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 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1933 if (!(flags & ENQUEUE_NOCLOCK))
1934 update_rq_clock(rq);
1936 if (!(flags & ENQUEUE_RESTORE)) {
1937 sched_info_enqueue(rq, p);
1938 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1941 uclamp_rq_inc(rq, p);
1942 p->sched_class->enqueue_task(rq, p, flags);
1944 if (sched_core_enabled(rq))
1945 sched_core_enqueue(rq, p);
1948 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1950 if (sched_core_enabled(rq))
1951 sched_core_dequeue(rq, p);
1953 if (!(flags & DEQUEUE_NOCLOCK))
1954 update_rq_clock(rq);
1956 if (!(flags & DEQUEUE_SAVE)) {
1957 sched_info_dequeue(rq, p);
1958 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1961 uclamp_rq_dec(rq, p);
1962 p->sched_class->dequeue_task(rq, p, flags);
1965 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1967 enqueue_task(rq, p, flags);
1969 p->on_rq = TASK_ON_RQ_QUEUED;
1972 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1974 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1976 dequeue_task(rq, p, flags);
1980 * __normal_prio - return the priority that is based on the static prio
1982 static inline int __normal_prio(struct task_struct *p)
1984 return p->static_prio;
1988 * Calculate the expected normal priority: i.e. priority
1989 * without taking RT-inheritance into account. Might be
1990 * boosted by interactivity modifiers. Changes upon fork,
1991 * setprio syscalls, and whenever the interactivity
1992 * estimator recalculates.
1994 static inline int normal_prio(struct task_struct *p)
1998 if (task_has_dl_policy(p))
1999 prio = MAX_DL_PRIO-1;
2000 else if (task_has_rt_policy(p))
2001 prio = MAX_RT_PRIO-1 - p->rt_priority;
2003 prio = __normal_prio(p);
2008 * Calculate the current priority, i.e. the priority
2009 * taken into account by the scheduler. This value might
2010 * be boosted by RT tasks, or might be boosted by
2011 * interactivity modifiers. Will be RT if the task got
2012 * RT-boosted. If not then it returns p->normal_prio.
2014 static int effective_prio(struct task_struct *p)
2016 p->normal_prio = normal_prio(p);
2018 * If we are RT tasks or we were boosted to RT priority,
2019 * keep the priority unchanged. Otherwise, update priority
2020 * to the normal priority:
2022 if (!rt_prio(p->prio))
2023 return p->normal_prio;
2028 * task_curr - is this task currently executing on a CPU?
2029 * @p: the task in question.
2031 * Return: 1 if the task is currently executing. 0 otherwise.
2033 inline int task_curr(const struct task_struct *p)
2035 return cpu_curr(task_cpu(p)) == p;
2039 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2040 * use the balance_callback list if you want balancing.
2042 * this means any call to check_class_changed() must be followed by a call to
2043 * balance_callback().
2045 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2046 const struct sched_class *prev_class,
2049 if (prev_class != p->sched_class) {
2050 if (prev_class->switched_from)
2051 prev_class->switched_from(rq, p);
2053 p->sched_class->switched_to(rq, p);
2054 } else if (oldprio != p->prio || dl_task(p))
2055 p->sched_class->prio_changed(rq, p, oldprio);
2058 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2060 if (p->sched_class == rq->curr->sched_class)
2061 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2062 else if (p->sched_class > rq->curr->sched_class)
2066 * A queue event has occurred, and we're going to schedule. In
2067 * this case, we can save a useless back to back clock update.
2069 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2070 rq_clock_skip_update(rq);
2076 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2078 static int __set_cpus_allowed_ptr(struct task_struct *p,
2079 const struct cpumask *new_mask,
2082 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2084 if (likely(!p->migration_disabled))
2087 if (p->cpus_ptr != &p->cpus_mask)
2091 * Violates locking rules! see comment in __do_set_cpus_allowed().
2093 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2096 void migrate_disable(void)
2098 struct task_struct *p = current;
2100 if (p->migration_disabled) {
2101 p->migration_disabled++;
2106 this_rq()->nr_pinned++;
2107 p->migration_disabled = 1;
2110 EXPORT_SYMBOL_GPL(migrate_disable);
2112 void migrate_enable(void)
2114 struct task_struct *p = current;
2116 if (p->migration_disabled > 1) {
2117 p->migration_disabled--;
2122 * Ensure stop_task runs either before or after this, and that
2123 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2126 if (p->cpus_ptr != &p->cpus_mask)
2127 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2129 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2130 * regular cpus_mask, otherwise things that race (eg.
2131 * select_fallback_rq) get confused.
2134 p->migration_disabled = 0;
2135 this_rq()->nr_pinned--;
2138 EXPORT_SYMBOL_GPL(migrate_enable);
2140 static inline bool rq_has_pinned_tasks(struct rq *rq)
2142 return rq->nr_pinned;
2146 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2147 * __set_cpus_allowed_ptr() and select_fallback_rq().
2149 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2151 /* When not in the task's cpumask, no point in looking further. */
2152 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2155 /* migrate_disabled() must be allowed to finish. */
2156 if (is_migration_disabled(p))
2157 return cpu_online(cpu);
2159 /* Non kernel threads are not allowed during either online or offline. */
2160 if (!(p->flags & PF_KTHREAD))
2161 return cpu_active(cpu);
2163 /* KTHREAD_IS_PER_CPU is always allowed. */
2164 if (kthread_is_per_cpu(p))
2165 return cpu_online(cpu);
2167 /* Regular kernel threads don't get to stay during offline. */
2171 /* But are allowed during online. */
2172 return cpu_online(cpu);
2176 * This is how migration works:
2178 * 1) we invoke migration_cpu_stop() on the target CPU using
2180 * 2) stopper starts to run (implicitly forcing the migrated thread
2182 * 3) it checks whether the migrated task is still in the wrong runqueue.
2183 * 4) if it's in the wrong runqueue then the migration thread removes
2184 * it and puts it into the right queue.
2185 * 5) stopper completes and stop_one_cpu() returns and the migration
2190 * move_queued_task - move a queued task to new rq.
2192 * Returns (locked) new rq. Old rq's lock is released.
2194 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2195 struct task_struct *p, int new_cpu)
2197 lockdep_assert_rq_held(rq);
2199 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2200 set_task_cpu(p, new_cpu);
2203 rq = cpu_rq(new_cpu);
2206 BUG_ON(task_cpu(p) != new_cpu);
2207 activate_task(rq, p, 0);
2208 check_preempt_curr(rq, p, 0);
2213 struct migration_arg {
2214 struct task_struct *task;
2216 struct set_affinity_pending *pending;
2220 * @refs: number of wait_for_completion()
2221 * @stop_pending: is @stop_work in use
2223 struct set_affinity_pending {
2225 unsigned int stop_pending;
2226 struct completion done;
2227 struct cpu_stop_work stop_work;
2228 struct migration_arg arg;
2232 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2233 * this because either it can't run here any more (set_cpus_allowed()
2234 * away from this CPU, or CPU going down), or because we're
2235 * attempting to rebalance this task on exec (sched_exec).
2237 * So we race with normal scheduler movements, but that's OK, as long
2238 * as the task is no longer on this CPU.
2240 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2241 struct task_struct *p, int dest_cpu)
2243 /* Affinity changed (again). */
2244 if (!is_cpu_allowed(p, dest_cpu))
2247 update_rq_clock(rq);
2248 rq = move_queued_task(rq, rf, p, dest_cpu);
2254 * migration_cpu_stop - this will be executed by a highprio stopper thread
2255 * and performs thread migration by bumping thread off CPU then
2256 * 'pushing' onto another runqueue.
2258 static int migration_cpu_stop(void *data)
2260 struct migration_arg *arg = data;
2261 struct set_affinity_pending *pending = arg->pending;
2262 struct task_struct *p = arg->task;
2263 struct rq *rq = this_rq();
2264 bool complete = false;
2268 * The original target CPU might have gone down and we might
2269 * be on another CPU but it doesn't matter.
2271 local_irq_save(rf.flags);
2273 * We need to explicitly wake pending tasks before running
2274 * __migrate_task() such that we will not miss enforcing cpus_ptr
2275 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2277 flush_smp_call_function_from_idle();
2279 raw_spin_lock(&p->pi_lock);
2283 * If we were passed a pending, then ->stop_pending was set, thus
2284 * p->migration_pending must have remained stable.
2286 WARN_ON_ONCE(pending && pending != p->migration_pending);
2289 * If task_rq(p) != rq, it cannot be migrated here, because we're
2290 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2291 * we're holding p->pi_lock.
2293 if (task_rq(p) == rq) {
2294 if (is_migration_disabled(p))
2298 p->migration_pending = NULL;
2301 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2305 if (task_on_rq_queued(p))
2306 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2308 p->wake_cpu = arg->dest_cpu;
2311 * XXX __migrate_task() can fail, at which point we might end
2312 * up running on a dodgy CPU, AFAICT this can only happen
2313 * during CPU hotplug, at which point we'll get pushed out
2314 * anyway, so it's probably not a big deal.
2317 } else if (pending) {
2319 * This happens when we get migrated between migrate_enable()'s
2320 * preempt_enable() and scheduling the stopper task. At that
2321 * point we're a regular task again and not current anymore.
2323 * A !PREEMPT kernel has a giant hole here, which makes it far
2328 * The task moved before the stopper got to run. We're holding
2329 * ->pi_lock, so the allowed mask is stable - if it got
2330 * somewhere allowed, we're done.
2332 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2333 p->migration_pending = NULL;
2339 * When migrate_enable() hits a rq mis-match we can't reliably
2340 * determine is_migration_disabled() and so have to chase after
2343 WARN_ON_ONCE(!pending->stop_pending);
2344 task_rq_unlock(rq, p, &rf);
2345 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2346 &pending->arg, &pending->stop_work);
2351 pending->stop_pending = false;
2352 task_rq_unlock(rq, p, &rf);
2355 complete_all(&pending->done);
2360 int push_cpu_stop(void *arg)
2362 struct rq *lowest_rq = NULL, *rq = this_rq();
2363 struct task_struct *p = arg;
2365 raw_spin_lock_irq(&p->pi_lock);
2366 raw_spin_rq_lock(rq);
2368 if (task_rq(p) != rq)
2371 if (is_migration_disabled(p)) {
2372 p->migration_flags |= MDF_PUSH;
2376 p->migration_flags &= ~MDF_PUSH;
2378 if (p->sched_class->find_lock_rq)
2379 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2384 // XXX validate p is still the highest prio task
2385 if (task_rq(p) == rq) {
2386 deactivate_task(rq, p, 0);
2387 set_task_cpu(p, lowest_rq->cpu);
2388 activate_task(lowest_rq, p, 0);
2389 resched_curr(lowest_rq);
2392 double_unlock_balance(rq, lowest_rq);
2395 rq->push_busy = false;
2396 raw_spin_rq_unlock(rq);
2397 raw_spin_unlock_irq(&p->pi_lock);
2404 * sched_class::set_cpus_allowed must do the below, but is not required to
2405 * actually call this function.
2407 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2409 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2410 p->cpus_ptr = new_mask;
2414 cpumask_copy(&p->cpus_mask, new_mask);
2415 p->nr_cpus_allowed = cpumask_weight(new_mask);
2419 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2421 struct rq *rq = task_rq(p);
2422 bool queued, running;
2425 * This here violates the locking rules for affinity, since we're only
2426 * supposed to change these variables while holding both rq->lock and
2429 * HOWEVER, it magically works, because ttwu() is the only code that
2430 * accesses these variables under p->pi_lock and only does so after
2431 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2432 * before finish_task().
2434 * XXX do further audits, this smells like something putrid.
2436 if (flags & SCA_MIGRATE_DISABLE)
2437 SCHED_WARN_ON(!p->on_cpu);
2439 lockdep_assert_held(&p->pi_lock);
2441 queued = task_on_rq_queued(p);
2442 running = task_current(rq, p);
2446 * Because __kthread_bind() calls this on blocked tasks without
2449 lockdep_assert_rq_held(rq);
2450 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2453 put_prev_task(rq, p);
2455 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2458 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2460 set_next_task(rq, p);
2463 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2465 __do_set_cpus_allowed(p, new_mask, 0);
2469 * This function is wildly self concurrent; here be dragons.
2472 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2473 * designated task is enqueued on an allowed CPU. If that task is currently
2474 * running, we have to kick it out using the CPU stopper.
2476 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2479 * Initial conditions: P0->cpus_mask = [0, 1]
2483 * migrate_disable();
2485 * set_cpus_allowed_ptr(P0, [1]);
2487 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2488 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2489 * This means we need the following scheme:
2493 * migrate_disable();
2495 * set_cpus_allowed_ptr(P0, [1]);
2499 * __set_cpus_allowed_ptr();
2500 * <wakes local stopper>
2501 * `--> <woken on migration completion>
2503 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2504 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2505 * task p are serialized by p->pi_lock, which we can leverage: the one that
2506 * should come into effect at the end of the Migrate-Disable region is the last
2507 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2508 * but we still need to properly signal those waiting tasks at the appropriate
2511 * This is implemented using struct set_affinity_pending. The first
2512 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2513 * setup an instance of that struct and install it on the targeted task_struct.
2514 * Any and all further callers will reuse that instance. Those then wait for
2515 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2516 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2519 * (1) In the cases covered above. There is one more where the completion is
2520 * signaled within affine_move_task() itself: when a subsequent affinity request
2521 * occurs after the stopper bailed out due to the targeted task still being
2522 * Migrate-Disable. Consider:
2524 * Initial conditions: P0->cpus_mask = [0, 1]
2528 * migrate_disable();
2530 * set_cpus_allowed_ptr(P0, [1]);
2533 * migration_cpu_stop()
2534 * is_migration_disabled()
2536 * set_cpus_allowed_ptr(P0, [0, 1]);
2537 * <signal completion>
2540 * Note that the above is safe vs a concurrent migrate_enable(), as any
2541 * pending affinity completion is preceded by an uninstallation of
2542 * p->migration_pending done with p->pi_lock held.
2544 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2545 int dest_cpu, unsigned int flags)
2547 struct set_affinity_pending my_pending = { }, *pending = NULL;
2548 bool stop_pending, complete = false;
2550 /* Can the task run on the task's current CPU? If so, we're done */
2551 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2552 struct task_struct *push_task = NULL;
2554 if ((flags & SCA_MIGRATE_ENABLE) &&
2555 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2556 rq->push_busy = true;
2557 push_task = get_task_struct(p);
2561 * If there are pending waiters, but no pending stop_work,
2562 * then complete now.
2564 pending = p->migration_pending;
2565 if (pending && !pending->stop_pending) {
2566 p->migration_pending = NULL;
2570 task_rq_unlock(rq, p, rf);
2573 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2578 complete_all(&pending->done);
2583 if (!(flags & SCA_MIGRATE_ENABLE)) {
2584 /* serialized by p->pi_lock */
2585 if (!p->migration_pending) {
2586 /* Install the request */
2587 refcount_set(&my_pending.refs, 1);
2588 init_completion(&my_pending.done);
2589 my_pending.arg = (struct migration_arg) {
2591 .dest_cpu = dest_cpu,
2592 .pending = &my_pending,
2595 p->migration_pending = &my_pending;
2597 pending = p->migration_pending;
2598 refcount_inc(&pending->refs);
2600 * Affinity has changed, but we've already installed a
2601 * pending. migration_cpu_stop() *must* see this, else
2602 * we risk a completion of the pending despite having a
2603 * task on a disallowed CPU.
2605 * Serialized by p->pi_lock, so this is safe.
2607 pending->arg.dest_cpu = dest_cpu;
2610 pending = p->migration_pending;
2612 * - !MIGRATE_ENABLE:
2613 * we'll have installed a pending if there wasn't one already.
2616 * we're here because the current CPU isn't matching anymore,
2617 * the only way that can happen is because of a concurrent
2618 * set_cpus_allowed_ptr() call, which should then still be
2619 * pending completion.
2621 * Either way, we really should have a @pending here.
2623 if (WARN_ON_ONCE(!pending)) {
2624 task_rq_unlock(rq, p, rf);
2628 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2630 * MIGRATE_ENABLE gets here because 'p == current', but for
2631 * anything else we cannot do is_migration_disabled(), punt
2632 * and have the stopper function handle it all race-free.
2634 stop_pending = pending->stop_pending;
2636 pending->stop_pending = true;
2638 if (flags & SCA_MIGRATE_ENABLE)
2639 p->migration_flags &= ~MDF_PUSH;
2641 task_rq_unlock(rq, p, rf);
2643 if (!stop_pending) {
2644 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2645 &pending->arg, &pending->stop_work);
2648 if (flags & SCA_MIGRATE_ENABLE)
2652 if (!is_migration_disabled(p)) {
2653 if (task_on_rq_queued(p))
2654 rq = move_queued_task(rq, rf, p, dest_cpu);
2656 if (!pending->stop_pending) {
2657 p->migration_pending = NULL;
2661 task_rq_unlock(rq, p, rf);
2664 complete_all(&pending->done);
2667 wait_for_completion(&pending->done);
2669 if (refcount_dec_and_test(&pending->refs))
2670 wake_up_var(&pending->refs); /* No UaF, just an address */
2673 * Block the original owner of &pending until all subsequent callers
2674 * have seen the completion and decremented the refcount
2676 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2679 WARN_ON_ONCE(my_pending.stop_pending);
2685 * Change a given task's CPU affinity. Migrate the thread to a
2686 * proper CPU and schedule it away if the CPU it's executing on
2687 * is removed from the allowed bitmask.
2689 * NOTE: the caller must have a valid reference to the task, the
2690 * task must not exit() & deallocate itself prematurely. The
2691 * call is not atomic; no spinlocks may be held.
2693 static int __set_cpus_allowed_ptr(struct task_struct *p,
2694 const struct cpumask *new_mask,
2697 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2698 unsigned int dest_cpu;
2703 rq = task_rq_lock(p, &rf);
2704 update_rq_clock(rq);
2706 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2708 * Kernel threads are allowed on online && !active CPUs,
2709 * however, during cpu-hot-unplug, even these might get pushed
2710 * away if not KTHREAD_IS_PER_CPU.
2712 * Specifically, migration_disabled() tasks must not fail the
2713 * cpumask_any_and_distribute() pick below, esp. so on
2714 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2715 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2717 cpu_valid_mask = cpu_online_mask;
2721 * Must re-check here, to close a race against __kthread_bind(),
2722 * sched_setaffinity() is not guaranteed to observe the flag.
2724 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2729 if (!(flags & SCA_MIGRATE_ENABLE)) {
2730 if (cpumask_equal(&p->cpus_mask, new_mask))
2733 if (WARN_ON_ONCE(p == current &&
2734 is_migration_disabled(p) &&
2735 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2742 * Picking a ~random cpu helps in cases where we are changing affinity
2743 * for groups of tasks (ie. cpuset), so that load balancing is not
2744 * immediately required to distribute the tasks within their new mask.
2746 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2747 if (dest_cpu >= nr_cpu_ids) {
2752 __do_set_cpus_allowed(p, new_mask, flags);
2754 return affine_move_task(rq, p, &rf, dest_cpu, flags);
2757 task_rq_unlock(rq, p, &rf);
2762 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2764 return __set_cpus_allowed_ptr(p, new_mask, 0);
2766 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2768 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2770 #ifdef CONFIG_SCHED_DEBUG
2771 unsigned int state = READ_ONCE(p->__state);
2774 * We should never call set_task_cpu() on a blocked task,
2775 * ttwu() will sort out the placement.
2777 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
2780 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2781 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2782 * time relying on p->on_rq.
2784 WARN_ON_ONCE(state == TASK_RUNNING &&
2785 p->sched_class == &fair_sched_class &&
2786 (p->on_rq && !task_on_rq_migrating(p)));
2788 #ifdef CONFIG_LOCKDEP
2790 * The caller should hold either p->pi_lock or rq->lock, when changing
2791 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2793 * sched_move_task() holds both and thus holding either pins the cgroup,
2796 * Furthermore, all task_rq users should acquire both locks, see
2799 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2800 lockdep_is_held(__rq_lockp(task_rq(p)))));
2803 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2805 WARN_ON_ONCE(!cpu_online(new_cpu));
2807 WARN_ON_ONCE(is_migration_disabled(p));
2810 trace_sched_migrate_task(p, new_cpu);
2812 if (task_cpu(p) != new_cpu) {
2813 if (p->sched_class->migrate_task_rq)
2814 p->sched_class->migrate_task_rq(p, new_cpu);
2815 p->se.nr_migrations++;
2817 perf_event_task_migrate(p);
2820 __set_task_cpu(p, new_cpu);
2823 #ifdef CONFIG_NUMA_BALANCING
2824 static void __migrate_swap_task(struct task_struct *p, int cpu)
2826 if (task_on_rq_queued(p)) {
2827 struct rq *src_rq, *dst_rq;
2828 struct rq_flags srf, drf;
2830 src_rq = task_rq(p);
2831 dst_rq = cpu_rq(cpu);
2833 rq_pin_lock(src_rq, &srf);
2834 rq_pin_lock(dst_rq, &drf);
2836 deactivate_task(src_rq, p, 0);
2837 set_task_cpu(p, cpu);
2838 activate_task(dst_rq, p, 0);
2839 check_preempt_curr(dst_rq, p, 0);
2841 rq_unpin_lock(dst_rq, &drf);
2842 rq_unpin_lock(src_rq, &srf);
2846 * Task isn't running anymore; make it appear like we migrated
2847 * it before it went to sleep. This means on wakeup we make the
2848 * previous CPU our target instead of where it really is.
2854 struct migration_swap_arg {
2855 struct task_struct *src_task, *dst_task;
2856 int src_cpu, dst_cpu;
2859 static int migrate_swap_stop(void *data)
2861 struct migration_swap_arg *arg = data;
2862 struct rq *src_rq, *dst_rq;
2865 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2868 src_rq = cpu_rq(arg->src_cpu);
2869 dst_rq = cpu_rq(arg->dst_cpu);
2871 double_raw_lock(&arg->src_task->pi_lock,
2872 &arg->dst_task->pi_lock);
2873 double_rq_lock(src_rq, dst_rq);
2875 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2878 if (task_cpu(arg->src_task) != arg->src_cpu)
2881 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2884 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2887 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2888 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2893 double_rq_unlock(src_rq, dst_rq);
2894 raw_spin_unlock(&arg->dst_task->pi_lock);
2895 raw_spin_unlock(&arg->src_task->pi_lock);
2901 * Cross migrate two tasks
2903 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2904 int target_cpu, int curr_cpu)
2906 struct migration_swap_arg arg;
2909 arg = (struct migration_swap_arg){
2911 .src_cpu = curr_cpu,
2913 .dst_cpu = target_cpu,
2916 if (arg.src_cpu == arg.dst_cpu)
2920 * These three tests are all lockless; this is OK since all of them
2921 * will be re-checked with proper locks held further down the line.
2923 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2926 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2929 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2932 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2933 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2938 #endif /* CONFIG_NUMA_BALANCING */
2941 * wait_task_inactive - wait for a thread to unschedule.
2943 * If @match_state is nonzero, it's the @p->state value just checked and
2944 * not expected to change. If it changes, i.e. @p might have woken up,
2945 * then return zero. When we succeed in waiting for @p to be off its CPU,
2946 * we return a positive number (its total switch count). If a second call
2947 * a short while later returns the same number, the caller can be sure that
2948 * @p has remained unscheduled the whole time.
2950 * The caller must ensure that the task *will* unschedule sometime soon,
2951 * else this function might spin for a *long* time. This function can't
2952 * be called with interrupts off, or it may introduce deadlock with
2953 * smp_call_function() if an IPI is sent by the same process we are
2954 * waiting to become inactive.
