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)
1219 bool update_load = !(READ_ONCE(p->__state) & TASK_NEW);
1220 int prio = p->static_prio - MAX_RT_PRIO;
1221 struct load_weight *load = &p->se.load;
1224 * SCHED_IDLE tasks get minimal weight:
1226 if (task_has_idle_policy(p)) {
1227 load->weight = scale_load(WEIGHT_IDLEPRIO);
1228 load->inv_weight = WMULT_IDLEPRIO;
1233 * SCHED_OTHER tasks have to update their load when changing their
1236 if (update_load && p->sched_class == &fair_sched_class) {
1237 reweight_task(p, prio);
1239 load->weight = scale_load(sched_prio_to_weight[prio]);
1240 load->inv_weight = sched_prio_to_wmult[prio];
1244 #ifdef CONFIG_UCLAMP_TASK
1246 * Serializes updates of utilization clamp values
1248 * The (slow-path) user-space triggers utilization clamp value updates which
1249 * can require updates on (fast-path) scheduler's data structures used to
1250 * support enqueue/dequeue operations.
1251 * While the per-CPU rq lock protects fast-path update operations, user-space
1252 * requests are serialized using a mutex to reduce the risk of conflicting
1253 * updates or API abuses.
1255 static DEFINE_MUTEX(uclamp_mutex);
1257 /* Max allowed minimum utilization */
1258 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1260 /* Max allowed maximum utilization */
1261 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1264 * By default RT tasks run at the maximum performance point/capacity of the
1265 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1266 * SCHED_CAPACITY_SCALE.
1268 * This knob allows admins to change the default behavior when uclamp is being
1269 * used. In battery powered devices, particularly, running at the maximum
1270 * capacity and frequency will increase energy consumption and shorten the
1273 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1275 * This knob will not override the system default sched_util_clamp_min defined
1278 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1280 /* All clamps are required to be less or equal than these values */
1281 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1284 * This static key is used to reduce the uclamp overhead in the fast path. It
1285 * primarily disables the call to uclamp_rq_{inc, dec}() in
1286 * enqueue/dequeue_task().
1288 * This allows users to continue to enable uclamp in their kernel config with
1289 * minimum uclamp overhead in the fast path.
1291 * As soon as userspace modifies any of the uclamp knobs, the static key is
1292 * enabled, since we have an actual users that make use of uclamp
1295 * The knobs that would enable this static key are:
1297 * * A task modifying its uclamp value with sched_setattr().
1298 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1299 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1301 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1303 /* Integer rounded range for each bucket */
1304 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1306 #define for_each_clamp_id(clamp_id) \
1307 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1309 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1311 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1314 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1316 if (clamp_id == UCLAMP_MIN)
1318 return SCHED_CAPACITY_SCALE;
1321 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1322 unsigned int value, bool user_defined)
1324 uc_se->value = value;
1325 uc_se->bucket_id = uclamp_bucket_id(value);
1326 uc_se->user_defined = user_defined;
1329 static inline unsigned int
1330 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1331 unsigned int clamp_value)
1334 * Avoid blocked utilization pushing up the frequency when we go
1335 * idle (which drops the max-clamp) by retaining the last known
1338 if (clamp_id == UCLAMP_MAX) {
1339 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1343 return uclamp_none(UCLAMP_MIN);
1346 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1347 unsigned int clamp_value)
1349 /* Reset max-clamp retention only on idle exit */
1350 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1353 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1357 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1358 unsigned int clamp_value)
1360 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1361 int bucket_id = UCLAMP_BUCKETS - 1;
1364 * Since both min and max clamps are max aggregated, find the
1365 * top most bucket with tasks in.
1367 for ( ; bucket_id >= 0; bucket_id--) {
1368 if (!bucket[bucket_id].tasks)
1370 return bucket[bucket_id].value;
1373 /* No tasks -- default clamp values */
1374 return uclamp_idle_value(rq, clamp_id, clamp_value);
1377 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1379 unsigned int default_util_min;
1380 struct uclamp_se *uc_se;
1382 lockdep_assert_held(&p->pi_lock);
1384 uc_se = &p->uclamp_req[UCLAMP_MIN];
1386 /* Only sync if user didn't override the default */
1387 if (uc_se->user_defined)
1390 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1391 uclamp_se_set(uc_se, default_util_min, false);
1394 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1402 /* Protect updates to p->uclamp_* */
1403 rq = task_rq_lock(p, &rf);
1404 __uclamp_update_util_min_rt_default(p);
1405 task_rq_unlock(rq, p, &rf);
1408 static void uclamp_sync_util_min_rt_default(void)
1410 struct task_struct *g, *p;
1413 * copy_process() sysctl_uclamp
1414 * uclamp_min_rt = X;
1415 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1416 * // link thread smp_mb__after_spinlock()
1417 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1418 * sched_post_fork() for_each_process_thread()
1419 * __uclamp_sync_rt() __uclamp_sync_rt()
1421 * Ensures that either sched_post_fork() will observe the new
1422 * uclamp_min_rt or for_each_process_thread() will observe the new
1425 read_lock(&tasklist_lock);
1426 smp_mb__after_spinlock();
1427 read_unlock(&tasklist_lock);
1430 for_each_process_thread(g, p)
1431 uclamp_update_util_min_rt_default(p);
1435 static inline struct uclamp_se
1436 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1438 /* Copy by value as we could modify it */
1439 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1440 #ifdef CONFIG_UCLAMP_TASK_GROUP
1441 unsigned int tg_min, tg_max, value;
1444 * Tasks in autogroups or root task group will be
1445 * restricted by system defaults.
1447 if (task_group_is_autogroup(task_group(p)))
1449 if (task_group(p) == &root_task_group)
1452 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1453 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1454 value = uc_req.value;
1455 value = clamp(value, tg_min, tg_max);
1456 uclamp_se_set(&uc_req, value, false);
1463 * The effective clamp bucket index of a task depends on, by increasing
1465 * - the task specific clamp value, when explicitly requested from userspace
1466 * - the task group effective clamp value, for tasks not either in the root
1467 * group or in an autogroup
1468 * - the system default clamp value, defined by the sysadmin
1470 static inline struct uclamp_se
1471 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1473 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1474 struct uclamp_se uc_max = uclamp_default[clamp_id];
1476 /* System default restrictions always apply */
1477 if (unlikely(uc_req.value > uc_max.value))
1483 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1485 struct uclamp_se uc_eff;
1487 /* Task currently refcounted: use back-annotated (effective) value */
1488 if (p->uclamp[clamp_id].active)
1489 return (unsigned long)p->uclamp[clamp_id].value;
1491 uc_eff = uclamp_eff_get(p, clamp_id);
1493 return (unsigned long)uc_eff.value;
1497 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1498 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1499 * updates the rq's clamp value if required.
1501 * Tasks can have a task-specific value requested from user-space, track
1502 * within each bucket the maximum value for tasks refcounted in it.
1503 * This "local max aggregation" allows to track the exact "requested" value
1504 * for each bucket when all its RUNNABLE tasks require the same clamp.
1506 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1507 enum uclamp_id clamp_id)
1509 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1510 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1511 struct uclamp_bucket *bucket;
1513 lockdep_assert_rq_held(rq);
1515 /* Update task effective clamp */
1516 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1518 bucket = &uc_rq->bucket[uc_se->bucket_id];
1520 uc_se->active = true;
1522 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1525 * Local max aggregation: rq buckets always track the max
1526 * "requested" clamp value of its RUNNABLE tasks.
1528 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1529 bucket->value = uc_se->value;
1531 if (uc_se->value > READ_ONCE(uc_rq->value))
1532 WRITE_ONCE(uc_rq->value, uc_se->value);
1536 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1537 * is released. If this is the last task reference counting the rq's max
1538 * active clamp value, then the rq's clamp value is updated.
1540 * Both refcounted tasks and rq's cached clamp values are expected to be
1541 * always valid. If it's detected they are not, as defensive programming,
1542 * enforce the expected state and warn.
1544 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1545 enum uclamp_id clamp_id)
1547 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1548 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1549 struct uclamp_bucket *bucket;
1550 unsigned int bkt_clamp;
1551 unsigned int rq_clamp;
1553 lockdep_assert_rq_held(rq);
1556 * If sched_uclamp_used was enabled after task @p was enqueued,
1557 * we could end up with unbalanced call to uclamp_rq_dec_id().
1559 * In this case the uc_se->active flag should be false since no uclamp
1560 * accounting was performed at enqueue time and we can just return
1563 * Need to be careful of the following enqueue/dequeue ordering
1567 * // sched_uclamp_used gets enabled
1570 * // Must not decrement bucket->tasks here
1573 * where we could end up with stale data in uc_se and
1574 * bucket[uc_se->bucket_id].
1576 * The following check here eliminates the possibility of such race.
1578 if (unlikely(!uc_se->active))
1581 bucket = &uc_rq->bucket[uc_se->bucket_id];
1583 SCHED_WARN_ON(!bucket->tasks);
1584 if (likely(bucket->tasks))
1587 uc_se->active = false;
1590 * Keep "local max aggregation" simple and accept to (possibly)
1591 * overboost some RUNNABLE tasks in the same bucket.
1592 * The rq clamp bucket value is reset to its base value whenever
1593 * there are no more RUNNABLE tasks refcounting it.
1595 if (likely(bucket->tasks))
1598 rq_clamp = READ_ONCE(uc_rq->value);
1600 * Defensive programming: this should never happen. If it happens,
1601 * e.g. due to future modification, warn and fixup the expected value.
1603 SCHED_WARN_ON(bucket->value > rq_clamp);
1604 if (bucket->value >= rq_clamp) {
1605 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1606 WRITE_ONCE(uc_rq->value, bkt_clamp);
1610 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1612 enum uclamp_id clamp_id;
1615 * Avoid any overhead until uclamp is actually used by the userspace.
1617 * The condition is constructed such that a NOP is generated when
1618 * sched_uclamp_used is disabled.
1620 if (!static_branch_unlikely(&sched_uclamp_used))
1623 if (unlikely(!p->sched_class->uclamp_enabled))
1626 for_each_clamp_id(clamp_id)
1627 uclamp_rq_inc_id(rq, p, clamp_id);
1629 /* Reset clamp idle holding when there is one RUNNABLE task */
1630 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1631 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1634 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1636 enum uclamp_id clamp_id;
1639 * Avoid any overhead until uclamp is actually used by the userspace.
1641 * The condition is constructed such that a NOP is generated when
1642 * sched_uclamp_used is disabled.
1644 if (!static_branch_unlikely(&sched_uclamp_used))
1647 if (unlikely(!p->sched_class->uclamp_enabled))
1650 for_each_clamp_id(clamp_id)
1651 uclamp_rq_dec_id(rq, p, clamp_id);
1654 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1655 enum uclamp_id clamp_id)
1657 if (!p->uclamp[clamp_id].active)
1660 uclamp_rq_dec_id(rq, p, clamp_id);
1661 uclamp_rq_inc_id(rq, p, clamp_id);
1664 * Make sure to clear the idle flag if we've transiently reached 0
1665 * active tasks on rq.
1667 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1668 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1672 uclamp_update_active(struct task_struct *p)
1674 enum uclamp_id clamp_id;
1679 * Lock the task and the rq where the task is (or was) queued.
1681 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1682 * price to pay to safely serialize util_{min,max} updates with
1683 * enqueues, dequeues and migration operations.
1684 * This is the same locking schema used by __set_cpus_allowed_ptr().
1686 rq = task_rq_lock(p, &rf);
1689 * Setting the clamp bucket is serialized by task_rq_lock().
1690 * If the task is not yet RUNNABLE and its task_struct is not
1691 * affecting a valid clamp bucket, the next time it's enqueued,
1692 * it will already see the updated clamp bucket value.
1694 for_each_clamp_id(clamp_id)
1695 uclamp_rq_reinc_id(rq, p, clamp_id);
1697 task_rq_unlock(rq, p, &rf);
1700 #ifdef CONFIG_UCLAMP_TASK_GROUP
1702 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1704 struct css_task_iter it;
1705 struct task_struct *p;
1707 css_task_iter_start(css, 0, &it);
1708 while ((p = css_task_iter_next(&it)))
1709 uclamp_update_active(p);
1710 css_task_iter_end(&it);
1713 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1714 static void uclamp_update_root_tg(void)
1716 struct task_group *tg = &root_task_group;
1718 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1719 sysctl_sched_uclamp_util_min, false);
1720 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1721 sysctl_sched_uclamp_util_max, false);
1724 cpu_util_update_eff(&root_task_group.css);
1728 static void uclamp_update_root_tg(void) { }
1731 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1732 void *buffer, size_t *lenp, loff_t *ppos)
1734 bool update_root_tg = false;
1735 int old_min, old_max, old_min_rt;
1738 mutex_lock(&uclamp_mutex);
1739 old_min = sysctl_sched_uclamp_util_min;
1740 old_max = sysctl_sched_uclamp_util_max;
1741 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1743 result = proc_dointvec(table, write, buffer, lenp, ppos);
1749 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1750 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1751 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1757 if (old_min != sysctl_sched_uclamp_util_min) {
1758 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1759 sysctl_sched_uclamp_util_min, false);
1760 update_root_tg = true;
1762 if (old_max != sysctl_sched_uclamp_util_max) {
1763 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1764 sysctl_sched_uclamp_util_max, false);
1765 update_root_tg = true;
1768 if (update_root_tg) {
1769 static_branch_enable(&sched_uclamp_used);
1770 uclamp_update_root_tg();
1773 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1774 static_branch_enable(&sched_uclamp_used);
1775 uclamp_sync_util_min_rt_default();
1779 * We update all RUNNABLE tasks only when task groups are in use.
1780 * Otherwise, keep it simple and do just a lazy update at each next
1781 * task enqueue time.
1787 sysctl_sched_uclamp_util_min = old_min;
1788 sysctl_sched_uclamp_util_max = old_max;
1789 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1791 mutex_unlock(&uclamp_mutex);
1796 static int uclamp_validate(struct task_struct *p,
1797 const struct sched_attr *attr)
1799 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1800 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1802 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1803 util_min = attr->sched_util_min;
1805 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1809 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1810 util_max = attr->sched_util_max;
1812 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1816 if (util_min != -1 && util_max != -1 && util_min > util_max)
1820 * We have valid uclamp attributes; make sure uclamp is enabled.
1822 * We need to do that here, because enabling static branches is a
1823 * blocking operation which obviously cannot be done while holding
1826 static_branch_enable(&sched_uclamp_used);
1831 static bool uclamp_reset(const struct sched_attr *attr,
1832 enum uclamp_id clamp_id,
1833 struct uclamp_se *uc_se)
1835 /* Reset on sched class change for a non user-defined clamp value. */
1836 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1837 !uc_se->user_defined)
1840 /* Reset on sched_util_{min,max} == -1. */
1841 if (clamp_id == UCLAMP_MIN &&
1842 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1843 attr->sched_util_min == -1) {
1847 if (clamp_id == UCLAMP_MAX &&
1848 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1849 attr->sched_util_max == -1) {
1856 static void __setscheduler_uclamp(struct task_struct *p,
1857 const struct sched_attr *attr)
1859 enum uclamp_id clamp_id;
1861 for_each_clamp_id(clamp_id) {
1862 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1865 if (!uclamp_reset(attr, clamp_id, uc_se))
1869 * RT by default have a 100% boost value that could be modified
1872 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1873 value = sysctl_sched_uclamp_util_min_rt_default;
1875 value = uclamp_none(clamp_id);
1877 uclamp_se_set(uc_se, value, false);
1881 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1884 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1885 attr->sched_util_min != -1) {
1886 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1887 attr->sched_util_min, true);
1890 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1891 attr->sched_util_max != -1) {
1892 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1893 attr->sched_util_max, true);
1897 static void uclamp_fork(struct task_struct *p)
1899 enum uclamp_id clamp_id;
1902 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1903 * as the task is still at its early fork stages.
1905 for_each_clamp_id(clamp_id)
1906 p->uclamp[clamp_id].active = false;
1908 if (likely(!p->sched_reset_on_fork))
1911 for_each_clamp_id(clamp_id) {
1912 uclamp_se_set(&p->uclamp_req[clamp_id],
1913 uclamp_none(clamp_id), false);
1917 static void uclamp_post_fork(struct task_struct *p)
1919 uclamp_update_util_min_rt_default(p);
1922 static void __init init_uclamp_rq(struct rq *rq)
1924 enum uclamp_id clamp_id;
1925 struct uclamp_rq *uc_rq = rq->uclamp;
1927 for_each_clamp_id(clamp_id) {
1928 uc_rq[clamp_id] = (struct uclamp_rq) {
1929 .value = uclamp_none(clamp_id)
1933 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1936 static void __init init_uclamp(void)
1938 struct uclamp_se uc_max = {};
1939 enum uclamp_id clamp_id;
1942 for_each_possible_cpu(cpu)
1943 init_uclamp_rq(cpu_rq(cpu));
1945 for_each_clamp_id(clamp_id) {
1946 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1947 uclamp_none(clamp_id), false);
1950 /* System defaults allow max clamp values for both indexes */
1951 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1952 for_each_clamp_id(clamp_id) {
1953 uclamp_default[clamp_id] = uc_max;
1954 #ifdef CONFIG_UCLAMP_TASK_GROUP
1955 root_task_group.uclamp_req[clamp_id] = uc_max;
1956 root_task_group.uclamp[clamp_id] = uc_max;
1961 #else /* CONFIG_UCLAMP_TASK */
1962 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1963 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1964 static inline int uclamp_validate(struct task_struct *p,
1965 const struct sched_attr *attr)
1969 static void __setscheduler_uclamp(struct task_struct *p,
1970 const struct sched_attr *attr) { }
1971 static inline void uclamp_fork(struct task_struct *p) { }
1972 static inline void uclamp_post_fork(struct task_struct *p) { }
1973 static inline void init_uclamp(void) { }
1974 #endif /* CONFIG_UCLAMP_TASK */
1976 bool sched_task_on_rq(struct task_struct *p)
1978 return task_on_rq_queued(p);
1981 unsigned long get_wchan(struct task_struct *p)
1983 unsigned long ip = 0;
1986 if (!p || p == current)
1989 /* Only get wchan if task is blocked and we can keep it that way. */
1990 raw_spin_lock_irq(&p->pi_lock);
1991 state = READ_ONCE(p->__state);
1992 smp_rmb(); /* see try_to_wake_up() */
1993 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
1994 ip = __get_wchan(p);
1995 raw_spin_unlock_irq(&p->pi_lock);
2000 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
2002 if (!(flags & ENQUEUE_NOCLOCK))
2003 update_rq_clock(rq);
2005 if (!(flags & ENQUEUE_RESTORE)) {
2006 sched_info_enqueue(rq, p);
2007 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
2010 uclamp_rq_inc(rq, p);
2011 p->sched_class->enqueue_task(rq, p, flags);
2013 if (sched_core_enabled(rq))
2014 sched_core_enqueue(rq, p);
2017 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
2019 if (sched_core_enabled(rq))
2020 sched_core_dequeue(rq, p, flags);
2022 if (!(flags & DEQUEUE_NOCLOCK))
2023 update_rq_clock(rq);
2025 if (!(flags & DEQUEUE_SAVE)) {
2026 sched_info_dequeue(rq, p);
2027 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2030 uclamp_rq_dec(rq, p);
2031 p->sched_class->dequeue_task(rq, p, flags);
2034 void activate_task(struct rq *rq, struct task_struct *p, int flags)
2036 enqueue_task(rq, p, flags);
2038 p->on_rq = TASK_ON_RQ_QUEUED;
2041 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
2043 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2045 dequeue_task(rq, p, flags);
2048 static inline int __normal_prio(int policy, int rt_prio, int nice)
2052 if (dl_policy(policy))
2053 prio = MAX_DL_PRIO - 1;
2054 else if (rt_policy(policy))
2055 prio = MAX_RT_PRIO - 1 - rt_prio;
2057 prio = NICE_TO_PRIO(nice);
2063 * Calculate the expected normal priority: i.e. priority
2064 * without taking RT-inheritance into account. Might be
2065 * boosted by interactivity modifiers. Changes upon fork,
2066 * setprio syscalls, and whenever the interactivity
2067 * estimator recalculates.
2069 static inline int normal_prio(struct task_struct *p)
2071 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2075 * Calculate the current priority, i.e. the priority
2076 * taken into account by the scheduler. This value might
2077 * be boosted by RT tasks, or might be boosted by
2078 * interactivity modifiers. Will be RT if the task got
2079 * RT-boosted. If not then it returns p->normal_prio.
