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);
1985 * __normal_prio - return the priority that is based on the static prio
1987 static inline int __normal_prio(struct task_struct *p)
1989 return p->static_prio;
1993 * Calculate the expected normal priority: i.e. priority
1994 * without taking RT-inheritance into account. Might be
1995 * boosted by interactivity modifiers. Changes upon fork,
1996 * setprio syscalls, and whenever the interactivity
1997 * estimator recalculates.
1999 static inline int normal_prio(struct task_struct *p)
2003 if (task_has_dl_policy(p))
2004 prio = MAX_DL_PRIO-1;
2005 else if (task_has_rt_policy(p))
2006 prio = MAX_RT_PRIO-1 - p->rt_priority;
2008 prio = __normal_prio(p);
2013 * Calculate the current priority, i.e. the priority
2014 * taken into account by the scheduler. This value might
2015 * be boosted by RT tasks, or might be boosted by
2016 * interactivity modifiers. Will be RT if the task got
2017 * RT-boosted. If not then it returns p->normal_prio.
2019 static int effective_prio(struct task_struct *p)
2021 p->normal_prio = normal_prio(p);
2023 * If we are RT tasks or we were boosted to RT priority,
2024 * keep the priority unchanged. Otherwise, update priority
2025 * to the normal priority:
2027 if (!rt_prio(p->prio))
2028 return p->normal_prio;
2033 * task_curr - is this task currently executing on a CPU?
2034 * @p: the task in question.
2036 * Return: 1 if the task is currently executing. 0 otherwise.
2038 inline int task_curr(const struct task_struct *p)
2040 return cpu_curr(task_cpu(p)) == p;
2044 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2045 * use the balance_callback list if you want balancing.
2047 * this means any call to check_class_changed() must be followed by a call to
2048 * balance_callback().
2050 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2051 const struct sched_class *prev_class,
2054 if (prev_class != p->sched_class) {
2055 if (prev_class->switched_from)
2056 prev_class->switched_from(rq, p);
2058 p->sched_class->switched_to(rq, p);
2059 } else if (oldprio != p->prio || dl_task(p))
2060 p->sched_class->prio_changed(rq, p, oldprio);
2063 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2065 if (p->sched_class == rq->curr->sched_class)
2066 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2067 else if (p->sched_class > rq->curr->sched_class)
2071 * A queue event has occurred, and we're going to schedule. In
2072 * this case, we can save a useless back to back clock update.
2074 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2075 rq_clock_skip_update(rq);
2081 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2083 static int __set_cpus_allowed_ptr(struct task_struct *p,
2084 const struct cpumask *new_mask,
2087 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2089 if (likely(!p->migration_disabled))
2092 if (p->cpus_ptr != &p->cpus_mask)
2096 * Violates locking rules! see comment in __do_set_cpus_allowed().
2098 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2101 void migrate_disable(void)
2103 struct task_struct *p = current;
2105 if (p->migration_disabled) {
2106 p->migration_disabled++;
2111 this_rq()->nr_pinned++;
2112 p->migration_disabled = 1;
2115 EXPORT_SYMBOL_GPL(migrate_disable);
2117 void migrate_enable(void)
2119 struct task_struct *p = current;
2121 if (p->migration_disabled > 1) {
2122 p->migration_disabled--;
2127 * Ensure stop_task runs either before or after this, and that
2128 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2131 if (p->cpus_ptr != &p->cpus_mask)
2132 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2134 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2135 * regular cpus_mask, otherwise things that race (eg.
2136 * select_fallback_rq) get confused.
2139 p->migration_disabled = 0;
2140 this_rq()->nr_pinned--;
2143 EXPORT_SYMBOL_GPL(migrate_enable);
2145 static inline bool rq_has_pinned_tasks(struct rq *rq)
2147 return rq->nr_pinned;
2151 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2152 * __set_cpus_allowed_ptr() and select_fallback_rq().
2154 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2156 /* When not in the task's cpumask, no point in looking further. */
2157 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2160 /* migrate_disabled() must be allowed to finish. */
2161 if (is_migration_disabled(p))
2162 return cpu_online(cpu);
2164 /* Non kernel threads are not allowed during either online or offline. */
2165 if (!(p->flags & PF_KTHREAD))
2166 return cpu_active(cpu);
2168 /* KTHREAD_IS_PER_CPU is always allowed. */
2169 if (kthread_is_per_cpu(p))
2170 return cpu_online(cpu);
2172 /* Regular kernel threads don't get to stay during offline. */
2176 /* But are allowed during online. */
2177 return cpu_online(cpu);
2181 * This is how migration works:
2183 * 1) we invoke migration_cpu_stop() on the target CPU using
2185 * 2) stopper starts to run (implicitly forcing the migrated thread
2187 * 3) it checks whether the migrated task is still in the wrong runqueue.
2188 * 4) if it's in the wrong runqueue then the migration thread removes
2189 * it and puts it into the right queue.
2190 * 5) stopper completes and stop_one_cpu() returns and the migration
2195 * move_queued_task - move a queued task to new rq.
2197 * Returns (locked) new rq. Old rq's lock is released.
2199 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2200 struct task_struct *p, int new_cpu)
2202 lockdep_assert_rq_held(rq);
2204 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2205 set_task_cpu(p, new_cpu);
2208 rq = cpu_rq(new_cpu);
2211 BUG_ON(task_cpu(p) != new_cpu);
2212 activate_task(rq, p, 0);
2213 check_preempt_curr(rq, p, 0);
2218 struct migration_arg {
2219 struct task_struct *task;
2221 struct set_affinity_pending *pending;
2225 * @refs: number of wait_for_completion()
2226 * @stop_pending: is @stop_work in use
2228 struct set_affinity_pending {
2230 unsigned int stop_pending;
2231 struct completion done;
2232 struct cpu_stop_work stop_work;
2233 struct migration_arg arg;
2237 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2238 * this because either it can't run here any more (set_cpus_allowed()
2239 * away from this CPU, or CPU going down), or because we're
2240 * attempting to rebalance this task on exec (sched_exec).
2242 * So we race with normal scheduler movements, but that's OK, as long
2243 * as the task is no longer on this CPU.
2245 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2246 struct task_struct *p, int dest_cpu)
2248 /* Affinity changed (again). */
2249 if (!is_cpu_allowed(p, dest_cpu))
2252 update_rq_clock(rq);
2253 rq = move_queued_task(rq, rf, p, dest_cpu);
2259 * migration_cpu_stop - this will be executed by a highprio stopper thread
2260 * and performs thread migration by bumping thread off CPU then
2261 * 'pushing' onto another runqueue.
2263 static int migration_cpu_stop(void *data)
2265 struct migration_arg *arg = data;
2266 struct set_affinity_pending *pending = arg->pending;
2267 struct task_struct *p = arg->task;
2268 struct rq *rq = this_rq();
2269 bool complete = false;
2273 * The original target CPU might have gone down and we might
2274 * be on another CPU but it doesn't matter.
2276 local_irq_save(rf.flags);
2278 * We need to explicitly wake pending tasks before running
2279 * __migrate_task() such that we will not miss enforcing cpus_ptr
2280 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2282 flush_smp_call_function_from_idle();
2284 raw_spin_lock(&p->pi_lock);
2288 * If we were passed a pending, then ->stop_pending was set, thus
2289 * p->migration_pending must have remained stable.
2291 WARN_ON_ONCE(pending && pending != p->migration_pending);
2294 * If task_rq(p) != rq, it cannot be migrated here, because we're
2295 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2296 * we're holding p->pi_lock.
2298 if (task_rq(p) == rq) {
2299 if (is_migration_disabled(p))
2303 p->migration_pending = NULL;
2306 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2310 if (task_on_rq_queued(p))
2311 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2313 p->wake_cpu = arg->dest_cpu;
2316 * XXX __migrate_task() can fail, at which point we might end
2317 * up running on a dodgy CPU, AFAICT this can only happen
2318 * during CPU hotplug, at which point we'll get pushed out
2319 * anyway, so it's probably not a big deal.
2322 } else if (pending) {
2324 * This happens when we get migrated between migrate_enable()'s
2325 * preempt_enable() and scheduling the stopper task. At that
2326 * point we're a regular task again and not current anymore.
2328 * A !PREEMPT kernel has a giant hole here, which makes it far
2333 * The task moved before the stopper got to run. We're holding
2334 * ->pi_lock, so the allowed mask is stable - if it got
2335 * somewhere allowed, we're done.
2337 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2338 p->migration_pending = NULL;
2344 * When migrate_enable() hits a rq mis-match we can't reliably
2345 * determine is_migration_disabled() and so have to chase after
2348 WARN_ON_ONCE(!pending->stop_pending);
2349 task_rq_unlock(rq, p, &rf);
2350 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2351 &pending->arg, &pending->stop_work);
2356 pending->stop_pending = false;
2357 task_rq_unlock(rq, p, &rf);
2360 complete_all(&pending->done);
2365 int push_cpu_stop(void *arg)
2367 struct rq *lowest_rq = NULL, *rq = this_rq();
2368 struct task_struct *p = arg;
2370 raw_spin_lock_irq(&p->pi_lock);
2371 raw_spin_rq_lock(rq);
2373 if (task_rq(p) != rq)
2376 if (is_migration_disabled(p)) {
2377 p->migration_flags |= MDF_PUSH;
2381 p->migration_flags &= ~MDF_PUSH;
2383 if (p->sched_class->find_lock_rq)
2384 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2389 // XXX validate p is still the highest prio task
2390 if (task_rq(p) == rq) {
2391 deactivate_task(rq, p, 0);
2392 set_task_cpu(p, lowest_rq->cpu);
2393 activate_task(lowest_rq, p, 0);
2394 resched_curr(lowest_rq);
2397 double_unlock_balance(rq, lowest_rq);
2400 rq->push_busy = false;
2401 raw_spin_rq_unlock(rq);
2402 raw_spin_unlock_irq(&p->pi_lock);
2409 * sched_class::set_cpus_allowed must do the below, but is not required to
2410 * actually call this function.
2412 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2414 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2415 p->cpus_ptr = new_mask;
2419 cpumask_copy(&p->cpus_mask, new_mask);
2420 p->nr_cpus_allowed = cpumask_weight(new_mask);
2424 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2426 struct rq *rq = task_rq(p);
2427 bool queued, running;
2430 * This here violates the locking rules for affinity, since we're only
2431 * supposed to change these variables while holding both rq->lock and
2434 * HOWEVER, it magically works, because ttwu() is the only code that
2435 * accesses these variables under p->pi_lock and only does so after
2436 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2437 * before finish_task().
2439 * XXX do further audits, this smells like something putrid.
2441 if (flags & SCA_MIGRATE_DISABLE)
2442 SCHED_WARN_ON(!p->on_cpu);
2444 lockdep_assert_held(&p->pi_lock);
2446 queued = task_on_rq_queued(p);
2447 running = task_current(rq, p);
2451 * Because __kthread_bind() calls this on blocked tasks without
2454 lockdep_assert_rq_held(rq);
2455 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2458 put_prev_task(rq, p);
2460 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2463 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2465 set_next_task(rq, p);
2468 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2470 __do_set_cpus_allowed(p, new_mask, 0);
2474 * This function is wildly self concurrent; here be dragons.
2477 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2478 * designated task is enqueued on an allowed CPU. If that task is currently
2479 * running, we have to kick it out using the CPU stopper.
2481 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2484 * Initial conditions: P0->cpus_mask = [0, 1]
2488 * migrate_disable();
2490 * set_cpus_allowed_ptr(P0, [1]);
2492 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2493 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2494 * This means we need the following scheme:
2498 * migrate_disable();
2500 * set_cpus_allowed_ptr(P0, [1]);
2504 * __set_cpus_allowed_ptr();
2505 * <wakes local stopper>
2506 * `--> <woken on migration completion>
2508 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2509 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2510 * task p are serialized by p->pi_lock, which we can leverage: the one that
2511 * should come into effect at the end of the Migrate-Disable region is the last
2512 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2513 * but we still need to properly signal those waiting tasks at the appropriate
2516 * This is implemented using struct set_affinity_pending. The first
2517 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2518 * setup an instance of that struct and install it on the targeted task_struct.
2519 * Any and all further callers will reuse that instance. Those then wait for
2520 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2521 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2524 * (1) In the cases covered above. There is one more where the completion is
2525 * signaled within affine_move_task() itself: when a subsequent affinity request
2526 * occurs after the stopper bailed out due to the targeted task still being
2527 * Migrate-Disable. Consider:
2529 * Initial conditions: P0->cpus_mask = [0, 1]
2533 * migrate_disable();
2535 * set_cpus_allowed_ptr(P0, [1]);
2538 * migration_cpu_stop()
2539 * is_migration_disabled()
2541 * set_cpus_allowed_ptr(P0, [0, 1]);
2542 * <signal completion>
2545 * Note that the above is safe vs a concurrent migrate_enable(), as any
2546 * pending affinity completion is preceded by an uninstallation of
2547 * p->migration_pending done with p->pi_lock held.
2549 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2550 int dest_cpu, unsigned int flags)
2552 struct set_affinity_pending my_pending = { }, *pending = NULL;
2553 bool stop_pending, complete = false;
2555 /* Can the task run on the task's current CPU? If so, we're done */
2556 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2557 struct task_struct *push_task = NULL;
2559 if ((flags & SCA_MIGRATE_ENABLE) &&
2560 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2561 rq->push_busy = true;
2562 push_task = get_task_struct(p);
2566 * If there are pending waiters, but no pending stop_work,
2567 * then complete now.
2569 pending = p->migration_pending;
2570 if (pending && !pending->stop_pending) {
2571 p->migration_pending = NULL;
2575 task_rq_unlock(rq, p, rf);
2578 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2583 complete_all(&pending->done);
2588 if (!(flags & SCA_MIGRATE_ENABLE)) {
2589 /* serialized by p->pi_lock */
2590 if (!p->migration_pending) {
2591 /* Install the request */
2592 refcount_set(&my_pending.refs, 1);
2593 init_completion(&my_pending.done);
2594 my_pending.arg = (struct migration_arg) {
2596 .dest_cpu = dest_cpu,
2597 .pending = &my_pending,
2600 p->migration_pending = &my_pending;
2602 pending = p->migration_pending;
2603 refcount_inc(&pending->refs);
2605 * Affinity has changed, but we've already installed a
2606 * pending. migration_cpu_stop() *must* see this, else
2607 * we risk a completion of the pending despite having a
2608 * task on a disallowed CPU.
2610 * Serialized by p->pi_lock, so this is safe.
2612 pending->arg.dest_cpu = dest_cpu;
2615 pending = p->migration_pending;
2617 * - !MIGRATE_ENABLE:
2618 * we'll have installed a pending if there wasn't one already.
2621 * we're here because the current CPU isn't matching anymore,
2622 * the only way that can happen is because of a concurrent
2623 * set_cpus_allowed_ptr() call, which should then still be
2624 * pending completion.
2626 * Either way, we really should have a @pending here.
2628 if (WARN_ON_ONCE(!pending)) {
2629 task_rq_unlock(rq, p, rf);
2633 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2635 * MIGRATE_ENABLE gets here because 'p == current', but for
2636 * anything else we cannot do is_migration_disabled(), punt
2637 * and have the stopper function handle it all race-free.
2639 stop_pending = pending->stop_pending;
2641 pending->stop_pending = true;
2643 if (flags & SCA_MIGRATE_ENABLE)
2644 p->migration_flags &= ~MDF_PUSH;
2646 task_rq_unlock(rq, p, rf);
2648 if (!stop_pending) {
2649 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2650 &pending->arg, &pending->stop_work);
2653 if (flags & SCA_MIGRATE_ENABLE)
2657 if (!is_migration_disabled(p)) {
2658 if (task_on_rq_queued(p))
2659 rq = move_queued_task(rq, rf, p, dest_cpu);
2661 if (!pending->stop_pending) {
2662 p->migration_pending = NULL;
2666 task_rq_unlock(rq, p, rf);
2669 complete_all(&pending->done);
2672 wait_for_completion(&pending->done);
2674 if (refcount_dec_and_test(&pending->refs))
2675 wake_up_var(&pending->refs); /* No UaF, just an address */
2678 * Block the original owner of &pending until all subsequent callers
2679 * have seen the completion and decremented the refcount
2681 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2684 WARN_ON_ONCE(my_pending.stop_pending);
2690 * Change a given task's CPU affinity. Migrate the thread to a
2691 * proper CPU and schedule it away if the CPU it's executing on
2692 * is removed from the allowed bitmask.
2694 * NOTE: the caller must have a valid reference to the task, the
2695 * task must not exit() & deallocate itself prematurely. The
2696 * call is not atomic; no spinlocks may be held.
2698 static int __set_cpus_allowed_ptr(struct task_struct *p,
2699 const struct cpumask *new_mask,
2702 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2703 unsigned int dest_cpu;
2708 rq = task_rq_lock(p, &rf);
2709 update_rq_clock(rq);
2711 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2713 * Kernel threads are allowed on online && !active CPUs,
2714 * however, during cpu-hot-unplug, even these might get pushed
2715 * away if not KTHREAD_IS_PER_CPU.
2717 * Specifically, migration_disabled() tasks must not fail the
2718 * cpumask_any_and_distribute() pick below, esp. so on
2719 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2720 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2722 cpu_valid_mask = cpu_online_mask;
2726 * Must re-check here, to close a race against __kthread_bind(),
2727 * sched_setaffinity() is not guaranteed to observe the flag.
2729 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2734 if (!(flags & SCA_MIGRATE_ENABLE)) {
2735 if (cpumask_equal(&p->cpus_mask, new_mask))
2738 if (WARN_ON_ONCE(p == current &&
2739 is_migration_disabled(p) &&
2740 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2747 * Picking a ~random cpu helps in cases where we are changing affinity
2748 * for groups of tasks (ie. cpuset), so that load balancing is not
2749 * immediately required to distribute the tasks within their new mask.
2751 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2752 if (dest_cpu >= nr_cpu_ids) {
2757 __do_set_cpus_allowed(p, new_mask, flags);
2759 return affine_move_task(rq, p, &rf, dest_cpu, flags);
2762 task_rq_unlock(rq, p, &rf);
2767 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2769 return __set_cpus_allowed_ptr(p, new_mask, 0);
2771 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2773 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2775 #ifdef CONFIG_SCHED_DEBUG
2776 unsigned int state = READ_ONCE(p->__state);
2779 * We should never call set_task_cpu() on a blocked task,
2780 * ttwu() will sort out the placement.
2782 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
2785 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2786 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2787 * time relying on p->on_rq.
2789 WARN_ON_ONCE(state == TASK_RUNNING &&
2790 p->sched_class == &fair_sched_class &&
2791 (p->on_rq && !task_on_rq_migrating(p)));
2793 #ifdef CONFIG_LOCKDEP
2795 * The caller should hold either p->pi_lock or rq->lock, when changing
2796 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2798 * sched_move_task() holds both and thus holding either pins the cgroup,
2801 * Furthermore, all task_rq users should acquire both locks, see
2804 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2805 lockdep_is_held(__rq_lockp(task_rq(p)))));
2808 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2810 WARN_ON_ONCE(!cpu_online(new_cpu));
2812 WARN_ON_ONCE(is_migration_disabled(p));
2815 trace_sched_migrate_task(p, new_cpu);
2817 if (task_cpu(p) != new_cpu) {
2818 if (p->sched_class->migrate_task_rq)
2819 p->sched_class->migrate_task_rq(p, new_cpu);
2820 p->se.nr_migrations++;
2822 perf_event_task_migrate(p);
2825 __set_task_cpu(p, new_cpu);
2828 #ifdef CONFIG_NUMA_BALANCING
2829 static void __migrate_swap_task(struct task_struct *p, int cpu)
2831 if (task_on_rq_queued(p)) {
2832 struct rq *src_rq, *dst_rq;
2833 struct rq_flags srf, drf;
2835 src_rq = task_rq(p);
2836 dst_rq = cpu_rq(cpu);
2838 rq_pin_lock(src_rq, &srf);
2839 rq_pin_lock(dst_rq, &drf);
2841 deactivate_task(src_rq, p, 0);
2842 set_task_cpu(p, cpu);
2843 activate_task(dst_rq, p, 0);
2844 check_preempt_curr(dst_rq, p, 0);
2846 rq_unpin_lock(dst_rq, &drf);
2847 rq_unpin_lock(src_rq, &srf);
2851 * Task isn't running anymore; make it appear like we migrated
2852 * it before it went to sleep. This means on wakeup we make the
2853 * previous CPU our target instead of where it really is.
2859 struct migration_swap_arg {
2860 struct task_struct *src_task, *dst_task;
2861 int src_cpu, dst_cpu;
2864 static int migrate_swap_stop(void *data)
2866 struct migration_swap_arg *arg = data;
2867 struct rq *src_rq, *dst_rq;
2870 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2873 src_rq = cpu_rq(arg->src_cpu);
2874 dst_rq = cpu_rq(arg->dst_cpu);
2876 double_raw_lock(&arg->src_task->pi_lock,
2877 &arg->dst_task->pi_lock);
2878 double_rq_lock(src_rq, dst_rq);
2880 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2883 if (task_cpu(arg->src_task) != arg->src_cpu)
2886 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2889 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2892 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2893 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2898 double_rq_unlock(src_rq, dst_rq);
2899 raw_spin_unlock(&arg->dst_task->pi_lock);
2900 raw_spin_unlock(&arg->src_task->pi_lock);
2906 * Cross migrate two tasks
2908 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2909 int target_cpu, int curr_cpu)
2911 struct migration_swap_arg arg;
2914 arg = (struct migration_swap_arg){
2916 .src_cpu = curr_cpu,
2918 .dst_cpu = target_cpu,
2921 if (arg.src_cpu == arg.dst_cpu)
2925 * These three tests are all lockless; this is OK since all of them
2926 * will be re-checked with proper locks held further down the line.
2928 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2931 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2934 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2937 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2938 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2943 #endif /* CONFIG_NUMA_BALANCING */
2946 * wait_task_inactive - wait for a thread to unschedule.
2948 * If @match_state is nonzero, it's the @p->state value just checked and
2949 * not expected to change. If it changes, i.e. @p might have woken up,
2950 * then return zero. When we succeed in waiting for @p to be off its CPU,
2951 * we return a positive number (its total switch count). If a second call
2952 * a short while later returns the same number, the caller can be sure that
2953 * @p has remained unscheduled the whole time.
2955 * The caller must ensure that the task *will* unschedule sometime soon,
2956 * else this function might spin for a *long* time. This function can't
2957 * be called with interrupts off, or it may introduce deadlock with
2958 * smp_call_function() if an IPI is sent by the same process we are
2959 * waiting to become inactive.
2961 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2963 int running, queued;
2970 * We do the initial early heuristics without holding
2971 * any task-queue locks at all. We'll only try to get
2972 * the runqueue lock when things look like they will
2978 * If the task is actively running on another CPU
2979 * still, just relax and busy-wait without holding
2982 * NOTE! Since we don't hold any locks, it's not
2983 * even sure that "rq" stays as the right runqueue!
2984 * But we don't care, since "task_running()" will
2985 * return false if the runqueue has changed and p
2986 * is actually now running somewhere else!
2988 while (task_running(rq, p)) {
2989 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
2995 * Ok, time to look more closely! We need the rq
2996 * lock now, to be *sure*. If we're wrong, we'll
2997 * just go back and repeat.
