_CPU_SYMBOL(kstat);
4819 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4822 * The function fair_sched_class.update_curr accesses the struct curr
4823 * and its field curr->exec_start; when called from task_sched_runtime(),
4824 * we observe a high rate of cache misses in practice.
4825 * Prefetching this data results in improved performance.
4827 static inline void prefetch_curr_exec_start(struct task_struct *p)
4829 #ifdef CONFIG_FAIR_GROUP_SCHED
4830 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4832 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4835 prefetch(&curr->exec_start);
4839 * Return accounted runtime for the task.
4840 * In case the task is currently running, return the runtime plus current's
4841 * pending runtime that have not been accounted yet.
4843 unsigned long long task_sched_runtime(struct task_struct *p)
4849 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4851 * 64-bit doesn't need locks to atomically read a 64-bit value.
4852 * So we have a optimization chance when the task's delta_exec is 0.
4853 * Reading ->on_cpu is racy, but this is ok.
4855 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4856 * If we race with it entering CPU, unaccounted time is 0. This is
4857 * indistinguishable from the read occurring a few cycles earlier.
4858 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4859 * been accounted, so we're correct here as well.
4861 if (!p->on_cpu || !task_on_rq_queued(p))
4862 return p->se.sum_exec_runtime;
4865 rq = task_rq_lock(p, &rf);
4867 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4868 * project cycles that may never be accounted to this
4869 * thread, breaking clock_gettime().
4871 if (task_current(rq, p) && task_on_rq_queued(p)) {
4872 prefetch_curr_exec_start(p);
4873 update_rq_clock(rq);
4874 p->sched_class->update_curr(rq);
4876 ns = p->se.sum_exec_runtime;
4877 task_rq_unlock(rq, p, &rf);
4882 #ifdef CONFIG_SCHED_DEBUG
4883 static u64 cpu_resched_latency(struct rq *rq)
4885 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
4886 u64 resched_latency, now = rq_clock(rq);
4887 static bool warned_once;
4889 if (sysctl_resched_latency_warn_once && warned_once)
4892 if (!need_resched() || !latency_warn_ms)
4895 if (system_state == SYSTEM_BOOTING)
4898 if (!rq->last_seen_need_resched_ns) {
4899 rq->last_seen_need_resched_ns = now;
4900 rq->ticks_without_resched = 0;
4904 rq->ticks_without_resched++;
4905 resched_latency = now - rq->last_seen_need_resched_ns;
4906 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
4911 return resched_latency;
4914 static int __init setup_resched_latency_warn_ms(char *str)
4918 if ((kstrtol(str, 0, &val))) {
4919 pr_warn("Unable to set resched_latency_warn_ms\n");
4923 sysctl_resched_latency_warn_ms = val;
4926 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
4928 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
4929 #endif /* CONFIG_SCHED_DEBUG */
4932 * This function gets called by the timer code, with HZ frequency.
4933 * We call it with interrupts disabled.
4935 void scheduler_tick(void)
4937 int cpu = smp_processor_id();
4938 struct rq *rq = cpu_rq(cpu);
4939 struct task_struct *curr = rq->curr;
4941 unsigned long thermal_pressure;
4942 u64 resched_latency;
4944 arch_scale_freq_tick();
4949 update_rq_clock(rq);
4950 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4951 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4952 curr->sched_class->task_tick(rq, curr, 0);
4953 if (sched_feat(LATENCY_WARN))
4954 resched_latency = cpu_resched_latency(rq);
4955 calc_global_load_tick(rq);
4959 if (sched_feat(LATENCY_WARN) && resched_latency)
4960 resched_latency_warn(cpu, resched_latency);
4962 perf_event_task_tick();
4965 rq->idle_balance = idle_cpu(cpu);
4966 trigger_load_balance(rq);
4970 #ifdef CONFIG_NO_HZ_FULL
4975 struct delayed_work work;
4977 /* Values for ->state, see diagram below. */
4978 #define TICK_SCHED_REMOTE_OFFLINE 0
4979 #define TICK_SCHED_REMOTE_OFFLINING 1
4980 #define TICK_SCHED_REMOTE_RUNNING 2
4983 * State diagram for ->state:
4986 * TICK_SCHED_REMOTE_OFFLINE
4989 * | | sched_tick_remote()
4992 * +--TICK_SCHED_REMOTE_OFFLINING
4995 * sched_tick_start() | | sched_tick_stop()
4998 * TICK_SCHED_REMOTE_RUNNING
5001 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5002 * and sched_tick_start() are happy to leave the state in RUNNING.
5005 static struct tick_work __percpu *tick_work_cpu;
5007 static void sched_tick_remote(struct work_struct *work)
5009 struct delayed_work *dwork = to_delayed_work(work);
5010 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5011 int cpu = twork->cpu;
5012 struct rq *rq = cpu_rq(cpu);
5013 struct task_struct *curr;
5019 * Handle the tick only if it appears the remote CPU is running in full
5020 * dynticks mode. The check is racy by nature, but missing a tick or
5021 * having one too much is no big deal because the scheduler tick updates
5022 * statistics and checks timeslices in a time-independent way, regardless
5023 * of when exactly it is running.
5025 if (!tick_nohz_tick_stopped_cpu(cpu))
5028 rq_lock_irq(rq, &rf);
5030 if (cpu_is_offline(cpu))
5033 update_rq_clock(rq);
5035 if (!is_idle_task(curr)) {
5037 * Make sure the next tick runs within a reasonable
5040 delta = rq_clock_task(rq) - curr->se.exec_start;
5041 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5043 curr->sched_class->task_tick(rq, curr, 0);
5045 calc_load_nohz_remote(rq);
5047 rq_unlock_irq(rq, &rf);
5051 * Run the remote tick once per second (1Hz). This arbitrary
5052 * frequency is large enough to avoid overload but short enough
5053 * to keep scheduler internal stats reasonably up to date. But
5054 * first update state to reflect hotplug activity if required.
5056 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5057 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5058 if (os == TICK_SCHED_REMOTE_RUNNING)
5059 queue_delayed_work(system_unbound_wq, dwork, HZ);
5062 static void sched_tick_start(int cpu)
5065 struct tick_work *twork;
5067 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5070 WARN_ON_ONCE(!tick_work_cpu);
5072 twork = per_cpu_ptr(tick_work_cpu, cpu);
5073 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5074 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5075 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5077 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5078 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5082 #ifdef CONFIG_HOTPLUG_CPU
5083 static void sched_tick_stop(int cpu)
5085 struct tick_work *twork;
5088 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5091 WARN_ON_ONCE(!tick_work_cpu);
5093 twork = per_cpu_ptr(tick_work_cpu, cpu);
5094 /* There cannot be competing actions, but don't rely on stop-machine. */
5095 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5096 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5097 /* Don't cancel, as this would mess up the state machine. */
5099 #endif /* CONFIG_HOTPLUG_CPU */
5101 int __init sched_tick_offload_init(void)
5103 tick_work_cpu = alloc_percpu(struct tick_work);
5104 BUG_ON(!tick_work_cpu);
5108 #else /* !CONFIG_NO_HZ_FULL */
5109 static inline void sched_tick_start(int cpu) { }
5110 static inline void sched_tick_stop(int cpu) { }
5113 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5114 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5116 * If the value passed in is equal to the current preempt count
5117 * then we just disabled preemption. Start timing the latency.
5119 static inline void preempt_latency_start(int val)
5121 if (preempt_count() == val) {
5122 unsigned long ip = get_lock_parent_ip();
5123 #ifdef CONFIG_DEBUG_PREEMPT
5124 current->preempt_disable_ip = ip;
5126 trace_preempt_off(CALLER_ADDR0, ip);
5130 void preempt_count_add(int val)
5132 #ifdef CONFIG_DEBUG_PREEMPT
5136 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5139 __preempt_count_add(val);
5140 #ifdef CONFIG_DEBUG_PREEMPT
5142 * Spinlock count overflowing soon?
5144 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5147 preempt_latency_start(val);
5149 EXPORT_SYMBOL(preempt_count_add);
5150 NOKPROBE_SYMBOL(preempt_count_add);
5153 * If the value passed in equals to the current preempt count
5154 * then we just enabled preemption. Stop timing the latency.
5156 static inline void preempt_latency_stop(int val)
5158 if (preempt_count() == val)
5159 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5162 void preempt_count_sub(int val)
5164 #ifdef CONFIG_DEBUG_PREEMPT
5168 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5171 * Is the spinlock portion underflowing?
5173 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5174 !(preempt_count() & PREEMPT_MASK)))
5178 preempt_latency_stop(val);
5179 __preempt_count_sub(val);
5181 EXPORT_SYMBOL(preempt_count_sub);
5182 NOKPROBE_SYMBOL(preempt_count_sub);
5185 static inline void preempt_latency_start(int val) { }
5186 static inline void preempt_latency_stop(int val) { }
5189 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5191 #ifdef CONFIG_DEBUG_PREEMPT
5192 return p->preempt_disable_ip;
5199 * Print scheduling while atomic bug:
5201 static noinline void __schedule_bug(struct task_struct *prev)
5203 /* Save this before calling printk(), since that will clobber it */
5204 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5206 if (oops_in_progress)
5209 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5210 prev->comm, prev->pid, preempt_count());
5212 debug_show_held_locks(prev);
5214 if (irqs_disabled())
5215 print_irqtrace_events(prev);
5216 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5217 && in_atomic_preempt_off()) {
5218 pr_err("Preemption disabled at:");
5219 print_ip_sym(KERN_ERR, preempt_disable_ip);
5222 panic("scheduling while atomic\n");
5225 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5229 * Various schedule()-time debugging checks and statistics:
5231 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5233 #ifdef CONFIG_SCHED_STACK_END_CHECK
5234 if (task_stack_end_corrupted(prev))
5235 panic("corrupted stack end detected inside scheduler\n");
5237 if (task_scs_end_corrupted(prev))
5238 panic("corrupted shadow stack detected inside scheduler\n");
5241 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5242 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5243 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5244 prev->comm, prev->pid, prev->non_block_count);
5246 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5250 if (unlikely(in_atomic_preempt_off())) {
5251 __schedule_bug(prev);
5252 preempt_count_set(PREEMPT_DISABLED);
5255 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5257 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5259 schedstat_inc(this_rq()->sched_count);
5262 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5263 struct rq_flags *rf)
5266 const struct sched_class *class;
5268 * We must do the balancing pass before put_prev_task(), such
5269 * that when we release the rq->lock the task is in the same
5270 * state as before we took rq->lock.
5272 * We can terminate the balance pass as soon as we know there is
5273 * a runnable task of @class priority or higher.
5275 for_class_range(class, prev->sched_class, &idle_sched_class) {
5276 if (class->balance(rq, prev, rf))
5281 put_prev_task(rq, prev);
5285 * Pick up the highest-prio task:
5287 static inline struct task_struct *
5288 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5290 const struct sched_class *class;
5291 struct task_struct *p;
5294 * Optimization: we know that if all tasks are in the fair class we can
5295 * call that function directly, but only if the @prev task wasn't of a
5296 * higher scheduling class, because otherwise those lose the
5297 * opportunity to pull in more work from other CPUs.
5299 if (likely(prev->sched_class <= &fair_sched_class &&
5300 rq->nr_running == rq->cfs.h_nr_running)) {
5302 p = pick_next_task_fair(rq, prev, rf);
5303 if (unlikely(p == RETRY_TASK))
5306 /* Assume the next prioritized class is idle_sched_class */
5308 put_prev_task(rq, prev);
5309 p = pick_next_task_idle(rq);
5316 put_prev_task_balance(rq, prev, rf);
5318 for_each_class(class) {
5319 p = class->pick_next_task(rq);
5324 /* The idle class should always have a runnable task: */
5328 #ifdef CONFIG_SCHED_CORE
5329 static inline bool is_task_rq_idle(struct task_struct *t)
5331 return (task_rq(t)->idle == t);
5334 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5336 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5339 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5341 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5344 return a->core_cookie == b->core_cookie;
5347 // XXX fairness/fwd progress conditions
5350 * - NULL if there is no runnable task for this class.
5351 * - the highest priority task for this runqueue if it matches
5352 * rq->core->core_cookie or its priority is greater than max.
5353 * - Else returns idle_task.
5355 static struct task_struct *
5356 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5358 struct task_struct *class_pick, *cookie_pick;
5359 unsigned long cookie = rq->core->core_cookie;
5361 class_pick = class->pick_task(rq);
5367 * If class_pick is tagged, return it only if it has
5368 * higher priority than max.
5370 if (max && class_pick->core_cookie &&
5371 prio_less(class_pick, max, in_fi))
5372 return idle_sched_class.pick_task(rq);
5378 * If class_pick is idle or matches cookie, return early.
5380 if (cookie_equals(class_pick, cookie))
5383 cookie_pick = sched_core_find(rq, cookie);
5386 * If class > max && class > cookie, it is the highest priority task on
5387 * the core (so far) and it must be selected, otherwise we must go with
5388 * the cookie pick in order to satisfy the constraint.
5390 if (prio_less(cookie_pick, class_pick, in_fi) &&
5391 (!max || prio_less(max, class_pick, in_fi)))
5397 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5399 static struct task_struct *
5400 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5402 struct task_struct *next, *max = NULL;
5403 const struct sched_class *class;
5404 const struct cpumask *smt_mask;
5405 bool fi_before = false;
5406 int i, j, cpu, occ = 0;
5409 if (!sched_core_enabled(rq))
5410 return __pick_next_task(rq, prev, rf);
5414 /* Stopper task is switching into idle, no need core-wide selection. */
5415 if (cpu_is_offline(cpu)) {
5417 * Reset core_pick so that we don't enter the fastpath when
5418 * coming online. core_pick would already be migrated to
5419 * another cpu during offline.
5421 rq->core_pick = NULL;
5422 return __pick_next_task(rq, prev, rf);
5426 * If there were no {en,de}queues since we picked (IOW, the task
5427 * pointers are all still valid), and we haven't scheduled the last
5428 * pick yet, do so now.
5430 * rq->core_pick can be NULL if no selection was made for a CPU because
5431 * it was either offline or went offline during a sibling's core-wide
5432 * selection. In this case, do a core-wide selection.
5434 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5435 rq->core->core_pick_seq != rq->core_sched_seq &&
5437 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5439 next = rq->core_pick;
5441 put_prev_task(rq, prev);
5442 set_next_task(rq, next);
5445 rq->core_pick = NULL;
5449 put_prev_task_balance(rq, prev, rf);
5451 smt_mask = cpu_smt_mask(cpu);
5452 need_sync = !!rq->core->core_cookie;
5455 rq->core->core_cookie = 0UL;
5456 if (rq->core->core_forceidle) {
5459 rq->core->core_forceidle = false;
5463 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5465 * @task_seq guards the task state ({en,de}queues)
5466 * @pick_seq is the @task_seq we did a selection on
5467 * @sched_seq is the @pick_seq we scheduled
5469 * However, preemptions can cause multiple picks on the same task set.
5470 * 'Fix' this by also increasing @task_seq for every pick.