2081 static int effective_prio(struct task_struct *p)
2083 p->normal_prio = normal_prio(p);
2085 * If we are RT tasks or we were boosted to RT priority,
2086 * keep the priority unchanged. Otherwise, update priority
2087 * to the normal priority:
2089 if (!rt_prio(p->prio))
2090 return p->normal_prio;
2095 * task_curr - is this task currently executing on a CPU?
2096 * @p: the task in question.
2098 * Return: 1 if the task is currently executing. 0 otherwise.
2100 inline int task_curr(const struct task_struct *p)
2102 return cpu_curr(task_cpu(p)) == p;
2106 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2107 * use the balance_callback list if you want balancing.
2109 * this means any call to check_class_changed() must be followed by a call to
2110 * balance_callback().
2112 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2113 const struct sched_class *prev_class,
2116 if (prev_class != p->sched_class) {
2117 if (prev_class->switched_from)
2118 prev_class->switched_from(rq, p);
2120 p->sched_class->switched_to(rq, p);
2121 } else if (oldprio != p->prio || dl_task(p))
2122 p->sched_class->prio_changed(rq, p, oldprio);
2125 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2127 if (p->sched_class == rq->curr->sched_class)
2128 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2129 else if (p->sched_class > rq->curr->sched_class)
2133 * A queue event has occurred, and we're going to schedule. In
2134 * this case, we can save a useless back to back clock update.
2136 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2137 rq_clock_skip_update(rq);
2143 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2145 static int __set_cpus_allowed_ptr(struct task_struct *p,
2146 const struct cpumask *new_mask,
2149 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2151 if (likely(!p->migration_disabled))
2154 if (p->cpus_ptr != &p->cpus_mask)
2158 * Violates locking rules! see comment in __do_set_cpus_allowed().
2160 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2163 void migrate_disable(void)
2165 struct task_struct *p = current;
2167 if (p->migration_disabled) {
2168 p->migration_disabled++;
2173 this_rq()->nr_pinned++;
2174 p->migration_disabled = 1;
2177 EXPORT_SYMBOL_GPL(migrate_disable);
2179 void migrate_enable(void)
2181 struct task_struct *p = current;
2183 if (p->migration_disabled > 1) {
2184 p->migration_disabled--;
2188 if (WARN_ON_ONCE(!p->migration_disabled))
2192 * Ensure stop_task runs either before or after this, and that
2193 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2196 if (p->cpus_ptr != &p->cpus_mask)
2197 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2199 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2200 * regular cpus_mask, otherwise things that race (eg.
2201 * select_fallback_rq) get confused.
2204 p->migration_disabled = 0;
2205 this_rq()->nr_pinned--;
2208 EXPORT_SYMBOL_GPL(migrate_enable);
2210 static inline bool rq_has_pinned_tasks(struct rq *rq)
2212 return rq->nr_pinned;
2216 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2217 * __set_cpus_allowed_ptr() and select_fallback_rq().
2219 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2221 /* When not in the task's cpumask, no point in looking further. */
2222 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2225 /* migrate_disabled() must be allowed to finish. */
2226 if (is_migration_disabled(p))
2227 return cpu_online(cpu);
2229 /* Non kernel threads are not allowed during either online or offline. */
2230 if (!(p->flags & PF_KTHREAD))
2231 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2233 /* KTHREAD_IS_PER_CPU is always allowed. */
2234 if (kthread_is_per_cpu(p))
2235 return cpu_online(cpu);
2237 /* Regular kernel threads don't get to stay during offline. */
2241 /* But are allowed during online. */
2242 return cpu_online(cpu);
2246 * This is how migration works:
2248 * 1) we invoke migration_cpu_stop() on the target CPU using
2250 * 2) stopper starts to run (implicitly forcing the migrated thread
2252 * 3) it checks whether the migrated task is still in the wrong runqueue.
2253 * 4) if it's in the wrong runqueue then the migration thread removes
2254 * it and puts it into the right queue.
2255 * 5) stopper completes and stop_one_cpu() returns and the migration
2260 * move_queued_task - move a queued task to new rq.
2262 * Returns (locked) new rq. Old rq's lock is released.
2264 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2265 struct task_struct *p, int new_cpu)
2267 lockdep_assert_rq_held(rq);
2269 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2270 set_task_cpu(p, new_cpu);
2273 rq = cpu_rq(new_cpu);
2276 BUG_ON(task_cpu(p) != new_cpu);
2277 activate_task(rq, p, 0);
2278 check_preempt_curr(rq, p, 0);
2283 struct migration_arg {
2284 struct task_struct *task;
2286 struct set_affinity_pending *pending;
2290 * @refs: number of wait_for_completion()
2291 * @stop_pending: is @stop_work in use
2293 struct set_affinity_pending {
2295 unsigned int stop_pending;
2296 struct completion done;
2297 struct cpu_stop_work stop_work;
2298 struct migration_arg arg;
2302 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2303 * this because either it can't run here any more (set_cpus_allowed()
2304 * away from this CPU, or CPU going down), or because we're
2305 * attempting to rebalance this task on exec (sched_exec).
2307 * So we race with normal scheduler movements, but that's OK, as long
2308 * as the task is no longer on this CPU.
2310 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2311 struct task_struct *p, int dest_cpu)
2313 /* Affinity changed (again). */
2314 if (!is_cpu_allowed(p, dest_cpu))
2317 update_rq_clock(rq);
2318 rq = move_queued_task(rq, rf, p, dest_cpu);
2324 * migration_cpu_stop - this will be executed by a highprio stopper thread
2325 * and performs thread migration by bumping thread off CPU then
2326 * 'pushing' onto another runqueue.
2328 static int migration_cpu_stop(void *data)
2330 struct migration_arg *arg = data;
2331 struct set_affinity_pending *pending = arg->pending;
2332 struct task_struct *p = arg->task;
2333 struct rq *rq = this_rq();
2334 bool complete = false;
2338 * The original target CPU might have gone down and we might
2339 * be on another CPU but it doesn't matter.
2341 local_irq_save(rf.flags);
2343 * We need to explicitly wake pending tasks before running
2344 * __migrate_task() such that we will not miss enforcing cpus_ptr
2345 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2347 flush_smp_call_function_from_idle();
2349 raw_spin_lock(&p->pi_lock);
2353 * If we were passed a pending, then ->stop_pending was set, thus
2354 * p->migration_pending must have remained stable.
2356 WARN_ON_ONCE(pending && pending != p->migration_pending);
2359 * If task_rq(p) != rq, it cannot be migrated here, because we're
2360 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2361 * we're holding p->pi_lock.
2363 if (task_rq(p) == rq) {
2364 if (is_migration_disabled(p))
2368 p->migration_pending = NULL;
2371 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2375 if (task_on_rq_queued(p))
2376 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2378 p->wake_cpu = arg->dest_cpu;
2381 * XXX __migrate_task() can fail, at which point we might end
2382 * up running on a dodgy CPU, AFAICT this can only happen
2383 * during CPU hotplug, at which point we'll get pushed out
2384 * anyway, so it's probably not a big deal.
2387 } else if (pending) {
2389 * This happens when we get migrated between migrate_enable()'s
2390 * preempt_enable() and scheduling the stopper task. At that
2391 * point we're a regular task again and not current anymore.
2393 * A !PREEMPT kernel has a giant hole here, which makes it far
2398 * The task moved before the stopper got to run. We're holding
2399 * ->pi_lock, so the allowed mask is stable - if it got
2400 * somewhere allowed, we're done.
2402 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2403 p->migration_pending = NULL;
2409 * When migrate_enable() hits a rq mis-match we can't reliably
2410 * determine is_migration_disabled() and so have to chase after
2413 WARN_ON_ONCE(!pending->stop_pending);
2414 task_rq_unlock(rq, p, &rf);
2415 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2416 &pending->arg, &pending->stop_work);
2421 pending->stop_pending = false;
2422 task_rq_unlock(rq, p, &rf);
2425 complete_all(&pending->done);
2430 int push_cpu_stop(void *arg)
2432 struct rq *lowest_rq = NULL, *rq = this_rq();
2433 struct task_struct *p = arg;
2435 raw_spin_lock_irq(&p->pi_lock);
2436 raw_spin_rq_lock(rq);
2438 if (task_rq(p) != rq)
2441 if (is_migration_disabled(p)) {
2442 p->migration_flags |= MDF_PUSH;
2446 p->migration_flags &= ~MDF_PUSH;
2448 if (p->sched_class->find_lock_rq)
2449 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2454 // XXX validate p is still the highest prio task
2455 if (task_rq(p) == rq) {
2456 deactivate_task(rq, p, 0);
2457 set_task_cpu(p, lowest_rq->cpu);
2458 activate_task(lowest_rq, p, 0);
2459 resched_curr(lowest_rq);
2462 double_unlock_balance(rq, lowest_rq);
2465 rq->push_busy = false;
2466 raw_spin_rq_unlock(rq);
2467 raw_spin_unlock_irq(&p->pi_lock);
2474 * sched_class::set_cpus_allowed must do the below, but is not required to
2475 * actually call this function.
2477 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2479 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2480 p->cpus_ptr = new_mask;
2484 cpumask_copy(&p->cpus_mask, new_mask);
2485 p->nr_cpus_allowed = cpumask_weight(new_mask);
2489 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2491 struct rq *rq = task_rq(p);
2492 bool queued, running;
2495 * This here violates the locking rules for affinity, since we're only
2496 * supposed to change these variables while holding both rq->lock and
2499 * HOWEVER, it magically works, because ttwu() is the only code that
2500 * accesses these variables under p->pi_lock and only does so after
2501 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2502 * before finish_task().
2504 * XXX do further audits, this smells like something putrid.
2506 if (flags & SCA_MIGRATE_DISABLE)
2507 SCHED_WARN_ON(!p->on_cpu);
2509 lockdep_assert_held(&p->pi_lock);
2511 queued = task_on_rq_queued(p);
2512 running = task_current(rq, p);
2516 * Because __kthread_bind() calls this on blocked tasks without
2519 lockdep_assert_rq_held(rq);
2520 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2523 put_prev_task(rq, p);
2525 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2528 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2530 set_next_task(rq, p);
2533 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2535 __do_set_cpus_allowed(p, new_mask, 0);
2538 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2541 if (!src->user_cpus_ptr)
2544 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2545 if (!dst->user_cpus_ptr)
2548 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2552 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2554 struct cpumask *user_mask = NULL;
2556 swap(p->user_cpus_ptr, user_mask);
2561 void release_user_cpus_ptr(struct task_struct *p)
2563 kfree(clear_user_cpus_ptr(p));
2567 * This function is wildly self concurrent; here be dragons.
2570 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2571 * designated task is enqueued on an allowed CPU. If that task is currently
2572 * running, we have to kick it out using the CPU stopper.
2574 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2577 * Initial conditions: P0->cpus_mask = [0, 1]
2581 * migrate_disable();
2583 * set_cpus_allowed_ptr(P0, [1]);
2585 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2586 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2587 * This means we need the following scheme:
2591 * migrate_disable();
2593 * set_cpus_allowed_ptr(P0, [1]);
2597 * __set_cpus_allowed_ptr();
2598 * <wakes local stopper>
2599 * `--> <woken on migration completion>
2601 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2602 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2603 * task p are serialized by p->pi_lock, which we can leverage: the one that
2604 * should come into effect at the end of the Migrate-Disable region is the last
2605 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2606 * but we still need to properly signal those waiting tasks at the appropriate
2609 * This is implemented using struct set_affinity_pending. The first
2610 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2611 * setup an instance of that struct and install it on the targeted task_struct.
2612 * Any and all further callers will reuse that instance. Those then wait for
2613 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2614 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2617 * (1) In the cases covered above. There is one more where the completion is
2618 * signaled within affine_move_task() itself: when a subsequent affinity request
2619 * occurs after the stopper bailed out due to the targeted task still being
2620 * Migrate-Disable. Consider:
2622 * Initial conditions: P0->cpus_mask = [0, 1]
2626 * migrate_disable();
2628 * set_cpus_allowed_ptr(P0, [1]);
2631 * migration_cpu_stop()
2632 * is_migration_disabled()
2634 * set_cpus_allowed_ptr(P0, [0, 1]);
2635 * <signal completion>
2638 * Note that the above is safe vs a concurrent migrate_enable(), as any
2639 * pending affinity completion is preceded by an uninstallation of
2640 * p->migration_pending done with p->pi_lock held.
2642 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2643 int dest_cpu, unsigned int flags)
2645 struct set_affinity_pending my_pending = { }, *pending = NULL;
2646 bool stop_pending, complete = false;
2648 /* Can the task run on the task's current CPU? If so, we're done */
2649 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2650 struct task_struct *push_task = NULL;
2652 if ((flags & SCA_MIGRATE_ENABLE) &&
2653 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2654 rq->push_busy = true;
2655 push_task = get_task_struct(p);
2659 * If there are pending waiters, but no pending stop_work,
2660 * then complete now.
2662 pending = p->migration_pending;
2663 if (pending && !pending->stop_pending) {
2664 p->migration_pending = NULL;
2668 task_rq_unlock(rq, p, rf);
2671 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2676 complete_all(&pending->done);
2681 if (!(flags & SCA_MIGRATE_ENABLE)) {
2682 /* serialized by p->pi_lock */
2683 if (!p->migration_pending) {
2684 /* Install the request */
2685 refcount_set(&my_pending.refs, 1);
2686 init_completion(&my_pending.done);
2687 my_pending.arg = (struct migration_arg) {
2689 .dest_cpu = dest_cpu,
2690 .pending = &my_pending,
2693 p->migration_pending = &my_pending;
2695 pending = p->migration_pending;
2696 refcount_inc(&pending->refs);
2698 * Affinity has changed, but we've already installed a
2699 * pending. migration_cpu_stop() *must* see this, else
2700 * we risk a completion of the pending despite having a
2701 * task on a disallowed CPU.
2703 * Serialized by p->pi_lock, so this is safe.
2705 pending->arg.dest_cpu = dest_cpu;
2708 pending = p->migration_pending;
2710 * - !MIGRATE_ENABLE:
2711 * we'll have installed a pending if there wasn't one already.
2714 * we're here because the current CPU isn't matching anymore,
2715 * the only way that can happen is because of a concurrent
2716 * set_cpus_allowed_ptr() call, which should then still be
2717 * pending completion.
2719 * Either way, we really should have a @pending here.
2721 if (WARN_ON_ONCE(!pending)) {
2722 task_rq_unlock(rq, p, rf);
2726 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2728 * MIGRATE_ENABLE gets here because 'p == current', but for
2729 * anything else we cannot do is_migration_disabled(), punt
2730 * and have the stopper function handle it all race-free.
2732 stop_pending = pending->stop_pending;
2734 pending->stop_pending = true;
2736 if (flags & SCA_MIGRATE_ENABLE)
2737 p->migration_flags &= ~MDF_PUSH;
2739 task_rq_unlock(rq, p, rf);
2741 if (!stop_pending) {
2742 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2743 &pending->arg, &pending->stop_work);
2746 if (flags & SCA_MIGRATE_ENABLE)
2750 if (!is_migration_disabled(p)) {
2751 if (task_on_rq_queued(p))
2752 rq = move_queued_task(rq, rf, p, dest_cpu);
2754 if (!pending->stop_pending) {
2755 p->migration_pending = NULL;
2759 task_rq_unlock(rq, p, rf);
2762 complete_all(&pending->done);
2765 wait_for_completion(&pending->done);
2767 if (refcount_dec_and_test(&pending->refs))
2768 wake_up_var(&pending->refs); /* No UaF, just an address */
2771 * Block the original owner of &pending until all subsequent callers
2772 * have seen the completion and decremented the refcount
2774 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2777 WARN_ON_ONCE(my_pending.stop_pending);
2783 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2785 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2786 const struct cpumask *new_mask,
2789 struct rq_flags *rf)
2790 __releases(rq->lock)
2791 __releases(p->pi_lock)
2793 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2794 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2795 bool kthread = p->flags & PF_KTHREAD;
2796 struct cpumask *user_mask = NULL;
2797 unsigned int dest_cpu;
2800 update_rq_clock(rq);
2802 if (kthread || is_migration_disabled(p)) {
2804 * Kernel threads are allowed on online && !active CPUs,
2805 * however, during cpu-hot-unplug, even these might get pushed
2806 * away if not KTHREAD_IS_PER_CPU.
2808 * Specifically, migration_disabled() tasks must not fail the
2809 * cpumask_any_and_distribute() pick below, esp. so on
2810 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2811 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2813 cpu_valid_mask = cpu_online_mask;
2816 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2822 * Must re-check here, to close a race against __kthread_bind(),
2823 * sched_setaffinity() is not guaranteed to observe the flag.
2825 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2830 if (!(flags & SCA_MIGRATE_ENABLE)) {
2831 if (cpumask_equal(&p->cpus_mask, new_mask))
2834 if (WARN_ON_ONCE(p == current &&
2835 is_migration_disabled(p) &&
2836 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2843 * Picking a ~random cpu helps in cases where we are changing affinity
2844 * for groups of tasks (ie. cpuset), so that load balancing is not
2845 * immediately required to distribute the tasks within their new mask.
2847 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2848 if (dest_cpu >= nr_cpu_ids) {
2853 __do_set_cpus_allowed(p, new_mask, flags);
2855 if (flags & SCA_USER)
2856 user_mask = clear_user_cpus_ptr(p);
2858 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2865 task_rq_unlock(rq, p, rf);
2871 * Change a given task's CPU affinity. Migrate the thread to a
2872 * proper CPU and schedule it away if the CPU it's executing on
2873 * is removed from the allowed bitmask.
2875 * NOTE: the caller must have a valid reference to the task, the
2876 * task must not exit() & deallocate itself prematurely. The
2877 * call is not atomic; no spinlocks may be held.
2879 static int __set_cpus_allowed_ptr(struct task_struct *p,
2880 const struct cpumask *new_mask, u32 flags)
2885 rq = task_rq_lock(p, &rf);
2886 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2889 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2891 return __set_cpus_allowed_ptr(p, new_mask, 0);
2893 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2896 * Change a given task's CPU affinity to the intersection of its current
2897 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2898 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2899 * If the resulting mask is empty, leave the affinity unchanged and return
2902 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2903 struct cpumask *new_mask,
2904 const struct cpumask *subset_mask)
2906 struct cpumask *user_mask = NULL;
2911 if (!p->user_cpus_ptr) {
2912 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2917 rq = task_rq_lock(p, &rf);
2920 * Forcefully restricting the affinity of a deadline task is
2921 * likely to cause problems, so fail and noisily override the
2924 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2929 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2935 * We're about to butcher the task affinity, so keep track of what
2936 * the user asked for in case we're able to restore it later on.
2939 cpumask_copy(user_mask, p->cpus_ptr);
2940 p->user_cpus_ptr = user_mask;
2943 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2946 task_rq_unlock(rq, p, &rf);
2952 * Restrict the CPU affinity of task @p so that it is a subset of
2953 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2954 * old affinity mask. If the resulting mask is empty, we warn and walk
2955 * up the cpuset hierarchy until we find a suitable mask.
2957 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2959 cpumask_var_t new_mask;
2960 const struct cpumask *override_mask = task_cpu_possible_mask(p);
2962 alloc_cpumask_var(&new_mask, GFP_KERNEL);
2965 * __migrate_task() can fail silently in the face of concurrent
2966 * offlining of the chosen destination CPU, so take the hotplug
2967 * lock to ensure that the migration succeeds.
2970 if (!cpumask_available(new_mask))
2973 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2977 * We failed to find a valid subset of the affinity mask for the
2978 * task, so override it based on its cpuset hierarchy.
2980 cpuset_cpus_allowed(p, new_mask);
2981 override_mask = new_mask;
2984 if (printk_ratelimit()) {
2985 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2986 task_pid_nr(p), p->comm,
2987 cpumask_pr_args(override_mask));
2990 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2993 free_cpumask_var(new_mask);
2997 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
3000 * Restore the affinity of a task @p which was previously restricted by a
3001 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
3002 * @p->user_cpus_ptr.
3004 * It is the caller's responsibility to serialise this with any calls to
3005 * force_compatible_cpus_allowed_ptr(@p).
3007 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3009 struct cpumask *user_mask = p->user_cpus_ptr;
3010 unsigned long flags;
3013 * Try to restore the old affinity mask. If this fails, then
3014 * we free the mask explicitly to avoid it being inherited across
3015 * a subsequent fork().
3017 if (!user_mask || !__sched_setaffinity(p, user_mask))
3020 raw_spin_lock_irqsave(&p->pi_lock, flags);
3021 user_mask = clear_user_cpus_ptr(p);
3022 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3027 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
3029 #ifdef CONFIG_SCHED_DEBUG
3030 unsigned int state = READ_ONCE(p->__state);
3033 * We should never call set_task_cpu() on a blocked task,
3034 * ttwu() will sort out the placement.