2999 rq = task_rq_lock(p, &rf);
3000 trace_sched_wait_task(p);
3001 running = task_running(rq, p);
3002 queued = task_on_rq_queued(p);
3004 if (!match_state || READ_ONCE(p->__state) == match_state)
3005 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3006 task_rq_unlock(rq, p, &rf);
3009 * If it changed from the expected state, bail out now.
3011 if (unlikely(!ncsw))
3015 * Was it really running after all now that we
3016 * checked with the proper locks actually held?
3018 * Oops. Go back and try again..
3020 if (unlikely(running)) {
3026 * It's not enough that it's not actively running,
3027 * it must be off the runqueue _entirely_, and not
3030 * So if it was still runnable (but just not actively
3031 * running right now), it's preempted, and we should
3032 * yield - it could be a while.
3034 if (unlikely(queued)) {
3035 ktime_t to = NSEC_PER_SEC / HZ;
3037 set_current_state(TASK_UNINTERRUPTIBLE);
3038 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
3043 * Ahh, all good. It wasn't running, and it wasn't
3044 * runnable, which means that it will never become
3045 * running in the future either. We're all done!
3054 * kick_process - kick a running thread to enter/exit the kernel
3055 * @p: the to-be-kicked thread
3057 * Cause a process which is running on another CPU to enter
3058 * kernel-mode, without any delay. (to get signals handled.)
3060 * NOTE: this function doesn't have to take the runqueue lock,
3061 * because all it wants to ensure is that the remote task enters
3062 * the kernel. If the IPI races and the task has been migrated
3063 * to another CPU then no harm is done and the purpose has been
3066 void kick_process(struct task_struct *p)
3072 if ((cpu != smp_processor_id()) && task_curr(p))
3073 smp_send_reschedule(cpu);
3076 EXPORT_SYMBOL_GPL(kick_process);
3079 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3081 * A few notes on cpu_active vs cpu_online:
3083 * - cpu_active must be a subset of cpu_online
3085 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3086 * see __set_cpus_allowed_ptr(). At this point the newly online
3087 * CPU isn't yet part of the sched domains, and balancing will not
3090 * - on CPU-down we clear cpu_active() to mask the sched domains and
3091 * avoid the load balancer to place new tasks on the to be removed
3092 * CPU. Existing tasks will remain running there and will be taken
3095 * This means that fallback selection must not select !active CPUs.
3096 * And can assume that any active CPU must be online. Conversely
3097 * select_task_rq() below may allow selection of !active CPUs in order
3098 * to satisfy the above rules.
3100 static int select_fallback_rq(int cpu, struct task_struct *p)
3102 int nid = cpu_to_node(cpu);
3103 const struct cpumask *nodemask = NULL;
3104 enum { cpuset, possible, fail } state = cpuset;
3108 * If the node that the CPU is on has been offlined, cpu_to_node()
3109 * will return -1. There is no CPU on the node, and we should
3110 * select the CPU on the other node.
3113 nodemask = cpumask_of_node(nid);
3115 /* Look for allowed, online CPU in same node. */
3116 for_each_cpu(dest_cpu, nodemask) {
3117 if (!cpu_active(dest_cpu))
3119 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
3125 /* Any allowed, online CPU? */
3126 for_each_cpu(dest_cpu, p->cpus_ptr) {
3127 if (!is_cpu_allowed(p, dest_cpu))
3133 /* No more Mr. Nice Guy. */
3136 if (IS_ENABLED(CONFIG_CPUSETS)) {
3137 cpuset_cpus_allowed_fallback(p);
3144 * XXX When called from select_task_rq() we only
3145 * hold p->pi_lock and again violate locking order.
3147 * More yuck to audit.
3149 do_set_cpus_allowed(p, cpu_possible_mask);
3160 if (state != cpuset) {
3162 * Don't tell them about moving exiting tasks or
3163 * kernel threads (both mm NULL), since they never
3166 if (p->mm && printk_ratelimit()) {
3167 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3168 task_pid_nr(p), p->comm, cpu);
3176 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3179 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3181 lockdep_assert_held(&p->pi_lock);
3183 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3184 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3186 cpu = cpumask_any(p->cpus_ptr);
3189 * In order not to call set_task_cpu() on a blocking task we need
3190 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3193 * Since this is common to all placement strategies, this lives here.
3195 * [ this allows ->select_task() to simply return task_cpu(p) and
3196 * not worry about this generic constraint ]
3198 if (unlikely(!is_cpu_allowed(p, cpu)))
3199 cpu = select_fallback_rq(task_cpu(p), p);
3204 void sched_set_stop_task(int cpu, struct task_struct *stop)
3206 static struct lock_class_key stop_pi_lock;
3207 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3208 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3212 * Make it appear like a SCHED_FIFO task, its something
3213 * userspace knows about and won't get confused about.
3215 * Also, it will make PI more or less work without too
3216 * much confusion -- but then, stop work should not
3217 * rely on PI working anyway.
3219 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3221 stop->sched_class = &stop_sched_class;
3224 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3225 * adjust the effective priority of a task. As a result,
3226 * rt_mutex_setprio() can trigger (RT) balancing operations,
3227 * which can then trigger wakeups of the stop thread to push
3228 * around the current task.
3230 * The stop task itself will never be part of the PI-chain, it
3231 * never blocks, therefore that ->pi_lock recursion is safe.
3232 * Tell lockdep about this by placing the stop->pi_lock in its
3235 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3238 cpu_rq(cpu)->stop = stop;
3242 * Reset it back to a normal scheduling class so that
3243 * it can die in pieces.
3245 old_stop->sched_class = &rt_sched_class;
3249 #else /* CONFIG_SMP */
3251 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3252 const struct cpumask *new_mask,
3255 return set_cpus_allowed_ptr(p, new_mask);
3258 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3260 static inline bool rq_has_pinned_tasks(struct rq *rq)
3265 #endif /* !CONFIG_SMP */
3268 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3272 if (!schedstat_enabled())
3278 if (cpu == rq->cpu) {
3279 __schedstat_inc(rq->ttwu_local);
3280 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3282 struct sched_domain *sd;
3284 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3286 for_each_domain(rq->cpu, sd) {
3287 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3288 __schedstat_inc(sd->ttwu_wake_remote);
3295 if (wake_flags & WF_MIGRATED)
3296 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3297 #endif /* CONFIG_SMP */
3299 __schedstat_inc(rq->ttwu_count);
3300 __schedstat_inc(p->se.statistics.nr_wakeups);
3302 if (wake_flags & WF_SYNC)
3303 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3307 * Mark the task runnable and perform wakeup-preemption.
3309 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3310 struct rq_flags *rf)
3312 check_preempt_curr(rq, p, wake_flags);
3313 WRITE_ONCE(p->__state, TASK_RUNNING);
3314 trace_sched_wakeup(p);
3317 if (p->sched_class->task_woken) {
3319 * Our task @p is fully woken up and running; so it's safe to
3320 * drop the rq->lock, hereafter rq is only used for statistics.
3322 rq_unpin_lock(rq, rf);
3323 p->sched_class->task_woken(rq, p);
3324 rq_repin_lock(rq, rf);
3327 if (rq->idle_stamp) {
3328 u64 delta = rq_clock(rq) - rq->idle_stamp;
3329 u64 max = 2*rq->max_idle_balance_cost;
3331 update_avg(&rq->avg_idle, delta);
3333 if (rq->avg_idle > max)
3336 rq->wake_stamp = jiffies;
3337 rq->wake_avg_idle = rq->avg_idle / 2;
3345 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3346 struct rq_flags *rf)
3348 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3350 lockdep_assert_rq_held(rq);
3352 if (p->sched_contributes_to_load)
3353 rq->nr_uninterruptible--;
3356 if (wake_flags & WF_MIGRATED)
3357 en_flags |= ENQUEUE_MIGRATED;
3361 delayacct_blkio_end(p);
3362 atomic_dec(&task_rq(p)->nr_iowait);
3365 activate_task(rq, p, en_flags);
3366 ttwu_do_wakeup(rq, p, wake_flags, rf);
3370 * Consider @p being inside a wait loop:
3373 * set_current_state(TASK_UNINTERRUPTIBLE);
3380 * __set_current_state(TASK_RUNNING);
3382 * between set_current_state() and schedule(). In this case @p is still
3383 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3386 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3387 * then schedule() must still happen and p->state can be changed to
3388 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3389 * need to do a full wakeup with enqueue.
3391 * Returns: %true when the wakeup is done,
3394 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3400 rq = __task_rq_lock(p, &rf);
3401 if (task_on_rq_queued(p)) {
3402 /* check_preempt_curr() may use rq clock */
3403 update_rq_clock(rq);
3404 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3407 __task_rq_unlock(rq, &rf);
3413 void sched_ttwu_pending(void *arg)
3415 struct llist_node *llist = arg;
3416 struct rq *rq = this_rq();
3417 struct task_struct *p, *t;
3424 * rq::ttwu_pending racy indication of out-standing wakeups.
3425 * Races such that false-negatives are possible, since they
3426 * are shorter lived that false-positives would be.
3428 WRITE_ONCE(rq->ttwu_pending, 0);
3430 rq_lock_irqsave(rq, &rf);
3431 update_rq_clock(rq);
3433 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3434 if (WARN_ON_ONCE(p->on_cpu))
3435 smp_cond_load_acquire(&p->on_cpu, !VAL);
3437 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3438 set_task_cpu(p, cpu_of(rq));
3440 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3443 rq_unlock_irqrestore(rq, &rf);
3446 void send_call_function_single_ipi(int cpu)
3448 struct rq *rq = cpu_rq(cpu);
3450 if (!set_nr_if_polling(rq->idle))
3451 arch_send_call_function_single_ipi(cpu);
3453 trace_sched_wake_idle_without_ipi(cpu);
3457 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3458 * necessary. The wakee CPU on receipt of the IPI will queue the task
3459 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3460 * of the wakeup instead of the waker.
3462 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3464 struct rq *rq = cpu_rq(cpu);
3466 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3468 WRITE_ONCE(rq->ttwu_pending, 1);
3469 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3472 void wake_up_if_idle(int cpu)
3474 struct rq *rq = cpu_rq(cpu);
3479 if (!is_idle_task(rcu_dereference(rq->curr)))
3482 if (set_nr_if_polling(rq->idle)) {
3483 trace_sched_wake_idle_without_ipi(cpu);
3485 rq_lock_irqsave(rq, &rf);
3486 if (is_idle_task(rq->curr))
3487 smp_send_reschedule(cpu);
3488 /* Else CPU is not idle, do nothing here: */
3489 rq_unlock_irqrestore(rq, &rf);
3496 bool cpus_share_cache(int this_cpu, int that_cpu)
3498 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3501 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3504 * Do not complicate things with the async wake_list while the CPU is
3507 if (!cpu_active(cpu))
3511 * If the CPU does not share cache, then queue the task on the
3512 * remote rqs wakelist to avoid accessing remote data.
3514 if (!cpus_share_cache(smp_processor_id(), cpu))
3518 * If the task is descheduling and the only running task on the
3519 * CPU then use the wakelist to offload the task activation to
3520 * the soon-to-be-idle CPU as the current CPU is likely busy.
3521 * nr_running is checked to avoid unnecessary task stacking.
3523 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3529 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3531 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3532 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3535 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3536 __ttwu_queue_wakelist(p, cpu, wake_flags);
3543 #else /* !CONFIG_SMP */
3545 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3550 #endif /* CONFIG_SMP */
3552 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3554 struct rq *rq = cpu_rq(cpu);
3557 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3561 update_rq_clock(rq);
3562 ttwu_do_activate(rq, p, wake_flags, &rf);
3567 * Notes on Program-Order guarantees on SMP systems.
3571 * The basic program-order guarantee on SMP systems is that when a task [t]
3572 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3573 * execution on its new CPU [c1].
3575 * For migration (of runnable tasks) this is provided by the following means:
3577 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3578 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3579 * rq(c1)->lock (if not at the same time, then in that order).
3580 * C) LOCK of the rq(c1)->lock scheduling in task
3582 * Release/acquire chaining guarantees that B happens after A and C after B.
3583 * Note: the CPU doing B need not be c0 or c1
3592 * UNLOCK rq(0)->lock
3594 * LOCK rq(0)->lock // orders against CPU0
3596 * UNLOCK rq(0)->lock
3600 * UNLOCK rq(1)->lock
3602 * LOCK rq(1)->lock // orders against CPU2
3605 * UNLOCK rq(1)->lock
3608 * BLOCKING -- aka. SLEEP + WAKEUP
3610 * For blocking we (obviously) need to provide the same guarantee as for
3611 * migration. However the means are completely different as there is no lock
3612 * chain to provide order. Instead we do:
3614 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3615 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3619 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3621 * LOCK rq(0)->lock LOCK X->pi_lock
3624 * smp_store_release(X->on_cpu, 0);
3626 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3632 * X->state = RUNNING
3633 * UNLOCK rq(2)->lock
3635 * LOCK rq(2)->lock // orders against CPU1
3638 * UNLOCK rq(2)->lock
3641 * UNLOCK rq(0)->lock
3644 * However, for wakeups there is a second guarantee we must provide, namely we
3645 * must ensure that CONDITION=1 done by the caller can not be reordered with
3646 * accesses to the task state; see try_to_wake_up() and set_current_state().
3650 * try_to_wake_up - wake up a thread
3651 * @p: the thread to be awakened
3652 * @state: the mask of task states that can be woken
3653 * @wake_flags: wake modifier flags (WF_*)
3655 * Conceptually does:
3657 * If (@state & @p->state) @p->state = TASK_RUNNING.
3659 * If the task was not queued/runnable, also place it back on a runqueue.
3661 * This function is atomic against schedule() which would dequeue the task.
3663 * It issues a full memory barrier before accessing @p->state, see the comment
3664 * with set_current_state().
3666 * Uses p->pi_lock to serialize against concurrent wake-ups.
3668 * Relies on p->pi_lock stabilizing:
3671 * - p->sched_task_group
3672 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3674 * Tries really hard to only take one task_rq(p)->lock for performance.
3675 * Takes rq->lock in:
3676 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3677 * - ttwu_queue() -- new rq, for enqueue of the task;
3678 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3680 * As a consequence we race really badly with just about everything. See the
3681 * many memory barriers and their comments for details.
3683 * Return: %true if @p->state changes (an actual wakeup was done),
3687 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3689 unsigned long flags;
3690 int cpu, success = 0;
3695 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3696 * == smp_processor_id()'. Together this means we can special
3697 * case the whole 'p->on_rq && ttwu_runnable()' case below
3698 * without taking any locks.
3701 * - we rely on Program-Order guarantees for all the ordering,
3702 * - we're serialized against set_special_state() by virtue of
3703 * it disabling IRQs (this allows not taking ->pi_lock).
3705 if (!(READ_ONCE(p->__state) & state))
3709 trace_sched_waking(p);
3710 WRITE_ONCE(p->__state, TASK_RUNNING);
3711 trace_sched_wakeup(p);
3716 * If we are going to wake up a thread waiting for CONDITION we
3717 * need to ensure that CONDITION=1 done by the caller can not be
3718 * reordered with p->state check below. This pairs with smp_store_mb()
3719 * in set_current_state() that the waiting thread does.
3721 raw_spin_lock_irqsave(&p->pi_lock, flags);
3722 smp_mb__after_spinlock();
3723 if (!(READ_ONCE(p->__state) & state))
3726 trace_sched_waking(p);
3728 /* We're going to change ->state: */
3732 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3733 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3734 * in smp_cond_load_acquire() below.
3736 * sched_ttwu_pending() try_to_wake_up()
3737 * STORE p->on_rq = 1 LOAD p->state
3740 * __schedule() (switch to task 'p')
3741 * LOCK rq->lock smp_rmb();
3742 * smp_mb__after_spinlock();
3746 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3748 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3749 * __schedule(). See the comment for smp_mb__after_spinlock().
3751 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3754 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3759 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3760 * possible to, falsely, observe p->on_cpu == 0.
3762 * One must be running (->on_cpu == 1) in order to remove oneself
3763 * from the runqueue.
3765 * __schedule() (switch to task 'p') try_to_wake_up()
3766 * STORE p->on_cpu = 1 LOAD p->on_rq
3769 * __schedule() (put 'p' to sleep)
3770 * LOCK rq->lock smp_rmb();
3771 * smp_mb__after_spinlock();
3772 * STORE p->on_rq = 0 LOAD p->on_cpu
3774 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3775 * __schedule(). See the comment for smp_mb__after_spinlock().
3777 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3778 * schedule()'s deactivate_task() has 'happened' and p will no longer
3779 * care about it's own p->state. See the comment in __schedule().
3781 smp_acquire__after_ctrl_dep();
3784 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3785 * == 0), which means we need to do an enqueue, change p->state to
3786 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3787 * enqueue, such as ttwu_queue_wakelist().
3789 WRITE_ONCE(p->__state, TASK_WAKING);
3792 * If the owning (remote) CPU is still in the middle of schedule() with
3793 * this task as prev, considering queueing p on the remote CPUs wake_list
3794 * which potentially sends an IPI instead of spinning on p->on_cpu to
3795 * let the waker make forward progress. This is safe because IRQs are
3796 * disabled and the IPI will deliver after on_cpu is cleared.
3798 * Ensure we load task_cpu(p) after p->on_cpu:
3800 * set_task_cpu(p, cpu);
3801 * STORE p->cpu = @cpu
3802 * __schedule() (switch to task 'p')
3804 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3805 * STORE p->on_cpu = 1 LOAD p->cpu
3807 * to ensure we observe the correct CPU on which the task is currently
3810 if (smp_load_acquire(&p->on_cpu) &&
3811 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3815 * If the owning (remote) CPU is still in the middle of schedule() with
3816 * this task as prev, wait until it's done referencing the task.
3818 * Pairs with the smp_store_release() in finish_task().
3820 * This ensures that tasks getting woken will be fully ordered against
3821 * their previous state and preserve Program Order.
3823 smp_cond_load_acquire(&p->on_cpu, !VAL);
3825 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3826 if (task_cpu(p) != cpu) {
3828 delayacct_blkio_end(p);
3829 atomic_dec(&task_rq(p)->nr_iowait);
3832 wake_flags |= WF_MIGRATED;
3833 psi_ttwu_dequeue(p);
3834 set_task_cpu(p, cpu);
3838 #endif /* CONFIG_SMP */
3840 ttwu_queue(p, cpu, wake_flags);
3842 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3845 ttwu_stat(p, task_cpu(p), wake_flags);
3852 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3853 * @p: Process for which the function is to be invoked, can be @current.
3854 * @func: Function to invoke.
3855 * @arg: Argument to function.
3857 * If the specified task can be quickly locked into a definite state
3858 * (either sleeping or on a given runqueue), arrange to keep it in that
3859 * state while invoking @func(@arg). This function can use ->on_rq and
3860 * task_curr() to work out what the state is, if required. Given that
3861 * @func can be invoked with a runqueue lock held, it had better be quite
3865 * @false if the task slipped out from under the locks.
3866 * @true if the task was locked onto a runqueue or is sleeping.
3867 * However, @func can override this by returning @false.
3869 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3875 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3877 rq = __task_rq_lock(p, &rf);
3878 if (task_rq(p) == rq)
3882 switch (READ_ONCE(p->__state)) {
3887 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3892 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
3897 * wake_up_process - Wake up a specific process
3898 * @p: The process to be woken up.
3900 * Attempt to wake up the nominated process and move it to the set of runnable
3903 * Return: 1 if the process was woken up, 0 if it was already running.
3905 * This function executes a full memory barrier before accessing the task state.
3907 int wake_up_process(struct task_struct *p)
3909 return try_to_wake_up(p, TASK_NORMAL, 0);
3911 EXPORT_SYMBOL(wake_up_process);
3913 int wake_up_state(struct task_struct *p, unsigned int state)
3915 return try_to_wake_up(p, state, 0);
3919 * Perform scheduler related setup for a newly forked process p.
3920 * p is forked by current.
3922 * __sched_fork() is basic setup used by init_idle() too:
3924 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3929 p->se.exec_start = 0;
3930 p->se.sum_exec_runtime = 0;
3931 p->se.prev_sum_exec_runtime = 0;
3932 p->se.nr_migrations = 0;
3934 INIT_LIST_HEAD(&p->se.group_node);
3936 #ifdef CONFIG_FAIR_GROUP_SCHED
3937 p->se.cfs_rq = NULL;
3940 #ifdef CONFIG_SCHEDSTATS
3941 /* Even if schedstat is disabled, there should not be garbage */
3942 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3945 RB_CLEAR_NODE(&p->dl.rb_node);
3946 init_dl_task_timer(&p->dl);
3947 init_dl_inactive_task_timer(&p->dl);
3948 __dl_clear_params(p);
3950 INIT_LIST_HEAD(&p->rt.run_list);
3952 p->rt.time_slice = sched_rr_timeslice;
3956 #ifdef CONFIG_PREEMPT_NOTIFIERS
3957 INIT_HLIST_HEAD(&p->preempt_notifiers);
3960 #ifdef CONFIG_COMPACTION
3961 p->capture_control = NULL;
3963 init_numa_balancing(clone_flags, p);
3965 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3966 p->migration_pending = NULL;
3970 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3972 #ifdef CONFIG_NUMA_BALANCING
3974 void set_numabalancing_state(bool enabled)
3977 static_branch_enable(&sched_numa_balancing);
3979 static_branch_disable(&sched_numa_balancing);
3982 #ifdef CONFIG_PROC_SYSCTL
3983 int sysctl_numa_balancing(struct ctl_table *table, int write,
3984 void *buffer, size_t *lenp, loff_t *ppos)
3988 int state = static_branch_likely(&sched_numa_balancing);
3990 if (write && !capable(CAP_SYS_ADMIN))
3995 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3999 set_numabalancing_state(state);
4005 #ifdef CONFIG_SCHEDSTATS
4007 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4009 static void set_schedstats(bool enabled)
4012 static_branch_enable(&sched_schedstats);
4014 static_branch_disable(&sched_schedstats);
4017 void force_schedstat_enabled(void)
4019 if (!schedstat_enabled()) {
4020 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4021 static_branch_enable(&sched_schedstats);
4025 static int __init setup_schedstats(char *str)
4031 if (!strcmp(str, "enable")) {
4032 set_schedstats(true);
4034 } else if (!strcmp(str, "disable")) {
4035 set_schedstats(false);
4040 pr_warn("Unable to parse schedstats=\n");
4044 __setup("schedstats=", setup_schedstats);
4046 #ifdef CONFIG_PROC_SYSCTL
4047 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4048 size_t *lenp, loff_t *ppos)
4052 int state = static_branch_likely(&sched_schedstats);
4054 if (write && !capable(CAP_SYS_ADMIN))
4059 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4063 set_schedstats(state);
4066 #endif /* CONFIG_PROC_SYSCTL */
4067 #endif /* CONFIG_SCHEDSTATS */
4070 * fork()/clone()-time setup:
4072 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4074 unsigned long flags;
4076 __sched_fork(clone_flags, p);
4078 * We mark the process as NEW here. This guarantees that
4079 * nobody will actually run it, and a signal or other external
4080 * event cannot wake it up and insert it on the runqueue either.
4082 p->__state = TASK_NEW;
4085 * Make sure we do not leak PI boosting priority to the child.
4087 p->prio = current->normal_prio;
4092 * Revert to default priority/policy on fork if requested.