5472 rq->core->core_task_seq++;
5475 * Optimize for common case where this CPU has no cookies
5476 * and there are no cookied tasks running on siblings.
5479 for_each_class(class) {
5480 next = class->pick_task(rq);
5485 if (!next->core_cookie) {
5486 rq->core_pick = NULL;
5488 * For robustness, update the min_vruntime_fi for
5489 * unconstrained picks as well.
5491 WARN_ON_ONCE(fi_before);
5492 task_vruntime_update(rq, next, false);
5497 for_each_cpu(i, smt_mask) {
5498 struct rq *rq_i = cpu_rq(i);
5500 rq_i->core_pick = NULL;
5503 update_rq_clock(rq_i);
5507 * Try and select tasks for each sibling in descending sched_class
5510 for_each_class(class) {
5512 for_each_cpu_wrap(i, smt_mask, cpu) {
5513 struct rq *rq_i = cpu_rq(i);
5514 struct task_struct *p;
5516 if (rq_i->core_pick)
5520 * If this sibling doesn't yet have a suitable task to
5521 * run; ask for the most eligible task, given the
5522 * highest priority task already selected for this
5525 p = pick_task(rq_i, class, max, fi_before);
5529 if (!is_task_rq_idle(p))
5532 rq_i->core_pick = p;
5533 if (rq_i->idle == p && rq_i->nr_running) {
5534 rq->core->core_forceidle = true;
5536 rq->core->core_forceidle_seq++;
5540 * If this new candidate is of higher priority than the
5541 * previous; and they're incompatible; we need to wipe
5542 * the slate and start over. pick_task makes sure that
5543 * p's priority is more than max if it doesn't match
5546 * NOTE: this is a linear max-filter and is thus bounded
5547 * in execution time.
5549 if (!max || !cookie_match(max, p)) {
5550 struct task_struct *old_max = max;
5552 rq->core->core_cookie = p->core_cookie;
5556 rq->core->core_forceidle = false;
5557 for_each_cpu(j, smt_mask) {
5561 cpu_rq(j)->core_pick = NULL;
5570 rq->core->core_pick_seq = rq->core->core_task_seq;
5571 next = rq->core_pick;
5572 rq->core_sched_seq = rq->core->core_pick_seq;
5574 /* Something should have been selected for current CPU */
5575 WARN_ON_ONCE(!next);
5578 * Reschedule siblings
5580 * NOTE: L1TF -- at this point we're no longer running the old task and
5581 * sending an IPI (below) ensures the sibling will no longer be running
5582 * their task. This ensures there is no inter-sibling overlap between
5583 * non-matching user state.
5585 for_each_cpu(i, smt_mask) {
5586 struct rq *rq_i = cpu_rq(i);
5589 * An online sibling might have gone offline before a task
5590 * could be picked for it, or it might be offline but later
5591 * happen to come online, but its too late and nothing was
5592 * picked for it. That's Ok - it will pick tasks for itself,
5595 if (!rq_i->core_pick)
5599 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5600 * fi_before fi update?
5606 if (!(fi_before && rq->core->core_forceidle))
5607 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5609 rq_i->core_pick->core_occupation = occ;
5612 rq_i->core_pick = NULL;
5616 /* Did we break L1TF mitigation requirements? */
5617 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5619 if (rq_i->curr == rq_i->core_pick) {
5620 rq_i->core_pick = NULL;
5628 set_next_task(rq, next);
5632 static bool try_steal_cookie(int this, int that)
5634 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5635 struct task_struct *p;
5636 unsigned long cookie;
5637 bool success = false;
5639 local_irq_disable();
5640 double_rq_lock(dst, src);
5642 cookie = dst->core->core_cookie;
5646 if (dst->curr != dst->idle)
5649 p = sched_core_find(src, cookie);
5654 if (p == src->core_pick || p == src->curr)
5657 if (!cpumask_test_cpu(this, &p->cpus_mask))
5660 if (p->core_occupation > dst->idle->core_occupation)
5663 p->on_rq = TASK_ON_RQ_MIGRATING;
5664 deactivate_task(src, p, 0);
5665 set_task_cpu(p, this);
5666 activate_task(dst, p, 0);
5667 p->on_rq = TASK_ON_RQ_QUEUED;
5675 p = sched_core_next(p, cookie);
5679 double_rq_unlock(dst, src);
5685 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5689 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5696 if (try_steal_cookie(cpu, i))
5703 static void sched_core_balance(struct rq *rq)
5705 struct sched_domain *sd;
5706 int cpu = cpu_of(rq);
5710 raw_spin_rq_unlock_irq(rq);
5711 for_each_domain(cpu, sd) {
5715 if (steal_cookie_task(cpu, sd))
5718 raw_spin_rq_lock_irq(rq);
5723 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5725 void queue_core_balance(struct rq *rq)
5727 if (!sched_core_enabled(rq))
5730 if (!rq->core->core_cookie)
5733 if (!rq->nr_running) /* not forced idle */
5736 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5739 static inline void sched_core_cpu_starting(unsigned int cpu)
5741 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5742 struct rq *rq, *core_rq = NULL;
5745 core_rq = cpu_rq(cpu)->core;
5748 for_each_cpu(i, smt_mask) {
5750 if (rq->core && rq->core == rq)
5755 core_rq = cpu_rq(cpu);
5757 for_each_cpu(i, smt_mask) {
5760 WARN_ON_ONCE(rq->core && rq->core != core_rq);
5765 #else /* !CONFIG_SCHED_CORE */
5767 static inline void sched_core_cpu_starting(unsigned int cpu) {}
5769 static struct task_struct *
5770 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5772 return __pick_next_task(rq, prev, rf);
5775 #endif /* CONFIG_SCHED_CORE */
5778 * __schedule() is the main scheduler function.
5780 * The main means of driving the scheduler and thus entering this function are:
5782 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
5784 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
5785 * paths. For example, see arch/x86/entry_64.S.
5787 * To drive preemption between tasks, the scheduler sets the flag in timer
5788 * interrupt handler scheduler_tick().
5790 * 3. Wakeups don't really cause entry into schedule(). They add a
5791 * task to the run-queue and that's it.
5793 * Now, if the new task added to the run-queue preempts the current
5794 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
5795 * called on the nearest possible occasion:
5797 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
5799 * - in syscall or exception context, at the next outmost
5800 * preempt_enable(). (this might be as soon as the wake_up()'s
5803 * - in IRQ context, return from interrupt-handler to
5804 * preemptible context
5806 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
5809 * - cond_resched() call
5810 * - explicit schedule() call
5811 * - return from syscall or exception to user-space
5812 * - return from interrupt-handler to user-space
5814 * WARNING: must be called with preemption disabled!
5816 static void __sched notrace __schedule(bool preempt)
5818 struct task_struct *prev, *next;
5819 unsigned long *switch_count;
5820 unsigned long prev_state;
5825 cpu = smp_processor_id();
5829 schedule_debug(prev, preempt);
5831 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
5834 local_irq_disable();
5835 rcu_note_context_switch(preempt);
5838 * Make sure that signal_pending_state()->signal_pending() below
5839 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
5840 * done by the caller to avoid the race with signal_wake_up():
5842 * __set_current_state(@state) signal_wake_up()
5843 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
5844 * wake_up_state(p, state)
5845 * LOCK rq->lock LOCK p->pi_state
5846 * smp_mb__after_spinlock() smp_mb__after_spinlock()
5847 * if (signal_pending_state()) if (p->state & @state)
5849 * Also, the membarrier system call requires a full memory barrier
5850 * after coming from user-space, before storing to rq->curr.
5853 smp_mb__after_spinlock();
5855 /* Promote REQ to ACT */
5856 rq->clock_update_flags <<= 1;
5857 update_rq_clock(rq);
5859 switch_count = &prev->nivcsw;
5862 * We must load prev->state once (task_struct::state is volatile), such
5865 * - we form a control dependency vs deactivate_task() below.
5866 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5868 prev_state = READ_ONCE(prev->__state);
5869 if (!preempt && prev_state) {
5870 if (signal_pending_state(prev_state, prev)) {
5871 WRITE_ONCE(prev->__state, TASK_RUNNING);
5873 prev->sched_contributes_to_load =
5874 (prev_state & TASK_UNINTERRUPTIBLE) &&
5875 !(prev_state & TASK_NOLOAD) &&
5876 !(prev->flags & PF_FROZEN);
5878 if (prev->sched_contributes_to_load)
5879 rq->nr_uninterruptible++;
5882 * __schedule() ttwu()
5883 * prev_state = prev->state; if (p->on_rq && ...)
5884 * if (prev_state) goto out;
5885 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5886 * p->state = TASK_WAKING
5888 * Where __schedule() and ttwu() have matching control dependencies.
5890 * After this, schedule() must not care about p->state any more.
5892 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
5894 if (prev->in_iowait) {
5895 atomic_inc(&rq->nr_iowait);
5896 delayacct_blkio_start();
5899 switch_count = &prev->nvcsw;
5902 next = pick_next_task(rq, prev, &rf);
5903 clear_tsk_need_resched(prev);
5904 clear_preempt_need_resched();
5905 #ifdef CONFIG_SCHED_DEBUG
5906 rq->last_seen_need_resched_ns = 0;
5909 if (likely(prev != next)) {
5912 * RCU users of rcu_dereference(rq->curr) may not see
5913 * changes to task_struct made by pick_next_task().
5915 RCU_INIT_POINTER(rq->curr, next);
5917 * The membarrier system call requires each architecture
5918 * to have a full memory barrier after updating
5919 * rq->curr, before returning to user-space.
5921 * Here are the schemes providing that barrier on the
5922 * various architectures:
5923 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5924 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5925 * - finish_lock_switch() for weakly-ordered
5926 * architectures where spin_unlock is a full barrier,
5927 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5928 * is a RELEASE barrier),
5932 migrate_disable_switch(rq, prev);
5933 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5935 trace_sched_switch(preempt, prev, next);
5937 /* Also unlocks the rq: */
5938 rq = context_switch(rq, prev, next, &rf);
5940 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5942 rq_unpin_lock(rq, &rf);
5943 __balance_callbacks(rq);
5944 raw_spin_rq_unlock_irq(rq);
5948 void __noreturn do_task_dead(void)
5950 /* Causes final put_task_struct in finish_task_switch(): */
5951 set_special_state(TASK_DEAD);
5953 /* Tell freezer to ignore us: */
5954 current->flags |= PF_NOFREEZE;
5959 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5964 static inline void sched_submit_work(struct task_struct *tsk)
5966 unsigned int task_flags;
5968 if (task_is_running(tsk))
5971 task_flags = tsk->flags;
5973 * If a worker went to sleep, notify and ask workqueue whether
5974 * it wants to wake up a task to maintain concurrency.
5975 * As this function is called inside the schedule() context,
5976 * we disable preemption to avoid it calling schedule() again
5977 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5980 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5982 if (task_flags & PF_WQ_WORKER)
5983 wq_worker_sleeping(tsk);
5985 io_wq_worker_sleeping(tsk);
5986 preempt_enable_no_resched();
5989 if (tsk_is_pi_blocked(tsk))
5993 * If we are going to sleep and we have plugged IO queued,
5994 * make sure to submit it to avoid deadlocks.
5996 if (blk_needs_flush_plug(tsk))
5997 blk_schedule_flush_plug(tsk);
6000 static void sched_update_worker(struct task_struct *tsk)
6002 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6003 if (tsk->flags & PF_WQ_WORKER)
6004 wq_worker_running(tsk);
6006 io_wq_worker_running(tsk);
6010 asmlinkage __visible void __sched schedule(void)
6012 struct task_struct *tsk = current;
6014 sched_submit_work(tsk);
6018 sched_preempt_enable_no_resched();
6019 } while (need_resched());
6020 sched_update_worker(tsk);
6022 EXPORT_SYMBOL(schedule);
6025 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6026 * state (have scheduled out non-voluntarily) by making sure that all
6027 * tasks have either left the run queue or have gone into user space.
6028 * As idle tasks do not do either, they must not ever be preempted
6029 * (schedule out non-voluntarily).
6031 * schedule_idle() is similar to schedule_preempt_disable() except that it
6032 * never enables preemption because it does not call sched_submit_work().
6034 void __sched schedule_idle(void)
6037 * As this skips calling sched_submit_work(), which the idle task does
6038 * regardless because that function is a nop when the task is in a
6039 * TASK_RUNNING state, make sure this isn't used someplace that the
6040 * current task can be in any other state. Note, idle is always in the
6041 * TASK_RUNNING state.
6043 WARN_ON_ONCE(current->__state);
6046 } while (need_resched());
6049 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6050 asmlinkage __visible void __sched schedule_user(void)
6053 * If we come here after a random call to set_need_resched(),
6054 * or we have been woken up remotely but the IPI has not yet arrived,
6055 * we haven't yet exited the RCU idle mode. Do it here manually until
6056 * we find a better solution.
6058 * NB: There are buggy callers of this function. Ideally we
6059 * should warn if prev_state != CONTEXT_USER, but that will trigger
6060 * too frequently to make sense yet.
6062 enum ctx_state prev_state = exception_enter();
6064 exception_exit(prev_state);
6069 * schedule_preempt_disabled - called with preemption disabled
6071 * Returns with preemption disabled. Note: preempt_count must be 1
6073 void __sched schedule_preempt_disabled(void)
6075 sched_preempt_enable_no_resched();
6080 static void __sched notrace preempt_schedule_common(void)
6084 * Because the function tracer can trace preempt_count_sub()
6085 * and it also uses preempt_enable/disable_notrace(), if
6086 * NEED_RESCHED is set, the preempt_enable_notrace() called
6087 * by the function tracer will call this function again and
6088 * cause infinite recursion.
6090 * Preemption must be disabled here before the function
6091 * tracer can trace. Break up preempt_disable() into two
6092 * calls. One to disable preemption without fear of being
6093 * traced. The other to still record the preemption latency,
6094 * which can also be traced by the function tracer.
6096 preempt_disable_notrace();
6097 preempt_latency_start(1);
6099 preempt_latency_stop(1);
6100 preempt_enable_no_resched_notrace();
6103 * Check again in case we missed a preemption opportunity
6104 * between schedule and now.
6106 } while (need_resched());
6109 #ifdef CONFIG_PREEMPTION
6111 * This is the entry point to schedule() from in-kernel preemption
6112 * off of preempt_enable.
6114 asmlinkage __visible void __sched notrace preempt_schedule(void)
6117 * If there is a non-zero preempt_count or interrupts are disabled,
6118 * we do not want to preempt the current task. Just return..
6120 if (likely(!preemptible()))
6123 preempt_schedule_common();
6125 NOKPROBE_SYMBOL(preempt_schedule);
6126 EXPORT_SYMBOL(preempt_schedule);
6128 #ifdef CONFIG_PREEMPT_DYNAMIC
6129 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6130 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6135 * preempt_schedule_notrace - preempt_schedule called by tracing
6137 * The tracing infrastructure uses preempt_enable_notrace to prevent
6138 * recursion and tracing preempt enabling caused by the tracing
6139 * infrastructure itself. But as tracing can happen in areas coming
6140 * from userspace or just about to enter userspace, a preempt enable
6141 * can occur before user_exit() is called. This will cause the scheduler
6142 * to be called when the system is still in usermode.