3036 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
3039 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3040 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3041 * time relying on p->on_rq.
3043 WARN_ON_ONCE(state == TASK_RUNNING &&
3044 p->sched_class == &fair_sched_class &&
3045 (p->on_rq && !task_on_rq_migrating(p)));
3047 #ifdef CONFIG_LOCKDEP
3049 * The caller should hold either p->pi_lock or rq->lock, when changing
3050 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3052 * sched_move_task() holds both and thus holding either pins the cgroup,
3055 * Furthermore, all task_rq users should acquire both locks, see
3058 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3059 lockdep_is_held(__rq_lockp(task_rq(p)))));
3062 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3064 WARN_ON_ONCE(!cpu_online(new_cpu));
3066 WARN_ON_ONCE(is_migration_disabled(p));
3069 trace_sched_migrate_task(p, new_cpu);
3071 if (task_cpu(p) != new_cpu) {
3072 if (p->sched_class->migrate_task_rq)
3073 p->sched_class->migrate_task_rq(p, new_cpu);
3074 p->se.nr_migrations++;
3076 perf_event_task_migrate(p);
3079 __set_task_cpu(p, new_cpu);
3082 #ifdef CONFIG_NUMA_BALANCING
3083 static void __migrate_swap_task(struct task_struct *p, int cpu)
3085 if (task_on_rq_queued(p)) {
3086 struct rq *src_rq, *dst_rq;
3087 struct rq_flags srf, drf;
3089 src_rq = task_rq(p);
3090 dst_rq = cpu_rq(cpu);
3092 rq_pin_lock(src_rq, &srf);
3093 rq_pin_lock(dst_rq, &drf);
3095 deactivate_task(src_rq, p, 0);
3096 set_task_cpu(p, cpu);
3097 activate_task(dst_rq, p, 0);
3098 check_preempt_curr(dst_rq, p, 0);
3100 rq_unpin_lock(dst_rq, &drf);
3101 rq_unpin_lock(src_rq, &srf);
3105 * Task isn't running anymore; make it appear like we migrated
3106 * it before it went to sleep. This means on wakeup we make the
3107 * previous CPU our target instead of where it really is.
3113 struct migration_swap_arg {
3114 struct task_struct *src_task, *dst_task;
3115 int src_cpu, dst_cpu;
3118 static int migrate_swap_stop(void *data)
3120 struct migration_swap_arg *arg = data;
3121 struct rq *src_rq, *dst_rq;
3124 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3127 src_rq = cpu_rq(arg->src_cpu);
3128 dst_rq = cpu_rq(arg->dst_cpu);
3130 double_raw_lock(&arg->src_task->pi_lock,
3131 &arg->dst_task->pi_lock);
3132 double_rq_lock(src_rq, dst_rq);
3134 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3137 if (task_cpu(arg->src_task) != arg->src_cpu)
3140 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3143 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3146 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3147 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3152 double_rq_unlock(src_rq, dst_rq);
3153 raw_spin_unlock(&arg->dst_task->pi_lock);
3154 raw_spin_unlock(&arg->src_task->pi_lock);
3160 * Cross migrate two tasks
3162 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3163 int target_cpu, int curr_cpu)
3165 struct migration_swap_arg arg;
3168 arg = (struct migration_swap_arg){
3170 .src_cpu = curr_cpu,
3172 .dst_cpu = target_cpu,
3175 if (arg.src_cpu == arg.dst_cpu)
3179 * These three tests are all lockless; this is OK since all of them
3180 * will be re-checked with proper locks held further down the line.
3182 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3185 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3188 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3191 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3192 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3197 #endif /* CONFIG_NUMA_BALANCING */
3200 * wait_task_inactive - wait for a thread to unschedule.
3202 * If @match_state is nonzero, it's the @p->state value just checked and
3203 * not expected to change. If it changes, i.e. @p might have woken up,
3204 * then return zero. When we succeed in waiting for @p to be off its CPU,
3205 * we return a positive number (its total switch count). If a second call
3206 * a short while later returns the same number, the caller can be sure that
3207 * @p has remained unscheduled the whole time.
3209 * The caller must ensure that the task *will* unschedule sometime soon,
3210 * else this function might spin for a *long* time. This function can't
3211 * be called with interrupts off, or it may introduce deadlock with
3212 * smp_call_function() if an IPI is sent by the same process we are
3213 * waiting to become inactive.
3215 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3217 int running, queued;
3224 * We do the initial early heuristics without holding
3225 * any task-queue locks at all. We'll only try to get
3226 * the runqueue lock when things look like they will
3232 * If the task is actively running on another CPU
3233 * still, just relax and busy-wait without holding
3236 * NOTE! Since we don't hold any locks, it's not
3237 * even sure that "rq" stays as the right runqueue!
3238 * But we don't care, since "task_running()" will
3239 * return false if the runqueue has changed and p
3240 * is actually now running somewhere else!
3242 while (task_running(rq, p)) {
3243 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3249 * Ok, time to look more closely! We need the rq
3250 * lock now, to be *sure*. If we're wrong, we'll
3251 * just go back and repeat.
3253 rq = task_rq_lock(p, &rf);
3254 trace_sched_wait_task(p);
3255 running = task_running(rq, p);
3256 queued = task_on_rq_queued(p);
3258 if (!match_state || READ_ONCE(p->__state) == match_state)
3259 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3260 task_rq_unlock(rq, p, &rf);
3263 * If it changed from the expected state, bail out now.
3265 if (unlikely(!ncsw))
3269 * Was it really running after all now that we
3270 * checked with the proper locks actually held?
3272 * Oops. Go back and try again..
3274 if (unlikely(running)) {
3280 * It's not enough that it's not actively running,
3281 * it must be off the runqueue _entirely_, and not
3284 * So if it was still runnable (but just not actively
3285 * running right now), it's preempted, and we should
3286 * yield - it could be a while.
3288 if (unlikely(queued)) {
3289 ktime_t to = NSEC_PER_SEC / HZ;
3291 set_current_state(TASK_UNINTERRUPTIBLE);
3292 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
3297 * Ahh, all good. It wasn't running, and it wasn't
3298 * runnable, which means that it will never become
3299 * running in the future either. We're all done!
3308 * kick_process - kick a running thread to enter/exit the kernel
3309 * @p: the to-be-kicked thread
3311 * Cause a process which is running on another CPU to enter
3312 * kernel-mode, without any delay. (to get signals handled.)
3314 * NOTE: this function doesn't have to take the runqueue lock,
3315 * because all it wants to ensure is that the remote task enters
3316 * the kernel. If the IPI races and the task has been migrated
3317 * to another CPU then no harm is done and the purpose has been
3320 void kick_process(struct task_struct *p)
3326 if ((cpu != smp_processor_id()) && task_curr(p))
3327 smp_send_reschedule(cpu);
3330 EXPORT_SYMBOL_GPL(kick_process);
3333 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3335 * A few notes on cpu_active vs cpu_online:
3337 * - cpu_active must be a subset of cpu_online
3339 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3340 * see __set_cpus_allowed_ptr(). At this point the newly online
3341 * CPU isn't yet part of the sched domains, and balancing will not
3344 * - on CPU-down we clear cpu_active() to mask the sched domains and
3345 * avoid the load balancer to place new tasks on the to be removed
3346 * CPU. Existing tasks will remain running there and will be taken
3349 * This means that fallback selection must not select !active CPUs.
3350 * And can assume that any active CPU must be online. Conversely
3351 * select_task_rq() below may allow selection of !active CPUs in order
3352 * to satisfy the above rules.
3354 static int select_fallback_rq(int cpu, struct task_struct *p)
3356 int nid = cpu_to_node(cpu);
3357 const struct cpumask *nodemask = NULL;
3358 enum { cpuset, possible, fail } state = cpuset;
3362 * If the node that the CPU is on has been offlined, cpu_to_node()
3363 * will return -1. There is no CPU on the node, and we should
3364 * select the CPU on the other node.
3367 nodemask = cpumask_of_node(nid);
3369 /* Look for allowed, online CPU in same node. */
3370 for_each_cpu(dest_cpu, nodemask) {
3371 if (is_cpu_allowed(p, dest_cpu))
3377 /* Any allowed, online CPU? */
3378 for_each_cpu(dest_cpu, p->cpus_ptr) {
3379 if (!is_cpu_allowed(p, dest_cpu))
3385 /* No more Mr. Nice Guy. */
3388 if (cpuset_cpus_allowed_fallback(p)) {
3395 * XXX When called from select_task_rq() we only
3396 * hold p->pi_lock and again violate locking order.
3398 * More yuck to audit.
3400 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3410 if (state != cpuset) {
3412 * Don't tell them about moving exiting tasks or
3413 * kernel threads (both mm NULL), since they never
3416 if (p->mm && printk_ratelimit()) {
3417 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3418 task_pid_nr(p), p->comm, cpu);
3426 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3429 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3431 lockdep_assert_held(&p->pi_lock);
3433 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3434 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3436 cpu = cpumask_any(p->cpus_ptr);
3439 * In order not to call set_task_cpu() on a blocking task we need
3440 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3443 * Since this is common to all placement strategies, this lives here.
3445 * [ this allows ->select_task() to simply return task_cpu(p) and
3446 * not worry about this generic constraint ]
3448 if (unlikely(!is_cpu_allowed(p, cpu)))
3449 cpu = select_fallback_rq(task_cpu(p), p);
3454 void sched_set_stop_task(int cpu, struct task_struct *stop)
3456 static struct lock_class_key stop_pi_lock;
3457 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3458 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3462 * Make it appear like a SCHED_FIFO task, its something
3463 * userspace knows about and won't get confused about.
3465 * Also, it will make PI more or less work without too
3466 * much confusion -- but then, stop work should not
3467 * rely on PI working anyway.
3469 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3471 stop->sched_class = &stop_sched_class;
3474 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3475 * adjust the effective priority of a task. As a result,
3476 * rt_mutex_setprio() can trigger (RT) balancing operations,
3477 * which can then trigger wakeups of the stop thread to push
3478 * around the current task.
3480 * The stop task itself will never be part of the PI-chain, it
3481 * never blocks, therefore that ->pi_lock recursion is safe.
3482 * Tell lockdep about this by placing the stop->pi_lock in its
3485 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3488 cpu_rq(cpu)->stop = stop;
3492 * Reset it back to a normal scheduling class so that
3493 * it can die in pieces.
3495 old_stop->sched_class = &rt_sched_class;
3499 #else /* CONFIG_SMP */
3501 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3502 const struct cpumask *new_mask,
3505 return set_cpus_allowed_ptr(p, new_mask);
3508 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3510 static inline bool rq_has_pinned_tasks(struct rq *rq)
3515 #endif /* !CONFIG_SMP */
3518 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3522 if (!schedstat_enabled())
3528 if (cpu == rq->cpu) {
3529 __schedstat_inc(rq->ttwu_local);
3530 __schedstat_inc(p->stats.nr_wakeups_local);
3532 struct sched_domain *sd;
3534 __schedstat_inc(p->stats.nr_wakeups_remote);
3536 for_each_domain(rq->cpu, sd) {
3537 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3538 __schedstat_inc(sd->ttwu_wake_remote);
3545 if (wake_flags & WF_MIGRATED)
3546 __schedstat_inc(p->stats.nr_wakeups_migrate);
3547 #endif /* CONFIG_SMP */
3549 __schedstat_inc(rq->ttwu_count);
3550 __schedstat_inc(p->stats.nr_wakeups);
3552 if (wake_flags & WF_SYNC)
3553 __schedstat_inc(p->stats.nr_wakeups_sync);
3557 * Mark the task runnable and perform wakeup-preemption.
3559 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3560 struct rq_flags *rf)
3562 check_preempt_curr(rq, p, wake_flags);
3563 WRITE_ONCE(p->__state, TASK_RUNNING);
3564 trace_sched_wakeup(p);
3567 if (p->sched_class->task_woken) {
3569 * Our task @p is fully woken up and running; so it's safe to
3570 * drop the rq->lock, hereafter rq is only used for statistics.
3572 rq_unpin_lock(rq, rf);
3573 p->sched_class->task_woken(rq, p);
3574 rq_repin_lock(rq, rf);
3577 if (rq->idle_stamp) {
3578 u64 delta = rq_clock(rq) - rq->idle_stamp;
3579 u64 max = 2*rq->max_idle_balance_cost;
3581 update_avg(&rq->avg_idle, delta);
3583 if (rq->avg_idle > max)
3586 rq->wake_stamp = jiffies;
3587 rq->wake_avg_idle = rq->avg_idle / 2;
3595 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3596 struct rq_flags *rf)
3598 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3600 lockdep_assert_rq_held(rq);
3602 if (p->sched_contributes_to_load)
3603 rq->nr_uninterruptible--;
3606 if (wake_flags & WF_MIGRATED)
3607 en_flags |= ENQUEUE_MIGRATED;
3611 delayacct_blkio_end(p);
3612 atomic_dec(&task_rq(p)->nr_iowait);
3615 activate_task(rq, p, en_flags);
3616 ttwu_do_wakeup(rq, p, wake_flags, rf);
3620 * Consider @p being inside a wait loop:
3623 * set_current_state(TASK_UNINTERRUPTIBLE);
3630 * __set_current_state(TASK_RUNNING);
3632 * between set_current_state() and schedule(). In this case @p is still
3633 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3636 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3637 * then schedule() must still happen and p->state can be changed to
3638 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3639 * need to do a full wakeup with enqueue.
3641 * Returns: %true when the wakeup is done,
3644 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3650 rq = __task_rq_lock(p, &rf);
3651 if (task_on_rq_queued(p)) {
3652 /* check_preempt_curr() may use rq clock */
3653 update_rq_clock(rq);
3654 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3657 __task_rq_unlock(rq, &rf);
3663 void sched_ttwu_pending(void *arg)
3665 struct llist_node *llist = arg;
3666 struct rq *rq = this_rq();
3667 struct task_struct *p, *t;
3674 * rq::ttwu_pending racy indication of out-standing wakeups.
3675 * Races such that false-negatives are possible, since they
3676 * are shorter lived that false-positives would be.
3678 WRITE_ONCE(rq->ttwu_pending, 0);
3680 rq_lock_irqsave(rq, &rf);
3681 update_rq_clock(rq);
3683 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3684 if (WARN_ON_ONCE(p->on_cpu))
3685 smp_cond_load_acquire(&p->on_cpu, !VAL);
3687 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3688 set_task_cpu(p, cpu_of(rq));
3690 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3693 rq_unlock_irqrestore(rq, &rf);
3696 void send_call_function_single_ipi(int cpu)
3698 struct rq *rq = cpu_rq(cpu);
3700 if (!set_nr_if_polling(rq->idle))
3701 arch_send_call_function_single_ipi(cpu);
3703 trace_sched_wake_idle_without_ipi(cpu);
3707 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3708 * necessary. The wakee CPU on receipt of the IPI will queue the task
3709 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3710 * of the wakeup instead of the waker.
3712 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3714 struct rq *rq = cpu_rq(cpu);
3716 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3718 WRITE_ONCE(rq->ttwu_pending, 1);
3719 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3722 void wake_up_if_idle(int cpu)
3724 struct rq *rq = cpu_rq(cpu);
3729 if (!is_idle_task(rcu_dereference(rq->curr)))
3732 rq_lock_irqsave(rq, &rf);
3733 if (is_idle_task(rq->curr))
3735 /* Else CPU is not idle, do nothing here: */
3736 rq_unlock_irqrestore(rq, &rf);
3742 bool cpus_share_cache(int this_cpu, int that_cpu)
3744 if (this_cpu == that_cpu)
3747 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3750 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3753 * Do not complicate things with the async wake_list while the CPU is
3756 if (!cpu_active(cpu))
3760 * If the CPU does not share cache, then queue the task on the
3761 * remote rqs wakelist to avoid accessing remote data.
3763 if (!cpus_share_cache(smp_processor_id(), cpu))
3767 * If the task is descheduling and the only running task on the
3768 * CPU then use the wakelist to offload the task activation to
3769 * the soon-to-be-idle CPU as the current CPU is likely busy.
3770 * nr_running is checked to avoid unnecessary task stacking.
3772 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3778 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3780 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3781 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3784 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3785 __ttwu_queue_wakelist(p, cpu, wake_flags);
3792 #else /* !CONFIG_SMP */
3794 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3799 #endif /* CONFIG_SMP */
3801 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3803 struct rq *rq = cpu_rq(cpu);
3806 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3810 update_rq_clock(rq);
3811 ttwu_do_activate(rq, p, wake_flags, &rf);
3816 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3818 * The caller holds p::pi_lock if p != current or has preemption
3819 * disabled when p == current.
3821 * The rules of PREEMPT_RT saved_state:
3823 * The related locking code always holds p::pi_lock when updating
3824 * p::saved_state, which means the code is fully serialized in both cases.
3826 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3827 * bits set. This allows to distinguish all wakeup scenarios.
3829 static __always_inline
3830 bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3832 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3833 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3834 state != TASK_RTLOCK_WAIT);
3837 if (READ_ONCE(p->__state) & state) {
3842 #ifdef CONFIG_PREEMPT_RT
3844 * Saved state preserves the task state across blocking on
3845 * an RT lock. If the state matches, set p::saved_state to
3846 * TASK_RUNNING, but do not wake the task because it waits
3847 * for a lock wakeup. Also indicate success because from
3848 * the regular waker's point of view this has succeeded.
3850 * After acquiring the lock the task will restore p::__state
3851 * from p::saved_state which ensures that the regular
3852 * wakeup is not lost. The restore will also set
3853 * p::saved_state to TASK_RUNNING so any further tests will
3854 * not result in false positives vs. @success
3856 if (p->saved_state & state) {
3857 p->saved_state = TASK_RUNNING;
3865 * Notes on Program-Order guarantees on SMP systems.
3869 * The basic program-order guarantee on SMP systems is that when a task [t]
3870 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3871 * execution on its new CPU [c1].
3873 * For migration (of runnable tasks) this is provided by the following means:
3875 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3876 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3877 * rq(c1)->lock (if not at the same time, then in that order).
3878 * C) LOCK of the rq(c1)->lock scheduling in task
3880 * Release/acquire chaining guarantees that B happens after A and C after B.
3881 * Note: the CPU doing B need not be c0 or c1
3890 * UNLOCK rq(0)->lock
3892 * LOCK rq(0)->lock // orders against CPU0
3894 * UNLOCK rq(0)->lock
3898 * UNLOCK rq(1)->lock
3900 * LOCK rq(1)->lock // orders against CPU2
3903 * UNLOCK rq(1)->lock
3906 * BLOCKING -- aka. SLEEP + WAKEUP
3908 * For blocking we (obviously) need to provide the same guarantee as for
3909 * migration. However the means are completely different as there is no lock
3910 * chain to provide order. Instead we do:
3912 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3913 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3917 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3919 * LOCK rq(0)->lock LOCK X->pi_lock
3922 * smp_store_release(X->on_cpu, 0);
3924 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3930 * X->state = RUNNING
3931 * UNLOCK rq(2)->lock
3933 * LOCK rq(2)->lock // orders against CPU1
3936 * UNLOCK rq(2)->lock
3939 * UNLOCK rq(0)->lock
3942 * However, for wakeups there is a second guarantee we must provide, namely we
3943 * must ensure that CONDITION=1 done by the caller can not be reordered with
3944 * accesses to the task state; see try_to_wake_up() and set_current_state().
3948 * try_to_wake_up - wake up a thread
3949 * @p: the thread to be awakened
3950 * @state: the mask of task states that can be woken
3951 * @wake_flags: wake modifier flags (WF_*)
3953 * Conceptually does:
3955 * If (@state & @p->state) @p->state = TASK_RUNNING.
3957 * If the task was not queued/runnable, also place it back on a runqueue.
3959 * This function is atomic against schedule() which would dequeue the task.
3961 * It issues a full memory barrier before accessing @p->state, see the comment
3962 * with set_current_state().
3964 * Uses p->pi_lock to serialize against concurrent wake-ups.
3966 * Relies on p->pi_lock stabilizing:
3969 * - p->sched_task_group
3970 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3972 * Tries really hard to only take one task_rq(p)->lock for performance.
3973 * Takes rq->lock in:
3974 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3975 * - ttwu_queue() -- new rq, for enqueue of the task;
3976 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3978 * As a consequence we race really badly with just about everything. See the
3979 * many memory barriers and their comments for details.
3981 * Return: %true if @p->state changes (an actual wakeup was done),
3985 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3987 unsigned long flags;
3988 int cpu, success = 0;
3993 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3994 * == smp_processor_id()'. Together this means we can special
3995 * case the whole 'p->on_rq && ttwu_runnable()' case below
3996 * without taking any locks.
