1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
11 #include <linux/nospec.h>
13 #include <linux/kcov.h>
14 #include <linux/scs.h>
16 #include <asm/switch_to.h>
19 #include "../workqueue_internal.h"
20 #include "../../fs/io-wq.h"
21 #include "../smpboot.h"
26 #define CREATE_TRACE_POINTS
27 #include <trace/events/sched.h>
30 * Export tracepoints that act as a bare tracehook (ie: have no trace event
31 * associated with them) to allow external modules to probe them.
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
40 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
42 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
44 * Debugging: various feature bits
46 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
47 * sysctl_sched_features, defined in sched.h, to allow constants propagation
48 * at compile time and compiler optimization based on features default.
50 #define SCHED_FEAT(name, enabled) \
51 (1UL << __SCHED_FEAT_##name) * enabled |
52 const_debug unsigned int sysctl_sched_features =
59 * Number of tasks to iterate in a single balance run.
60 * Limited because this is done with IRQs disabled.
62 const_debug unsigned int sysctl_sched_nr_migrate = 32;
65 * period over which we measure -rt task CPU usage in us.
68 unsigned int sysctl_sched_rt_period = 1000000;
70 __read_mostly int scheduler_running;
73 * part of the period that we allow rt tasks to run in us.
76 int sysctl_sched_rt_runtime = 950000;
79 * __task_rq_lock - lock the rq @p resides on.
81 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
86 lockdep_assert_held(&p->pi_lock);
90 raw_spin_lock(&rq->lock);
91 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
95 raw_spin_unlock(&rq->lock);
97 while (unlikely(task_on_rq_migrating(p)))
103 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
105 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
106 __acquires(p->pi_lock)
112 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
114 raw_spin_lock(&rq->lock);
116 * move_queued_task() task_rq_lock()
119 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
120 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
121 * [S] ->cpu = new_cpu [L] task_rq()
125 * If we observe the old CPU in task_rq_lock(), the acquire of
126 * the old rq->lock will fully serialize against the stores.
128 * If we observe the new CPU in task_rq_lock(), the address
129 * dependency headed by '[L] rq = task_rq()' and the acquire
130 * will pair with the WMB to ensure we then also see migrating.
132 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
136 raw_spin_unlock(&rq->lock);
137 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
139 while (unlikely(task_on_rq_migrating(p)))
145 * RQ-clock updating methods:
148 static void update_rq_clock_task(struct rq *rq, s64 delta)
151 * In theory, the compile should just see 0 here, and optimize out the call
152 * to sched_rt_avg_update. But I don't trust it...
154 s64 __maybe_unused steal = 0, irq_delta = 0;
156 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
157 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
160 * Since irq_time is only updated on {soft,}irq_exit, we might run into
161 * this case when a previous update_rq_clock() happened inside a
164 * When this happens, we stop ->clock_task and only update the
165 * prev_irq_time stamp to account for the part that fit, so that a next
166 * update will consume the rest. This ensures ->clock_task is
169 * It does however cause some slight miss-attribution of {soft,}irq
170 * time, a more accurate solution would be to update the irq_time using
171 * the current rq->clock timestamp, except that would require using
174 if (irq_delta > delta)
177 rq->prev_irq_time += irq_delta;
180 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
181 if (static_key_false((¶virt_steal_rq_enabled))) {
182 steal = paravirt_steal_clock(cpu_of(rq));
183 steal -= rq->prev_steal_time_rq;
185 if (unlikely(steal > delta))
188 rq->prev_steal_time_rq += steal;
193 rq->clock_task += delta;
195 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
196 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
197 update_irq_load_avg(rq, irq_delta + steal);
199 update_rq_clock_pelt(rq, delta);
202 void update_rq_clock(struct rq *rq)
206 lockdep_assert_held(&rq->lock);
208 if (rq->clock_update_flags & RQCF_ACT_SKIP)
211 #ifdef CONFIG_SCHED_DEBUG
212 if (sched_feat(WARN_DOUBLE_CLOCK))
213 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
214 rq->clock_update_flags |= RQCF_UPDATED;
217 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
221 update_rq_clock_task(rq, delta);
225 rq_csd_init(struct rq *rq, call_single_data_t *csd, smp_call_func_t func)
232 #ifdef CONFIG_SCHED_HRTICK
234 * Use HR-timers to deliver accurate preemption points.
237 static void hrtick_clear(struct rq *rq)
239 if (hrtimer_active(&rq->hrtick_timer))
240 hrtimer_cancel(&rq->hrtick_timer);
244 * High-resolution timer tick.
245 * Runs from hardirq context with interrupts disabled.
247 static enum hrtimer_restart hrtick(struct hrtimer *timer)
249 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
252 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
256 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
259 return HRTIMER_NORESTART;
264 static void __hrtick_restart(struct rq *rq)
266 struct hrtimer *timer = &rq->hrtick_timer;
268 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
272 * called from hardirq (IPI) context
274 static void __hrtick_start(void *arg)
280 __hrtick_restart(rq);
285 * Called to set the hrtick timer state.
287 * called with rq->lock held and irqs disabled
289 void hrtick_start(struct rq *rq, u64 delay)
291 struct hrtimer *timer = &rq->hrtick_timer;
296 * Don't schedule slices shorter than 10000ns, that just
297 * doesn't make sense and can cause timer DoS.
299 delta = max_t(s64, delay, 10000LL);
300 time = ktime_add_ns(timer->base->get_time(), delta);
302 hrtimer_set_expires(timer, time);
305 __hrtick_restart(rq);
307 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
312 * Called to set the hrtick timer state.
314 * called with rq->lock held and irqs disabled
316 void hrtick_start(struct rq *rq, u64 delay)
319 * Don't schedule slices shorter than 10000ns, that just
320 * doesn't make sense. Rely on vruntime for fairness.
322 delay = max_t(u64, delay, 10000LL);
323 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
324 HRTIMER_MODE_REL_PINNED_HARD);
327 #endif /* CONFIG_SMP */
329 static void hrtick_rq_init(struct rq *rq)
332 rq_csd_init(rq, &rq->hrtick_csd, __hrtick_start);
334 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
335 rq->hrtick_timer.function = hrtick;
337 #else /* CONFIG_SCHED_HRTICK */
338 static inline void hrtick_clear(struct rq *rq)
342 static inline void hrtick_rq_init(struct rq *rq)
345 #endif /* CONFIG_SCHED_HRTICK */
348 * cmpxchg based fetch_or, macro so it works for different integer types
350 #define fetch_or(ptr, mask) \
352 typeof(ptr) _ptr = (ptr); \
353 typeof(mask) _mask = (mask); \
354 typeof(*_ptr) _old, _val = *_ptr; \
357 _old = cmpxchg(_ptr, _val, _val | _mask); \
365 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
367 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
368 * this avoids any races wrt polling state changes and thereby avoids
371 static bool set_nr_and_not_polling(struct task_struct *p)
373 struct thread_info *ti = task_thread_info(p);
374 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
378 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
380 * If this returns true, then the idle task promises to call
381 * sched_ttwu_pending() and reschedule soon.
383 static bool set_nr_if_polling(struct task_struct *p)
385 struct thread_info *ti = task_thread_info(p);
386 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
389 if (!(val & _TIF_POLLING_NRFLAG))
391 if (val & _TIF_NEED_RESCHED)
393 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
402 static bool set_nr_and_not_polling(struct task_struct *p)
404 set_tsk_need_resched(p);
409 static bool set_nr_if_polling(struct task_struct *p)
416 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
418 struct wake_q_node *node = &task->wake_q;
421 * Atomically grab the task, if ->wake_q is !nil already it means
422 * its already queued (either by us or someone else) and will get the
423 * wakeup due to that.
425 * In order to ensure that a pending wakeup will observe our pending
426 * state, even in the failed case, an explicit smp_mb() must be used.
428 smp_mb__before_atomic();
429 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
433 * The head is context local, there can be no concurrency.
436 head->lastp = &node->next;
441 * wake_q_add() - queue a wakeup for 'later' waking.
442 * @head: the wake_q_head to add @task to
443 * @task: the task to queue for 'later' wakeup
445 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
446 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
449 * This function must be used as-if it were wake_up_process(); IOW the task
450 * must be ready to be woken at this location.
452 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
454 if (__wake_q_add(head, task))
455 get_task_struct(task);
459 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
460 * @head: the wake_q_head to add @task to
461 * @task: the task to queue for 'later' wakeup
463 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
464 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
467 * This function must be used as-if it were wake_up_process(); IOW the task
468 * must be ready to be woken at this location.
470 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
471 * that already hold reference to @task can call the 'safe' version and trust
472 * wake_q to do the right thing depending whether or not the @task is already
475 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
477 if (!__wake_q_add(head, task))
478 put_task_struct(task);
481 void wake_up_q(struct wake_q_head *head)
483 struct wake_q_node *node = head->first;
485 while (node != WAKE_Q_TAIL) {
486 struct task_struct *task;
488 task = container_of(node, struct task_struct, wake_q);
490 /* Task can safely be re-inserted now: */
492 task->wake_q.next = NULL;
495 * wake_up_process() executes a full barrier, which pairs with
496 * the queueing in wake_q_add() so as not to miss wakeups.
498 wake_up_process(task);
499 put_task_struct(task);
504 * resched_curr - mark rq's current task 'to be rescheduled now'.
506 * On UP this means the setting of the need_resched flag, on SMP it
507 * might also involve a cross-CPU call to trigger the scheduler on
510 void resched_curr(struct rq *rq)
512 struct task_struct *curr = rq->curr;
515 lockdep_assert_held(&rq->lock);
517 if (test_tsk_need_resched(curr))
522 if (cpu == smp_processor_id()) {
523 set_tsk_need_resched(curr);
524 set_preempt_need_resched();
528 if (set_nr_and_not_polling(curr))
529 smp_send_reschedule(cpu);
531 trace_sched_wake_idle_without_ipi(cpu);
534 void resched_cpu(int cpu)
536 struct rq *rq = cpu_rq(cpu);
539 raw_spin_lock_irqsave(&rq->lock, flags);
540 if (cpu_online(cpu) || cpu == smp_processor_id())
542 raw_spin_unlock_irqrestore(&rq->lock, flags);
546 #ifdef CONFIG_NO_HZ_COMMON
548 * In the semi idle case, use the nearest busy CPU for migrating timers
549 * from an idle CPU. This is good for power-savings.
551 * We don't do similar optimization for completely idle system, as
552 * selecting an idle CPU will add more delays to the timers than intended
553 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
555 int get_nohz_timer_target(void)
557 int i, cpu = smp_processor_id(), default_cpu = -1;
558 struct sched_domain *sd;
560 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
567 for_each_domain(cpu, sd) {
568 for_each_cpu_and(i, sched_domain_span(sd),
569 housekeeping_cpumask(HK_FLAG_TIMER)) {
580 if (default_cpu == -1)
581 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
589 * When add_timer_on() enqueues a timer into the timer wheel of an
590 * idle CPU then this timer might expire before the next timer event
591 * which is scheduled to wake up that CPU. In case of a completely
592 * idle system the next event might even be infinite time into the
593 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
594 * leaves the inner idle loop so the newly added timer is taken into
595 * account when the CPU goes back to idle and evaluates the timer
596 * wheel for the next timer event.
598 static void wake_up_idle_cpu(int cpu)
600 struct rq *rq = cpu_rq(cpu);
602 if (cpu == smp_processor_id())
605 if (set_nr_and_not_polling(rq->idle))
606 smp_send_reschedule(cpu);
608 trace_sched_wake_idle_without_ipi(cpu);
611 static bool wake_up_full_nohz_cpu(int cpu)
614 * We just need the target to call irq_exit() and re-evaluate
615 * the next tick. The nohz full kick at least implies that.
616 * If needed we can still optimize that later with an
619 if (cpu_is_offline(cpu))
620 return true; /* Don't try to wake offline CPUs. */
621 if (tick_nohz_full_cpu(cpu)) {
622 if (cpu != smp_processor_id() ||
623 tick_nohz_tick_stopped())
624 tick_nohz_full_kick_cpu(cpu);
632 * Wake up the specified CPU. If the CPU is going offline, it is the
633 * caller's responsibility to deal with the lost wakeup, for example,
634 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
636 void wake_up_nohz_cpu(int cpu)
638 if (!wake_up_full_nohz_cpu(cpu))
639 wake_up_idle_cpu(cpu);
642 static void nohz_csd_func(void *info)
644 struct rq *rq = info;
645 int cpu = cpu_of(rq);
649 * Release the rq::nohz_csd.
651 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
652 WARN_ON(!(flags & NOHZ_KICK_MASK));
654 rq->idle_balance = idle_cpu(cpu);
655 if (rq->idle_balance && !need_resched()) {
656 rq->nohz_idle_balance = flags;
657 raise_softirq_irqoff(SCHED_SOFTIRQ);
661 #endif /* CONFIG_NO_HZ_COMMON */
663 #ifdef CONFIG_NO_HZ_FULL
664 bool sched_can_stop_tick(struct rq *rq)
668 /* Deadline tasks, even if single, need the tick */
669 if (rq->dl.dl_nr_running)
673 * If there are more than one RR tasks, we need the tick to effect the
674 * actual RR behaviour.
676 if (rq->rt.rr_nr_running) {
677 if (rq->rt.rr_nr_running == 1)
684 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
685 * forced preemption between FIFO tasks.
687 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
692 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
693 * if there's more than one we need the tick for involuntary
696 if (rq->nr_running > 1)
701 #endif /* CONFIG_NO_HZ_FULL */
702 #endif /* CONFIG_SMP */
704 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
705 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
707 * Iterate task_group tree rooted at *from, calling @down when first entering a
708 * node and @up when leaving it for the final time.
710 * Caller must hold rcu_lock or sufficient equivalent.
712 int walk_tg_tree_from(struct task_group *from,
713 tg_visitor down, tg_visitor up, void *data)
715 struct task_group *parent, *child;
721 ret = (*down)(parent, data);
724 list_for_each_entry_rcu(child, &parent->children, siblings) {
731 ret = (*up)(parent, data);
732 if (ret || parent == from)
736 parent = parent->parent;
743 int tg_nop(struct task_group *tg, void *data)
749 static void set_load_weight(struct task_struct *p, bool update_load)
751 int prio = p->static_prio - MAX_RT_PRIO;
752 struct load_weight *load = &p->se.load;
755 * SCHED_IDLE tasks get minimal weight:
757 if (task_has_idle_policy(p)) {
758 load->weight = scale_load(WEIGHT_IDLEPRIO);
759 load->inv_weight = WMULT_IDLEPRIO;
764 * SCHED_OTHER tasks have to update their load when changing their
767 if (update_load && p->sched_class == &fair_sched_class) {
768 reweight_task(p, prio);
770 load->weight = scale_load(sched_prio_to_weight[prio]);
771 load->inv_weight = sched_prio_to_wmult[prio];
775 #ifdef CONFIG_UCLAMP_TASK
777 * Serializes updates of utilization clamp values
779 * The (slow-path) user-space triggers utilization clamp value updates which
780 * can require updates on (fast-path) scheduler's data structures used to
781 * support enqueue/dequeue operations.
782 * While the per-CPU rq lock protects fast-path update operations, user-space
783 * requests are serialized using a mutex to reduce the risk of conflicting
784 * updates or API abuses.
786 static DEFINE_MUTEX(uclamp_mutex);
788 /* Max allowed minimum utilization */
789 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
791 /* Max allowed maximum utilization */
792 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
794 /* All clamps are required to be less or equal than these values */
795 static struct uclamp_se uclamp_default[UCLAMP_CNT];
797 /* Integer rounded range for each bucket */
798 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
800 #define for_each_clamp_id(clamp_id) \
801 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
803 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
805 return clamp_value / UCLAMP_BUCKET_DELTA;
808 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
810 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
813 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
815 if (clamp_id == UCLAMP_MIN)
817 return SCHED_CAPACITY_SCALE;
820 static inline void uclamp_se_set(struct uclamp_se *uc_se,
821 unsigned int value, bool user_defined)
823 uc_se->value = value;
824 uc_se->bucket_id = uclamp_bucket_id(value);
825 uc_se->user_defined = user_defined;
828 static inline unsigned int
829 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
830 unsigned int clamp_value)
833 * Avoid blocked utilization pushing up the frequency when we go
834 * idle (which drops the max-clamp) by retaining the last known
837 if (clamp_id == UCLAMP_MAX) {
838 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
842 return uclamp_none(UCLAMP_MIN);
845 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
846 unsigned int clamp_value)
848 /* Reset max-clamp retention only on idle exit */
849 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
852 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
856 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
857 unsigned int clamp_value)
859 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
860 int bucket_id = UCLAMP_BUCKETS - 1;
863 * Since both min and max clamps are max aggregated, find the
864 * top most bucket with tasks in.
866 for ( ; bucket_id >= 0; bucket_id--) {
867 if (!bucket[bucket_id].tasks)
869 return bucket[bucket_id].value;
872 /* No tasks -- default clamp values */
873 return uclamp_idle_value(rq, clamp_id, clamp_value);
876 static inline struct uclamp_se
877 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
879 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
880 #ifdef CONFIG_UCLAMP_TASK_GROUP
881 struct uclamp_se uc_max;
884 * Tasks in autogroups or root task group will be
885 * restricted by system defaults.
887 if (task_group_is_autogroup(task_group(p)))
889 if (task_group(p) == &root_task_group)
892 uc_max = task_group(p)->uclamp[clamp_id];
893 if (uc_req.value > uc_max.value || !uc_req.user_defined)
901 * The effective clamp bucket index of a task depends on, by increasing
903 * - the task specific clamp value, when explicitly requested from userspace
904 * - the task group effective clamp value, for tasks not either in the root
905 * group or in an autogroup
906 * - the system default clamp value, defined by the sysadmin
908 static inline struct uclamp_se
909 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
911 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
912 struct uclamp_se uc_max = uclamp_default[clamp_id];
914 /* System default restrictions always apply */
915 if (unlikely(uc_req.value > uc_max.value))
921 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
923 struct uclamp_se uc_eff;
925 /* Task currently refcounted: use back-annotated (effective) value */
926 if (p->uclamp[clamp_id].active)
927 return (unsigned long)p->uclamp[clamp_id].value;
929 uc_eff = uclamp_eff_get(p, clamp_id);
931 return (unsigned long)uc_eff.value;
935 * When a task is enqueued on a rq, the clamp bucket currently defined by the
936 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
937 * updates the rq's clamp value if required.
939 * Tasks can have a task-specific value requested from user-space, track
940 * within each bucket the maximum value for tasks refcounted in it.
941 * This "local max aggregation" allows to track the exact "requested" value
942 * for each bucket when all its RUNNABLE tasks require the same clamp.
944 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
945 enum uclamp_id clamp_id)
947 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
948 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
949 struct uclamp_bucket *bucket;
951 lockdep_assert_held(&rq->lock);
953 /* Update task effective clamp */
954 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
956 bucket = &uc_rq->bucket[uc_se->bucket_id];
958 uc_se->active = true;
960 uclamp_idle_reset(rq, clamp_id, uc_se->value);
963 * Local max aggregation: rq buckets always track the max
964 * "requested" clamp value of its RUNNABLE tasks.
966 if (bucket->tasks == 1 || uc_se->value > bucket->value)
967 bucket->value = uc_se->value;
969 if (uc_se->value > READ_ONCE(uc_rq->value))
970 WRITE_ONCE(uc_rq->value, uc_se->value);
974 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
975 * is released. If this is the last task reference counting the rq's max
976 * active clamp value, then the rq's clamp value is updated.
978 * Both refcounted tasks and rq's cached clamp values are expected to be
979 * always valid. If it's detected they are not, as defensive programming,
980 * enforce the expected state and warn.
982 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
983 enum uclamp_id clamp_id)
985 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
986 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
987 struct uclamp_bucket *bucket;
988 unsigned int bkt_clamp;
989 unsigned int rq_clamp;
991 lockdep_assert_held(&rq->lock);
993 bucket = &uc_rq->bucket[uc_se->bucket_id];
994 SCHED_WARN_ON(!bucket->tasks);
995 if (likely(bucket->tasks))
997 uc_se->active = false;
1000 * Keep "local max aggregation" simple and accept to (possibly)
1001 * overboost some RUNNABLE tasks in the same bucket.
1002 * The rq clamp bucket value is reset to its base value whenever
1003 * there are no more RUNNABLE tasks refcounting it.
1005 if (likely(bucket->tasks))
1008 rq_clamp = READ_ONCE(uc_rq->value);
1010 * Defensive programming: this should never happen. If it happens,
1011 * e.g. due to future modification, warn and fixup the expected value.
1013 SCHED_WARN_ON(bucket->value > rq_clamp);
1014 if (bucket->value >= rq_clamp) {
1015 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1016 WRITE_ONCE(uc_rq->value, bkt_clamp);
1020 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1022 enum uclamp_id clamp_id;
1024 if (unlikely(!p->sched_class->uclamp_enabled))
1027 for_each_clamp_id(clamp_id)
1028 uclamp_rq_inc_id(rq, p, clamp_id);
1030 /* Reset clamp idle holding when there is one RUNNABLE task */
1031 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1032 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1035 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1037 enum uclamp_id clamp_id;
1039 if (unlikely(!p->sched_class->uclamp_enabled))
1042 for_each_clamp_id(clamp_id)
1043 uclamp_rq_dec_id(rq, p, clamp_id);
1047 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1053 * Lock the task and the rq where the task is (or was) queued.
1055 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1056 * price to pay to safely serialize util_{min,max} updates with
1057 * enqueues, dequeues and migration operations.
1058 * This is the same locking schema used by __set_cpus_allowed_ptr().
1060 rq = task_rq_lock(p, &rf);
1063 * Setting the clamp bucket is serialized by task_rq_lock().
1064 * If the task is not yet RUNNABLE and its task_struct is not
1065 * affecting a valid clamp bucket, the next time it's enqueued,
1066 * it will already see the updated clamp bucket value.