6144 * To prevent this, the preempt_enable_notrace will use this function
6145 * instead of preempt_schedule() to exit user context if needed before
6146 * calling the scheduler.
6148 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6150 enum ctx_state prev_ctx;
6152 if (likely(!preemptible()))
6157 * Because the function tracer can trace preempt_count_sub()
6158 * and it also uses preempt_enable/disable_notrace(), if
6159 * NEED_RESCHED is set, the preempt_enable_notrace() called
6160 * by the function tracer will call this function again and
6161 * cause infinite recursion.
6163 * Preemption must be disabled here before the function
6164 * tracer can trace. Break up preempt_disable() into two
6165 * calls. One to disable preemption without fear of being
6166 * traced. The other to still record the preemption latency,
6167 * which can also be traced by the function tracer.
6169 preempt_disable_notrace();
6170 preempt_latency_start(1);
6172 * Needs preempt disabled in case user_exit() is traced
6173 * and the tracer calls preempt_enable_notrace() causing
6174 * an infinite recursion.
6176 prev_ctx = exception_enter();
6178 exception_exit(prev_ctx);
6180 preempt_latency_stop(1);
6181 preempt_enable_no_resched_notrace();
6182 } while (need_resched());
6184 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6186 #ifdef CONFIG_PREEMPT_DYNAMIC
6187 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6188 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6191 #endif /* CONFIG_PREEMPTION */
6193 #ifdef CONFIG_PREEMPT_DYNAMIC
6195 #include <linux/entry-common.h>
6200 * SC:preempt_schedule
6201 * SC:preempt_schedule_notrace
6202 * SC:irqentry_exit_cond_resched
6206 * cond_resched <- __cond_resched
6207 * might_resched <- RET0
6208 * preempt_schedule <- NOP
6209 * preempt_schedule_notrace <- NOP
6210 * irqentry_exit_cond_resched <- NOP
6213 * cond_resched <- __cond_resched
6214 * might_resched <- __cond_resched
6215 * preempt_schedule <- NOP
6216 * preempt_schedule_notrace <- NOP
6217 * irqentry_exit_cond_resched <- NOP
6220 * cond_resched <- RET0
6221 * might_resched <- RET0
6222 * preempt_schedule <- preempt_schedule
6223 * preempt_schedule_notrace <- preempt_schedule_notrace
6224 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6228 preempt_dynamic_none = 0,
6229 preempt_dynamic_voluntary,
6230 preempt_dynamic_full,
6233 int preempt_dynamic_mode = preempt_dynamic_full;
6235 int sched_dynamic_mode(const char *str)
6237 if (!strcmp(str, "none"))
6238 return preempt_dynamic_none;
6240 if (!strcmp(str, "voluntary"))
6241 return preempt_dynamic_voluntary;
6243 if (!strcmp(str, "full"))
6244 return preempt_dynamic_full;
6249 void sched_dynamic_update(int mode)
6252 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6253 * the ZERO state, which is invalid.
6255 static_call_update(cond_resched, __cond_resched);
6256 static_call_update(might_resched, __cond_resched);
6257 static_call_update(preempt_schedule, __preempt_schedule_func);
6258 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6259 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6262 case preempt_dynamic_none:
6263 static_call_update(cond_resched, __cond_resched);
6264 static_call_update(might_resched, (void *)&__static_call_return0);
6265 static_call_update(preempt_schedule, NULL);
6266 static_call_update(preempt_schedule_notrace, NULL);
6267 static_call_update(irqentry_exit_cond_resched, NULL);
6268 pr_info("Dynamic Preempt: none\n");
6271 case preempt_dynamic_voluntary:
6272 static_call_update(cond_resched, __cond_resched);
6273 static_call_update(might_resched, __cond_resched);
6274 static_call_update(preempt_schedule, NULL);
6275 static_call_update(preempt_schedule_notrace, NULL);
6276 static_call_update(irqentry_exit_cond_resched, NULL);
6277 pr_info("Dynamic Preempt: voluntary\n");
6280 case preempt_dynamic_full:
6281 static_call_update(cond_resched, (void *)&__static_call_return0);
6282 static_call_update(might_resched, (void *)&__static_call_return0);
6283 static_call_update(preempt_schedule, __preempt_schedule_func);
6284 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6285 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6286 pr_info("Dynamic Preempt: full\n");
6290 preempt_dynamic_mode = mode;
6293 static int __init setup_preempt_mode(char *str)
6295 int mode = sched_dynamic_mode(str);
6297 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6301 sched_dynamic_update(mode);
6304 __setup("preempt=", setup_preempt_mode);
6306 #endif /* CONFIG_PREEMPT_DYNAMIC */
6309 * This is the entry point to schedule() from kernel preemption
6310 * off of irq context.
6311 * Note, that this is called and return with irqs disabled. This will
6312 * protect us against recursive calling from irq.
6314 asmlinkage __visible void __sched preempt_schedule_irq(void)
6316 enum ctx_state prev_state;
6318 /* Catch callers which need to be fixed */
6319 BUG_ON(preempt_count() || !irqs_disabled());
6321 prev_state = exception_enter();
6327 local_irq_disable();
6328 sched_preempt_enable_no_resched();
6329 } while (need_resched());
6331 exception_exit(prev_state);
6334 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6337 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6338 return try_to_wake_up(curr->private, mode, wake_flags);
6340 EXPORT_SYMBOL(default_wake_function);
6342 static void __setscheduler_prio(struct task_struct *p, int prio)
6345 p->sched_class = &dl_sched_class;
6346 else if (rt_prio(prio))
6347 p->sched_class = &rt_sched_class;
6349 p->sched_class = &fair_sched_class;
6354 #ifdef CONFIG_RT_MUTEXES
6356 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6359 prio = min(prio, pi_task->prio);
6364 static inline int rt_effective_prio(struct task_struct *p, int prio)
6366 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6368 return __rt_effective_prio(pi_task, prio);
6372 * rt_mutex_setprio - set the current priority of a task
6374 * @pi_task: donor task
6376 * This function changes the 'effective' priority of a task. It does
6377 * not touch ->normal_prio like __setscheduler().
6379 * Used by the rt_mutex code to implement priority inheritance
6380 * logic. Call site only calls if the priority of the task changed.
6382 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6384 int prio, oldprio, queued, running, queue_flag =
6385 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6386 const struct sched_class *prev_class;
6390 /* XXX used to be waiter->prio, not waiter->task->prio */
6391 prio = __rt_effective_prio(pi_task, p->normal_prio);
6394 * If nothing changed; bail early.
6396 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6399 rq = __task_rq_lock(p, &rf);
6400 update_rq_clock(rq);
6402 * Set under pi_lock && rq->lock, such that the value can be used under
6405 * Note that there is loads of tricky to make this pointer cache work
6406 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6407 * ensure a task is de-boosted (pi_task is set to NULL) before the
6408 * task is allowed to run again (and can exit). This ensures the pointer
6409 * points to a blocked task -- which guarantees the task is present.
6411 p->pi_top_task = pi_task;
6414 * For FIFO/RR we only need to set prio, if that matches we're done.
6416 if (prio == p->prio && !dl_prio(prio))
6420 * Idle task boosting is a nono in general. There is one
6421 * exception, when PREEMPT_RT and NOHZ is active:
6423 * The idle task calls get_next_timer_interrupt() and holds
6424 * the timer wheel base->lock on the CPU and another CPU wants
6425 * to access the timer (probably to cancel it). We can safely
6426 * ignore the boosting request, as the idle CPU runs this code
6427 * with interrupts disabled and will complete the lock
6428 * protected section without being interrupted. So there is no
6429 * real need to boost.
6431 if (unlikely(p == rq->idle)) {
6432 WARN_ON(p != rq->curr);
6433 WARN_ON(p->pi_blocked_on);
6437 trace_sched_pi_setprio(p, pi_task);
6440 if (oldprio == prio)
6441 queue_flag &= ~DEQUEUE_MOVE;
6443 prev_class = p->sched_class;
6444 queued = task_on_rq_queued(p);
6445 running = task_current(rq, p);
6447 dequeue_task(rq, p, queue_flag);
6449 put_prev_task(rq, p);
6452 * Boosting condition are:
6453 * 1. -rt task is running and holds mutex A
6454 * --> -dl task blocks on mutex A
6456 * 2. -dl task is running and holds mutex A
6457 * --> -dl task blocks on mutex A and could preempt the
6460 if (dl_prio(prio)) {
6461 if (!dl_prio(p->normal_prio) ||
6462 (pi_task && dl_prio(pi_task->prio) &&
6463 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6464 p->dl.pi_se = pi_task->dl.pi_se;
6465 queue_flag |= ENQUEUE_REPLENISH;
6467 p->dl.pi_se = &p->dl;
6469 } else if (rt_prio(prio)) {
6470 if (dl_prio(oldprio))
6471 p->dl.pi_se = &p->dl;
6473 queue_flag |= ENQUEUE_HEAD;
6475 if (dl_prio(oldprio))
6476 p->dl.pi_se = &p->dl;
6477 if (rt_prio(oldprio))
6481 __setscheduler_prio(p, prio);
6484 enqueue_task(rq, p, queue_flag);
6486 set_next_task(rq, p);
6488 check_class_changed(rq, p, prev_class, oldprio);
6490 /* Avoid rq from going away on us: */
6493 rq_unpin_lock(rq, &rf);
6494 __balance_callbacks(rq);
6495 raw_spin_rq_unlock(rq);
6500 static inline int rt_effective_prio(struct task_struct *p, int prio)
6506 void set_user_nice(struct task_struct *p, long nice)
6508 bool queued, running;
6513 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6516 * We have to be careful, if called from sys_setpriority(),
6517 * the task might be in the middle of scheduling on another CPU.
6519 rq = task_rq_lock(p, &rf);
6520 update_rq_clock(rq);
6523 * The RT priorities are set via sched_setscheduler(), but we still
6524 * allow the 'normal' nice value to be set - but as expected
6525 * it won't have any effect on scheduling until the task is
6526 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6528 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6529 p->static_prio = NICE_TO_PRIO(nice);
6532 queued = task_on_rq_queued(p);
6533 running = task_current(rq, p);
6535 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6537 put_prev_task(rq, p);
6539 p->static_prio = NICE_TO_PRIO(nice);
6540 set_load_weight(p, true);
6542 p->prio = effective_prio(p);
6545 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6547 set_next_task(rq, p);
6550 * If the task increased its priority or is running and
6551 * lowered its priority, then reschedule its CPU:
6553 p->sched_class->prio_changed(rq, p, old_prio);
6556 task_rq_unlock(rq, p, &rf);
6558 EXPORT_SYMBOL(set_user_nice);
6561 * can_nice - check if a task can reduce its nice value
6565 int can_nice(const struct task_struct *p, const int nice)
6567 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6568 int nice_rlim = nice_to_rlimit(nice);
6570 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6571 capable(CAP_SYS_NICE));
6574 #ifdef __ARCH_WANT_SYS_NICE
6577 * sys_nice - change the priority of the current process.
6578 * @increment: priority increment
6580 * sys_setpriority is a more generic, but much slower function that
6581 * does similar things.
6583 SYSCALL_DEFINE1(nice, int, increment)
6588 * Setpriority might change our priority at the same moment.
6589 * We don't have to worry. Conceptually one call occurs first
6590 * and we have a single winner.
6592 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6593 nice = task_nice(current) + increment;
6595 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6596 if (increment < 0 && !can_nice(current, nice))
6599 retval = security_task_setnice(current, nice);
6603 set_user_nice(current, nice);
6610 * task_prio - return the priority value of a given task.
6611 * @p: the task in question.
6613 * Return: The priority value as seen by users in /proc.
6615 * sched policy return value kernel prio user prio/nice
6617 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6618 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6619 * deadline -101 -1 0
6621 int task_prio(const struct task_struct *p)
6623 return p->prio - MAX_RT_PRIO;
6627 * idle_cpu - is a given CPU idle currently?
6628 * @cpu: the processor in question.
6630 * Return: 1 if the CPU is currently idle. 0 otherwise.
6632 int idle_cpu(int cpu)
6634 struct rq *rq = cpu_rq(cpu);
6636 if (rq->curr != rq->idle)
6643 if (rq->ttwu_pending)
6651 * available_idle_cpu - is a given CPU idle for enqueuing work.
6652 * @cpu: the CPU in question.
6654 * Return: 1 if the CPU is currently idle. 0 otherwise.
6656 int available_idle_cpu(int cpu)
6661 if (vcpu_is_preempted(cpu))
6668 * idle_task - return the idle task for a given CPU.
6669 * @cpu: the processor in question.
6671 * Return: The idle task for the CPU @cpu.
6673 struct task_struct *idle_task(int cpu)
6675 return cpu_rq(cpu)->idle;
6680 * This function computes an effective utilization for the given CPU, to be
6681 * used for frequency selection given the linear relation: f = u * f_max.
6683 * The scheduler tracks the following metrics:
6685 * cpu_util_{cfs,rt,dl,irq}()
6688 * Where the cfs,rt and dl util numbers are tracked with the same metric and
6689 * synchronized windows and are thus directly comparable.
6691 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
6692 * which excludes things like IRQ and steal-time. These latter are then accrued
6693 * in the irq utilization.
6695 * The DL bandwidth number otoh is not a measured metric but a value computed
6696 * based on the task model parameters and gives the minimal utilization
6697 * required to meet deadlines.
6699 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
6700 unsigned long max, enum cpu_util_type type,
6701 struct task_struct *p)
6703 unsigned long dl_util, util, irq;
6704 struct rq *rq = cpu_rq(cpu);
6706 if (!uclamp_is_used() &&
6707 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
6712 * Early check to see if IRQ/steal time saturates the CPU, can be
6713 * because of inaccuracies in how we track these -- see
6714 * update_irq_load_avg().
6716 irq = cpu_util_irq(rq);
6717 if (unlikely(irq >= max))
6721 * Because the time spend on RT/DL tasks is visible as 'lost' time to
6722 * CFS tasks and we use the same metric to track the effective
6723 * utilization (PELT windows are synchronized) we can directly add them
6724 * to obtain the CPU's actual utilization.
6726 * CFS and RT utilization can be boosted or capped, depending on
6727 * utilization clamp constraints requested by currently RUNNABLE
6729 * When there are no CFS RUNNABLE tasks, clamps are released and
6730 * frequency will be gracefully reduced with the utilization decay.
6732 util = util_cfs + cpu_util_rt(rq);
6733 if (type == FREQUENCY_UTIL)
6734 util = uclamp_rq_util_with(rq, util, p);
6736 dl_util = cpu_util_dl(rq);
6739 * For frequency selection we do not make cpu_util_dl() a permanent part
6740 * of this sum because we want to use cpu_bw_dl() later on, but we need
6741 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
6742 * that we select f_max when there is no idle time.
6744 * NOTE: numerical errors or stop class might cause us to not quite hit
6745 * saturation when we should -- something for later.