4094 if (unlikely(p->sched_reset_on_fork)) {
4095 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4096 p->policy = SCHED_NORMAL;
4097 p->static_prio = NICE_TO_PRIO(0);
4099 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4100 p->static_prio = NICE_TO_PRIO(0);
4102 p->prio = p->normal_prio = __normal_prio(p);
4103 set_load_weight(p, false);
4106 * We don't need the reset flag anymore after the fork. It has
4107 * fulfilled its duty:
4109 p->sched_reset_on_fork = 0;
4112 if (dl_prio(p->prio))
4114 else if (rt_prio(p->prio))
4115 p->sched_class = &rt_sched_class;
4117 p->sched_class = &fair_sched_class;
4119 init_entity_runnable_average(&p->se);
4122 * The child is not yet in the pid-hash so no cgroup attach races,
4123 * and the cgroup is pinned to this child due to cgroup_fork()
4124 * is ran before sched_fork().
4126 * Silence PROVE_RCU.
4128 raw_spin_lock_irqsave(&p->pi_lock, flags);
4131 * We're setting the CPU for the first time, we don't migrate,
4132 * so use __set_task_cpu().
4134 __set_task_cpu(p, smp_processor_id());
4135 if (p->sched_class->task_fork)
4136 p->sched_class->task_fork(p);
4137 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4139 #ifdef CONFIG_SCHED_INFO
4140 if (likely(sched_info_on()))
4141 memset(&p->sched_info, 0, sizeof(p->sched_info));
4143 #if defined(CONFIG_SMP)
4146 init_task_preempt_count(p);
4148 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4149 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4154 void sched_post_fork(struct task_struct *p)
4156 uclamp_post_fork(p);
4159 unsigned long to_ratio(u64 period, u64 runtime)
4161 if (runtime == RUNTIME_INF)
4165 * Doing this here saves a lot of checks in all
4166 * the calling paths, and returning zero seems
4167 * safe for them anyway.
4172 return div64_u64(runtime << BW_SHIFT, period);
4176 * wake_up_new_task - wake up a newly created task for the first time.
4178 * This function will do some initial scheduler statistics housekeeping
4179 * that must be done for every newly created context, then puts the task
4180 * on the runqueue and wakes it.
4182 void wake_up_new_task(struct task_struct *p)
4187 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4188 WRITE_ONCE(p->__state, TASK_RUNNING);
4191 * Fork balancing, do it here and not earlier because:
4192 * - cpus_ptr can change in the fork path
4193 * - any previously selected CPU might disappear through hotplug
4195 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4196 * as we're not fully set-up yet.
4198 p->recent_used_cpu = task_cpu(p);
4200 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4202 rq = __task_rq_lock(p, &rf);
4203 update_rq_clock(rq);
4204 post_init_entity_util_avg(p);
4206 activate_task(rq, p, ENQUEUE_NOCLOCK);
4207 trace_sched_wakeup_new(p);
4208 check_preempt_curr(rq, p, WF_FORK);
4210 if (p->sched_class->task_woken) {
4212 * Nothing relies on rq->lock after this, so it's fine to
4215 rq_unpin_lock(rq, &rf);
4216 p->sched_class->task_woken(rq, p);
4217 rq_repin_lock(rq, &rf);
4220 task_rq_unlock(rq, p, &rf);
4223 #ifdef CONFIG_PREEMPT_NOTIFIERS
4225 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4227 void preempt_notifier_inc(void)
4229 static_branch_inc(&preempt_notifier_key);
4231 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4233 void preempt_notifier_dec(void)
4235 static_branch_dec(&preempt_notifier_key);
4237 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4240 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4241 * @notifier: notifier struct to register
4243 void preempt_notifier_register(struct preempt_notifier *notifier)
4245 if (!static_branch_unlikely(&preempt_notifier_key))
4246 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4248 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4250 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4253 * preempt_notifier_unregister - no longer interested in preemption notifications
4254 * @notifier: notifier struct to unregister
4256 * This is *not* safe to call from within a preemption notifier.
4258 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4260 hlist_del(¬ifier->link);
4262 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4264 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4266 struct preempt_notifier *notifier;
4268 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4269 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4272 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4274 if (static_branch_unlikely(&preempt_notifier_key))
4275 __fire_sched_in_preempt_notifiers(curr);
4279 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4280 struct task_struct *next)
4282 struct preempt_notifier *notifier;
4284 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4285 notifier->ops->sched_out(notifier, next);
4288 static __always_inline void
4289 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4290 struct task_struct *next)
4292 if (static_branch_unlikely(&preempt_notifier_key))
4293 __fire_sched_out_preempt_notifiers(curr, next);
4296 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4298 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4303 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4304 struct task_struct *next)
4308 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4310 static inline void prepare_task(struct task_struct *next)
4314 * Claim the task as running, we do this before switching to it
4315 * such that any running task will have this set.
4317 * See the ttwu() WF_ON_CPU case and its ordering comment.
4319 WRITE_ONCE(next->on_cpu, 1);
4323 static inline void finish_task(struct task_struct *prev)
4327 * This must be the very last reference to @prev from this CPU. After
4328 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4329 * must ensure this doesn't happen until the switch is completely
4332 * In particular, the load of prev->state in finish_task_switch() must
4333 * happen before this.
4335 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4337 smp_store_release(&prev->on_cpu, 0);
4343 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4345 void (*func)(struct rq *rq);
4346 struct callback_head *next;
4348 lockdep_assert_rq_held(rq);
4351 func = (void (*)(struct rq *))head->func;
4360 static void balance_push(struct rq *rq);
4362 struct callback_head balance_push_callback = {
4364 .func = (void (*)(struct callback_head *))balance_push,
4367 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4369 struct callback_head *head = rq->balance_callback;
4371 lockdep_assert_rq_held(rq);
4373 rq->balance_callback = NULL;
4378 static void __balance_callbacks(struct rq *rq)
4380 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4383 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4385 unsigned long flags;
4387 if (unlikely(head)) {
4388 raw_spin_rq_lock_irqsave(rq, flags);
4389 do_balance_callbacks(rq, head);
4390 raw_spin_rq_unlock_irqrestore(rq, flags);
4396 static inline void __balance_callbacks(struct rq *rq)
4400 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4405 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4412 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4415 * Since the runqueue lock will be released by the next
4416 * task (which is an invalid locking op but in the case
4417 * of the scheduler it's an obvious special-case), so we
4418 * do an early lockdep release here:
4420 rq_unpin_lock(rq, rf);
4421 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4422 #ifdef CONFIG_DEBUG_SPINLOCK
4423 /* this is a valid case when another task releases the spinlock */
4424 rq_lockp(rq)->owner = next;
4428 static inline void finish_lock_switch(struct rq *rq)
4431 * If we are tracking spinlock dependencies then we have to
4432 * fix up the runqueue lock - which gets 'carried over' from
4433 * prev into current:
4435 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4436 __balance_callbacks(rq);
4437 raw_spin_rq_unlock_irq(rq);
4441 * NOP if the arch has not defined these:
4444 #ifndef prepare_arch_switch
4445 # define prepare_arch_switch(next) do { } while (0)
4448 #ifndef finish_arch_post_lock_switch
4449 # define finish_arch_post_lock_switch() do { } while (0)
4452 static inline void kmap_local_sched_out(void)
4454 #ifdef CONFIG_KMAP_LOCAL
4455 if (unlikely(current->kmap_ctrl.idx))
4456 __kmap_local_sched_out();
4460 static inline void kmap_local_sched_in(void)
4462 #ifdef CONFIG_KMAP_LOCAL
4463 if (unlikely(current->kmap_ctrl.idx))
4464 __kmap_local_sched_in();
4469 * prepare_task_switch - prepare to switch tasks
4470 * @rq: the runqueue preparing to switch
4471 * @prev: the current task that is being switched out
4472 * @next: the task we are going to switch to.
4474 * This is called with the rq lock held and interrupts off. It must
4475 * be paired with a subsequent finish_task_switch after the context
4478 * prepare_task_switch sets up locking and calls architecture specific
4482 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4483 struct task_struct *next)
4485 kcov_prepare_switch(prev);
4486 sched_info_switch(rq, prev, next);
4487 perf_event_task_sched_out(prev, next);
4489 fire_sched_out_preempt_notifiers(prev, next);
4490 kmap_local_sched_out();
4492 prepare_arch_switch(next);
4496 * finish_task_switch - clean up after a task-switch
4497 * @prev: the thread we just switched away from.
4499 * finish_task_switch must be called after the context switch, paired
4500 * with a prepare_task_switch call before the context switch.
4501 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4502 * and do any other architecture-specific cleanup actions.
4504 * Note that we may have delayed dropping an mm in context_switch(). If
4505 * so, we finish that here outside of the runqueue lock. (Doing it
4506 * with the lock held can cause deadlocks; see schedule() for
4509 * The context switch have flipped the stack from under us and restored the
4510 * local variables which were saved when this task called schedule() in the
4511 * past. prev == current is still correct but we need to recalculate this_rq
4512 * because prev may have moved to another CPU.
4514 static struct rq *finish_task_switch(struct task_struct *prev)
4515 __releases(rq->lock)
4517 struct rq *rq = this_rq();
4518 struct mm_struct *mm = rq->prev_mm;
4522 * The previous task will have left us with a preempt_count of 2
4523 * because it left us after:
4526 * preempt_disable(); // 1
4528 * raw_spin_lock_irq(&rq->lock) // 2
4530 * Also, see FORK_PREEMPT_COUNT.
4532 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4533 "corrupted preempt_count: %s/%d/0x%x\n",
4534 current->comm, current->pid, preempt_count()))
4535 preempt_count_set(FORK_PREEMPT_COUNT);
4540 * A task struct has one reference for the use as "current".
4541 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4542 * schedule one last time. The schedule call will never return, and
4543 * the scheduled task must drop that reference.
4545 * We must observe prev->state before clearing prev->on_cpu (in
4546 * finish_task), otherwise a concurrent wakeup can get prev
4547 * running on another CPU and we could rave with its RUNNING -> DEAD
4548 * transition, resulting in a double drop.
4550 prev_state = READ_ONCE(prev->__state);
4551 vtime_task_switch(prev);
4552 perf_event_task_sched_in(prev, current);
4554 tick_nohz_task_switch();
4555 finish_lock_switch(rq);
4556 finish_arch_post_lock_switch();
4557 kcov_finish_switch(current);
4559 * kmap_local_sched_out() is invoked with rq::lock held and
4560 * interrupts disabled. There is no requirement for that, but the
4561 * sched out code does not have an interrupt enabled section.
4562 * Restoring the maps on sched in does not require interrupts being
4565 kmap_local_sched_in();
4567 fire_sched_in_preempt_notifiers(current);
4569 * When switching through a kernel thread, the loop in
4570 * membarrier_{private,global}_expedited() may have observed that
4571 * kernel thread and not issued an IPI. It is therefore possible to
4572 * schedule between user->kernel->user threads without passing though
4573 * switch_mm(). Membarrier requires a barrier after storing to
4574 * rq->curr, before returning to userspace, so provide them here:
4576 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4577 * provided by mmdrop(),
4578 * - a sync_core for SYNC_CORE.
4581 membarrier_mm_sync_core_before_usermode(mm);
4584 if (unlikely(prev_state == TASK_DEAD)) {
4585 if (prev->sched_class->task_dead)
4586 prev->sched_class->task_dead(prev);
4589 * Remove function-return probe instances associated with this
4590 * task and put them back on the free list.
4592 kprobe_flush_task(prev);
4594 /* Task is done with its stack. */
4595 put_task_stack(prev);
4597 put_task_struct_rcu_user(prev);
4604 * schedule_tail - first thing a freshly forked thread must call.
4605 * @prev: the thread we just switched away from.
4607 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4608 __releases(rq->lock)
4611 * New tasks start with FORK_PREEMPT_COUNT, see there and
4612 * finish_task_switch() for details.
4614 * finish_task_switch() will drop rq->lock() and lower preempt_count
4615 * and the preempt_enable() will end up enabling preemption (on
4616 * PREEMPT_COUNT kernels).
4619 finish_task_switch(prev);
4622 if (current->set_child_tid)
4623 put_user(task_pid_vnr(current), current->set_child_tid);
4625 calculate_sigpending();
4629 * context_switch - switch to the new MM and the new thread's register state.
4631 static __always_inline struct rq *
4632 context_switch(struct rq *rq, struct task_struct *prev,
4633 struct task_struct *next, struct rq_flags *rf)
4635 prepare_task_switch(rq, prev, next);
4638 * For paravirt, this is coupled with an exit in switch_to to
4639 * combine the page table reload and the switch backend into
4642 arch_start_context_switch(prev);
4645 * kernel -> kernel lazy + transfer active
4646 * user -> kernel lazy + mmgrab() active
4648 * kernel -> user switch + mmdrop() active
4649 * user -> user switch
4651 if (!next->mm) { // to kernel
4652 enter_lazy_tlb(prev->active_mm, next);
4654 next->active_mm = prev->active_mm;
4655 if (prev->mm) // from user
4656 mmgrab(prev->active_mm);
4658 prev->active_mm = NULL;
4660 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4662 * sys_membarrier() requires an smp_mb() between setting
4663 * rq->curr / membarrier_switch_mm() and returning to userspace.
4665 * The below provides this either through switch_mm(), or in
4666 * case 'prev->active_mm == next->mm' through
4667 * finish_task_switch()'s mmdrop().
4669 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4671 if (!prev->mm) { // from kernel
4672 /* will mmdrop() in finish_task_switch(). */
4673 rq->prev_mm = prev->active_mm;
4674 prev->active_mm = NULL;
4678 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4680 prepare_lock_switch(rq, next, rf);
4682 /* Here we just switch the register state and the stack. */
4683 switch_to(prev, next, prev);
4686 return finish_task_switch(prev);
4690 * nr_running and nr_context_switches:
4692 * externally visible scheduler statistics: current number of runnable
4693 * threads, total number of context switches performed since bootup.
4695 unsigned int nr_running(void)
4697 unsigned int i, sum = 0;
4699 for_each_online_cpu(i)
4700 sum += cpu_rq(i)->nr_running;
4706 * Check if only the current task is running on the CPU.
4708 * Caution: this function does not check that the caller has disabled
4709 * preemption, thus the result might have a time-of-check-to-time-of-use
4710 * race. The caller is responsible to use it correctly, for example:
4712 * - from a non-preemptible section (of course)
4714 * - from a thread that is bound to a single CPU
4716 * - in a loop with very short iterations (e.g. a polling loop)
4718 bool single_task_running(void)
4720 return raw_rq()->nr_running == 1;
4722 EXPORT_SYMBOL(single_task_running);
4724 unsigned long long nr_context_switches(void)
4727 unsigned long long sum = 0;
4729 for_each_possible_cpu(i)
4730 sum += cpu_rq(i)->nr_switches;
4736 * Consumers of these two interfaces, like for example the cpuidle menu
4737 * governor, are using nonsensical data. Preferring shallow idle state selection
4738 * for a CPU that has IO-wait which might not even end up running the task when
4739 * it does become runnable.
4742 unsigned int nr_iowait_cpu(int cpu)
4744 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4748 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4750 * The idea behind IO-wait account is to account the idle time that we could
4751 * have spend running if it were not for IO. That is, if we were to improve the
4752 * storage performance, we'd have a proportional reduction in IO-wait time.
4754 * This all works nicely on UP, where, when a task blocks on IO, we account
4755 * idle time as IO-wait, because if the storage were faster, it could've been
4756 * running and we'd not be idle.
4758 * This has been extended to SMP, by doing the same for each CPU. This however
4761 * Imagine for instance the case where two tasks block on one CPU, only the one
4762 * CPU will have IO-wait accounted, while the other has regular idle. Even
4763 * though, if the storage were faster, both could've ran at the same time,
4764 * utilising both CPUs.
4766 * This means, that when looking globally, the current IO-wait accounting on
4767 * SMP is a lower bound, by reason of under accounting.
4769 * Worse, since the numbers are provided per CPU, they are sometimes
4770 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4771 * associated with any one particular CPU, it can wake to another CPU than it
4772 * blocked on. This means the per CPU IO-wait number is meaningless.
4774 * Task CPU affinities can make all that even more 'interesting'.
4777 unsigned int nr_iowait(void)
4779 unsigned int i, sum = 0;
4781 for_each_possible_cpu(i)
4782 sum += nr_iowait_cpu(i);
4790 * sched_exec - execve() is a valuable balancing opportunity, because at
4791 * this point the task has the smallest effective memory and cache footprint.
4793 void sched_exec(void)
4795 struct task_struct *p = current;
4796 unsigned long flags;
4799 raw_spin_lock_irqsave(&p->pi_lock, flags);
4800 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4801 if (dest_cpu == smp_processor_id())
4804 if (likely(cpu_active(dest_cpu))) {
4805 struct migration_arg arg = { p, dest_cpu };
4807 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4808 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4812 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4817 DEFINE_PER_CPU(struct kernel_stat, kstat);
4818 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4820 EXPORT_PER_CPU_SYMBOL(kstat);
4821 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4824 * The function fair_sched_class.update_curr accesses the struct curr
4825 * and its field curr->exec_start; when called from task_sched_runtime(),
4826 * we observe a high rate of cache misses in practice.
4827 * Prefetching this data results in improved performance.
4829 static inline void prefetch_curr_exec_start(struct task_struct *p)
4831 #ifdef CONFIG_FAIR_GROUP_SCHED
4832 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4834 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4837 prefetch(&curr->exec_start);
4841 * Return accounted runtime for the task.
4842 * In case the task is currently running, return the runtime plus current's
4843 * pending runtime that have not been accounted yet.
4845 unsigned long long task_sched_runtime(struct task_struct *p)
4851 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4853 * 64-bit doesn't need locks to atomically read a 64-bit value.
4854 * So we have a optimization chance when the task's delta_exec is 0.
4855 * Reading ->on_cpu is racy, but this is ok.
4857 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4858 * If we race with it entering CPU, unaccounted time is 0. This is
4859 * indistinguishable from the read occurring a few cycles earlier.
4860 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4861 * been accounted, so we're correct here as well.
4863 if (!p->on_cpu || !task_on_rq_queued(p))
4864 return p->se.sum_exec_runtime;
4867 rq = task_rq_lock(p, &rf);
4869 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4870 * project cycles that may never be accounted to this
4871 * thread, breaking clock_gettime().
4873 if (task_current(rq, p) && task_on_rq_queued(p)) {
4874 prefetch_curr_exec_start(p);
4875 update_rq_clock(rq);
4876 p->sched_class->update_curr(rq);
4878 ns = p->se.sum_exec_runtime;
4879 task_rq_unlock(rq, p, &rf);
4884 #ifdef CONFIG_SCHED_DEBUG
4885 static u64 cpu_resched_latency(struct rq *rq)
4887 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
4888 u64 resched_latency, now = rq_clock(rq);
4889 static bool warned_once;
4891 if (sysctl_resched_latency_warn_once && warned_once)
4894 if (!need_resched() || !latency_warn_ms)
4897 if (system_state == SYSTEM_BOOTING)
4900 if (!rq->last_seen_need_resched_ns) {
4901 rq->last_seen_need_resched_ns = now;
4902 rq->ticks_without_resched = 0;
4906 rq->ticks_without_resched++;
4907 resched_latency = now - rq->last_seen_need_resched_ns;
4908 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
4913 return resched_latency;
4916 static int __init setup_resched_latency_warn_ms(char *str)
4920 if ((kstrtol(str, 0, &val))) {
4921 pr_warn("Unable to set resched_latency_warn_ms\n");
4925 sysctl_resched_latency_warn_ms = val;
4928 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
4930 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
4931 #endif /* CONFIG_SCHED_DEBUG */
4934 * This function gets called by the timer code, with HZ frequency.
4935 * We call it with interrupts disabled.
4937 void scheduler_tick(void)
4939 int cpu = smp_processor_id();
4940 struct rq *rq = cpu_rq(cpu);
4941 struct task_struct *curr = rq->curr;
4943 unsigned long thermal_pressure;
4944 u64 resched_latency;
4946 arch_scale_freq_tick();
4951 update_rq_clock(rq);
4952 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4953 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4954 curr->sched_class->task_tick(rq, curr, 0);
4955 if (sched_feat(LATENCY_WARN))
4956 resched_latency = cpu_resched_latency(rq);
4957 calc_global_load_tick(rq);
4961 if (sched_feat(LATENCY_WARN) && resched_latency)
4962 resched_latency_warn(cpu, resched_latency);
4964 perf_event_task_tick();
4967 rq->idle_balance = idle_cpu(cpu);
4968 trigger_load_balance(rq);
4972 #ifdef CONFIG_NO_HZ_FULL
4977 struct delayed_work work;
4979 /* Values for ->state, see diagram below. */
4980 #define TICK_SCHED_REMOTE_OFFLINE 0
4981 #define TICK_SCHED_REMOTE_OFFLINING 1
4982 #define TICK_SCHED_REMOTE_RUNNING 2
4985 * State diagram for ->state:
4988 * TICK_SCHED_REMOTE_OFFLINE
4991 * | | sched_tick_remote()
4994 * +--TICK_SCHED_REMOTE_OFFLINING
4997 * sched_tick_start() | | sched_tick_stop()
5000 * TICK_SCHED_REMOTE_RUNNING
5003 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5004 * and sched_tick_start() are happy to leave the state in RUNNING.
5007 static struct tick_work __percpu *tick_work_cpu;
5009 static void sched_tick_remote(struct work_struct *work)
5011 struct delayed_work *dwork = to_delayed_work(work);
5012 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5013 int cpu = twork->cpu;
5014 struct rq *rq = cpu_rq(cpu);
5015 struct task_struct *curr;
5021 * Handle the tick only if it appears the remote CPU is running in full
5022 * dynticks mode. The check is racy by nature, but missing a tick or
5023 * having one too much is no big deal because the scheduler tick updates
5024 * statistics and checks timeslices in a time-independent way, regardless
5025 * of when exactly it is running.
5027 if (!tick_nohz_tick_stopped_cpu(cpu))
5030 rq_lock_irq(rq, &rf);
5032 if (cpu_is_offline(cpu))
5035 update_rq_clock(rq);
5037 if (!is_idle_task(curr)) {
5039 * Make sure the next tick runs within a reasonable
5042 delta = rq_clock_task(rq) - curr->se.exec_start;
5043 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5045 curr->sched_class->task_tick(rq, curr, 0);
5047 calc_load_nohz_remote(rq);
5049 rq_unlock_irq(rq, &rf);
5053 * Run the remote tick once per second (1Hz). This arbitrary
5054 * frequency is large enough to avoid overload but short enough
5055 * to keep scheduler internal stats reasonably up to date. But
5056 * first update state to reflect hotplug activity if required.
5058 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5059 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5060 if (os == TICK_SCHED_REMOTE_RUNNING)
5061 queue_delayed_work(system_unbound_wq, dwork, HZ);
5064 static void sched_tick_start(int cpu)
5067 struct tick_work *twork;
5069 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5072 WARN_ON_ONCE(!tick_work_cpu);
5074 twork = per_cpu_ptr(tick_work_cpu, cpu);
5075 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5076 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5077 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5079 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5080 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5084 #ifdef CONFIG_HOTPLUG_CPU
5085 static void sched_tick_stop(int cpu)
5087 struct tick_work *twork;
5090 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5093 WARN_ON_ONCE(!tick_work_cpu);
5095 twork = per_cpu_ptr(tick_work_cpu, cpu);
5096 /* There cannot be competing actions, but don't rely on stop-machine. */
5097 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5098 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5099 /* Don't cancel, as this would mess up the state machine. */
5101 #endif /* CONFIG_HOTPLUG_CPU */
5103 int __init sched_tick_offload_init(void)
5105 tick_work_cpu = alloc_percpu(struct tick_work);
5106 BUG_ON(!tick_work_cpu);
5110 #else /* !CONFIG_NO_HZ_FULL */
5111 static inline void sched_tick_start(int cpu) { }
5112 static inline void sched_tick_stop(int cpu) { }
5115 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5116 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5118 * If the value passed in is equal to the current preempt count
5119 * then we just disabled preemption. Start timing the latency.
