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>
15 #include <asm/switch_to.h>
18 #include "../workqueue_internal.h"
19 #include "../smpboot.h"
23 #define CREATE_TRACE_POINTS
24 #include <trace/events/sched.h>
27 * Export tracepoints that act as a bare tracehook (ie: have no trace event
28 * associated with them) to allow external modules to probe them.
30 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
31 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
32 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
37 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
39 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
41 * Debugging: various feature bits
43 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
44 * sysctl_sched_features, defined in sched.h, to allow constants propagation
45 * at compile time and compiler optimization based on features default.
47 #define SCHED_FEAT(name, enabled) \
48 (1UL << __SCHED_FEAT_##name) * enabled |
49 const_debug unsigned int sysctl_sched_features =
56 * Number of tasks to iterate in a single balance run.
57 * Limited because this is done with IRQs disabled.
59 const_debug unsigned int sysctl_sched_nr_migrate = 32;
62 * period over which we measure -rt task CPU usage in us.
65 unsigned int sysctl_sched_rt_period = 1000000;
67 __read_mostly int scheduler_running;
70 * part of the period that we allow rt tasks to run in us.
73 int sysctl_sched_rt_runtime = 950000;
76 * __task_rq_lock - lock the rq @p resides on.
78 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
83 lockdep_assert_held(&p->pi_lock);
87 raw_spin_lock(&rq->lock);
88 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
92 raw_spin_unlock(&rq->lock);
94 while (unlikely(task_on_rq_migrating(p)))
100 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
102 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
103 __acquires(p->pi_lock)
109 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
111 raw_spin_lock(&rq->lock);
113 * move_queued_task() task_rq_lock()
116 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
117 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
118 * [S] ->cpu = new_cpu [L] task_rq()
122 * If we observe the old CPU in task_rq_lock(), the acquire of
123 * the old rq->lock will fully serialize against the stores.
125 * If we observe the new CPU in task_rq_lock(), the address
126 * dependency headed by '[L] rq = task_rq()' and the acquire
127 * will pair with the WMB to ensure we then also see migrating.
129 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
133 raw_spin_unlock(&rq->lock);
134 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
136 while (unlikely(task_on_rq_migrating(p)))
142 * RQ-clock updating methods:
145 static void update_rq_clock_task(struct rq *rq, s64 delta)
148 * In theory, the compile should just see 0 here, and optimize out the call
149 * to sched_rt_avg_update. But I don't trust it...
151 s64 __maybe_unused steal = 0, irq_delta = 0;
153 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
154 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
157 * Since irq_time is only updated on {soft,}irq_exit, we might run into
158 * this case when a previous update_rq_clock() happened inside a
161 * When this happens, we stop ->clock_task and only update the
162 * prev_irq_time stamp to account for the part that fit, so that a next
163 * update will consume the rest. This ensures ->clock_task is
166 * It does however cause some slight miss-attribution of {soft,}irq
167 * time, a more accurate solution would be to update the irq_time using
168 * the current rq->clock timestamp, except that would require using
171 if (irq_delta > delta)
174 rq->prev_irq_time += irq_delta;
177 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
178 if (static_key_false((¶virt_steal_rq_enabled))) {
179 steal = paravirt_steal_clock(cpu_of(rq));
180 steal -= rq->prev_steal_time_rq;
182 if (unlikely(steal > delta))
185 rq->prev_steal_time_rq += steal;
190 rq->clock_task += delta;
192 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
193 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
194 update_irq_load_avg(rq, irq_delta + steal);
196 update_rq_clock_pelt(rq, delta);
199 void update_rq_clock(struct rq *rq)
203 lockdep_assert_held(&rq->lock);
205 if (rq->clock_update_flags & RQCF_ACT_SKIP)
208 #ifdef CONFIG_SCHED_DEBUG
209 if (sched_feat(WARN_DOUBLE_CLOCK))
210 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
211 rq->clock_update_flags |= RQCF_UPDATED;
214 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
218 update_rq_clock_task(rq, delta);
222 #ifdef CONFIG_SCHED_HRTICK
224 * Use HR-timers to deliver accurate preemption points.
227 static void hrtick_clear(struct rq *rq)
229 if (hrtimer_active(&rq->hrtick_timer))
230 hrtimer_cancel(&rq->hrtick_timer);
234 * High-resolution timer tick.
235 * Runs from hardirq context with interrupts disabled.
237 static enum hrtimer_restart hrtick(struct hrtimer *timer)
239 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
242 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
246 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
249 return HRTIMER_NORESTART;
254 static void __hrtick_restart(struct rq *rq)
256 struct hrtimer *timer = &rq->hrtick_timer;
258 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
262 * called from hardirq (IPI) context
264 static void __hrtick_start(void *arg)
270 __hrtick_restart(rq);
271 rq->hrtick_csd_pending = 0;
276 * Called to set the hrtick timer state.
278 * called with rq->lock held and irqs disabled
280 void hrtick_start(struct rq *rq, u64 delay)
282 struct hrtimer *timer = &rq->hrtick_timer;
287 * Don't schedule slices shorter than 10000ns, that just
288 * doesn't make sense and can cause timer DoS.
