1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
63 * Print a warning if need_resched is set for the given duration (if
64 * LATENCY_WARN is enabled).
66 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
69 __read_mostly int sysctl_resched_latency_warn_ms = 100;
70 __read_mostly int sysctl_resched_latency_warn_once = 1;
71 #endif /* CONFIG_SCHED_DEBUG */
74 * Number of tasks to iterate in a single balance run.
75 * Limited because this is done with IRQs disabled.
77 const_debug unsigned int sysctl_sched_nr_migrate = 32;
80 * period over which we measure -rt task CPU usage in us.
83 unsigned int sysctl_sched_rt_period = 1000000;
85 __read_mostly int scheduler_running;
87 #ifdef CONFIG_SCHED_CORE
89 DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
91 /* kernel prio, less is more */
92 static inline int __task_prio(struct task_struct *p)
94 if (p->sched_class == &stop_sched_class) /* trumps deadline */
97 if (rt_prio(p->prio)) /* includes deadline */
98 return p->prio; /* [-1, 99] */
100 if (p->sched_class == &idle_sched_class)
101 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
103 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
113 /* real prio, less is less */
114 static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
117 int pa = __task_prio(a), pb = __task_prio(b);
125 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
126 return !dl_time_before(a->dl.deadline, b->dl.deadline);
128 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
129 return cfs_prio_less(a, b, in_fi);
134 static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
136 if (a->core_cookie < b->core_cookie)
139 if (a->core_cookie > b->core_cookie)
142 /* flip prio, so high prio is leftmost */
143 if (prio_less(b, a, task_rq(a)->core->core_forceidle))
149 #define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
151 static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
153 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
156 static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
158 const struct task_struct *p = __node_2_sc(node);
159 unsigned long cookie = (unsigned long)key;
161 if (cookie < p->core_cookie)
164 if (cookie > p->core_cookie)
170 void sched_core_enqueue(struct rq *rq, struct task_struct *p)
172 rq->core->core_task_seq++;
177 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
180 void sched_core_dequeue(struct rq *rq, struct task_struct *p)
182 rq->core->core_task_seq++;
184 if (!sched_core_enqueued(p))
187 rb_erase(&p->core_node, &rq->core_tree);
188 RB_CLEAR_NODE(&p->core_node);
192 * Find left-most (aka, highest priority) task matching @cookie.
194 static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
196 struct rb_node *node;
198 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
200 * The idle task always matches any cookie!
203 return idle_sched_class.pick_task(rq);
205 return __node_2_sc(node);
208 static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
210 struct rb_node *node = &p->core_node;
212 node = rb_next(node);
216 p = container_of(node, struct task_struct, core_node);
217 if (p->core_cookie != cookie)
224 * Magic required such that:
226 * raw_spin_rq_lock(rq);
228 * raw_spin_rq_unlock(rq);
230 * ends up locking and unlocking the _same_ lock, and all CPUs
231 * always agree on what rq has what lock.
233 * XXX entirely possible to selectively enable cores, don't bother for now.
236 static DEFINE_MUTEX(sched_core_mutex);
237 static atomic_t sched_core_count;
238 static struct cpumask sched_core_mask;
240 static void __sched_core_flip(bool enabled)
247 * Toggle the online cores, one by one.
249 cpumask_copy(&sched_core_mask, cpu_online_mask);
250 for_each_cpu(cpu, &sched_core_mask) {
251 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
255 for_each_cpu(t, smt_mask) {
256 /* supports up to SMT8 */
257 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
260 for_each_cpu(t, smt_mask)
261 cpu_rq(t)->core_enabled = enabled;
263 for_each_cpu(t, smt_mask)
264 raw_spin_unlock(&cpu_rq(t)->__lock);
267 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
271 * Toggle the offline CPUs.
273 cpumask_copy(&sched_core_mask, cpu_possible_mask);
274 cpumask_andnot(&sched_core_mask, &sched_core_mask, cpu_online_mask);
276 for_each_cpu(cpu, &sched_core_mask)
277 cpu_rq(cpu)->core_enabled = enabled;
282 static void sched_core_assert_empty(void)
286 for_each_possible_cpu(cpu)
287 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
290 static void __sched_core_enable(void)
292 static_branch_enable(&__sched_core_enabled);
294 * Ensure all previous instances of raw_spin_rq_*lock() have finished
295 * and future ones will observe !sched_core_disabled().
298 __sched_core_flip(true);
299 sched_core_assert_empty();
302 static void __sched_core_disable(void)
304 sched_core_assert_empty();
305 __sched_core_flip(false);
306 static_branch_disable(&__sched_core_enabled);
309 void sched_core_get(void)
311 if (atomic_inc_not_zero(&sched_core_count))
314 mutex_lock(&sched_core_mutex);
315 if (!atomic_read(&sched_core_count))
316 __sched_core_enable();
318 smp_mb__before_atomic();
319 atomic_inc(&sched_core_count);
320 mutex_unlock(&sched_core_mutex);
323 static void __sched_core_put(struct work_struct *work)
325 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
326 __sched_core_disable();
327 mutex_unlock(&sched_core_mutex);
331 void sched_core_put(void)
333 static DECLARE_WORK(_work, __sched_core_put);
336 * "There can be only one"
338 * Either this is the last one, or we don't actually need to do any
339 * 'work'. If it is the last *again*, we rely on
340 * WORK_STRUCT_PENDING_BIT.
342 if (!atomic_add_unless(&sched_core_count, -1, 1))
343 schedule_work(&_work);
346 #else /* !CONFIG_SCHED_CORE */
348 static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
349 static inline void sched_core_dequeue(struct rq *rq, struct task_struct *p) { }
351 #endif /* CONFIG_SCHED_CORE */
354 * part of the period that we allow rt tasks to run in us.
357 int sysctl_sched_rt_runtime = 950000;
361 * Serialization rules:
367 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
370 * rq2->lock where: rq1 < rq2
374 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
375 * local CPU's rq->lock, it optionally removes the task from the runqueue and
376 * always looks at the local rq data structures to find the most eligible task
379 * Task enqueue is also under rq->lock, possibly taken from another CPU.
380 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
381 * the local CPU to avoid bouncing the runqueue state around [ see
382 * ttwu_queue_wakelist() ]
384 * Task wakeup, specifically wakeups that involve migration, are horribly
385 * complicated to avoid having to take two rq->locks.
389 * System-calls and anything external will use task_rq_lock() which acquires
390 * both p->pi_lock and rq->lock. As a consequence the state they change is
391 * stable while holding either lock:
393 * - sched_setaffinity()/
394 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
395 * - set_user_nice(): p->se.load, p->*prio
396 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
397 * p->se.load, p->rt_priority,
398 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
399 * - sched_setnuma(): p->numa_preferred_nid
400 * - sched_move_task()/
401 * cpu_cgroup_fork(): p->sched_task_group
402 * - uclamp_update_active() p->uclamp*
404 * p->state <- TASK_*:
406 * is changed locklessly using set_current_state(), __set_current_state() or
407 * set_special_state(), see their respective comments, or by
408 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
411 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
413 * is set by activate_task() and cleared by deactivate_task(), under
414 * rq->lock. Non-zero indicates the task is runnable, the special
415 * ON_RQ_MIGRATING state is used for migration without holding both
416 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
418 * p->on_cpu <- { 0, 1 }:
420 * is set by prepare_task() and cleared by finish_task() such that it will be
421 * set before p is scheduled-in and cleared after p is scheduled-out, both
422 * under rq->lock. Non-zero indicates the task is running on its CPU.
424 * [ The astute reader will observe that it is possible for two tasks on one
425 * CPU to have ->on_cpu = 1 at the same time. ]
427 * task_cpu(p): is changed by set_task_cpu(), the rules are:
429 * - Don't call set_task_cpu() on a blocked task:
431 * We don't care what CPU we're not running on, this simplifies hotplug,
432 * the CPU assignment of blocked tasks isn't required to be valid.
434 * - for try_to_wake_up(), called under p->pi_lock:
436 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
438 * - for migration called under rq->lock:
439 * [ see task_on_rq_migrating() in task_rq_lock() ]
441 * o move_queued_task()
444 * - for migration called under double_rq_lock():
446 * o __migrate_swap_task()
447 * o push_rt_task() / pull_rt_task()
448 * o push_dl_task() / pull_dl_task()
449 * o dl_task_offline_migration()
453 void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
455 raw_spinlock_t *lock;
457 /* Matches synchronize_rcu() in __sched_core_enable() */
459 if (sched_core_disabled()) {
460 raw_spin_lock_nested(&rq->__lock, subclass);
461 /* preempt_count *MUST* be > 1 */
462 preempt_enable_no_resched();
467 lock = __rq_lockp(rq);
468 raw_spin_lock_nested(lock, subclass);
469 if (likely(lock == __rq_lockp(rq))) {
470 /* preempt_count *MUST* be > 1 */
471 preempt_enable_no_resched();
474 raw_spin_unlock(lock);
478 bool raw_spin_rq_trylock(struct rq *rq)
480 raw_spinlock_t *lock;
483 /* Matches synchronize_rcu() in __sched_core_enable() */
485 if (sched_core_disabled()) {
486 ret = raw_spin_trylock(&rq->__lock);
492 lock = __rq_lockp(rq);
493 ret = raw_spin_trylock(lock);
494 if (!ret || (likely(lock == __rq_lockp(rq)))) {
498 raw_spin_unlock(lock);
502 void raw_spin_rq_unlock(struct rq *rq)
504 raw_spin_unlock(rq_lockp(rq));
509 * double_rq_lock - safely lock two runqueues
511 void double_rq_lock(struct rq *rq1, struct rq *rq2)
513 lockdep_assert_irqs_disabled();
515 if (rq_order_less(rq2, rq1))
518 raw_spin_rq_lock(rq1);
519 if (__rq_lockp(rq1) == __rq_lockp(rq2))
522 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
527 * __task_rq_lock - lock the rq @p resides on.
529 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
534 lockdep_assert_held(&p->pi_lock);
538 raw_spin_rq_lock(rq);
539 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
543 raw_spin_rq_unlock(rq);
545 while (unlikely(task_on_rq_migrating(p)))
551 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
553 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
554 __acquires(p->pi_lock)
560 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
562 raw_spin_rq_lock(rq);
564 * move_queued_task() task_rq_lock()
567 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
568 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
569 * [S] ->cpu = new_cpu [L] task_rq()
573 * If we observe the old CPU in task_rq_lock(), the acquire of
574 * the old rq->lock will fully serialize against the stores.
576 * If we observe the new CPU in task_rq_lock(), the address
577 * dependency headed by '[L] rq = task_rq()' and the acquire
578 * will pair with the WMB to ensure we then also see migrating.
580 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
584 raw_spin_rq_unlock(rq);
585 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
587 while (unlikely(task_on_rq_migrating(p)))
593 * RQ-clock updating methods:
596 static void update_rq_clock_task(struct rq *rq, s64 delta)
599 * In theory, the compile should just see 0 here, and optimize out the call
600 * to sched_rt_avg_update. But I don't trust it...
602 s64 __maybe_unused steal = 0, irq_delta = 0;
604 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
605 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
608 * Since irq_time is only updated on {soft,}irq_exit, we might run into
609 * this case when a previous update_rq_clock() happened inside a
612 * When this happens, we stop ->clock_task and only update the
613 * prev_irq_time stamp to account for the part that fit, so that a next
614 * update will consume the rest. This ensures ->clock_task is
617 * It does however cause some slight miss-attribution of {soft,}irq
618 * time, a more accurate solution would be to update the irq_time using
619 * the current rq->clock timestamp, except that would require using
622 if (irq_delta > delta)
625 rq->prev_irq_time += irq_delta;
628 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
629 if (static_key_false((¶virt_steal_rq_enabled))) {
630 steal = paravirt_steal_clock(cpu_of(rq));
631 steal -= rq->prev_steal_time_rq;
633 if (unlikely(steal > delta))
636 rq->prev_steal_time_rq += steal;
641 rq->clock_task += delta;
643 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
644 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
645 update_irq_load_avg(rq, irq_delta + steal);
647 update_rq_clock_pelt(rq, delta);
650 void update_rq_clock(struct rq *rq)
654 lockdep_assert_rq_held(rq);
656 if (rq->clock_update_flags & RQCF_ACT_SKIP)
659 #ifdef CONFIG_SCHED_DEBUG
660 if (sched_feat(WARN_DOUBLE_CLOCK))
661 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
662 rq->clock_update_flags |= RQCF_UPDATED;
665 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
669 update_rq_clock_task(rq, delta);
672 #ifdef CONFIG_SCHED_HRTICK
674 * Use HR-timers to deliver accurate preemption points.
677 static void hrtick_clear(struct rq *rq)
679 if (hrtimer_active(&rq->hrtick_timer))
680 hrtimer_cancel(&rq->hrtick_timer);
684 * High-resolution timer tick.
685 * Runs from hardirq context with interrupts disabled.
687 static enum hrtimer_restart hrtick(struct hrtimer *timer)
689 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
692 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
696 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
699 return HRTIMER_NORESTART;
704 static void __hrtick_restart(struct rq *rq)
706 struct hrtimer *timer = &rq->hrtick_timer;
707 ktime_t time = rq->hrtick_time;
709 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
713 * called from hardirq (IPI) context
715 static void __hrtick_start(void *arg)
721 __hrtick_restart(rq);
726 * Called to set the hrtick timer state.
728 * called with rq->lock held and irqs disabled
730 void hrtick_start(struct rq *rq, u64 delay)
732 struct hrtimer *timer = &rq->hrtick_timer;
736 * Don't schedule slices shorter than 10000ns, that just
737 * doesn't make sense and can cause timer DoS.
739 delta = max_t(s64, delay, 10000LL);
740 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
743 __hrtick_restart(rq);
745 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
750 * Called to set the hrtick timer state.
752 * called with rq->lock held and irqs disabled
754 void hrtick_start(struct rq *rq, u64 delay)
757 * Don't schedule slices shorter than 10000ns, that just
758 * doesn't make sense. Rely on vruntime for fairness.
760 delay = max_t(u64, delay, 10000LL);
761 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
762 HRTIMER_MODE_REL_PINNED_HARD);
765 #endif /* CONFIG_SMP */
767 static void hrtick_rq_init(struct rq *rq)
770 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
772 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
773 rq->hrtick_timer.function = hrtick;
775 #else /* CONFIG_SCHED_HRTICK */
776 static inline void hrtick_clear(struct rq *rq)
780 static inline void hrtick_rq_init(struct rq *rq)
783 #endif /* CONFIG_SCHED_HRTICK */
786 * cmpxchg based fetch_or, macro so it works for different integer types
788 #define fetch_or(ptr, mask) \
790 typeof(ptr) _ptr = (ptr); \
791 typeof(mask) _mask = (mask); \
792 typeof(*_ptr) _old, _val = *_ptr; \
795 _old = cmpxchg(_ptr, _val, _val | _mask); \
803 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
805 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
806 * this avoids any races wrt polling state changes and thereby avoids
809 static bool set_nr_and_not_polling(struct task_struct *p)
811 struct thread_info *ti = task_thread_info(p);
812 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
816 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
818 * If this returns true, then the idle task promises to call
819 * sched_ttwu_pending() and reschedule soon.
821 static bool set_nr_if_polling(struct task_struct *p)
823 struct thread_info *ti = task_thread_info(p);
824 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
827 if (!(val & _TIF_POLLING_NRFLAG))
829 if (val & _TIF_NEED_RESCHED)
831 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
840 static bool set_nr_and_not_polling(struct task_struct *p)
842 set_tsk_need_resched(p);
847 static bool set_nr_if_polling(struct task_struct *p)
854 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
856 struct wake_q_node *node = &task->wake_q;
859 * Atomically grab the task, if ->wake_q is !nil already it means
860 * it's already queued (either by us or someone else) and will get the
861 * wakeup due to that.
863 * In order to ensure that a pending wakeup will observe our pending
864 * state, even in the failed case, an explicit smp_mb() must be used.
866 smp_mb__before_atomic();
867 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
871 * The head is context local, there can be no concurrency.
874 head->lastp = &node->next;
879 * wake_q_add() - queue a wakeup for 'later' waking.
880 * @head: the wake_q_head to add @task to
881 * @task: the task to queue for 'later' wakeup
883 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
884 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
887 * This function must be used as-if it were wake_up_process(); IOW the task
888 * must be ready to be woken at this location.
890 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
892 if (__wake_q_add(head, task))
893 get_task_struct(task);
897 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
898 * @head: the wake_q_head to add @task to
899 * @task: the task to queue for 'later' wakeup
901 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
902 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
905 * This function must be used as-if it were wake_up_process(); IOW the task
906 * must be ready to be woken at this location.
908 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
909 * that already hold reference to @task can call the 'safe' version and trust
910 * wake_q to do the right thing depending whether or not the @task is already
913 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
915 if (!__wake_q_add(head, task))
916 put_task_struct(task);
919 void wake_up_q(struct wake_q_head *head)
921 struct wake_q_node *node = head->first;
923 while (node != WAKE_Q_TAIL) {
924 struct task_struct *task;
926 task = container_of(node, struct task_struct, wake_q);
927 /* Task can safely be re-inserted now: */
929 task->wake_q.next = NULL;
932 * wake_up_process() executes a full barrier, which pairs with
933 * the queueing in wake_q_add() so as not to miss wakeups.
935 wake_up_process(task);
936 put_task_struct(task);
941 * resched_curr - mark rq's current task 'to be rescheduled now'.
943 * On UP this means the setting of the need_resched flag, on SMP it
944 * might also involve a cross-CPU call to trigger the scheduler on
947 void resched_curr(struct rq *rq)
949 struct task_struct *curr = rq->curr;
952 lockdep_assert_rq_held(rq);
954 if (test_tsk_need_resched(curr))
959 if (cpu == smp_processor_id()) {
960 set_tsk_need_resched(curr);
961 set_preempt_need_resched();
965 if (set_nr_and_not_polling(curr))
966 smp_send_reschedule(cpu);
968 trace_sched_wake_idle_without_ipi(cpu);
971 void resched_cpu(int cpu)
973 struct rq *rq = cpu_rq(cpu);
976 raw_spin_rq_lock_irqsave(rq, flags);
977 if (cpu_online(cpu) || cpu == smp_processor_id())
979 raw_spin_rq_unlock_irqrestore(rq, flags);
983 #ifdef CONFIG_NO_HZ_COMMON
985 * In the semi idle case, use the nearest busy CPU for migrating timers
986 * from an idle CPU. This is good for power-savings.
988 * We don't do similar optimization for completely idle system, as
989 * selecting an idle CPU will add more delays to the timers than intended
990 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
992 int get_nohz_timer_target(void)
994 int i, cpu = smp_processor_id(), default_cpu = -1;
995 struct sched_domain *sd;
996 const struct cpumask *hk_mask;
998 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
1004 hk_mask = housekeeping_cpumask(HK_FLAG_TIMER);
1007 for_each_domain(cpu, sd) {
1008 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
1019 if (default_cpu == -1)
1020 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
1028 * When add_timer_on() enqueues a timer into the timer wheel of an
1029 * idle CPU then this timer might expire before the next timer event
1030 * which is scheduled to wake up that CPU. In case of a completely
1031 * idle system the next event might even be infinite time into the
1032 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1033 * leaves the inner idle loop so the newly added timer is taken into
1034 * account when the CPU goes back to idle and evaluates the timer
1035 * wheel for the next timer event.
1037 static void wake_up_idle_cpu(int cpu)
1039 struct rq *rq = cpu_rq(cpu);
1041 if (cpu == smp_processor_id())
1044 if (set_nr_and_not_polling(rq->idle))
1045 smp_send_reschedule(cpu);
1047 trace_sched_wake_idle_without_ipi(cpu);
1050 static bool wake_up_full_nohz_cpu(int cpu)
1053 * We just need the target to call irq_exit() and re-evaluate
1054 * the next tick. The nohz full kick at least implies that.
1055 * If needed we can still optimize that later with an
1058 if (cpu_is_offline(cpu))
1059 return true; /* Don't try to wake offline CPUs. */
1060 if (tick_nohz_full_cpu(cpu)) {
1061 if (cpu != smp_processor_id() ||
1062 tick_nohz_tick_stopped())
1063 tick_nohz_full_kick_cpu(cpu);
1071 * Wake up the specified CPU. If the CPU is going offline, it is the
1072 * caller's responsibility to deal with the lost wakeup, for example,
1073 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1075 void wake_up_nohz_cpu(int cpu)
1077 if (!wake_up_full_nohz_cpu(cpu))
1078 wake_up_idle_cpu(cpu);
1081 static void nohz_csd_func(void *info)
1083 struct rq *rq = info;
1084 int cpu = cpu_of(rq);
1088 * Release the rq::nohz_csd.
1090 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
1091 WARN_ON(!(flags & NOHZ_KICK_MASK));
1093 rq->idle_balance = idle_cpu(cpu);
1094 if (rq->idle_balance && !need_resched()) {
1095 rq->nohz_idle_balance = flags;
1096 raise_softirq_irqoff(SCHED_SOFTIRQ);
1100 #endif /* CONFIG_NO_HZ_COMMON */
1102 #ifdef CONFIG_NO_HZ_FULL
1103 bool sched_can_stop_tick(struct rq *rq)
1105 int fifo_nr_running;
1107 /* Deadline tasks, even if single, need the tick */
1108 if (rq->dl.dl_nr_running)
1112 * If there are more than one RR tasks, we need the tick to affect the
1113 * actual RR behaviour.
1115 if (rq->rt.rr_nr_running) {
1116 if (rq->rt.rr_nr_running == 1)
1123 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1124 * forced preemption between FIFO tasks.
1126 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1127 if (fifo_nr_running)
1131 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1132 * if there's more than one we need the tick for involuntary
1135 if (rq->nr_running > 1)
1140 #endif /* CONFIG_NO_HZ_FULL */
1141 #endif /* CONFIG_SMP */
1143 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1144 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1146 * Iterate task_group tree rooted at *from, calling @down when first entering a
1147 * node and @up when leaving it for the final time.
1149 * Caller must hold rcu_lock or sufficient equivalent.
1151 int walk_tg_tree_from(struct task_group *from,
1152 tg_visitor down, tg_visitor up, void *data)
1154 struct task_group *parent, *child;
1160 ret = (*down)(parent, data);
1163 list_for_each_entry_rcu(child, &parent->children, siblings) {
1170 ret = (*up)(parent, data);
1171 if (ret || parent == from)
1175 parent = parent->parent;
1182 int tg_nop(struct task_group *tg, void *data)
1188 static void set_load_weight(struct task_struct *p, bool update_load)
1190 int prio = p->static_prio - MAX_RT_PRIO;
1191 struct load_weight *load = &p->se.load;
1194 * SCHED_IDLE tasks get minimal weight:
1196 if (task_has_idle_policy(p)) {
1197 load->weight = scale_load(WEIGHT_IDLEPRIO);
1198 load->inv_weight = WMULT_IDLEPRIO;
1203 * SCHED_OTHER tasks have to update their load when changing their
1206 if (update_load && p->sched_class == &fair_sched_class) {
1207 reweight_task(p, prio);
1209 load->weight = scale_load(sched_prio_to_weight[prio]);
1210 load->inv_weight = sched_prio_to_wmult[prio];
1214 #ifdef CONFIG_UCLAMP_TASK
1216 * Serializes updates of utilization clamp values
1218 * The (slow-path) user-space triggers utilization clamp value updates which
1219 * can require updates on (fast-path) scheduler's data structures used to
1220 * support enqueue/dequeue operations.
1221 * While the per-CPU rq lock protects fast-path update operations, user-space
1222 * requests are serialized using a mutex to reduce the risk of conflicting
1223 * updates or API abuses.
1225 static DEFINE_MUTEX(uclamp_mutex);
1227 /* Max allowed minimum utilization */
1228 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
1230 /* Max allowed maximum utilization */
1231 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
1234 * By default RT tasks run at the maximum performance point/capacity of the
1235 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1236 * SCHED_CAPACITY_SCALE.
1238 * This knob allows admins to change the default behavior when uclamp is being
1239 * used. In battery powered devices, particularly, running at the maximum
1240 * capacity and frequency will increase energy consumption and shorten the
1243 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1245 * This knob will not override the system default sched_util_clamp_min defined
1248 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
1250 /* All clamps are required to be less or equal than these values */
1251 static struct uclamp_se uclamp_default[UCLAMP_CNT];
1254 * This static key is used to reduce the uclamp overhead in the fast path. It
1255 * primarily disables the call to uclamp_rq_{inc, dec}() in
1256 * enqueue/dequeue_task().
1258 * This allows users to continue to enable uclamp in their kernel config with
1259 * minimum uclamp overhead in the fast path.
1261 * As soon as userspace modifies any of the uclamp knobs, the static key is
1262 * enabled, since we have an actual users that make use of uclamp
1265 * The knobs that would enable this static key are:
1267 * * A task modifying its uclamp value with sched_setattr().
1268 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1269 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1271 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1273 /* Integer rounded range for each bucket */
1274 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1276 #define for_each_clamp_id(clamp_id) \
1277 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1279 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1281 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
1284 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
1286 if (clamp_id == UCLAMP_MIN)
1288 return SCHED_CAPACITY_SCALE;
1291 static inline void uclamp_se_set(struct uclamp_se *uc_se,
1292 unsigned int value, bool user_defined)
1294 uc_se->value = value;
1295 uc_se->bucket_id = uclamp_bucket_id(value);
1296 uc_se->user_defined = user_defined;
1299 static inline unsigned int
1300 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
1301 unsigned int clamp_value)
1304 * Avoid blocked utilization pushing up the frequency when we go
1305 * idle (which drops the max-clamp) by retaining the last known
1308 if (clamp_id == UCLAMP_MAX) {
1309 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1313 return uclamp_none(UCLAMP_MIN);
1316 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
1317 unsigned int clamp_value)
1319 /* Reset max-clamp retention only on idle exit */
1320 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1323 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
1327 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
1328 unsigned int clamp_value)
1330 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1331 int bucket_id = UCLAMP_BUCKETS - 1;
1334 * Since both min and max clamps are max aggregated, find the
1335 * top most bucket with tasks in.
1337 for ( ; bucket_id >= 0; bucket_id--) {
1338 if (!bucket[bucket_id].tasks)
1340 return bucket[bucket_id].value;
1343 /* No tasks -- default clamp values */
1344 return uclamp_idle_value(rq, clamp_id, clamp_value);
1347 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1349 unsigned int default_util_min;
1350 struct uclamp_se *uc_se;
1352 lockdep_assert_held(&p->pi_lock);
1354 uc_se = &p->uclamp_req[UCLAMP_MIN];
1356 /* Only sync if user didn't override the default */
1357 if (uc_se->user_defined)
1360 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1361 uclamp_se_set(uc_se, default_util_min, false);
1364 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1372 /* Protect updates to p->uclamp_* */
1373 rq = task_rq_lock(p, &rf);
1374 __uclamp_update_util_min_rt_default(p);
1375 task_rq_unlock(rq, p, &rf);
1378 static void uclamp_sync_util_min_rt_default(void)
1380 struct task_struct *g, *p;
1383 * copy_process() sysctl_uclamp
1384 * uclamp_min_rt = X;
1385 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1386 * // link thread smp_mb__after_spinlock()
1387 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1388 * sched_post_fork() for_each_process_thread()
1389 * __uclamp_sync_rt() __uclamp_sync_rt()
1391 * Ensures that either sched_post_fork() will observe the new
1392 * uclamp_min_rt or for_each_process_thread() will observe the new
1395 read_lock(&tasklist_lock);
1396 smp_mb__after_spinlock();
1397 read_unlock(&tasklist_lock);
1400 for_each_process_thread(g, p)
1401 uclamp_update_util_min_rt_default(p);
1405 static inline struct uclamp_se
1406 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1408 /* Copy by value as we could modify it */
1409 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1410 #ifdef CONFIG_UCLAMP_TASK_GROUP
1411 unsigned int tg_min, tg_max, value;
1414 * Tasks in autogroups or root task group will be
1415 * restricted by system defaults.
1417 if (task_group_is_autogroup(task_group(p)))
1419 if (task_group(p) == &root_task_group)
1422 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1423 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1424 value = uc_req.value;
1425 value = clamp(value, tg_min, tg_max);
1426 uclamp_se_set(&uc_req, value, false);
1433 * The effective clamp bucket index of a task depends on, by increasing
1435 * - the task specific clamp value, when explicitly requested from userspace
1436 * - the task group effective clamp value, for tasks not either in the root
1437 * group or in an autogroup
1438 * - the system default clamp value, defined by the sysadmin
1440 static inline struct uclamp_se
1441 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1443 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1444 struct uclamp_se uc_max = uclamp_default[clamp_id];
1446 /* System default restrictions always apply */
1447 if (unlikely(uc_req.value > uc_max.value))
1453 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1455 struct uclamp_se uc_eff;
1457 /* Task currently refcounted: use back-annotated (effective) value */
1458 if (p->uclamp[clamp_id].active)
1459 return (unsigned long)p->uclamp[clamp_id].value;
1461 uc_eff = uclamp_eff_get(p, clamp_id);
1463 return (unsigned long)uc_eff.value;
1467 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1468 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1469 * updates the rq's clamp value if required.
1471 * Tasks can have a task-specific value requested from user-space, track
1472 * within each bucket the maximum value for tasks refcounted in it.
1473 * This "local max aggregation" allows to track the exact "requested" value
1474 * for each bucket when all its RUNNABLE tasks require the same clamp.