5121 static inline void preempt_latency_start(int val)
5123 if (preempt_count() == val) {
5124 unsigned long ip = get_lock_parent_ip();
5125 #ifdef CONFIG_DEBUG_PREEMPT
5126 current->preempt_disable_ip = ip;
5128 trace_preempt_off(CALLER_ADDR0, ip);
5132 void preempt_count_add(int val)
5134 #ifdef CONFIG_DEBUG_PREEMPT
5138 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5141 __preempt_count_add(val);
5142 #ifdef CONFIG_DEBUG_PREEMPT
5144 * Spinlock count overflowing soon?
5146 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5149 preempt_latency_start(val);
5151 EXPORT_SYMBOL(preempt_count_add);
5152 NOKPROBE_SYMBOL(preempt_count_add);
5155 * If the value passed in equals to the current preempt count
5156 * then we just enabled preemption. Stop timing the latency.
5158 static inline void preempt_latency_stop(int val)
5160 if (preempt_count() == val)
5161 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5164 void preempt_count_sub(int val)
5166 #ifdef CONFIG_DEBUG_PREEMPT
5170 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5173 * Is the spinlock portion underflowing?
5175 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5176 !(preempt_count() & PREEMPT_MASK)))
5180 preempt_latency_stop(val);
5181 __preempt_count_sub(val);
5183 EXPORT_SYMBOL(preempt_count_sub);
5184 NOKPROBE_SYMBOL(preempt_count_sub);
5187 static inline void preempt_latency_start(int val) { }
5188 static inline void preempt_latency_stop(int val) { }
5191 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5193 #ifdef CONFIG_DEBUG_PREEMPT
5194 return p->preempt_disable_ip;
5201 * Print scheduling while atomic bug:
5203 static noinline void __schedule_bug(struct task_struct *prev)
5205 /* Save this before calling printk(), since that will clobber it */
5206 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5208 if (oops_in_progress)
5211 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5212 prev->comm, prev->pid, preempt_count());
5214 debug_show_held_locks(prev);
5216 if (irqs_disabled())
5217 print_irqtrace_events(prev);
5218 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5219 && in_atomic_preempt_off()) {
5220 pr_err("Preemption disabled at:");
5221 print_ip_sym(KERN_ERR, preempt_disable_ip);
5224 panic("scheduling while atomic\n");
5227 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5231 * Various schedule()-time debugging checks and statistics:
5233 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5235 #ifdef CONFIG_SCHED_STACK_END_CHECK
5236 if (task_stack_end_corrupted(prev))
5237 panic("corrupted stack end detected inside scheduler\n");
5239 if (task_scs_end_corrupted(prev))
5240 panic("corrupted shadow stack detected inside scheduler\n");
5243 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5244 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5245 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5246 prev->comm, prev->pid, prev->non_block_count);
5248 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5252 if (unlikely(in_atomic_preempt_off())) {
5253 __schedule_bug(prev);
5254 preempt_count_set(PREEMPT_DISABLED);
5257 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5259 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5261 schedstat_inc(this_rq()->sched_count);
5264 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5265 struct rq_flags *rf)
5268 const struct sched_class *class;
5270 * We must do the balancing pass before put_prev_task(), such
5271 * that when we release the rq->lock the task is in the same
5272 * state as before we took rq->lock.
5274 * We can terminate the balance pass as soon as we know there is
5275 * a runnable task of @class priority or higher.
5277 for_class_range(class, prev->sched_class, &idle_sched_class) {
5278 if (class->balance(rq, prev, rf))
5283 put_prev_task(rq, prev);
5287 * Pick up the highest-prio task:
5289 static inline struct task_struct *
5290 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5292 const struct sched_class *class;
5293 struct task_struct *p;
5296 * Optimization: we know that if all tasks are in the fair class we can
5297 * call that function directly, but only if the @prev task wasn't of a
5298 * higher scheduling class, because otherwise those lose the
5299 * opportunity to pull in more work from other CPUs.
5301 if (likely(prev->sched_class <= &fair_sched_class &&
5302 rq->nr_running == rq->cfs.h_nr_running)) {
5304 p = pick_next_task_fair(rq, prev, rf);
5305 if (unlikely(p == RETRY_TASK))
5308 /* Assume the next prioritized class is idle_sched_class */
5310 put_prev_task(rq, prev);
5311 p = pick_next_task_idle(rq);
5318 put_prev_task_balance(rq, prev, rf);
5320 for_each_class(class) {
5321 p = class->pick_next_task(rq);
5326 /* The idle class should always have a runnable task: */
5330 #ifdef CONFIG_SCHED_CORE
5331 static inline bool is_task_rq_idle(struct task_struct *t)
5333 return (task_rq(t)->idle == t);
5336 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5338 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5341 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5343 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5346 return a->core_cookie == b->core_cookie;
5349 // XXX fairness/fwd progress conditions
5352 * - NULL if there is no runnable task for this class.
5353 * - the highest priority task for this runqueue if it matches
5354 * rq->core->core_cookie or its priority is greater than max.
5355 * - Else returns idle_task.
5357 static struct task_struct *
5358 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5360 struct task_struct *class_pick, *cookie_pick;
5361 unsigned long cookie = rq->core->core_cookie;
5363 class_pick = class->pick_task(rq);
5369 * If class_pick is tagged, return it only if it has
5370 * higher priority than max.
5372 if (max && class_pick->core_cookie &&
5373 prio_less(class_pick, max, in_fi))
5374 return idle_sched_class.pick_task(rq);
5380 * If class_pick is idle or matches cookie, return early.
5382 if (cookie_equals(class_pick, cookie))
5385 cookie_pick = sched_core_find(rq, cookie);
5388 * If class > max && class > cookie, it is the highest priority task on
5389 * the core (so far) and it must be selected, otherwise we must go with
5390 * the cookie pick in order to satisfy the constraint.
5392 if (prio_less(cookie_pick, class_pick, in_fi) &&
5393 (!max || prio_less(max, class_pick, in_fi)))
5399 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5401 static struct task_struct *
5402 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5404 struct task_struct *next, *max = NULL;
5405 const struct sched_class *class;
5406 const struct cpumask *smt_mask;
5407 bool fi_before = false;
5408 int i, j, cpu, occ = 0;
5411 if (!sched_core_enabled(rq))
5412 return __pick_next_task(rq, prev, rf);
5416 /* Stopper task is switching into idle, no need core-wide selection. */
5417 if (cpu_is_offline(cpu)) {
5419 * Reset core_pick so that we don't enter the fastpath when
5420 * coming online. core_pick would already be migrated to
5421 * another cpu during offline.
5423 rq->core_pick = NULL;
5424 return __pick_next_task(rq, prev, rf);
5428 * If there were no {en,de}queues since we picked (IOW, the task
5429 * pointers are all still valid), and we haven't scheduled the last
5430 * pick yet, do so now.
5432 * rq->core_pick can be NULL if no selection was made for a CPU because
5433 * it was either offline or went offline during a sibling's core-wide
5434 * selection. In this case, do a core-wide selection.
5436 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5437 rq->core->core_pick_seq != rq->core_sched_seq &&
5439 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5441 next = rq->core_pick;
5443 put_prev_task(rq, prev);
5444 set_next_task(rq, next);
5447 rq->core_pick = NULL;
5451 put_prev_task_balance(rq, prev, rf);
5453 smt_mask = cpu_smt_mask(cpu);
5454 need_sync = !!rq->core->core_cookie;
5457 rq->core->core_cookie = 0UL;
5458 if (rq->core->core_forceidle) {
5461 rq->core->core_forceidle = false;
5465 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5467 * @task_seq guards the task state ({en,de}queues)
5468 * @pick_seq is the @task_seq we did a selection on
5469 * @sched_seq is the @pick_seq we scheduled
5471 * However, preemptions can cause multiple picks on the same task set.
5472 * 'Fix' this by also increasing @task_seq for every pick.
5474 rq->core->core_task_seq++;
5477 * Optimize for common case where this CPU has no cookies
5478 * and there are no cookied tasks running on siblings.
5481 for_each_class(class) {
5482 next = class->pick_task(rq);
5487 if (!next->core_cookie) {
5488 rq->core_pick = NULL;
5490 * For robustness, update the min_vruntime_fi for
5491 * unconstrained picks as well.
5493 WARN_ON_ONCE(fi_before);
5494 task_vruntime_update(rq, next, false);
5499 for_each_cpu(i, smt_mask) {
5500 struct rq *rq_i = cpu_rq(i);
5502 rq_i->core_pick = NULL;
5505 update_rq_clock(rq_i);
5509 * Try and select tasks for each sibling in descending sched_class
5512 for_each_class(class) {
5514 for_each_cpu_wrap(i, smt_mask, cpu) {
5515 struct rq *rq_i = cpu_rq(i);
5516 struct task_struct *p;
5518 if (rq_i->core_pick)
5522 * If this sibling doesn't yet have a suitable task to
5523 * run; ask for the most eligible task, given the
5524 * highest priority task already selected for this
5527 p = pick_task(rq_i, class, max, fi_before);
5531 if (!is_task_rq_idle(p))
5534 rq_i->core_pick = p;
5535 if (rq_i->idle == p && rq_i->nr_running) {
5536 rq->core->core_forceidle = true;
5538 rq->core->core_forceidle_seq++;
5542 * If this new candidate is of higher priority than the
5543 * previous; and they're incompatible; we need to wipe
5544 * the slate and start over. pick_task makes sure that
5545 * p's priority is more than max if it doesn't match
5548 * NOTE: this is a linear max-filter and is thus bounded
5549 * in execution time.
5551 if (!max || !cookie_match(max, p)) {
5552 struct task_struct *old_max = max;
5554 rq->core->core_cookie = p->core_cookie;
5558 rq->core->core_forceidle = false;
5559 for_each_cpu(j, smt_mask) {
5563 cpu_rq(j)->core_pick = NULL;
5572 rq->core->core_pick_seq = rq->core->core_task_seq;
5573 next = rq->core_pick;
5574 rq->core_sched_seq = rq->core->core_pick_seq;
5576 /* Something should have been selected for current CPU */
5577 WARN_ON_ONCE(!next);
5580 * Reschedule siblings
5582 * NOTE: L1TF -- at this point we're no longer running the old task and
5583 * sending an IPI (below) ensures the sibling will no longer be running
5584 * their task. This ensures there is no inter-sibling overlap between
5585 * non-matching user state.
5587 for_each_cpu(i, smt_mask) {
5588 struct rq *rq_i = cpu_rq(i);
5591 * An online sibling might have gone offline before a task
5592 * could be picked for it, or it might be offline but later
5593 * happen to come online, but its too late and nothing was
5594 * picked for it. That's Ok - it will pick tasks for itself,
5597 if (!rq_i->core_pick)
5601 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5602 * fi_before fi update?
5608 if (!(fi_before && rq->core->core_forceidle))
5609 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5611 rq_i->core_pick->core_occupation = occ;
5614 rq_i->core_pick = NULL;
5618 /* Did we break L1TF mitigation requirements? */
5619 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5621 if (rq_i->curr == rq_i->core_pick) {
5622 rq_i->core_pick = NULL;
5630 set_next_task(rq, next);
5634 static bool try_steal_cookie(int this, int that)
5636 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5637 struct task_struct *p;
5638 unsigned long cookie;
5639 bool success = false;
5641 local_irq_disable();
5642 double_rq_lock(dst, src);
5644 cookie = dst->core->core_cookie;
5648 if (dst->curr != dst->idle)
5651 p = sched_core_find(src, cookie);
5656 if (p == src->core_pick || p == src->curr)
5659 if (!cpumask_test_cpu(this, &p->cpus_mask))
5662 if (p->core_occupation > dst->idle->core_occupation)
5665 p->on_rq = TASK_ON_RQ_MIGRATING;
5666 deactivate_task(src, p, 0);
5667 set_task_cpu(p, this);
5668 activate_task(dst, p, 0);
5669 p->on_rq = TASK_ON_RQ_QUEUED;
5677 p = sched_core_next(p, cookie);
5681 double_rq_unlock(dst, src);
5687 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5691 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5698 if (try_steal_cookie(cpu, i))
5705 static void sched_core_balance(struct rq *rq)
5707 struct sched_domain *sd;
5708 int cpu = cpu_of(rq);
5712 raw_spin_rq_unlock_irq(rq);
5713 for_each_domain(cpu, sd) {
5717 if (steal_cookie_task(cpu, sd))
5720 raw_spin_rq_lock_irq(rq);
5725 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5727 void queue_core_balance(struct rq *rq)
5729 if (!sched_core_enabled(rq))
5732 if (!rq->core->core_cookie)
5735 if (!rq->nr_running) /* not forced idle */
5738 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5741 static inline void sched_core_cpu_starting(unsigned int cpu)
5743 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5744 struct rq *rq, *core_rq = NULL;
5747 core_rq = cpu_rq(cpu)->core;
5750 for_each_cpu(i, smt_mask) {
5752 if (rq->core && rq->core == rq)
5757 core_rq = cpu_rq(cpu);
5759 for_each_cpu(i, smt_mask) {
5762 WARN_ON_ONCE(rq->core && rq->core != core_rq);
5767 #else /* !CONFIG_SCHED_CORE */
5769 static inline void sched_core_cpu_starting(unsigned int cpu) {}
5771 static struct task_struct *
5772 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5774 return __pick_next_task(rq, prev, rf);
5777 #endif /* CONFIG_SCHED_CORE */
5780 * __schedule() is the main scheduler function.
5782 * The main means of driving the scheduler and thus entering this function are:
5784 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
5786 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
5787 * paths. For example, see arch/x86/entry_64.S.
5789 * To drive preemption between tasks, the scheduler sets the flag in timer
5790 * interrupt handler scheduler_tick().
5792 * 3. Wakeups don't really cause entry into schedule(). They add a
5793 * task to the run-queue and that's it.
5795 * Now, if the new task added to the run-queue preempts the current
5796 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
5797 * called on the nearest possible occasion:
5799 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
5801 * - in syscall or exception context, at the next outmost
5802 * preempt_enable(). (this might be as soon as the wake_up()'s
5805 * - in IRQ context, return from interrupt-handler to
5806 * preemptible context
5808 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
5811 * - cond_resched() call
5812 * - explicit schedule() call
5813 * - return from syscall or exception to user-space
5814 * - return from interrupt-handler to user-space
5816 * WARNING: must be called with preemption disabled!
5818 static void __sched notrace __schedule(bool preempt)
5820 struct task_struct *prev, *next;
5821 unsigned long *switch_count;
5822 unsigned long prev_state;
5827 cpu = smp_processor_id();
5831 schedule_debug(prev, preempt);
5833 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
5836 local_irq_disable();
5837 rcu_note_context_switch(preempt);
5840 * Make sure that signal_pending_state()->signal_pending() below
5841 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
5842 * done by the caller to avoid the race with signal_wake_up():
5844 * __set_current_state(@state) signal_wake_up()
5845 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
5846 * wake_up_state(p, state)
5847 * LOCK rq->lock LOCK p->pi_state
5848 * smp_mb__after_spinlock() smp_mb__after_spinlock()
5849 * if (signal_pending_state()) if (p->state & @state)
5851 * Also, the membarrier system call requires a full memory barrier
5852 * after coming from user-space, before storing to rq->curr.
5855 smp_mb__after_spinlock();
5857 /* Promote REQ to ACT */
5858 rq->clock_update_flags <<= 1;
5859 update_rq_clock(rq);
5861 switch_count = &prev->nivcsw;
5864 * We must load prev->state once (task_struct::state is volatile), such
5867 * - we form a control dependency vs deactivate_task() below.
5868 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5870 prev_state = READ_ONCE(prev->__state);
5871 if (!preempt && prev_state) {
5872 if (signal_pending_state(prev_state, prev)) {
5873 WRITE_ONCE(prev->__state, TASK_RUNNING);
5875 prev->sched_contributes_to_load =
5876 (prev_state & TASK_UNINTERRUPTIBLE) &&
5877 !(prev_state & TASK_NOLOAD) &&
5878 !(prev->flags & PF_FROZEN);
5880 if (prev->sched_contributes_to_load)
5881 rq->nr_uninterruptible++;
5884 * __schedule() ttwu()
5885 * prev_state = prev->state; if (p->on_rq && ...)
5886 * if (prev_state) goto out;
5887 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5888 * p->state = TASK_WAKING
5890 * Where __schedule() and ttwu() have matching control dependencies.
5892 * After this, schedule() must not care about p->state any more.
5894 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
5896 if (prev->in_iowait) {
5897 atomic_inc(&rq->nr_iowait);
5898 delayacct_blkio_start();
5901 switch_count = &prev->nvcsw;
5904 next = pick_next_task(rq, prev, &rf);
5905 clear_tsk_need_resched(prev);
5906 clear_preempt_need_resched();
5907 #ifdef CONFIG_SCHED_DEBUG
5908 rq->last_seen_need_resched_ns = 0;
5911 if (likely(prev != next)) {
5914 * RCU users of rcu_dereference(rq->curr) may not see
5915 * changes to task_struct made by pick_next_task().
5917 RCU_INIT_POINTER(rq->curr, next);
5919 * The membarrier system call requires each architecture
5920 * to have a full memory barrier after updating
5921 * rq->curr, before returning to user-space.
5923 * Here are the schemes providing that barrier on the
5924 * various architectures:
5925 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5926 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5927 * - finish_lock_switch() for weakly-ordered
5928 * architectures where spin_unlock is a full barrier,
5929 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5930 * is a RELEASE barrier),
5934 migrate_disable_switch(rq, prev);
5935 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5937 trace_sched_switch(preempt, prev, next);
5939 /* Also unlocks the rq: */
5940 rq = context_switch(rq, prev, next, &rf);
5942 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5944 rq_unpin_lock(rq, &rf);
5945 __balance_callbacks(rq);
5946 raw_spin_rq_unlock_irq(rq);
5950 void __noreturn do_task_dead(void)
5952 /* Causes final put_task_struct in finish_task_switch(): */
5953 set_special_state(TASK_DEAD);
5955 /* Tell freezer to ignore us: */
5956 current->flags |= PF_NOFREEZE;
5961 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5966 static inline void sched_submit_work(struct task_struct *tsk)
5968 unsigned int task_flags;
5970 if (task_is_running(tsk))
5973 task_flags = tsk->flags;
5975 * If a worker went to sleep, notify and ask workqueue whether
5976 * it wants to wake up a task to maintain concurrency.
5977 * As this function is called inside the schedule() context,
5978 * we disable preemption to avoid it calling schedule() again
5979 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5982 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5984 if (task_flags & PF_WQ_WORKER)
5985 wq_worker_sleeping(tsk);
5987 io_wq_worker_sleeping(tsk);
5988 preempt_enable_no_resched();
5991 if (tsk_is_pi_blocked(tsk))
5995 * If we are going to sleep and we have plugged IO queued,
5996 * make sure to submit it to avoid deadlocks.
5998 if (blk_needs_flush_plug(tsk))
5999 blk_schedule_flush_plug(tsk);
6002 static void sched_update_worker(struct task_struct *tsk)
6004 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6005 if (tsk->flags & PF_WQ_WORKER)
6006 wq_worker_running(tsk);
6008 io_wq_worker_running(tsk);
6012 asmlinkage __visible void __sched schedule(void)
6014 struct task_struct *tsk = current;
6016 sched_submit_work(tsk);
6020 sched_preempt_enable_no_resched();
6021 } while (need_resched());
6022 sched_update_worker(tsk);
6024 EXPORT_SYMBOL(schedule);
6027 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6028 * state (have scheduled out non-voluntarily) by making sure that all
6029 * tasks have either left the run queue or have gone into user space.
6030 * As idle tasks do not do either, they must not ever be preempted
6031 * (schedule out non-voluntarily).
6033 * schedule_idle() is similar to schedule_preempt_disable() except that it
6034 * never enables preemption because it does not call sched_submit_work().
6036 void __sched schedule_idle(void)
6039 * As this skips calling sched_submit_work(), which the idle task does
6040 * regardless because that function is a nop when the task is in a
6041 * TASK_RUNNING state, make sure this isn't used someplace that the
6042 * current task can be in any other state. Note, idle is always in the
6043 * TASK_RUNNING state.
6045 WARN_ON_ONCE(current->__state);
6048 } while (need_resched());
6051 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6052 asmlinkage __visible void __sched schedule_user(void)
6055 * If we come here after a random call to set_need_resched(),
6056 * or we have been woken up remotely but the IPI has not yet arrived,
6057 * we haven't yet exited the RCU idle mode. Do it here manually until
6058 * we find a better solution.
6060 * NB: There are buggy callers of this function. Ideally we
6061 * should warn if prev_state != CONTEXT_USER, but that will trigger
6062 * too frequently to make sense yet.
6064 enum ctx_state prev_state = exception_enter();
6066 exception_exit(prev_state);
6071 * schedule_preempt_disabled - called with preemption disabled
6073 * Returns with preemption disabled. Note: preempt_count must be 1
6075 void __sched schedule_preempt_disabled(void)
6077 sched_preempt_enable_no_resched();
6082 static void __sched notrace preempt_schedule_common(void)
6086 * Because the function tracer can trace preempt_count_sub()
6087 * and it also uses preempt_enable/disable_notrace(), if
6088 * NEED_RESCHED is set, the preempt_enable_notrace() called
6089 * by the function tracer will call this function again and
6090 * cause infinite recursion.
6092 * Preemption must be disabled here before the function
6093 * tracer can trace. Break up preempt_disable() into two
6094 * calls. One to disable preemption without fear of being
6095 * traced. The other to still record the preemption latency,
6096 * which can also be traced by the function tracer.
6098 preempt_disable_notrace();
6099 preempt_latency_start(1);
6101 preempt_latency_stop(1);
6102 preempt_enable_no_resched_notrace();
6105 * Check again in case we missed a preemption opportunity
6106 * between schedule and now.
6108 } while (need_resched());
6111 #ifdef CONFIG_PREEMPTION
6113 * This is the entry point to schedule() from in-kernel preemption
6114 * off of preempt_enable.
6116 asmlinkage __visible void __sched notrace preempt_schedule(void)
6119 * If there is a non-zero preempt_count or interrupts are disabled,
6120 * we do not want to preempt the current task. Just return..
6122 if (likely(!preemptible()))
6125 preempt_schedule_common();
6127 NOKPROBE_SYMBOL(preempt_schedule);
6128 EXPORT_SYMBOL(preempt_schedule);
6130 #ifdef CONFIG_PREEMPT_DYNAMIC
6131 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6132 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6137 * preempt_schedule_notrace - preempt_schedule called by tracing
6139 * The tracing infrastructure uses preempt_enable_notrace to prevent
6140 * recursion and tracing preempt enabling caused by the tracing
6141 * infrastructure itself. But as tracing can happen in areas coming
6142 * from userspace or just about to enter userspace, a preempt enable
6143 * can occur before user_exit() is called. This will cause the scheduler
6144 * to be called when the system is still in usermode.
6146 * To prevent this, the preempt_enable_notrace will use this function
6147 * instead of preempt_schedule() to exit user context if needed before
6148 * calling the scheduler.