6747 if (util + dl_util >= max)
6751 * OTOH, for energy computation we need the estimated running time, so
6752 * include util_dl and ignore dl_bw.
6754 if (type == ENERGY_UTIL)
6758 * There is still idle time; further improve the number by using the
6759 * irq metric. Because IRQ/steal time is hidden from the task clock we
6760 * need to scale the task numbers:
6763 * U' = irq + --------- * U
6766 util = scale_irq_capacity(util, irq, max);
6770 * Bandwidth required by DEADLINE must always be granted while, for
6771 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
6772 * to gracefully reduce the frequency when no tasks show up for longer
6775 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
6776 * bw_dl as requested freq. However, cpufreq is not yet ready for such
6777 * an interface. So, we only do the latter for now.
6779 if (type == FREQUENCY_UTIL)
6780 util += cpu_bw_dl(rq);
6782 return min(max, util);
6785 unsigned long sched_cpu_util(int cpu, unsigned long max)
6787 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
6790 #endif /* CONFIG_SMP */
6793 * find_process_by_pid - find a process with a matching PID value.
6794 * @pid: the pid in question.
6796 * The task of @pid, if found. %NULL otherwise.
6798 static struct task_struct *find_process_by_pid(pid_t pid)
6800 return pid ? find_task_by_vpid(pid) : current;
6804 * sched_setparam() passes in -1 for its policy, to let the functions
6805 * it calls know not to change it.
6807 #define SETPARAM_POLICY -1
6809 static void __setscheduler_params(struct task_struct *p,
6810 const struct sched_attr *attr)
6812 int policy = attr->sched_policy;
6814 if (policy == SETPARAM_POLICY)
6819 if (dl_policy(policy))
6820 __setparam_dl(p, attr);
6821 else if (fair_policy(policy))
6822 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6825 * __sched_setscheduler() ensures attr->sched_priority == 0 when
6826 * !rt_policy. Always setting this ensures that things like
6827 * getparam()/getattr() don't report silly values for !rt tasks.
6829 p->rt_priority = attr->sched_priority;
6830 p->normal_prio = normal_prio(p);
6831 set_load_weight(p, true);
6835 * Check the target process has a UID that matches the current process's:
6837 static bool check_same_owner(struct task_struct *p)
6839 const struct cred *cred = current_cred(), *pcred;
6843 pcred = __task_cred(p);
6844 match = (uid_eq(cred->euid, pcred->euid) ||
6845 uid_eq(cred->euid, pcred->uid));
6850 static int __sched_setscheduler(struct task_struct *p,
6851 const struct sched_attr *attr,
6854 int oldpolicy = -1, policy = attr->sched_policy;
6855 int retval, oldprio, newprio, queued, running;
6856 const struct sched_class *prev_class;
6857 struct callback_head *head;
6860 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6863 /* The pi code expects interrupts enabled */
6864 BUG_ON(pi && in_interrupt());
6866 /* Double check policy once rq lock held: */
6868 reset_on_fork = p->sched_reset_on_fork;
6869 policy = oldpolicy = p->policy;
6871 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
6873 if (!valid_policy(policy))
6877 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
6881 * Valid priorities for SCHED_FIFO and SCHED_RR are
6882 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6883 * SCHED_BATCH and SCHED_IDLE is 0.
6885 if (attr->sched_priority > MAX_RT_PRIO-1)
6887 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
6888 (rt_policy(policy) != (attr->sched_priority != 0)))
6892 * Allow unprivileged RT tasks to decrease priority:
6894 if (user && !capable(CAP_SYS_NICE)) {
6895 if (fair_policy(policy)) {
6896 if (attr->sched_nice < task_nice(p) &&
6897 !can_nice(p, attr->sched_nice))
6901 if (rt_policy(policy)) {
6902 unsigned long rlim_rtprio =
6903 task_rlimit(p, RLIMIT_RTPRIO);
6905 /* Can't set/change the rt policy: */
6906 if (policy != p->policy && !rlim_rtprio)
6909 /* Can't increase priority: */
6910 if (attr->sched_priority > p->rt_priority &&
6911 attr->sched_priority > rlim_rtprio)
6916 * Can't set/change SCHED_DEADLINE policy at all for now
6917 * (safest behavior); in the future we would like to allow
6918 * unprivileged DL tasks to increase their relative deadline
6919 * or reduce their runtime (both ways reducing utilization)
6921 if (dl_policy(policy))
6925 * Treat SCHED_IDLE as nice 20. Only allow a switch to
6926 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6928 if (task_has_idle_policy(p) && !idle_policy(policy)) {
6929 if (!can_nice(p, task_nice(p)))
6933 /* Can't change other user's priorities: */
6934 if (!check_same_owner(p))
6937 /* Normal users shall not reset the sched_reset_on_fork flag: */
6938 if (p->sched_reset_on_fork && !reset_on_fork)
6943 if (attr->sched_flags & SCHED_FLAG_SUGOV)
6946 retval = security_task_setscheduler(p);
6951 /* Update task specific "requested" clamps */
6952 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
6953 retval = uclamp_validate(p, attr);
6962 * Make sure no PI-waiters arrive (or leave) while we are
6963 * changing the priority of the task:
6965 * To be able to change p->policy safely, the appropriate
6966 * runqueue lock must be held.
6968 rq = task_rq_lock(p, &rf);
6969 update_rq_clock(rq);
6972 * Changing the policy of the stop threads its a very bad idea:
6974 if (p == rq->stop) {
6980 * If not changing anything there's no need to proceed further,
6981 * but store a possible modification of reset_on_fork.
6983 if (unlikely(policy == p->policy)) {
6984 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
6986 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
6988 if (dl_policy(policy) && dl_param_changed(p, attr))
6990 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
6993 p->sched_reset_on_fork = reset_on_fork;
7000 #ifdef CONFIG_RT_GROUP_SCHED
7002 * Do not allow realtime tasks into groups that have no runtime
7005 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7006 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7007 !task_group_is_autogroup(task_group(p))) {
7013 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7014 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7015 cpumask_t *span = rq->rd->span;
7018 * Don't allow tasks with an affinity mask smaller than
7019 * the entire root_domain to become SCHED_DEADLINE. We
7020 * will also fail if there's no bandwidth available.
7022 if (!cpumask_subset(span, p->cpus_ptr) ||
7023 rq->rd->dl_bw.bw == 0) {
7031 /* Re-check policy now with rq lock held: */
7032 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7033 policy = oldpolicy = -1;
7034 task_rq_unlock(rq, p, &rf);
7036 cpuset_read_unlock();
7041 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7042 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7045 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7050 p->sched_reset_on_fork = reset_on_fork;
7053 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7056 * Take priority boosted tasks into account. If the new
7057 * effective priority is unchanged, we just store the new
7058 * normal parameters and do not touch the scheduler class and
7059 * the runqueue. This will be done when the task deboost
7062 newprio = rt_effective_prio(p, newprio);
7063 if (newprio == oldprio)
7064 queue_flags &= ~DEQUEUE_MOVE;
7067 queued = task_on_rq_queued(p);
7068 running = task_current(rq, p);
7070 dequeue_task(rq, p, queue_flags);
7072 put_prev_task(rq, p);
7074 prev_class = p->sched_class;
7076 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7077 __setscheduler_params(p, attr);
7078 __setscheduler_prio(p, newprio);
7080 __setscheduler_uclamp(p, attr);
7084 * We enqueue to tail when the priority of a task is
7085 * increased (user space view).
7087 if (oldprio < p->prio)
7088 queue_flags |= ENQUEUE_HEAD;
7090 enqueue_task(rq, p, queue_flags);
7093 set_next_task(rq, p);
7095 check_class_changed(rq, p, prev_class, oldprio);
7097 /* Avoid rq from going away on us: */
7099 head = splice_balance_callbacks(rq);
7100 task_rq_unlock(rq, p, &rf);
7103 cpuset_read_unlock();
7104 rt_mutex_adjust_pi(p);
7107 /* Run balance callbacks after we've adjusted the PI chain: */
7108 balance_callbacks(rq, head);
7114 task_rq_unlock(rq, p, &rf);
7116 cpuset_read_unlock();
7120 static int _sched_setscheduler(struct task_struct *p, int policy,
7121 const struct sched_param *param, bool check)
7123 struct sched_attr attr = {
7124 .sched_policy = policy,
7125 .sched_priority = param->sched_priority,
7126 .sched_nice = PRIO_TO_NICE(p->static_prio),
7129 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7130 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7131 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7132 policy &= ~SCHED_RESET_ON_FORK;
7133 attr.sched_policy = policy;
7136 return __sched_setscheduler(p, &attr, check, true);
7139 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7140 * @p: the task in question.
7141 * @policy: new policy.
7142 * @param: structure containing the new RT priority.
7144 * Use sched_set_fifo(), read its comment.
7146 * Return: 0 on success. An error code otherwise.
7148 * NOTE that the task may be already dead.
7150 int sched_setscheduler(struct task_struct *p, int policy,
7151 const struct sched_param *param)
7153 return _sched_setscheduler(p, policy, param, true);
7156 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7158 return __sched_setscheduler(p, attr, true, true);
7161 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7163 return __sched_setscheduler(p, attr, false, true);
7165 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7168 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7169 * @p: the task in question.
7170 * @policy: new policy.
7171 * @param: structure containing the new RT priority.
7173 * Just like sched_setscheduler, only don't bother checking if the
7174 * current context has permission. For example, this is needed in
7175 * stop_machine(): we create temporary high priority worker threads,
7176 * but our caller might not have that capability.
7178 * Return: 0 on success. An error code otherwise.
7180 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7181 const struct sched_param *param)
7183 return _sched_setscheduler(p, policy, param, false);
7187 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7188 * incapable of resource management, which is the one thing an OS really should
7191 * This is of course the reason it is limited to privileged users only.
7193 * Worse still; it is fundamentally impossible to compose static priority
7194 * workloads. You cannot take two correctly working static prio workloads
7195 * and smash them together and still expect them to work.
7197 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7201 * The administrator _MUST_ configure the system, the kernel simply doesn't
7202 * know enough information to make a sensible choice.
7204 void sched_set_fifo(struct task_struct *p)
7206 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7207 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7209 EXPORT_SYMBOL_GPL(sched_set_fifo);
7212 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7214 void sched_set_fifo_low(struct task_struct *p)
7216 struct sched_param sp = { .sched_priority = 1 };
7217 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7219 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7221 void sched_set_normal(struct task_struct *p, int nice)
7223 struct sched_attr attr = {
7224 .sched_policy = SCHED_NORMAL,
7227 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7229 EXPORT_SYMBOL_GPL(sched_set_normal);
7232 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7234 struct sched_param lparam;
7235 struct task_struct *p;
7238 if (!param || pid < 0)
7240 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7245 p = find_process_by_pid(pid);
7251 retval = sched_setscheduler(p, policy, &lparam);
7259 * Mimics kernel/events/core.c perf_copy_attr().
7261 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7266 /* Zero the full structure, so that a short copy will be nice: */
7267 memset(attr, 0, sizeof(*attr));
7269 ret = get_user(size, &uattr->size);
7273 /* ABI compatibility quirk: */
7275 size = SCHED_ATTR_SIZE_VER0;
7276 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7279 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7286 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7287 size < SCHED_ATTR_SIZE_VER1)
7291 * XXX: Do we want to be lenient like existing syscalls; or do we want
7292 * to be strict and return an error on out-of-bounds values?
7294 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7299 put_user(sizeof(*attr), &uattr->size);
7304 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7305 * @pid: the pid in question.
7306 * @policy: new policy.
7307 * @param: structure containing the new RT priority.
7309 * Return: 0 on success. An error code otherwise.
7311 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7316 return do_sched_setscheduler(pid, policy, param);
7320 * sys_sched_setparam - set/change the RT priority of a thread
7321 * @pid: the pid in question.
7322 * @param: structure containing the new RT priority.
7324 * Return: 0 on success. An error code otherwise.
7326 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7328 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7332 * sys_sched_setattr - same as above, but with extended sched_attr
7333 * @pid: the pid in question.
7334 * @uattr: structure containing the extended parameters.
7335 * @flags: for future extension.
7337 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7338 unsigned int, flags)
7340 struct sched_attr attr;
7341 struct task_struct *p;
7344 if (!uattr || pid < 0 || flags)
7347 retval = sched_copy_attr(uattr, &attr);
7351 if ((int)attr.sched_policy < 0)
7353 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7354 attr.sched_policy = SETPARAM_POLICY;
7358 p = find_process_by_pid(pid);
7364 retval = sched_setattr(p, &attr);
7372 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7373 * @pid: the pid in question.
7375 * Return: On success, the policy of the thread. Otherwise, a negative error
7378 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7380 struct task_struct *p;
7388 p = find_process_by_pid(pid);
7390 retval = security_task_getscheduler(p);
7393 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7400 * sys_sched_getparam - get the RT priority of a thread
7401 * @pid: the pid in question.
7402 * @param: structure containing the RT priority.
7404 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7407 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7409 struct sched_param lp = { .sched_priority = 0 };
7410 struct task_struct *p;
7413 if (!param || pid < 0)
7417 p = find_process_by_pid(pid);
7422 retval = security_task_getscheduler(p);
7426 if (task_has_rt_policy(p))
7427 lp.sched_priority = p->rt_priority;
7431 * This one might sleep, we cannot do it with a spinlock held ...
7433 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7443 * Copy the kernel size attribute structure (which might be larger
7444 * than what user-space knows about) to user-space.
7446 * Note that all cases are valid: user-space buffer can be larger or
7447 * smaller than the kernel-space buffer. The usual case is that both
7448 * have the same size.
7451 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7452 struct sched_attr *kattr,
7455 unsigned int ksize = sizeof(*kattr);
7457 if (!access_ok(uattr, usize))
7461 * sched_getattr() ABI forwards and backwards compatibility:
7463 * If usize == ksize then we just copy everything to user-space and all is good.
7465 * If usize < ksize then we only copy as much as user-space has space for,
7466 * this keeps ABI compatibility as well. We skip the rest.
7468 * If usize > ksize then user-space is using a newer version of the ABI,
7469 * which part the kernel doesn't know about. Just ignore it - tooling can
7470 * detect the kernel's knowledge of attributes from the attr->size value
7471 * which is set to ksize in this case.
7473 kattr->size = min(usize, ksize);
7475 if (copy_to_user(uattr, kattr, kattr->size))
7482 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7483 * @pid: the pid in question.
7484 * @uattr: structure containing the extended parameters.
7485 * @usize: sizeof(attr) for fwd/bwd comp.
7486 * @flags: for future extension.