290 delta = max_t(s64, delay, 10000LL);
291 time = ktime_add_ns(timer->base->get_time(), delta);
293 hrtimer_set_expires(timer, time);
295 if (rq == this_rq()) {
296 __hrtick_restart(rq);
297 } else if (!rq->hrtick_csd_pending) {
298 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
299 rq->hrtick_csd_pending = 1;
305 * Called to set the hrtick timer state.
307 * called with rq->lock held and irqs disabled
309 void hrtick_start(struct rq *rq, u64 delay)
312 * Don't schedule slices shorter than 10000ns, that just
313 * doesn't make sense. Rely on vruntime for fairness.
315 delay = max_t(u64, delay, 10000LL);
316 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
317 HRTIMER_MODE_REL_PINNED);
319 #endif /* CONFIG_SMP */
321 static void hrtick_rq_init(struct rq *rq)
324 rq->hrtick_csd_pending = 0;
326 rq->hrtick_csd.flags = 0;
327 rq->hrtick_csd.func = __hrtick_start;
328 rq->hrtick_csd.info = rq;
331 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
332 rq->hrtick_timer.function = hrtick;
334 #else /* CONFIG_SCHED_HRTICK */
335 static inline void hrtick_clear(struct rq *rq)
339 static inline void hrtick_rq_init(struct rq *rq)
342 #endif /* CONFIG_SCHED_HRTICK */
345 * cmpxchg based fetch_or, macro so it works for different integer types
347 #define fetch_or(ptr, mask) \
349 typeof(ptr) _ptr = (ptr); \
350 typeof(mask) _mask = (mask); \
351 typeof(*_ptr) _old, _val = *_ptr; \
354 _old = cmpxchg(_ptr, _val, _val | _mask); \
362 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
364 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
365 * this avoids any races wrt polling state changes and thereby avoids
368 static bool set_nr_and_not_polling(struct task_struct *p)
370 struct thread_info *ti = task_thread_info(p);
371 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
375 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
377 * If this returns true, then the idle task promises to call
378 * sched_ttwu_pending() and reschedule soon.
380 static bool set_nr_if_polling(struct task_struct *p)
382 struct thread_info *ti = task_thread_info(p);
383 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
386 if (!(val & _TIF_POLLING_NRFLAG))
388 if (val & _TIF_NEED_RESCHED)
390 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
399 static bool set_nr_and_not_polling(struct task_struct *p)
401 set_tsk_need_resched(p);
406 static bool set_nr_if_polling(struct task_struct *p)
413 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
415 struct wake_q_node *node = &task->wake_q;
418 * Atomically grab the task, if ->wake_q is !nil already it means
419 * its already queued (either by us or someone else) and will get the
420 * wakeup due to that.
422 * In order to ensure that a pending wakeup will observe our pending
423 * state, even in the failed case, an explicit smp_mb() must be used.
425 smp_mb__before_atomic();
426 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
430 * The head is context local, there can be no concurrency.
433 head->lastp = &node->next;
438 * wake_q_add() - queue a wakeup for 'later' waking.
439 * @head: the wake_q_head to add @task to
440 * @task: the task to queue for 'later' wakeup
442 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
443 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
446 * This function must be used as-if it were wake_up_process(); IOW the task
447 * must be ready to be woken at this location.
449 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
451 if (__wake_q_add(head, task))
452 get_task_struct(task);
456 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
457 * @head: the wake_q_head to add @task to
458 * @task: the task to queue for 'later' wakeup
460 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
461 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
464 * This function must be used as-if it were wake_up_process(); IOW the task
465 * must be ready to be woken at this location.
467 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
468 * that already hold reference to @task can call the 'safe' version and trust
469 * wake_q to do the right thing depending whether or not the @task is already
472 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
474 if (!__wake_q_add(head, task))
475 put_task_struct(task);
478 void wake_up_q(struct wake_q_head *head)
480 struct wake_q_node *node = head->first;
482 while (node != WAKE_Q_TAIL) {
483 struct task_struct *task;
485 task = container_of(node, struct task_struct, wake_q);
487 /* Task can safely be re-inserted now: */
489 task->wake_q.next = NULL;
492 * wake_up_process() executes a full barrier, which pairs with
493 * the queueing in wake_q_add() so as not to miss wakeups.
495 wake_up_process(task);
496 put_task_struct(task);
501 * resched_curr - mark rq's current task 'to be rescheduled now'.
503 * On UP this means the setting of the need_resched flag, on SMP it
504 * might also involve a cross-CPU call to trigger the scheduler on
507 void resched_curr(struct rq *rq)
509 struct task_struct *curr = rq->curr;
512 lockdep_assert_held(&rq->lock);
514 if (test_tsk_need_resched(curr))
519 if (cpu == smp_processor_id()) {
520 set_tsk_need_resched(curr);
521 set_preempt_need_resched();
525 if (set_nr_and_not_polling(curr))
526 smp_send_reschedule(cpu);
528 trace_sched_wake_idle_without_ipi(cpu);
531 void resched_cpu(int cpu)
533 struct rq *rq = cpu_rq(cpu);
536 raw_spin_lock_irqsave(&rq->lock, flags);
537 if (cpu_online(cpu) || cpu == smp_processor_id())
539 raw_spin_unlock_irqrestore(&rq->lock, flags);
543 #ifdef CONFIG_NO_HZ_COMMON
545 * In the semi idle case, use the nearest busy CPU for migrating timers
546 * from an idle CPU. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle CPU will add more delays to the timers than intended
550 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int i, cpu = smp_processor_id();
555 struct sched_domain *sd;
557 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
561 for_each_domain(cpu, sd) {
562 for_each_cpu(i, sched_domain_span(sd)) {
566 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
573 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
574 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 static void wake_up_idle_cpu(int cpu)
592 struct rq *rq = cpu_rq(cpu);
594 if (cpu == smp_processor_id())
597 if (set_nr_and_not_polling(rq->idle))
598 smp_send_reschedule(cpu);
600 trace_sched_wake_idle_without_ipi(cpu);
603 static bool wake_up_full_nohz_cpu(int cpu)
606 * We just need the target to call irq_exit() and re-evaluate
607 * the next tick. The nohz full kick at least implies that.