1476 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1477 enum uclamp_id clamp_id)
1479 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1480 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1481 struct uclamp_bucket *bucket;
1483 lockdep_assert_rq_held(rq);
1485 /* Update task effective clamp */
1486 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1488 bucket = &uc_rq->bucket[uc_se->bucket_id];
1490 uc_se->active = true;
1492 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1495 * Local max aggregation: rq buckets always track the max
1496 * "requested" clamp value of its RUNNABLE tasks.
1498 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1499 bucket->value = uc_se->value;
1501 if (uc_se->value > READ_ONCE(uc_rq->value))
1502 WRITE_ONCE(uc_rq->value, uc_se->value);
1506 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1507 * is released. If this is the last task reference counting the rq's max
1508 * active clamp value, then the rq's clamp value is updated.
1510 * Both refcounted tasks and rq's cached clamp values are expected to be
1511 * always valid. If it's detected they are not, as defensive programming,
1512 * enforce the expected state and warn.
1514 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1515 enum uclamp_id clamp_id)
1517 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1518 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1519 struct uclamp_bucket *bucket;
1520 unsigned int bkt_clamp;
1521 unsigned int rq_clamp;
1523 lockdep_assert_rq_held(rq);
1526 * If sched_uclamp_used was enabled after task @p was enqueued,
1527 * we could end up with unbalanced call to uclamp_rq_dec_id().
1529 * In this case the uc_se->active flag should be false since no uclamp
1530 * accounting was performed at enqueue time and we can just return
1533 * Need to be careful of the following enqueue/dequeue ordering
1537 * // sched_uclamp_used gets enabled
1540 * // Must not decrement bucket->tasks here
1543 * where we could end up with stale data in uc_se and
1544 * bucket[uc_se->bucket_id].
1546 * The following check here eliminates the possibility of such race.
1548 if (unlikely(!uc_se->active))
1551 bucket = &uc_rq->bucket[uc_se->bucket_id];
1553 SCHED_WARN_ON(!bucket->tasks);
1554 if (likely(bucket->tasks))
1557 uc_se->active = false;
1560 * Keep "local max aggregation" simple and accept to (possibly)
1561 * overboost some RUNNABLE tasks in the same bucket.
1562 * The rq clamp bucket value is reset to its base value whenever
1563 * there are no more RUNNABLE tasks refcounting it.
1565 if (likely(bucket->tasks))
1568 rq_clamp = READ_ONCE(uc_rq->value);
1570 * Defensive programming: this should never happen. If it happens,
1571 * e.g. due to future modification, warn and fixup the expected value.
1573 SCHED_WARN_ON(bucket->value > rq_clamp);
1574 if (bucket->value >= rq_clamp) {
1575 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1576 WRITE_ONCE(uc_rq->value, bkt_clamp);
1580 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1582 enum uclamp_id clamp_id;
1585 * Avoid any overhead until uclamp is actually used by the userspace.
1587 * The condition is constructed such that a NOP is generated when
1588 * sched_uclamp_used is disabled.
1590 if (!static_branch_unlikely(&sched_uclamp_used))
1593 if (unlikely(!p->sched_class->uclamp_enabled))
1596 for_each_clamp_id(clamp_id)
1597 uclamp_rq_inc_id(rq, p, clamp_id);
1599 /* Reset clamp idle holding when there is one RUNNABLE task */
1600 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1601 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1604 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1606 enum uclamp_id clamp_id;
1609 * Avoid any overhead until uclamp is actually used by the userspace.
1611 * The condition is constructed such that a NOP is generated when
1612 * sched_uclamp_used is disabled.
1614 if (!static_branch_unlikely(&sched_uclamp_used))
1617 if (unlikely(!p->sched_class->uclamp_enabled))
1620 for_each_clamp_id(clamp_id)
1621 uclamp_rq_dec_id(rq, p, clamp_id);
1624 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1625 enum uclamp_id clamp_id)
1627 if (!p->uclamp[clamp_id].active)
1630 uclamp_rq_dec_id(rq, p, clamp_id);
1631 uclamp_rq_inc_id(rq, p, clamp_id);
1634 * Make sure to clear the idle flag if we've transiently reached 0
1635 * active tasks on rq.
1637 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1638 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1642 uclamp_update_active(struct task_struct *p)
1644 enum uclamp_id clamp_id;
1649 * Lock the task and the rq where the task is (or was) queued.
1651 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1652 * price to pay to safely serialize util_{min,max} updates with
1653 * enqueues, dequeues and migration operations.
1654 * This is the same locking schema used by __set_cpus_allowed_ptr().
1656 rq = task_rq_lock(p, &rf);
1659 * Setting the clamp bucket is serialized by task_rq_lock().
1660 * If the task is not yet RUNNABLE and its task_struct is not
1661 * affecting a valid clamp bucket, the next time it's enqueued,
1662 * it will already see the updated clamp bucket value.
1664 for_each_clamp_id(clamp_id)
1665 uclamp_rq_reinc_id(rq, p, clamp_id);
1667 task_rq_unlock(rq, p, &rf);
1670 #ifdef CONFIG_UCLAMP_TASK_GROUP
1672 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1674 struct css_task_iter it;
1675 struct task_struct *p;
1677 css_task_iter_start(css, 0, &it);
1678 while ((p = css_task_iter_next(&it)))
1679 uclamp_update_active(p);
1680 css_task_iter_end(&it);
1683 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1684 static void uclamp_update_root_tg(void)
1686 struct task_group *tg = &root_task_group;
1688 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1689 sysctl_sched_uclamp_util_min, false);
1690 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1691 sysctl_sched_uclamp_util_max, false);
1694 cpu_util_update_eff(&root_task_group.css);
1698 static void uclamp_update_root_tg(void) { }
1701 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1702 void *buffer, size_t *lenp, loff_t *ppos)
1704 bool update_root_tg = false;
1705 int old_min, old_max, old_min_rt;
1708 mutex_lock(&uclamp_mutex);
1709 old_min = sysctl_sched_uclamp_util_min;
1710 old_max = sysctl_sched_uclamp_util_max;
1711 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1713 result = proc_dointvec(table, write, buffer, lenp, ppos);
1719 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1720 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1721 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1727 if (old_min != sysctl_sched_uclamp_util_min) {
1728 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1729 sysctl_sched_uclamp_util_min, false);
1730 update_root_tg = true;
1732 if (old_max != sysctl_sched_uclamp_util_max) {
1733 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1734 sysctl_sched_uclamp_util_max, false);
1735 update_root_tg = true;
1738 if (update_root_tg) {
1739 static_branch_enable(&sched_uclamp_used);
1740 uclamp_update_root_tg();
1743 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1744 static_branch_enable(&sched_uclamp_used);
1745 uclamp_sync_util_min_rt_default();
1749 * We update all RUNNABLE tasks only when task groups are in use.
1750 * Otherwise, keep it simple and do just a lazy update at each next
1751 * task enqueue time.
1757 sysctl_sched_uclamp_util_min = old_min;
1758 sysctl_sched_uclamp_util_max = old_max;
1759 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1761 mutex_unlock(&uclamp_mutex);
1766 static int uclamp_validate(struct task_struct *p,
1767 const struct sched_attr *attr)
1769 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1770 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1772 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1773 util_min = attr->sched_util_min;
1775 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1779 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1780 util_max = attr->sched_util_max;
1782 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1786 if (util_min != -1 && util_max != -1 && util_min > util_max)
1790 * We have valid uclamp attributes; make sure uclamp is enabled.
1792 * We need to do that here, because enabling static branches is a
1793 * blocking operation which obviously cannot be done while holding
1796 static_branch_enable(&sched_uclamp_used);
1801 static bool uclamp_reset(const struct sched_attr *attr,
1802 enum uclamp_id clamp_id,
1803 struct uclamp_se *uc_se)
1805 /* Reset on sched class change for a non user-defined clamp value. */
1806 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1807 !uc_se->user_defined)
1810 /* Reset on sched_util_{min,max} == -1. */
1811 if (clamp_id == UCLAMP_MIN &&
1812 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1813 attr->sched_util_min == -1) {
1817 if (clamp_id == UCLAMP_MAX &&
1818 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1819 attr->sched_util_max == -1) {
1826 static void __setscheduler_uclamp(struct task_struct *p,
1827 const struct sched_attr *attr)
1829 enum uclamp_id clamp_id;
1831 for_each_clamp_id(clamp_id) {
1832 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1835 if (!uclamp_reset(attr, clamp_id, uc_se))
1839 * RT by default have a 100% boost value that could be modified
1842 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1843 value = sysctl_sched_uclamp_util_min_rt_default;
1845 value = uclamp_none(clamp_id);
1847 uclamp_se_set(uc_se, value, false);
1851 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1854 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1855 attr->sched_util_min != -1) {
1856 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1857 attr->sched_util_min, true);
1860 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1861 attr->sched_util_max != -1) {
1862 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1863 attr->sched_util_max, true);
1867 static void uclamp_fork(struct task_struct *p)
1869 enum uclamp_id clamp_id;
1872 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1873 * as the task is still at its early fork stages.
1875 for_each_clamp_id(clamp_id)
1876 p->uclamp[clamp_id].active = false;
1878 if (likely(!p->sched_reset_on_fork))
1881 for_each_clamp_id(clamp_id) {
1882 uclamp_se_set(&p->uclamp_req[clamp_id],
1883 uclamp_none(clamp_id), false);
1887 static void uclamp_post_fork(struct task_struct *p)
1889 uclamp_update_util_min_rt_default(p);
1892 static void __init init_uclamp_rq(struct rq *rq)
1894 enum uclamp_id clamp_id;
1895 struct uclamp_rq *uc_rq = rq->uclamp;
1897 for_each_clamp_id(clamp_id) {
1898 uc_rq[clamp_id] = (struct uclamp_rq) {
1899 .value = uclamp_none(clamp_id)
1903 rq->uclamp_flags = 0;
1906 static void __init init_uclamp(void)
1908 struct uclamp_se uc_max = {};
1909 enum uclamp_id clamp_id;
1912 for_each_possible_cpu(cpu)
1913 init_uclamp_rq(cpu_rq(cpu));
1915 for_each_clamp_id(clamp_id) {
1916 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1917 uclamp_none(clamp_id), false);
1920 /* System defaults allow max clamp values for both indexes */
1921 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1922 for_each_clamp_id(clamp_id) {
1923 uclamp_default[clamp_id] = uc_max;
1924 #ifdef CONFIG_UCLAMP_TASK_GROUP
1925 root_task_group.uclamp_req[clamp_id] = uc_max;
1926 root_task_group.uclamp[clamp_id] = uc_max;
1931 #else /* CONFIG_UCLAMP_TASK */
1932 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1933 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1934 static inline int uclamp_validate(struct task_struct *p,
1935 const struct sched_attr *attr)
1939 static void __setscheduler_uclamp(struct task_struct *p,
1940 const struct sched_attr *attr) { }
1941 static inline void uclamp_fork(struct task_struct *p) { }
1942 static inline void uclamp_post_fork(struct task_struct *p) { }
1943 static inline void init_uclamp(void) { }
1944 #endif /* CONFIG_UCLAMP_TASK */
1946 bool sched_task_on_rq(struct task_struct *p)
1948 return task_on_rq_queued(p);
1951 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1953 if (!(flags & ENQUEUE_NOCLOCK))
1954 update_rq_clock(rq);
1956 if (!(flags & ENQUEUE_RESTORE)) {
1957 sched_info_enqueue(rq, p);
1958 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1961 uclamp_rq_inc(rq, p);
1962 p->sched_class->enqueue_task(rq, p, flags);
1964 if (sched_core_enabled(rq))
1965 sched_core_enqueue(rq, p);
1968 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1970 if (sched_core_enabled(rq))
1971 sched_core_dequeue(rq, p);
1973 if (!(flags & DEQUEUE_NOCLOCK))
1974 update_rq_clock(rq);
1976 if (!(flags & DEQUEUE_SAVE)) {
1977 sched_info_dequeue(rq, p);
1978 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1981 uclamp_rq_dec(rq, p);
1982 p->sched_class->dequeue_task(rq, p, flags);
1985 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1987 enqueue_task(rq, p, flags);
1989 p->on_rq = TASK_ON_RQ_QUEUED;
1992 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1994 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1996 dequeue_task(rq, p, flags);
1999 static inline int __normal_prio(int policy, int rt_prio, int nice)
2003 if (dl_policy(policy))
2004 prio = MAX_DL_PRIO - 1;
2005 else if (rt_policy(policy))
2006 prio = MAX_RT_PRIO - 1 - rt_prio;
2008 prio = NICE_TO_PRIO(nice);
2014 * Calculate the expected normal priority: i.e. priority
2015 * without taking RT-inheritance into account. Might be
2016 * boosted by interactivity modifiers. Changes upon fork,
2017 * setprio syscalls, and whenever the interactivity
2018 * estimator recalculates.
2020 static inline int normal_prio(struct task_struct *p)
2022 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
2026 * Calculate the current priority, i.e. the priority
2027 * taken into account by the scheduler. This value might
2028 * be boosted by RT tasks, or might be boosted by
2029 * interactivity modifiers. Will be RT if the task got
2030 * RT-boosted. If not then it returns p->normal_prio.
2032 static int effective_prio(struct task_struct *p)
2034 p->normal_prio = normal_prio(p);
2036 * If we are RT tasks or we were boosted to RT priority,
2037 * keep the priority unchanged. Otherwise, update priority
2038 * to the normal priority:
2040 if (!rt_prio(p->prio))
2041 return p->normal_prio;
2046 * task_curr - is this task currently executing on a CPU?
2047 * @p: the task in question.
2049 * Return: 1 if the task is currently executing. 0 otherwise.
2051 inline int task_curr(const struct task_struct *p)
2053 return cpu_curr(task_cpu(p)) == p;
2057 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2058 * use the balance_callback list if you want balancing.
2060 * this means any call to check_class_changed() must be followed by a call to
2061 * balance_callback().
2063 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2064 const struct sched_class *prev_class,
2067 if (prev_class != p->sched_class) {
2068 if (prev_class->switched_from)
2069 prev_class->switched_from(rq, p);
2071 p->sched_class->switched_to(rq, p);
2072 } else if (oldprio != p->prio || dl_task(p))
2073 p->sched_class->prio_changed(rq, p, oldprio);
2076 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2078 if (p->sched_class == rq->curr->sched_class)
2079 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2080 else if (p->sched_class > rq->curr->sched_class)
2084 * A queue event has occurred, and we're going to schedule. In
2085 * this case, we can save a useless back to back clock update.
2087 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
2088 rq_clock_skip_update(rq);
2094 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
2096 static int __set_cpus_allowed_ptr(struct task_struct *p,
2097 const struct cpumask *new_mask,
2100 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2102 if (likely(!p->migration_disabled))
2105 if (p->cpus_ptr != &p->cpus_mask)
2109 * Violates locking rules! see comment in __do_set_cpus_allowed().
2111 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
2114 void migrate_disable(void)
2116 struct task_struct *p = current;
2118 if (p->migration_disabled) {
2119 p->migration_disabled++;
2124 this_rq()->nr_pinned++;
2125 p->migration_disabled = 1;
2128 EXPORT_SYMBOL_GPL(migrate_disable);
2130 void migrate_enable(void)
2132 struct task_struct *p = current;
2134 if (p->migration_disabled > 1) {
2135 p->migration_disabled--;
2140 * Ensure stop_task runs either before or after this, and that
2141 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2144 if (p->cpus_ptr != &p->cpus_mask)
2145 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
2147 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2148 * regular cpus_mask, otherwise things that race (eg.
2149 * select_fallback_rq) get confused.
2152 p->migration_disabled = 0;
2153 this_rq()->nr_pinned--;
2156 EXPORT_SYMBOL_GPL(migrate_enable);
2158 static inline bool rq_has_pinned_tasks(struct rq *rq)
2160 return rq->nr_pinned;
2164 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
2165 * __set_cpus_allowed_ptr() and select_fallback_rq().
2167 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2169 /* When not in the task's cpumask, no point in looking further. */
2170 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
2173 /* migrate_disabled() must be allowed to finish. */
2174 if (is_migration_disabled(p))
2175 return cpu_online(cpu);
2177 /* Non kernel threads are not allowed during either online or offline. */
2178 if (!(p->flags & PF_KTHREAD))
2179 return cpu_active(cpu) && task_cpu_possible(cpu, p);
2181 /* KTHREAD_IS_PER_CPU is always allowed. */
2182 if (kthread_is_per_cpu(p))
2183 return cpu_online(cpu);
2185 /* Regular kernel threads don't get to stay during offline. */
2189 /* But are allowed during online. */
2190 return cpu_online(cpu);
2194 * This is how migration works:
2196 * 1) we invoke migration_cpu_stop() on the target CPU using
2198 * 2) stopper starts to run (implicitly forcing the migrated thread
2200 * 3) it checks whether the migrated task is still in the wrong runqueue.
2201 * 4) if it's in the wrong runqueue then the migration thread removes
2202 * it and puts it into the right queue.
2203 * 5) stopper completes and stop_one_cpu() returns and the migration
2208 * move_queued_task - move a queued task to new rq.
2210 * Returns (locked) new rq. Old rq's lock is released.
2212 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2213 struct task_struct *p, int new_cpu)
2215 lockdep_assert_rq_held(rq);
2217 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
2218 set_task_cpu(p, new_cpu);
2221 rq = cpu_rq(new_cpu);
2224 BUG_ON(task_cpu(p) != new_cpu);
2225 activate_task(rq, p, 0);
2226 check_preempt_curr(rq, p, 0);
2231 struct migration_arg {
2232 struct task_struct *task;
2234 struct set_affinity_pending *pending;
2238 * @refs: number of wait_for_completion()
2239 * @stop_pending: is @stop_work in use
2241 struct set_affinity_pending {
2243 unsigned int stop_pending;
2244 struct completion done;
2245 struct cpu_stop_work stop_work;
2246 struct migration_arg arg;
2250 * Move (not current) task off this CPU, onto the destination CPU. We're doing
2251 * this because either it can't run here any more (set_cpus_allowed()
2252 * away from this CPU, or CPU going down), or because we're
2253 * attempting to rebalance this task on exec (sched_exec).
2255 * So we race with normal scheduler movements, but that's OK, as long
2256 * as the task is no longer on this CPU.
2258 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2259 struct task_struct *p, int dest_cpu)
2261 /* Affinity changed (again). */
2262 if (!is_cpu_allowed(p, dest_cpu))
2265 update_rq_clock(rq);
2266 rq = move_queued_task(rq, rf, p, dest_cpu);
2272 * migration_cpu_stop - this will be executed by a highprio stopper thread
2273 * and performs thread migration by bumping thread off CPU then
2274 * 'pushing' onto another runqueue.
2276 static int migration_cpu_stop(void *data)
2278 struct migration_arg *arg = data;
2279 struct set_affinity_pending *pending = arg->pending;
2280 struct task_struct *p = arg->task;
2281 struct rq *rq = this_rq();
2282 bool complete = false;
2286 * The original target CPU might have gone down and we might
2287 * be on another CPU but it doesn't matter.
2289 local_irq_save(rf.flags);
2291 * We need to explicitly wake pending tasks before running
2292 * __migrate_task() such that we will not miss enforcing cpus_ptr
2293 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2295 flush_smp_call_function_from_idle();
2297 raw_spin_lock(&p->pi_lock);
2301 * If we were passed a pending, then ->stop_pending was set, thus
2302 * p->migration_pending must have remained stable.
2304 WARN_ON_ONCE(pending && pending != p->migration_pending);
2307 * If task_rq(p) != rq, it cannot be migrated here, because we're
2308 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2309 * we're holding p->pi_lock.
2311 if (task_rq(p) == rq) {
2312 if (is_migration_disabled(p))
2316 p->migration_pending = NULL;
2319 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2323 if (task_on_rq_queued(p))
2324 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
2326 p->wake_cpu = arg->dest_cpu;
2329 * XXX __migrate_task() can fail, at which point we might end
2330 * up running on a dodgy CPU, AFAICT this can only happen
2331 * during CPU hotplug, at which point we'll get pushed out
2332 * anyway, so it's probably not a big deal.
2335 } else if (pending) {
2337 * This happens when we get migrated between migrate_enable()'s
2338 * preempt_enable() and scheduling the stopper task. At that
2339 * point we're a regular task again and not current anymore.
2341 * A !PREEMPT kernel has a giant hole here, which makes it far
2346 * The task moved before the stopper got to run. We're holding
2347 * ->pi_lock, so the allowed mask is stable - if it got
2348 * somewhere allowed, we're done.
2350 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
2351 p->migration_pending = NULL;
2357 * When migrate_enable() hits a rq mis-match we can't reliably
2358 * determine is_migration_disabled() and so have to chase after
2361 WARN_ON_ONCE(!pending->stop_pending);
2362 task_rq_unlock(rq, p, &rf);
2363 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2364 &pending->arg, &pending->stop_work);
2369 pending->stop_pending = false;
2370 task_rq_unlock(rq, p, &rf);
2373 complete_all(&pending->done);
2378 int push_cpu_stop(void *arg)
2380 struct rq *lowest_rq = NULL, *rq = this_rq();
2381 struct task_struct *p = arg;
2383 raw_spin_lock_irq(&p->pi_lock);
2384 raw_spin_rq_lock(rq);
2386 if (task_rq(p) != rq)
2389 if (is_migration_disabled(p)) {
2390 p->migration_flags |= MDF_PUSH;
2394 p->migration_flags &= ~MDF_PUSH;
2396 if (p->sched_class->find_lock_rq)
2397 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2402 // XXX validate p is still the highest prio task
2403 if (task_rq(p) == rq) {
2404 deactivate_task(rq, p, 0);
2405 set_task_cpu(p, lowest_rq->cpu);
2406 activate_task(lowest_rq, p, 0);
2407 resched_curr(lowest_rq);
2410 double_unlock_balance(rq, lowest_rq);
2413 rq->push_busy = false;
2414 raw_spin_rq_unlock(rq);
2415 raw_spin_unlock_irq(&p->pi_lock);
2422 * sched_class::set_cpus_allowed must do the below, but is not required to
2423 * actually call this function.
2425 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2427 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2428 p->cpus_ptr = new_mask;
2432 cpumask_copy(&p->cpus_mask, new_mask);
2433 p->nr_cpus_allowed = cpumask_weight(new_mask);
2437 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2439 struct rq *rq = task_rq(p);
2440 bool queued, running;
2443 * This here violates the locking rules for affinity, since we're only
2444 * supposed to change these variables while holding both rq->lock and
2447 * HOWEVER, it magically works, because ttwu() is the only code that
2448 * accesses these variables under p->pi_lock and only does so after
2449 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2450 * before finish_task().
2452 * XXX do further audits, this smells like something putrid.
2454 if (flags & SCA_MIGRATE_DISABLE)
2455 SCHED_WARN_ON(!p->on_cpu);
2457 lockdep_assert_held(&p->pi_lock);
2459 queued = task_on_rq_queued(p);
2460 running = task_current(rq, p);
2464 * Because __kthread_bind() calls this on blocked tasks without
2467 lockdep_assert_rq_held(rq);
2468 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2471 put_prev_task(rq, p);
2473 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2476 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2478 set_next_task(rq, p);
2481 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2483 __do_set_cpus_allowed(p, new_mask, 0);
2486 int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2489 if (!src->user_cpus_ptr)
2492 dst->user_cpus_ptr = kmalloc_node(cpumask_size(), GFP_KERNEL, node);
2493 if (!dst->user_cpus_ptr)
2496 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2500 static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2502 struct cpumask *user_mask = NULL;
2504 swap(p->user_cpus_ptr, user_mask);
2509 void release_user_cpus_ptr(struct task_struct *p)
2511 kfree(clear_user_cpus_ptr(p));
2515 * This function is wildly self concurrent; here be dragons.
2518 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2519 * designated task is enqueued on an allowed CPU. If that task is currently
2520 * running, we have to kick it out using the CPU stopper.
2522 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2525 * Initial conditions: P0->cpus_mask = [0, 1]
2529 * migrate_disable();
2531 * set_cpus_allowed_ptr(P0, [1]);
2533 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2534 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2535 * This means we need the following scheme:
2539 * migrate_disable();
2541 * set_cpus_allowed_ptr(P0, [1]);
2545 * __set_cpus_allowed_ptr();
2546 * <wakes local stopper>
2547 * `--> <woken on migration completion>
2549 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2550 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2551 * task p are serialized by p->pi_lock, which we can leverage: the one that
2552 * should come into effect at the end of the Migrate-Disable region is the last
2553 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2554 * but we still need to properly signal those waiting tasks at the appropriate
2557 * This is implemented using struct set_affinity_pending. The first
2558 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2559 * setup an instance of that struct and install it on the targeted task_struct.
2560 * Any and all further callers will reuse that instance. Those then wait for
2561 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2562 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2565 * (1) In the cases covered above. There is one more where the completion is
2566 * signaled within affine_move_task() itself: when a subsequent affinity request
2567 * occurs after the stopper bailed out due to the targeted task still being
2568 * Migrate-Disable. Consider:
2570 * Initial conditions: P0->cpus_mask = [0, 1]
2574 * migrate_disable();
2576 * set_cpus_allowed_ptr(P0, [1]);
2579 * migration_cpu_stop()
2580 * is_migration_disabled()
2582 * set_cpus_allowed_ptr(P0, [0, 1]);
2583 * <signal completion>
2586 * Note that the above is safe vs a concurrent migrate_enable(), as any
2587 * pending affinity completion is preceded by an uninstallation of
2588 * p->migration_pending done with p->pi_lock held.
2590 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2591 int dest_cpu, unsigned int flags)
2593 struct set_affinity_pending my_pending = { }, *pending = NULL;
2594 bool stop_pending, complete = false;
2596 /* Can the task run on the task's current CPU? If so, we're done */
2597 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2598 struct task_struct *push_task = NULL;
2600 if ((flags & SCA_MIGRATE_ENABLE) &&
2601 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2602 rq->push_busy = true;
2603 push_task = get_task_struct(p);
2607 * If there are pending waiters, but no pending stop_work,
2608 * then complete now.
2610 pending = p->migration_pending;
2611 if (pending && !pending->stop_pending) {
2612 p->migration_pending = NULL;
2616 task_rq_unlock(rq, p, rf);
2619 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2624 complete_all(&pending->done);
2629 if (!(flags & SCA_MIGRATE_ENABLE)) {
2630 /* serialized by p->pi_lock */
2631 if (!p->migration_pending) {
2632 /* Install the request */
2633 refcount_set(&my_pending.refs, 1);
2634 init_completion(&my_pending.done);
2635 my_pending.arg = (struct migration_arg) {
2637 .dest_cpu = dest_cpu,
2638 .pending = &my_pending,
2641 p->migration_pending = &my_pending;
2643 pending = p->migration_pending;
2644 refcount_inc(&pending->refs);
2646 * Affinity has changed, but we've already installed a
2647 * pending. migration_cpu_stop() *must* see this, else
2648 * we risk a completion of the pending despite having a
2649 * task on a disallowed CPU.
2651 * Serialized by p->pi_lock, so this is safe.
2653 pending->arg.dest_cpu = dest_cpu;
2656 pending = p->migration_pending;
2658 * - !MIGRATE_ENABLE:
2659 * we'll have installed a pending if there wasn't one already.
2662 * we're here because the current CPU isn't matching anymore,
2663 * the only way that can happen is because of a concurrent
2664 * set_cpus_allowed_ptr() call, which should then still be
2665 * pending completion.
2667 * Either way, we really should have a @pending here.
2669 if (WARN_ON_ONCE(!pending)) {
2670 task_rq_unlock(rq, p, rf);
2674 if (task_running(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
2676 * MIGRATE_ENABLE gets here because 'p == current', but for
2677 * anything else we cannot do is_migration_disabled(), punt
2678 * and have the stopper function handle it all race-free.
2680 stop_pending = pending->stop_pending;
2682 pending->stop_pending = true;
2684 if (flags & SCA_MIGRATE_ENABLE)
2685 p->migration_flags &= ~MDF_PUSH;
2687 task_rq_unlock(rq, p, rf);
2689 if (!stop_pending) {
2690 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2691 &pending->arg, &pending->stop_work);
2694 if (flags & SCA_MIGRATE_ENABLE)
2698 if (!is_migration_disabled(p)) {
2699 if (task_on_rq_queued(p))
2700 rq = move_queued_task(rq, rf, p, dest_cpu);
2702 if (!pending->stop_pending) {
2703 p->migration_pending = NULL;
2707 task_rq_unlock(rq, p, rf);
2710 complete_all(&pending->done);
2713 wait_for_completion(&pending->done);
2715 if (refcount_dec_and_test(&pending->refs))
2716 wake_up_var(&pending->refs); /* No UaF, just an address */
2719 * Block the original owner of &pending until all subsequent callers
2720 * have seen the completion and decremented the refcount
2722 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2725 WARN_ON_ONCE(my_pending.stop_pending);
2731 * Called with both p->pi_lock and rq->lock held; drops both before returning.
2733 static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
2734 const struct cpumask *new_mask,
2737 struct rq_flags *rf)
2738 __releases(rq->lock)
2739 __releases(p->pi_lock)
2741 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
2742 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2743 bool kthread = p->flags & PF_KTHREAD;
2744 struct cpumask *user_mask = NULL;
2745 unsigned int dest_cpu;
2748 update_rq_clock(rq);
2750 if (kthread || is_migration_disabled(p)) {
2752 * Kernel threads are allowed on online && !active CPUs,
2753 * however, during cpu-hot-unplug, even these might get pushed
2754 * away if not KTHREAD_IS_PER_CPU.
2756 * Specifically, migration_disabled() tasks must not fail the
2757 * cpumask_any_and_distribute() pick below, esp. so on
2758 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2759 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2761 cpu_valid_mask = cpu_online_mask;
2764 if (!kthread && !cpumask_subset(new_mask, cpu_allowed_mask)) {
2770 * Must re-check here, to close a race against __kthread_bind(),
2771 * sched_setaffinity() is not guaranteed to observe the flag.
