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>
16 #include <linux/blkdev.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 #ifdef CONFIG_PREEMPT_RT
78 const_debug unsigned int sysctl_sched_nr_migrate = 8;
80 const_debug unsigned int sysctl_sched_nr_migrate = 32;
84 * period over which we measure -rt task CPU usage in us.
87 unsigned int sysctl_sched_rt_period = 1000000;
89 __read_mostly int scheduler_running;
91 #ifdef CONFIG_SCHED_CORE
93 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
95 /* kernel prio, less is more */
96 static inline int __task_prio(struct task_struct *p)
98 if (p->sched_class == &stop_sched_class) /* trumps deadline */
101 if (rt_prio(p->prio)) /* includes deadline */
102 return p->prio; /* [-1, 99] */
104 if (p->sched_class == &idle_sched_class)
105 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
107 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
117 /* real prio, less is less */
118 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
121 int pa = __task_prio(a), pb = __task_prio(b);
129 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
130 return !dl_time_before(a->dl.deadline, b->dl.deadline);
132 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
133 return cfs_prio_less(a, b, in_fi);
138 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
140 if (a->core_cookie < b->core_cookie)
143 if (a->core_cookie > b->core_cookie)
146 /* flip prio, so high prio is leftmost */
147 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
153 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
155 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
157 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
160 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
162 const struct task_struct *p = __node_2_sc(node);
163 unsigned long cookie = (unsigned long)key;
165 if (cookie < p->core_cookie)
168 if (cookie > p->core_cookie)
174 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
176 rq->core->core_task_seq++;
181 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
184 void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
186 rq->core->core_task_seq++;
188 if (sched_core_enqueued(p)) {
189 rb_erase(&p->core_node, &rq->core_tree);
190 RB_CLEAR_NODE(&p->core_node);
194 * Migrating the last task off the cpu, with the cpu in forced idle
195 * state. Reschedule to create an accounting edge for forced idle,
196 * and re-examine whether the core is still in forced idle state.
198 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
199 rq->core->core_forceidle_count && rq->curr == rq->idle)
204 * Find left-most (aka, highest priority) task matching @cookie.
206 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
208 struct rb_node *node;
210 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
212 * The idle task always matches any cookie!
215 return idle_sched_class.pick_task(rq);
217 return __node_2_sc(node);
220 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
222 struct rb_node *node = &p->core_node;
224 node = rb_next(node);
228 p = container_of(node, struct task_struct, core_node);
229 if (p->core_cookie != cookie)
236 * Magic required such that:
238 * raw_spin_rq_lock(rq);
240 * raw_spin_rq_unlock(rq);
242 * ends up locking and unlocking the _same_ lock, and all CPUs
243 * always agree on what rq has what lock.
245 * XXX entirely possible to selectively enable cores, don't bother for now.
248 static DEFINE_MUTEX(sched_core_mutex);
249 static atomic_t sched_core_count;
250 static struct cpumask sched_core_mask;
252 static void sched_core_lock(int cpu, unsigned long *flags)
254 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
257 local_irq_save(*flags);
258 for_each_cpu(t, smt_mask)
259 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
262 static void sched_core_unlock(int cpu, unsigned long *flags)
264 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
267 for_each_cpu(t, smt_mask)
268 raw_spin_unlock(&cpu_rq(t)->__lock);
269 local_irq_restore(*flags);
272 static void __sched_core_flip(bool enabled)
280 * Toggle the online cores, one by one.
282 cpumask_copy(&sched_core_mask, cpu_online_mask);
283 for_each_cpu(cpu, &sched_core_mask) {
284 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
286 sched_core_lock(cpu, &flags);
288 for_each_cpu(t, smt_mask)
289 cpu_rq(t)->core_enabled = enabled;
291 cpu_rq(cpu)->core->core_forceidle_start = 0;
293 sched_core_unlock(cpu, &flags);
295 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
299 * Toggle the offline CPUs.
301 cpumask_copy(&sched_core_mask, cpu_possible_mask);
302 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
304 for_each_cpu(cpu, &sched_core_mask)
305 cpu_rq(cpu)->core_enabled = enabled;
310 static void sched_core_assert_empty(void)
314 for_each_possible_cpu(cpu)
315 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
318 static void __sched_core_enable(void)
320 static_branch_enable(&__sched_core_enabled);
322 * Ensure all previous instances of raw_spin_rq_*lock() have finished
323 * and future ones will observe !sched_core_disabled().
326 __sched_core_flip(true);
327 sched_core_assert_empty();
330 static void __sched_core_disable(void)
332 sched_core_assert_empty();
333 __sched_core_flip(false);
334 static_branch_disable(&__sched_core_enabled);
337 void sched_core_get(void)
339 if (atomic_inc_not_zero(&sched_core_count))
342 mutex_lock(&sched_core_mutex);
343 if (!atomic_read(&sched_core_count))
344 __sched_core_enable();
346 smp_mb__before_atomic();
347 atomic_inc(&sched_core_count);
348 mutex_unlock(&sched_core_mutex);
351 static void __sched_core_put(struct work_struct *work)
353 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
354 __sched_core_disable();
355 mutex_unlock(&sched_core_mutex);
359 void sched_core_put(void)
361 static DECLARE_WORK(_work, __sched_core_put);
364 * "There can be only one"
366 * Either this is the last one, or we don't actually need to do any
367 * 'work'. If it is the last *again*, we rely on
368 * WORK_STRUCT_PENDING_BIT.
370 if (!atomic_add_unless(&sched_core_count, -1, 1))
371 schedule_work(&_work);
374 #else /* !CONFIG_SCHED_CORE */
376 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
378 sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
380 #endif /* CONFIG_SCHED_CORE */
383 * part of the period that we allow rt tasks to run in us.
386 int sysctl_sched_rt_runtime = 950000;
390 * Serialization rules:
396 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
399 * rq2->lock where: rq1 < rq2
403 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
404 * local CPU's rq->lock, it optionally removes the task from the runqueue and
405 * always looks at the local rq data structures to find the most eligible task
408 * Task enqueue is also under rq->lock, possibly taken from another CPU.
409 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
410 * the local CPU to avoid bouncing the runqueue state around [ see
411 * ttwu_queue_wakelist() ]
413 * Task wakeup, specifically wakeups that involve migration, are horribly
414 * complicated to avoid having to take two rq->locks.
418 * System-calls and anything external will use task_rq_lock() which acquires
419 * both p->pi_lock and rq->lock. As a consequence the state they change is
420 * stable while holding either lock:
422 * - sched_setaffinity()/
423 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
424 * - set_user_nice(): p->se.load, p->*prio
425 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
426 * p->se.load, p->rt_priority,
427 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
428 * - sched_setnuma(): p->numa_preferred_nid
429 * - sched_move_task()/
430 * cpu_cgroup_fork(): p->sched_task_group
431 * - uclamp_update_active() p->uclamp*
433 * p->state <- TASK_*:
435 * is changed locklessly using set_current_state(), __set_current_state() or
436 * set_special_state(), see their respective comments, or by
437 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
440 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
442 * is set by activate_task() and cleared by deactivate_task(), under
443 * rq->lock. Non-zero indicates the task is runnable, the special
444 * ON_RQ_MIGRATING state is used for migration without holding both
445 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
447 * p->on_cpu <- { 0, 1 }:
449 * is set by prepare_task() and cleared by finish_task() such that it will be
450 * set before p is scheduled-in and cleared after p is scheduled-out, both
451 * under rq->lock. Non-zero indicates the task is running on its CPU.
453 * [ The astute reader will observe that it is possible for two tasks on one
454 * CPU to have ->on_cpu = 1 at the same time. ]
456 * task_cpu(p): is changed by set_task_cpu(), the rules are:
458 * - Don't call set_task_cpu() on a blocked task:
460 * We don't care what CPU we're not running on, this simplifies hotplug,
461 * the CPU assignment of blocked tasks isn't required to be valid.
463 * - for try_to_wake_up(), called under p->pi_lock:
465 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
467 * - for migration called under rq->lock:
468 * [ see task_on_rq_migrating() in task_rq_lock() ]
470 * o move_queued_task()
473 * - for migration called under double_rq_lock():
475 * o __migrate_swap_task()
476 * o push_rt_task() / pull_rt_task()
477 * o push_dl_task() / pull_dl_task()
478 * o dl_task_offline_migration()
482 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
484 raw_spinlock_t *lock;
486 /* Matches synchronize_rcu() in __sched_core_enable() */
488 if (sched_core_disabled()) {
489 raw_spin_lock_nested(&rq->__lock, subclass);
490 /* preempt_count *MUST* be > 1 */
491 preempt_enable_no_resched();
496 lock = __rq_lockp(rq);
497 raw_spin_lock_nested(lock, subclass);
498 if (likely(lock == __rq_lockp(rq))) {
499 /* preempt_count *MUST* be > 1 */
500 preempt_enable_no_resched();
503 raw_spin_unlock(lock);
507 bool raw_spin_rq_trylock(struct rq *rq)
509 raw_spinlock_t *lock;
512 /* Matches synchronize_rcu() in __sched_core_enable() */
514 if (sched_core_disabled()) {
515 ret = raw_spin_trylock(&rq->__lock);
521 lock = __rq_lockp(rq);
522 ret = raw_spin_trylock(lock);
523 if (!ret || (likely(lock == __rq_lockp(rq)))) {
527 raw_spin_unlock(lock);
531 void raw_spin_rq_unlock(struct rq *rq)
533 raw_spin_unlock(rq_lockp(rq));
538 * double_rq_lock - safely lock two runqueues
540 void double_rq_lock(struct rq *rq1, struct rq *rq2)
542 lockdep_assert_irqs_disabled();
544 if (rq_order_less(rq2, rq1))
547 raw_spin_rq_lock(rq1);
548 if (__rq_lockp(rq1) == __rq_lockp(rq2))
551 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
556 * __task_rq_lock - lock the rq @p resides on.
558 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
563 lockdep_assert_held(&p->pi_lock);
567 raw_spin_rq_lock(rq);
568 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
572 raw_spin_rq_unlock(rq);
574 while (unlikely(task_on_rq_migrating(p)))
580 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
582 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
583 __acquires(p->pi_lock)
589 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
591 raw_spin_rq_lock(rq);
593 * move_queued_task() task_rq_lock()
596 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
597 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
598 * [S] ->cpu = new_cpu [L] task_rq()
602 * If we observe the old CPU in task_rq_lock(), the acquire of
603 * the old rq->lock will fully serialize against the stores.
605 * If we observe the new CPU in task_rq_lock(), the address
606 * dependency headed by '[L] rq = task_rq()' and the acquire
607 * will pair with the WMB to ensure we then also see migrating.
609 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
613 raw_spin_rq_unlock(rq);
614 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
616 while (unlikely(task_on_rq_migrating(p)))
622 * RQ-clock updating methods:
625 static void update_rq_clock_task(struct rq *rq, s64 delta)
628 * In theory, the compile should just see 0 here, and optimize out the call
629 * to sched_rt_avg_update. But I don't trust it...
631 s64 __maybe_unused steal = 0, irq_delta = 0;
633 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
634 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
637 * Since irq_time is only updated on {soft,}irq_exit, we might run into
638 * this case when a previous update_rq_clock() happened inside a
641 * When this happens, we stop ->clock_task and only update the
642 * prev_irq_time stamp to account for the part that fit, so that a next
643 * update will consume the rest. This ensures ->clock_task is
646 * It does however cause some slight miss-attribution of {soft,}irq
647 * time, a more accurate solution would be to update the irq_time using
648 * the current rq->clock timestamp, except that would require using
651 if (irq_delta > delta)
654 rq->prev_irq_time += irq_delta;
657 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
658 if (static_key_false((¶virt_steal_rq_enabled))) {
659 steal = paravirt_steal_clock(cpu_of(rq));
660 steal -= rq->prev_steal_time_rq;
662 if (unlikely(steal > delta))
665 rq->prev_steal_time_rq += steal;
670 rq->clock_task += delta;
672 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
673 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
674 update_irq_load_avg(rq, irq_delta + steal);
676 update_rq_clock_pelt(rq, delta);
679 void update_rq_clock(struct rq *rq)
683 lockdep_assert_rq_held(rq);
685 if (rq->clock_update_flags & RQCF_ACT_SKIP)
688 #ifdef CONFIG_SCHED_DEBUG
689 if (sched_feat(WARN_DOUBLE_CLOCK))
690 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
691 rq->clock_update_flags |= RQCF_UPDATED;
694 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
698 update_rq_clock_task(rq, delta);
701 #ifdef CONFIG_SCHED_HRTICK
703 * Use HR-timers to deliver accurate preemption points.
706 static void hrtick_clear(struct rq *rq)
708 if (hrtimer_active(&rq->hrtick_timer))
709 hrtimer_cancel(&rq->hrtick_timer);
713 * High-resolution timer tick.
714 * Runs from hardirq context with interrupts disabled.
716 static enum hrtimer_restart hrtick(struct hrtimer *timer)
718 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
721 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
725 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
728 return HRTIMER_NORESTART;
733 static void __hrtick_restart(struct rq *rq)
735 struct hrtimer *timer = &rq->hrtick_timer;
736 ktime_t time = rq->hrtick_time;
738 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
742 * called from hardirq (IPI) context
744 static void __hrtick_start(void *arg)
750 __hrtick_restart(rq);
755 * Called to set the hrtick timer state.
757 * called with rq->lock held and irqs disabled
759 void hrtick_start(struct rq *rq, u64 delay)
761 struct hrtimer *timer = &rq->hrtick_timer;
765 * Don't schedule slices shorter than 10000ns, that just
766 * doesn't make sense and can cause timer DoS.
768 delta = max_t(s64, delay, 10000LL);
769 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
772 __hrtick_restart(rq);
774 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
779 * Called to set the hrtick timer state.
781 * called with rq->lock held and irqs disabled
783 void hrtick_start(struct rq *rq, u64 delay)
786 * Don't schedule slices shorter than 10000ns, that just
787 * doesn't make sense. Rely on vruntime for fairness.
789 delay = max_t(u64, delay, 10000LL);
790 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
791 HRTIMER_MODE_REL_PINNED_HARD);
794 #endif /* CONFIG_SMP */
796 static void hrtick_rq_init(struct rq *rq)
799 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
801 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
802 rq->hrtick_timer.function = hrtick;
804 #else /* CONFIG_SCHED_HRTICK */
805 static inline void hrtick_clear(struct rq *rq)
809 static inline void hrtick_rq_init(struct rq *rq)
812 #endif /* CONFIG_SCHED_HRTICK */
815 * cmpxchg based fetch_or, macro so it works for different integer types
817 #define fetch_or(ptr, mask) \
819 typeof(ptr) _ptr = (ptr); \
820 typeof(mask) _mask = (mask); \
821 typeof(*_ptr) _old, _val = *_ptr; \
824 _old = cmpxchg(_ptr, _val, _val | _mask); \
832 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
834 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
835 * this avoids any races wrt polling state changes and thereby avoids
838 static bool set_nr_and_not_polling(struct task_struct *p)
840 struct thread_info *ti = task_thread_info(p);
841 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
845 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
847 * If this returns true, then the idle task promises to call
848 * sched_ttwu_pending() and reschedule soon.
850 static bool set_nr_if_polling(struct task_struct *p)
852 struct thread_info *ti = task_thread_info(p);
853 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
856 if (!(val & _TIF_POLLING_NRFLAG))
858 if (val & _TIF_NEED_RESCHED)
860 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
869 static bool set_nr_and_not_polling(struct task_struct *p)
871 set_tsk_need_resched(p);
876 static bool set_nr_if_polling(struct task_struct *p)
883 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
885 struct wake_q_node *node = &task->wake_q;
888 * Atomically grab the task, if ->wake_q is !nil already it means
889 * it's already queued (either by us or someone else) and will get the
890 * wakeup due to that.
892 * In order to ensure that a pending wakeup will observe our pending
893 * state, even in the failed case, an explicit smp_mb() must be used.
895 smp_mb__before_atomic();
896 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
900 * The head is context local, there can be no concurrency.
903 head->lastp = &node->next;
908 * wake_q_add() - queue a wakeup for 'later' waking.
909 * @head: the wake_q_head to add @task to
910 * @task: the task to queue for 'later' wakeup
912 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
913 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
916 * This function must be used as-if it were wake_up_process(); IOW the task
917 * must be ready to be woken at this location.
919 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
921 if (__wake_q_add(head, task))
922 get_task_struct(task);
926 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
927 * @head: the wake_q_head to add @task to
928 * @task: the task to queue for 'later' wakeup
930 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
931 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
934 * This function must be used as-if it were wake_up_process(); IOW the task
935 * must be ready to be woken at this location.
937 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
938 * that already hold reference to @task can call the 'safe' version and trust
939 * wake_q to do the right thing depending whether or not the @task is already
942 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
944 if (!__wake_q_add(head, task))
945 put_task_struct(task);
948 void wake_up_q(struct wake_q_head *head)
950 struct wake_q_node *node = head->first;
952 while (node != WAKE_Q_TAIL) {
953 struct task_struct *task;
955 task = container_of(node, struct task_struct, wake_q);
956 /* Task can safely be re-inserted now: */
958 task->wake_q.next = NULL;
961 * wake_up_process() executes a full barrier, which pairs with
962 * the queueing in wake_q_add() so as not to miss wakeups.
964 wake_up_process(task);
965 put_task_struct(task);
970 * resched_curr - mark rq's current task 'to be rescheduled now'.
972 * On UP this means the setting of the need_resched flag, on SMP it
973 * might also involve a cross-CPU call to trigger the scheduler on
976 void resched_curr(struct rq *rq)
978 struct task_struct *curr = rq->curr;
981 lockdep_assert_rq_held(rq);
983 if (test_tsk_need_resched(curr))
988 if (cpu == smp_processor_id()) {
989 set_tsk_need_resched(curr);
990 set_preempt_need_resched();
994 if (set_nr_and_not_polling(curr))
995 smp_send_reschedule(cpu);
997 trace_sched_wake_idle_without_ipi(cpu);
1000 void resched_cpu(int cpu)
1002 struct rq *rq = cpu_rq(cpu);
1003 unsigned long flags;
1005 raw_spin_rq_lock_irqsave(rq, flags);
1006 if (cpu_online(cpu) || cpu == smp_processor_id())
1008 raw_spin_rq_unlock_irqrestore(rq, flags);
1012 #ifdef CONFIG_NO_HZ_COMMON
1014 * In the semi idle case, use the nearest busy CPU for migrating timers
1015 * from an idle CPU. This is good for power-savings.
1017 * We don't do similar optimization for completely idle system, as
1018 * selecting an idle CPU will add more delays to the timers than intended
1019 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
1021 int get_nohz_timer_target(void)
1023 int i, cpu = smp_processor_id(), default_cpu = -1;
1024 struct sched_domain *sd;
1025 const struct cpumask *hk_mask;
1027 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1033 hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1036 for_each_domain(cpu, sd) {
1037 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1048 if (default_cpu == -1)
1049 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1057 * When add_timer_on() enqueues a timer into the timer wheel of an
1058 * idle CPU then this timer might expire before the next timer event
1059 * which is scheduled to wake up that CPU. In case of a completely
1060 * idle system the next event might even be infinite time into the
1061 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1062 * leaves the inner idle loop so the newly added timer is taken into
1063 * account when the CPU goes back to idle and evaluates the timer
1064 * wheel for the next timer event.
1066 static void wake_up_idle_cpu(int cpu)
1068 struct rq *rq = cpu_rq(cpu);
1070 if (cpu == smp_processor_id())
1073 if (set_nr_and_not_polling(rq->idle))
1074 smp_send_reschedule(cpu);
1076 trace_sched_wake_idle_without_ipi(cpu);
1079 static bool wake_up_full_nohz_cpu(int cpu)
1082 * We just need the target to call irq_exit() and re-evaluate
1083 * the next tick. The nohz full kick at least implies that.
1084 * If needed we can still optimize that later with an
1087 if (cpu_is_offline(cpu))
1088 return true; /* Don't try to wake offline CPUs. */
1089 if (tick_nohz_full_cpu(cpu)) {
1090 if (cpu != smp_processor_id() ||
1091 tick_nohz_tick_stopped())
1092 tick_nohz_full_kick_cpu(cpu);
1100 * Wake up the specified CPU. If the CPU is going offline, it is the
1101 * caller's responsibility to deal with the lost wakeup, for example,
1102 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1104 void wake_up_nohz_cpu(int cpu)
1106 if (!wake_up_full_nohz_cpu(cpu))
1107 wake_up_idle_cpu(cpu);
1110 static void nohz_csd_func(void *info)
1112 struct rq *rq = info;
1113 int cpu = cpu_of(rq);
1117 * Release the rq::nohz_csd.
1119 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1120 WARN_ON(!(flags & NOHZ_KICK_MASK));
1122 rq->idle_balance = idle_cpu(cpu);
1123 if (rq->idle_balance && !need_resched()) {
1124 rq->nohz_idle_balance = flags;
1125 raise_softirq_irqoff(SCHED_SOFTIRQ);
1129 #endif /* CONFIG_NO_HZ_COMMON */
1131 #ifdef CONFIG_NO_HZ_FULL
1132 bool sched_can_stop_tick(struct rq *rq)
1134 int fifo_nr_running;
1136 /* Deadline tasks, even if single, need the tick */
1137 if (rq->dl.dl_nr_running)
1141 * If there are more than one RR tasks, we need the tick to affect the
1142 * actual RR behaviour.
1144 if (rq->rt.rr_nr_running) {
1145 if (rq->rt.rr_nr_running == 1)
1152 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1153 * forced preemption between FIFO tasks.
1155 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1156 if (fifo_nr_running)
1160 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1161 * if there's more than one we need the tick for involuntary
1164 if (rq->nr_running > 1)
1169 #endif /* CONFIG_NO_HZ_FULL */
1170 #endif /* CONFIG_SMP */
1172 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1173 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1175 * Iterate task_group tree rooted at *from, calling @down when first entering a
1176 * node and @up when leaving it for the final time.
1178 * Caller must hold rcu_lock or sufficient equivalent.
1180 int walk_tg_tree_from(struct task_group *from,
1181 tg_visitor down, tg_visitor up, void *data)
1183 struct task_group *parent, *child;
1189 ret = (*down)(parent, data);
1192 list_for_each_entry_rcu(child, &parent->children, siblings) {
1199 ret = (*up)(parent, data);
1200 if (ret || parent == from)
1204 parent = parent->parent;
1211 int tg_nop(struct task_group *tg, void *data)
1217 static void set_load_weight(struct task_struct *p, bool update_load)
1219 int prio = p->static_prio - MAX_RT_PRIO;
1220 struct load_weight *load = &p->se.load;
1223 * SCHED_IDLE tasks get minimal weight:
1225 if (task_has_idle_policy(p)) {
1226 load->weight = scale_load(WEIGHT_IDLEPRIO);
1227 load->inv_weight = WMULT_IDLEPRIO;
1232 * SCHED_OTHER tasks have to update their load when changing their
1235 if (update_load && p->sched_class == &fair_sched_class) {
1236 reweight_task(p, prio);
1238 load->weight = scale_load(sched_prio_to_weight[prio]);
1239 load->inv_weight = sched_prio_to_wmult[prio];
1243 #ifdef CONFIG_UCLAMP_TASK
1245 * Serializes updates of utilization clamp values
1247 * The (slow-path) user-space triggers utilization clamp value updates which
1248 * can require updates on (fast-path) scheduler's data structures used to
1249 * support enqueue/dequeue operations.
1250 * While the per-CPU rq lock protects fast-path update operations, user-space
1251 * requests are serialized using a mutex to reduce the risk of conflicting
1252 * updates or API abuses.
1254 static DEFINE_MUTEX(uclamp_mutex);
1256 /* Max allowed minimum utilization */
1257 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1259 /* Max allowed maximum utilization */
1260 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1263 * By default RT tasks run at the maximum performance point/capacity of the
1264 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1265 * SCHED_CAPACITY_SCALE.
1267 * This knob allows admins to change the default behavior when uclamp is being
1268 * used. In battery powered devices, particularly, running at the maximum
1269 * capacity and frequency will increase energy consumption and shorten the
1272 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1274 * This knob will not override the system default sched_util_clamp_min defined
1277 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1279 /* All clamps are required to be less or equal than these values */
1280 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1283 * This static key is used to reduce the uclamp overhead in the fast path. It
1284 * primarily disables the call to uclamp_rq_{inc, dec}() in
1285 * enqueue/dequeue_task().
1287 * This allows users to continue to enable uclamp in their kernel config with
1288 * minimum uclamp overhead in the fast path.
1290 * As soon as userspace modifies any of the uclamp knobs, the static key is
1291 * enabled, since we have an actual users that make use of uclamp
1294 * The knobs that would enable this static key are:
1296 * * A task modifying its uclamp value with sched_setattr().
1297 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1298 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1300 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1302 /* Integer rounded range for each bucket */
1303 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1305 #define for_each_clamp_id(clamp_id) \
1306 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1308 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1310 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1313 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1315 if (clamp_id == UCLAMP_MIN)
1317 return SCHED_CAPACITY_SCALE;
1320 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1321 unsigned int value, bool user_defined)
1323 uc_se->value = value;
1324 uc_se->bucket_id = uclamp_bucket_id(value);
1325 uc_se->user_defined = user_defined;
1328 static inline unsigned int
1329 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1330 unsigned int clamp_value)
1333 * Avoid blocked utilization pushing up the frequency when we go
1334 * idle (which drops the max-clamp) by retaining the last known
1337 if (clamp_id == UCLAMP_MAX) {
1338 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1342 return uclamp_none(UCLAMP_MIN);
1345 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1346 unsigned int clamp_value)
1348 /* Reset max-clamp retention only on idle exit */
1349 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1352 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1356 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1357 unsigned int clamp_value)
1359 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1360 int bucket_id = UCLAMP_BUCKETS - 1;
1363 * Since both min and max clamps are max aggregated, find the
1364 * top most bucket with tasks in.
1366 for ( ; bucket_id >= 0; bucket_id--) {
1367 if (!bucket[bucket_id].tasks)
1369 return bucket[bucket_id].value;
1372 /* No tasks -- default clamp values */
1373 return uclamp_idle_value(rq, clamp_id, clamp_value);
1376 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1378 unsigned int default_util_min;
1379 struct uclamp_se *uc_se;
1381 lockdep_assert_held(&p->pi_lock);
1383 uc_se = &p->uclamp_req[UCLAMP_MIN];
1385 /* Only sync if user didn't override the default */
1386 if (uc_se->user_defined)
1389 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1390 uclamp_se_set(uc_se, default_util_min, false);
1393 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1401 /* Protect updates to p->uclamp_* */
1402 rq = task_rq_lock(p, &rf);
1403 __uclamp_update_util_min_rt_default(p);
1404 task_rq_unlock(rq, p, &rf);
1407 static void uclamp_sync_util_min_rt_default(void)
1409 struct task_struct *g, *p;
1412 * copy_process() sysctl_uclamp
1413 * uclamp_min_rt = X;
1414 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1415 * // link thread smp_mb__after_spinlock()
1416 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1417 * sched_post_fork() for_each_process_thread()
1418 * __uclamp_sync_rt() __uclamp_sync_rt()
1420 * Ensures that either sched_post_fork() will observe the new
1421 * uclamp_min_rt or for_each_process_thread() will observe the new
1424 read_lock(&tasklist_lock);
1425 smp_mb__after_spinlock();
1426 read_unlock(&tasklist_lock);
1429 for_each_process_thread(g, p)
1430 uclamp_update_util_min_rt_default(p);
1434 static inline struct uclamp_se
1435 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1437 /* Copy by value as we could modify it */
1438 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1439 #ifdef CONFIG_UCLAMP_TASK_GROUP
1440 unsigned int tg_min, tg_max, value;
1443 * Tasks in autogroups or root task group will be
1444 * restricted by system defaults.
1446 if (task_group_is_autogroup(task_group(p)))
1448 if (task_group(p) == &root_task_group)
1451 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1452 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1453 value = uc_req.value;
1454 value = clamp(value, tg_min, tg_max);
1455 uclamp_se_set(&uc_req, value, false);
1462 * The effective clamp bucket index of a task depends on, by increasing
1464 * - the task specific clamp value, when explicitly requested from userspace
1465 * - the task group effective clamp value, for tasks not either in the root
1466 * group or in an autogroup
1467 * - the system default clamp value, defined by the sysadmin
1469 static inline struct uclamp_se
1470 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1472 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1473 struct uclamp_se uc_max = uclamp_default[clamp_id];
1475 /* System default restrictions always apply */
1476 if (unlikely(uc_req.value > uc_max.value))
1482 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1484 struct uclamp_se uc_eff;
1486 /* Task currently refcounted: use back-annotated (effective) value */
1487 if (p->uclamp[clamp_id].active)
1488 return (unsigned long)p->uclamp[clamp_id].value;
1490 uc_eff = uclamp_eff_get(p, clamp_id);
1492 return (unsigned long)uc_eff.value;
1496 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1497 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1498 * updates the rq's clamp value if required.
1500 * Tasks can have a task-specific value requested from user-space, track
1501 * within each bucket the maximum value for tasks refcounted in it.
1502 * This "local max aggregation" allows to track the exact "requested" value
1503 * for each bucket when all its RUNNABLE tasks require the same clamp.
1505 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1506 enum uclamp_id clamp_id)
1508 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1509 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1510 struct uclamp_bucket *bucket;
1512 lockdep_assert_rq_held(rq);
1514 /* Update task effective clamp */
1515 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1517 bucket = &uc_rq->bucket[uc_se->bucket_id];
1519 uc_se->active = true;
1521 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1524 * Local max aggregation: rq buckets always track the max
1525 * "requested" clamp value of its RUNNABLE tasks.