6150 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6152 enum ctx_state prev_ctx;
6154 if (likely(!preemptible()))
6159 * Because the function tracer can trace preempt_count_sub()
6160 * and it also uses preempt_enable/disable_notrace(), if
6161 * NEED_RESCHED is set, the preempt_enable_notrace() called
6162 * by the function tracer will call this function again and
6163 * cause infinite recursion.
6165 * Preemption must be disabled here before the function
6166 * tracer can trace. Break up preempt_disable() into two
6167 * calls. One to disable preemption without fear of being
6168 * traced. The other to still record the preemption latency,
6169 * which can also be traced by the function tracer.
6171 preempt_disable_notrace();
6172 preempt_latency_start(1);
6174 * Needs preempt disabled in case user_exit() is traced
6175 * and the tracer calls preempt_enable_notrace() causing
6176 * an infinite recursion.
6178 prev_ctx = exception_enter();
6180 exception_exit(prev_ctx);
6182 preempt_latency_stop(1);
6183 preempt_enable_no_resched_notrace();
6184 } while (need_resched());
6186 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6188 #ifdef CONFIG_PREEMPT_DYNAMIC
6189 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6190 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6193 #endif /* CONFIG_PREEMPTION */
6195 #ifdef CONFIG_PREEMPT_DYNAMIC
6197 #include <linux/entry-common.h>
6202 * SC:preempt_schedule
6203 * SC:preempt_schedule_notrace
6204 * SC:irqentry_exit_cond_resched
6208 * cond_resched <- __cond_resched
6209 * might_resched <- RET0
6210 * preempt_schedule <- NOP
6211 * preempt_schedule_notrace <- NOP
6212 * irqentry_exit_cond_resched <- NOP
6215 * cond_resched <- __cond_resched
6216 * might_resched <- __cond_resched
6217 * preempt_schedule <- NOP
6218 * preempt_schedule_notrace <- NOP
6219 * irqentry_exit_cond_resched <- NOP
6222 * cond_resched <- RET0
6223 * might_resched <- RET0
6224 * preempt_schedule <- preempt_schedule
6225 * preempt_schedule_notrace <- preempt_schedule_notrace
6226 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6230 preempt_dynamic_none = 0,
6231 preempt_dynamic_voluntary,
6232 preempt_dynamic_full,
6235 int preempt_dynamic_mode = preempt_dynamic_full;
6237 int sched_dynamic_mode(const char *str)
6239 if (!strcmp(str, "none"))
6240 return preempt_dynamic_none;
6242 if (!strcmp(str, "voluntary"))
6243 return preempt_dynamic_voluntary;
6245 if (!strcmp(str, "full"))
6246 return preempt_dynamic_full;
6251 void sched_dynamic_update(int mode)
6254 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6255 * the ZERO state, which is invalid.
6257 static_call_update(cond_resched, __cond_resched);
6258 static_call_update(might_resched, __cond_resched);
6259 static_call_update(preempt_schedule, __preempt_schedule_func);
6260 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6261 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6264 case preempt_dynamic_none:
6265 static_call_update(cond_resched, __cond_resched);
6266 static_call_update(might_resched, (void *)&__static_call_return0);
6267 static_call_update(preempt_schedule, NULL);
6268 static_call_update(preempt_schedule_notrace, NULL);
6269 static_call_update(irqentry_exit_cond_resched, NULL);
6270 pr_info("Dynamic Preempt: none\n");
6273 case preempt_dynamic_voluntary:
6274 static_call_update(cond_resched, __cond_resched);
6275 static_call_update(might_resched, __cond_resched);
6276 static_call_update(preempt_schedule, NULL);
6277 static_call_update(preempt_schedule_notrace, NULL);
6278 static_call_update(irqentry_exit_cond_resched, NULL);
6279 pr_info("Dynamic Preempt: voluntary\n");
6282 case preempt_dynamic_full:
6283 static_call_update(cond_resched, (void *)&__static_call_return0);
6284 static_call_update(might_resched, (void *)&__static_call_return0);
6285 static_call_update(preempt_schedule, __preempt_schedule_func);
6286 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6287 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6288 pr_info("Dynamic Preempt: full\n");
6292 preempt_dynamic_mode = mode;
6295 static int __init setup_preempt_mode(char *str)
6297 int mode = sched_dynamic_mode(str);
6299 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6303 sched_dynamic_update(mode);
6306 __setup("preempt=", setup_preempt_mode);
6308 #endif /* CONFIG_PREEMPT_DYNAMIC */
6311 * This is the entry point to schedule() from kernel preemption
6312 * off of irq context.
6313 * Note, that this is called and return with irqs disabled. This will
6314 * protect us against recursive calling from irq.
6316 asmlinkage __visible void __sched preempt_schedule_irq(void)
6318 enum ctx_state prev_state;
6320 /* Catch callers which need to be fixed */
6321 BUG_ON(preempt_count() || !irqs_disabled());
6323 prev_state = exception_enter();
6329 local_irq_disable();
6330 sched_preempt_enable_no_resched();
6331 } while (need_resched());
6333 exception_exit(prev_state);
6336 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6339 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6340 return try_to_wake_up(curr->private, mode, wake_flags);
6342 EXPORT_SYMBOL(default_wake_function);
6344 #ifdef CONFIG_RT_MUTEXES
6346 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6349 prio = min(prio, pi_task->prio);
6354 static inline int rt_effective_prio(struct task_struct *p, int prio)
6356 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6358 return __rt_effective_prio(pi_task, prio);
6362 * rt_mutex_setprio - set the current priority of a task
6364 * @pi_task: donor task
6366 * This function changes the 'effective' priority of a task. It does
6367 * not touch ->normal_prio like __setscheduler().
6369 * Used by the rt_mutex code to implement priority inheritance
6370 * logic. Call site only calls if the priority of the task changed.
6372 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6374 int prio, oldprio, queued, running, queue_flag =
6375 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6376 const struct sched_class *prev_class;
6380 /* XXX used to be waiter->prio, not waiter->task->prio */
6381 prio = __rt_effective_prio(pi_task, p->normal_prio);
6384 * If nothing changed; bail early.
6386 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6389 rq = __task_rq_lock(p, &rf);
6390 update_rq_clock(rq);
6392 * Set under pi_lock && rq->lock, such that the value can be used under
6395 * Note that there is loads of tricky to make this pointer cache work
6396 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6397 * ensure a task is de-boosted (pi_task is set to NULL) before the
6398 * task is allowed to run again (and can exit). This ensures the pointer
6399 * points to a blocked task -- which guarantees the task is present.
6401 p->pi_top_task = pi_task;
6404 * For FIFO/RR we only need to set prio, if that matches we're done.
6406 if (prio == p->prio && !dl_prio(prio))
6410 * Idle task boosting is a nono in general. There is one
6411 * exception, when PREEMPT_RT and NOHZ is active:
6413 * The idle task calls get_next_timer_interrupt() and holds
6414 * the timer wheel base->lock on the CPU and another CPU wants
6415 * to access the timer (probably to cancel it). We can safely
6416 * ignore the boosting request, as the idle CPU runs this code
6417 * with interrupts disabled and will complete the lock
6418 * protected section without being interrupted. So there is no
6419 * real need to boost.
6421 if (unlikely(p == rq->idle)) {
6422 WARN_ON(p != rq->curr);
6423 WARN_ON(p->pi_blocked_on);
6427 trace_sched_pi_setprio(p, pi_task);
6430 if (oldprio == prio)
6431 queue_flag &= ~DEQUEUE_MOVE;
6433 prev_class = p->sched_class;
6434 queued = task_on_rq_queued(p);
6435 running = task_current(rq, p);
6437 dequeue_task(rq, p, queue_flag);
6439 put_prev_task(rq, p);
6442 * Boosting condition are:
6443 * 1. -rt task is running and holds mutex A
6444 * --> -dl task blocks on mutex A
6446 * 2. -dl task is running and holds mutex A
6447 * --> -dl task blocks on mutex A and could preempt the
6450 if (dl_prio(prio)) {
6451 if (!dl_prio(p->normal_prio) ||
6452 (pi_task && dl_prio(pi_task->prio) &&
6453 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6454 p->dl.pi_se = pi_task->dl.pi_se;
6455 queue_flag |= ENQUEUE_REPLENISH;
6457 p->dl.pi_se = &p->dl;
6459 p->sched_class = &dl_sched_class;
6460 } else if (rt_prio(prio)) {
6461 if (dl_prio(oldprio))
6462 p->dl.pi_se = &p->dl;
6464 queue_flag |= ENQUEUE_HEAD;
6465 p->sched_class = &rt_sched_class;
6467 if (dl_prio(oldprio))
6468 p->dl.pi_se = &p->dl;
6469 if (rt_prio(oldprio))
6471 p->sched_class = &fair_sched_class;
6477 enqueue_task(rq, p, queue_flag);
6479 set_next_task(rq, p);
6481 check_class_changed(rq, p, prev_class, oldprio);
6483 /* Avoid rq from going away on us: */
6486 rq_unpin_lock(rq, &rf);
6487 __balance_callbacks(rq);
6488 raw_spin_rq_unlock(rq);
6493 static inline int rt_effective_prio(struct task_struct *p, int prio)
6499 void set_user_nice(struct task_struct *p, long nice)
6501 bool queued, running;
6506 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6509 * We have to be careful, if called from sys_setpriority(),
6510 * the task might be in the middle of scheduling on another CPU.
6512 rq = task_rq_lock(p, &rf);
6513 update_rq_clock(rq);
6516 * The RT priorities are set via sched_setscheduler(), but we still
6517 * allow the 'normal' nice value to be set - but as expected
6518 * it won't have any effect on scheduling until the task is
6519 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6521 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6522 p->static_prio = NICE_TO_PRIO(nice);
6525 queued = task_on_rq_queued(p);
6526 running = task_current(rq, p);
6528 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6530 put_prev_task(rq, p);
6532 p->static_prio = NICE_TO_PRIO(nice);
6533 set_load_weight(p, true);
6535 p->prio = effective_prio(p);
6538 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6540 set_next_task(rq, p);
6543 * If the task increased its priority or is running and
6544 * lowered its priority, then reschedule its CPU:
6546 p->sched_class->prio_changed(rq, p, old_prio);
6549 task_rq_unlock(rq, p, &rf);
6551 EXPORT_SYMBOL(set_user_nice);
6554 * can_nice - check if a task can reduce its nice value
6558 int can_nice(const struct task_struct *p, const int nice)
6560 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6561 int nice_rlim = nice_to_rlimit(nice);
6563 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6564 capable(CAP_SYS_NICE));
6567 #ifdef __ARCH_WANT_SYS_NICE
6570 * sys_nice - change the priority of the current process.
6571 * @increment: priority increment
6573 * sys_setpriority is a more generic, but much slower function that
6574 * does similar things.
6576 SYSCALL_DEFINE1(nice, int, increment)
6581 * Setpriority might change our priority at the same moment.
6582 * We don't have to worry. Conceptually one call occurs first
6583 * and we have a single winner.
6585 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6586 nice = task_nice(current) + increment;
6588 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6589 if (increment < 0 && !can_nice(current, nice))
6592 retval = security_task_setnice(current, nice);
6596 set_user_nice(current, nice);
6603 * task_prio - return the priority value of a given task.
6604 * @p: the task in question.
6606 * Return: The priority value as seen by users in /proc.
6608 * sched policy return value kernel prio user prio/nice
6610 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6611 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6612 * deadline -101 -1 0
6614 int task_prio(const struct task_struct *p)
6616 return p->prio - MAX_RT_PRIO;
6620 * idle_cpu - is a given CPU idle currently?
6621 * @cpu: the processor in question.
6623 * Return: 1 if the CPU is currently idle. 0 otherwise.
6625 int idle_cpu(int cpu)
6627 struct rq *rq = cpu_rq(cpu);
6629 if (rq->curr != rq->idle)
6636 if (rq->ttwu_pending)
6644 * available_idle_cpu - is a given CPU idle for enqueuing work.
6645 * @cpu: the CPU in question.
6647 * Return: 1 if the CPU is currently idle. 0 otherwise.
6649 int available_idle_cpu(int cpu)
6654 if (vcpu_is_preempted(cpu))
6661 * idle_task - return the idle task for a given CPU.
6662 * @cpu: the processor in question.
6664 * Return: The idle task for the CPU @cpu.
6666 struct task_struct *idle_task(int cpu)
6668 return cpu_rq(cpu)->idle;
6673 * This function computes an effective utilization for the given CPU, to be
6674 * used for frequency selection given the linear relation: f = u * f_max.
6676 * The scheduler tracks the following metrics:
6678 * cpu_util_{cfs,rt,dl,irq}()
6681 * Where the cfs,rt and dl util numbers are tracked with the same metric and
6682 * synchronized windows and are thus directly comparable.
6684 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
6685 * which excludes things like IRQ and steal-time. These latter are then accrued
6686 * in the irq utilization.
6688 * The DL bandwidth number otoh is not a measured metric but a value computed
6689 * based on the task model parameters and gives the minimal utilization
6690 * required to meet deadlines.
6692 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
6693 unsigned long max, enum cpu_util_type type,
6694 struct task_struct *p)
6696 unsigned long dl_util, util, irq;
6697 struct rq *rq = cpu_rq(cpu);
6699 if (!uclamp_is_used() &&
6700 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
6705 * Early check to see if IRQ/steal time saturates the CPU, can be
6706 * because of inaccuracies in how we track these -- see
6707 * update_irq_load_avg().
6709 irq = cpu_util_irq(rq);
6710 if (unlikely(irq >= max))
6714 * Because the time spend on RT/DL tasks is visible as 'lost' time to
6715 * CFS tasks and we use the same metric to track the effective
6716 * utilization (PELT windows are synchronized) we can directly add them
6717 * to obtain the CPU's actual utilization.
6719 * CFS and RT utilization can be boosted or capped, depending on
6720 * utilization clamp constraints requested by currently RUNNABLE
6722 * When there are no CFS RUNNABLE tasks, clamps are released and
6723 * frequency will be gracefully reduced with the utilization decay.
6725 util = util_cfs + cpu_util_rt(rq);
6726 if (type == FREQUENCY_UTIL)
6727 util = uclamp_rq_util_with(rq, util, p);
6729 dl_util = cpu_util_dl(rq);
6732 * For frequency selection we do not make cpu_util_dl() a permanent part
6733 * of this sum because we want to use cpu_bw_dl() later on, but we need
6734 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
6735 * that we select f_max when there is no idle time.
6737 * NOTE: numerical errors or stop class might cause us to not quite hit
6738 * saturation when we should -- something for later.
6740 if (util + dl_util >= max)
6744 * OTOH, for energy computation we need the estimated running time, so
6745 * include util_dl and ignore dl_bw.
6747 if (type == ENERGY_UTIL)
6751 * There is still idle time; further improve the number by using the
6752 * irq metric. Because IRQ/steal time is hidden from the task clock we
6753 * need to scale the task numbers:
6756 * U' = irq + --------- * U
6759 util = scale_irq_capacity(util, irq, max);
6763 * Bandwidth required by DEADLINE must always be granted while, for
6764 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
6765 * to gracefully reduce the frequency when no tasks show up for longer
6768 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
6769 * bw_dl as requested freq. However, cpufreq is not yet ready for such
6770 * an interface. So, we only do the latter for now.
6772 if (type == FREQUENCY_UTIL)
6773 util += cpu_bw_dl(rq);
6775 return min(max, util);
6778 unsigned long sched_cpu_util(int cpu, unsigned long max)
6780 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
6783 #endif /* CONFIG_SMP */
6786 * find_process_by_pid - find a process with a matching PID value.
6787 * @pid: the pid in question.
6789 * The task of @pid, if found. %NULL otherwise.
6791 static struct task_struct *find_process_by_pid(pid_t pid)
6793 return pid ? find_task_by_vpid(pid) : current;
6797 * sched_setparam() passes in -1 for its policy, to let the functions
6798 * it calls know not to change it.
6800 #define SETPARAM_POLICY -1
6802 static void __setscheduler_params(struct task_struct *p,
6803 const struct sched_attr *attr)
6805 int policy = attr->sched_policy;
6807 if (policy == SETPARAM_POLICY)
6812 if (dl_policy(policy))
6813 __setparam_dl(p, attr);
6814 else if (fair_policy(policy))
6815 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6818 * __sched_setscheduler() ensures attr->sched_priority == 0 when
6819 * !rt_policy. Always setting this ensures that things like
6820 * getparam()/getattr() don't report silly values for !rt tasks.
6822 p->rt_priority = attr->sched_priority;
6823 p->normal_prio = normal_prio(p);
6824 set_load_weight(p, true);
6827 /* Actually do priority change: must hold pi & rq lock. */
6828 static void __setscheduler(struct rq *rq, struct task_struct *p,
6829 const struct sched_attr *attr, bool keep_boost)
6832 * If params can't change scheduling class changes aren't allowed
6835 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
6838 __setscheduler_params(p, attr);
6841 * Keep a potential priority boosting if called from
6842 * sched_setscheduler().
6844 p->prio = normal_prio(p);
6846 p->prio = rt_effective_prio(p, p->prio);
6848 if (dl_prio(p->prio))
6849 p->sched_class = &dl_sched_class;
6850 else if (rt_prio(p->prio))
6851 p->sched_class = &rt_sched_class;
6853 p->sched_class = &fair_sched_class;
6857 * Check the target process has a UID that matches the current process's:
6859 static bool check_same_owner(struct task_struct *p)
6861 const struct cred *cred = current_cred(), *pcred;
6865 pcred = __task_cred(p);
6866 match = (uid_eq(cred->euid, pcred->euid) ||
6867 uid_eq(cred->euid, pcred->uid));
6872 static int __sched_setscheduler(struct task_struct *p,
6873 const struct sched_attr *attr,
6876 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
6877 MAX_RT_PRIO - 1 - attr->sched_priority;
6878 int retval, oldprio, oldpolicy = -1, queued, running;
6879 int new_effective_prio, policy = attr->sched_policy;
6880 const struct sched_class *prev_class;
6881 struct callback_head *head;
6884 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6887 /* The pi code expects interrupts enabled */
6888 BUG_ON(pi && in_interrupt());
6890 /* Double check policy once rq lock held: */
6892 reset_on_fork = p->sched_reset_on_fork;
6893 policy = oldpolicy = p->policy;
6895 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
6897 if (!valid_policy(policy))
6901 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
6905 * Valid priorities for SCHED_FIFO and SCHED_RR are
6906 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6907 * SCHED_BATCH and SCHED_IDLE is 0.
6909 if (attr->sched_priority > MAX_RT_PRIO-1)
6911 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
6912 (rt_policy(policy) != (attr->sched_priority != 0)))
6916 * Allow unprivileged RT tasks to decrease priority:
6918 if (user && !capable(CAP_SYS_NICE)) {
6919 if (fair_policy(policy)) {
6920 if (attr->sched_nice < task_nice(p) &&
6921 !can_nice(p, attr->sched_nice))
6925 if (rt_policy(policy)) {
6926 unsigned long rlim_rtprio =
6927 task_rlimit(p, RLIMIT_RTPRIO);
6929 /* Can't set/change the rt policy: */
6930 if (policy != p->policy && !rlim_rtprio)
6933 /* Can't increase priority: */
6934 if (attr->sched_priority > p->rt_priority &&
6935 attr->sched_priority > rlim_rtprio)
6940 * Can't set/change SCHED_DEADLINE policy at all for now
6941 * (safest behavior); in the future we would like to allow
6942 * unprivileged DL tasks to increase their relative deadline
6943 * or reduce their runtime (both ways reducing utilization)
6945 if (dl_policy(policy))
6949 * Treat SCHED_IDLE as nice 20. Only allow a switch to
6950 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6952 if (task_has_idle_policy(p) && !idle_policy(policy)) {
6953 if (!can_nice(p, task_nice(p)))
6957 /* Can't change other user's priorities: */
6958 if (!check_same_owner(p))
6961 /* Normal users shall not reset the sched_reset_on_fork flag: */
6962 if (p->sched_reset_on_fork && !reset_on_fork)
6967 if (attr->sched_flags & SCHED_FLAG_SUGOV)
6970 retval = security_task_setscheduler(p);
6975 /* Update task specific "requested" clamps */
6976 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
6977 retval = uclamp_validate(p, attr);
6986 * Make sure no PI-waiters arrive (or leave) while we are
6987 * changing the priority of the task:
6989 * To be able to change p->policy safely, the appropriate
6990 * runqueue lock must be held.
6992 rq = task_rq_lock(p, &rf);
6993 update_rq_clock(rq);
6996 * Changing the policy of the stop threads its a very bad idea:
6998 if (p == rq->stop) {
7004 * If not changing anything there's no need to proceed further,
7005 * but store a possible modification of reset_on_fork.
7007 if (unlikely(policy == p->policy)) {
7008 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7010 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7012 if (dl_policy(policy) && dl_param_changed(p, attr))
7014 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7017 p->sched_reset_on_fork = reset_on_fork;
7024 #ifdef CONFIG_RT_GROUP_SCHED
7026 * Do not allow realtime tasks into groups that have no runtime
7029 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7030 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7031 !task_group_is_autogroup(task_group(p))) {
7037 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7038 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7039 cpumask_t *span = rq->rd->span;
7042 * Don't allow tasks with an affinity mask smaller than
7043 * the entire root_domain to become SCHED_DEADLINE. We
7044 * will also fail if there's no bandwidth available.
7046 if (!cpumask_subset(span, p->cpus_ptr) ||
7047 rq->rd->dl_bw.bw == 0) {
7055 /* Re-check policy now with rq lock held: */
7056 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7057 policy = oldpolicy = -1;
7058 task_rq_unlock(rq, p, &rf);
7060 cpuset_read_unlock();
7065 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7066 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7069 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7074 p->sched_reset_on_fork = reset_on_fork;
7079 * Take priority boosted tasks into account. If the new
7080 * effective priority is unchanged, we just store the new
7081 * normal parameters and do not touch the scheduler class and
7082 * the runqueue. This will be done when the task deboost
7085 new_effective_prio = rt_effective_prio(p, newprio);
7086 if (new_effective_prio == oldprio)
7087 queue_flags &= ~DEQUEUE_MOVE;
7090 queued = task_on_rq_queued(p);
7091 running = task_current(rq, p);
7093 dequeue_task(rq, p, queue_flags);
7095 put_prev_task(rq, p);
7097 prev_class = p->sched_class;
7099 __setscheduler(rq, p, attr, pi);
7100 __setscheduler_uclamp(p, attr);
7104 * We enqueue to tail when the priority of a task is
7105 * increased (user space view).
7107 if (oldprio < p->prio)
7108 queue_flags |= ENQUEUE_HEAD;
7110 enqueue_task(rq, p, queue_flags);
7113 set_next_task(rq, p);
7115 check_class_changed(rq, p, prev_class, oldprio);
7117 /* Avoid rq from going away on us: */
7119 head = splice_balance_callbacks(rq);
7120 task_rq_unlock(rq, p, &rf);
7123 cpuset_read_unlock();
7124 rt_mutex_adjust_pi(p);
7127 /* Run balance callbacks after we've adjusted the PI chain: */
7128 balance_callbacks(rq, head);
7134 task_rq_unlock(rq, p, &rf);
7136 cpuset_read_unlock();
7140 static int _sched_setscheduler(struct task_struct *p, int policy,
7141 const struct sched_param *param, bool check)
7143 struct sched_attr attr = {
7144 .sched_policy = policy,
7145 .sched_priority = param->sched_priority,
7146 .sched_nice = PRIO_TO_NICE(p->static_prio),
7149 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7150 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7151 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7152 policy &= ~SCHED_RESET_ON_FORK;
7153 attr.sched_policy = policy;
7156 return __sched_setscheduler(p, &attr, check, true);
7159 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7160 * @p: the task in question.