2773 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2778 if (!(flags & SCA_MIGRATE_ENABLE)) {
2779 if (cpumask_equal(&p->cpus_mask, new_mask))
2782 if (WARN_ON_ONCE(p == current &&
2783 is_migration_disabled(p) &&
2784 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2791 * Picking a ~random cpu helps in cases where we are changing affinity
2792 * for groups of tasks (ie. cpuset), so that load balancing is not
2793 * immediately required to distribute the tasks within their new mask.
2795 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2796 if (dest_cpu >= nr_cpu_ids) {
2801 __do_set_cpus_allowed(p, new_mask, flags);
2803 if (flags & SCA_USER)
2804 user_mask = clear_user_cpus_ptr(p);
2806 ret = affine_move_task(rq, p, rf, dest_cpu, flags);
2813 task_rq_unlock(rq, p, rf);
2819 * Change a given task's CPU affinity. Migrate the thread to a
2820 * proper CPU and schedule it away if the CPU it's executing on
2821 * is removed from the allowed bitmask.
2823 * NOTE: the caller must have a valid reference to the task, the
2824 * task must not exit() & deallocate itself prematurely. The
2825 * call is not atomic; no spinlocks may be held.
2827 static int __set_cpus_allowed_ptr(struct task_struct *p,
2828 const struct cpumask *new_mask, u32 flags)
2833 rq = task_rq_lock(p, &rf);
2834 return __set_cpus_allowed_ptr_locked(p, new_mask, flags, rq, &rf);
2837 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2839 return __set_cpus_allowed_ptr(p, new_mask, 0);
2841 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2844 * Change a given task's CPU affinity to the intersection of its current
2845 * affinity mask and @subset_mask, writing the resulting mask to @new_mask
2846 * and pointing @p->user_cpus_ptr to a copy of the old mask.
2847 * If the resulting mask is empty, leave the affinity unchanged and return
2850 static int restrict_cpus_allowed_ptr(struct task_struct *p,
2851 struct cpumask *new_mask,
2852 const struct cpumask *subset_mask)
2854 struct cpumask *user_mask = NULL;
2859 if (!p->user_cpus_ptr) {
2860 user_mask = kmalloc(cpumask_size(), GFP_KERNEL);
2865 rq = task_rq_lock(p, &rf);
2868 * Forcefully restricting the affinity of a deadline task is
2869 * likely to cause problems, so fail and noisily override the
2872 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
2877 if (!cpumask_and(new_mask, &p->cpus_mask, subset_mask)) {
2883 * We're about to butcher the task affinity, so keep track of what
2884 * the user asked for in case we're able to restore it later on.
2887 cpumask_copy(user_mask, p->cpus_ptr);
2888 p->user_cpus_ptr = user_mask;
2891 return __set_cpus_allowed_ptr_locked(p, new_mask, 0, rq, &rf);
2894 task_rq_unlock(rq, p, &rf);
2900 * Restrict the CPU affinity of task @p so that it is a subset of
2901 * task_cpu_possible_mask() and point @p->user_cpu_ptr to a copy of the
2902 * old affinity mask. If the resulting mask is empty, we warn and walk
2903 * up the cpuset hierarchy until we find a suitable mask.
2905 void force_compatible_cpus_allowed_ptr(struct task_struct *p)
2907 cpumask_var_t new_mask;
2908 const struct cpumask *override_mask = task_cpu_possible_mask(p);
2910 alloc_cpumask_var(&new_mask, GFP_KERNEL);
2913 * __migrate_task() can fail silently in the face of concurrent
2914 * offlining of the chosen destination CPU, so take the hotplug
2915 * lock to ensure that the migration succeeds.
2918 if (!cpumask_available(new_mask))
2921 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
2925 * We failed to find a valid subset of the affinity mask for the
2926 * task, so override it based on its cpuset hierarchy.
2928 cpuset_cpus_allowed(p, new_mask);
2929 override_mask = new_mask;
2932 if (printk_ratelimit()) {
2933 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
2934 task_pid_nr(p), p->comm,
2935 cpumask_pr_args(override_mask));
2938 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
2941 free_cpumask_var(new_mask);
2945 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask);
2948 * Restore the affinity of a task @p which was previously restricted by a
2949 * call to force_compatible_cpus_allowed_ptr(). This will clear (and free)
2950 * @p->user_cpus_ptr.
2952 * It is the caller's responsibility to serialise this with any calls to
2953 * force_compatible_cpus_allowed_ptr(@p).
2955 void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
2957 struct cpumask *user_mask = p->user_cpus_ptr;
2958 unsigned long flags;
2961 * Try to restore the old affinity mask. If this fails, then
2962 * we free the mask explicitly to avoid it being inherited across
2963 * a subsequent fork().
2965 if (!user_mask || !__sched_setaffinity(p, user_mask))
2968 raw_spin_lock_irqsave(&p->pi_lock, flags);
2969 user_mask = clear_user_cpus_ptr(p);
2970 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2975 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2977 #ifdef CONFIG_SCHED_DEBUG
2978 unsigned int state = READ_ONCE(p->__state);
2981 * We should never call set_task_cpu() on a blocked task,
2982 * ttwu() will sort out the placement.
2984 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
2987 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2988 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2989 * time relying on p->on_rq.
2991 WARN_ON_ONCE(state == TASK_RUNNING &&
2992 p->sched_class == &fair_sched_class &&
2993 (p->on_rq && !task_on_rq_migrating(p)));
2995 #ifdef CONFIG_LOCKDEP
2997 * The caller should hold either p->pi_lock or rq->lock, when changing
2998 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3000 * sched_move_task() holds both and thus holding either pins the cgroup,
3003 * Furthermore, all task_rq users should acquire both locks, see
3006 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
3007 lockdep_is_held(__rq_lockp(task_rq(p)))));
3010 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3012 WARN_ON_ONCE(!cpu_online(new_cpu));
3014 WARN_ON_ONCE(is_migration_disabled(p));
3017 trace_sched_migrate_task(p, new_cpu);
3019 if (task_cpu(p) != new_cpu) {
3020 if (p->sched_class->migrate_task_rq)
3021 p->sched_class->migrate_task_rq(p, new_cpu);
3022 p->se.nr_migrations++;
3024 perf_event_task_migrate(p);
3027 __set_task_cpu(p, new_cpu);
3030 #ifdef CONFIG_NUMA_BALANCING
3031 static void __migrate_swap_task(struct task_struct *p, int cpu)
3033 if (task_on_rq_queued(p)) {
3034 struct rq *src_rq, *dst_rq;
3035 struct rq_flags srf, drf;
3037 src_rq = task_rq(p);
3038 dst_rq = cpu_rq(cpu);
3040 rq_pin_lock(src_rq, &srf);
3041 rq_pin_lock(dst_rq, &drf);
3043 deactivate_task(src_rq, p, 0);
3044 set_task_cpu(p, cpu);
3045 activate_task(dst_rq, p, 0);
3046 check_preempt_curr(dst_rq, p, 0);
3048 rq_unpin_lock(dst_rq, &drf);
3049 rq_unpin_lock(src_rq, &srf);
3053 * Task isn't running anymore; make it appear like we migrated
3054 * it before it went to sleep. This means on wakeup we make the
3055 * previous CPU our target instead of where it really is.
3061 struct migration_swap_arg {
3062 struct task_struct *src_task, *dst_task;
3063 int src_cpu, dst_cpu;
3066 static int migrate_swap_stop(void *data)
3068 struct migration_swap_arg *arg = data;
3069 struct rq *src_rq, *dst_rq;
3072 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3075 src_rq = cpu_rq(arg->src_cpu);
3076 dst_rq = cpu_rq(arg->dst_cpu);
3078 double_raw_lock(&arg->src_task->pi_lock,
3079 &arg->dst_task->pi_lock);
3080 double_rq_lock(src_rq, dst_rq);
3082 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3085 if (task_cpu(arg->src_task) != arg->src_cpu)
3088 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
3091 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
3094 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3095 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3100 double_rq_unlock(src_rq, dst_rq);
3101 raw_spin_unlock(&arg->dst_task->pi_lock);
3102 raw_spin_unlock(&arg->src_task->pi_lock);
3108 * Cross migrate two tasks
3110 int migrate_swap(struct task_struct *cur, struct task_struct *p,
3111 int target_cpu, int curr_cpu)
3113 struct migration_swap_arg arg;
3116 arg = (struct migration_swap_arg){
3118 .src_cpu = curr_cpu,
3120 .dst_cpu = target_cpu,
3123 if (arg.src_cpu == arg.dst_cpu)
3127 * These three tests are all lockless; this is OK since all of them
3128 * will be re-checked with proper locks held further down the line.
3130 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3133 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
3136 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
3139 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
3140 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3145 #endif /* CONFIG_NUMA_BALANCING */
3148 * wait_task_inactive - wait for a thread to unschedule.
3150 * If @match_state is nonzero, it's the @p->state value just checked and
3151 * not expected to change. If it changes, i.e. @p might have woken up,
3152 * then return zero. When we succeed in waiting for @p to be off its CPU,
3153 * we return a positive number (its total switch count). If a second call
3154 * a short while later returns the same number, the caller can be sure that
3155 * @p has remained unscheduled the whole time.
3157 * The caller must ensure that the task *will* unschedule sometime soon,
3158 * else this function might spin for a *long* time. This function can't
3159 * be called with interrupts off, or it may introduce deadlock with
3160 * smp_call_function() if an IPI is sent by the same process we are
3161 * waiting to become inactive.
3163 unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
3165 int running, queued;
3172 * We do the initial early heuristics without holding
3173 * any task-queue locks at all. We'll only try to get
3174 * the runqueue lock when things look like they will
3180 * If the task is actively running on another CPU
3181 * still, just relax and busy-wait without holding
3184 * NOTE! Since we don't hold any locks, it's not
3185 * even sure that "rq" stays as the right runqueue!
3186 * But we don't care, since "task_running()" will
3187 * return false if the runqueue has changed and p
3188 * is actually now running somewhere else!
3190 while (task_running(rq, p)) {
3191 if (match_state && unlikely(READ_ONCE(p->__state) != match_state))
3197 * Ok, time to look more closely! We need the rq
3198 * lock now, to be *sure*. If we're wrong, we'll
3199 * just go back and repeat.
3201 rq = task_rq_lock(p, &rf);
3202 trace_sched_wait_task(p);
3203 running = task_running(rq, p);
3204 queued = task_on_rq_queued(p);
3206 if (!match_state || READ_ONCE(p->__state) == match_state)
3207 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3208 task_rq_unlock(rq, p, &rf);
3211 * If it changed from the expected state, bail out now.
3213 if (unlikely(!ncsw))
3217 * Was it really running after all now that we
3218 * checked with the proper locks actually held?
3220 * Oops. Go back and try again..
3222 if (unlikely(running)) {
3228 * It's not enough that it's not actively running,
3229 * it must be off the runqueue _entirely_, and not
3232 * So if it was still runnable (but just not actively
3233 * running right now), it's preempted, and we should
3234 * yield - it could be a while.
3236 if (unlikely(queued)) {
3237 ktime_t to = NSEC_PER_SEC / HZ;
3239 set_current_state(TASK_UNINTERRUPTIBLE);
3240 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
3245 * Ahh, all good. It wasn't running, and it wasn't
3246 * runnable, which means that it will never become
3247 * running in the future either. We're all done!
3256 * kick_process - kick a running thread to enter/exit the kernel
3257 * @p: the to-be-kicked thread
3259 * Cause a process which is running on another CPU to enter
3260 * kernel-mode, without any delay. (to get signals handled.)
3262 * NOTE: this function doesn't have to take the runqueue lock,
3263 * because all it wants to ensure is that the remote task enters
3264 * the kernel. If the IPI races and the task has been migrated
3265 * to another CPU then no harm is done and the purpose has been
3268 void kick_process(struct task_struct *p)
3274 if ((cpu != smp_processor_id()) && task_curr(p))
3275 smp_send_reschedule(cpu);
3278 EXPORT_SYMBOL_GPL(kick_process);
3281 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
3283 * A few notes on cpu_active vs cpu_online:
3285 * - cpu_active must be a subset of cpu_online
3287 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
3288 * see __set_cpus_allowed_ptr(). At this point the newly online
3289 * CPU isn't yet part of the sched domains, and balancing will not
3292 * - on CPU-down we clear cpu_active() to mask the sched domains and
3293 * avoid the load balancer to place new tasks on the to be removed
3294 * CPU. Existing tasks will remain running there and will be taken
3297 * This means that fallback selection must not select !active CPUs.
3298 * And can assume that any active CPU must be online. Conversely
3299 * select_task_rq() below may allow selection of !active CPUs in order
3300 * to satisfy the above rules.
3302 static int select_fallback_rq(int cpu, struct task_struct *p)
3304 int nid = cpu_to_node(cpu);
3305 const struct cpumask *nodemask = NULL;
3306 enum { cpuset, possible, fail } state = cpuset;
3310 * If the node that the CPU is on has been offlined, cpu_to_node()
3311 * will return -1. There is no CPU on the node, and we should
3312 * select the CPU on the other node.
3315 nodemask = cpumask_of_node(nid);
3317 /* Look for allowed, online CPU in same node. */
3318 for_each_cpu(dest_cpu, nodemask) {
3319 if (is_cpu_allowed(p, dest_cpu))
3325 /* Any allowed, online CPU? */
3326 for_each_cpu(dest_cpu, p->cpus_ptr) {
3327 if (!is_cpu_allowed(p, dest_cpu))
3333 /* No more Mr. Nice Guy. */
3336 if (cpuset_cpus_allowed_fallback(p)) {
3343 * XXX When called from select_task_rq() we only
3344 * hold p->pi_lock and again violate locking order.
3346 * More yuck to audit.
3348 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
3358 if (state != cpuset) {
3360 * Don't tell them about moving exiting tasks or
3361 * kernel threads (both mm NULL), since they never
3364 if (p->mm && printk_ratelimit()) {
3365 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
3366 task_pid_nr(p), p->comm, cpu);
3374 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
3377 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
3379 lockdep_assert_held(&p->pi_lock);
3381 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
3382 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
3384 cpu = cpumask_any(p->cpus_ptr);
3387 * In order not to call set_task_cpu() on a blocking task we need
3388 * to rely on ttwu() to place the task on a valid ->cpus_ptr
3391 * Since this is common to all placement strategies, this lives here.
3393 * [ this allows ->select_task() to simply return task_cpu(p) and
3394 * not worry about this generic constraint ]
3396 if (unlikely(!is_cpu_allowed(p, cpu)))
3397 cpu = select_fallback_rq(task_cpu(p), p);
3402 void sched_set_stop_task(int cpu, struct task_struct *stop)
3404 static struct lock_class_key stop_pi_lock;
3405 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3406 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3410 * Make it appear like a SCHED_FIFO task, its something
3411 * userspace knows about and won't get confused about.
3413 * Also, it will make PI more or less work without too
3414 * much confusion -- but then, stop work should not
3415 * rely on PI working anyway.
3417 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
3419 stop->sched_class = &stop_sched_class;
3422 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3423 * adjust the effective priority of a task. As a result,
3424 * rt_mutex_setprio() can trigger (RT) balancing operations,
3425 * which can then trigger wakeups of the stop thread to push
3426 * around the current task.
3428 * The stop task itself will never be part of the PI-chain, it
3429 * never blocks, therefore that ->pi_lock recursion is safe.
3430 * Tell lockdep about this by placing the stop->pi_lock in its
3433 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
3436 cpu_rq(cpu)->stop = stop;
3440 * Reset it back to a normal scheduling class so that
3441 * it can die in pieces.
3443 old_stop->sched_class = &rt_sched_class;
3447 #else /* CONFIG_SMP */
3449 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
3450 const struct cpumask *new_mask,
3453 return set_cpus_allowed_ptr(p, new_mask);
3456 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3458 static inline bool rq_has_pinned_tasks(struct rq *rq)
3463 #endif /* !CONFIG_SMP */
3466 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
3470 if (!schedstat_enabled())
3476 if (cpu == rq->cpu) {
3477 __schedstat_inc(rq->ttwu_local);
3478 __schedstat_inc(p->se.statistics.nr_wakeups_local);
3480 struct sched_domain *sd;
3482 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
3484 for_each_domain(rq->cpu, sd) {
3485 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
3486 __schedstat_inc(sd->ttwu_wake_remote);
3493 if (wake_flags & WF_MIGRATED)
3494 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
3495 #endif /* CONFIG_SMP */
3497 __schedstat_inc(rq->ttwu_count);
3498 __schedstat_inc(p->se.statistics.nr_wakeups);
3500 if (wake_flags & WF_SYNC)
3501 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
3505 * Mark the task runnable and perform wakeup-preemption.
3507 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
3508 struct rq_flags *rf)
3510 check_preempt_curr(rq, p, wake_flags);
3511 WRITE_ONCE(p->__state, TASK_RUNNING);
3512 trace_sched_wakeup(p);
3515 if (p->sched_class->task_woken) {
3517 * Our task @p is fully woken up and running; so it's safe to
3518 * drop the rq->lock, hereafter rq is only used for statistics.
3520 rq_unpin_lock(rq, rf);
3521 p->sched_class->task_woken(rq, p);
3522 rq_repin_lock(rq, rf);
3525 if (rq->idle_stamp) {
3526 u64 delta = rq_clock(rq) - rq->idle_stamp;
3527 u64 max = 2*rq->max_idle_balance_cost;
3529 update_avg(&rq->avg_idle, delta);
3531 if (rq->avg_idle > max)
3534 rq->wake_stamp = jiffies;
3535 rq->wake_avg_idle = rq->avg_idle / 2;
3543 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
3544 struct rq_flags *rf)
3546 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
3548 lockdep_assert_rq_held(rq);
3550 if (p->sched_contributes_to_load)
3551 rq->nr_uninterruptible--;
3554 if (wake_flags & WF_MIGRATED)
3555 en_flags |= ENQUEUE_MIGRATED;
3559 delayacct_blkio_end(p);
3560 atomic_dec(&task_rq(p)->nr_iowait);
3563 activate_task(rq, p, en_flags);
3564 ttwu_do_wakeup(rq, p, wake_flags, rf);
3568 * Consider @p being inside a wait loop:
3571 * set_current_state(TASK_UNINTERRUPTIBLE);
3578 * __set_current_state(TASK_RUNNING);
3580 * between set_current_state() and schedule(). In this case @p is still
3581 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3584 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3585 * then schedule() must still happen and p->state can be changed to
3586 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3587 * need to do a full wakeup with enqueue.
3589 * Returns: %true when the wakeup is done,
3592 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3598 rq = __task_rq_lock(p, &rf);
3599 if (task_on_rq_queued(p)) {
3600 /* check_preempt_curr() may use rq clock */
3601 update_rq_clock(rq);
3602 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3605 __task_rq_unlock(rq, &rf);
3611 void sched_ttwu_pending(void *arg)
3613 struct llist_node *llist = arg;
3614 struct rq *rq = this_rq();
3615 struct task_struct *p, *t;
3622 * rq::ttwu_pending racy indication of out-standing wakeups.
3623 * Races such that false-negatives are possible, since they
3624 * are shorter lived that false-positives would be.
3626 WRITE_ONCE(rq->ttwu_pending, 0);
3628 rq_lock_irqsave(rq, &rf);
3629 update_rq_clock(rq);
3631 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3632 if (WARN_ON_ONCE(p->on_cpu))
3633 smp_cond_load_acquire(&p->on_cpu, !VAL);
3635 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3636 set_task_cpu(p, cpu_of(rq));
3638 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3641 rq_unlock_irqrestore(rq, &rf);
3644 void send_call_function_single_ipi(int cpu)
3646 struct rq *rq = cpu_rq(cpu);
3648 if (!set_nr_if_polling(rq->idle))
3649 arch_send_call_function_single_ipi(cpu);
3651 trace_sched_wake_idle_without_ipi(cpu);
3655 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3656 * necessary. The wakee CPU on receipt of the IPI will queue the task
3657 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3658 * of the wakeup instead of the waker.
3660 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3662 struct rq *rq = cpu_rq(cpu);
3664 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3666 WRITE_ONCE(rq->ttwu_pending, 1);
3667 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3670 void wake_up_if_idle(int cpu)
3672 struct rq *rq = cpu_rq(cpu);
3677 if (!is_idle_task(rcu_dereference(rq->curr)))
3680 if (set_nr_if_polling(rq->idle)) {
3681 trace_sched_wake_idle_without_ipi(cpu);
3683 rq_lock_irqsave(rq, &rf);
3684 if (is_idle_task(rq->curr))
3685 smp_send_reschedule(cpu);
3686 /* Else CPU is not idle, do nothing here: */
3687 rq_unlock_irqrestore(rq, &rf);
3694 bool cpus_share_cache(int this_cpu, int that_cpu)
3696 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3699 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3702 * Do not complicate things with the async wake_list while the CPU is
3705 if (!cpu_active(cpu))
3709 * If the CPU does not share cache, then queue the task on the
3710 * remote rqs wakelist to avoid accessing remote data.
3712 if (!cpus_share_cache(smp_processor_id(), cpu))
3716 * If the task is descheduling and the only running task on the
3717 * CPU then use the wakelist to offload the task activation to
3718 * the soon-to-be-idle CPU as the current CPU is likely busy.
3719 * nr_running is checked to avoid unnecessary task stacking.
3721 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3727 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3729 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3730 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3733 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3734 __ttwu_queue_wakelist(p, cpu, wake_flags);
3741 #else /* !CONFIG_SMP */
3743 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3748 #endif /* CONFIG_SMP */
3750 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3752 struct rq *rq = cpu_rq(cpu);
3755 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3759 update_rq_clock(rq);
3760 ttwu_do_activate(rq, p, wake_flags, &rf);
3765 * Notes on Program-Order guarantees on SMP systems.
3769 * The basic program-order guarantee on SMP systems is that when a task [t]
3770 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3771 * execution on its new CPU [c1].
3773 * For migration (of runnable tasks) this is provided by the following means:
3775 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3776 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3777 * rq(c1)->lock (if not at the same time, then in that order).
3778 * C) LOCK of the rq(c1)->lock scheduling in task
3780 * Release/acquire chaining guarantees that B happens after A and C after B.
3781 * Note: the CPU doing B need not be c0 or c1
3790 * UNLOCK rq(0)->lock
3792 * LOCK rq(0)->lock // orders against CPU0
3794 * UNLOCK rq(0)->lock
3798 * UNLOCK rq(1)->lock
3800 * LOCK rq(1)->lock // orders against CPU2
3803 * UNLOCK rq(1)->lock
3806 * BLOCKING -- aka. SLEEP + WAKEUP
3808 * For blocking we (obviously) need to provide the same guarantee as for
3809 * migration. However the means are completely different as there is no lock
3810 * chain to provide order. Instead we do:
3812 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3813 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3817 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3819 * LOCK rq(0)->lock LOCK X->pi_lock
3822 * smp_store_release(X->on_cpu, 0);
3824 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3830 * X->state = RUNNING
3831 * UNLOCK rq(2)->lock
3833 * LOCK rq(2)->lock // orders against CPU1
3836 * UNLOCK rq(2)->lock
3839 * UNLOCK rq(0)->lock
3842 * However, for wakeups there is a second guarantee we must provide, namely we
3843 * must ensure that CONDITION=1 done by the caller can not be reordered with
3844 * accesses to the task state; see try_to_wake_up() and set_current_state().
3848 * try_to_wake_up - wake up a thread
3849 * @p: the thread to be awakened
3850 * @state: the mask of task states that can be woken
3851 * @wake_flags: wake modifier flags (WF_*)
3853 * Conceptually does:
3855 * If (@state & @p->state) @p->state = TASK_RUNNING.
3857 * If the task was not queued/runnable, also place it back on a runqueue.
3859 * This function is atomic against schedule() which would dequeue the task.
3861 * It issues a full memory barrier before accessing @p->state, see the comment
3862 * with set_current_state().
3864 * Uses p->pi_lock to serialize against concurrent wake-ups.
3866 * Relies on p->pi_lock stabilizing:
3869 * - p->sched_task_group
3870 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3872 * Tries really hard to only take one task_rq(p)->lock for performance.
3873 * Takes rq->lock in:
3874 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3875 * - ttwu_queue() -- new rq, for enqueue of the task;
3876 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3878 * As a consequence we race really badly with just about everything. See the
3879 * many memory barriers and their comments for details.
3881 * Return: %true if @p->state changes (an actual wakeup was done),
3885 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3887 unsigned long flags;
3888 int cpu, success = 0;
3893 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3894 * == smp_processor_id()'. Together this means we can special
3895 * case the whole 'p->on_rq && ttwu_runnable()' case below
3896 * without taking any locks.
3899 * - we rely on Program-Order guarantees for all the ordering,
3900 * - we're serialized against set_special_state() by virtue of
3901 * it disabling IRQs (this allows not taking ->pi_lock).
3903 if (!(READ_ONCE(p->__state) & state))
3907 trace_sched_waking(p);
3908 WRITE_ONCE(p->__state, TASK_RUNNING);
3909 trace_sched_wakeup(p);
3914 * If we are going to wake up a thread waiting for CONDITION we
3915 * need to ensure that CONDITION=1 done by the caller can not be
3916 * reordered with p->state check below. This pairs with smp_store_mb()
3917 * in set_current_state() that the waiting thread does.
3919 raw_spin_lock_irqsave(&p->pi_lock, flags);
3920 smp_mb__after_spinlock();
3921 if (!(READ_ONCE(p->__state) & state))
3924 trace_sched_waking(p);
3926 /* We're going to change ->state: */
3930 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3931 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3932 * in smp_cond_load_acquire() below.
3934 * sched_ttwu_pending() try_to_wake_up()
3935 * STORE p->on_rq = 1 LOAD p->state
3938 * __schedule() (switch to task 'p')
3939 * LOCK rq->lock smp_rmb();
3940 * smp_mb__after_spinlock();
3944 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3946 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3947 * __schedule(). See the comment for smp_mb__after_spinlock().
3949 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3952 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3957 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3958 * possible to, falsely, observe p->on_cpu == 0.
3960 * One must be running (->on_cpu == 1) in order to remove oneself
3961 * from the runqueue.
3963 * __schedule() (switch to task 'p') try_to_wake_up()
3964 * STORE p->on_cpu = 1 LOAD p->on_rq
3967 * __schedule() (put 'p' to sleep)
3968 * LOCK rq->lock smp_rmb();
3969 * smp_mb__after_spinlock();
3970 * STORE p->on_rq = 0 LOAD p->on_cpu
3972 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3973 * __schedule(). See the comment for smp_mb__after_spinlock().
3975 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3976 * schedule()'s deactivate_task() has 'happened' and p will no longer
3977 * care about it's own p->state. See the comment in __schedule().
3979 smp_acquire__after_ctrl_dep();
3982 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3983 * == 0), which means we need to do an enqueue, change p->state to
3984 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3985 * enqueue, such as ttwu_queue_wakelist().
3987 WRITE_ONCE(p->__state, TASK_WAKING);
3990 * If the owning (remote) CPU is still in the middle of schedule() with
3991 * this task as prev, considering queueing p on the remote CPUs wake_list
3992 * which potentially sends an IPI instead of spinning on p->on_cpu to
3993 * let the waker make forward progress. This is safe because IRQs are
3994 * disabled and the IPI will deliver after on_cpu is cleared.
3996 * Ensure we load task_cpu(p) after p->on_cpu:
3998 * set_task_cpu(p, cpu);
3999 * STORE p->cpu = @cpu
4000 * __schedule() (switch to task 'p')
4002 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4003 * STORE p->on_cpu = 1 LOAD p->cpu
4005 * to ensure we observe the correct CPU on which the task is currently
4008 if (smp_load_acquire(&p->on_cpu) &&
4009 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
4013 * If the owning (remote) CPU is still in the middle of schedule() with
4014 * this task as prev, wait until it's done referencing the task.
4016 * Pairs with the smp_store_release() in finish_task().
4018 * This ensures that tasks getting woken will be fully ordered against
4019 * their previous state and preserve Program Order.
4021 smp_cond_load_acquire(&p->on_cpu, !VAL);
4023 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
4024 if (task_cpu(p) != cpu) {
4026 delayacct_blkio_end(p);
4027 atomic_dec(&task_rq(p)->nr_iowait);
4030 wake_flags |= WF_MIGRATED;
4031 psi_ttwu_dequeue(p);
4032 set_task_cpu(p, cpu);
4036 #endif /* CONFIG_SMP */
4038 ttwu_queue(p, cpu, wake_flags);
4040 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4043 ttwu_stat(p, task_cpu(p), wake_flags);
4050 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
4051 * @p: Process for which the function is to be invoked, can be @current.
4052 * @func: Function to invoke.
4053 * @arg: Argument to function.
4055 * If the specified task can be quickly locked into a definite state
4056 * (either sleeping or on a given runqueue), arrange to keep it in that
4057 * state while invoking @func(@arg). This function can use ->on_rq and
4058 * task_curr() to work out what the state is, if required. Given that
4059 * @func can be invoked with a runqueue lock held, it had better be quite
4063 * @false if the task slipped out from under the locks.
4064 * @true if the task was locked onto a runqueue or is sleeping.
4065 * However, @func can override this by returning @false.
4067 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
4073 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4075 rq = __task_rq_lock(p, &rf);
4076 if (task_rq(p) == rq)
4080 switch (READ_ONCE(p->__state)) {
4085 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
4090 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
4095 * wake_up_process - Wake up a specific process
4096 * @p: The process to be woken up.
4098 * Attempt to wake up the nominated process and move it to the set of runnable
4101 * Return: 1 if the process was woken up, 0 if it was already running.