2956 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
2958 int running, queued;
2965 * We do the initial early heuristics without holding
2966 * any task-queue locks at all. We'll only try to get
2967 * the runqueue lock when things look like they will
2973 * If the task is actively running on another CPU
2974 * still, just relax and busy-wait without holding
2977 * NOTE! Since we don't hold any locks, it's not
2978 * even sure that "rq" stays as the right runqueue!
2979 * But we don't care, since "task_running()" will
2980 * return false if the runqueue has changed and p
2981 * is actually now running somewhere else!
2983 while (task_running(rq, p)) {
2984 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
2990 * Ok, time to look more closely! We need the rq
2991 * lock now, to be *sure*. If we're wrong, we'll
2992 * just go back and repeat.
2994 rq = task_rq_lock(p, &rf);
2995 trace_sched_wait_task(p);
2996 running = task_running(rq, p);
2997 queued = task_on_rq_queued(p);
2999 if (!match_state || READ_ONCE(p->__state) == match_state)
3000 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3001 task_rq_unlock(rq, p, &rf);
3004 * If it changed from the expected state, bail out now.
3006 if (unlikely(!ncsw))
3010 * Was it really running after all now that we
3011 * checked with the proper locks actually held?
3013 * Oops. Go back and try again..
3015 if (unlikely(running)) {
3021 * It's not enough that it's not actively running,
3022 * it must be off the runqueue _entirely_, and not
3025 * So if it was still runnable (but just not actively
3026 * running right now), it's preempted, and we should
3027 * yield - it could be a while.
3029 if (unlikely(queued)) {
3030 ktime_t to = NSEC_PER_SEC / HZ;
3032 set_current_state(TASK_UNINTERRUPTIBLE);
3033 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
3038 * Ahh, all good. It wasn't running, and it wasn't
3039 * runnable, which means that it will never become
3040 * running in the future either. We're all done!
3049 * kick_process - kick a running thread to enter/exit the kernel
3050 * @p: the to-be-kicked thread
3052 * Cause a process which is running on another CPU to enter
3053 * kernel-mode, without any delay. (to get signals handled.)
3055 * NOTE: this function doesn't have to take the runqueue lock,
3056 * because all it wants to ensure is that the remote task enters
3057 * the kernel. If the IPI races and the task has been migrated
3058 * to another CPU then no harm is done and the purpose has been
3061 void kick_process(struct task_struct *p)
3067 if ((cpu != smp_processor_id()) && task_curr(p))
3068 smp_send_reschedule(cpu);
3071 EXPORT_SYMBOL_GPL(kick_process);
3074 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3076 * A few notes on cpu_active vs cpu_online:
3078 * - cpu_active must be a subset of cpu_online
3080 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3081 * see __set_cpus_allowed_ptr(). At this point the newly online
3082 * CPU isn't yet part of the sched domains, and balancing will not
3085 * - on CPU-down we clear cpu_active() to mask the sched domains and
3086 * avoid the load balancer to place new tasks on the to be removed
3087 * CPU. Existing tasks will remain running there and will be taken
3090 * This means that fallback selection must not select !active CPUs.
3091 * And can assume that any active CPU must be online. Conversely
3092 * select_task_rq() below may allow selection of !active CPUs in order
3093 * to satisfy the above rules.
3095 static int select_fallback_rq(int cpu, struct task_struct *p)
3097 int nid = cpu_to_node(cpu);
3098 const struct cpumask *nodemask = NULL;
3099 enum { cpuset, possible, fail } state = cpuset;
3103 * If the node that the CPU is on has been offlined, cpu_to_node()
3104 * will return -1. There is no CPU on the node, and we should
3105 * select the CPU on the other node.
3108 nodemask = cpumask_of_node(nid);
3110 /* Look for allowed, online CPU in same node. */
3111 for_each_cpu(dest_cpu, nodemask) {
3112 if (!cpu_active(dest_cpu))
3114 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
3120 /* Any allowed, online CPU? */
3121 for_each_cpu(dest_cpu, p->cpus_ptr) {
3122 if (!is_cpu_allowed(p, dest_cpu))
3128 /* No more Mr. Nice Guy. */
3131 if (IS_ENABLED(CONFIG_CPUSETS)) {
3132 cpuset_cpus_allowed_fallback(p);
3139 * XXX When called from select_task_rq() we only
3140 * hold p->pi_lock and again violate locking order.
3142 * More yuck to audit.
3144 do_set_cpus_allowed(p, cpu_possible_mask);
3155 if (state != cpuset) {
3157 * Don't tell them about moving exiting tasks or
3158 * kernel threads (both mm NULL), since they never
3161 if (p->mm && printk_ratelimit()) {
3162 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3163 task_pid_nr(p), p->comm, cpu);
3171 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3174 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3176 lockdep_assert_held(&p->pi_lock);
3178 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3179 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3181 cpu = cpumask_any(p->cpus_ptr);
3184 * In order not to call set_task_cpu() on a blocking task we need
3185 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3188 * Since this is common to all placement strategies, this lives here.
3190 * [ this allows ->select_task() to simply return task_cpu(p) and
3191 * not worry about this generic constraint ]
3193 if (unlikely(!is_cpu_allowed(p, cpu)))
3194 cpu = select_fallback_rq(task_cpu(p), p);
3199 void sched_set_stop_task(int cpu, struct task_struct *stop)
3201 static struct lock_class_key stop_pi_lock;
3202 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3203 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3207 * Make it appear like a SCHED_FIFO task, its something
3208 * userspace knows about and won't get confused about.
3210 * Also, it will make PI more or less work without too
3211 * much confusion -- but then, stop work should not
3212 * rely on PI working anyway.
3214 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3216 stop->sched_class = &stop_sched_class;
3219 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3220 * adjust the effective priority of a task. As a result,
3221 * rt_mutex_setprio() can trigger (RT) balancing operations,
3222 * which can then trigger wakeups of the stop thread to push
3223 * around the current task.
3225 * The stop task itself will never be part of the PI-chain, it
3226 * never blocks, therefore that ->pi_lock recursion is safe.
3227 * Tell lockdep about this by placing the stop->pi_lock in its
3230 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3233 cpu_rq(cpu)->stop = stop;
3237 * Reset it back to a normal scheduling class so that
3238 * it can die in pieces.
3240 old_stop->sched_class = &rt_sched_class;
3244 #else /* CONFIG_SMP */
3246 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3247 const struct cpumask *new_mask,
3250 return set_cpus_allowed_ptr(p, new_mask);
3253 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3255 static inline bool rq_has_pinned_tasks(struct rq *rq)
3260 #endif /* !CONFIG_SMP */
3263 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3267 if (!schedstat_enabled())
3273 if (cpu == rq->cpu) {
3274 __schedstat_inc(rq->ttwu_local);
3275 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3277 struct sched_domain *sd;
3279 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3281 for_each_domain(rq->cpu, sd) {
3282 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3283 __schedstat_inc(sd->ttwu_wake_remote);
3290 if (wake_flags & WF_MIGRATED)
3291 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3292 #endif /* CONFIG_SMP */
3294 __schedstat_inc(rq->ttwu_count);
3295 __schedstat_inc(p->se.statistics.nr_wakeups);
3297 if (wake_flags & WF_SYNC)
3298 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3302 * Mark the task runnable and perform wakeup-preemption.
3304 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3305 struct rq_flags *rf)
3307 check_preempt_curr(rq, p, wake_flags);
3308 WRITE_ONCE(p->__state, TASK_RUNNING);
3309 trace_sched_wakeup(p);
3312 if (p->sched_class->task_woken) {
3314 * Our task @p is fully woken up and running; so it's safe to
3315 * drop the rq->lock, hereafter rq is only used for statistics.
3317 rq_unpin_lock(rq, rf);
3318 p->sched_class->task_woken(rq, p);
3319 rq_repin_lock(rq, rf);
3322 if (rq->idle_stamp) {
3323 u64 delta = rq_clock(rq) - rq->idle_stamp;
3324 u64 max = 2*rq->max_idle_balance_cost;
3326 update_avg(&rq->avg_idle, delta);
3328 if (rq->avg_idle > max)
3331 rq->wake_stamp = jiffies;
3332 rq->wake_avg_idle = rq->avg_idle / 2;
3340 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3341 struct rq_flags *rf)
3343 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3345 lockdep_assert_rq_held(rq);
3347 if (p->sched_contributes_to_load)
3348 rq->nr_uninterruptible--;
3351 if (wake_flags & WF_MIGRATED)
3352 en_flags |= ENQUEUE_MIGRATED;
3356 delayacct_blkio_end(p);
3357 atomic_dec(&task_rq(p)->nr_iowait);
3360 activate_task(rq, p, en_flags);
3361 ttwu_do_wakeup(rq, p, wake_flags, rf);
3365 * Consider @p being inside a wait loop:
3368 * set_current_state(TASK_UNINTERRUPTIBLE);
3375 * __set_current_state(TASK_RUNNING);
3377 * between set_current_state() and schedule(). In this case @p is still
3378 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3381 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3382 * then schedule() must still happen and p->state can be changed to
3383 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3384 * need to do a full wakeup with enqueue.
3386 * Returns: %true when the wakeup is done,
3389 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3395 rq = __task_rq_lock(p, &rf);
3396 if (task_on_rq_queued(p)) {
3397 /* check_preempt_curr() may use rq clock */
3398 update_rq_clock(rq);
3399 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3402 __task_rq_unlock(rq, &rf);
3408 void sched_ttwu_pending(void *arg)
3410 struct llist_node *llist = arg;
3411 struct rq *rq = this_rq();
3412 struct task_struct *p, *t;
3419 * rq::ttwu_pending racy indication of out-standing wakeups.
3420 * Races such that false-negatives are possible, since they
3421 * are shorter lived that false-positives would be.
3423 WRITE_ONCE(rq->ttwu_pending, 0);
3425 rq_lock_irqsave(rq, &rf);
3426 update_rq_clock(rq);
3428 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3429 if (WARN_ON_ONCE(p->on_cpu))
3430 smp_cond_load_acquire(&p->on_cpu, !VAL);
3432 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3433 set_task_cpu(p, cpu_of(rq));
3435 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3438 rq_unlock_irqrestore(rq, &rf);
3441 void send_call_function_single_ipi(int cpu)
3443 struct rq *rq = cpu_rq(cpu);
3445 if (!set_nr_if_polling(rq->idle))
3446 arch_send_call_function_single_ipi(cpu);
3448 trace_sched_wake_idle_without_ipi(cpu);
3452 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3453 * necessary. The wakee CPU on receipt of the IPI will queue the task
3454 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3455 * of the wakeup instead of the waker.
3457 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3459 struct rq *rq = cpu_rq(cpu);
3461 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3463 WRITE_ONCE(rq->ttwu_pending, 1);
3464 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3467 void wake_up_if_idle(int cpu)
3469 struct rq *rq = cpu_rq(cpu);
3474 if (!is_idle_task(rcu_dereference(rq->curr)))
3477 if (set_nr_if_polling(rq->idle)) {
3478 trace_sched_wake_idle_without_ipi(cpu);
3480 rq_lock_irqsave(rq, &rf);
3481 if (is_idle_task(rq->curr))
3482 smp_send_reschedule(cpu);
3483 /* Else CPU is not idle, do nothing here: */
3484 rq_unlock_irqrestore(rq, &rf);
3491 bool cpus_share_cache(int this_cpu, int that_cpu)
3493 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3496 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3499 * Do not complicate things with the async wake_list while the CPU is
3502 if (!cpu_active(cpu))
3506 * If the CPU does not share cache, then queue the task on the
3507 * remote rqs wakelist to avoid accessing remote data.
3509 if (!cpus_share_cache(smp_processor_id(), cpu))
3513 * If the task is descheduling and the only running task on the
3514 * CPU then use the wakelist to offload the task activation to
3515 * the soon-to-be-idle CPU as the current CPU is likely busy.
3516 * nr_running is checked to avoid unnecessary task stacking.
3518 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3524 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3526 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3527 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3530 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3531 __ttwu_queue_wakelist(p, cpu, wake_flags);
3538 #else /* !CONFIG_SMP */
3540 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3545 #endif /* CONFIG_SMP */
3547 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3549 struct rq *rq = cpu_rq(cpu);
3552 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3556 update_rq_clock(rq);
3557 ttwu_do_activate(rq, p, wake_flags, &rf);
3562 * Notes on Program-Order guarantees on SMP systems.
3566 * The basic program-order guarantee on SMP systems is that when a task [t]
3567 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3568 * execution on its new CPU [c1].
3570 * For migration (of runnable tasks) this is provided by the following means:
3572 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3573 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3574 * rq(c1)->lock (if not at the same time, then in that order).
3575 * C) LOCK of the rq(c1)->lock scheduling in task
3577 * Release/acquire chaining guarantees that B happens after A and C after B.
3578 * Note: the CPU doing B need not be c0 or c1
3587 * UNLOCK rq(0)->lock
3589 * LOCK rq(0)->lock // orders against CPU0
3591 * UNLOCK rq(0)->lock
3595 * UNLOCK rq(1)->lock
3597 * LOCK rq(1)->lock // orders against CPU2
3600 * UNLOCK rq(1)->lock
3603 * BLOCKING -- aka. SLEEP + WAKEUP
3605 * For blocking we (obviously) need to provide the same guarantee as for
3606 * migration. However the means are completely different as there is no lock
3607 * chain to provide order. Instead we do:
3609 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3610 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3614 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3616 * LOCK rq(0)->lock LOCK X->pi_lock
3619 * smp_store_release(X->on_cpu, 0);
3621 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3627 * X->state = RUNNING
3628 * UNLOCK rq(2)->lock
3630 * LOCK rq(2)->lock // orders against CPU1
3633 * UNLOCK rq(2)->lock
3636 * UNLOCK rq(0)->lock
3639 * However, for wakeups there is a second guarantee we must provide, namely we
3640 * must ensure that CONDITION=1 done by the caller can not be reordered with
3641 * accesses to the task state; see try_to_wake_up() and set_current_state().
3645 * try_to_wake_up - wake up a thread
3646 * @p: the thread to be awakened
3647 * @state: the mask of task states that can be woken
3648 * @wake_flags: wake modifier flags (WF_*)
3650 * Conceptually does:
3652 * If (@state & @p->state) @p->state = TASK_RUNNING.
3654 * If the task was not queued/runnable, also place it back on a runqueue.
3656 * This function is atomic against schedule() which would dequeue the task.
3658 * It issues a full memory barrier before accessing @p->state, see the comment
3659 * with set_current_state().
3661 * Uses p->pi_lock to serialize against concurrent wake-ups.
3663 * Relies on p->pi_lock stabilizing:
3666 * - p->sched_task_group
3667 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3669 * Tries really hard to only take one task_rq(p)->lock for performance.
3670 * Takes rq->lock in:
3671 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3672 * - ttwu_queue() -- new rq, for enqueue of the task;
3673 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3675 * As a consequence we race really badly with just about everything. See the
3676 * many memory barriers and their comments for details.
3678 * Return: %true if @p->state changes (an actual wakeup was done),
3682 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3684 unsigned long flags;
3685 int cpu, success = 0;
3690 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3691 * == smp_processor_id()'. Together this means we can special
3692 * case the whole 'p->on_rq && ttwu_runnable()' case below
3693 * without taking any locks.
3696 * - we rely on Program-Order guarantees for all the ordering,
3697 * - we're serialized against set_special_state() by virtue of
3698 * it disabling IRQs (this allows not taking ->pi_lock).
3700 if (!(READ_ONCE(p->__state) & state))
3704 trace_sched_waking(p);
3705 WRITE_ONCE(p->__state, TASK_RUNNING);
3706 trace_sched_wakeup(p);
3711 * If we are going to wake up a thread waiting for CONDITION we
3712 * need to ensure that CONDITION=1 done by the caller can not be
3713 * reordered with p->state check below. This pairs with smp_store_mb()
3714 * in set_current_state() that the waiting thread does.
3716 raw_spin_lock_irqsave(&p->pi_lock, flags);
3717 smp_mb__after_spinlock();
3718 if (!(READ_ONCE(p->__state) & state))
3721 trace_sched_waking(p);
3723 /* We're going to change ->state: */
3727 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3728 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3729 * in smp_cond_load_acquire() below.
3731 * sched_ttwu_pending() try_to_wake_up()
3732 * STORE p->on_rq = 1 LOAD p->state
3735 * __schedule() (switch to task 'p')
3736 * LOCK rq->lock smp_rmb();
3737 * smp_mb__after_spinlock();
3741 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3743 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3744 * __schedule(). See the comment for smp_mb__after_spinlock().
3746 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3749 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3754 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3755 * possible to, falsely, observe p->on_cpu == 0.
3757 * One must be running (->on_cpu == 1) in order to remove oneself
3758 * from the runqueue.
3760 * __schedule() (switch to task 'p') try_to_wake_up()
3761 * STORE p->on_cpu = 1 LOAD p->on_rq
3764 * __schedule() (put 'p' to sleep)
3765 * LOCK rq->lock smp_rmb();
3766 * smp_mb__after_spinlock();
3767 * STORE p->on_rq = 0 LOAD p->on_cpu
3769 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3770 * __schedule(). See the comment for smp_mb__after_spinlock().
3772 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3773 * schedule()'s deactivate_task() has 'happened' and p will no longer
3774 * care about it's own p->state. See the comment in __schedule().
3776 smp_acquire__after_ctrl_dep();
3779 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3780 * == 0), which means we need to do an enqueue, change p->state to
3781 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3782 * enqueue, such as ttwu_queue_wakelist().
3784 WRITE_ONCE(p->__state, TASK_WAKING);
3787 * If the owning (remote) CPU is still in the middle of schedule() with
3788 * this task as prev, considering queueing p on the remote CPUs wake_list
3789 * which potentially sends an IPI instead of spinning on p->on_cpu to
3790 * let the waker make forward progress. This is safe because IRQs are
3791 * disabled and the IPI will deliver after on_cpu is cleared.
3793 * Ensure we load task_cpu(p) after p->on_cpu:
3795 * set_task_cpu(p, cpu);
3796 * STORE p->cpu = @cpu
3797 * __schedule() (switch to task 'p')
3799 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3800 * STORE p->on_cpu = 1 LOAD p->cpu
3802 * to ensure we observe the correct CPU on which the task is currently
3805 if (smp_load_acquire(&p->on_cpu) &&
3806 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3810 * If the owning (remote) CPU is still in the middle of schedule() with
3811 * this task as prev, wait until it's done referencing the task.
3813 * Pairs with the smp_store_release() in finish_task().
3815 * This ensures that tasks getting woken will be fully ordered against
3816 * their previous state and preserve Program Order.
3818 smp_cond_load_acquire(&p->on_cpu, !VAL);
3820 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3821 if (task_cpu(p) != cpu) {
3823 delayacct_blkio_end(p);
3824 atomic_dec(&task_rq(p)->nr_iowait);
3827 wake_flags |= WF_MIGRATED;
3828 psi_ttwu_dequeue(p);
3829 set_task_cpu(p, cpu);
3833 #endif /* CONFIG_SMP */
3835 ttwu_queue(p, cpu, wake_flags);
3837 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3840 ttwu_stat(p, task_cpu(p), wake_flags);
3847 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3848 * @p: Process for which the function is to be invoked, can be @current.
3849 * @func: Function to invoke.
3850 * @arg: Argument to function.
3852 * If the specified task can be quickly locked into a definite state
3853 * (either sleeping or on a given runqueue), arrange to keep it in that
3854 * state while invoking @func(@arg). This function can use ->on_rq and
3855 * task_curr() to work out what the state is, if required. Given that
3856 * @func can be invoked with a runqueue lock held, it had better be quite
3860 * @false if the task slipped out from under the locks.
3861 * @true if the task was locked onto a runqueue or is sleeping.
3862 * However, @func can override this by returning @false.
3864 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3870 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3872 rq = __task_rq_lock(p, &rf);
3873 if (task_rq(p) == rq)
3877 switch (READ_ONCE(p->__state)) {
3882 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3887 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
3892 * wake_up_process - Wake up a specific process
3893 * @p: The process to be woken up.
3895 * Attempt to wake up the nominated process and move it to the set of runnable
3898 * Return: 1 if the process was woken up, 0 if it was already running.
3900 * This function executes a full memory barrier before accessing the task state.
3902 int wake_up_process(struct task_struct *p)
3904 return try_to_wake_up(p, TASK_NORMAL, 0);
3906 EXPORT_SYMBOL(wake_up_process);
3908 int wake_up_state(struct task_struct *p, unsigned int state)
3910 return try_to_wake_up(p, state, 0);
3914 * Perform scheduler related setup for a newly forked process p.
3915 * p is forked by current.
3917 * __sched_fork() is basic setup used by init_idle() too:
3919 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3924 p->se.exec_start = 0;
3925 p->se.sum_exec_runtime = 0;
3926 p->se.prev_sum_exec_runtime = 0;
3927 p->se.nr_migrations = 0;
3929 INIT_LIST_HEAD(&p->se.group_node);
3931 #ifdef CONFIG_FAIR_GROUP_SCHED
3932 p->se.cfs_rq = NULL;
3935 #ifdef CONFIG_SCHEDSTATS
3936 /* Even if schedstat is disabled, there should not be garbage */
3937 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3940 RB_CLEAR_NODE(&p->dl.rb_node);
3941 init_dl_task_timer(&p->dl);
3942 init_dl_inactive_task_timer(&p->dl);
3943 __dl_clear_params(p);
3945 INIT_LIST_HEAD(&p->rt.run_list);
3947 p->rt.time_slice = sched_rr_timeslice;
3951 #ifdef CONFIG_PREEMPT_NOTIFIERS
3952 INIT_HLIST_HEAD(&p->preempt_notifiers);
3955 #ifdef CONFIG_COMPACTION
3956 p->capture_control = NULL;
3958 init_numa_balancing(clone_flags, p);
3960 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3961 p->migration_pending = NULL;
3965 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3967 #ifdef CONFIG_NUMA_BALANCING
3969 void set_numabalancing_state(bool enabled)
3972 static_branch_enable(&sched_numa_balancing);
3974 static_branch_disable(&sched_numa_balancing);
3977 #ifdef CONFIG_PROC_SYSCTL
3978 int sysctl_numa_balancing(struct ctl_table *table, int write,
3979 void *buffer, size_t *lenp, loff_t *ppos)
3983 int state = static_branch_likely(&sched_numa_balancing);
3985 if (write && !capable(CAP_SYS_ADMIN))
3990 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3994 set_numabalancing_state(state);
4000 #ifdef CONFIG_SCHEDSTATS
4002 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4004 static void set_schedstats(bool enabled)
4007 static_branch_enable(&sched_schedstats);
4009 static_branch_disable(&sched_schedstats);
4012 void force_schedstat_enabled(void)
4014 if (!schedstat_enabled()) {
4015 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4016 static_branch_enable(&sched_schedstats);
4020 static int __init setup_schedstats(char *str)
4026 if (!strcmp(str, "enable")) {
4027 set_schedstats(true);
4029 } else if (!strcmp(str, "disable")) {
4030 set_schedstats(false);
4035 pr_warn("Unable to parse schedstats=\n");
4039 __setup("schedstats=", setup_schedstats);
4041 #ifdef CONFIG_PROC_SYSCTL
4042 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4043 size_t *lenp, loff_t *ppos)
4047 int state = static_branch_likely(&sched_schedstats);
4049 if (write && !capable(CAP_SYS_ADMIN))
4054 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4058 set_schedstats(state);
4061 #endif /* CONFIG_PROC_SYSCTL */
4062 #endif /* CONFIG_SCHEDSTATS */
4065 * fork()/clone()-time setup:
4067 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4069 unsigned long flags;
4071 __sched_fork(clone_flags, p);
4073 * We mark the process as NEW here. This guarantees that
4074 * nobody will actually run it, and a signal or other external
4075 * event cannot wake it up and insert it on the runqueue either.