7161 * @policy: new policy.
7162 * @param: structure containing the new RT priority.
7164 * Use sched_set_fifo(), read its comment.
7166 * Return: 0 on success. An error code otherwise.
7168 * NOTE that the task may be already dead.
7170 int sched_setscheduler(struct task_struct *p, int policy,
7171 const struct sched_param *param)
7173 return _sched_setscheduler(p, policy, param, true);
7176 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7178 return __sched_setscheduler(p, attr, true, true);
7181 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7183 return __sched_setscheduler(p, attr, false, true);
7185 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7188 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7189 * @p: the task in question.
7190 * @policy: new policy.
7191 * @param: structure containing the new RT priority.
7193 * Just like sched_setscheduler, only don't bother checking if the
7194 * current context has permission. For example, this is needed in
7195 * stop_machine(): we create temporary high priority worker threads,
7196 * but our caller might not have that capability.
7198 * Return: 0 on success. An error code otherwise.
7200 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7201 const struct sched_param *param)
7203 return _sched_setscheduler(p, policy, param, false);
7207 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7208 * incapable of resource management, which is the one thing an OS really should
7211 * This is of course the reason it is limited to privileged users only.
7213 * Worse still; it is fundamentally impossible to compose static priority
7214 * workloads. You cannot take two correctly working static prio workloads
7215 * and smash them together and still expect them to work.
7217 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7221 * The administrator _MUST_ configure the system, the kernel simply doesn't
7222 * know enough information to make a sensible choice.
7224 void sched_set_fifo(struct task_struct *p)
7226 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7227 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7229 EXPORT_SYMBOL_GPL(sched_set_fifo);
7232 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7234 void sched_set_fifo_low(struct task_struct *p)
7236 struct sched_param sp = { .sched_priority = 1 };
7237 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7239 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7241 void sched_set_normal(struct task_struct *p, int nice)
7243 struct sched_attr attr = {
7244 .sched_policy = SCHED_NORMAL,
7247 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7249 EXPORT_SYMBOL_GPL(sched_set_normal);
7252 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7254 struct sched_param lparam;
7255 struct task_struct *p;
7258 if (!param || pid < 0)
7260 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7265 p = find_process_by_pid(pid);
7271 retval = sched_setscheduler(p, policy, &lparam);
7279 * Mimics kernel/events/core.c perf_copy_attr().
7281 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7286 /* Zero the full structure, so that a short copy will be nice: */
7287 memset(attr, 0, sizeof(*attr));
7289 ret = get_user(size, &uattr->size);
7293 /* ABI compatibility quirk: */
7295 size = SCHED_ATTR_SIZE_VER0;
7296 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7299 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7306 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7307 size < SCHED_ATTR_SIZE_VER1)
7311 * XXX: Do we want to be lenient like existing syscalls; or do we want
7312 * to be strict and return an error on out-of-bounds values?
7314 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7319 put_user(sizeof(*attr), &uattr->size);
7324 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7325 * @pid: the pid in question.
7326 * @policy: new policy.
7327 * @param: structure containing the new RT priority.
7329 * Return: 0 on success. An error code otherwise.
7331 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7336 return do_sched_setscheduler(pid, policy, param);
7340 * sys_sched_setparam - set/change the RT priority of a thread
7341 * @pid: the pid in question.
7342 * @param: structure containing the new RT priority.
7344 * Return: 0 on success. An error code otherwise.
7346 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7348 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7352 * sys_sched_setattr - same as above, but with extended sched_attr
7353 * @pid: the pid in question.
7354 * @uattr: structure containing the extended parameters.
7355 * @flags: for future extension.
7357 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7358 unsigned int, flags)
7360 struct sched_attr attr;
7361 struct task_struct *p;
7364 if (!uattr || pid < 0 || flags)
7367 retval = sched_copy_attr(uattr, &attr);
7371 if ((int)attr.sched_policy < 0)
7373 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7374 attr.sched_policy = SETPARAM_POLICY;
7378 p = find_process_by_pid(pid);
7384 retval = sched_setattr(p, &attr);
7392 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7393 * @pid: the pid in question.
7395 * Return: On success, the policy of the thread. Otherwise, a negative error
7398 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7400 struct task_struct *p;
7408 p = find_process_by_pid(pid);
7410 retval = security_task_getscheduler(p);
7413 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7420 * sys_sched_getparam - get the RT priority of a thread
7421 * @pid: the pid in question.
7422 * @param: structure containing the RT priority.
7424 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7427 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7429 struct sched_param lp = { .sched_priority = 0 };
7430 struct task_struct *p;
7433 if (!param || pid < 0)
7437 p = find_process_by_pid(pid);
7442 retval = security_task_getscheduler(p);
7446 if (task_has_rt_policy(p))
7447 lp.sched_priority = p->rt_priority;
7451 * This one might sleep, we cannot do it with a spinlock held ...
7453 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7463 * Copy the kernel size attribute structure (which might be larger
7464 * than what user-space knows about) to user-space.
7466 * Note that all cases are valid: user-space buffer can be larger or
7467 * smaller than the kernel-space buffer. The usual case is that both
7468 * have the same size.
7471 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7472 struct sched_attr *kattr,
7475 unsigned int ksize = sizeof(*kattr);
7477 if (!access_ok(uattr, usize))
7481 * sched_getattr() ABI forwards and backwards compatibility:
7483 * If usize == ksize then we just copy everything to user-space and all is good.
7485 * If usize < ksize then we only copy as much as user-space has space for,
7486 * this keeps ABI compatibility as well. We skip the rest.
7488 * If usize > ksize then user-space is using a newer version of the ABI,
7489 * which part the kernel doesn't know about. Just ignore it - tooling can
7490 * detect the kernel's knowledge of attributes from the attr->size value
7491 * which is set to ksize in this case.
7493 kattr->size = min(usize, ksize);
7495 if (copy_to_user(uattr, kattr, kattr->size))
7502 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7503 * @pid: the pid in question.
7504 * @uattr: structure containing the extended parameters.
7505 * @usize: sizeof(attr) for fwd/bwd comp.
7506 * @flags: for future extension.
7508 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7509 unsigned int, usize, unsigned int, flags)
7511 struct sched_attr kattr = { };
7512 struct task_struct *p;
7515 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7516 usize < SCHED_ATTR_SIZE_VER0 || flags)
7520 p = find_process_by_pid(pid);
7525 retval = security_task_getscheduler(p);
7529 kattr.sched_policy = p->policy;
7530 if (p->sched_reset_on_fork)
7531 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7532 if (task_has_dl_policy(p))
7533 __getparam_dl(p, &kattr);
7534 else if (task_has_rt_policy(p))
7535 kattr.sched_priority = p->rt_priority;
7537 kattr.sched_nice = task_nice(p);
7539 #ifdef CONFIG_UCLAMP_TASK
7541 * This could race with another potential updater, but this is fine
7542 * because it'll correctly read the old or the new value. We don't need
7543 * to guarantee who wins the race as long as it doesn't return garbage.
7545 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7546 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7551 return sched_attr_copy_to_user(uattr, &kattr, usize);
7558 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7560 cpumask_var_t cpus_allowed, new_mask;
7561 struct task_struct *p;
7566 p = find_process_by_pid(pid);
7572 /* Prevent p going away */
7576 if (p->flags & PF_NO_SETAFFINITY) {
7580 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
7584 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7586 goto out_free_cpus_allowed;
7589 if (!check_same_owner(p)) {
7591 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7593 goto out_free_new_mask;
7598 retval = security_task_setscheduler(p);
7600 goto out_free_new_mask;
7603 cpuset_cpus_allowed(p, cpus_allowed);
7604 cpumask_and(new_mask, in_mask, cpus_allowed);
7607 * Since bandwidth control happens on root_domain basis,
7608 * if admission test is enabled, we only admit -deadline
7609 * tasks allowed to run on all the CPUs in the task's
7613 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
7615 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
7618 goto out_free_new_mask;
7624 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
7627 cpuset_cpus_allowed(p, cpus_allowed);
7628 if (!cpumask_subset(new_mask, cpus_allowed)) {
7630 * We must have raced with a concurrent cpuset
7631 * update. Just reset the cpus_allowed to the
7632 * cpuset's cpus_allowed
7634 cpumask_copy(new_mask, cpus_allowed);
7639 free_cpumask_var(new_mask);
7640 out_free_cpus_allowed:
7641 free_cpumask_var(cpus_allowed);
7647 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7648 struct cpumask *new_mask)
7650 if (len < cpumask_size())
7651 cpumask_clear(new_mask);
7652 else if (len > cpumask_size())
7653 len = cpumask_size();
7655 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
7659 * sys_sched_setaffinity - set the CPU affinity of a process
7660 * @pid: pid of the process
7661 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7662 * @user_mask_ptr: user-space pointer to the new CPU mask
7664 * Return: 0 on success. An error code otherwise.
7666 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
7667 unsigned long __user *, user_mask_ptr)
7669 cpumask_var_t new_mask;
7672 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
7675 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
7677 retval = sched_setaffinity(pid, new_mask);
7678 free_cpumask_var(new_mask);
7682 long sched_getaffinity(pid_t pid, struct cpumask *mask)
7684 struct task_struct *p;
7685 unsigned long flags;
7691 p = find_process_by_pid(pid);
7695 retval = security_task_getscheduler(p);
7699 raw_spin_lock_irqsave(&p->pi_lock, flags);
7700 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
7701 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
7710 * sys_sched_getaffinity - get the CPU affinity of a process
7711 * @pid: pid of the process
7712 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7713 * @user_mask_ptr: user-space pointer to hold the current CPU mask
7715 * Return: size of CPU mask copied to user_mask_ptr on success. An
7716 * error code otherwise.
7718 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
7719 unsigned long __user *, user_mask_ptr)
7724 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
7726 if (len & (sizeof(unsigned long)-1))
7729 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
7732 ret = sched_getaffinity(pid, mask);
7734 unsigned int retlen = min(len, cpumask_size());
7736 if (copy_to_user(user_mask_ptr, mask, retlen))
7741 free_cpumask_var(mask);
7746 static void do_sched_yield(void)
7751 rq = this_rq_lock_irq(&rf);
7753 schedstat_inc(rq->yld_count);
7754 current->sched_class->yield_task(rq);
7757 rq_unlock_irq(rq, &rf);
7758 sched_preempt_enable_no_resched();
7764 * sys_sched_yield - yield the current processor to other threads.
7766 * This function yields the current CPU to other tasks. If there are no
7767 * other threads running on this CPU then this function will return.
7771 SYSCALL_DEFINE0(sched_yield)
7777 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
7778 int __sched __cond_resched(void)
7780 if (should_resched(0)) {
7781 preempt_schedule_common();
7784 #ifndef CONFIG_PREEMPT_RCU
7789 EXPORT_SYMBOL(__cond_resched);
7792 #ifdef CONFIG_PREEMPT_DYNAMIC
7793 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
7794 EXPORT_STATIC_CALL_TRAMP(cond_resched);
7796 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
7797 EXPORT_STATIC_CALL_TRAMP(might_resched);
7801 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7802 * call schedule, and on return reacquire the lock.
7804 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7805 * operations here to prevent schedule() from being called twice (once via
7806 * spin_unlock(), once by hand).
7808 int __cond_resched_lock(spinlock_t *lock)
7810 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7813 lockdep_assert_held(lock);
7815 if (spin_needbreak(lock) || resched) {
7818 preempt_schedule_common();
7826 EXPORT_SYMBOL(__cond_resched_lock);
7828 int __cond_resched_rwlock_read(rwlock_t *lock)
7830 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7833 lockdep_assert_held_read(lock);
7835 if (rwlock_needbreak(lock) || resched) {
7838 preempt_schedule_common();
7846 EXPORT_SYMBOL(__cond_resched_rwlock_read);
7848 int __cond_resched_rwlock_write(rwlock_t *lock)
7850 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7853 lockdep_assert_held_write(lock);
7855 if (rwlock_needbreak(lock) || resched) {
7858 preempt_schedule_common();
7866 EXPORT_SYMBOL(__cond_resched_rwlock_write);
7869 * yield - yield the current processor to other threads.
7871 * Do not ever use this function, there's a 99% chance you're doing it wrong.
7873 * The scheduler is at all times free to pick the calling task as the most
7874 * eligible task to run, if removing the yield() call from your code breaks
7875 * it, it's already broken.
7877 * Typical broken usage is:
7882 * where one assumes that yield() will let 'the other' process run that will
7883 * make event true. If the current task is a SCHED_FIFO task that will never
7884 * happen. Never use yield() as a progress guarantee!!
7886 * If you want to use yield() to wait for something, use wait_event().
7887 * If you want to use yield() to be 'nice' for others, use cond_resched().
7888 * If you still want to use yield(), do not!
7890 void __sched yield(void)
7892 set_current_state(TASK_RUNNING);
7895 EXPORT_SYMBOL(yield);
7898 * yield_to - yield the current processor to another thread in
7899 * your thread group, or accelerate that thread toward the
7900 * processor it's on.
7902 * @preempt: whether task preemption is allowed or not
7904 * It's the caller's job to ensure that the target task struct
7905 * can't go away on us before we can do any checks.
7908 * true (>0) if we indeed boosted the target task.
7909 * false (0) if we failed to boost the target.
7910 * -ESRCH if there's no task to yield to.
7912 int __sched yield_to(struct task_struct *p, bool preempt)
7914 struct task_struct *curr = current;
7915 struct rq *rq, *p_rq;
7916 unsigned long flags;
7919 local_irq_save(flags);
7925 * If we're the only runnable task on the rq and target rq also
7926 * has only one task, there's absolutely no point in yielding.
7928 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
7933 double_rq_lock(rq, p_rq);
7934 if (task_rq(p) != p_rq) {
7935 double_rq_unlock(rq, p_rq);
7939 if (!curr->sched_class->yield_to_task)
7942 if (curr->sched_class != p->sched_class)
7945 if (task_running(p_rq, p) || !task_is_running(p))
7948 yielded = curr->sched_class->yield_to_task(rq, p);
7950 schedstat_inc(rq->yld_count);
7952 * Make p's CPU reschedule; pick_next_entity takes care of
7955 if (preempt && rq != p_rq)
7960 double_rq_unlock(rq, p_rq);
7962 local_irq_restore(flags);
7969 EXPORT_SYMBOL_GPL(yield_to);
7971 int io_schedule_prepare(void)
7973 int old_iowait = current->in_iowait;
7975 current->in_iowait = 1;
7976 blk_schedule_flush_plug(current);
7981 void io_schedule_finish(int token)
7983 current->in_iowait = token;
7987 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7988 * that process accounting knows that this is a task in IO wait state.
7990 long __sched io_schedule_timeout(long timeout)
7995 token = io_schedule_prepare();
7996 ret = schedule_timeout(timeout);
7997 io_schedule_finish(token);
8001 EXPORT_SYMBOL(io_schedule_timeout);
8003 void __sched io_schedule(void)
8007 token = io_schedule_prepare();
8009 io_schedule_finish(token);
8011 EXPORT_SYMBOL(io_schedule);
8014 * sys_sched_get_priority_max - return maximum RT priority.
8015 * @policy: scheduling class.
8017 * Return: On success, this syscall returns the maximum
8018 * rt_priority that can be used by a given scheduling class.
8019 * On failure, a negative error code is returned.
8021 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8028 ret = MAX_RT_PRIO-1;
8030 case SCHED_DEADLINE:
8041 * sys_sched_get_priority_min - return minimum RT priority.
8042 * @policy: scheduling class.
8044 * Return: On success, this syscall returns the minimum
8045 * rt_priority that can be used by a given scheduling class.
8046 * On failure, a negative error code is returned.
8048 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8057 case SCHED_DEADLINE:
8066 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8068 struct task_struct *p;
8069 unsigned int time_slice;
8079 p = find_process_by_pid(pid);
8083 retval = security_task_getscheduler(p);
8087 rq = task_rq_lock(p, &rf);
8089 if (p->sched_class->get_rr_interval)
8090 time_slice = p->sched_class->get_rr_interval(rq, p);
8091 task_rq_unlock(rq, p, &rf);
8094 jiffies_to_timespec64(time_slice, t);
8103 * sys_sched_rr_get_interval - return the default timeslice of a process.
8104 * @pid: pid of the process.
8105 * @interval: userspace pointer to the timeslice value.
8107 * this syscall writes the default timeslice value of a given process
8108 * into the user-space timespec buffer. A value of '0' means infinity.
8110 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8113 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8114 struct __kernel_timespec __user *, interval)
8116 struct timespec64 t;
8117 int retval = sched_rr_get_interval(pid, &t);
8120 retval = put_timespec64(&t, interval);
8125 #ifdef CONFIG_COMPAT_32BIT_TIME
8126 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8127 struct old_timespec32 __user *, interval)
8129 struct timespec64 t;
8130 int retval = sched_rr_get_interval(pid, &t);
8133 retval = put_old_timespec32(&t, interval);
8138 void sched_show_task(struct task_struct *p)
8140 unsigned long free = 0;
8143 if (!try_get_task_stack(p))
8146 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8148 if (task_is_running(p))
8149 pr_cont(" running task ");
8150 #ifdef CONFIG_DEBUG_STACK_USAGE
8151 free = stack_not_used(p);
8156 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8158 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8159 free, task_pid_nr(p), ppid,
8160 (unsigned long)task_thread_info(p)->flags);
8162 print_worker_info(KERN_INFO, p);
8163 print_stop_info(KERN_INFO, p);
8164 show_stack(p, NULL, KERN_INFO);
8167 EXPORT_SYMBOL_GPL(sched_show_task);
8170 state_filter_match(unsigned long state_filter, struct task_struct *p)
8172 unsigned int state = READ_ONCE(p->__state);
8174 /* no filter, everything matches */
8178 /* filter, but doesn't match */
8179 if (!(state & state_filter))
8183 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8186 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8193 void show_state_filter(unsigned int state_filter)
8195 struct task_struct *g, *p;
8198 for_each_process_thread(g, p) {
8200 * reset the NMI-timeout, listing all files on a slow
8201 * console might take a lot of time:
8202 * Also, reset softlockup watchdogs on all CPUs, because
8203 * another CPU might be blocked waiting for us to process
8206 touch_nmi_watchdog();
8207 touch_all_softlockup_watchdogs();
8208 if (state_filter_match(state_filter, p))
8212 #ifdef CONFIG_SCHED_DEBUG
8214 sysrq_sched_debug_show();
8218 * Only show locks if all tasks are dumped:
8221 debug_show_all_locks();
8225 * init_idle - set up an idle thread for a given CPU
8226 * @idle: task in question
8227 * @cpu: CPU the idle task belongs to
8229 * NOTE: this function does not set the idle thread's NEED_RESCHED
8230 * flag, to make booting more robust.
8232 void __init init_idle(struct task_struct *idle, int cpu)
8234 struct rq *rq = cpu_rq(cpu);
8235 unsigned long flags;
8237 __sched_fork(0, idle);
8240 * The idle task doesn't need the kthread struct to function, but it
8241 * is dressed up as a per-CPU kthread and thus needs to play the part
8242 * if we want to avoid special-casing it in code that deals with per-CPU
8245 set_kthread_struct(idle);
8247 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8248 raw_spin_rq_lock(rq);
8250 idle->__state = TASK_RUNNING;
8251 idle->se.exec_start = sched_clock();
8253 * PF_KTHREAD should already be set at this point; regardless, make it
8254 * look like a proper per-CPU kthread.
8256 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8257 kthread_set_per_cpu(idle, cpu);
8259 scs_task_reset(idle);
8260 kasan_unpoison_task_stack(idle);
8264 * It's possible that init_idle() gets called multiple times on a task,
8265 * in that case do_set_cpus_allowed() will not do the right thing.
8267 * And since this is boot we can forgo the serialization.
8269 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8272 * We're having a chicken and egg problem, even though we are
8273 * holding rq->lock, the CPU isn't yet set to this CPU so the
8274 * lockdep check in task_group() will fail.
8276 * Similar case to sched_fork(). / Alternatively we could
8277 * use task_rq_lock() here and obtain the other rq->lock.
8282 __set_task_cpu(idle, cpu);
8286 rcu_assign_pointer(rq->curr, idle);
8287 idle->on_rq = TASK_ON_RQ_QUEUED;
8291 raw_spin_rq_unlock(rq);
8292 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8294 /* Set the preempt count _outside_ the spinlocks! */
8295 init_idle_preempt_count(idle, cpu);
8298 * The idle tasks have their own, simple scheduling class:
8300 idle->sched_class = &idle_sched_class;
8301 ftrace_graph_init_idle_task(idle, cpu);
8302 vtime_init_idle(idle, cpu);
8304 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8310 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8311 const struct cpumask *trial)
8315 if (!cpumask_weight(cur))
8318 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8323 int task_can_attach(struct task_struct *p,
8324 const struct cpumask *cs_cpus_allowed)
8329 * Kthreads which disallow setaffinity shouldn't be moved
8330 * to a new cpuset; we don't want to change their CPU
8331 * affinity and isolating such threads by their set of
8332 * allowed nodes is unnecessary. Thus, cpusets are not
8333 * applicable for such threads. This prevents checking for
8334 * success of set_cpus_allowed_ptr() on all attached tasks
8335 * before cpus_mask may be changed.
8337 if (p->flags & PF_NO_SETAFFINITY) {
8342 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8344 ret = dl_task_can_attach(p, cs_cpus_allowed);
8350 bool sched_smp_initialized __read_mostly;
8352 #ifdef CONFIG_NUMA_BALANCING
8353 /* Migrate current task p to target_cpu */
8354 int migrate_task_to(struct task_struct *p, int target_cpu)
8356 struct migration_arg arg = { p, target_cpu };
8357 int curr_cpu = task_cpu(p);
8359 if (curr_cpu == target_cpu)
8362 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8365 /* TODO: This is not properly updating schedstats */
8367 trace_sched_move_numa(p, curr_cpu, target_cpu);
8368 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8372 * Requeue a task on a given node and accurately track the number of NUMA
8373 * tasks on the runqueues
8375 void sched_setnuma(struct task_struct *p, int nid)
8377 bool queued, running;
8381 rq = task_rq_lock(p, &rf);
8382 queued = task_on_rq_queued(p);
8383 running = task_current(rq, p);
8386 dequeue_task(rq, p, DEQUEUE_SAVE);
8388 put_prev_task(rq, p);
8390 p->numa_preferred_nid = nid;
8393 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8395 set_next_task(rq, p);
8396 task_rq_unlock(rq, p, &rf);
8398 #endif /* CONFIG_NUMA_BALANCING */
8400 #ifdef CONFIG_HOTPLUG_CPU
8402 * Ensure that the idle task is using init_mm right before its CPU goes
8405 void idle_task_exit(void)
8407 struct mm_struct *mm = current->active_mm;
8409 BUG_ON(cpu_online(smp_processor_id()));
8410 BUG_ON(current != this_rq()->idle);
8412 if (mm != &init_mm) {
8413 switch_mm(mm, &init_mm, current);
8414 finish_arch_post_lock_switch();
8417 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8420 static int __balance_push_cpu_stop(void *arg)
8422 struct task_struct *p = arg;
8423 struct rq *rq = this_rq();
8427 raw_spin_lock_irq(&p->pi_lock);
8430 update_rq_clock(rq);
8432 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8433 cpu = select_fallback_rq(rq->cpu, p);
8434 rq = __migrate_task(rq, &rf, p, cpu);
8438 raw_spin_unlock_irq(&p->pi_lock);
8445 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8448 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8450 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8451 * effective when the hotplug motion is down.