1527 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1528 bucket->value = uc_se->value;
1530 if (uc_se->value > READ_ONCE(uc_rq->value))
1531 WRITE_ONCE(uc_rq->value, uc_se->value);
1535 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1536 * is released. If this is the last task reference counting the rq's max
1537 * active clamp value, then the rq's clamp value is updated.
1539 * Both refcounted tasks and rq's cached clamp values are expected to be
1540 * always valid. If it's detected they are not, as defensive programming,
1541 * enforce the expected state and warn.
1543 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1544 enum uclamp_id clamp_id)
1546 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1547 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1548 struct uclamp_bucket *bucket;
1549 unsigned int bkt_clamp;
1550 unsigned int rq_clamp;
1552 lockdep_assert_rq_held(rq);
1555 * If sched_uclamp_used was enabled after task @p was enqueued,
1556 * we could end up with unbalanced call to uclamp_rq_dec_id().
1558 * In this case the uc_se->active flag should be false since no uclamp
1559 * accounting was performed at enqueue time and we can just return
1562 * Need to be careful of the following enqueue/dequeue ordering
1566 * // sched_uclamp_used gets enabled
1569 * // Must not decrement bucket->tasks here
1572 * where we could end up with stale data in uc_se and
1573 * bucket[uc_se->bucket_id].
1575 * The following check here eliminates the possibility of such race.
1577 if (unlikely(!uc_se->active))
1580 bucket = &uc_rq->bucket[uc_se->bucket_id];
1582 SCHED_WARN_ON(!bucket->tasks);
1583 if (likely(bucket->tasks))
1586 uc_se->active = false;
1589 * Keep "local max aggregation" simple and accept to (possibly)
1590 * overboost some RUNNABLE tasks in the same bucket.
1591 * The rq clamp bucket value is reset to its base value whenever
1592 * there are no more RUNNABLE tasks refcounting it.
1594 if (likely(bucket->tasks))
1597 rq_clamp = READ_ONCE(uc_rq->value);
1599 * Defensive programming: this should never happen. If it happens,
1600 * e.g. due to future modification, warn and fixup the expected value.
1602 SCHED_WARN_ON(bucket->value > rq_clamp);
1603 if (bucket->value >= rq_clamp) {
1604 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1605 WRITE_ONCE(uc_rq->value, bkt_clamp);
1609 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1611 enum uclamp_id clamp_id;
1614 * Avoid any overhead until uclamp is actually used by the userspace.
1616 * The condition is constructed such that a NOP is generated when
1617 * sched_uclamp_used is disabled.
1619 if (!static_branch_unlikely(&sched_uclamp_used))
1622 if (unlikely(!p->sched_class->uclamp_enabled))
1625 for_each_clamp_id(clamp_id)
1626 uclamp_rq_inc_id(rq, p, clamp_id);
1628 /* Reset clamp idle holding when there is one RUNNABLE task */
1629 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1630 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1633 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1635 enum uclamp_id clamp_id;
1638 * Avoid any overhead until uclamp is actually used by the userspace.
1640 * The condition is constructed such that a NOP is generated when
1641 * sched_uclamp_used is disabled.
1643 if (!static_branch_unlikely(&sched_uclamp_used))
1646 if (unlikely(!p->sched_class->uclamp_enabled))
1649 for_each_clamp_id(clamp_id)
1650 uclamp_rq_dec_id(rq, p, clamp_id);
1653 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1654 enum uclamp_id clamp_id)
1656 if (!p->uclamp[clamp_id].active)
1659 uclamp_rq_dec_id(rq, p, clamp_id);
1660 uclamp_rq_inc_id(rq, p, clamp_id);
1663 * Make sure to clear the idle flag if we've transiently reached 0
1664 * active tasks on rq.
1666 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1667 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1671 uclamp_update_active(struct task_struct *p)
1673 enum uclamp_id clamp_id;
1678 * Lock the task and the rq where the task is (or was) queued.
1680 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1681 * price to pay to safely serialize util_{min,max} updates with
1682 * enqueues, dequeues and migration operations.
1683 * This is the same locking schema used by __set_cpus_allowed_ptr().
1685 rq = task_rq_lock(p, &rf);
1688 * Setting the clamp bucket is serialized by task_rq_lock().
1689 * If the task is not yet RUNNABLE and its task_struct is not
1690 * affecting a valid clamp bucket, the next time it's enqueued,
1691 * it will already see the updated clamp bucket value.
1693 for_each_clamp_id(clamp_id)
1694 uclamp_rq_reinc_id(rq, p, clamp_id);
1696 task_rq_unlock(rq, p, &rf);
1699 #ifdef CONFIG_UCLAMP_TASK_GROUP
1701 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1703 struct css_task_iter it;
1704 struct task_struct *p;
1706 css_task_iter_start(css, 0, &it);
1707 while ((p = css_task_iter_next(&it)))
1708 uclamp_update_active(p);
1709 css_task_iter_end(&it);
1712 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1713 static void uclamp_update_root_tg(void)
1715 struct task_group *tg = &root_task_group;
1717 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1718 sysctl_sched_uclamp_util_min, false);
1719 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1720 sysctl_sched_uclamp_util_max, false);
1723 cpu_util_update_eff(&root_task_group.css);
1727 static void uclamp_update_root_tg(void) { }
1730 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1731 void *buffer, size_t *lenp, loff_t *ppos)
1733 bool update_root_tg = false;
1734 int old_min, old_max, old_min_rt;
1737 mutex_lock(&uclamp_mutex);
1738 old_min = sysctl_sched_uclamp_util_min;
1739 old_max = sysctl_sched_uclamp_util_max;
1740 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1742 result = proc_dointvec(table, write, buffer, lenp, ppos);
1748 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1749 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1750 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1756 if (old_min != sysctl_sched_uclamp_util_min) {
1757 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1758 sysctl_sched_uclamp_util_min, false);
1759 update_root_tg = true;
1761 if (old_max != sysctl_sched_uclamp_util_max) {
1762 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1763 sysctl_sched_uclamp_util_max, false);
1764 update_root_tg = true;
1767 if (update_root_tg) {
1768 static_branch_enable(&sched_uclamp_used);
1769 uclamp_update_root_tg();
1772 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1773 static_branch_enable(&sched_uclamp_used);
1774 uclamp_sync_util_min_rt_default();
1778 * We update all RUNNABLE tasks only when task groups are in use.
1779 * Otherwise, keep it simple and do just a lazy update at each next
1780 * task enqueue time.
1786 sysctl_sched_uclamp_util_min = old_min;
1787 sysctl_sched_uclamp_util_max = old_max;
1788 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1790 mutex_unlock(&uclamp_mutex);
1795 static int uclamp_validate(struct task_struct *p,
1796 const struct sched_attr *attr)
1798 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1799 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1801 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1802 util_min = attr->sched_util_min;
1804 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1808 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1809 util_max = attr->sched_util_max;
1811 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1815 if (util_min != -1 && util_max != -1 && util_min > util_max)
1819 * We have valid uclamp attributes; make sure uclamp is enabled.
1821 * We need to do that here, because enabling static branches is a
1822 * blocking operation which obviously cannot be done while holding
1825 static_branch_enable(&sched_uclamp_used);
1830 static bool uclamp_reset(const struct sched_attr *attr,
1831 enum uclamp_id clamp_id,
1832 struct uclamp_se *uc_se)
1834 /* Reset on sched class change for a non user-defined clamp value. */
1835 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1836 !uc_se->user_defined)
1839 /* Reset on sched_util_{min,max} == -1. */
1840 if (clamp_id == UCLAMP_MIN &&
1841 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1842 attr->sched_util_min == -1) {
1846 if (clamp_id == UCLAMP_MAX &&
1847 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1848 attr->sched_util_max == -1) {
1855 static void __setscheduler_uclamp(struct task_struct *p,
1856 const struct sched_attr *attr)
1858 enum uclamp_id clamp_id;
1860 for_each_clamp_id(clamp_id) {
1861 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1864 if (!uclamp_reset(attr, clamp_id, uc_se))
1868 * RT by default have a 100% boost value that could be modified
1871 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1872 value = sysctl_sched_uclamp_util_min_rt_default;
1874 value = uclamp_none(clamp_id);
1876 uclamp_se_set(uc_se, value, false);
1880 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1883 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1884 attr->sched_util_min != -1) {
1885 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1886 attr->sched_util_min, true);
1889 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1890 attr->sched_util_max != -1) {
1891 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1892 attr->sched_util_max, true);
1896 static void uclamp_fork(struct task_struct *p)
1898 enum uclamp_id clamp_id;
1901 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1902 * as the task is still at its early fork stages.
1904 for_each_clamp_id(clamp_id)
1905 p->uclamp[clamp_id].active = false;
1907 if (likely(!p->sched_reset_on_fork))
1910 for_each_clamp_id(clamp_id) {
1911 uclamp_se_set(&p->uclamp_req[clamp_id],
1912 uclamp_none(clamp_id), false);
1916 static void uclamp_post_fork(struct task_struct *p)
1918 uclamp_update_util_min_rt_default(p);
1921 static void __init init_uclamp_rq(struct rq *rq)
1923 enum uclamp_id clamp_id;
1924 struct uclamp_rq *uc_rq = rq->uclamp;
1926 for_each_clamp_id(clamp_id) {
1927 uc_rq[clamp_id] = (struct uclamp_rq) {
1928 .value = uclamp_none(clamp_id)
1932 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1935 static void __init init_uclamp(void)
1937 struct uclamp_se uc_max = {};
1938 enum uclamp_id clamp_id;
1941 for_each_possible_cpu(cpu)
1942 init_uclamp_rq(cpu_rq(cpu));
1944 for_each_clamp_id(clamp_id) {
1945 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1946 uclamp_none(clamp_id), false);
1949 /* System defaults allow max clamp values for both indexes */
1950 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1951 for_each_clamp_id(clamp_id) {
1952 uclamp_default[clamp_id] = uc_max;
1953 #ifdef CONFIG_UCLAMP_TASK_GROUP
1954 root_task_group.uclamp_req[clamp_id] = uc_max;
1955 root_task_group.uclamp[clamp_id] = uc_max;
1960 #else /* CONFIG_UCLAMP_TASK */
1961 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1962 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1963 static inline int uclamp_validate(struct task_struct *p,
1964 const struct sched_attr *attr)
1968 static void __setscheduler_uclamp(struct task_struct *p,
1969 const struct sched_attr *attr) { }
1970 static inline void uclamp_fork(struct task_struct *p) { }
1971 static inline void uclamp_post_fork(struct task_struct *p) { }
1972 static inline void init_uclamp(void) { }
1973 #endif /* CONFIG_UCLAMP_TASK */
1975 bool sched_task_on_rq(struct task_struct *p)
1977 return task_on_rq_queued(p);
1980 unsigned long get_wchan(struct task_struct *p)
1982 unsigned long ip = 0;
1985 if (!p || p == current)
1988 /* Only get wchan if task is blocked and we can keep it that way. */
1989 raw_spin_lock_irq(&p->pi_lock);
1990 state = READ_ONCE(p->__state);
1991 smp_rmb(); /* see try_to_wake_up() */
1992 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
1993 ip = __get_wchan(p);
1994 raw_spin_unlock_irq(&p->pi_lock);
1999 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2001 if (!(flags & ENQUEUE_NOCLOCK))
2002 update_rq_clock(rq);
2004 if (!(flags & ENQUEUE_RESTORE)) {
2005 sched_info_enqueue(rq, p);
2006 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2009 uclamp_rq_inc(rq, p);
2010 p->sched_class->enqueue_task(rq, p, flags);
2012 if (sched_core_enabled(rq))
2013 sched_core_enqueue(rq, p);
2016 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2018 if (sched_core_enabled(rq))
2019 sched_core_dequeue(rq, p, flags);
2021 if (!(flags & DEQUEUE_NOCLOCK))
2022 update_rq_clock(rq);
2024 if (!(flags & DEQUEUE_SAVE)) {
2025 sched_info_dequeue(rq, p);
2026 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2029 uclamp_rq_dec(rq, p);
2030 p->sched_class->dequeue_task(rq, p, flags);
2033 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2035 enqueue_task(rq, p, flags);
2037 p->on_rq = TASK_ON_RQ_QUEUED;
2040 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2042 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2044 dequeue_task(rq, p, flags);
2047 static inline int __normal_prio(int policy, int rt_prio, int nice)
2051 if (dl_policy(policy))
2052 prio = MAX_DL_PRIO - 1;
2053 else if (rt_policy(policy))
2054 prio = MAX_RT_PRIO - 1 - rt_prio;
2056 prio = NICE_TO_PRIO(nice);
2062 * Calculate the expected normal priority: i.e. priority
2063 * without taking RT-inheritance into account. Might be
2064 * boosted by interactivity modifiers. Changes upon fork,
2065 * setprio syscalls, and whenever the interactivity
2066 * estimator recalculates.
2068 static inline int normal_prio(struct task_struct *p)
2070 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2074 * Calculate the current priority, i.e. the priority
2075 * taken into account by the scheduler. This value might
2076 * be boosted by RT tasks, or might be boosted by
2077 * interactivity modifiers. Will be RT if the task got
2078 * RT-boosted. If not then it returns p->normal_prio.
2080 static int effective_prio(struct task_struct *p)
2082 p->normal_prio = normal_prio(p);
2084 * If we are RT tasks or we were boosted to RT priority,
2085 * keep the priority unchanged. Otherwise, update priority
2086 * to the normal priority:
2088 if (!rt_prio(p->prio))
2089 return p->normal_prio;
2094 * task_curr - is this task currently executing on a CPU?
2095 * @p: the task in question.
2097 * Return: 1 if the task is currently executing. 0 otherwise.
2099 inline int task_curr(const struct task_struct *p)
2101 return cpu_curr(task_cpu(p)) == p;
2105 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2106 * use the balance_callback list if you want balancing.
2108 * this means any call to check_class_changed() must be followed by a call to
2109 * balance_callback().
2111 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2112 const struct sched_class *prev_class,
2115 if (prev_class != p->sched_class) {
2116 if (prev_class->switched_from)
2117 prev_class->switched_from(rq, p);
2119 p->sched_class->switched_to(rq, p);
2120 } else if (oldprio != p->prio || dl_task(p))
2121 p->sched_class->prio_changed(rq, p, oldprio);
2124 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2126 if (p->sched_class == rq->curr->sched_class)
2127 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2128 else if (p->sched_class > rq->curr->sched_class)
2132 * A queue event has occurred, and we're going to schedule. In
2133 * this case, we can save a useless back to back clock update.
2135 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2136 rq_clock_skip_update(rq);
2142 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2144 static int __set_cpus_allowed_ptr(struct task_struct *p,
2145 const struct cpumask *new_mask,
2148 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2150 if (likely(!p->migration_disabled))
2153 if (p->cpus_ptr != &p->cpus_mask)
2157 * Violates locking rules! see comment in __do_set_cpus_allowed().
2159 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2162 void migrate_disable(void)
2164 struct task_struct *p = current;
2166 if (p->migration_disabled) {
2167 p->migration_disabled++;
2172 this_rq()->nr_pinned++;
2173 p->migration_disabled = 1;
2176 EXPORT_SYMBOL_GPL(migrate_disable);
2178 void migrate_enable(void)
2180 struct task_struct *p = current;
2182 if (p->migration_disabled > 1) {
2183 p->migration_disabled--;
2187 if (WARN_ON_ONCE(!p->migration_disabled))
2191 * Ensure stop_task runs either before or after this, and that
2192 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2195 if (p->cpus_ptr != &p->cpus_mask)
2196 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2198 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2199 * regular cpus_mask, otherwise things that race (eg.
2200 * select_fallback_rq) get confused.
2203 p->migration_disabled = 0;
2204 this_rq()->nr_pinned--;
2207 EXPORT_SYMBOL_GPL(migrate_enable);
2209 static inline bool rq_has_pinned_tasks(struct rq *rq)
2211 return rq->nr_pinned;
2215 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2216 * __set_cpus_allowed_ptr() and select_fallback_rq().
2218 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2220 /* When not in the task's cpumask, no point in looking further. */
2221 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2224 /* migrate_disabled() must be allowed to finish. */
2225 if (is_migration_disabled(p))
2226 return cpu_online(cpu);
2228 /* Non kernel threads are not allowed during either online or offline. */
2229 if (!(p->flags & PF_KTHREAD))
2230 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2232 /* KTHREAD_IS_PER_CPU is always allowed. */
2233 if (kthread_is_per_cpu(p))
2234 return cpu_online(cpu);
2236 /* Regular kernel threads don't get to stay during offline. */
2240 /* But are allowed during online. */
2241 return cpu_online(cpu);
2245 * This is how migration works:
2247 * 1) we invoke migration_cpu_stop() on the target CPU using
2249 * 2) stopper starts to run (implicitly forcing the migrated thread
2251 * 3) it checks whether the migrated task is still in the wrong runqueue.
2252 * 4) if it's in the wrong runqueue then the migration thread removes
2253 * it and puts it into the right queue.
2254 * 5) stopper completes and stop_one_cpu() returns and the migration
2259 * move_queued_task - move a queued task to new rq.
2261 * Returns (locked) new rq. Old rq's lock is released.
2263 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2264 struct task_struct *p, int new_cpu)
2266 lockdep_assert_rq_held(rq);
2268 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2269 set_task_cpu(p, new_cpu);
2272 rq = cpu_rq(new_cpu);
2275 BUG_ON(task_cpu(p) != new_cpu);
2276 activate_task(rq, p, 0);
2277 check_preempt_curr(rq, p, 0);
2282 struct migration_arg {
2283 struct task_struct *task;
2285 struct set_affinity_pending *pending;
2289 * @refs: number of wait_for_completion()
2290 * @stop_pending: is @stop_work in use
2292 struct set_affinity_pending {
2294 unsigned int stop_pending;
2295 struct completion done;
2296 struct cpu_stop_work stop_work;
2297 struct migration_arg arg;
2301 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2302 * this because either it can't run here any more (set_cpus_allowed()
2303 * away from this CPU, or CPU going down), or because we're
2304 * attempting to rebalance this task on exec (sched_exec).
2306 * So we race with normal scheduler movements, but that's OK, as long
2307 * as the task is no longer on this CPU.
2309 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2310 struct task_struct *p, int dest_cpu)
2312 /* Affinity changed (again). */
2313 if (!is_cpu_allowed(p, dest_cpu))
2316 update_rq_clock(rq);
2317 rq = move_queued_task(rq, rf, p, dest_cpu);
2323 * migration_cpu_stop - this will be executed by a highprio stopper thread
2324 * and performs thread migration by bumping thread off CPU then
2325 * 'pushing' onto another runqueue.
2327 static int migration_cpu_stop(void *data)
2329 struct migration_arg *arg = data;
2330 struct set_affinity_pending *pending = arg->pending;
2331 struct task_struct *p = arg->task;
2332 struct rq *rq = this_rq();
2333 bool complete = false;
2337 * The original target CPU might have gone down and we might
2338 * be on another CPU but it doesn't matter.
2340 local_irq_save(rf.flags);
2342 * We need to explicitly wake pending tasks before running
2343 * __migrate_task() such that we will not miss enforcing cpus_ptr
2344 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2346 flush_smp_call_function_from_idle();
2348 raw_spin_lock(&p->pi_lock);
2352 * If we were passed a pending, then ->stop_pending was set, thus
2353 * p->migration_pending must have remained stable.
2355 WARN_ON_ONCE(pending && pending != p->migration_pending);
2358 * If task_rq(p) != rq, it cannot be migrated here, because we're
2359 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2360 * we're holding p->pi_lock.
2362 if (task_rq(p) == rq) {
2363 if (is_migration_disabled(p))
2367 p->migration_pending = NULL;
2370 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2374 if (task_on_rq_queued(p))
2375 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2377 p->wake_cpu = arg->dest_cpu;
2380 * XXX __migrate_task() can fail, at which point we might end
2381 * up running on a dodgy CPU, AFAICT this can only happen
2382 * during CPU hotplug, at which point we'll get pushed out
2383 * anyway, so it's probably not a big deal.
2386 } else if (pending) {
2388 * This happens when we get migrated between migrate_enable()'s
2389 * preempt_enable() and scheduling the stopper task. At that
2390 * point we're a regular task again and not current anymore.
2392 * A !PREEMPT kernel has a giant hole here, which makes it far
2397 * The task moved before the stopper got to run. We're holding
2398 * ->pi_lock, so the allowed mask is stable - if it got
2399 * somewhere allowed, we're done.
2401 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2402 p->migration_pending = NULL;
2408 * When migrate_enable() hits a rq mis-match we can't reliably
2409 * determine is_migration_disabled() and so have to chase after
2412 WARN_ON_ONCE(!pending->stop_pending);
2413 task_rq_unlock(rq, p, &rf);
2414 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2415 &pending->arg, &pending->stop_work);
2420 pending->stop_pending = false;
2421 task_rq_unlock(rq, p, &rf);
2424 complete_all(&pending->done);
2429 int push_cpu_stop(void *arg)
2431 struct rq *lowest_rq = NULL, *rq = this_rq();
2432 struct task_struct *p = arg;
2434 raw_spin_lock_irq(&p->pi_lock);
2435 raw_spin_rq_lock(rq);
2437 if (task_rq(p) != rq)
2440 if (is_migration_disabled(p)) {
2441 p->migration_flags |= MDF_PUSH;
2445 p->migration_flags &= ~MDF_PUSH;
2447 if (p->sched_class->find_lock_rq)
2448 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2453 // XXX validate p is still the highest prio task
2454 if (task_rq(p) == rq) {
2455 deactivate_task(rq, p, 0);
2456 set_task_cpu(p, lowest_rq->cpu);
2457 activate_task(lowest_rq, p, 0);
2458 resched_curr(lowest_rq);
2461 double_unlock_balance(rq, lowest_rq);
2464 rq->push_busy = false;
2465 raw_spin_rq_unlock(rq);
2466 raw_spin_unlock_irq(&p->pi_lock);
2473 * sched_class::set_cpus_allowed must do the below, but is not required to
2474 * actually call this function.
2476 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2478 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2479 p->cpus_ptr = new_mask;
2483 cpumask_copy(&p->cpus_mask, new_mask);
2484 p->nr_cpus_allowed = cpumask_weight(new_mask);
2488 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2490 struct rq *rq = task_rq(p);
2491 bool queued, running;
2494 * This here violates the locking rules for affinity, since we're only
2495 * supposed to change these variables while holding both rq->lock and
2498 * HOWEVER, it magically works, because ttwu() is the only code that
2499 * accesses these variables under p->pi_lock and only does so after
2500 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2501 * before finish_task().
2503 * XXX do further audits, this smells like something putrid.
2505 if (flags & SCA_MIGRATE_DISABLE)
2506 SCHED_WARN_ON(!p->on_cpu);
2508 lockdep_assert_held(&p->pi_lock);
2510 queued = task_on_rq_queued(p);
2511 running = task_current(rq, p);
2515 * Because __kthread_bind() calls this on blocked tasks without
2518 lockdep_assert_rq_held(rq);
2519 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2522 put_prev_task(rq, p);
2524 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2527 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2529 set_next_task(rq, p);
2532 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2534 __do_set_cpus_allowed(p, new_mask, 0);
2537 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2540 if (!src->user_cpus_ptr)
2543 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2544 if (!dst->user_cpus_ptr)
2547 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2551 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2553 struct cpumask *user_mask = NULL;
2555 swap(p->user_cpus_ptr, user_mask);
2560 void release_user_cpus_ptr(struct task_struct *p)
2562 kfree(clear_user_cpus_ptr(p));
2566 * This function is wildly self concurrent; here be dragons.
2569 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2570 * designated task is enqueued on an allowed CPU. If that task is currently
2571 * running, we have to kick it out using the CPU stopper.
2573 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2576 * Initial conditions: P0->cpus_mask = [0, 1]
2580 * migrate_disable();
2582 * set_cpus_allowed_ptr(P0, [1]);
2584 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2585 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2586 * This means we need the following scheme:
2590 * migrate_disable();
2592 * set_cpus_allowed_ptr(P0, [1]);
2596 * __set_cpus_allowed_ptr();
2597 * <wakes local stopper>
2598 * `--> <woken on migration completion>
2600 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2601 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2602 * task p are serialized by p->pi_lock, which we can leverage: the one that
2603 * should come into effect at the end of the Migrate-Disable region is the last
2604 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2605 * but we still need to properly signal those waiting tasks at the appropriate
2608 * This is implemented using struct set_affinity_pending. The first
2609 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2610 * setup an instance of that struct and install it on the targeted task_struct.
2611 * Any and all further callers will reuse that instance. Those then wait for
2612 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2613 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2616 * (1) In the cases covered above. There is one more where the completion is
2617 * signaled within affine_move_task() itself: when a subsequent affinity request
2618 * occurs after the stopper bailed out due to the targeted task still being
2619 * Migrate-Disable. Consider:
2621 * Initial conditions: P0->cpus_mask = [0, 1]
2625 * migrate_disable();
2627 * set_cpus_allowed_ptr(P0, [1]);
2630 * migration_cpu_stop()
2631 * is_migration_disabled()
2633 * set_cpus_allowed_ptr(P0, [0, 1]);
2634 * <signal completion>
2637 * Note that the above is safe vs a concurrent migrate_enable(), as any
2638 * pending affinity completion is preceded by an uninstallation of
2639 * p->migration_pending done with p->pi_lock held.
2641 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2642 int dest_cpu, unsigned int flags)
2644 struct set_affinity_pending my_pending = { }, *pending = NULL;
2645 bool stop_pending, complete = false;
2647 /* Can the task run on the task's current CPU? If so, we're done */
2648 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2649 struct task_struct *push_task = NULL;
2651 if ((flags & SCA_MIGRATE_ENABLE) &&
2652 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2653 rq->push_busy = true;
2654 push_task = get_task_struct(p);
2658 * If there are pending waiters, but no pending stop_work,
2659 * then complete now.
2661 pending = p->migration_pending;
2662 if (pending && !pending->stop_pending) {
2663 p->migration_pending = NULL;
2667 task_rq_unlock(rq, p, rf);
2670 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2675 complete_all(&pending->done);
2680 if (!(flags & SCA_MIGRATE_ENABLE)) {
2681 /* serialized by p->pi_lock */
2682 if (!p->migration_pending) {
2683 /* Install the request */
2684 refcount_set(&my_pending.refs, 1);
2685 init_completion(&my_pending.done);
2686 my_pending.arg = (struct migration_arg) {
2688 .dest_cpu = dest_cpu,
2689 .pending = &my_pending,
2692 p->migration_pending = &my_pending;
2694 pending = p->migration_pending;
2695 refcount_inc(&pending->refs);
2697 * Affinity has changed, but we've already installed a
2698 * pending. migration_cpu_stop() *must* see this, else
2699 * we risk a completion of the pending despite having a
2700 * task on a disallowed CPU.
2702 * Serialized by p->pi_lock, so this is safe.
2704 pending->arg.dest_cpu = dest_cpu;
2707 pending = p->migration_pending;
2709 * - !MIGRATE_ENABLE:
2710 * we'll have installed a pending if there wasn't one already.
2713 * we're here because the current CPU isn't matching anymore,
2714 * the only way that can happen is because of a concurrent
2715 * set_cpus_allowed_ptr() call, which should then still be
2716 * pending completion.
2718 * Either way, we really should have a @pending here.
2720 if (WARN_ON_ONCE(!pending)) {
2721 task_rq_unlock(rq, p, rf);
2725 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2727 * MIGRATE_ENABLE gets here because 'p == current', but for
2728 * anything else we cannot do is_migration_disabled(), punt
2729 * and have the stopper function handle it all race-free.
2731 stop_pending = pending->stop_pending;
2733 pending->stop_pending = true;
2735 if (flags & SCA_MIGRATE_ENABLE)
2736 p->migration_flags &= ~MDF_PUSH;
2738 task_rq_unlock(rq, p, rf);
2740 if (!stop_pending) {
2741 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2742 &pending->arg, &pending->stop_work);
2745 if (flags & SCA_MIGRATE_ENABLE)
2749 if (!is_migration_disabled(p)) {
2750 if (task_on_rq_queued(p))
2751 rq = move_queued_task(rq, rf, p, dest_cpu);
2753 if (!pending->stop_pending) {
2754 p->migration_pending = NULL;
2758 task_rq_unlock(rq, p, rf);
2761 complete_all(&pending->done);
2764 wait_for_completion(&pending->done);
2766 if (refcount_dec_and_test(&pending->refs))
2767 wake_up_var(&pending->refs); /* No UaF, just an address */
2770 * Block the original owner of &pending until all subsequent callers
2771 * have seen the completion and decremented the refcount
2773 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2776 WARN_ON_ONCE(my_pending.stop_pending);
2782 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2784 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2785 const struct cpumask *new_mask,
2788 struct rq_flags *rf)
2789 __releases(rq->lock)
2790 __releases(p->pi_lock)
2792 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2793 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2794 bool kthread = p->flags & PF_KTHREAD;
2795 struct cpumask *user_mask = NULL;
2796 unsigned int dest_cpu;
2799 update_rq_clock(rq);
2801 if (kthread || is_migration_disabled(p)) {
2803 * Kernel threads are allowed on online && !active CPUs,
2804 * however, during cpu-hot-unplug, even these might get pushed
2805 * away if not KTHREAD_IS_PER_CPU.
2807 * Specifically, migration_disabled() tasks must not fail the
2808 * cpumask_any_and_distribute() pick below, esp. so on
2809 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2810 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2812 cpu_valid_mask = cpu_online_mask;
2815 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2821 * Must re-check here, to close a race against __kthread_bind(),
2822 * sched_setaffinity() is not guaranteed to observe the flag.