4077 p->__state = TASK_NEW;
4080 * Make sure we do not leak PI boosting priority to the child.
4082 p->prio = current->normal_prio;
4087 * Revert to default priority/policy on fork if requested.
4089 if (unlikely(p->sched_reset_on_fork)) {
4090 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4091 p->policy = SCHED_NORMAL;
4092 p->static_prio = NICE_TO_PRIO(0);
4094 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4095 p->static_prio = NICE_TO_PRIO(0);
4097 p->prio = p->normal_prio = __normal_prio(p);
4098 set_load_weight(p, false);
4101 * We don't need the reset flag anymore after the fork. It has
4102 * fulfilled its duty:
4104 p->sched_reset_on_fork = 0;
4107 if (dl_prio(p->prio))
4109 else if (rt_prio(p->prio))
4110 p->sched_class = &rt_sched_class;
4112 p->sched_class = &fair_sched_class;
4114 init_entity_runnable_average(&p->se);
4117 * The child is not yet in the pid-hash so no cgroup attach races,
4118 * and the cgroup is pinned to this child due to cgroup_fork()
4119 * is ran before sched_fork().
4121 * Silence PROVE_RCU.
4123 raw_spin_lock_irqsave(&p->pi_lock, flags);
4126 * We're setting the CPU for the first time, we don't migrate,
4127 * so use __set_task_cpu().
4129 __set_task_cpu(p, smp_processor_id());
4130 if (p->sched_class->task_fork)
4131 p->sched_class->task_fork(p);
4132 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4134 #ifdef CONFIG_SCHED_INFO
4135 if (likely(sched_info_on()))
4136 memset(&p->sched_info, 0, sizeof(p->sched_info));
4138 #if defined(CONFIG_SMP)
4141 init_task_preempt_count(p);
4143 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4144 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4149 void sched_post_fork(struct task_struct *p)
4151 uclamp_post_fork(p);
4154 unsigned long to_ratio(u64 period, u64 runtime)
4156 if (runtime == RUNTIME_INF)
4160 * Doing this here saves a lot of checks in all
4161 * the calling paths, and returning zero seems
4162 * safe for them anyway.
4167 return div64_u64(runtime << BW_SHIFT, period);
4171 * wake_up_new_task - wake up a newly created task for the first time.
4173 * This function will do some initial scheduler statistics housekeeping
4174 * that must be done for every newly created context, then puts the task
4175 * on the runqueue and wakes it.
4177 void wake_up_new_task(struct task_struct *p)
4182 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4183 WRITE_ONCE(p->__state, TASK_RUNNING);
4186 * Fork balancing, do it here and not earlier because:
4187 * - cpus_ptr can change in the fork path
4188 * - any previously selected CPU might disappear through hotplug
4190 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4191 * as we're not fully set-up yet.
4193 p->recent_used_cpu = task_cpu(p);
4195 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4197 rq = __task_rq_lock(p, &rf);
4198 update_rq_clock(rq);
4199 post_init_entity_util_avg(p);
4201 activate_task(rq, p, ENQUEUE_NOCLOCK);
4202 trace_sched_wakeup_new(p);
4203 check_preempt_curr(rq, p, WF_FORK);
4205 if (p->sched_class->task_woken) {
4207 * Nothing relies on rq->lock after this, so it's fine to
4210 rq_unpin_lock(rq, &rf);
4211 p->sched_class->task_woken(rq, p);
4212 rq_repin_lock(rq, &rf);
4215 task_rq_unlock(rq, p, &rf);
4218 #ifdef CONFIG_PREEMPT_NOTIFIERS
4220 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4222 void preempt_notifier_inc(void)
4224 static_branch_inc(&preempt_notifier_key);
4226 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4228 void preempt_notifier_dec(void)
4230 static_branch_dec(&preempt_notifier_key);
4232 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4235 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4236 * @notifier: notifier struct to register
4238 void preempt_notifier_register(struct preempt_notifier *notifier)
4240 if (!static_branch_unlikely(&preempt_notifier_key))
4241 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4243 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4245 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4248 * preempt_notifier_unregister - no longer interested in preemption notifications
4249 * @notifier: notifier struct to unregister
4251 * This is *not* safe to call from within a preemption notifier.
4253 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4255 hlist_del(¬ifier->link);
4257 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4259 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4261 struct preempt_notifier *notifier;
4263 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4264 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4267 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4269 if (static_branch_unlikely(&preempt_notifier_key))
4270 __fire_sched_in_preempt_notifiers(curr);
4274 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4275 struct task_struct *next)
4277 struct preempt_notifier *notifier;
4279 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4280 notifier->ops->sched_out(notifier, next);
4283 static __always_inline void
4284 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4285 struct task_struct *next)
4287 if (static_branch_unlikely(&preempt_notifier_key))
4288 __fire_sched_out_preempt_notifiers(curr, next);
4291 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4293 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4298 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4299 struct task_struct *next)
4303 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4305 static inline void prepare_task(struct task_struct *next)
4309 * Claim the task as running, we do this before switching to it
4310 * such that any running task will have this set.
4312 * See the ttwu() WF_ON_CPU case and its ordering comment.
4314 WRITE_ONCE(next->on_cpu, 1);
4318 static inline void finish_task(struct task_struct *prev)
4322 * This must be the very last reference to @prev from this CPU. After
4323 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4324 * must ensure this doesn't happen until the switch is completely
4327 * In particular, the load of prev->state in finish_task_switch() must
4328 * happen before this.
4330 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4332 smp_store_release(&prev->on_cpu, 0);
4338 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4340 void (*func)(struct rq *rq);
4341 struct callback_head *next;
4343 lockdep_assert_rq_held(rq);
4346 func = (void (*)(struct rq *))head->func;
4355 static void balance_push(struct rq *rq);
4357 struct callback_head balance_push_callback = {
4359 .func = (void (*)(struct callback_head *))balance_push,
4362 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4364 struct callback_head *head = rq->balance_callback;
4366 lockdep_assert_rq_held(rq);
4368 rq->balance_callback = NULL;
4373 static void __balance_callbacks(struct rq *rq)
4375 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4378 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4380 unsigned long flags;
4382 if (unlikely(head)) {
4383 raw_spin_rq_lock_irqsave(rq, flags);
4384 do_balance_callbacks(rq, head);
4385 raw_spin_rq_unlock_irqrestore(rq, flags);
4391 static inline void __balance_callbacks(struct rq *rq)
4395 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4400 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4407 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4410 * Since the runqueue lock will be released by the next
4411 * task (which is an invalid locking op but in the case
4412 * of the scheduler it's an obvious special-case), so we
4413 * do an early lockdep release here:
4415 rq_unpin_lock(rq, rf);
4416 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4417 #ifdef CONFIG_DEBUG_SPINLOCK
4418 /* this is a valid case when another task releases the spinlock */
4419 rq_lockp(rq)->owner = next;
4423 static inline void finish_lock_switch(struct rq *rq)
4426 * If we are tracking spinlock dependencies then we have to
4427 * fix up the runqueue lock - which gets 'carried over' from
4428 * prev into current:
4430 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4431 __balance_callbacks(rq);
4432 raw_spin_rq_unlock_irq(rq);
4436 * NOP if the arch has not defined these:
4439 #ifndef prepare_arch_switch
4440 # define prepare_arch_switch(next) do { } while (0)
4443 #ifndef finish_arch_post_lock_switch
4444 # define finish_arch_post_lock_switch() do { } while (0)
4447 static inline void kmap_local_sched_out(void)
4449 #ifdef CONFIG_KMAP_LOCAL
4450 if (unlikely(current->kmap_ctrl.idx))
4451 __kmap_local_sched_out();
4455 static inline void kmap_local_sched_in(void)
4457 #ifdef CONFIG_KMAP_LOCAL
4458 if (unlikely(current->kmap_ctrl.idx))
4459 __kmap_local_sched_in();
4464 * prepare_task_switch - prepare to switch tasks
4465 * @rq: the runqueue preparing to switch
4466 * @prev: the current task that is being switched out
4467 * @next: the task we are going to switch to.
4469 * This is called with the rq lock held and interrupts off. It must
4470 * be paired with a subsequent finish_task_switch after the context
4473 * prepare_task_switch sets up locking and calls architecture specific
4477 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4478 struct task_struct *next)
4480 kcov_prepare_switch(prev);
4481 sched_info_switch(rq, prev, next);
4482 perf_event_task_sched_out(prev, next);
4484 fire_sched_out_preempt_notifiers(prev, next);
4485 kmap_local_sched_out();
4487 prepare_arch_switch(next);
4491 * finish_task_switch - clean up after a task-switch
4492 * @prev: the thread we just switched away from.
4494 * finish_task_switch must be called after the context switch, paired
4495 * with a prepare_task_switch call before the context switch.
4496 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4497 * and do any other architecture-specific cleanup actions.
4499 * Note that we may have delayed dropping an mm in context_switch(). If
4500 * so, we finish that here outside of the runqueue lock. (Doing it
4501 * with the lock held can cause deadlocks; see schedule() for
4504 * The context switch have flipped the stack from under us and restored the
4505 * local variables which were saved when this task called schedule() in the
4506 * past. prev == current is still correct but we need to recalculate this_rq
4507 * because prev may have moved to another CPU.
4509 static struct rq *finish_task_switch(struct task_struct *prev)
4510 __releases(rq->lock)
4512 struct rq *rq = this_rq();
4513 struct mm_struct *mm = rq->prev_mm;
4517 * The previous task will have left us with a preempt_count of 2
4518 * because it left us after:
4521 * preempt_disable(); // 1
4523 * raw_spin_lock_irq(&rq->lock) // 2
4525 * Also, see FORK_PREEMPT_COUNT.
4527 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4528 "corrupted preempt_count: %s/%d/0x%x\n",
4529 current->comm, current->pid, preempt_count()))
4530 preempt_count_set(FORK_PREEMPT_COUNT);
4535 * A task struct has one reference for the use as "current".
4536 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4537 * schedule one last time. The schedule call will never return, and
4538 * the scheduled task must drop that reference.
4540 * We must observe prev->state before clearing prev->on_cpu (in
4541 * finish_task), otherwise a concurrent wakeup can get prev
4542 * running on another CPU and we could rave with its RUNNING -> DEAD
4543 * transition, resulting in a double drop.
4545 prev_state = READ_ONCE(prev->__state);
4546 vtime_task_switch(prev);
4547 perf_event_task_sched_in(prev, current);
4549 finish_lock_switch(rq);
4550 finish_arch_post_lock_switch();
4551 kcov_finish_switch(current);
4553 * kmap_local_sched_out() is invoked with rq::lock held and
4554 * interrupts disabled. There is no requirement for that, but the
4555 * sched out code does not have an interrupt enabled section.
4556 * Restoring the maps on sched in does not require interrupts being
4559 kmap_local_sched_in();
4561 fire_sched_in_preempt_notifiers(current);
4563 * When switching through a kernel thread, the loop in
4564 * membarrier_{private,global}_expedited() may have observed that
4565 * kernel thread and not issued an IPI. It is therefore possible to
4566 * schedule between user->kernel->user threads without passing though
4567 * switch_mm(). Membarrier requires a barrier after storing to
4568 * rq->curr, before returning to userspace, so provide them here:
4570 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4571 * provided by mmdrop(),
4572 * - a sync_core for SYNC_CORE.
4575 membarrier_mm_sync_core_before_usermode(mm);
4578 if (unlikely(prev_state == TASK_DEAD)) {
4579 if (prev->sched_class->task_dead)
4580 prev->sched_class->task_dead(prev);
4583 * Remove function-return probe instances associated with this
4584 * task and put them back on the free list.
4586 kprobe_flush_task(prev);
4588 /* Task is done with its stack. */
4589 put_task_stack(prev);
4591 put_task_struct_rcu_user(prev);
4594 tick_nohz_task_switch();
4599 * schedule_tail - first thing a freshly forked thread must call.
4600 * @prev: the thread we just switched away from.
4602 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4603 __releases(rq->lock)
4606 * New tasks start with FORK_PREEMPT_COUNT, see there and
4607 * finish_task_switch() for details.
4609 * finish_task_switch() will drop rq->lock() and lower preempt_count
4610 * and the preempt_enable() will end up enabling preemption (on
4611 * PREEMPT_COUNT kernels).
4614 finish_task_switch(prev);
4617 if (current->set_child_tid)
4618 put_user(task_pid_vnr(current), current->set_child_tid);
4620 calculate_sigpending();
4624 * context_switch - switch to the new MM and the new thread's register state.
4626 static __always_inline struct rq *
4627 context_switch(struct rq *rq, struct task_struct *prev,
4628 struct task_struct *next, struct rq_flags *rf)
4630 prepare_task_switch(rq, prev, next);
4633 * For paravirt, this is coupled with an exit in switch_to to
4634 * combine the page table reload and the switch backend into
4637 arch_start_context_switch(prev);
4640 * kernel -> kernel lazy + transfer active
4641 * user -> kernel lazy + mmgrab() active
4643 * kernel -> user switch + mmdrop() active
4644 * user -> user switch
4646 if (!next->mm) { // to kernel
4647 enter_lazy_tlb(prev->active_mm, next);
4649 next->active_mm = prev->active_mm;
4650 if (prev->mm) // from user
4651 mmgrab(prev->active_mm);
4653 prev->active_mm = NULL;
4655 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4657 * sys_membarrier() requires an smp_mb() between setting
4658 * rq->curr / membarrier_switch_mm() and returning to userspace.
4660 * The below provides this either through switch_mm(), or in
4661 * case 'prev->active_mm == next->mm' through
4662 * finish_task_switch()'s mmdrop().
4664 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4666 if (!prev->mm) { // from kernel
4667 /* will mmdrop() in finish_task_switch(). */
4668 rq->prev_mm = prev->active_mm;
4669 prev->active_mm = NULL;
4673 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4675 prepare_lock_switch(rq, next, rf);
4677 /* Here we just switch the register state and the stack. */
4678 switch_to(prev, next, prev);
4681 return finish_task_switch(prev);
4685 * nr_running and nr_context_switches:
4687 * externally visible scheduler statistics: current number of runnable
4688 * threads, total number of context switches performed since bootup.
4690 unsigned int nr_running(void)
4692 unsigned int i, sum = 0;
4694 for_each_online_cpu(i)
4695 sum += cpu_rq(i)->nr_running;
4701 * Check if only the current task is running on the CPU.
4703 * Caution: this function does not check that the caller has disabled
4704 * preemption, thus the result might have a time-of-check-to-time-of-use
4705 * race. The caller is responsible to use it correctly, for example:
4707 * - from a non-preemptible section (of course)
4709 * - from a thread that is bound to a single CPU
4711 * - in a loop with very short iterations (e.g. a polling loop)
4713 bool single_task_running(void)
4715 return raw_rq()->nr_running == 1;
4717 EXPORT_SYMBOL(single_task_running);
4719 unsigned long long nr_context_switches(void)
4722 unsigned long long sum = 0;
4724 for_each_possible_cpu(i)
4725 sum += cpu_rq(i)->nr_switches;
4731 * Consumers of these two interfaces, like for example the cpuidle menu
4732 * governor, are using nonsensical data. Preferring shallow idle state selection
4733 * for a CPU that has IO-wait which might not even end up running the task when
4734 * it does become runnable.
4737 unsigned int nr_iowait_cpu(int cpu)
4739 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4743 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4745 * The idea behind IO-wait account is to account the idle time that we could
4746 * have spend running if it were not for IO. That is, if we were to improve the
4747 * storage performance, we'd have a proportional reduction in IO-wait time.
4749 * This all works nicely on UP, where, when a task blocks on IO, we account
4750 * idle time as IO-wait, because if the storage were faster, it could've been
4751 * running and we'd not be idle.
4753 * This has been extended to SMP, by doing the same for each CPU. This however
4756 * Imagine for instance the case where two tasks block on one CPU, only the one
4757 * CPU will have IO-wait accounted, while the other has regular idle. Even
4758 * though, if the storage were faster, both could've ran at the same time,
4759 * utilising both CPUs.
4761 * This means, that when looking globally, the current IO-wait accounting on
4762 * SMP is a lower bound, by reason of under accounting.
4764 * Worse, since the numbers are provided per CPU, they are sometimes
4765 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4766 * associated with any one particular CPU, it can wake to another CPU than it
4767 * blocked on. This means the per CPU IO-wait number is meaningless.
4769 * Task CPU affinities can make all that even more 'interesting'.
4772 unsigned int nr_iowait(void)
4774 unsigned int i, sum = 0;
4776 for_each_possible_cpu(i)
4777 sum += nr_iowait_cpu(i);
4785 * sched_exec - execve() is a valuable balancing opportunity, because at
4786 * this point the task has the smallest effective memory and cache footprint.
4788 void sched_exec(void)
4790 struct task_struct *p = current;
4791 unsigned long flags;
4794 raw_spin_lock_irqsave(&p->pi_lock, flags);
4795 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4796 if (dest_cpu == smp_processor_id())
4799 if (likely(cpu_active(dest_cpu))) {
4800 struct migration_arg arg = { p, dest_cpu };
4802 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4803 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4807 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4812 DEFINE_PER_CPU(struct kernel_stat, kstat);
4813 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4815 EXPORT_PER_CPU_SYMBOL(kstat);
4816 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4819 * The function fair_sched_class.update_curr accesses the struct curr
4820 * and its field curr->exec_start; when called from task_sched_runtime(),
4821 * we observe a high rate of cache misses in practice.
4822 * Prefetching this data results in improved performance.
4824 static inline void prefetch_curr_exec_start(struct task_struct *p)
4826 #ifdef CONFIG_FAIR_GROUP_SCHED
4827 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4829 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4832 prefetch(&curr->exec_start);
4836 * Return accounted runtime for the task.
4837 * In case the task is currently running, return the runtime plus current's
4838 * pending runtime that have not been accounted yet.
4840 unsigned long long task_sched_runtime(struct task_struct *p)
4846 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4848 * 64-bit doesn't need locks to atomically read a 64-bit value.
4849 * So we have a optimization chance when the task's delta_exec is 0.
4850 * Reading ->on_cpu is racy, but this is ok.
4852 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4853 * If we race with it entering CPU, unaccounted time is 0. This is
4854 * indistinguishable from the read occurring a few cycles earlier.
4855 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4856 * been accounted, so we're correct here as well.
4858 if (!p->on_cpu || !task_on_rq_queued(p))
4859 return p->se.sum_exec_runtime;
4862 rq = task_rq_lock(p, &rf);
4864 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4865 * project cycles that may never be accounted to this
4866 * thread, breaking clock_gettime().
4868 if (task_current(rq, p) && task_on_rq_queued(p)) {
4869 prefetch_curr_exec_start(p);
4870 update_rq_clock(rq);
4871 p->sched_class->update_curr(rq);
4873 ns = p->se.sum_exec_runtime;
4874 task_rq_unlock(rq, p, &rf);
4879 #ifdef CONFIG_SCHED_DEBUG
4880 static u64 cpu_resched_latency(struct rq *rq)
4882 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
4883 u64 resched_latency, now = rq_clock(rq);
4884 static bool warned_once;
4886 if (sysctl_resched_latency_warn_once && warned_once)
4889 if (!need_resched() || !latency_warn_ms)
4892 if (system_state == SYSTEM_BOOTING)
4895 if (!rq->last_seen_need_resched_ns) {
4896 rq->last_seen_need_resched_ns = now;
4897 rq->ticks_without_resched = 0;
4901 rq->ticks_without_resched++;
4902 resched_latency = now - rq->last_seen_need_resched_ns;
4903 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
4908 return resched_latency;
4911 static int __init setup_resched_latency_warn_ms(char *str)
4915 if ((kstrtol(str, 0, &val))) {
4916 pr_warn("Unable to set resched_latency_warn_ms\n");
4920 sysctl_resched_latency_warn_ms = val;
4923 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
4925 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
4926 #endif /* CONFIG_SCHED_DEBUG */
4929 * This function gets called by the timer code, with HZ frequency.
4930 * We call it with interrupts disabled.
4932 void scheduler_tick(void)
4934 int cpu = smp_processor_id();
4935 struct rq *rq = cpu_rq(cpu);
4936 struct task_struct *curr = rq->curr;
4938 unsigned long thermal_pressure;
4939 u64 resched_latency;
4941 arch_scale_freq_tick();
4946 update_rq_clock(rq);
4947 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4948 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4949 curr->sched_class->task_tick(rq, curr, 0);
4950 if (sched_feat(LATENCY_WARN))
4951 resched_latency = cpu_resched_latency(rq);
4952 calc_global_load_tick(rq);
4956 if (sched_feat(LATENCY_WARN) && resched_latency)
4957 resched_latency_warn(cpu, resched_latency);
4959 perf_event_task_tick();
4962 rq->idle_balance = idle_cpu(cpu);
4963 trigger_load_balance(rq);
4967 #ifdef CONFIG_NO_HZ_FULL
4972 struct delayed_work work;
4974 /* Values for ->state, see diagram below. */
4975 #define TICK_SCHED_REMOTE_OFFLINE 0
4976 #define TICK_SCHED_REMOTE_OFFLINING 1
4977 #define TICK_SCHED_REMOTE_RUNNING 2
4980 * State diagram for ->state:
4983 * TICK_SCHED_REMOTE_OFFLINE
4986 * | | sched_tick_remote()
4989 * +--TICK_SCHED_REMOTE_OFFLINING
4992 * sched_tick_start() | | sched_tick_stop()
4995 * TICK_SCHED_REMOTE_RUNNING
4998 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4999 * and sched_tick_start() are happy to leave the state in RUNNING.
5002 static struct tick_work __percpu *tick_work_cpu;
5004 static void sched_tick_remote(struct work_struct *work)
5006 struct delayed_work *dwork = to_delayed_work(work);
5007 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5008 int cpu = twork->cpu;
5009 struct rq *rq = cpu_rq(cpu);
5010 struct task_struct *curr;
5016 * Handle the tick only if it appears the remote CPU is running in full
5017 * dynticks mode. The check is racy by nature, but missing a tick or
5018 * having one too much is no big deal because the scheduler tick updates
5019 * statistics and checks timeslices in a time-independent way, regardless
5020 * of when exactly it is running.
5022 if (!tick_nohz_tick_stopped_cpu(cpu))
5025 rq_lock_irq(rq, &rf);
5027 if (cpu_is_offline(cpu))
5030 update_rq_clock(rq);
5032 if (!is_idle_task(curr)) {
5034 * Make sure the next tick runs within a reasonable
5037 delta = rq_clock_task(rq) - curr->se.exec_start;
5038 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5040 curr->sched_class->task_tick(rq, curr, 0);
5042 calc_load_nohz_remote(rq);
5044 rq_unlock_irq(rq, &rf);
5048 * Run the remote tick once per second (1Hz). This arbitrary
5049 * frequency is large enough to avoid overload but short enough
5050 * to keep scheduler internal stats reasonably up to date. But
5051 * first update state to reflect hotplug activity if required.