7488 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7489 unsigned int, usize, unsigned int, flags)
7491 struct sched_attr kattr = { };
7492 struct task_struct *p;
7495 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7496 usize < SCHED_ATTR_SIZE_VER0 || flags)
7500 p = find_process_by_pid(pid);
7505 retval = security_task_getscheduler(p);
7509 kattr.sched_policy = p->policy;
7510 if (p->sched_reset_on_fork)
7511 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7512 if (task_has_dl_policy(p))
7513 __getparam_dl(p, &kattr);
7514 else if (task_has_rt_policy(p))
7515 kattr.sched_priority = p->rt_priority;
7517 kattr.sched_nice = task_nice(p);
7519 #ifdef CONFIG_UCLAMP_TASK
7521 * This could race with another potential updater, but this is fine
7522 * because it'll correctly read the old or the new value. We don't need
7523 * to guarantee who wins the race as long as it doesn't return garbage.
7525 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7526 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7531 return sched_attr_copy_to_user(uattr, &kattr, usize);
7538 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7540 cpumask_var_t cpus_allowed, new_mask;
7541 struct task_struct *p;
7546 p = find_process_by_pid(pid);
7552 /* Prevent p going away */
7556 if (p->flags & PF_NO_SETAFFINITY) {
7560 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
7564 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7566 goto out_free_cpus_allowed;
7569 if (!check_same_owner(p)) {
7571 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7573 goto out_free_new_mask;
7578 retval = security_task_setscheduler(p);
7580 goto out_free_new_mask;
7583 cpuset_cpus_allowed(p, cpus_allowed);
7584 cpumask_and(new_mask, in_mask, cpus_allowed);
7587 * Since bandwidth control happens on root_domain basis,
7588 * if admission test is enabled, we only admit -deadline
7589 * tasks allowed to run on all the CPUs in the task's
7593 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
7595 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
7598 goto out_free_new_mask;
7604 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
7607 cpuset_cpus_allowed(p, cpus_allowed);
7608 if (!cpumask_subset(new_mask, cpus_allowed)) {
7610 * We must have raced with a concurrent cpuset
7611 * update. Just reset the cpus_allowed to the
7612 * cpuset's cpus_allowed
7614 cpumask_copy(new_mask, cpus_allowed);
7619 free_cpumask_var(new_mask);
7620 out_free_cpus_allowed:
7621 free_cpumask_var(cpus_allowed);
7627 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7628 struct cpumask *new_mask)
7630 if (len < cpumask_size())
7631 cpumask_clear(new_mask);
7632 else if (len > cpumask_size())
7633 len = cpumask_size();
7635 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
7639 * sys_sched_setaffinity - set the CPU affinity of a process
7640 * @pid: pid of the process
7641 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7642 * @user_mask_ptr: user-space pointer to the new CPU mask
7644 * Return: 0 on success. An error code otherwise.
7646 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
7647 unsigned long __user *, user_mask_ptr)
7649 cpumask_var_t new_mask;
7652 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
7655 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
7657 retval = sched_setaffinity(pid, new_mask);
7658 free_cpumask_var(new_mask);
7662 long sched_getaffinity(pid_t pid, struct cpumask *mask)
7664 struct task_struct *p;
7665 unsigned long flags;
7671 p = find_process_by_pid(pid);
7675 retval = security_task_getscheduler(p);
7679 raw_spin_lock_irqsave(&p->pi_lock, flags);
7680 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
7681 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
7690 * sys_sched_getaffinity - get the CPU affinity of a process
7691 * @pid: pid of the process
7692 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7693 * @user_mask_ptr: user-space pointer to hold the current CPU mask
7695 * Return: size of CPU mask copied to user_mask_ptr on success. An
7696 * error code otherwise.
7698 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
7699 unsigned long __user *, user_mask_ptr)
7704 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
7706 if (len & (sizeof(unsigned long)-1))
7709 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
7712 ret = sched_getaffinity(pid, mask);
7714 unsigned int retlen = min(len, cpumask_size());
7716 if (copy_to_user(user_mask_ptr, mask, retlen))
7721 free_cpumask_var(mask);
7726 static void do_sched_yield(void)
7731 rq = this_rq_lock_irq(&rf);
7733 schedstat_inc(rq->yld_count);
7734 current->sched_class->yield_task(rq);
7737 rq_unlock_irq(rq, &rf);
7738 sched_preempt_enable_no_resched();
7744 * sys_sched_yield - yield the current processor to other threads.
7746 * This function yields the current CPU to other tasks. If there are no
7747 * other threads running on this CPU then this function will return.
7751 SYSCALL_DEFINE0(sched_yield)
7757 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
7758 int __sched __cond_resched(void)
7760 if (should_resched(0)) {
7761 preempt_schedule_common();
7764 #ifndef CONFIG_PREEMPT_RCU
7769 EXPORT_SYMBOL(__cond_resched);
7772 #ifdef CONFIG_PREEMPT_DYNAMIC
7773 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
7774 EXPORT_STATIC_CALL_TRAMP(cond_resched);
7776 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
7777 EXPORT_STATIC_CALL_TRAMP(might_resched);
7781 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7782 * call schedule, and on return reacquire the lock.
7784 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7785 * operations here to prevent schedule() from being called twice (once via
7786 * spin_unlock(), once by hand).
7788 int __cond_resched_lock(spinlock_t *lock)
7790 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7793 lockdep_assert_held(lock);
7795 if (spin_needbreak(lock) || resched) {
7798 preempt_schedule_common();
7806 EXPORT_SYMBOL(__cond_resched_lock);
7808 int __cond_resched_rwlock_read(rwlock_t *lock)
7810 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7813 lockdep_assert_held_read(lock);
7815 if (rwlock_needbreak(lock) || resched) {
7818 preempt_schedule_common();
7826 EXPORT_SYMBOL(__cond_resched_rwlock_read);
7828 int __cond_resched_rwlock_write(rwlock_t *lock)
7830 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7833 lockdep_assert_held_write(lock);
7835 if (rwlock_needbreak(lock) || resched) {
7838 preempt_schedule_common();
7846 EXPORT_SYMBOL(__cond_resched_rwlock_write);
7849 * yield - yield the current processor to other threads.
7851 * Do not ever use this function, there's a 99% chance you're doing it wrong.
7853 * The scheduler is at all times free to pick the calling task as the most
7854 * eligible task to run, if removing the yield() call from your code breaks
7855 * it, it's already broken.
7857 * Typical broken usage is:
7862 * where one assumes that yield() will let 'the other' process run that will
7863 * make event true. If the current task is a SCHED_FIFO task that will never
7864 * happen. Never use yield() as a progress guarantee!!
7866 * If you want to use yield() to wait for something, use wait_event().
7867 * If you want to use yield() to be 'nice' for others, use cond_resched().
7868 * If you still want to use yield(), do not!
7870 void __sched yield(void)
7872 set_current_state(TASK_RUNNING);
7875 EXPORT_SYMBOL(yield);
7878 * yield_to - yield the current processor to another thread in
7879 * your thread group, or accelerate that thread toward the
7880 * processor it's on.
7882 * @preempt: whether task preemption is allowed or not
7884 * It's the caller's job to ensure that the target task struct
7885 * can't go away on us before we can do any checks.
7888 * true (>0) if we indeed boosted the target task.
7889 * false (0) if we failed to boost the target.
7890 * -ESRCH if there's no task to yield to.
7892 int __sched yield_to(struct task_struct *p, bool preempt)
7894 struct task_struct *curr = current;
7895 struct rq *rq, *p_rq;
7896 unsigned long flags;
7899 local_irq_save(flags);
7905 * If we're the only runnable task on the rq and target rq also
7906 * has only one task, there's absolutely no point in yielding.
7908 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
7913 double_rq_lock(rq, p_rq);
7914 if (task_rq(p) != p_rq) {
7915 double_rq_unlock(rq, p_rq);
7919 if (!curr->sched_class->yield_to_task)
7922 if (curr->sched_class != p->sched_class)
7925 if (task_running(p_rq, p) || !task_is_running(p))
7928 yielded = curr->sched_class->yield_to_task(rq, p);
7930 schedstat_inc(rq->yld_count);
7932 * Make p's CPU reschedule; pick_next_entity takes care of
7935 if (preempt && rq != p_rq)
7940 double_rq_unlock(rq, p_rq);
7942 local_irq_restore(flags);
7949 EXPORT_SYMBOL_GPL(yield_to);
7951 int io_schedule_prepare(void)
7953 int old_iowait = current->in_iowait;
7955 current->in_iowait = 1;
7956 blk_schedule_flush_plug(current);
7961 void io_schedule_finish(int token)
7963 current->in_iowait = token;
7967 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7968 * that process accounting knows that this is a task in IO wait state.
7970 long __sched io_schedule_timeout(long timeout)
7975 token = io_schedule_prepare();
7976 ret = schedule_timeout(timeout);
7977 io_schedule_finish(token);
7981 EXPORT_SYMBOL(io_schedule_timeout);
7983 void __sched io_schedule(void)
7987 token = io_schedule_prepare();
7989 io_schedule_finish(token);
7991 EXPORT_SYMBOL(io_schedule);
7994 * sys_sched_get_priority_max - return maximum RT priority.
7995 * @policy: scheduling class.
7997 * Return: On success, this syscall returns the maximum
7998 * rt_priority that can be used by a given scheduling class.
7999 * On failure, a negative error code is returned.
8001 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8008 ret = MAX_RT_PRIO-1;
8010 case SCHED_DEADLINE:
8021 * sys_sched_get_priority_min - return minimum RT priority.
8022 * @policy: scheduling class.
8024 * Return: On success, this syscall returns the minimum
8025 * rt_priority that can be used by a given scheduling class.
8026 * On failure, a negative error code is returned.
8028 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8037 case SCHED_DEADLINE:
8046 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8048 struct task_struct *p;
8049 unsigned int time_slice;
8059 p = find_process_by_pid(pid);
8063 retval = security_task_getscheduler(p);
8067 rq = task_rq_lock(p, &rf);
8069 if (p->sched_class->get_rr_interval)
8070 time_slice = p->sched_class->get_rr_interval(rq, p);
8071 task_rq_unlock(rq, p, &rf);
8074 jiffies_to_timespec64(time_slice, t);
8083 * sys_sched_rr_get_interval - return the default timeslice of a process.
8084 * @pid: pid of the process.
8085 * @interval: userspace pointer to the timeslice value.
8087 * this syscall writes the default timeslice value of a given process
8088 * into the user-space timespec buffer. A value of '0' means infinity.
8090 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8093 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8094 struct __kernel_timespec __user *, interval)
8096 struct timespec64 t;
8097 int retval = sched_rr_get_interval(pid, &t);
8100 retval = put_timespec64(&t, interval);
8105 #ifdef CONFIG_COMPAT_32BIT_TIME
8106 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8107 struct old_timespec32 __user *, interval)
8109 struct timespec64 t;
8110 int retval = sched_rr_get_interval(pid, &t);
8113 retval = put_old_timespec32(&t, interval);
8118 void sched_show_task(struct task_struct *p)
8120 unsigned long free = 0;
8123 if (!try_get_task_stack(p))
8126 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8128 if (task_is_running(p))
8129 pr_cont(" running task ");
8130 #ifdef CONFIG_DEBUG_STACK_USAGE
8131 free = stack_not_used(p);
8136 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8138 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8139 free, task_pid_nr(p), ppid,
8140 (unsigned long)task_thread_info(p)->flags);
8142 print_worker_info(KERN_INFO, p);
8143 print_stop_info(KERN_INFO, p);
8144 show_stack(p, NULL, KERN_INFO);
8147 EXPORT_SYMBOL_GPL(sched_show_task);
8150 state_filter_match(unsigned long state_filter, struct task_struct *p)
8152 unsigned int state = READ_ONCE(p->__state);
8154 /* no filter, everything matches */
8158 /* filter, but doesn't match */
8159 if (!(state & state_filter))
8163 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8166 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8173 void show_state_filter(unsigned int state_filter)
8175 struct task_struct *g, *p;
8178 for_each_process_thread(g, p) {
8180 * reset the NMI-timeout, listing all files on a slow
8181 * console might take a lot of time:
8182 * Also, reset softlockup watchdogs on all CPUs, because
8183 * another CPU might be blocked waiting for us to process
8186 touch_nmi_watchdog();
8187 touch_all_softlockup_watchdogs();
8188 if (state_filter_match(state_filter, p))
8192 #ifdef CONFIG_SCHED_DEBUG
8194 sysrq_sched_debug_show();
8198 * Only show locks if all tasks are dumped:
8201 debug_show_all_locks();
8205 * init_idle - set up an idle thread for a given CPU
8206 * @idle: task in question
8207 * @cpu: CPU the idle task belongs to
8209 * NOTE: this function does not set the idle thread's NEED_RESCHED
8210 * flag, to make booting more robust.
8212 void __init init_idle(struct task_struct *idle, int cpu)
8214 struct rq *rq = cpu_rq(cpu);
8215 unsigned long flags;
8217 __sched_fork(0, idle);
8220 * The idle task doesn't need the kthread struct to function, but it
8221 * is dressed up as a per-CPU kthread and thus needs to play the part
8222 * if we want to avoid special-casing it in code that deals with per-CPU
8225 set_kthread_struct(idle);
8227 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8228 raw_spin_rq_lock(rq);
8230 idle->__state = TASK_RUNNING;
8231 idle->se.exec_start = sched_clock();
8233 * PF_KTHREAD should already be set at this point; regardless, make it
8234 * look like a proper per-CPU kthread.
8236 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8237 kthread_set_per_cpu(idle, cpu);
8239 scs_task_reset(idle);
8240 kasan_unpoison_task_stack(idle);
8244 * It's possible that init_idle() gets called multiple times on a task,
8245 * in that case do_set_cpus_allowed() will not do the right thing.
8247 * And since this is boot we can forgo the serialization.
8249 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8252 * We're having a chicken and egg problem, even though we are
8253 * holding rq->lock, the CPU isn't yet set to this CPU so the
8254 * lockdep check in task_group() will fail.
8256 * Similar case to sched_fork(). / Alternatively we could
8257 * use task_rq_lock() here and obtain the other rq->lock.
8262 __set_task_cpu(idle, cpu);
8266 rcu_assign_pointer(rq->curr, idle);
8267 idle->on_rq = TASK_ON_RQ_QUEUED;
8271 raw_spin_rq_unlock(rq);
8272 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8274 /* Set the preempt count _outside_ the spinlocks! */
8275 init_idle_preempt_count(idle, cpu);
8278 * The idle tasks have their own, simple scheduling class:
8280 idle->sched_class = &idle_sched_class;
8281 ftrace_graph_init_idle_task(idle, cpu);
8282 vtime_init_idle(idle, cpu);
8284 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8290 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8291 const struct cpumask *trial)
8295 if (!cpumask_weight(cur))
8298 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8303 int task_can_attach(struct task_struct *p,
8304 const struct cpumask *cs_cpus_allowed)
8309 * Kthreads which disallow setaffinity shouldn't be moved
8310 * to a new cpuset; we don't want to change their CPU
8311 * affinity and isolating such threads by their set of
8312 * allowed nodes is unnecessary. Thus, cpusets are not
8313 * applicable for such threads. This prevents checking for
8314 * success of set_cpus_allowed_ptr() on all attached tasks
8315 * before cpus_mask may be changed.