8453 static void balance_push(struct rq *rq)
8455 struct task_struct *push_task = rq->curr;
8457 lockdep_assert_rq_held(rq);
8458 SCHED_WARN_ON(rq->cpu != smp_processor_id());
8461 * Ensure the thing is persistent until balance_push_set(.on = false);
8463 rq->balance_callback = &balance_push_callback;
8466 * Only active while going offline.
8468 if (!cpu_dying(rq->cpu))
8472 * Both the cpu-hotplug and stop task are in this case and are
8473 * required to complete the hotplug process.
8475 if (kthread_is_per_cpu(push_task) ||
8476 is_migration_disabled(push_task)) {
8479 * If this is the idle task on the outgoing CPU try to wake
8480 * up the hotplug control thread which might wait for the
8481 * last task to vanish. The rcuwait_active() check is
8482 * accurate here because the waiter is pinned on this CPU
8483 * and can't obviously be running in parallel.
8485 * On RT kernels this also has to check whether there are
8486 * pinned and scheduled out tasks on the runqueue. They
8487 * need to leave the migrate disabled section first.
8489 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8490 rcuwait_active(&rq->hotplug_wait)) {
8491 raw_spin_rq_unlock(rq);
8492 rcuwait_wake_up(&rq->hotplug_wait);
8493 raw_spin_rq_lock(rq);
8498 get_task_struct(push_task);
8500 * Temporarily drop rq->lock such that we can wake-up the stop task.
8501 * Both preemption and IRQs are still disabled.
8503 raw_spin_rq_unlock(rq);
8504 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8505 this_cpu_ptr(&push_work));
8507 * At this point need_resched() is true and we'll take the loop in
8508 * schedule(). The next pick is obviously going to be the stop task
8509 * which kthread_is_per_cpu() and will push this task away.
8511 raw_spin_rq_lock(rq);
8514 static void balance_push_set(int cpu, bool on)
8516 struct rq *rq = cpu_rq(cpu);
8519 rq_lock_irqsave(rq, &rf);
8521 WARN_ON_ONCE(rq->balance_callback);
8522 rq->balance_callback = &balance_push_callback;
8523 } else if (rq->balance_callback == &balance_push_callback) {
8524 rq->balance_callback = NULL;
8526 rq_unlock_irqrestore(rq, &rf);
8530 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8531 * inactive. All tasks which are not per CPU kernel threads are either
8532 * pushed off this CPU now via balance_push() or placed on a different CPU
8533 * during wakeup. Wait until the CPU is quiescent.
8535 static void balance_hotplug_wait(void)
8537 struct rq *rq = this_rq();
8539 rcuwait_wait_event(&rq->hotplug_wait,
8540 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8541 TASK_UNINTERRUPTIBLE);
8546 static inline void balance_push(struct rq *rq)
8550 static inline void balance_push_set(int cpu, bool on)
8554 static inline void balance_hotplug_wait(void)
8558 #endif /* CONFIG_HOTPLUG_CPU */
8560 void set_rq_online(struct rq *rq)
8563 const struct sched_class *class;
8565 cpumask_set_cpu(rq->cpu, rq->rd->online);
8568 for_each_class(class) {
8569 if (class->rq_online)
8570 class->rq_online(rq);
8575 void set_rq_offline(struct rq *rq)
8578 const struct sched_class *class;
8580 for_each_class(class) {
8581 if (class->rq_offline)
8582 class->rq_offline(rq);
8585 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8591 * used to mark begin/end of suspend/resume:
8593 static int num_cpus_frozen;
8596 * Update cpusets according to cpu_active mask. If cpusets are
8597 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8598 * around partition_sched_domains().
8600 * If we come here as part of a suspend/resume, don't touch cpusets because we
8601 * want to restore it back to its original state upon resume anyway.
8603 static void cpuset_cpu_active(void)
8605 if (cpuhp_tasks_frozen) {
8607 * num_cpus_frozen tracks how many CPUs are involved in suspend
8608 * resume sequence. As long as this is not the last online
8609 * operation in the resume sequence, just build a single sched
8610 * domain, ignoring cpusets.
8612 partition_sched_domains(1, NULL, NULL);
8613 if (--num_cpus_frozen)
8616 * This is the last CPU online operation. So fall through and
8617 * restore the original sched domains by considering the
8618 * cpuset configurations.
8620 cpuset_force_rebuild();
8622 cpuset_update_active_cpus();
8625 static int cpuset_cpu_inactive(unsigned int cpu)
8627 if (!cpuhp_tasks_frozen) {
8628 if (dl_cpu_busy(cpu))
8630 cpuset_update_active_cpus();
8633 partition_sched_domains(1, NULL, NULL);
8638 int sched_cpu_activate(unsigned int cpu)
8640 struct rq *rq = cpu_rq(cpu);
8644 * Clear the balance_push callback and prepare to schedule
8647 balance_push_set(cpu, false);
8649 #ifdef CONFIG_SCHED_SMT
8651 * When going up, increment the number of cores with SMT present.
8653 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8654 static_branch_inc_cpuslocked(&sched_smt_present);
8656 set_cpu_active(cpu, true);
8658 if (sched_smp_initialized) {
8659 sched_domains_numa_masks_set(cpu);
8660 cpuset_cpu_active();
8664 * Put the rq online, if not already. This happens:
8666 * 1) In the early boot process, because we build the real domains
8667 * after all CPUs have been brought up.
8669 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
8672 rq_lock_irqsave(rq, &rf);
8674 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8677 rq_unlock_irqrestore(rq, &rf);
8682 int sched_cpu_deactivate(unsigned int cpu)
8684 struct rq *rq = cpu_rq(cpu);
8689 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
8690 * load balancing when not active
8692 nohz_balance_exit_idle(rq);
8694 set_cpu_active(cpu, false);
8697 * From this point forward, this CPU will refuse to run any task that
8698 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
8699 * push those tasks away until this gets cleared, see
8700 * sched_cpu_dying().
8702 balance_push_set(cpu, true);
8705 * We've cleared cpu_active_mask / set balance_push, wait for all
8706 * preempt-disabled and RCU users of this state to go away such that
8707 * all new such users will observe it.
8709 * Specifically, we rely on ttwu to no longer target this CPU, see
8710 * ttwu_queue_cond() and is_cpu_allowed().
8712 * Do sync before park smpboot threads to take care the rcu boost case.
8716 rq_lock_irqsave(rq, &rf);
8718 update_rq_clock(rq);
8719 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8722 rq_unlock_irqrestore(rq, &rf);
8724 #ifdef CONFIG_SCHED_SMT
8726 * When going down, decrement the number of cores with SMT present.
8728 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8729 static_branch_dec_cpuslocked(&sched_smt_present);
8732 if (!sched_smp_initialized)
8735 ret = cpuset_cpu_inactive(cpu);
8737 balance_push_set(cpu, false);
8738 set_cpu_active(cpu, true);
8741 sched_domains_numa_masks_clear(cpu);
8745 static void sched_rq_cpu_starting(unsigned int cpu)
8747 struct rq *rq = cpu_rq(cpu);
8749 rq->calc_load_update = calc_load_update;
8750 update_max_interval();
8753 int sched_cpu_starting(unsigned int cpu)
8755 sched_core_cpu_starting(cpu);
8756 sched_rq_cpu_starting(cpu);
8757 sched_tick_start(cpu);
8761 #ifdef CONFIG_HOTPLUG_CPU
8764 * Invoked immediately before the stopper thread is invoked to bring the
8765 * CPU down completely. At this point all per CPU kthreads except the
8766 * hotplug thread (current) and the stopper thread (inactive) have been
8767 * either parked or have been unbound from the outgoing CPU. Ensure that
8768 * any of those which might be on the way out are gone.
8770 * If after this point a bound task is being woken on this CPU then the
8771 * responsible hotplug callback has failed to do it's job.
8772 * sched_cpu_dying() will catch it with the appropriate fireworks.
8774 int sched_cpu_wait_empty(unsigned int cpu)
8776 balance_hotplug_wait();
8781 * Since this CPU is going 'away' for a while, fold any nr_active delta we
8782 * might have. Called from the CPU stopper task after ensuring that the
8783 * stopper is the last running task on the CPU, so nr_active count is
8784 * stable. We need to take the teardown thread which is calling this into
8785 * account, so we hand in adjust = 1 to the load calculation.
8787 * Also see the comment "Global load-average calculations".
8789 static void calc_load_migrate(struct rq *rq)
8791 long delta = calc_load_fold_active(rq, 1);
8794 atomic_long_add(delta, &calc_load_tasks);
8797 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
8799 struct task_struct *g, *p;
8800 int cpu = cpu_of(rq);
8802 lockdep_assert_rq_held(rq);
8804 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
8805 for_each_process_thread(g, p) {
8806 if (task_cpu(p) != cpu)
8809 if (!task_on_rq_queued(p))
8812 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
8816 int sched_cpu_dying(unsigned int cpu)
8818 struct rq *rq = cpu_rq(cpu);
8821 /* Handle pending wakeups and then migrate everything off */
8822 sched_tick_stop(cpu);
8824 rq_lock_irqsave(rq, &rf);
8825 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
8826 WARN(true, "Dying CPU not properly vacated!");
8827 dump_rq_tasks(rq, KERN_WARNING);
8829 rq_unlock_irqrestore(rq, &rf);
8831 calc_load_migrate(rq);
8832 update_max_interval();
8838 void __init sched_init_smp(void)
8843 * There's no userspace yet to cause hotplug operations; hence all the
8844 * CPU masks are stable and all blatant races in the below code cannot
8847 mutex_lock(&sched_domains_mutex);
8848 sched_init_domains(cpu_active_mask);
8849 mutex_unlock(&sched_domains_mutex);
8851 /* Move init over to a non-isolated CPU */
8852 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
8854 current->flags &= ~PF_NO_SETAFFINITY;
8855 sched_init_granularity();
8857 init_sched_rt_class();
8858 init_sched_dl_class();
8860 sched_smp_initialized = true;
8863 static int __init migration_init(void)
8865 sched_cpu_starting(smp_processor_id());
8868 early_initcall(migration_init);
8871 void __init sched_init_smp(void)
8873 sched_init_granularity();
8875 #endif /* CONFIG_SMP */
8877 int in_sched_functions(unsigned long addr)
8879 return in_lock_functions(addr) ||
8880 (addr >= (unsigned long)__sched_text_start
8881 && addr < (unsigned long)__sched_text_end);
8884 #ifdef CONFIG_CGROUP_SCHED
8886 * Default task group.
8887 * Every task in system belongs to this group at bootup.
8889 struct task_group root_task_group;
8890 LIST_HEAD(task_groups);
8892 /* Cacheline aligned slab cache for task_group */
8893 static struct kmem_cache *task_group_cache __read_mostly;
8896 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
8897 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
8899 void __init sched_init(void)
8901 unsigned long ptr = 0;
8904 /* Make sure the linker didn't screw up */
8905 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
8906 &fair_sched_class + 1 != &rt_sched_class ||
8907 &rt_sched_class + 1 != &dl_sched_class);
8909 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
8914 #ifdef CONFIG_FAIR_GROUP_SCHED
8915 ptr += 2 * nr_cpu_ids * sizeof(void **);
8917 #ifdef CONFIG_RT_GROUP_SCHED
8918 ptr += 2 * nr_cpu_ids * sizeof(void **);
8921 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
8923 #ifdef CONFIG_FAIR_GROUP_SCHED
8924 root_task_group.se = (struct sched_entity **)ptr;
8925 ptr += nr_cpu_ids * sizeof(void **);
8927 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8928 ptr += nr_cpu_ids * sizeof(void **);
8930 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
8931 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8932 #endif /* CONFIG_FAIR_GROUP_SCHED */
8933 #ifdef CONFIG_RT_GROUP_SCHED
8934 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8935 ptr += nr_cpu_ids * sizeof(void **);
8937 root_task_group.rt_rq = (struct rt_rq **)ptr;
8938 ptr += nr_cpu_ids * sizeof(void **);
8940 #endif /* CONFIG_RT_GROUP_SCHED */
8942 #ifdef CONFIG_CPUMASK_OFFSTACK
8943 for_each_possible_cpu(i) {
8944 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
8945 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8946 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
8947 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8949 #endif /* CONFIG_CPUMASK_OFFSTACK */
8951 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
8952 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
8955 init_defrootdomain();
8958 #ifdef CONFIG_RT_GROUP_SCHED
8959 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8960 global_rt_period(), global_rt_runtime());
8961 #endif /* CONFIG_RT_GROUP_SCHED */
8963 #ifdef CONFIG_CGROUP_SCHED
8964 task_group_cache = KMEM_CACHE(task_group, 0);
8966 list_add(&root_task_group.list, &task_groups);
8967 INIT_LIST_HEAD(&root_task_group.children);
8968 INIT_LIST_HEAD(&root_task_group.siblings);
8969 autogroup_init(&init_task);
8970 #endif /* CONFIG_CGROUP_SCHED */
8972 for_each_possible_cpu(i) {
8976 raw_spin_lock_init(&rq->__lock);
8978 rq->calc_load_active = 0;
8979 rq->calc_load_update = jiffies + LOAD_FREQ;
8980 init_cfs_rq(&rq->cfs);
8981 init_rt_rq(&rq->rt);
8982 init_dl_rq(&rq->dl);
8983 #ifdef CONFIG_FAIR_GROUP_SCHED
8984 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8985 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
8987 * How much CPU bandwidth does root_task_group get?
8989 * In case of task-groups formed thr' the cgroup filesystem, it
8990 * gets 100% of the CPU resources in the system. This overall
8991 * system CPU resource is divided among the tasks of
8992 * root_task_group and its child task-groups in a fair manner,
8993 * based on each entity's (task or task-group's) weight
8994 * (se->load.weight).
8996 * In other words, if root_task_group has 10 tasks of weight
8997 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8998 * then A0's share of the CPU resource is:
9000 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9002 * We achieve this by letting root_task_group's tasks sit
9003 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9005 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9006 #endif /* CONFIG_FAIR_GROUP_SCHED */
9008 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9009 #ifdef CONFIG_RT_GROUP_SCHED
9010 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9015 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9016 rq->balance_callback = &balance_push_callback;
9017 rq->active_balance = 0;
9018 rq->next_balance = jiffies;
9023 rq->avg_idle = 2*sysctl_sched_migration_cost;
9024 rq->wake_stamp = jiffies;
9025 rq->wake_avg_idle = rq->avg_idle;
9026 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9028 INIT_LIST_HEAD(&rq->cfs_tasks);
9030 rq_attach_root(rq, &def_root_domain);
9031 #ifdef CONFIG_NO_HZ_COMMON
9032 rq->last_blocked_load_update_tick = jiffies;
9033 atomic_set(&rq->nohz_flags, 0);
9035 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9037 #ifdef CONFIG_HOTPLUG_CPU
9038 rcuwait_init(&rq->hotplug_wait);
9040 #endif /* CONFIG_SMP */
9042 atomic_set(&rq->nr_iowait, 0);
9044 #ifdef CONFIG_SCHED_CORE
9046 rq->core_pick = NULL;
9047 rq->core_enabled = 0;
9048 rq->core_tree = RB_ROOT;
9049 rq->core_forceidle = false;
9051 rq->core_cookie = 0UL;
9055 set_load_weight(&init_task, false);
9058 * The boot idle thread does lazy MMU switching as well:
9061 enter_lazy_tlb(&init_mm, current);
9064 * Make us the idle thread. Technically, schedule() should not be
9065 * called from this thread, however somewhere below it might be,
9066 * but because we are the idle thread, we just pick up running again
9067 * when this runqueue becomes "idle".
9069 init_idle(current, smp_processor_id());
9071 calc_load_update = jiffies + LOAD_FREQ;
9074 idle_thread_set_boot_cpu();
9075 balance_push_set(smp_processor_id(), false);
9077 init_sched_fair_class();
9083 scheduler_running = 1;
9086 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9087 static inline int preempt_count_equals(int preempt_offset)
9089 int nested = preempt_count() + rcu_preempt_depth();
9091 return (nested == preempt_offset);
9094 void __might_sleep(const char *file, int line, int preempt_offset)
9096 unsigned int state = get_current_state();
9098 * Blocking primitives will set (and therefore destroy) current->state,
9099 * since we will exit with TASK_RUNNING make sure we enter with it,
9100 * otherwise we will destroy state.
9102 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9103 "do not call blocking ops when !TASK_RUNNING; "
9104 "state=%x set at [<%p>] %pS\n", state,
9105 (void *)current->task_state_change,
9106 (void *)current->task_state_change);
9108 ___might_sleep(file, line, preempt_offset);
9110 EXPORT_SYMBOL(__might_sleep);
9112 void ___might_sleep(const char *file, int line, int preempt_offset)
9114 /* Ratelimiting timestamp: */
9115 static unsigned long prev_jiffy;
9117 unsigned long preempt_disable_ip;
9119 /* WARN_ON_ONCE() by default, no rate limit required: */
9122 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
9123 !is_idle_task(current) && !current->non_block_count) ||
9124 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9128 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9130 prev_jiffy = jiffies;
9132 /* Save this before calling printk(), since that will clobber it: */
9133 preempt_disable_ip = get_preempt_disable_ip(current);
9136 "BUG: sleeping function called from invalid context at %s:%d\n",
9139 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9140 in_atomic(), irqs_disabled(), current->non_block_count,
9141 current->pid, current->comm);
9143 if (task_stack_end_corrupted(current))
9144 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
9146 debug_show_held_locks(current);
9147 if (irqs_disabled())
9148 print_irqtrace_events(current);
9149 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
9150 && !preempt_count_equals(preempt_offset)) {
9151 pr_err("Preemption disabled at:");
9152 print_ip_sym(KERN_ERR, preempt_disable_ip);
9155 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9157 EXPORT_SYMBOL(___might_sleep);
9159 void __cant_sleep(const char *file, int line, int preempt_offset)
9161 static unsigned long prev_jiffy;
9163 if (irqs_disabled())
9166 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9169 if (preempt_count() > preempt_offset)
9172 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9174 prev_jiffy = jiffies;
9176 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9177 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9178 in_atomic(), irqs_disabled(),
9179 current->pid, current->comm);
9181 debug_show_held_locks(current);
9183 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9185 EXPORT_SYMBOL_GPL(__cant_sleep);
9188 void __cant_migrate(const char *file, int line)
9190 static unsigned long prev_jiffy;
9192 if (irqs_disabled())
9195 if (is_migration_disabled(current))
9198 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9201 if (preempt_count() > 0)
9204 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9206 prev_jiffy = jiffies;
9208 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9209 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9210 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9211 current->pid, current->comm);
9213 debug_show_held_locks(current);
9215 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9217 EXPORT_SYMBOL_GPL(__cant_migrate);
9221 #ifdef CONFIG_MAGIC_SYSRQ
9222 void normalize_rt_tasks(void)
9224 struct task_struct *g, *p;
9225 struct sched_attr attr = {
9226 .sched_policy = SCHED_NORMAL,
9229 read_lock(&tasklist_lock);
9230 for_each_process_thread(g, p) {
9232 * Only normalize user tasks:
9234 if (p->flags & PF_KTHREAD)
9237 p->se.exec_start = 0;
9238 schedstat_set(p->se.statistics.wait_start, 0);
9239 schedstat_set(p->se.statistics.sleep_start, 0);
9240 schedstat_set(p->se.statistics.block_start, 0);
9242 if (!dl_task(p) && !rt_task(p)) {
9244 * Renice negative nice level userspace
9247 if (task_nice(p) < 0)
9248 set_user_nice(p, 0);
9252 __sched_setscheduler(p, &attr, false, false);
9254 read_unlock(&tasklist_lock);
9257 #endif /* CONFIG_MAGIC_SYSRQ */
9259 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9261 * These functions are only useful for the IA64 MCA handling, or kdb.
9263 * They can only be called when the whole system has been
9264 * stopped - every CPU needs to be quiescent, and no scheduling
9265 * activity can take place. Using them for anything else would
9266 * be a serious bug, and as a result, they aren't even visible
9267 * under any other configuration.
9271 * curr_task - return the current task for a given CPU.
9272 * @cpu: the processor in question.
9274 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9276 * Return: The current task for @cpu.
9278 struct task_struct *curr_task(int cpu)
9280 return cpu_curr(cpu);
9283 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9287 * ia64_set_curr_task - set the current task for a given CPU.
9288 * @cpu: the processor in question.
9289 * @p: the task pointer to set.
9291 * Description: This function must only be used when non-maskable interrupts
9292 * are serviced on a separate stack. It allows the architecture to switch the
9293 * notion of the current task on a CPU in a non-blocking manner. This function
9294 * must be called with all CPU's synchronized, and interrupts disabled, the
9295 * and caller must save the original value of the current task (see
9296 * curr_task() above) and restore that value before reenabling interrupts and
9297 * re-starting the system.
9299 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9301 void ia64_set_curr_task(int cpu, struct task_struct *p)
9308 #ifdef CONFIG_CGROUP_SCHED
9309 /* task_group_lock serializes the addition/removal of task groups */
9310 static DEFINE_SPINLOCK(task_group_lock);
9312 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9313 struct task_group *parent)
9315 #ifdef CONFIG_UCLAMP_TASK_GROUP
9316 enum uclamp_id clamp_id;
9318 for_each_clamp_id(clamp_id) {
9319 uclamp_se_set(&tg->uclamp_req[clamp_id],
9320 uclamp_none(clamp_id), false);
9321 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9326 static void sched_free_group(struct task_group *tg)
9328 free_fair_sched_group(tg);
9329 free_rt_sched_group(tg);
9331 kmem_cache_free(task_group_cache, tg);
9334 /* allocate runqueue etc for a new task group */
9335 struct task_group *sched_create_group(struct task_group *parent)
9337 struct task_group *tg;
9339 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9341 return ERR_PTR(-ENOMEM);
9343 if (!alloc_fair_sched_group(tg, parent))
9346 if (!alloc_rt_sched_group(tg, parent))
9349 alloc_uclamp_sched_group(tg, parent);
9354 sched_free_group(tg);
9355 return ERR_PTR(-ENOMEM);
9358 void sched_online_group(struct task_group *tg, struct task_group *parent)
9360 unsigned long flags;
9362 spin_lock_irqsave(&task_group_lock, flags);
9363 list_add_rcu(&tg->list, &task_groups);
9365 /* Root should already exist: */
9368 tg->parent = parent;
9369 INIT_LIST_HEAD(&tg->children);
9370 list_add_rcu(&tg->siblings, &parent->children);
9371 spin_unlock_irqrestore(&task_group_lock, flags);
9373 online_fair_sched_group(tg);
9376 /* rcu callback to free various structures associated with a task group */
9377 static void sched_free_group_rcu(struct rcu_head *rhp)
9379 /* Now it should be safe to free those cfs_rqs: */
9380 sched_free_group(container_of(rhp, struct task_group, rcu));
9383 void sched_destroy_group(struct task_group *tg)
9385 /* Wait for possible concurrent references to cfs_rqs complete: */
9386 call_rcu(&tg->rcu, sched_free_group_rcu);
9389 void sched_offline_group(struct task_group *tg)
9391 unsigned long flags;
9393 /* End participation in shares distribution: */
9394 unregister_fair_sched_group(tg);
9396 spin_lock_irqsave(&task_group_lock, flags);
9397 list_del_rcu(&tg->list);
9398 list_del_rcu(&tg->siblings);
9399 spin_unlock_irqrestore(&task_group_lock, flags);
9402 static void sched_change_group(struct task_struct *tsk, int type)
9404 struct task_group *tg;
9407 * All callers are synchronized by task_rq_lock(); we do not use RCU
9408 * which is pointless here. Thus, we pass "true" to task_css_check()
9409 * to prevent lockdep warnings.