2824 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2829 if (!(flags & SCA_MIGRATE_ENABLE)) {
2830 if (cpumask_equal(&p->cpus_mask, new_mask))
2833 if (WARN_ON_ONCE(p == current &&
2834 is_migration_disabled(p) &&
2835 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2842 * Picking a ~random cpu helps in cases where we are changing affinity
2843 * for groups of tasks (ie. cpuset), so that load balancing is not
2844 * immediately required to distribute the tasks within their new mask.
2846 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2847 if (dest_cpu >= nr_cpu_ids) {
2852 __do_set_cpus_allowed(p, new_mask, flags);
2854 if (flags & SCA_USER)
2855 user_mask = clear_user_cpus_ptr(p);
2857 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2864 task_rq_unlock(rq, p, rf);
2870 * Change a given task's CPU affinity. Migrate the thread to a
2871 * proper CPU and schedule it away if the CPU it's executing on
2872 * is removed from the allowed bitmask.
2874 * NOTE: the caller must have a valid reference to the task, the
2875 * task must not exit() & deallocate itself prematurely. The
2876 * call is not atomic; no spinlocks may be held.
2878 static int __set_cpus_allowed_ptr(struct task_struct *p,
2879 const struct cpumask *new_mask, u32 flags)
2884 rq = task_rq_lock(p, &rf);
2885 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2888 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2890 return __set_cpus_allowed_ptr(p, new_mask, 0);
2892 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2895 * Change a given task's CPU affinity to the intersection of its current
2896 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2897 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2898 * If the resulting mask is empty, leave the affinity unchanged and return
2901 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2902 struct cpumask *new_mask,
2903 const struct cpumask *subset_mask)
2905 struct cpumask *user_mask = NULL;
2910 if (!p->user_cpus_ptr) {
2911 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2916 rq = task_rq_lock(p, &rf);
2919 * Forcefully restricting the affinity of a deadline task is
2920 * likely to cause problems, so fail and noisily override the
2923 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2928 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2934 * We're about to butcher the task affinity, so keep track of what
2935 * the user asked for in case we're able to restore it later on.
2938 cpumask_copy(user_mask, p->cpus_ptr);
2939 p->user_cpus_ptr = user_mask;
2942 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2945 task_rq_unlock(rq, p, &rf);
2951 * Restrict the CPU affinity of task @p so that it is a subset of
2952 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2953 * old affinity mask. If the resulting mask is empty, we warn and walk
2954 * up the cpuset hierarchy until we find a suitable mask.
2956 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2958 cpumask_var_t new_mask;
2959 const struct cpumask *override_mask = task_cpu_possible_mask(p);
2961 alloc_cpumask_var(&new_mask, GFP_KERNEL);
2964 * __migrate_task() can fail silently in the face of concurrent
2965 * offlining of the chosen destination CPU, so take the hotplug
2966 * lock to ensure that the migration succeeds.
2969 if (!cpumask_available(new_mask))
2972 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2976 * We failed to find a valid subset of the affinity mask for the
2977 * task, so override it based on its cpuset hierarchy.
2979 cpuset_cpus_allowed(p, new_mask);
2980 override_mask = new_mask;
2983 if (printk_ratelimit()) {
2984 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2985 task_pid_nr(p), p->comm,
2986 cpumask_pr_args(override_mask));
2989 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2992 free_cpumask_var(new_mask);
2996 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
2999 * Restore the affinity of a task @p which was previously restricted by a
3000 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3001 * @p->user_cpus_ptr.
3003 * It is the caller's responsibility to serialise this with any calls to
3004 * force_compatible_cpus_allowed_ptr(@p).
3006 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3008 struct cpumask *user_mask = p->user_cpus_ptr;
3009 unsigned long flags;
3012 * Try to restore the old affinity mask. If this fails, then
3013 * we free the mask explicitly to avoid it being inherited across
3014 * a subsequent fork().
3016 if (!user_mask || !__sched_setaffinity(p, user_mask))
3019 raw_spin_lock_irqsave(&p->pi_lock, flags);
3020 user_mask = clear_user_cpus_ptr(p);
3021 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3026 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3028 #ifdef CONFIG_SCHED_DEBUG
3029 unsigned int state = READ_ONCE(p->__state);
3032 * We should never call set_task_cpu() on a blocked task,
3033 * ttwu() will sort out the placement.
3035 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3038 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3039 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3040 * time relying on p->on_rq.
3042 WARN_ON_ONCE(state == TASK_RUNNING &&
3043 p->sched_class == &fair_sched_class &&
3044 (p->on_rq && !task_on_rq_migrating(p)));
3046 #ifdef CONFIG_LOCKDEP
3048 * The caller should hold either p->pi_lock or rq->lock, when changing
3049 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3051 * sched_move_task() holds both and thus holding either pins the cgroup,
3054 * Furthermore, all task_rq users should acquire both locks, see
3057 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3058 lockdep_is_held(__rq_lockp(task_rq(p)))));
3061 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3063 WARN_ON_ONCE(!cpu_online(new_cpu));
3065 WARN_ON_ONCE(is_migration_disabled(p));
3068 trace_sched_migrate_task(p, new_cpu);
3070 if (task_cpu(p) != new_cpu) {
3071 if (p->sched_class->migrate_task_rq)
3072 p->sched_class->migrate_task_rq(p, new_cpu);
3073 p->se.nr_migrations++;
3075 perf_event_task_migrate(p);
3078 __set_task_cpu(p, new_cpu);
3081 #ifdef CONFIG_NUMA_BALANCING
3082 static void __migrate_swap_task(struct task_struct *p, int cpu)
3084 if (task_on_rq_queued(p)) {
3085 struct rq *src_rq, *dst_rq;
3086 struct rq_flags srf, drf;
3088 src_rq = task_rq(p);
3089 dst_rq = cpu_rq(cpu);
3091 rq_pin_lock(src_rq, &srf);
3092 rq_pin_lock(dst_rq, &drf);
3094 deactivate_task(src_rq, p, 0);
3095 set_task_cpu(p, cpu);
3096 activate_task(dst_rq, p, 0);
3097 check_preempt_curr(dst_rq, p, 0);
3099 rq_unpin_lock(dst_rq, &drf);
3100 rq_unpin_lock(src_rq, &srf);
3104 * Task isn't running anymore; make it appear like we migrated
3105 * it before it went to sleep. This means on wakeup we make the
3106 * previous CPU our target instead of where it really is.
3112 struct migration_swap_arg {
3113 struct task_struct *src_task, *dst_task;
3114 int src_cpu, dst_cpu;
3117 static int migrate_swap_stop(void *data)
3119 struct migration_swap_arg *arg = data;
3120 struct rq *src_rq, *dst_rq;
3123 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3126 src_rq = cpu_rq(arg->src_cpu);
3127 dst_rq = cpu_rq(arg->dst_cpu);
3129 double_raw_lock(&arg->src_task->pi_lock,
3130 &arg->dst_task->pi_lock);
3131 double_rq_lock(src_rq, dst_rq);
3133 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3136 if (task_cpu(arg->src_task) != arg->src_cpu)
3139 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3142 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3145 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3146 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3151 double_rq_unlock(src_rq, dst_rq);
3152 raw_spin_unlock(&arg->dst_task->pi_lock);
3153 raw_spin_unlock(&arg->src_task->pi_lock);
3159 * Cross migrate two tasks
3161 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3162 int target_cpu, int curr_cpu)
3164 struct migration_swap_arg arg;
3167 arg = (struct migration_swap_arg){
3169 .src_cpu = curr_cpu,
3171 .dst_cpu = target_cpu,
3174 if (arg.src_cpu == arg.dst_cpu)
3178 * These three tests are all lockless; this is OK since all of them
3179 * will be re-checked with proper locks held further down the line.
3181 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3184 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3187 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3190 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3191 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3196 #endif /* CONFIG_NUMA_BALANCING */
3199 * wait_task_inactive - wait for a thread to unschedule.
3201 * If @match_state is nonzero, it's the @p->state value just checked and
3202 * not expected to change. If it changes, i.e. @p might have woken up,
3203 * then return zero. When we succeed in waiting for @p to be off its CPU,
3204 * we return a positive number (its total switch count). If a second call
3205 * a short while later returns the same number, the caller can be sure that
3206 * @p has remained unscheduled the whole time.
3208 * The caller must ensure that the task *will* unschedule sometime soon,
3209 * else this function might spin for a *long* time. This function can't
3210 * be called with interrupts off, or it may introduce deadlock with
3211 * smp_call_function() if an IPI is sent by the same process we are
3212 * waiting to become inactive.
3214 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3216 int running, queued;
3223 * We do the initial early heuristics without holding
3224 * any task-queue locks at all. We'll only try to get
3225 * the runqueue lock when things look like they will
3231 * If the task is actively running on another CPU
3232 * still, just relax and busy-wait without holding
3235 * NOTE! Since we don't hold any locks, it's not
3236 * even sure that "rq" stays as the right runqueue!
3237 * But we don't care, since "task_running()" will
3238 * return false if the runqueue has changed and p
3239 * is actually now running somewhere else!
3241 while (task_running(rq, p)) {
3242 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3248 * Ok, time to look more closely! We need the rq
3249 * lock now, to be *sure*. If we're wrong, we'll
3250 * just go back and repeat.
3252 rq = task_rq_lock(p, &rf);
3253 trace_sched_wait_task(p);
3254 running = task_running(rq, p);
3255 queued = task_on_rq_queued(p);
3257 if (!match_state || READ_ONCE(p->__state) == match_state)
3258 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3259 task_rq_unlock(rq, p, &rf);
3262 * If it changed from the expected state, bail out now.
3264 if (unlikely(!ncsw))
3268 * Was it really running after all now that we
3269 * checked with the proper locks actually held?
3271 * Oops. Go back and try again..
3273 if (unlikely(running)) {
3279 * It's not enough that it's not actively running,
3280 * it must be off the runqueue _entirely_, and not
3283 * So if it was still runnable (but just not actively
3284 * running right now), it's preempted, and we should
3285 * yield - it could be a while.
3287 if (unlikely(queued)) {
3288 ktime_t to = NSEC_PER_SEC / HZ;
3290 set_current_state(TASK_UNINTERRUPTIBLE);
3291 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3296 * Ahh, all good. It wasn't running, and it wasn't
3297 * runnable, which means that it will never become
3298 * running in the future either. We're all done!
3307 * kick_process - kick a running thread to enter/exit the kernel
3308 * @p: the to-be-kicked thread
3310 * Cause a process which is running on another CPU to enter
3311 * kernel-mode, without any delay. (to get signals handled.)
3313 * NOTE: this function doesn't have to take the runqueue lock,
3314 * because all it wants to ensure is that the remote task enters
3315 * the kernel. If the IPI races and the task has been migrated
3316 * to another CPU then no harm is done and the purpose has been
3319 void kick_process(struct task_struct *p)
3325 if ((cpu != smp_processor_id()) && task_curr(p))
3326 smp_send_reschedule(cpu);
3329 EXPORT_SYMBOL_GPL(kick_process);
3332 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3334 * A few notes on cpu_active vs cpu_online:
3336 * - cpu_active must be a subset of cpu_online
3338 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3339 * see __set_cpus_allowed_ptr(). At this point the newly online
3340 * CPU isn't yet part of the sched domains, and balancing will not
3343 * - on CPU-down we clear cpu_active() to mask the sched domains and
3344 * avoid the load balancer to place new tasks on the to be removed
3345 * CPU. Existing tasks will remain running there and will be taken
3348 * This means that fallback selection must not select !active CPUs.
3349 * And can assume that any active CPU must be online. Conversely
3350 * select_task_rq() below may allow selection of !active CPUs in order
3351 * to satisfy the above rules.
3353 static int select_fallback_rq(int cpu, struct task_struct *p)
3355 int nid = cpu_to_node(cpu);
3356 const struct cpumask *nodemask = NULL;
3357 enum { cpuset, possible, fail } state = cpuset;
3361 * If the node that the CPU is on has been offlined, cpu_to_node()
3362 * will return -1. There is no CPU on the node, and we should
3363 * select the CPU on the other node.
3366 nodemask = cpumask_of_node(nid);
3368 /* Look for allowed, online CPU in same node. */
3369 for_each_cpu(dest_cpu, nodemask) {
3370 if (is_cpu_allowed(p, dest_cpu))
3376 /* Any allowed, online CPU? */
3377 for_each_cpu(dest_cpu, p->cpus_ptr) {
3378 if (!is_cpu_allowed(p, dest_cpu))
3384 /* No more Mr. Nice Guy. */
3387 if (cpuset_cpus_allowed_fallback(p)) {
3394 * XXX When called from select_task_rq() we only
3395 * hold p->pi_lock and again violate locking order.
3397 * More yuck to audit.
3399 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3409 if (state != cpuset) {
3411 * Don't tell them about moving exiting tasks or
3412 * kernel threads (both mm NULL), since they never
3415 if (p->mm && printk_ratelimit()) {
3416 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3417 task_pid_nr(p), p->comm, cpu);
3425 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3428 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3430 lockdep_assert_held(&p->pi_lock);
3432 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3433 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3435 cpu = cpumask_any(p->cpus_ptr);
3438 * In order not to call set_task_cpu() on a blocking task we need
3439 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3442 * Since this is common to all placement strategies, this lives here.
3444 * [ this allows ->select_task() to simply return task_cpu(p) and
3445 * not worry about this generic constraint ]
3447 if (unlikely(!is_cpu_allowed(p, cpu)))
3448 cpu = select_fallback_rq(task_cpu(p), p);
3453 void sched_set_stop_task(int cpu, struct task_struct *stop)
3455 static struct lock_class_key stop_pi_lock;
3456 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3457 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3461 * Make it appear like a SCHED_FIFO task, its something
3462 * userspace knows about and won't get confused about.
3464 * Also, it will make PI more or less work without too
3465 * much confusion -- but then, stop work should not
3466 * rely on PI working anyway.
3468 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3470 stop->sched_class = &stop_sched_class;
3473 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3474 * adjust the effective priority of a task. As a result,
3475 * rt_mutex_setprio() can trigger (RT) balancing operations,
3476 * which can then trigger wakeups of the stop thread to push
3477 * around the current task.
3479 * The stop task itself will never be part of the PI-chain, it
3480 * never blocks, therefore that ->pi_lock recursion is safe.
3481 * Tell lockdep about this by placing the stop->pi_lock in its
3484 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3487 cpu_rq(cpu)->stop = stop;
3491 * Reset it back to a normal scheduling class so that
3492 * it can die in pieces.
3494 old_stop->sched_class = &rt_sched_class;
3498 #else /* CONFIG_SMP */
3500 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3501 const struct cpumask *new_mask,
3504 return set_cpus_allowed_ptr(p, new_mask);
3507 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3509 static inline bool rq_has_pinned_tasks(struct rq *rq)
3514 #endif /* !CONFIG_SMP */
3517 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3521 if (!schedstat_enabled())
3527 if (cpu == rq->cpu) {
3528 __schedstat_inc(rq->ttwu_local);
3529 __schedstat_inc(p->stats.nr_wakeups_local);
3531 struct sched_domain *sd;
3533 __schedstat_inc(p->stats.nr_wakeups_remote);
3535 for_each_domain(rq->cpu, sd) {
3536 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3537 __schedstat_inc(sd->ttwu_wake_remote);
3544 if (wake_flags & WF_MIGRATED)
3545 __schedstat_inc(p->stats.nr_wakeups_migrate);
3546 #endif /* CONFIG_SMP */
3548 __schedstat_inc(rq->ttwu_count);
3549 __schedstat_inc(p->stats.nr_wakeups);
3551 if (wake_flags & WF_SYNC)
3552 __schedstat_inc(p->stats.nr_wakeups_sync);
3556 * Mark the task runnable and perform wakeup-preemption.
3558 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3559 struct rq_flags *rf)
3561 check_preempt_curr(rq, p, wake_flags);
3562 WRITE_ONCE(p->__state, TASK_RUNNING);
3563 trace_sched_wakeup(p);
3566 if (p->sched_class->task_woken) {
3568 * Our task @p is fully woken up and running; so it's safe to
3569 * drop the rq->lock, hereafter rq is only used for statistics.
3571 rq_unpin_lock(rq, rf);
3572 p->sched_class->task_woken(rq, p);
3573 rq_repin_lock(rq, rf);
3576 if (rq->idle_stamp) {
3577 u64 delta = rq_clock(rq) - rq->idle_stamp;
3578 u64 max = 2*rq->max_idle_balance_cost;
3580 update_avg(&rq->avg_idle, delta);
3582 if (rq->avg_idle > max)
3585 rq->wake_stamp = jiffies;
3586 rq->wake_avg_idle = rq->avg_idle / 2;
3594 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3595 struct rq_flags *rf)
3597 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3599 lockdep_assert_rq_held(rq);
3601 if (p->sched_contributes_to_load)
3602 rq->nr_uninterruptible--;
3605 if (wake_flags & WF_MIGRATED)
3606 en_flags |= ENQUEUE_MIGRATED;
3610 delayacct_blkio_end(p);
3611 atomic_dec(&task_rq(p)->nr_iowait);
3614 activate_task(rq, p, en_flags);
3615 ttwu_do_wakeup(rq, p, wake_flags, rf);
3619 * Consider @p being inside a wait loop:
3622 * set_current_state(TASK_UNINTERRUPTIBLE);
3629 * __set_current_state(TASK_RUNNING);
3631 * between set_current_state() and schedule(). In this case @p is still
3632 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3635 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3636 * then schedule() must still happen and p->state can be changed to
3637 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3638 * need to do a full wakeup with enqueue.
3640 * Returns: %true when the wakeup is done,
3643 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3649 rq = __task_rq_lock(p, &rf);
3650 if (task_on_rq_queued(p)) {
3651 /* check_preempt_curr() may use rq clock */
3652 update_rq_clock(rq);
3653 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3656 __task_rq_unlock(rq, &rf);
3662 void sched_ttwu_pending(void *arg)
3664 struct llist_node *llist = arg;
3665 struct rq *rq = this_rq();
3666 struct task_struct *p, *t;
3673 * rq::ttwu_pending racy indication of out-standing wakeups.
3674 * Races such that false-negatives are possible, since they
3675 * are shorter lived that false-positives would be.
3677 WRITE_ONCE(rq->ttwu_pending, 0);
3679 rq_lock_irqsave(rq, &rf);
3680 update_rq_clock(rq);
3682 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3683 if (WARN_ON_ONCE(p->on_cpu))
3684 smp_cond_load_acquire(&p->on_cpu, !VAL);
3686 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3687 set_task_cpu(p, cpu_of(rq));
3689 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3692 rq_unlock_irqrestore(rq, &rf);
3695 void send_call_function_single_ipi(int cpu)
3697 struct rq *rq = cpu_rq(cpu);
3699 if (!set_nr_if_polling(rq->idle))
3700 arch_send_call_function_single_ipi(cpu);
3702 trace_sched_wake_idle_without_ipi(cpu);
3706 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3707 * necessary. The wakee CPU on receipt of the IPI will queue the task
3708 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3709 * of the wakeup instead of the waker.
3711 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3713 struct rq *rq = cpu_rq(cpu);
3715 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3717 WRITE_ONCE(rq->ttwu_pending, 1);
3718 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3721 void wake_up_if_idle(int cpu)
3723 struct rq *rq = cpu_rq(cpu);
3728 if (!is_idle_task(rcu_dereference(rq->curr)))
3731 rq_lock_irqsave(rq, &rf);
3732 if (is_idle_task(rq->curr))
3734 /* Else CPU is not idle, do nothing here: */
3735 rq_unlock_irqrestore(rq, &rf);
3741 bool cpus_share_cache(int this_cpu, int that_cpu)
3743 if (this_cpu == that_cpu)
3746 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3749 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3752 * Do not complicate things with the async wake_list while the CPU is
3755 if (!cpu_active(cpu))
3759 * If the CPU does not share cache, then queue the task on the
3760 * remote rqs wakelist to avoid accessing remote data.
3762 if (!cpus_share_cache(smp_processor_id(), cpu))
3766 * If the task is descheduling and the only running task on the
3767 * CPU then use the wakelist to offload the task activation to
3768 * the soon-to-be-idle CPU as the current CPU is likely busy.
3769 * nr_running is checked to avoid unnecessary task stacking.
3771 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3777 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3779 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3780 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3783 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3784 __ttwu_queue_wakelist(p, cpu, wake_flags);
3791 #else /* !CONFIG_SMP */
3793 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3798 #endif /* CONFIG_SMP */
3800 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3802 struct rq *rq = cpu_rq(cpu);
3805 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3809 update_rq_clock(rq);
3810 ttwu_do_activate(rq, p, wake_flags, &rf);
3815 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3817 * The caller holds p::pi_lock if p != current or has preemption
3818 * disabled when p == current.
3820 * The rules of PREEMPT_RT saved_state:
3822 * The related locking code always holds p::pi_lock when updating
3823 * p::saved_state, which means the code is fully serialized in both cases.
3825 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3826 * bits set. This allows to distinguish all wakeup scenarios.
3828 static __always_inline
3829 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3831 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3832 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3833 state != TASK_RTLOCK_WAIT);
3836 if (READ_ONCE(p->__state) & state) {
3841 #ifdef CONFIG_PREEMPT_RT
3843 * Saved state preserves the task state across blocking on
3844 * an RT lock. If the state matches, set p::saved_state to
3845 * TASK_RUNNING, but do not wake the task because it waits
3846 * for a lock wakeup. Also indicate success because from
3847 * the regular waker's point of view this has succeeded.
3849 * After acquiring the lock the task will restore p::__state
3850 * from p::saved_state which ensures that the regular
3851 * wakeup is not lost. The restore will also set
3852 * p::saved_state to TASK_RUNNING so any further tests will
3853 * not result in false positives vs. @success
3855 if (p->saved_state & state) {
3856 p->saved_state = TASK_RUNNING;
3864 * Notes on Program-Order guarantees on SMP systems.
3868 * The basic program-order guarantee on SMP systems is that when a task [t]
3869 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3870 * execution on its new CPU [c1].
3872 * For migration (of runnable tasks) this is provided by the following means:
3874 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3875 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3876 * rq(c1)->lock (if not at the same time, then in that order).
3877 * C) LOCK of the rq(c1)->lock scheduling in task
3879 * Release/acquire chaining guarantees that B happens after A and C after B.
3880 * Note: the CPU doing B need not be c0 or c1
3889 * UNLOCK rq(0)->lock
3891 * LOCK rq(0)->lock // orders against CPU0
3893 * UNLOCK rq(0)->lock
3897 * UNLOCK rq(1)->lock
3899 * LOCK rq(1)->lock // orders against CPU2
3902 * UNLOCK rq(1)->lock
3905 * BLOCKING -- aka. SLEEP + WAKEUP
3907 * For blocking we (obviously) need to provide the same guarantee as for
3908 * migration. However the means are completely different as there is no lock
3909 * chain to provide order. Instead we do:
3911 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3912 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3916 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3918 * LOCK rq(0)->lock LOCK X->pi_lock
3921 * smp_store_release(X->on_cpu, 0);
3923 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3929 * X->state = RUNNING
3930 * UNLOCK rq(2)->lock
3932 * LOCK rq(2)->lock // orders against CPU1
3935 * UNLOCK rq(2)->lock
3938 * UNLOCK rq(0)->lock
3941 * However, for wakeups there is a second guarantee we must provide, namely we
3942 * must ensure that CONDITION=1 done by the caller can not be reordered with
3943 * accesses to the task state; see try_to_wake_up() and set_current_state().
3947 * try_to_wake_up - wake up a thread
3948 * @p: the thread to be awakened
3949 * @state: the mask of task states that can be woken
3950 * @wake_flags: wake modifier flags (WF_*)
3952 * Conceptually does:
3954 * If (@state & @p->state) @p->state = TASK_RUNNING.
3956 * If the task was not queued/runnable, also place it back on a runqueue.
3958 * This function is atomic against schedule() which would dequeue the task.
3960 * It issues a full memory barrier before accessing @p->state, see the comment
3961 * with set_current_state().
3963 * Uses p->pi_lock to serialize against concurrent wake-ups.
3965 * Relies on p->pi_lock stabilizing:
3968 * - p->sched_task_group
3969 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3971 * Tries really hard to only take one task_rq(p)->lock for performance.
3972 * Takes rq->lock in:
3973 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3974 * - ttwu_queue() -- new rq, for enqueue of the task;
3975 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3977 * As a consequence we race really badly with just about everything. See the
3978 * many memory barriers and their comments for details.
3980 * Return: %true if @p->state changes (an actual wakeup was done),
3984 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3986 unsigned long flags;
3987 int cpu, success = 0;
3992 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3993 * == smp_processor_id()'. Together this means we can special
3994 * case the whole 'p->on_rq && ttwu_runnable()' case below
3995 * without taking any locks.
3998 * - we rely on Program-Order guarantees for all the ordering,
3999 * - we're serialized against set_special_state() by virtue of
4000 * it disabling IRQs (this allows not taking ->pi_lock).
4002 if (!ttwu_state_match(p, state, &success))
4005 trace_sched_waking(p);
4006 WRITE_ONCE(p->__state, TASK_RUNNING);
4007 trace_sched_wakeup(p);
4012 * If we are going to wake up a thread waiting for CONDITION we
4013 * need to ensure that CONDITION=1 done by the caller can not be
4014 * reordered with p->state check below. This pairs with smp_store_mb()
4015 * in set_current_state() that the waiting thread does.
4017 raw_spin_lock_irqsave(&p->pi_lock, flags);
4018 smp_mb__after_spinlock();
4019 if (!ttwu_state_match(p, state, &success))
4022 trace_sched_waking(p);
4025 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4026 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4027 * in smp_cond_load_acquire() below.
4029 * sched_ttwu_pending() try_to_wake_up()
4030 * STORE p->on_rq = 1 LOAD p->state
4033 * __schedule() (switch to task 'p')
4034 * LOCK rq->lock smp_rmb();
4035 * smp_mb__after_spinlock();
4039 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4041 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4042 * __schedule(). See the comment for smp_mb__after_spinlock().
4044 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4047 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4052 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4053 * possible to, falsely, observe p->on_cpu == 0.
4055 * One must be running (->on_cpu == 1) in order to remove oneself
4056 * from the runqueue.
4058 * __schedule() (switch to task 'p') try_to_wake_up()
4059 * STORE p->on_cpu = 1 LOAD p->on_rq
4062 * __schedule() (put 'p' to sleep)
4063 * LOCK rq->lock smp_rmb();
4064 * smp_mb__after_spinlock();
4065 * STORE p->on_rq = 0 LOAD p->on_cpu
4067 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4068 * __schedule(). See the comment for smp_mb__after_spinlock().
4070 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4071 * schedule()'s deactivate_task() has 'happened' and p will no longer
4072 * care about it's own p->state. See the comment in __schedule().
4074 smp_acquire__after_ctrl_dep();
4077 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4078 * == 0), which means we need to do an enqueue, change p->state to
4079 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4080 * enqueue, such as ttwu_queue_wakelist().
4082 WRITE_ONCE(p->__state, TASK_WAKING);
4085 * If the owning (remote) CPU is still in the middle of schedule() with
4086 * this task as prev, considering queueing p on the remote CPUs wake_list
4087 * which potentially sends an IPI instead of spinning on p->on_cpu to
4088 * let the waker make forward progress. This is safe because IRQs are
4089 * disabled and the IPI will deliver after on_cpu is cleared.
4091 * Ensure we load task_cpu(p) after p->on_cpu:
4093 * set_task_cpu(p, cpu);
4094 * STORE p->cpu = @cpu
4095 * __schedule() (switch to task 'p')
4097 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4098 * STORE p->on_cpu = 1 LOAD p->cpu
4100 * to ensure we observe the correct CPU on which the task is currently
4103 if (smp_load_acquire(&p->on_cpu) &&
4104 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4108 * If the owning (remote) CPU is still in the middle of schedule() with
4109 * this task as prev, wait until it's done referencing the task.
4111 * Pairs with the smp_store_release() in finish_task().
4113 * This ensures that tasks getting woken will be fully ordered against
4114 * their previous state and preserve Program Order.
4116 smp_cond_load_acquire(&p->on_cpu, !VAL);
4118 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4119 if (task_cpu(p) != cpu) {
4121 delayacct_blkio_end(p);
4122 atomic_dec(&task_rq(p)->nr_iowait);
4125 wake_flags |= WF_MIGRATED;
4126 psi_ttwu_dequeue(p);
4127 set_task_cpu(p, cpu);
4131 #endif /* CONFIG_SMP */
4133 ttwu_queue(p, cpu, wake_flags);
4135 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4138 ttwu_stat(p, task_cpu(p), wake_flags);
4145 * task_call_func - Invoke a function on task in fixed state
4146 * @p: Process for which the function is to be invoked, can be @current.
4147 * @func: Function to invoke.
4148 * @arg: Argument to function.
4150 * Fix the task in it's current state by avoiding wakeups and or rq operations
4151 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4152 * to work out what the state is, if required. Given that @func can be invoked
4153 * with a runqueue lock held, it had better be quite lightweight.
4156 * Whatever @func returns
4158 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4160 struct rq *rq = NULL;
4165 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4167 state = READ_ONCE(p->__state);
4170 * Ensure we load p->on_rq after p->__state, otherwise it would be
4171 * possible to, falsely, observe p->on_rq == 0.
4173 * See try_to_wake_up() for a longer comment.
4178 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4179 * the task is blocked. Make sure to check @state since ttwu() can drop
4180 * locks at the end, see ttwu_queue_wakelist().