4103 * This function executes a full memory barrier before accessing the task state.
4105 int wake_up_process(struct task_struct *p)
4107 return try_to_wake_up(p, TASK_NORMAL, 0);
4109 EXPORT_SYMBOL(wake_up_process);
4111 int wake_up_state(struct task_struct *p, unsigned int state)
4113 return try_to_wake_up(p, state, 0);
4117 * Perform scheduler related setup for a newly forked process p.
4118 * p is forked by current.
4120 * __sched_fork() is basic setup used by init_idle() too:
4122 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
4127 p->se.exec_start = 0;
4128 p->se.sum_exec_runtime = 0;
4129 p->se.prev_sum_exec_runtime = 0;
4130 p->se.nr_migrations = 0;
4132 INIT_LIST_HEAD(&p->se.group_node);
4134 #ifdef CONFIG_FAIR_GROUP_SCHED
4135 p->se.cfs_rq = NULL;
4138 #ifdef CONFIG_SCHEDSTATS
4139 /* Even if schedstat is disabled, there should not be garbage */
4140 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
4143 RB_CLEAR_NODE(&p->dl.rb_node);
4144 init_dl_task_timer(&p->dl);
4145 init_dl_inactive_task_timer(&p->dl);
4146 __dl_clear_params(p);
4148 INIT_LIST_HEAD(&p->rt.run_list);
4150 p->rt.time_slice = sched_rr_timeslice;
4154 #ifdef CONFIG_PREEMPT_NOTIFIERS
4155 INIT_HLIST_HEAD(&p->preempt_notifiers);
4158 #ifdef CONFIG_COMPACTION
4159 p->capture_control = NULL;
4161 init_numa_balancing(clone_flags, p);
4163 p->wake_entry.u_flags = CSD_TYPE_TTWU;
4164 p->migration_pending = NULL;
4168 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4170 #ifdef CONFIG_NUMA_BALANCING
4172 void set_numabalancing_state(bool enabled)
4175 static_branch_enable(&sched_numa_balancing);
4177 static_branch_disable(&sched_numa_balancing);
4180 #ifdef CONFIG_PROC_SYSCTL
4181 int sysctl_numa_balancing(struct ctl_table *table, int write,
4182 void *buffer, size_t *lenp, loff_t *ppos)
4186 int state = static_branch_likely(&sched_numa_balancing);
4188 if (write && !capable(CAP_SYS_ADMIN))
4193 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4197 set_numabalancing_state(state);
4203 #ifdef CONFIG_SCHEDSTATS
4205 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
4207 static void set_schedstats(bool enabled)
4210 static_branch_enable(&sched_schedstats);
4212 static_branch_disable(&sched_schedstats);
4215 void force_schedstat_enabled(void)
4217 if (!schedstat_enabled()) {
4218 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4219 static_branch_enable(&sched_schedstats);
4223 static int __init setup_schedstats(char *str)
4229 if (!strcmp(str, "enable")) {
4230 set_schedstats(true);
4232 } else if (!strcmp(str, "disable")) {
4233 set_schedstats(false);
4238 pr_warn("Unable to parse schedstats=\n");
4242 __setup("schedstats=", setup_schedstats);
4244 #ifdef CONFIG_PROC_SYSCTL
4245 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
4246 size_t *lenp, loff_t *ppos)
4250 int state = static_branch_likely(&sched_schedstats);
4252 if (write && !capable(CAP_SYS_ADMIN))
4257 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4261 set_schedstats(state);
4264 #endif /* CONFIG_PROC_SYSCTL */
4265 #endif /* CONFIG_SCHEDSTATS */
4268 * fork()/clone()-time setup:
4270 int sched_fork(unsigned long clone_flags, struct task_struct *p)
4272 unsigned long flags;
4274 __sched_fork(clone_flags, p);
4276 * We mark the process as NEW here. This guarantees that
4277 * nobody will actually run it, and a signal or other external
4278 * event cannot wake it up and insert it on the runqueue either.
4280 p->__state = TASK_NEW;
4283 * Make sure we do not leak PI boosting priority to the child.
4285 p->prio = current->normal_prio;
4290 * Revert to default priority/policy on fork if requested.
4292 if (unlikely(p->sched_reset_on_fork)) {
4293 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4294 p->policy = SCHED_NORMAL;
4295 p->static_prio = NICE_TO_PRIO(0);
4297 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4298 p->static_prio = NICE_TO_PRIO(0);
4300 p->prio = p->normal_prio = p->static_prio;
4301 set_load_weight(p, false);
4304 * We don't need the reset flag anymore after the fork. It has
4305 * fulfilled its duty:
4307 p->sched_reset_on_fork = 0;
4310 if (dl_prio(p->prio))
4312 else if (rt_prio(p->prio))
4313 p->sched_class = &rt_sched_class;
4315 p->sched_class = &fair_sched_class;
4317 init_entity_runnable_average(&p->se);
4320 * The child is not yet in the pid-hash so no cgroup attach races,
4321 * and the cgroup is pinned to this child due to cgroup_fork()
4322 * is ran before sched_fork().
4324 * Silence PROVE_RCU.
4326 raw_spin_lock_irqsave(&p->pi_lock, flags);
4329 * We're setting the CPU for the first time, we don't migrate,
4330 * so use __set_task_cpu().
4332 __set_task_cpu(p, smp_processor_id());
4333 if (p->sched_class->task_fork)
4334 p->sched_class->task_fork(p);
4335 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4337 #ifdef CONFIG_SCHED_INFO
4338 if (likely(sched_info_on()))
4339 memset(&p->sched_info, 0, sizeof(p->sched_info));
4341 #if defined(CONFIG_SMP)
4344 init_task_preempt_count(p);
4346 plist_node_init(&p->pushable_tasks, MAX_PRIO);
4347 RB_CLEAR_NODE(&p->pushable_dl_tasks);
4352 void sched_post_fork(struct task_struct *p)
4354 uclamp_post_fork(p);
4357 unsigned long to_ratio(u64 period, u64 runtime)
4359 if (runtime == RUNTIME_INF)
4363 * Doing this here saves a lot of checks in all
4364 * the calling paths, and returning zero seems
4365 * safe for them anyway.
4370 return div64_u64(runtime << BW_SHIFT, period);
4374 * wake_up_new_task - wake up a newly created task for the first time.
4376 * This function will do some initial scheduler statistics housekeeping
4377 * that must be done for every newly created context, then puts the task
4378 * on the runqueue and wakes it.
4380 void wake_up_new_task(struct task_struct *p)
4385 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
4386 WRITE_ONCE(p->__state, TASK_RUNNING);
4389 * Fork balancing, do it here and not earlier because:
4390 * - cpus_ptr can change in the fork path
4391 * - any previously selected CPU might disappear through hotplug
4393 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4394 * as we're not fully set-up yet.
4396 p->recent_used_cpu = task_cpu(p);
4398 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
4400 rq = __task_rq_lock(p, &rf);
4401 update_rq_clock(rq);
4402 post_init_entity_util_avg(p);
4404 activate_task(rq, p, ENQUEUE_NOCLOCK);
4405 trace_sched_wakeup_new(p);
4406 check_preempt_curr(rq, p, WF_FORK);
4408 if (p->sched_class->task_woken) {
4410 * Nothing relies on rq->lock after this, so it's fine to
4413 rq_unpin_lock(rq, &rf);
4414 p->sched_class->task_woken(rq, p);
4415 rq_repin_lock(rq, &rf);
4418 task_rq_unlock(rq, p, &rf);
4421 #ifdef CONFIG_PREEMPT_NOTIFIERS
4423 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
4425 void preempt_notifier_inc(void)
4427 static_branch_inc(&preempt_notifier_key);
4429 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4431 void preempt_notifier_dec(void)
4433 static_branch_dec(&preempt_notifier_key);
4435 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4438 * preempt_notifier_register - tell me when current is being preempted & rescheduled
4439 * @notifier: notifier struct to register
4441 void preempt_notifier_register(struct preempt_notifier *notifier)
4443 if (!static_branch_unlikely(&preempt_notifier_key))
4444 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4446 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
4448 EXPORT_SYMBOL_GPL(preempt_notifier_register);
4451 * preempt_notifier_unregister - no longer interested in preemption notifications
4452 * @notifier: notifier struct to unregister
4454 * This is *not* safe to call from within a preemption notifier.
4456 void preempt_notifier_unregister(struct preempt_notifier *notifier)
4458 hlist_del(¬ifier->link);
4460 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4462 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
4464 struct preempt_notifier *notifier;
4466 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4467 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4470 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4472 if (static_branch_unlikely(&preempt_notifier_key))
4473 __fire_sched_in_preempt_notifiers(curr);
4477 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
4478 struct task_struct *next)
4480 struct preempt_notifier *notifier;
4482 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
4483 notifier->ops->sched_out(notifier, next);
4486 static __always_inline void
4487 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4488 struct task_struct *next)
4490 if (static_branch_unlikely(&preempt_notifier_key))
4491 __fire_sched_out_preempt_notifiers(curr, next);
4494 #else /* !CONFIG_PREEMPT_NOTIFIERS */
4496 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4501 fire_sched_out_preempt_notifiers(struct task_struct *curr,
4502 struct task_struct *next)
4506 #endif /* CONFIG_PREEMPT_NOTIFIERS */
4508 static inline void prepare_task(struct task_struct *next)
4512 * Claim the task as running, we do this before switching to it
4513 * such that any running task will have this set.
4515 * See the ttwu() WF_ON_CPU case and its ordering comment.
4517 WRITE_ONCE(next->on_cpu, 1);
4521 static inline void finish_task(struct task_struct *prev)
4525 * This must be the very last reference to @prev from this CPU. After
4526 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4527 * must ensure this doesn't happen until the switch is completely
4530 * In particular, the load of prev->state in finish_task_switch() must
4531 * happen before this.
4533 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4535 smp_store_release(&prev->on_cpu, 0);
4541 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
4543 void (*func)(struct rq *rq);
4544 struct callback_head *next;
4546 lockdep_assert_rq_held(rq);
4549 func = (void (*)(struct rq *))head->func;
4558 static void balance_push(struct rq *rq);
4560 struct callback_head balance_push_callback = {
4562 .func = (void (*)(struct callback_head *))balance_push,
4565 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4567 struct callback_head *head = rq->balance_callback;
4569 lockdep_assert_rq_held(rq);
4571 rq->balance_callback = NULL;
4576 static void __balance_callbacks(struct rq *rq)
4578 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4581 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4583 unsigned long flags;
4585 if (unlikely(head)) {
4586 raw_spin_rq_lock_irqsave(rq, flags);
4587 do_balance_callbacks(rq, head);
4588 raw_spin_rq_unlock_irqrestore(rq, flags);
4594 static inline void __balance_callbacks(struct rq *rq)
4598 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4603 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4610 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4613 * Since the runqueue lock will be released by the next
4614 * task (which is an invalid locking op but in the case
4615 * of the scheduler it's an obvious special-case), so we
4616 * do an early lockdep release here:
4618 rq_unpin_lock(rq, rf);
4619 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
4620 #ifdef CONFIG_DEBUG_SPINLOCK
4621 /* this is a valid case when another task releases the spinlock */
4622 rq_lockp(rq)->owner = next;
4626 static inline void finish_lock_switch(struct rq *rq)
4629 * If we are tracking spinlock dependencies then we have to
4630 * fix up the runqueue lock - which gets 'carried over' from
4631 * prev into current:
4633 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
4634 __balance_callbacks(rq);
4635 raw_spin_rq_unlock_irq(rq);
4639 * NOP if the arch has not defined these:
4642 #ifndef prepare_arch_switch
4643 # define prepare_arch_switch(next) do { } while (0)
4646 #ifndef finish_arch_post_lock_switch
4647 # define finish_arch_post_lock_switch() do { } while (0)
4650 static inline void kmap_local_sched_out(void)
4652 #ifdef CONFIG_KMAP_LOCAL
4653 if (unlikely(current->kmap_ctrl.idx))
4654 __kmap_local_sched_out();
4658 static inline void kmap_local_sched_in(void)
4660 #ifdef CONFIG_KMAP_LOCAL
4661 if (unlikely(current->kmap_ctrl.idx))
4662 __kmap_local_sched_in();
4667 * prepare_task_switch - prepare to switch tasks
4668 * @rq: the runqueue preparing to switch
4669 * @prev: the current task that is being switched out
4670 * @next: the task we are going to switch to.
4672 * This is called with the rq lock held and interrupts off. It must
4673 * be paired with a subsequent finish_task_switch after the context
4676 * prepare_task_switch sets up locking and calls architecture specific
4680 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4681 struct task_struct *next)
4683 kcov_prepare_switch(prev);
4684 sched_info_switch(rq, prev, next);
4685 perf_event_task_sched_out(prev, next);
4687 fire_sched_out_preempt_notifiers(prev, next);
4688 kmap_local_sched_out();
4690 prepare_arch_switch(next);
4694 * finish_task_switch - clean up after a task-switch
4695 * @prev: the thread we just switched away from.
4697 * finish_task_switch must be called after the context switch, paired
4698 * with a prepare_task_switch call before the context switch.
4699 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4700 * and do any other architecture-specific cleanup actions.
4702 * Note that we may have delayed dropping an mm in context_switch(). If
4703 * so, we finish that here outside of the runqueue lock. (Doing it
4704 * with the lock held can cause deadlocks; see schedule() for
4707 * The context switch have flipped the stack from under us and restored the
4708 * local variables which were saved when this task called schedule() in the
4709 * past. prev == current is still correct but we need to recalculate this_rq
4710 * because prev may have moved to another CPU.
4712 static struct rq *finish_task_switch(struct task_struct *prev)
4713 __releases(rq->lock)
4715 struct rq *rq = this_rq();
4716 struct mm_struct *mm = rq->prev_mm;
4720 * The previous task will have left us with a preempt_count of 2
4721 * because it left us after:
4724 * preempt_disable(); // 1
4726 * raw_spin_lock_irq(&rq->lock) // 2
4728 * Also, see FORK_PREEMPT_COUNT.
4730 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4731 "corrupted preempt_count: %s/%d/0x%x\n",
4732 current->comm, current->pid, preempt_count()))
4733 preempt_count_set(FORK_PREEMPT_COUNT);
4738 * A task struct has one reference for the use as "current".
4739 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4740 * schedule one last time. The schedule call will never return, and
4741 * the scheduled task must drop that reference.
4743 * We must observe prev->state before clearing prev->on_cpu (in
4744 * finish_task), otherwise a concurrent wakeup can get prev
4745 * running on another CPU and we could rave with its RUNNING -> DEAD
4746 * transition, resulting in a double drop.
4748 prev_state = READ_ONCE(prev->__state);
4749 vtime_task_switch(prev);
4750 perf_event_task_sched_in(prev, current);
4752 tick_nohz_task_switch();
4753 finish_lock_switch(rq);
4754 finish_arch_post_lock_switch();
4755 kcov_finish_switch(current);
4757 * kmap_local_sched_out() is invoked with rq::lock held and
4758 * interrupts disabled. There is no requirement for that, but the
4759 * sched out code does not have an interrupt enabled section.
4760 * Restoring the maps on sched in does not require interrupts being
4763 kmap_local_sched_in();
4765 fire_sched_in_preempt_notifiers(current);
4767 * When switching through a kernel thread, the loop in
4768 * membarrier_{private,global}_expedited() may have observed that
4769 * kernel thread and not issued an IPI. It is therefore possible to
4770 * schedule between user->kernel->user threads without passing though
4771 * switch_mm(). Membarrier requires a barrier after storing to
4772 * rq->curr, before returning to userspace, so provide them here:
4774 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4775 * provided by mmdrop(),
4776 * - a sync_core for SYNC_CORE.
4779 membarrier_mm_sync_core_before_usermode(mm);
4782 if (unlikely(prev_state == TASK_DEAD)) {
4783 if (prev->sched_class->task_dead)
4784 prev->sched_class->task_dead(prev);
4787 * Remove function-return probe instances associated with this
4788 * task and put them back on the free list.
4790 kprobe_flush_task(prev);
4792 /* Task is done with its stack. */
4793 put_task_stack(prev);
4795 put_task_struct_rcu_user(prev);
4802 * schedule_tail - first thing a freshly forked thread must call.
4803 * @prev: the thread we just switched away from.
4805 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4806 __releases(rq->lock)
4809 * New tasks start with FORK_PREEMPT_COUNT, see there and
4810 * finish_task_switch() for details.
4812 * finish_task_switch() will drop rq->lock() and lower preempt_count
4813 * and the preempt_enable() will end up enabling preemption (on
4814 * PREEMPT_COUNT kernels).
4817 finish_task_switch(prev);
4820 if (current->set_child_tid)
4821 put_user(task_pid_vnr(current), current->set_child_tid);
4823 calculate_sigpending();
4827 * context_switch - switch to the new MM and the new thread's register state.
4829 static __always_inline struct rq *
4830 context_switch(struct rq *rq, struct task_struct *prev,
4831 struct task_struct *next, struct rq_flags *rf)
4833 prepare_task_switch(rq, prev, next);
4836 * For paravirt, this is coupled with an exit in switch_to to
4837 * combine the page table reload and the switch backend into
4840 arch_start_context_switch(prev);
4843 * kernel -> kernel lazy + transfer active
4844 * user -> kernel lazy + mmgrab() active
4846 * kernel -> user switch + mmdrop() active
4847 * user -> user switch
4849 if (!next->mm) { // to kernel
4850 enter_lazy_tlb(prev->active_mm, next);
4852 next->active_mm = prev->active_mm;
4853 if (prev->mm) // from user
4854 mmgrab(prev->active_mm);
4856 prev->active_mm = NULL;
4858 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4860 * sys_membarrier() requires an smp_mb() between setting
4861 * rq->curr / membarrier_switch_mm() and returning to userspace.
4863 * The below provides this either through switch_mm(), or in
4864 * case 'prev->active_mm == next->mm' through
4865 * finish_task_switch()'s mmdrop().
4867 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4869 if (!prev->mm) { // from kernel
4870 /* will mmdrop() in finish_task_switch(). */
4871 rq->prev_mm = prev->active_mm;
4872 prev->active_mm = NULL;
4876 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4878 prepare_lock_switch(rq, next, rf);
4880 /* Here we just switch the register state and the stack. */
4881 switch_to(prev, next, prev);
4884 return finish_task_switch(prev);
4888 * nr_running and nr_context_switches:
4890 * externally visible scheduler statistics: current number of runnable
4891 * threads, total number of context switches performed since bootup.
4893 unsigned int nr_running(void)
4895 unsigned int i, sum = 0;
4897 for_each_online_cpu(i)
4898 sum += cpu_rq(i)->nr_running;
4904 * Check if only the current task is running on the CPU.
4906 * Caution: this function does not check that the caller has disabled
4907 * preemption, thus the result might have a time-of-check-to-time-of-use
4908 * race. The caller is responsible to use it correctly, for example:
4910 * - from a non-preemptible section (of course)
4912 * - from a thread that is bound to a single CPU
4914 * - in a loop with very short iterations (e.g. a polling loop)
4916 bool single_task_running(void)
4918 return raw_rq()->nr_running == 1;
4920 EXPORT_SYMBOL(single_task_running);
4922 unsigned long long nr_context_switches(void)
4925 unsigned long long sum = 0;
4927 for_each_possible_cpu(i)
4928 sum += cpu_rq(i)->nr_switches;
4934 * Consumers of these two interfaces, like for example the cpuidle menu
4935 * governor, are using nonsensical data. Preferring shallow idle state selection
4936 * for a CPU that has IO-wait which might not even end up running the task when
4937 * it does become runnable.
4940 unsigned int nr_iowait_cpu(int cpu)
4942 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4946 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4948 * The idea behind IO-wait account is to account the idle time that we could
4949 * have spend running if it were not for IO. That is, if we were to improve the
4950 * storage performance, we'd have a proportional reduction in IO-wait time.
4952 * This all works nicely on UP, where, when a task blocks on IO, we account
4953 * idle time as IO-wait, because if the storage were faster, it could've been
4954 * running and we'd not be idle.
4956 * This has been extended to SMP, by doing the same for each CPU. This however
4959 * Imagine for instance the case where two tasks block on one CPU, only the one
4960 * CPU will have IO-wait accounted, while the other has regular idle. Even
4961 * though, if the storage were faster, both could've ran at the same time,
4962 * utilising both CPUs.
4964 * This means, that when looking globally, the current IO-wait accounting on
4965 * SMP is a lower bound, by reason of under accounting.
4967 * Worse, since the numbers are provided per CPU, they are sometimes
4968 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4969 * associated with any one particular CPU, it can wake to another CPU than it
4970 * blocked on. This means the per CPU IO-wait number is meaningless.
4972 * Task CPU affinities can make all that even more 'interesting'.
4975 unsigned int nr_iowait(void)
4977 unsigned int i, sum = 0;
4979 for_each_possible_cpu(i)
4980 sum += nr_iowait_cpu(i);
4988 * sched_exec - execve() is a valuable balancing opportunity, because at
4989 * this point the task has the smallest effective memory and cache footprint.
4991 void sched_exec(void)
4993 struct task_struct *p = current;
4994 unsigned long flags;
4997 raw_spin_lock_irqsave(&p->pi_lock, flags);
4998 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4999 if (dest_cpu == smp_processor_id())
5002 if (likely(cpu_active(dest_cpu))) {
5003 struct migration_arg arg = { p, dest_cpu };
5005 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5006 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
5010 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5015 DEFINE_PER_CPU(struct kernel_stat, kstat);
5016 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
5018 EXPORT_PER_CPU_SYMBOL(kstat);
5019 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
5022 * The function fair_sched_class.update_curr accesses the struct curr
5023 * and its field curr->exec_start; when called from task_sched_runtime(),
5024 * we observe a high rate of cache misses in practice.
5025 * Prefetching this data results in improved performance.
5027 static inline void prefetch_curr_exec_start(struct task_struct *p)
5029 #ifdef CONFIG_FAIR_GROUP_SCHED
5030 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5032 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5035 prefetch(&curr->exec_start);
5039 * Return accounted runtime for the task.
5040 * In case the task is currently running, return the runtime plus current's
5041 * pending runtime that have not been accounted yet.
5043 unsigned long long task_sched_runtime(struct task_struct *p)
5049 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5051 * 64-bit doesn't need locks to atomically read a 64-bit value.
5052 * So we have a optimization chance when the task's delta_exec is 0.
5053 * Reading ->on_cpu is racy, but this is ok.
5055 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5056 * If we race with it entering CPU, unaccounted time is 0. This is
5057 * indistinguishable from the read occurring a few cycles earlier.
5058 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5059 * been accounted, so we're correct here as well.
5061 if (!p->on_cpu || !task_on_rq_queued(p))
5062 return p->se.sum_exec_runtime;
5065 rq = task_rq_lock(p, &rf);
5067 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5068 * project cycles that may never be accounted to this
5069 * thread, breaking clock_gettime().
5071 if (task_current(rq, p) && task_on_rq_queued(p)) {
5072 prefetch_curr_exec_start(p);
5073 update_rq_clock(rq);
5074 p->sched_class->update_curr(rq);
5076 ns = p->se.sum_exec_runtime;
5077 task_rq_unlock(rq, p, &rf);
5082 #ifdef CONFIG_SCHED_DEBUG
5083 static u64 cpu_resched_latency(struct rq *rq)
5085 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5086 u64 resched_latency, now = rq_clock(rq);
5087 static bool warned_once;
5089 if (sysctl_resched_latency_warn_once && warned_once)
5092 if (!need_resched() || !latency_warn_ms)
5095 if (system_state == SYSTEM_BOOTING)
5098 if (!rq->last_seen_need_resched_ns) {
5099 rq->last_seen_need_resched_ns = now;
5100 rq->ticks_without_resched = 0;
5104 rq->ticks_without_resched++;
5105 resched_latency = now - rq->last_seen_need_resched_ns;
5106 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5111 return resched_latency;
5114 static int __init setup_resched_latency_warn_ms(char *str)
5118 if ((kstrtol(str, 0, &val))) {
5119 pr_warn("Unable to set resched_latency_warn_ms\n");
5123 sysctl_resched_latency_warn_ms = val;
5126 __setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5128 static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5129 #endif /* CONFIG_SCHED_DEBUG */
5132 * This function gets called by the timer code, with HZ frequency.
5133 * We call it with interrupts disabled.
5135 void scheduler_tick(void)
5137 int cpu = smp_processor_id();
5138 struct rq *rq = cpu_rq(cpu);
5139 struct task_struct *curr = rq->curr;
5141 unsigned long thermal_pressure;
5142 u64 resched_latency;
5144 arch_scale_freq_tick();
5149 update_rq_clock(rq);
5150 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
5151 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
5152 curr->sched_class->task_tick(rq, curr, 0);
5153 if (sched_feat(LATENCY_WARN))
5154 resched_latency = cpu_resched_latency(rq);
5155 calc_global_load_tick(rq);
5159 if (sched_feat(LATENCY_WARN) && resched_latency)
5160 resched_latency_warn(cpu, resched_latency);
5162 perf_event_task_tick();
5165 rq->idle_balance = idle_cpu(cpu);
5166 trigger_load_balance(rq);
5170 #ifdef CONFIG_NO_HZ_FULL
5175 struct delayed_work work;
5177 /* Values for ->state, see diagram below. */
5178 #define TICK_SCHED_REMOTE_OFFLINE 0
5179 #define TICK_SCHED_REMOTE_OFFLINING 1
5180 #define TICK_SCHED_REMOTE_RUNNING 2
5183 * State diagram for ->state:
5186 * TICK_SCHED_REMOTE_OFFLINE
5189 * | | sched_tick_remote()
5192 * +--TICK_SCHED_REMOTE_OFFLINING
5195 * sched_tick_start() | | sched_tick_stop()
5198 * TICK_SCHED_REMOTE_RUNNING
5201 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5202 * and sched_tick_start() are happy to leave the state in RUNNING.
5205 static struct tick_work __percpu *tick_work_cpu;
5207 static void sched_tick_remote(struct work_struct *work)
5209 struct delayed_work *dwork = to_delayed_work(work);
5210 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5211 int cpu = twork->cpu;
5212 struct rq *rq = cpu_rq(cpu);
5213 struct task_struct *curr;
5219 * Handle the tick only if it appears the remote CPU is running in full
5220 * dynticks mode. The check is racy by nature, but missing a tick or
5221 * having one too much is no big deal because the scheduler tick updates
5222 * statistics and checks timeslices in a time-independent way, regardless
5223 * of when exactly it is running.
5225 if (!tick_nohz_tick_stopped_cpu(cpu))
5228 rq_lock_irq(rq, &rf);
5230 if (cpu_is_offline(cpu))
5233 update_rq_clock(rq);
5235 if (!is_idle_task(curr)) {
5237 * Make sure the next tick runs within a reasonable
5240 delta = rq_clock_task(rq) - curr->se.exec_start;
5241 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5243 curr->sched_class->task_tick(rq, curr, 0);
5245 calc_load_nohz_remote(rq);
5247 rq_unlock_irq(rq, &rf);
5251 * Run the remote tick once per second (1Hz). This arbitrary
5252 * frequency is large enough to avoid overload but short enough
5253 * to keep scheduler internal stats reasonably up to date. But
5254 * first update state to reflect hotplug activity if required.
5256 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5257 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5258 if (os == TICK_SCHED_REMOTE_RUNNING)
5259 queue_delayed_work(system_unbound_wq, dwork, HZ);
5262 static void sched_tick_start(int cpu)
5265 struct tick_work *twork;
5267 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5270 WARN_ON_ONCE(!tick_work_cpu);
5272 twork = per_cpu_ptr(tick_work_cpu, cpu);
5273 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5274 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5275 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5277 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5278 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5282 #ifdef CONFIG_HOTPLUG_CPU
5283 static void sched_tick_stop(int cpu)
5285 struct tick_work *twork;
5288 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
5291 WARN_ON_ONCE(!tick_work_cpu);
5293 twork = per_cpu_ptr(tick_work_cpu, cpu);
5294 /* There cannot be competing actions, but don't rely on stop-machine. */
5295 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5296 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5297 /* Don't cancel, as this would mess up the state machine. */
5299 #endif /* CONFIG_HOTPLUG_CPU */
5301 int __init sched_tick_offload_init(void)
5303 tick_work_cpu = alloc_percpu(struct tick_work);
5304 BUG_ON(!tick_work_cpu);
5308 #else /* !CONFIG_NO_HZ_FULL */
5309 static inline void sched_tick_start(int cpu) { }
5310 static inline void sched_tick_stop(int cpu) { }
5313 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
5314 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
5316 * If the value passed in is equal to the current preempt count
5317 * then we just disabled preemption. Start timing the latency.
5319 static inline void preempt_latency_start(int val)
5321 if (preempt_count() == val) {
5322 unsigned long ip = get_lock_parent_ip();
5323 #ifdef CONFIG_DEBUG_PREEMPT
5324 current->preempt_disable_ip = ip;
5326 trace_preempt_off(CALLER_ADDR0, ip);
5330 void preempt_count_add(int val)
5332 #ifdef CONFIG_DEBUG_PREEMPT
5336 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5339 __preempt_count_add(val);
5340 #ifdef CONFIG_DEBUG_PREEMPT
5342 * Spinlock count overflowing soon?
5344 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5347 preempt_latency_start(val);
5349 EXPORT_SYMBOL(preempt_count_add);
5350 NOKPROBE_SYMBOL(preempt_count_add);
5353 * If the value passed in equals to the current preempt count
5354 * then we just enabled preemption. Stop timing the latency.