608 * If needed we can still optimize that later with an
611 if (cpu_is_offline(cpu))
612 return true; /* Don't try to wake offline CPUs. */
613 if (tick_nohz_full_cpu(cpu)) {
614 if (cpu != smp_processor_id() ||
615 tick_nohz_tick_stopped())
616 tick_nohz_full_kick_cpu(cpu);
624 * Wake up the specified CPU. If the CPU is going offline, it is the
625 * caller's responsibility to deal with the lost wakeup, for example,
626 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
628 void wake_up_nohz_cpu(int cpu)
630 if (!wake_up_full_nohz_cpu(cpu))
631 wake_up_idle_cpu(cpu);
634 static inline bool got_nohz_idle_kick(void)
636 int cpu = smp_processor_id();
638 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
641 if (idle_cpu(cpu) && !need_resched())
645 * We can't run Idle Load Balance on this CPU for this time so we
646 * cancel it and clear NOHZ_BALANCE_KICK
648 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
652 #else /* CONFIG_NO_HZ_COMMON */
654 static inline bool got_nohz_idle_kick(void)
659 #endif /* CONFIG_NO_HZ_COMMON */
661 #ifdef CONFIG_NO_HZ_FULL
662 bool sched_can_stop_tick(struct rq *rq)
666 /* Deadline tasks, even if single, need the tick */
667 if (rq->dl.dl_nr_running)
671 * If there are more than one RR tasks, we need the tick to effect the
672 * actual RR behaviour.
674 if (rq->rt.rr_nr_running) {
675 if (rq->rt.rr_nr_running == 1)
682 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
683 * forced preemption between FIFO tasks.
685 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
690 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
691 * if there's more than one we need the tick for involuntary
694 if (rq->nr_running > 1)
699 #endif /* CONFIG_NO_HZ_FULL */
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
713 struct task_group *parent, *child;
719 ret = (*down)(parent, data);
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
734 parent = parent->parent;
741 int tg_nop(struct task_group *tg, void *data)
747 static void set_load_weight(struct task_struct *p, bool update_load)
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
753 * SCHED_IDLE tasks get minimal weight:
755 if (task_has_idle_policy(p)) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
758 p->se.runnable_weight = load->weight;
763 * SCHED_OTHER tasks have to update their load when changing their
766 if (update_load && p->sched_class == &fair_sched_class) {
767 reweight_task(p, prio);
769 load->weight = scale_load(sched_prio_to_weight[prio]);
770 load->inv_weight = sched_prio_to_wmult[prio];
771 p->se.runnable_weight = load->weight;
775 #ifdef CONFIG_UCLAMP_TASK
776 /* Max allowed minimum utilization */
777 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
779 /* Max allowed maximum utilization */
780 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
782 /* All clamps are required to be less or equal than these values */
783 static struct uclamp_se uclamp_default[UCLAMP_CNT];
785 /* Integer rounded range for each bucket */
786 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
788 #define for_each_clamp_id(clamp_id) \
789 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
791 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
793 return clamp_value / UCLAMP_BUCKET_DELTA;
796 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
798 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
801 static inline unsigned int uclamp_none(int clamp_id)
803 if (clamp_id == UCLAMP_MIN)
805 return SCHED_CAPACITY_SCALE;
808 static inline void uclamp_se_set(struct uclamp_se *uc_se,
809 unsigned int value, bool user_defined)
811 uc_se->value = value;
812 uc_se->bucket_id = uclamp_bucket_id(value);
813 uc_se->user_defined = user_defined;
816 static inline unsigned int
817 uclamp_idle_value(struct rq *rq, unsigned int clamp_id,
818 unsigned int clamp_value)
821 * Avoid blocked utilization pushing up the frequency when we go
822 * idle (which drops the max-clamp) by retaining the last known
825 if (clamp_id == UCLAMP_MAX) {
826 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
830 return uclamp_none(UCLAMP_MIN);
833 static inline void uclamp_idle_reset(struct rq *rq, unsigned int clamp_id,
834 unsigned int clamp_value)
836 /* Reset max-clamp retention only on idle exit */
837 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
840 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
844 unsigned int uclamp_rq_max_value(struct rq *rq, unsigned int clamp_id,
845 unsigned int clamp_value)
847 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
848 int bucket_id = UCLAMP_BUCKETS - 1;
851 * Since both min and max clamps are max aggregated, find the
852 * top most bucket with tasks in.