4182 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4183 rq = __task_rq_lock(p, &rf);
4186 * At this point the task is pinned; either:
4187 * - blocked and we're holding off wakeups (pi->lock)
4188 * - woken, and we're holding off enqueue (rq->lock)
4189 * - queued, and we're holding off schedule (rq->lock)
4190 * - running, and we're holding off de-schedule (rq->lock)
4192 * The called function (@func) can use: task_curr(), p->on_rq and
4193 * p->__state to differentiate between these states.
4200 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4205 * wake_up_process - Wake up a specific process
4206 * @p: The process to be woken up.
4208 * Attempt to wake up the nominated process and move it to the set of runnable
4211 * Return: 1 if the process was woken up, 0 if it was already running.
4213 * This function executes a full memory barrier before accessing the task state.
4215 int wake_up_process(struct task_struct *p)
4217 return try_to_wake_up(p, TASK_NORMAL, 0);
4219 EXPORT_SYMBOL(wake_up_process);
4221 int wake_up_state(struct task_struct *p, unsigned int state)
4223 return try_to_wake_up(p, state, 0);
4227 * Perform scheduler related setup for a newly forked process p.
4228 * p is forked by current.
4230 * __sched_fork() is basic setup used by init_idle() too:
4232 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4237 p->se.exec_start = 0;
4238 p->se.sum_exec_runtime = 0;
4239 p->se.prev_sum_exec_runtime = 0;
4240 p->se.nr_migrations = 0;
4242 INIT_LIST_HEAD(&p->se.group_node);
4244 #ifdef CONFIG_FAIR_GROUP_SCHED
4245 p->se.cfs_rq = NULL;
4248 #ifdef CONFIG_SCHEDSTATS
4249 /* Even if schedstat is disabled, there should not be garbage */
4250 memset(&p->stats, 0, sizeof(p->stats));
4253 RB_CLEAR_NODE(&p->dl.rb_node);
4254 init_dl_task_timer(&p->dl);
4255 init_dl_inactive_task_timer(&p->dl);
4256 __dl_clear_params(p);
4258 INIT_LIST_HEAD(&p->rt.run_list);
4260 p->rt.time_slice = sched_rr_timeslice;
4264 #ifdef CONFIG_PREEMPT_NOTIFIERS
4265 INIT_HLIST_HEAD(&p->preempt_notifiers);
4268 #ifdef CONFIG_COMPACTION
4269 p->capture_control = NULL;
4271 init_numa_balancing(clone_flags, p);
4273 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4274 p->migration_pending = NULL;
4278 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4280 #ifdef CONFIG_NUMA_BALANCING
4282 void set_numabalancing_state(bool enabled)
4285 static_branch_enable(&sched_numa_balancing);
4287 static_branch_disable(&sched_numa_balancing);
4290 #ifdef CONFIG_PROC_SYSCTL
4291 int sysctl_numa_balancing(struct ctl_table *table, int write,
4292 void *buffer, size_t *lenp, loff_t *ppos)
4296 int state = static_branch_likely(&sched_numa_balancing);
4298 if (write && !capable(CAP_SYS_ADMIN))
4303 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4307 set_numabalancing_state(state);
4313 #ifdef CONFIG_SCHEDSTATS
4315 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4317 static void set_schedstats(bool enabled)
4320 static_branch_enable(&sched_schedstats);
4322 static_branch_disable(&sched_schedstats);
4325 void force_schedstat_enabled(void)
4327 if (!schedstat_enabled()) {
4328 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4329 static_branch_enable(&sched_schedstats);
4333 static int __init setup_schedstats(char *str)
4339 if (!strcmp(str, "enable")) {
4340 set_schedstats(true);
4342 } else if (!strcmp(str, "disable")) {
4343 set_schedstats(false);
4348 pr_warn("Unable to parse schedstats=\n");
4352 __setup("schedstats=", setup_schedstats);
4354 #ifdef CONFIG_PROC_SYSCTL
4355 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4356 size_t *lenp, loff_t *ppos)
4360 int state = static_branch_likely(&sched_schedstats);
4362 if (write && !capable(CAP_SYS_ADMIN))
4367 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4371 set_schedstats(state);
4374 #endif /* CONFIG_PROC_SYSCTL */
4375 #endif /* CONFIG_SCHEDSTATS */
4378 * fork()/clone()-time setup:
4380 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4382 __sched_fork(clone_flags, p);
4384 * We mark the process as NEW here. This guarantees that
4385 * nobody will actually run it, and a signal or other external
4386 * event cannot wake it up and insert it on the runqueue either.
4388 p->__state = TASK_NEW;
4391 * Make sure we do not leak PI boosting priority to the child.
4393 p->prio = current->normal_prio;
4398 * Revert to default priority/policy on fork if requested.
4400 if (unlikely(p->sched_reset_on_fork)) {
4401 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4402 p->policy = SCHED_NORMAL;
4403 p->static_prio = NICE_TO_PRIO(0);
4405 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4406 p->static_prio = NICE_TO_PRIO(0);
4408 p->prio = p->normal_prio = p->static_prio;
4409 set_load_weight(p, false);
4412 * We don't need the reset flag anymore after the fork. It has
4413 * fulfilled its duty:
4415 p->sched_reset_on_fork = 0;
4418 if (dl_prio(p->prio))
4420 else if (rt_prio(p->prio))
4421 p->sched_class = &rt_sched_class;
4423 p->sched_class = &fair_sched_class;
4425 init_entity_runnable_average(&p->se);
4427 #ifdef CONFIG_SCHED_INFO
4428 if (likely(sched_info_on()))
4429 memset(&p->sched_info, 0, sizeof(p->sched_info));
4431 #if defined(CONFIG_SMP)
4434 init_task_preempt_count(p);
4436 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4437 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4442 void sched_post_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4444 unsigned long flags;
4445 #ifdef CONFIG_CGROUP_SCHED
4446 struct task_group *tg;
4449 raw_spin_lock_irqsave(&p->pi_lock, flags);
4450 #ifdef CONFIG_CGROUP_SCHED
4451 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4452 struct task_group, css);
4453 p->sched_task_group = autogroup_task_group(p, tg);
4457 * We're setting the CPU for the first time, we don't migrate,
4458 * so use __set_task_cpu().
4460 __set_task_cpu(p, smp_processor_id());
4461 if (p->sched_class->task_fork)
4462 p->sched_class->task_fork(p);
4463 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4465 uclamp_post_fork(p);
4468 unsigned long to_ratio(u64 period, u64 runtime)
4470 if (runtime == RUNTIME_INF)
4474 * Doing this here saves a lot of checks in all
4475 * the calling paths, and returning zero seems
4476 * safe for them anyway.
4481 return div64_u64(runtime << BW_SHIFT, period);
4485 * wake_up_new_task - wake up a newly created task for the first time.
4487 * This function will do some initial scheduler statistics housekeeping
4488 * that must be done for every newly created context, then puts the task
4489 * on the runqueue and wakes it.
4491 void wake_up_new_task(struct task_struct *p)
4496 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4497 WRITE_ONCE(p->__state, TASK_RUNNING);
4500 * Fork balancing, do it here and not earlier because:
4501 * - cpus_ptr can change in the fork path
4502 * - any previously selected CPU might disappear through hotplug
4504 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4505 * as we're not fully set-up yet.
4507 p->recent_used_cpu = task_cpu(p);
4509 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4511 rq = __task_rq_lock(p, &rf);
4512 update_rq_clock(rq);
4513 post_init_entity_util_avg(p);
4515 activate_task(rq, p, ENQUEUE_NOCLOCK);
4516 trace_sched_wakeup_new(p);
4517 check_preempt_curr(rq, p, WF_FORK);
4519 if (p->sched_class->task_woken) {
4521 * Nothing relies on rq->lock after this, so it's fine to
4524 rq_unpin_lock(rq, &rf);
4525 p->sched_class->task_woken(rq, p);
4526 rq_repin_lock(rq, &rf);
4529 task_rq_unlock(rq, p, &rf);
4532 #ifdef CONFIG_PREEMPT_NOTIFIERS
4534 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4536 void preempt_notifier_inc(void)
4538 static_branch_inc(&preempt_notifier_key);
4540 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4542 void preempt_notifier_dec(void)
4544 static_branch_dec(&preempt_notifier_key);
4546 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4549 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4550 * @notifier: notifier struct to register
4552 void preempt_notifier_register(struct preempt_notifier *notifier)
4554 if (!static_branch_unlikely(&preempt_notifier_key))
4555 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4557 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4559 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4562 * preempt_notifier_unregister - no longer interested in preemption notifications
4563 * @notifier: notifier struct to unregister
4565 * This is *not* safe to call from within a preemption notifier.
4567 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4569 hlist_del(¬ifier->link);
4571 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4573 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4575 struct preempt_notifier *notifier;
4577 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4578 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4581 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4583 if (static_branch_unlikely(&preempt_notifier_key))
4584 __fire_sched_in_preempt_notifiers(curr);
4588 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4589 struct task_struct *next)
4591 struct preempt_notifier *notifier;
4593 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4594 notifier->ops->sched_out(notifier, next);
4597 static __always_inline void
4598 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4599 struct task_struct *next)
4601 if (static_branch_unlikely(&preempt_notifier_key))
4602 __fire_sched_out_preempt_notifiers(curr, next);
4605 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4607 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4612 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4613 struct task_struct *next)
4617 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4619 static inline void prepare_task(struct task_struct *next)
4623 * Claim the task as running, we do this before switching to it
4624 * such that any running task will have this set.
4626 * See the ttwu() WF_ON_CPU case and its ordering comment.
4628 WRITE_ONCE(next->on_cpu, 1);
4632 static inline void finish_task(struct task_struct *prev)
4636 * This must be the very last reference to @prev from this CPU. After
4637 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4638 * must ensure this doesn't happen until the switch is completely
4641 * In particular, the load of prev->state in finish_task_switch() must
4642 * happen before this.
4644 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4646 smp_store_release(&prev->on_cpu, 0);
4652 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4654 void (*func)(struct rq *rq);
4655 struct callback_head *next;
4657 lockdep_assert_rq_held(rq);
4660 func = (void (*)(struct rq *))head->func;
4669 static void balance_push(struct rq *rq);
4671 struct callback_head balance_push_callback = {
4673 .func = (void (*)(struct callback_head *))balance_push,
4676 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4678 struct callback_head *head = rq->balance_callback;
4680 lockdep_assert_rq_held(rq);
4682 rq->balance_callback = NULL;
4687 static void __balance_callbacks(struct rq *rq)
4689 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4692 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4694 unsigned long flags;
4696 if (unlikely(head)) {
4697 raw_spin_rq_lock_irqsave(rq, flags);
4698 do_balance_callbacks(rq, head);
4699 raw_spin_rq_unlock_irqrestore(rq, flags);
4705 static inline void __balance_callbacks(struct rq *rq)
4709 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4714 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4721 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4724 * Since the runqueue lock will be released by the next
4725 * task (which is an invalid locking op but in the case
4726 * of the scheduler it's an obvious special-case), so we
4727 * do an early lockdep release here:
4729 rq_unpin_lock(rq, rf);
4730 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4731 #ifdef CONFIG_DEBUG_SPINLOCK
4732 /* this is a valid case when another task releases the spinlock */
4733 rq_lockp(rq)->owner = next;
4737 static inline void finish_lock_switch(struct rq *rq)
4740 * If we are tracking spinlock dependencies then we have to
4741 * fix up the runqueue lock - which gets 'carried over' from
4742 * prev into current:
4744 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4745 __balance_callbacks(rq);
4746 raw_spin_rq_unlock_irq(rq);
4750 * NOP if the arch has not defined these:
4753 #ifndef prepare_arch_switch
4754 # define prepare_arch_switch(next) do { } while (0)
4757 #ifndef finish_arch_post_lock_switch
4758 # define finish_arch_post_lock_switch() do { } while (0)
4761 static inline void kmap_local_sched_out(void)
4763 #ifdef CONFIG_KMAP_LOCAL
4764 if (unlikely(current->kmap_ctrl.idx))
4765 __kmap_local_sched_out();
4769 static inline void kmap_local_sched_in(void)
4771 #ifdef CONFIG_KMAP_LOCAL
4772 if (unlikely(current->kmap_ctrl.idx))
4773 __kmap_local_sched_in();
4778 * prepare_task_switch - prepare to switch tasks
4779 * @rq: the runqueue preparing to switch
4780 * @prev: the current task that is being switched out
4781 * @next: the task we are going to switch to.
4783 * This is called with the rq lock held and interrupts off. It must
4784 * be paired with a subsequent finish_task_switch after the context
4787 * prepare_task_switch sets up locking and calls architecture specific
4791 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4792 struct task_struct *next)
4794 kcov_prepare_switch(prev);
4795 sched_info_switch(rq, prev, next);
4796 perf_event_task_sched_out(prev, next);
4798 fire_sched_out_preempt_notifiers(prev, next);
4799 kmap_local_sched_out();
4801 prepare_arch_switch(next);
4805 * finish_task_switch - clean up after a task-switch
4806 * @prev: the thread we just switched away from.
4808 * finish_task_switch must be called after the context switch, paired
4809 * with a prepare_task_switch call before the context switch.
4810 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4811 * and do any other architecture-specific cleanup actions.
4813 * Note that we may have delayed dropping an mm in context_switch(). If
4814 * so, we finish that here outside of the runqueue lock. (Doing it
4815 * with the lock held can cause deadlocks; see schedule() for
4818 * The context switch have flipped the stack from under us and restored the
4819 * local variables which were saved when this task called schedule() in the
4820 * past. prev == current is still correct but we need to recalculate this_rq
4821 * because prev may have moved to another CPU.
4823 static struct rq *finish_task_switch(struct task_struct *prev)
4824 __releases(rq->lock)
4826 struct rq *rq = this_rq();
4827 struct mm_struct *mm = rq->prev_mm;
4831 * The previous task will have left us with a preempt_count of 2
4832 * because it left us after:
4835 * preempt_disable(); // 1
4837 * raw_spin_lock_irq(&rq->lock) // 2
4839 * Also, see FORK_PREEMPT_COUNT.
4841 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4842 "corrupted preempt_count: %s/%d/0x%x\n",
4843 current->comm, current->pid, preempt_count()))
4844 preempt_count_set(FORK_PREEMPT_COUNT);
4849 * A task struct has one reference for the use as "current".
4850 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4851 * schedule one last time. The schedule call will never return, and
4852 * the scheduled task must drop that reference.
4854 * We must observe prev->state before clearing prev->on_cpu (in
4855 * finish_task), otherwise a concurrent wakeup can get prev
4856 * running on another CPU and we could rave with its RUNNING -> DEAD
4857 * transition, resulting in a double drop.
4859 prev_state = READ_ONCE(prev->__state);
4860 vtime_task_switch(prev);
4861 perf_event_task_sched_in(prev, current);
4863 tick_nohz_task_switch();
4864 finish_lock_switch(rq);
4865 finish_arch_post_lock_switch();
4866 kcov_finish_switch(current);
4868 * kmap_local_sched_out() is invoked with rq::lock held and
4869 * interrupts disabled. There is no requirement for that, but the
4870 * sched out code does not have an interrupt enabled section.
4871 * Restoring the maps on sched in does not require interrupts being
4874 kmap_local_sched_in();
4876 fire_sched_in_preempt_notifiers(current);
4878 * When switching through a kernel thread, the loop in
4879 * membarrier_{private,global}_expedited() may have observed that
4880 * kernel thread and not issued an IPI. It is therefore possible to
4881 * schedule between user->kernel->user threads without passing though
4882 * switch_mm(). Membarrier requires a barrier after storing to
4883 * rq->curr, before returning to userspace, so provide them here:
4885 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4886 * provided by mmdrop(),
4887 * - a sync_core for SYNC_CORE.
4890 membarrier_mm_sync_core_before_usermode(mm);
4893 if (unlikely(prev_state == TASK_DEAD)) {
4894 if (prev->sched_class->task_dead)
4895 prev->sched_class->task_dead(prev);
4897 /* Task is done with its stack. */
4898 put_task_stack(prev);
4900 put_task_struct_rcu_user(prev);
4907 * schedule_tail - first thing a freshly forked thread must call.
4908 * @prev: the thread we just switched away from.
4910 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4911 __releases(rq->lock)
4914 * New tasks start with FORK_PREEMPT_COUNT, see there and
4915 * finish_task_switch() for details.
4917 * finish_task_switch() will drop rq->lock() and lower preempt_count
4918 * and the preempt_enable() will end up enabling preemption (on
4919 * PREEMPT_COUNT kernels).
4922 finish_task_switch(prev);
4925 if (current->set_child_tid)
4926 put_user(task_pid_vnr(current), current->set_child_tid);
4928 calculate_sigpending();
4932 * context_switch - switch to the new MM and the new thread's register state.
4934 static __always_inline struct rq *
4935 context_switch(struct rq *rq, struct task_struct *prev,
4936 struct task_struct *next, struct rq_flags *rf)
4938 prepare_task_switch(rq, prev, next);
4941 * For paravirt, this is coupled with an exit in switch_to to
4942 * combine the page table reload and the switch backend into
4945 arch_start_context_switch(prev);
4948 * kernel -> kernel lazy + transfer active
4949 * user -> kernel lazy + mmgrab() active
4951 * kernel -> user switch + mmdrop() active
4952 * user -> user switch
4954 if (!next->mm) { // to kernel
4955 enter_lazy_tlb(prev->active_mm, next);
4957 next->active_mm = prev->active_mm;
4958 if (prev->mm) // from user
4959 mmgrab(prev->active_mm);
4961 prev->active_mm = NULL;
4963 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4965 * sys_membarrier() requires an smp_mb() between setting
4966 * rq->curr / membarrier_switch_mm() and returning to userspace.
4968 * The below provides this either through switch_mm(), or in
4969 * case 'prev->active_mm == next->mm' through
4970 * finish_task_switch()'s mmdrop().
4972 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4974 if (!prev->mm) { // from kernel
4975 /* will mmdrop() in finish_task_switch(). */
4976 rq->prev_mm = prev->active_mm;
4977 prev->active_mm = NULL;
4981 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4983 prepare_lock_switch(rq, next, rf);
4985 /* Here we just switch the register state and the stack. */
4986 switch_to(prev, next, prev);
4989 return finish_task_switch(prev);
4993 * nr_running and nr_context_switches:
4995 * externally visible scheduler statistics: current number of runnable
4996 * threads, total number of context switches performed since bootup.
4998 unsigned int nr_running(void)
5000 unsigned int i, sum = 0;
5002 for_each_online_cpu(i)
5003 sum += cpu_rq(i)->nr_running;
5009 * Check if only the current task is running on the CPU.
5011 * Caution: this function does not check that the caller has disabled
5012 * preemption, thus the result might have a time-of-check-to-time-of-use
5013 * race. The caller is responsible to use it correctly, for example:
5015 * - from a non-preemptible section (of course)
5017 * - from a thread that is bound to a single CPU
5019 * - in a loop with very short iterations (e.g. a polling loop)
5021 bool single_task_running(void)
5023 return raw_rq()->nr_running == 1;
5025 EXPORT_SYMBOL(single_task_running);
5027 unsigned long long nr_context_switches(void)
5030 unsigned long long sum = 0;
5032 for_each_possible_cpu(i)
5033 sum += cpu_rq(i)->nr_switches;
5039 * Consumers of these two interfaces, like for example the cpuidle menu
5040 * governor, are using nonsensical data. Preferring shallow idle state selection
5041 * for a CPU that has IO-wait which might not even end up running the task when
5042 * it does become runnable.
5045 unsigned int nr_iowait_cpu(int cpu)
5047 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5051 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5053 * The idea behind IO-wait account is to account the idle time that we could
5054 * have spend running if it were not for IO. That is, if we were to improve the
5055 * storage performance, we'd have a proportional reduction in IO-wait time.
5057 * This all works nicely on UP, where, when a task blocks on IO, we account
5058 * idle time as IO-wait, because if the storage were faster, it could've been
5059 * running and we'd not be idle.
5061 * This has been extended to SMP, by doing the same for each CPU. This however
5064 * Imagine for instance the case where two tasks block on one CPU, only the one
5065 * CPU will have IO-wait accounted, while the other has regular idle. Even
5066 * though, if the storage were faster, both could've ran at the same time,
5067 * utilising both CPUs.
5069 * This means, that when looking globally, the current IO-wait accounting on
5070 * SMP is a lower bound, by reason of under accounting.
5072 * Worse, since the numbers are provided per CPU, they are sometimes
5073 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5074 * associated with any one particular CPU, it can wake to another CPU than it
5075 * blocked on. This means the per CPU IO-wait number is meaningless.
5077 * Task CPU affinities can make all that even more 'interesting'.
5080 unsigned int nr_iowait(void)
5082 unsigned int i, sum = 0;
5084 for_each_possible_cpu(i)
5085 sum += nr_iowait_cpu(i);
5093 * sched_exec - execve() is a valuable balancing opportunity, because at
5094 * this point the task has the smallest effective memory and cache footprint.
5096 void sched_exec(void)
5098 struct task_struct *p = current;
5099 unsigned long flags;
5102 raw_spin_lock_irqsave(&p->pi_lock, flags);
5103 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5104 if (dest_cpu == smp_processor_id())
5107 if (likely(cpu_active(dest_cpu))) {
5108 struct migration_arg arg = { p, dest_cpu };
5110 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5111 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5115 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5120 DEFINE_PER_CPU(struct kernel_stat, kstat);
5121 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5123 EXPORT_PER_CPU_SYMBOL(kstat);
5124 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5127 * The function fair_sched_class.update_curr accesses the struct curr
5128 * and its field curr->exec_start; when called from task_sched_runtime(),
5129 * we observe a high rate of cache misses in practice.
5130 * Prefetching this data results in improved performance.
5132 static inline void prefetch_curr_exec_start(struct task_struct *p)
5134 #ifdef CONFIG_FAIR_GROUP_SCHED
5135 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5137 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5140 prefetch(&curr->exec_start);
5144 * Return accounted runtime for the task.
5145 * In case the task is currently running, return the runtime plus current's
5146 * pending runtime that have not been accounted yet.
5148 unsigned long long task_sched_runtime(struct task_struct *p)
5154 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5156 * 64-bit doesn't need locks to atomically read a 64-bit value.
5157 * So we have a optimization chance when the task's delta_exec is 0.
5158 * Reading ->on_cpu is racy, but this is ok.
5160 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5161 * If we race with it entering CPU, unaccounted time is 0. This is
5162 * indistinguishable from the read occurring a few cycles earlier.
5163 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5164 * been accounted, so we're correct here as well.
5166 if (!p->on_cpu || !task_on_rq_queued(p))
5167 return p->se.sum_exec_runtime;
5170 rq = task_rq_lock(p, &rf);
5172 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5173 * project cycles that may never be accounted to this
5174 * thread, breaking clock_gettime().
5176 if (task_current(rq, p) && task_on_rq_queued(p)) {
5177 prefetch_curr_exec_start(p);
5178 update_rq_clock(rq);
5179 p->sched_class->update_curr(rq);
5181 ns = p->se.sum_exec_runtime;
5182 task_rq_unlock(rq, p, &rf);
5187 #ifdef CONFIG_SCHED_DEBUG
5188 static u64 cpu_resched_latency(struct rq *rq)
5190 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5191 u64 resched_latency, now = rq_clock(rq);
5192 static bool warned_once;
5194 if (sysctl_resched_latency_warn_once && warned_once)
5197 if (!need_resched() || !latency_warn_ms)
5200 if (system_state == SYSTEM_BOOTING)
5203 if (!rq->last_seen_need_resched_ns) {
5204 rq->last_seen_need_resched_ns = now;
5205 rq->ticks_without_resched = 0;
5209 rq->ticks_without_resched++;
5210 resched_latency = now - rq->last_seen_need_resched_ns;
5211 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5216 return resched_latency;
5219 static int __init setup_resched_latency_warn_ms(char *str)
5223 if ((kstrtol(str, 0, &val))) {
5224 pr_warn("Unable to set resched_latency_warn_ms\n");
5228 sysctl_resched_latency_warn_ms = val;
5231 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5233 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5234 #endif /* CONFIG_SCHED_DEBUG */
5237 * This function gets called by the timer code, with HZ frequency.
5238 * We call it with interrupts disabled.
5240 void scheduler_tick(void)
5242 int cpu = smp_processor_id();
5243 struct rq *rq = cpu_rq(cpu);
5244 struct task_struct *curr = rq->curr;
5246 unsigned long thermal_pressure;
5247 u64 resched_latency;
5249 arch_scale_freq_tick();
5254 update_rq_clock(rq);
5255 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5256 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5257 curr->sched_class->task_tick(rq, curr, 0);
5258 if (sched_feat(LATENCY_WARN))
5259 resched_latency = cpu_resched_latency(rq);
5260 calc_global_load_tick(rq);
5261 sched_core_tick(rq);
5265 if (sched_feat(LATENCY_WARN) && resched_latency)
5266 resched_latency_warn(cpu, resched_latency);
5268 perf_event_task_tick();
5271 rq->idle_balance = idle_cpu(cpu);
5272 trigger_load_balance(rq);
5276 #ifdef CONFIG_NO_HZ_FULL
5281 struct delayed_work work;
5283 /* Values for ->state, see diagram below. */
5284 #define TICK_SCHED_REMOTE_OFFLINE 0
5285 #define TICK_SCHED_REMOTE_OFFLINING 1
5286 #define TICK_SCHED_REMOTE_RUNNING 2
5289 * State diagram for ->state:
5292 * TICK_SCHED_REMOTE_OFFLINE
5295 * | | sched_tick_remote()
5298 * +--TICK_SCHED_REMOTE_OFFLINING
5301 * sched_tick_start() | | sched_tick_stop()
5304 * TICK_SCHED_REMOTE_RUNNING
5307 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5308 * and sched_tick_start() are happy to leave the state in RUNNING.
5311 static struct tick_work __percpu *tick_work_cpu;
5313 static void sched_tick_remote(struct work_struct *work)
5315 struct delayed_work *dwork = to_delayed_work(work);
5316 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5317 int cpu = twork->cpu;
5318 struct rq *rq = cpu_rq(cpu);
5319 struct task_struct *curr;
5325 * Handle the tick only if it appears the remote CPU is running in full
5326 * dynticks mode. The check is racy by nature, but missing a tick or
5327 * having one too much is no big deal because the scheduler tick updates
5328 * statistics and checks timeslices in a time-independent way, regardless
5329 * of when exactly it is running.
5331 if (!tick_nohz_tick_stopped_cpu(cpu))
5334 rq_lock_irq(rq, &rf);
5336 if (cpu_is_offline(cpu))
5339 update_rq_clock(rq);
5341 if (!is_idle_task(curr)) {
5343 * Make sure the next tick runs within a reasonable
5346 delta = rq_clock_task(rq) - curr->se.exec_start;
5347 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5349 curr->sched_class->task_tick(rq, curr, 0);
5351 calc_load_nohz_remote(rq);
5353 rq_unlock_irq(rq, &rf);
5357 * Run the remote tick once per second (1Hz). This arbitrary
5358 * frequency is large enough to avoid overload but short enough
5359 * to keep scheduler internal stats reasonably up to date. But
5360 * first update state to reflect hotplug activity if required.
5362 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5363 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5364 if (os == TICK_SCHED_REMOTE_RUNNING)
5365 queue_delayed_work(system_unbound_wq, dwork, HZ);
5368 static void sched_tick_start(int cpu)
5371 struct tick_work *twork;
5373 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5376 WARN_ON_ONCE(!tick_work_cpu);
5378 twork = per_cpu_ptr(tick_work_cpu, cpu);
5379 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5380 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5381 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5383 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5384 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5388 #ifdef CONFIG_HOTPLUG_CPU
5389 static void sched_tick_stop(int cpu)
5391 struct tick_work *twork;
5394 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5397 WARN_ON_ONCE(!tick_work_cpu);
5399 twork = per_cpu_ptr(tick_work_cpu, cpu);
5400 /* There cannot be competing actions, but don't rely on stop-machine. */
5401 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5402 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5403 /* Don't cancel, as this would mess up the state machine. */
5405 #endif /* CONFIG_HOTPLUG_CPU */
5407 int __init sched_tick_offload_init(void)
5409 tick_work_cpu = alloc_percpu(struct tick_work);
5410 BUG_ON(!tick_work_cpu);
5414 #else /* !CONFIG_NO_HZ_FULL */
5415 static inline void sched_tick_start(int cpu) { }
5416 static inline void sched_tick_stop(int cpu) { }
5419 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5420 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5422 * If the value passed in is equal to the current preempt count
5423 * then we just disabled preemption. Start timing the latency.
5425 static inline void preempt_latency_start(int val)
5427 if (preempt_count() == val) {
5428 unsigned long ip = get_lock_parent_ip();
5429 #ifdef CONFIG_DEBUG_PREEMPT
5430 current->preempt_disable_ip = ip;
5432 trace_preempt_off(CALLER_ADDR0, ip);
5436 void preempt_count_add(int val)
5438 #ifdef CONFIG_DEBUG_PREEMPT
5442 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5445 __preempt_count_add(val);
5446 #ifdef CONFIG_DEBUG_PREEMPT
5448 * Spinlock count overflowing soon?
5450 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5453 preempt_latency_start(val);
5455 EXPORT_SYMBOL(preempt_count_add);
5456 NOKPROBE_SYMBOL(preempt_count_add);
5459 * If the value passed in equals to the current preempt count
5460 * then we just enabled preemption. Stop timing the latency.
5462 static inline void preempt_latency_stop(int val)
5464 if (preempt_count() == val)
5465 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5468 void preempt_count_sub(int val)
5470 #ifdef CONFIG_DEBUG_PREEMPT
5474 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5477 * Is the spinlock portion underflowing?