5356 static inline void preempt_latency_stop(int val)
5358 if (preempt_count() == val)
5359 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5362 void preempt_count_sub(int val)
5364 #ifdef CONFIG_DEBUG_PREEMPT
5368 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5371 * Is the spinlock portion underflowing?
5373 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5374 !(preempt_count() & PREEMPT_MASK)))
5378 preempt_latency_stop(val);
5379 __preempt_count_sub(val);
5381 EXPORT_SYMBOL(preempt_count_sub);
5382 NOKPROBE_SYMBOL(preempt_count_sub);
5385 static inline void preempt_latency_start(int val) { }
5386 static inline void preempt_latency_stop(int val) { }
5389 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5391 #ifdef CONFIG_DEBUG_PREEMPT
5392 return p->preempt_disable_ip;
5399 * Print scheduling while atomic bug:
5401 static noinline void __schedule_bug(struct task_struct *prev)
5403 /* Save this before calling printk(), since that will clobber it */
5404 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5406 if (oops_in_progress)
5409 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5410 prev->comm, prev->pid, preempt_count());
5412 debug_show_held_locks(prev);
5414 if (irqs_disabled())
5415 print_irqtrace_events(prev);
5416 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5417 && in_atomic_preempt_off()) {
5418 pr_err("Preemption disabled at:");
5419 print_ip_sym(KERN_ERR, preempt_disable_ip);
5422 panic("scheduling while atomic\n");
5425 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5429 * Various schedule()-time debugging checks and statistics:
5431 static inline void schedule_debug(struct task_struct *prev, bool preempt)
5433 #ifdef CONFIG_SCHED_STACK_END_CHECK
5434 if (task_stack_end_corrupted(prev))
5435 panic("corrupted stack end detected inside scheduler\n");
5437 if (task_scs_end_corrupted(prev))
5438 panic("corrupted shadow stack detected inside scheduler\n");
5441 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5442 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
5443 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5444 prev->comm, prev->pid, prev->non_block_count);
5446 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5450 if (unlikely(in_atomic_preempt_off())) {
5451 __schedule_bug(prev);
5452 preempt_count_set(PREEMPT_DISABLED);
5455 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
5457 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5459 schedstat_inc(this_rq()->sched_count);
5462 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5463 struct rq_flags *rf)
5466 const struct sched_class *class;
5468 * We must do the balancing pass before put_prev_task(), such
5469 * that when we release the rq->lock the task is in the same
5470 * state as before we took rq->lock.
5472 * We can terminate the balance pass as soon as we know there is
5473 * a runnable task of @class priority or higher.
5475 for_class_range(class, prev->sched_class, &idle_sched_class) {
5476 if (class->balance(rq, prev, rf))
5481 put_prev_task(rq, prev);
5485 * Pick up the highest-prio task:
5487 static inline struct task_struct *
5488 __pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5490 const struct sched_class *class;
5491 struct task_struct *p;
5494 * Optimization: we know that if all tasks are in the fair class we can
5495 * call that function directly, but only if the @prev task wasn't of a
5496 * higher scheduling class, because otherwise those lose the
5497 * opportunity to pull in more work from other CPUs.
5499 if (likely(prev->sched_class <= &fair_sched_class &&
5500 rq->nr_running == rq->cfs.h_nr_running)) {
5502 p = pick_next_task_fair(rq, prev, rf);
5503 if (unlikely(p == RETRY_TASK))
5506 /* Assume the next prioritized class is idle_sched_class */
5508 put_prev_task(rq, prev);
5509 p = pick_next_task_idle(rq);
5516 put_prev_task_balance(rq, prev, rf);
5518 for_each_class(class) {
5519 p = class->pick_next_task(rq);
5524 /* The idle class should always have a runnable task: */
5528 #ifdef CONFIG_SCHED_CORE
5529 static inline bool is_task_rq_idle(struct task_struct *t)
5531 return (task_rq(t)->idle == t);
5534 static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5536 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5539 static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5541 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5544 return a->core_cookie == b->core_cookie;
5547 // XXX fairness/fwd progress conditions
5550 * - NULL if there is no runnable task for this class.
5551 * - the highest priority task for this runqueue if it matches
5552 * rq->core->core_cookie or its priority is greater than max.
5553 * - Else returns idle_task.
5555 static struct task_struct *
5556 pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max, bool in_fi)
5558 struct task_struct *class_pick, *cookie_pick;
5559 unsigned long cookie = rq->core->core_cookie;
5561 class_pick = class->pick_task(rq);
5567 * If class_pick is tagged, return it only if it has
5568 * higher priority than max.
5570 if (max && class_pick->core_cookie &&
5571 prio_less(class_pick, max, in_fi))
5572 return idle_sched_class.pick_task(rq);
5578 * If class_pick is idle or matches cookie, return early.
5580 if (cookie_equals(class_pick, cookie))
5583 cookie_pick = sched_core_find(rq, cookie);
5586 * If class > max && class > cookie, it is the highest priority task on
5587 * the core (so far) and it must be selected, otherwise we must go with
5588 * the cookie pick in order to satisfy the constraint.
5590 if (prio_less(cookie_pick, class_pick, in_fi) &&
5591 (!max || prio_less(max, class_pick, in_fi)))
5597 extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5599 static struct task_struct *
5600 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5602 struct task_struct *next, *max = NULL;
5603 const struct sched_class *class;
5604 const struct cpumask *smt_mask;
5605 bool fi_before = false;
5606 int i, j, cpu, occ = 0;
5609 if (!sched_core_enabled(rq))
5610 return __pick_next_task(rq, prev, rf);
5614 /* Stopper task is switching into idle, no need core-wide selection. */
5615 if (cpu_is_offline(cpu)) {
5617 * Reset core_pick so that we don't enter the fastpath when
5618 * coming online. core_pick would already be migrated to
5619 * another cpu during offline.
5621 rq->core_pick = NULL;
5622 return __pick_next_task(rq, prev, rf);
5626 * If there were no {en,de}queues since we picked (IOW, the task
5627 * pointers are all still valid), and we haven't scheduled the last
5628 * pick yet, do so now.
5630 * rq->core_pick can be NULL if no selection was made for a CPU because
5631 * it was either offline or went offline during a sibling's core-wide
5632 * selection. In this case, do a core-wide selection.
5634 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
5635 rq->core->core_pick_seq != rq->core_sched_seq &&
5637 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
5639 next = rq->core_pick;
5641 put_prev_task(rq, prev);
5642 set_next_task(rq, next);
5645 rq->core_pick = NULL;
5649 put_prev_task_balance(rq, prev, rf);
5651 smt_mask = cpu_smt_mask(cpu);
5652 need_sync = !!rq->core->core_cookie;
5655 rq->core->core_cookie = 0UL;
5656 if (rq->core->core_forceidle) {
5659 rq->core->core_forceidle = false;
5663 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
5665 * @task_seq guards the task state ({en,de}queues)
5666 * @pick_seq is the @task_seq we did a selection on
5667 * @sched_seq is the @pick_seq we scheduled
5669 * However, preemptions can cause multiple picks on the same task set.
5670 * 'Fix' this by also increasing @task_seq for every pick.
5672 rq->core->core_task_seq++;
5675 * Optimize for common case where this CPU has no cookies
5676 * and there are no cookied tasks running on siblings.
5679 for_each_class(class) {
5680 next = class->pick_task(rq);
5685 if (!next->core_cookie) {
5686 rq->core_pick = NULL;
5688 * For robustness, update the min_vruntime_fi for
5689 * unconstrained picks as well.
5691 WARN_ON_ONCE(fi_before);
5692 task_vruntime_update(rq, next, false);
5697 for_each_cpu(i, smt_mask) {
5698 struct rq *rq_i = cpu_rq(i);
5700 rq_i->core_pick = NULL;
5703 update_rq_clock(rq_i);
5707 * Try and select tasks for each sibling in descending sched_class
5710 for_each_class(class) {
5712 for_each_cpu_wrap(i, smt_mask, cpu) {
5713 struct rq *rq_i = cpu_rq(i);
5714 struct task_struct *p;
5716 if (rq_i->core_pick)
5720 * If this sibling doesn't yet have a suitable task to
5721 * run; ask for the most eligible task, given the
5722 * highest priority task already selected for this
5725 p = pick_task(rq_i, class, max, fi_before);
5729 if (!is_task_rq_idle(p))
5732 rq_i->core_pick = p;
5733 if (rq_i->idle == p && rq_i->nr_running) {
5734 rq->core->core_forceidle = true;
5736 rq->core->core_forceidle_seq++;
5740 * If this new candidate is of higher priority than the
5741 * previous; and they're incompatible; we need to wipe
5742 * the slate and start over. pick_task makes sure that
5743 * p's priority is more than max if it doesn't match
5746 * NOTE: this is a linear max-filter and is thus bounded
5747 * in execution time.
5749 if (!max || !cookie_match(max, p)) {
5750 struct task_struct *old_max = max;
5752 rq->core->core_cookie = p->core_cookie;
5756 rq->core->core_forceidle = false;
5757 for_each_cpu(j, smt_mask) {
5761 cpu_rq(j)->core_pick = NULL;
5770 rq->core->core_pick_seq = rq->core->core_task_seq;
5771 next = rq->core_pick;
5772 rq->core_sched_seq = rq->core->core_pick_seq;
5774 /* Something should have been selected for current CPU */
5775 WARN_ON_ONCE(!next);
5778 * Reschedule siblings
5780 * NOTE: L1TF -- at this point we're no longer running the old task and
5781 * sending an IPI (below) ensures the sibling will no longer be running
5782 * their task. This ensures there is no inter-sibling overlap between
5783 * non-matching user state.
5785 for_each_cpu(i, smt_mask) {
5786 struct rq *rq_i = cpu_rq(i);
5789 * An online sibling might have gone offline before a task
5790 * could be picked for it, or it might be offline but later
5791 * happen to come online, but its too late and nothing was
5792 * picked for it. That's Ok - it will pick tasks for itself,
5795 if (!rq_i->core_pick)
5799 * Update for new !FI->FI transitions, or if continuing to be in !FI:
5800 * fi_before fi update?
5806 if (!(fi_before && rq->core->core_forceidle))
5807 task_vruntime_update(rq_i, rq_i->core_pick, rq->core->core_forceidle);
5809 rq_i->core_pick->core_occupation = occ;
5812 rq_i->core_pick = NULL;
5816 /* Did we break L1TF mitigation requirements? */
5817 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
5819 if (rq_i->curr == rq_i->core_pick) {
5820 rq_i->core_pick = NULL;
5828 set_next_task(rq, next);
5832 static bool try_steal_cookie(int this, int that)
5834 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
5835 struct task_struct *p;
5836 unsigned long cookie;
5837 bool success = false;
5839 local_irq_disable();
5840 double_rq_lock(dst, src);
5842 cookie = dst->core->core_cookie;
5846 if (dst->curr != dst->idle)
5849 p = sched_core_find(src, cookie);
5854 if (p == src->core_pick || p == src->curr)
5857 if (!cpumask_test_cpu(this, &p->cpus_mask))
5860 if (p->core_occupation > dst->idle->core_occupation)
5863 deactivate_task(src, p, 0);
5864 set_task_cpu(p, this);
5865 activate_task(dst, p, 0);
5873 p = sched_core_next(p, cookie);
5877 double_rq_unlock(dst, src);
5883 static bool steal_cookie_task(int cpu, struct sched_domain *sd)
5887 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
5894 if (try_steal_cookie(cpu, i))
5901 static void sched_core_balance(struct rq *rq)
5903 struct sched_domain *sd;
5904 int cpu = cpu_of(rq);
5908 raw_spin_rq_unlock_irq(rq);
5909 for_each_domain(cpu, sd) {
5913 if (steal_cookie_task(cpu, sd))
5916 raw_spin_rq_lock_irq(rq);
5921 static DEFINE_PER_CPU(struct callback_head, core_balance_head);
5923 void queue_core_balance(struct rq *rq)
5925 if (!sched_core_enabled(rq))
5928 if (!rq->core->core_cookie)
5931 if (!rq->nr_running) /* not forced idle */
5934 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
5937 static inline void sched_core_cpu_starting(unsigned int cpu)
5939 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
5940 struct rq *rq, *core_rq = NULL;
5943 core_rq = cpu_rq(cpu)->core;
5946 for_each_cpu(i, smt_mask) {
5948 if (rq->core && rq->core == rq)
5953 core_rq = cpu_rq(cpu);
5955 for_each_cpu(i, smt_mask) {
5958 WARN_ON_ONCE(rq->core && rq->core != core_rq);
5963 #else /* !CONFIG_SCHED_CORE */
5965 static inline void sched_core_cpu_starting(unsigned int cpu) {}
5967 static struct task_struct *
5968 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5970 return __pick_next_task(rq, prev, rf);
5973 #endif /* CONFIG_SCHED_CORE */
5976 * __schedule() is the main scheduler function.
5978 * The main means of driving the scheduler and thus entering this function are:
5980 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
5982 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
5983 * paths. For example, see arch/x86/entry_64.S.
5985 * To drive preemption between tasks, the scheduler sets the flag in timer
5986 * interrupt handler scheduler_tick().
5988 * 3. Wakeups don't really cause entry into schedule(). They add a
5989 * task to the run-queue and that's it.
5991 * Now, if the new task added to the run-queue preempts the current
5992 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
5993 * called on the nearest possible occasion:
5995 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
5997 * - in syscall or exception context, at the next outmost
5998 * preempt_enable(). (this might be as soon as the wake_up()'s
6001 * - in IRQ context, return from interrupt-handler to
6002 * preemptible context
6004 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
6007 * - cond_resched() call
6008 * - explicit schedule() call
6009 * - return from syscall or exception to user-space
6010 * - return from interrupt-handler to user-space
6012 * WARNING: must be called with preemption disabled!
6014 static void __sched notrace __schedule(bool preempt)
6016 struct task_struct *prev, *next;
6017 unsigned long *switch_count;
6018 unsigned long prev_state;
6023 cpu = smp_processor_id();
6027 schedule_debug(prev, preempt);
6029 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
6032 local_irq_disable();
6033 rcu_note_context_switch(preempt);
6036 * Make sure that signal_pending_state()->signal_pending() below
6037 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
6038 * done by the caller to avoid the race with signal_wake_up():
6040 * __set_current_state(@state) signal_wake_up()
6041 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6042 * wake_up_state(p, state)
6043 * LOCK rq->lock LOCK p->pi_state
6044 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6045 * if (signal_pending_state()) if (p->state & @state)
6047 * Also, the membarrier system call requires a full memory barrier
6048 * after coming from user-space, before storing to rq->curr.
6051 smp_mb__after_spinlock();
6053 /* Promote REQ to ACT */
6054 rq->clock_update_flags <<= 1;
6055 update_rq_clock(rq);
6057 switch_count = &prev->nivcsw;
6060 * We must load prev->state once (task_struct::state is volatile), such
6063 * - we form a control dependency vs deactivate_task() below.
6064 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
6066 prev_state = READ_ONCE(prev->__state);
6067 if (!preempt && prev_state) {
6068 if (signal_pending_state(prev_state, prev)) {
6069 WRITE_ONCE(prev->__state, TASK_RUNNING);
6071 prev->sched_contributes_to_load =
6072 (prev_state & TASK_UNINTERRUPTIBLE) &&
6073 !(prev_state & TASK_NOLOAD) &&
6074 !(prev->flags & PF_FROZEN);
6076 if (prev->sched_contributes_to_load)
6077 rq->nr_uninterruptible++;
6080 * __schedule() ttwu()
6081 * prev_state = prev->state; if (p->on_rq && ...)
6082 * if (prev_state) goto out;
6083 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6084 * p->state = TASK_WAKING
6086 * Where __schedule() and ttwu() have matching control dependencies.
6088 * After this, schedule() must not care about p->state any more.
6090 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
6092 if (prev->in_iowait) {
6093 atomic_inc(&rq->nr_iowait);
6094 delayacct_blkio_start();
6097 switch_count = &prev->nvcsw;
6100 next = pick_next_task(rq, prev, &rf);
6101 clear_tsk_need_resched(prev);
6102 clear_preempt_need_resched();
6103 #ifdef CONFIG_SCHED_DEBUG
6104 rq->last_seen_need_resched_ns = 0;
6107 if (likely(prev != next)) {
6110 * RCU users of rcu_dereference(rq->curr) may not see
6111 * changes to task_struct made by pick_next_task().
6113 RCU_INIT_POINTER(rq->curr, next);
6115 * The membarrier system call requires each architecture
6116 * to have a full memory barrier after updating
6117 * rq->curr, before returning to user-space.
6119 * Here are the schemes providing that barrier on the
6120 * various architectures:
6121 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6122 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6123 * - finish_lock_switch() for weakly-ordered
6124 * architectures where spin_unlock is a full barrier,
6125 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6126 * is a RELEASE barrier),
6130 migrate_disable_switch(rq, prev);
6131 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6133 trace_sched_switch(preempt, prev, next);
6135 /* Also unlocks the rq: */
6136 rq = context_switch(rq, prev, next, &rf);
6138 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
6140 rq_unpin_lock(rq, &rf);
6141 __balance_callbacks(rq);
6142 raw_spin_rq_unlock_irq(rq);
6146 void __noreturn do_task_dead(void)
6148 /* Causes final put_task_struct in finish_task_switch(): */
6149 set_special_state(TASK_DEAD);
6151 /* Tell freezer to ignore us: */
6152 current->flags |= PF_NOFREEZE;
6157 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
6162 static inline void sched_submit_work(struct task_struct *tsk)
6164 unsigned int task_flags;
6166 if (task_is_running(tsk))
6169 task_flags = tsk->flags;
6171 * If a worker went to sleep, notify and ask workqueue whether
6172 * it wants to wake up a task to maintain concurrency.
6173 * As this function is called inside the schedule() context,
6174 * we disable preemption to avoid it calling schedule() again
6175 * in the possible wakeup of a kworker and because wq_worker_sleeping()
6178 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6180 if (task_flags & PF_WQ_WORKER)
6181 wq_worker_sleeping(tsk);
6183 io_wq_worker_sleeping(tsk);
6184 preempt_enable_no_resched();
6187 if (tsk_is_pi_blocked(tsk))
6191 * If we are going to sleep and we have plugged IO queued,
6192 * make sure to submit it to avoid deadlocks.
6194 if (blk_needs_flush_plug(tsk))
6195 blk_schedule_flush_plug(tsk);
6198 static void sched_update_worker(struct task_struct *tsk)
6200 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6201 if (tsk->flags & PF_WQ_WORKER)
6202 wq_worker_running(tsk);
6204 io_wq_worker_running(tsk);
6208 asmlinkage __visible void __sched schedule(void)
6210 struct task_struct *tsk = current;
6212 sched_submit_work(tsk);
6216 sched_preempt_enable_no_resched();
6217 } while (need_resched());
6218 sched_update_worker(tsk);
6220 EXPORT_SYMBOL(schedule);
6223 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6224 * state (have scheduled out non-voluntarily) by making sure that all
6225 * tasks have either left the run queue or have gone into user space.
6226 * As idle tasks do not do either, they must not ever be preempted
6227 * (schedule out non-voluntarily).
6229 * schedule_idle() is similar to schedule_preempt_disable() except that it
6230 * never enables preemption because it does not call sched_submit_work().
6232 void __sched schedule_idle(void)
6235 * As this skips calling sched_submit_work(), which the idle task does
6236 * regardless because that function is a nop when the task is in a
6237 * TASK_RUNNING state, make sure this isn't used someplace that the
6238 * current task can be in any other state. Note, idle is always in the
6239 * TASK_RUNNING state.
6241 WARN_ON_ONCE(current->__state);
6244 } while (need_resched());
6247 #if defined(CONFIG_CONTEXT_TRACKING) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_OFFSTACK)
6248 asmlinkage __visible void __sched schedule_user(void)
6251 * If we come here after a random call to set_need_resched(),
6252 * or we have been woken up remotely but the IPI has not yet arrived,
6253 * we haven't yet exited the RCU idle mode. Do it here manually until
6254 * we find a better solution.
6256 * NB: There are buggy callers of this function. Ideally we
6257 * should warn if prev_state != CONTEXT_USER, but that will trigger
6258 * too frequently to make sense yet.
6260 enum ctx_state prev_state = exception_enter();
6262 exception_exit(prev_state);
6267 * schedule_preempt_disabled - called with preemption disabled
6269 * Returns with preemption disabled. Note: preempt_count must be 1
6271 void __sched schedule_preempt_disabled(void)
6273 sched_preempt_enable_no_resched();
6278 static void __sched notrace preempt_schedule_common(void)
6282 * Because the function tracer can trace preempt_count_sub()
6283 * and it also uses preempt_enable/disable_notrace(), if
6284 * NEED_RESCHED is set, the preempt_enable_notrace() called
6285 * by the function tracer will call this function again and
6286 * cause infinite recursion.
6288 * Preemption must be disabled here before the function
6289 * tracer can trace. Break up preempt_disable() into two
6290 * calls. One to disable preemption without fear of being
6291 * traced. The other to still record the preemption latency,
6292 * which can also be traced by the function tracer.
6294 preempt_disable_notrace();
6295 preempt_latency_start(1);
6297 preempt_latency_stop(1);
6298 preempt_enable_no_resched_notrace();
6301 * Check again in case we missed a preemption opportunity
6302 * between schedule and now.
6304 } while (need_resched());
6307 #ifdef CONFIG_PREEMPTION
6309 * This is the entry point to schedule() from in-kernel preemption
6310 * off of preempt_enable.
6312 asmlinkage __visible void __sched notrace preempt_schedule(void)
6315 * If there is a non-zero preempt_count or interrupts are disabled,
6316 * we do not want to preempt the current task. Just return..
6318 if (likely(!preemptible()))
6321 preempt_schedule_common();
6323 NOKPROBE_SYMBOL(preempt_schedule);
6324 EXPORT_SYMBOL(preempt_schedule);
6326 #ifdef CONFIG_PREEMPT_DYNAMIC
6327 DEFINE_STATIC_CALL(preempt_schedule, __preempt_schedule_func);
6328 EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
6333 * preempt_schedule_notrace - preempt_schedule called by tracing
6335 * The tracing infrastructure uses preempt_enable_notrace to prevent
6336 * recursion and tracing preempt enabling caused by the tracing
6337 * infrastructure itself. But as tracing can happen in areas coming
6338 * from userspace or just about to enter userspace, a preempt enable
6339 * can occur before user_exit() is called. This will cause the scheduler
6340 * to be called when the system is still in usermode.
6342 * To prevent this, the preempt_enable_notrace will use this function
6343 * instead of preempt_schedule() to exit user context if needed before
6344 * calling the scheduler.
6346 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
6348 enum ctx_state prev_ctx;
6350 if (likely(!preemptible()))
6355 * Because the function tracer can trace preempt_count_sub()
6356 * and it also uses preempt_enable/disable_notrace(), if
6357 * NEED_RESCHED is set, the preempt_enable_notrace() called
6358 * by the function tracer will call this function again and
6359 * cause infinite recursion.
6361 * Preemption must be disabled here before the function
6362 * tracer can trace. Break up preempt_disable() into two
6363 * calls. One to disable preemption without fear of being
6364 * traced. The other to still record the preemption latency,
6365 * which can also be traced by the function tracer.
6367 preempt_disable_notrace();
6368 preempt_latency_start(1);
6370 * Needs preempt disabled in case user_exit() is traced
6371 * and the tracer calls preempt_enable_notrace() causing
6372 * an infinite recursion.
6374 prev_ctx = exception_enter();
6376 exception_exit(prev_ctx);
6378 preempt_latency_stop(1);
6379 preempt_enable_no_resched_notrace();
6380 } while (need_resched());
6382 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
6384 #ifdef CONFIG_PREEMPT_DYNAMIC
6385 DEFINE_STATIC_CALL(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6386 EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
6389 #endif /* CONFIG_PREEMPTION */
6391 #ifdef CONFIG_PREEMPT_DYNAMIC
6393 #include <linux/entry-common.h>
6398 * SC:preempt_schedule
6399 * SC:preempt_schedule_notrace
6400 * SC:irqentry_exit_cond_resched
6404 * cond_resched <- __cond_resched
6405 * might_resched <- RET0
6406 * preempt_schedule <- NOP
6407 * preempt_schedule_notrace <- NOP
6408 * irqentry_exit_cond_resched <- NOP
6411 * cond_resched <- __cond_resched
6412 * might_resched <- __cond_resched
6413 * preempt_schedule <- NOP
6414 * preempt_schedule_notrace <- NOP
6415 * irqentry_exit_cond_resched <- NOP
6418 * cond_resched <- RET0
6419 * might_resched <- RET0
6420 * preempt_schedule <- preempt_schedule
6421 * preempt_schedule_notrace <- preempt_schedule_notrace
6422 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
6426 preempt_dynamic_none = 0,
6427 preempt_dynamic_voluntary,
6428 preempt_dynamic_full,
6431 int preempt_dynamic_mode = preempt_dynamic_full;
6433 int sched_dynamic_mode(const char *str)
6435 if (!strcmp(str, "none"))
6436 return preempt_dynamic_none;
6438 if (!strcmp(str, "voluntary"))
6439 return preempt_dynamic_voluntary;
6441 if (!strcmp(str, "full"))
6442 return preempt_dynamic_full;
6447 void sched_dynamic_update(int mode)
6450 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
6451 * the ZERO state, which is invalid.
6453 static_call_update(cond_resched, __cond_resched);
6454 static_call_update(might_resched, __cond_resched);
6455 static_call_update(preempt_schedule, __preempt_schedule_func);
6456 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6457 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6460 case preempt_dynamic_none:
6461 static_call_update(cond_resched, __cond_resched);
6462 static_call_update(might_resched, (void *)&__static_call_return0);
6463 static_call_update(preempt_schedule, NULL);
6464 static_call_update(preempt_schedule_notrace, NULL);
6465 static_call_update(irqentry_exit_cond_resched, NULL);
6466 pr_info("Dynamic Preempt: none\n");
6469 case preempt_dynamic_voluntary:
6470 static_call_update(cond_resched, __cond_resched);
6471 static_call_update(might_resched, __cond_resched);
6472 static_call_update(preempt_schedule, NULL);
6473 static_call_update(preempt_schedule_notrace, NULL);
6474 static_call_update(irqentry_exit_cond_resched, NULL);
6475 pr_info("Dynamic Preempt: voluntary\n");
6478 case preempt_dynamic_full:
6479 static_call_update(cond_resched, (void *)&__static_call_return0);
6480 static_call_update(might_resched, (void *)&__static_call_return0);
6481 static_call_update(preempt_schedule, __preempt_schedule_func);
6482 static_call_update(preempt_schedule_notrace, __preempt_schedule_notrace_func);
6483 static_call_update(irqentry_exit_cond_resched, irqentry_exit_cond_resched);
6484 pr_info("Dynamic Preempt: full\n");
6488 preempt_dynamic_mode = mode;
6491 static int __init setup_preempt_mode(char *str)
6493 int mode = sched_dynamic_mode(str);
6495 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
6499 sched_dynamic_update(mode);
6502 __setup("preempt=", setup_preempt_mode);
6504 #endif /* CONFIG_PREEMPT_DYNAMIC */
6507 * This is the entry point to schedule() from kernel preemption
6508 * off of irq context.
6509 * Note, that this is called and return with irqs disabled. This will
6510 * protect us against recursive calling from irq.
6512 asmlinkage __visible void __sched preempt_schedule_irq(void)
6514 enum ctx_state prev_state;
6516 /* Catch callers which need to be fixed */
6517 BUG_ON(preempt_count() || !irqs_disabled());
6519 prev_state = exception_enter();
6525 local_irq_disable();
6526 sched_preempt_enable_no_resched();
6527 } while (need_resched());
6529 exception_exit(prev_state);
6532 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
6535 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
6536 return try_to_wake_up(curr->private, mode, wake_flags);
6538 EXPORT_SYMBOL(default_wake_function);
6540 static void __setscheduler_prio(struct task_struct *p, int prio)
6543 p->sched_class = &dl_sched_class;
6544 else if (rt_prio(prio))
6545 p->sched_class = &rt_sched_class;
6547 p->sched_class = &fair_sched_class;
6552 #ifdef CONFIG_RT_MUTEXES
6554 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6557 prio = min(prio, pi_task->prio);
6562 static inline int rt_effective_prio(struct task_struct *p, int prio)
6564 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6566 return __rt_effective_prio(pi_task, prio);
6570 * rt_mutex_setprio - set the current priority of a task
6572 * @pi_task: donor task
6574 * This function changes the 'effective' priority of a task. It does
6575 * not touch ->normal_prio like __setscheduler().
6577 * Used by the rt_mutex code to implement priority inheritance
6578 * logic. Call site only calls if the priority of the task changed.
6580 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
6582 int prio, oldprio, queued, running, queue_flag =
6583 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6584 const struct sched_class *prev_class;
6588 /* XXX used to be waiter->prio, not waiter->task->prio */
6589 prio = __rt_effective_prio(pi_task, p->normal_prio);
6592 * If nothing changed; bail early.
6594 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6597 rq = __task_rq_lock(p, &rf);
6598 update_rq_clock(rq);
6600 * Set under pi_lock && rq->lock, such that the value can be used under
6603 * Note that there is loads of tricky to make this pointer cache work
6604 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6605 * ensure a task is de-boosted (pi_task is set to NULL) before the
6606 * task is allowed to run again (and can exit). This ensures the pointer
6607 * points to a blocked task -- which guarantees the task is present.
6609 p->pi_top_task = pi_task;
6612 * For FIFO/RR we only need to set prio, if that matches we're done.