854 for ( ; bucket_id >= 0; bucket_id--) {
855 if (!bucket[bucket_id].tasks)
857 return bucket[bucket_id].value;
860 /* No tasks -- default clamp values */
861 return uclamp_idle_value(rq, clamp_id, clamp_value);
865 * The effective clamp bucket index of a task depends on, by increasing
867 * - the task specific clamp value, when explicitly requested from userspace
868 * - the system default clamp value, defined by the sysadmin
870 static inline struct uclamp_se
871 uclamp_eff_get(struct task_struct *p, unsigned int clamp_id)
873 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
874 struct uclamp_se uc_max = uclamp_default[clamp_id];
876 /* System default restrictions always apply */
877 if (unlikely(uc_req.value > uc_max.value))
883 unsigned int uclamp_eff_value(struct task_struct *p, unsigned int clamp_id)
885 struct uclamp_se uc_eff;
887 /* Task currently refcounted: use back-annotated (effective) value */
888 if (p->uclamp[clamp_id].active)
889 return p->uclamp[clamp_id].value;
891 uc_eff = uclamp_eff_get(p, clamp_id);
897 * When a task is enqueued on a rq, the clamp bucket currently defined by the
898 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
899 * updates the rq's clamp value if required.
901 * Tasks can have a task-specific value requested from user-space, track
902 * within each bucket the maximum value for tasks refcounted in it.
903 * This "local max aggregation" allows to track the exact "requested" value
904 * for each bucket when all its RUNNABLE tasks require the same clamp.
906 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
907 unsigned int clamp_id)
909 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
910 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
911 struct uclamp_bucket *bucket;
913 lockdep_assert_held(&rq->lock);
915 /* Update task effective clamp */
916 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
918 bucket = &uc_rq->bucket[uc_se->bucket_id];
920 uc_se->active = true;
922 uclamp_idle_reset(rq, clamp_id, uc_se->value);
925 * Local max aggregation: rq buckets always track the max
926 * "requested" clamp value of its RUNNABLE tasks.
928 if (bucket->tasks == 1 || uc_se->value > bucket->value)
929 bucket->value = uc_se->value;
931 if (uc_se->value > READ_ONCE(uc_rq->value))
932 WRITE_ONCE(uc_rq->value, uc_se->value);
936 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
937 * is released. If this is the last task reference counting the rq's max
938 * active clamp value, then the rq's clamp value is updated.
940 * Both refcounted tasks and rq's cached clamp values are expected to be
941 * always valid. If it's detected they are not, as defensive programming,
942 * enforce the expected state and warn.
944 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
945 unsigned int 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;
950 unsigned int bkt_clamp;
951 unsigned int rq_clamp;
953 lockdep_assert_held(&rq->lock);
955 bucket = &uc_rq->bucket[uc_se->bucket_id];
956 SCHED_WARN_ON(!bucket->tasks);
957 if (likely(bucket->tasks))
959 uc_se->active = false;
962 * Keep "local max aggregation" simple and accept to (possibly)
963 * overboost some RUNNABLE tasks in the same bucket.
964 * The rq clamp bucket value is reset to its base value whenever
965 * there are no more RUNNABLE tasks refcounting it.
967 if (likely(bucket->tasks))
970 rq_clamp = READ_ONCE(uc_rq->value);
972 * Defensive programming: this should never happen. If it happens,
973 * e.g. due to future modification, warn and fixup the expected value.
975 SCHED_WARN_ON(bucket->value > rq_clamp);
976 if (bucket->value >= rq_clamp) {
977 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
978 WRITE_ONCE(uc_rq->value, bkt_clamp);
982 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
984 unsigned int clamp_id;
986 if (unlikely(!p->sched_class->uclamp_enabled))
989 for_each_clamp_id(clamp_id)
990 uclamp_rq_inc_id(rq, p, clamp_id);
992 /* Reset clamp idle holding when there is one RUNNABLE task */
993 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
994 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
997 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
999 unsigned int clamp_id;
1001 if (unlikely(!p->sched_class->uclamp_enabled))
1004 for_each_clamp_id(clamp_id)
1005 uclamp_rq_dec_id(rq, p, clamp_id);
1008 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1009 void __user *buffer, size_t *lenp,
1012 int old_min, old_max;
1013 static DEFINE_MUTEX(mutex);
1017 old_min = sysctl_sched_uclamp_util_min;
1018 old_max = sysctl_sched_uclamp_util_max;
1020 result = proc_dointvec(table, write, buffer, lenp, ppos);
1026 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1027 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1032 if (old_min != sysctl_sched_uclamp_util_min) {
1033 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1034 sysctl_sched_uclamp_util_min, false);
1036 if (old_max != sysctl_sched_uclamp_util_max) {
1037 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1038 sysctl_sched_uclamp_util_max, false);
1042 * Updating all the RUNNABLE task is expensive, keep it simple and do
1043 * just a lazy update at each next enqueue time.
1048 sysctl_sched_uclamp_util_min = old_min;
1049 sysctl_sched_uclamp_util_max = old_max;
1051 mutex_unlock(&mutex);
1056 static int uclamp_validate(struct task_struct *p,
1057 const struct sched_attr *attr)
1059 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1060 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1062 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1063 lower_bound = attr->sched_util_min;
1064 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1065 upper_bound = attr->sched_util_max;
1067 if (lower_bound > upper_bound)
1069 if (upper_bound > SCHED_CAPACITY_SCALE)
1075 static void __setscheduler_uclamp(struct task_struct *p,
1076 const struct sched_attr *attr)
1078 unsigned int clamp_id;
1081 * On scheduling class change, reset to default clamps for tasks
1082 * without a task-specific value.