5053 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5054 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5055 if (os == TICK_SCHED_REMOTE_RUNNING)
5056 queue_delayed_work(system_unbound_wq, dwork, HZ);
5059 static void sched_tick_start(int cpu)
5062 struct tick_work *twork;
5064 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5067 WARN_ON_ONCE(!tick_work_cpu);
5069 twork = per_cpu_ptr(tick_work_cpu, cpu);
5070 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5071 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5072 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5074 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5075 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5079 #ifdef CONFIG_HOTPLUG_CPU
5080 static void sched_tick_stop(int cpu)
5082 struct tick_work *twork;
5085 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5088 WARN_ON_ONCE(!tick_work_cpu);
5090 twork = per_cpu_ptr(tick_work_cpu, cpu);
5091 /* There cannot be competing actions, but don't rely on stop-machine. */
5092 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5093 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5094 /* Don't cancel, as this would mess up the state machine. */
5096 #endif /* CONFIG_HOTPLUG_CPU */
5098 int __init sched_tick_offload_init(void)
5100 tick_work_cpu = alloc_percpu(struct tick_work);
5101 BUG_ON(!tick_work_cpu);
5105 #else /* !CONFIG_NO_HZ_FULL */
5106 static inline void sched_tick_start(int cpu) { }
5107 static inline void sched_tick_stop(int cpu) { }
5110 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5111 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5113 * If the value passed in is equal to the current preempt count
5114 * then we just disabled preemption. Start timing the latency.
5116 static inline void preempt_latency_start(int val)
5118 if (preempt_count() == val) {
5119 unsigned long ip = get_lock_parent_ip();
5120 #ifdef CONFIG_DEBUG_PREEMPT
5121 current->preempt_disable_ip = ip;
5123 trace_preempt_off(CALLER_ADDR0, ip);
5127 void preempt_count_add(int val)
5129 #ifdef CONFIG_DEBUG_PREEMPT
5133 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5136 __preempt_count_add(val);
5137 #ifdef CONFIG_DEBUG_PREEMPT
5139 * Spinlock count overflowing soon?
5141 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5144 preempt_latency_start(val);
5146 EXPORT_SYMBOL(preempt_count_add);
5147 NOKPROBE_SYMBOL(preempt_count_add);
5150 * If the value passed in equals to the current preempt count
5151 * then we just enabled preemption. Stop timing the latency.
5153 static inline void preempt_latency_stop(int val)
5155 if (preempt_count() == val)
5156 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5159 void preempt_count_sub(int val)
5161 #ifdef CONFIG_DEBUG_PREEMPT
5165 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5168 * Is the spinlock portion underflowing?
5170 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5171 !(preempt_count() & PREEMPT_MASK)))
5175 preempt_latency_stop(val);
5176 __preempt_count_sub(val);
5178 EXPORT_SYMBOL(preempt_count_sub);
5179 NOKPROBE_SYMBOL(preempt_count_sub);
5182 static inline void preempt_latency_start(int val) { }
5183 static inline void preempt_latency_stop(int val) { }
5186 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5188 #ifdef CONFIG_DEBUG_PREEMPT
5189 return p->preempt_disable_ip;
5196 * Print scheduling while atomic bug:
5198 static noinline void __schedule_bug(struct task_struct *prev)
5200 /* Save this before calling printk(), since that will clobber it */
5201 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5203 if (oops_in_progress)
5206 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5207 prev->comm, prev->pid, preempt_count());
5209 debug_show_held_locks(prev);
5211 if (irqs_disabled())
5212 print_irqtrace_events(prev);
5213 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5214 && in_atomic_preempt_off()) {
5215 pr_err("Preemption disabled at:");
5216 print_ip_sym(KERN_ERR, preempt_disable_ip);
5219 panic("scheduling while atomic\n");
5222 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5226 * Various schedule()-time debugging checks and statistics:
5228 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5230 #ifdef CONFIG_SCHED_STACK_END_CHECK
5231 if (task_stack_end_corrupted(prev))
5232 panic("corrupted stack end detected inside scheduler\n");
5234 if (task_scs_end_corrupted(prev))
5235 panic("corrupted shadow stack detected inside scheduler\n");
5238 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5239 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5240 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5241 prev->comm, prev->pid, prev->non_block_count);
5243 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5247 if (unlikely(in_atomic_preempt_off())) {
5248 __schedule_bug(prev);
5249 preempt_count_set(PREEMPT_DISABLED);
5252 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5254 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5256 schedstat_inc(this_rq()->sched_count);
5259 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5260 struct rq_flags *rf)
5263 const struct sched_class *class;
5265 * We must do the balancing pass before put_prev_task(), such
5266 * that when we release the rq->lock the task is in the same
5267 * state as before we took rq->lock.
5269 * We can terminate the balance pass as soon as we know there is
5270 * a runnable task of @class priority or higher.
5272 for_class_range(class, prev->sched_class, &idle_sched_class) {
5273 if (class->balance(rq, prev, rf))
5278 put_prev_task(rq, prev);
5282 * Pick up the highest-prio task:
5284 static inline struct task_struct *
5285 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5287 const struct sched_class *class;
5288 struct task_struct *p;
5291 * Optimization: we know that if all tasks are in the fair class we can
5292 * call that function directly, but only if the @prev task wasn't of a
5293 * higher scheduling class, because otherwise those lose the
5294 * opportunity to pull in more work from other CPUs.
5296 if (likely(prev->sched_class <= &fair_sched_class &&
5297 rq->nr_running == rq->cfs.h_nr_running)) {
5299 p = pick_next_task_fair(rq, prev, rf);
5300 if (unlikely(p == RETRY_TASK))
5303 /* Assume the next prioritized class is idle_sched_class */
5305 put_prev_task(rq, prev);
5306 p = pick_next_task_idle(rq);
5313 put_prev_task_balance(rq, prev, rf);
5315 for_each_class(class) {
5316 p = class->pick_next_task(rq);
5321 /* The idle class should always have a runnable task: */
5325 #ifdef CONFIG_SCHED_CORE
5326 static inline bool is_task_rq_idle(struct task_struct *t)
5328 return (task_rq(t)->idle == t);
5331 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5333 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5336 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5338 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5341 return a->core_cookie == b->core_cookie;
5344 // XXX fairness/fwd progress conditions
5347 * - NULL if there is no runnable task for this class.
5348 * - the highest priority task for this runqueue if it matches
5349 * rq->core->core_cookie or its priority is greater than max.
5350 * - Else returns idle_task.
5352 static struct task_struct *
5353 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5355 struct task_struct *class_pick, *cookie_pick;
5356 unsigned long cookie = rq->core->core_cookie;
5358 class_pick = class->pick_task(rq);
5364 * If class_pick is tagged, return it only if it has
5365 * higher priority than max.
5367 if (max && class_pick->core_cookie &&
5368 prio_less(class_pick, max, in_fi))
5369 return idle_sched_class.pick_task(rq);
5375 * If class_pick is idle or matches cookie, return early.
5377 if (cookie_equals(class_pick, cookie))
5380 cookie_pick = sched_core_find(rq, cookie);
5383 * If class > max && class > cookie, it is the highest priority task on
5384 * the core (so far) and it must be selected, otherwise we must go with
5385 * the cookie pick in order to satisfy the constraint.
5387 if (prio_less(cookie_pick, class_pick, in_fi) &&
5388 (!max || prio_less(max, class_pick, in_fi)))
5394 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5396 static struct task_struct *
5397 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5399 struct task_struct *next, *max = NULL;
5400 const struct sched_class *class;
5401 const struct cpumask *smt_mask;
5402 bool fi_before = false;
5403 int i, j, cpu, occ = 0;
5406 if (!sched_core_enabled(rq))
5407 return __pick_next_task(rq, prev, rf);
5411 /* Stopper task is switching into idle, no need core-wide selection. */
5412 if (cpu_is_offline(cpu)) {
5414 * Reset core_pick so that we don't enter the fastpath when
5415 * coming online. core_pick would already be migrated to
5416 * another cpu during offline.
5418 rq->core_pick = NULL;
5419 return __pick_next_task(rq, prev, rf);
5423 * If there were no {en,de}queues since we picked (IOW, the task
5424 * pointers are all still valid), and we haven't scheduled the last
5425 * pick yet, do so now.
5427 * rq->core_pick can be NULL if no selection was made for a CPU because
5428 * it was either offline or went offline during a sibling's core-wide
5429 * selection. In this case, do a core-wide selection.
5431 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5432 rq->core->core_pick_seq != rq->core_sched_seq &&
5434 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5436 next = rq->core_pick;
5438 put_prev_task(rq, prev);
5439 set_next_task(rq, next);
5442 rq->core_pick = NULL;
5446 put_prev_task_balance(rq, prev, rf);
5448 smt_mask = cpu_smt_mask(cpu);
5449 need_sync = !!rq->core->core_cookie;
5452 rq->core->core_cookie = 0UL;
5453 if (rq->core->core_forceidle) {
5456 rq->core->core_forceidle = false;
5460 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5462 * @task_seq guards the task state ({en,de}queues)
5463 * @pick_seq is the @task_seq we did a selection on
5464 * @sched_seq is the @pick_seq we scheduled
5466 * However, preemptions can cause multiple picks on the same task set.
5467 * 'Fix' this by also increasing @task_seq for every pick.
5469 rq->core->core_task_seq++;
5472 * Optimize for common case where this CPU has no cookies
5473 * and there are no cookied tasks running on siblings.
5476 for_each_class(class) {
5477 next = class->pick_task(rq);
5482 if (!next->core_cookie) {
5483 rq->core_pick = NULL;
5485 * For robustness, update the min_vruntime_fi for
5486 * unconstrained picks as well.
5488 WARN_ON_ONCE(fi_before);
5489 task_vruntime_update(rq, next, false);
5494 for_each_cpu(i, smt_mask) {
5495 struct rq *rq_i = cpu_rq(i);
5497 rq_i->core_pick = NULL;
5500 update_rq_clock(rq_i);
5504 * Try and select tasks for each sibling in descending sched_class
5507 for_each_class(class) {
5509 for_each_cpu_wrap(i, smt_mask, cpu) {
5510 struct rq *rq_i = cpu_rq(i);
5511 struct task_struct *p;
5513 if (rq_i->core_pick)
5517 * If this sibling doesn't yet have a suitable task to
5518 * run; ask for the most eligible task, given the
5519 * highest priority task already selected for this
5522 p = pick_task(rq_i, class, max, fi_before);
5526 if (!is_task_rq_idle(p))
5529 rq_i->core_pick = p;
5530 if (rq_i->idle == p && rq_i->nr_running) {
5531 rq->core->core_forceidle = true;
5533 rq->core->core_forceidle_seq++;
5537 * If this new candidate is of higher priority than the
5538 * previous; and they're incompatible; we need to wipe
5539 * the slate and start over. pick_task makes sure that
5540 * p's priority is more than max if it doesn't match
5543 * NOTE: this is a linear max-filter and is thus bounded
5544 * in execution time.
5546 if (!max || !cookie_match(max, p)) {
5547 struct task_struct *old_max = max;
5549 rq->core->core_cookie = p->core_cookie;
5553 rq->core->core_forceidle = false;
5554 for_each_cpu(j, smt_mask) {
5558 cpu_rq(j)->core_pick = NULL;
5567 rq->core->core_pick_seq = rq->core->core_task_seq;
5568 next = rq->core_pick;
5569 rq->core_sched_seq = rq->core->core_pick_seq;
5571 /* Something should have been selected for current CPU */
5572 WARN_ON_ONCE(!next);
5575 * Reschedule siblings
5577 * NOTE: L1TF -- at this point we're no longer running the old task and
5578 * sending an IPI (below) ensures the sibling will no longer be running
5579 * their task. This ensures there is no inter-sibling overlap between
5580 * non-matching user state.
5582 for_each_cpu(i, smt_mask) {
5583 struct rq *rq_i = cpu_rq(i);
5586 * An online sibling might have gone offline before a task
5587 * could be picked for it, or it might be offline but later
5588 * happen to come online, but its too late and nothing was
5589 * picked for it. That's Ok - it will pick tasks for itself,
5592 if (!rq_i->core_pick)
5596 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5597 * fi_before fi update?
5603 if (!(fi_before && rq->core->core_forceidle))
5604 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5606 rq_i->core_pick->core_occupation = occ;
5609 rq_i->core_pick = NULL;
5613 /* Did we break L1TF mitigation requirements? */
5614 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5616 if (rq_i->curr == rq_i->core_pick) {
5617 rq_i->core_pick = NULL;
5625 set_next_task(rq, next);
5629 static bool try_steal_cookie(int this, int that)
5631 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5632 struct task_struct *p;
5633 unsigned long cookie;
5634 bool success = false;
5636 local_irq_disable();
5637 double_rq_lock(dst, src);
5639 cookie = dst->core->core_cookie;
5643 if (dst->curr != dst->idle)
5646 p = sched_core_find(src, cookie);
5651 if (p == src->core_pick || p == src->curr)
5654 if (!cpumask_test_cpu(this, &p->cpus_mask))
5657 if (p->core_occupation > dst->idle->core_occupation)
5660 p->on_rq = TASK_ON_RQ_MIGRATING;
5661 deactivate_task(src, p, 0);
5662 set_task_cpu(p, this);
5663 activate_task(dst, p, 0);
5664 p->on_rq = TASK_ON_RQ_QUEUED;
5672 p = sched_core_next(p, cookie);
5676 double_rq_unlock(dst, src);
5682 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5686 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5693 if (try_steal_cookie(cpu, i))
5700 static void sched_core_balance(struct rq *rq)
5702 struct sched_domain *sd;
5703 int cpu = cpu_of(rq);
5707 raw_spin_rq_unlock_irq(rq);
5708 for_each_domain(cpu, sd) {
5712 if (steal_cookie_task(cpu, sd))
5715 raw_spin_rq_lock_irq(rq);
5720 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5722 void queue_core_balance(struct rq *rq)
5724 if (!sched_core_enabled(rq))
5727 if (!rq->core->core_cookie)
5730 if (!rq->nr_running) /* not forced idle */
5733 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5736 static inline void sched_core_cpu_starting(unsigned int cpu)
5738 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5739 struct rq *rq, *core_rq = NULL;
5742 core_rq = cpu_rq(cpu)->core;
5745 for_each_cpu(i, smt_mask) {
5747 if (rq->core && rq->core == rq)
5752 core_rq = cpu_rq(cpu);
5754 for_each_cpu(i, smt_mask) {
5757 WARN_ON_ONCE(rq->core && rq->core != core_rq);
5762 #else /* !CONFIG_SCHED_CORE */
5764 static inline void sched_core_cpu_starting(unsigned int cpu) {}
5766 static struct task_struct *
5767 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5769 return __pick_next_task(rq, prev, rf);
5772 #endif /* CONFIG_SCHED_CORE */
5775 * __schedule() is the main scheduler function.
5777 * The main means of driving the scheduler and thus entering this function are:
5779 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
5781 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
5782 * paths. For example, see arch/x86/entry_64.S.
5784 * To drive preemption between tasks, the scheduler sets the flag in timer
5785 * interrupt handler scheduler_tick().
5787 * 3. Wakeups don't really cause entry into schedule(). They add a
5788 * task to the run-queue and that's it.
5790 * Now, if the new task added to the run-queue preempts the current
5791 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
5792 * called on the nearest possible occasion:
5794 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
5796 * - in syscall or exception context, at the next outmost
5797 * preempt_enable(). (this might be as soon as the wake_up()'s
5800 * - in IRQ context, return from interrupt-handler to
5801 * preemptible context
5803 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
5806 * - cond_resched() call
5807 * - explicit schedule() call
5808 * - return from syscall or exception to user-space
5809 * - return from interrupt-handler to user-space
5811 * WARNING: must be called with preemption disabled!
5813 static void __sched notrace __schedule(bool preempt)
5815 struct task_struct *prev, *next;
5816 unsigned long *switch_count;
5817 unsigned long prev_state;
5822 cpu = smp_processor_id();
5826 schedule_debug(prev, preempt);
5828 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
5831 local_irq_disable();
5832 rcu_note_context_switch(preempt);
5835 * Make sure that signal_pending_state()->signal_pending() below
5836 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
5837 * done by the caller to avoid the race with signal_wake_up():
5839 * __set_current_state(@state) signal_wake_up()
5840 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
5841 * wake_up_state(p, state)
5842 * LOCK rq->lock LOCK p->pi_state
5843 * smp_mb__after_spinlock() smp_mb__after_spinlock()
5844 * if (signal_pending_state()) if (p->state & @state)
5846 * Also, the membarrier system call requires a full memory barrier
5847 * after coming from user-space, before storing to rq->curr.
5850 smp_mb__after_spinlock();
5852 /* Promote REQ to ACT */
5853 rq->clock_update_flags <<= 1;
5854 update_rq_clock(rq);
5856 switch_count = &prev->nivcsw;
5859 * We must load prev->state once (task_struct::state is volatile), such
5862 * - we form a control dependency vs deactivate_task() below.
5863 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
5865 prev_state = READ_ONCE(prev->__state);
5866 if (!preempt && prev_state) {
5867 if (signal_pending_state(prev_state, prev)) {
5868 WRITE_ONCE(prev->__state, TASK_RUNNING);
5870 prev->sched_contributes_to_load =
5871 (prev_state & TASK_UNINTERRUPTIBLE) &&
5872 !(prev_state & TASK_NOLOAD) &&
5873 !(prev->flags & PF_FROZEN);
5875 if (prev->sched_contributes_to_load)
5876 rq->nr_uninterruptible++;
5879 * __schedule() ttwu()
5880 * prev_state = prev->state; if (p->on_rq && ...)
5881 * if (prev_state) goto out;
5882 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
5883 * p->state = TASK_WAKING
5885 * Where __schedule() and ttwu() have matching control dependencies.
5887 * After this, schedule() must not care about p->state any more.
5889 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
5891 if (prev->in_iowait) {
5892 atomic_inc(&rq->nr_iowait);
5893 delayacct_blkio_start();
5896 switch_count = &prev->nvcsw;
5899 next = pick_next_task(rq, prev, &rf);
5900 clear_tsk_need_resched(prev);
5901 clear_preempt_need_resched();
5902 #ifdef CONFIG_SCHED_DEBUG
5903 rq->last_seen_need_resched_ns = 0;
5906 if (likely(prev != next)) {
5909 * RCU users of rcu_dereference(rq->curr) may not see
5910 * changes to task_struct made by pick_next_task().
5912 RCU_INIT_POINTER(rq->curr, next);
5914 * The membarrier system call requires each architecture
5915 * to have a full memory barrier after updating
5916 * rq->curr, before returning to user-space.
5918 * Here are the schemes providing that barrier on the
5919 * various architectures:
5920 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5921 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5922 * - finish_lock_switch() for weakly-ordered
5923 * architectures where spin_unlock is a full barrier,
5924 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5925 * is a RELEASE barrier),
5929 migrate_disable_switch(rq, prev);
5930 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5932 trace_sched_switch(preempt, prev, next);
5934 /* Also unlocks the rq: */
5935 rq = context_switch(rq, prev, next, &rf);
5937 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5939 rq_unpin_lock(rq, &rf);
5940 __balance_callbacks(rq);
5941 raw_spin_rq_unlock_irq(rq);
5945 void __noreturn do_task_dead(void)
5947 /* Causes final put_task_struct in finish_task_switch(): */
5948 set_special_state(TASK_DEAD);
5950 /* Tell freezer to ignore us: */
5951 current->flags |= PF_NOFREEZE;
5956 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5961 static inline void sched_submit_work(struct task_struct *tsk)
5963 unsigned int task_flags;
5965 if (task_is_running(tsk))
5968 task_flags = tsk->flags;
5970 * If a worker went to sleep, notify and ask workqueue whether
5971 * it wants to wake up a task to maintain concurrency.
5972 * As this function is called inside the schedule() context,
5973 * we disable preemption to avoid it calling schedule() again
5974 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5977 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5979 if (task_flags & PF_WQ_WORKER)
5980 wq_worker_sleeping(tsk);
5982 io_wq_worker_sleeping(tsk);
5983 preempt_enable_no_resched();
5986 if (tsk_is_pi_blocked(tsk))
5990 * If we are going to sleep and we have plugged IO queued,
5991 * make sure to submit it to avoid deadlocks.
5993 if (blk_needs_flush_plug(tsk))
5994 blk_schedule_flush_plug(tsk);
5997 static void sched_update_worker(struct task_struct *tsk)
5999 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6000 if (tsk->flags & PF_WQ_WORKER)
6001 wq_worker_running(tsk);
6003 io_wq_worker_running(tsk);
6007 asmlinkage __visible void __sched schedule(void)
6009 struct task_struct *tsk = current;
6011 sched_submit_work(tsk);
6015 sched_preempt_enable_no_resched();
6016 } while (need_resched());
6017 sched_update_worker(tsk);
6019 EXPORT_SYMBOL(schedule);
6022 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6023 * state (have scheduled out non-voluntarily) by making sure that all
6024 * tasks have either left the run queue or have gone into user space.
6025 * As idle tasks do not do either, they must not ever be preempted
6026 * (schedule out non-voluntarily).
6028 * schedule_idle() is similar to schedule_preempt_disable() except that it
6029 * never enables preemption because it does not call sched_submit_work().
6031 void __sched schedule_idle(void)
6034 * As this skips calling sched_submit_work(), which the idle task does
6035 * regardless because that function is a nop when the task is in a
6036 * TASK_RUNNING state, make sure this isn't used someplace that the
6037 * current task can be in any other state. Note, idle is always in the
6038 * TASK_RUNNING state.
6040 WARN_ON_ONCE(current->__state);
6043 } while (need_resched());
6046 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6047 asmlinkage __visible void __sched schedule_user(void)
6050 * If we come here after a random call to set_need_resched(),
6051 * or we have been woken up remotely but the IPI has not yet arrived,
6052 * we haven't yet exited the RCU idle mode. Do it here manually until
6053 * we find a better solution.
6055 * NB: There are buggy callers of this function. Ideally we
6056 * should warn if prev_state != CONTEXT_USER, but that will trigger
6057 * too frequently to make sense yet.
6059 enum ctx_state prev_state = exception_enter();
6061 exception_exit(prev_state);
6066 * schedule_preempt_disabled - called with preemption disabled
6068 * Returns with preemption disabled. Note: preempt_count must be 1
6070 void __sched schedule_preempt_disabled(void)
6072 sched_preempt_enable_no_resched();
6077 static void __sched notrace preempt_schedule_common(void)
6081 * Because the function tracer can trace preempt_count_sub()
6082 * and it also uses preempt_enable/disable_notrace(), if
6083 * NEED_RESCHED is set, the preempt_enable_notrace() called
6084 * by the function tracer will call this function again and
6085 * cause infinite recursion.
6087 * Preemption must be disabled here before the function
6088 * tracer can trace. Break up preempt_disable() into two
6089 * calls. One to disable preemption without fear of being
6090 * traced. The other to still record the preemption latency,
6091 * which can also be traced by the function tracer.
6093 preempt_disable_notrace();
6094 preempt_latency_start(1);
6096 preempt_latency_stop(1);
6097 preempt_enable_no_resched_notrace();
6100 * Check again in case we missed a preemption opportunity
6101 * between schedule and now.
6103 } while (need_resched());
6106 #ifdef CONFIG_PREEMPTION
6108 * This is the entry point to schedule() from in-kernel preemption
6109 * off of preempt_enable.
6111 asmlinkage __visible void __sched notrace preempt_schedule(void)
6114 * If there is a non-zero preempt_count or interrupts are disabled,
6115 * we do not want to preempt the current task. Just return..
6117 if (likely(!preemptible()))
6120 preempt_schedule_common();
6122 NOKPROBE_SYMBOL(preempt_schedule);
6123 EXPORT_SYMBOL(preempt_schedule);
6125 #ifdef CONFIG_PREEMPT_DYNAMIC
6126 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6127 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6132 * preempt_schedule_notrace - preempt_schedule called by tracing
6134 * The tracing infrastructure uses preempt_enable_notrace to prevent
6135 * recursion and tracing preempt enabling caused by the tracing
6136 * infrastructure itself. But as tracing can happen in areas coming
6137 * from userspace or just about to enter userspace, a preempt enable
6138 * can occur before user_exit() is called. This will cause the scheduler
6139 * to be called when the system is still in usermode.