1068 if (p->uclamp[clamp_id].active) {
1069 uclamp_rq_dec_id(rq, p, clamp_id);
1070 uclamp_rq_inc_id(rq, p, clamp_id);
1073 task_rq_unlock(rq, p, &rf);
1076 #ifdef CONFIG_UCLAMP_TASK_GROUP
1078 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1079 unsigned int clamps)
1081 enum uclamp_id clamp_id;
1082 struct css_task_iter it;
1083 struct task_struct *p;
1085 css_task_iter_start(css, 0, &it);
1086 while ((p = css_task_iter_next(&it))) {
1087 for_each_clamp_id(clamp_id) {
1088 if ((0x1 << clamp_id) & clamps)
1089 uclamp_update_active(p, clamp_id);
1092 css_task_iter_end(&it);
1095 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1096 static void uclamp_update_root_tg(void)
1098 struct task_group *tg = &root_task_group;
1100 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1101 sysctl_sched_uclamp_util_min, false);
1102 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1103 sysctl_sched_uclamp_util_max, false);
1106 cpu_util_update_eff(&root_task_group.css);
1110 static void uclamp_update_root_tg(void) { }
1113 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1114 void *buffer, size_t *lenp, loff_t *ppos)
1116 bool update_root_tg = false;
1117 int old_min, old_max;
1120 mutex_lock(&uclamp_mutex);
1121 old_min = sysctl_sched_uclamp_util_min;
1122 old_max = sysctl_sched_uclamp_util_max;
1124 result = proc_dointvec(table, write, buffer, lenp, ppos);
1130 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1131 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1136 if (old_min != sysctl_sched_uclamp_util_min) {
1137 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1138 sysctl_sched_uclamp_util_min, false);
1139 update_root_tg = true;
1141 if (old_max != sysctl_sched_uclamp_util_max) {
1142 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1143 sysctl_sched_uclamp_util_max, false);
1144 update_root_tg = true;
1148 uclamp_update_root_tg();
1151 * We update all RUNNABLE tasks only when task groups are in use.
1152 * Otherwise, keep it simple and do just a lazy update at each next
1153 * task enqueue time.
1159 sysctl_sched_uclamp_util_min = old_min;
1160 sysctl_sched_uclamp_util_max = old_max;
1162 mutex_unlock(&uclamp_mutex);
1167 static int uclamp_validate(struct task_struct *p,
1168 const struct sched_attr *attr)
1170 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1171 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1173 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1174 lower_bound = attr->sched_util_min;
1175 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1176 upper_bound = attr->sched_util_max;
1178 if (lower_bound > upper_bound)
1180 if (upper_bound > SCHED_CAPACITY_SCALE)
1186 static void __setscheduler_uclamp(struct task_struct *p,
1187 const struct sched_attr *attr)
1189 enum uclamp_id clamp_id;
1192 * On scheduling class change, reset to default clamps for tasks
1193 * without a task-specific value.
1195 for_each_clamp_id(clamp_id) {
1196 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1197 unsigned int clamp_value = uclamp_none(clamp_id);
1199 /* Keep using defined clamps across class changes */
1200 if (uc_se->user_defined)
1203 /* By default, RT tasks always get 100% boost */
1204 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1205 clamp_value = uclamp_none(UCLAMP_MAX);
1207 uclamp_se_set(uc_se, clamp_value, false);
1210 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1213 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1214 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1215 attr->sched_util_min, true);
1218 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1219 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1220 attr->sched_util_max, true);
1224 static void uclamp_fork(struct task_struct *p)
1226 enum uclamp_id clamp_id;
1228 for_each_clamp_id(clamp_id)
1229 p->uclamp[clamp_id].active = false;
1231 if (likely(!p->sched_reset_on_fork))
1234 for_each_clamp_id(clamp_id) {
1235 uclamp_se_set(&p->uclamp_req[clamp_id],
1236 uclamp_none(clamp_id), false);
1240 static void __init init_uclamp(void)
1242 struct uclamp_se uc_max = {};
1243 enum uclamp_id clamp_id;
1246 mutex_init(&uclamp_mutex);
1248 for_each_possible_cpu(cpu) {
1249 memset(&cpu_rq(cpu)->uclamp, 0,
1250 sizeof(struct uclamp_rq)*UCLAMP_CNT);
1251 cpu_rq(cpu)->uclamp_flags = 0;
1254 for_each_clamp_id(clamp_id) {
1255 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1256 uclamp_none(clamp_id), false);
1259 /* System defaults allow max clamp values for both indexes */
1260 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1261 for_each_clamp_id(clamp_id) {
1262 uclamp_default[clamp_id] = uc_max;
1263 #ifdef CONFIG_UCLAMP_TASK_GROUP
1264 root_task_group.uclamp_req[clamp_id] = uc_max;
1265 root_task_group.uclamp[clamp_id] = uc_max;
1270 #else /* CONFIG_UCLAMP_TASK */
1271 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1272 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1273 static inline int uclamp_validate(struct task_struct *p,
1274 const struct sched_attr *attr)
1278 static void __setscheduler_uclamp(struct task_struct *p,
1279 const struct sched_attr *attr) { }
1280 static inline void uclamp_fork(struct task_struct *p) { }
1281 static inline void init_uclamp(void) { }
1282 #endif /* CONFIG_UCLAMP_TASK */
1284 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1286 if (!(flags & ENQUEUE_NOCLOCK))
1287 update_rq_clock(rq);
1289 if (!(flags & ENQUEUE_RESTORE)) {
1290 sched_info_queued(rq, p);
1291 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1294 uclamp_rq_inc(rq, p);
1295 p->sched_class->enqueue_task(rq, p, flags);
1298 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1300 if (!(flags & DEQUEUE_NOCLOCK))
1301 update_rq_clock(rq);
1303 if (!(flags & DEQUEUE_SAVE)) {
1304 sched_info_dequeued(rq, p);
1305 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1308 uclamp_rq_dec(rq, p);
1309 p->sched_class->dequeue_task(rq, p, flags);
1312 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1314 enqueue_task(rq, p, flags);
1316 p->on_rq = TASK_ON_RQ_QUEUED;
1319 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1321 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1323 dequeue_task(rq, p, flags);
1327 * __normal_prio - return the priority that is based on the static prio
1329 static inline int __normal_prio(struct task_struct *p)
1331 return p->static_prio;
1335 * Calculate the expected normal priority: i.e. priority
1336 * without taking RT-inheritance into account. Might be
1337 * boosted by interactivity modifiers. Changes upon fork,
1338 * setprio syscalls, and whenever the interactivity
1339 * estimator recalculates.
1341 static inline int normal_prio(struct task_struct *p)
1345 if (task_has_dl_policy(p))
1346 prio = MAX_DL_PRIO-1;
1347 else if (task_has_rt_policy(p))
1348 prio = MAX_RT_PRIO-1 - p->rt_priority;
1350 prio = __normal_prio(p);
1355 * Calculate the current priority, i.e. the priority
1356 * taken into account by the scheduler. This value might
1357 * be boosted by RT tasks, or might be boosted by
1358 * interactivity modifiers. Will be RT if the task got
1359 * RT-boosted. If not then it returns p->normal_prio.
1361 static int effective_prio(struct task_struct *p)
1363 p->normal_prio = normal_prio(p);
1365 * If we are RT tasks or we were boosted to RT priority,
1366 * keep the priority unchanged. Otherwise, update priority
1367 * to the normal priority:
1369 if (!rt_prio(p->prio))
1370 return p->normal_prio;
1375 * task_curr - is this task currently executing on a CPU?
1376 * @p: the task in question.
1378 * Return: 1 if the task is currently executing. 0 otherwise.
1380 inline int task_curr(const struct task_struct *p)
1382 return cpu_curr(task_cpu(p)) == p;
1386 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1387 * use the balance_callback list if you want balancing.
1389 * this means any call to check_class_changed() must be followed by a call to
1390 * balance_callback().
1392 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1393 const struct sched_class *prev_class,
1396 if (prev_class != p->sched_class) {
1397 if (prev_class->switched_from)
1398 prev_class->switched_from(rq, p);
1400 p->sched_class->switched_to(rq, p);
1401 } else if (oldprio != p->prio || dl_task(p))
1402 p->sched_class->prio_changed(rq, p, oldprio);
1405 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1407 const struct sched_class *class;
1409 if (p->sched_class == rq->curr->sched_class) {
1410 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1412 for_each_class(class) {
1413 if (class == rq->curr->sched_class)
1415 if (class == p->sched_class) {
1423 * A queue event has occurred, and we're going to schedule. In
1424 * this case, we can save a useless back to back clock update.
1426 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1427 rq_clock_skip_update(rq);
1433 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1434 * __set_cpus_allowed_ptr() and select_fallback_rq().
1436 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1438 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1441 if (is_per_cpu_kthread(p))
1442 return cpu_online(cpu);
1444 return cpu_active(cpu);
1448 * This is how migration works:
1450 * 1) we invoke migration_cpu_stop() on the target CPU using
1452 * 2) stopper starts to run (implicitly forcing the migrated thread
1454 * 3) it checks whether the migrated task is still in the wrong runqueue.
1455 * 4) if it's in the wrong runqueue then the migration thread removes
1456 * it and puts it into the right queue.
1457 * 5) stopper completes and stop_one_cpu() returns and the migration
1462 * move_queued_task - move a queued task to new rq.
1464 * Returns (locked) new rq. Old rq's lock is released.
1466 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1467 struct task_struct *p, int new_cpu)
1469 lockdep_assert_held(&rq->lock);
1471 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1472 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1473 set_task_cpu(p, new_cpu);
1476 rq = cpu_rq(new_cpu);
1479 BUG_ON(task_cpu(p) != new_cpu);
1480 enqueue_task(rq, p, 0);
1481 p->on_rq = TASK_ON_RQ_QUEUED;
1482 check_preempt_curr(rq, p, 0);
1487 struct migration_arg {
1488 struct task_struct *task;
1493 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1494 * this because either it can't run here any more (set_cpus_allowed()
1495 * away from this CPU, or CPU going down), or because we're
1496 * attempting to rebalance this task on exec (sched_exec).
1498 * So we race with normal scheduler movements, but that's OK, as long
1499 * as the task is no longer on this CPU.
1501 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1502 struct task_struct *p, int dest_cpu)
1504 /* Affinity changed (again). */
1505 if (!is_cpu_allowed(p, dest_cpu))
1508 update_rq_clock(rq);
1509 rq = move_queued_task(rq, rf, p, dest_cpu);
1515 * migration_cpu_stop - this will be executed by a highprio stopper thread
1516 * and performs thread migration by bumping thread off CPU then
1517 * 'pushing' onto another runqueue.
1519 static int migration_cpu_stop(void *data)
1521 struct migration_arg *arg = data;
1522 struct task_struct *p = arg->task;
1523 struct rq *rq = this_rq();
1527 * The original target CPU might have gone down and we might
1528 * be on another CPU but it doesn't matter.
1530 local_irq_disable();
1532 * We need to explicitly wake pending tasks before running
1533 * __migrate_task() such that we will not miss enforcing cpus_ptr
1534 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1536 flush_smp_call_function_from_idle();
1538 raw_spin_lock(&p->pi_lock);
1541 * If task_rq(p) != rq, it cannot be migrated here, because we're
1542 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1543 * we're holding p->pi_lock.
1545 if (task_rq(p) == rq) {
1546 if (task_on_rq_queued(p))
1547 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1549 p->wake_cpu = arg->dest_cpu;
1552 raw_spin_unlock(&p->pi_lock);
1559 * sched_class::set_cpus_allowed must do the below, but is not required to
1560 * actually call this function.
1562 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1564 cpumask_copy(&p->cpus_mask, new_mask);
1565 p->nr_cpus_allowed = cpumask_weight(new_mask);
1568 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1570 struct rq *rq = task_rq(p);
1571 bool queued, running;
1573 lockdep_assert_held(&p->pi_lock);
1575 queued = task_on_rq_queued(p);
1576 running = task_current(rq, p);
1580 * Because __kthread_bind() calls this on blocked tasks without
1583 lockdep_assert_held(&rq->lock);
1584 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1587 put_prev_task(rq, p);
1589 p->sched_class->set_cpus_allowed(p, new_mask);
1592 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1594 set_next_task(rq, p);
1598 * Change a given task's CPU affinity. Migrate the thread to a
1599 * proper CPU and schedule it away if the CPU it's executing on
1600 * is removed from the allowed bitmask.
1602 * NOTE: the caller must have a valid reference to the task, the
1603 * task must not exit() & deallocate itself prematurely. The
1604 * call is not atomic; no spinlocks may be held.
1606 static int __set_cpus_allowed_ptr(struct task_struct *p,
1607 const struct cpumask *new_mask, bool check)
1609 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1610 unsigned int dest_cpu;
1615 rq = task_rq_lock(p, &rf);
1616 update_rq_clock(rq);
1618 if (p->flags & PF_KTHREAD) {
1620 * Kernel threads are allowed on online && !active CPUs
1622 cpu_valid_mask = cpu_online_mask;
1626 * Must re-check here, to close a race against __kthread_bind(),
1627 * sched_setaffinity() is not guaranteed to observe the flag.
1629 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1634 if (cpumask_equal(&p->cpus_mask, new_mask))
1638 * Picking a ~random cpu helps in cases where we are changing affinity
1639 * for groups of tasks (ie. cpuset), so that load balancing is not
1640 * immediately required to distribute the tasks within their new mask.
1642 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
1643 if (dest_cpu >= nr_cpu_ids) {
1648 do_set_cpus_allowed(p, new_mask);
1650 if (p->flags & PF_KTHREAD) {
1652 * For kernel threads that do indeed end up on online &&
1653 * !active we want to ensure they are strict per-CPU threads.
1655 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1656 !cpumask_intersects(new_mask, cpu_active_mask) &&
1657 p->nr_cpus_allowed != 1);
1660 /* Can the task run on the task's current CPU? If so, we're done */
1661 if (cpumask_test_cpu(task_cpu(p), new_mask))
1664 if (task_running(rq, p) || p->state == TASK_WAKING) {
1665 struct migration_arg arg = { p, dest_cpu };
1666 /* Need help from migration thread: drop lock and wait. */
1667 task_rq_unlock(rq, p, &rf);
1668 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1670 } else if (task_on_rq_queued(p)) {
1672 * OK, since we're going to drop the lock immediately
1673 * afterwards anyway.
1675 rq = move_queued_task(rq, &rf, p, dest_cpu);
1678 task_rq_unlock(rq, p, &rf);
1683 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1685 return __set_cpus_allowed_ptr(p, new_mask, false);
1687 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1689 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1691 #ifdef CONFIG_SCHED_DEBUG
1693 * We should never call set_task_cpu() on a blocked task,
1694 * ttwu() will sort out the placement.
1696 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1700 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1701 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1702 * time relying on p->on_rq.
1704 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1705 p->sched_class == &fair_sched_class &&
1706 (p->on_rq && !task_on_rq_migrating(p)));
1708 #ifdef CONFIG_LOCKDEP
1710 * The caller should hold either p->pi_lock or rq->lock, when changing
1711 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1713 * sched_move_task() holds both and thus holding either pins the cgroup,
1716 * Furthermore, all task_rq users should acquire both locks, see
1719 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1720 lockdep_is_held(&task_rq(p)->lock)));
1723 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1725 WARN_ON_ONCE(!cpu_online(new_cpu));
1728 trace_sched_migrate_task(p, new_cpu);
1730 if (task_cpu(p) != new_cpu) {
1731 if (p->sched_class->migrate_task_rq)
1732 p->sched_class->migrate_task_rq(p, new_cpu);
1733 p->se.nr_migrations++;
1735 perf_event_task_migrate(p);
1738 __set_task_cpu(p, new_cpu);
1741 #ifdef CONFIG_NUMA_BALANCING
1742 static void __migrate_swap_task(struct task_struct *p, int cpu)
1744 if (task_on_rq_queued(p)) {
1745 struct rq *src_rq, *dst_rq;
1746 struct rq_flags srf, drf;
1748 src_rq = task_rq(p);
1749 dst_rq = cpu_rq(cpu);
1751 rq_pin_lock(src_rq, &srf);
1752 rq_pin_lock(dst_rq, &drf);
1754 deactivate_task(src_rq, p, 0);
1755 set_task_cpu(p, cpu);
1756 activate_task(dst_rq, p, 0);
1757 check_preempt_curr(dst_rq, p, 0);
1759 rq_unpin_lock(dst_rq, &drf);
1760 rq_unpin_lock(src_rq, &srf);
1764 * Task isn't running anymore; make it appear like we migrated
1765 * it before it went to sleep. This means on wakeup we make the
1766 * previous CPU our target instead of where it really is.
1772 struct migration_swap_arg {
1773 struct task_struct *src_task, *dst_task;
1774 int src_cpu, dst_cpu;
1777 static int migrate_swap_stop(void *data)
1779 struct migration_swap_arg *arg = data;
1780 struct rq *src_rq, *dst_rq;
1783 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1786 src_rq = cpu_rq(arg->src_cpu);
1787 dst_rq = cpu_rq(arg->dst_cpu);
1789 double_raw_lock(&arg->src_task->pi_lock,
1790 &arg->dst_task->pi_lock);
1791 double_rq_lock(src_rq, dst_rq);
1793 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1796 if (task_cpu(arg->src_task) != arg->src_cpu)
1799 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1802 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1805 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1806 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1811 double_rq_unlock(src_rq, dst_rq);
1812 raw_spin_unlock(&arg->dst_task->pi_lock);
1813 raw_spin_unlock(&arg->src_task->pi_lock);
1819 * Cross migrate two tasks
1821 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1822 int target_cpu, int curr_cpu)
1824 struct migration_swap_arg arg;
1827 arg = (struct migration_swap_arg){
1829 .src_cpu = curr_cpu,
1831 .dst_cpu = target_cpu,
1834 if (arg.src_cpu == arg.dst_cpu)
1838 * These three tests are all lockless; this is OK since all of them
1839 * will be re-checked with proper locks held further down the line.
1841 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1844 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1847 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1850 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1851 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1856 #endif /* CONFIG_NUMA_BALANCING */
1859 * wait_task_inactive - wait for a thread to unschedule.
1861 * If @match_state is nonzero, it's the @p->state value just checked and
1862 * not expected to change. If it changes, i.e. @p might have woken up,
1863 * then return zero. When we succeed in waiting for @p to be off its CPU,
1864 * we return a positive number (its total switch count). If a second call
1865 * a short while later returns the same number, the caller can be sure that
1866 * @p has remained unscheduled the whole time.
1868 * The caller must ensure that the task *will* unschedule sometime soon,
1869 * else this function might spin for a *long* time. This function can't
1870 * be called with interrupts off, or it may introduce deadlock with
1871 * smp_call_function() if an IPI is sent by the same process we are
1872 * waiting to become inactive.
1874 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1876 int running, queued;
1883 * We do the initial early heuristics without holding
1884 * any task-queue locks at all. We'll only try to get
1885 * the runqueue lock when things look like they will
1891 * If the task is actively running on another CPU
1892 * still, just relax and busy-wait without holding
1895 * NOTE! Since we don't hold any locks, it's not
1896 * even sure that "rq" stays as the right runqueue!
1897 * But we don't care, since "task_running()" will
1898 * return false if the runqueue has changed and p
1899 * is actually now running somewhere else!
1901 while (task_running(rq, p)) {
1902 if (match_state && unlikely(p->state != match_state))
1908 * Ok, time to look more closely! We need the rq
1909 * lock now, to be *sure*. If we're wrong, we'll
1910 * just go back and repeat.
1912 rq = task_rq_lock(p, &rf);
1913 trace_sched_wait_task(p);
1914 running = task_running(rq, p);
1915 queued = task_on_rq_queued(p);
1917 if (!match_state || p->state == match_state)
1918 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1919 task_rq_unlock(rq, p, &rf);
1922 * If it changed from the expected state, bail out now.
1924 if (unlikely(!ncsw))
1928 * Was it really running after all now that we
1929 * checked with the proper locks actually held?
1931 * Oops. Go back and try again..
1933 if (unlikely(running)) {
1939 * It's not enough that it's not actively running,
1940 * it must be off the runqueue _entirely_, and not
1943 * So if it was still runnable (but just not actively
1944 * running right now), it's preempted, and we should
1945 * yield - it could be a while.
1947 if (unlikely(queued)) {
1948 ktime_t to = NSEC_PER_SEC / HZ;
1950 set_current_state(TASK_UNINTERRUPTIBLE);
1951 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1956 * Ahh, all good. It wasn't running, and it wasn't
1957 * runnable, which means that it will never become
1958 * running in the future either. We're all done!
1967 * kick_process - kick a running thread to enter/exit the kernel
1968 * @p: the to-be-kicked thread
1970 * Cause a process which is running on another CPU to enter
1971 * kernel-mode, without any delay. (to get signals handled.)
1973 * NOTE: this function doesn't have to take the runqueue lock,
1974 * because all it wants to ensure is that the remote task enters
1975 * the kernel. If the IPI races and the task has been migrated
1976 * to another CPU then no harm is done and the purpose has been
1979 void kick_process(struct task_struct *p)
1985 if ((cpu != smp_processor_id()) && task_curr(p))
1986 smp_send_reschedule(cpu);
1989 EXPORT_SYMBOL_GPL(kick_process);
1992 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
1994 * A few notes on cpu_active vs cpu_online:
1996 * - cpu_active must be a subset of cpu_online
1998 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1999 * see __set_cpus_allowed_ptr(). At this point the newly online
2000 * CPU isn't yet part of the sched domains, and balancing will not
2003 * - on CPU-down we clear cpu_active() to mask the sched domains and
2004 * avoid the load balancer to place new tasks on the to be removed
2005 * CPU. Existing tasks will remain running there and will be taken
2008 * This means that fallback selection must not select !active CPUs.
2009 * And can assume that any active CPU must be online. Conversely
2010 * select_task_rq() below may allow selection of !active CPUs in order
2011 * to satisfy the above rules.
2013 static int select_fallback_rq(int cpu, struct task_struct *p)
2015 int nid = cpu_to_node(cpu);
2016 const struct cpumask *nodemask = NULL;
2017 enum { cpuset, possible, fail } state = cpuset;
2021 * If the node that the CPU is on has been offlined, cpu_to_node()
2022 * will return -1. There is no CPU on the node, and we should
2023 * select the CPU on the other node.