9411 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9412 struct task_group, css);
9413 tg = autogroup_task_group(tsk, tg);
9414 tsk->sched_task_group = tg;
9416 #ifdef CONFIG_FAIR_GROUP_SCHED
9417 if (tsk->sched_class->task_change_group)
9418 tsk->sched_class->task_change_group(tsk, type);
9421 set_task_rq(tsk, task_cpu(tsk));
9425 * Change task's runqueue when it moves between groups.
9427 * The caller of this function should have put the task in its new group by
9428 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9431 void sched_move_task(struct task_struct *tsk)
9433 int queued, running, queue_flags =
9434 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9438 rq = task_rq_lock(tsk, &rf);
9439 update_rq_clock(rq);
9441 running = task_current(rq, tsk);
9442 queued = task_on_rq_queued(tsk);
9445 dequeue_task(rq, tsk, queue_flags);
9447 put_prev_task(rq, tsk);
9449 sched_change_group(tsk, TASK_MOVE_GROUP);
9452 enqueue_task(rq, tsk, queue_flags);
9454 set_next_task(rq, tsk);
9456 * After changing group, the running task may have joined a
9457 * throttled one but it's still the running task. Trigger a
9458 * resched to make sure that task can still run.
9463 task_rq_unlock(rq, tsk, &rf);
9466 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9468 return css ? container_of(css, struct task_group, css) : NULL;
9471 static struct cgroup_subsys_state *
9472 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9474 struct task_group *parent = css_tg(parent_css);
9475 struct task_group *tg;
9478 /* This is early initialization for the top cgroup */
9479 return &root_task_group.css;
9482 tg = sched_create_group(parent);
9484 return ERR_PTR(-ENOMEM);
9489 /* Expose task group only after completing cgroup initialization */
9490 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9492 struct task_group *tg = css_tg(css);
9493 struct task_group *parent = css_tg(css->parent);
9496 sched_online_group(tg, parent);
9498 #ifdef CONFIG_UCLAMP_TASK_GROUP
9499 /* Propagate the effective uclamp value for the new group */
9500 mutex_lock(&uclamp_mutex);
9502 cpu_util_update_eff(css);
9504 mutex_unlock(&uclamp_mutex);
9510 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9512 struct task_group *tg = css_tg(css);
9514 sched_offline_group(tg);
9517 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9519 struct task_group *tg = css_tg(css);
9522 * Relies on the RCU grace period between css_released() and this.
9524 sched_free_group(tg);
9528 * This is called before wake_up_new_task(), therefore we really only
9529 * have to set its group bits, all the other stuff does not apply.
9531 static void cpu_cgroup_fork(struct task_struct *task)
9536 rq = task_rq_lock(task, &rf);
9538 update_rq_clock(rq);
9539 sched_change_group(task, TASK_SET_GROUP);
9541 task_rq_unlock(rq, task, &rf);
9544 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9546 struct task_struct *task;
9547 struct cgroup_subsys_state *css;
9550 cgroup_taskset_for_each(task, css, tset) {
9551 #ifdef CONFIG_RT_GROUP_SCHED
9552 if (!sched_rt_can_attach(css_tg(css), task))
9556 * Serialize against wake_up_new_task() such that if it's
9557 * running, we're sure to observe its full state.
9559 raw_spin_lock_irq(&task->pi_lock);
9561 * Avoid calling sched_move_task() before wake_up_new_task()
9562 * has happened. This would lead to problems with PELT, due to
9563 * move wanting to detach+attach while we're not attached yet.
9565 if (READ_ONCE(task->__state) == TASK_NEW)
9567 raw_spin_unlock_irq(&task->pi_lock);
9575 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9577 struct task_struct *task;
9578 struct cgroup_subsys_state *css;
9580 cgroup_taskset_for_each(task, css, tset)
9581 sched_move_task(task);
9584 #ifdef CONFIG_UCLAMP_TASK_GROUP
9585 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9587 struct cgroup_subsys_state *top_css = css;
9588 struct uclamp_se *uc_parent = NULL;
9589 struct uclamp_se *uc_se = NULL;
9590 unsigned int eff[UCLAMP_CNT];
9591 enum uclamp_id clamp_id;
9592 unsigned int clamps;
9594 lockdep_assert_held(&uclamp_mutex);
9595 SCHED_WARN_ON(!rcu_read_lock_held());
9597 css_for_each_descendant_pre(css, top_css) {
9598 uc_parent = css_tg(css)->parent
9599 ? css_tg(css)->parent->uclamp : NULL;
9601 for_each_clamp_id(clamp_id) {
9602 /* Assume effective clamps matches requested clamps */
9603 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
9604 /* Cap effective clamps with parent's effective clamps */
9606 eff[clamp_id] > uc_parent[clamp_id].value) {
9607 eff[clamp_id] = uc_parent[clamp_id].value;
9610 /* Ensure protection is always capped by limit */
9611 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
9613 /* Propagate most restrictive effective clamps */
9615 uc_se = css_tg(css)->uclamp;
9616 for_each_clamp_id(clamp_id) {
9617 if (eff[clamp_id] == uc_se[clamp_id].value)
9619 uc_se[clamp_id].value = eff[clamp_id];
9620 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
9621 clamps |= (0x1 << clamp_id);
9624 css = css_rightmost_descendant(css);
9628 /* Immediately update descendants RUNNABLE tasks */
9629 uclamp_update_active_tasks(css);
9634 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
9635 * C expression. Since there is no way to convert a macro argument (N) into a
9636 * character constant, use two levels of macros.
9638 #define _POW10(exp) ((unsigned int)1e##exp)
9639 #define POW10(exp) _POW10(exp)
9641 struct uclamp_request {
9642 #define UCLAMP_PERCENT_SHIFT 2
9643 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
9649 static inline struct uclamp_request
9650 capacity_from_percent(char *buf)
9652 struct uclamp_request req = {
9653 .percent = UCLAMP_PERCENT_SCALE,
9654 .util = SCHED_CAPACITY_SCALE,
9659 if (strcmp(buf, "max")) {
9660 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
9664 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
9669 req.util = req.percent << SCHED_CAPACITY_SHIFT;
9670 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
9676 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
9677 size_t nbytes, loff_t off,
9678 enum uclamp_id clamp_id)
9680 struct uclamp_request req;
9681 struct task_group *tg;
9683 req = capacity_from_percent(buf);
9687 static_branch_enable(&sched_uclamp_used);
9689 mutex_lock(&uclamp_mutex);
9692 tg = css_tg(of_css(of));
9693 if (tg->uclamp_req[clamp_id].value != req.util)
9694 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
9697 * Because of not recoverable conversion rounding we keep track of the
9698 * exact requested value
9700 tg->uclamp_pct[clamp_id] = req.percent;
9702 /* Update effective clamps to track the most restrictive value */
9703 cpu_util_update_eff(of_css(of));
9706 mutex_unlock(&uclamp_mutex);
9711 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
9712 char *buf, size_t nbytes,
9715 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
9718 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
9719 char *buf, size_t nbytes,
9722 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
9725 static inline void cpu_uclamp_print(struct seq_file *sf,
9726 enum uclamp_id clamp_id)
9728 struct task_group *tg;
9734 tg = css_tg(seq_css(sf));
9735 util_clamp = tg->uclamp_req[clamp_id].value;
9738 if (util_clamp == SCHED_CAPACITY_SCALE) {
9739 seq_puts(sf, "max\n");
9743 percent = tg->uclamp_pct[clamp_id];
9744 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
9745 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
9748 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
9750 cpu_uclamp_print(sf, UCLAMP_MIN);
9754 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
9756 cpu_uclamp_print(sf, UCLAMP_MAX);
9759 #endif /* CONFIG_UCLAMP_TASK_GROUP */
9761 #ifdef CONFIG_FAIR_GROUP_SCHED
9762 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
9763 struct cftype *cftype, u64 shareval)
9765 if (shareval > scale_load_down(ULONG_MAX))
9766 shareval = MAX_SHARES;
9767 return sched_group_set_shares(css_tg(css), scale_load(shareval));
9770 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
9773 struct task_group *tg = css_tg(css);
9775 return (u64) scale_load_down(tg->shares);
9778 #ifdef CONFIG_CFS_BANDWIDTH
9779 static DEFINE_MUTEX(cfs_constraints_mutex);
9781 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9782 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9783 /* More than 203 days if BW_SHIFT equals 20. */
9784 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
9786 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9788 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
9791 int i, ret = 0, runtime_enabled, runtime_was_enabled;
9792 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9794 if (tg == &root_task_group)
9798 * Ensure we have at some amount of bandwidth every period. This is
9799 * to prevent reaching a state of large arrears when throttled via
9800 * entity_tick() resulting in prolonged exit starvation.
9802 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9806 * Likewise, bound things on the other side by preventing insane quota
9807 * periods. This also allows us to normalize in computing quota
9810 if (period > max_cfs_quota_period)
9814 * Bound quota to defend quota against overflow during bandwidth shift.
9816 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
9819 if (quota != RUNTIME_INF && (burst > quota ||
9820 burst + quota > max_cfs_runtime))
9824 * Prevent race between setting of cfs_rq->runtime_enabled and
9825 * unthrottle_offline_cfs_rqs().
9828 mutex_lock(&cfs_constraints_mutex);
9829 ret = __cfs_schedulable(tg, period, quota);
9833 runtime_enabled = quota != RUNTIME_INF;
9834 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
9836 * If we need to toggle cfs_bandwidth_used, off->on must occur
9837 * before making related changes, and on->off must occur afterwards
9839 if (runtime_enabled && !runtime_was_enabled)
9840 cfs_bandwidth_usage_inc();
9841 raw_spin_lock_irq(&cfs_b->lock);
9842 cfs_b->period = ns_to_ktime(period);
9843 cfs_b->quota = quota;
9844 cfs_b->burst = burst;
9846 __refill_cfs_bandwidth_runtime(cfs_b);
9848 /* Restart the period timer (if active) to handle new period expiry: */
9849 if (runtime_enabled)
9850 start_cfs_bandwidth(cfs_b);
9852 raw_spin_unlock_irq(&cfs_b->lock);
9854 for_each_online_cpu(i) {
9855 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9856 struct rq *rq = cfs_rq->rq;
9859 rq_lock_irq(rq, &rf);
9860 cfs_rq->runtime_enabled = runtime_enabled;
9861 cfs_rq->runtime_remaining = 0;
9863 if (cfs_rq->throttled)
9864 unthrottle_cfs_rq(cfs_rq);
9865 rq_unlock_irq(rq, &rf);
9867 if (runtime_was_enabled && !runtime_enabled)
9868 cfs_bandwidth_usage_dec();
9870 mutex_unlock(&cfs_constraints_mutex);
9876 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9878 u64 quota, period, burst;
9880 period = ktime_to_ns(tg->cfs_bandwidth.period);
9881 burst = tg->cfs_bandwidth.burst;
9882 if (cfs_quota_us < 0)
9883 quota = RUNTIME_INF;
9884 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
9885 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9889 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9892 static long tg_get_cfs_quota(struct task_group *tg)
9896 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
9899 quota_us = tg->cfs_bandwidth.quota;
9900 do_div(quota_us, NSEC_PER_USEC);
9905 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9907 u64 quota, period, burst;
9909 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
9912 period = (u64)cfs_period_us * NSEC_PER_USEC;
9913 quota = tg->cfs_bandwidth.quota;
9914 burst = tg->cfs_bandwidth.burst;
9916 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9919 static long tg_get_cfs_period(struct task_group *tg)
9923 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
9924 do_div(cfs_period_us, NSEC_PER_USEC);
9926 return cfs_period_us;
9929 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
9931 u64 quota, period, burst;
9933 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
9936 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
9937 period = ktime_to_ns(tg->cfs_bandwidth.period);
9938 quota = tg->cfs_bandwidth.quota;
9940 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9943 static long tg_get_cfs_burst(struct task_group *tg)
9947 burst_us = tg->cfs_bandwidth.burst;
9948 do_div(burst_us, NSEC_PER_USEC);
9953 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
9956 return tg_get_cfs_quota(css_tg(css));
9959 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
9960 struct cftype *cftype, s64 cfs_quota_us)
9962 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
9965 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
9968 return tg_get_cfs_period(css_tg(css));
9971 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
9972 struct cftype *cftype, u64 cfs_period_us)
9974 return tg_set_cfs_period(css_tg(css), cfs_period_us);
9977 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
9980 return tg_get_cfs_burst(css_tg(css));
9983 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
9984 struct cftype *cftype, u64 cfs_burst_us)
9986 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
9989 struct cfs_schedulable_data {
9990 struct task_group *tg;
9995 * normalize group quota/period to be quota/max_period
9996 * note: units are usecs
9998 static u64 normalize_cfs_quota(struct task_group *tg,
9999 struct cfs_schedulable_data *d)
10004 period = d->period;
10007 period = tg_get_cfs_period(tg);
10008 quota = tg_get_cfs_quota(tg);
10011 /* note: these should typically be equivalent */
10012 if (quota == RUNTIME_INF || quota == -1)
10013 return RUNTIME_INF;
10015 return to_ratio(period, quota);
10018 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10020 struct cfs_schedulable_data *d = data;
10021 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10022 s64 quota = 0, parent_quota = -1;
10025 quota = RUNTIME_INF;
10027 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10029 quota = normalize_cfs_quota(tg, d);
10030 parent_quota = parent_b->hierarchical_quota;
10033 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10034 * always take the min. On cgroup1, only inherit when no
10037 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10038 quota = min(quota, parent_quota);
10040 if (quota == RUNTIME_INF)
10041 quota = parent_quota;
10042 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10046 cfs_b->hierarchical_quota = quota;
10051 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10054 struct cfs_schedulable_data data = {
10060 if (quota != RUNTIME_INF) {
10061 do_div(data.period, NSEC_PER_USEC);
10062 do_div(data.quota, NSEC_PER_USEC);
10066 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10072 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10074 struct task_group *tg = css_tg(seq_css(sf));
10075 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10077 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10078 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10079 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10081 if (schedstat_enabled() && tg != &root_task_group) {
10085 for_each_possible_cpu(i)
10086 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10088 seq_printf(sf, "wait_sum %llu\n", ws);
10093 #endif /* CONFIG_CFS_BANDWIDTH */
10094 #endif /* CONFIG_FAIR_GROUP_SCHED */
10096 #ifdef CONFIG_RT_GROUP_SCHED
10097 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10098 struct cftype *cft, s64 val)
10100 return sched_group_set_rt_runtime(css_tg(css), val);
10103 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10104 struct cftype *cft)
10106 return sched_group_rt_runtime(css_tg(css));
10109 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10110 struct cftype *cftype, u64 rt_period_us)
10112 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10115 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10116 struct cftype *cft)
10118 return sched_group_rt_period(css_tg(css));
10120 #endif /* CONFIG_RT_GROUP_SCHED */
10122 static struct cftype cpu_legacy_files[] = {
10123 #ifdef CONFIG_FAIR_GROUP_SCHED
10126 .read_u64 = cpu_shares_read_u64,
10127 .write_u64 = cpu_shares_write_u64,
10130 #ifdef CONFIG_CFS_BANDWIDTH
10132 .name = "cfs_quota_us",
10133 .read_s64 = cpu_cfs_quota_read_s64,
10134 .write_s64 = cpu_cfs_quota_write_s64,
10137 .name = "cfs_period_us",
10138 .read_u64 = cpu_cfs_period_read_u64,
10139 .write_u64 = cpu_cfs_period_write_u64,
10142 .name = "cfs_burst_us",
10143 .read_u64 = cpu_cfs_burst_read_u64,
10144 .write_u64 = cpu_cfs_burst_write_u64,
10148 .seq_show = cpu_cfs_stat_show,
10151 #ifdef CONFIG_RT_GROUP_SCHED
10153 .name = "rt_runtime_us",
10154 .read_s64 = cpu_rt_runtime_read,
10155 .write_s64 = cpu_rt_runtime_write,
10158 .name = "rt_period_us",
10159 .read_u64 = cpu_rt_period_read_uint,
10160 .write_u64 = cpu_rt_period_write_uint,
10163 #ifdef CONFIG_UCLAMP_TASK_GROUP
10165 .name = "uclamp.min",
10166 .flags = CFTYPE_NOT_ON_ROOT,
10167 .seq_show = cpu_uclamp_min_show,
10168 .write = cpu_uclamp_min_write,
10171 .name = "uclamp.max",
10172 .flags = CFTYPE_NOT_ON_ROOT,
10173 .seq_show = cpu_uclamp_max_show,
10174 .write = cpu_uclamp_max_write,
10177 { } /* Terminate */
10180 static int cpu_extra_stat_show(struct seq_file *sf,
10181 struct cgroup_subsys_state *css)
10183 #ifdef CONFIG_CFS_BANDWIDTH
10185 struct task_group *tg = css_tg(css);
10186 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10187 u64 throttled_usec;
10189 throttled_usec = cfs_b->throttled_time;
10190 do_div(throttled_usec, NSEC_PER_USEC);
10192 seq_printf(sf, "nr_periods %d\n"
10193 "nr_throttled %d\n"
10194 "throttled_usec %llu\n",
10195 cfs_b->nr_periods, cfs_b->nr_throttled,
10202 #ifdef CONFIG_FAIR_GROUP_SCHED
10203 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10204 struct cftype *cft)
10206 struct task_group *tg = css_tg(css);
10207 u64 weight = scale_load_down(tg->shares);
10209 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10212 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10213 struct cftype *cft, u64 weight)
10216 * cgroup weight knobs should use the common MIN, DFL and MAX
10217 * values which are 1, 100 and 10000 respectively. While it loses
10218 * a bit of range on both ends, it maps pretty well onto the shares
10219 * value used by scheduler and the round-trip conversions preserve
10220 * the original value over the entire range.
10222 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10225 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10227 return sched_group_set_shares(css_tg(css), scale_load(weight));
10230 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10231 struct cftype *cft)
10233 unsigned long weight = scale_load_down(css_tg(css)->shares);
10234 int last_delta = INT_MAX;
10237 /* find the closest nice value to the current weight */
10238 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10239 delta = abs(sched_prio_to_weight[prio] - weight);
10240 if (delta >= last_delta)
10242 last_delta = delta;
10245 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10248 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10249 struct cftype *cft, s64 nice)
10251 unsigned long weight;
10254 if (nice < MIN_NICE || nice > MAX_NICE)
10257 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10258 idx = array_index_nospec(idx, 40);
10259 weight = sched_prio_to_weight[idx];
10261 return sched_group_set_shares(css_tg(css), scale_load(weight));
10265 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10266 long period, long quota)
10269 seq_puts(sf, "max");
10271 seq_printf(sf, "%ld", quota);
10273 seq_printf(sf, " %ld\n", period);
10276 /* caller should put the current value in *@periodp before calling */
10277 static int __maybe_unused cpu_period_quota_parse(char *buf,
10278 u64 *periodp, u64 *quotap)
10280 char tok[21]; /* U64_MAX */
10282 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10285 *periodp *= NSEC_PER_USEC;
10287 if (sscanf(tok, "%llu", quotap))
10288 *quotap *= NSEC_PER_USEC;
10289 else if (!strcmp(tok, "max"))
10290 *quotap = RUNTIME_INF;
10297 #ifdef CONFIG_CFS_BANDWIDTH
10298 static int cpu_max_show(struct seq_file *sf, void *v)
10300 struct task_group *tg = css_tg(seq_css(sf));
10302 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10306 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10307 char *buf, size_t nbytes, loff_t off)
10309 struct task_group *tg = css_tg(of_css(of));
10310 u64 period = tg_get_cfs_period(tg);
10311 u64 burst = tg_get_cfs_burst(tg);
10315 ret = cpu_period_quota_parse(buf, &period, "a);
10317 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10318 return ret ?: nbytes;
10322 static struct cftype cpu_files[] = {
10323 #ifdef CONFIG_FAIR_GROUP_SCHED
10326 .flags = CFTYPE_NOT_ON_ROOT,
10327 .read_u64 = cpu_weight_read_u64,
10328 .write_u64 = cpu_weight_write_u64,
10331 .name = "weight.nice",
10332 .flags = CFTYPE_NOT_ON_ROOT,
10333 .read_s64 = cpu_weight_nice_read_s64,
10334 .write_s64 = cpu_weight_nice_write_s64,
10337 #ifdef CONFIG_CFS_BANDWIDTH
10340 .flags = CFTYPE_NOT_ON_ROOT,
10341 .seq_show = cpu_max_show,
10342 .write = cpu_max_write,
10345 .name = "max.burst",
10346 .flags = CFTYPE_NOT_ON_ROOT,
10347 .read_u64 = cpu_cfs_burst_read_u64,
10348 .write_u64 = cpu_cfs_burst_write_u64,
10351 #ifdef CONFIG_UCLAMP_TASK_GROUP
10353 .name = "uclamp.min",
10354 .flags = CFTYPE_NOT_ON_ROOT,
10355 .seq_show = cpu_uclamp_min_show,
10356 .write = cpu_uclamp_min_write,
10359 .name = "uclamp.max",
10360 .flags = CFTYPE_NOT_ON_ROOT,
10361 .seq_show = cpu_uclamp_max_show,
10362 .write = cpu_uclamp_max_write,
10365 { } /* terminate */
10368 struct cgroup_subsys cpu_cgrp_subsys = {
10369 .css_alloc = cpu_cgroup_css_alloc,
10370 .css_online = cpu_cgroup_css_online,
10371 .css_released = cpu_cgroup_css_released,
10372 .css_free = cpu_cgroup_css_free,
10373 .css_extra_stat_show = cpu_extra_stat_show,
10374 .fork = cpu_cgroup_fork,
10375 .can_attach = cpu_cgroup_can_attach,
10376 .attach = cpu_cgroup_attach,
10377 .legacy_cftypes = cpu_legacy_files,
10378 .dfl_cftypes = cpu_files,
10379 .early_init = true,
10383 #endif /* CONFIG_CGROUP_SCHED */
10385 void dump_cpu_task(int cpu)
10387 pr_info("Task dump for CPU %d:\n", cpu);
10388 sched_show_task(cpu_curr(cpu));
10392 * Nice levels are multiplicative, with a gentle 10% change for every
10393 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10394 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10395 * that remained on nice 0.
10397 * The "10% effect" is relative and cumulative: from _any_ nice level,
10398 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10399 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10400 * If a task goes up by ~10% and another task goes down by ~10% then
10401 * the relative distance between them is ~25%.)
10403 const int sched_prio_to_weight[40] = {
10404 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10405 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10406 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10407 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10408 /* 0 */ 1024, 820, 655, 526, 423,
10409 /* 5 */ 335, 272, 215, 172, 137,
10410 /* 10 */ 110, 87, 70, 56, 45,
10411 /* 15 */ 36, 29, 23, 18, 15,
10415 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10417 * In cases where the weight does not change often, we can use the
10418 * precalculated inverse to speed up arithmetics by turning divisions
10419 * into multiplications:
10421 const u32 sched_prio_to_wmult[40] = {
10422 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10423 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10424 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10425 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10426 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10427 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10428 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10429 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10432 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10434 trace_sched_update_nr_running_tp(rq, count);