1084 for_each_clamp_id(clamp_id) {
1085 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1086 unsigned int clamp_value = uclamp_none(clamp_id);
1088 /* Keep using defined clamps across class changes */
1089 if (uc_se->user_defined)
1092 /* By default, RT tasks always get 100% boost */
1093 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1094 clamp_value = uclamp_none(UCLAMP_MAX);
1096 uclamp_se_set(uc_se, clamp_value, false);
1099 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1102 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1103 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1104 attr->sched_util_min, true);
1107 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1108 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1109 attr->sched_util_max, true);
1113 static void uclamp_fork(struct task_struct *p)
1115 unsigned int clamp_id;
1117 for_each_clamp_id(clamp_id)
1118 p->uclamp[clamp_id].active = false;
1120 if (likely(!p->sched_reset_on_fork))
1123 for_each_clamp_id(clamp_id) {
1124 unsigned int clamp_value = uclamp_none(clamp_id);
1126 /* By default, RT tasks always get 100% boost */
1127 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1128 clamp_value = uclamp_none(UCLAMP_MAX);
1130 uclamp_se_set(&p->uclamp_req[clamp_id], clamp_value, false);
1134 static void __init init_uclamp(void)
1136 struct uclamp_se uc_max = {};
1137 unsigned int clamp_id;
1140 for_each_possible_cpu(cpu) {
1141 memset(&cpu_rq(cpu)->uclamp, 0, sizeof(struct uclamp_rq));
1142 cpu_rq(cpu)->uclamp_flags = 0;
1145 for_each_clamp_id(clamp_id) {
1146 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1147 uclamp_none(clamp_id), false);
1150 /* System defaults allow max clamp values for both indexes */
1151 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1152 for_each_clamp_id(clamp_id)
1153 uclamp_default[clamp_id] = uc_max;
1156 #else /* CONFIG_UCLAMP_TASK */
1157 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1158 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1159 static inline int uclamp_validate(struct task_struct *p,
1160 const struct sched_attr *attr)
1164 static void __setscheduler_uclamp(struct task_struct *p,
1165 const struct sched_attr *attr) { }
1166 static inline void uclamp_fork(struct task_struct *p) { }
1167 static inline void init_uclamp(void) { }
1168 #endif /* CONFIG_UCLAMP_TASK */
1170 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1172 if (!(flags & ENQUEUE_NOCLOCK))
1173 update_rq_clock(rq);
1175 if (!(flags & ENQUEUE_RESTORE)) {
1176 sched_info_queued(rq, p);
1177 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1180 uclamp_rq_inc(rq, p);
1181 p->sched_class->enqueue_task(rq, p, flags);
1184 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1186 if (!(flags & DEQUEUE_NOCLOCK))
1187 update_rq_clock(rq);
1189 if (!(flags & DEQUEUE_SAVE)) {
1190 sched_info_dequeued(rq, p);
1191 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1194 uclamp_rq_dec(rq, p);
1195 p->sched_class->dequeue_task(rq, p, flags);
1198 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1200 if (task_contributes_to_load(p))
1201 rq->nr_uninterruptible--;
1203 enqueue_task(rq, p, flags);
1205 p->on_rq = TASK_ON_RQ_QUEUED;
1208 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1210 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1212 if (task_contributes_to_load(p))
1213 rq->nr_uninterruptible++;
1215 dequeue_task(rq, p, flags);
1219 * __normal_prio - return the priority that is based on the static prio
1221 static inline int __normal_prio(struct task_struct *p)
1223 return p->static_prio;
1227 * Calculate the expected normal priority: i.e. priority
1228 * without taking RT-inheritance into account. Might be
1229 * boosted by interactivity modifiers. Changes upon fork,
1230 * setprio syscalls, and whenever the interactivity
1231 * estimator recalculates.
1233 static inline int normal_prio(struct task_struct *p)
1237 if (task_has_dl_policy(p))
1238 prio = MAX_DL_PRIO-1;
1239 else if (task_has_rt_policy(p))
1240 prio = MAX_RT_PRIO-1 - p->rt_priority;
1242 prio = __normal_prio(p);
1247 * Calculate the current priority, i.e. the priority
1248 * taken into account by the scheduler. This value might
1249 * be boosted by RT tasks, or might be boosted by
1250 * interactivity modifiers. Will be RT if the task got
1251 * RT-boosted. If not then it returns p->normal_prio.
1253 static int effective_prio(struct task_struct *p)
1255 p->normal_prio = normal_prio(p);
1257 * If we are RT tasks or we were boosted to RT priority,
1258 * keep the priority unchanged. Otherwise, update priority
1259 * to the normal priority:
1261 if (!rt_prio(p->prio))
1262 return p->normal_prio;
1267 * task_curr - is this task currently executing on a CPU?
1268 * @p: the task in question.
1270 * Return: 1 if the task is currently executing. 0 otherwise.