2026 nodemask = cpumask_of_node(nid);
2028 /* Look for allowed, online CPU in same node. */
2029 for_each_cpu(dest_cpu, nodemask) {
2030 if (!cpu_active(dest_cpu))
2032 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2038 /* Any allowed, online CPU? */
2039 for_each_cpu(dest_cpu, p->cpus_ptr) {
2040 if (!is_cpu_allowed(p, dest_cpu))
2046 /* No more Mr. Nice Guy. */
2049 if (IS_ENABLED(CONFIG_CPUSETS)) {
2050 cpuset_cpus_allowed_fallback(p);
2056 do_set_cpus_allowed(p, cpu_possible_mask);
2067 if (state != cpuset) {
2069 * Don't tell them about moving exiting tasks or
2070 * kernel threads (both mm NULL), since they never
2073 if (p->mm && printk_ratelimit()) {
2074 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2075 task_pid_nr(p), p->comm, cpu);
2083 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2086 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2088 lockdep_assert_held(&p->pi_lock);
2090 if (p->nr_cpus_allowed > 1)
2091 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2093 cpu = cpumask_any(p->cpus_ptr);
2096 * In order not to call set_task_cpu() on a blocking task we need
2097 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2100 * Since this is common to all placement strategies, this lives here.
2102 * [ this allows ->select_task() to simply return task_cpu(p) and
2103 * not worry about this generic constraint ]
2105 if (unlikely(!is_cpu_allowed(p, cpu)))
2106 cpu = select_fallback_rq(task_cpu(p), p);
2111 void sched_set_stop_task(int cpu, struct task_struct *stop)
2113 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2114 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2118 * Make it appear like a SCHED_FIFO task, its something
2119 * userspace knows about and won't get confused about.
2121 * Also, it will make PI more or less work without too
2122 * much confusion -- but then, stop work should not
2123 * rely on PI working anyway.
2125 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2127 stop->sched_class = &stop_sched_class;
2130 cpu_rq(cpu)->stop = stop;
2134 * Reset it back to a normal scheduling class so that
2135 * it can die in pieces.
2137 old_stop->sched_class = &rt_sched_class;
2143 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2144 const struct cpumask *new_mask, bool check)
2146 return set_cpus_allowed_ptr(p, new_mask);
2149 #endif /* CONFIG_SMP */
2152 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2156 if (!schedstat_enabled())
2162 if (cpu == rq->cpu) {
2163 __schedstat_inc(rq->ttwu_local);
2164 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2166 struct sched_domain *sd;
2168 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2170 for_each_domain(rq->cpu, sd) {
2171 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2172 __schedstat_inc(sd->ttwu_wake_remote);
2179 if (wake_flags & WF_MIGRATED)
2180 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2181 #endif /* CONFIG_SMP */
2183 __schedstat_inc(rq->ttwu_count);
2184 __schedstat_inc(p->se.statistics.nr_wakeups);
2186 if (wake_flags & WF_SYNC)
2187 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2191 * Mark the task runnable and perform wakeup-preemption.
2193 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2194 struct rq_flags *rf)
2196 check_preempt_curr(rq, p, wake_flags);
2197 p->state = TASK_RUNNING;
2198 trace_sched_wakeup(p);
2201 if (p->sched_class->task_woken) {
2203 * Our task @p is fully woken up and running; so its safe to
2204 * drop the rq->lock, hereafter rq is only used for statistics.
2206 rq_unpin_lock(rq, rf);
2207 p->sched_class->task_woken(rq, p);
2208 rq_repin_lock(rq, rf);
2211 if (rq->idle_stamp) {
2212 u64 delta = rq_clock(rq) - rq->idle_stamp;
2213 u64 max = 2*rq->max_idle_balance_cost;
2215 update_avg(&rq->avg_idle, delta);
2217 if (rq->avg_idle > max)
2226 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2227 struct rq_flags *rf)
2229 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2231 lockdep_assert_held(&rq->lock);
2233 if (p->sched_contributes_to_load)
2234 rq->nr_uninterruptible--;
2237 if (wake_flags & WF_MIGRATED)
2238 en_flags |= ENQUEUE_MIGRATED;
2241 activate_task(rq, p, en_flags);
2242 ttwu_do_wakeup(rq, p, wake_flags, rf);
2246 * Called in case the task @p isn't fully descheduled from its runqueue,
2247 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2248 * since all we need to do is flip p->state to TASK_RUNNING, since
2249 * the task is still ->on_rq.
2251 static int ttwu_remote(struct task_struct *p, int wake_flags)
2257 rq = __task_rq_lock(p, &rf);
2258 if (task_on_rq_queued(p)) {
2259 /* check_preempt_curr() may use rq clock */
2260 update_rq_clock(rq);
2261 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2264 __task_rq_unlock(rq, &rf);
2270 void sched_ttwu_pending(void *arg)
2272 struct llist_node *llist = arg;
2273 struct rq *rq = this_rq();
2274 struct task_struct *p, *t;
2281 * rq::ttwu_pending racy indication of out-standing wakeups.
2282 * Races such that false-negatives are possible, since they
2283 * are shorter lived that false-positives would be.
2285 WRITE_ONCE(rq->ttwu_pending, 0);
2287 rq_lock_irqsave(rq, &rf);
2288 update_rq_clock(rq);
2290 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
2291 if (WARN_ON_ONCE(p->on_cpu))
2292 smp_cond_load_acquire(&p->on_cpu, !VAL);
2294 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
2295 set_task_cpu(p, cpu_of(rq));
2297 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2300 rq_unlock_irqrestore(rq, &rf);
2303 void send_call_function_single_ipi(int cpu)
2305 struct rq *rq = cpu_rq(cpu);
2307 if (!set_nr_if_polling(rq->idle))
2308 arch_send_call_function_single_ipi(cpu);
2310 trace_sched_wake_idle_without_ipi(cpu);
2314 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2315 * necessary. The wakee CPU on receipt of the IPI will queue the task
2316 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2317 * of the wakeup instead of the waker.
2319 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2321 struct rq *rq = cpu_rq(cpu);
2323 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2325 WRITE_ONCE(rq->ttwu_pending, 1);
2326 __smp_call_single_queue(cpu, &p->wake_entry.llist);
2329 void wake_up_if_idle(int cpu)
2331 struct rq *rq = cpu_rq(cpu);
2336 if (!is_idle_task(rcu_dereference(rq->curr)))
2339 if (set_nr_if_polling(rq->idle)) {
2340 trace_sched_wake_idle_without_ipi(cpu);
2342 rq_lock_irqsave(rq, &rf);
2343 if (is_idle_task(rq->curr))
2344 smp_send_reschedule(cpu);
2345 /* Else CPU is not idle, do nothing here: */
2346 rq_unlock_irqrestore(rq, &rf);
2353 bool cpus_share_cache(int this_cpu, int that_cpu)
2355 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2358 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
2361 * If the CPU does not share cache, then queue the task on the
2362 * remote rqs wakelist to avoid accessing remote data.
2364 if (!cpus_share_cache(smp_processor_id(), cpu))
2368 * If the task is descheduling and the only running task on the
2369 * CPU then use the wakelist to offload the task activation to
2370 * the soon-to-be-idle CPU as the current CPU is likely busy.
2371 * nr_running is checked to avoid unnecessary task stacking.
2373 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
2379 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2381 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
2382 if (WARN_ON_ONCE(cpu == smp_processor_id()))
2385 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2386 __ttwu_queue_wakelist(p, cpu, wake_flags);
2392 #endif /* CONFIG_SMP */
2394 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2396 struct rq *rq = cpu_rq(cpu);
2399 #if defined(CONFIG_SMP)
2400 if (ttwu_queue_wakelist(p, cpu, wake_flags))
2405 update_rq_clock(rq);
2406 ttwu_do_activate(rq, p, wake_flags, &rf);
2411 * Notes on Program-Order guarantees on SMP systems.
2415 * The basic program-order guarantee on SMP systems is that when a task [t]
2416 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2417 * execution on its new CPU [c1].
2419 * For migration (of runnable tasks) this is provided by the following means:
2421 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2422 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2423 * rq(c1)->lock (if not at the same time, then in that order).
2424 * C) LOCK of the rq(c1)->lock scheduling in task
2426 * Release/acquire chaining guarantees that B happens after A and C after B.
2427 * Note: the CPU doing B need not be c0 or c1
2436 * UNLOCK rq(0)->lock
2438 * LOCK rq(0)->lock // orders against CPU0
2440 * UNLOCK rq(0)->lock
2444 * UNLOCK rq(1)->lock
2446 * LOCK rq(1)->lock // orders against CPU2
2449 * UNLOCK rq(1)->lock
2452 * BLOCKING -- aka. SLEEP + WAKEUP
2454 * For blocking we (obviously) need to provide the same guarantee as for
2455 * migration. However the means are completely different as there is no lock
2456 * chain to provide order. Instead we do:
2458 * 1) smp_store_release(X->on_cpu, 0)
2459 * 2) smp_cond_load_acquire(!X->on_cpu)
2463 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2465 * LOCK rq(0)->lock LOCK X->pi_lock
2468 * smp_store_release(X->on_cpu, 0);
2470 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2476 * X->state = RUNNING
2477 * UNLOCK rq(2)->lock
2479 * LOCK rq(2)->lock // orders against CPU1
2482 * UNLOCK rq(2)->lock
2485 * UNLOCK rq(0)->lock
2488 * However, for wakeups there is a second guarantee we must provide, namely we
2489 * must ensure that CONDITION=1 done by the caller can not be reordered with
2490 * accesses to the task state; see try_to_wake_up() and set_current_state().
2494 * try_to_wake_up - wake up a thread
2495 * @p: the thread to be awakened
2496 * @state: the mask of task states that can be woken
2497 * @wake_flags: wake modifier flags (WF_*)
2499 * If (@state & @p->state) @p->state = TASK_RUNNING.
2501 * If the task was not queued/runnable, also place it back on a runqueue.
2503 * Atomic against schedule() which would dequeue a task, also see
2504 * set_current_state().
2506 * This function executes a full memory barrier before accessing the task
2507 * state; see set_current_state().
2509 * Return: %true if @p->state changes (an actual wakeup was done),
2513 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2515 unsigned long flags;
2516 int cpu, success = 0;
2521 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2522 * == smp_processor_id()'. Together this means we can special
2523 * case the whole 'p->on_rq && ttwu_remote()' case below
2524 * without taking any locks.
2527 * - we rely on Program-Order guarantees for all the ordering,
2528 * - we're serialized against set_special_state() by virtue of
2529 * it disabling IRQs (this allows not taking ->pi_lock).
2531 if (!(p->state & state))
2535 trace_sched_waking(p);
2536 p->state = TASK_RUNNING;
2537 trace_sched_wakeup(p);
2542 * If we are going to wake up a thread waiting for CONDITION we
2543 * need to ensure that CONDITION=1 done by the caller can not be
2544 * reordered with p->state check below. This pairs with mb() in
2545 * set_current_state() the waiting thread does.
2547 raw_spin_lock_irqsave(&p->pi_lock, flags);
2548 smp_mb__after_spinlock();
2549 if (!(p->state & state))
2552 trace_sched_waking(p);
2554 /* We're going to change ->state: */
2558 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2559 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2560 * in smp_cond_load_acquire() below.
2562 * sched_ttwu_pending() try_to_wake_up()
2563 * STORE p->on_rq = 1 LOAD p->state
2566 * __schedule() (switch to task 'p')
2567 * LOCK rq->lock smp_rmb();
2568 * smp_mb__after_spinlock();
2572 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2574 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2575 * __schedule(). See the comment for smp_mb__after_spinlock().
2577 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2580 if (READ_ONCE(p->on_rq) && ttwu_remote(p, wake_flags))
2584 delayacct_blkio_end(p);
2585 atomic_dec(&task_rq(p)->nr_iowait);
2590 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2591 * possible to, falsely, observe p->on_cpu == 0.
2593 * One must be running (->on_cpu == 1) in order to remove oneself
2594 * from the runqueue.
2596 * __schedule() (switch to task 'p') try_to_wake_up()
2597 * STORE p->on_cpu = 1 LOAD p->on_rq
2600 * __schedule() (put 'p' to sleep)
2601 * LOCK rq->lock smp_rmb();
2602 * smp_mb__after_spinlock();
2603 * STORE p->on_rq = 0 LOAD p->on_cpu
2605 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2606 * __schedule(). See the comment for smp_mb__after_spinlock().
2608 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
2609 * schedule()'s deactivate_task() has 'happened' and p will no longer
2610 * care about it's own p->state. See the comment in __schedule().
2612 smp_acquire__after_ctrl_dep();
2615 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
2616 * == 0), which means we need to do an enqueue, change p->state to
2617 * TASK_WAKING such that we can unlock p->pi_lock before doing the
2618 * enqueue, such as ttwu_queue_wakelist().
2620 p->state = TASK_WAKING;
2623 * If the owning (remote) CPU is still in the middle of schedule() with
2624 * this task as prev, considering queueing p on the remote CPUs wake_list
2625 * which potentially sends an IPI instead of spinning on p->on_cpu to
2626 * let the waker make forward progress. This is safe because IRQs are
2627 * disabled and the IPI will deliver after on_cpu is cleared.
2629 * Ensure we load task_cpu(p) after p->on_cpu:
2631 * set_task_cpu(p, cpu);
2632 * STORE p->cpu = @cpu
2633 * __schedule() (switch to task 'p')
2635 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
2636 * STORE p->on_cpu = 1 LOAD p->cpu
2638 * to ensure we observe the correct CPU on which the task is currently
2641 if (smp_load_acquire(&p->on_cpu) &&
2642 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
2646 * If the owning (remote) CPU is still in the middle of schedule() with
2647 * this task as prev, wait until its done referencing the task.
2649 * Pairs with the smp_store_release() in finish_task().
2651 * This ensures that tasks getting woken will be fully ordered against
2652 * their previous state and preserve Program Order.
2654 smp_cond_load_acquire(&p->on_cpu, !VAL);
2656 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2657 if (task_cpu(p) != cpu) {
2658 wake_flags |= WF_MIGRATED;
2659 psi_ttwu_dequeue(p);
2660 set_task_cpu(p, cpu);
2664 #endif /* CONFIG_SMP */
2666 ttwu_queue(p, cpu, wake_flags);
2668 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2671 ttwu_stat(p, task_cpu(p), wake_flags);
2678 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2679 * @p: Process for which the function is to be invoked.
2680 * @func: Function to invoke.
2681 * @arg: Argument to function.
2683 * If the specified task can be quickly locked into a definite state
2684 * (either sleeping or on a given runqueue), arrange to keep it in that
2685 * state while invoking @func(@arg). This function can use ->on_rq and
2686 * task_curr() to work out what the state is, if required. Given that
2687 * @func can be invoked with a runqueue lock held, it had better be quite
2691 * @false if the task slipped out from under the locks.
2692 * @true if the task was locked onto a runqueue or is sleeping.
2693 * However, @func can override this by returning @false.
2695 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
2701 lockdep_assert_irqs_enabled();
2702 raw_spin_lock_irq(&p->pi_lock);
2704 rq = __task_rq_lock(p, &rf);
2705 if (task_rq(p) == rq)
2714 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
2719 raw_spin_unlock_irq(&p->pi_lock);
2724 * wake_up_process - Wake up a specific process
2725 * @p: The process to be woken up.
2727 * Attempt to wake up the nominated process and move it to the set of runnable
2730 * Return: 1 if the process was woken up, 0 if it was already running.
2732 * This function executes a full memory barrier before accessing the task state.
2734 int wake_up_process(struct task_struct *p)
2736 return try_to_wake_up(p, TASK_NORMAL, 0);
2738 EXPORT_SYMBOL(wake_up_process);
2740 int wake_up_state(struct task_struct *p, unsigned int state)
2742 return try_to_wake_up(p, state, 0);
2746 * Perform scheduler related setup for a newly forked process p.
2747 * p is forked by current.
2749 * __sched_fork() is basic setup used by init_idle() too:
2751 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2756 p->se.exec_start = 0;
2757 p->se.sum_exec_runtime = 0;
2758 p->se.prev_sum_exec_runtime = 0;
2759 p->se.nr_migrations = 0;
2761 INIT_LIST_HEAD(&p->se.group_node);
2763 #ifdef CONFIG_FAIR_GROUP_SCHED
2764 p->se.cfs_rq = NULL;
2767 #ifdef CONFIG_SCHEDSTATS
2768 /* Even if schedstat is disabled, there should not be garbage */
2769 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2772 RB_CLEAR_NODE(&p->dl.rb_node);
2773 init_dl_task_timer(&p->dl);
2774 init_dl_inactive_task_timer(&p->dl);
2775 __dl_clear_params(p);
2777 INIT_LIST_HEAD(&p->rt.run_list);
2779 p->rt.time_slice = sched_rr_timeslice;
2783 #ifdef CONFIG_PREEMPT_NOTIFIERS
2784 INIT_HLIST_HEAD(&p->preempt_notifiers);
2787 #ifdef CONFIG_COMPACTION
2788 p->capture_control = NULL;
2790 init_numa_balancing(clone_flags, p);
2792 p->wake_entry.u_flags = CSD_TYPE_TTWU;
2796 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2798 #ifdef CONFIG_NUMA_BALANCING
2800 void set_numabalancing_state(bool enabled)
2803 static_branch_enable(&sched_numa_balancing);
2805 static_branch_disable(&sched_numa_balancing);
2808 #ifdef CONFIG_PROC_SYSCTL
2809 int sysctl_numa_balancing(struct ctl_table *table, int write,
2810 void *buffer, size_t *lenp, loff_t *ppos)
2814 int state = static_branch_likely(&sched_numa_balancing);
2816 if (write && !capable(CAP_SYS_ADMIN))
2821 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2825 set_numabalancing_state(state);
2831 #ifdef CONFIG_SCHEDSTATS
2833 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2834 static bool __initdata __sched_schedstats = false;
2836 static void set_schedstats(bool enabled)
2839 static_branch_enable(&sched_schedstats);
2841 static_branch_disable(&sched_schedstats);
2844 void force_schedstat_enabled(void)
2846 if (!schedstat_enabled()) {
2847 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2848 static_branch_enable(&sched_schedstats);
2852 static int __init setup_schedstats(char *str)
2859 * This code is called before jump labels have been set up, so we can't
2860 * change the static branch directly just yet. Instead set a temporary
2861 * variable so init_schedstats() can do it later.
2863 if (!strcmp(str, "enable")) {
2864 __sched_schedstats = true;
2866 } else if (!strcmp(str, "disable")) {
2867 __sched_schedstats = false;
2872 pr_warn("Unable to parse schedstats=\n");
2876 __setup("schedstats=", setup_schedstats);
2878 static void __init init_schedstats(void)
2880 set_schedstats(__sched_schedstats);
2883 #ifdef CONFIG_PROC_SYSCTL
2884 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
2885 size_t *lenp, loff_t *ppos)
2889 int state = static_branch_likely(&sched_schedstats);
2891 if (write && !capable(CAP_SYS_ADMIN))
2896 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2900 set_schedstats(state);
2903 #endif /* CONFIG_PROC_SYSCTL */
2904 #else /* !CONFIG_SCHEDSTATS */
2905 static inline void init_schedstats(void) {}
2906 #endif /* CONFIG_SCHEDSTATS */
2909 * fork()/clone()-time setup:
2911 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2913 unsigned long flags;
2915 __sched_fork(clone_flags, p);
2917 * We mark the process as NEW here. This guarantees that
2918 * nobody will actually run it, and a signal or other external
2919 * event cannot wake it up and insert it on the runqueue either.
2921 p->state = TASK_NEW;
2924 * Make sure we do not leak PI boosting priority to the child.
2926 p->prio = current->normal_prio;
2931 * Revert to default priority/policy on fork if requested.
2933 if (unlikely(p->sched_reset_on_fork)) {
2934 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2935 p->policy = SCHED_NORMAL;
2936 p->static_prio = NICE_TO_PRIO(0);
2938 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2939 p->static_prio = NICE_TO_PRIO(0);
2941 p->prio = p->normal_prio = __normal_prio(p);
2942 set_load_weight(p, false);
2945 * We don't need the reset flag anymore after the fork. It has
2946 * fulfilled its duty:
2948 p->sched_reset_on_fork = 0;
2951 if (dl_prio(p->prio))
2953 else if (rt_prio(p->prio))
2954 p->sched_class = &rt_sched_class;
2956 p->sched_class = &fair_sched_class;
2958 init_entity_runnable_average(&p->se);
2961 * The child is not yet in the pid-hash so no cgroup attach races,
2962 * and the cgroup is pinned to this child due to cgroup_fork()
2963 * is ran before sched_fork().
2965 * Silence PROVE_RCU.
2967 raw_spin_lock_irqsave(&p->pi_lock, flags);
2970 * We're setting the CPU for the first time, we don't migrate,
2971 * so use __set_task_cpu().
2973 __set_task_cpu(p, smp_processor_id());
2974 if (p->sched_class->task_fork)
2975 p->sched_class->task_fork(p);
2976 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2978 #ifdef CONFIG_SCHED_INFO
2979 if (likely(sched_info_on()))
2980 memset(&p->sched_info, 0, sizeof(p->sched_info));
2982 #if defined(CONFIG_SMP)
2985 init_task_preempt_count(p);
2987 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2988 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2993 unsigned long to_ratio(u64 period, u64 runtime)
2995 if (runtime == RUNTIME_INF)
2999 * Doing this here saves a lot of checks in all
3000 * the calling paths, and returning zero seems
3001 * safe for them anyway.
3006 return div64_u64(runtime << BW_SHIFT, period);
3010 * wake_up_new_task - wake up a newly created task for the first time.
3012 * This function will do some initial scheduler statistics housekeeping
3013 * that must be done for every newly created context, then puts the task
3014 * on the runqueue and wakes it.
3016 void wake_up_new_task(struct task_struct *p)
3021 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3022 p->state = TASK_RUNNING;
3025 * Fork balancing, do it here and not earlier because:
3026 * - cpus_ptr can change in the fork path
3027 * - any previously selected CPU might disappear through hotplug
3029 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3030 * as we're not fully set-up yet.