5479 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5480 !(preempt_count() & PREEMPT_MASK)))
5484 preempt_latency_stop(val);
5485 __preempt_count_sub(val);
5487 EXPORT_SYMBOL(preempt_count_sub);
5488 NOKPROBE_SYMBOL(preempt_count_sub);
5491 static inline void preempt_latency_start(int val) { }
5492 static inline void preempt_latency_stop(int val) { }
5495 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5497 #ifdef CONFIG_DEBUG_PREEMPT
5498 return p->preempt_disable_ip;
5505 * Print scheduling while atomic bug:
5507 static noinline void __schedule_bug(struct task_struct *prev)
5509 /* Save this before calling printk(), since that will clobber it */
5510 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5512 if (oops_in_progress)
5515 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5516 prev->comm, prev->pid, preempt_count());
5518 debug_show_held_locks(prev);
5520 if (irqs_disabled())
5521 print_irqtrace_events(prev);
5522 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5523 && in_atomic_preempt_off()) {
5524 pr_err("Preemption disabled at:");
5525 print_ip_sym(KERN_ERR, preempt_disable_ip);
5528 panic("scheduling while atomic\n");
5531 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5535 * Various schedule()-time debugging checks and statistics:
5537 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5539 #ifdef CONFIG_SCHED_STACK_END_CHECK
5540 if (task_stack_end_corrupted(prev))
5541 panic("corrupted stack end detected inside scheduler\n");
5543 if (task_scs_end_corrupted(prev))
5544 panic("corrupted shadow stack detected inside scheduler\n");
5547 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5548 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5549 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5550 prev->comm, prev->pid, prev->non_block_count);
5552 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5556 if (unlikely(in_atomic_preempt_off())) {
5557 __schedule_bug(prev);
5558 preempt_count_set(PREEMPT_DISABLED);
5561 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5563 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5565 schedstat_inc(this_rq()->sched_count);
5568 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5569 struct rq_flags *rf)
5572 const struct sched_class *class;
5574 * We must do the balancing pass before put_prev_task(), such
5575 * that when we release the rq->lock the task is in the same
5576 * state as before we took rq->lock.
5578 * We can terminate the balance pass as soon as we know there is
5579 * a runnable task of @class priority or higher.
5581 for_class_range(class, prev->sched_class, &idle_sched_class) {
5582 if (class->balance(rq, prev, rf))
5587 put_prev_task(rq, prev);
5591 * Pick up the highest-prio task:
5593 static inline struct task_struct *
5594 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5596 const struct sched_class *class;
5597 struct task_struct *p;
5600 * Optimization: we know that if all tasks are in the fair class we can
5601 * call that function directly, but only if the @prev task wasn't of a
5602 * higher scheduling class, because otherwise those lose the
5603 * opportunity to pull in more work from other CPUs.
5605 if (likely(prev->sched_class <= &fair_sched_class &&
5606 rq->nr_running == rq->cfs.h_nr_running)) {
5608 p = pick_next_task_fair(rq, prev, rf);
5609 if (unlikely(p == RETRY_TASK))
5612 /* Assume the next prioritized class is idle_sched_class */
5614 put_prev_task(rq, prev);
5615 p = pick_next_task_idle(rq);
5622 put_prev_task_balance(rq, prev, rf);
5624 for_each_class(class) {
5625 p = class->pick_next_task(rq);
5630 BUG(); /* The idle class should always have a runnable task. */
5633 #ifdef CONFIG_SCHED_CORE
5634 static inline bool is_task_rq_idle(struct task_struct *t)
5636 return (task_rq(t)->idle == t);
5639 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5641 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5644 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5646 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5649 return a->core_cookie == b->core_cookie;
5652 static inline struct task_struct *pick_task(struct rq *rq)
5654 const struct sched_class *class;
5655 struct task_struct *p;
5657 for_each_class(class) {
5658 p = class->pick_task(rq);
5663 BUG(); /* The idle class should always have a runnable task. */
5666 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5668 static struct task_struct *
5669 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5671 struct task_struct *next, *p, *max = NULL;
5672 const struct cpumask *smt_mask;
5673 bool fi_before = false;
5674 bool core_clock_updated = (rq == rq->core);
5675 unsigned long cookie;
5676 int i, cpu, occ = 0;
5680 if (!sched_core_enabled(rq))
5681 return __pick_next_task(rq, prev, rf);
5685 /* Stopper task is switching into idle, no need core-wide selection. */
5686 if (cpu_is_offline(cpu)) {
5688 * Reset core_pick so that we don't enter the fastpath when
5689 * coming online. core_pick would already be migrated to
5690 * another cpu during offline.
5692 rq->core_pick = NULL;
5693 return __pick_next_task(rq, prev, rf);
5697 * If there were no {en,de}queues since we picked (IOW, the task
5698 * pointers are all still valid), and we haven't scheduled the last
5699 * pick yet, do so now.
5701 * rq->core_pick can be NULL if no selection was made for a CPU because
5702 * it was either offline or went offline during a sibling's core-wide
5703 * selection. In this case, do a core-wide selection.
5705 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5706 rq->core->core_pick_seq != rq->core_sched_seq &&
5708 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5710 next = rq->core_pick;
5712 put_prev_task(rq, prev);
5713 set_next_task(rq, next);
5716 rq->core_pick = NULL;
5720 put_prev_task_balance(rq, prev, rf);
5722 smt_mask = cpu_smt_mask(cpu);
5723 need_sync = !!rq->core->core_cookie;
5726 rq->core->core_cookie = 0UL;
5727 if (rq->core->core_forceidle_count) {
5728 if (!core_clock_updated) {
5729 update_rq_clock(rq->core);
5730 core_clock_updated = true;
5732 sched_core_account_forceidle(rq);
5733 /* reset after accounting force idle */
5734 rq->core->core_forceidle_start = 0;
5735 rq->core->core_forceidle_count = 0;
5736 rq->core->core_forceidle_occupation = 0;
5742 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5744 * @task_seq guards the task state ({en,de}queues)
5745 * @pick_seq is the @task_seq we did a selection on
5746 * @sched_seq is the @pick_seq we scheduled
5748 * However, preemptions can cause multiple picks on the same task set.
5749 * 'Fix' this by also increasing @task_seq for every pick.
5751 rq->core->core_task_seq++;
5754 * Optimize for common case where this CPU has no cookies
5755 * and there are no cookied tasks running on siblings.
5758 next = pick_task(rq);
5759 if (!next->core_cookie) {
5760 rq->core_pick = NULL;
5762 * For robustness, update the min_vruntime_fi for
5763 * unconstrained picks as well.
5765 WARN_ON_ONCE(fi_before);
5766 task_vruntime_update(rq, next, false);
5772 * For each thread: do the regular task pick and find the max prio task
5775 * Tie-break prio towards the current CPU
5777 for_each_cpu_wrap(i, smt_mask, cpu) {
5781 * Current cpu always has its clock updated on entrance to
5782 * pick_next_task(). If the current cpu is not the core,
5783 * the core may also have been updated above.
5785 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5786 update_rq_clock(rq_i);
5788 p = rq_i->core_pick = pick_task(rq_i);
5789 if (!max || prio_less(max, p, fi_before))
5793 cookie = rq->core->core_cookie = max->core_cookie;
5796 * For each thread: try and find a runnable task that matches @max or
5799 for_each_cpu(i, smt_mask) {
5801 p = rq_i->core_pick;
5803 if (!cookie_equals(p, cookie)) {
5806 p = sched_core_find(rq_i, cookie);
5808 p = idle_sched_class.pick_task(rq_i);
5811 rq_i->core_pick = p;
5813 if (p == rq_i->idle) {
5814 if (rq_i->nr_running) {
5815 rq->core->core_forceidle_count++;
5817 rq->core->core_forceidle_seq++;
5824 if (schedstat_enabled() && rq->core->core_forceidle_count) {
5826 rq->core->core_forceidle_start = rq_clock(rq->core);
5827 rq->core->core_forceidle_occupation = occ;
5830 rq->core->core_pick_seq = rq->core->core_task_seq;
5831 next = rq->core_pick;
5832 rq->core_sched_seq = rq->core->core_pick_seq;
5834 /* Something should have been selected for current CPU */
5835 WARN_ON_ONCE(!next);
5838 * Reschedule siblings
5840 * NOTE: L1TF -- at this point we're no longer running the old task and
5841 * sending an IPI (below) ensures the sibling will no longer be running
5842 * their task. This ensures there is no inter-sibling overlap between
5843 * non-matching user state.
5845 for_each_cpu(i, smt_mask) {
5849 * An online sibling might have gone offline before a task
5850 * could be picked for it, or it might be offline but later
5851 * happen to come online, but its too late and nothing was
5852 * picked for it. That's Ok - it will pick tasks for itself,
5855 if (!rq_i->core_pick)
5859 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5860 * fi_before fi update?
5866 if (!(fi_before && rq->core->core_forceidle_count))
5867 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
5869 rq_i->core_pick->core_occupation = occ;
5872 rq_i->core_pick = NULL;
5876 /* Did we break L1TF mitigation requirements? */
5877 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5879 if (rq_i->curr == rq_i->core_pick) {
5880 rq_i->core_pick = NULL;
5888 set_next_task(rq, next);
5892 static bool try_steal_cookie(int this, int that)
5894 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5895 struct task_struct *p;
5896 unsigned long cookie;
5897 bool success = false;
5899 local_irq_disable();
5900 double_rq_lock(dst, src);
5902 cookie = dst->core->core_cookie;
5906 if (dst->curr != dst->idle)
5909 p = sched_core_find(src, cookie);
5914 if (p == src->core_pick || p == src->curr)
5917 if (!cpumask_test_cpu(this, &p->cpus_mask))
5920 if (p->core_occupation > dst->idle->core_occupation)
5923 deactivate_task(src, p, 0);
5924 set_task_cpu(p, this);
5925 activate_task(dst, p, 0);
5933 p = sched_core_next(p, cookie);
5937 double_rq_unlock(dst, src);
5943 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5947 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5954 if (try_steal_cookie(cpu, i))
5961 static void sched_core_balance(struct rq *rq)
5963 struct sched_domain *sd;
5964 int cpu = cpu_of(rq);
5968 raw_spin_rq_unlock_irq(rq);
5969 for_each_domain(cpu, sd) {
5973 if (steal_cookie_task(cpu, sd))
5976 raw_spin_rq_lock_irq(rq);
5981 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5983 void queue_core_balance(struct rq *rq)
5985 if (!sched_core_enabled(rq))
5988 if (!rq->core->core_cookie)
5991 if (!rq->nr_running) /* not forced idle */
5994 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5997 static void sched_core_cpu_starting(unsigned int cpu)
5999 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6000 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6001 unsigned long flags;
6004 sched_core_lock(cpu, &flags);
6006 WARN_ON_ONCE(rq->core != rq);
6008 /* if we're the first, we'll be our own leader */
6009 if (cpumask_weight(smt_mask) == 1)
6012 /* find the leader */
6013 for_each_cpu(t, smt_mask) {
6017 if (rq->core == rq) {
6023 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6026 /* install and validate core_rq */
6027 for_each_cpu(t, smt_mask) {
6033 WARN_ON_ONCE(rq->core != core_rq);
6037 sched_core_unlock(cpu, &flags);
6040 static void sched_core_cpu_deactivate(unsigned int cpu)
6042 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6043 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6044 unsigned long flags;
6047 sched_core_lock(cpu, &flags);
6049 /* if we're the last man standing, nothing to do */
6050 if (cpumask_weight(smt_mask) == 1) {
6051 WARN_ON_ONCE(rq->core != rq);
6055 /* if we're not the leader, nothing to do */
6059 /* find a new leader */
6060 for_each_cpu(t, smt_mask) {
6063 core_rq = cpu_rq(t);
6067 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6070 /* copy the shared state to the new leader */
6071 core_rq->core_task_seq = rq->core_task_seq;
6072 core_rq->core_pick_seq = rq->core_pick_seq;
6073 core_rq->core_cookie = rq->core_cookie;
6074 core_rq->core_forceidle_count = rq->core_forceidle_count;
6075 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6076 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6079 * Accounting edge for forced idle is handled in pick_next_task().
6080 * Don't need another one here, since the hotplug thread shouldn't
6083 core_rq->core_forceidle_start = 0;
6085 /* install new leader */
6086 for_each_cpu(t, smt_mask) {
6092 sched_core_unlock(cpu, &flags);
6095 static inline void sched_core_cpu_dying(unsigned int cpu)
6097 struct rq *rq = cpu_rq(cpu);
6103 #else /* !CONFIG_SCHED_CORE */
6105 static inline void sched_core_cpu_starting(unsigned int cpu) {}
6106 static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6107 static inline void sched_core_cpu_dying(unsigned int cpu) {}
6109 static struct task_struct *
6110 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6112 return __pick_next_task(rq, prev, rf);
6115 #endif /* CONFIG_SCHED_CORE */
6118 * Constants for the sched_mode argument of __schedule().
6120 * The mode argument allows RT enabled kernels to differentiate a
6121 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6122 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6123 * optimize the AND operation out and just check for zero.
6126 #define SM_PREEMPT 0x1
6127 #define SM_RTLOCK_WAIT 0x2
6129 #ifndef CONFIG_PREEMPT_RT
6130 # define SM_MASK_PREEMPT (~0U)
6132 # define SM_MASK_PREEMPT SM_PREEMPT
6136 * __schedule() is the main scheduler function.
6138 * The main means of driving the scheduler and thus entering this function are:
6140 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6142 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6143 * paths. For example, see arch/x86/entry_64.S.
6145 * To drive preemption between tasks, the scheduler sets the flag in timer
6146 * interrupt handler scheduler_tick().
6148 * 3. Wakeups don't really cause entry into schedule(). They add a
6149 * task to the run-queue and that's it.
6151 * Now, if the new task added to the run-queue preempts the current
6152 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6153 * called on the nearest possible occasion:
6155 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
6157 * - in syscall or exception context, at the next outmost
6158 * preempt_enable(). (this might be as soon as the wake_up()'s
6161 * - in IRQ context, return from interrupt-handler to
6162 * preemptible context
6164 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6167 * - cond_resched() call
6168 * - explicit schedule() call
6169 * - return from syscall or exception to user-space
6170 * - return from interrupt-handler to user-space
6172 * WARNING: must be called with preemption disabled!
6174 static void __sched notrace __schedule(unsigned int sched_mode)
6176 struct task_struct *prev, *next;
6177 unsigned long *switch_count;
6178 unsigned long prev_state;
6183 cpu = smp_processor_id();
6187 schedule_debug(prev, !!sched_mode);
6189 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6192 local_irq_disable();
6193 rcu_note_context_switch(!!sched_mode);
6196 * Make sure that signal_pending_state()->signal_pending() below
6197 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6198 * done by the caller to avoid the race with signal_wake_up():
6200 * __set_current_state(@state) signal_wake_up()
6201 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6202 * wake_up_state(p, state)
6203 * LOCK rq->lock LOCK p->pi_state
6204 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6205 * if (signal_pending_state()) if (p->state & @state)
6207 * Also, the membarrier system call requires a full memory barrier
6208 * after coming from user-space, before storing to rq->curr.
6211 smp_mb__after_spinlock();
6213 /* Promote REQ to ACT */
6214 rq->clock_update_flags <<= 1;
6215 update_rq_clock(rq);
6217 switch_count = &prev->nivcsw;
6220 * We must load prev->state once (task_struct::state is volatile), such
6223 * - we form a control dependency vs deactivate_task() below.
6224 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
6226 prev_state = READ_ONCE(prev->__state);
6227 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
6228 if (signal_pending_state(prev_state, prev)) {
6229 WRITE_ONCE(prev->__state, TASK_RUNNING);
6231 prev->sched_contributes_to_load =
6232 (prev_state & TASK_UNINTERRUPTIBLE) &&
6233 !(prev_state & TASK_NOLOAD) &&
6234 !(prev->flags & PF_FROZEN);
6236 if (prev->sched_contributes_to_load)
6237 rq->nr_uninterruptible++;
6240 * __schedule() ttwu()
6241 * prev_state = prev->state; if (p->on_rq && ...)
6242 * if (prev_state) goto out;
6243 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6244 * p->state = TASK_WAKING
6246 * Where __schedule() and ttwu() have matching control dependencies.
6248 * After this, schedule() must not care about p->state any more.
6250 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6252 if (prev->in_iowait) {
6253 atomic_inc(&rq->nr_iowait);
6254 delayacct_blkio_start();
6257 switch_count = &prev->nvcsw;
6260 next = pick_next_task(rq, prev, &rf);
6261 clear_tsk_need_resched(prev);
6262 clear_preempt_need_resched();
6263 #ifdef CONFIG_SCHED_DEBUG
6264 rq->last_seen_need_resched_ns = 0;
6267 if (likely(prev != next)) {
6270 * RCU users of rcu_dereference(rq->curr) may not see
6271 * changes to task_struct made by pick_next_task().
6273 RCU_INIT_POINTER(rq->curr, next);
6275 * The membarrier system call requires each architecture
6276 * to have a full memory barrier after updating
6277 * rq->curr, before returning to user-space.
6279 * Here are the schemes providing that barrier on the
6280 * various architectures:
6281 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6282 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6283 * - finish_lock_switch() for weakly-ordered
6284 * architectures where spin_unlock is a full barrier,
6285 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6286 * is a RELEASE barrier),
6290 migrate_disable_switch(rq, prev);
6291 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6293 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next);
6295 /* Also unlocks the rq: */
6296 rq = context_switch(rq, prev, next, &rf);
6298 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6300 rq_unpin_lock(rq, &rf);
6301 __balance_callbacks(rq);
6302 raw_spin_rq_unlock_irq(rq);
6306 void __noreturn do_task_dead(void)
6308 /* Causes final put_task_struct in finish_task_switch(): */
6309 set_special_state(TASK_DEAD);
6311 /* Tell freezer to ignore us: */
6312 current->flags |= PF_NOFREEZE;
6314 __schedule(SM_NONE);
6317 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6322 static inline void sched_submit_work(struct task_struct *tsk)
6324 unsigned int task_flags;
6326 if (task_is_running(tsk))
6329 task_flags = tsk->flags;
6331 * If a worker goes to sleep, notify and ask workqueue whether it
6332 * wants to wake up a task to maintain concurrency.
6334 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6335 if (task_flags & PF_WQ_WORKER)
6336 wq_worker_sleeping(tsk);
6338 io_wq_worker_sleeping(tsk);
6341 if (tsk_is_pi_blocked(tsk))
6345 * If we are going to sleep and we have plugged IO queued,
6346 * make sure to submit it to avoid deadlocks.
6348 if (blk_needs_flush_plug(tsk))
6349 blk_flush_plug(tsk->plug, true);
6352 static void sched_update_worker(struct task_struct *tsk)
6354 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6355 if (tsk->flags & PF_WQ_WORKER)
6356 wq_worker_running(tsk);
6358 io_wq_worker_running(tsk);
6362 asmlinkage __visible void __sched schedule(void)
6364 struct task_struct *tsk = current;
6366 sched_submit_work(tsk);
6369 __schedule(SM_NONE);
6370 sched_preempt_enable_no_resched();
6371 } while (need_resched());
6372 sched_update_worker(tsk);
6374 EXPORT_SYMBOL(schedule);
6377 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6378 * state (have scheduled out non-voluntarily) by making sure that all
6379 * tasks have either left the run queue or have gone into user space.
6380 * As idle tasks do not do either, they must not ever be preempted
6381 * (schedule out non-voluntarily).
6383 * schedule_idle() is similar to schedule_preempt_disable() except that it
6384 * never enables preemption because it does not call sched_submit_work().
6386 void __sched schedule_idle(void)
6389 * As this skips calling sched_submit_work(), which the idle task does
6390 * regardless because that function is a nop when the task is in a
6391 * TASK_RUNNING state, make sure this isn't used someplace that the
6392 * current task can be in any other state. Note, idle is always in the
6393 * TASK_RUNNING state.
6395 WARN_ON_ONCE(current->__state);
6397 __schedule(SM_NONE);
6398 } while (need_resched());
6401 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6402 asmlinkage __visible void __sched schedule_user(void)
6405 * If we come here after a random call to set_need_resched(),
6406 * or we have been woken up remotely but the IPI has not yet arrived,
6407 * we haven't yet exited the RCU idle mode. Do it here manually until
6408 * we find a better solution.
6410 * NB: There are buggy callers of this function. Ideally we
6411 * should warn if prev_state != CONTEXT_USER, but that will trigger
6412 * too frequently to make sense yet.
6414 enum ctx_state prev_state = exception_enter();
6416 exception_exit(prev_state);
6421 * schedule_preempt_disabled - called with preemption disabled
6423 * Returns with preemption disabled. Note: preempt_count must be 1
6425 void __sched schedule_preempt_disabled(void)
6427 sched_preempt_enable_no_resched();
6432 #ifdef CONFIG_PREEMPT_RT
6433 void __sched notrace schedule_rtlock(void)
6437 __schedule(SM_RTLOCK_WAIT);
6438 sched_preempt_enable_no_resched();
6439 } while (need_resched());
6441 NOKPROBE_SYMBOL(schedule_rtlock);
6444 static void __sched notrace preempt_schedule_common(void)
6448 * Because the function tracer can trace preempt_count_sub()
6449 * and it also uses preempt_enable/disable_notrace(), if
6450 * NEED_RESCHED is set, the preempt_enable_notrace() called
6451 * by the function tracer will call this function again and
6452 * cause infinite recursion.
6454 * Preemption must be disabled here before the function
6455 * tracer can trace. Break up preempt_disable() into two
6456 * calls. One to disable preemption without fear of being
6457 * traced. The other to still record the preemption latency,
6458 * which can also be traced by the function tracer.
6460 preempt_disable_notrace();
6461 preempt_latency_start(1);
6462 __schedule(SM_PREEMPT);
6463 preempt_latency_stop(1);
6464 preempt_enable_no_resched_notrace();
6467 * Check again in case we missed a preemption opportunity
6468 * between schedule and now.
6470 } while (need_resched());
6473 #ifdef CONFIG_PREEMPTION
6475 * This is the entry point to schedule() from in-kernel preemption
6476 * off of preempt_enable.
6478 asmlinkage __visible void __sched notrace preempt_schedule(void)
6481 * If there is a non-zero preempt_count or interrupts are disabled,
6482 * we do not want to preempt the current task. Just return..
6484 if (likely(!preemptible()))
6487 preempt_schedule_common();
6489 NOKPROBE_SYMBOL(preempt_schedule);
6490 EXPORT_SYMBOL(preempt_schedule);
6492 #ifdef CONFIG_PREEMPT_DYNAMIC
6493 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6494 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6499 * preempt_schedule_notrace - preempt_schedule called by tracing
6501 * The tracing infrastructure uses preempt_enable_notrace to prevent
6502 * recursion and tracing preempt enabling caused by the tracing
6503 * infrastructure itself. But as tracing can happen in areas coming
6504 * from userspace or just about to enter userspace, a preempt enable
6505 * can occur before user_exit() is called. This will cause the scheduler
6506 * to be called when the system is still in usermode.
6508 * To prevent this, the preempt_enable_notrace will use this function
6509 * instead of preempt_schedule() to exit user context if needed before
6510 * calling the scheduler.
6512 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6514 enum ctx_state prev_ctx;
6516 if (likely(!preemptible()))
6521 * Because the function tracer can trace preempt_count_sub()
6522 * and it also uses preempt_enable/disable_notrace(), if
6523 * NEED_RESCHED is set, the preempt_enable_notrace() called
6524 * by the function tracer will call this function again and
6525 * cause infinite recursion.
6527 * Preemption must be disabled here before the function
6528 * tracer can trace. Break up preempt_disable() into two
6529 * calls. One to disable preemption without fear of being
6530 * traced. The other to still record the preemption latency,
6531 * which can also be traced by the function tracer.
6533 preempt_disable_notrace();
6534 preempt_latency_start(1);
6536 * Needs preempt disabled in case user_exit() is traced
6537 * and the tracer calls preempt_enable_notrace() causing
6538 * an infinite recursion.
6540 prev_ctx = exception_enter();
6541 __schedule(SM_PREEMPT);
6542 exception_exit(prev_ctx);
6544 preempt_latency_stop(1);
6545 preempt_enable_no_resched_notrace();
6546 } while (need_resched());
6548 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6550 #ifdef CONFIG_PREEMPT_DYNAMIC
6551 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6552 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6555 #endif /* CONFIG_PREEMPTION */
6557 #ifdef CONFIG_PREEMPT_DYNAMIC
6559 #include <linux/entry-common.h>
6564 * SC:preempt_schedule
6565 * SC:preempt_schedule_notrace
6566 * SC:irqentry_exit_cond_resched
6570 * cond_resched <- __cond_resched
6571 * might_resched <- RET0
6572 * preempt_schedule <- NOP
6573 * preempt_schedule_notrace <- NOP
6574 * irqentry_exit_cond_resched <- NOP
6577 * cond_resched <- __cond_resched
6578 * might_resched <- __cond_resched
6579 * preempt_schedule <- NOP
6580 * preempt_schedule_notrace <- NOP
6581 * irqentry_exit_cond_resched <- NOP
6584 * cond_resched <- RET0
6585 * might_resched <- RET0
6586 * preempt_schedule <- preempt_schedule
6587 * preempt_schedule_notrace <- preempt_schedule_notrace
6588 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6592 preempt_dynamic_undefined = -1,
6593 preempt_dynamic_none,
6594 preempt_dynamic_voluntary,
6595 preempt_dynamic_full,
6598 int preempt_dynamic_mode = preempt_dynamic_undefined;
6600 int sched_dynamic_mode(const char *str)
6602 if (!strcmp(str, "none"))
6603 return preempt_dynamic_none;
6605 if (!strcmp(str, "voluntary"))
6606 return preempt_dynamic_voluntary;
6608 if (!strcmp(str, "full"))
6609 return preempt_dynamic_full;
6614 void sched_dynamic_update(int mode)
6617 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6618 * the ZERO state, which is invalid.
6620 static_call_update(cond_resched, __cond_resched);
6621 static_call_update(might_resched, __cond_resched);
6622 static_call_update(preempt_schedule, __preempt_schedule_func);
6623 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6624 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6627 case preempt_dynamic_none:
6628 static_call_update(cond_resched, __cond_resched);
6629 static_call_update(might_resched, (void *)&__static_call_return0);
6630 static_call_update(preempt_schedule, NULL);
6631 static_call_update(preempt_schedule_notrace, NULL);
6632 static_call_update(irqentry_exit_cond_resched, NULL);
6633 pr_info("Dynamic Preempt: none\n");
6636 case preempt_dynamic_voluntary:
6637 static_call_update(cond_resched, __cond_resched);
6638 static_call_update(might_resched, __cond_resched);
6639 static_call_update(preempt_schedule, NULL);
6640 static_call_update(preempt_schedule_notrace, NULL);
6641 static_call_update(irqentry_exit_cond_resched, NULL);
6642 pr_info("Dynamic Preempt: voluntary\n");
6645 case preempt_dynamic_full:
6646 static_call_update(cond_resched, (void *)&__static_call_return0);
6647 static_call_update(might_resched, (void *)&__static_call_return0);
6648 static_call_update(preempt_schedule, __preempt_schedule_func);
6649 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6650 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6651 pr_info("Dynamic Preempt: full\n");
6655 preempt_dynamic_mode = mode;
6658 static int __init setup_preempt_mode(char *str)
6660 int mode = sched_dynamic_mode(str);
6662 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6666 sched_dynamic_update(mode);
6669 __setup("preempt=", setup_preempt_mode);
6671 static void __init preempt_dynamic_init(void)
6673 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
6674 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
6675 sched_dynamic_update(preempt_dynamic_none);
6676 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
6677 sched_dynamic_update(preempt_dynamic_voluntary);
6679 /* Default static call setting, nothing to do */
6680 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
6681 preempt_dynamic_mode = preempt_dynamic_full;
6682 pr_info("Dynamic Preempt: full\n");
6687 #else /* !CONFIG_PREEMPT_DYNAMIC */
6689 static inline void preempt_dynamic_init(void) { }
6691 #endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
6694 * This is the entry point to schedule() from kernel preemption
6695 * off of irq context.
6696 * Note, that this is called and return with irqs disabled. This will
6697 * protect us against recursive calling from irq.
6699 asmlinkage __visible void __sched preempt_schedule_irq(void)
6701 enum ctx_state prev_state;
6703 /* Catch callers which need to be fixed */
6704 BUG_ON(preempt_count() || !irqs_disabled());
6706 prev_state = exception_enter();
6711 __schedule(SM_PREEMPT);
6712 local_irq_disable();
6713 sched_preempt_enable_no_resched();
6714 } while (need_resched());
6716 exception_exit(prev_state);
6719 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6722 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6723 return try_to_wake_up(curr->private, mode, wake_flags);
6725 EXPORT_SYMBOL(default_wake_function);
6727 static void __setscheduler_prio(struct task_struct *p, int prio)
6730 p->sched_class = &dl_sched_class;
6731 else if (rt_prio(prio))
6732 p->sched_class = &rt_sched_class;
6734 p->sched_class = &fair_sched_class;
6739 #ifdef CONFIG_RT_MUTEXES
6741 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6744 prio = min(prio, pi_task->prio);
6749 static inline int rt_effective_prio(struct task_struct *p, int prio)
6751 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6753 return __rt_effective_prio(pi_task, prio);
6757 * rt_mutex_setprio - set the current priority of a task
6759 * @pi_task: donor task
6761 * This function changes the 'effective' priority of a task. It does
6762 * not touch ->normal_prio like __setscheduler().
6764 * Used by the rt_mutex code to implement priority inheritance
6765 * logic. Call site only calls if the priority of the task changed.