8317 if (p->flags & PF_NO_SETAFFINITY) {
8322 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8324 ret = dl_task_can_attach(p, cs_cpus_allowed);
8330 bool sched_smp_initialized __read_mostly;
8332 #ifdef CONFIG_NUMA_BALANCING
8333 /* Migrate current task p to target_cpu */
8334 int migrate_task_to(struct task_struct *p, int target_cpu)
8336 struct migration_arg arg = { p, target_cpu };
8337 int curr_cpu = task_cpu(p);
8339 if (curr_cpu == target_cpu)
8342 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8345 /* TODO: This is not properly updating schedstats */
8347 trace_sched_move_numa(p, curr_cpu, target_cpu);
8348 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8352 * Requeue a task on a given node and accurately track the number of NUMA
8353 * tasks on the runqueues
8355 void sched_setnuma(struct task_struct *p, int nid)
8357 bool queued, running;
8361 rq = task_rq_lock(p, &rf);
8362 queued = task_on_rq_queued(p);
8363 running = task_current(rq, p);
8366 dequeue_task(rq, p, DEQUEUE_SAVE);
8368 put_prev_task(rq, p);
8370 p->numa_preferred_nid = nid;
8373 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8375 set_next_task(rq, p);
8376 task_rq_unlock(rq, p, &rf);
8378 #endif /* CONFIG_NUMA_BALANCING */
8380 #ifdef CONFIG_HOTPLUG_CPU
8382 * Ensure that the idle task is using init_mm right before its CPU goes
8385 void idle_task_exit(void)
8387 struct mm_struct *mm = current->active_mm;
8389 BUG_ON(cpu_online(smp_processor_id()));
8390 BUG_ON(current != this_rq()->idle);
8392 if (mm != &init_mm) {
8393 switch_mm(mm, &init_mm, current);
8394 finish_arch_post_lock_switch();
8397 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8400 static int __balance_push_cpu_stop(void *arg)
8402 struct task_struct *p = arg;
8403 struct rq *rq = this_rq();
8407 raw_spin_lock_irq(&p->pi_lock);
8410 update_rq_clock(rq);
8412 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8413 cpu = select_fallback_rq(rq->cpu, p);
8414 rq = __migrate_task(rq, &rf, p, cpu);
8418 raw_spin_unlock_irq(&p->pi_lock);
8425 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8428 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8430 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8431 * effective when the hotplug motion is down.
8433 static void balance_push(struct rq *rq)
8435 struct task_struct *push_task = rq->curr;
8437 lockdep_assert_rq_held(rq);
8438 SCHED_WARN_ON(rq->cpu != smp_processor_id());
8441 * Ensure the thing is persistent until balance_push_set(.on = false);
8443 rq->balance_callback = &balance_push_callback;
8446 * Only active while going offline.
8448 if (!cpu_dying(rq->cpu))
8452 * Both the cpu-hotplug and stop task are in this case and are
8453 * required to complete the hotplug process.
8455 if (kthread_is_per_cpu(push_task) ||
8456 is_migration_disabled(push_task)) {
8459 * If this is the idle task on the outgoing CPU try to wake
8460 * up the hotplug control thread which might wait for the
8461 * last task to vanish. The rcuwait_active() check is
8462 * accurate here because the waiter is pinned on this CPU
8463 * and can't obviously be running in parallel.
8465 * On RT kernels this also has to check whether there are
8466 * pinned and scheduled out tasks on the runqueue. They
8467 * need to leave the migrate disabled section first.
8469 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8470 rcuwait_active(&rq->hotplug_wait)) {
8471 raw_spin_rq_unlock(rq);
8472 rcuwait_wake_up(&rq->hotplug_wait);
8473 raw_spin_rq_lock(rq);
8478 get_task_struct(push_task);
8480 * Temporarily drop rq->lock such that we can wake-up the stop task.
8481 * Both preemption and IRQs are still disabled.
8483 raw_spin_rq_unlock(rq);
8484 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8485 this_cpu_ptr(&push_work));
8487 * At this point need_resched() is true and we'll take the loop in
8488 * schedule(). The next pick is obviously going to be the stop task
8489 * which kthread_is_per_cpu() and will push this task away.
8491 raw_spin_rq_lock(rq);
8494 static void balance_push_set(int cpu, bool on)
8496 struct rq *rq = cpu_rq(cpu);
8499 rq_lock_irqsave(rq, &rf);
8501 WARN_ON_ONCE(rq->balance_callback);
8502 rq->balance_callback = &balance_push_callback;
8503 } else if (rq->balance_callback == &balance_push_callback) {
8504 rq->balance_callback = NULL;
8506 rq_unlock_irqrestore(rq, &rf);
8510 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8511 * inactive. All tasks which are not per CPU kernel threads are either
8512 * pushed off this CPU now via balance_push() or placed on a different CPU
8513 * during wakeup. Wait until the CPU is quiescent.
8515 static void balance_hotplug_wait(void)
8517 struct rq *rq = this_rq();
8519 rcuwait_wait_event(&rq->hotplug_wait,
8520 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8521 TASK_UNINTERRUPTIBLE);
8526 static inline void balance_push(struct rq *rq)
8530 static inline void balance_push_set(int cpu, bool on)
8534 static inline void balance_hotplug_wait(void)
8538 #endif /* CONFIG_HOTPLUG_CPU */
8540 void set_rq_online(struct rq *rq)
8543 const struct sched_class *class;
8545 cpumask_set_cpu(rq->cpu, rq->rd->online);
8548 for_each_class(class) {
8549 if (class->rq_online)
8550 class->rq_online(rq);
8555 void set_rq_offline(struct rq *rq)
8558 const struct sched_class *class;
8560 for_each_class(class) {
8561 if (class->rq_offline)
8562 class->rq_offline(rq);
8565 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8571 * used to mark begin/end of suspend/resume:
8573 static int num_cpus_frozen;
8576 * Update cpusets according to cpu_active mask. If cpusets are
8577 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8578 * around partition_sched_domains().
8580 * If we come here as part of a suspend/resume, don't touch cpusets because we
8581 * want to restore it back to its original state upon resume anyway.
8583 static void cpuset_cpu_active(void)
8585 if (cpuhp_tasks_frozen) {
8587 * num_cpus_frozen tracks how many CPUs are involved in suspend
8588 * resume sequence. As long as this is not the last online
8589 * operation in the resume sequence, just build a single sched
8590 * domain, ignoring cpusets.
8592 partition_sched_domains(1, NULL, NULL);
8593 if (--num_cpus_frozen)
8596 * This is the last CPU online operation. So fall through and
8597 * restore the original sched domains by considering the
8598 * cpuset configurations.
8600 cpuset_force_rebuild();
8602 cpuset_update_active_cpus();
8605 static int cpuset_cpu_inactive(unsigned int cpu)
8607 if (!cpuhp_tasks_frozen) {
8608 if (dl_cpu_busy(cpu))
8610 cpuset_update_active_cpus();
8613 partition_sched_domains(1, NULL, NULL);
8618 int sched_cpu_activate(unsigned int cpu)
8620 struct rq *rq = cpu_rq(cpu);
8624 * Clear the balance_push callback and prepare to schedule
8627 balance_push_set(cpu, false);
8629 #ifdef CONFIG_SCHED_SMT
8631 * When going up, increment the number of cores with SMT present.
8633 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8634 static_branch_inc_cpuslocked(&sched_smt_present);
8636 set_cpu_active(cpu, true);
8638 if (sched_smp_initialized) {
8639 sched_domains_numa_masks_set(cpu);
8640 cpuset_cpu_active();
8644 * Put the rq online, if not already. This happens:
8646 * 1) In the early boot process, because we build the real domains
8647 * after all CPUs have been brought up.
8649 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
8652 rq_lock_irqsave(rq, &rf);
8654 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8657 rq_unlock_irqrestore(rq, &rf);
8662 int sched_cpu_deactivate(unsigned int cpu)
8664 struct rq *rq = cpu_rq(cpu);
8669 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
8670 * load balancing when not active
8672 nohz_balance_exit_idle(rq);
8674 set_cpu_active(cpu, false);
8677 * From this point forward, this CPU will refuse to run any task that
8678 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
8679 * push those tasks away until this gets cleared, see
8680 * sched_cpu_dying().
8682 balance_push_set(cpu, true);
8685 * We've cleared cpu_active_mask / set balance_push, wait for all
8686 * preempt-disabled and RCU users of this state to go away such that
8687 * all new such users will observe it.
8689 * Specifically, we rely on ttwu to no longer target this CPU, see
8690 * ttwu_queue_cond() and is_cpu_allowed().
8692 * Do sync before park smpboot threads to take care the rcu boost case.
8696 rq_lock_irqsave(rq, &rf);
8698 update_rq_clock(rq);
8699 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8702 rq_unlock_irqrestore(rq, &rf);
8704 #ifdef CONFIG_SCHED_SMT
8706 * When going down, decrement the number of cores with SMT present.
8708 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8709 static_branch_dec_cpuslocked(&sched_smt_present);
8712 if (!sched_smp_initialized)
8715 ret = cpuset_cpu_inactive(cpu);
8717 balance_push_set(cpu, false);
8718 set_cpu_active(cpu, true);
8721 sched_domains_numa_masks_clear(cpu);
8725 static void sched_rq_cpu_starting(unsigned int cpu)
8727 struct rq *rq = cpu_rq(cpu);
8729 rq->calc_load_update = calc_load_update;
8730 update_max_interval();
8733 int sched_cpu_starting(unsigned int cpu)
8735 sched_core_cpu_starting(cpu);
8736 sched_rq_cpu_starting(cpu);
8737 sched_tick_start(cpu);
8741 #ifdef CONFIG_HOTPLUG_CPU
8744 * Invoked immediately before the stopper thread is invoked to bring the
8745 * CPU down completely. At this point all per CPU kthreads except the
8746 * hotplug thread (current) and the stopper thread (inactive) have been
8747 * either parked or have been unbound from the outgoing CPU. Ensure that
8748 * any of those which might be on the way out are gone.
8750 * If after this point a bound task is being woken on this CPU then the
8751 * responsible hotplug callback has failed to do it's job.
8752 * sched_cpu_dying() will catch it with the appropriate fireworks.
8754 int sched_cpu_wait_empty(unsigned int cpu)
8756 balance_hotplug_wait();
8761 * Since this CPU is going 'away' for a while, fold any nr_active delta we
8762 * might have. Called from the CPU stopper task after ensuring that the
8763 * stopper is the last running task on the CPU, so nr_active count is
8764 * stable. We need to take the teardown thread which is calling this into
8765 * account, so we hand in adjust = 1 to the load calculation.
8767 * Also see the comment "Global load-average calculations".
8769 static void calc_load_migrate(struct rq *rq)
8771 long delta = calc_load_fold_active(rq, 1);
8774 atomic_long_add(delta, &calc_load_tasks);
8777 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
8779 struct task_struct *g, *p;
8780 int cpu = cpu_of(rq);
8782 lockdep_assert_rq_held(rq);
8784 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
8785 for_each_process_thread(g, p) {
8786 if (task_cpu(p) != cpu)
8789 if (!task_on_rq_queued(p))
8792 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
8796 int sched_cpu_dying(unsigned int cpu)
8798 struct rq *rq = cpu_rq(cpu);
8801 /* Handle pending wakeups and then migrate everything off */
8802 sched_tick_stop(cpu);
8804 rq_lock_irqsave(rq, &rf);
8805 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
8806 WARN(true, "Dying CPU not properly vacated!");
8807 dump_rq_tasks(rq, KERN_WARNING);
8809 rq_unlock_irqrestore(rq, &rf);
8811 calc_load_migrate(rq);
8812 update_max_interval();
8818 void __init sched_init_smp(void)
8823 * There's no userspace yet to cause hotplug operations; hence all the
8824 * CPU masks are stable and all blatant races in the below code cannot
8827 mutex_lock(&sched_domains_mutex);
8828 sched_init_domains(cpu_active_mask);
8829 mutex_unlock(&sched_domains_mutex);
8831 /* Move init over to a non-isolated CPU */
8832 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
8834 current->flags &= ~PF_NO_SETAFFINITY;
8835 sched_init_granularity();
8837 init_sched_rt_class();
8838 init_sched_dl_class();
8840 sched_smp_initialized = true;
8843 static int __init migration_init(void)
8845 sched_cpu_starting(smp_processor_id());
8848 early_initcall(migration_init);
8851 void __init sched_init_smp(void)
8853 sched_init_granularity();
8855 #endif /* CONFIG_SMP */
8857 int in_sched_functions(unsigned long addr)
8859 return in_lock_functions(addr) ||
8860 (addr >= (unsigned long)__sched_text_start
8861 && addr < (unsigned long)__sched_text_end);
8864 #ifdef CONFIG_CGROUP_SCHED
8866 * Default task group.
8867 * Every task in system belongs to this group at bootup.
8869 struct task_group root_task_group;
8870 LIST_HEAD(task_groups);
8872 /* Cacheline aligned slab cache for task_group */
8873 static struct kmem_cache *task_group_cache __read_mostly;
8876 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
8877 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
8879 void __init sched_init(void)
8881 unsigned long ptr = 0;
8884 /* Make sure the linker didn't screw up */
8885 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
8886 &fair_sched_class + 1 != &rt_sched_class ||
8887 &rt_sched_class + 1 != &dl_sched_class);
8889 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
8894 #ifdef CONFIG_FAIR_GROUP_SCHED
8895 ptr += 2 * nr_cpu_ids * sizeof(void **);
8897 #ifdef CONFIG_RT_GROUP_SCHED
8898 ptr += 2 * nr_cpu_ids * sizeof(void **);
8901 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
8903 #ifdef CONFIG_FAIR_GROUP_SCHED
8904 root_task_group.se = (struct sched_entity **)ptr;
8905 ptr += nr_cpu_ids * sizeof(void **);
8907 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8908 ptr += nr_cpu_ids * sizeof(void **);
8910 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
8911 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8912 #endif /* CONFIG_FAIR_GROUP_SCHED */
8913 #ifdef CONFIG_RT_GROUP_SCHED
8914 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8915 ptr += nr_cpu_ids * sizeof(void **);
8917 root_task_group.rt_rq = (struct rt_rq **)ptr;
8918 ptr += nr_cpu_ids * sizeof(void **);
8920 #endif /* CONFIG_RT_GROUP_SCHED */
8922 #ifdef CONFIG_CPUMASK_OFFSTACK
8923 for_each_possible_cpu(i) {
8924 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
8925 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8926 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
8927 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8929 #endif /* CONFIG_CPUMASK_OFFSTACK */
8931 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
8932 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
8935 init_defrootdomain();
8938 #ifdef CONFIG_RT_GROUP_SCHED
8939 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8940 global_rt_period(), global_rt_runtime());
8941 #endif /* CONFIG_RT_GROUP_SCHED */
8943 #ifdef CONFIG_CGROUP_SCHED
8944 task_group_cache = KMEM_CACHE(task_group, 0);
8946 list_add(&root_task_group.list, &task_groups);
8947 INIT_LIST_HEAD(&root_task_group.children);
8948 INIT_LIST_HEAD(&root_task_group.siblings);
8949 autogroup_init(&init_task);
8950 #endif /* CONFIG_CGROUP_SCHED */
8952 for_each_possible_cpu(i) {
8956 raw_spin_lock_init(&rq->__lock);
8958 rq->calc_load_active = 0;
8959 rq->calc_load_update = jiffies + LOAD_FREQ;
8960 init_cfs_rq(&rq->cfs);
8961 init_rt_rq(&rq->rt);
8962 init_dl_rq(&rq->dl);
8963 #ifdef CONFIG_FAIR_GROUP_SCHED
8964 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8965 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
8967 * How much CPU bandwidth does root_task_group get?