3032 p->recent_used_cpu = task_cpu(p);
3034 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
3036 rq = __task_rq_lock(p, &rf);
3037 update_rq_clock(rq);
3038 post_init_entity_util_avg(p);
3040 activate_task(rq, p, ENQUEUE_NOCLOCK);
3041 trace_sched_wakeup_new(p);
3042 check_preempt_curr(rq, p, WF_FORK);
3044 if (p->sched_class->task_woken) {
3046 * Nothing relies on rq->lock after this, so its fine to
3049 rq_unpin_lock(rq, &rf);
3050 p->sched_class->task_woken(rq, p);
3051 rq_repin_lock(rq, &rf);
3054 task_rq_unlock(rq, p, &rf);
3057 #ifdef CONFIG_PREEMPT_NOTIFIERS
3059 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3061 void preempt_notifier_inc(void)
3063 static_branch_inc(&preempt_notifier_key);
3065 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3067 void preempt_notifier_dec(void)
3069 static_branch_dec(&preempt_notifier_key);
3071 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3074 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3075 * @notifier: notifier struct to register
3077 void preempt_notifier_register(struct preempt_notifier *notifier)
3079 if (!static_branch_unlikely(&preempt_notifier_key))
3080 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3082 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3084 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3087 * preempt_notifier_unregister - no longer interested in preemption notifications
3088 * @notifier: notifier struct to unregister
3090 * This is *not* safe to call from within a preemption notifier.
3092 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3094 hlist_del(¬ifier->link);
3096 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3098 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3100 struct preempt_notifier *notifier;
3102 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3103 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3106 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3108 if (static_branch_unlikely(&preempt_notifier_key))
3109 __fire_sched_in_preempt_notifiers(curr);
3113 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3114 struct task_struct *next)
3116 struct preempt_notifier *notifier;
3118 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3119 notifier->ops->sched_out(notifier, next);
3122 static __always_inline void
3123 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3124 struct task_struct *next)
3126 if (static_branch_unlikely(&preempt_notifier_key))
3127 __fire_sched_out_preempt_notifiers(curr, next);
3130 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3132 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3137 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3138 struct task_struct *next)
3142 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3144 static inline void prepare_task(struct task_struct *next)
3148 * Claim the task as running, we do this before switching to it
3149 * such that any running task will have this set.
3155 static inline void finish_task(struct task_struct *prev)
3159 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3160 * We must ensure this doesn't happen until the switch is completely
3163 * In particular, the load of prev->state in finish_task_switch() must
3164 * happen before this.
3166 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3168 smp_store_release(&prev->on_cpu, 0);
3173 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3176 * Since the runqueue lock will be released by the next
3177 * task (which is an invalid locking op but in the case
3178 * of the scheduler it's an obvious special-case), so we
3179 * do an early lockdep release here:
3181 rq_unpin_lock(rq, rf);
3182 spin_release(&rq->lock.dep_map, _THIS_IP_);
3183 #ifdef CONFIG_DEBUG_SPINLOCK
3184 /* this is a valid case when another task releases the spinlock */
3185 rq->lock.owner = next;
3189 static inline void finish_lock_switch(struct rq *rq)
3192 * If we are tracking spinlock dependencies then we have to
3193 * fix up the runqueue lock - which gets 'carried over' from
3194 * prev into current:
3196 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3197 raw_spin_unlock_irq(&rq->lock);
3201 * NOP if the arch has not defined these:
3204 #ifndef prepare_arch_switch
3205 # define prepare_arch_switch(next) do { } while (0)
3208 #ifndef finish_arch_post_lock_switch
3209 # define finish_arch_post_lock_switch() do { } while (0)
3213 * prepare_task_switch - prepare to switch tasks
3214 * @rq: the runqueue preparing to switch
3215 * @prev: the current task that is being switched out
3216 * @next: the task we are going to switch to.
3218 * This is called with the rq lock held and interrupts off. It must
3219 * be paired with a subsequent finish_task_switch after the context
3222 * prepare_task_switch sets up locking and calls architecture specific
3226 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3227 struct task_struct *next)
3229 kcov_prepare_switch(prev);
3230 sched_info_switch(rq, prev, next);
3231 perf_event_task_sched_out(prev, next);
3233 fire_sched_out_preempt_notifiers(prev, next);
3235 prepare_arch_switch(next);
3239 * finish_task_switch - clean up after a task-switch
3240 * @prev: the thread we just switched away from.
3242 * finish_task_switch must be called after the context switch, paired
3243 * with a prepare_task_switch call before the context switch.
3244 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3245 * and do any other architecture-specific cleanup actions.
3247 * Note that we may have delayed dropping an mm in context_switch(). If
3248 * so, we finish that here outside of the runqueue lock. (Doing it
3249 * with the lock held can cause deadlocks; see schedule() for
3252 * The context switch have flipped the stack from under us and restored the
3253 * local variables which were saved when this task called schedule() in the
3254 * past. prev == current is still correct but we need to recalculate this_rq
3255 * because prev may have moved to another CPU.
3257 static struct rq *finish_task_switch(struct task_struct *prev)
3258 __releases(rq->lock)
3260 struct rq *rq = this_rq();
3261 struct mm_struct *mm = rq->prev_mm;
3265 * The previous task will have left us with a preempt_count of 2
3266 * because it left us after:
3269 * preempt_disable(); // 1
3271 * raw_spin_lock_irq(&rq->lock) // 2
3273 * Also, see FORK_PREEMPT_COUNT.
3275 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3276 "corrupted preempt_count: %s/%d/0x%x\n",
3277 current->comm, current->pid, preempt_count()))
3278 preempt_count_set(FORK_PREEMPT_COUNT);
3283 * A task struct has one reference for the use as "current".
3284 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3285 * schedule one last time. The schedule call will never return, and
3286 * the scheduled task must drop that reference.
3288 * We must observe prev->state before clearing prev->on_cpu (in
3289 * finish_task), otherwise a concurrent wakeup can get prev
3290 * running on another CPU and we could rave with its RUNNING -> DEAD
3291 * transition, resulting in a double drop.
3293 prev_state = prev->state;
3294 vtime_task_switch(prev);
3295 perf_event_task_sched_in(prev, current);
3297 finish_lock_switch(rq);
3298 finish_arch_post_lock_switch();
3299 kcov_finish_switch(current);
3301 fire_sched_in_preempt_notifiers(current);
3303 * When switching through a kernel thread, the loop in
3304 * membarrier_{private,global}_expedited() may have observed that
3305 * kernel thread and not issued an IPI. It is therefore possible to
3306 * schedule between user->kernel->user threads without passing though
3307 * switch_mm(). Membarrier requires a barrier after storing to
3308 * rq->curr, before returning to userspace, so provide them here:
3310 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3311 * provided by mmdrop(),
3312 * - a sync_core for SYNC_CORE.
3315 membarrier_mm_sync_core_before_usermode(mm);
3318 if (unlikely(prev_state == TASK_DEAD)) {
3319 if (prev->sched_class->task_dead)
3320 prev->sched_class->task_dead(prev);
3323 * Remove function-return probe instances associated with this
3324 * task and put them back on the free list.
3326 kprobe_flush_task(prev);
3328 /* Task is done with its stack. */
3329 put_task_stack(prev);
3331 put_task_struct_rcu_user(prev);
3334 tick_nohz_task_switch();
3340 /* rq->lock is NOT held, but preemption is disabled */
3341 static void __balance_callback(struct rq *rq)
3343 struct callback_head *head, *next;
3344 void (*func)(struct rq *rq);
3345 unsigned long flags;
3347 raw_spin_lock_irqsave(&rq->lock, flags);
3348 head = rq->balance_callback;
3349 rq->balance_callback = NULL;
3351 func = (void (*)(struct rq *))head->func;
3358 raw_spin_unlock_irqrestore(&rq->lock, flags);
3361 static inline void balance_callback(struct rq *rq)
3363 if (unlikely(rq->balance_callback))
3364 __balance_callback(rq);
3369 static inline void balance_callback(struct rq *rq)
3376 * schedule_tail - first thing a freshly forked thread must call.
3377 * @prev: the thread we just switched away from.
3379 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3380 __releases(rq->lock)
3385 * New tasks start with FORK_PREEMPT_COUNT, see there and
3386 * finish_task_switch() for details.
3388 * finish_task_switch() will drop rq->lock() and lower preempt_count
3389 * and the preempt_enable() will end up enabling preemption (on
3390 * PREEMPT_COUNT kernels).
3393 rq = finish_task_switch(prev);
3394 balance_callback(rq);
3397 if (current->set_child_tid)
3398 put_user(task_pid_vnr(current), current->set_child_tid);
3400 calculate_sigpending();
3404 * context_switch - switch to the new MM and the new thread's register state.
3406 static __always_inline struct rq *
3407 context_switch(struct rq *rq, struct task_struct *prev,
3408 struct task_struct *next, struct rq_flags *rf)
3410 prepare_task_switch(rq, prev, next);
3413 * For paravirt, this is coupled with an exit in switch_to to
3414 * combine the page table reload and the switch backend into
3417 arch_start_context_switch(prev);
3420 * kernel -> kernel lazy + transfer active
3421 * user -> kernel lazy + mmgrab() active
3423 * kernel -> user switch + mmdrop() active
3424 * user -> user switch
3426 if (!next->mm) { // to kernel
3427 enter_lazy_tlb(prev->active_mm, next);
3429 next->active_mm = prev->active_mm;
3430 if (prev->mm) // from user
3431 mmgrab(prev->active_mm);
3433 prev->active_mm = NULL;
3435 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3437 * sys_membarrier() requires an smp_mb() between setting
3438 * rq->curr / membarrier_switch_mm() and returning to userspace.
3440 * The below provides this either through switch_mm(), or in
3441 * case 'prev->active_mm == next->mm' through
3442 * finish_task_switch()'s mmdrop().
3444 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3446 if (!prev->mm) { // from kernel
3447 /* will mmdrop() in finish_task_switch(). */
3448 rq->prev_mm = prev->active_mm;
3449 prev->active_mm = NULL;
3453 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3455 prepare_lock_switch(rq, next, rf);
3457 /* Here we just switch the register state and the stack. */
3458 switch_to(prev, next, prev);
3461 return finish_task_switch(prev);
3465 * nr_running and nr_context_switches:
3467 * externally visible scheduler statistics: current number of runnable
3468 * threads, total number of context switches performed since bootup.
3470 unsigned long nr_running(void)
3472 unsigned long i, sum = 0;
3474 for_each_online_cpu(i)
3475 sum += cpu_rq(i)->nr_running;
3481 * Check if only the current task is running on the CPU.
3483 * Caution: this function does not check that the caller has disabled
3484 * preemption, thus the result might have a time-of-check-to-time-of-use
3485 * race. The caller is responsible to use it correctly, for example:
3487 * - from a non-preemptible section (of course)
3489 * - from a thread that is bound to a single CPU
3491 * - in a loop with very short iterations (e.g. a polling loop)
3493 bool single_task_running(void)
3495 return raw_rq()->nr_running == 1;
3497 EXPORT_SYMBOL(single_task_running);
3499 unsigned long long nr_context_switches(void)
3502 unsigned long long sum = 0;
3504 for_each_possible_cpu(i)
3505 sum += cpu_rq(i)->nr_switches;
3511 * Consumers of these two interfaces, like for example the cpuidle menu
3512 * governor, are using nonsensical data. Preferring shallow idle state selection
3513 * for a CPU that has IO-wait which might not even end up running the task when
3514 * it does become runnable.
3517 unsigned long nr_iowait_cpu(int cpu)
3519 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3523 * IO-wait accounting, and how its mostly bollocks (on SMP).
3525 * The idea behind IO-wait account is to account the idle time that we could
3526 * have spend running if it were not for IO. That is, if we were to improve the
3527 * storage performance, we'd have a proportional reduction in IO-wait time.
3529 * This all works nicely on UP, where, when a task blocks on IO, we account
3530 * idle time as IO-wait, because if the storage were faster, it could've been
3531 * running and we'd not be idle.
3533 * This has been extended to SMP, by doing the same for each CPU. This however
3536 * Imagine for instance the case where two tasks block on one CPU, only the one
3537 * CPU will have IO-wait accounted, while the other has regular idle. Even
3538 * though, if the storage were faster, both could've ran at the same time,
3539 * utilising both CPUs.
3541 * This means, that when looking globally, the current IO-wait accounting on
3542 * SMP is a lower bound, by reason of under accounting.
3544 * Worse, since the numbers are provided per CPU, they are sometimes
3545 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3546 * associated with any one particular CPU, it can wake to another CPU than it
3547 * blocked on. This means the per CPU IO-wait number is meaningless.
3549 * Task CPU affinities can make all that even more 'interesting'.
3552 unsigned long nr_iowait(void)
3554 unsigned long i, sum = 0;
3556 for_each_possible_cpu(i)
3557 sum += nr_iowait_cpu(i);
3565 * sched_exec - execve() is a valuable balancing opportunity, because at
3566 * this point the task has the smallest effective memory and cache footprint.
3568 void sched_exec(void)
3570 struct task_struct *p = current;
3571 unsigned long flags;
3574 raw_spin_lock_irqsave(&p->pi_lock, flags);
3575 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3576 if (dest_cpu == smp_processor_id())
3579 if (likely(cpu_active(dest_cpu))) {
3580 struct migration_arg arg = { p, dest_cpu };
3582 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3583 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3587 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3592 DEFINE_PER_CPU(struct kernel_stat, kstat);
3593 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3595 EXPORT_PER_CPU_SYMBOL(kstat);
3596 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3599 * The function fair_sched_class.update_curr accesses the struct curr
3600 * and its field curr->exec_start; when called from task_sched_runtime(),
3601 * we observe a high rate of cache misses in practice.
3602 * Prefetching this data results in improved performance.
3604 static inline void prefetch_curr_exec_start(struct task_struct *p)
3606 #ifdef CONFIG_FAIR_GROUP_SCHED
3607 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3609 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3612 prefetch(&curr->exec_start);
3616 * Return accounted runtime for the task.
3617 * In case the task is currently running, return the runtime plus current's
3618 * pending runtime that have not been accounted yet.
3620 unsigned long long task_sched_runtime(struct task_struct *p)
3626 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3628 * 64-bit doesn't need locks to atomically read a 64-bit value.
3629 * So we have a optimization chance when the task's delta_exec is 0.
3630 * Reading ->on_cpu is racy, but this is ok.
3632 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3633 * If we race with it entering CPU, unaccounted time is 0. This is
3634 * indistinguishable from the read occurring a few cycles earlier.
3635 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3636 * been accounted, so we're correct here as well.
3638 if (!p->on_cpu || !task_on_rq_queued(p))
3639 return p->se.sum_exec_runtime;
3642 rq = task_rq_lock(p, &rf);
3644 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3645 * project cycles that may never be accounted to this
3646 * thread, breaking clock_gettime().
3648 if (task_current(rq, p) && task_on_rq_queued(p)) {
3649 prefetch_curr_exec_start(p);
3650 update_rq_clock(rq);
3651 p->sched_class->update_curr(rq);
3653 ns = p->se.sum_exec_runtime;
3654 task_rq_unlock(rq, p, &rf);
3659 DEFINE_PER_CPU(unsigned long, thermal_pressure);
3661 void arch_set_thermal_pressure(struct cpumask *cpus,
3662 unsigned long th_pressure)
3666 for_each_cpu(cpu, cpus)
3667 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
3671 * This function gets called by the timer code, with HZ frequency.
3672 * We call it with interrupts disabled.
3674 void scheduler_tick(void)
3676 int cpu = smp_processor_id();
3677 struct rq *rq = cpu_rq(cpu);
3678 struct task_struct *curr = rq->curr;
3680 unsigned long thermal_pressure;
3682 arch_scale_freq_tick();
3687 update_rq_clock(rq);
3688 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
3689 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
3690 curr->sched_class->task_tick(rq, curr, 0);
3691 calc_global_load_tick(rq);
3696 perf_event_task_tick();
3699 rq->idle_balance = idle_cpu(cpu);
3700 trigger_load_balance(rq);
3704 #ifdef CONFIG_NO_HZ_FULL
3709 struct delayed_work work;
3711 /* Values for ->state, see diagram below. */
3712 #define TICK_SCHED_REMOTE_OFFLINE 0
3713 #define TICK_SCHED_REMOTE_OFFLINING 1
3714 #define TICK_SCHED_REMOTE_RUNNING 2
3717 * State diagram for ->state:
3720 * TICK_SCHED_REMOTE_OFFLINE
3723 * | | sched_tick_remote()
3726 * +--TICK_SCHED_REMOTE_OFFLINING
3729 * sched_tick_start() | | sched_tick_stop()
3732 * TICK_SCHED_REMOTE_RUNNING
3735 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3736 * and sched_tick_start() are happy to leave the state in RUNNING.
3739 static struct tick_work __percpu *tick_work_cpu;
3741 static void sched_tick_remote(struct work_struct *work)
3743 struct delayed_work *dwork = to_delayed_work(work);
3744 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3745 int cpu = twork->cpu;
3746 struct rq *rq = cpu_rq(cpu);
3747 struct task_struct *curr;
3753 * Handle the tick only if it appears the remote CPU is running in full
3754 * dynticks mode. The check is racy by nature, but missing a tick or
3755 * having one too much is no big deal because the scheduler tick updates
3756 * statistics and checks timeslices in a time-independent way, regardless
3757 * of when exactly it is running.
3759 if (!tick_nohz_tick_stopped_cpu(cpu))
3762 rq_lock_irq(rq, &rf);
3764 if (cpu_is_offline(cpu))
3767 update_rq_clock(rq);
3769 if (!is_idle_task(curr)) {
3771 * Make sure the next tick runs within a reasonable
3774 delta = rq_clock_task(rq) - curr->se.exec_start;
3775 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3777 curr->sched_class->task_tick(rq, curr, 0);
3779 calc_load_nohz_remote(rq);
3781 rq_unlock_irq(rq, &rf);
3785 * Run the remote tick once per second (1Hz). This arbitrary
3786 * frequency is large enough to avoid overload but short enough
3787 * to keep scheduler internal stats reasonably up to date. But
3788 * first update state to reflect hotplug activity if required.
3790 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3791 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3792 if (os == TICK_SCHED_REMOTE_RUNNING)
3793 queue_delayed_work(system_unbound_wq, dwork, HZ);
3796 static void sched_tick_start(int cpu)
3799 struct tick_work *twork;
3801 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3804 WARN_ON_ONCE(!tick_work_cpu);
3806 twork = per_cpu_ptr(tick_work_cpu, cpu);
3807 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3808 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3809 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3811 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3812 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3816 #ifdef CONFIG_HOTPLUG_CPU
3817 static void sched_tick_stop(int cpu)
3819 struct tick_work *twork;
3822 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3825 WARN_ON_ONCE(!tick_work_cpu);
3827 twork = per_cpu_ptr(tick_work_cpu, cpu);
3828 /* There cannot be competing actions, but don't rely on stop-machine. */
3829 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3830 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3831 /* Don't cancel, as this would mess up the state machine. */
3833 #endif /* CONFIG_HOTPLUG_CPU */
3835 int __init sched_tick_offload_init(void)
3837 tick_work_cpu = alloc_percpu(struct tick_work);
3838 BUG_ON(!tick_work_cpu);
3842 #else /* !CONFIG_NO_HZ_FULL */
3843 static inline void sched_tick_start(int cpu) { }
3844 static inline void sched_tick_stop(int cpu) { }
3847 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3848 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3850 * If the value passed in is equal to the current preempt count
3851 * then we just disabled preemption. Start timing the latency.
3853 static inline void preempt_latency_start(int val)
3855 if (preempt_count() == val) {
3856 unsigned long ip = get_lock_parent_ip();
3857 #ifdef CONFIG_DEBUG_PREEMPT
3858 current->preempt_disable_ip = ip;
3860 trace_preempt_off(CALLER_ADDR0, ip);
3864 void preempt_count_add(int val)
3866 #ifdef CONFIG_DEBUG_PREEMPT
3870 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3873 __preempt_count_add(val);
3874 #ifdef CONFIG_DEBUG_PREEMPT
3876 * Spinlock count overflowing soon?
3878 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3881 preempt_latency_start(val);
3883 EXPORT_SYMBOL(preempt_count_add);
3884 NOKPROBE_SYMBOL(preempt_count_add);
3887 * If the value passed in equals to the current preempt count
3888 * then we just enabled preemption. Stop timing the latency.
3890 static inline void preempt_latency_stop(int val)
3892 if (preempt_count() == val)
3893 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3896 void preempt_count_sub(int val)
3898 #ifdef CONFIG_DEBUG_PREEMPT
3902 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3905 * Is the spinlock portion underflowing?
3907 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3908 !(preempt_count() & PREEMPT_MASK)))
3912 preempt_latency_stop(val);
3913 __preempt_count_sub(val);
3915 EXPORT_SYMBOL(preempt_count_sub);
3916 NOKPROBE_SYMBOL(preempt_count_sub);
3919 static inline void preempt_latency_start(int val) { }
3920 static inline void preempt_latency_stop(int val) { }
3923 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3925 #ifdef CONFIG_DEBUG_PREEMPT
3926 return p->preempt_disable_ip;
3933 * Print scheduling while atomic bug:
3935 static noinline void __schedule_bug(struct task_struct *prev)
3937 /* Save this before calling printk(), since that will clobber it */
3938 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3940 if (oops_in_progress)
3943 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3944 prev->comm, prev->pid, preempt_count());
3946 debug_show_held_locks(prev);
3948 if (irqs_disabled())
3949 print_irqtrace_events(prev);
3950 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3951 && in_atomic_preempt_off()) {
3952 pr_err("Preemption disabled at:");
3953 print_ip_sym(KERN_ERR, preempt_disable_ip);
3956 panic("scheduling while atomic\n");
3959 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3963 * Various schedule()-time debugging checks and statistics:
3965 static inline void schedule_debug(struct task_struct *prev, bool preempt)
3967 #ifdef CONFIG_SCHED_STACK_END_CHECK
3968 if (task_stack_end_corrupted(prev))
3969 panic("corrupted stack end detected inside scheduler\n");
3971 if (task_scs_end_corrupted(prev))
3972 panic("corrupted shadow stack detected inside scheduler\n");
3975 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3976 if (!preempt && prev->state && prev->non_block_count) {
3977 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3978 prev->comm, prev->pid, prev->non_block_count);
3980 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3984 if (unlikely(in_atomic_preempt_off())) {
3985 __schedule_bug(prev);
3986 preempt_count_set(PREEMPT_DISABLED);
3990 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3992 schedstat_inc(this_rq()->sched_count);
3995 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
3996 struct rq_flags *rf)
3999 const struct sched_class *class;
4001 * We must do the balancing pass before put_prev_task(), such
4002 * that when we release the rq->lock the task is in the same
4003 * state as before we took rq->lock.