6767 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6769 int prio, oldprio, queued, running, queue_flag =
6770 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6771 const struct sched_class *prev_class;
6775 /* XXX used to be waiter->prio, not waiter->task->prio */
6776 prio = __rt_effective_prio(pi_task, p->normal_prio);
6779 * If nothing changed; bail early.
6781 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6784 rq = __task_rq_lock(p, &rf);
6785 update_rq_clock(rq);
6787 * Set under pi_lock && rq->lock, such that the value can be used under
6790 * Note that there is loads of tricky to make this pointer cache work
6791 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6792 * ensure a task is de-boosted (pi_task is set to NULL) before the
6793 * task is allowed to run again (and can exit). This ensures the pointer
6794 * points to a blocked task -- which guarantees the task is present.
6796 p->pi_top_task = pi_task;
6799 * For FIFO/RR we only need to set prio, if that matches we're done.
6801 if (prio == p->prio && !dl_prio(prio))
6805 * Idle task boosting is a nono in general. There is one
6806 * exception, when PREEMPT_RT and NOHZ is active:
6808 * The idle task calls get_next_timer_interrupt() and holds
6809 * the timer wheel base->lock on the CPU and another CPU wants
6810 * to access the timer (probably to cancel it). We can safely
6811 * ignore the boosting request, as the idle CPU runs this code
6812 * with interrupts disabled and will complete the lock
6813 * protected section without being interrupted. So there is no
6814 * real need to boost.
6816 if (unlikely(p == rq->idle)) {
6817 WARN_ON(p != rq->curr);
6818 WARN_ON(p->pi_blocked_on);
6822 trace_sched_pi_setprio(p, pi_task);
6825 if (oldprio == prio)
6826 queue_flag &= ~DEQUEUE_MOVE;
6828 prev_class = p->sched_class;
6829 queued = task_on_rq_queued(p);
6830 running = task_current(rq, p);
6832 dequeue_task(rq, p, queue_flag);
6834 put_prev_task(rq, p);
6837 * Boosting condition are:
6838 * 1. -rt task is running and holds mutex A
6839 * --> -dl task blocks on mutex A
6841 * 2. -dl task is running and holds mutex A
6842 * --> -dl task blocks on mutex A and could preempt the
6845 if (dl_prio(prio)) {
6846 if (!dl_prio(p->normal_prio) ||
6847 (pi_task && dl_prio(pi_task->prio) &&
6848 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6849 p->dl.pi_se = pi_task->dl.pi_se;
6850 queue_flag |= ENQUEUE_REPLENISH;
6852 p->dl.pi_se = &p->dl;
6854 } else if (rt_prio(prio)) {
6855 if (dl_prio(oldprio))
6856 p->dl.pi_se = &p->dl;
6858 queue_flag |= ENQUEUE_HEAD;
6860 if (dl_prio(oldprio))
6861 p->dl.pi_se = &p->dl;
6862 if (rt_prio(oldprio))
6866 __setscheduler_prio(p, prio);
6869 enqueue_task(rq, p, queue_flag);
6871 set_next_task(rq, p);
6873 check_class_changed(rq, p, prev_class, oldprio);
6875 /* Avoid rq from going away on us: */
6878 rq_unpin_lock(rq, &rf);
6879 __balance_callbacks(rq);
6880 raw_spin_rq_unlock(rq);
6885 static inline int rt_effective_prio(struct task_struct *p, int prio)
6891 void set_user_nice(struct task_struct *p, long nice)
6893 bool queued, running;
6898 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6901 * We have to be careful, if called from sys_setpriority(),
6902 * the task might be in the middle of scheduling on another CPU.
6904 rq = task_rq_lock(p, &rf);
6905 update_rq_clock(rq);
6908 * The RT priorities are set via sched_setscheduler(), but we still
6909 * allow the 'normal' nice value to be set - but as expected
6910 * it won't have any effect on scheduling until the task is
6911 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6913 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6914 p->static_prio = NICE_TO_PRIO(nice);
6917 queued = task_on_rq_queued(p);
6918 running = task_current(rq, p);
6920 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6922 put_prev_task(rq, p);
6924 p->static_prio = NICE_TO_PRIO(nice);
6925 set_load_weight(p, true);
6927 p->prio = effective_prio(p);
6930 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6932 set_next_task(rq, p);
6935 * If the task increased its priority or is running and
6936 * lowered its priority, then reschedule its CPU:
6938 p->sched_class->prio_changed(rq, p, old_prio);
6941 task_rq_unlock(rq, p, &rf);
6943 EXPORT_SYMBOL(set_user_nice);
6946 * can_nice - check if a task can reduce its nice value
6950 int can_nice(const struct task_struct *p, const int nice)
6952 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6953 int nice_rlim = nice_to_rlimit(nice);
6955 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6956 capable(CAP_SYS_NICE));
6959 #ifdef __ARCH_WANT_SYS_NICE
6962 * sys_nice - change the priority of the current process.
6963 * @increment: priority increment
6965 * sys_setpriority is a more generic, but much slower function that
6966 * does similar things.
6968 SYSCALL_DEFINE1(nice, int, increment)
6973 * Setpriority might change our priority at the same moment.
6974 * We don't have to worry. Conceptually one call occurs first
6975 * and we have a single winner.
6977 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6978 nice = task_nice(current) + increment;
6980 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6981 if (increment < 0 && !can_nice(current, nice))
6984 retval = security_task_setnice(current, nice);
6988 set_user_nice(current, nice);
6995 * task_prio - return the priority value of a given task.
6996 * @p: the task in question.
6998 * Return: The priority value as seen by users in /proc.
7000 * sched policy return value kernel prio user prio/nice
7002 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7003 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7004 * deadline -101 -1 0
7006 int task_prio(const struct task_struct *p)
7008 return p->prio - MAX_RT_PRIO;
7012 * idle_cpu - is a given CPU idle currently?
7013 * @cpu: the processor in question.
7015 * Return: 1 if the CPU is currently idle. 0 otherwise.
7017 int idle_cpu(int cpu)
7019 struct rq *rq = cpu_rq(cpu);
7021 if (rq->curr != rq->idle)
7028 if (rq->ttwu_pending)
7036 * available_idle_cpu - is a given CPU idle for enqueuing work.
7037 * @cpu: the CPU in question.
7039 * Return: 1 if the CPU is currently idle. 0 otherwise.
7041 int available_idle_cpu(int cpu)
7046 if (vcpu_is_preempted(cpu))
7053 * idle_task - return the idle task for a given CPU.
7054 * @cpu: the processor in question.
7056 * Return: The idle task for the CPU @cpu.
7058 struct task_struct *idle_task(int cpu)
7060 return cpu_rq(cpu)->idle;
7065 * This function computes an effective utilization for the given CPU, to be
7066 * used for frequency selection given the linear relation: f = u * f_max.
7068 * The scheduler tracks the following metrics:
7070 * cpu_util_{cfs,rt,dl,irq}()
7073 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7074 * synchronized windows and are thus directly comparable.
7076 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7077 * which excludes things like IRQ and steal-time. These latter are then accrued
7078 * in the irq utilization.
7080 * The DL bandwidth number otoh is not a measured metric but a value computed
7081 * based on the task model parameters and gives the minimal utilization
7082 * required to meet deadlines.
7084 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
7085 unsigned long max, enum cpu_util_type type,
7086 struct task_struct *p)
7088 unsigned long dl_util, util, irq;
7089 struct rq *rq = cpu_rq(cpu);
7091 if (!uclamp_is_used() &&
7092 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7097 * Early check to see if IRQ/steal time saturates the CPU, can be
7098 * because of inaccuracies in how we track these -- see
7099 * update_irq_load_avg().
7101 irq = cpu_util_irq(rq);
7102 if (unlikely(irq >= max))
7106 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7107 * CFS tasks and we use the same metric to track the effective
7108 * utilization (PELT windows are synchronized) we can directly add them
7109 * to obtain the CPU's actual utilization.
7111 * CFS and RT utilization can be boosted or capped, depending on
7112 * utilization clamp constraints requested by currently RUNNABLE
7114 * When there are no CFS RUNNABLE tasks, clamps are released and
7115 * frequency will be gracefully reduced with the utilization decay.
7117 util = util_cfs + cpu_util_rt(rq);
7118 if (type == FREQUENCY_UTIL)
7119 util = uclamp_rq_util_with(rq, util, p);
7121 dl_util = cpu_util_dl(rq);
7124 * For frequency selection we do not make cpu_util_dl() a permanent part
7125 * of this sum because we want to use cpu_bw_dl() later on, but we need
7126 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7127 * that we select f_max when there is no idle time.
7129 * NOTE: numerical errors or stop class might cause us to not quite hit
7130 * saturation when we should -- something for later.
7132 if (util + dl_util >= max)
7136 * OTOH, for energy computation we need the estimated running time, so
7137 * include util_dl and ignore dl_bw.
7139 if (type == ENERGY_UTIL)
7143 * There is still idle time; further improve the number by using the
7144 * irq metric. Because IRQ/steal time is hidden from the task clock we
7145 * need to scale the task numbers:
7148 * U' = irq + --------- * U
7151 util = scale_irq_capacity(util, irq, max);
7155 * Bandwidth required by DEADLINE must always be granted while, for
7156 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7157 * to gracefully reduce the frequency when no tasks show up for longer
7160 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7161 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7162 * an interface. So, we only do the latter for now.
7164 if (type == FREQUENCY_UTIL)
7165 util += cpu_bw_dl(rq);
7167 return min(max, util);
7170 unsigned long sched_cpu_util(int cpu, unsigned long max)
7172 return effective_cpu_util(cpu, cpu_util_cfs(cpu), max,
7175 #endif /* CONFIG_SMP */
7178 * find_process_by_pid - find a process with a matching PID value.
7179 * @pid: the pid in question.
7181 * The task of @pid, if found. %NULL otherwise.
7183 static struct task_struct *find_process_by_pid(pid_t pid)
7185 return pid ? find_task_by_vpid(pid) : current;
7189 * sched_setparam() passes in -1 for its policy, to let the functions
7190 * it calls know not to change it.
7192 #define SETPARAM_POLICY -1
7194 static void __setscheduler_params(struct task_struct *p,
7195 const struct sched_attr *attr)
7197 int policy = attr->sched_policy;
7199 if (policy == SETPARAM_POLICY)
7204 if (dl_policy(policy))
7205 __setparam_dl(p, attr);
7206 else if (fair_policy(policy))
7207 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7210 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7211 * !rt_policy. Always setting this ensures that things like
7212 * getparam()/getattr() don't report silly values for !rt tasks.
7214 p->rt_priority = attr->sched_priority;
7215 p->normal_prio = normal_prio(p);
7216 set_load_weight(p, true);
7220 * Check the target process has a UID that matches the current process's:
7222 static bool check_same_owner(struct task_struct *p)
7224 const struct cred *cred = current_cred(), *pcred;
7228 pcred = __task_cred(p);
7229 match = (uid_eq(cred->euid, pcred->euid) ||
7230 uid_eq(cred->euid, pcred->uid));
7235 static int __sched_setscheduler(struct task_struct *p,
7236 const struct sched_attr *attr,
7239 int oldpolicy = -1, policy = attr->sched_policy;
7240 int retval, oldprio, newprio, queued, running;
7241 const struct sched_class *prev_class;
7242 struct callback_head *head;
7245 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7248 /* The pi code expects interrupts enabled */
7249 BUG_ON(pi && in_interrupt());
7251 /* Double check policy once rq lock held: */
7253 reset_on_fork = p->sched_reset_on_fork;
7254 policy = oldpolicy = p->policy;
7256 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7258 if (!valid_policy(policy))
7262 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7266 * Valid priorities for SCHED_FIFO and SCHED_RR are
7267 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7268 * SCHED_BATCH and SCHED_IDLE is 0.
7270 if (attr->sched_priority > MAX_RT_PRIO-1)
7272 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7273 (rt_policy(policy) != (attr->sched_priority != 0)))
7277 * Allow unprivileged RT tasks to decrease priority:
7279 if (user && !capable(CAP_SYS_NICE)) {
7280 if (fair_policy(policy)) {
7281 if (attr->sched_nice < task_nice(p) &&
7282 !can_nice(p, attr->sched_nice))
7286 if (rt_policy(policy)) {
7287 unsigned long rlim_rtprio =
7288 task_rlimit(p, RLIMIT_RTPRIO);
7290 /* Can't set/change the rt policy: */
7291 if (policy != p->policy && !rlim_rtprio)
7294 /* Can't increase priority: */
7295 if (attr->sched_priority > p->rt_priority &&
7296 attr->sched_priority > rlim_rtprio)
7301 * Can't set/change SCHED_DEADLINE policy at all for now
7302 * (safest behavior); in the future we would like to allow
7303 * unprivileged DL tasks to increase their relative deadline
7304 * or reduce their runtime (both ways reducing utilization)
7306 if (dl_policy(policy))
7310 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7311 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7313 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7314 if (!can_nice(p, task_nice(p)))
7318 /* Can't change other user's priorities: */
7319 if (!check_same_owner(p))
7322 /* Normal users shall not reset the sched_reset_on_fork flag: */
7323 if (p->sched_reset_on_fork && !reset_on_fork)
7328 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7331 retval = security_task_setscheduler(p);
7336 /* Update task specific "requested" clamps */
7337 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7338 retval = uclamp_validate(p, attr);
7347 * Make sure no PI-waiters arrive (or leave) while we are
7348 * changing the priority of the task:
7350 * To be able to change p->policy safely, the appropriate
7351 * runqueue lock must be held.
7353 rq = task_rq_lock(p, &rf);
7354 update_rq_clock(rq);
7357 * Changing the policy of the stop threads its a very bad idea:
7359 if (p == rq->stop) {
7365 * If not changing anything there's no need to proceed further,
7366 * but store a possible modification of reset_on_fork.
7368 if (unlikely(policy == p->policy)) {
7369 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7371 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7373 if (dl_policy(policy) && dl_param_changed(p, attr))
7375 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7378 p->sched_reset_on_fork = reset_on_fork;
7385 #ifdef CONFIG_RT_GROUP_SCHED
7387 * Do not allow realtime tasks into groups that have no runtime
7390 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7391 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7392 !task_group_is_autogroup(task_group(p))) {
7398 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7399 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7400 cpumask_t *span = rq->rd->span;
7403 * Don't allow tasks with an affinity mask smaller than
7404 * the entire root_domain to become SCHED_DEADLINE. We
7405 * will also fail if there's no bandwidth available.
7407 if (!cpumask_subset(span, p->cpus_ptr) ||
7408 rq->rd->dl_bw.bw == 0) {
7416 /* Re-check policy now with rq lock held: */
7417 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7418 policy = oldpolicy = -1;
7419 task_rq_unlock(rq, p, &rf);
7421 cpuset_read_unlock();
7426 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7427 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7430 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7435 p->sched_reset_on_fork = reset_on_fork;
7438 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7441 * Take priority boosted tasks into account. If the new
7442 * effective priority is unchanged, we just store the new
7443 * normal parameters and do not touch the scheduler class and
7444 * the runqueue. This will be done when the task deboost
7447 newprio = rt_effective_prio(p, newprio);
7448 if (newprio == oldprio)
7449 queue_flags &= ~DEQUEUE_MOVE;
7452 queued = task_on_rq_queued(p);
7453 running = task_current(rq, p);
7455 dequeue_task(rq, p, queue_flags);
7457 put_prev_task(rq, p);
7459 prev_class = p->sched_class;
7461 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7462 __setscheduler_params(p, attr);
7463 __setscheduler_prio(p, newprio);
7465 __setscheduler_uclamp(p, attr);
7469 * We enqueue to tail when the priority of a task is
7470 * increased (user space view).
7472 if (oldprio < p->prio)
7473 queue_flags |= ENQUEUE_HEAD;
7475 enqueue_task(rq, p, queue_flags);
7478 set_next_task(rq, p);
7480 check_class_changed(rq, p, prev_class, oldprio);
7482 /* Avoid rq from going away on us: */
7484 head = splice_balance_callbacks(rq);
7485 task_rq_unlock(rq, p, &rf);
7488 cpuset_read_unlock();
7489 rt_mutex_adjust_pi(p);
7492 /* Run balance callbacks after we've adjusted the PI chain: */
7493 balance_callbacks(rq, head);
7499 task_rq_unlock(rq, p, &rf);
7501 cpuset_read_unlock();
7505 static int _sched_setscheduler(struct task_struct *p, int policy,
7506 const struct sched_param *param, bool check)
7508 struct sched_attr attr = {
7509 .sched_policy = policy,
7510 .sched_priority = param->sched_priority,
7511 .sched_nice = PRIO_TO_NICE(p->static_prio),
7514 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7515 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7516 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7517 policy &= ~SCHED_RESET_ON_FORK;
7518 attr.sched_policy = policy;
7521 return __sched_setscheduler(p, &attr, check, true);
7524 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7525 * @p: the task in question.
7526 * @policy: new policy.
7527 * @param: structure containing the new RT priority.
7529 * Use sched_set_fifo(), read its comment.
7531 * Return: 0 on success. An error code otherwise.
7533 * NOTE that the task may be already dead.
7535 int sched_setscheduler(struct task_struct *p, int policy,
7536 const struct sched_param *param)
7538 return _sched_setscheduler(p, policy, param, true);
7541 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7543 return __sched_setscheduler(p, attr, true, true);
7546 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7548 return __sched_setscheduler(p, attr, false, true);
7550 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7553 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7554 * @p: the task in question.
7555 * @policy: new policy.
7556 * @param: structure containing the new RT priority.
7558 * Just like sched_setscheduler, only don't bother checking if the
7559 * current context has permission. For example, this is needed in
7560 * stop_machine(): we create temporary high priority worker threads,
7561 * but our caller might not have that capability.
7563 * Return: 0 on success. An error code otherwise.
7565 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7566 const struct sched_param *param)
7568 return _sched_setscheduler(p, policy, param, false);
7572 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7573 * incapable of resource management, which is the one thing an OS really should
7576 * This is of course the reason it is limited to privileged users only.
7578 * Worse still; it is fundamentally impossible to compose static priority
7579 * workloads. You cannot take two correctly working static prio workloads
7580 * and smash them together and still expect them to work.
7582 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7586 * The administrator _MUST_ configure the system, the kernel simply doesn't
7587 * know enough information to make a sensible choice.
7589 void sched_set_fifo(struct task_struct *p)
7591 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7592 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7594 EXPORT_SYMBOL_GPL(sched_set_fifo);
7597 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7599 void sched_set_fifo_low(struct task_struct *p)
7601 struct sched_param sp = { .sched_priority = 1 };
7602 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7604 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7606 void sched_set_normal(struct task_struct *p, int nice)
7608 struct sched_attr attr = {
7609 .sched_policy = SCHED_NORMAL,
7612 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7614 EXPORT_SYMBOL_GPL(sched_set_normal);
7617 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7619 struct sched_param lparam;
7620 struct task_struct *p;
7623 if (!param || pid < 0)
7625 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7630 p = find_process_by_pid(pid);
7636 retval = sched_setscheduler(p, policy, &lparam);
7644 * Mimics kernel/events/core.c perf_copy_attr().
7646 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7651 /* Zero the full structure, so that a short copy will be nice: */
7652 memset(attr, 0, sizeof(*attr));
7654 ret = get_user(size, &uattr->size);
7658 /* ABI compatibility quirk: */
7660 size = SCHED_ATTR_SIZE_VER0;
7661 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7664 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7671 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7672 size < SCHED_ATTR_SIZE_VER1)
7676 * XXX: Do we want to be lenient like existing syscalls; or do we want
7677 * to be strict and return an error on out-of-bounds values?
7679 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7684 put_user(sizeof(*attr), &uattr->size);
7688 static void get_params(struct task_struct *p, struct sched_attr *attr)
7690 if (task_has_dl_policy(p))
7691 __getparam_dl(p, attr);
7692 else if (task_has_rt_policy(p))
7693 attr->sched_priority = p->rt_priority;
7695 attr->sched_nice = task_nice(p);
7699 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7700 * @pid: the pid in question.
7701 * @policy: new policy.
7702 * @param: structure containing the new RT priority.
7704 * Return: 0 on success. An error code otherwise.
7706 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7711 return do_sched_setscheduler(pid, policy, param);
7715 * sys_sched_setparam - set/change the RT priority of a thread
7716 * @pid: the pid in question.
7717 * @param: structure containing the new RT priority.
7719 * Return: 0 on success. An error code otherwise.
7721 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7723 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7727 * sys_sched_setattr - same as above, but with extended sched_attr
7728 * @pid: the pid in question.
7729 * @uattr: structure containing the extended parameters.
7730 * @flags: for future extension.
7732 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7733 unsigned int, flags)
7735 struct sched_attr attr;
7736 struct task_struct *p;
7739 if (!uattr || pid < 0 || flags)
7742 retval = sched_copy_attr(uattr, &attr);
7746 if ((int)attr.sched_policy < 0)
7748 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7749 attr.sched_policy = SETPARAM_POLICY;
7753 p = find_process_by_pid(pid);
7759 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7760 get_params(p, &attr);
7761 retval = sched_setattr(p, &attr);
7769 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7770 * @pid: the pid in question.
7772 * Return: On success, the policy of the thread. Otherwise, a negative error
7775 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7777 struct task_struct *p;
7785 p = find_process_by_pid(pid);
7787 retval = security_task_getscheduler(p);
7790 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7797 * sys_sched_getparam - get the RT priority of a thread
7798 * @pid: the pid in question.
7799 * @param: structure containing the RT priority.
7801 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7804 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7806 struct sched_param lp = { .sched_priority = 0 };
7807 struct task_struct *p;
7810 if (!param || pid < 0)
7814 p = find_process_by_pid(pid);
7819 retval = security_task_getscheduler(p);
7823 if (task_has_rt_policy(p))
7824 lp.sched_priority = p->rt_priority;
7828 * This one might sleep, we cannot do it with a spinlock held ...
7830 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7840 * Copy the kernel size attribute structure (which might be larger
7841 * than what user-space knows about) to user-space.
7843 * Note that all cases are valid: user-space buffer can be larger or
7844 * smaller than the kernel-space buffer. The usual case is that both
7845 * have the same size.
7848 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7849 struct sched_attr *kattr,
7852 unsigned int ksize = sizeof(*kattr);
7854 if (!access_ok(uattr, usize))
7858 * sched_getattr() ABI forwards and backwards compatibility:
7860 * If usize == ksize then we just copy everything to user-space and all is good.
7862 * If usize < ksize then we only copy as much as user-space has space for,
7863 * this keeps ABI compatibility as well. We skip the rest.
7865 * If usize > ksize then user-space is using a newer version of the ABI,
7866 * which part the kernel doesn't know about. Just ignore it - tooling can
7867 * detect the kernel's knowledge of attributes from the attr->size value
7868 * which is set to ksize in this case.
7870 kattr->size = min(usize, ksize);
7872 if (copy_to_user(uattr, kattr, kattr->size))
7879 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7880 * @pid: the pid in question.
7881 * @uattr: structure containing the extended parameters.
7882 * @usize: sizeof(attr) for fwd/bwd comp.
7883 * @flags: for future extension.
7885 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7886 unsigned int, usize, unsigned int, flags)
7888 struct sched_attr kattr = { };
7889 struct task_struct *p;
7892 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7893 usize < SCHED_ATTR_SIZE_VER0 || flags)
7897 p = find_process_by_pid(pid);
7902 retval = security_task_getscheduler(p);
7906 kattr.sched_policy = p->policy;
7907 if (p->sched_reset_on_fork)
7908 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7909 get_params(p, &kattr);
7910 kattr.sched_flags &= SCHED_FLAG_ALL;
7912 #ifdef CONFIG_UCLAMP_TASK
7914 * This could race with another potential updater, but this is fine
7915 * because it'll correctly read the old or the new value. We don't need
7916 * to guarantee who wins the race as long as it doesn't return garbage.
7918 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7919 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7924 return sched_attr_copy_to_user(uattr, &kattr, usize);
7932 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7937 * If the task isn't a deadline task or admission control is
7938 * disabled then we don't care about affinity changes.
7940 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7944 * Since bandwidth control happens on root_domain basis,
7945 * if admission test is enabled, we only admit -deadline
7946 * tasks allowed to run on all the CPUs in the task's
7950 if (!cpumask_subset(task_rq(p)->rd->span, mask))
7958 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
7961 cpumask_var_t cpus_allowed, new_mask;
7963 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
7966 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7968 goto out_free_cpus_allowed;
7971 cpuset_cpus_allowed(p, cpus_allowed);
7972 cpumask_and(new_mask, mask, cpus_allowed);
7974 retval = dl_task_check_affinity(p, new_mask);
7976 goto out_free_new_mask;
7978 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
7980 goto out_free_new_mask;
7982 cpuset_cpus_allowed(p, cpus_allowed);
7983 if (!cpumask_subset(new_mask, cpus_allowed)) {
7985 * We must have raced with a concurrent cpuset update.
7986 * Just reset the cpumask to the cpuset's cpus_allowed.
7988 cpumask_copy(new_mask, cpus_allowed);
7993 free_cpumask_var(new_mask);
7994 out_free_cpus_allowed:
7995 free_cpumask_var(cpus_allowed);
7999 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8001 struct task_struct *p;
8006 p = find_process_by_pid(pid);
8012 /* Prevent p going away */
8016 if (p->flags & PF_NO_SETAFFINITY) {
8021 if (!check_same_owner(p)) {
8023 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8031 retval = security_task_setscheduler(p);
8035 retval = __sched_setaffinity(p, in_mask);
8041 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
8042 struct cpumask *new_mask)
8044 if (len < cpumask_size())
8045 cpumask_clear(new_mask);
8046 else if (len > cpumask_size())
8047 len = cpumask_size();
8049 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8053 * sys_sched_setaffinity - set the CPU affinity of a process
8054 * @pid: pid of the process
8055 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8056 * @user_mask_ptr: user-space pointer to the new CPU mask
8058 * Return: 0 on success. An error code otherwise.
8060 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8061 unsigned long __user *, user_mask_ptr)
8063 cpumask_var_t new_mask;
8066 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8069 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8071 retval = sched_setaffinity(pid, new_mask);
8072 free_cpumask_var(new_mask);
8076 long sched_getaffinity(pid_t pid, struct cpumask *mask)
8078 struct task_struct *p;
8079 unsigned long flags;
8085 p = find_process_by_pid(pid);
8089 retval = security_task_getscheduler(p);
8093 raw_spin_lock_irqsave(&p->pi_lock, flags);
8094 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
8095 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
8104 * sys_sched_getaffinity - get the CPU affinity of a process
8105 * @pid: pid of the process
8106 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
8107 * @user_mask_ptr: user-space pointer to hold the current CPU mask
8109 * Return: size of CPU mask copied to user_mask_ptr on success. An
8110 * error code otherwise.
8112 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8113 unsigned long __user *, user_mask_ptr)
8118 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
8120 if (len & (sizeof(unsigned long)-1))
8123 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8126 ret = sched_getaffinity(pid, mask);
8128 unsigned int retlen = min(len, cpumask_size());
8130 if (copy_to_user(user_mask_ptr, mask, retlen))
8135 free_cpumask_var(mask);
8140 static void do_sched_yield(void)
8145 rq = this_rq_lock_irq(&rf);
8147 schedstat_inc(rq->yld_count);
8148 current->sched_class->yield_task(rq);
8151 rq_unlock_irq(rq, &rf);
8152 sched_preempt_enable_no_resched();
8158 * sys_sched_yield - yield the current processor to other threads.
8160 * This function yields the current CPU to other tasks. If there are no
8161 * other threads running on this CPU then this function will return.
8165 SYSCALL_DEFINE0(sched_yield)
8171 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8172 int __sched __cond_resched(void)
8174 if (should_resched(0)) {
8175 preempt_schedule_common();
8179 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8180 * whether the current CPU is in an RCU read-side critical section,
8181 * so the tick can report quiescent states even for CPUs looping
8182 * in kernel context. In contrast, in non-preemptible kernels,
8183 * RCU readers leave no in-memory hints, which means that CPU-bound
8184 * processes executing in kernel context might never report an
8185 * RCU quiescent state. Therefore, the following code causes
8186 * cond_resched() to report a quiescent state, but only when RCU
8187 * is in urgent need of one.
8189 #ifndef CONFIG_PREEMPT_RCU
8194 EXPORT_SYMBOL(__cond_resched);
8197 #ifdef CONFIG_PREEMPT_DYNAMIC
8198 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8199 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8201 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8202 EXPORT_STATIC_CALL_TRAMP(might_resched);
8206 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8207 * call schedule, and on return reacquire the lock.
8209 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8210 * operations here to prevent schedule() from being called twice (once via
8211 * spin_unlock(), once by hand).
8213 int __cond_resched_lock(spinlock_t *lock)
8215 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8218 lockdep_assert_held(lock);
8220 if (spin_needbreak(lock) || resched) {
8223 preempt_schedule_common();
8231 EXPORT_SYMBOL(__cond_resched_lock);
8233 int __cond_resched_rwlock_read(rwlock_t *lock)
8235 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8238 lockdep_assert_held_read(lock);
8240 if (rwlock_needbreak(lock) || resched) {
8243 preempt_schedule_common();
8251 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8253 int __cond_resched_rwlock_write(rwlock_t *lock)
8255 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8258 lockdep_assert_held_write(lock);
8260 if (rwlock_needbreak(lock) || resched) {
8263 preempt_schedule_common();
8271 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8274 * yield - yield the current processor to other threads.
8276 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8278 * The scheduler is at all times free to pick the calling task as the most
8279 * eligible task to run, if removing the yield() call from your code breaks
8280 * it, it's already broken.
8282 * Typical broken usage is:
8287 * where one assumes that yield() will let 'the other' process run that will
8288 * make event true. If the current task is a SCHED_FIFO task that will never
8289 * happen. Never use yield() as a progress guarantee!!