6141 * To prevent this, the preempt_enable_notrace will use this function
6142 * instead of preempt_schedule() to exit user context if needed before
6143 * calling the scheduler.
6145 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6147 enum ctx_state prev_ctx;
6149 if (likely(!preemptible()))
6154 * Because the function tracer can trace preempt_count_sub()
6155 * and it also uses preempt_enable/disable_notrace(), if
6156 * NEED_RESCHED is set, the preempt_enable_notrace() called
6157 * by the function tracer will call this function again and
6158 * cause infinite recursion.
6160 * Preemption must be disabled here before the function
6161 * tracer can trace. Break up preempt_disable() into two
6162 * calls. One to disable preemption without fear of being
6163 * traced. The other to still record the preemption latency,
6164 * which can also be traced by the function tracer.
6166 preempt_disable_notrace();
6167 preempt_latency_start(1);
6169 * Needs preempt disabled in case user_exit() is traced
6170 * and the tracer calls preempt_enable_notrace() causing
6171 * an infinite recursion.
6173 prev_ctx = exception_enter();
6175 exception_exit(prev_ctx);
6177 preempt_latency_stop(1);
6178 preempt_enable_no_resched_notrace();
6179 } while (need_resched());
6181 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6183 #ifdef CONFIG_PREEMPT_DYNAMIC
6184 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6185 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6188 #endif /* CONFIG_PREEMPTION */
6190 #ifdef CONFIG_PREEMPT_DYNAMIC
6192 #include <linux/entry-common.h>
6197 * SC:preempt_schedule
6198 * SC:preempt_schedule_notrace
6199 * SC:irqentry_exit_cond_resched
6203 * cond_resched <- __cond_resched
6204 * might_resched <- RET0
6205 * preempt_schedule <- NOP
6206 * preempt_schedule_notrace <- NOP
6207 * irqentry_exit_cond_resched <- NOP
6210 * cond_resched <- __cond_resched
6211 * might_resched <- __cond_resched
6212 * preempt_schedule <- NOP
6213 * preempt_schedule_notrace <- NOP
6214 * irqentry_exit_cond_resched <- NOP
6217 * cond_resched <- RET0
6218 * might_resched <- RET0
6219 * preempt_schedule <- preempt_schedule
6220 * preempt_schedule_notrace <- preempt_schedule_notrace
6221 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6225 preempt_dynamic_none = 0,
6226 preempt_dynamic_voluntary,
6227 preempt_dynamic_full,
6230 int preempt_dynamic_mode = preempt_dynamic_full;
6232 int sched_dynamic_mode(const char *str)
6234 if (!strcmp(str, "none"))
6235 return preempt_dynamic_none;
6237 if (!strcmp(str, "voluntary"))
6238 return preempt_dynamic_voluntary;
6240 if (!strcmp(str, "full"))
6241 return preempt_dynamic_full;
6246 void sched_dynamic_update(int mode)
6249 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6250 * the ZERO state, which is invalid.
6252 static_call_update(cond_resched, __cond_resched);
6253 static_call_update(might_resched, __cond_resched);
6254 static_call_update(preempt_schedule, __preempt_schedule_func);
6255 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6256 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6259 case preempt_dynamic_none:
6260 static_call_update(cond_resched, __cond_resched);
6261 static_call_update(might_resched, (void *)&__static_call_return0);
6262 static_call_update(preempt_schedule, NULL);
6263 static_call_update(preempt_schedule_notrace, NULL);
6264 static_call_update(irqentry_exit_cond_resched, NULL);
6265 pr_info("Dynamic Preempt: none\n");
6268 case preempt_dynamic_voluntary:
6269 static_call_update(cond_resched, __cond_resched);
6270 static_call_update(might_resched, __cond_resched);
6271 static_call_update(preempt_schedule, NULL);
6272 static_call_update(preempt_schedule_notrace, NULL);
6273 static_call_update(irqentry_exit_cond_resched, NULL);
6274 pr_info("Dynamic Preempt: voluntary\n");
6277 case preempt_dynamic_full:
6278 static_call_update(cond_resched, (void *)&__static_call_return0);
6279 static_call_update(might_resched, (void *)&__static_call_return0);
6280 static_call_update(preempt_schedule, __preempt_schedule_func);
6281 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6282 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6283 pr_info("Dynamic Preempt: full\n");
6287 preempt_dynamic_mode = mode;
6290 static int __init setup_preempt_mode(char *str)
6292 int mode = sched_dynamic_mode(str);
6294 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6298 sched_dynamic_update(mode);
6301 __setup("preempt=", setup_preempt_mode);
6303 #endif /* CONFIG_PREEMPT_DYNAMIC */
6306 * This is the entry point to schedule() from kernel preemption
6307 * off of irq context.
6308 * Note, that this is called and return with irqs disabled. This will
6309 * protect us against recursive calling from irq.
6311 asmlinkage __visible void __sched preempt_schedule_irq(void)
6313 enum ctx_state prev_state;
6315 /* Catch callers which need to be fixed */
6316 BUG_ON(preempt_count() || !irqs_disabled());
6318 prev_state = exception_enter();
6324 local_irq_disable();
6325 sched_preempt_enable_no_resched();
6326 } while (need_resched());
6328 exception_exit(prev_state);
6331 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6334 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6335 return try_to_wake_up(curr->private, mode, wake_flags);
6337 EXPORT_SYMBOL(default_wake_function);
6339 #ifdef CONFIG_RT_MUTEXES
6341 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6344 prio = min(prio, pi_task->prio);
6349 static inline int rt_effective_prio(struct task_struct *p, int prio)
6351 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6353 return __rt_effective_prio(pi_task, prio);
6357 * rt_mutex_setprio - set the current priority of a task
6359 * @pi_task: donor task
6361 * This function changes the 'effective' priority of a task. It does
6362 * not touch ->normal_prio like __setscheduler().
6364 * Used by the rt_mutex code to implement priority inheritance
6365 * logic. Call site only calls if the priority of the task changed.
6367 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6369 int prio, oldprio, queued, running, queue_flag =
6370 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6371 const struct sched_class *prev_class;
6375 /* XXX used to be waiter->prio, not waiter->task->prio */
6376 prio = __rt_effective_prio(pi_task, p->normal_prio);
6379 * If nothing changed; bail early.
6381 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6384 rq = __task_rq_lock(p, &rf);
6385 update_rq_clock(rq);
6387 * Set under pi_lock && rq->lock, such that the value can be used under
6390 * Note that there is loads of tricky to make this pointer cache work
6391 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6392 * ensure a task is de-boosted (pi_task is set to NULL) before the
6393 * task is allowed to run again (and can exit). This ensures the pointer
6394 * points to a blocked task -- which guarantees the task is present.
6396 p->pi_top_task = pi_task;
6399 * For FIFO/RR we only need to set prio, if that matches we're done.
6401 if (prio == p->prio && !dl_prio(prio))
6405 * Idle task boosting is a nono in general. There is one
6406 * exception, when PREEMPT_RT and NOHZ is active:
6408 * The idle task calls get_next_timer_interrupt() and holds
6409 * the timer wheel base->lock on the CPU and another CPU wants
6410 * to access the timer (probably to cancel it). We can safely
6411 * ignore the boosting request, as the idle CPU runs this code
6412 * with interrupts disabled and will complete the lock
6413 * protected section without being interrupted. So there is no
6414 * real need to boost.
6416 if (unlikely(p == rq->idle)) {
6417 WARN_ON(p != rq->curr);
6418 WARN_ON(p->pi_blocked_on);
6422 trace_sched_pi_setprio(p, pi_task);
6425 if (oldprio == prio)
6426 queue_flag &= ~DEQUEUE_MOVE;
6428 prev_class = p->sched_class;
6429 queued = task_on_rq_queued(p);
6430 running = task_current(rq, p);
6432 dequeue_task(rq, p, queue_flag);
6434 put_prev_task(rq, p);
6437 * Boosting condition are:
6438 * 1. -rt task is running and holds mutex A
6439 * --> -dl task blocks on mutex A
6441 * 2. -dl task is running and holds mutex A
6442 * --> -dl task blocks on mutex A and could preempt the
6445 if (dl_prio(prio)) {
6446 if (!dl_prio(p->normal_prio) ||
6447 (pi_task && dl_prio(pi_task->prio) &&
6448 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6449 p->dl.pi_se = pi_task->dl.pi_se;
6450 queue_flag |= ENQUEUE_REPLENISH;
6452 p->dl.pi_se = &p->dl;
6454 p->sched_class = &dl_sched_class;
6455 } else if (rt_prio(prio)) {
6456 if (dl_prio(oldprio))
6457 p->dl.pi_se = &p->dl;
6459 queue_flag |= ENQUEUE_HEAD;
6460 p->sched_class = &rt_sched_class;
6462 if (dl_prio(oldprio))
6463 p->dl.pi_se = &p->dl;
6464 if (rt_prio(oldprio))
6466 p->sched_class = &fair_sched_class;
6472 enqueue_task(rq, p, queue_flag);
6474 set_next_task(rq, p);
6476 check_class_changed(rq, p, prev_class, oldprio);
6478 /* Avoid rq from going away on us: */
6481 rq_unpin_lock(rq, &rf);
6482 __balance_callbacks(rq);
6483 raw_spin_rq_unlock(rq);
6488 static inline int rt_effective_prio(struct task_struct *p, int prio)
6494 void set_user_nice(struct task_struct *p, long nice)
6496 bool queued, running;
6501 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6504 * We have to be careful, if called from sys_setpriority(),
6505 * the task might be in the middle of scheduling on another CPU.
6507 rq = task_rq_lock(p, &rf);
6508 update_rq_clock(rq);
6511 * The RT priorities are set via sched_setscheduler(), but we still
6512 * allow the 'normal' nice value to be set - but as expected
6513 * it won't have any effect on scheduling until the task is
6514 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6516 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6517 p->static_prio = NICE_TO_PRIO(nice);
6520 queued = task_on_rq_queued(p);
6521 running = task_current(rq, p);
6523 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6525 put_prev_task(rq, p);
6527 p->static_prio = NICE_TO_PRIO(nice);
6528 set_load_weight(p, true);
6530 p->prio = effective_prio(p);
6533 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6535 set_next_task(rq, p);
6538 * If the task increased its priority or is running and
6539 * lowered its priority, then reschedule its CPU:
6541 p->sched_class->prio_changed(rq, p, old_prio);
6544 task_rq_unlock(rq, p, &rf);
6546 EXPORT_SYMBOL(set_user_nice);
6549 * can_nice - check if a task can reduce its nice value
6553 int can_nice(const struct task_struct *p, const int nice)
6555 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6556 int nice_rlim = nice_to_rlimit(nice);
6558 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6559 capable(CAP_SYS_NICE));
6562 #ifdef __ARCH_WANT_SYS_NICE
6565 * sys_nice - change the priority of the current process.
6566 * @increment: priority increment
6568 * sys_setpriority is a more generic, but much slower function that
6569 * does similar things.
6571 SYSCALL_DEFINE1(nice, int, increment)
6576 * Setpriority might change our priority at the same moment.
6577 * We don't have to worry. Conceptually one call occurs first
6578 * and we have a single winner.
6580 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6581 nice = task_nice(current) + increment;
6583 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6584 if (increment < 0 && !can_nice(current, nice))
6587 retval = security_task_setnice(current, nice);
6591 set_user_nice(current, nice);
6598 * task_prio - return the priority value of a given task.
6599 * @p: the task in question.
6601 * Return: The priority value as seen by users in /proc.
6603 * sched policy return value kernel prio user prio/nice
6605 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6606 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6607 * deadline -101 -1 0
6609 int task_prio(const struct task_struct *p)
6611 return p->prio - MAX_RT_PRIO;
6615 * idle_cpu - is a given CPU idle currently?
6616 * @cpu: the processor in question.
6618 * Return: 1 if the CPU is currently idle. 0 otherwise.
6620 int idle_cpu(int cpu)
6622 struct rq *rq = cpu_rq(cpu);
6624 if (rq->curr != rq->idle)
6631 if (rq->ttwu_pending)
6639 * available_idle_cpu - is a given CPU idle for enqueuing work.
6640 * @cpu: the CPU in question.
6642 * Return: 1 if the CPU is currently idle. 0 otherwise.
6644 int available_idle_cpu(int cpu)
6649 if (vcpu_is_preempted(cpu))
6656 * idle_task - return the idle task for a given CPU.
6657 * @cpu: the processor in question.
6659 * Return: The idle task for the CPU @cpu.
6661 struct task_struct *idle_task(int cpu)
6663 return cpu_rq(cpu)->idle;
6668 * This function computes an effective utilization for the given CPU, to be
6669 * used for frequency selection given the linear relation: f = u * f_max.
6671 * The scheduler tracks the following metrics:
6673 * cpu_util_{cfs,rt,dl,irq}()
6676 * Where the cfs,rt and dl util numbers are tracked with the same metric and
6677 * synchronized windows and are thus directly comparable.
6679 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
6680 * which excludes things like IRQ and steal-time. These latter are then accrued
6681 * in the irq utilization.
6683 * The DL bandwidth number otoh is not a measured metric but a value computed
6684 * based on the task model parameters and gives the minimal utilization
6685 * required to meet deadlines.
6687 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
6688 unsigned long max, enum cpu_util_type type,
6689 struct task_struct *p)
6691 unsigned long dl_util, util, irq;
6692 struct rq *rq = cpu_rq(cpu);
6694 if (!uclamp_is_used() &&
6695 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
6700 * Early check to see if IRQ/steal time saturates the CPU, can be
6701 * because of inaccuracies in how we track these -- see
6702 * update_irq_load_avg().
6704 irq = cpu_util_irq(rq);
6705 if (unlikely(irq >= max))
6709 * Because the time spend on RT/DL tasks is visible as 'lost' time to
6710 * CFS tasks and we use the same metric to track the effective
6711 * utilization (PELT windows are synchronized) we can directly add them
6712 * to obtain the CPU's actual utilization.
6714 * CFS and RT utilization can be boosted or capped, depending on
6715 * utilization clamp constraints requested by currently RUNNABLE
6717 * When there are no CFS RUNNABLE tasks, clamps are released and
6718 * frequency will be gracefully reduced with the utilization decay.
6720 util = util_cfs + cpu_util_rt(rq);
6721 if (type == FREQUENCY_UTIL)
6722 util = uclamp_rq_util_with(rq, util, p);
6724 dl_util = cpu_util_dl(rq);
6727 * For frequency selection we do not make cpu_util_dl() a permanent part
6728 * of this sum because we want to use cpu_bw_dl() later on, but we need
6729 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
6730 * that we select f_max when there is no idle time.
6732 * NOTE: numerical errors or stop class might cause us to not quite hit
6733 * saturation when we should -- something for later.
6735 if (util + dl_util >= max)
6739 * OTOH, for energy computation we need the estimated running time, so
6740 * include util_dl and ignore dl_bw.
6742 if (type == ENERGY_UTIL)
6746 * There is still idle time; further improve the number by using the
6747 * irq metric. Because IRQ/steal time is hidden from the task clock we
6748 * need to scale the task numbers:
6751 * U' = irq + --------- * U
6754 util = scale_irq_capacity(util, irq, max);
6758 * Bandwidth required by DEADLINE must always be granted while, for
6759 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
6760 * to gracefully reduce the frequency when no tasks show up for longer
6763 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
6764 * bw_dl as requested freq. However, cpufreq is not yet ready for such
6765 * an interface. So, we only do the latter for now.
6767 if (type == FREQUENCY_UTIL)
6768 util += cpu_bw_dl(rq);
6770 return min(max, util);
6773 unsigned long sched_cpu_util(int cpu, unsigned long max)
6775 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
6778 #endif /* CONFIG_SMP */
6781 * find_process_by_pid - find a process with a matching PID value.
6782 * @pid: the pid in question.
6784 * The task of @pid, if found. %NULL otherwise.
6786 static struct task_struct *find_process_by_pid(pid_t pid)
6788 return pid ? find_task_by_vpid(pid) : current;
6792 * sched_setparam() passes in -1 for its policy, to let the functions
6793 * it calls know not to change it.
6795 #define SETPARAM_POLICY -1
6797 static void __setscheduler_params(struct task_struct *p,
6798 const struct sched_attr *attr)
6800 int policy = attr->sched_policy;
6802 if (policy == SETPARAM_POLICY)
6807 if (dl_policy(policy))
6808 __setparam_dl(p, attr);
6809 else if (fair_policy(policy))
6810 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
6813 * __sched_setscheduler() ensures attr->sched_priority == 0 when
6814 * !rt_policy. Always setting this ensures that things like
6815 * getparam()/getattr() don't report silly values for !rt tasks.
6817 p->rt_priority = attr->sched_priority;
6818 p->normal_prio = normal_prio(p);
6819 set_load_weight(p, true);
6822 /* Actually do priority change: must hold pi & rq lock. */
6823 static void __setscheduler(struct rq *rq, struct task_struct *p,
6824 const struct sched_attr *attr, bool keep_boost)
6827 * If params can't change scheduling class changes aren't allowed
6830 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
6833 __setscheduler_params(p, attr);
6836 * Keep a potential priority boosting if called from
6837 * sched_setscheduler().
6839 p->prio = normal_prio(p);
6841 p->prio = rt_effective_prio(p, p->prio);
6843 if (dl_prio(p->prio))
6844 p->sched_class = &dl_sched_class;
6845 else if (rt_prio(p->prio))
6846 p->sched_class = &rt_sched_class;
6848 p->sched_class = &fair_sched_class;
6852 * Check the target process has a UID that matches the current process's:
6854 static bool check_same_owner(struct task_struct *p)
6856 const struct cred *cred = current_cred(), *pcred;
6860 pcred = __task_cred(p);
6861 match = (uid_eq(cred->euid, pcred->euid) ||
6862 uid_eq(cred->euid, pcred->uid));
6867 static int __sched_setscheduler(struct task_struct *p,
6868 const struct sched_attr *attr,
6871 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
6872 MAX_RT_PRIO - 1 - attr->sched_priority;
6873 int retval, oldprio, oldpolicy = -1, queued, running;
6874 int new_effective_prio, policy = attr->sched_policy;
6875 const struct sched_class *prev_class;
6876 struct callback_head *head;
6879 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6882 /* The pi code expects interrupts enabled */
6883 BUG_ON(pi && in_interrupt());
6885 /* Double check policy once rq lock held: */
6887 reset_on_fork = p->sched_reset_on_fork;
6888 policy = oldpolicy = p->policy;
6890 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
6892 if (!valid_policy(policy))
6896 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
6900 * Valid priorities for SCHED_FIFO and SCHED_RR are
6901 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
6902 * SCHED_BATCH and SCHED_IDLE is 0.
6904 if (attr->sched_priority > MAX_RT_PRIO-1)
6906 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
6907 (rt_policy(policy) != (attr->sched_priority != 0)))
6911 * Allow unprivileged RT tasks to decrease priority:
6913 if (user && !capable(CAP_SYS_NICE)) {
6914 if (fair_policy(policy)) {
6915 if (attr->sched_nice < task_nice(p) &&
6916 !can_nice(p, attr->sched_nice))
6920 if (rt_policy(policy)) {
6921 unsigned long rlim_rtprio =
6922 task_rlimit(p, RLIMIT_RTPRIO);
6924 /* Can't set/change the rt policy: */
6925 if (policy != p->policy && !rlim_rtprio)
6928 /* Can't increase priority: */
6929 if (attr->sched_priority > p->rt_priority &&
6930 attr->sched_priority > rlim_rtprio)
6935 * Can't set/change SCHED_DEADLINE policy at all for now
6936 * (safest behavior); in the future we would like to allow
6937 * unprivileged DL tasks to increase their relative deadline
6938 * or reduce their runtime (both ways reducing utilization)
6940 if (dl_policy(policy))
6944 * Treat SCHED_IDLE as nice 20. Only allow a switch to
6945 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
6947 if (task_has_idle_policy(p) && !idle_policy(policy)) {
6948 if (!can_nice(p, task_nice(p)))
6952 /* Can't change other user's priorities: */
6953 if (!check_same_owner(p))
6956 /* Normal users shall not reset the sched_reset_on_fork flag: */
6957 if (p->sched_reset_on_fork && !reset_on_fork)
6962 if (attr->sched_flags & SCHED_FLAG_SUGOV)
6965 retval = security_task_setscheduler(p);
6970 /* Update task specific "requested" clamps */
6971 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
6972 retval = uclamp_validate(p, attr);
6981 * Make sure no PI-waiters arrive (or leave) while we are
6982 * changing the priority of the task:
6984 * To be able to change p->policy safely, the appropriate
6985 * runqueue lock must be held.
6987 rq = task_rq_lock(p, &rf);
6988 update_rq_clock(rq);
6991 * Changing the policy of the stop threads its a very bad idea:
6993 if (p == rq->stop) {
6999 * If not changing anything there's no need to proceed further,
7000 * but store a possible modification of reset_on_fork.
7002 if (unlikely(policy == p->policy)) {
7003 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7005 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7007 if (dl_policy(policy) && dl_param_changed(p, attr))
7009 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7012 p->sched_reset_on_fork = reset_on_fork;
7019 #ifdef CONFIG_RT_GROUP_SCHED
7021 * Do not allow realtime tasks into groups that have no runtime
7024 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7025 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7026 !task_group_is_autogroup(task_group(p))) {
7032 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7033 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7034 cpumask_t *span = rq->rd->span;
7037 * Don't allow tasks with an affinity mask smaller than
7038 * the entire root_domain to become SCHED_DEADLINE. We
7039 * will also fail if there's no bandwidth available.
7041 if (!cpumask_subset(span, p->cpus_ptr) ||
7042 rq->rd->dl_bw.bw == 0) {
7050 /* Re-check policy now with rq lock held: */
7051 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7052 policy = oldpolicy = -1;
7053 task_rq_unlock(rq, p, &rf);
7055 cpuset_read_unlock();
7060 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7061 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7064 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7069 p->sched_reset_on_fork = reset_on_fork;
7074 * Take priority boosted tasks into account. If the new
7075 * effective priority is unchanged, we just store the new
7076 * normal parameters and do not touch the scheduler class and
7077 * the runqueue. This will be done when the task deboost
7080 new_effective_prio = rt_effective_prio(p, newprio);
7081 if (new_effective_prio == oldprio)
7082 queue_flags &= ~DEQUEUE_MOVE;
7085 queued = task_on_rq_queued(p);
7086 running = task_current(rq, p);
7088 dequeue_task(rq, p, queue_flags);
7090 put_prev_task(rq, p);
7092 prev_class = p->sched_class;
7094 __setscheduler(rq, p, attr, pi);
7095 __setscheduler_uclamp(p, attr);
7099 * We enqueue to tail when the priority of a task is
7100 * increased (user space view).
7102 if (oldprio < p->prio)
7103 queue_flags |= ENQUEUE_HEAD;
7105 enqueue_task(rq, p, queue_flags);
7108 set_next_task(rq, p);
7110 check_class_changed(rq, p, prev_class, oldprio);
7112 /* Avoid rq from going away on us: */
7114 head = splice_balance_callbacks(rq);
7115 task_rq_unlock(rq, p, &rf);
7118 cpuset_read_unlock();
7119 rt_mutex_adjust_pi(p);
7122 /* Run balance callbacks after we've adjusted the PI chain: */
7123 balance_callbacks(rq, head);
7129 task_rq_unlock(rq, p, &rf);
7131 cpuset_read_unlock();
7135 static int _sched_setscheduler(struct task_struct *p, int policy,
7136 const struct sched_param *param, bool check)
7138 struct sched_attr attr = {
7139 .sched_policy = policy,
7140 .sched_priority = param->sched_priority,
7141 .sched_nice = PRIO_TO_NICE(p->static_prio),
7144 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7145 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7146 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7147 policy &= ~SCHED_RESET_ON_FORK;
7148 attr.sched_policy = policy;
7151 return __sched_setscheduler(p, &attr, check, true);
7154 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7155 * @p: the task in question.