4005 * We can terminate the balance pass as soon as we know there is
4006 * a runnable task of @class priority or higher.
4008 for_class_range(class, prev->sched_class, &idle_sched_class) {
4009 if (class->balance(rq, prev, rf))
4014 put_prev_task(rq, prev);
4018 * Pick up the highest-prio task:
4020 static inline struct task_struct *
4021 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4023 const struct sched_class *class;
4024 struct task_struct *p;
4027 * Optimization: we know that if all tasks are in the fair class we can
4028 * call that function directly, but only if the @prev task wasn't of a
4029 * higher scheduling class, because otherwise those loose the
4030 * opportunity to pull in more work from other CPUs.
4032 if (likely((prev->sched_class == &idle_sched_class ||
4033 prev->sched_class == &fair_sched_class) &&
4034 rq->nr_running == rq->cfs.h_nr_running)) {
4036 p = pick_next_task_fair(rq, prev, rf);
4037 if (unlikely(p == RETRY_TASK))
4040 /* Assumes fair_sched_class->next == idle_sched_class */
4042 put_prev_task(rq, prev);
4043 p = pick_next_task_idle(rq);
4050 put_prev_task_balance(rq, prev, rf);
4052 for_each_class(class) {
4053 p = class->pick_next_task(rq);
4058 /* The idle class should always have a runnable task: */
4063 * __schedule() is the main scheduler function.
4065 * The main means of driving the scheduler and thus entering this function are:
4067 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4069 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4070 * paths. For example, see arch/x86/entry_64.S.
4072 * To drive preemption between tasks, the scheduler sets the flag in timer
4073 * interrupt handler scheduler_tick().
4075 * 3. Wakeups don't really cause entry into schedule(). They add a
4076 * task to the run-queue and that's it.
4078 * Now, if the new task added to the run-queue preempts the current
4079 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4080 * called on the nearest possible occasion:
4082 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4084 * - in syscall or exception context, at the next outmost
4085 * preempt_enable(). (this might be as soon as the wake_up()'s
4088 * - in IRQ context, return from interrupt-handler to
4089 * preemptible context
4091 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4094 * - cond_resched() call
4095 * - explicit schedule() call
4096 * - return from syscall or exception to user-space
4097 * - return from interrupt-handler to user-space
4099 * WARNING: must be called with preemption disabled!
4101 static void __sched notrace __schedule(bool preempt)
4103 struct task_struct *prev, *next;
4104 unsigned long *switch_count;
4105 unsigned long prev_state;
4110 cpu = smp_processor_id();
4114 schedule_debug(prev, preempt);
4116 if (sched_feat(HRTICK))
4119 local_irq_disable();
4120 rcu_note_context_switch(preempt);
4122 /* See deactivate_task() below. */
4123 prev_state = prev->state;
4126 * Make sure that signal_pending_state()->signal_pending() below
4127 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4128 * done by the caller to avoid the race with signal_wake_up():
4130 * __set_current_state(@state) signal_wake_up()
4131 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4132 * wake_up_state(p, state)
4133 * LOCK rq->lock LOCK p->pi_state
4134 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4135 * if (signal_pending_state()) if (p->state & @state)
4137 * Also, the membarrier system call requires a full memory barrier
4138 * after coming from user-space, before storing to rq->curr.
4141 smp_mb__after_spinlock();
4143 /* Promote REQ to ACT */
4144 rq->clock_update_flags <<= 1;
4145 update_rq_clock(rq);
4147 switch_count = &prev->nivcsw;
4149 * We must re-load prev->state in case ttwu_remote() changed it
4150 * before we acquired rq->lock.
4152 if (!preempt && prev_state && prev_state == prev->state) {
4153 if (signal_pending_state(prev_state, prev)) {
4154 prev->state = TASK_RUNNING;
4156 prev->sched_contributes_to_load =
4157 (prev_state & TASK_UNINTERRUPTIBLE) &&
4158 !(prev_state & TASK_NOLOAD) &&
4159 !(prev->flags & PF_FROZEN);
4161 if (prev->sched_contributes_to_load)
4162 rq->nr_uninterruptible++;
4165 * __schedule() ttwu()
4166 * prev_state = prev->state; if (READ_ONCE(p->on_rq) && ...)
4167 * LOCK rq->lock goto out;
4168 * smp_mb__after_spinlock(); smp_acquire__after_ctrl_dep();
4169 * p->on_rq = 0; p->state = TASK_WAKING;
4171 * After this, schedule() must not care about p->state any more.
4173 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4175 if (prev->in_iowait) {
4176 atomic_inc(&rq->nr_iowait);
4177 delayacct_blkio_start();
4180 switch_count = &prev->nvcsw;
4183 next = pick_next_task(rq, prev, &rf);
4184 clear_tsk_need_resched(prev);
4185 clear_preempt_need_resched();
4187 if (likely(prev != next)) {
4190 * RCU users of rcu_dereference(rq->curr) may not see
4191 * changes to task_struct made by pick_next_task().
4193 RCU_INIT_POINTER(rq->curr, next);
4195 * The membarrier system call requires each architecture
4196 * to have a full memory barrier after updating
4197 * rq->curr, before returning to user-space.
4199 * Here are the schemes providing that barrier on the
4200 * various architectures:
4201 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4202 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4203 * - finish_lock_switch() for weakly-ordered
4204 * architectures where spin_unlock is a full barrier,
4205 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4206 * is a RELEASE barrier),
4210 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
4212 trace_sched_switch(preempt, prev, next);
4214 /* Also unlocks the rq: */
4215 rq = context_switch(rq, prev, next, &rf);
4217 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4218 rq_unlock_irq(rq, &rf);
4221 balance_callback(rq);
4224 void __noreturn do_task_dead(void)
4226 /* Causes final put_task_struct in finish_task_switch(): */
4227 set_special_state(TASK_DEAD);
4229 /* Tell freezer to ignore us: */
4230 current->flags |= PF_NOFREEZE;
4235 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4240 static inline void sched_submit_work(struct task_struct *tsk)
4246 * If a worker went to sleep, notify and ask workqueue whether
4247 * it wants to wake up a task to maintain concurrency.
4248 * As this function is called inside the schedule() context,
4249 * we disable preemption to avoid it calling schedule() again
4250 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4253 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4255 if (tsk->flags & PF_WQ_WORKER)
4256 wq_worker_sleeping(tsk);
4258 io_wq_worker_sleeping(tsk);
4259 preempt_enable_no_resched();
4262 if (tsk_is_pi_blocked(tsk))
4266 * If we are going to sleep and we have plugged IO queued,
4267 * make sure to submit it to avoid deadlocks.
4269 if (blk_needs_flush_plug(tsk))
4270 blk_schedule_flush_plug(tsk);
4273 static void sched_update_worker(struct task_struct *tsk)
4275 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4276 if (tsk->flags & PF_WQ_WORKER)
4277 wq_worker_running(tsk);
4279 io_wq_worker_running(tsk);
4283 asmlinkage __visible void __sched schedule(void)
4285 struct task_struct *tsk = current;
4287 sched_submit_work(tsk);
4291 sched_preempt_enable_no_resched();
4292 } while (need_resched());
4293 sched_update_worker(tsk);
4295 EXPORT_SYMBOL(schedule);
4298 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4299 * state (have scheduled out non-voluntarily) by making sure that all
4300 * tasks have either left the run queue or have gone into user space.
4301 * As idle tasks do not do either, they must not ever be preempted
4302 * (schedule out non-voluntarily).
4304 * schedule_idle() is similar to schedule_preempt_disable() except that it
4305 * never enables preemption because it does not call sched_submit_work().
4307 void __sched schedule_idle(void)
4310 * As this skips calling sched_submit_work(), which the idle task does
4311 * regardless because that function is a nop when the task is in a
4312 * TASK_RUNNING state, make sure this isn't used someplace that the
4313 * current task can be in any other state. Note, idle is always in the
4314 * TASK_RUNNING state.
4316 WARN_ON_ONCE(current->state);
4319 } while (need_resched());
4322 #ifdef CONFIG_CONTEXT_TRACKING
4323 asmlinkage __visible void __sched schedule_user(void)
4326 * If we come here after a random call to set_need_resched(),
4327 * or we have been woken up remotely but the IPI has not yet arrived,
4328 * we haven't yet exited the RCU idle mode. Do it here manually until
4329 * we find a better solution.
4331 * NB: There are buggy callers of this function. Ideally we
4332 * should warn if prev_state != CONTEXT_USER, but that will trigger
4333 * too frequently to make sense yet.
4335 enum ctx_state prev_state = exception_enter();
4337 exception_exit(prev_state);
4342 * schedule_preempt_disabled - called with preemption disabled
4344 * Returns with preemption disabled. Note: preempt_count must be 1
4346 void __sched schedule_preempt_disabled(void)
4348 sched_preempt_enable_no_resched();
4353 static void __sched notrace preempt_schedule_common(void)
4357 * Because the function tracer can trace preempt_count_sub()
4358 * and it also uses preempt_enable/disable_notrace(), if
4359 * NEED_RESCHED is set, the preempt_enable_notrace() called
4360 * by the function tracer will call this function again and
4361 * cause infinite recursion.
4363 * Preemption must be disabled here before the function
4364 * tracer can trace. Break up preempt_disable() into two
4365 * calls. One to disable preemption without fear of being
4366 * traced. The other to still record the preemption latency,
4367 * which can also be traced by the function tracer.
4369 preempt_disable_notrace();
4370 preempt_latency_start(1);
4372 preempt_latency_stop(1);
4373 preempt_enable_no_resched_notrace();
4376 * Check again in case we missed a preemption opportunity
4377 * between schedule and now.
4379 } while (need_resched());
4382 #ifdef CONFIG_PREEMPTION
4384 * This is the entry point to schedule() from in-kernel preemption
4385 * off of preempt_enable.
4387 asmlinkage __visible void __sched notrace preempt_schedule(void)
4390 * If there is a non-zero preempt_count or interrupts are disabled,
4391 * we do not want to preempt the current task. Just return..
4393 if (likely(!preemptible()))
4396 preempt_schedule_common();
4398 NOKPROBE_SYMBOL(preempt_schedule);
4399 EXPORT_SYMBOL(preempt_schedule);
4402 * preempt_schedule_notrace - preempt_schedule called by tracing
4404 * The tracing infrastructure uses preempt_enable_notrace to prevent
4405 * recursion and tracing preempt enabling caused by the tracing
4406 * infrastructure itself. But as tracing can happen in areas coming
4407 * from userspace or just about to enter userspace, a preempt enable
4408 * can occur before user_exit() is called. This will cause the scheduler
4409 * to be called when the system is still in usermode.
4411 * To prevent this, the preempt_enable_notrace will use this function
4412 * instead of preempt_schedule() to exit user context if needed before
4413 * calling the scheduler.
4415 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4417 enum ctx_state prev_ctx;
4419 if (likely(!preemptible()))
4424 * Because the function tracer can trace preempt_count_sub()
4425 * and it also uses preempt_enable/disable_notrace(), if
4426 * NEED_RESCHED is set, the preempt_enable_notrace() called
4427 * by the function tracer will call this function again and
4428 * cause infinite recursion.
4430 * Preemption must be disabled here before the function
4431 * tracer can trace. Break up preempt_disable() into two
4432 * calls. One to disable preemption without fear of being
4433 * traced. The other to still record the preemption latency,
4434 * which can also be traced by the function tracer.
4436 preempt_disable_notrace();
4437 preempt_latency_start(1);
4439 * Needs preempt disabled in case user_exit() is traced
4440 * and the tracer calls preempt_enable_notrace() causing
4441 * an infinite recursion.
4443 prev_ctx = exception_enter();
4445 exception_exit(prev_ctx);
4447 preempt_latency_stop(1);
4448 preempt_enable_no_resched_notrace();
4449 } while (need_resched());
4451 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4453 #endif /* CONFIG_PREEMPTION */
4456 * This is the entry point to schedule() from kernel preemption
4457 * off of irq context.
4458 * Note, that this is called and return with irqs disabled. This will
4459 * protect us against recursive calling from irq.
4461 asmlinkage __visible void __sched preempt_schedule_irq(void)
4463 enum ctx_state prev_state;
4465 /* Catch callers which need to be fixed */
4466 BUG_ON(preempt_count() || !irqs_disabled());
4468 prev_state = exception_enter();
4474 local_irq_disable();
4475 sched_preempt_enable_no_resched();
4476 } while (need_resched());
4478 exception_exit(prev_state);
4481 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4484 return try_to_wake_up(curr->private, mode, wake_flags);
4486 EXPORT_SYMBOL(default_wake_function);
4488 #ifdef CONFIG_RT_MUTEXES
4490 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4493 prio = min(prio, pi_task->prio);
4498 static inline int rt_effective_prio(struct task_struct *p, int prio)
4500 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4502 return __rt_effective_prio(pi_task, prio);
4506 * rt_mutex_setprio - set the current priority of a task
4508 * @pi_task: donor task
4510 * This function changes the 'effective' priority of a task. It does
4511 * not touch ->normal_prio like __setscheduler().
4513 * Used by the rt_mutex code to implement priority inheritance
4514 * logic. Call site only calls if the priority of the task changed.
4516 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4518 int prio, oldprio, queued, running, queue_flag =
4519 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4520 const struct sched_class *prev_class;
4524 /* XXX used to be waiter->prio, not waiter->task->prio */
4525 prio = __rt_effective_prio(pi_task, p->normal_prio);
4528 * If nothing changed; bail early.
4530 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4533 rq = __task_rq_lock(p, &rf);
4534 update_rq_clock(rq);
4536 * Set under pi_lock && rq->lock, such that the value can be used under
4539 * Note that there is loads of tricky to make this pointer cache work
4540 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4541 * ensure a task is de-boosted (pi_task is set to NULL) before the
4542 * task is allowed to run again (and can exit). This ensures the pointer
4543 * points to a blocked task -- which guaratees the task is present.
4545 p->pi_top_task = pi_task;
4548 * For FIFO/RR we only need to set prio, if that matches we're done.
4550 if (prio == p->prio && !dl_prio(prio))
4554 * Idle task boosting is a nono in general. There is one
4555 * exception, when PREEMPT_RT and NOHZ is active:
4557 * The idle task calls get_next_timer_interrupt() and holds
4558 * the timer wheel base->lock on the CPU and another CPU wants
4559 * to access the timer (probably to cancel it). We can safely
4560 * ignore the boosting request, as the idle CPU runs this code
4561 * with interrupts disabled and will complete the lock
4562 * protected section without being interrupted. So there is no
4563 * real need to boost.
4565 if (unlikely(p == rq->idle)) {
4566 WARN_ON(p != rq->curr);
4567 WARN_ON(p->pi_blocked_on);
4571 trace_sched_pi_setprio(p, pi_task);
4574 if (oldprio == prio)
4575 queue_flag &= ~DEQUEUE_MOVE;
4577 prev_class = p->sched_class;
4578 queued = task_on_rq_queued(p);
4579 running = task_current(rq, p);
4581 dequeue_task(rq, p, queue_flag);
4583 put_prev_task(rq, p);
4586 * Boosting condition are:
4587 * 1. -rt task is running and holds mutex A
4588 * --> -dl task blocks on mutex A
4590 * 2. -dl task is running and holds mutex A
4591 * --> -dl task blocks on mutex A and could preempt the
4594 if (dl_prio(prio)) {
4595 if (!dl_prio(p->normal_prio) ||
4596 (pi_task && dl_prio(pi_task->prio) &&
4597 dl_entity_preempt(&pi_task->dl, &p->dl))) {
4598 p->dl.dl_boosted = 1;
4599 queue_flag |= ENQUEUE_REPLENISH;
4601 p->dl.dl_boosted = 0;
4602 p->sched_class = &dl_sched_class;
4603 } else if (rt_prio(prio)) {
4604 if (dl_prio(oldprio))
4605 p->dl.dl_boosted = 0;
4607 queue_flag |= ENQUEUE_HEAD;
4608 p->sched_class = &rt_sched_class;
4610 if (dl_prio(oldprio))
4611 p->dl.dl_boosted = 0;
4612 if (rt_prio(oldprio))
4614 p->sched_class = &fair_sched_class;
4620 enqueue_task(rq, p, queue_flag);
4622 set_next_task(rq, p);
4624 check_class_changed(rq, p, prev_class, oldprio);
4626 /* Avoid rq from going away on us: */
4628 __task_rq_unlock(rq, &rf);
4630 balance_callback(rq);
4634 static inline int rt_effective_prio(struct task_struct *p, int prio)
4640 void set_user_nice(struct task_struct *p, long nice)
4642 bool queued, running;
4647 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4650 * We have to be careful, if called from sys_setpriority(),
4651 * the task might be in the middle of scheduling on another CPU.
4653 rq = task_rq_lock(p, &rf);
4654 update_rq_clock(rq);
4657 * The RT priorities are set via sched_setscheduler(), but we still
4658 * allow the 'normal' nice value to be set - but as expected
4659 * it wont have any effect on scheduling until the task is
4660 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4662 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4663 p->static_prio = NICE_TO_PRIO(nice);
4666 queued = task_on_rq_queued(p);
4667 running = task_current(rq, p);
4669 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4671 put_prev_task(rq, p);
4673 p->static_prio = NICE_TO_PRIO(nice);
4674 set_load_weight(p, true);
4676 p->prio = effective_prio(p);
4679 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4681 set_next_task(rq, p);
4684 * If the task increased its priority or is running and
4685 * lowered its priority, then reschedule its CPU:
4687 p->sched_class->prio_changed(rq, p, old_prio);
4690 task_rq_unlock(rq, p, &rf);
4692 EXPORT_SYMBOL(set_user_nice);
4695 * can_nice - check if a task can reduce its nice value
4699 int can_nice(const struct task_struct *p, const int nice)
4701 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4702 int nice_rlim = nice_to_rlimit(nice);
4704 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4705 capable(CAP_SYS_NICE));
4708 #ifdef __ARCH_WANT_SYS_NICE
4711 * sys_nice - change the priority of the current process.
4712 * @increment: priority increment
4714 * sys_setpriority is a more generic, but much slower function that
4715 * does similar things.
4717 SYSCALL_DEFINE1(nice, int, increment)
4722 * Setpriority might change our priority at the same moment.
4723 * We don't have to worry. Conceptually one call occurs first
4724 * and we have a single winner.
4726 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4727 nice = task_nice(current) + increment;
4729 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4730 if (increment < 0 && !can_nice(current, nice))
4733 retval = security_task_setnice(current, nice);
4737 set_user_nice(current, nice);
4744 * task_prio - return the priority value of a given task.
4745 * @p: the task in question.
4747 * Return: The priority value as seen by users in /proc.
4748 * RT tasks are offset by -200. Normal tasks are centered
4749 * around 0, value goes from -16 to +15.
4751 int task_prio(const struct task_struct *p)
4753 return p->prio - MAX_RT_PRIO;
4757 * idle_cpu - is a given CPU idle currently?
4758 * @cpu: the processor in question.
4760 * Return: 1 if the CPU is currently idle. 0 otherwise.
4762 int idle_cpu(int cpu)
4764 struct rq *rq = cpu_rq(cpu);
4766 if (rq->curr != rq->idle)
4773 if (rq->ttwu_pending)
4781 * available_idle_cpu - is a given CPU idle for enqueuing work.
4782 * @cpu: the CPU in question.
4784 * Return: 1 if the CPU is currently idle. 0 otherwise.
4786 int available_idle_cpu(int cpu)
4791 if (vcpu_is_preempted(cpu))
4798 * idle_task - return the idle task for a given CPU.
4799 * @cpu: the processor in question.
4801 * Return: The idle task for the CPU @cpu.
4803 struct task_struct *idle_task(int cpu)
4805 return cpu_rq(cpu)->idle;
4809 * find_process_by_pid - find a process with a matching PID value.
4810 * @pid: the pid in question.
4812 * The task of @pid, if found. %NULL otherwise.
4814 static struct task_struct *find_process_by_pid(pid_t pid)
4816 return pid ? find_task_by_vpid(pid) : current;
4820 * sched_setparam() passes in -1 for its policy, to let the functions
4821 * it calls know not to change it.
4823 #define SETPARAM_POLICY -1
4825 static void __setscheduler_params(struct task_struct *p,
4826 const struct sched_attr *attr)
4828 int policy = attr->sched_policy;
4830 if (policy == SETPARAM_POLICY)
4835 if (dl_policy(policy))
4836 __setparam_dl(p, attr);
4837 else if (fair_policy(policy))
4838 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4841 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4842 * !rt_policy. Always setting this ensures that things like
4843 * getparam()/getattr() don't report silly values for !rt tasks.
4845 p->rt_priority = attr->sched_priority;
4846 p->normal_prio = normal_prio(p);
4847 set_load_weight(p, true);
4850 /* Actually do priority change: must hold pi & rq lock. */
4851 static void __setscheduler(struct rq *rq, struct task_struct *p,
4852 const struct sched_attr *attr, bool keep_boost)
4855 * If params can't change scheduling class changes aren't allowed
4858 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4861 __setscheduler_params(p, attr);
4864 * Keep a potential priority boosting if called from
4865 * sched_setscheduler().
4867 p->prio = normal_prio(p);
4869 p->prio = rt_effective_prio(p, p->prio);
4871 if (dl_prio(p->prio))
4872 p->sched_class = &dl_sched_class;
4873 else if (rt_prio(p->prio))
4874 p->sched_class = &rt_sched_class;
4876 p->sched_class = &fair_sched_class;
4880 * Check the target process has a UID that matches the current process's:
4882 static bool check_same_owner(struct task_struct *p)
4884 const struct cred *cred = current_cred(), *pcred;
4888 pcred = __task_cred(p);
4889 match = (uid_eq(cred->euid, pcred->euid) ||
4890 uid_eq(cred->euid, pcred->uid));
4895 static int __sched_setscheduler(struct task_struct *p,
4896 const struct sched_attr *attr,
4899 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4900 MAX_RT_PRIO - 1 - attr->sched_priority;
4901 int retval, oldprio, oldpolicy = -1, queued, running;
4902 int new_effective_prio, policy = attr->sched_policy;
4903 const struct sched_class *prev_class;
4906 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4909 /* The pi code expects interrupts enabled */
4910 BUG_ON(pi && in_interrupt());
4912 /* Double check policy once rq lock held: */
4914 reset_on_fork = p->sched_reset_on_fork;
4915 policy = oldpolicy = p->policy;
4917 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4919 if (!valid_policy(policy))
4923 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4927 * Valid priorities for SCHED_FIFO and SCHED_RR are
4928 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4929 * SCHED_BATCH and SCHED_IDLE is 0.