8291 * If you want to use yield() to wait for something, use wait_event().
8292 * If you want to use yield() to be 'nice' for others, use cond_resched().
8293 * If you still want to use yield(), do not!
8295 void __sched yield(void)
8297 set_current_state(TASK_RUNNING);
8300 EXPORT_SYMBOL(yield);
8303 * yield_to - yield the current processor to another thread in
8304 * your thread group, or accelerate that thread toward the
8305 * processor it's on.
8307 * @preempt: whether task preemption is allowed or not
8309 * It's the caller's job to ensure that the target task struct
8310 * can't go away on us before we can do any checks.
8313 * true (>0) if we indeed boosted the target task.
8314 * false (0) if we failed to boost the target.
8315 * -ESRCH if there's no task to yield to.
8317 int __sched yield_to(struct task_struct *p, bool preempt)
8319 struct task_struct *curr = current;
8320 struct rq *rq, *p_rq;
8321 unsigned long flags;
8324 local_irq_save(flags);
8330 * If we're the only runnable task on the rq and target rq also
8331 * has only one task, there's absolutely no point in yielding.
8333 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8338 double_rq_lock(rq, p_rq);
8339 if (task_rq(p) != p_rq) {
8340 double_rq_unlock(rq, p_rq);
8344 if (!curr->sched_class->yield_to_task)
8347 if (curr->sched_class != p->sched_class)
8350 if (task_running(p_rq, p) || !task_is_running(p))
8353 yielded = curr->sched_class->yield_to_task(rq, p);
8355 schedstat_inc(rq->yld_count);
8357 * Make p's CPU reschedule; pick_next_entity takes care of
8360 if (preempt && rq != p_rq)
8365 double_rq_unlock(rq, p_rq);
8367 local_irq_restore(flags);
8374 EXPORT_SYMBOL_GPL(yield_to);
8376 int io_schedule_prepare(void)
8378 int old_iowait = current->in_iowait;
8380 current->in_iowait = 1;
8382 blk_flush_plug(current->plug, true);
8387 void io_schedule_finish(int token)
8389 current->in_iowait = token;
8393 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8394 * that process accounting knows that this is a task in IO wait state.
8396 long __sched io_schedule_timeout(long timeout)
8401 token = io_schedule_prepare();
8402 ret = schedule_timeout(timeout);
8403 io_schedule_finish(token);
8407 EXPORT_SYMBOL(io_schedule_timeout);
8409 void __sched io_schedule(void)
8413 token = io_schedule_prepare();
8415 io_schedule_finish(token);
8417 EXPORT_SYMBOL(io_schedule);
8420 * sys_sched_get_priority_max - return maximum RT priority.
8421 * @policy: scheduling class.
8423 * Return: On success, this syscall returns the maximum
8424 * rt_priority that can be used by a given scheduling class.
8425 * On failure, a negative error code is returned.
8427 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8434 ret = MAX_RT_PRIO-1;
8436 case SCHED_DEADLINE:
8447 * sys_sched_get_priority_min - return minimum RT priority.
8448 * @policy: scheduling class.
8450 * Return: On success, this syscall returns the minimum
8451 * rt_priority that can be used by a given scheduling class.
8452 * On failure, a negative error code is returned.
8454 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8463 case SCHED_DEADLINE:
8472 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8474 struct task_struct *p;
8475 unsigned int time_slice;
8485 p = find_process_by_pid(pid);
8489 retval = security_task_getscheduler(p);
8493 rq = task_rq_lock(p, &rf);
8495 if (p->sched_class->get_rr_interval)
8496 time_slice = p->sched_class->get_rr_interval(rq, p);
8497 task_rq_unlock(rq, p, &rf);
8500 jiffies_to_timespec64(time_slice, t);
8509 * sys_sched_rr_get_interval - return the default timeslice of a process.
8510 * @pid: pid of the process.
8511 * @interval: userspace pointer to the timeslice value.
8513 * this syscall writes the default timeslice value of a given process
8514 * into the user-space timespec buffer. A value of '0' means infinity.
8516 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8519 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8520 struct __kernel_timespec __user *, interval)
8522 struct timespec64 t;
8523 int retval = sched_rr_get_interval(pid, &t);
8526 retval = put_timespec64(&t, interval);
8531 #ifdef CONFIG_COMPAT_32BIT_TIME
8532 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8533 struct old_timespec32 __user *, interval)
8535 struct timespec64 t;
8536 int retval = sched_rr_get_interval(pid, &t);
8539 retval = put_old_timespec32(&t, interval);
8544 void sched_show_task(struct task_struct *p)
8546 unsigned long free = 0;
8549 if (!try_get_task_stack(p))
8552 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8554 if (task_is_running(p))
8555 pr_cont(" running task ");
8556 #ifdef CONFIG_DEBUG_STACK_USAGE
8557 free = stack_not_used(p);
8562 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8564 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8565 free, task_pid_nr(p), ppid,
8566 read_task_thread_flags(p));
8568 print_worker_info(KERN_INFO, p);
8569 print_stop_info(KERN_INFO, p);
8570 show_stack(p, NULL, KERN_INFO);
8573 EXPORT_SYMBOL_GPL(sched_show_task);
8576 state_filter_match(unsigned long state_filter, struct task_struct *p)
8578 unsigned int state = READ_ONCE(p->__state);
8580 /* no filter, everything matches */
8584 /* filter, but doesn't match */
8585 if (!(state & state_filter))
8589 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8592 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8599 void show_state_filter(unsigned int state_filter)
8601 struct task_struct *g, *p;
8604 for_each_process_thread(g, p) {
8606 * reset the NMI-timeout, listing all files on a slow
8607 * console might take a lot of time:
8608 * Also, reset softlockup watchdogs on all CPUs, because
8609 * another CPU might be blocked waiting for us to process
8612 touch_nmi_watchdog();
8613 touch_all_softlockup_watchdogs();
8614 if (state_filter_match(state_filter, p))
8618 #ifdef CONFIG_SCHED_DEBUG
8620 sysrq_sched_debug_show();
8624 * Only show locks if all tasks are dumped:
8627 debug_show_all_locks();
8631 * init_idle - set up an idle thread for a given CPU
8632 * @idle: task in question
8633 * @cpu: CPU the idle task belongs to
8635 * NOTE: this function does not set the idle thread's NEED_RESCHED
8636 * flag, to make booting more robust.
8638 void __init init_idle(struct task_struct *idle, int cpu)
8640 struct rq *rq = cpu_rq(cpu);
8641 unsigned long flags;
8643 __sched_fork(0, idle);
8646 * The idle task doesn't need the kthread struct to function, but it
8647 * is dressed up as a per-CPU kthread and thus needs to play the part
8648 * if we want to avoid special-casing it in code that deals with per-CPU
8651 set_kthread_struct(idle);
8653 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8654 raw_spin_rq_lock(rq);
8656 idle->__state = TASK_RUNNING;
8657 idle->se.exec_start = sched_clock();
8659 * PF_KTHREAD should already be set at this point; regardless, make it
8660 * look like a proper per-CPU kthread.
8662 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8663 kthread_set_per_cpu(idle, cpu);
8667 * It's possible that init_idle() gets called multiple times on a task,
8668 * in that case do_set_cpus_allowed() will not do the right thing.
8670 * And since this is boot we can forgo the serialization.
8672 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8675 * We're having a chicken and egg problem, even though we are
8676 * holding rq->lock, the CPU isn't yet set to this CPU so the
8677 * lockdep check in task_group() will fail.
8679 * Similar case to sched_fork(). / Alternatively we could
8680 * use task_rq_lock() here and obtain the other rq->lock.
8685 __set_task_cpu(idle, cpu);
8689 rcu_assign_pointer(rq->curr, idle);
8690 idle->on_rq = TASK_ON_RQ_QUEUED;
8694 raw_spin_rq_unlock(rq);
8695 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8697 /* Set the preempt count _outside_ the spinlocks! */
8698 init_idle_preempt_count(idle, cpu);
8701 * The idle tasks have their own, simple scheduling class:
8703 idle->sched_class = &idle_sched_class;
8704 ftrace_graph_init_idle_task(idle, cpu);
8705 vtime_init_idle(idle, cpu);
8707 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8713 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8714 const struct cpumask *trial)
8718 if (!cpumask_weight(cur))
8721 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8726 int task_can_attach(struct task_struct *p,
8727 const struct cpumask *cs_cpus_allowed)
8732 * Kthreads which disallow setaffinity shouldn't be moved
8733 * to a new cpuset; we don't want to change their CPU
8734 * affinity and isolating such threads by their set of
8735 * allowed nodes is unnecessary. Thus, cpusets are not
8736 * applicable for such threads. This prevents checking for
8737 * success of set_cpus_allowed_ptr() on all attached tasks
8738 * before cpus_mask may be changed.
8740 if (p->flags & PF_NO_SETAFFINITY) {
8745 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8747 ret = dl_task_can_attach(p, cs_cpus_allowed);
8753 bool sched_smp_initialized __read_mostly;
8755 #ifdef CONFIG_NUMA_BALANCING
8756 /* Migrate current task p to target_cpu */
8757 int migrate_task_to(struct task_struct *p, int target_cpu)
8759 struct migration_arg arg = { p, target_cpu };
8760 int curr_cpu = task_cpu(p);
8762 if (curr_cpu == target_cpu)
8765 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8768 /* TODO: This is not properly updating schedstats */
8770 trace_sched_move_numa(p, curr_cpu, target_cpu);
8771 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8775 * Requeue a task on a given node and accurately track the number of NUMA
8776 * tasks on the runqueues
8778 void sched_setnuma(struct task_struct *p, int nid)
8780 bool queued, running;
8784 rq = task_rq_lock(p, &rf);
8785 queued = task_on_rq_queued(p);
8786 running = task_current(rq, p);
8789 dequeue_task(rq, p, DEQUEUE_SAVE);
8791 put_prev_task(rq, p);
8793 p->numa_preferred_nid = nid;
8796 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8798 set_next_task(rq, p);
8799 task_rq_unlock(rq, p, &rf);
8801 #endif /* CONFIG_NUMA_BALANCING */
8803 #ifdef CONFIG_HOTPLUG_CPU
8805 * Ensure that the idle task is using init_mm right before its CPU goes
8808 void idle_task_exit(void)
8810 struct mm_struct *mm = current->active_mm;
8812 BUG_ON(cpu_online(smp_processor_id()));
8813 BUG_ON(current != this_rq()->idle);
8815 if (mm != &init_mm) {
8816 switch_mm(mm, &init_mm, current);
8817 finish_arch_post_lock_switch();
8820 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8823 static int __balance_push_cpu_stop(void *arg)
8825 struct task_struct *p = arg;
8826 struct rq *rq = this_rq();
8830 raw_spin_lock_irq(&p->pi_lock);
8833 update_rq_clock(rq);
8835 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8836 cpu = select_fallback_rq(rq->cpu, p);
8837 rq = __migrate_task(rq, &rf, p, cpu);
8841 raw_spin_unlock_irq(&p->pi_lock);
8848 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8851 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8853 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8854 * effective when the hotplug motion is down.
8856 static void balance_push(struct rq *rq)
8858 struct task_struct *push_task = rq->curr;
8860 lockdep_assert_rq_held(rq);
8863 * Ensure the thing is persistent until balance_push_set(.on = false);
8865 rq->balance_callback = &balance_push_callback;
8868 * Only active while going offline and when invoked on the outgoing
8871 if (!cpu_dying(rq->cpu) || rq != this_rq())
8875 * Both the cpu-hotplug and stop task are in this case and are
8876 * required to complete the hotplug process.
8878 if (kthread_is_per_cpu(push_task) ||
8879 is_migration_disabled(push_task)) {
8882 * If this is the idle task on the outgoing CPU try to wake
8883 * up the hotplug control thread which might wait for the
8884 * last task to vanish. The rcuwait_active() check is
8885 * accurate here because the waiter is pinned on this CPU
8886 * and can't obviously be running in parallel.
8888 * On RT kernels this also has to check whether there are
8889 * pinned and scheduled out tasks on the runqueue. They
8890 * need to leave the migrate disabled section first.
8892 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8893 rcuwait_active(&rq->hotplug_wait)) {
8894 raw_spin_rq_unlock(rq);
8895 rcuwait_wake_up(&rq->hotplug_wait);
8896 raw_spin_rq_lock(rq);
8901 get_task_struct(push_task);
8903 * Temporarily drop rq->lock such that we can wake-up the stop task.
8904 * Both preemption and IRQs are still disabled.
8906 raw_spin_rq_unlock(rq);
8907 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8908 this_cpu_ptr(&push_work));
8910 * At this point need_resched() is true and we'll take the loop in
8911 * schedule(). The next pick is obviously going to be the stop task
8912 * which kthread_is_per_cpu() and will push this task away.
8914 raw_spin_rq_lock(rq);
8917 static void balance_push_set(int cpu, bool on)
8919 struct rq *rq = cpu_rq(cpu);
8922 rq_lock_irqsave(rq, &rf);
8924 WARN_ON_ONCE(rq->balance_callback);
8925 rq->balance_callback = &balance_push_callback;
8926 } else if (rq->balance_callback == &balance_push_callback) {
8927 rq->balance_callback = NULL;
8929 rq_unlock_irqrestore(rq, &rf);
8933 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8934 * inactive. All tasks which are not per CPU kernel threads are either
8935 * pushed off this CPU now via balance_push() or placed on a different CPU
8936 * during wakeup. Wait until the CPU is quiescent.
8938 static void balance_hotplug_wait(void)
8940 struct rq *rq = this_rq();
8942 rcuwait_wait_event(&rq->hotplug_wait,
8943 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8944 TASK_UNINTERRUPTIBLE);
8949 static inline void balance_push(struct rq *rq)
8953 static inline void balance_push_set(int cpu, bool on)
8957 static inline void balance_hotplug_wait(void)
8961 #endif /* CONFIG_HOTPLUG_CPU */
8963 void set_rq_online(struct rq *rq)
8966 const struct sched_class *class;
8968 cpumask_set_cpu(rq->cpu, rq->rd->online);
8971 for_each_class(class) {
8972 if (class->rq_online)
8973 class->rq_online(rq);
8978 void set_rq_offline(struct rq *rq)
8981 const struct sched_class *class;
8983 for_each_class(class) {
8984 if (class->rq_offline)
8985 class->rq_offline(rq);
8988 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8994 * used to mark begin/end of suspend/resume:
8996 static int num_cpus_frozen;
8999 * Update cpusets according to cpu_active mask. If cpusets are
9000 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9001 * around partition_sched_domains().
9003 * If we come here as part of a suspend/resume, don't touch cpusets because we
9004 * want to restore it back to its original state upon resume anyway.
9006 static void cpuset_cpu_active(void)
9008 if (cpuhp_tasks_frozen) {
9010 * num_cpus_frozen tracks how many CPUs are involved in suspend
9011 * resume sequence. As long as this is not the last online
9012 * operation in the resume sequence, just build a single sched
9013 * domain, ignoring cpusets.
9015 partition_sched_domains(1, NULL, NULL);
9016 if (--num_cpus_frozen)
9019 * This is the last CPU online operation. So fall through and
9020 * restore the original sched domains by considering the
9021 * cpuset configurations.
9023 cpuset_force_rebuild();
9025 cpuset_update_active_cpus();
9028 static int cpuset_cpu_inactive(unsigned int cpu)
9030 if (!cpuhp_tasks_frozen) {
9031 if (dl_cpu_busy(cpu))
9033 cpuset_update_active_cpus();
9036 partition_sched_domains(1, NULL, NULL);
9041 int sched_cpu_activate(unsigned int cpu)
9043 struct rq *rq = cpu_rq(cpu);
9047 * Clear the balance_push callback and prepare to schedule
9050 balance_push_set(cpu, false);
9052 #ifdef CONFIG_SCHED_SMT
9054 * When going up, increment the number of cores with SMT present.
9056 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9057 static_branch_inc_cpuslocked(&sched_smt_present);
9059 set_cpu_active(cpu, true);
9061 if (sched_smp_initialized) {
9062 sched_domains_numa_masks_set(cpu);
9063 cpuset_cpu_active();
9067 * Put the rq online, if not already. This happens:
9069 * 1) In the early boot process, because we build the real domains
9070 * after all CPUs have been brought up.
9072 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9075 rq_lock_irqsave(rq, &rf);
9077 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9080 rq_unlock_irqrestore(rq, &rf);
9085 int sched_cpu_deactivate(unsigned int cpu)
9087 struct rq *rq = cpu_rq(cpu);
9092 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9093 * load balancing when not active
9095 nohz_balance_exit_idle(rq);
9097 set_cpu_active(cpu, false);
9100 * From this point forward, this CPU will refuse to run any task that
9101 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9102 * push those tasks away until this gets cleared, see
9103 * sched_cpu_dying().
9105 balance_push_set(cpu, true);
9108 * We've cleared cpu_active_mask / set balance_push, wait for all
9109 * preempt-disabled and RCU users of this state to go away such that
9110 * all new such users will observe it.
9112 * Specifically, we rely on ttwu to no longer target this CPU, see
9113 * ttwu_queue_cond() and is_cpu_allowed().
9115 * Do sync before park smpboot threads to take care the rcu boost case.
9119 rq_lock_irqsave(rq, &rf);
9121 update_rq_clock(rq);
9122 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9125 rq_unlock_irqrestore(rq, &rf);
9127 #ifdef CONFIG_SCHED_SMT
9129 * When going down, decrement the number of cores with SMT present.
9131 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9132 static_branch_dec_cpuslocked(&sched_smt_present);
9134 sched_core_cpu_deactivate(cpu);
9137 if (!sched_smp_initialized)
9140 ret = cpuset_cpu_inactive(cpu);
9142 balance_push_set(cpu, false);
9143 set_cpu_active(cpu, true);
9146 sched_domains_numa_masks_clear(cpu);
9150 static void sched_rq_cpu_starting(unsigned int cpu)
9152 struct rq *rq = cpu_rq(cpu);
9154 rq->calc_load_update = calc_load_update;
9155 update_max_interval();
9158 int sched_cpu_starting(unsigned int cpu)
9160 sched_core_cpu_starting(cpu);
9161 sched_rq_cpu_starting(cpu);
9162 sched_tick_start(cpu);
9166 #ifdef CONFIG_HOTPLUG_CPU
9169 * Invoked immediately before the stopper thread is invoked to bring the
9170 * CPU down completely. At this point all per CPU kthreads except the
9171 * hotplug thread (current) and the stopper thread (inactive) have been
9172 * either parked or have been unbound from the outgoing CPU. Ensure that
9173 * any of those which might be on the way out are gone.
9175 * If after this point a bound task is being woken on this CPU then the
9176 * responsible hotplug callback has failed to do it's job.
9177 * sched_cpu_dying() will catch it with the appropriate fireworks.
9179 int sched_cpu_wait_empty(unsigned int cpu)
9181 balance_hotplug_wait();
9186 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9187 * might have. Called from the CPU stopper task after ensuring that the
9188 * stopper is the last running task on the CPU, so nr_active count is
9189 * stable. We need to take the teardown thread which is calling this into
9190 * account, so we hand in adjust = 1 to the load calculation.
9192 * Also see the comment "Global load-average calculations".
9194 static void calc_load_migrate(struct rq *rq)
9196 long delta = calc_load_fold_active(rq, 1);
9199 atomic_long_add(delta, &calc_load_tasks);
9202 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9204 struct task_struct *g, *p;
9205 int cpu = cpu_of(rq);
9207 lockdep_assert_rq_held(rq);
9209 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9210 for_each_process_thread(g, p) {
9211 if (task_cpu(p) != cpu)
9214 if (!task_on_rq_queued(p))
9217 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9221 int sched_cpu_dying(unsigned int cpu)
9223 struct rq *rq = cpu_rq(cpu);
9226 /* Handle pending wakeups and then migrate everything off */
9227 sched_tick_stop(cpu);
9229 rq_lock_irqsave(rq, &rf);
9230 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9231 WARN(true, "Dying CPU not properly vacated!");
9232 dump_rq_tasks(rq, KERN_WARNING);
9234 rq_unlock_irqrestore(rq, &rf);
9236 calc_load_migrate(rq);
9237 update_max_interval();
9239 sched_core_cpu_dying(cpu);
9244 void __init sched_init_smp(void)
9249 * There's no userspace yet to cause hotplug operations; hence all the
9250 * CPU masks are stable and all blatant races in the below code cannot
9253 mutex_lock(&sched_domains_mutex);
9254 sched_init_domains(cpu_active_mask);
9255 mutex_unlock(&sched_domains_mutex);
9257 /* Move init over to a non-isolated CPU */
9258 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9260 current->flags &= ~PF_NO_SETAFFINITY;
9261 sched_init_granularity();
9263 init_sched_rt_class();
9264 init_sched_dl_class();
9266 sched_smp_initialized = true;
9269 static int __init migration_init(void)
9271 sched_cpu_starting(smp_processor_id());
9274 early_initcall(migration_init);
9277 void __init sched_init_smp(void)
9279 sched_init_granularity();
9281 #endif /* CONFIG_SMP */
9283 int in_sched_functions(unsigned long addr)
9285 return in_lock_functions(addr) ||
9286 (addr >= (unsigned long)__sched_text_start
9287 && addr < (unsigned long)__sched_text_end);
9290 #ifdef CONFIG_CGROUP_SCHED
9292 * Default task group.
9293 * Every task in system belongs to this group at bootup.
9295 struct task_group root_task_group;
9296 LIST_HEAD(task_groups);
9298 /* Cacheline aligned slab cache for task_group */
9299 static struct kmem_cache *task_group_cache __read_mostly;
9302 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9303 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9305 void __init sched_init(void)
9307 unsigned long ptr = 0;
9310 /* Make sure the linker didn't screw up */
9311 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9312 &fair_sched_class + 1 != &rt_sched_class ||
9313 &rt_sched_class + 1 != &dl_sched_class);
9315 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9320 #ifdef CONFIG_FAIR_GROUP_SCHED
9321 ptr += 2 * nr_cpu_ids * sizeof(void **);
9323 #ifdef CONFIG_RT_GROUP_SCHED
9324 ptr += 2 * nr_cpu_ids * sizeof(void **);
9327 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9329 #ifdef CONFIG_FAIR_GROUP_SCHED
9330 root_task_group.se = (struct sched_entity **)ptr;
9331 ptr += nr_cpu_ids * sizeof(void **);
9333 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9334 ptr += nr_cpu_ids * sizeof(void **);
9336 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9337 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9338 #endif /* CONFIG_FAIR_GROUP_SCHED */
9339 #ifdef CONFIG_RT_GROUP_SCHED
9340 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9341 ptr += nr_cpu_ids * sizeof(void **);
9343 root_task_group.rt_rq = (struct rt_rq **)ptr;
9344 ptr += nr_cpu_ids * sizeof(void **);
9346 #endif /* CONFIG_RT_GROUP_SCHED */
9348 #ifdef CONFIG_CPUMASK_OFFSTACK
9349 for_each_possible_cpu(i) {
9350 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9351 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9352 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9353 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9355 #endif /* CONFIG_CPUMASK_OFFSTACK */
9357 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9358 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9361 init_defrootdomain();
9364 #ifdef CONFIG_RT_GROUP_SCHED
9365 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9366 global_rt_period(), global_rt_runtime());
9367 #endif /* CONFIG_RT_GROUP_SCHED */
9369 #ifdef CONFIG_CGROUP_SCHED
9370 task_group_cache = KMEM_CACHE(task_group, 0);
9372 list_add(&root_task_group.list, &task_groups);
9373 INIT_LIST_HEAD(&root_task_group.children);
9374 INIT_LIST_HEAD(&root_task_group.siblings);
9375 autogroup_init(&init_task);
9376 #endif /* CONFIG_CGROUP_SCHED */
9378 for_each_possible_cpu(i) {
9382 raw_spin_lock_init(&rq->__lock);
9384 rq->calc_load_active = 0;
9385 rq->calc_load_update = jiffies + LOAD_FREQ;
9386 init_cfs_rq(&rq->cfs);
9387 init_rt_rq(&rq->rt);
9388 init_dl_rq(&rq->dl);
9389 #ifdef CONFIG_FAIR_GROUP_SCHED
9390 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9391 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9393 * How much CPU bandwidth does root_task_group get?
9395 * In case of task-groups formed thr' the cgroup filesystem, it
9396 * gets 100% of the CPU resources in the system. This overall
9397 * system CPU resource is divided among the tasks of
9398 * root_task_group and its child task-groups in a fair manner,
9399 * based on each entity's (task or task-group's) weight
9400 * (se->load.weight).
9402 * In other words, if root_task_group has 10 tasks of weight
9403 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9404 * then A0's share of the CPU resource is:
9406 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9408 * We achieve this by letting root_task_group's tasks sit
9409 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9411 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9412 #endif /* CONFIG_FAIR_GROUP_SCHED */
9414 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9415 #ifdef CONFIG_RT_GROUP_SCHED
9416 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9421 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9422 rq->balance_callback = &balance_push_callback;
9423 rq->active_balance = 0;
9424 rq->next_balance = jiffies;
9429 rq->avg_idle = 2*sysctl_sched_migration_cost;
9430 rq->wake_stamp = jiffies;
9431 rq->wake_avg_idle = rq->avg_idle;
9432 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9434 INIT_LIST_HEAD(&rq->cfs_tasks);
9436 rq_attach_root(rq, &def_root_domain);
9437 #ifdef CONFIG_NO_HZ_COMMON
9438 rq->last_blocked_load_update_tick = jiffies;
9439 atomic_set(&rq->nohz_flags, 0);
9441 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9443 #ifdef CONFIG_HOTPLUG_CPU
9444 rcuwait_init(&rq->hotplug_wait);
9446 #endif /* CONFIG_SMP */
9448 atomic_set(&rq->nr_iowait, 0);
9450 #ifdef CONFIG_SCHED_CORE
9452 rq->core_pick = NULL;
9453 rq->core_enabled = 0;
9454 rq->core_tree = RB_ROOT;
9455 rq->core_forceidle_count = 0;
9456 rq->core_forceidle_occupation = 0;
9457 rq->core_forceidle_start = 0;
9459 rq->core_cookie = 0UL;
9463 set_load_weight(&init_task, false);
9466 * The boot idle thread does lazy MMU switching as well:
9469 enter_lazy_tlb(&init_mm, current);
9472 * Make us the idle thread. Technically, schedule() should not be
9473 * called from this thread, however somewhere below it might be,
9474 * but because we are the idle thread, we just pick up running again
9475 * when this runqueue becomes "idle".
9477 init_idle(current, smp_processor_id());
9479 calc_load_update = jiffies + LOAD_FREQ;
9482 idle_thread_set_boot_cpu();
9483 balance_push_set(smp_processor_id(), false);
9485 init_sched_fair_class();
9491 preempt_dynamic_init();
9493 scheduler_running = 1;
9496 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9498 void __might_sleep(const char *file, int line)
9500 unsigned int state = get_current_state();
9502 * Blocking primitives will set (and therefore destroy) current->state,
9503 * since we will exit with TASK_RUNNING make sure we enter with it,
9504 * otherwise we will destroy state.
9506 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9507 "do not call blocking ops when !TASK_RUNNING; "
9508 "state=%x set at [<%p>] %pS\n", state,
9509 (void *)current->task_state_change,
9510 (void *)current->task_state_change);
9512 __might_resched(file, line, 0);
9514 EXPORT_SYMBOL(__might_sleep);
9516 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9518 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9521 if (preempt_count() == preempt_offset)
9524 pr_err("Preemption disabled at:");
9525 print_ip_sym(KERN_ERR, ip);
9528 static inline bool resched_offsets_ok(unsigned int offsets)
9530 unsigned int nested = preempt_count();
9532 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9534 return nested == offsets;
9537 void __might_resched(const char *file, int line, unsigned int offsets)
9539 /* Ratelimiting timestamp: */
9540 static unsigned long prev_jiffy;
9542 unsigned long preempt_disable_ip;
9544 /* WARN_ON_ONCE() by default, no rate limit required: */
9547 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9548 !is_idle_task(current) && !current->non_block_count) ||
9549 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9553 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9555 prev_jiffy = jiffies;
9557 /* Save this before calling printk(), since that will clobber it: */
9558 preempt_disable_ip = get_preempt_disable_ip(current);
9560 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9562 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9563 in_atomic(), irqs_disabled(), current->non_block_count,
9564 current->pid, current->comm);
9565 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9566 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9568 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9569 pr_err("RCU nest depth: %d, expected: %u\n",
9570 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9573 if (task_stack_end_corrupted(current))
9574 pr_emerg("Thread overran stack, or stack corrupted\n");
9576 debug_show_held_locks(current);
9577 if (irqs_disabled())
9578 print_irqtrace_events(current);
9580 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9581 preempt_disable_ip);
9584 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9586 EXPORT_SYMBOL(__might_resched);
9588 void __cant_sleep(const char *file, int line, int preempt_offset)
9590 static unsigned long prev_jiffy;
9592 if (irqs_disabled())
9595 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9598 if (preempt_count() > preempt_offset)
9601 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9603 prev_jiffy = jiffies;
9605 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9606 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9607 in_atomic(), irqs_disabled(),
9608 current->pid, current->comm);
9610 debug_show_held_locks(current);
9612 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9614 EXPORT_SYMBOL_GPL(__cant_sleep);
9617 void __cant_migrate(const char *file, int line)
9619 static unsigned long prev_jiffy;
9621 if (irqs_disabled())
9624 if (is_migration_disabled(current))
9627 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9630 if (preempt_count() > 0)
9633 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9635 prev_jiffy = jiffies;
9637 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9638 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9639 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9640 current->pid, current->comm);
9642 debug_show_held_locks(current);
9644 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9646 EXPORT_SYMBOL_GPL(__cant_migrate);
9650 #ifdef CONFIG_MAGIC_SYSRQ
9651 void normalize_rt_tasks(void)
9653 struct task_struct *g, *p;
9654 struct sched_attr attr = {
9655 .sched_policy = SCHED_NORMAL,
9658 read_lock(&tasklist_lock);
9659 for_each_process_thread(g, p) {
9661 * Only normalize user tasks:
9663 if (p->flags & PF_KTHREAD)
9666 p->se.exec_start = 0;
9667 schedstat_set(p->stats.wait_start, 0);
9668 schedstat_set(p->stats.sleep_start, 0);
9669 schedstat_set(p->stats.block_start, 0);
9671 if (!dl_task(p) && !rt_task(p)) {
9673 * Renice negative nice level userspace
9676 if (task_nice(p) < 0)
9677 set_user_nice(p, 0);
9681 __sched_setscheduler(p, &attr, false, false);
9683 read_unlock(&tasklist_lock);
9686 #endif /* CONFIG_MAGIC_SYSRQ */
9688 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9690 * These functions are only useful for the IA64 MCA handling, or kdb.