7156 * @policy: new policy.
7157 * @param: structure containing the new RT priority.
7159 * Use sched_set_fifo(), read its comment.
7161 * Return: 0 on success. An error code otherwise.
7163 * NOTE that the task may be already dead.
7165 int sched_setscheduler(struct task_struct *p, int policy,
7166 const struct sched_param *param)
7168 return _sched_setscheduler(p, policy, param, true);
7171 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7173 return __sched_setscheduler(p, attr, true, true);
7176 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7178 return __sched_setscheduler(p, attr, false, true);
7182 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7183 * @p: the task in question.
7184 * @policy: new policy.
7185 * @param: structure containing the new RT priority.
7187 * Just like sched_setscheduler, only don't bother checking if the
7188 * current context has permission. For example, this is needed in
7189 * stop_machine(): we create temporary high priority worker threads,
7190 * but our caller might not have that capability.
7192 * Return: 0 on success. An error code otherwise.
7194 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7195 const struct sched_param *param)
7197 return _sched_setscheduler(p, policy, param, false);
7201 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7202 * incapable of resource management, which is the one thing an OS really should
7205 * This is of course the reason it is limited to privileged users only.
7207 * Worse still; it is fundamentally impossible to compose static priority
7208 * workloads. You cannot take two correctly working static prio workloads
7209 * and smash them together and still expect them to work.
7211 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7215 * The administrator _MUST_ configure the system, the kernel simply doesn't
7216 * know enough information to make a sensible choice.
7218 void sched_set_fifo(struct task_struct *p)
7220 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7221 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7223 EXPORT_SYMBOL_GPL(sched_set_fifo);
7226 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7228 void sched_set_fifo_low(struct task_struct *p)
7230 struct sched_param sp = { .sched_priority = 1 };
7231 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7233 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7235 void sched_set_normal(struct task_struct *p, int nice)
7237 struct sched_attr attr = {
7238 .sched_policy = SCHED_NORMAL,
7241 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7243 EXPORT_SYMBOL_GPL(sched_set_normal);
7246 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7248 struct sched_param lparam;
7249 struct task_struct *p;
7252 if (!param || pid < 0)
7254 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7259 p = find_process_by_pid(pid);
7265 retval = sched_setscheduler(p, policy, &lparam);
7273 * Mimics kernel/events/core.c perf_copy_attr().
7275 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7280 /* Zero the full structure, so that a short copy will be nice: */
7281 memset(attr, 0, sizeof(*attr));
7283 ret = get_user(size, &uattr->size);
7287 /* ABI compatibility quirk: */
7289 size = SCHED_ATTR_SIZE_VER0;
7290 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7293 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7300 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7301 size < SCHED_ATTR_SIZE_VER1)
7305 * XXX: Do we want to be lenient like existing syscalls; or do we want
7306 * to be strict and return an error on out-of-bounds values?
7308 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7313 put_user(sizeof(*attr), &uattr->size);
7318 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7319 * @pid: the pid in question.
7320 * @policy: new policy.
7321 * @param: structure containing the new RT priority.
7323 * Return: 0 on success. An error code otherwise.
7325 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7330 return do_sched_setscheduler(pid, policy, param);
7334 * sys_sched_setparam - set/change the RT priority of a thread
7335 * @pid: the pid in question.
7336 * @param: structure containing the new RT priority.
7338 * Return: 0 on success. An error code otherwise.
7340 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7342 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7346 * sys_sched_setattr - same as above, but with extended sched_attr
7347 * @pid: the pid in question.
7348 * @uattr: structure containing the extended parameters.
7349 * @flags: for future extension.
7351 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7352 unsigned int, flags)
7354 struct sched_attr attr;
7355 struct task_struct *p;
7358 if (!uattr || pid < 0 || flags)
7361 retval = sched_copy_attr(uattr, &attr);
7365 if ((int)attr.sched_policy < 0)
7367 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7368 attr.sched_policy = SETPARAM_POLICY;
7372 p = find_process_by_pid(pid);
7378 retval = sched_setattr(p, &attr);
7386 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7387 * @pid: the pid in question.
7389 * Return: On success, the policy of the thread. Otherwise, a negative error
7392 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7394 struct task_struct *p;
7402 p = find_process_by_pid(pid);
7404 retval = security_task_getscheduler(p);
7407 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7414 * sys_sched_getparam - get the RT priority of a thread
7415 * @pid: the pid in question.
7416 * @param: structure containing the RT priority.
7418 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7421 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7423 struct sched_param lp = { .sched_priority = 0 };
7424 struct task_struct *p;
7427 if (!param || pid < 0)
7431 p = find_process_by_pid(pid);
7436 retval = security_task_getscheduler(p);
7440 if (task_has_rt_policy(p))
7441 lp.sched_priority = p->rt_priority;
7445 * This one might sleep, we cannot do it with a spinlock held ...
7447 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7457 * Copy the kernel size attribute structure (which might be larger
7458 * than what user-space knows about) to user-space.
7460 * Note that all cases are valid: user-space buffer can be larger or
7461 * smaller than the kernel-space buffer. The usual case is that both
7462 * have the same size.
7465 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7466 struct sched_attr *kattr,
7469 unsigned int ksize = sizeof(*kattr);
7471 if (!access_ok(uattr, usize))
7475 * sched_getattr() ABI forwards and backwards compatibility:
7477 * If usize == ksize then we just copy everything to user-space and all is good.
7479 * If usize < ksize then we only copy as much as user-space has space for,
7480 * this keeps ABI compatibility as well. We skip the rest.
7482 * If usize > ksize then user-space is using a newer version of the ABI,
7483 * which part the kernel doesn't know about. Just ignore it - tooling can
7484 * detect the kernel's knowledge of attributes from the attr->size value
7485 * which is set to ksize in this case.
7487 kattr->size = min(usize, ksize);
7489 if (copy_to_user(uattr, kattr, kattr->size))
7496 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7497 * @pid: the pid in question.
7498 * @uattr: structure containing the extended parameters.
7499 * @usize: sizeof(attr) for fwd/bwd comp.
7500 * @flags: for future extension.
7502 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7503 unsigned int, usize, unsigned int, flags)
7505 struct sched_attr kattr = { };
7506 struct task_struct *p;
7509 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7510 usize < SCHED_ATTR_SIZE_VER0 || flags)
7514 p = find_process_by_pid(pid);
7519 retval = security_task_getscheduler(p);
7523 kattr.sched_policy = p->policy;
7524 if (p->sched_reset_on_fork)
7525 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7526 if (task_has_dl_policy(p))
7527 __getparam_dl(p, &kattr);
7528 else if (task_has_rt_policy(p))
7529 kattr.sched_priority = p->rt_priority;
7531 kattr.sched_nice = task_nice(p);
7533 #ifdef CONFIG_UCLAMP_TASK
7535 * This could race with another potential updater, but this is fine
7536 * because it'll correctly read the old or the new value. We don't need
7537 * to guarantee who wins the race as long as it doesn't return garbage.
7539 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7540 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7545 return sched_attr_copy_to_user(uattr, &kattr, usize);
7552 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7554 cpumask_var_t cpus_allowed, new_mask;
7555 struct task_struct *p;
7560 p = find_process_by_pid(pid);
7566 /* Prevent p going away */
7570 if (p->flags & PF_NO_SETAFFINITY) {
7574 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
7578 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7580 goto out_free_cpus_allowed;
7583 if (!check_same_owner(p)) {
7585 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7587 goto out_free_new_mask;
7592 retval = security_task_setscheduler(p);
7594 goto out_free_new_mask;
7597 cpuset_cpus_allowed(p, cpus_allowed);
7598 cpumask_and(new_mask, in_mask, cpus_allowed);
7601 * Since bandwidth control happens on root_domain basis,
7602 * if admission test is enabled, we only admit -deadline
7603 * tasks allowed to run on all the CPUs in the task's
7607 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
7609 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
7612 goto out_free_new_mask;
7618 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
7621 cpuset_cpus_allowed(p, cpus_allowed);
7622 if (!cpumask_subset(new_mask, cpus_allowed)) {
7624 * We must have raced with a concurrent cpuset
7625 * update. Just reset the cpus_allowed to the
7626 * cpuset's cpus_allowed
7628 cpumask_copy(new_mask, cpus_allowed);
7633 free_cpumask_var(new_mask);
7634 out_free_cpus_allowed:
7635 free_cpumask_var(cpus_allowed);
7641 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7642 struct cpumask *new_mask)
7644 if (len < cpumask_size())
7645 cpumask_clear(new_mask);
7646 else if (len > cpumask_size())
7647 len = cpumask_size();
7649 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
7653 * sys_sched_setaffinity - set the CPU affinity of a process
7654 * @pid: pid of the process
7655 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7656 * @user_mask_ptr: user-space pointer to the new CPU mask
7658 * Return: 0 on success. An error code otherwise.
7660 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
7661 unsigned long __user *, user_mask_ptr)
7663 cpumask_var_t new_mask;
7666 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
7669 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
7671 retval = sched_setaffinity(pid, new_mask);
7672 free_cpumask_var(new_mask);
7676 long sched_getaffinity(pid_t pid, struct cpumask *mask)
7678 struct task_struct *p;
7679 unsigned long flags;
7685 p = find_process_by_pid(pid);
7689 retval = security_task_getscheduler(p);
7693 raw_spin_lock_irqsave(&p->pi_lock, flags);
7694 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
7695 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
7704 * sys_sched_getaffinity - get the CPU affinity of a process
7705 * @pid: pid of the process
7706 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7707 * @user_mask_ptr: user-space pointer to hold the current CPU mask
7709 * Return: size of CPU mask copied to user_mask_ptr on success. An
7710 * error code otherwise.
7712 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
7713 unsigned long __user *, user_mask_ptr)
7718 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
7720 if (len & (sizeof(unsigned long)-1))
7723 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
7726 ret = sched_getaffinity(pid, mask);
7728 unsigned int retlen = min(len, cpumask_size());
7730 if (copy_to_user(user_mask_ptr, mask, retlen))
7735 free_cpumask_var(mask);
7740 static void do_sched_yield(void)
7745 rq = this_rq_lock_irq(&rf);
7747 schedstat_inc(rq->yld_count);
7748 current->sched_class->yield_task(rq);
7751 rq_unlock_irq(rq, &rf);
7752 sched_preempt_enable_no_resched();
7758 * sys_sched_yield - yield the current processor to other threads.
7760 * This function yields the current CPU to other tasks. If there are no
7761 * other threads running on this CPU then this function will return.
7765 SYSCALL_DEFINE0(sched_yield)
7771 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
7772 int __sched __cond_resched(void)
7774 if (should_resched(0)) {
7775 preempt_schedule_common();
7778 #ifndef CONFIG_PREEMPT_RCU
7783 EXPORT_SYMBOL(__cond_resched);
7786 #ifdef CONFIG_PREEMPT_DYNAMIC
7787 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
7788 EXPORT_STATIC_CALL_TRAMP(cond_resched);
7790 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
7791 EXPORT_STATIC_CALL_TRAMP(might_resched);
7795 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7796 * call schedule, and on return reacquire the lock.
7798 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
7799 * operations here to prevent schedule() from being called twice (once via
7800 * spin_unlock(), once by hand).
7802 int __cond_resched_lock(spinlock_t *lock)
7804 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7807 lockdep_assert_held(lock);
7809 if (spin_needbreak(lock) || resched) {
7812 preempt_schedule_common();
7820 EXPORT_SYMBOL(__cond_resched_lock);
7822 int __cond_resched_rwlock_read(rwlock_t *lock)
7824 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7827 lockdep_assert_held_read(lock);
7829 if (rwlock_needbreak(lock) || resched) {
7832 preempt_schedule_common();
7840 EXPORT_SYMBOL(__cond_resched_rwlock_read);
7842 int __cond_resched_rwlock_write(rwlock_t *lock)
7844 int resched = should_resched(PREEMPT_LOCK_OFFSET);
7847 lockdep_assert_held_write(lock);
7849 if (rwlock_needbreak(lock) || resched) {
7852 preempt_schedule_common();
7860 EXPORT_SYMBOL(__cond_resched_rwlock_write);
7863 * yield - yield the current processor to other threads.
7865 * Do not ever use this function, there's a 99% chance you're doing it wrong.
7867 * The scheduler is at all times free to pick the calling task as the most
7868 * eligible task to run, if removing the yield() call from your code breaks
7869 * it, it's already broken.
7871 * Typical broken usage is:
7876 * where one assumes that yield() will let 'the other' process run that will
7877 * make event true. If the current task is a SCHED_FIFO task that will never
7878 * happen. Never use yield() as a progress guarantee!!
7880 * If you want to use yield() to wait for something, use wait_event().
7881 * If you want to use yield() to be 'nice' for others, use cond_resched().
7882 * If you still want to use yield(), do not!
7884 void __sched yield(void)
7886 set_current_state(TASK_RUNNING);
7889 EXPORT_SYMBOL(yield);
7892 * yield_to - yield the current processor to another thread in
7893 * your thread group, or accelerate that thread toward the
7894 * processor it's on.
7896 * @preempt: whether task preemption is allowed or not
7898 * It's the caller's job to ensure that the target task struct
7899 * can't go away on us before we can do any checks.
7902 * true (>0) if we indeed boosted the target task.
7903 * false (0) if we failed to boost the target.
7904 * -ESRCH if there's no task to yield to.
7906 int __sched yield_to(struct task_struct *p, bool preempt)
7908 struct task_struct *curr = current;
7909 struct rq *rq, *p_rq;
7910 unsigned long flags;
7913 local_irq_save(flags);
7919 * If we're the only runnable task on the rq and target rq also
7920 * has only one task, there's absolutely no point in yielding.
7922 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
7927 double_rq_lock(rq, p_rq);
7928 if (task_rq(p) != p_rq) {
7929 double_rq_unlock(rq, p_rq);
7933 if (!curr->sched_class->yield_to_task)
7936 if (curr->sched_class != p->sched_class)
7939 if (task_running(p_rq, p) || !task_is_running(p))
7942 yielded = curr->sched_class->yield_to_task(rq, p);
7944 schedstat_inc(rq->yld_count);
7946 * Make p's CPU reschedule; pick_next_entity takes care of
7949 if (preempt && rq != p_rq)
7954 double_rq_unlock(rq, p_rq);
7956 local_irq_restore(flags);
7963 EXPORT_SYMBOL_GPL(yield_to);
7965 int io_schedule_prepare(void)
7967 int old_iowait = current->in_iowait;
7969 current->in_iowait = 1;
7970 blk_schedule_flush_plug(current);
7975 void io_schedule_finish(int token)
7977 current->in_iowait = token;
7981 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7982 * that process accounting knows that this is a task in IO wait state.
7984 long __sched io_schedule_timeout(long timeout)
7989 token = io_schedule_prepare();
7990 ret = schedule_timeout(timeout);
7991 io_schedule_finish(token);
7995 EXPORT_SYMBOL(io_schedule_timeout);
7997 void __sched io_schedule(void)
8001 token = io_schedule_prepare();
8003 io_schedule_finish(token);
8005 EXPORT_SYMBOL(io_schedule);
8008 * sys_sched_get_priority_max - return maximum RT priority.
8009 * @policy: scheduling class.
8011 * Return: On success, this syscall returns the maximum
8012 * rt_priority that can be used by a given scheduling class.
8013 * On failure, a negative error code is returned.
8015 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8022 ret = MAX_RT_PRIO-1;
8024 case SCHED_DEADLINE:
8035 * sys_sched_get_priority_min - return minimum RT priority.
8036 * @policy: scheduling class.
8038 * Return: On success, this syscall returns the minimum
8039 * rt_priority that can be used by a given scheduling class.
8040 * On failure, a negative error code is returned.
8042 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8051 case SCHED_DEADLINE:
8060 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8062 struct task_struct *p;
8063 unsigned int time_slice;
8073 p = find_process_by_pid(pid);
8077 retval = security_task_getscheduler(p);
8081 rq = task_rq_lock(p, &rf);
8083 if (p->sched_class->get_rr_interval)
8084 time_slice = p->sched_class->get_rr_interval(rq, p);
8085 task_rq_unlock(rq, p, &rf);
8088 jiffies_to_timespec64(time_slice, t);
8097 * sys_sched_rr_get_interval - return the default timeslice of a process.
8098 * @pid: pid of the process.
8099 * @interval: userspace pointer to the timeslice value.
8101 * this syscall writes the default timeslice value of a given process
8102 * into the user-space timespec buffer. A value of '0' means infinity.
8104 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8107 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8108 struct __kernel_timespec __user *, interval)
8110 struct timespec64 t;
8111 int retval = sched_rr_get_interval(pid, &t);
8114 retval = put_timespec64(&t, interval);
8119 #ifdef CONFIG_COMPAT_32BIT_TIME
8120 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8121 struct old_timespec32 __user *, interval)
8123 struct timespec64 t;
8124 int retval = sched_rr_get_interval(pid, &t);
8127 retval = put_old_timespec32(&t, interval);
8132 void sched_show_task(struct task_struct *p)
8134 unsigned long free = 0;
8137 if (!try_get_task_stack(p))
8140 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8142 if (task_is_running(p))
8143 pr_cont(" running task ");
8144 #ifdef CONFIG_DEBUG_STACK_USAGE
8145 free = stack_not_used(p);
8150 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8152 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8153 free, task_pid_nr(p), ppid,
8154 (unsigned long)task_thread_info(p)->flags);
8156 print_worker_info(KERN_INFO, p);
8157 print_stop_info(KERN_INFO, p);
8158 show_stack(p, NULL, KERN_INFO);
8161 EXPORT_SYMBOL_GPL(sched_show_task);
8164 state_filter_match(unsigned long state_filter, struct task_struct *p)
8166 unsigned int state = READ_ONCE(p->__state);
8168 /* no filter, everything matches */
8172 /* filter, but doesn't match */
8173 if (!(state & state_filter))
8177 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8180 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8187 void show_state_filter(unsigned int state_filter)
8189 struct task_struct *g, *p;
8192 for_each_process_thread(g, p) {
8194 * reset the NMI-timeout, listing all files on a slow
8195 * console might take a lot of time:
8196 * Also, reset softlockup watchdogs on all CPUs, because
8197 * another CPU might be blocked waiting for us to process
8200 touch_nmi_watchdog();
8201 touch_all_softlockup_watchdogs();
8202 if (state_filter_match(state_filter, p))
8206 #ifdef CONFIG_SCHED_DEBUG
8208 sysrq_sched_debug_show();
8212 * Only show locks if all tasks are dumped:
8215 debug_show_all_locks();
8219 * init_idle - set up an idle thread for a given CPU
8220 * @idle: task in question
8221 * @cpu: CPU the idle task belongs to
8223 * NOTE: this function does not set the idle thread's NEED_RESCHED
8224 * flag, to make booting more robust.
8226 void __init init_idle(struct task_struct *idle, int cpu)
8228 struct rq *rq = cpu_rq(cpu);
8229 unsigned long flags;
8231 __sched_fork(0, idle);
8234 * The idle task doesn't need the kthread struct to function, but it
8235 * is dressed up as a per-CPU kthread and thus needs to play the part
8236 * if we want to avoid special-casing it in code that deals with per-CPU
8239 set_kthread_struct(idle);
8241 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8242 raw_spin_rq_lock(rq);
8244 idle->__state = TASK_RUNNING;
8245 idle->se.exec_start = sched_clock();
8247 * PF_KTHREAD should already be set at this point; regardless, make it
8248 * look like a proper per-CPU kthread.
8250 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8251 kthread_set_per_cpu(idle, cpu);
8253 scs_task_reset(idle);
8254 kasan_unpoison_task_stack(idle);
8258 * It's possible that init_idle() gets called multiple times on a task,
8259 * in that case do_set_cpus_allowed() will not do the right thing.
8261 * And since this is boot we can forgo the serialization.
8263 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8266 * We're having a chicken and egg problem, even though we are
8267 * holding rq->lock, the CPU isn't yet set to this CPU so the
8268 * lockdep check in task_group() will fail.
8270 * Similar case to sched_fork(). / Alternatively we could
8271 * use task_rq_lock() here and obtain the other rq->lock.
8276 __set_task_cpu(idle, cpu);
8280 rcu_assign_pointer(rq->curr, idle);
8281 idle->on_rq = TASK_ON_RQ_QUEUED;
8285 raw_spin_rq_unlock(rq);
8286 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8288 /* Set the preempt count _outside_ the spinlocks! */
8289 init_idle_preempt_count(idle, cpu);
8292 * The idle tasks have their own, simple scheduling class:
8294 idle->sched_class = &idle_sched_class;
8295 ftrace_graph_init_idle_task(idle, cpu);
8296 vtime_init_idle(idle, cpu);
8298 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8304 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8305 const struct cpumask *trial)
8309 if (!cpumask_weight(cur))
8312 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8317 int task_can_attach(struct task_struct *p,
8318 const struct cpumask *cs_cpus_allowed)
8323 * Kthreads which disallow setaffinity shouldn't be moved
8324 * to a new cpuset; we don't want to change their CPU
8325 * affinity and isolating such threads by their set of
8326 * allowed nodes is unnecessary. Thus, cpusets are not
8327 * applicable for such threads. This prevents checking for
8328 * success of set_cpus_allowed_ptr() on all attached tasks
8329 * before cpus_mask may be changed.
8331 if (p->flags & PF_NO_SETAFFINITY) {
8336 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8338 ret = dl_task_can_attach(p, cs_cpus_allowed);
8344 bool sched_smp_initialized __read_mostly;
8346 #ifdef CONFIG_NUMA_BALANCING
8347 /* Migrate current task p to target_cpu */
8348 int migrate_task_to(struct task_struct *p, int target_cpu)
8350 struct migration_arg arg = { p, target_cpu };
8351 int curr_cpu = task_cpu(p);
8353 if (curr_cpu == target_cpu)
8356 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8359 /* TODO: This is not properly updating schedstats */
8361 trace_sched_move_numa(p, curr_cpu, target_cpu);
8362 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8366 * Requeue a task on a given node and accurately track the number of NUMA
8367 * tasks on the runqueues
8369 void sched_setnuma(struct task_struct *p, int nid)
8371 bool queued, running;
8375 rq = task_rq_lock(p, &rf);
8376 queued = task_on_rq_queued(p);
8377 running = task_current(rq, p);
8380 dequeue_task(rq, p, DEQUEUE_SAVE);
8382 put_prev_task(rq, p);
8384 p->numa_preferred_nid = nid;
8387 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8389 set_next_task(rq, p);
8390 task_rq_unlock(rq, p, &rf);
8392 #endif /* CONFIG_NUMA_BALANCING */
8394 #ifdef CONFIG_HOTPLUG_CPU
8396 * Ensure that the idle task is using init_mm right before its CPU goes
8399 void idle_task_exit(void)
8401 struct mm_struct *mm = current->active_mm;
8403 BUG_ON(cpu_online(smp_processor_id()));
8404 BUG_ON(current != this_rq()->idle);
8406 if (mm != &init_mm) {
8407 switch_mm(mm, &init_mm, current);
8408 finish_arch_post_lock_switch();
8411 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8414 static int __balance_push_cpu_stop(void *arg)
8416 struct task_struct *p = arg;
8417 struct rq *rq = this_rq();
8421 raw_spin_lock_irq(&p->pi_lock);
8424 update_rq_clock(rq);
8426 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8427 cpu = select_fallback_rq(rq->cpu, p);
8428 rq = __migrate_task(rq, &rf, p, cpu);
8432 raw_spin_unlock_irq(&p->pi_lock);
8439 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8442 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8444 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8445 * effective when the hotplug motion is down.