3999 * - we rely on Program-Order guarantees for all the ordering,
4000 * - we're serialized against set_special_state() by virtue of
4001 * it disabling IRQs (this allows not taking ->pi_lock).
4003 if (!ttwu_state_match(p, state, &success))
4006 trace_sched_waking(p);
4007 WRITE_ONCE(p->__state, TASK_RUNNING);
4008 trace_sched_wakeup(p);
4013 * If we are going to wake up a thread waiting for CONDITION we
4014 * need to ensure that CONDITION=1 done by the caller can not be
4015 * reordered with p->state check below. This pairs with smp_store_mb()
4016 * in set_current_state() that the waiting thread does.
4018 raw_spin_lock_irqsave(&p->pi_lock, flags);
4019 smp_mb__after_spinlock();
4020 if (!ttwu_state_match(p, state, &success))
4023 trace_sched_waking(p);
4026 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4027 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4028 * in smp_cond_load_acquire() below.
4030 * sched_ttwu_pending() try_to_wake_up()
4031 * STORE p->on_rq = 1 LOAD p->state
4034 * __schedule() (switch to task 'p')
4035 * LOCK rq->lock smp_rmb();
4036 * smp_mb__after_spinlock();
4040 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
4042 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4043 * __schedule(). See the comment for smp_mb__after_spinlock().
4045 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
4048 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
4053 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4054 * possible to, falsely, observe p->on_cpu == 0.
4056 * One must be running (->on_cpu == 1) in order to remove oneself
4057 * from the runqueue.
4059 * __schedule() (switch to task 'p') try_to_wake_up()
4060 * STORE p->on_cpu = 1 LOAD p->on_rq
4063 * __schedule() (put 'p' to sleep)
4064 * LOCK rq->lock smp_rmb();
4065 * smp_mb__after_spinlock();
4066 * STORE p->on_rq = 0 LOAD p->on_cpu
4068 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4069 * __schedule(). See the comment for smp_mb__after_spinlock().
4071 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4072 * schedule()'s deactivate_task() has 'happened' and p will no longer
4073 * care about it's own p->state. See the comment in __schedule().
4075 smp_acquire__after_ctrl_dep();
4078 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4079 * == 0), which means we need to do an enqueue, change p->state to
4080 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4081 * enqueue, such as ttwu_queue_wakelist().
4083 WRITE_ONCE(p->__state, TASK_WAKING);
4086 * If the owning (remote) CPU is still in the middle of schedule() with
4087 * this task as prev, considering queueing p on the remote CPUs wake_list
4088 * which potentially sends an IPI instead of spinning on p->on_cpu to
4089 * let the waker make forward progress. This is safe because IRQs are
4090 * disabled and the IPI will deliver after on_cpu is cleared.
4092 * Ensure we load task_cpu(p) after p->on_cpu:
4094 * set_task_cpu(p, cpu);
4095 * STORE p->cpu = @cpu
4096 * __schedule() (switch to task 'p')
4098 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4099 * STORE p->on_cpu = 1 LOAD p->cpu
4101 * to ensure we observe the correct CPU on which the task is currently
4104 if (smp_load_acquire(&p->on_cpu) &&
4105 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4109 * If the owning (remote) CPU is still in the middle of schedule() with
4110 * this task as prev, wait until it's done referencing the task.
4112 * Pairs with the smp_store_release() in finish_task().
4114 * This ensures that tasks getting woken will be fully ordered against
4115 * their previous state and preserve Program Order.
4117 smp_cond_load_acquire(&p->on_cpu, !VAL);
4119 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4120 if (task_cpu(p) != cpu) {
4122 delayacct_blkio_end(p);
4123 atomic_dec(&task_rq(p)->nr_iowait);
4126 wake_flags |= WF_MIGRATED;
4127 psi_ttwu_dequeue(p);
4128 set_task_cpu(p, cpu);
4132 #endif /* CONFIG_SMP */
4134 ttwu_queue(p, cpu, wake_flags);
4136 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4139 ttwu_stat(p, task_cpu(p), wake_flags);
4146 * task_call_func - Invoke a function on task in fixed state
4147 * @p: Process for which the function is to be invoked, can be @current.
4148 * @func: Function to invoke.
4149 * @arg: Argument to function.
4151 * Fix the task in it's current state by avoiding wakeups and or rq operations
4152 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4153 * to work out what the state is, if required. Given that @func can be invoked
4154 * with a runqueue lock held, it had better be quite lightweight.
4157 * Whatever @func returns
4159 int task_call_func(struct task_struct *p, task_call_f func, void *arg)
4161 struct rq *rq = NULL;
4166 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4168 state = READ_ONCE(p->__state);
4171 * Ensure we load p->on_rq after p->__state, otherwise it would be
4172 * possible to, falsely, observe p->on_rq == 0.
4174 * See try_to_wake_up() for a longer comment.
4179 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4180 * the task is blocked. Make sure to check @state since ttwu() can drop
4181 * locks at the end, see ttwu_queue_wakelist().
4183 if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
4184 rq = __task_rq_lock(p, &rf);
4187 * At this point the task is pinned; either:
4188 * - blocked and we're holding off wakeups (pi->lock)
4189 * - woken, and we're holding off enqueue (rq->lock)
4190 * - queued, and we're holding off schedule (rq->lock)
4191 * - running, and we're holding off de-schedule (rq->lock)
4193 * The called function (@func) can use: task_curr(), p->on_rq and
4194 * p->__state to differentiate between these states.
4201 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4206 * wake_up_process - Wake up a specific process
4207 * @p: The process to be woken up.
4209 * Attempt to wake up the nominated process and move it to the set of runnable
4212 * Return: 1 if the process was woken up, 0 if it was already running.
4214 * This function executes a full memory barrier before accessing the task state.
4216 int wake_up_process(struct task_struct *p)
4218 return try_to_wake_up(p, TASK_NORMAL, 0);
4220 EXPORT_SYMBOL(wake_up_process);
4222 int wake_up_state(struct task_struct *p, unsigned int state)
4224 return try_to_wake_up(p, state, 0);
4228 * Perform scheduler related setup for a newly forked process p.
4229 * p is forked by current.
4231 * __sched_fork() is basic setup used by init_idle() too:
4233 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4238 p->se.exec_start = 0;
4239 p->se.sum_exec_runtime = 0;
4240 p->se.prev_sum_exec_runtime = 0;
4241 p->se.nr_migrations = 0;
4243 INIT_LIST_HEAD(&p->se.group_node);
4245 #ifdef CONFIG_FAIR_GROUP_SCHED
4246 p->se.cfs_rq = NULL;
4249 #ifdef CONFIG_SCHEDSTATS
4250 /* Even if schedstat is disabled, there should not be garbage */
4251 memset(&p->stats, 0, sizeof(p->stats));
4254 RB_CLEAR_NODE(&p->dl.rb_node);
4255 init_dl_task_timer(&p->dl);
4256 init_dl_inactive_task_timer(&p->dl);
4257 __dl_clear_params(p);
4259 INIT_LIST_HEAD(&p->rt.run_list);
4261 p->rt.time_slice = sched_rr_timeslice;
4265 #ifdef CONFIG_PREEMPT_NOTIFIERS
4266 INIT_HLIST_HEAD(&p->preempt_notifiers);
4269 #ifdef CONFIG_COMPACTION
4270 p->capture_control = NULL;
4272 init_numa_balancing(clone_flags, p);
4274 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4275 p->migration_pending = NULL;
4279 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4281 #ifdef CONFIG_NUMA_BALANCING
4283 void set_numabalancing_state(bool enabled)
4286 static_branch_enable(&sched_numa_balancing);
4288 static_branch_disable(&sched_numa_balancing);
4291 #ifdef CONFIG_PROC_SYSCTL
4292 int sysctl_numa_balancing(struct ctl_table *table, int write,
4293 void *buffer, size_t *lenp, loff_t *ppos)
4297 int state = static_branch_likely(&sched_numa_balancing);
4299 if (write && !capable(CAP_SYS_ADMIN))
4304 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4308 set_numabalancing_state(state);
4314 #ifdef CONFIG_SCHEDSTATS
4316 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4318 static void set_schedstats(bool enabled)
4321 static_branch_enable(&sched_schedstats);
4323 static_branch_disable(&sched_schedstats);
4326 void force_schedstat_enabled(void)
4328 if (!schedstat_enabled()) {
4329 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4330 static_branch_enable(&sched_schedstats);
4334 static int __init setup_schedstats(char *str)
4340 if (!strcmp(str, "enable")) {
4341 set_schedstats(true);
4343 } else if (!strcmp(str, "disable")) {
4344 set_schedstats(false);
4349 pr_warn("Unable to parse schedstats=\n");
4353 __setup("schedstats=", setup_schedstats);
4355 #ifdef CONFIG_PROC_SYSCTL
4356 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4357 size_t *lenp, loff_t *ppos)
4361 int state = static_branch_likely(&sched_schedstats);
4363 if (write && !capable(CAP_SYS_ADMIN))
4368 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4372 set_schedstats(state);
4375 #endif /* CONFIG_PROC_SYSCTL */
4376 #endif /* CONFIG_SCHEDSTATS */
4379 * fork()/clone()-time setup:
4381 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4383 __sched_fork(clone_flags, p);
4385 * We mark the process as NEW here. This guarantees that
4386 * nobody will actually run it, and a signal or other external
4387 * event cannot wake it up and insert it on the runqueue either.
4389 p->__state = TASK_NEW;
4392 * Make sure we do not leak PI boosting priority to the child.
4394 p->prio = current->normal_prio;
4399 * Revert to default priority/policy on fork if requested.
4401 if (unlikely(p->sched_reset_on_fork)) {
4402 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4403 p->policy = SCHED_NORMAL;
4404 p->static_prio = NICE_TO_PRIO(0);
4406 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4407 p->static_prio = NICE_TO_PRIO(0);
4409 p->prio = p->normal_prio = p->static_prio;
4413 * We don't need the reset flag anymore after the fork. It has
4414 * fulfilled its duty:
4416 p->sched_reset_on_fork = 0;
4419 if (dl_prio(p->prio))
4421 else if (rt_prio(p->prio))
4422 p->sched_class = &rt_sched_class;
4424 p->sched_class = &fair_sched_class;
4426 init_entity_runnable_average(&p->se);
4428 #ifdef CONFIG_SCHED_INFO
4429 if (likely(sched_info_on()))
4430 memset(&p->sched_info, 0, sizeof(p->sched_info));
4432 #if defined(CONFIG_SMP)
4435 init_task_preempt_count(p);
4437 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4438 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4443 void sched_post_fork(struct task_struct *p, struct kernel_clone_args *kargs)
4445 unsigned long flags;
4446 #ifdef CONFIG_CGROUP_SCHED
4447 struct task_group *tg;
4450 raw_spin_lock_irqsave(&p->pi_lock, flags);
4451 #ifdef CONFIG_CGROUP_SCHED
4452 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4453 struct task_group, css);
4454 p->sched_task_group = autogroup_task_group(p, tg);
4458 * We're setting the CPU for the first time, we don't migrate,
4459 * so use __set_task_cpu().
4461 __set_task_cpu(p, smp_processor_id());
4462 if (p->sched_class->task_fork)
4463 p->sched_class->task_fork(p);
4464 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4466 uclamp_post_fork(p);
4469 unsigned long to_ratio(u64 period, u64 runtime)
4471 if (runtime == RUNTIME_INF)
4475 * Doing this here saves a lot of checks in all
4476 * the calling paths, and returning zero seems
4477 * safe for them anyway.
4482 return div64_u64(runtime << BW_SHIFT, period);
4486 * wake_up_new_task - wake up a newly created task for the first time.
4488 * This function will do some initial scheduler statistics housekeeping
4489 * that must be done for every newly created context, then puts the task
4490 * on the runqueue and wakes it.
4492 void wake_up_new_task(struct task_struct *p)
4497 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4498 WRITE_ONCE(p->__state, TASK_RUNNING);
4501 * Fork balancing, do it here and not earlier because:
4502 * - cpus_ptr can change in the fork path
4503 * - any previously selected CPU might disappear through hotplug
4505 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4506 * as we're not fully set-up yet.
4508 p->recent_used_cpu = task_cpu(p);
4510 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4512 rq = __task_rq_lock(p, &rf);
4513 update_rq_clock(rq);
4514 post_init_entity_util_avg(p);
4516 activate_task(rq, p, ENQUEUE_NOCLOCK);
4517 trace_sched_wakeup_new(p);
4518 check_preempt_curr(rq, p, WF_FORK);
4520 if (p->sched_class->task_woken) {
4522 * Nothing relies on rq->lock after this, so it's fine to
4525 rq_unpin_lock(rq, &rf);
4526 p->sched_class->task_woken(rq, p);
4527 rq_repin_lock(rq, &rf);
4530 task_rq_unlock(rq, p, &rf);
4533 #ifdef CONFIG_PREEMPT_NOTIFIERS
4535 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4537 void preempt_notifier_inc(void)
4539 static_branch_inc(&preempt_notifier_key);
4541 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4543 void preempt_notifier_dec(void)
4545 static_branch_dec(&preempt_notifier_key);
4547 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4550 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4551 * @notifier: notifier struct to register
4553 void preempt_notifier_register(struct preempt_notifier *notifier)
4555 if (!static_branch_unlikely(&preempt_notifier_key))
4556 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4558 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4560 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4563 * preempt_notifier_unregister - no longer interested in preemption notifications
4564 * @notifier: notifier struct to unregister
4566 * This is *not* safe to call from within a preemption notifier.
4568 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4570 hlist_del(¬ifier->link);
4572 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4574 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4576 struct preempt_notifier *notifier;
4578 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4579 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4582 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4584 if (static_branch_unlikely(&preempt_notifier_key))
4585 __fire_sched_in_preempt_notifiers(curr);
4589 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4590 struct task_struct *next)
4592 struct preempt_notifier *notifier;
4594 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4595 notifier->ops->sched_out(notifier, next);
4598 static __always_inline void
4599 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4600 struct task_struct *next)
4602 if (static_branch_unlikely(&preempt_notifier_key))
4603 __fire_sched_out_preempt_notifiers(curr, next);
4606 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4608 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4613 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4614 struct task_struct *next)
4618 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4620 static inline void prepare_task(struct task_struct *next)
4624 * Claim the task as running, we do this before switching to it
4625 * such that any running task will have this set.
4627 * See the ttwu() WF_ON_CPU case and its ordering comment.
4629 WRITE_ONCE(next->on_cpu, 1);
4633 static inline void finish_task(struct task_struct *prev)
4637 * This must be the very last reference to @prev from this CPU. After
4638 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4639 * must ensure this doesn't happen until the switch is completely
4642 * In particular, the load of prev->state in finish_task_switch() must
4643 * happen before this.
4645 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4647 smp_store_release(&prev->on_cpu, 0);
4653 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4655 void (*func)(struct rq *rq);
4656 struct callback_head *next;
4658 lockdep_assert_rq_held(rq);
4661 func = (void (*)(struct rq *))head->func;
4670 static void balance_push(struct rq *rq);
4672 struct callback_head balance_push_callback = {
4674 .func = (void (*)(struct callback_head *))balance_push,
4677 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4679 struct callback_head *head = rq->balance_callback;
4681 lockdep_assert_rq_held(rq);
4683 rq->balance_callback = NULL;
4688 static void __balance_callbacks(struct rq *rq)
4690 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4693 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4695 unsigned long flags;
4697 if (unlikely(head)) {
4698 raw_spin_rq_lock_irqsave(rq, flags);
4699 do_balance_callbacks(rq, head);
4700 raw_spin_rq_unlock_irqrestore(rq, flags);
4706 static inline void __balance_callbacks(struct rq *rq)
4710 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4715 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4722 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4725 * Since the runqueue lock will be released by the next
4726 * task (which is an invalid locking op but in the case
4727 * of the scheduler it's an obvious special-case), so we
4728 * do an early lockdep release here:
4730 rq_unpin_lock(rq, rf);
4731 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4732 #ifdef CONFIG_DEBUG_SPINLOCK
4733 /* this is a valid case when another task releases the spinlock */
4734 rq_lockp(rq)->owner = next;
4738 static inline void finish_lock_switch(struct rq *rq)
4741 * If we are tracking spinlock dependencies then we have to
4742 * fix up the runqueue lock - which gets 'carried over' from
4743 * prev into current:
4745 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4746 __balance_callbacks(rq);
4747 raw_spin_rq_unlock_irq(rq);
4751 * NOP if the arch has not defined these:
4754 #ifndef prepare_arch_switch
4755 # define prepare_arch_switch(next) do { } while (0)
4758 #ifndef finish_arch_post_lock_switch
4759 # define finish_arch_post_lock_switch() do { } while (0)
4762 static inline void kmap_local_sched_out(void)
4764 #ifdef CONFIG_KMAP_LOCAL
4765 if (unlikely(current->kmap_ctrl.idx))
4766 __kmap_local_sched_out();
4770 static inline void kmap_local_sched_in(void)
4772 #ifdef CONFIG_KMAP_LOCAL
4773 if (unlikely(current->kmap_ctrl.idx))
4774 __kmap_local_sched_in();
4779 * prepare_task_switch - prepare to switch tasks
4780 * @rq: the runqueue preparing to switch
4781 * @prev: the current task that is being switched out
4782 * @next: the task we are going to switch to.
4784 * This is called with the rq lock held and interrupts off. It must
4785 * be paired with a subsequent finish_task_switch after the context
4788 * prepare_task_switch sets up locking and calls architecture specific
4792 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4793 struct task_struct *next)
4795 kcov_prepare_switch(prev);
4796 sched_info_switch(rq, prev, next);
4797 perf_event_task_sched_out(prev, next);
4799 fire_sched_out_preempt_notifiers(prev, next);
4800 kmap_local_sched_out();
4802 prepare_arch_switch(next);
4806 * finish_task_switch - clean up after a task-switch
4807 * @prev: the thread we just switched away from.
4809 * finish_task_switch must be called after the context switch, paired
4810 * with a prepare_task_switch call before the context switch.
4811 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4812 * and do any other architecture-specific cleanup actions.
4814 * Note that we may have delayed dropping an mm in context_switch(). If
4815 * so, we finish that here outside of the runqueue lock. (Doing it
4816 * with the lock held can cause deadlocks; see schedule() for
4819 * The context switch have flipped the stack from under us and restored the
4820 * local variables which were saved when this task called schedule() in the
4821 * past. prev == current is still correct but we need to recalculate this_rq
4822 * because prev may have moved to another CPU.
4824 static struct rq *finish_task_switch(struct task_struct *prev)
4825 __releases(rq->lock)
4827 struct rq *rq = this_rq();
4828 struct mm_struct *mm = rq->prev_mm;
4832 * The previous task will have left us with a preempt_count of 2
4833 * because it left us after:
4836 * preempt_disable(); // 1
4838 * raw_spin_lock_irq(&rq->lock) // 2
4840 * Also, see FORK_PREEMPT_COUNT.
4842 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4843 "corrupted preempt_count: %s/%d/0x%x\n",
4844 current->comm, current->pid, preempt_count()))
4845 preempt_count_set(FORK_PREEMPT_COUNT);
4850 * A task struct has one reference for the use as "current".
4851 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4852 * schedule one last time. The schedule call will never return, and
4853 * the scheduled task must drop that reference.
4855 * We must observe prev->state before clearing prev->on_cpu (in
4856 * finish_task), otherwise a concurrent wakeup can get prev
4857 * running on another CPU and we could rave with its RUNNING -> DEAD
4858 * transition, resulting in a double drop.
4860 prev_state = READ_ONCE(prev->__state);
4861 vtime_task_switch(prev);
4862 perf_event_task_sched_in(prev, current);
4864 tick_nohz_task_switch();
4865 finish_lock_switch(rq);
4866 finish_arch_post_lock_switch();
4867 kcov_finish_switch(current);
4869 * kmap_local_sched_out() is invoked with rq::lock held and
4870 * interrupts disabled. There is no requirement for that, but the
4871 * sched out code does not have an interrupt enabled section.
4872 * Restoring the maps on sched in does not require interrupts being
4875 kmap_local_sched_in();
4877 fire_sched_in_preempt_notifiers(current);
4879 * When switching through a kernel thread, the loop in
4880 * membarrier_{private,global}_expedited() may have observed that
4881 * kernel thread and not issued an IPI. It is therefore possible to
4882 * schedule between user->kernel->user threads without passing though
4883 * switch_mm(). Membarrier requires a barrier after storing to
4884 * rq->curr, before returning to userspace, so provide them here:
4886 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4887 * provided by mmdrop(),
4888 * - a sync_core for SYNC_CORE.