6614 if (prio == p->prio && !dl_prio(prio))
6618 * Idle task boosting is a nono in general. There is one
6619 * exception, when PREEMPT_RT and NOHZ is active:
6621 * The idle task calls get_next_timer_interrupt() and holds
6622 * the timer wheel base->lock on the CPU and another CPU wants
6623 * to access the timer (probably to cancel it). We can safely
6624 * ignore the boosting request, as the idle CPU runs this code
6625 * with interrupts disabled and will complete the lock
6626 * protected section without being interrupted. So there is no
6627 * real need to boost.
6629 if (unlikely(p == rq->idle)) {
6630 WARN_ON(p != rq->curr);
6631 WARN_ON(p->pi_blocked_on);
6635 trace_sched_pi_setprio(p, pi_task);
6638 if (oldprio == prio)
6639 queue_flag &= ~DEQUEUE_MOVE;
6641 prev_class = p->sched_class;
6642 queued = task_on_rq_queued(p);
6643 running = task_current(rq, p);
6645 dequeue_task(rq, p, queue_flag);
6647 put_prev_task(rq, p);
6650 * Boosting condition are:
6651 * 1. -rt task is running and holds mutex A
6652 * --> -dl task blocks on mutex A
6654 * 2. -dl task is running and holds mutex A
6655 * --> -dl task blocks on mutex A and could preempt the
6658 if (dl_prio(prio)) {
6659 if (!dl_prio(p->normal_prio) ||
6660 (pi_task && dl_prio(pi_task->prio) &&
6661 dl_entity_preempt(&pi_task->dl, &p->dl))) {
6662 p->dl.pi_se = pi_task->dl.pi_se;
6663 queue_flag |= ENQUEUE_REPLENISH;
6665 p->dl.pi_se = &p->dl;
6667 } else if (rt_prio(prio)) {
6668 if (dl_prio(oldprio))
6669 p->dl.pi_se = &p->dl;
6671 queue_flag |= ENQUEUE_HEAD;
6673 if (dl_prio(oldprio))
6674 p->dl.pi_se = &p->dl;
6675 if (rt_prio(oldprio))
6679 __setscheduler_prio(p, prio);
6682 enqueue_task(rq, p, queue_flag);
6684 set_next_task(rq, p);
6686 check_class_changed(rq, p, prev_class, oldprio);
6688 /* Avoid rq from going away on us: */
6691 rq_unpin_lock(rq, &rf);
6692 __balance_callbacks(rq);
6693 raw_spin_rq_unlock(rq);
6698 static inline int rt_effective_prio(struct task_struct *p, int prio)
6704 void set_user_nice(struct task_struct *p, long nice)
6706 bool queued, running;
6711 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
6714 * We have to be careful, if called from sys_setpriority(),
6715 * the task might be in the middle of scheduling on another CPU.
6717 rq = task_rq_lock(p, &rf);
6718 update_rq_clock(rq);
6721 * The RT priorities are set via sched_setscheduler(), but we still
6722 * allow the 'normal' nice value to be set - but as expected
6723 * it won't have any effect on scheduling until the task is
6724 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
6726 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
6727 p->static_prio = NICE_TO_PRIO(nice);
6730 queued = task_on_rq_queued(p);
6731 running = task_current(rq, p);
6733 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
6735 put_prev_task(rq, p);
6737 p->static_prio = NICE_TO_PRIO(nice);
6738 set_load_weight(p, true);
6740 p->prio = effective_prio(p);
6743 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6745 set_next_task(rq, p);
6748 * If the task increased its priority or is running and
6749 * lowered its priority, then reschedule its CPU:
6751 p->sched_class->prio_changed(rq, p, old_prio);
6754 task_rq_unlock(rq, p, &rf);
6756 EXPORT_SYMBOL(set_user_nice);
6759 * can_nice - check if a task can reduce its nice value
6763 int can_nice(const struct task_struct *p, const int nice)
6765 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
6766 int nice_rlim = nice_to_rlimit(nice);
6768 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
6769 capable(CAP_SYS_NICE));
6772 #ifdef __ARCH_WANT_SYS_NICE
6775 * sys_nice - change the priority of the current process.
6776 * @increment: priority increment
6778 * sys_setpriority is a more generic, but much slower function that
6779 * does similar things.
6781 SYSCALL_DEFINE1(nice, int, increment)
6786 * Setpriority might change our priority at the same moment.
6787 * We don't have to worry. Conceptually one call occurs first
6788 * and we have a single winner.
6790 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
6791 nice = task_nice(current) + increment;
6793 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
6794 if (increment < 0 && !can_nice(current, nice))
6797 retval = security_task_setnice(current, nice);
6801 set_user_nice(current, nice);
6808 * task_prio - return the priority value of a given task.
6809 * @p: the task in question.
6811 * Return: The priority value as seen by users in /proc.
6813 * sched policy return value kernel prio user prio/nice
6815 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
6816 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
6817 * deadline -101 -1 0
6819 int task_prio(const struct task_struct *p)
6821 return p->prio - MAX_RT_PRIO;
6825 * idle_cpu - is a given CPU idle currently?
6826 * @cpu: the processor in question.
6828 * Return: 1 if the CPU is currently idle. 0 otherwise.
6830 int idle_cpu(int cpu)
6832 struct rq *rq = cpu_rq(cpu);
6834 if (rq->curr != rq->idle)
6841 if (rq->ttwu_pending)
6849 * available_idle_cpu - is a given CPU idle for enqueuing work.
6850 * @cpu: the CPU in question.
6852 * Return: 1 if the CPU is currently idle. 0 otherwise.
6854 int available_idle_cpu(int cpu)
6859 if (vcpu_is_preempted(cpu))
6866 * idle_task - return the idle task for a given CPU.
6867 * @cpu: the processor in question.
6869 * Return: The idle task for the CPU @cpu.
6871 struct task_struct *idle_task(int cpu)
6873 return cpu_rq(cpu)->idle;
6878 * This function computes an effective utilization for the given CPU, to be
6879 * used for frequency selection given the linear relation: f = u * f_max.
6881 * The scheduler tracks the following metrics:
6883 * cpu_util_{cfs,rt,dl,irq}()
6886 * Where the cfs,rt and dl util numbers are tracked with the same metric and
6887 * synchronized windows and are thus directly comparable.
6889 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
6890 * which excludes things like IRQ and steal-time. These latter are then accrued
6891 * in the irq utilization.
6893 * The DL bandwidth number otoh is not a measured metric but a value computed
6894 * based on the task model parameters and gives the minimal utilization
6895 * required to meet deadlines.
6897 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
6898 unsigned long max, enum cpu_util_type type,
6899 struct task_struct *p)
6901 unsigned long dl_util, util, irq;
6902 struct rq *rq = cpu_rq(cpu);
6904 if (!uclamp_is_used() &&
6905 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
6910 * Early check to see if IRQ/steal time saturates the CPU, can be
6911 * because of inaccuracies in how we track these -- see
6912 * update_irq_load_avg().
6914 irq = cpu_util_irq(rq);
6915 if (unlikely(irq >= max))
6919 * Because the time spend on RT/DL tasks is visible as 'lost' time to
6920 * CFS tasks and we use the same metric to track the effective
6921 * utilization (PELT windows are synchronized) we can directly add them
6922 * to obtain the CPU's actual utilization.
6924 * CFS and RT utilization can be boosted or capped, depending on
6925 * utilization clamp constraints requested by currently RUNNABLE
6927 * When there are no CFS RUNNABLE tasks, clamps are released and
6928 * frequency will be gracefully reduced with the utilization decay.
6930 util = util_cfs + cpu_util_rt(rq);
6931 if (type == FREQUENCY_UTIL)
6932 util = uclamp_rq_util_with(rq, util, p);
6934 dl_util = cpu_util_dl(rq);
6937 * For frequency selection we do not make cpu_util_dl() a permanent part
6938 * of this sum because we want to use cpu_bw_dl() later on, but we need
6939 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
6940 * that we select f_max when there is no idle time.
6942 * NOTE: numerical errors or stop class might cause us to not quite hit
6943 * saturation when we should -- something for later.
6945 if (util + dl_util >= max)
6949 * OTOH, for energy computation we need the estimated running time, so
6950 * include util_dl and ignore dl_bw.
6952 if (type == ENERGY_UTIL)
6956 * There is still idle time; further improve the number by using the
6957 * irq metric. Because IRQ/steal time is hidden from the task clock we
6958 * need to scale the task numbers:
6961 * U' = irq + --------- * U
6964 util = scale_irq_capacity(util, irq, max);
6968 * Bandwidth required by DEADLINE must always be granted while, for
6969 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
6970 * to gracefully reduce the frequency when no tasks show up for longer
6973 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
6974 * bw_dl as requested freq. However, cpufreq is not yet ready for such
6975 * an interface. So, we only do the latter for now.
6977 if (type == FREQUENCY_UTIL)
6978 util += cpu_bw_dl(rq);
6980 return min(max, util);
6983 unsigned long sched_cpu_util(int cpu, unsigned long max)
6985 return effective_cpu_util(cpu, cpu_util_cfs(cpu_rq(cpu)), max,
6988 #endif /* CONFIG_SMP */
6991 * find_process_by_pid - find a process with a matching PID value.
6992 * @pid: the pid in question.
6994 * The task of @pid, if found. %NULL otherwise.
6996 static struct task_struct *find_process_by_pid(pid_t pid)
6998 return pid ? find_task_by_vpid(pid) : current;
7002 * sched_setparam() passes in -1 for its policy, to let the functions
7003 * it calls know not to change it.
7005 #define SETPARAM_POLICY -1
7007 static void __setscheduler_params(struct task_struct *p,
7008 const struct sched_attr *attr)
7010 int policy = attr->sched_policy;
7012 if (policy == SETPARAM_POLICY)
7017 if (dl_policy(policy))
7018 __setparam_dl(p, attr);
7019 else if (fair_policy(policy))
7020 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7023 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7024 * !rt_policy. Always setting this ensures that things like
7025 * getparam()/getattr() don't report silly values for !rt tasks.
7027 p->rt_priority = attr->sched_priority;
7028 p->normal_prio = normal_prio(p);
7029 set_load_weight(p, true);
7033 * Check the target process has a UID that matches the current process's:
7035 static bool check_same_owner(struct task_struct *p)
7037 const struct cred *cred = current_cred(), *pcred;
7041 pcred = __task_cred(p);
7042 match = (uid_eq(cred->euid, pcred->euid) ||
7043 uid_eq(cred->euid, pcred->uid));
7048 static int __sched_setscheduler(struct task_struct *p,
7049 const struct sched_attr *attr,
7052 int oldpolicy = -1, policy = attr->sched_policy;
7053 int retval, oldprio, newprio, queued, running;
7054 const struct sched_class *prev_class;
7055 struct callback_head *head;
7058 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7061 /* The pi code expects interrupts enabled */
7062 BUG_ON(pi && in_interrupt());
7064 /* Double check policy once rq lock held: */
7066 reset_on_fork = p->sched_reset_on_fork;
7067 policy = oldpolicy = p->policy;
7069 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
7071 if (!valid_policy(policy))
7075 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
7079 * Valid priorities for SCHED_FIFO and SCHED_RR are
7080 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
7081 * SCHED_BATCH and SCHED_IDLE is 0.
7083 if (attr->sched_priority > MAX_RT_PRIO-1)
7085 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7086 (rt_policy(policy) != (attr->sched_priority != 0)))
7090 * Allow unprivileged RT tasks to decrease priority:
7092 if (user && !capable(CAP_SYS_NICE)) {
7093 if (fair_policy(policy)) {
7094 if (attr->sched_nice < task_nice(p) &&
7095 !can_nice(p, attr->sched_nice))
7099 if (rt_policy(policy)) {
7100 unsigned long rlim_rtprio =
7101 task_rlimit(p, RLIMIT_RTPRIO);
7103 /* Can't set/change the rt policy: */
7104 if (policy != p->policy && !rlim_rtprio)
7107 /* Can't increase priority: */
7108 if (attr->sched_priority > p->rt_priority &&
7109 attr->sched_priority > rlim_rtprio)
7114 * Can't set/change SCHED_DEADLINE policy at all for now
7115 * (safest behavior); in the future we would like to allow
7116 * unprivileged DL tasks to increase their relative deadline
7117 * or reduce their runtime (both ways reducing utilization)
7119 if (dl_policy(policy))
7123 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7124 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7126 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7127 if (!can_nice(p, task_nice(p)))
7131 /* Can't change other user's priorities: */
7132 if (!check_same_owner(p))
7135 /* Normal users shall not reset the sched_reset_on_fork flag: */
7136 if (p->sched_reset_on_fork && !reset_on_fork)
7141 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7144 retval = security_task_setscheduler(p);
7149 /* Update task specific "requested" clamps */
7150 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7151 retval = uclamp_validate(p, attr);
7160 * Make sure no PI-waiters arrive (or leave) while we are
7161 * changing the priority of the task:
7163 * To be able to change p->policy safely, the appropriate
7164 * runqueue lock must be held.
7166 rq = task_rq_lock(p, &rf);
7167 update_rq_clock(rq);
7170 * Changing the policy of the stop threads its a very bad idea:
7172 if (p == rq->stop) {
7178 * If not changing anything there's no need to proceed further,
7179 * but store a possible modification of reset_on_fork.
7181 if (unlikely(policy == p->policy)) {
7182 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
7184 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7186 if (dl_policy(policy) && dl_param_changed(p, attr))
7188 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7191 p->sched_reset_on_fork = reset_on_fork;
7198 #ifdef CONFIG_RT_GROUP_SCHED
7200 * Do not allow realtime tasks into groups that have no runtime
7203 if (rt_bandwidth_enabled() && rt_policy(policy) &&
7204 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7205 !task_group_is_autogroup(task_group(p))) {
7211 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7212 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
7213 cpumask_t *span = rq->rd->span;
7216 * Don't allow tasks with an affinity mask smaller than
7217 * the entire root_domain to become SCHED_DEADLINE. We
7218 * will also fail if there's no bandwidth available.
7220 if (!cpumask_subset(span, p->cpus_ptr) ||
7221 rq->rd->dl_bw.bw == 0) {
7229 /* Re-check policy now with rq lock held: */
7230 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7231 policy = oldpolicy = -1;
7232 task_rq_unlock(rq, p, &rf);
7234 cpuset_read_unlock();
7239 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7240 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7243 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
7248 p->sched_reset_on_fork = reset_on_fork;
7251 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
7254 * Take priority boosted tasks into account. If the new
7255 * effective priority is unchanged, we just store the new
7256 * normal parameters and do not touch the scheduler class and
7257 * the runqueue. This will be done when the task deboost
7260 newprio = rt_effective_prio(p, newprio);
7261 if (newprio == oldprio)
7262 queue_flags &= ~DEQUEUE_MOVE;
7265 queued = task_on_rq_queued(p);
7266 running = task_current(rq, p);
7268 dequeue_task(rq, p, queue_flags);
7270 put_prev_task(rq, p);
7272 prev_class = p->sched_class;
7274 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7275 __setscheduler_params(p, attr);
7276 __setscheduler_prio(p, newprio);
7278 __setscheduler_uclamp(p, attr);
7282 * We enqueue to tail when the priority of a task is
7283 * increased (user space view).
7285 if (oldprio < p->prio)
7286 queue_flags |= ENQUEUE_HEAD;
7288 enqueue_task(rq, p, queue_flags);
7291 set_next_task(rq, p);
7293 check_class_changed(rq, p, prev_class, oldprio);
7295 /* Avoid rq from going away on us: */
7297 head = splice_balance_callbacks(rq);
7298 task_rq_unlock(rq, p, &rf);
7301 cpuset_read_unlock();
7302 rt_mutex_adjust_pi(p);
7305 /* Run balance callbacks after we've adjusted the PI chain: */
7306 balance_callbacks(rq, head);
7312 task_rq_unlock(rq, p, &rf);
7314 cpuset_read_unlock();
7318 static int _sched_setscheduler(struct task_struct *p, int policy,
7319 const struct sched_param *param, bool check)
7321 struct sched_attr attr = {
7322 .sched_policy = policy,
7323 .sched_priority = param->sched_priority,
7324 .sched_nice = PRIO_TO_NICE(p->static_prio),
7327 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7328 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
7329 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7330 policy &= ~SCHED_RESET_ON_FORK;
7331 attr.sched_policy = policy;
7334 return __sched_setscheduler(p, &attr, check, true);
7337 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7338 * @p: the task in question.
7339 * @policy: new policy.
7340 * @param: structure containing the new RT priority.
7342 * Use sched_set_fifo(), read its comment.
7344 * Return: 0 on success. An error code otherwise.
7346 * NOTE that the task may be already dead.
7348 int sched_setscheduler(struct task_struct *p, int policy,
7349 const struct sched_param *param)
7351 return _sched_setscheduler(p, policy, param, true);
7354 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7356 return __sched_setscheduler(p, attr, true, true);
7359 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7361 return __sched_setscheduler(p, attr, false, true);
7363 EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
7366 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7367 * @p: the task in question.
7368 * @policy: new policy.
7369 * @param: structure containing the new RT priority.
7371 * Just like sched_setscheduler, only don't bother checking if the
7372 * current context has permission. For example, this is needed in
7373 * stop_machine(): we create temporary high priority worker threads,
7374 * but our caller might not have that capability.
7376 * Return: 0 on success. An error code otherwise.
7378 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
7379 const struct sched_param *param)
7381 return _sched_setscheduler(p, policy, param, false);
7385 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7386 * incapable of resource management, which is the one thing an OS really should
7389 * This is of course the reason it is limited to privileged users only.
7391 * Worse still; it is fundamentally impossible to compose static priority
7392 * workloads. You cannot take two correctly working static prio workloads
7393 * and smash them together and still expect them to work.
7395 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7399 * The administrator _MUST_ configure the system, the kernel simply doesn't
7400 * know enough information to make a sensible choice.
7402 void sched_set_fifo(struct task_struct *p)
7404 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
7405 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7407 EXPORT_SYMBOL_GPL(sched_set_fifo);
7410 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7412 void sched_set_fifo_low(struct task_struct *p)
7414 struct sched_param sp = { .sched_priority = 1 };
7415 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
7417 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7419 void sched_set_normal(struct task_struct *p, int nice)
7421 struct sched_attr attr = {
7422 .sched_policy = SCHED_NORMAL,
7425 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
7427 EXPORT_SYMBOL_GPL(sched_set_normal);
7430 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7432 struct sched_param lparam;
7433 struct task_struct *p;
7436 if (!param || pid < 0)
7438 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7443 p = find_process_by_pid(pid);
7449 retval = sched_setscheduler(p, policy, &lparam);
7457 * Mimics kernel/events/core.c perf_copy_attr().
7459 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
7464 /* Zero the full structure, so that a short copy will be nice: */
7465 memset(attr, 0, sizeof(*attr));
7467 ret = get_user(size, &uattr->size);
7471 /* ABI compatibility quirk: */
7473 size = SCHED_ATTR_SIZE_VER0;
7474 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
7477 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7484 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7485 size < SCHED_ATTR_SIZE_VER1)
7489 * XXX: Do we want to be lenient like existing syscalls; or do we want
7490 * to be strict and return an error on out-of-bounds values?
7492 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
7497 put_user(sizeof(*attr), &uattr->size);
7501 static void get_params(struct task_struct *p, struct sched_attr *attr)
7503 if (task_has_dl_policy(p))
7504 __getparam_dl(p, attr);
7505 else if (task_has_rt_policy(p))
7506 attr->sched_priority = p->rt_priority;
7508 attr->sched_nice = task_nice(p);
7512 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7513 * @pid: the pid in question.
7514 * @policy: new policy.
7515 * @param: structure containing the new RT priority.
7517 * Return: 0 on success. An error code otherwise.
7519 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
7524 return do_sched_setscheduler(pid, policy, param);
7528 * sys_sched_setparam - set/change the RT priority of a thread
7529 * @pid: the pid in question.
7530 * @param: structure containing the new RT priority.
7532 * Return: 0 on success. An error code otherwise.
7534 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
7536 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
7540 * sys_sched_setattr - same as above, but with extended sched_attr
7541 * @pid: the pid in question.
7542 * @uattr: structure containing the extended parameters.
7543 * @flags: for future extension.
7545 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7546 unsigned int, flags)
7548 struct sched_attr attr;
7549 struct task_struct *p;
7552 if (!uattr || pid < 0 || flags)
7555 retval = sched_copy_attr(uattr, &attr);
7559 if ((int)attr.sched_policy < 0)
7561 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7562 attr.sched_policy = SETPARAM_POLICY;
7566 p = find_process_by_pid(pid);
7572 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7573 get_params(p, &attr);
7574 retval = sched_setattr(p, &attr);
7582 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
7583 * @pid: the pid in question.
7585 * Return: On success, the policy of the thread. Otherwise, a negative error
7588 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
7590 struct task_struct *p;
7598 p = find_process_by_pid(pid);
7600 retval = security_task_getscheduler(p);
7603 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
7610 * sys_sched_getparam - get the RT priority of a thread
7611 * @pid: the pid in question.
7612 * @param: structure containing the RT priority.
7614 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
7617 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
7619 struct sched_param lp = { .sched_priority = 0 };
7620 struct task_struct *p;
7623 if (!param || pid < 0)
7627 p = find_process_by_pid(pid);
7632 retval = security_task_getscheduler(p);
7636 if (task_has_rt_policy(p))
7637 lp.sched_priority = p->rt_priority;
7641 * This one might sleep, we cannot do it with a spinlock held ...
7643 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
7653 * Copy the kernel size attribute structure (which might be larger
7654 * than what user-space knows about) to user-space.
7656 * Note that all cases are valid: user-space buffer can be larger or
7657 * smaller than the kernel-space buffer. The usual case is that both
7658 * have the same size.
7661 sched_attr_copy_to_user(struct sched_attr __user *uattr,
7662 struct sched_attr *kattr,
7665 unsigned int ksize = sizeof(*kattr);
7667 if (!access_ok(uattr, usize))
7671 * sched_getattr() ABI forwards and backwards compatibility:
7673 * If usize == ksize then we just copy everything to user-space and all is good.
7675 * If usize < ksize then we only copy as much as user-space has space for,
7676 * this keeps ABI compatibility as well. We skip the rest.
7678 * If usize > ksize then user-space is using a newer version of the ABI,
7679 * which part the kernel doesn't know about. Just ignore it - tooling can
7680 * detect the kernel's knowledge of attributes from the attr->size value
7681 * which is set to ksize in this case.
7683 kattr->size = min(usize, ksize);
7685 if (copy_to_user(uattr, kattr, kattr->size))
7692 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
7693 * @pid: the pid in question.
7694 * @uattr: structure containing the extended parameters.
7695 * @usize: sizeof(attr) for fwd/bwd comp.
7696 * @flags: for future extension.
7698 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
7699 unsigned int, usize, unsigned int, flags)
7701 struct sched_attr kattr = { };
7702 struct task_struct *p;
7705 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
7706 usize < SCHED_ATTR_SIZE_VER0 || flags)
7710 p = find_process_by_pid(pid);
7715 retval = security_task_getscheduler(p);
7719 kattr.sched_policy = p->policy;
7720 if (p->sched_reset_on_fork)
7721 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7722 get_params(p, &kattr);
7723 kattr.sched_flags &= SCHED_FLAG_ALL;
7725 #ifdef CONFIG_UCLAMP_TASK
7727 * This could race with another potential updater, but this is fine
7728 * because it'll correctly read the old or the new value. We don't need
7729 * to guarantee who wins the race as long as it doesn't return garbage.
7731 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
7732 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
7737 return sched_attr_copy_to_user(uattr, &kattr, usize);
7745 int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
7750 * If the task isn't a deadline task or admission control is
7751 * disabled then we don't care about affinity changes.
7753 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
7757 * Since bandwidth control happens on root_domain basis,
7758 * if admission test is enabled, we only admit -deadline
7759 * tasks allowed to run on all the CPUs in the task's
7763 if (!cpumask_subset(task_rq(p)->rd->span, mask))
7771 __sched_setaffinity(struct task_struct *p, const struct cpumask *mask)
7774 cpumask_var_t cpus_allowed, new_mask;
7776 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
7779 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
7781 goto out_free_cpus_allowed;
7784 cpuset_cpus_allowed(p, cpus_allowed);
7785 cpumask_and(new_mask, mask, cpus_allowed);
7787 retval = dl_task_check_affinity(p, new_mask);
7789 goto out_free_new_mask;
7791 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK | SCA_USER);
7793 goto out_free_new_mask;
7795 cpuset_cpus_allowed(p, cpus_allowed);
7796 if (!cpumask_subset(new_mask, cpus_allowed)) {
7798 * We must have raced with a concurrent cpuset update.
7799 * Just reset the cpumask to the cpuset's cpus_allowed.
7801 cpumask_copy(new_mask, cpus_allowed);
7806 free_cpumask_var(new_mask);
7807 out_free_cpus_allowed:
7808 free_cpumask_var(cpus_allowed);
7812 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
7814 struct task_struct *p;
7819 p = find_process_by_pid(pid);
7825 /* Prevent p going away */
7829 if (p->flags & PF_NO_SETAFFINITY) {
7834 if (!check_same_owner(p)) {
7836 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
7844 retval = security_task_setscheduler(p);
7848 retval = __sched_setaffinity(p, in_mask);
7854 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
7855 struct cpumask *new_mask)
7857 if (len < cpumask_size())
7858 cpumask_clear(new_mask);
7859 else if (len > cpumask_size())
7860 len = cpumask_size();
7862 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
7866 * sys_sched_setaffinity - set the CPU affinity of a process
7867 * @pid: pid of the process
7868 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7869 * @user_mask_ptr: user-space pointer to the new CPU mask
7871 * Return: 0 on success. An error code otherwise.
7873 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
7874 unsigned long __user *, user_mask_ptr)
7876 cpumask_var_t new_mask;
7879 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
7882 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
7884 retval = sched_setaffinity(pid, new_mask);
7885 free_cpumask_var(new_mask);
7889 long sched_getaffinity(pid_t pid, struct cpumask *mask)
7891 struct task_struct *p;
7892 unsigned long flags;
7898 p = find_process_by_pid(pid);
7902 retval = security_task_getscheduler(p);
7906 raw_spin_lock_irqsave(&p->pi_lock, flags);
7907 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
7908 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
7917 * sys_sched_getaffinity - get the CPU affinity of a process
7918 * @pid: pid of the process
7919 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
7920 * @user_mask_ptr: user-space pointer to hold the current CPU mask
7922 * Return: size of CPU mask copied to user_mask_ptr on success. An
7923 * error code otherwise.
7925 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
7926 unsigned long __user *, user_mask_ptr)
7931 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
7933 if (len & (sizeof(unsigned long)-1))
7936 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
7939 ret = sched_getaffinity(pid, mask);
7941 unsigned int retlen = min(len, cpumask_size());
7943 if (copy_to_user(user_mask_ptr, mask, retlen))
7948 free_cpumask_var(mask);
7953 static void do_sched_yield(void)
7958 rq = this_rq_lock_irq(&rf);
7960 schedstat_inc(rq->yld_count);
7961 current->sched_class->yield_task(rq);
7964 rq_unlock_irq(rq, &rf);
7965 sched_preempt_enable_no_resched();
7971 * sys_sched_yield - yield the current processor to other threads.
7973 * This function yields the current CPU to other tasks. If there are no
7974 * other threads running on this CPU then this function will return.
7978 SYSCALL_DEFINE0(sched_yield)
7984 #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
7985 int __sched __cond_resched(void)
7987 if (should_resched(0)) {
7988 preempt_schedule_common();
7991 #ifndef CONFIG_PREEMPT_RCU
7996 EXPORT_SYMBOL(__cond_resched);
7999 #ifdef CONFIG_PREEMPT_DYNAMIC
8000 DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
8001 EXPORT_STATIC_CALL_TRAMP(cond_resched);
8003 DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
8004 EXPORT_STATIC_CALL_TRAMP(might_resched);
8008 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
8009 * call schedule, and on return reacquire the lock.
8011 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
8012 * operations here to prevent schedule() from being called twice (once via
8013 * spin_unlock(), once by hand).
8015 int __cond_resched_lock(spinlock_t *lock)
8017 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8020 lockdep_assert_held(lock);
8022 if (spin_needbreak(lock) || resched) {
8025 preempt_schedule_common();
8033 EXPORT_SYMBOL(__cond_resched_lock);
8035 int __cond_resched_rwlock_read(rwlock_t *lock)
8037 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8040 lockdep_assert_held_read(lock);
8042 if (rwlock_needbreak(lock) || resched) {
8045 preempt_schedule_common();
8053 EXPORT_SYMBOL(__cond_resched_rwlock_read);
8055 int __cond_resched_rwlock_write(rwlock_t *lock)
8057 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8060 lockdep_assert_held_write(lock);
8062 if (rwlock_needbreak(lock) || resched) {
8065 preempt_schedule_common();
8073 EXPORT_SYMBOL(__cond_resched_rwlock_write);
8076 * yield - yield the current processor to other threads.
8078 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8080 * The scheduler is at all times free to pick the calling task as the most
8081 * eligible task to run, if removing the yield() call from your code breaks
8082 * it, it's already broken.
8084 * Typical broken usage is:
8089 * where one assumes that yield() will let 'the other' process run that will
8090 * make event true. If the current task is a SCHED_FIFO task that will never
8091 * happen. Never use yield() as a progress guarantee!!
8093 * If you want to use yield() to wait for something, use wait_event().
8094 * If you want to use yield() to be 'nice' for others, use cond_resched().
8095 * If you still want to use yield(), do not!
8097 void __sched yield(void)
8099 set_current_state(TASK_RUNNING);
8102 EXPORT_SYMBOL(yield);
8105 * yield_to - yield the current processor to another thread in
8106 * your thread group, or accelerate that thread toward the
8107 * processor it's on.