8447 static void balance_push(struct rq *rq)
8449 struct task_struct *push_task = rq->curr;
8451 lockdep_assert_rq_held(rq);
8452 SCHED_WARN_ON(rq->cpu != smp_processor_id());
8455 * Ensure the thing is persistent until balance_push_set(.on = false);
8457 rq->balance_callback = &balance_push_callback;
8460 * Only active while going offline.
8462 if (!cpu_dying(rq->cpu))
8466 * Both the cpu-hotplug and stop task are in this case and are
8467 * required to complete the hotplug process.
8469 if (kthread_is_per_cpu(push_task) ||
8470 is_migration_disabled(push_task)) {
8473 * If this is the idle task on the outgoing CPU try to wake
8474 * up the hotplug control thread which might wait for the
8475 * last task to vanish. The rcuwait_active() check is
8476 * accurate here because the waiter is pinned on this CPU
8477 * and can't obviously be running in parallel.
8479 * On RT kernels this also has to check whether there are
8480 * pinned and scheduled out tasks on the runqueue. They
8481 * need to leave the migrate disabled section first.
8483 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8484 rcuwait_active(&rq->hotplug_wait)) {
8485 raw_spin_rq_unlock(rq);
8486 rcuwait_wake_up(&rq->hotplug_wait);
8487 raw_spin_rq_lock(rq);
8492 get_task_struct(push_task);
8494 * Temporarily drop rq->lock such that we can wake-up the stop task.
8495 * Both preemption and IRQs are still disabled.
8497 raw_spin_rq_unlock(rq);
8498 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8499 this_cpu_ptr(&push_work));
8501 * At this point need_resched() is true and we'll take the loop in
8502 * schedule(). The next pick is obviously going to be the stop task
8503 * which kthread_is_per_cpu() and will push this task away.
8505 raw_spin_rq_lock(rq);
8508 static void balance_push_set(int cpu, bool on)
8510 struct rq *rq = cpu_rq(cpu);
8513 rq_lock_irqsave(rq, &rf);
8515 WARN_ON_ONCE(rq->balance_callback);
8516 rq->balance_callback = &balance_push_callback;
8517 } else if (rq->balance_callback == &balance_push_callback) {
8518 rq->balance_callback = NULL;
8520 rq_unlock_irqrestore(rq, &rf);
8524 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8525 * inactive. All tasks which are not per CPU kernel threads are either
8526 * pushed off this CPU now via balance_push() or placed on a different CPU
8527 * during wakeup. Wait until the CPU is quiescent.
8529 static void balance_hotplug_wait(void)
8531 struct rq *rq = this_rq();
8533 rcuwait_wait_event(&rq->hotplug_wait,
8534 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8535 TASK_UNINTERRUPTIBLE);
8540 static inline void balance_push(struct rq *rq)
8544 static inline void balance_push_set(int cpu, bool on)
8548 static inline void balance_hotplug_wait(void)
8552 #endif /* CONFIG_HOTPLUG_CPU */
8554 void set_rq_online(struct rq *rq)
8557 const struct sched_class *class;
8559 cpumask_set_cpu(rq->cpu, rq->rd->online);
8562 for_each_class(class) {
8563 if (class->rq_online)
8564 class->rq_online(rq);
8569 void set_rq_offline(struct rq *rq)
8572 const struct sched_class *class;
8574 for_each_class(class) {
8575 if (class->rq_offline)
8576 class->rq_offline(rq);
8579 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8585 * used to mark begin/end of suspend/resume:
8587 static int num_cpus_frozen;
8590 * Update cpusets according to cpu_active mask. If cpusets are
8591 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8592 * around partition_sched_domains().
8594 * If we come here as part of a suspend/resume, don't touch cpusets because we
8595 * want to restore it back to its original state upon resume anyway.
8597 static void cpuset_cpu_active(void)
8599 if (cpuhp_tasks_frozen) {
8601 * num_cpus_frozen tracks how many CPUs are involved in suspend
8602 * resume sequence. As long as this is not the last online
8603 * operation in the resume sequence, just build a single sched
8604 * domain, ignoring cpusets.
8606 partition_sched_domains(1, NULL, NULL);
8607 if (--num_cpus_frozen)
8610 * This is the last CPU online operation. So fall through and
8611 * restore the original sched domains by considering the
8612 * cpuset configurations.
8614 cpuset_force_rebuild();
8616 cpuset_update_active_cpus();
8619 static int cpuset_cpu_inactive(unsigned int cpu)
8621 if (!cpuhp_tasks_frozen) {
8622 if (dl_cpu_busy(cpu))
8624 cpuset_update_active_cpus();
8627 partition_sched_domains(1, NULL, NULL);
8632 int sched_cpu_activate(unsigned int cpu)
8634 struct rq *rq = cpu_rq(cpu);
8638 * Clear the balance_push callback and prepare to schedule
8641 balance_push_set(cpu, false);
8643 #ifdef CONFIG_SCHED_SMT
8645 * When going up, increment the number of cores with SMT present.
8647 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8648 static_branch_inc_cpuslocked(&sched_smt_present);
8650 set_cpu_active(cpu, true);
8652 if (sched_smp_initialized) {
8653 sched_domains_numa_masks_set(cpu);
8654 cpuset_cpu_active();
8658 * Put the rq online, if not already. This happens:
8660 * 1) In the early boot process, because we build the real domains
8661 * after all CPUs have been brought up.
8663 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
8666 rq_lock_irqsave(rq, &rf);
8668 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8671 rq_unlock_irqrestore(rq, &rf);
8676 int sched_cpu_deactivate(unsigned int cpu)
8678 struct rq *rq = cpu_rq(cpu);
8683 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
8684 * load balancing when not active
8686 nohz_balance_exit_idle(rq);
8688 set_cpu_active(cpu, false);
8691 * From this point forward, this CPU will refuse to run any task that
8692 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
8693 * push those tasks away until this gets cleared, see
8694 * sched_cpu_dying().
8696 balance_push_set(cpu, true);
8699 * We've cleared cpu_active_mask / set balance_push, wait for all
8700 * preempt-disabled and RCU users of this state to go away such that
8701 * all new such users will observe it.
8703 * Specifically, we rely on ttwu to no longer target this CPU, see
8704 * ttwu_queue_cond() and is_cpu_allowed().
8706 * Do sync before park smpboot threads to take care the rcu boost case.
8710 rq_lock_irqsave(rq, &rf);
8712 update_rq_clock(rq);
8713 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8716 rq_unlock_irqrestore(rq, &rf);
8718 #ifdef CONFIG_SCHED_SMT
8720 * When going down, decrement the number of cores with SMT present.
8722 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8723 static_branch_dec_cpuslocked(&sched_smt_present);
8726 if (!sched_smp_initialized)
8729 ret = cpuset_cpu_inactive(cpu);
8731 balance_push_set(cpu, false);
8732 set_cpu_active(cpu, true);
8735 sched_domains_numa_masks_clear(cpu);
8739 static void sched_rq_cpu_starting(unsigned int cpu)
8741 struct rq *rq = cpu_rq(cpu);
8743 rq->calc_load_update = calc_load_update;
8744 update_max_interval();
8747 int sched_cpu_starting(unsigned int cpu)
8749 sched_core_cpu_starting(cpu);
8750 sched_rq_cpu_starting(cpu);
8751 sched_tick_start(cpu);
8755 #ifdef CONFIG_HOTPLUG_CPU
8758 * Invoked immediately before the stopper thread is invoked to bring the
8759 * CPU down completely. At this point all per CPU kthreads except the
8760 * hotplug thread (current) and the stopper thread (inactive) have been
8761 * either parked or have been unbound from the outgoing CPU. Ensure that
8762 * any of those which might be on the way out are gone.
8764 * If after this point a bound task is being woken on this CPU then the
8765 * responsible hotplug callback has failed to do it's job.
8766 * sched_cpu_dying() will catch it with the appropriate fireworks.
8768 int sched_cpu_wait_empty(unsigned int cpu)
8770 balance_hotplug_wait();
8775 * Since this CPU is going 'away' for a while, fold any nr_active delta we
8776 * might have. Called from the CPU stopper task after ensuring that the
8777 * stopper is the last running task on the CPU, so nr_active count is
8778 * stable. We need to take the teardown thread which is calling this into
8779 * account, so we hand in adjust = 1 to the load calculation.
8781 * Also see the comment "Global load-average calculations".
8783 static void calc_load_migrate(struct rq *rq)
8785 long delta = calc_load_fold_active(rq, 1);
8788 atomic_long_add(delta, &calc_load_tasks);
8791 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
8793 struct task_struct *g, *p;
8794 int cpu = cpu_of(rq);
8796 lockdep_assert_rq_held(rq);
8798 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
8799 for_each_process_thread(g, p) {
8800 if (task_cpu(p) != cpu)
8803 if (!task_on_rq_queued(p))
8806 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
8810 int sched_cpu_dying(unsigned int cpu)
8812 struct rq *rq = cpu_rq(cpu);
8815 /* Handle pending wakeups and then migrate everything off */
8816 sched_tick_stop(cpu);
8818 rq_lock_irqsave(rq, &rf);
8819 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
8820 WARN(true, "Dying CPU not properly vacated!");
8821 dump_rq_tasks(rq, KERN_WARNING);
8823 rq_unlock_irqrestore(rq, &rf);
8825 calc_load_migrate(rq);
8826 update_max_interval();
8832 void __init sched_init_smp(void)
8837 * There's no userspace yet to cause hotplug operations; hence all the
8838 * CPU masks are stable and all blatant races in the below code cannot
8841 mutex_lock(&sched_domains_mutex);
8842 sched_init_domains(cpu_active_mask);
8843 mutex_unlock(&sched_domains_mutex);
8845 /* Move init over to a non-isolated CPU */
8846 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
8848 current->flags &= ~PF_NO_SETAFFINITY;
8849 sched_init_granularity();
8851 init_sched_rt_class();
8852 init_sched_dl_class();
8854 sched_smp_initialized = true;
8857 static int __init migration_init(void)
8859 sched_cpu_starting(smp_processor_id());
8862 early_initcall(migration_init);
8865 void __init sched_init_smp(void)
8867 sched_init_granularity();
8869 #endif /* CONFIG_SMP */
8871 int in_sched_functions(unsigned long addr)
8873 return in_lock_functions(addr) ||
8874 (addr >= (unsigned long)__sched_text_start
8875 && addr < (unsigned long)__sched_text_end);
8878 #ifdef CONFIG_CGROUP_SCHED
8880 * Default task group.
8881 * Every task in system belongs to this group at bootup.
8883 struct task_group root_task_group;
8884 LIST_HEAD(task_groups);
8886 /* Cacheline aligned slab cache for task_group */
8887 static struct kmem_cache *task_group_cache __read_mostly;
8890 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
8891 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
8893 void __init sched_init(void)
8895 unsigned long ptr = 0;
8898 /* Make sure the linker didn't screw up */
8899 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
8900 &fair_sched_class + 1 != &rt_sched_class ||
8901 &rt_sched_class + 1 != &dl_sched_class);
8903 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
8908 #ifdef CONFIG_FAIR_GROUP_SCHED
8909 ptr += 2 * nr_cpu_ids * sizeof(void **);
8911 #ifdef CONFIG_RT_GROUP_SCHED
8912 ptr += 2 * nr_cpu_ids * sizeof(void **);
8915 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
8917 #ifdef CONFIG_FAIR_GROUP_SCHED
8918 root_task_group.se = (struct sched_entity **)ptr;
8919 ptr += nr_cpu_ids * sizeof(void **);
8921 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8922 ptr += nr_cpu_ids * sizeof(void **);
8924 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
8925 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8926 #endif /* CONFIG_FAIR_GROUP_SCHED */
8927 #ifdef CONFIG_RT_GROUP_SCHED
8928 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8929 ptr += nr_cpu_ids * sizeof(void **);
8931 root_task_group.rt_rq = (struct rt_rq **)ptr;
8932 ptr += nr_cpu_ids * sizeof(void **);
8934 #endif /* CONFIG_RT_GROUP_SCHED */
8936 #ifdef CONFIG_CPUMASK_OFFSTACK
8937 for_each_possible_cpu(i) {
8938 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
8939 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8940 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
8941 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
8943 #endif /* CONFIG_CPUMASK_OFFSTACK */
8945 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
8946 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
8949 init_defrootdomain();
8952 #ifdef CONFIG_RT_GROUP_SCHED
8953 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8954 global_rt_period(), global_rt_runtime());
8955 #endif /* CONFIG_RT_GROUP_SCHED */
8957 #ifdef CONFIG_CGROUP_SCHED
8958 task_group_cache = KMEM_CACHE(task_group, 0);
8960 list_add(&root_task_group.list, &task_groups);
8961 INIT_LIST_HEAD(&root_task_group.children);
8962 INIT_LIST_HEAD(&root_task_group.siblings);
8963 autogroup_init(&init_task);
8964 #endif /* CONFIG_CGROUP_SCHED */
8966 for_each_possible_cpu(i) {
8970 raw_spin_lock_init(&rq->__lock);
8972 rq->calc_load_active = 0;
8973 rq->calc_load_update = jiffies + LOAD_FREQ;
8974 init_cfs_rq(&rq->cfs);
8975 init_rt_rq(&rq->rt);
8976 init_dl_rq(&rq->dl);
8977 #ifdef CONFIG_FAIR_GROUP_SCHED
8978 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8979 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
8981 * How much CPU bandwidth does root_task_group get?
8983 * In case of task-groups formed thr' the cgroup filesystem, it
8984 * gets 100% of the CPU resources in the system. This overall
8985 * system CPU resource is divided among the tasks of
8986 * root_task_group and its child task-groups in a fair manner,
8987 * based on each entity's (task or task-group's) weight
8988 * (se->load.weight).
8990 * In other words, if root_task_group has 10 tasks of weight
8991 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8992 * then A0's share of the CPU resource is:
8994 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8996 * We achieve this by letting root_task_group's tasks sit
8997 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8999 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9000 #endif /* CONFIG_FAIR_GROUP_SCHED */
9002 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9003 #ifdef CONFIG_RT_GROUP_SCHED
9004 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9009 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9010 rq->balance_callback = &balance_push_callback;
9011 rq->active_balance = 0;
9012 rq->next_balance = jiffies;
9017 rq->avg_idle = 2*sysctl_sched_migration_cost;
9018 rq->wake_stamp = jiffies;
9019 rq->wake_avg_idle = rq->avg_idle;
9020 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9022 INIT_LIST_HEAD(&rq->cfs_tasks);
9024 rq_attach_root(rq, &def_root_domain);
9025 #ifdef CONFIG_NO_HZ_COMMON
9026 rq->last_blocked_load_update_tick = jiffies;
9027 atomic_set(&rq->nohz_flags, 0);
9029 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9031 #ifdef CONFIG_HOTPLUG_CPU
9032 rcuwait_init(&rq->hotplug_wait);
9034 #endif /* CONFIG_SMP */
9036 atomic_set(&rq->nr_iowait, 0);
9038 #ifdef CONFIG_SCHED_CORE
9040 rq->core_pick = NULL;
9041 rq->core_enabled = 0;
9042 rq->core_tree = RB_ROOT;
9043 rq->core_forceidle = false;
9045 rq->core_cookie = 0UL;
9049 set_load_weight(&init_task, false);
9052 * The boot idle thread does lazy MMU switching as well:
9055 enter_lazy_tlb(&init_mm, current);
9058 * Make us the idle thread. Technically, schedule() should not be
9059 * called from this thread, however somewhere below it might be,
9060 * but because we are the idle thread, we just pick up running again
9061 * when this runqueue becomes "idle".
9063 init_idle(current, smp_processor_id());
9065 calc_load_update = jiffies + LOAD_FREQ;
9068 idle_thread_set_boot_cpu();
9069 balance_push_set(smp_processor_id(), false);
9071 init_sched_fair_class();
9077 scheduler_running = 1;
9080 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9081 static inline int preempt_count_equals(int preempt_offset)
9083 int nested = preempt_count() + rcu_preempt_depth();
9085 return (nested == preempt_offset);
9088 void __might_sleep(const char *file, int line, int preempt_offset)
9090 unsigned int state = get_current_state();
9092 * Blocking primitives will set (and therefore destroy) current->state,
9093 * since we will exit with TASK_RUNNING make sure we enter with it,
9094 * otherwise we will destroy state.
9096 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9097 "do not call blocking ops when !TASK_RUNNING; "
9098 "state=%x set at [<%p>] %pS\n", state,
9099 (void *)current->task_state_change,
9100 (void *)current->task_state_change);
9102 ___might_sleep(file, line, preempt_offset);
9104 EXPORT_SYMBOL(__might_sleep);
9106 void ___might_sleep(const char *file, int line, int preempt_offset)
9108 /* Ratelimiting timestamp: */
9109 static unsigned long prev_jiffy;
9111 unsigned long preempt_disable_ip;
9113 /* WARN_ON_ONCE() by default, no rate limit required: */
9116 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
9117 !is_idle_task(current) && !current->non_block_count) ||
9118 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9122 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9124 prev_jiffy = jiffies;
9126 /* Save this before calling printk(), since that will clobber it: */
9127 preempt_disable_ip = get_preempt_disable_ip(current);
9130 "BUG: sleeping function called from invalid context at %s:%d\n",
9133 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9134 in_atomic(), irqs_disabled(), current->non_block_count,
9135 current->pid, current->comm);
9137 if (task_stack_end_corrupted(current))
9138 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
9140 debug_show_held_locks(current);
9141 if (irqs_disabled())
9142 print_irqtrace_events(current);
9143 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
9144 && !preempt_count_equals(preempt_offset)) {
9145 pr_err("Preemption disabled at:");
9146 print_ip_sym(KERN_ERR, preempt_disable_ip);
9149 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9151 EXPORT_SYMBOL(___might_sleep);
9153 void __cant_sleep(const char *file, int line, int preempt_offset)
9155 static unsigned long prev_jiffy;
9157 if (irqs_disabled())
9160 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9163 if (preempt_count() > preempt_offset)
9166 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9168 prev_jiffy = jiffies;
9170 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9171 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9172 in_atomic(), irqs_disabled(),
9173 current->pid, current->comm);
9175 debug_show_held_locks(current);
9177 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9179 EXPORT_SYMBOL_GPL(__cant_sleep);
9182 void __cant_migrate(const char *file, int line)
9184 static unsigned long prev_jiffy;
9186 if (irqs_disabled())
9189 if (is_migration_disabled(current))
9192 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9195 if (preempt_count() > 0)
9198 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9200 prev_jiffy = jiffies;
9202 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9203 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9204 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9205 current->pid, current->comm);
9207 debug_show_held_locks(current);
9209 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9211 EXPORT_SYMBOL_GPL(__cant_migrate);
9215 #ifdef CONFIG_MAGIC_SYSRQ
9216 void normalize_rt_tasks(void)
9218 struct task_struct *g, *p;
9219 struct sched_attr attr = {
9220 .sched_policy = SCHED_NORMAL,
9223 read_lock(&tasklist_lock);
9224 for_each_process_thread(g, p) {
9226 * Only normalize user tasks:
9228 if (p->flags & PF_KTHREAD)
9231 p->se.exec_start = 0;
9232 schedstat_set(p->se.statistics.wait_start, 0);
9233 schedstat_set(p->se.statistics.sleep_start, 0);
9234 schedstat_set(p->se.statistics.block_start, 0);
9236 if (!dl_task(p) && !rt_task(p)) {
9238 * Renice negative nice level userspace
9241 if (task_nice(p) < 0)
9242 set_user_nice(p, 0);
9246 __sched_setscheduler(p, &attr, false, false);
9248 read_unlock(&tasklist_lock);
9251 #endif /* CONFIG_MAGIC_SYSRQ */
9253 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9255 * These functions are only useful for the IA64 MCA handling, or kdb.
9257 * They can only be called when the whole system has been
9258 * stopped - every CPU needs to be quiescent, and no scheduling
9259 * activity can take place. Using them for anything else would
9260 * be a serious bug, and as a result, they aren't even visible
9261 * under any other configuration.
9265 * curr_task - return the current task for a given CPU.
9266 * @cpu: the processor in question.
9268 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9270 * Return: The current task for @cpu.
9272 struct task_struct *curr_task(int cpu)
9274 return cpu_curr(cpu);
9277 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9281 * ia64_set_curr_task - set the current task for a given CPU.
9282 * @cpu: the processor in question.
9283 * @p: the task pointer to set.
9285 * Description: This function must only be used when non-maskable interrupts
9286 * are serviced on a separate stack. It allows the architecture to switch the
9287 * notion of the current task on a CPU in a non-blocking manner. This function
9288 * must be called with all CPU's synchronized, and interrupts disabled, the
9289 * and caller must save the original value of the current task (see
9290 * curr_task() above) and restore that value before reenabling interrupts and
9291 * re-starting the system.
9293 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9295 void ia64_set_curr_task(int cpu, struct task_struct *p)
9302 #ifdef CONFIG_CGROUP_SCHED
9303 /* task_group_lock serializes the addition/removal of task groups */
9304 static DEFINE_SPINLOCK(task_group_lock);
9306 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9307 struct task_group *parent)
9309 #ifdef CONFIG_UCLAMP_TASK_GROUP
9310 enum uclamp_id clamp_id;
9312 for_each_clamp_id(clamp_id) {
9313 uclamp_se_set(&tg->uclamp_req[clamp_id],
9314 uclamp_none(clamp_id), false);
9315 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9320 static void sched_free_group(struct task_group *tg)
9322 free_fair_sched_group(tg);
9323 free_rt_sched_group(tg);
9325 kmem_cache_free(task_group_cache, tg);
9328 /* allocate runqueue etc for a new task group */
9329 struct task_group *sched_create_group(struct task_group *parent)
9331 struct task_group *tg;
9333 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9335 return ERR_PTR(-ENOMEM);
9337 if (!alloc_fair_sched_group(tg, parent))
9340 if (!alloc_rt_sched_group(tg, parent))
9343 alloc_uclamp_sched_group(tg, parent);
9348 sched_free_group(tg);
9349 return ERR_PTR(-ENOMEM);
9352 void sched_online_group(struct task_group *tg, struct task_group *parent)
9354 unsigned long flags;
9356 spin_lock_irqsave(&task_group_lock, flags);
9357 list_add_rcu(&tg->list, &task_groups);
9359 /* Root should already exist: */
9362 tg->parent = parent;
9363 INIT_LIST_HEAD(&tg->children);
9364 list_add_rcu(&tg->siblings, &parent->children);
9365 spin_unlock_irqrestore(&task_group_lock, flags);
9367 online_fair_sched_group(tg);
9370 /* rcu callback to free various structures associated with a task group */
9371 static void sched_free_group_rcu(struct rcu_head *rhp)
9373 /* Now it should be safe to free those cfs_rqs: */
9374 sched_free_group(container_of(rhp, struct task_group, rcu));
9377 void sched_destroy_group(struct task_group *tg)
9379 /* Wait for possible concurrent references to cfs_rqs complete: */
9380 call_rcu(&tg->rcu, sched_free_group_rcu);
9383 void sched_offline_group(struct task_group *tg)
9385 unsigned long flags;
9387 /* End participation in shares distribution: */
9388 unregister_fair_sched_group(tg);
9390 spin_lock_irqsave(&task_group_lock, flags);
9391 list_del_rcu(&tg->list);
9392 list_del_rcu(&tg->siblings);
9393 spin_unlock_irqrestore(&task_group_lock, flags);
9396 static void sched_change_group(struct task_struct *tsk, int type)
9398 struct task_group *tg;
9401 * All callers are synchronized by task_rq_lock(); we do not use RCU
9402 * which is pointless here. Thus, we pass "true" to task_css_check()
9403 * to prevent lockdep warnings.