4931 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4932 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4934 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4935 (rt_policy(policy) != (attr->sched_priority != 0)))
4939 * Allow unprivileged RT tasks to decrease priority:
4941 if (user && !capable(CAP_SYS_NICE)) {
4942 if (fair_policy(policy)) {
4943 if (attr->sched_nice < task_nice(p) &&
4944 !can_nice(p, attr->sched_nice))
4948 if (rt_policy(policy)) {
4949 unsigned long rlim_rtprio =
4950 task_rlimit(p, RLIMIT_RTPRIO);
4952 /* Can't set/change the rt policy: */
4953 if (policy != p->policy && !rlim_rtprio)
4956 /* Can't increase priority: */
4957 if (attr->sched_priority > p->rt_priority &&
4958 attr->sched_priority > rlim_rtprio)
4963 * Can't set/change SCHED_DEADLINE policy at all for now
4964 * (safest behavior); in the future we would like to allow
4965 * unprivileged DL tasks to increase their relative deadline
4966 * or reduce their runtime (both ways reducing utilization)
4968 if (dl_policy(policy))
4972 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4973 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4975 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4976 if (!can_nice(p, task_nice(p)))
4980 /* Can't change other user's priorities: */
4981 if (!check_same_owner(p))
4984 /* Normal users shall not reset the sched_reset_on_fork flag: */
4985 if (p->sched_reset_on_fork && !reset_on_fork)
4990 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4993 retval = security_task_setscheduler(p);
4998 /* Update task specific "requested" clamps */
4999 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
5000 retval = uclamp_validate(p, attr);
5009 * Make sure no PI-waiters arrive (or leave) while we are
5010 * changing the priority of the task:
5012 * To be able to change p->policy safely, the appropriate
5013 * runqueue lock must be held.
5015 rq = task_rq_lock(p, &rf);
5016 update_rq_clock(rq);
5019 * Changing the policy of the stop threads its a very bad idea:
5021 if (p == rq->stop) {
5027 * If not changing anything there's no need to proceed further,
5028 * but store a possible modification of reset_on_fork.
5030 if (unlikely(policy == p->policy)) {
5031 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
5033 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
5035 if (dl_policy(policy) && dl_param_changed(p, attr))
5037 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
5040 p->sched_reset_on_fork = reset_on_fork;
5047 #ifdef CONFIG_RT_GROUP_SCHED
5049 * Do not allow realtime tasks into groups that have no runtime
5052 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5053 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5054 !task_group_is_autogroup(task_group(p))) {
5060 if (dl_bandwidth_enabled() && dl_policy(policy) &&
5061 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
5062 cpumask_t *span = rq->rd->span;
5065 * Don't allow tasks with an affinity mask smaller than
5066 * the entire root_domain to become SCHED_DEADLINE. We
5067 * will also fail if there's no bandwidth available.
5069 if (!cpumask_subset(span, p->cpus_ptr) ||
5070 rq->rd->dl_bw.bw == 0) {
5078 /* Re-check policy now with rq lock held: */
5079 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5080 policy = oldpolicy = -1;
5081 task_rq_unlock(rq, p, &rf);
5083 cpuset_read_unlock();
5088 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5089 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5092 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
5097 p->sched_reset_on_fork = reset_on_fork;
5102 * Take priority boosted tasks into account. If the new
5103 * effective priority is unchanged, we just store the new
5104 * normal parameters and do not touch the scheduler class and
5105 * the runqueue. This will be done when the task deboost
5108 new_effective_prio = rt_effective_prio(p, newprio);
5109 if (new_effective_prio == oldprio)
5110 queue_flags &= ~DEQUEUE_MOVE;
5113 queued = task_on_rq_queued(p);
5114 running = task_current(rq, p);
5116 dequeue_task(rq, p, queue_flags);
5118 put_prev_task(rq, p);
5120 prev_class = p->sched_class;
5122 __setscheduler(rq, p, attr, pi);
5123 __setscheduler_uclamp(p, attr);
5127 * We enqueue to tail when the priority of a task is
5128 * increased (user space view).
5130 if (oldprio < p->prio)
5131 queue_flags |= ENQUEUE_HEAD;
5133 enqueue_task(rq, p, queue_flags);
5136 set_next_task(rq, p);
5138 check_class_changed(rq, p, prev_class, oldprio);
5140 /* Avoid rq from going away on us: */
5142 task_rq_unlock(rq, p, &rf);
5145 cpuset_read_unlock();
5146 rt_mutex_adjust_pi(p);
5149 /* Run balance callbacks after we've adjusted the PI chain: */
5150 balance_callback(rq);
5156 task_rq_unlock(rq, p, &rf);
5158 cpuset_read_unlock();
5162 static int _sched_setscheduler(struct task_struct *p, int policy,
5163 const struct sched_param *param, bool check)
5165 struct sched_attr attr = {
5166 .sched_policy = policy,
5167 .sched_priority = param->sched_priority,
5168 .sched_nice = PRIO_TO_NICE(p->static_prio),
5171 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5172 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5173 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5174 policy &= ~SCHED_RESET_ON_FORK;
5175 attr.sched_policy = policy;
5178 return __sched_setscheduler(p, &attr, check, true);
5181 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5182 * @p: the task in question.
5183 * @policy: new policy.
5184 * @param: structure containing the new RT priority.
5186 * Return: 0 on success. An error code otherwise.
5188 * NOTE that the task may be already dead.
5190 int sched_setscheduler(struct task_struct *p, int policy,
5191 const struct sched_param *param)
5193 return _sched_setscheduler(p, policy, param, true);
5195 EXPORT_SYMBOL_GPL(sched_setscheduler);
5197 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5199 return __sched_setscheduler(p, attr, true, true);
5201 EXPORT_SYMBOL_GPL(sched_setattr);
5203 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5205 return __sched_setscheduler(p, attr, false, true);
5209 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5210 * @p: the task in question.
5211 * @policy: new policy.
5212 * @param: structure containing the new RT priority.
5214 * Just like sched_setscheduler, only don't bother checking if the
5215 * current context has permission. For example, this is needed in
5216 * stop_machine(): we create temporary high priority worker threads,
5217 * but our caller might not have that capability.
5219 * Return: 0 on success. An error code otherwise.
5221 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5222 const struct sched_param *param)
5224 return _sched_setscheduler(p, policy, param, false);
5226 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5229 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5231 struct sched_param lparam;
5232 struct task_struct *p;
5235 if (!param || pid < 0)
5237 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5242 p = find_process_by_pid(pid);
5248 retval = sched_setscheduler(p, policy, &lparam);
5256 * Mimics kernel/events/core.c perf_copy_attr().
5258 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5263 /* Zero the full structure, so that a short copy will be nice: */
5264 memset(attr, 0, sizeof(*attr));
5266 ret = get_user(size, &uattr->size);
5270 /* ABI compatibility quirk: */
5272 size = SCHED_ATTR_SIZE_VER0;
5273 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5276 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5283 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5284 size < SCHED_ATTR_SIZE_VER1)
5288 * XXX: Do we want to be lenient like existing syscalls; or do we want
5289 * to be strict and return an error on out-of-bounds values?
5291 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5296 put_user(sizeof(*attr), &uattr->size);
5301 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5302 * @pid: the pid in question.
5303 * @policy: new policy.
5304 * @param: structure containing the new RT priority.
5306 * Return: 0 on success. An error code otherwise.
5308 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5313 return do_sched_setscheduler(pid, policy, param);
5317 * sys_sched_setparam - set/change the RT priority of a thread
5318 * @pid: the pid in question.
5319 * @param: structure containing the new RT priority.
5321 * Return: 0 on success. An error code otherwise.
5323 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5325 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5329 * sys_sched_setattr - same as above, but with extended sched_attr
5330 * @pid: the pid in question.
5331 * @uattr: structure containing the extended parameters.
5332 * @flags: for future extension.
5334 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5335 unsigned int, flags)
5337 struct sched_attr attr;
5338 struct task_struct *p;
5341 if (!uattr || pid < 0 || flags)
5344 retval = sched_copy_attr(uattr, &attr);
5348 if ((int)attr.sched_policy < 0)
5350 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5351 attr.sched_policy = SETPARAM_POLICY;
5355 p = find_process_by_pid(pid);
5361 retval = sched_setattr(p, &attr);
5369 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5370 * @pid: the pid in question.
5372 * Return: On success, the policy of the thread. Otherwise, a negative error
5375 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5377 struct task_struct *p;
5385 p = find_process_by_pid(pid);
5387 retval = security_task_getscheduler(p);
5390 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5397 * sys_sched_getparam - get the RT priority of a thread
5398 * @pid: the pid in question.
5399 * @param: structure containing the RT priority.
5401 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5404 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5406 struct sched_param lp = { .sched_priority = 0 };
5407 struct task_struct *p;
5410 if (!param || pid < 0)
5414 p = find_process_by_pid(pid);
5419 retval = security_task_getscheduler(p);
5423 if (task_has_rt_policy(p))
5424 lp.sched_priority = p->rt_priority;
5428 * This one might sleep, we cannot do it with a spinlock held ...
5430 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5440 * Copy the kernel size attribute structure (which might be larger
5441 * than what user-space knows about) to user-space.
5443 * Note that all cases are valid: user-space buffer can be larger or
5444 * smaller than the kernel-space buffer. The usual case is that both
5445 * have the same size.
5448 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5449 struct sched_attr *kattr,
5452 unsigned int ksize = sizeof(*kattr);
5454 if (!access_ok(uattr, usize))
5458 * sched_getattr() ABI forwards and backwards compatibility:
5460 * If usize == ksize then we just copy everything to user-space and all is good.
5462 * If usize < ksize then we only copy as much as user-space has space for,
5463 * this keeps ABI compatibility as well. We skip the rest.
5465 * If usize > ksize then user-space is using a newer version of the ABI,
5466 * which part the kernel doesn't know about. Just ignore it - tooling can
5467 * detect the kernel's knowledge of attributes from the attr->size value
5468 * which is set to ksize in this case.
5470 kattr->size = min(usize, ksize);
5472 if (copy_to_user(uattr, kattr, kattr->size))
5479 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5480 * @pid: the pid in question.
5481 * @uattr: structure containing the extended parameters.
5482 * @usize: sizeof(attr) for fwd/bwd comp.
5483 * @flags: for future extension.
5485 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5486 unsigned int, usize, unsigned int, flags)
5488 struct sched_attr kattr = { };
5489 struct task_struct *p;
5492 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5493 usize < SCHED_ATTR_SIZE_VER0 || flags)
5497 p = find_process_by_pid(pid);
5502 retval = security_task_getscheduler(p);
5506 kattr.sched_policy = p->policy;
5507 if (p->sched_reset_on_fork)
5508 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5509 if (task_has_dl_policy(p))
5510 __getparam_dl(p, &kattr);
5511 else if (task_has_rt_policy(p))
5512 kattr.sched_priority = p->rt_priority;
5514 kattr.sched_nice = task_nice(p);
5516 #ifdef CONFIG_UCLAMP_TASK
5517 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5518 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5523 return sched_attr_copy_to_user(uattr, &kattr, usize);
5530 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5532 cpumask_var_t cpus_allowed, new_mask;
5533 struct task_struct *p;
5538 p = find_process_by_pid(pid);
5544 /* Prevent p going away */
5548 if (p->flags & PF_NO_SETAFFINITY) {
5552 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5556 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5558 goto out_free_cpus_allowed;
5561 if (!check_same_owner(p)) {
5563 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5565 goto out_free_new_mask;
5570 retval = security_task_setscheduler(p);
5572 goto out_free_new_mask;
5575 cpuset_cpus_allowed(p, cpus_allowed);
5576 cpumask_and(new_mask, in_mask, cpus_allowed);
5579 * Since bandwidth control happens on root_domain basis,
5580 * if admission test is enabled, we only admit -deadline
5581 * tasks allowed to run on all the CPUs in the task's
5585 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5587 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5590 goto out_free_new_mask;
5596 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5599 cpuset_cpus_allowed(p, cpus_allowed);
5600 if (!cpumask_subset(new_mask, cpus_allowed)) {
5602 * We must have raced with a concurrent cpuset
5603 * update. Just reset the cpus_allowed to the
5604 * cpuset's cpus_allowed
5606 cpumask_copy(new_mask, cpus_allowed);
5611 free_cpumask_var(new_mask);
5612 out_free_cpus_allowed:
5613 free_cpumask_var(cpus_allowed);
5619 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5620 struct cpumask *new_mask)
5622 if (len < cpumask_size())
5623 cpumask_clear(new_mask);
5624 else if (len > cpumask_size())
5625 len = cpumask_size();
5627 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5631 * sys_sched_setaffinity - set the CPU affinity of a process
5632 * @pid: pid of the process
5633 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5634 * @user_mask_ptr: user-space pointer to the new CPU mask
5636 * Return: 0 on success. An error code otherwise.
5638 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5639 unsigned long __user *, user_mask_ptr)
5641 cpumask_var_t new_mask;
5644 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5647 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5649 retval = sched_setaffinity(pid, new_mask);
5650 free_cpumask_var(new_mask);
5654 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5656 struct task_struct *p;
5657 unsigned long flags;
5663 p = find_process_by_pid(pid);
5667 retval = security_task_getscheduler(p);
5671 raw_spin_lock_irqsave(&p->pi_lock, flags);
5672 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5673 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5682 * sys_sched_getaffinity - get the CPU affinity of a process
5683 * @pid: pid of the process
5684 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5685 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5687 * Return: size of CPU mask copied to user_mask_ptr on success. An
5688 * error code otherwise.
5690 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5691 unsigned long __user *, user_mask_ptr)
5696 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5698 if (len & (sizeof(unsigned long)-1))
5701 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5704 ret = sched_getaffinity(pid, mask);
5706 unsigned int retlen = min(len, cpumask_size());
5708 if (copy_to_user(user_mask_ptr, mask, retlen))
5713 free_cpumask_var(mask);
5719 * sys_sched_yield - yield the current processor to other threads.
5721 * This function yields the current CPU to other tasks. If there are no
5722 * other threads running on this CPU then this function will return.
5726 static void do_sched_yield(void)
5731 rq = this_rq_lock_irq(&rf);
5733 schedstat_inc(rq->yld_count);
5734 current->sched_class->yield_task(rq);
5737 * Since we are going to call schedule() anyway, there's
5738 * no need to preempt or enable interrupts:
5742 sched_preempt_enable_no_resched();
5747 SYSCALL_DEFINE0(sched_yield)
5753 #ifndef CONFIG_PREEMPTION
5754 int __sched _cond_resched(void)
5756 if (should_resched(0)) {
5757 preempt_schedule_common();
5763 EXPORT_SYMBOL(_cond_resched);
5767 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5768 * call schedule, and on return reacquire the lock.
5770 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5771 * operations here to prevent schedule() from being called twice (once via
5772 * spin_unlock(), once by hand).
5774 int __cond_resched_lock(spinlock_t *lock)
5776 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5779 lockdep_assert_held(lock);
5781 if (spin_needbreak(lock) || resched) {
5784 preempt_schedule_common();
5792 EXPORT_SYMBOL(__cond_resched_lock);
5795 * yield - yield the current processor to other threads.
5797 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5799 * The scheduler is at all times free to pick the calling task as the most
5800 * eligible task to run, if removing the yield() call from your code breaks
5801 * it, its already broken.
5803 * Typical broken usage is:
5808 * where one assumes that yield() will let 'the other' process run that will
5809 * make event true. If the current task is a SCHED_FIFO task that will never
5810 * happen. Never use yield() as a progress guarantee!!
5812 * If you want to use yield() to wait for something, use wait_event().
5813 * If you want to use yield() to be 'nice' for others, use cond_resched().
5814 * If you still want to use yield(), do not!
5816 void __sched yield(void)
5818 set_current_state(TASK_RUNNING);
5821 EXPORT_SYMBOL(yield);
5824 * yield_to - yield the current processor to another thread in
5825 * your thread group, or accelerate that thread toward the
5826 * processor it's on.
5828 * @preempt: whether task preemption is allowed or not
5830 * It's the caller's job to ensure that the target task struct
5831 * can't go away on us before we can do any checks.
5834 * true (>0) if we indeed boosted the target task.
5835 * false (0) if we failed to boost the target.
5836 * -ESRCH if there's no task to yield to.
5838 int __sched yield_to(struct task_struct *p, bool preempt)
5840 struct task_struct *curr = current;
5841 struct rq *rq, *p_rq;
5842 unsigned long flags;
5845 local_irq_save(flags);
5851 * If we're the only runnable task on the rq and target rq also
5852 * has only one task, there's absolutely no point in yielding.
5854 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5859 double_rq_lock(rq, p_rq);
5860 if (task_rq(p) != p_rq) {
5861 double_rq_unlock(rq, p_rq);
5865 if (!curr->sched_class->yield_to_task)
5868 if (curr->sched_class != p->sched_class)
5871 if (task_running(p_rq, p) || p->state)
5874 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5876 schedstat_inc(rq->yld_count);
5878 * Make p's CPU reschedule; pick_next_entity takes care of
5881 if (preempt && rq != p_rq)
5886 double_rq_unlock(rq, p_rq);
5888 local_irq_restore(flags);
5895 EXPORT_SYMBOL_GPL(yield_to);
5897 int io_schedule_prepare(void)
5899 int old_iowait = current->in_iowait;
5901 current->in_iowait = 1;
5902 blk_schedule_flush_plug(current);
5907 void io_schedule_finish(int token)
5909 current->in_iowait = token;
5913 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5914 * that process accounting knows that this is a task in IO wait state.
5916 long __sched io_schedule_timeout(long timeout)
5921 token = io_schedule_prepare();
5922 ret = schedule_timeout(timeout);
5923 io_schedule_finish(token);
5927 EXPORT_SYMBOL(io_schedule_timeout);
5929 void __sched io_schedule(void)
5933 token = io_schedule_prepare();
5935 io_schedule_finish(token);
5937 EXPORT_SYMBOL(io_schedule);
5940 * sys_sched_get_priority_max - return maximum RT priority.
5941 * @policy: scheduling class.
5943 * Return: On success, this syscall returns the maximum
5944 * rt_priority that can be used by a given scheduling class.
5945 * On failure, a negative error code is returned.
5947 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5954 ret = MAX_USER_RT_PRIO-1;
5956 case SCHED_DEADLINE:
5967 * sys_sched_get_priority_min - return minimum RT priority.
5968 * @policy: scheduling class.
5970 * Return: On success, this syscall returns the minimum
5971 * rt_priority that can be used by a given scheduling class.
5972 * On failure, a negative error code is returned.
5974 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5983 case SCHED_DEADLINE:
5992 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5994 struct task_struct *p;
5995 unsigned int time_slice;
6005 p = find_process_by_pid(pid);
6009 retval = security_task_getscheduler(p);
6013 rq = task_rq_lock(p, &rf);
6015 if (p->sched_class->get_rr_interval)
6016 time_slice = p->sched_class->get_rr_interval(rq, p);
6017 task_rq_unlock(rq, p, &rf);
6020 jiffies_to_timespec64(time_slice, t);
6029 * sys_sched_rr_get_interval - return the default timeslice of a process.
6030 * @pid: pid of the process.
6031 * @interval: userspace pointer to the timeslice value.
6033 * this syscall writes the default timeslice value of a given process
6034 * into the user-space timespec buffer. A value of '0' means infinity.
6036 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6039 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6040 struct __kernel_timespec __user *, interval)
6042 struct timespec64 t;
6043 int retval = sched_rr_get_interval(pid, &t);
6046 retval = put_timespec64(&t, interval);
6051 #ifdef CONFIG_COMPAT_32BIT_TIME
6052 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
6053 struct old_timespec32 __user *, interval)
6055 struct timespec64 t;
6056 int retval = sched_rr_get_interval(pid, &t);
6059 retval = put_old_timespec32(&t, interval);
6064 void sched_show_task(struct task_struct *p)
6066 unsigned long free = 0;
6069 if (!try_get_task_stack(p))
6072 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
6074 if (p->state == TASK_RUNNING)
6075 printk(KERN_CONT " running task ");
6076 #ifdef CONFIG_DEBUG_STACK_USAGE
6077 free = stack_not_used(p);
6082 ppid = task_pid_nr(rcu_dereference(p->real_parent));
6084 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6085 task_pid_nr(p), ppid,
6086 (unsigned long)task_thread_info(p)->flags);
6088 print_worker_info(KERN_INFO, p);
6089 show_stack(p, NULL, KERN_INFO);
6092 EXPORT_SYMBOL_GPL(sched_show_task);
6095 state_filter_match(unsigned long state_filter, struct task_struct *p)
6097 /* no filter, everything matches */
6101 /* filter, but doesn't match */
6102 if (!(p->state & state_filter))
6106 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6109 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
6116 void show_state_filter(unsigned long state_filter)
6118 struct task_struct *g, *p;
6120 #if BITS_PER_LONG == 32
6122 " task PC stack pid father\n");
6125 " task PC stack pid father\n");
6128 for_each_process_thread(g, p) {
6130 * reset the NMI-timeout, listing all files on a slow
6131 * console might take a lot of time:
6132 * Also, reset softlockup watchdogs on all CPUs, because
6133 * another CPU might be blocked waiting for us to process
6136 touch_nmi_watchdog();
6137 touch_all_softlockup_watchdogs();
6138 if (state_filter_match(state_filter, p))
6142 #ifdef CONFIG_SCHED_DEBUG
6144 sysrq_sched_debug_show();
6148 * Only show locks if all tasks are dumped:
6151 debug_show_all_locks();
6155 * init_idle - set up an idle thread for a given CPU
6156 * @idle: task in question
6157 * @cpu: CPU the idle task belongs to
6159 * NOTE: this function does not set the idle thread's NEED_RESCHED
6160 * flag, to make booting more robust.