1272 inline int task_curr(const struct task_struct *p)
1274 return cpu_curr(task_cpu(p)) == p;
1278 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1279 * use the balance_callback list if you want balancing.
1281 * this means any call to check_class_changed() must be followed by a call to
1282 * balance_callback().
1284 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1285 const struct sched_class *prev_class,
1288 if (prev_class != p->sched_class) {
1289 if (prev_class->switched_from)
1290 prev_class->switched_from(rq, p);
1292 p->sched_class->switched_to(rq, p);
1293 } else if (oldprio != p->prio || dl_task(p))
1294 p->sched_class->prio_changed(rq, p, oldprio);
1297 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1299 const struct sched_class *class;
1301 if (p->sched_class == rq->curr->sched_class) {
1302 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1304 for_each_class(class) {
1305 if (class == rq->curr->sched_class)
1307 if (class == p->sched_class) {
1315 * A queue event has occurred, and we're going to schedule. In
1316 * this case, we can save a useless back to back clock update.
1318 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1319 rq_clock_skip_update(rq);
1324 static inline bool is_per_cpu_kthread(struct task_struct *p)
1326 if (!(p->flags & PF_KTHREAD))
1329 if (p->nr_cpus_allowed != 1)
1336 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1337 * __set_cpus_allowed_ptr() and select_fallback_rq().
1339 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1341 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1344 if (is_per_cpu_kthread(p))
1345 return cpu_online(cpu);
1347 return cpu_active(cpu);
1351 * This is how migration works:
1353 * 1) we invoke migration_cpu_stop() on the target CPU using
1355 * 2) stopper starts to run (implicitly forcing the migrated thread
1357 * 3) it checks whether the migrated task is still in the wrong runqueue.
1358 * 4) if it's in the wrong runqueue then the migration thread removes
1359 * it and puts it into the right queue.
1360 * 5) stopper completes and stop_one_cpu() returns and the migration
1365 * move_queued_task - move a queued task to new rq.
1367 * Returns (locked) new rq. Old rq's lock is released.
1369 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1370 struct task_struct *p, int new_cpu)
1372 lockdep_assert_held(&rq->lock);
1374 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1375 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1376 set_task_cpu(p, new_cpu);
1379 rq = cpu_rq(new_cpu);
1382 BUG_ON(task_cpu(p) != new_cpu);
1383 enqueue_task(rq, p, 0);
1384 p->on_rq = TASK_ON_RQ_QUEUED;
1385 check_preempt_curr(rq, p, 0);
1390 struct migration_arg {
1391 struct task_struct *task;
1396 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1397 * this because either it can't run here any more (set_cpus_allowed()
1398 * away from this CPU, or CPU going down), or because we're
1399 * attempting to rebalance this task on exec (sched_exec).
1401 * So we race with normal scheduler movements, but that's OK, as long
1402 * as the task is no longer on this CPU.
1404 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1405 struct task_struct *p, int dest_cpu)
1407 /* Affinity changed (again). */
1408 if (!is_cpu_allowed(p, dest_cpu))
1411 update_rq_clock(rq);
1412 rq = move_queued_task(rq, rf, p, dest_cpu);
1418 * migration_cpu_stop - this will be executed by a highprio stopper thread
1419 * and performs thread migration by bumping thread off CPU then
1420 * 'pushing' onto another runqueue.
1422 static int migration_cpu_stop(void *data)
1424 struct migration_arg *arg = data;
1425 struct task_struct *p = arg->task;
1426 struct rq *rq = this_rq();
1430 * The original target CPU might have gone down and we might
1431 * be on another CPU but it doesn't matter.
1433 local_irq_disable();
1435 * We need to explicitly wake pending tasks before running
1436 * __migrate_task() such that we will not miss enforcing cpus_ptr
1437 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1439 sched_ttwu_pending();
1441 raw_spin_lock(&p->pi_lock);
1444 * If task_rq(p) != rq, it cannot be migrated here, because we're
1445 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1446 * we're holding p->pi_lock.
1448 if (task_rq(p) == rq) {
1449 if (task_on_rq_queued(p))
1450 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1452 p->wake_cpu = arg->dest_cpu;
1455 raw_spin_unlock(&p->pi_lock);
1462 * sched_class::set_cpus_allowed must do the below, but is not required to
1463 * actually call this function.
1465 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1467 cpumask_copy(&p->cpus_mask, new_mask);
1468 p->nr_cpus_allowed = cpumask_weight(new_mask);
1471 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1473 struct rq *rq = task_rq(p);
1474 bool queued, running;
1476 lockdep_assert_held(&p->pi_lock);
1478 queued = task_on_rq_queued(p);
1479 running = task_current(rq, p);
1483 * Because __kthread_bind() calls this on blocked tasks without
1486 lockdep_assert_held(&rq->lock);
1487 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1490 put_prev_task(rq, p);
1492 p->sched_class->set_cpus_allowed(p, new_mask);
1495 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1497 set_curr_task(rq, p);
1501 * Change a given task's CPU affinity. Migrate the thread to a
1502 * proper CPU and schedule it away if the CPU it's executing on
1503 * is removed from the allowed bitmask.
1505 * NOTE: the caller must have a valid reference to the task, the
1506 * task must not exit() & deallocate itself prematurely. The
1507 * call is not atomic; no spinlocks may be held.