9405 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9406 struct task_group, css);
9407 tg = autogroup_task_group(tsk, tg);
9408 tsk->sched_task_group = tg;
9410 #ifdef CONFIG_FAIR_GROUP_SCHED
9411 if (tsk->sched_class->task_change_group)
9412 tsk->sched_class->task_change_group(tsk, type);
9415 set_task_rq(tsk, task_cpu(tsk));
9419 * Change task's runqueue when it moves between groups.
9421 * The caller of this function should have put the task in its new group by
9422 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9425 void sched_move_task(struct task_struct *tsk)
9427 int queued, running, queue_flags =
9428 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9432 rq = task_rq_lock(tsk, &rf);
9433 update_rq_clock(rq);
9435 running = task_current(rq, tsk);
9436 queued = task_on_rq_queued(tsk);
9439 dequeue_task(rq, tsk, queue_flags);
9441 put_prev_task(rq, tsk);
9443 sched_change_group(tsk, TASK_MOVE_GROUP);
9446 enqueue_task(rq, tsk, queue_flags);
9448 set_next_task(rq, tsk);
9450 * After changing group, the running task may have joined a
9451 * throttled one but it's still the running task. Trigger a
9452 * resched to make sure that task can still run.
9457 task_rq_unlock(rq, tsk, &rf);
9460 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9462 return css ? container_of(css, struct task_group, css) : NULL;
9465 static struct cgroup_subsys_state *
9466 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9468 struct task_group *parent = css_tg(parent_css);
9469 struct task_group *tg;
9472 /* This is early initialization for the top cgroup */
9473 return &root_task_group.css;
9476 tg = sched_create_group(parent);
9478 return ERR_PTR(-ENOMEM);
9483 /* Expose task group only after completing cgroup initialization */
9484 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9486 struct task_group *tg = css_tg(css);
9487 struct task_group *parent = css_tg(css->parent);
9490 sched_online_group(tg, parent);
9492 #ifdef CONFIG_UCLAMP_TASK_GROUP
9493 /* Propagate the effective uclamp value for the new group */
9494 mutex_lock(&uclamp_mutex);
9496 cpu_util_update_eff(css);
9498 mutex_unlock(&uclamp_mutex);
9504 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9506 struct task_group *tg = css_tg(css);
9508 sched_offline_group(tg);
9511 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9513 struct task_group *tg = css_tg(css);
9516 * Relies on the RCU grace period between css_released() and this.
9518 sched_free_group(tg);
9522 * This is called before wake_up_new_task(), therefore we really only
9523 * have to set its group bits, all the other stuff does not apply.
9525 static void cpu_cgroup_fork(struct task_struct *task)
9530 rq = task_rq_lock(task, &rf);
9532 update_rq_clock(rq);
9533 sched_change_group(task, TASK_SET_GROUP);
9535 task_rq_unlock(rq, task, &rf);
9538 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9540 struct task_struct *task;
9541 struct cgroup_subsys_state *css;
9544 cgroup_taskset_for_each(task, css, tset) {
9545 #ifdef CONFIG_RT_GROUP_SCHED
9546 if (!sched_rt_can_attach(css_tg(css), task))
9550 * Serialize against wake_up_new_task() such that if it's
9551 * running, we're sure to observe its full state.
9553 raw_spin_lock_irq(&task->pi_lock);
9555 * Avoid calling sched_move_task() before wake_up_new_task()
9556 * has happened. This would lead to problems with PELT, due to
9557 * move wanting to detach+attach while we're not attached yet.
9559 if (READ_ONCE(task->__state) == TASK_NEW)
9561 raw_spin_unlock_irq(&task->pi_lock);
9569 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9571 struct task_struct *task;
9572 struct cgroup_subsys_state *css;
9574 cgroup_taskset_for_each(task, css, tset)
9575 sched_move_task(task);
9578 #ifdef CONFIG_UCLAMP_TASK_GROUP
9579 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9581 struct cgroup_subsys_state *top_css = css;
9582 struct uclamp_se *uc_parent = NULL;
9583 struct uclamp_se *uc_se = NULL;
9584 unsigned int eff[UCLAMP_CNT];
9585 enum uclamp_id clamp_id;
9586 unsigned int clamps;
9588 lockdep_assert_held(&uclamp_mutex);
9589 SCHED_WARN_ON(!rcu_read_lock_held());
9591 css_for_each_descendant_pre(css, top_css) {
9592 uc_parent = css_tg(css)->parent
9593 ? css_tg(css)->parent->uclamp : NULL;
9595 for_each_clamp_id(clamp_id) {
9596 /* Assume effective clamps matches requested clamps */
9597 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
9598 /* Cap effective clamps with parent's effective clamps */
9600 eff[clamp_id] > uc_parent[clamp_id].value) {
9601 eff[clamp_id] = uc_parent[clamp_id].value;
9604 /* Ensure protection is always capped by limit */
9605 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
9607 /* Propagate most restrictive effective clamps */
9609 uc_se = css_tg(css)->uclamp;
9610 for_each_clamp_id(clamp_id) {
9611 if (eff[clamp_id] == uc_se[clamp_id].value)
9613 uc_se[clamp_id].value = eff[clamp_id];
9614 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
9615 clamps |= (0x1 << clamp_id);
9618 css = css_rightmost_descendant(css);
9622 /* Immediately update descendants RUNNABLE tasks */
9623 uclamp_update_active_tasks(css);
9628 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
9629 * C expression. Since there is no way to convert a macro argument (N) into a
9630 * character constant, use two levels of macros.
9632 #define _POW10(exp) ((unsigned int)1e##exp)
9633 #define POW10(exp) _POW10(exp)
9635 struct uclamp_request {
9636 #define UCLAMP_PERCENT_SHIFT 2
9637 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
9643 static inline struct uclamp_request
9644 capacity_from_percent(char *buf)
9646 struct uclamp_request req = {
9647 .percent = UCLAMP_PERCENT_SCALE,
9648 .util = SCHED_CAPACITY_SCALE,
9653 if (strcmp(buf, "max")) {
9654 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
9658 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
9663 req.util = req.percent << SCHED_CAPACITY_SHIFT;
9664 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
9670 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
9671 size_t nbytes, loff_t off,
9672 enum uclamp_id clamp_id)
9674 struct uclamp_request req;
9675 struct task_group *tg;
9677 req = capacity_from_percent(buf);
9681 static_branch_enable(&sched_uclamp_used);
9683 mutex_lock(&uclamp_mutex);
9686 tg = css_tg(of_css(of));
9687 if (tg->uclamp_req[clamp_id].value != req.util)
9688 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
9691 * Because of not recoverable conversion rounding we keep track of the
9692 * exact requested value
9694 tg->uclamp_pct[clamp_id] = req.percent;
9696 /* Update effective clamps to track the most restrictive value */
9697 cpu_util_update_eff(of_css(of));
9700 mutex_unlock(&uclamp_mutex);
9705 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
9706 char *buf, size_t nbytes,
9709 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
9712 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
9713 char *buf, size_t nbytes,
9716 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
9719 static inline void cpu_uclamp_print(struct seq_file *sf,
9720 enum uclamp_id clamp_id)
9722 struct task_group *tg;
9728 tg = css_tg(seq_css(sf));
9729 util_clamp = tg->uclamp_req[clamp_id].value;
9732 if (util_clamp == SCHED_CAPACITY_SCALE) {
9733 seq_puts(sf, "max\n");
9737 percent = tg->uclamp_pct[clamp_id];
9738 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
9739 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
9742 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
9744 cpu_uclamp_print(sf, UCLAMP_MIN);
9748 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
9750 cpu_uclamp_print(sf, UCLAMP_MAX);
9753 #endif /* CONFIG_UCLAMP_TASK_GROUP */
9755 #ifdef CONFIG_FAIR_GROUP_SCHED
9756 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
9757 struct cftype *cftype, u64 shareval)
9759 if (shareval > scale_load_down(ULONG_MAX))
9760 shareval = MAX_SHARES;
9761 return sched_group_set_shares(css_tg(css), scale_load(shareval));
9764 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
9767 struct task_group *tg = css_tg(css);
9769 return (u64) scale_load_down(tg->shares);
9772 #ifdef CONFIG_CFS_BANDWIDTH
9773 static DEFINE_MUTEX(cfs_constraints_mutex);
9775 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9776 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9777 /* More than 203 days if BW_SHIFT equals 20. */
9778 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
9780 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9782 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
9785 int i, ret = 0, runtime_enabled, runtime_was_enabled;
9786 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9788 if (tg == &root_task_group)
9792 * Ensure we have at some amount of bandwidth every period. This is
9793 * to prevent reaching a state of large arrears when throttled via
9794 * entity_tick() resulting in prolonged exit starvation.
9796 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9800 * Likewise, bound things on the other side by preventing insane quota
9801 * periods. This also allows us to normalize in computing quota
9804 if (period > max_cfs_quota_period)
9808 * Bound quota to defend quota against overflow during bandwidth shift.
9810 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
9813 if (quota != RUNTIME_INF && (burst > quota ||
9814 burst + quota > max_cfs_runtime))
9818 * Prevent race between setting of cfs_rq->runtime_enabled and
9819 * unthrottle_offline_cfs_rqs().
9822 mutex_lock(&cfs_constraints_mutex);
9823 ret = __cfs_schedulable(tg, period, quota);
9827 runtime_enabled = quota != RUNTIME_INF;
9828 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
9830 * If we need to toggle cfs_bandwidth_used, off->on must occur
9831 * before making related changes, and on->off must occur afterwards
9833 if (runtime_enabled && !runtime_was_enabled)
9834 cfs_bandwidth_usage_inc();
9835 raw_spin_lock_irq(&cfs_b->lock);
9836 cfs_b->period = ns_to_ktime(period);
9837 cfs_b->quota = quota;
9838 cfs_b->burst = burst;
9840 __refill_cfs_bandwidth_runtime(cfs_b);
9842 /* Restart the period timer (if active) to handle new period expiry: */
9843 if (runtime_enabled)
9844 start_cfs_bandwidth(cfs_b);
9846 raw_spin_unlock_irq(&cfs_b->lock);
9848 for_each_online_cpu(i) {
9849 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9850 struct rq *rq = cfs_rq->rq;
9853 rq_lock_irq(rq, &rf);
9854 cfs_rq->runtime_enabled = runtime_enabled;
9855 cfs_rq->runtime_remaining = 0;
9857 if (cfs_rq->throttled)
9858 unthrottle_cfs_rq(cfs_rq);
9859 rq_unlock_irq(rq, &rf);
9861 if (runtime_was_enabled && !runtime_enabled)
9862 cfs_bandwidth_usage_dec();
9864 mutex_unlock(&cfs_constraints_mutex);
9870 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9872 u64 quota, period, burst;
9874 period = ktime_to_ns(tg->cfs_bandwidth.period);
9875 burst = tg->cfs_bandwidth.burst;
9876 if (cfs_quota_us < 0)
9877 quota = RUNTIME_INF;
9878 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
9879 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9883 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9886 static long tg_get_cfs_quota(struct task_group *tg)
9890 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
9893 quota_us = tg->cfs_bandwidth.quota;
9894 do_div(quota_us, NSEC_PER_USEC);
9899 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9901 u64 quota, period, burst;
9903 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
9906 period = (u64)cfs_period_us * NSEC_PER_USEC;
9907 quota = tg->cfs_bandwidth.quota;
9908 burst = tg->cfs_bandwidth.burst;
9910 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9913 static long tg_get_cfs_period(struct task_group *tg)
9917 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
9918 do_div(cfs_period_us, NSEC_PER_USEC);
9920 return cfs_period_us;
9923 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
9925 u64 quota, period, burst;
9927 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
9930 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
9931 period = ktime_to_ns(tg->cfs_bandwidth.period);
9932 quota = tg->cfs_bandwidth.quota;
9934 return tg_set_cfs_bandwidth(tg, period, quota, burst);
9937 static long tg_get_cfs_burst(struct task_group *tg)
9941 burst_us = tg->cfs_bandwidth.burst;
9942 do_div(burst_us, NSEC_PER_USEC);
9947 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
9950 return tg_get_cfs_quota(css_tg(css));
9953 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
9954 struct cftype *cftype, s64 cfs_quota_us)
9956 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
9959 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
9962 return tg_get_cfs_period(css_tg(css));
9965 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
9966 struct cftype *cftype, u64 cfs_period_us)
9968 return tg_set_cfs_period(css_tg(css), cfs_period_us);
9971 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
9974 return tg_get_cfs_burst(css_tg(css));
9977 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
9978 struct cftype *cftype, u64 cfs_burst_us)
9980 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
9983 struct cfs_schedulable_data {
9984 struct task_group *tg;
9989 * normalize group quota/period to be quota/max_period
9990 * note: units are usecs
9992 static u64 normalize_cfs_quota(struct task_group *tg,
9993 struct cfs_schedulable_data *d)
10001 period = tg_get_cfs_period(tg);
10002 quota = tg_get_cfs_quota(tg);
10005 /* note: these should typically be equivalent */
10006 if (quota == RUNTIME_INF || quota == -1)
10007 return RUNTIME_INF;
10009 return to_ratio(period, quota);
10012 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10014 struct cfs_schedulable_data *d = data;
10015 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10016 s64 quota = 0, parent_quota = -1;
10019 quota = RUNTIME_INF;
10021 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10023 quota = normalize_cfs_quota(tg, d);
10024 parent_quota = parent_b->hierarchical_quota;
10027 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10028 * always take the min. On cgroup1, only inherit when no
10031 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10032 quota = min(quota, parent_quota);
10034 if (quota == RUNTIME_INF)
10035 quota = parent_quota;
10036 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10040 cfs_b->hierarchical_quota = quota;
10045 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10048 struct cfs_schedulable_data data = {
10054 if (quota != RUNTIME_INF) {
10055 do_div(data.period, NSEC_PER_USEC);
10056 do_div(data.quota, NSEC_PER_USEC);
10060 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10066 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10068 struct task_group *tg = css_tg(seq_css(sf));
10069 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10071 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10072 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10073 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10075 if (schedstat_enabled() && tg != &root_task_group) {
10079 for_each_possible_cpu(i)
10080 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10082 seq_printf(sf, "wait_sum %llu\n", ws);
10087 #endif /* CONFIG_CFS_BANDWIDTH */
10088 #endif /* CONFIG_FAIR_GROUP_SCHED */
10090 #ifdef CONFIG_RT_GROUP_SCHED
10091 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10092 struct cftype *cft, s64 val)
10094 return sched_group_set_rt_runtime(css_tg(css), val);
10097 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10098 struct cftype *cft)
10100 return sched_group_rt_runtime(css_tg(css));
10103 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10104 struct cftype *cftype, u64 rt_period_us)
10106 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10109 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10110 struct cftype *cft)
10112 return sched_group_rt_period(css_tg(css));
10114 #endif /* CONFIG_RT_GROUP_SCHED */
10116 static struct cftype cpu_legacy_files[] = {
10117 #ifdef CONFIG_FAIR_GROUP_SCHED
10120 .read_u64 = cpu_shares_read_u64,
10121 .write_u64 = cpu_shares_write_u64,
10124 #ifdef CONFIG_CFS_BANDWIDTH
10126 .name = "cfs_quota_us",
10127 .read_s64 = cpu_cfs_quota_read_s64,
10128 .write_s64 = cpu_cfs_quota_write_s64,
10131 .name = "cfs_period_us",
10132 .read_u64 = cpu_cfs_period_read_u64,
10133 .write_u64 = cpu_cfs_period_write_u64,
10136 .name = "cfs_burst_us",
10137 .read_u64 = cpu_cfs_burst_read_u64,
10138 .write_u64 = cpu_cfs_burst_write_u64,
10142 .seq_show = cpu_cfs_stat_show,
10145 #ifdef CONFIG_RT_GROUP_SCHED
10147 .name = "rt_runtime_us",
10148 .read_s64 = cpu_rt_runtime_read,
10149 .write_s64 = cpu_rt_runtime_write,
10152 .name = "rt_period_us",
10153 .read_u64 = cpu_rt_period_read_uint,
10154 .write_u64 = cpu_rt_period_write_uint,
10157 #ifdef CONFIG_UCLAMP_TASK_GROUP
10159 .name = "uclamp.min",
10160 .flags = CFTYPE_NOT_ON_ROOT,
10161 .seq_show = cpu_uclamp_min_show,
10162 .write = cpu_uclamp_min_write,
10165 .name = "uclamp.max",
10166 .flags = CFTYPE_NOT_ON_ROOT,
10167 .seq_show = cpu_uclamp_max_show,
10168 .write = cpu_uclamp_max_write,
10171 { } /* Terminate */
10174 static int cpu_extra_stat_show(struct seq_file *sf,
10175 struct cgroup_subsys_state *css)
10177 #ifdef CONFIG_CFS_BANDWIDTH
10179 struct task_group *tg = css_tg(css);
10180 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10181 u64 throttled_usec;
10183 throttled_usec = cfs_b->throttled_time;
10184 do_div(throttled_usec, NSEC_PER_USEC);
10186 seq_printf(sf, "nr_periods %d\n"
10187 "nr_throttled %d\n"
10188 "throttled_usec %llu\n",
10189 cfs_b->nr_periods, cfs_b->nr_throttled,
10196 #ifdef CONFIG_FAIR_GROUP_SCHED
10197 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10198 struct cftype *cft)
10200 struct task_group *tg = css_tg(css);
10201 u64 weight = scale_load_down(tg->shares);
10203 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10206 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10207 struct cftype *cft, u64 weight)
10210 * cgroup weight knobs should use the common MIN, DFL and MAX
10211 * values which are 1, 100 and 10000 respectively. While it loses
10212 * a bit of range on both ends, it maps pretty well onto the shares
10213 * value used by scheduler and the round-trip conversions preserve
10214 * the original value over the entire range.
10216 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10219 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10221 return sched_group_set_shares(css_tg(css), scale_load(weight));
10224 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10225 struct cftype *cft)
10227 unsigned long weight = scale_load_down(css_tg(css)->shares);
10228 int last_delta = INT_MAX;
10231 /* find the closest nice value to the current weight */
10232 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10233 delta = abs(sched_prio_to_weight[prio] - weight);
10234 if (delta >= last_delta)
10236 last_delta = delta;
10239 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10242 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10243 struct cftype *cft, s64 nice)
10245 unsigned long weight;
10248 if (nice < MIN_NICE || nice > MAX_NICE)
10251 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10252 idx = array_index_nospec(idx, 40);
10253 weight = sched_prio_to_weight[idx];
10255 return sched_group_set_shares(css_tg(css), scale_load(weight));
10259 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10260 long period, long quota)
10263 seq_puts(sf, "max");
10265 seq_printf(sf, "%ld", quota);
10267 seq_printf(sf, " %ld\n", period);
10270 /* caller should put the current value in *@periodp before calling */
10271 static int __maybe_unused cpu_period_quota_parse(char *buf,
10272 u64 *periodp, u64 *quotap)
10274 char tok[21]; /* U64_MAX */
10276 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10279 *periodp *= NSEC_PER_USEC;
10281 if (sscanf(tok, "%llu", quotap))
10282 *quotap *= NSEC_PER_USEC;
10283 else if (!strcmp(tok, "max"))
10284 *quotap = RUNTIME_INF;
10291 #ifdef CONFIG_CFS_BANDWIDTH
10292 static int cpu_max_show(struct seq_file *sf, void *v)
10294 struct task_group *tg = css_tg(seq_css(sf));
10296 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10300 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10301 char *buf, size_t nbytes, loff_t off)
10303 struct task_group *tg = css_tg(of_css(of));
10304 u64 period = tg_get_cfs_period(tg);
10305 u64 burst = tg_get_cfs_burst(tg);
10309 ret = cpu_period_quota_parse(buf, &period, "a);
10311 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10312 return ret ?: nbytes;
10316 static struct cftype cpu_files[] = {
10317 #ifdef CONFIG_FAIR_GROUP_SCHED
10320 .flags = CFTYPE_NOT_ON_ROOT,
10321 .read_u64 = cpu_weight_read_u64,
10322 .write_u64 = cpu_weight_write_u64,
10325 .name = "weight.nice",
10326 .flags = CFTYPE_NOT_ON_ROOT,
10327 .read_s64 = cpu_weight_nice_read_s64,
10328 .write_s64 = cpu_weight_nice_write_s64,
10331 #ifdef CONFIG_CFS_BANDWIDTH
10334 .flags = CFTYPE_NOT_ON_ROOT,
10335 .seq_show = cpu_max_show,
10336 .write = cpu_max_write,
10339 .name = "max.burst",
10340 .flags = CFTYPE_NOT_ON_ROOT,
10341 .read_u64 = cpu_cfs_burst_read_u64,
10342 .write_u64 = cpu_cfs_burst_write_u64,
10345 #ifdef CONFIG_UCLAMP_TASK_GROUP
10347 .name = "uclamp.min",
10348 .flags = CFTYPE_NOT_ON_ROOT,
10349 .seq_show = cpu_uclamp_min_show,
10350 .write = cpu_uclamp_min_write,
10353 .name = "uclamp.max",
10354 .flags = CFTYPE_NOT_ON_ROOT,
10355 .seq_show = cpu_uclamp_max_show,
10356 .write = cpu_uclamp_max_write,
10359 { } /* terminate */
10362 struct cgroup_subsys cpu_cgrp_subsys = {
10363 .css_alloc = cpu_cgroup_css_alloc,
10364 .css_online = cpu_cgroup_css_online,
10365 .css_released = cpu_cgroup_css_released,
10366 .css_free = cpu_cgroup_css_free,
10367 .css_extra_stat_show = cpu_extra_stat_show,
10368 .fork = cpu_cgroup_fork,
10369 .can_attach = cpu_cgroup_can_attach,
10370 .attach = cpu_cgroup_attach,
10371 .legacy_cftypes = cpu_legacy_files,
10372 .dfl_cftypes = cpu_files,
10373 .early_init = true,
10377 #endif /* CONFIG_CGROUP_SCHED */
10379 void dump_cpu_task(int cpu)
10381 pr_info("Task dump for CPU %d:\n", cpu);
10382 sched_show_task(cpu_curr(cpu));
10386 * Nice levels are multiplicative, with a gentle 10% change for every
10387 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10388 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10389 * that remained on nice 0.
10391 * The "10% effect" is relative and cumulative: from _any_ nice level,
10392 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10393 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10394 * If a task goes up by ~10% and another task goes down by ~10% then
10395 * the relative distance between them is ~25%.)
10397 const int sched_prio_to_weight[40] = {
10398 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10399 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10400 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10401 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10402 /* 0 */ 1024, 820, 655, 526, 423,
10403 /* 5 */ 335, 272, 215, 172, 137,
10404 /* 10 */ 110, 87, 70, 56, 45,
10405 /* 15 */ 36, 29, 23, 18, 15,
10409 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10411 * In cases where the weight does not change often, we can use the
10412 * precalculated inverse to speed up arithmetics by turning divisions
10413 * into multiplications:
10415 const u32 sched_prio_to_wmult[40] = {
10416 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10417 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10418 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10419 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10420 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10421 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10422 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10423 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10426 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10428 trace_sched_update_nr_running_tp(rq, count);