8109 * @preempt: whether task preemption is allowed or not
8111 * It's the caller's job to ensure that the target task struct
8112 * can't go away on us before we can do any checks.
8115 * true (>0) if we indeed boosted the target task.
8116 * false (0) if we failed to boost the target.
8117 * -ESRCH if there's no task to yield to.
8119 int __sched yield_to(struct task_struct *p, bool preempt)
8121 struct task_struct *curr = current;
8122 struct rq *rq, *p_rq;
8123 unsigned long flags;
8126 local_irq_save(flags);
8132 * If we're the only runnable task on the rq and target rq also
8133 * has only one task, there's absolutely no point in yielding.
8135 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8140 double_rq_lock(rq, p_rq);
8141 if (task_rq(p) != p_rq) {
8142 double_rq_unlock(rq, p_rq);
8146 if (!curr->sched_class->yield_to_task)
8149 if (curr->sched_class != p->sched_class)
8152 if (task_running(p_rq, p) || !task_is_running(p))
8155 yielded = curr->sched_class->yield_to_task(rq, p);
8157 schedstat_inc(rq->yld_count);
8159 * Make p's CPU reschedule; pick_next_entity takes care of
8162 if (preempt && rq != p_rq)
8167 double_rq_unlock(rq, p_rq);
8169 local_irq_restore(flags);
8176 EXPORT_SYMBOL_GPL(yield_to);
8178 int io_schedule_prepare(void)
8180 int old_iowait = current->in_iowait;
8182 current->in_iowait = 1;
8183 blk_schedule_flush_plug(current);
8188 void io_schedule_finish(int token)
8190 current->in_iowait = token;
8194 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
8195 * that process accounting knows that this is a task in IO wait state.
8197 long __sched io_schedule_timeout(long timeout)
8202 token = io_schedule_prepare();
8203 ret = schedule_timeout(timeout);
8204 io_schedule_finish(token);
8208 EXPORT_SYMBOL(io_schedule_timeout);
8210 void __sched io_schedule(void)
8214 token = io_schedule_prepare();
8216 io_schedule_finish(token);
8218 EXPORT_SYMBOL(io_schedule);
8221 * sys_sched_get_priority_max - return maximum RT priority.
8222 * @policy: scheduling class.
8224 * Return: On success, this syscall returns the maximum
8225 * rt_priority that can be used by a given scheduling class.
8226 * On failure, a negative error code is returned.
8228 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
8235 ret = MAX_RT_PRIO-1;
8237 case SCHED_DEADLINE:
8248 * sys_sched_get_priority_min - return minimum RT priority.
8249 * @policy: scheduling class.
8251 * Return: On success, this syscall returns the minimum
8252 * rt_priority that can be used by a given scheduling class.
8253 * On failure, a negative error code is returned.
8255 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
8264 case SCHED_DEADLINE:
8273 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
8275 struct task_struct *p;
8276 unsigned int time_slice;
8286 p = find_process_by_pid(pid);
8290 retval = security_task_getscheduler(p);
8294 rq = task_rq_lock(p, &rf);
8296 if (p->sched_class->get_rr_interval)
8297 time_slice = p->sched_class->get_rr_interval(rq, p);
8298 task_rq_unlock(rq, p, &rf);
8301 jiffies_to_timespec64(time_slice, t);
8310 * sys_sched_rr_get_interval - return the default timeslice of a process.
8311 * @pid: pid of the process.
8312 * @interval: userspace pointer to the timeslice value.
8314 * this syscall writes the default timeslice value of a given process
8315 * into the user-space timespec buffer. A value of '0' means infinity.
8317 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8320 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
8321 struct __kernel_timespec __user *, interval)
8323 struct timespec64 t;
8324 int retval = sched_rr_get_interval(pid, &t);
8327 retval = put_timespec64(&t, interval);
8332 #ifdef CONFIG_COMPAT_32BIT_TIME
8333 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8334 struct old_timespec32 __user *, interval)
8336 struct timespec64 t;
8337 int retval = sched_rr_get_interval(pid, &t);
8340 retval = put_old_timespec32(&t, interval);
8345 void sched_show_task(struct task_struct *p)
8347 unsigned long free = 0;
8350 if (!try_get_task_stack(p))
8353 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
8355 if (task_is_running(p))
8356 pr_cont(" running task ");
8357 #ifdef CONFIG_DEBUG_STACK_USAGE
8358 free = stack_not_used(p);
8363 ppid = task_pid_nr(rcu_dereference(p->real_parent));
8365 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
8366 free, task_pid_nr(p), ppid,
8367 (unsigned long)task_thread_info(p)->flags);
8369 print_worker_info(KERN_INFO, p);
8370 print_stop_info(KERN_INFO, p);
8371 show_stack(p, NULL, KERN_INFO);
8374 EXPORT_SYMBOL_GPL(sched_show_task);
8377 state_filter_match(unsigned long state_filter, struct task_struct *p)
8379 unsigned int state = READ_ONCE(p->__state);
8381 /* no filter, everything matches */
8385 /* filter, but doesn't match */
8386 if (!(state & state_filter))
8390 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
8393 if (state_filter == TASK_UNINTERRUPTIBLE && state == TASK_IDLE)
8400 void show_state_filter(unsigned int state_filter)
8402 struct task_struct *g, *p;
8405 for_each_process_thread(g, p) {
8407 * reset the NMI-timeout, listing all files on a slow
8408 * console might take a lot of time:
8409 * Also, reset softlockup watchdogs on all CPUs, because
8410 * another CPU might be blocked waiting for us to process
8413 touch_nmi_watchdog();
8414 touch_all_softlockup_watchdogs();
8415 if (state_filter_match(state_filter, p))
8419 #ifdef CONFIG_SCHED_DEBUG
8421 sysrq_sched_debug_show();
8425 * Only show locks if all tasks are dumped:
8428 debug_show_all_locks();
8432 * init_idle - set up an idle thread for a given CPU
8433 * @idle: task in question
8434 * @cpu: CPU the idle task belongs to
8436 * NOTE: this function does not set the idle thread's NEED_RESCHED
8437 * flag, to make booting more robust.
8439 void __init init_idle(struct task_struct *idle, int cpu)
8441 struct rq *rq = cpu_rq(cpu);
8442 unsigned long flags;
8444 __sched_fork(0, idle);
8447 * The idle task doesn't need the kthread struct to function, but it
8448 * is dressed up as a per-CPU kthread and thus needs to play the part
8449 * if we want to avoid special-casing it in code that deals with per-CPU
8452 set_kthread_struct(idle);
8454 raw_spin_lock_irqsave(&idle->pi_lock, flags);
8455 raw_spin_rq_lock(rq);
8457 idle->__state = TASK_RUNNING;
8458 idle->se.exec_start = sched_clock();
8460 * PF_KTHREAD should already be set at this point; regardless, make it
8461 * look like a proper per-CPU kthread.
8463 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
8464 kthread_set_per_cpu(idle, cpu);
8466 scs_task_reset(idle);
8467 kasan_unpoison_task_stack(idle);
8471 * It's possible that init_idle() gets called multiple times on a task,
8472 * in that case do_set_cpus_allowed() will not do the right thing.
8474 * And since this is boot we can forgo the serialization.
8476 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
8479 * We're having a chicken and egg problem, even though we are
8480 * holding rq->lock, the CPU isn't yet set to this CPU so the
8481 * lockdep check in task_group() will fail.
8483 * Similar case to sched_fork(). / Alternatively we could
8484 * use task_rq_lock() here and obtain the other rq->lock.
8489 __set_task_cpu(idle, cpu);
8493 rcu_assign_pointer(rq->curr, idle);
8494 idle->on_rq = TASK_ON_RQ_QUEUED;
8498 raw_spin_rq_unlock(rq);
8499 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
8501 /* Set the preempt count _outside_ the spinlocks! */
8502 init_idle_preempt_count(idle, cpu);
8505 * The idle tasks have their own, simple scheduling class:
8507 idle->sched_class = &idle_sched_class;
8508 ftrace_graph_init_idle_task(idle, cpu);
8509 vtime_init_idle(idle, cpu);
8511 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
8517 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
8518 const struct cpumask *trial)
8522 if (!cpumask_weight(cur))
8525 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
8530 int task_can_attach(struct task_struct *p,
8531 const struct cpumask *cs_cpus_allowed)
8536 * Kthreads which disallow setaffinity shouldn't be moved
8537 * to a new cpuset; we don't want to change their CPU
8538 * affinity and isolating such threads by their set of
8539 * allowed nodes is unnecessary. Thus, cpusets are not
8540 * applicable for such threads. This prevents checking for
8541 * success of set_cpus_allowed_ptr() on all attached tasks
8542 * before cpus_mask may be changed.
8544 if (p->flags & PF_NO_SETAFFINITY) {
8549 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
8551 ret = dl_task_can_attach(p, cs_cpus_allowed);
8557 bool sched_smp_initialized __read_mostly;
8559 #ifdef CONFIG_NUMA_BALANCING
8560 /* Migrate current task p to target_cpu */
8561 int migrate_task_to(struct task_struct *p, int target_cpu)
8563 struct migration_arg arg = { p, target_cpu };
8564 int curr_cpu = task_cpu(p);
8566 if (curr_cpu == target_cpu)
8569 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
8572 /* TODO: This is not properly updating schedstats */
8574 trace_sched_move_numa(p, curr_cpu, target_cpu);
8575 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
8579 * Requeue a task on a given node and accurately track the number of NUMA
8580 * tasks on the runqueues
8582 void sched_setnuma(struct task_struct *p, int nid)
8584 bool queued, running;
8588 rq = task_rq_lock(p, &rf);
8589 queued = task_on_rq_queued(p);
8590 running = task_current(rq, p);
8593 dequeue_task(rq, p, DEQUEUE_SAVE);
8595 put_prev_task(rq, p);
8597 p->numa_preferred_nid = nid;
8600 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
8602 set_next_task(rq, p);
8603 task_rq_unlock(rq, p, &rf);
8605 #endif /* CONFIG_NUMA_BALANCING */
8607 #ifdef CONFIG_HOTPLUG_CPU
8609 * Ensure that the idle task is using init_mm right before its CPU goes
8612 void idle_task_exit(void)
8614 struct mm_struct *mm = current->active_mm;
8616 BUG_ON(cpu_online(smp_processor_id()));
8617 BUG_ON(current != this_rq()->idle);
8619 if (mm != &init_mm) {
8620 switch_mm(mm, &init_mm, current);
8621 finish_arch_post_lock_switch();
8624 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
8627 static int __balance_push_cpu_stop(void *arg)
8629 struct task_struct *p = arg;
8630 struct rq *rq = this_rq();
8634 raw_spin_lock_irq(&p->pi_lock);
8637 update_rq_clock(rq);
8639 if (task_rq(p) == rq && task_on_rq_queued(p)) {
8640 cpu = select_fallback_rq(rq->cpu, p);
8641 rq = __migrate_task(rq, &rf, p, cpu);
8645 raw_spin_unlock_irq(&p->pi_lock);
8652 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
8655 * Ensure we only run per-cpu kthreads once the CPU goes !active.
8657 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
8658 * effective when the hotplug motion is down.
8660 static void balance_push(struct rq *rq)
8662 struct task_struct *push_task = rq->curr;
8664 lockdep_assert_rq_held(rq);
8665 SCHED_WARN_ON(rq->cpu != smp_processor_id());
8668 * Ensure the thing is persistent until balance_push_set(.on = false);
8670 rq->balance_callback = &balance_push_callback;
8673 * Only active while going offline.
8675 if (!cpu_dying(rq->cpu))
8679 * Both the cpu-hotplug and stop task are in this case and are
8680 * required to complete the hotplug process.
8682 if (kthread_is_per_cpu(push_task) ||
8683 is_migration_disabled(push_task)) {
8686 * If this is the idle task on the outgoing CPU try to wake
8687 * up the hotplug control thread which might wait for the
8688 * last task to vanish. The rcuwait_active() check is
8689 * accurate here because the waiter is pinned on this CPU
8690 * and can't obviously be running in parallel.
8692 * On RT kernels this also has to check whether there are
8693 * pinned and scheduled out tasks on the runqueue. They
8694 * need to leave the migrate disabled section first.
8696 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
8697 rcuwait_active(&rq->hotplug_wait)) {
8698 raw_spin_rq_unlock(rq);
8699 rcuwait_wake_up(&rq->hotplug_wait);
8700 raw_spin_rq_lock(rq);
8705 get_task_struct(push_task);
8707 * Temporarily drop rq->lock such that we can wake-up the stop task.
8708 * Both preemption and IRQs are still disabled.
8710 raw_spin_rq_unlock(rq);
8711 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
8712 this_cpu_ptr(&push_work));
8714 * At this point need_resched() is true and we'll take the loop in
8715 * schedule(). The next pick is obviously going to be the stop task
8716 * which kthread_is_per_cpu() and will push this task away.
8718 raw_spin_rq_lock(rq);
8721 static void balance_push_set(int cpu, bool on)
8723 struct rq *rq = cpu_rq(cpu);
8726 rq_lock_irqsave(rq, &rf);
8728 WARN_ON_ONCE(rq->balance_callback);
8729 rq->balance_callback = &balance_push_callback;
8730 } else if (rq->balance_callback == &balance_push_callback) {
8731 rq->balance_callback = NULL;
8733 rq_unlock_irqrestore(rq, &rf);
8737 * Invoked from a CPUs hotplug control thread after the CPU has been marked
8738 * inactive. All tasks which are not per CPU kernel threads are either
8739 * pushed off this CPU now via balance_push() or placed on a different CPU
8740 * during wakeup. Wait until the CPU is quiescent.
8742 static void balance_hotplug_wait(void)
8744 struct rq *rq = this_rq();
8746 rcuwait_wait_event(&rq->hotplug_wait,
8747 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
8748 TASK_UNINTERRUPTIBLE);
8753 static inline void balance_push(struct rq *rq)
8757 static inline void balance_push_set(int cpu, bool on)
8761 static inline void balance_hotplug_wait(void)
8765 #endif /* CONFIG_HOTPLUG_CPU */
8767 void set_rq_online(struct rq *rq)
8770 const struct sched_class *class;
8772 cpumask_set_cpu(rq->cpu, rq->rd->online);
8775 for_each_class(class) {
8776 if (class->rq_online)
8777 class->rq_online(rq);
8782 void set_rq_offline(struct rq *rq)
8785 const struct sched_class *class;
8787 for_each_class(class) {
8788 if (class->rq_offline)
8789 class->rq_offline(rq);
8792 cpumask_clear_cpu(rq->cpu, rq->rd->online);
8798 * used to mark begin/end of suspend/resume:
8800 static int num_cpus_frozen;
8803 * Update cpusets according to cpu_active mask. If cpusets are
8804 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8805 * around partition_sched_domains().
8807 * If we come here as part of a suspend/resume, don't touch cpusets because we
8808 * want to restore it back to its original state upon resume anyway.
8810 static void cpuset_cpu_active(void)
8812 if (cpuhp_tasks_frozen) {
8814 * num_cpus_frozen tracks how many CPUs are involved in suspend
8815 * resume sequence. As long as this is not the last online
8816 * operation in the resume sequence, just build a single sched
8817 * domain, ignoring cpusets.
8819 partition_sched_domains(1, NULL, NULL);
8820 if (--num_cpus_frozen)
8823 * This is the last CPU online operation. So fall through and
8824 * restore the original sched domains by considering the
8825 * cpuset configurations.
8827 cpuset_force_rebuild();
8829 cpuset_update_active_cpus();
8832 static int cpuset_cpu_inactive(unsigned int cpu)
8834 if (!cpuhp_tasks_frozen) {
8835 if (dl_cpu_busy(cpu))
8837 cpuset_update_active_cpus();
8840 partition_sched_domains(1, NULL, NULL);
8845 int sched_cpu_activate(unsigned int cpu)
8847 struct rq *rq = cpu_rq(cpu);
8851 * Clear the balance_push callback and prepare to schedule
8854 balance_push_set(cpu, false);
8856 #ifdef CONFIG_SCHED_SMT
8858 * When going up, increment the number of cores with SMT present.
8860 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8861 static_branch_inc_cpuslocked(&sched_smt_present);
8863 set_cpu_active(cpu, true);
8865 if (sched_smp_initialized) {
8866 sched_domains_numa_masks_set(cpu);
8867 cpuset_cpu_active();
8871 * Put the rq online, if not already. This happens:
8873 * 1) In the early boot process, because we build the real domains
8874 * after all CPUs have been brought up.
8876 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
8879 rq_lock_irqsave(rq, &rf);
8881 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8884 rq_unlock_irqrestore(rq, &rf);
8889 int sched_cpu_deactivate(unsigned int cpu)
8891 struct rq *rq = cpu_rq(cpu);
8896 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
8897 * load balancing when not active
8899 nohz_balance_exit_idle(rq);
8901 set_cpu_active(cpu, false);
8904 * From this point forward, this CPU will refuse to run any task that
8905 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
8906 * push those tasks away until this gets cleared, see
8907 * sched_cpu_dying().
8909 balance_push_set(cpu, true);
8912 * We've cleared cpu_active_mask / set balance_push, wait for all
8913 * preempt-disabled and RCU users of this state to go away such that
8914 * all new such users will observe it.
8916 * Specifically, we rely on ttwu to no longer target this CPU, see
8917 * ttwu_queue_cond() and is_cpu_allowed().
8919 * Do sync before park smpboot threads to take care the rcu boost case.
8923 rq_lock_irqsave(rq, &rf);
8925 update_rq_clock(rq);
8926 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8929 rq_unlock_irqrestore(rq, &rf);
8931 #ifdef CONFIG_SCHED_SMT
8933 * When going down, decrement the number of cores with SMT present.
8935 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
8936 static_branch_dec_cpuslocked(&sched_smt_present);
8939 if (!sched_smp_initialized)
8942 ret = cpuset_cpu_inactive(cpu);
8944 balance_push_set(cpu, false);
8945 set_cpu_active(cpu, true);
8948 sched_domains_numa_masks_clear(cpu);
8952 static void sched_rq_cpu_starting(unsigned int cpu)
8954 struct rq *rq = cpu_rq(cpu);
8956 rq->calc_load_update = calc_load_update;
8957 update_max_interval();
8960 int sched_cpu_starting(unsigned int cpu)
8962 sched_core_cpu_starting(cpu);
8963 sched_rq_cpu_starting(cpu);
8964 sched_tick_start(cpu);
8968 #ifdef CONFIG_HOTPLUG_CPU
8971 * Invoked immediately before the stopper thread is invoked to bring the
8972 * CPU down completely. At this point all per CPU kthreads except the
8973 * hotplug thread (current) and the stopper thread (inactive) have been
8974 * either parked or have been unbound from the outgoing CPU. Ensure that
8975 * any of those which might be on the way out are gone.
8977 * If after this point a bound task is being woken on this CPU then the
8978 * responsible hotplug callback has failed to do it's job.
8979 * sched_cpu_dying() will catch it with the appropriate fireworks.
8981 int sched_cpu_wait_empty(unsigned int cpu)
8983 balance_hotplug_wait();
8988 * Since this CPU is going 'away' for a while, fold any nr_active delta we
8989 * might have. Called from the CPU stopper task after ensuring that the
8990 * stopper is the last running task on the CPU, so nr_active count is
8991 * stable. We need to take the teardown thread which is calling this into
8992 * account, so we hand in adjust = 1 to the load calculation.
8994 * Also see the comment "Global load-average calculations".
8996 static void calc_load_migrate(struct rq *rq)
8998 long delta = calc_load_fold_active(rq, 1);
9001 atomic_long_add(delta, &calc_load_tasks);
9004 static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9006 struct task_struct *g, *p;
9007 int cpu = cpu_of(rq);
9009 lockdep_assert_rq_held(rq);
9011 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9012 for_each_process_thread(g, p) {
9013 if (task_cpu(p) != cpu)
9016 if (!task_on_rq_queued(p))
9019 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9023 int sched_cpu_dying(unsigned int cpu)
9025 struct rq *rq = cpu_rq(cpu);
9028 /* Handle pending wakeups and then migrate everything off */
9029 sched_tick_stop(cpu);
9031 rq_lock_irqsave(rq, &rf);
9032 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9033 WARN(true, "Dying CPU not properly vacated!");
9034 dump_rq_tasks(rq, KERN_WARNING);
9036 rq_unlock_irqrestore(rq, &rf);
9038 calc_load_migrate(rq);
9039 update_max_interval();
9045 void __init sched_init_smp(void)
9050 * There's no userspace yet to cause hotplug operations; hence all the
9051 * CPU masks are stable and all blatant races in the below code cannot
9054 mutex_lock(&sched_domains_mutex);
9055 sched_init_domains(cpu_active_mask);
9056 mutex_unlock(&sched_domains_mutex);
9058 /* Move init over to a non-isolated CPU */
9059 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
9061 current->flags &= ~PF_NO_SETAFFINITY;
9062 sched_init_granularity();
9064 init_sched_rt_class();
9065 init_sched_dl_class();
9067 sched_smp_initialized = true;
9070 static int __init migration_init(void)
9072 sched_cpu_starting(smp_processor_id());
9075 early_initcall(migration_init);
9078 void __init sched_init_smp(void)
9080 sched_init_granularity();
9082 #endif /* CONFIG_SMP */
9084 int in_sched_functions(unsigned long addr)
9086 return in_lock_functions(addr) ||
9087 (addr >= (unsigned long)__sched_text_start
9088 && addr < (unsigned long)__sched_text_end);
9091 #ifdef CONFIG_CGROUP_SCHED
9093 * Default task group.
9094 * Every task in system belongs to this group at bootup.
9096 struct task_group root_task_group;
9097 LIST_HEAD(task_groups);
9099 /* Cacheline aligned slab cache for task_group */
9100 static struct kmem_cache *task_group_cache __read_mostly;
9103 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
9104 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
9106 void __init sched_init(void)
9108 unsigned long ptr = 0;
9111 /* Make sure the linker didn't screw up */
9112 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
9113 &fair_sched_class + 1 != &rt_sched_class ||
9114 &rt_sched_class + 1 != &dl_sched_class);
9116 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
9121 #ifdef CONFIG_FAIR_GROUP_SCHED
9122 ptr += 2 * nr_cpu_ids * sizeof(void **);
9124 #ifdef CONFIG_RT_GROUP_SCHED
9125 ptr += 2 * nr_cpu_ids * sizeof(void **);
9128 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
9130 #ifdef CONFIG_FAIR_GROUP_SCHED
9131 root_task_group.se = (struct sched_entity **)ptr;
9132 ptr += nr_cpu_ids * sizeof(void **);
9134 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9135 ptr += nr_cpu_ids * sizeof(void **);
9137 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9138 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
9139 #endif /* CONFIG_FAIR_GROUP_SCHED */
9140 #ifdef CONFIG_RT_GROUP_SCHED
9141 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9142 ptr += nr_cpu_ids * sizeof(void **);
9144 root_task_group.rt_rq = (struct rt_rq **)ptr;
9145 ptr += nr_cpu_ids * sizeof(void **);
9147 #endif /* CONFIG_RT_GROUP_SCHED */
9149 #ifdef CONFIG_CPUMASK_OFFSTACK
9150 for_each_possible_cpu(i) {
9151 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
9152 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9153 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
9154 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
9156 #endif /* CONFIG_CPUMASK_OFFSTACK */
9158 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
9159 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
9162 init_defrootdomain();
9165 #ifdef CONFIG_RT_GROUP_SCHED
9166 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9167 global_rt_period(), global_rt_runtime());
9168 #endif /* CONFIG_RT_GROUP_SCHED */
9170 #ifdef CONFIG_CGROUP_SCHED
9171 task_group_cache = KMEM_CACHE(task_group, 0);
9173 list_add(&root_task_group.list, &task_groups);
9174 INIT_LIST_HEAD(&root_task_group.children);
9175 INIT_LIST_HEAD(&root_task_group.siblings);
9176 autogroup_init(&init_task);
9177 #endif /* CONFIG_CGROUP_SCHED */
9179 for_each_possible_cpu(i) {
9183 raw_spin_lock_init(&rq->__lock);
9185 rq->calc_load_active = 0;
9186 rq->calc_load_update = jiffies + LOAD_FREQ;
9187 init_cfs_rq(&rq->cfs);
9188 init_rt_rq(&rq->rt);
9189 init_dl_rq(&rq->dl);
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9191 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9192 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
9194 * How much CPU bandwidth does root_task_group get?
9196 * In case of task-groups formed thr' the cgroup filesystem, it
9197 * gets 100% of the CPU resources in the system. This overall
9198 * system CPU resource is divided among the tasks of
9199 * root_task_group and its child task-groups in a fair manner,
9200 * based on each entity's (task or task-group's) weight
9201 * (se->load.weight).
9203 * In other words, if root_task_group has 10 tasks of weight
9204 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9205 * then A0's share of the CPU resource is:
9207 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9209 * We achieve this by letting root_task_group's tasks sit
9210 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
9212 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
9213 #endif /* CONFIG_FAIR_GROUP_SCHED */
9215 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9216 #ifdef CONFIG_RT_GROUP_SCHED
9217 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
9222 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
9223 rq->balance_callback = &balance_push_callback;
9224 rq->active_balance = 0;
9225 rq->next_balance = jiffies;
9230 rq->avg_idle = 2*sysctl_sched_migration_cost;
9231 rq->wake_stamp = jiffies;
9232 rq->wake_avg_idle = rq->avg_idle;
9233 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
9235 INIT_LIST_HEAD(&rq->cfs_tasks);
9237 rq_attach_root(rq, &def_root_domain);
9238 #ifdef CONFIG_NO_HZ_COMMON
9239 rq->last_blocked_load_update_tick = jiffies;
9240 atomic_set(&rq->nohz_flags, 0);
9242 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
9244 #ifdef CONFIG_HOTPLUG_CPU
9245 rcuwait_init(&rq->hotplug_wait);
9247 #endif /* CONFIG_SMP */
9249 atomic_set(&rq->nr_iowait, 0);
9251 #ifdef CONFIG_SCHED_CORE
9253 rq->core_pick = NULL;
9254 rq->core_enabled = 0;
9255 rq->core_tree = RB_ROOT;
9256 rq->core_forceidle = false;
9258 rq->core_cookie = 0UL;
9262 set_load_weight(&init_task, false);
9265 * The boot idle thread does lazy MMU switching as well:
9268 enter_lazy_tlb(&init_mm, current);
9271 * Make us the idle thread. Technically, schedule() should not be
9272 * called from this thread, however somewhere below it might be,
9273 * but because we are the idle thread, we just pick up running again
9274 * when this runqueue becomes "idle".
9276 init_idle(current, smp_processor_id());
9278 calc_load_update = jiffies + LOAD_FREQ;
9281 idle_thread_set_boot_cpu();
9282 balance_push_set(smp_processor_id(), false);
9284 init_sched_fair_class();
9290 scheduler_running = 1;
9293 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
9294 static inline int preempt_count_equals(int preempt_offset)
9296 int nested = preempt_count() + rcu_preempt_depth();
9298 return (nested == preempt_offset);
9301 void __might_sleep(const char *file, int line, int preempt_offset)
9303 unsigned int state = get_current_state();
9305 * Blocking primitives will set (and therefore destroy) current->state,
9306 * since we will exit with TASK_RUNNING make sure we enter with it,
9307 * otherwise we will destroy state.
9309 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
9310 "do not call blocking ops when !TASK_RUNNING; "
9311 "state=%x set at [<%p>] %pS\n", state,
9312 (void *)current->task_state_change,
9313 (void *)current->task_state_change);
9315 ___might_sleep(file, line, preempt_offset);
9317 EXPORT_SYMBOL(__might_sleep);
9319 void ___might_sleep(const char *file, int line, int preempt_offset)
9321 /* Ratelimiting timestamp: */
9322 static unsigned long prev_jiffy;
9324 unsigned long preempt_disable_ip;
9326 /* WARN_ON_ONCE() by default, no rate limit required: */
9329 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
9330 !is_idle_task(current) && !current->non_block_count) ||
9331 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
9335 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9337 prev_jiffy = jiffies;
9339 /* Save this before calling printk(), since that will clobber it: */
9340 preempt_disable_ip = get_preempt_disable_ip(current);
9343 "BUG: sleeping function called from invalid context at %s:%d\n",
9346 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
9347 in_atomic(), irqs_disabled(), current->non_block_count,
9348 current->pid, current->comm);
9350 if (task_stack_end_corrupted(current))
9351 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
9353 debug_show_held_locks(current);
9354 if (irqs_disabled())
9355 print_irqtrace_events(current);
9356 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
9357 && !preempt_count_equals(preempt_offset)) {
9358 pr_err("Preemption disabled at:");
9359 print_ip_sym(KERN_ERR, preempt_disable_ip);
9362 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9364 EXPORT_SYMBOL(___might_sleep);
9366 void __cant_sleep(const char *file, int line, int preempt_offset)
9368 static unsigned long prev_jiffy;
9370 if (irqs_disabled())
9373 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9376 if (preempt_count() > preempt_offset)
9379 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9381 prev_jiffy = jiffies;
9383 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
9384 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9385 in_atomic(), irqs_disabled(),
9386 current->pid, current->comm);
9388 debug_show_held_locks(current);
9390 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9392 EXPORT_SYMBOL_GPL(__cant_sleep);
9395 void __cant_migrate(const char *file, int line)
9397 static unsigned long prev_jiffy;
9399 if (irqs_disabled())
9402 if (is_migration_disabled(current))
9405 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
9408 if (preempt_count() > 0)
9411 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9413 prev_jiffy = jiffies;
9415 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
9416 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
9417 in_atomic(), irqs_disabled(), is_migration_disabled(current),
9418 current->pid, current->comm);
9420 debug_show_held_locks(current);
9422 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
9424 EXPORT_SYMBOL_GPL(__cant_migrate);
9428 #ifdef CONFIG_MAGIC_SYSRQ
9429 void normalize_rt_tasks(void)
9431 struct task_struct *g, *p;
9432 struct sched_attr attr = {
9433 .sched_policy = SCHED_NORMAL,
9436 read_lock(&tasklist_lock);
9437 for_each_process_thread(g, p) {
9439 * Only normalize user tasks:
9441 if (p->flags & PF_KTHREAD)
9444 p->se.exec_start = 0;
9445 schedstat_set(p->se.statistics.wait_start, 0);
9446 schedstat_set(p->se.statistics.sleep_start, 0);
9447 schedstat_set(p->se.statistics.block_start, 0);
9449 if (!dl_task(p) && !rt_task(p)) {
9451 * Renice negative nice level userspace
9454 if (task_nice(p) < 0)
9455 set_user_nice(p, 0);
9459 __sched_setscheduler(p, &attr, false, false);
9461 read_unlock(&tasklist_lock);
9464 #endif /* CONFIG_MAGIC_SYSRQ */
9466 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
9468 * These functions are only useful for the IA64 MCA handling, or kdb.