6162 void init_idle(struct task_struct *idle, int cpu)
6164 struct rq *rq = cpu_rq(cpu);
6165 unsigned long flags;
6167 __sched_fork(0, idle);
6169 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6170 raw_spin_lock(&rq->lock);
6172 idle->state = TASK_RUNNING;
6173 idle->se.exec_start = sched_clock();
6174 idle->flags |= PF_IDLE;
6176 scs_task_reset(idle);
6177 kasan_unpoison_task_stack(idle);
6181 * Its possible that init_idle() gets called multiple times on a task,
6182 * in that case do_set_cpus_allowed() will not do the right thing.
6184 * And since this is boot we can forgo the serialization.
6186 set_cpus_allowed_common(idle, cpumask_of(cpu));
6189 * We're having a chicken and egg problem, even though we are
6190 * holding rq->lock, the CPU isn't yet set to this CPU so the
6191 * lockdep check in task_group() will fail.
6193 * Similar case to sched_fork(). / Alternatively we could
6194 * use task_rq_lock() here and obtain the other rq->lock.
6199 __set_task_cpu(idle, cpu);
6203 rcu_assign_pointer(rq->curr, idle);
6204 idle->on_rq = TASK_ON_RQ_QUEUED;
6208 raw_spin_unlock(&rq->lock);
6209 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6211 /* Set the preempt count _outside_ the spinlocks! */
6212 init_idle_preempt_count(idle, cpu);
6215 * The idle tasks have their own, simple scheduling class:
6217 idle->sched_class = &idle_sched_class;
6218 ftrace_graph_init_idle_task(idle, cpu);
6219 vtime_init_idle(idle, cpu);
6221 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6227 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6228 const struct cpumask *trial)
6232 if (!cpumask_weight(cur))
6235 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6240 int task_can_attach(struct task_struct *p,
6241 const struct cpumask *cs_cpus_allowed)
6246 * Kthreads which disallow setaffinity shouldn't be moved
6247 * to a new cpuset; we don't want to change their CPU
6248 * affinity and isolating such threads by their set of
6249 * allowed nodes is unnecessary. Thus, cpusets are not
6250 * applicable for such threads. This prevents checking for
6251 * success of set_cpus_allowed_ptr() on all attached tasks
6252 * before cpus_mask may be changed.
6254 if (p->flags & PF_NO_SETAFFINITY) {
6259 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6261 ret = dl_task_can_attach(p, cs_cpus_allowed);
6267 bool sched_smp_initialized __read_mostly;
6269 #ifdef CONFIG_NUMA_BALANCING
6270 /* Migrate current task p to target_cpu */
6271 int migrate_task_to(struct task_struct *p, int target_cpu)
6273 struct migration_arg arg = { p, target_cpu };
6274 int curr_cpu = task_cpu(p);
6276 if (curr_cpu == target_cpu)
6279 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6282 /* TODO: This is not properly updating schedstats */
6284 trace_sched_move_numa(p, curr_cpu, target_cpu);
6285 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6289 * Requeue a task on a given node and accurately track the number of NUMA
6290 * tasks on the runqueues
6292 void sched_setnuma(struct task_struct *p, int nid)
6294 bool queued, running;
6298 rq = task_rq_lock(p, &rf);
6299 queued = task_on_rq_queued(p);
6300 running = task_current(rq, p);
6303 dequeue_task(rq, p, DEQUEUE_SAVE);
6305 put_prev_task(rq, p);
6307 p->numa_preferred_nid = nid;
6310 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6312 set_next_task(rq, p);
6313 task_rq_unlock(rq, p, &rf);
6315 #endif /* CONFIG_NUMA_BALANCING */
6317 #ifdef CONFIG_HOTPLUG_CPU
6319 * Ensure that the idle task is using init_mm right before its CPU goes
6322 void idle_task_exit(void)
6324 struct mm_struct *mm = current->active_mm;
6326 BUG_ON(cpu_online(smp_processor_id()));
6327 BUG_ON(current != this_rq()->idle);
6329 if (mm != &init_mm) {
6330 switch_mm(mm, &init_mm, current);
6331 finish_arch_post_lock_switch();
6334 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6338 * Since this CPU is going 'away' for a while, fold any nr_active delta
6339 * we might have. Assumes we're called after migrate_tasks() so that the
6340 * nr_active count is stable. We need to take the teardown thread which
6341 * is calling this into account, so we hand in adjust = 1 to the load
6344 * Also see the comment "Global load-average calculations".
6346 static void calc_load_migrate(struct rq *rq)
6348 long delta = calc_load_fold_active(rq, 1);
6350 atomic_long_add(delta, &calc_load_tasks);
6353 static struct task_struct *__pick_migrate_task(struct rq *rq)
6355 const struct sched_class *class;
6356 struct task_struct *next;
6358 for_each_class(class) {
6359 next = class->pick_next_task(rq);
6361 next->sched_class->put_prev_task(rq, next);
6366 /* The idle class should always have a runnable task */
6371 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6372 * try_to_wake_up()->select_task_rq().
6374 * Called with rq->lock held even though we'er in stop_machine() and
6375 * there's no concurrency possible, we hold the required locks anyway
6376 * because of lock validation efforts.
6378 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6380 struct rq *rq = dead_rq;
6381 struct task_struct *next, *stop = rq->stop;
6382 struct rq_flags orf = *rf;
6386 * Fudge the rq selection such that the below task selection loop
6387 * doesn't get stuck on the currently eligible stop task.
6389 * We're currently inside stop_machine() and the rq is either stuck
6390 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6391 * either way we should never end up calling schedule() until we're
6397 * put_prev_task() and pick_next_task() sched
6398 * class method both need to have an up-to-date
6399 * value of rq->clock[_task]
6401 update_rq_clock(rq);
6405 * There's this thread running, bail when that's the only
6408 if (rq->nr_running == 1)
6411 next = __pick_migrate_task(rq);
6414 * Rules for changing task_struct::cpus_mask are holding
6415 * both pi_lock and rq->lock, such that holding either
6416 * stabilizes the mask.
6418 * Drop rq->lock is not quite as disastrous as it usually is
6419 * because !cpu_active at this point, which means load-balance
6420 * will not interfere. Also, stop-machine.
6423 raw_spin_lock(&next->pi_lock);
6427 * Since we're inside stop-machine, _nothing_ should have
6428 * changed the task, WARN if weird stuff happened, because in
6429 * that case the above rq->lock drop is a fail too.
6431 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6432 raw_spin_unlock(&next->pi_lock);
6436 /* Find suitable destination for @next, with force if needed. */
6437 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6438 rq = __migrate_task(rq, rf, next, dest_cpu);
6439 if (rq != dead_rq) {
6445 raw_spin_unlock(&next->pi_lock);
6450 #endif /* CONFIG_HOTPLUG_CPU */
6452 void set_rq_online(struct rq *rq)
6455 const struct sched_class *class;
6457 cpumask_set_cpu(rq->cpu, rq->rd->online);
6460 for_each_class(class) {
6461 if (class->rq_online)
6462 class->rq_online(rq);
6467 void set_rq_offline(struct rq *rq)
6470 const struct sched_class *class;
6472 for_each_class(class) {
6473 if (class->rq_offline)
6474 class->rq_offline(rq);
6477 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6483 * used to mark begin/end of suspend/resume:
6485 static int num_cpus_frozen;
6488 * Update cpusets according to cpu_active mask. If cpusets are
6489 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6490 * around partition_sched_domains().
6492 * If we come here as part of a suspend/resume, don't touch cpusets because we
6493 * want to restore it back to its original state upon resume anyway.
6495 static void cpuset_cpu_active(void)
6497 if (cpuhp_tasks_frozen) {
6499 * num_cpus_frozen tracks how many CPUs are involved in suspend
6500 * resume sequence. As long as this is not the last online
6501 * operation in the resume sequence, just build a single sched
6502 * domain, ignoring cpusets.
6504 partition_sched_domains(1, NULL, NULL);
6505 if (--num_cpus_frozen)
6508 * This is the last CPU online operation. So fall through and
6509 * restore the original sched domains by considering the
6510 * cpuset configurations.
6512 cpuset_force_rebuild();
6514 cpuset_update_active_cpus();
6517 static int cpuset_cpu_inactive(unsigned int cpu)
6519 if (!cpuhp_tasks_frozen) {
6520 if (dl_cpu_busy(cpu))
6522 cpuset_update_active_cpus();
6525 partition_sched_domains(1, NULL, NULL);
6530 int sched_cpu_activate(unsigned int cpu)
6532 struct rq *rq = cpu_rq(cpu);
6535 #ifdef CONFIG_SCHED_SMT
6537 * When going up, increment the number of cores with SMT present.
6539 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6540 static_branch_inc_cpuslocked(&sched_smt_present);
6542 set_cpu_active(cpu, true);
6544 if (sched_smp_initialized) {
6545 sched_domains_numa_masks_set(cpu);
6546 cpuset_cpu_active();
6550 * Put the rq online, if not already. This happens:
6552 * 1) In the early boot process, because we build the real domains
6553 * after all CPUs have been brought up.
6555 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6558 rq_lock_irqsave(rq, &rf);
6560 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6563 rq_unlock_irqrestore(rq, &rf);
6568 int sched_cpu_deactivate(unsigned int cpu)
6572 set_cpu_active(cpu, false);
6574 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6575 * users of this state to go away such that all new such users will
6578 * Do sync before park smpboot threads to take care the rcu boost case.
6582 #ifdef CONFIG_SCHED_SMT
6584 * When going down, decrement the number of cores with SMT present.
6586 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6587 static_branch_dec_cpuslocked(&sched_smt_present);
6590 if (!sched_smp_initialized)
6593 ret = cpuset_cpu_inactive(cpu);
6595 set_cpu_active(cpu, true);
6598 sched_domains_numa_masks_clear(cpu);
6602 static void sched_rq_cpu_starting(unsigned int cpu)
6604 struct rq *rq = cpu_rq(cpu);
6606 rq->calc_load_update = calc_load_update;
6607 update_max_interval();
6610 int sched_cpu_starting(unsigned int cpu)
6612 sched_rq_cpu_starting(cpu);
6613 sched_tick_start(cpu);
6617 #ifdef CONFIG_HOTPLUG_CPU
6618 int sched_cpu_dying(unsigned int cpu)
6620 struct rq *rq = cpu_rq(cpu);
6623 /* Handle pending wakeups and then migrate everything off */
6624 sched_tick_stop(cpu);
6626 rq_lock_irqsave(rq, &rf);
6628 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6631 migrate_tasks(rq, &rf);
6632 BUG_ON(rq->nr_running != 1);
6633 rq_unlock_irqrestore(rq, &rf);
6635 calc_load_migrate(rq);
6636 update_max_interval();
6637 nohz_balance_exit_idle(rq);
6643 void __init sched_init_smp(void)
6648 * There's no userspace yet to cause hotplug operations; hence all the
6649 * CPU masks are stable and all blatant races in the below code cannot
6652 mutex_lock(&sched_domains_mutex);
6653 sched_init_domains(cpu_active_mask);
6654 mutex_unlock(&sched_domains_mutex);
6656 /* Move init over to a non-isolated CPU */
6657 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6659 sched_init_granularity();
6661 init_sched_rt_class();
6662 init_sched_dl_class();
6664 sched_smp_initialized = true;
6667 static int __init migration_init(void)
6669 sched_cpu_starting(smp_processor_id());
6672 early_initcall(migration_init);
6675 void __init sched_init_smp(void)
6677 sched_init_granularity();
6679 #endif /* CONFIG_SMP */
6681 int in_sched_functions(unsigned long addr)
6683 return in_lock_functions(addr) ||
6684 (addr >= (unsigned long)__sched_text_start
6685 && addr < (unsigned long)__sched_text_end);
6688 #ifdef CONFIG_CGROUP_SCHED
6690 * Default task group.
6691 * Every task in system belongs to this group at bootup.
6693 struct task_group root_task_group;
6694 LIST_HEAD(task_groups);
6696 /* Cacheline aligned slab cache for task_group */
6697 static struct kmem_cache *task_group_cache __read_mostly;
6700 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6701 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6703 void __init sched_init(void)
6705 unsigned long ptr = 0;
6710 #ifdef CONFIG_FAIR_GROUP_SCHED
6711 ptr += 2 * nr_cpu_ids * sizeof(void **);
6713 #ifdef CONFIG_RT_GROUP_SCHED
6714 ptr += 2 * nr_cpu_ids * sizeof(void **);
6717 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6719 #ifdef CONFIG_FAIR_GROUP_SCHED
6720 root_task_group.se = (struct sched_entity **)ptr;
6721 ptr += nr_cpu_ids * sizeof(void **);
6723 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6724 ptr += nr_cpu_ids * sizeof(void **);
6726 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6727 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6728 #endif /* CONFIG_FAIR_GROUP_SCHED */
6729 #ifdef CONFIG_RT_GROUP_SCHED
6730 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6731 ptr += nr_cpu_ids * sizeof(void **);
6733 root_task_group.rt_rq = (struct rt_rq **)ptr;
6734 ptr += nr_cpu_ids * sizeof(void **);
6736 #endif /* CONFIG_RT_GROUP_SCHED */
6738 #ifdef CONFIG_CPUMASK_OFFSTACK
6739 for_each_possible_cpu(i) {
6740 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6741 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6742 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6743 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6745 #endif /* CONFIG_CPUMASK_OFFSTACK */
6747 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6748 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6751 init_defrootdomain();
6754 #ifdef CONFIG_RT_GROUP_SCHED
6755 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6756 global_rt_period(), global_rt_runtime());
6757 #endif /* CONFIG_RT_GROUP_SCHED */
6759 #ifdef CONFIG_CGROUP_SCHED
6760 task_group_cache = KMEM_CACHE(task_group, 0);
6762 list_add(&root_task_group.list, &task_groups);
6763 INIT_LIST_HEAD(&root_task_group.children);
6764 INIT_LIST_HEAD(&root_task_group.siblings);
6765 autogroup_init(&init_task);
6766 #endif /* CONFIG_CGROUP_SCHED */
6768 for_each_possible_cpu(i) {
6772 raw_spin_lock_init(&rq->lock);
6774 rq->calc_load_active = 0;
6775 rq->calc_load_update = jiffies + LOAD_FREQ;
6776 init_cfs_rq(&rq->cfs);
6777 init_rt_rq(&rq->rt);
6778 init_dl_rq(&rq->dl);
6779 #ifdef CONFIG_FAIR_GROUP_SCHED
6780 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6781 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6783 * How much CPU bandwidth does root_task_group get?
6785 * In case of task-groups formed thr' the cgroup filesystem, it
6786 * gets 100% of the CPU resources in the system. This overall
6787 * system CPU resource is divided among the tasks of
6788 * root_task_group and its child task-groups in a fair manner,
6789 * based on each entity's (task or task-group's) weight
6790 * (se->load.weight).
6792 * In other words, if root_task_group has 10 tasks of weight
6793 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6794 * then A0's share of the CPU resource is:
6796 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6798 * We achieve this by letting root_task_group's tasks sit
6799 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6801 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6802 #endif /* CONFIG_FAIR_GROUP_SCHED */
6804 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6805 #ifdef CONFIG_RT_GROUP_SCHED
6806 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6811 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6812 rq->balance_callback = NULL;
6813 rq->active_balance = 0;
6814 rq->next_balance = jiffies;
6819 rq->avg_idle = 2*sysctl_sched_migration_cost;
6820 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6822 INIT_LIST_HEAD(&rq->cfs_tasks);
6824 rq_attach_root(rq, &def_root_domain);
6825 #ifdef CONFIG_NO_HZ_COMMON
6826 rq->last_blocked_load_update_tick = jiffies;
6827 atomic_set(&rq->nohz_flags, 0);
6829 rq_csd_init(rq, &rq->nohz_csd, nohz_csd_func);
6831 #endif /* CONFIG_SMP */
6833 atomic_set(&rq->nr_iowait, 0);
6836 set_load_weight(&init_task, false);
6839 * The boot idle thread does lazy MMU switching as well:
6842 enter_lazy_tlb(&init_mm, current);
6845 * Make us the idle thread. Technically, schedule() should not be
6846 * called from this thread, however somewhere below it might be,
6847 * but because we are the idle thread, we just pick up running again
6848 * when this runqueue becomes "idle".
6850 init_idle(current, smp_processor_id());
6852 calc_load_update = jiffies + LOAD_FREQ;
6855 idle_thread_set_boot_cpu();
6857 init_sched_fair_class();
6865 scheduler_running = 1;
6868 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6869 static inline int preempt_count_equals(int preempt_offset)
6871 int nested = preempt_count() + rcu_preempt_depth();
6873 return (nested == preempt_offset);
6876 void __might_sleep(const char *file, int line, int preempt_offset)
6879 * Blocking primitives will set (and therefore destroy) current->state,
6880 * since we will exit with TASK_RUNNING make sure we enter with it,
6881 * otherwise we will destroy state.
6883 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6884 "do not call blocking ops when !TASK_RUNNING; "
6885 "state=%lx set at [<%p>] %pS\n",
6887 (void *)current->task_state_change,
6888 (void *)current->task_state_change);
6890 ___might_sleep(file, line, preempt_offset);
6892 EXPORT_SYMBOL(__might_sleep);
6894 void ___might_sleep(const char *file, int line, int preempt_offset)
6896 /* Ratelimiting timestamp: */
6897 static unsigned long prev_jiffy;
6899 unsigned long preempt_disable_ip;
6901 /* WARN_ON_ONCE() by default, no rate limit required: */
6904 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6905 !is_idle_task(current) && !current->non_block_count) ||
6906 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6910 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6912 prev_jiffy = jiffies;
6914 /* Save this before calling printk(), since that will clobber it: */
6915 preempt_disable_ip = get_preempt_disable_ip(current);
6918 "BUG: sleeping function called from invalid context at %s:%d\n",
6921 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6922 in_atomic(), irqs_disabled(), current->non_block_count,
6923 current->pid, current->comm);
6925 if (task_stack_end_corrupted(current))
6926 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6928 debug_show_held_locks(current);
6929 if (irqs_disabled())
6930 print_irqtrace_events(current);
6931 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6932 && !preempt_count_equals(preempt_offset)) {
6933 pr_err("Preemption disabled at:");
6934 print_ip_sym(KERN_ERR, preempt_disable_ip);
6937 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6939 EXPORT_SYMBOL(___might_sleep);
6941 void __cant_sleep(const char *file, int line, int preempt_offset)
6943 static unsigned long prev_jiffy;
6945 if (irqs_disabled())
6948 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6951 if (preempt_count() > preempt_offset)
6954 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6956 prev_jiffy = jiffies;
6958 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6959 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6960 in_atomic(), irqs_disabled(),
6961 current->pid, current->comm);
6963 debug_show_held_locks(current);
6965 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6967 EXPORT_SYMBOL_GPL(__cant_sleep);
6970 #ifdef CONFIG_MAGIC_SYSRQ
6971 void normalize_rt_tasks(void)
6973 struct task_struct *g, *p;
6974 struct sched_attr attr = {
6975 .sched_policy = SCHED_NORMAL,
6978 read_lock(&tasklist_lock);
6979 for_each_process_thread(g, p) {
6981 * Only normalize user tasks:
6983 if (p->flags & PF_KTHREAD)
6986 p->se.exec_start = 0;
6987 schedstat_set(p->se.statistics.wait_start, 0);
6988 schedstat_set(p->se.statistics.sleep_start, 0);
6989 schedstat_set(p->se.statistics.block_start, 0);
6991 if (!dl_task(p) && !rt_task(p)) {
6993 * Renice negative nice level userspace
6996 if (task_nice(p) < 0)
6997 set_user_nice(p, 0);
7001 __sched_setscheduler(p, &attr, false, false);
7003 read_unlock(&tasklist_lock);
7006 #endif /* CONFIG_MAGIC_SYSRQ */
7008 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7010 * These functions are only useful for the IA64 MCA handling, or kdb.
7012 * They can only be called when the whole system has been
7013 * stopped - every CPU needs to be quiescent, and no scheduling
7014 * activity can take place. Using them for anything else would
7015 * be a serious bug, and as a result, they aren't even visible
7016 * under any other configuration.
7020 * curr_task - return the current task for a given CPU.
7021 * @cpu: the processor in question.
7023 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7025 * Return: The current task for @cpu.
7027 struct task_struct *curr_task(int cpu)
7029 return cpu_curr(cpu);
7032 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7036 * ia64_set_curr_task - set the current task for a given CPU.
7037 * @cpu: the processor in question.
7038 * @p: the task pointer to set.
7040 * Description: This function must only be used when non-maskable interrupts
7041 * are serviced on a separate stack. It allows the architecture to switch the
7042 * notion of the current task on a CPU in a non-blocking manner. This function
7043 * must be called with all CPU's synchronized, and interrupts disabled, the
7044 * and caller must save the original value of the current task (see
7045 * curr_task() above) and restore that value before reenabling interrupts and
7046 * re-starting the system.
7048 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7050 void ia64_set_curr_task(int cpu, struct task_struct *p)
7057 #ifdef CONFIG_CGROUP_SCHED
7058 /* task_group_lock serializes the addition/removal of task groups */
7059 static DEFINE_SPINLOCK(task_group_lock);
7061 static inline void alloc_uclamp_sched_group(struct task_group *tg,
7062 struct task_group *parent)
7064 #ifdef CONFIG_UCLAMP_TASK_GROUP
7065 enum uclamp_id clamp_id;
7067 for_each_clamp_id(clamp_id) {
7068 uclamp_se_set(&tg->uclamp_req[clamp_id],
7069 uclamp_none(clamp_id), false);
7070 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
7075 static void sched_free_group(struct task_group *tg)
7077 free_fair_sched_group(tg);
7078 free_rt_sched_group(tg);
7080 kmem_cache_free(task_group_cache, tg);
7083 /* allocate runqueue etc for a new task group */
7084 struct task_group *sched_create_group(struct task_group *parent)
7086 struct task_group *tg;
7088 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7090 return ERR_PTR(-ENOMEM);
7092 if (!alloc_fair_sched_group(tg, parent))
7095 if (!alloc_rt_sched_group(tg, parent))
7098 alloc_uclamp_sched_group(tg, parent);
7103 sched_free_group(tg);
7104 return ERR_PTR(-ENOMEM);
7107 void sched_online_group(struct task_group *tg, struct task_group *parent)
7109 unsigned long flags;
7111 spin_lock_irqsave(&task_group_lock, flags);
7112 list_add_rcu(&tg->list, &task_groups);
7114 /* Root should already exist: */
7117 tg->parent = parent;
7118 INIT_LIST_HEAD(&tg->children);
7119 list_add_rcu(&tg->siblings, &parent->children);
7120 spin_unlock_irqrestore(&task_group_lock, flags);
7122 online_fair_sched_group(tg);
7125 /* rcu callback to free various structures associated with a task group */
7126 static void sched_free_group_rcu(struct rcu_head *rhp)
7128 /* Now it should be safe to free those cfs_rqs: */
7129 sched_free_group(container_of(rhp, struct task_group, rcu));
7132 void sched_destroy_group(struct task_group *tg)
7134 /* Wait for possible concurrent references to cfs_rqs complete: */
7135 call_rcu(&tg->rcu, sched_free_group_rcu);
7138 void sched_offline_group(struct task_group *tg)
7140 unsigned long flags;
7142 /* End participation in shares distribution: */
7143 unregister_fair_sched_group(tg);
7145 spin_lock_irqsave(&task_group_lock, flags);
7146 list_del_rcu(&tg->list);
7147 list_del_rcu(&tg->siblings);
7148 spin_unlock_irqrestore(&task_group_lock, flags);
7151 static void sched_change_group(struct task_struct *tsk, int type)
7153 struct task_group *tg;
7156 * All callers are synchronized by task_rq_lock(); we do not use RCU
7157 * which is pointless here. Thus, we pass "true" to task_css_check()
7158 * to prevent lockdep warnings.