1509 static int __set_cpus_allowed_ptr(struct task_struct *p,
1510 const struct cpumask *new_mask, bool check)
1512 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1513 unsigned int dest_cpu;
1518 rq = task_rq_lock(p, &rf);
1519 update_rq_clock(rq);
1521 if (p->flags & PF_KTHREAD) {
1523 * Kernel threads are allowed on online && !active CPUs
1525 cpu_valid_mask = cpu_online_mask;
1529 * Must re-check here, to close a race against __kthread_bind(),
1530 * sched_setaffinity() is not guaranteed to observe the flag.
1532 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1537 if (cpumask_equal(p->cpus_ptr, new_mask))
1540 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1545 do_set_cpus_allowed(p, new_mask);
1547 if (p->flags & PF_KTHREAD) {
1549 * For kernel threads that do indeed end up on online &&
1550 * !active we want to ensure they are strict per-CPU threads.
1552 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1553 !cpumask_intersects(new_mask, cpu_active_mask) &&
1554 p->nr_cpus_allowed != 1);
1557 /* Can the task run on the task's current CPU? If so, we're done */
1558 if (cpumask_test_cpu(task_cpu(p), new_mask))
1561 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1562 if (task_running(rq, p) || p->state == TASK_WAKING) {
1563 struct migration_arg arg = { p, dest_cpu };
1564 /* Need help from migration thread: drop lock and wait. */
1565 task_rq_unlock(rq, p, &rf);
1566 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1568 } else if (task_on_rq_queued(p)) {
1570 * OK, since we're going to drop the lock immediately
1571 * afterwards anyway.
1573 rq = move_queued_task(rq, &rf, p, dest_cpu);
1576 task_rq_unlock(rq, p, &rf);
1581 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1583 return __set_cpus_allowed_ptr(p, new_mask, false);
1585 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1587 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1589 #ifdef CONFIG_SCHED_DEBUG
1591 * We should never call set_task_cpu() on a blocked task,
1592 * ttwu() will sort out the placement.
1594 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1598 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1599 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1600 * time relying on p->on_rq.
1602 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1603 p->sched_class == &fair_sched_class &&
1604 (p->on_rq && !task_on_rq_migrating(p)));
1606 #ifdef CONFIG_LOCKDEP
1608 * The caller should hold either p->pi_lock or rq->lock, when changing
1609 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1611 * sched_move_task() holds both and thus holding either pins the cgroup,
1614 * Furthermore, all task_rq users should acquire both locks, see
1617 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1618 lockdep_is_held(&task_rq(p)->lock)));
1621 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1623 WARN_ON_ONCE(!cpu_online(new_cpu));
1626 trace_sched_migrate_task(p, new_cpu);
1628 if (task_cpu(p) != new_cpu) {
1629 if (p->sched_class->migrate_task_rq)
1630 p->sched_class->migrate_task_rq(p, new_cpu);
1631 p->se.nr_migrations++;
1633 perf_event_task_migrate(p);
1636 __set_task_cpu(p, new_cpu);
1639 #ifdef CONFIG_NUMA_BALANCING
1640 static void __migrate_swap_task(struct task_struct *p, int cpu)
1642 if (task_on_rq_queued(p)) {
1643 struct rq *src_rq, *dst_rq;
1644 struct rq_flags srf, drf;
1646 src_rq = task_rq(p);
1647 dst_rq = cpu_rq(cpu);
1649 rq_pin_lock(src_rq, &srf);
1650 rq_pin_lock(dst_rq, &drf);
1652 deactivate_task(src_rq, p, 0);
1653 set_task_cpu(p, cpu);
1654 activate_task(dst_rq, p, 0);
1655 check_preempt_curr(dst_rq, p, 0);
1657 rq_unpin_lock(dst_rq, &drf);
1658 rq_unpin_lock(src_rq, &srf);
1662 * Task isn't running anymore; make it appear like we migrated
1663 * it before it went to sleep. This means on wakeup we make the
1664 * previous CPU our target instead of where it really is.
1670 struct migration_swap_arg {
1671 struct task_struct *src_task, *dst_task;
1672 int src_cpu, dst_cpu;
1675 static int migrate_swap_stop(void *data)
1677 struct migration_swap_arg *arg = data;
1678 struct rq *src_rq, *dst_rq;
1681 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1684 src_rq = cpu_rq(arg->src_cpu);
1685 dst_rq = cpu_rq(arg->dst_cpu);
1687 double_raw_lock(&arg->src_task->pi_lock,
1688 &arg->dst_task->pi_lock);
1689 double_rq_lock(src_rq, dst_rq);
1691 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1694 if (task_cpu(arg->src_task) != arg->src_cpu)
1697 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1700 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1703 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1704 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1709 double_rq_unlock(src_rq, dst_rq);
1710 raw_spin_unlock(&arg->dst_task->pi_lock);
1711 raw_spin_unlock(&arg->src_task->pi_lock);
1717 * Cross migrate two tasks
1719 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1720 int target_cpu, int curr_cpu)
1722 struct migration_swap_arg arg;
1725 arg = (struct migration_swap_arg){
1727 .src_cpu = curr_cpu,
1729 .dst_cpu = target_cpu,
1732 if (arg.src_cpu == arg.dst_cpu)
1736 * These three tests are all lockless; this is OK since all of them
1737 * will be re-checked with proper locks held further down the line.