4891 membarrier_mm_sync_core_before_usermode(mm);
4894 if (unlikely(prev_state == TASK_DEAD)) {
4895 if (prev->sched_class->task_dead)
4896 prev->sched_class->task_dead(prev);
4898 /* Task is done with its stack. */
4899 put_task_stack(prev);
4901 put_task_struct_rcu_user(prev);
4908 * schedule_tail - first thing a freshly forked thread must call.
4909 * @prev: the thread we just switched away from.
4911 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4912 __releases(rq->lock)
4915 * New tasks start with FORK_PREEMPT_COUNT, see there and
4916 * finish_task_switch() for details.
4918 * finish_task_switch() will drop rq->lock() and lower preempt_count
4919 * and the preempt_enable() will end up enabling preemption (on
4920 * PREEMPT_COUNT kernels).
4923 finish_task_switch(prev);
4926 if (current->set_child_tid)
4927 put_user(task_pid_vnr(current), current->set_child_tid);
4929 calculate_sigpending();
4933 * context_switch - switch to the new MM and the new thread's register state.
4935 static __always_inline struct rq *
4936 context_switch(struct rq *rq, struct task_struct *prev,
4937 struct task_struct *next, struct rq_flags *rf)
4939 prepare_task_switch(rq, prev, next);
4942 * For paravirt, this is coupled with an exit in switch_to to
4943 * combine the page table reload and the switch backend into
4946 arch_start_context_switch(prev);
4949 * kernel -> kernel lazy + transfer active
4950 * user -> kernel lazy + mmgrab() active
4952 * kernel -> user switch + mmdrop() active
4953 * user -> user switch
4955 if (!next->mm) { // to kernel
4956 enter_lazy_tlb(prev->active_mm, next);
4958 next->active_mm = prev->active_mm;
4959 if (prev->mm) // from user
4960 mmgrab(prev->active_mm);
4962 prev->active_mm = NULL;
4964 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4966 * sys_membarrier() requires an smp_mb() between setting
4967 * rq->curr / membarrier_switch_mm() and returning to userspace.
4969 * The below provides this either through switch_mm(), or in
4970 * case 'prev->active_mm == next->mm' through
4971 * finish_task_switch()'s mmdrop().
4973 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4975 if (!prev->mm) { // from kernel
4976 /* will mmdrop() in finish_task_switch(). */
4977 rq->prev_mm = prev->active_mm;
4978 prev->active_mm = NULL;
4982 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4984 prepare_lock_switch(rq, next, rf);
4986 /* Here we just switch the register state and the stack. */
4987 switch_to(prev, next, prev);
4990 return finish_task_switch(prev);
4994 * nr_running and nr_context_switches:
4996 * externally visible scheduler statistics: current number of runnable
4997 * threads, total number of context switches performed since bootup.
4999 unsigned int nr_running(void)
5001 unsigned int i, sum = 0;
5003 for_each_online_cpu(i)
5004 sum += cpu_rq(i)->nr_running;
5010 * Check if only the current task is running on the CPU.
5012 * Caution: this function does not check that the caller has disabled
5013 * preemption, thus the result might have a time-of-check-to-time-of-use
5014 * race. The caller is responsible to use it correctly, for example:
5016 * - from a non-preemptible section (of course)
5018 * - from a thread that is bound to a single CPU
5020 * - in a loop with very short iterations (e.g. a polling loop)
5022 bool single_task_running(void)
5024 return raw_rq()->nr_running == 1;
5026 EXPORT_SYMBOL(single_task_running);
5028 unsigned long long nr_context_switches(void)
5031 unsigned long long sum = 0;
5033 for_each_possible_cpu(i)
5034 sum += cpu_rq(i)->nr_switches;
5040 * Consumers of these two interfaces, like for example the cpuidle menu
5041 * governor, are using nonsensical data. Preferring shallow idle state selection
5042 * for a CPU that has IO-wait which might not even end up running the task when
5043 * it does become runnable.
5046 unsigned int nr_iowait_cpu(int cpu)
5048 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5052 * IO-wait accounting, and how it's mostly bollocks (on SMP).
5054 * The idea behind IO-wait account is to account the idle time that we could
5055 * have spend running if it were not for IO. That is, if we were to improve the
5056 * storage performance, we'd have a proportional reduction in IO-wait time.
5058 * This all works nicely on UP, where, when a task blocks on IO, we account
5059 * idle time as IO-wait, because if the storage were faster, it could've been
5060 * running and we'd not be idle.
5062 * This has been extended to SMP, by doing the same for each CPU. This however
5065 * Imagine for instance the case where two tasks block on one CPU, only the one
5066 * CPU will have IO-wait accounted, while the other has regular idle. Even
5067 * though, if the storage were faster, both could've ran at the same time,
5068 * utilising both CPUs.
5070 * This means, that when looking globally, the current IO-wait accounting on
5071 * SMP is a lower bound, by reason of under accounting.
5073 * Worse, since the numbers are provided per CPU, they are sometimes
5074 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5075 * associated with any one particular CPU, it can wake to another CPU than it
5076 * blocked on. This means the per CPU IO-wait number is meaningless.
5078 * Task CPU affinities can make all that even more 'interesting'.
5081 unsigned int nr_iowait(void)
5083 unsigned int i, sum = 0;
5085 for_each_possible_cpu(i)
5086 sum += nr_iowait_cpu(i);
5094 * sched_exec - execve() is a valuable balancing opportunity, because at
5095 * this point the task has the smallest effective memory and cache footprint.
5097 void sched_exec(void)
5099 struct task_struct *p = current;
5100 unsigned long flags;
5103 raw_spin_lock_irqsave(&p->pi_lock, flags);
5104 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
5105 if (dest_cpu == smp_processor_id())
5108 if (likely(cpu_active(dest_cpu))) {
5109 struct migration_arg arg = { p, dest_cpu };
5111 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5112 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5116 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5121 DEFINE_PER_CPU(struct kernel_stat, kstat);
5122 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5124 EXPORT_PER_CPU_SYMBOL(kstat);
5125 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5128 * The function fair_sched_class.update_curr accesses the struct curr
5129 * and its field curr->exec_start; when called from task_sched_runtime(),
5130 * we observe a high rate of cache misses in practice.
5131 * Prefetching this data results in improved performance.
5133 static inline void prefetch_curr_exec_start(struct task_struct *p)
5135 #ifdef CONFIG_FAIR_GROUP_SCHED
5136 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5138 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5141 prefetch(&curr->exec_start);
5145 * Return accounted runtime for the task.
5146 * In case the task is currently running, return the runtime plus current's
5147 * pending runtime that have not been accounted yet.
5149 unsigned long long task_sched_runtime(struct task_struct *p)
5155 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5157 * 64-bit doesn't need locks to atomically read a 64-bit value.
5158 * So we have a optimization chance when the task's delta_exec is 0.
5159 * Reading ->on_cpu is racy, but this is ok.
5161 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5162 * If we race with it entering CPU, unaccounted time is 0. This is
5163 * indistinguishable from the read occurring a few cycles earlier.
5164 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5165 * been accounted, so we're correct here as well.
5167 if (!p->on_cpu || !task_on_rq_queued(p))
5168 return p->se.sum_exec_runtime;
5171 rq = task_rq_lock(p, &rf);
5173 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5174 * project cycles that may never be accounted to this
5175 * thread, breaking clock_gettime().
5177 if (task_current(rq, p) && task_on_rq_queued(p)) {
5178 prefetch_curr_exec_start(p);
5179 update_rq_clock(rq);
5180 p->sched_class->update_curr(rq);
5182 ns = p->se.sum_exec_runtime;
5183 task_rq_unlock(rq, p, &rf);
5188 #ifdef CONFIG_SCHED_DEBUG
5189 static u64 cpu_resched_latency(struct rq *rq)
5191 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5192 u64 resched_latency, now = rq_clock(rq);
5193 static bool warned_once;
5195 if (sysctl_resched_latency_warn_once && warned_once)
5198 if (!need_resched() || !latency_warn_ms)
5201 if (system_state == SYSTEM_BOOTING)
5204 if (!rq->last_seen_need_resched_ns) {
5205 rq->last_seen_need_resched_ns = now;
5206 rq->ticks_without_resched = 0;
5210 rq->ticks_without_resched++;
5211 resched_latency = now - rq->last_seen_need_resched_ns;
5212 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5217 return resched_latency;
5220 static int __init setup_resched_latency_warn_ms(char *str)
5224 if ((kstrtol(str, 0, &val))) {
5225 pr_warn("Unable to set resched_latency_warn_ms\n");
5229 sysctl_resched_latency_warn_ms = val;
5232 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5234 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5235 #endif /* CONFIG_SCHED_DEBUG */
5238 * This function gets called by the timer code, with HZ frequency.
5239 * We call it with interrupts disabled.
5241 void scheduler_tick(void)
5243 int cpu = smp_processor_id();
5244 struct rq *rq = cpu_rq(cpu);
5245 struct task_struct *curr = rq->curr;
5247 unsigned long thermal_pressure;
5248 u64 resched_latency;
5250 arch_scale_freq_tick();
5255 update_rq_clock(rq);
5256 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5257 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5258 curr->sched_class->task_tick(rq, curr, 0);
5259 if (sched_feat(LATENCY_WARN))
5260 resched_latency = cpu_resched_latency(rq);
5261 calc_global_load_tick(rq);
5262 sched_core_tick(rq);
5266 if (sched_feat(LATENCY_WARN) && resched_latency)
5267 resched_latency_warn(cpu, resched_latency);
5269 perf_event_task_tick();
5272 rq->idle_balance = idle_cpu(cpu);
5273 trigger_load_balance(rq);
5277 #ifdef CONFIG_NO_HZ_FULL
5282 struct delayed_work work;
5284 /* Values for ->state, see diagram below. */
5285 #define TICK_SCHED_REMOTE_OFFLINE 0
5286 #define TICK_SCHED_REMOTE_OFFLINING 1
5287 #define TICK_SCHED_REMOTE_RUNNING 2
5290 * State diagram for ->state:
5293 * TICK_SCHED_REMOTE_OFFLINE
5296 * | | sched_tick_remote()
5299 * +--TICK_SCHED_REMOTE_OFFLINING
5302 * sched_tick_start() | | sched_tick_stop()
5305 * TICK_SCHED_REMOTE_RUNNING
5308 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5309 * and sched_tick_start() are happy to leave the state in RUNNING.
5312 static struct tick_work __percpu *tick_work_cpu;
5314 static void sched_tick_remote(struct work_struct *work)
5316 struct delayed_work *dwork = to_delayed_work(work);
5317 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5318 int cpu = twork->cpu;
5319 struct rq *rq = cpu_rq(cpu);
5320 struct task_struct *curr;
5326 * Handle the tick only if it appears the remote CPU is running in full
5327 * dynticks mode. The check is racy by nature, but missing a tick or
5328 * having one too much is no big deal because the scheduler tick updates
5329 * statistics and checks timeslices in a time-independent way, regardless
5330 * of when exactly it is running.
5332 if (!tick_nohz_tick_stopped_cpu(cpu))
5335 rq_lock_irq(rq, &rf);
5337 if (cpu_is_offline(cpu))
5340 update_rq_clock(rq);
5342 if (!is_idle_task(curr)) {
5344 * Make sure the next tick runs within a reasonable
5347 delta = rq_clock_task(rq) - curr->se.exec_start;
5348 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5350 curr->sched_class->task_tick(rq, curr, 0);
5352 calc_load_nohz_remote(rq);
5354 rq_unlock_irq(rq, &rf);
5358 * Run the remote tick once per second (1Hz). This arbitrary
5359 * frequency is large enough to avoid overload but short enough
5360 * to keep scheduler internal stats reasonably up to date. But
5361 * first update state to reflect hotplug activity if required.
5363 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5364 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5365 if (os == TICK_SCHED_REMOTE_RUNNING)
5366 queue_delayed_work(system_unbound_wq, dwork, HZ);
5369 static void sched_tick_start(int cpu)
5372 struct tick_work *twork;
5374 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5377 WARN_ON_ONCE(!tick_work_cpu);
5379 twork = per_cpu_ptr(tick_work_cpu, cpu);
5380 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5381 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5382 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5384 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5385 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5389 #ifdef CONFIG_HOTPLUG_CPU
5390 static void sched_tick_stop(int cpu)
5392 struct tick_work *twork;
5395 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5398 WARN_ON_ONCE(!tick_work_cpu);
5400 twork = per_cpu_ptr(tick_work_cpu, cpu);
5401 /* There cannot be competing actions, but don't rely on stop-machine. */
5402 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5403 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5404 /* Don't cancel, as this would mess up the state machine. */
5406 #endif /* CONFIG_HOTPLUG_CPU */
5408 int __init sched_tick_offload_init(void)
5410 tick_work_cpu = alloc_percpu(struct tick_work);
5411 BUG_ON(!tick_work_cpu);
5415 #else /* !CONFIG_NO_HZ_FULL */
5416 static inline void sched_tick_start(int cpu) { }
5417 static inline void sched_tick_stop(int cpu) { }
5420 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5421 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5423 * If the value passed in is equal to the current preempt count
5424 * then we just disabled preemption. Start timing the latency.
5426 static inline void preempt_latency_start(int val)
5428 if (preempt_count() == val) {
5429 unsigned long ip = get_lock_parent_ip();
5430 #ifdef CONFIG_DEBUG_PREEMPT
5431 current->preempt_disable_ip = ip;
5433 trace_preempt_off(CALLER_ADDR0, ip);
5437 void preempt_count_add(int val)
5439 #ifdef CONFIG_DEBUG_PREEMPT
5443 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5446 __preempt_count_add(val);
5447 #ifdef CONFIG_DEBUG_PREEMPT
5449 * Spinlock count overflowing soon?
5451 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5454 preempt_latency_start(val);
5456 EXPORT_SYMBOL(preempt_count_add);
5457 NOKPROBE_SYMBOL(preempt_count_add);
5460 * If the value passed in equals to the current preempt count
5461 * then we just enabled preemption. Stop timing the latency.
5463 static inline void preempt_latency_stop(int val)
5465 if (preempt_count() == val)
5466 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5469 void preempt_count_sub(int val)
5471 #ifdef CONFIG_DEBUG_PREEMPT
5475 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5478 * Is the spinlock portion underflowing?
5480 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5481 !(preempt_count() & PREEMPT_MASK)))
5485 preempt_latency_stop(val);
5486 __preempt_count_sub(val);
5488 EXPORT_SYMBOL(preempt_count_sub);
5489 NOKPROBE_SYMBOL(preempt_count_sub);
5492 static inline void preempt_latency_start(int val) { }
5493 static inline void preempt_latency_stop(int val) { }
5496 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5498 #ifdef CONFIG_DEBUG_PREEMPT
5499 return p->preempt_disable_ip;
5506 * Print scheduling while atomic bug:
5508 static noinline void __schedule_bug(struct task_struct *prev)
5510 /* Save this before calling printk(), since that will clobber it */
5511 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5513 if (oops_in_progress)
5516 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5517 prev->comm, prev->pid, preempt_count());
5519 debug_show_held_locks(prev);
5521 if (irqs_disabled())
5522 print_irqtrace_events(prev);
5523 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5524 && in_atomic_preempt_off()) {
5525 pr_err("Preemption disabled at:");
5526 print_ip_sym(KERN_ERR, preempt_disable_ip);
5529 panic("scheduling while atomic\n");
5532 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5536 * Various schedule()-time debugging checks and statistics:
5538 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5540 #ifdef CONFIG_SCHED_STACK_END_CHECK
5541 if (task_stack_end_corrupted(prev))
5542 panic("corrupted stack end detected inside scheduler\n");
5544 if (task_scs_end_corrupted(prev))
5545 panic("corrupted shadow stack detected inside scheduler\n");
5548 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5549 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5550 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5551 prev->comm, prev->pid, prev->non_block_count);
5553 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5557 if (unlikely(in_atomic_preempt_off())) {
5558 __schedule_bug(prev);
5559 preempt_count_set(PREEMPT_DISABLED);
5562 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5564 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5566 schedstat_inc(this_rq()->sched_count);
5569 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5570 struct rq_flags *rf)
5573 const struct sched_class *class;
5575 * We must do the balancing pass before put_prev_task(), such
5576 * that when we release the rq->lock the task is in the same
5577 * state as before we took rq->lock.
5579 * We can terminate the balance pass as soon as we know there is
5580 * a runnable task of @class priority or higher.
5582 for_class_range(class, prev->sched_class, &idle_sched_class) {
5583 if (class->balance(rq, prev, rf))
5588 put_prev_task(rq, prev);
5592 * Pick up the highest-prio task:
5594 static inline struct task_struct *
5595 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5597 const struct sched_class *class;
5598 struct task_struct *p;
5601 * Optimization: we know that if all tasks are in the fair class we can
5602 * call that function directly, but only if the @prev task wasn't of a
5603 * higher scheduling class, because otherwise those lose the
5604 * opportunity to pull in more work from other CPUs.
5606 if (likely(prev->sched_class <= &fair_sched_class &&
5607 rq->nr_running == rq->cfs.h_nr_running)) {
5609 p = pick_next_task_fair(rq, prev, rf);
5610 if (unlikely(p == RETRY_TASK))
5613 /* Assume the next prioritized class is idle_sched_class */
5615 put_prev_task(rq, prev);
5616 p = pick_next_task_idle(rq);
5623 put_prev_task_balance(rq, prev, rf);
5625 for_each_class(class) {
5626 p = class->pick_next_task(rq);
5631 BUG(); /* The idle class should always have a runnable task. */
5634 #ifdef CONFIG_SCHED_CORE
5635 static inline bool is_task_rq_idle(struct task_struct *t)
5637 return (task_rq(t)->idle == t);
5640 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5642 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5645 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5647 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5650 return a->core_cookie == b->core_cookie;
5653 static inline struct task_struct *pick_task(struct rq *rq)
5655 const struct sched_class *class;
5656 struct task_struct *p;
5658 for_each_class(class) {
5659 p = class->pick_task(rq);
5664 BUG(); /* The idle class should always have a runnable task. */
5667 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5669 static struct task_struct *
5670 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5672 struct task_struct *next, *p, *max = NULL;
5673 const struct cpumask *smt_mask;
5674 bool fi_before = false;
5675 bool core_clock_updated = (rq == rq->core);
5676 unsigned long cookie;
5677 int i, cpu, occ = 0;
5681 if (!sched_core_enabled(rq))
5682 return __pick_next_task(rq, prev, rf);
5686 /* Stopper task is switching into idle, no need core-wide selection. */
5687 if (cpu_is_offline(cpu)) {
5689 * Reset core_pick so that we don't enter the fastpath when
5690 * coming online. core_pick would already be migrated to
5691 * another cpu during offline.
5693 rq->core_pick = NULL;
5694 return __pick_next_task(rq, prev, rf);
5698 * If there were no {en,de}queues since we picked (IOW, the task
5699 * pointers are all still valid), and we haven't scheduled the last
5700 * pick yet, do so now.
5702 * rq->core_pick can be NULL if no selection was made for a CPU because
5703 * it was either offline or went offline during a sibling's core-wide
5704 * selection. In this case, do a core-wide selection.
5706 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5707 rq->core->core_pick_seq != rq->core_sched_seq &&
5709 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5711 next = rq->core_pick;
5713 put_prev_task(rq, prev);
5714 set_next_task(rq, next);
5717 rq->core_pick = NULL;
5721 put_prev_task_balance(rq, prev, rf);
5723 smt_mask = cpu_smt_mask(cpu);
5724 need_sync = !!rq->core->core_cookie;
5727 rq->core->core_cookie = 0UL;
5728 if (rq->core->core_forceidle_count) {
5729 if (!core_clock_updated) {
5730 update_rq_clock(rq->core);
5731 core_clock_updated = true;
5733 sched_core_account_forceidle(rq);
5734 /* reset after accounting force idle */
5735 rq->core->core_forceidle_start = 0;
5736 rq->core->core_forceidle_count = 0;
5737 rq->core->core_forceidle_occupation = 0;
5743 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5745 * @task_seq guards the task state ({en,de}queues)
5746 * @pick_seq is the @task_seq we did a selection on
5747 * @sched_seq is the @pick_seq we scheduled
5749 * However, preemptions can cause multiple picks on the same task set.
5750 * 'Fix' this by also increasing @task_seq for every pick.
5752 rq->core->core_task_seq++;
5755 * Optimize for common case where this CPU has no cookies
5756 * and there are no cookied tasks running on siblings.
5759 next = pick_task(rq);
5760 if (!next->core_cookie) {
5761 rq->core_pick = NULL;
5763 * For robustness, update the min_vruntime_fi for
5764 * unconstrained picks as well.