9470 * They can only be called when the whole system has been
9471 * stopped - every CPU needs to be quiescent, and no scheduling
9472 * activity can take place. Using them for anything else would
9473 * be a serious bug, and as a result, they aren't even visible
9474 * under any other configuration.
9478 * curr_task - return the current task for a given CPU.
9479 * @cpu: the processor in question.
9481 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9483 * Return: The current task for @cpu.
9485 struct task_struct *curr_task(int cpu)
9487 return cpu_curr(cpu);
9490 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
9494 * ia64_set_curr_task - set the current task for a given CPU.
9495 * @cpu: the processor in question.
9496 * @p: the task pointer to set.
9498 * Description: This function must only be used when non-maskable interrupts
9499 * are serviced on a separate stack. It allows the architecture to switch the
9500 * notion of the current task on a CPU in a non-blocking manner. This function
9501 * must be called with all CPU's synchronized, and interrupts disabled, the
9502 * and caller must save the original value of the current task (see
9503 * curr_task() above) and restore that value before reenabling interrupts and
9504 * re-starting the system.
9506 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9508 void ia64_set_curr_task(int cpu, struct task_struct *p)
9515 #ifdef CONFIG_CGROUP_SCHED
9516 /* task_group_lock serializes the addition/removal of task groups */
9517 static DEFINE_SPINLOCK(task_group_lock);
9519 static inline void alloc_uclamp_sched_group(struct task_group *tg,
9520 struct task_group *parent)
9522 #ifdef CONFIG_UCLAMP_TASK_GROUP
9523 enum uclamp_id clamp_id;
9525 for_each_clamp_id(clamp_id) {
9526 uclamp_se_set(&tg->uclamp_req[clamp_id],
9527 uclamp_none(clamp_id), false);
9528 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
9533 static void sched_free_group(struct task_group *tg)
9535 free_fair_sched_group(tg);
9536 free_rt_sched_group(tg);
9538 kmem_cache_free(task_group_cache, tg);
9541 /* allocate runqueue etc for a new task group */
9542 struct task_group *sched_create_group(struct task_group *parent)
9544 struct task_group *tg;
9546 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
9548 return ERR_PTR(-ENOMEM);
9550 if (!alloc_fair_sched_group(tg, parent))
9553 if (!alloc_rt_sched_group(tg, parent))
9556 alloc_uclamp_sched_group(tg, parent);
9561 sched_free_group(tg);
9562 return ERR_PTR(-ENOMEM);
9565 void sched_online_group(struct task_group *tg, struct task_group *parent)
9567 unsigned long flags;
9569 spin_lock_irqsave(&task_group_lock, flags);
9570 list_add_rcu(&tg->list, &task_groups);
9572 /* Root should already exist: */
9575 tg->parent = parent;
9576 INIT_LIST_HEAD(&tg->children);
9577 list_add_rcu(&tg->siblings, &parent->children);
9578 spin_unlock_irqrestore(&task_group_lock, flags);
9580 online_fair_sched_group(tg);
9583 /* rcu callback to free various structures associated with a task group */
9584 static void sched_free_group_rcu(struct rcu_head *rhp)
9586 /* Now it should be safe to free those cfs_rqs: */
9587 sched_free_group(container_of(rhp, struct task_group, rcu));
9590 void sched_destroy_group(struct task_group *tg)
9592 /* Wait for possible concurrent references to cfs_rqs complete: */
9593 call_rcu(&tg->rcu, sched_free_group_rcu);
9596 void sched_offline_group(struct task_group *tg)
9598 unsigned long flags;
9600 /* End participation in shares distribution: */
9601 unregister_fair_sched_group(tg);
9603 spin_lock_irqsave(&task_group_lock, flags);
9604 list_del_rcu(&tg->list);
9605 list_del_rcu(&tg->siblings);
9606 spin_unlock_irqrestore(&task_group_lock, flags);
9609 static void sched_change_group(struct task_struct *tsk, int type)
9611 struct task_group *tg;
9614 * All callers are synchronized by task_rq_lock(); we do not use RCU
9615 * which is pointless here. Thus, we pass "true" to task_css_check()
9616 * to prevent lockdep warnings.
9618 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
9619 struct task_group, css);
9620 tg = autogroup_task_group(tsk, tg);
9621 tsk->sched_task_group = tg;
9623 #ifdef CONFIG_FAIR_GROUP_SCHED
9624 if (tsk->sched_class->task_change_group)
9625 tsk->sched_class->task_change_group(tsk, type);
9628 set_task_rq(tsk, task_cpu(tsk));
9632 * Change task's runqueue when it moves between groups.
9634 * The caller of this function should have put the task in its new group by
9635 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
9638 void sched_move_task(struct task_struct *tsk)
9640 int queued, running, queue_flags =
9641 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
9645 rq = task_rq_lock(tsk, &rf);
9646 update_rq_clock(rq);
9648 running = task_current(rq, tsk);
9649 queued = task_on_rq_queued(tsk);
9652 dequeue_task(rq, tsk, queue_flags);
9654 put_prev_task(rq, tsk);
9656 sched_change_group(tsk, TASK_MOVE_GROUP);
9659 enqueue_task(rq, tsk, queue_flags);
9661 set_next_task(rq, tsk);
9663 * After changing group, the running task may have joined a
9664 * throttled one but it's still the running task. Trigger a
9665 * resched to make sure that task can still run.
9670 task_rq_unlock(rq, tsk, &rf);
9673 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
9675 return css ? container_of(css, struct task_group, css) : NULL;
9678 static struct cgroup_subsys_state *
9679 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9681 struct task_group *parent = css_tg(parent_css);
9682 struct task_group *tg;
9685 /* This is early initialization for the top cgroup */
9686 return &root_task_group.css;
9689 tg = sched_create_group(parent);
9691 return ERR_PTR(-ENOMEM);
9696 /* Expose task group only after completing cgroup initialization */
9697 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
9699 struct task_group *tg = css_tg(css);
9700 struct task_group *parent = css_tg(css->parent);
9703 sched_online_group(tg, parent);
9705 #ifdef CONFIG_UCLAMP_TASK_GROUP
9706 /* Propagate the effective uclamp value for the new group */
9707 mutex_lock(&uclamp_mutex);
9709 cpu_util_update_eff(css);
9711 mutex_unlock(&uclamp_mutex);
9717 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
9719 struct task_group *tg = css_tg(css);
9721 sched_offline_group(tg);
9724 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
9726 struct task_group *tg = css_tg(css);
9729 * Relies on the RCU grace period between css_released() and this.
9731 sched_free_group(tg);
9735 * This is called before wake_up_new_task(), therefore we really only
9736 * have to set its group bits, all the other stuff does not apply.
9738 static void cpu_cgroup_fork(struct task_struct *task)
9743 rq = task_rq_lock(task, &rf);
9745 update_rq_clock(rq);
9746 sched_change_group(task, TASK_SET_GROUP);
9748 task_rq_unlock(rq, task, &rf);
9751 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
9753 struct task_struct *task;
9754 struct cgroup_subsys_state *css;
9757 cgroup_taskset_for_each(task, css, tset) {
9758 #ifdef CONFIG_RT_GROUP_SCHED
9759 if (!sched_rt_can_attach(css_tg(css), task))
9763 * Serialize against wake_up_new_task() such that if it's
9764 * running, we're sure to observe its full state.
9766 raw_spin_lock_irq(&task->pi_lock);
9768 * Avoid calling sched_move_task() before wake_up_new_task()
9769 * has happened. This would lead to problems with PELT, due to
9770 * move wanting to detach+attach while we're not attached yet.
9772 if (READ_ONCE(task->__state) == TASK_NEW)
9774 raw_spin_unlock_irq(&task->pi_lock);
9782 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
9784 struct task_struct *task;
9785 struct cgroup_subsys_state *css;
9787 cgroup_taskset_for_each(task, css, tset)
9788 sched_move_task(task);
9791 #ifdef CONFIG_UCLAMP_TASK_GROUP
9792 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
9794 struct cgroup_subsys_state *top_css = css;
9795 struct uclamp_se *uc_parent = NULL;
9796 struct uclamp_se *uc_se = NULL;
9797 unsigned int eff[UCLAMP_CNT];
9798 enum uclamp_id clamp_id;
9799 unsigned int clamps;
9801 lockdep_assert_held(&uclamp_mutex);
9802 SCHED_WARN_ON(!rcu_read_lock_held());
9804 css_for_each_descendant_pre(css, top_css) {
9805 uc_parent = css_tg(css)->parent
9806 ? css_tg(css)->parent->uclamp : NULL;
9808 for_each_clamp_id(clamp_id) {
9809 /* Assume effective clamps matches requested clamps */
9810 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
9811 /* Cap effective clamps with parent's effective clamps */
9813 eff[clamp_id] > uc_parent[clamp_id].value) {
9814 eff[clamp_id] = uc_parent[clamp_id].value;
9817 /* Ensure protection is always capped by limit */
9818 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
9820 /* Propagate most restrictive effective clamps */
9822 uc_se = css_tg(css)->uclamp;
9823 for_each_clamp_id(clamp_id) {
9824 if (eff[clamp_id] == uc_se[clamp_id].value)
9826 uc_se[clamp_id].value = eff[clamp_id];
9827 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
9828 clamps |= (0x1 << clamp_id);
9831 css = css_rightmost_descendant(css);
9835 /* Immediately update descendants RUNNABLE tasks */
9836 uclamp_update_active_tasks(css);
9841 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
9842 * C expression. Since there is no way to convert a macro argument (N) into a
9843 * character constant, use two levels of macros.
9845 #define _POW10(exp) ((unsigned int)1e##exp)
9846 #define POW10(exp) _POW10(exp)
9848 struct uclamp_request {
9849 #define UCLAMP_PERCENT_SHIFT 2
9850 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
9856 static inline struct uclamp_request
9857 capacity_from_percent(char *buf)
9859 struct uclamp_request req = {
9860 .percent = UCLAMP_PERCENT_SCALE,
9861 .util = SCHED_CAPACITY_SCALE,
9866 if (strcmp(buf, "max")) {
9867 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
9871 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
9876 req.util = req.percent << SCHED_CAPACITY_SHIFT;
9877 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
9883 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
9884 size_t nbytes, loff_t off,
9885 enum uclamp_id clamp_id)
9887 struct uclamp_request req;
9888 struct task_group *tg;
9890 req = capacity_from_percent(buf);
9894 static_branch_enable(&sched_uclamp_used);
9896 mutex_lock(&uclamp_mutex);
9899 tg = css_tg(of_css(of));
9900 if (tg->uclamp_req[clamp_id].value != req.util)
9901 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
9904 * Because of not recoverable conversion rounding we keep track of the
9905 * exact requested value
9907 tg->uclamp_pct[clamp_id] = req.percent;
9909 /* Update effective clamps to track the most restrictive value */
9910 cpu_util_update_eff(of_css(of));
9913 mutex_unlock(&uclamp_mutex);
9918 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
9919 char *buf, size_t nbytes,
9922 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
9925 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
9926 char *buf, size_t nbytes,
9929 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
9932 static inline void cpu_uclamp_print(struct seq_file *sf,
9933 enum uclamp_id clamp_id)
9935 struct task_group *tg;
9941 tg = css_tg(seq_css(sf));
9942 util_clamp = tg->uclamp_req[clamp_id].value;
9945 if (util_clamp == SCHED_CAPACITY_SCALE) {
9946 seq_puts(sf, "max\n");
9950 percent = tg->uclamp_pct[clamp_id];
9951 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
9952 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
9955 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
9957 cpu_uclamp_print(sf, UCLAMP_MIN);
9961 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
9963 cpu_uclamp_print(sf, UCLAMP_MAX);
9966 #endif /* CONFIG_UCLAMP_TASK_GROUP */
9968 #ifdef CONFIG_FAIR_GROUP_SCHED
9969 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
9970 struct cftype *cftype, u64 shareval)
9972 if (shareval > scale_load_down(ULONG_MAX))
9973 shareval = MAX_SHARES;
9974 return sched_group_set_shares(css_tg(css), scale_load(shareval));
9977 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
9980 struct task_group *tg = css_tg(css);
9982 return (u64) scale_load_down(tg->shares);
9985 #ifdef CONFIG_CFS_BANDWIDTH
9986 static DEFINE_MUTEX(cfs_constraints_mutex);
9988 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9989 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9990 /* More than 203 days if BW_SHIFT equals 20. */
9991 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
9993 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9995 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
9998 int i, ret = 0, runtime_enabled, runtime_was_enabled;
9999 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10001 if (tg == &root_task_group)
10005 * Ensure we have at some amount of bandwidth every period. This is
10006 * to prevent reaching a state of large arrears when throttled via
10007 * entity_tick() resulting in prolonged exit starvation.
10009 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10013 * Likewise, bound things on the other side by preventing insane quota
10014 * periods. This also allows us to normalize in computing quota
10017 if (period > max_cfs_quota_period)
10021 * Bound quota to defend quota against overflow during bandwidth shift.
10023 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10026 if (quota != RUNTIME_INF && (burst > quota ||
10027 burst + quota > max_cfs_runtime))
10031 * Prevent race between setting of cfs_rq->runtime_enabled and
10032 * unthrottle_offline_cfs_rqs().
10035 mutex_lock(&cfs_constraints_mutex);
10036 ret = __cfs_schedulable(tg, period, quota);
10040 runtime_enabled = quota != RUNTIME_INF;
10041 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
10043 * If we need to toggle cfs_bandwidth_used, off->on must occur
10044 * before making related changes, and on->off must occur afterwards
10046 if (runtime_enabled && !runtime_was_enabled)
10047 cfs_bandwidth_usage_inc();
10048 raw_spin_lock_irq(&cfs_b->lock);
10049 cfs_b->period = ns_to_ktime(period);
10050 cfs_b->quota = quota;
10051 cfs_b->burst = burst;
10053 __refill_cfs_bandwidth_runtime(cfs_b);
10055 /* Restart the period timer (if active) to handle new period expiry: */
10056 if (runtime_enabled)
10057 start_cfs_bandwidth(cfs_b);
10059 raw_spin_unlock_irq(&cfs_b->lock);
10061 for_each_online_cpu(i) {
10062 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
10063 struct rq *rq = cfs_rq->rq;
10064 struct rq_flags rf;
10066 rq_lock_irq(rq, &rf);
10067 cfs_rq->runtime_enabled = runtime_enabled;
10068 cfs_rq->runtime_remaining = 0;
10070 if (cfs_rq->throttled)
10071 unthrottle_cfs_rq(cfs_rq);
10072 rq_unlock_irq(rq, &rf);
10074 if (runtime_was_enabled && !runtime_enabled)
10075 cfs_bandwidth_usage_dec();
10077 mutex_unlock(&cfs_constraints_mutex);
10078 cpus_read_unlock();
10083 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
10085 u64 quota, period, burst;
10087 period = ktime_to_ns(tg->cfs_bandwidth.period);
10088 burst = tg->cfs_bandwidth.burst;
10089 if (cfs_quota_us < 0)
10090 quota = RUNTIME_INF;
10091 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
10092 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
10096 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10099 static long tg_get_cfs_quota(struct task_group *tg)
10103 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
10106 quota_us = tg->cfs_bandwidth.quota;
10107 do_div(quota_us, NSEC_PER_USEC);
10112 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
10114 u64 quota, period, burst;
10116 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10119 period = (u64)cfs_period_us * NSEC_PER_USEC;
10120 quota = tg->cfs_bandwidth.quota;
10121 burst = tg->cfs_bandwidth.burst;
10123 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10126 static long tg_get_cfs_period(struct task_group *tg)
10130 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
10131 do_div(cfs_period_us, NSEC_PER_USEC);
10133 return cfs_period_us;
10136 static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10138 u64 quota, period, burst;
10140 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10143 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10144 period = ktime_to_ns(tg->cfs_bandwidth.period);
10145 quota = tg->cfs_bandwidth.quota;
10147 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10150 static long tg_get_cfs_burst(struct task_group *tg)
10154 burst_us = tg->cfs_bandwidth.burst;
10155 do_div(burst_us, NSEC_PER_USEC);
10160 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10161 struct cftype *cft)
10163 return tg_get_cfs_quota(css_tg(css));
10166 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10167 struct cftype *cftype, s64 cfs_quota_us)
10169 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
10172 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10173 struct cftype *cft)
10175 return tg_get_cfs_period(css_tg(css));
10178 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10179 struct cftype *cftype, u64 cfs_period_us)
10181 return tg_set_cfs_period(css_tg(css), cfs_period_us);
10184 static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10185 struct cftype *cft)
10187 return tg_get_cfs_burst(css_tg(css));
10190 static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10191 struct cftype *cftype, u64 cfs_burst_us)
10193 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10196 struct cfs_schedulable_data {
10197 struct task_group *tg;
10202 * normalize group quota/period to be quota/max_period
10203 * note: units are usecs
10205 static u64 normalize_cfs_quota(struct task_group *tg,
10206 struct cfs_schedulable_data *d)
10211 period = d->period;
10214 period = tg_get_cfs_period(tg);
10215 quota = tg_get_cfs_quota(tg);
10218 /* note: these should typically be equivalent */
10219 if (quota == RUNTIME_INF || quota == -1)
10220 return RUNTIME_INF;
10222 return to_ratio(period, quota);
10225 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10227 struct cfs_schedulable_data *d = data;
10228 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10229 s64 quota = 0, parent_quota = -1;
10232 quota = RUNTIME_INF;
10234 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
10236 quota = normalize_cfs_quota(tg, d);
10237 parent_quota = parent_b->hierarchical_quota;
10240 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10241 * always take the min. On cgroup1, only inherit when no
10244 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10245 quota = min(quota, parent_quota);
10247 if (quota == RUNTIME_INF)
10248 quota = parent_quota;
10249 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10253 cfs_b->hierarchical_quota = quota;
10258 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10261 struct cfs_schedulable_data data = {
10267 if (quota != RUNTIME_INF) {
10268 do_div(data.period, NSEC_PER_USEC);
10269 do_div(data.quota, NSEC_PER_USEC);
10273 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10279 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
10281 struct task_group *tg = css_tg(seq_css(sf));
10282 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10284 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10285 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10286 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
10288 if (schedstat_enabled() && tg != &root_task_group) {
10292 for_each_possible_cpu(i)
10293 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
10295 seq_printf(sf, "wait_sum %llu\n", ws);
10300 #endif /* CONFIG_CFS_BANDWIDTH */
10301 #endif /* CONFIG_FAIR_GROUP_SCHED */
10303 #ifdef CONFIG_RT_GROUP_SCHED
10304 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10305 struct cftype *cft, s64 val)
10307 return sched_group_set_rt_runtime(css_tg(css), val);
10310 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10311 struct cftype *cft)
10313 return sched_group_rt_runtime(css_tg(css));
10316 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10317 struct cftype *cftype, u64 rt_period_us)
10319 return sched_group_set_rt_period(css_tg(css), rt_period_us);
10322 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
10323 struct cftype *cft)
10325 return sched_group_rt_period(css_tg(css));
10327 #endif /* CONFIG_RT_GROUP_SCHED */
10329 #ifdef CONFIG_FAIR_GROUP_SCHED
10330 static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
10331 struct cftype *cft)
10333 return css_tg(css)->idle;
10336 static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
10337 struct cftype *cft, s64 idle)
10339 return sched_group_set_idle(css_tg(css), idle);
10343 static struct cftype cpu_legacy_files[] = {
10344 #ifdef CONFIG_FAIR_GROUP_SCHED
10347 .read_u64 = cpu_shares_read_u64,
10348 .write_u64 = cpu_shares_write_u64,
10352 .read_s64 = cpu_idle_read_s64,
10353 .write_s64 = cpu_idle_write_s64,
10356 #ifdef CONFIG_CFS_BANDWIDTH
10358 .name = "cfs_quota_us",
10359 .read_s64 = cpu_cfs_quota_read_s64,
10360 .write_s64 = cpu_cfs_quota_write_s64,
10363 .name = "cfs_period_us",
10364 .read_u64 = cpu_cfs_period_read_u64,
10365 .write_u64 = cpu_cfs_period_write_u64,
10368 .name = "cfs_burst_us",
10369 .read_u64 = cpu_cfs_burst_read_u64,
10370 .write_u64 = cpu_cfs_burst_write_u64,
10374 .seq_show = cpu_cfs_stat_show,
10377 #ifdef CONFIG_RT_GROUP_SCHED
10379 .name = "rt_runtime_us",
10380 .read_s64 = cpu_rt_runtime_read,
10381 .write_s64 = cpu_rt_runtime_write,
10384 .name = "rt_period_us",
10385 .read_u64 = cpu_rt_period_read_uint,
10386 .write_u64 = cpu_rt_period_write_uint,
10389 #ifdef CONFIG_UCLAMP_TASK_GROUP
10391 .name = "uclamp.min",
10392 .flags = CFTYPE_NOT_ON_ROOT,
10393 .seq_show = cpu_uclamp_min_show,
10394 .write = cpu_uclamp_min_write,
10397 .name = "uclamp.max",
10398 .flags = CFTYPE_NOT_ON_ROOT,
10399 .seq_show = cpu_uclamp_max_show,
10400 .write = cpu_uclamp_max_write,
10403 { } /* Terminate */
10406 static int cpu_extra_stat_show(struct seq_file *sf,
10407 struct cgroup_subsys_state *css)
10409 #ifdef CONFIG_CFS_BANDWIDTH
10411 struct task_group *tg = css_tg(css);
10412 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
10413 u64 throttled_usec;
10415 throttled_usec = cfs_b->throttled_time;
10416 do_div(throttled_usec, NSEC_PER_USEC);
10418 seq_printf(sf, "nr_periods %d\n"
10419 "nr_throttled %d\n"
10420 "throttled_usec %llu\n",
10421 cfs_b->nr_periods, cfs_b->nr_throttled,
10428 #ifdef CONFIG_FAIR_GROUP_SCHED
10429 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
10430 struct cftype *cft)
10432 struct task_group *tg = css_tg(css);
10433 u64 weight = scale_load_down(tg->shares);
10435 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
10438 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
10439 struct cftype *cft, u64 weight)
10442 * cgroup weight knobs should use the common MIN, DFL and MAX
10443 * values which are 1, 100 and 10000 respectively. While it loses
10444 * a bit of range on both ends, it maps pretty well onto the shares
10445 * value used by scheduler and the round-trip conversions preserve
10446 * the original value over the entire range.
10448 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
10451 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
10453 return sched_group_set_shares(css_tg(css), scale_load(weight));
10456 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
10457 struct cftype *cft)
10459 unsigned long weight = scale_load_down(css_tg(css)->shares);
10460 int last_delta = INT_MAX;
10463 /* find the closest nice value to the current weight */
10464 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
10465 delta = abs(sched_prio_to_weight[prio] - weight);
10466 if (delta >= last_delta)
10468 last_delta = delta;
10471 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
10474 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
10475 struct cftype *cft, s64 nice)
10477 unsigned long weight;
10480 if (nice < MIN_NICE || nice > MAX_NICE)
10483 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
10484 idx = array_index_nospec(idx, 40);
10485 weight = sched_prio_to_weight[idx];
10487 return sched_group_set_shares(css_tg(css), scale_load(weight));
10491 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
10492 long period, long quota)
10495 seq_puts(sf, "max");
10497 seq_printf(sf, "%ld", quota);
10499 seq_printf(sf, " %ld\n", period);
10502 /* caller should put the current value in *@periodp before calling */
10503 static int __maybe_unused cpu_period_quota_parse(char *buf,
10504 u64 *periodp, u64 *quotap)
10506 char tok[21]; /* U64_MAX */
10508 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
10511 *periodp *= NSEC_PER_USEC;
10513 if (sscanf(tok, "%llu", quotap))
10514 *quotap *= NSEC_PER_USEC;
10515 else if (!strcmp(tok, "max"))
10516 *quotap = RUNTIME_INF;
10523 #ifdef CONFIG_CFS_BANDWIDTH
10524 static int cpu_max_show(struct seq_file *sf, void *v)
10526 struct task_group *tg = css_tg(seq_css(sf));
10528 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
10532 static ssize_t cpu_max_write(struct kernfs_open_file *of,
10533 char *buf, size_t nbytes, loff_t off)
10535 struct task_group *tg = css_tg(of_css(of));
10536 u64 period = tg_get_cfs_period(tg);
10537 u64 burst = tg_get_cfs_burst(tg);
10541 ret = cpu_period_quota_parse(buf, &period, "a);
10543 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
10544 return ret ?: nbytes;
10548 static struct cftype cpu_files[] = {
10549 #ifdef CONFIG_FAIR_GROUP_SCHED
10552 .flags = CFTYPE_NOT_ON_ROOT,
10553 .read_u64 = cpu_weight_read_u64,
10554 .write_u64 = cpu_weight_write_u64,
10557 .name = "weight.nice",
10558 .flags = CFTYPE_NOT_ON_ROOT,
10559 .read_s64 = cpu_weight_nice_read_s64,
10560 .write_s64 = cpu_weight_nice_write_s64,
10564 .flags = CFTYPE_NOT_ON_ROOT,
10565 .read_s64 = cpu_idle_read_s64,
10566 .write_s64 = cpu_idle_write_s64,
10569 #ifdef CONFIG_CFS_BANDWIDTH
10572 .flags = CFTYPE_NOT_ON_ROOT,
10573 .seq_show = cpu_max_show,
10574 .write = cpu_max_write,
10577 .name = "max.burst",
10578 .flags = CFTYPE_NOT_ON_ROOT,
10579 .read_u64 = cpu_cfs_burst_read_u64,
10580 .write_u64 = cpu_cfs_burst_write_u64,
10583 #ifdef CONFIG_UCLAMP_TASK_GROUP
10585 .name = "uclamp.min",
10586 .flags = CFTYPE_NOT_ON_ROOT,
10587 .seq_show = cpu_uclamp_min_show,
10588 .write = cpu_uclamp_min_write,
10591 .name = "uclamp.max",
10592 .flags = CFTYPE_NOT_ON_ROOT,
10593 .seq_show = cpu_uclamp_max_show,
10594 .write = cpu_uclamp_max_write,
10597 { } /* terminate */
10600 struct cgroup_subsys cpu_cgrp_subsys = {
10601 .css_alloc = cpu_cgroup_css_alloc,
10602 .css_online = cpu_cgroup_css_online,
10603 .css_released = cpu_cgroup_css_released,
10604 .css_free = cpu_cgroup_css_free,
10605 .css_extra_stat_show = cpu_extra_stat_show,
10606 .fork = cpu_cgroup_fork,
10607 .can_attach = cpu_cgroup_can_attach,
10608 .attach = cpu_cgroup_attach,
10609 .legacy_cftypes = cpu_legacy_files,
10610 .dfl_cftypes = cpu_files,
10611 .early_init = true,
10615 #endif /* CONFIG_CGROUP_SCHED */
10617 void dump_cpu_task(int cpu)
10619 pr_info("Task dump for CPU %d:\n", cpu);
10620 sched_show_task(cpu_curr(cpu));
10624 * Nice levels are multiplicative, with a gentle 10% change for every
10625 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
10626 * nice 1, it will get ~10% less CPU time than another CPU-bound task
10627 * that remained on nice 0.
10629 * The "10% effect" is relative and cumulative: from _any_ nice level,
10630 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
10631 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
10632 * If a task goes up by ~10% and another task goes down by ~10% then
10633 * the relative distance between them is ~25%.)
10635 const int sched_prio_to_weight[40] = {
10636 /* -20 */ 88761, 71755, 56483, 46273, 36291,
10637 /* -15 */ 29154, 23254, 18705, 14949, 11916,
10638 /* -10 */ 9548, 7620, 6100, 4904, 3906,
10639 /* -5 */ 3121, 2501, 1991, 1586, 1277,
10640 /* 0 */ 1024, 820, 655, 526, 423,
10641 /* 5 */ 335, 272, 215, 172, 137,
10642 /* 10 */ 110, 87, 70, 56, 45,
10643 /* 15 */ 36, 29, 23, 18, 15,
10647 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
10649 * In cases where the weight does not change often, we can use the
10650 * precalculated inverse to speed up arithmetics by turning divisions
10651 * into multiplications:
10653 const u32 sched_prio_to_wmult[40] = {
10654 /* -20 */ 48388, 59856, 76040, 92818, 118348,
10655 /* -15 */ 147320, 184698, 229616, 287308, 360437,
10656 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
10657 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
10658 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
10659 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
10660 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
10661 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10664 void call_trace_sched_update_nr_running(struct rq *rq, int count)
10666 trace_sched_update_nr_running_tp(rq, count);