8969 * In case of task-groups formed thr' the cgroup filesystem, it
8970 * gets 100% of the CPU resources in the system. This overall
8971 * system CPU resource is divided among the tasks of
8972 * root_task_group and its child task-groups in a fair manner,
8973 * based on each entity's (task or task-group's) weight
8974 * (se->load.weight).
8976 * In other words, if root_task_group has 10 tasks of weight
8977 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8978 * then A0's share of the CPU resource is:
8980 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8982 * We achieve this by letting root_task_group's tasks sit
8983 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8985 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8986 #endif /* CONFIG_FAIR_GROUP_SCHED */
8988 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8989 #ifdef CONFIG_RT_GROUP_SCHED
8990 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8995 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
8996 rq->balance_callback = &balance_push_callback;
8997 rq->active_balance = 0;
8998 rq->next_balance = jiffies;
9003 rq->avg_idle = 2*sysctl_sched_migration_cost;
9004 rq->wake_stamp = jiffies;
9005 rq->wake_avg_idle = rq->avg_idle;
9006 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9008 INIT_LIST_HEAD(&rq->cfs_tasks);
9010 rq_attach_root(rq, &def_root_domain);
9011 #ifdef CONFIG_NO_HZ_COMMON
9012 rq->last_blocked_load_update_tick = jiffies;
9013 atomic_set(&rq->nohz_flags, 0);
9015 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9017 #ifdef CONFIG_HOTPLUG_CPU
9018 rcuwait_init(&rq->hotplug_wait);
9020 #endif /* CONFIG_SMP */
9022 atomic_set(&rq->nr_iowait, 0);
9024 #ifdef CONFIG_SCHED_CORE
9026 rq->core_pick = NULL;
9027 rq->core_enabled = 0;
9028 rq->core_tree = RB_ROOT;
9029 rq->core_forceidle = false;
9031 rq->core_cookie = 0UL;
9035 set_load_weight(&init_task, false);
9038 * The boot idle thread does lazy MMU switching as well:
9041 enter_lazy_tlb(&init_mm, current);
9044 * Make us the idle thread. Technically, schedule() should not be
9045 * called from this thread, however somewhere below it might be,
9046 * but because we are the idle thread, we just pick up running again
9047 * when this runqueue becomes "idle".
9049 init_idle(current, smp_processor_id());
9051 calc_load_update = jiffies + LOAD_FREQ;
9054 idle_thread_set_boot_cpu();
9055 balance_push_set(smp_processor_id(), false);
9057 init_sched_fair_class();
9063 scheduler_running = 1;
9066 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9067 static inline int preempt_count_equals(int preempt_offset)
9069 int nested = preempt_count() + rcu_preempt_depth();
9071 return (nested == preempt_offset);
9074 void __might_sleep(const char *file, int line, int preempt_offset)
9076 unsigned int state = get_current_state();
9078 * Blocking primitives will set (and therefore destroy) current->state,
9079 * since we will exit with TASK_RUNNING make sure we enter with it,
9080 * otherwise we will destroy state.
9082 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9083 "do not call blocking ops when !TASK_RUNNING; "
9084 "state=%x set at [<%p>] %pS\n", state,
9085 (void *)current->task_state_change,
9086 (void *)current->task_state_change);
9088 ___might_sleep(file, line, preempt_offset);
9090 EXPORT_SYMBOL(__might_sleep);
9092 void ___might_sleep(const char *file, int line, int preempt_offset)
9094 /* Ratelimiting timestamp: */
9095 static unsigned long prev_jiffy;
9097 unsigned long preempt_disable_ip;
9099 /* WARN_ON_ONCE() by default, no rate limit required: */
9102 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
9103 !is_idle_task(current) && !current->non_block_count) ||
9104 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9108 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9110 prev_jiffy = jiffies;
9112 /* Save this before calling printk(), since that will clobber it: */
9113 preempt_disable_ip = get_preempt_disable_ip(current);
9116 "BUG: sleeping function called from invalid context at %s:%d\n",
9119 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9120 in_atomic(), irqs_disabled(), current->non_block_count,
9121 current->pid, current->comm);
9123 if (task_stack_end_corrupted(current))
9124 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
9126 debug_show_held_locks(current);
9127 if (irqs_disabled())
9128 print_irqtrace_events(current);
9129 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
9130 && !preempt_count_equals(preempt_offset)) {
9131 pr_err("Preemption disabled at:");
9132 print_ip_sym(KERN_ERR, preempt_disable_ip);
9135 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9137 EXPORT_SYMBOL(___might_sleep);
9139 void __cant_sleep(const char *file, int line, int preempt_offset)
9141 static unsigned long prev_jiffy;
9143 if (irqs_disabled())
9146 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9149 if (preempt_count() > preempt_offset)
9152 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9154 prev_jiffy = jiffies;
9156 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9157 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9158 in_atomic(), irqs_disabled(),
9159 current->pid, current->comm);
9161 debug_show_held_locks(current);
9163 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9165 EXPORT_SYMBOL_GPL(__cant_sleep);
9168 void __cant_migrate(const char *file, int line)
9170 static unsigned long prev_jiffy;
9172 if (irqs_disabled())
9175 if (is_migration_disabled(current))
9178 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9181 if (preempt_count() > 0)
9184 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9186 prev_jiffy = jiffies;
9188 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9189 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9190 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9191 current->pid, current->comm);
9193 debug_show_held_locks(current);
9195 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9197 EXPORT_SYMBOL_GPL(__cant_migrate);
9201 #ifdef CONFIG_MAGIC_SYSRQ
9202 void normalize_rt_tasks(void)
9204 struct task_struct *g, *p;
9205 struct sched_attr attr = {
9206 .sched_policy = SCHED_NORMAL,
9209 read_lock(&tasklist_lock);
9210 for_each_process_thread(g, p) {
9212 * Only normalize user tasks:
9214 if (p->flags & PF_KTHREAD)
9217 p->se.exec_start = 0;
9218 schedstat_set(p->se.statistics.wait_start, 0);
9219 schedstat_set(p->se.statistics.sleep_start, 0);
9220 schedstat_set(p->se.statistics.block_start, 0);
9222 if (!dl_task(p) && !rt_task(p)) {
9224 * Renice negative nice level userspace
9227 if (task_nice(p) < 0)
9228 set_user_nice(p, 0);
9232 __sched_setscheduler(p, &attr, false, false);
9234 read_unlock(&tasklist_lock);
9237 #endif /* CONFIG_MAGIC_SYSRQ */
9239 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9241 * These functions are only useful for the IA64 MCA handling, or kdb.
9243 * They can only be called when the whole system has been
9244 * stopped - every CPU needs to be quiescent, and no scheduling
9245 * activity can take place. Using them for anything else would
9246 * be a serious bug, and as a result, they aren't even visible
9247 * under any other configuration.
9251 * curr_task - return the current task for a given CPU.
9252 * @cpu: the processor in question.
9254 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9256 * Return: The current task for @cpu.
9258 struct task_struct *curr_task(int cpu)
9260 return cpu_curr(cpu);
9263 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9267 * ia64_set_curr_task - set the current task for a given CPU.
9268 * @cpu: the processor in question.
9269 * @p: the task pointer to set.
9271 * Description: This function must only be used when non-maskable interrupts
9272 * are serviced on a separate stack. It allows the architecture to switch the
9273 * notion of the current task on a CPU in a non-blocking manner. This function
9274 * must be called with all CPU's synchronized, and interrupts disabled, the
9275 * and caller must save the original value of the current task (see
9276 * curr_task() above) and restore that value before reenabling interrupts and
9277 * re-starting the system.
9279 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9281 void ia64_set_curr_task(int cpu, struct task_struct *p)
9288 #ifdef CONFIG_CGROUP_SCHED
9289 /* task_group_lock serializes the addition/removal of task groups */
9290 static DEFINE_SPINLOCK(task_group_lock);
9292 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9293 struct task_group *parent)
9295 #ifdef CONFIG_UCLAMP_TASK_GROUP
9296 enum uclamp_id clamp_id;
9298 for_each_clamp_id(clamp_id) {
9299 uclamp_se_set(&tg->uclamp_req[clamp_id],
9300 uclamp_none(clamp_id), false);
9301 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9306 static void sched_free_group(struct task_group *tg)
9308 free_fair_sched_group(tg);
9309 free_rt_sched_group(tg);
9311 kmem_cache_free(task_group_cache, tg);
9314 /* allocate runqueue etc for a new task group */
9315 struct task_group *sched_create_group(struct task_group *parent)
9317 struct task_group *tg;
9319 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9321 return ERR_PTR(-ENOMEM);
9323 if (!alloc_fair_sched_group(tg, parent))
9326 if (!alloc_rt_sched_group(tg, parent))
9329 alloc_uclamp_sched_group(tg, parent);
9334 sched_free_group(tg);
9335 return ERR_PTR(-ENOMEM);
9338 void sched_online_group(struct task_group *tg, struct task_group *parent)
9340 unsigned long flags;
9342 spin_lock_irqsave(&task_group_lock, flags);
9343 list_add_rcu(&tg->list, &task_groups);
9345 /* Root should already exist: */
9348 tg->parent = parent;
9349 INIT_LIST_HEAD(&tg->children);
9350 list_add_rcu(&tg->siblings, &parent->children);
9351 spin_unlock_irqrestore(&task_group_lock, flags);
9353 online_fair_sched_group(tg);
9356 /* rcu callback to free various structures associated with a task group */
9357 static void sched_free_group_rcu(struct rcu_head *rhp)
9359 /* Now it should be safe to free those cfs_rqs: */
9360 sched_free_group(container_of(rhp, struct task_group, rcu));
9363 void sched_destroy_group(struct task_group *tg)
9365 /* Wait for possible concurrent references to cfs_rqs complete: */
9366 call_rcu(&tg->rcu, sched_free_group_rcu);
9369 void sched_offline_group(struct task_group *tg)
9371 unsigned long flags;
9373 /* End participation in shares distribution: */
9374 unregister_fair_sched_group(tg);
9376 spin_lock_irqsave(&task_group_lock, flags);
9377 list_del_rcu(&tg->list);
9378 list_del_rcu(&tg->siblings);
9379 spin_unlock_irqrestore(&task_group_lock, flags);
9382 static void sched_change_group(struct task_struct *tsk, int type)
9384 struct task_group *tg;
9387 * All callers are synchronized by task_rq_lock(); we do not use RCU
9388 * which is pointless here. Thus, we pass "true" to task_css_check()
9389 * to prevent lockdep warnings.
9391 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9392 struct task_group, css);
9393 tg = autogroup_task_group(tsk, tg);
9394 tsk->sched_task_group = tg;
9396 #ifdef CONFIG_FAIR_GROUP_SCHED
9397 if (tsk->sched_class->task_change_group)
9398 tsk->sched_class->task_change_group(tsk, type);
9401 set_task_rq(tsk, task_cpu(tsk));
9405 * Change task's runqueue when it moves between groups.
9407 * The caller of this function should have put the task in its new group by
9408 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9411 void sched_move_task(struct task_struct *tsk)
9413 int queued, running, queue_flags =
9414 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9418 rq = task_rq_lock(tsk, &rf);
9419 update_rq_clock(rq);
9421 running = task_current(rq, tsk);
9422 queued = task_on_rq_queued(tsk);
9425 dequeue_task(rq, tsk, queue_flags);
9427 put_prev_task(rq, tsk);
9429 sched_change_group(tsk, TASK_MOVE_GROUP);
9432 enqueue_task(rq, tsk, queue_flags);
9434 set_next_task(rq, tsk);
9436 * After changing group, the running task may have joined a
9437 * throttled one but it's still the running task. Trigger a
9438 * resched to make sure that task can still run.
9443 task_rq_unlock(rq, tsk, &rf);
9446 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9448 return css ? container_of(css, struct task_group, css) : NULL;
9451 static struct cgroup_subsys_state *
9452 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9454 struct task_group *parent = css_tg(parent_css);
9455 struct task_group *tg;
9458 /* This is early initialization for the top cgroup */
9459 return &root_task_group.css;
9462 tg = sched_create_group(parent);
9464 return ERR_PTR(-ENOMEM);
9469 /* Expose task group only after completing cgroup initialization */
9470 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9472 struct task_group *tg = css_tg(css);
9473 struct task_group *parent = css_tg(css->parent);
9476 sched_online_group(tg, parent);
9478 #ifdef CONFIG_UCLAMP_TASK_GROUP
9479 /* Propagate the effective uclamp value for the new group */
9480 mutex_lock(&uclamp_mutex);
9482 cpu_util_update_eff(css);
9484 mutex_unlock(&uclamp_mutex);
9490 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9492 struct task_group *tg = css_tg(css);
9494 sched_offline_group(tg);
9497 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9499 struct task_group *tg = css_tg(css);
9502 * Relies on the RCU grace period between css_released() and this.
9504 sched_free_group(tg);
9508 * This is called before wake_up_new_task(), therefore we really only
9509 * have to set its group bits, all the other stuff does not apply.
9511 static void cpu_cgroup_fork(struct task_struct *task)
9516 rq = task_rq_lock(task, &rf);
9518 update_rq_clock(rq);
9519 sched_change_group(task, TASK_SET_GROUP);
9521 task_rq_unlock(rq, task, &rf);
9524 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9526 struct task_struct *task;
9527 struct cgroup_subsys_state *css;
9530 cgroup_taskset_for_each(task, css, tset) {
9531 #ifdef CONFIG_RT_GROUP_SCHED
9532 if (!sched_rt_can_attach(css_tg(css), task))
9536 * Serialize against wake_up_new_task() such that if it's
9537 * running, we're sure to observe its full state.
9539 raw_spin_lock_irq(&task->pi_lock);
9541 * Avoid calling sched_move_task() before wake_up_new_task()
9542 * has happened. This would lead to problems with PELT, due to
9543 * move wanting to detach+attach while we're not attached yet.
9545 if (READ_ONCE(task->__state) == TASK_NEW)
9547 raw_spin_unlock_irq(&task->pi_lock);
9555 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9557 struct task_struct *task;
9558 struct cgroup_subsys_state *css;
9560 cgroup_taskset_for_each(task, css, tset)
9561 sched_move_task(task);
9564 #ifdef CONFIG_UCLAMP_TASK_GROUP
9565 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9567 struct cgroup_subsys_state *top_css = css;
9568 struct uclamp_se *uc_parent = NULL;
9569 struct uclamp_se *uc_se = NULL;
9570 unsigned int eff[UCLAMP_CNT];
9571 enum uclamp_id clamp_id;
9572 unsigned int clamps;
9574 lockdep_assert_held(&uclamp_mutex);
9575 SCHED_WARN_ON(!rcu_read_lock_held());
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);
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,
9771 int i, ret = 0, runtime_enabled, runtime_was_enabled;
9772 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9774 if (tg == &root_task_group)
9778 * Ensure we have at some amount of bandwidth every period. This is
9779 * to prevent reaching a state of large arrears when throttled via
9780 * entity_tick() resulting in prolonged exit starvation.