5766 WARN_ON_ONCE(fi_before);
5767 task_vruntime_update(rq, next, false);
5773 * For each thread: do the regular task pick and find the max prio task
5776 * Tie-break prio towards the current CPU
5778 for_each_cpu_wrap(i, smt_mask, cpu) {
5782 * Current cpu always has its clock updated on entrance to
5783 * pick_next_task(). If the current cpu is not the core,
5784 * the core may also have been updated above.
5786 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
5787 update_rq_clock(rq_i);
5789 p = rq_i->core_pick = pick_task(rq_i);
5790 if (!max || prio_less(max, p, fi_before))
5794 cookie = rq->core->core_cookie = max->core_cookie;
5797 * For each thread: try and find a runnable task that matches @max or
5800 for_each_cpu(i, smt_mask) {
5802 p = rq_i->core_pick;
5804 if (!cookie_equals(p, cookie)) {
5807 p = sched_core_find(rq_i, cookie);
5809 p = idle_sched_class.pick_task(rq_i);
5812 rq_i->core_pick = p;
5814 if (p == rq_i->idle) {
5815 if (rq_i->nr_running) {
5816 rq->core->core_forceidle_count++;
5818 rq->core->core_forceidle_seq++;
5825 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);
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);
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) {
8222 if (!_cond_resched())
8229 EXPORT_SYMBOL(__cond_resched_lock);
8231 int __cond_resched_rwlock_read(rwlock_t *lock)
8233 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8236 lockdep_assert_held_read(lock);
8238 if (rwlock_needbreak(lock) || resched) {
8240 if (!_cond_resched())
8247 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8249 int __cond_resched_rwlock_write(rwlock_t *lock)
8251 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8254 lockdep_assert_held_write(lock);
8256 if (rwlock_needbreak(lock) || resched) {
8258 if (!_cond_resched())
8265 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8268 * yield - yield the current processor to other threads.
8270 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8272 * The scheduler is at all times free to pick the calling task as the most
8273 * eligible task to run, if removing the yield() call from your code breaks
8274 * it, it's already broken.
8276 * Typical broken usage is:
8281 * where one assumes that yield() will let 'the other' process run that will
8282 * make event true. If the current task is a SCHED_FIFO task that will never
8283 * happen. Never use yield() as a progress guarantee!!
8285 * If you want to use yield() to wait for something, use wait_event().
8286 * If you want to use yield() to be 'nice' for others, use cond_resched().
8287 * If you still want to use yield(), do not!
8289 void __sched yield(void)
8291 set_current_state(TASK_RUNNING);
8294 EXPORT_SYMBOL(yield);
8297 * yield_to - yield the current processor to another thread in
8298 * your thread group, or accelerate that thread toward the
8299 * processor it's on.
8301 * @preempt: whether task preemption is allowed or not
8303 * It's the caller's job to ensure that the target task struct
8304 * can't go away on us before we can do any checks.
8307 * true (>0) if we indeed boosted the target task.
8308 * false (0) if we failed to boost the target.
8309 * -ESRCH if there's no task to yield to.
8311 int __sched yield_to(struct task_struct *p, bool preempt)
8313 struct task_struct *curr = current;
8314 struct rq *rq, *p_rq;
8315 unsigned long flags;
8318 local_irq_save(flags);
8324 * If we're the only runnable task on the rq and target rq also
8325 * has only one task, there's absolutely no point in yielding.
8327 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8332 double_rq_lock(rq, p_rq);
8333 if (task_rq(p) != p_rq) {
8334 double_rq_unlock(rq, p_rq);
8338 if (!curr->sched_class->yield_to_task)
8341 if (curr->sched_class != p->sched_class)
8344 if (task_running(p_rq, p) || !task_is_running(p))
8347 yielded = curr->sched_class->yield_to_task(rq, p);
8349 schedstat_inc(rq->yld_count);
8351 * Make p's CPU reschedule; pick_next_entity takes care of
8354 if (preempt && rq != p_rq)
8359 double_rq_unlock(rq, p_rq);
8361 local_irq_restore(flags);
8368 EXPORT_SYMBOL_GPL(yield_to);
8370 int io_schedule_prepare(void)
8372 int old_iowait = current->in_iowait;
8374 current->in_iowait = 1;
8376 blk_flush_plug(current->plug, true);
8381 void io_schedule_finish(int token)
8383 current->in_iowait = token;
8387 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8388 * that process accounting knows that this is a task in IO wait state.
8390 long __sched io_schedule_timeout(long timeout)
8395 token = io_schedule_prepare();
8396 ret = schedule_timeout(timeout);
8397 io_schedule_finish(token);
8401 EXPORT_SYMBOL(io_schedule_timeout);
8403 void __sched io_schedule(void)
8407 token = io_schedule_prepare();
8409 io_schedule_finish(token);
8411 EXPORT_SYMBOL(io_schedule);
8414 * sys_sched_get_priority_max - return maximum RT priority.
8415 * @policy: scheduling class.
8417 * Return: On success, this syscall returns the maximum
8418 * rt_priority that can be used by a given scheduling class.
8419 * On failure, a negative error code is returned.
8421 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8428 ret = MAX_RT_PRIO-1;
8430 case SCHED_DEADLINE:
8441 * sys_sched_get_priority_min - return minimum RT priority.
8442 * @policy: scheduling class.
8444 * Return: On success, this syscall returns the minimum
8445 * rt_priority that can be used by a given scheduling class.
8446 * On failure, a negative error code is returned.
8448 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8457 case SCHED_DEADLINE:
8466 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8468 struct task_struct *p;
8469 unsigned int time_slice;
8479 p = find_process_by_pid(pid);
8483 retval = security_task_getscheduler(p);
8487 rq = task_rq_lock(p, &rf);
8489 if (p->sched_class->get_rr_interval)
8490 time_slice = p->sched_class->get_rr_interval(rq, p);
8491 task_rq_unlock(rq, p, &rf);
8494 jiffies_to_timespec64(time_slice, t);
8503 * sys_sched_rr_get_interval - return the default timeslice of a process.
8504 * @pid: pid of the process.
8505 * @interval: userspace pointer to the timeslice value.
8507 * this syscall writes the default timeslice value of a given process
8508 * into the user-space timespec buffer. A value of '0' means infinity.
8510 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8513 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8514 struct __kernel_timespec __user *, interval)
8516 struct timespec64 t;
8517 int retval = sched_rr_get_interval(pid, &t);
8520 retval = put_timespec64(&t, interval);
8525 #ifdef CONFIG_COMPAT_32BIT_TIME
8526 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8527 struct old_timespec32 __user *, interval)
8529 struct timespec64 t;
8530 int retval = sched_rr_get_interval(pid, &t);
8533 retval = put_old_timespec32(&t, interval);
8538 void sched_show_task(struct task_struct *p)
8540 unsigned long free = 0;
8543 if (!try_get_task_stack(p))
8546 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8548 if (task_is_running(p))
8549 pr_cont(" running task ");
8550 #ifdef CONFIG_DEBUG_STACK_USAGE
8551 free = stack_not_used(p);
8556 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8558 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8559 free, task_pid_nr(p), ppid,
8560 read_task_thread_flags(p));
8562 print_worker_info(KERN_INFO, p);
8563 print_stop_info(KERN_INFO, p);
8564 show_stack(p, NULL, KERN_INFO);
8567 EXPORT_SYMBOL_GPL(sched_show_task);
8570 state_filter_match(unsigned long state_filter, struct task_struct *p)
8572 unsigned int state = READ_ONCE(p->__state);
8574 /* no filter, everything matches */
8578 /* filter, but doesn't match */
8579 if (!(state & state_filter))
8583 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8586 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8593 void show_state_filter(unsigned int state_filter)
8595 struct task_struct *g, *p;
8598 for_each_process_thread(g, p) {
8600 * reset the NMI-timeout, listing all files on a slow
8601 * console might take a lot of time:
8602 * Also, reset softlockup watchdogs on all CPUs, because
8603 * another CPU might be blocked waiting for us to process
8606 touch_nmi_watchdog();
8607 touch_all_softlockup_watchdogs();
8608 if (state_filter_match(state_filter, p))
8612 #ifdef CONFIG_SCHED_DEBUG
8614 sysrq_sched_debug_show();
8618 * Only show locks if all tasks are dumped:
8621 debug_show_all_locks();
8625 * init_idle - set up an idle thread for a given CPU
8626 * @idle: task in question
8627 * @cpu: CPU the idle task belongs to
8629 * NOTE: this function does not set the idle thread's NEED_RESCHED
8630 * flag, to make booting more robust.
8632 void __init init_idle(struct task_struct *idle, int cpu)
8634 struct rq *rq = cpu_rq(cpu);
8635 unsigned long flags;
8637 __sched_fork(0, idle);
8639 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8640 raw_spin_rq_lock(rq);
8642 idle->__state = TASK_RUNNING;
8643 idle->se.exec_start = sched_clock();
8645 * PF_KTHREAD should already be set at this point; regardless, make it
8646 * look like a proper per-CPU kthread.
8648 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8649 kthread_set_per_cpu(idle, cpu);
8653 * It's possible that init_idle() gets called multiple times on a task,
8654 * in that case do_set_cpus_allowed() will not do the right thing.
8656 * And since this is boot we can forgo the serialization.
8658 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8661 * We're having a chicken and egg problem, even though we are
8662 * holding rq->lock, the CPU isn't yet set to this CPU so the
8663 * lockdep check in task_group() will fail.
8665 * Similar case to sched_fork(). / Alternatively we could
8666 * use task_rq_lock() here and obtain the other rq->lock.
8671 __set_task_cpu(idle, cpu);
8675 rcu_assign_pointer(rq->curr, idle);
8676 idle->on_rq = TASK_ON_RQ_QUEUED;
8680 raw_spin_rq_unlock(rq);
8681 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8683 /* Set the preempt count _outside_ the spinlocks! */
8684 init_idle_preempt_count(idle, cpu);
8687 * The idle tasks have their own, simple scheduling class:
8689 idle->sched_class = &idle_sched_class;
8690 ftrace_graph_init_idle_task(idle, cpu);
8691 vtime_init_idle(idle, cpu);
8693 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8699 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8700 const struct cpumask *trial)
8704 if (!cpumask_weight(cur))
8707 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8712 int task_can_attach(struct task_struct *p,
8713 const struct cpumask *cs_cpus_allowed)
8718 * Kthreads which disallow setaffinity shouldn't be moved
8719 * to a new cpuset; we don't want to change their CPU
8720 * affinity and isolating such threads by their set of
8721 * allowed nodes is unnecessary. Thus, cpusets are not
8722 * applicable for such threads. This prevents checking for
8723 * success of set_cpus_allowed_ptr() on all attached tasks
8724 * before cpus_mask may be changed.
8726 if (p->flags & PF_NO_SETAFFINITY) {
8731 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8733 ret = dl_task_can_attach(p, cs_cpus_allowed);
8739 bool sched_smp_initialized __read_mostly;
8741 #ifdef CONFIG_NUMA_BALANCING
8742 /* Migrate current task p to target_cpu */
8743 int migrate_task_to(struct task_struct *p, int target_cpu)
8745 struct migration_arg arg = { p, target_cpu };
8746 int curr_cpu = task_cpu(p);
8748 if (curr_cpu == target_cpu)
8751 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8754 /* TODO: This is not properly updating schedstats */
8756 trace_sched_move_numa(p, curr_cpu, target_cpu);
8757 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8761 * Requeue a task on a given node and accurately track the number of NUMA
8762 * tasks on the runqueues
8764 void sched_setnuma(struct task_struct *p, int nid)
8766 bool queued, running;
8770 rq = task_rq_lock(p, &rf);
8771 queued = task_on_rq_queued(p);
8772 running = task_current(rq, p);
8775 dequeue_task(rq, p, DEQUEUE_SAVE);
8777 put_prev_task(rq, p);
8779 p->numa_preferred_nid = nid;
8782 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8784 set_next_task(rq, p);
8785 task_rq_unlock(rq, p, &rf);
8787 #endif /* CONFIG_NUMA_BALANCING */
8789 #ifdef CONFIG_HOTPLUG_CPU
8791 * Ensure that the idle task is using init_mm right before its CPU goes
8794 void idle_task_exit(void)
8796 struct mm_struct *mm = current->active_mm;
8798 BUG_ON(cpu_online(smp_processor_id()));
8799 BUG_ON(current != this_rq()->idle);
8801 if (mm != &init_mm) {
8802 switch_mm(mm, &init_mm, current);
8803 finish_arch_post_lock_switch();
8806 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8809 static int __balance_push_cpu_stop(void *arg)
8811 struct task_struct *p = arg;
8812 struct rq *rq = this_rq();
8816 raw_spin_lock_irq(&p->pi_lock);
8819 update_rq_clock(rq);
8821 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8822 cpu = select_fallback_rq(rq->cpu, p);
8823 rq = __migrate_task(rq, &rf, p, cpu);
8827 raw_spin_unlock_irq(&p->pi_lock);
8834 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8837 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8839 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8840 * effective when the hotplug motion is down.
8842 static void balance_push(struct rq *rq)
8844 struct task_struct *push_task = rq->curr;
8846 lockdep_assert_rq_held(rq);
8849 * Ensure the thing is persistent until balance_push_set(.on = false);
8851 rq->balance_callback = &balance_push_callback;
8854 * Only active while going offline and when invoked on the outgoing
8857 if (!cpu_dying(rq->cpu) || rq != this_rq())
8861 * Both the cpu-hotplug and stop task are in this case and are
8862 * required to complete the hotplug process.
8864 if (kthread_is_per_cpu(push_task) ||
8865 is_migration_disabled(push_task)) {
8868 * If this is the idle task on the outgoing CPU try to wake
8869 * up the hotplug control thread which might wait for the
8870 * last task to vanish. The rcuwait_active() check is
8871 * accurate here because the waiter is pinned on this CPU
8872 * and can't obviously be running in parallel.
8874 * On RT kernels this also has to check whether there are
8875 * pinned and scheduled out tasks on the runqueue. They
8876 * need to leave the migrate disabled section first.
8878 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8879 rcuwait_active(&rq->hotplug_wait)) {
8880 raw_spin_rq_unlock(rq);
8881 rcuwait_wake_up(&rq->hotplug_wait);
8882 raw_spin_rq_lock(rq);
8887 get_task_struct(push_task);
8889 * Temporarily drop rq->lock such that we can wake-up the stop task.
8890 * Both preemption and IRQs are still disabled.
8892 raw_spin_rq_unlock(rq);
8893 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8894 this_cpu_ptr(&push_work));
8896 * At this point need_resched() is true and we'll take the loop in
8897 * schedule(). The next pick is obviously going to be the stop task
8898 * which kthread_is_per_cpu() and will push this task away.
8900 raw_spin_rq_lock(rq);
8903 static void balance_push_set(int cpu, bool on)
8905 struct rq *rq = cpu_rq(cpu);
8908 rq_lock_irqsave(rq, &rf);
8910 WARN_ON_ONCE(rq->balance_callback);
8911 rq->balance_callback = &balance_push_callback;
8912 } else if (rq->balance_callback == &balance_push_callback) {
8913 rq->balance_callback = NULL;
8915 rq_unlock_irqrestore(rq, &rf);
8919 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8920 * inactive. All tasks which are not per CPU kernel threads are either
8921 * pushed off this CPU now via balance_push() or placed on a different CPU
8922 * during wakeup. Wait until the CPU is quiescent.
8924 static void balance_hotplug_wait(void)
8926 struct rq *rq = this_rq();
8928 rcuwait_wait_event(&rq->hotplug_wait,
8929 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8930 TASK_UNINTERRUPTIBLE);
8935 static inline void balance_push(struct rq *rq)
8939 static inline void balance_push_set(int cpu, bool on)
8943 static inline void balance_hotplug_wait(void)
8947 #endif /* CONFIG_HOTPLUG_CPU */
8949 void set_rq_online(struct rq *rq)
8952 const struct sched_class *class;
8954 cpumask_set_cpu(rq->cpu, rq->rd->online);
8957 for_each_class(class) {
8958 if (class->rq_online)
8959 class->rq_online(rq);
8964 void set_rq_offline(struct rq *rq)
8967 const struct sched_class *class;
8969 for_each_class(class) {
8970 if (class->rq_offline)
8971 class->rq_offline(rq);
8974 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8980 * used to mark begin/end of suspend/resume:
8982 static int num_cpus_frozen;
8985 * Update cpusets according to cpu_active mask. If cpusets are
8986 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8987 * around partition_sched_domains().
8989 * If we come here as part of a suspend/resume, don't touch cpusets because we
8990 * want to restore it back to its original state upon resume anyway.
8992 static void cpuset_cpu_active(void)
8994 if (cpuhp_tasks_frozen) {
8996 * num_cpus_frozen tracks how many CPUs are involved in suspend
8997 * resume sequence. As long as this is not the last online
8998 * operation in the resume sequence, just build a single sched
8999 * domain, ignoring cpusets.
9001 partition_sched_domains(1, NULL, NULL);
9002 if (--num_cpus_frozen)
9005 * This is the last CPU online operation. So fall through and
9006 * restore the original sched domains by considering the
9007 * cpuset configurations.
9009 cpuset_force_rebuild();
9011 cpuset_update_active_cpus();
9014 static int cpuset_cpu_inactive(unsigned int cpu)
9016 if (!cpuhp_tasks_frozen) {
9017 if (dl_cpu_busy(cpu))
9019 cpuset_update_active_cpus();
9022 partition_sched_domains(1, NULL, NULL);
9027 int sched_cpu_activate(unsigned int cpu)
9029 struct rq *rq = cpu_rq(cpu);
9033 * Clear the balance_push callback and prepare to schedule
9036 balance_push_set(cpu, false);
9038 #ifdef CONFIG_SCHED_SMT
9040 * When going up, increment the number of cores with SMT present.
9042 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9043 static_branch_inc_cpuslocked(&sched_smt_present);
9045 set_cpu_active(cpu, true);
9047 if (sched_smp_initialized) {
9048 sched_domains_numa_masks_set(cpu);
9049 cpuset_cpu_active();
9053 * Put the rq online, if not already. This happens:
9055 * 1) In the early boot process, because we build the real domains
9056 * after all CPUs have been brought up.
9058 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9061 rq_lock_irqsave(rq, &rf);
9063 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9066 rq_unlock_irqrestore(rq, &rf);
9071 int sched_cpu_deactivate(unsigned int cpu)
9073 struct rq *rq = cpu_rq(cpu);
9078 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9079 * load balancing when not active
9081 nohz_balance_exit_idle(rq);
9083 set_cpu_active(cpu, false);
9086 * From this point forward, this CPU will refuse to run any task that
9087 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9088 * push those tasks away until this gets cleared, see
9089 * sched_cpu_dying().
9091 balance_push_set(cpu, true);
9094 * We've cleared cpu_active_mask / set balance_push, wait for all
9095 * preempt-disabled and RCU users of this state to go away such that
9096 * all new such users will observe it.
9098 * Specifically, we rely on ttwu to no longer target this CPU, see
9099 * ttwu_queue_cond() and is_cpu_allowed().
9101 * Do sync before park smpboot threads to take care the rcu boost case.
9105 rq_lock_irqsave(rq, &rf);
9107 update_rq_clock(rq);
9108 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9111 rq_unlock_irqrestore(rq, &rf);
9113 #ifdef CONFIG_SCHED_SMT
9115 * When going down, decrement the number of cores with SMT present.
9117 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9118 static_branch_dec_cpuslocked(&sched_smt_present);
9120 sched_core_cpu_deactivate(cpu);
9123 if (!sched_smp_initialized)
9126 ret = cpuset_cpu_inactive(cpu);
9128 balance_push_set(cpu, false);
9129 set_cpu_active(cpu, true);
9132 sched_domains_numa_masks_clear(cpu);
9136 static void sched_rq_cpu_starting(unsigned int cpu)
9138 struct rq *rq = cpu_rq(cpu);
9140 rq->calc_load_update = calc_load_update;
9141 update_max_interval();
9144 int sched_cpu_starting(unsigned int cpu)
9146 sched_core_cpu_starting(cpu);
9147 sched_rq_cpu_starting(cpu);
9148 sched_tick_start(cpu);
9152 #ifdef CONFIG_HOTPLUG_CPU
9155 * Invoked immediately before the stopper thread is invoked to bring the
9156 * CPU down completely. At this point all per CPU kthreads except the
9157 * hotplug thread (current) and the stopper thread (inactive) have been
9158 * either parked or have been unbound from the outgoing CPU. Ensure that
9159 * any of those which might be on the way out are gone.
9161 * If after this point a bound task is being woken on this CPU then the
9162 * responsible hotplug callback has failed to do it's job.
9163 * sched_cpu_dying() will catch it with the appropriate fireworks.