9692 * They can only be called when the whole system has been
9693 * stopped - every CPU needs to be quiescent, and no scheduling
9694 * activity can take place. Using them for anything else would
9695 * be a serious bug, and as a result, they aren't even visible
9696 * under any other configuration.
9700 * curr_task - return the current task for a given CPU.
9701 * @cpu: the processor in question.
9703 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9705 * Return: The current task for @cpu.
9707 struct task_struct *curr_task(int cpu)
9709 return cpu_curr(cpu);
9712 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9716 * ia64_set_curr_task - set the current task for a given CPU.
9717 * @cpu: the processor in question.
9718 * @p: the task pointer to set.
9720 * Description: This function must only be used when non-maskable interrupts
9721 * are serviced on a separate stack. It allows the architecture to switch the
9722 * notion of the current task on a CPU in a non-blocking manner. This function
9723 * must be called with all CPU's synchronized, and interrupts disabled, the
9724 * and caller must save the original value of the current task (see
9725 * curr_task() above) and restore that value before reenabling interrupts and
9726 * re-starting the system.
9728 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9730 void ia64_set_curr_task(int cpu, struct task_struct *p)
9737 #ifdef CONFIG_CGROUP_SCHED
9738 /* task_group_lock serializes the addition/removal of task groups */
9739 static DEFINE_SPINLOCK(task_group_lock);
9741 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9742 struct task_group *parent)
9744 #ifdef CONFIG_UCLAMP_TASK_GROUP
9745 enum uclamp_id clamp_id;
9747 for_each_clamp_id(clamp_id) {
9748 uclamp_se_set(&tg->uclamp_req[clamp_id],
9749 uclamp_none(clamp_id), false);
9750 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9755 static void sched_free_group(struct task_group *tg)
9757 free_fair_sched_group(tg);
9758 free_rt_sched_group(tg);
9760 kmem_cache_free(task_group_cache, tg);
9763 static void sched_free_group_rcu(struct rcu_head *rcu)
9765 sched_free_group(container_of(rcu, struct task_group, rcu));
9768 static void sched_unregister_group(struct task_group *tg)
9770 unregister_fair_sched_group(tg);
9771 unregister_rt_sched_group(tg);
9773 * We have to wait for yet another RCU grace period to expire, as
9774 * print_cfs_stats() might run concurrently.
9776 call_rcu(&tg->rcu, sched_free_group_rcu);
9779 /* allocate runqueue etc for a new task group */
9780 struct task_group *sched_create_group(struct task_group *parent)
9782 struct task_group *tg;
9784 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9786 return ERR_PTR(-ENOMEM);
9788 if (!alloc_fair_sched_group(tg, parent))
9791 if (!alloc_rt_sched_group(tg, parent))
9794 alloc_uclamp_sched_group(tg, parent);
9799 sched_free_group(tg);
9800 return ERR_PTR(-ENOMEM);
9803 void sched_online_group(struct task_group *tg, struct task_group *parent)
9805 unsigned long flags;
9807 spin_lock_irqsave(&task_group_lock, flags);
9808 list_add_rcu(&tg->list, &task_groups);
9810 /* Root should already exist: */
9813 tg->parent = parent;
9814 INIT_LIST_HEAD(&tg->children);
9815 list_add_rcu(&tg->siblings, &parent->children);
9816 spin_unlock_irqrestore(&task_group_lock, flags);
9818 online_fair_sched_group(tg);
9821 /* rcu callback to free various structures associated with a task group */
9822 static void sched_unregister_group_rcu(struct rcu_head *rhp)
9824 /* Now it should be safe to free those cfs_rqs: */
9825 sched_unregister_group(container_of(rhp, struct task_group, rcu));
9828 void sched_destroy_group(struct task_group *tg)
9830 /* Wait for possible concurrent references to cfs_rqs complete: */
9831 call_rcu(&tg->rcu, sched_unregister_group_rcu);
9834 void sched_release_group(struct task_group *tg)
9836 unsigned long flags;
9839 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
9840 * sched_cfs_period_timer()).
9842 * For this to be effective, we have to wait for all pending users of
9843 * this task group to leave their RCU critical section to ensure no new
9844 * user will see our dying task group any more. Specifically ensure
9845 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
9847 * We therefore defer calling unregister_fair_sched_group() to
9848 * sched_unregister_group() which is guarantied to get called only after the
9849 * current RCU grace period has expired.
9851 spin_lock_irqsave(&task_group_lock, flags);
9852 list_del_rcu(&tg->list);
9853 list_del_rcu(&tg->siblings);
9854 spin_unlock_irqrestore(&task_group_lock, flags);
9857 static void sched_change_group(struct task_struct *tsk, int type)
9859 struct task_group *tg;
9862 * All callers are synchronized by task_rq_lock(); we do not use RCU
9863 * which is pointless here. Thus, we pass "true" to task_css_check()
9864 * to prevent lockdep warnings.
9866 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9867 struct task_group, css);
9868 tg = autogroup_task_group(tsk, tg);
9869 tsk->sched_task_group = tg;
9871 #ifdef CONFIG_FAIR_GROUP_SCHED
9872 if (tsk->sched_class->task_change_group)
9873 tsk->sched_class->task_change_group(tsk, type);
9876 set_task_rq(tsk, task_cpu(tsk));
9880 * Change task's runqueue when it moves between groups.
9882 * The caller of this function should have put the task in its new group by
9883 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9886 void sched_move_task(struct task_struct *tsk)
9888 int queued, running, queue_flags =
9889 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9893 rq = task_rq_lock(tsk, &rf);
9894 update_rq_clock(rq);
9896 running = task_current(rq, tsk);
9897 queued = task_on_rq_queued(tsk);
9900 dequeue_task(rq, tsk, queue_flags);
9902 put_prev_task(rq, tsk);
9904 sched_change_group(tsk, TASK_MOVE_GROUP);
9907 enqueue_task(rq, tsk, queue_flags);
9909 set_next_task(rq, tsk);
9911 * After changing group, the running task may have joined a
9912 * throttled one but it's still the running task. Trigger a
9913 * resched to make sure that task can still run.
9918 task_rq_unlock(rq, tsk, &rf);
9921 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9923 return css ? container_of(css, struct task_group, css) : NULL;
9926 static struct cgroup_subsys_state *
9927 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9929 struct task_group *parent = css_tg(parent_css);
9930 struct task_group *tg;
9933 /* This is early initialization for the top cgroup */
9934 return &root_task_group.css;
9937 tg = sched_create_group(parent);
9939 return ERR_PTR(-ENOMEM);
9944 /* Expose task group only after completing cgroup initialization */
9945 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9947 struct task_group *tg = css_tg(css);
9948 struct task_group *parent = css_tg(css->parent);
9951 sched_online_group(tg, parent);
9953 #ifdef CONFIG_UCLAMP_TASK_GROUP
9954 /* Propagate the effective uclamp value for the new group */
9955 mutex_lock(&uclamp_mutex);
9957 cpu_util_update_eff(css);
9959 mutex_unlock(&uclamp_mutex);
9965 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9967 struct task_group *tg = css_tg(css);
9969 sched_release_group(tg);
9972 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9974 struct task_group *tg = css_tg(css);
9977 * Relies on the RCU grace period between css_released() and this.
9979 sched_unregister_group(tg);
9983 * This is called before wake_up_new_task(), therefore we really only
9984 * have to set its group bits, all the other stuff does not apply.
9986 static void cpu_cgroup_fork(struct task_struct *task)
9991 rq = task_rq_lock(task, &rf);
9993 update_rq_clock(rq);
9994 sched_change_group(task, TASK_SET_GROUP);
9996 task_rq_unlock(rq, task, &rf);
9999 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
10001 struct task_struct *task;
10002 struct cgroup_subsys_state *css;
10005 cgroup_taskset_for_each(task, css, tset) {
10006 #ifdef CONFIG_RT_GROUP_SCHED
10007 if (!sched_rt_can_attach(css_tg(css), task))
10011 * Serialize against wake_up_new_task() such that if it's
10012 * running, we're sure to observe its full state.
10014 raw_spin_lock_irq(&task->pi_lock);
10016 * Avoid calling sched_move_task() before wake_up_new_task()
10017 * has happened. This would lead to problems with PELT, due to
10018 * move wanting to detach+attach while we're not attached yet.
10020 if (READ_ONCE(task->__state) == TASK_NEW)
10022 raw_spin_unlock_irq(&task->pi_lock);
10030 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10032 struct task_struct *task;
10033 struct cgroup_subsys_state *css;
10035 cgroup_taskset_for_each(task, css, tset)
10036 sched_move_task(task);
10039 #ifdef CONFIG_UCLAMP_TASK_GROUP
10040 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10042 struct cgroup_subsys_state *top_css = css;
10043 struct uclamp_se *uc_parent = NULL;
10044 struct uclamp_se *uc_se = NULL;
10045 unsigned int eff[UCLAMP_CNT];
10046 enum uclamp_id clamp_id;
10047 unsigned int clamps;
10049 lockdep_assert_held(&uclamp_mutex);
10050 SCHED_WARN_ON(!rcu_read_lock_held());
10052 css_for_each_descendant_pre(css, top_css) {
10053 uc_parent = css_tg(css)->parent
10054 ? css_tg(css)->parent->uclamp : NULL;
10056 for_each_clamp_id(clamp_id) {
10057 /* Assume effective clamps matches requested clamps */
10058 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10059 /* Cap effective clamps with parent's effective clamps */
10061 eff[clamp_id] > uc_parent[clamp_id].value) {
10062 eff[clamp_id] = uc_parent[clamp_id].value;
10065 /* Ensure protection is always capped by limit */
10066 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10068 /* Propagate most restrictive effective clamps */
10070 uc_se = css_tg(css)->uclamp;
10071 for_each_clamp_id(clamp_id) {
10072 if (eff[clamp_id] == uc_se[clamp_id].value)
10074 uc_se[clamp_id].value = eff[clamp_id];
10075 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10076 clamps |= (0x1 << clamp_id);
10079 css = css_rightmost_descendant(css);
10083 /* Immediately update descendants RUNNABLE tasks */
10084 uclamp_update_active_tasks(css);
10089 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10090 * C expression. Since there is no way to convert a macro argument (N) into a
10091 * character constant, use two levels of macros.
10093 #define _POW10(exp) ((unsigned int)1e##exp)
10094 #define POW10(exp) _POW10(exp)
10096 struct uclamp_request {
10097 #define UCLAMP_PERCENT_SHIFT 2
10098 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10104 static inline struct uclamp_request
10105 capacity_from_percent(char *buf)
10107 struct uclamp_request req = {
10108 .percent = UCLAMP_PERCENT_SCALE,
10109 .util = SCHED_CAPACITY_SCALE,
10114 if (strcmp(buf, "max")) {
10115 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10119 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10124 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10125 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10131 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10132 size_t nbytes, loff_t off,
10133 enum uclamp_id clamp_id)
10135 struct uclamp_request req;
10136 struct task_group *tg;
10138 req = capacity_from_percent(buf);
10142 static_branch_enable(&sched_uclamp_used);
10144 mutex_lock(&uclamp_mutex);
10147 tg = css_tg(of_css(of));
10148 if (tg->uclamp_req[clamp_id].value != req.util)
10149 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10152 * Because of not recoverable conversion rounding we keep track of the
10153 * exact requested value
10155 tg->uclamp_pct[clamp_id] = req.percent;
10157 /* Update effective clamps to track the most restrictive value */
10158 cpu_util_update_eff(of_css(of));
10161 mutex_unlock(&uclamp_mutex);
10166 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10167 char *buf, size_t nbytes,
10170 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10173 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10174 char *buf, size_t nbytes,
10177 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10180 static inline void cpu_uclamp_print(struct seq_file *sf,
10181 enum uclamp_id clamp_id)
10183 struct task_group *tg;
10189 tg = css_tg(seq_css(sf));
10190 util_clamp = tg->uclamp_req[clamp_id].value;
10193 if (util_clamp == SCHED_CAPACITY_SCALE) {
10194 seq_puts(sf, "max\n");
10198 percent = tg->uclamp_pct[clamp_id];
10199 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10200 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10203 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10205 cpu_uclamp_print(sf, UCLAMP_MIN);
10209 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10211 cpu_uclamp_print(sf, UCLAMP_MAX);
10214 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10216 #ifdef CONFIG_FAIR_GROUP_SCHED
10217 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10218 struct cftype *cftype, u64 shareval)
10220 if (shareval > scale_load_down(ULONG_MAX))
10221 shareval = MAX_SHARES;
10222 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10225 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10226 struct cftype *cft)
10228 struct task_group *tg = css_tg(css);
10230 return (u64) scale_load_down(tg->shares);
10233 #ifdef CONFIG_CFS_BANDWIDTH
10234 static DEFINE_MUTEX(cfs_constraints_mutex);
10236 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10237 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10238 /* More than 203 days if BW_SHIFT equals 20. */
10239 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10241 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10243 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10246 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10247 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10249 if (tg == &root_task_group)
10253 * Ensure we have at some amount of bandwidth every period. This is
10254 * to prevent reaching a state of large arrears when throttled via
10255 * entity_tick() resulting in prolonged exit starvation.
10257 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10261 * Likewise, bound things on the other side by preventing insane quota
10262 * periods. This also allows us to normalize in computing quota
10265 if (period > max_cfs_quota_period)
10269 * Bound quota to defend quota against overflow during bandwidth shift.
10271 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10274 if (quota != RUNTIME_INF && (burst > quota ||
10275 burst + quota > max_cfs_runtime))
10279 * Prevent race between setting of cfs_rq->runtime_enabled and
10280 * unthrottle_offline_cfs_rqs().
10283 mutex_lock(&cfs_constraints_mutex);
10284 ret = __cfs_schedulable(tg, period, quota);
10288 runtime_enabled = quota != RUNTIME_INF;
10289 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10291 * If we need to toggle cfs_bandwidth_used, off->on must occur
10292 * before making related changes, and on->off must occur afterwards
10294 if (runtime_enabled && !runtime_was_enabled)
10295 cfs_bandwidth_usage_inc();
10296 raw_spin_lock_irq(&cfs_b->lock);
10297 cfs_b->period = ns_to_ktime(period);
10298 cfs_b->quota = quota;
10299 cfs_b->burst = burst;
10301 __refill_cfs_bandwidth_runtime(cfs_b);
10303 /* Restart the period timer (if active) to handle new period expiry: */
10304 if (runtime_enabled)
10305 start_cfs_bandwidth(cfs_b);
10307 raw_spin_unlock_irq(&cfs_b->lock);
10309 for_each_online_cpu(i) {
10310 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10311 struct rq *rq = cfs_rq->rq;
10312 struct rq_flags rf;
10314 rq_lock_irq(rq, &rf);
10315 cfs_rq->runtime_enabled = runtime_enabled;
10316 cfs_rq->runtime_remaining = 0;
10318 if (cfs_rq->throttled)
10319 unthrottle_cfs_rq(cfs_rq);
10320 rq_unlock_irq(rq, &rf);
10322 if (runtime_was_enabled && !runtime_enabled)
10323 cfs_bandwidth_usage_dec();
10325 mutex_unlock(&cfs_constraints_mutex);
10326 cpus_read_unlock();
10331 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10333 u64 quota, period, burst;
10335 period = ktime_to_ns(tg->cfs_bandwidth.period);
10336 burst = tg->cfs_bandwidth.burst;
10337 if (cfs_quota_us < 0)
10338 quota = RUNTIME_INF;
10339 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10340 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10344 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10347 static long tg_get_cfs_quota(struct task_group *tg)
10351 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10354 quota_us = tg->cfs_bandwidth.quota;
10355 do_div(quota_us, NSEC_PER_USEC);
10360 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10362 u64 quota, period, burst;
10364 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10367 period = (u64)cfs_period_us * NSEC_PER_USEC;
10368 quota = tg->cfs_bandwidth.quota;
10369 burst = tg->cfs_bandwidth.burst;
10371 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10374 static long tg_get_cfs_period(struct task_group *tg)
10378 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10379 do_div(cfs_period_us, NSEC_PER_USEC);
10381 return cfs_period_us;
10384 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10386 u64 quota, period, burst;
10388 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10391 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10392 period = ktime_to_ns(tg->cfs_bandwidth.period);
10393 quota = tg->cfs_bandwidth.quota;
10395 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10398 static long tg_get_cfs_burst(struct task_group *tg)
10402 burst_us = tg->cfs_bandwidth.burst;
10403 do_div(burst_us, NSEC_PER_USEC);
10408 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10409 struct cftype *cft)
10411 return tg_get_cfs_quota(css_tg(css));
10414 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10415 struct cftype *cftype, s64 cfs_quota_us)
10417 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10420 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10421 struct cftype *cft)
10423 return tg_get_cfs_period(css_tg(css));
10426 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10427 struct cftype *cftype, u64 cfs_period_us)
10429 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10432 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10433 struct cftype *cft)
10435 return tg_get_cfs_burst(css_tg(css));
10438 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10439 struct cftype *cftype, u64 cfs_burst_us)
10441 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10444 struct cfs_schedulable_data {
10445 struct task_group *tg;
10450 * normalize group quota/period to be quota/max_period
10451 * note: units are usecs
10453 static u64 normalize_cfs_quota(struct task_group *tg,
10454 struct cfs_schedulable_data *d)
10459 period = d->period;
10462 period = tg_get_cfs_period(tg);
10463 quota = tg_get_cfs_quota(tg);
10466 /* note: these should typically be equivalent */
10467 if (quota == RUNTIME_INF || quota == -1)
10468 return RUNTIME_INF;
10470 return to_ratio(period, quota);
10473 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10475 struct cfs_schedulable_data *d = data;
10476 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10477 s64 quota = 0, parent_quota = -1;
10480 quota = RUNTIME_INF;
10482 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10484 quota = normalize_cfs_quota(tg, d);
10485 parent_quota = parent_b->hierarchical_quota;
10488 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10489 * always take the min. On cgroup1, only inherit when no
10492 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10493 quota = min(quota, parent_quota);
10495 if (quota == RUNTIME_INF)
10496 quota = parent_quota;
10497 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10501 cfs_b->hierarchical_quota = quota;
10506 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10509 struct cfs_schedulable_data data = {
10515 if (quota != RUNTIME_INF) {
10516 do_div(data.period, NSEC_PER_USEC);
10517 do_div(data.quota, NSEC_PER_USEC);
10521 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10527 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10529 struct task_group *tg = css_tg(seq_css(sf));
10530 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10532 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10533 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10534 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10536 if (schedstat_enabled() && tg != &root_task_group) {
10537 struct sched_statistics *stats;
10541 for_each_possible_cpu(i) {
10542 stats = __schedstats_from_se(tg->se[i]);
10543 ws += schedstat_val(stats->wait_sum);
10546 seq_printf(sf, "wait_sum %llu\n", ws);
10549 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10550 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10554 #endif /* CONFIG_CFS_BANDWIDTH */
10555 #endif /* CONFIG_FAIR_GROUP_SCHED */
10557 #ifdef CONFIG_RT_GROUP_SCHED
10558 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10559 struct cftype *cft, s64 val)
10561 return sched_group_set_rt_runtime(css_tg(css), val);
10564 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10565 struct cftype *cft)
10567 return sched_group_rt_runtime(css_tg(css));
10570 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10571 struct cftype *cftype, u64 rt_period_us)
10573 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10576 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10577 struct cftype *cft)
10579 return sched_group_rt_period(css_tg(css));
10581 #endif /* CONFIG_RT_GROUP_SCHED */
10583 #ifdef CONFIG_FAIR_GROUP_SCHED
10584 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10585 struct cftype *cft)
10587 return css_tg(css)->idle;
10590 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10591 struct cftype *cft, s64 idle)
10593 return sched_group_set_idle(css_tg(css), idle);
10597 static struct cftype cpu_legacy_files[] = {
10598 #ifdef CONFIG_FAIR_GROUP_SCHED
10601 .read_u64 = cpu_shares_read_u64,
10602 .write_u64 = cpu_shares_write_u64,
10606 .read_s64 = cpu_idle_read_s64,
10607 .write_s64 = cpu_idle_write_s64,
10610 #ifdef CONFIG_CFS_BANDWIDTH
10612 .name = "cfs_quota_us",
10613 .read_s64 = cpu_cfs_quota_read_s64,
10614 .write_s64 = cpu_cfs_quota_write_s64,
10617 .name = "cfs_period_us",
10618 .read_u64 = cpu_cfs_period_read_u64,
10619 .write_u64 = cpu_cfs_period_write_u64,
10622 .name = "cfs_burst_us",
10623 .read_u64 = cpu_cfs_burst_read_u64,
10624 .write_u64 = cpu_cfs_burst_write_u64,
10628 .seq_show = cpu_cfs_stat_show,
10631 #ifdef CONFIG_RT_GROUP_SCHED
10633 .name = "rt_runtime_us",
10634 .read_s64 = cpu_rt_runtime_read,
10635 .write_s64 = cpu_rt_runtime_write,
10638 .name = "rt_period_us",
10639 .read_u64 = cpu_rt_period_read_uint,
10640 .write_u64 = cpu_rt_period_write_uint,
10643 #ifdef CONFIG_UCLAMP_TASK_GROUP
10645 .name = "uclamp.min",
10646 .flags = CFTYPE_NOT_ON_ROOT,
10647 .seq_show = cpu_uclamp_min_show,
10648 .write = cpu_uclamp_min_write,
10651 .name = "uclamp.max",
10652 .flags = CFTYPE_NOT_ON_ROOT,
10653 .seq_show = cpu_uclamp_max_show,
10654 .write = cpu_uclamp_max_write,
10657 { } /* Terminate */
10660 static int cpu_extra_stat_show(struct seq_file *sf,
10661 struct cgroup_subsys_state *css)
10663 #ifdef CONFIG_CFS_BANDWIDTH
10665 struct task_group *tg = css_tg(css);
10666 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10667 u64 throttled_usec, burst_usec;
10669 throttled_usec = cfs_b->throttled_time;
10670 do_div(throttled_usec, NSEC_PER_USEC);
10671 burst_usec = cfs_b->burst_time;
10672 do_div(burst_usec, NSEC_PER_USEC);
10674 seq_printf(sf, "nr_periods %d\n"
10675 "nr_throttled %d\n"
10676 "throttled_usec %llu\n"
10678 "burst_usec %llu\n",
10679 cfs_b->nr_periods, cfs_b->nr_throttled,
10680 throttled_usec, cfs_b->nr_burst, burst_usec);
10686 #ifdef CONFIG_FAIR_GROUP_SCHED
10687 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10688 struct cftype *cft)
10690 struct task_group *tg = css_tg(css);
10691 u64 weight = scale_load_down(tg->shares);
10693 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10696 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10697 struct cftype *cft, u64 weight)
10700 * cgroup weight knobs should use the common MIN, DFL and MAX
10701 * values which are 1, 100 and 10000 respectively. While it loses
10702 * a bit of range on both ends, it maps pretty well onto the shares
10703 * value used by scheduler and the round-trip conversions preserve
10704 * the original value over the entire range.
10706 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10709 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10711 return sched_group_set_shares(css_tg(css), scale_load(weight));
10714 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10715 struct cftype *cft)
10717 unsigned long weight = scale_load_down(css_tg(css)->shares);
10718 int last_delta = INT_MAX;
10721 /* find the closest nice value to the current weight */
10722 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10723 delta = abs(sched_prio_to_weight[prio] - weight);
10724 if (delta >= last_delta)
10726 last_delta = delta;
10729 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10732 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10733 struct cftype *cft, s64 nice)
10735 unsigned long weight;
10738 if (nice < MIN_NICE || nice > MAX_NICE)
10741 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10742 idx = array_index_nospec(idx, 40);
10743 weight = sched_prio_to_weight[idx];
10745 return sched_group_set_shares(css_tg(css), scale_load(weight));
10749 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10750 long period, long quota)
10753 seq_puts(sf, "max");
10755 seq_printf(sf, "%ld", quota);
10757 seq_printf(sf, " %ld\n", period);
10760 /* caller should put the current value in *@periodp before calling */
10761 static int __maybe_unused cpu_period_quota_parse(char *buf,
10762 u64 *periodp, u64 *quotap)
10764 char tok[21]; /* U64_MAX */
10766 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10769 *periodp *= NSEC_PER_USEC;
10771 if (sscanf(tok, "%llu", quotap))
10772 *quotap *= NSEC_PER_USEC;
10773 else if (!strcmp(tok, "max"))
10774 *quotap = RUNTIME_INF;
10781 #ifdef CONFIG_CFS_BANDWIDTH
10782 static int cpu_max_show(struct seq_file *sf, void *v)
10784 struct task_group *tg = css_tg(seq_css(sf));
10786 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10790 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10791 char *buf, size_t nbytes, loff_t off)
10793 struct task_group *tg = css_tg(of_css(of));
10794 u64 period = tg_get_cfs_period(tg);
10795 u64 burst = tg_get_cfs_burst(tg);
10799 ret = cpu_period_quota_parse(buf, &period, "a);
10801 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10802 return ret ?: nbytes;
10806 static struct cftype cpu_files[] = {
10807 #ifdef CONFIG_FAIR_GROUP_SCHED
10810 .flags = CFTYPE_NOT_ON_ROOT,
10811 .read_u64 = cpu_weight_read_u64,
10812 .write_u64 = cpu_weight_write_u64,
10815 .name = "weight.nice",
10816 .flags = CFTYPE_NOT_ON_ROOT,
10817 .read_s64 = cpu_weight_nice_read_s64,
10818 .write_s64 = cpu_weight_nice_write_s64,
10822 .flags = CFTYPE_NOT_ON_ROOT,
10823 .read_s64 = cpu_idle_read_s64,
10824 .write_s64 = cpu_idle_write_s64,
10827 #ifdef CONFIG_CFS_BANDWIDTH
10830 .flags = CFTYPE_NOT_ON_ROOT,
10831 .seq_show = cpu_max_show,
10832 .write = cpu_max_write,
10835 .name = "max.burst",
10836 .flags = CFTYPE_NOT_ON_ROOT,
10837 .read_u64 = cpu_cfs_burst_read_u64,
10838 .write_u64 = cpu_cfs_burst_write_u64,
10841 #ifdef CONFIG_UCLAMP_TASK_GROUP
10843 .name = "uclamp.min",
10844 .flags = CFTYPE_NOT_ON_ROOT,
10845 .seq_show = cpu_uclamp_min_show,
10846 .write = cpu_uclamp_min_write,
10849 .name = "uclamp.max",
10850 .flags = CFTYPE_NOT_ON_ROOT,
10851 .seq_show = cpu_uclamp_max_show,
10852 .write = cpu_uclamp_max_write,
10855 { } /* terminate */
10858 struct cgroup_subsys cpu_cgrp_subsys = {
10859 .css_alloc = cpu_cgroup_css_alloc,
10860 .css_online = cpu_cgroup_css_online,
10861 .css_released = cpu_cgroup_css_released,
10862 .css_free = cpu_cgroup_css_free,
10863 .css_extra_stat_show = cpu_extra_stat_show,
10864 .fork = cpu_cgroup_fork,
10865 .can_attach = cpu_cgroup_can_attach,
10866 .attach = cpu_cgroup_attach,
10867 .legacy_cftypes = cpu_legacy_files,
10868 .dfl_cftypes = cpu_files,
10869 .early_init = true,
10873 #endif /* CONFIG_CGROUP_SCHED */
10875 void dump_cpu_task(int cpu)
10877 pr_info("Task dump for CPU %d:\n", cpu);
10878 sched_show_task(cpu_curr(cpu));
10882 * Nice levels are multiplicative, with a gentle 10% change for every
10883 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10884 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10885 * that remained on nice 0.
10887 * The "10% effect" is relative and cumulative: from _any_ nice level,
10888 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10889 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10890 * If a task goes up by ~10% and another task goes down by ~10% then
10891 * the relative distance between them is ~25%.)
10893 const int sched_prio_to_weight[40] = {
10894 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10895 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10896 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10897 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10898 /* 0 */ 1024, 820, 655, 526, 423,
10899 /* 5 */ 335, 272, 215, 172, 137,
10900 /* 10 */ 110, 87, 70, 56, 45,
10901 /* 15 */ 36, 29, 23, 18, 15,
10905 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10907 * In cases where the weight does not change often, we can use the
10908 * precalculated inverse to speed up arithmetics by turning divisions
10909 * into multiplications:
10911 const u32 sched_prio_to_wmult[40] = {
10912 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10913 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10914 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10915 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10916 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10917 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10918 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10919 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10922 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10924 trace_sched_update_nr_running_tp(rq, count);