9782 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9786 * Likewise, bound things on the other side by preventing insane quota
9787 * periods. This also allows us to normalize in computing quota
9790 if (period > max_cfs_quota_period)
9794 * Bound quota to defend quota against overflow during bandwidth shift.
9796 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
9799 if (quota != RUNTIME_INF && (burst > quota ||
9800 burst + quota > max_cfs_runtime))
9804 * Prevent race between setting of cfs_rq->runtime_enabled and
9805 * unthrottle_offline_cfs_rqs().
9808 mutex_lock(&cfs_constraints_mutex);
9809 ret = __cfs_schedulable(tg, period, quota);
9813 runtime_enabled = quota != RUNTIME_INF;
9814 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
9816 * If we need to toggle cfs_bandwidth_used, off->on must occur
9817 * before making related changes, and on->off must occur afterwards
9819 if (runtime_enabled && !runtime_was_enabled)
9820 cfs_bandwidth_usage_inc();
9821 raw_spin_lock_irq(&cfs_b->lock);
9822 cfs_b->period = ns_to_ktime(period);
9823 cfs_b->quota = quota;
9824 cfs_b->burst = burst;
9826 __refill_cfs_bandwidth_runtime(cfs_b);
9828 /* Restart the period timer (if active) to handle new period expiry: */
9829 if (runtime_enabled)
9830 start_cfs_bandwidth(cfs_b);
9832 raw_spin_unlock_irq(&cfs_b->lock);
9834 for_each_online_cpu(i) {
9835 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9836 struct rq *rq = cfs_rq->rq;
9839 rq_lock_irq(rq, &rf);
9840 cfs_rq->runtime_enabled = runtime_enabled;
9841 cfs_rq->runtime_remaining = 0;
9843 if (cfs_rq->throttled)
9844 unthrottle_cfs_rq(cfs_rq);
9845 rq_unlock_irq(rq, &rf);
9847 if (runtime_was_enabled && !runtime_enabled)
9848 cfs_bandwidth_usage_dec();
9850 mutex_unlock(&cfs_constraints_mutex);
9856 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9858 u64 quota, period, burst;
9860 period = ktime_to_ns(tg->cfs_bandwidth.period);
9861 burst = tg->cfs_bandwidth.burst;
9862 if (cfs_quota_us < 0)
9863 quota = RUNTIME_INF;
9864 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
9865 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9869 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9872 static long tg_get_cfs_quota(struct task_group *tg)
9876 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
9879 quota_us = tg->cfs_bandwidth.quota;
9880 do_div(quota_us, NSEC_PER_USEC);
9885 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9887 u64 quota, period, burst;
9889 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
9892 period = (u64)cfs_period_us * NSEC_PER_USEC;
9893 quota = tg->cfs_bandwidth.quota;
9894 burst = tg->cfs_bandwidth.burst;
9896 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9899 static long tg_get_cfs_period(struct task_group *tg)
9903 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
9904 do_div(cfs_period_us, NSEC_PER_USEC);
9906 return cfs_period_us;
9909 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
9911 u64 quota, period, burst;
9913 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
9916 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
9917 period = ktime_to_ns(tg->cfs_bandwidth.period);
9918 quota = tg->cfs_bandwidth.quota;
9920 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9923 static long tg_get_cfs_burst(struct task_group *tg)
9927 burst_us = tg->cfs_bandwidth.burst;
9928 do_div(burst_us, NSEC_PER_USEC);
9933 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
9936 return tg_get_cfs_quota(css_tg(css));
9939 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
9940 struct cftype *cftype, s64 cfs_quota_us)
9942 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
9945 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
9948 return tg_get_cfs_period(css_tg(css));
9951 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
9952 struct cftype *cftype, u64 cfs_period_us)
9954 return tg_set_cfs_period(css_tg(css), cfs_period_us);
9957 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
9960 return tg_get_cfs_burst(css_tg(css));
9963 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
9964 struct cftype *cftype, u64 cfs_burst_us)
9966 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
9969 struct cfs_schedulable_data {
9970 struct task_group *tg;
9975 * normalize group quota/period to be quota/max_period
9976 * note: units are usecs
9978 static u64 normalize_cfs_quota(struct task_group *tg,
9979 struct cfs_schedulable_data *d)
9987 period = tg_get_cfs_period(tg);
9988 quota = tg_get_cfs_quota(tg);
9991 /* note: these should typically be equivalent */
9992 if (quota == RUNTIME_INF || quota == -1)
9995 return to_ratio(period, quota);
9998 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10000 struct cfs_schedulable_data *d = data;
10001 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10002 s64 quota = 0, parent_quota = -1;
10005 quota = RUNTIME_INF;
10007 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10009 quota = normalize_cfs_quota(tg, d);
10010 parent_quota = parent_b->hierarchical_quota;
10013 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10014 * always take the min. On cgroup1, only inherit when no
10017 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10018 quota = min(quota, parent_quota);
10020 if (quota == RUNTIME_INF)
10021 quota = parent_quota;
10022 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10026 cfs_b->hierarchical_quota = quota;
10031 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10034 struct cfs_schedulable_data data = {
10040 if (quota != RUNTIME_INF) {
10041 do_div(data.period, NSEC_PER_USEC);
10042 do_div(data.quota, NSEC_PER_USEC);
10046 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10052 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10054 struct task_group *tg = css_tg(seq_css(sf));
10055 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10057 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10058 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10059 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10061 if (schedstat_enabled() && tg != &root_task_group) {
10065 for_each_possible_cpu(i)
10066 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10068 seq_printf(sf, "wait_sum %llu\n", ws);
10073 #endif /* CONFIG_CFS_BANDWIDTH */
10074 #endif /* CONFIG_FAIR_GROUP_SCHED */
10076 #ifdef CONFIG_RT_GROUP_SCHED
10077 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10078 struct cftype *cft, s64 val)
10080 return sched_group_set_rt_runtime(css_tg(css), val);
10083 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10084 struct cftype *cft)
10086 return sched_group_rt_runtime(css_tg(css));
10089 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10090 struct cftype *cftype, u64 rt_period_us)
10092 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10095 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10096 struct cftype *cft)
10098 return sched_group_rt_period(css_tg(css));
10100 #endif /* CONFIG_RT_GROUP_SCHED */
10102 static struct cftype cpu_legacy_files[] = {
10103 #ifdef CONFIG_FAIR_GROUP_SCHED
10106 .read_u64 = cpu_shares_read_u64,
10107 .write_u64 = cpu_shares_write_u64,
10110 #ifdef CONFIG_CFS_BANDWIDTH
10112 .name = "cfs_quota_us",
10113 .read_s64 = cpu_cfs_quota_read_s64,
10114 .write_s64 = cpu_cfs_quota_write_s64,
10117 .name = "cfs_period_us",
10118 .read_u64 = cpu_cfs_period_read_u64,
10119 .write_u64 = cpu_cfs_period_write_u64,
10122 .name = "cfs_burst_us",
10123 .read_u64 = cpu_cfs_burst_read_u64,
10124 .write_u64 = cpu_cfs_burst_write_u64,
10128 .seq_show = cpu_cfs_stat_show,
10131 #ifdef CONFIG_RT_GROUP_SCHED
10133 .name = "rt_runtime_us",
10134 .read_s64 = cpu_rt_runtime_read,
10135 .write_s64 = cpu_rt_runtime_write,
10138 .name = "rt_period_us",
10139 .read_u64 = cpu_rt_period_read_uint,
10140 .write_u64 = cpu_rt_period_write_uint,
10143 #ifdef CONFIG_UCLAMP_TASK_GROUP
10145 .name = "uclamp.min",
10146 .flags = CFTYPE_NOT_ON_ROOT,
10147 .seq_show = cpu_uclamp_min_show,
10148 .write = cpu_uclamp_min_write,
10151 .name = "uclamp.max",
10152 .flags = CFTYPE_NOT_ON_ROOT,
10153 .seq_show = cpu_uclamp_max_show,
10154 .write = cpu_uclamp_max_write,
10157 { } /* Terminate */
10160 static int cpu_extra_stat_show(struct seq_file *sf,
10161 struct cgroup_subsys_state *css)
10163 #ifdef CONFIG_CFS_BANDWIDTH
10165 struct task_group *tg = css_tg(css);
10166 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10167 u64 throttled_usec;
10169 throttled_usec = cfs_b->throttled_time;
10170 do_div(throttled_usec, NSEC_PER_USEC);
10172 seq_printf(sf, "nr_periods %d\n"
10173 "nr_throttled %d\n"
10174 "throttled_usec %llu\n",
10175 cfs_b->nr_periods, cfs_b->nr_throttled,
10182 #ifdef CONFIG_FAIR_GROUP_SCHED
10183 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10184 struct cftype *cft)
10186 struct task_group *tg = css_tg(css);
10187 u64 weight = scale_load_down(tg->shares);
10189 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10192 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10193 struct cftype *cft, u64 weight)
10196 * cgroup weight knobs should use the common MIN, DFL and MAX
10197 * values which are 1, 100 and 10000 respectively. While it loses
10198 * a bit of range on both ends, it maps pretty well onto the shares
10199 * value used by scheduler and the round-trip conversions preserve
10200 * the original value over the entire range.
10202 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10205 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10207 return sched_group_set_shares(css_tg(css), scale_load(weight));
10210 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10211 struct cftype *cft)
10213 unsigned long weight = scale_load_down(css_tg(css)->shares);
10214 int last_delta = INT_MAX;
10217 /* find the closest nice value to the current weight */
10218 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10219 delta = abs(sched_prio_to_weight[prio] - weight);
10220 if (delta >= last_delta)
10222 last_delta = delta;
10225 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10228 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10229 struct cftype *cft, s64 nice)
10231 unsigned long weight;
10234 if (nice < MIN_NICE || nice > MAX_NICE)
10237 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10238 idx = array_index_nospec(idx, 40);
10239 weight = sched_prio_to_weight[idx];
10241 return sched_group_set_shares(css_tg(css), scale_load(weight));
10245 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10246 long period, long quota)
10249 seq_puts(sf, "max");
10251 seq_printf(sf, "%ld", quota);
10253 seq_printf(sf, " %ld\n", period);
10256 /* caller should put the current value in *@periodp before calling */
10257 static int __maybe_unused cpu_period_quota_parse(char *buf,
10258 u64 *periodp, u64 *quotap)
10260 char tok[21]; /* U64_MAX */
10262 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10265 *periodp *= NSEC_PER_USEC;
10267 if (sscanf(tok, "%llu", quotap))
10268 *quotap *= NSEC_PER_USEC;
10269 else if (!strcmp(tok, "max"))
10270 *quotap = RUNTIME_INF;
10277 #ifdef CONFIG_CFS_BANDWIDTH
10278 static int cpu_max_show(struct seq_file *sf, void *v)
10280 struct task_group *tg = css_tg(seq_css(sf));
10282 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10286 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10287 char *buf, size_t nbytes, loff_t off)
10289 struct task_group *tg = css_tg(of_css(of));
10290 u64 period = tg_get_cfs_period(tg);
10291 u64 burst = tg_get_cfs_burst(tg);
10295 ret = cpu_period_quota_parse(buf, &period, "a);
10297 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10298 return ret ?: nbytes;
10302 static struct cftype cpu_files[] = {
10303 #ifdef CONFIG_FAIR_GROUP_SCHED
10306 .flags = CFTYPE_NOT_ON_ROOT,
10307 .read_u64 = cpu_weight_read_u64,
10308 .write_u64 = cpu_weight_write_u64,
10311 .name = "weight.nice",
10312 .flags = CFTYPE_NOT_ON_ROOT,
10313 .read_s64 = cpu_weight_nice_read_s64,
10314 .write_s64 = cpu_weight_nice_write_s64,
10317 #ifdef CONFIG_CFS_BANDWIDTH
10320 .flags = CFTYPE_NOT_ON_ROOT,
10321 .seq_show = cpu_max_show,
10322 .write = cpu_max_write,
10325 .name = "max.burst",
10326 .flags = CFTYPE_NOT_ON_ROOT,
10327 .read_u64 = cpu_cfs_burst_read_u64,
10328 .write_u64 = cpu_cfs_burst_write_u64,
10331 #ifdef CONFIG_UCLAMP_TASK_GROUP
10333 .name = "uclamp.min",
10334 .flags = CFTYPE_NOT_ON_ROOT,
10335 .seq_show = cpu_uclamp_min_show,
10336 .write = cpu_uclamp_min_write,
10339 .name = "uclamp.max",
10340 .flags = CFTYPE_NOT_ON_ROOT,
10341 .seq_show = cpu_uclamp_max_show,
10342 .write = cpu_uclamp_max_write,
10345 { } /* terminate */
10348 struct cgroup_subsys cpu_cgrp_subsys = {
10349 .css_alloc = cpu_cgroup_css_alloc,
10350 .css_online = cpu_cgroup_css_online,
10351 .css_released = cpu_cgroup_css_released,
10352 .css_free = cpu_cgroup_css_free,
10353 .css_extra_stat_show = cpu_extra_stat_show,
10354 .fork = cpu_cgroup_fork,
10355 .can_attach = cpu_cgroup_can_attach,
10356 .attach = cpu_cgroup_attach,
10357 .legacy_cftypes = cpu_legacy_files,
10358 .dfl_cftypes = cpu_files,
10359 .early_init = true,
10363 #endif /* CONFIG_CGROUP_SCHED */
10365 void dump_cpu_task(int cpu)
10367 pr_info("Task dump for CPU %d:\n", cpu);
10368 sched_show_task(cpu_curr(cpu));
10372 * Nice levels are multiplicative, with a gentle 10% change for every
10373 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10374 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10375 * that remained on nice 0.
10377 * The "10% effect" is relative and cumulative: from _any_ nice level,
10378 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10379 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10380 * If a task goes up by ~10% and another task goes down by ~10% then
10381 * the relative distance between them is ~25%.)
10383 const int sched_prio_to_weight[40] = {
10384 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10385 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10386 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10387 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10388 /* 0 */ 1024, 820, 655, 526, 423,
10389 /* 5 */ 335, 272, 215, 172, 137,
10390 /* 10 */ 110, 87, 70, 56, 45,
10391 /* 15 */ 36, 29, 23, 18, 15,
10395 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10397 * In cases where the weight does not change often, we can use the
10398 * precalculated inverse to speed up arithmetics by turning divisions
10399 * into multiplications:
10401 const u32 sched_prio_to_wmult[40] = {
10402 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10403 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10404 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10405 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10406 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10407 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10408 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10409 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10412 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10414 trace_sched_update_nr_running_tp(rq, count);