9165 int sched_cpu_wait_empty(unsigned int cpu)
9167 balance_hotplug_wait();
9172 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9173 * might have. Called from the CPU stopper task after ensuring that the
9174 * stopper is the last running task on the CPU, so nr_active count is
9175 * stable. We need to take the teardown thread which is calling this into
9176 * account, so we hand in adjust = 1 to the load calculation.
9178 * Also see the comment "Global load-average calculations".
9180 static void calc_load_migrate(struct rq *rq)
9182 long delta = calc_load_fold_active(rq, 1);
9185 atomic_long_add(delta, &calc_load_tasks);
9188 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9190 struct task_struct *g, *p;
9191 int cpu = cpu_of(rq);
9193 lockdep_assert_rq_held(rq);
9195 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9196 for_each_process_thread(g, p) {
9197 if (task_cpu(p) != cpu)
9200 if (!task_on_rq_queued(p))
9203 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9207 int sched_cpu_dying(unsigned int cpu)
9209 struct rq *rq = cpu_rq(cpu);
9212 /* Handle pending wakeups and then migrate everything off */
9213 sched_tick_stop(cpu);
9215 rq_lock_irqsave(rq, &rf);
9216 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9217 WARN(true, "Dying CPU not properly vacated!");
9218 dump_rq_tasks(rq, KERN_WARNING);
9220 rq_unlock_irqrestore(rq, &rf);
9222 calc_load_migrate(rq);
9223 update_max_interval();
9225 sched_core_cpu_dying(cpu);
9230 void __init sched_init_smp(void)
9235 * There's no userspace yet to cause hotplug operations; hence all the
9236 * CPU masks are stable and all blatant races in the below code cannot
9239 mutex_lock(&sched_domains_mutex);
9240 sched_init_domains(cpu_active_mask);
9241 mutex_unlock(&sched_domains_mutex);
9243 /* Move init over to a non-isolated CPU */
9244 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9246 current->flags &= ~PF_NO_SETAFFINITY;
9247 sched_init_granularity();
9249 init_sched_rt_class();
9250 init_sched_dl_class();
9252 sched_smp_initialized = true;
9255 static int __init migration_init(void)
9257 sched_cpu_starting(smp_processor_id());
9260 early_initcall(migration_init);
9263 void __init sched_init_smp(void)
9265 sched_init_granularity();
9267 #endif /* CONFIG_SMP */
9269 int in_sched_functions(unsigned long addr)
9271 return in_lock_functions(addr) ||
9272 (addr >= (unsigned long)__sched_text_start
9273 && addr < (unsigned long)__sched_text_end);
9276 #ifdef CONFIG_CGROUP_SCHED
9278 * Default task group.
9279 * Every task in system belongs to this group at bootup.
9281 struct task_group root_task_group;
9282 LIST_HEAD(task_groups);
9284 /* Cacheline aligned slab cache for task_group */
9285 static struct kmem_cache *task_group_cache __read_mostly;
9288 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9289 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9291 void __init sched_init(void)
9293 unsigned long ptr = 0;
9296 /* Make sure the linker didn't screw up */
9297 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9298 &fair_sched_class + 1 != &rt_sched_class ||
9299 &rt_sched_class + 1 != &dl_sched_class);
9301 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9306 #ifdef CONFIG_FAIR_GROUP_SCHED
9307 ptr += 2 * nr_cpu_ids * sizeof(void **);
9309 #ifdef CONFIG_RT_GROUP_SCHED
9310 ptr += 2 * nr_cpu_ids * sizeof(void **);
9313 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9315 #ifdef CONFIG_FAIR_GROUP_SCHED
9316 root_task_group.se = (struct sched_entity **)ptr;
9317 ptr += nr_cpu_ids * sizeof(void **);
9319 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9320 ptr += nr_cpu_ids * sizeof(void **);
9322 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9323 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9324 #endif /* CONFIG_FAIR_GROUP_SCHED */
9325 #ifdef CONFIG_RT_GROUP_SCHED
9326 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9327 ptr += nr_cpu_ids * sizeof(void **);
9329 root_task_group.rt_rq = (struct rt_rq **)ptr;
9330 ptr += nr_cpu_ids * sizeof(void **);
9332 #endif /* CONFIG_RT_GROUP_SCHED */
9334 #ifdef CONFIG_CPUMASK_OFFSTACK
9335 for_each_possible_cpu(i) {
9336 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9337 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9338 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9339 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9341 #endif /* CONFIG_CPUMASK_OFFSTACK */
9343 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9344 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9347 init_defrootdomain();
9350 #ifdef CONFIG_RT_GROUP_SCHED
9351 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9352 global_rt_period(), global_rt_runtime());
9353 #endif /* CONFIG_RT_GROUP_SCHED */
9355 #ifdef CONFIG_CGROUP_SCHED
9356 task_group_cache = KMEM_CACHE(task_group, 0);
9358 list_add(&root_task_group.list, &task_groups);
9359 INIT_LIST_HEAD(&root_task_group.children);
9360 INIT_LIST_HEAD(&root_task_group.siblings);
9361 autogroup_init(&init_task);
9362 #endif /* CONFIG_CGROUP_SCHED */
9364 for_each_possible_cpu(i) {
9368 raw_spin_lock_init(&rq->__lock);
9370 rq->calc_load_active = 0;
9371 rq->calc_load_update = jiffies + LOAD_FREQ;
9372 init_cfs_rq(&rq->cfs);
9373 init_rt_rq(&rq->rt);
9374 init_dl_rq(&rq->dl);
9375 #ifdef CONFIG_FAIR_GROUP_SCHED
9376 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9377 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9379 * How much CPU bandwidth does root_task_group get?
9381 * In case of task-groups formed thr' the cgroup filesystem, it
9382 * gets 100% of the CPU resources in the system. This overall
9383 * system CPU resource is divided among the tasks of
9384 * root_task_group and its child task-groups in a fair manner,
9385 * based on each entity's (task or task-group's) weight
9386 * (se->load.weight).
9388 * In other words, if root_task_group has 10 tasks of weight
9389 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9390 * then A0's share of the CPU resource is:
9392 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9394 * We achieve this by letting root_task_group's tasks sit
9395 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9397 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9398 #endif /* CONFIG_FAIR_GROUP_SCHED */
9400 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9401 #ifdef CONFIG_RT_GROUP_SCHED
9402 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9407 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9408 rq->balance_callback = &balance_push_callback;
9409 rq->active_balance = 0;
9410 rq->next_balance = jiffies;
9415 rq->avg_idle = 2*sysctl_sched_migration_cost;
9416 rq->wake_stamp = jiffies;
9417 rq->wake_avg_idle = rq->avg_idle;
9418 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9420 INIT_LIST_HEAD(&rq->cfs_tasks);
9422 rq_attach_root(rq, &def_root_domain);
9423 #ifdef CONFIG_NO_HZ_COMMON
9424 rq->last_blocked_load_update_tick = jiffies;
9425 atomic_set(&rq->nohz_flags, 0);
9427 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9429 #ifdef CONFIG_HOTPLUG_CPU
9430 rcuwait_init(&rq->hotplug_wait);
9432 #endif /* CONFIG_SMP */
9434 atomic_set(&rq->nr_iowait, 0);
9436 #ifdef CONFIG_SCHED_CORE
9438 rq->core_pick = NULL;
9439 rq->core_enabled = 0;
9440 rq->core_tree = RB_ROOT;
9441 rq->core_forceidle_count = 0;
9442 rq->core_forceidle_occupation = 0;
9443 rq->core_forceidle_start = 0;
9445 rq->core_cookie = 0UL;
9449 set_load_weight(&init_task);
9452 * The boot idle thread does lazy MMU switching as well:
9455 enter_lazy_tlb(&init_mm, current);
9458 * The idle task doesn't need the kthread struct to function, but it
9459 * is dressed up as a per-CPU kthread and thus needs to play the part
9460 * if we want to avoid special-casing it in code that deals with per-CPU
9463 WARN_ON(!set_kthread_struct(current));
9466 * Make us the idle thread. Technically, schedule() should not be
9467 * called from this thread, however somewhere below it might be,
9468 * but because we are the idle thread, we just pick up running again
9469 * when this runqueue becomes "idle".
9471 init_idle(current, smp_processor_id());
9473 calc_load_update = jiffies + LOAD_FREQ;
9476 idle_thread_set_boot_cpu();
9477 balance_push_set(smp_processor_id(), false);
9479 init_sched_fair_class();
9485 preempt_dynamic_init();
9487 scheduler_running = 1;
9490 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9492 void __might_sleep(const char *file, int line)
9494 unsigned int state = get_current_state();
9496 * Blocking primitives will set (and therefore destroy) current->state,
9497 * since we will exit with TASK_RUNNING make sure we enter with it,
9498 * otherwise we will destroy state.
9500 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9501 "do not call blocking ops when !TASK_RUNNING; "
9502 "state=%x set at [<%p>] %pS\n", state,
9503 (void *)current->task_state_change,
9504 (void *)current->task_state_change);
9506 __might_resched(file, line, 0);
9508 EXPORT_SYMBOL(__might_sleep);
9510 static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9512 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9515 if (preempt_count() == preempt_offset)
9518 pr_err("Preemption disabled at:");
9519 print_ip_sym(KERN_ERR, ip);
9522 static inline bool resched_offsets_ok(unsigned int offsets)
9524 unsigned int nested = preempt_count();
9526 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9528 return nested == offsets;
9531 void __might_resched(const char *file, int line, unsigned int offsets)
9533 /* Ratelimiting timestamp: */
9534 static unsigned long prev_jiffy;
9536 unsigned long preempt_disable_ip;
9538 /* WARN_ON_ONCE() by default, no rate limit required: */
9541 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
9542 !is_idle_task(current) && !current->non_block_count) ||
9543 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9547 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9549 prev_jiffy = jiffies;
9551 /* Save this before calling printk(), since that will clobber it: */
9552 preempt_disable_ip = get_preempt_disable_ip(current);
9554 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9556 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9557 in_atomic(), irqs_disabled(), current->non_block_count,
9558 current->pid, current->comm);
9559 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
9560 offsets & MIGHT_RESCHED_PREEMPT_MASK);
9562 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
9563 pr_err("RCU nest depth: %d, expected: %u\n",
9564 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
9567 if (task_stack_end_corrupted(current))
9568 pr_emerg("Thread overran stack, or stack corrupted\n");
9570 debug_show_held_locks(current);
9571 if (irqs_disabled())
9572 print_irqtrace_events(current);
9574 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
9575 preempt_disable_ip);
9578 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9580 EXPORT_SYMBOL(__might_resched);
9582 void __cant_sleep(const char *file, int line, int preempt_offset)
9584 static unsigned long prev_jiffy;
9586 if (irqs_disabled())
9589 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9592 if (preempt_count() > preempt_offset)
9595 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9597 prev_jiffy = jiffies;
9599 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9600 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9601 in_atomic(), irqs_disabled(),
9602 current->pid, current->comm);
9604 debug_show_held_locks(current);
9606 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9608 EXPORT_SYMBOL_GPL(__cant_sleep);
9611 void __cant_migrate(const char *file, int line)
9613 static unsigned long prev_jiffy;
9615 if (irqs_disabled())
9618 if (is_migration_disabled(current))
9621 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9624 if (preempt_count() > 0)
9627 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9629 prev_jiffy = jiffies;
9631 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9632 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9633 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9634 current->pid, current->comm);
9636 debug_show_held_locks(current);
9638 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9640 EXPORT_SYMBOL_GPL(__cant_migrate);
9644 #ifdef CONFIG_MAGIC_SYSRQ
9645 void normalize_rt_tasks(void)
9647 struct task_struct *g, *p;
9648 struct sched_attr attr = {
9649 .sched_policy = SCHED_NORMAL,
9652 read_lock(&tasklist_lock);
9653 for_each_process_thread(g, p) {
9655 * Only normalize user tasks:
9657 if (p->flags & PF_KTHREAD)
9660 p->se.exec_start = 0;
9661 schedstat_set(p->stats.wait_start, 0);
9662 schedstat_set(p->stats.sleep_start, 0);
9663 schedstat_set(p->stats.block_start, 0);
9665 if (!dl_task(p) && !rt_task(p)) {
9667 * Renice negative nice level userspace
9670 if (task_nice(p) < 0)
9671 set_user_nice(p, 0);
9675 __sched_setscheduler(p, &attr, false, false);
9677 read_unlock(&tasklist_lock);
9680 #endif /* CONFIG_MAGIC_SYSRQ */
9682 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9684 * These functions are only useful for the IA64 MCA handling, or kdb.
9686 * They can only be called when the whole system has been
9687 * stopped - every CPU needs to be quiescent, and no scheduling
9688 * activity can take place. Using them for anything else would
9689 * be a serious bug, and as a result, they aren't even visible
9690 * under any other configuration.
9694 * curr_task - return the current task for a given CPU.
9695 * @cpu: the processor in question.
9697 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9699 * Return: The current task for @cpu.
9701 struct task_struct *curr_task(int cpu)
9703 return cpu_curr(cpu);
9706 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9710 * ia64_set_curr_task - set the current task for a given CPU.
9711 * @cpu: the processor in question.
9712 * @p: the task pointer to set.
9714 * Description: This function must only be used when non-maskable interrupts
9715 * are serviced on a separate stack. It allows the architecture to switch the
9716 * notion of the current task on a CPU in a non-blocking manner. This function
9717 * must be called with all CPU's synchronized, and interrupts disabled, the
9718 * and caller must save the original value of the current task (see
9719 * curr_task() above) and restore that value before reenabling interrupts and
9720 * re-starting the system.
9722 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9724 void ia64_set_curr_task(int cpu, struct task_struct *p)
9731 #ifdef CONFIG_CGROUP_SCHED
9732 /* task_group_lock serializes the addition/removal of task groups */
9733 static DEFINE_SPINLOCK(task_group_lock);
9735 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9736 struct task_group *parent)
9738 #ifdef CONFIG_UCLAMP_TASK_GROUP
9739 enum uclamp_id clamp_id;
9741 for_each_clamp_id(clamp_id) {
9742 uclamp_se_set(&tg->uclamp_req[clamp_id],
9743 uclamp_none(clamp_id), false);
9744 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9749 static void sched_free_group(struct task_group *tg)
9751 free_fair_sched_group(tg);
9752 free_rt_sched_group(tg);
9754 kmem_cache_free(task_group_cache, tg);
9757 static void sched_free_group_rcu(struct rcu_head *rcu)
9759 sched_free_group(container_of(rcu, struct task_group, rcu));
9762 static void sched_unregister_group(struct task_group *tg)
9764 unregister_fair_sched_group(tg);
9765 unregister_rt_sched_group(tg);
9767 * We have to wait for yet another RCU grace period to expire, as
9768 * print_cfs_stats() might run concurrently.
9770 call_rcu(&tg->rcu, sched_free_group_rcu);
9773 /* allocate runqueue etc for a new task group */
9774 struct task_group *sched_create_group(struct task_group *parent)
9776 struct task_group *tg;
9778 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9780 return ERR_PTR(-ENOMEM);
9782 if (!alloc_fair_sched_group(tg, parent))
9785 if (!alloc_rt_sched_group(tg, parent))
9788 alloc_uclamp_sched_group(tg, parent);
9793 sched_free_group(tg);
9794 return ERR_PTR(-ENOMEM);
9797 void sched_online_group(struct task_group *tg, struct task_group *parent)
9799 unsigned long flags;
9801 spin_lock_irqsave(&task_group_lock, flags);
9802 list_add_rcu(&tg->list, &task_groups);
9804 /* Root should already exist: */
9807 tg->parent = parent;
9808 INIT_LIST_HEAD(&tg->children);
9809 list_add_rcu(&tg->siblings, &parent->children);
9810 spin_unlock_irqrestore(&task_group_lock, flags);
9812 online_fair_sched_group(tg);
9815 /* rcu callback to free various structures associated with a task group */
9816 static void sched_unregister_group_rcu(struct rcu_head *rhp)
9818 /* Now it should be safe to free those cfs_rqs: */
9819 sched_unregister_group(container_of(rhp, struct task_group, rcu));
9822 void sched_destroy_group(struct task_group *tg)
9824 /* Wait for possible concurrent references to cfs_rqs complete: */
9825 call_rcu(&tg->rcu, sched_unregister_group_rcu);
9828 void sched_release_group(struct task_group *tg)
9830 unsigned long flags;
9833 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
9834 * sched_cfs_period_timer()).
9836 * For this to be effective, we have to wait for all pending users of
9837 * this task group to leave their RCU critical section to ensure no new
9838 * user will see our dying task group any more. Specifically ensure
9839 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
9841 * We therefore defer calling unregister_fair_sched_group() to
9842 * sched_unregister_group() which is guarantied to get called only after the
9843 * current RCU grace period has expired.
9845 spin_lock_irqsave(&task_group_lock, flags);
9846 list_del_rcu(&tg->list);
9847 list_del_rcu(&tg->siblings);
9848 spin_unlock_irqrestore(&task_group_lock, flags);
9851 static void sched_change_group(struct task_struct *tsk, int type)
9853 struct task_group *tg;
9856 * All callers are synchronized by task_rq_lock(); we do not use RCU
9857 * which is pointless here. Thus, we pass "true" to task_css_check()
9858 * to prevent lockdep warnings.
9860 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9861 struct task_group, css);
9862 tg = autogroup_task_group(tsk, tg);
9863 tsk->sched_task_group = tg;
9865 #ifdef CONFIG_FAIR_GROUP_SCHED
9866 if (tsk->sched_class->task_change_group)
9867 tsk->sched_class->task_change_group(tsk, type);
9870 set_task_rq(tsk, task_cpu(tsk));
9874 * Change task's runqueue when it moves between groups.
9876 * The caller of this function should have put the task in its new group by
9877 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9880 void sched_move_task(struct task_struct *tsk)
9882 int queued, running, queue_flags =
9883 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9887 rq = task_rq_lock(tsk, &rf);
9888 update_rq_clock(rq);
9890 running = task_current(rq, tsk);
9891 queued = task_on_rq_queued(tsk);
9894 dequeue_task(rq, tsk, queue_flags);
9896 put_prev_task(rq, tsk);
9898 sched_change_group(tsk, TASK_MOVE_GROUP);
9901 enqueue_task(rq, tsk, queue_flags);
9903 set_next_task(rq, tsk);
9905 * After changing group, the running task may have joined a
9906 * throttled one but it's still the running task. Trigger a
9907 * resched to make sure that task can still run.
9912 task_rq_unlock(rq, tsk, &rf);
9915 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9917 return css ? container_of(css, struct task_group, css) : NULL;
9920 static struct cgroup_subsys_state *
9921 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9923 struct task_group *parent = css_tg(parent_css);
9924 struct task_group *tg;
9927 /* This is early initialization for the top cgroup */
9928 return &root_task_group.css;
9931 tg = sched_create_group(parent);
9933 return ERR_PTR(-ENOMEM);
9938 /* Expose task group only after completing cgroup initialization */
9939 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9941 struct task_group *tg = css_tg(css);
9942 struct task_group *parent = css_tg(css->parent);
9945 sched_online_group(tg, parent);
9947 #ifdef CONFIG_UCLAMP_TASK_GROUP
9948 /* Propagate the effective uclamp value for the new group */
9949 mutex_lock(&uclamp_mutex);
9951 cpu_util_update_eff(css);
9953 mutex_unlock(&uclamp_mutex);
9959 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9961 struct task_group *tg = css_tg(css);
9963 sched_release_group(tg);
9966 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9968 struct task_group *tg = css_tg(css);
9971 * Relies on the RCU grace period between css_released() and this.
9973 sched_unregister_group(tg);
9977 * This is called before wake_up_new_task(), therefore we really only
9978 * have to set its group bits, all the other stuff does not apply.
9980 static void cpu_cgroup_fork(struct task_struct *task)
9985 rq = task_rq_lock(task, &rf);
9987 update_rq_clock(rq);
9988 sched_change_group(task, TASK_SET_GROUP);
9990 task_rq_unlock(rq, task, &rf);
9993 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9995 struct task_struct *task;
9996 struct cgroup_subsys_state *css;
9999 cgroup_taskset_for_each(task, css, tset) {
10000 #ifdef CONFIG_RT_GROUP_SCHED
10001 if (!sched_rt_can_attach(css_tg(css), task))
10005 * Serialize against wake_up_new_task() such that if it's
10006 * running, we're sure to observe its full state.
10008 raw_spin_lock_irq(&task->pi_lock);
10010 * Avoid calling sched_move_task() before wake_up_new_task()
10011 * has happened. This would lead to problems with PELT, due to
10012 * move wanting to detach+attach while we're not attached yet.