7160 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7161 struct task_group, css);
7162 tg = autogroup_task_group(tsk, tg);
7163 tsk->sched_task_group = tg;
7165 #ifdef CONFIG_FAIR_GROUP_SCHED
7166 if (tsk->sched_class->task_change_group)
7167 tsk->sched_class->task_change_group(tsk, type);
7170 set_task_rq(tsk, task_cpu(tsk));
7174 * Change task's runqueue when it moves between groups.
7176 * The caller of this function should have put the task in its new group by
7177 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7180 void sched_move_task(struct task_struct *tsk)
7182 int queued, running, queue_flags =
7183 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7187 rq = task_rq_lock(tsk, &rf);
7188 update_rq_clock(rq);
7190 running = task_current(rq, tsk);
7191 queued = task_on_rq_queued(tsk);
7194 dequeue_task(rq, tsk, queue_flags);
7196 put_prev_task(rq, tsk);
7198 sched_change_group(tsk, TASK_MOVE_GROUP);
7201 enqueue_task(rq, tsk, queue_flags);
7203 set_next_task(rq, tsk);
7205 * After changing group, the running task may have joined a
7206 * throttled one but it's still the running task. Trigger a
7207 * resched to make sure that task can still run.
7212 task_rq_unlock(rq, tsk, &rf);
7215 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7217 return css ? container_of(css, struct task_group, css) : NULL;
7220 static struct cgroup_subsys_state *
7221 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7223 struct task_group *parent = css_tg(parent_css);
7224 struct task_group *tg;
7227 /* This is early initialization for the top cgroup */
7228 return &root_task_group.css;
7231 tg = sched_create_group(parent);
7233 return ERR_PTR(-ENOMEM);
7238 /* Expose task group only after completing cgroup initialization */
7239 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7241 struct task_group *tg = css_tg(css);
7242 struct task_group *parent = css_tg(css->parent);
7245 sched_online_group(tg, parent);
7247 #ifdef CONFIG_UCLAMP_TASK_GROUP
7248 /* Propagate the effective uclamp value for the new group */
7249 cpu_util_update_eff(css);
7255 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7257 struct task_group *tg = css_tg(css);
7259 sched_offline_group(tg);
7262 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7264 struct task_group *tg = css_tg(css);
7267 * Relies on the RCU grace period between css_released() and this.
7269 sched_free_group(tg);
7273 * This is called before wake_up_new_task(), therefore we really only
7274 * have to set its group bits, all the other stuff does not apply.
7276 static void cpu_cgroup_fork(struct task_struct *task)
7281 rq = task_rq_lock(task, &rf);
7283 update_rq_clock(rq);
7284 sched_change_group(task, TASK_SET_GROUP);
7286 task_rq_unlock(rq, task, &rf);
7289 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7291 struct task_struct *task;
7292 struct cgroup_subsys_state *css;
7295 cgroup_taskset_for_each(task, css, tset) {
7296 #ifdef CONFIG_RT_GROUP_SCHED
7297 if (!sched_rt_can_attach(css_tg(css), task))
7301 * Serialize against wake_up_new_task() such that if its
7302 * running, we're sure to observe its full state.
7304 raw_spin_lock_irq(&task->pi_lock);
7306 * Avoid calling sched_move_task() before wake_up_new_task()
7307 * has happened. This would lead to problems with PELT, due to
7308 * move wanting to detach+attach while we're not attached yet.
7310 if (task->state == TASK_NEW)
7312 raw_spin_unlock_irq(&task->pi_lock);
7320 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7322 struct task_struct *task;
7323 struct cgroup_subsys_state *css;
7325 cgroup_taskset_for_each(task, css, tset)
7326 sched_move_task(task);
7329 #ifdef CONFIG_UCLAMP_TASK_GROUP
7330 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7332 struct cgroup_subsys_state *top_css = css;
7333 struct uclamp_se *uc_parent = NULL;
7334 struct uclamp_se *uc_se = NULL;
7335 unsigned int eff[UCLAMP_CNT];
7336 enum uclamp_id clamp_id;
7337 unsigned int clamps;
7339 css_for_each_descendant_pre(css, top_css) {
7340 uc_parent = css_tg(css)->parent
7341 ? css_tg(css)->parent->uclamp : NULL;
7343 for_each_clamp_id(clamp_id) {
7344 /* Assume effective clamps matches requested clamps */
7345 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7346 /* Cap effective clamps with parent's effective clamps */
7348 eff[clamp_id] > uc_parent[clamp_id].value) {
7349 eff[clamp_id] = uc_parent[clamp_id].value;
7352 /* Ensure protection is always capped by limit */
7353 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7355 /* Propagate most restrictive effective clamps */
7357 uc_se = css_tg(css)->uclamp;
7358 for_each_clamp_id(clamp_id) {
7359 if (eff[clamp_id] == uc_se[clamp_id].value)
7361 uc_se[clamp_id].value = eff[clamp_id];
7362 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7363 clamps |= (0x1 << clamp_id);
7366 css = css_rightmost_descendant(css);
7370 /* Immediately update descendants RUNNABLE tasks */
7371 uclamp_update_active_tasks(css, clamps);
7376 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7377 * C expression. Since there is no way to convert a macro argument (N) into a
7378 * character constant, use two levels of macros.
7380 #define _POW10(exp) ((unsigned int)1e##exp)
7381 #define POW10(exp) _POW10(exp)
7383 struct uclamp_request {
7384 #define UCLAMP_PERCENT_SHIFT 2
7385 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7391 static inline struct uclamp_request
7392 capacity_from_percent(char *buf)
7394 struct uclamp_request req = {
7395 .percent = UCLAMP_PERCENT_SCALE,
7396 .util = SCHED_CAPACITY_SCALE,
7401 if (strcmp(buf, "max")) {
7402 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7406 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7411 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7412 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7418 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7419 size_t nbytes, loff_t off,
7420 enum uclamp_id clamp_id)
7422 struct uclamp_request req;
7423 struct task_group *tg;
7425 req = capacity_from_percent(buf);
7429 mutex_lock(&uclamp_mutex);
7432 tg = css_tg(of_css(of));
7433 if (tg->uclamp_req[clamp_id].value != req.util)
7434 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7437 * Because of not recoverable conversion rounding we keep track of the
7438 * exact requested value
7440 tg->uclamp_pct[clamp_id] = req.percent;
7442 /* Update effective clamps to track the most restrictive value */
7443 cpu_util_update_eff(of_css(of));
7446 mutex_unlock(&uclamp_mutex);
7451 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7452 char *buf, size_t nbytes,
7455 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7458 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7459 char *buf, size_t nbytes,
7462 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7465 static inline void cpu_uclamp_print(struct seq_file *sf,
7466 enum uclamp_id clamp_id)
7468 struct task_group *tg;
7474 tg = css_tg(seq_css(sf));
7475 util_clamp = tg->uclamp_req[clamp_id].value;
7478 if (util_clamp == SCHED_CAPACITY_SCALE) {
7479 seq_puts(sf, "max\n");
7483 percent = tg->uclamp_pct[clamp_id];
7484 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7485 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7488 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7490 cpu_uclamp_print(sf, UCLAMP_MIN);
7494 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7496 cpu_uclamp_print(sf, UCLAMP_MAX);
7499 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7501 #ifdef CONFIG_FAIR_GROUP_SCHED
7502 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7503 struct cftype *cftype, u64 shareval)
7505 if (shareval > scale_load_down(ULONG_MAX))
7506 shareval = MAX_SHARES;
7507 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7510 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7513 struct task_group *tg = css_tg(css);
7515 return (u64) scale_load_down(tg->shares);
7518 #ifdef CONFIG_CFS_BANDWIDTH
7519 static DEFINE_MUTEX(cfs_constraints_mutex);
7521 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7522 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7523 /* More than 203 days if BW_SHIFT equals 20. */
7524 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
7526 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7528 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7530 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7531 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7533 if (tg == &root_task_group)
7537 * Ensure we have at some amount of bandwidth every period. This is
7538 * to prevent reaching a state of large arrears when throttled via
7539 * entity_tick() resulting in prolonged exit starvation.
7541 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7545 * Likewise, bound things on the otherside by preventing insane quota
7546 * periods. This also allows us to normalize in computing quota
7549 if (period > max_cfs_quota_period)
7553 * Bound quota to defend quota against overflow during bandwidth shift.
7555 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
7559 * Prevent race between setting of cfs_rq->runtime_enabled and
7560 * unthrottle_offline_cfs_rqs().
7563 mutex_lock(&cfs_constraints_mutex);
7564 ret = __cfs_schedulable(tg, period, quota);
7568 runtime_enabled = quota != RUNTIME_INF;
7569 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7571 * If we need to toggle cfs_bandwidth_used, off->on must occur
7572 * before making related changes, and on->off must occur afterwards
7574 if (runtime_enabled && !runtime_was_enabled)
7575 cfs_bandwidth_usage_inc();
7576 raw_spin_lock_irq(&cfs_b->lock);
7577 cfs_b->period = ns_to_ktime(period);
7578 cfs_b->quota = quota;
7580 __refill_cfs_bandwidth_runtime(cfs_b);
7582 /* Restart the period timer (if active) to handle new period expiry: */
7583 if (runtime_enabled)
7584 start_cfs_bandwidth(cfs_b);
7586 raw_spin_unlock_irq(&cfs_b->lock);
7588 for_each_online_cpu(i) {
7589 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7590 struct rq *rq = cfs_rq->rq;
7593 rq_lock_irq(rq, &rf);
7594 cfs_rq->runtime_enabled = runtime_enabled;
7595 cfs_rq->runtime_remaining = 0;
7597 if (cfs_rq->throttled)
7598 unthrottle_cfs_rq(cfs_rq);
7599 rq_unlock_irq(rq, &rf);
7601 if (runtime_was_enabled && !runtime_enabled)
7602 cfs_bandwidth_usage_dec();
7604 mutex_unlock(&cfs_constraints_mutex);
7610 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7614 period = ktime_to_ns(tg->cfs_bandwidth.period);
7615 if (cfs_quota_us < 0)
7616 quota = RUNTIME_INF;
7617 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7618 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7622 return tg_set_cfs_bandwidth(tg, period, quota);
7625 static long tg_get_cfs_quota(struct task_group *tg)
7629 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7632 quota_us = tg->cfs_bandwidth.quota;
7633 do_div(quota_us, NSEC_PER_USEC);
7638 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7642 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7645 period = (u64)cfs_period_us * NSEC_PER_USEC;
7646 quota = tg->cfs_bandwidth.quota;
7648 return tg_set_cfs_bandwidth(tg, period, quota);
7651 static long tg_get_cfs_period(struct task_group *tg)
7655 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7656 do_div(cfs_period_us, NSEC_PER_USEC);
7658 return cfs_period_us;
7661 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7664 return tg_get_cfs_quota(css_tg(css));
7667 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7668 struct cftype *cftype, s64 cfs_quota_us)
7670 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7673 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7676 return tg_get_cfs_period(css_tg(css));
7679 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7680 struct cftype *cftype, u64 cfs_period_us)
7682 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7685 struct cfs_schedulable_data {
7686 struct task_group *tg;
7691 * normalize group quota/period to be quota/max_period
7692 * note: units are usecs
7694 static u64 normalize_cfs_quota(struct task_group *tg,
7695 struct cfs_schedulable_data *d)
7703 period = tg_get_cfs_period(tg);
7704 quota = tg_get_cfs_quota(tg);
7707 /* note: these should typically be equivalent */
7708 if (quota == RUNTIME_INF || quota == -1)
7711 return to_ratio(period, quota);
7714 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7716 struct cfs_schedulable_data *d = data;
7717 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7718 s64 quota = 0, parent_quota = -1;
7721 quota = RUNTIME_INF;
7723 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7725 quota = normalize_cfs_quota(tg, d);
7726 parent_quota = parent_b->hierarchical_quota;
7729 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7730 * always take the min. On cgroup1, only inherit when no
7733 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7734 quota = min(quota, parent_quota);
7736 if (quota == RUNTIME_INF)
7737 quota = parent_quota;
7738 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7742 cfs_b->hierarchical_quota = quota;
7747 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7750 struct cfs_schedulable_data data = {
7756 if (quota != RUNTIME_INF) {
7757 do_div(data.period, NSEC_PER_USEC);
7758 do_div(data.quota, NSEC_PER_USEC);
7762 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7768 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7770 struct task_group *tg = css_tg(seq_css(sf));
7771 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7773 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7774 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7775 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7777 if (schedstat_enabled() && tg != &root_task_group) {
7781 for_each_possible_cpu(i)
7782 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7784 seq_printf(sf, "wait_sum %llu\n", ws);
7789 #endif /* CONFIG_CFS_BANDWIDTH */
7790 #endif /* CONFIG_FAIR_GROUP_SCHED */
7792 #ifdef CONFIG_RT_GROUP_SCHED
7793 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7794 struct cftype *cft, s64 val)
7796 return sched_group_set_rt_runtime(css_tg(css), val);
7799 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7802 return sched_group_rt_runtime(css_tg(css));
7805 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7806 struct cftype *cftype, u64 rt_period_us)
7808 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7811 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7814 return sched_group_rt_period(css_tg(css));
7816 #endif /* CONFIG_RT_GROUP_SCHED */
7818 static struct cftype cpu_legacy_files[] = {
7819 #ifdef CONFIG_FAIR_GROUP_SCHED
7822 .read_u64 = cpu_shares_read_u64,
7823 .write_u64 = cpu_shares_write_u64,
7826 #ifdef CONFIG_CFS_BANDWIDTH
7828 .name = "cfs_quota_us",
7829 .read_s64 = cpu_cfs_quota_read_s64,
7830 .write_s64 = cpu_cfs_quota_write_s64,
7833 .name = "cfs_period_us",
7834 .read_u64 = cpu_cfs_period_read_u64,
7835 .write_u64 = cpu_cfs_period_write_u64,
7839 .seq_show = cpu_cfs_stat_show,
7842 #ifdef CONFIG_RT_GROUP_SCHED
7844 .name = "rt_runtime_us",
7845 .read_s64 = cpu_rt_runtime_read,
7846 .write_s64 = cpu_rt_runtime_write,
7849 .name = "rt_period_us",
7850 .read_u64 = cpu_rt_period_read_uint,
7851 .write_u64 = cpu_rt_period_write_uint,
7854 #ifdef CONFIG_UCLAMP_TASK_GROUP
7856 .name = "uclamp.min",
7857 .flags = CFTYPE_NOT_ON_ROOT,
7858 .seq_show = cpu_uclamp_min_show,
7859 .write = cpu_uclamp_min_write,
7862 .name = "uclamp.max",
7863 .flags = CFTYPE_NOT_ON_ROOT,
7864 .seq_show = cpu_uclamp_max_show,
7865 .write = cpu_uclamp_max_write,
7871 static int cpu_extra_stat_show(struct seq_file *sf,
7872 struct cgroup_subsys_state *css)
7874 #ifdef CONFIG_CFS_BANDWIDTH
7876 struct task_group *tg = css_tg(css);
7877 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7880 throttled_usec = cfs_b->throttled_time;
7881 do_div(throttled_usec, NSEC_PER_USEC);
7883 seq_printf(sf, "nr_periods %d\n"
7885 "throttled_usec %llu\n",
7886 cfs_b->nr_periods, cfs_b->nr_throttled,
7893 #ifdef CONFIG_FAIR_GROUP_SCHED
7894 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7897 struct task_group *tg = css_tg(css);
7898 u64 weight = scale_load_down(tg->shares);
7900 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7903 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7904 struct cftype *cft, u64 weight)
7907 * cgroup weight knobs should use the common MIN, DFL and MAX
7908 * values which are 1, 100 and 10000 respectively. While it loses
7909 * a bit of range on both ends, it maps pretty well onto the shares
7910 * value used by scheduler and the round-trip conversions preserve
7911 * the original value over the entire range.
7913 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7916 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7918 return sched_group_set_shares(css_tg(css), scale_load(weight));
7921 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7924 unsigned long weight = scale_load_down(css_tg(css)->shares);
7925 int last_delta = INT_MAX;
7928 /* find the closest nice value to the current weight */
7929 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7930 delta = abs(sched_prio_to_weight[prio] - weight);
7931 if (delta >= last_delta)
7936 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7939 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7940 struct cftype *cft, s64 nice)
7942 unsigned long weight;
7945 if (nice < MIN_NICE || nice > MAX_NICE)
7948 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7949 idx = array_index_nospec(idx, 40);
7950 weight = sched_prio_to_weight[idx];
7952 return sched_group_set_shares(css_tg(css), scale_load(weight));
7956 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7957 long period, long quota)
7960 seq_puts(sf, "max");
7962 seq_printf(sf, "%ld", quota);
7964 seq_printf(sf, " %ld\n", period);
7967 /* caller should put the current value in *@periodp before calling */
7968 static int __maybe_unused cpu_period_quota_parse(char *buf,
7969 u64 *periodp, u64 *quotap)
7971 char tok[21]; /* U64_MAX */
7973 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7976 *periodp *= NSEC_PER_USEC;
7978 if (sscanf(tok, "%llu", quotap))
7979 *quotap *= NSEC_PER_USEC;
7980 else if (!strcmp(tok, "max"))
7981 *quotap = RUNTIME_INF;
7988 #ifdef CONFIG_CFS_BANDWIDTH
7989 static int cpu_max_show(struct seq_file *sf, void *v)
7991 struct task_group *tg = css_tg(seq_css(sf));
7993 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7997 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7998 char *buf, size_t nbytes, loff_t off)
8000 struct task_group *tg = css_tg(of_css(of));
8001 u64 period = tg_get_cfs_period(tg);
8005 ret = cpu_period_quota_parse(buf, &period, "a);
8007 ret = tg_set_cfs_bandwidth(tg, period, quota);
8008 return ret ?: nbytes;
8012 static struct cftype cpu_files[] = {
8013 #ifdef CONFIG_FAIR_GROUP_SCHED
8016 .flags = CFTYPE_NOT_ON_ROOT,
8017 .read_u64 = cpu_weight_read_u64,
8018 .write_u64 = cpu_weight_write_u64,
8021 .name = "weight.nice",
8022 .flags = CFTYPE_NOT_ON_ROOT,
8023 .read_s64 = cpu_weight_nice_read_s64,
8024 .write_s64 = cpu_weight_nice_write_s64,
8027 #ifdef CONFIG_CFS_BANDWIDTH
8030 .flags = CFTYPE_NOT_ON_ROOT,
8031 .seq_show = cpu_max_show,
8032 .write = cpu_max_write,
8035 #ifdef CONFIG_UCLAMP_TASK_GROUP
8037 .name = "uclamp.min",
8038 .flags = CFTYPE_NOT_ON_ROOT,
8039 .seq_show = cpu_uclamp_min_show,
8040 .write = cpu_uclamp_min_write,
8043 .name = "uclamp.max",
8044 .flags = CFTYPE_NOT_ON_ROOT,
8045 .seq_show = cpu_uclamp_max_show,
8046 .write = cpu_uclamp_max_write,
8052 struct cgroup_subsys cpu_cgrp_subsys = {
8053 .css_alloc = cpu_cgroup_css_alloc,
8054 .css_online = cpu_cgroup_css_online,
8055 .css_released = cpu_cgroup_css_released,
8056 .css_free = cpu_cgroup_css_free,
8057 .css_extra_stat_show = cpu_extra_stat_show,
8058 .fork = cpu_cgroup_fork,
8059 .can_attach = cpu_cgroup_can_attach,
8060 .attach = cpu_cgroup_attach,
8061 .legacy_cftypes = cpu_legacy_files,
8062 .dfl_cftypes = cpu_files,
8067 #endif /* CONFIG_CGROUP_SCHED */
8069 void dump_cpu_task(int cpu)
8071 pr_info("Task dump for CPU %d:\n", cpu);
8072 sched_show_task(cpu_curr(cpu));
8076 * Nice levels are multiplicative, with a gentle 10% change for every
8077 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8078 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8079 * that remained on nice 0.
8081 * The "10% effect" is relative and cumulative: from _any_ nice level,
8082 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8083 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8084 * If a task goes up by ~10% and another task goes down by ~10% then
8085 * the relative distance between them is ~25%.)
8087 const int sched_prio_to_weight[40] = {
8088 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8089 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8090 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8091 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8092 /* 0 */ 1024, 820, 655, 526, 423,
8093 /* 5 */ 335, 272, 215, 172, 137,
8094 /* 10 */ 110, 87, 70, 56, 45,
8095 /* 15 */ 36, 29, 23, 18, 15,
8099 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8101 * In cases where the weight does not change often, we can use the
8102 * precalculated inverse to speed up arithmetics by turning divisions
8103 * into multiplications:
8105 const u32 sched_prio_to_wmult[40] = {
8106 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8107 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8108 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8109 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8110 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8111 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8112 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8113 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8116 #undef CREATE_TRACE_POINTS