10014 if (READ_ONCE(task->__state) == TASK_NEW)
10016 raw_spin_unlock_irq(&task->pi_lock);
10024 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
10026 struct task_struct *task;
10027 struct cgroup_subsys_state *css;
10029 cgroup_taskset_for_each(task, css, tset)
10030 sched_move_task(task);
10033 #ifdef CONFIG_UCLAMP_TASK_GROUP
10034 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10036 struct cgroup_subsys_state *top_css = css;
10037 struct uclamp_se *uc_parent = NULL;
10038 struct uclamp_se *uc_se = NULL;
10039 unsigned int eff[UCLAMP_CNT];
10040 enum uclamp_id clamp_id;
10041 unsigned int clamps;
10043 lockdep_assert_held(&uclamp_mutex);
10044 SCHED_WARN_ON(!rcu_read_lock_held());
10046 css_for_each_descendant_pre(css, top_css) {
10047 uc_parent = css_tg(css)->parent
10048 ? css_tg(css)->parent->uclamp : NULL;
10050 for_each_clamp_id(clamp_id) {
10051 /* Assume effective clamps matches requested clamps */
10052 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10053 /* Cap effective clamps with parent's effective clamps */
10055 eff[clamp_id] > uc_parent[clamp_id].value) {
10056 eff[clamp_id] = uc_parent[clamp_id].value;
10059 /* Ensure protection is always capped by limit */
10060 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10062 /* Propagate most restrictive effective clamps */
10064 uc_se = css_tg(css)->uclamp;
10065 for_each_clamp_id(clamp_id) {
10066 if (eff[clamp_id] == uc_se[clamp_id].value)
10068 uc_se[clamp_id].value = eff[clamp_id];
10069 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10070 clamps |= (0x1 << clamp_id);
10073 css = css_rightmost_descendant(css);
10077 /* Immediately update descendants RUNNABLE tasks */
10078 uclamp_update_active_tasks(css);
10083 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10084 * C expression. Since there is no way to convert a macro argument (N) into a
10085 * character constant, use two levels of macros.
10087 #define _POW10(exp) ((unsigned int)1e##exp)
10088 #define POW10(exp) _POW10(exp)
10090 struct uclamp_request {
10091 #define UCLAMP_PERCENT_SHIFT 2
10092 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10098 static inline struct uclamp_request
10099 capacity_from_percent(char *buf)
10101 struct uclamp_request req = {
10102 .percent = UCLAMP_PERCENT_SCALE,
10103 .util = SCHED_CAPACITY_SCALE,
10108 if (strcmp(buf, "max")) {
10109 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10113 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
10118 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10119 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10125 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10126 size_t nbytes, loff_t off,
10127 enum uclamp_id clamp_id)
10129 struct uclamp_request req;
10130 struct task_group *tg;
10132 req = capacity_from_percent(buf);
10136 static_branch_enable(&sched_uclamp_used);
10138 mutex_lock(&uclamp_mutex);
10141 tg = css_tg(of_css(of));
10142 if (tg->uclamp_req[clamp_id].value != req.util)
10143 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10146 * Because of not recoverable conversion rounding we keep track of the
10147 * exact requested value
10149 tg->uclamp_pct[clamp_id] = req.percent;
10151 /* Update effective clamps to track the most restrictive value */
10152 cpu_util_update_eff(of_css(of));
10155 mutex_unlock(&uclamp_mutex);
10160 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10161 char *buf, size_t nbytes,
10164 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10167 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10168 char *buf, size_t nbytes,
10171 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10174 static inline void cpu_uclamp_print(struct seq_file *sf,
10175 enum uclamp_id clamp_id)
10177 struct task_group *tg;
10183 tg = css_tg(seq_css(sf));
10184 util_clamp = tg->uclamp_req[clamp_id].value;
10187 if (util_clamp == SCHED_CAPACITY_SCALE) {
10188 seq_puts(sf, "max\n");
10192 percent = tg->uclamp_pct[clamp_id];
10193 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10194 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10197 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10199 cpu_uclamp_print(sf, UCLAMP_MIN);
10203 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10205 cpu_uclamp_print(sf, UCLAMP_MAX);
10208 #endif /* CONFIG_UCLAMP_TASK_GROUP */
10210 #ifdef CONFIG_FAIR_GROUP_SCHED
10211 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10212 struct cftype *cftype, u64 shareval)
10214 if (shareval > scale_load_down(ULONG_MAX))
10215 shareval = MAX_SHARES;
10216 return sched_group_set_shares(css_tg(css), scale_load(shareval));
10219 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10220 struct cftype *cft)
10222 struct task_group *tg = css_tg(css);
10224 return (u64) scale_load_down(tg->shares);
10227 #ifdef CONFIG_CFS_BANDWIDTH
10228 static DEFINE_MUTEX(cfs_constraints_mutex);
10230 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
10231 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
10232 /* More than 203 days if BW_SHIFT equals 20. */
10233 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
10235 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10237 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10240 int i, ret = 0, runtime_enabled, runtime_was_enabled;
10241 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10243 if (tg == &root_task_group)
10247 * Ensure we have at some amount of bandwidth every period. This is
10248 * to prevent reaching a state of large arrears when throttled via
10249 * entity_tick() resulting in prolonged exit starvation.
10251 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10255 * Likewise, bound things on the other side by preventing insane quota
10256 * periods. This also allows us to normalize in computing quota
10259 if (period > max_cfs_quota_period)
10263 * Bound quota to defend quota against overflow during bandwidth shift.
10265 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10268 if (quota != RUNTIME_INF && (burst > quota ||
10269 burst + quota > max_cfs_runtime))
10273 * Prevent race between setting of cfs_rq->runtime_enabled and
10274 * unthrottle_offline_cfs_rqs().
10277 mutex_lock(&cfs_constraints_mutex);
10278 ret = __cfs_schedulable(tg, period, quota);
10282 runtime_enabled = quota != RUNTIME_INF;
10283 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10285 * If we need to toggle cfs_bandwidth_used, off->on must occur
10286 * before making related changes, and on->off must occur afterwards
10288 if (runtime_enabled && !runtime_was_enabled)
10289 cfs_bandwidth_usage_inc();
10290 raw_spin_lock_irq(&cfs_b->lock);
10291 cfs_b->period = ns_to_ktime(period);
10292 cfs_b->quota = quota;
10293 cfs_b->burst = burst;
10295 __refill_cfs_bandwidth_runtime(cfs_b);
10297 /* Restart the period timer (if active) to handle new period expiry: */
10298 if (runtime_enabled)
10299 start_cfs_bandwidth(cfs_b);
10301 raw_spin_unlock_irq(&cfs_b->lock);
10303 for_each_online_cpu(i) {
10304 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10305 struct rq *rq = cfs_rq->rq;
10306 struct rq_flags rf;
10308 rq_lock_irq(rq, &rf);
10309 cfs_rq->runtime_enabled = runtime_enabled;
10310 cfs_rq->runtime_remaining = 0;
10312 if (cfs_rq->throttled)
10313 unthrottle_cfs_rq(cfs_rq);
10314 rq_unlock_irq(rq, &rf);
10316 if (runtime_was_enabled && !runtime_enabled)
10317 cfs_bandwidth_usage_dec();
10319 mutex_unlock(&cfs_constraints_mutex);
10320 cpus_read_unlock();
10325 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10327 u64 quota, period, burst;
10329 period = ktime_to_ns(tg->cfs_bandwidth.period);
10330 burst = tg->cfs_bandwidth.burst;
10331 if (cfs_quota_us < 0)
10332 quota = RUNTIME_INF;
10333 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10334 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10338 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10341 static long tg_get_cfs_quota(struct task_group *tg)
10345 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10348 quota_us = tg->cfs_bandwidth.quota;
10349 do_div(quota_us, NSEC_PER_USEC);
10354 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10356 u64 quota, period, burst;
10358 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10361 period = (u64)cfs_period_us * NSEC_PER_USEC;
10362 quota = tg->cfs_bandwidth.quota;
10363 burst = tg->cfs_bandwidth.burst;
10365 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10368 static long tg_get_cfs_period(struct task_group *tg)
10372 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10373 do_div(cfs_period_us, NSEC_PER_USEC);
10375 return cfs_period_us;
10378 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10380 u64 quota, period, burst;
10382 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10385 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10386 period = ktime_to_ns(tg->cfs_bandwidth.period);
10387 quota = tg->cfs_bandwidth.quota;
10389 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10392 static long tg_get_cfs_burst(struct task_group *tg)
10396 burst_us = tg->cfs_bandwidth.burst;
10397 do_div(burst_us, NSEC_PER_USEC);
10402 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10403 struct cftype *cft)
10405 return tg_get_cfs_quota(css_tg(css));
10408 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10409 struct cftype *cftype, s64 cfs_quota_us)
10411 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10414 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10415 struct cftype *cft)
10417 return tg_get_cfs_period(css_tg(css));
10420 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10421 struct cftype *cftype, u64 cfs_period_us)
10423 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10426 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10427 struct cftype *cft)
10429 return tg_get_cfs_burst(css_tg(css));
10432 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10433 struct cftype *cftype, u64 cfs_burst_us)
10435 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10438 struct cfs_schedulable_data {
10439 struct task_group *tg;
10444 * normalize group quota/period to be quota/max_period
10445 * note: units are usecs
10447 static u64 normalize_cfs_quota(struct task_group *tg,
10448 struct cfs_schedulable_data *d)
10453 period = d->period;
10456 period = tg_get_cfs_period(tg);
10457 quota = tg_get_cfs_quota(tg);
10460 /* note: these should typically be equivalent */
10461 if (quota == RUNTIME_INF || quota == -1)
10462 return RUNTIME_INF;
10464 return to_ratio(period, quota);
10467 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10469 struct cfs_schedulable_data *d = data;
10470 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10471 s64 quota = 0, parent_quota = -1;
10474 quota = RUNTIME_INF;
10476 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10478 quota = normalize_cfs_quota(tg, d);
10479 parent_quota = parent_b->hierarchical_quota;
10482 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10483 * always take the min. On cgroup1, only inherit when no
10486 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10487 quota = min(quota, parent_quota);
10489 if (quota == RUNTIME_INF)
10490 quota = parent_quota;
10491 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10495 cfs_b->hierarchical_quota = quota;
10500 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10503 struct cfs_schedulable_data data = {
10509 if (quota != RUNTIME_INF) {
10510 do_div(data.period, NSEC_PER_USEC);
10511 do_div(data.quota, NSEC_PER_USEC);
10515 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10521 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10523 struct task_group *tg = css_tg(seq_css(sf));
10524 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10526 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10527 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10528 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10530 if (schedstat_enabled() && tg != &root_task_group) {
10531 struct sched_statistics *stats;
10535 for_each_possible_cpu(i) {
10536 stats = __schedstats_from_se(tg->se[i]);
10537 ws += schedstat_val(stats->wait_sum);
10540 seq_printf(sf, "wait_sum %llu\n", ws);
10543 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10544 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10548 #endif /* CONFIG_CFS_BANDWIDTH */
10549 #endif /* CONFIG_FAIR_GROUP_SCHED */
10551 #ifdef CONFIG_RT_GROUP_SCHED
10552 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10553 struct cftype *cft, s64 val)
10555 return sched_group_set_rt_runtime(css_tg(css), val);
10558 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10559 struct cftype *cft)
10561 return sched_group_rt_runtime(css_tg(css));
10564 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10565 struct cftype *cftype, u64 rt_period_us)
10567 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10570 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10571 struct cftype *cft)
10573 return sched_group_rt_period(css_tg(css));
10575 #endif /* CONFIG_RT_GROUP_SCHED */
10577 #ifdef CONFIG_FAIR_GROUP_SCHED
10578 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10579 struct cftype *cft)
10581 return css_tg(css)->idle;
10584 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10585 struct cftype *cft, s64 idle)
10587 return sched_group_set_idle(css_tg(css), idle);
10591 static struct cftype cpu_legacy_files[] = {
10592 #ifdef CONFIG_FAIR_GROUP_SCHED
10595 .read_u64 = cpu_shares_read_u64,
10596 .write_u64 = cpu_shares_write_u64,
10600 .read_s64 = cpu_idle_read_s64,
10601 .write_s64 = cpu_idle_write_s64,
10604 #ifdef CONFIG_CFS_BANDWIDTH
10606 .name = "cfs_quota_us",
10607 .read_s64 = cpu_cfs_quota_read_s64,
10608 .write_s64 = cpu_cfs_quota_write_s64,
10611 .name = "cfs_period_us",
10612 .read_u64 = cpu_cfs_period_read_u64,
10613 .write_u64 = cpu_cfs_period_write_u64,
10616 .name = "cfs_burst_us",
10617 .read_u64 = cpu_cfs_burst_read_u64,
10618 .write_u64 = cpu_cfs_burst_write_u64,
10622 .seq_show = cpu_cfs_stat_show,
10625 #ifdef CONFIG_RT_GROUP_SCHED
10627 .name = "rt_runtime_us",
10628 .read_s64 = cpu_rt_runtime_read,
10629 .write_s64 = cpu_rt_runtime_write,
10632 .name = "rt_period_us",
10633 .read_u64 = cpu_rt_period_read_uint,
10634 .write_u64 = cpu_rt_period_write_uint,
10637 #ifdef CONFIG_UCLAMP_TASK_GROUP
10639 .name = "uclamp.min",
10640 .flags = CFTYPE_NOT_ON_ROOT,
10641 .seq_show = cpu_uclamp_min_show,
10642 .write = cpu_uclamp_min_write,
10645 .name = "uclamp.max",
10646 .flags = CFTYPE_NOT_ON_ROOT,
10647 .seq_show = cpu_uclamp_max_show,
10648 .write = cpu_uclamp_max_write,
10651 { } /* Terminate */
10654 static int cpu_extra_stat_show(struct seq_file *sf,
10655 struct cgroup_subsys_state *css)
10657 #ifdef CONFIG_CFS_BANDWIDTH
10659 struct task_group *tg = css_tg(css);
10660 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10661 u64 throttled_usec, burst_usec;
10663 throttled_usec = cfs_b->throttled_time;
10664 do_div(throttled_usec, NSEC_PER_USEC);
10665 burst_usec = cfs_b->burst_time;
10666 do_div(burst_usec, NSEC_PER_USEC);
10668 seq_printf(sf, "nr_periods %d\n"
10669 "nr_throttled %d\n"
10670 "throttled_usec %llu\n"
10672 "burst_usec %llu\n",
10673 cfs_b->nr_periods, cfs_b->nr_throttled,
10674 throttled_usec, cfs_b->nr_burst, burst_usec);
10680 #ifdef CONFIG_FAIR_GROUP_SCHED
10681 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10682 struct cftype *cft)
10684 struct task_group *tg = css_tg(css);
10685 u64 weight = scale_load_down(tg->shares);
10687 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10690 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10691 struct cftype *cft, u64 weight)
10694 * cgroup weight knobs should use the common MIN, DFL and MAX
10695 * values which are 1, 100 and 10000 respectively. While it loses
10696 * a bit of range on both ends, it maps pretty well onto the shares
10697 * value used by scheduler and the round-trip conversions preserve
10698 * the original value over the entire range.
10700 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10703 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10705 return sched_group_set_shares(css_tg(css), scale_load(weight));
10708 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10709 struct cftype *cft)
10711 unsigned long weight = scale_load_down(css_tg(css)->shares);
10712 int last_delta = INT_MAX;
10715 /* find the closest nice value to the current weight */
10716 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10717 delta = abs(sched_prio_to_weight[prio] - weight);
10718 if (delta >= last_delta)
10720 last_delta = delta;
10723 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10726 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10727 struct cftype *cft, s64 nice)
10729 unsigned long weight;
10732 if (nice < MIN_NICE || nice > MAX_NICE)
10735 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10736 idx = array_index_nospec(idx, 40);
10737 weight = sched_prio_to_weight[idx];
10739 return sched_group_set_shares(css_tg(css), scale_load(weight));
10743 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10744 long period, long quota)
10747 seq_puts(sf, "max");
10749 seq_printf(sf, "%ld", quota);
10751 seq_printf(sf, " %ld\n", period);
10754 /* caller should put the current value in *@periodp before calling */
10755 static int __maybe_unused cpu_period_quota_parse(char *buf,
10756 u64 *periodp, u64 *quotap)
10758 char tok[21]; /* U64_MAX */
10760 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10763 *periodp *= NSEC_PER_USEC;
10765 if (sscanf(tok, "%llu", quotap))
10766 *quotap *= NSEC_PER_USEC;
10767 else if (!strcmp(tok, "max"))
10768 *quotap = RUNTIME_INF;
10775 #ifdef CONFIG_CFS_BANDWIDTH
10776 static int cpu_max_show(struct seq_file *sf, void *v)
10778 struct task_group *tg = css_tg(seq_css(sf));
10780 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10784 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10785 char *buf, size_t nbytes, loff_t off)
10787 struct task_group *tg = css_tg(of_css(of));
10788 u64 period = tg_get_cfs_period(tg);
10789 u64 burst = tg_get_cfs_burst(tg);
10793 ret = cpu_period_quota_parse(buf, &period, "a);
10795 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10796 return ret ?: nbytes;
10800 static struct cftype cpu_files[] = {
10801 #ifdef CONFIG_FAIR_GROUP_SCHED
10804 .flags = CFTYPE_NOT_ON_ROOT,
10805 .read_u64 = cpu_weight_read_u64,
10806 .write_u64 = cpu_weight_write_u64,
10809 .name = "weight.nice",
10810 .flags = CFTYPE_NOT_ON_ROOT,
10811 .read_s64 = cpu_weight_nice_read_s64,
10812 .write_s64 = cpu_weight_nice_write_s64,
10816 .flags = CFTYPE_NOT_ON_ROOT,
10817 .read_s64 = cpu_idle_read_s64,
10818 .write_s64 = cpu_idle_write_s64,
10821 #ifdef CONFIG_CFS_BANDWIDTH
10824 .flags = CFTYPE_NOT_ON_ROOT,
10825 .seq_show = cpu_max_show,
10826 .write = cpu_max_write,
10829 .name = "max.burst",
10830 .flags = CFTYPE_NOT_ON_ROOT,
10831 .read_u64 = cpu_cfs_burst_read_u64,
10832 .write_u64 = cpu_cfs_burst_write_u64,
10835 #ifdef CONFIG_UCLAMP_TASK_GROUP
10837 .name = "uclamp.min",
10838 .flags = CFTYPE_NOT_ON_ROOT,
10839 .seq_show = cpu_uclamp_min_show,
10840 .write = cpu_uclamp_min_write,
10843 .name = "uclamp.max",
10844 .flags = CFTYPE_NOT_ON_ROOT,
10845 .seq_show = cpu_uclamp_max_show,
10846 .write = cpu_uclamp_max_write,
10849 { } /* terminate */
10852 struct cgroup_subsys cpu_cgrp_subsys = {
10853 .css_alloc = cpu_cgroup_css_alloc,
10854 .css_online = cpu_cgroup_css_online,
10855 .css_released = cpu_cgroup_css_released,
10856 .css_free = cpu_cgroup_css_free,
10857 .css_extra_stat_show = cpu_extra_stat_show,
10858 .fork = cpu_cgroup_fork,
10859 .can_attach = cpu_cgroup_can_attach,
10860 .attach = cpu_cgroup_attach,
10861 .legacy_cftypes = cpu_legacy_files,
10862 .dfl_cftypes = cpu_files,
10863 .early_init = true,
10867 #endif /* CONFIG_CGROUP_SCHED */
10869 void dump_cpu_task(int cpu)
10871 pr_info("Task dump for CPU %d:\n", cpu);
10872 sched_show_task(cpu_curr(cpu));
10876 * Nice levels are multiplicative, with a gentle 10% change for every
10877 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10878 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10879 * that remained on nice 0.
10881 * The "10% effect" is relative and cumulative: from _any_ nice level,
10882 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10883 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10884 * If a task goes up by ~10% and another task goes down by ~10% then
10885 * the relative distance between them is ~25%.)
10887 const int sched_prio_to_weight[40] = {
10888 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10889 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10890 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10891 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10892 /* 0 */ 1024, 820, 655, 526, 423,
10893 /* 5 */ 335, 272, 215, 172, 137,
10894 /* 10 */ 110, 87, 70, 56, 45,
10895 /* 15 */ 36, 29, 23, 18, 15,
10899 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10901 * In cases where the weight does not change often, we can use the
10902 * precalculated inverse to speed up arithmetics by turning divisions
10903 * into multiplications:
10905 const u32 sched_prio_to_wmult[40] = {
10906 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10907 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10908 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10909 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10910 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10911 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10912 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10913 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10916 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10918 trace_sched_update_nr_running_tp(rq, count);