1739 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1742 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1745 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1748 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1749 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1754 #endif /* CONFIG_NUMA_BALANCING */
1757 * wait_task_inactive - wait for a thread to unschedule.
1759 * If @match_state is nonzero, it's the @p->state value just checked and
1760 * not expected to change. If it changes, i.e. @p might have woken up,
1761 * then return zero. When we succeed in waiting for @p to be off its CPU,
1762 * we return a positive number (its total switch count). If a second call
1763 * a short while later returns the same number, the caller can be sure that
1764 * @p has remained unscheduled the whole time.
1766 * The caller must ensure that the task *will* unschedule sometime soon,
1767 * else this function might spin for a *long* time. This function can't
1768 * be called with interrupts off, or it may introduce deadlock with
1769 * smp_call_function() if an IPI is sent by the same process we are
1770 * waiting to become inactive.
1772 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1774 int running, queued;
1781 * We do the initial early heuristics without holding
1782 * any task-queue locks at all. We'll only try to get
1783 * the runqueue lock when things look like they will
1789 * If the task is actively running on another CPU
1790 * still, just relax and busy-wait without holding
1793 * NOTE! Since we don't hold any locks, it's not
1794 * even sure that "rq" stays as the right runqueue!
1795 * But we don't care, since "task_running()" will
1796 * return false if the runqueue has changed and p
1797 * is actually now running somewhere else!
1799 while (task_running(rq, p)) {
1800 if (match_state && unlikely(p->state != match_state))
1806 * Ok, time to look more closely! We need the rq
1807 * lock now, to be *sure*. If we're wrong, we'll
1808 * just go back and repeat.
1810 rq = task_rq_lock(p, &rf);
1811 trace_sched_wait_task(p);
1812 running = task_running(rq, p);
1813 queued = task_on_rq_queued(p);
1815 if (!match_state || p->state == match_state)
1816 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1817 task_rq_unlock(rq, p, &rf);
1820 * If it changed from the expected state, bail out now.
1822 if (unlikely(!ncsw))
1826 * Was it really running after all now that we
1827 * checked with the proper locks actually held?
1829 * Oops. Go back and try again..
1831 if (unlikely(running)) {
1837 * It's not enough that it's not actively running,
1838 * it must be off the runqueue _entirely_, and not
1841 * So if it was still runnable (but just not actively
1842 * running right now), it's preempted, and we should
1843 * yield - it could be a while.
1845 if (unlikely(queued)) {
1846 ktime_t to = NSEC_PER_SEC / HZ;
1848 set_current_state(TASK_UNINTERRUPTIBLE);
1849 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1854 * Ahh, all good. It wasn't running, and it wasn't
1855 * runnable, which means that it will never become
1856 * running in the future either. We're all done!
1865 * kick_process - kick a running thread to enter/exit the kernel
1866 * @p: the to-be-kicked thread
1868 * Cause a process which is running on another CPU to enter
1869 * kernel-mode, without any delay. (to get signals handled.)
1871 * NOTE: this function doesn't have to take the runqueue lock,
1872 * because all it wants to ensure is that the remote task enters
1873 * the kernel. If the IPI races and the task has been migrated
1874 * to another CPU then no harm is done and the purpose has been
1877 void kick_process(struct task_struct *p)
1883 if ((cpu != smp_processor_id()) && task_curr(p))
1884 smp_send_reschedule(cpu);
1887 EXPORT_SYMBOL_GPL(kick_process);
1890 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
1892 * A few notes on cpu_active vs cpu_online:
1894 * - cpu_active must be a subset of cpu_online
1896 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1897 * see __set_cpus_allowed_ptr(). At this point the newly online
1898 * CPU isn't yet part of the sched domains, and balancing will not
1901 * - on CPU-down we clear cpu_active() to mask the sched domains and
1902 * avoid the load balancer to place new tasks on the to be removed
1903 * CPU. Existing tasks will remain running there and will be taken
1906 * This means that fallback selection must not select !active CPUs.
1907 * And can assume that any active CPU must be online. Conversely
1908 * select_task_rq() below may allow selection of !active CPUs in order
1909 * to satisfy the above rules.
1911 static int select_fallback_rq(int cpu, struct task_struct *p)
1913 int nid = cpu_to_node(cpu);
1914 const struct cpumask *nodemask = NULL;
1915 enum { cpuset, possible, fail } state = cpuset;
1919 * If the node that the CPU is on has been offlined, cpu_to_node()
1920 * will return -1. There is no CPU on the node, and we should
1921 * select the CPU on the other node.
1924 nodemask = cpumask_of_node(nid);
1926 /* Look for allowed, online CPU in same node. */
1927 for_each_cpu(dest_cpu, nodemask) {
1928 if (!cpu_active(dest_cpu))
1930 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
1936 /* Any allowed, online CPU? */
1937 for_each_cpu(dest_cpu, p->cpus_ptr) {
1938 if (!is_cpu_allowed(p, dest_cpu))
1944 /* No more Mr. Nice Guy. */
1947 if (IS_ENABLED(CONFIG_CPUSETS)) {
1948 cpuset_cpus_allowed_fallback(p);
1954 do_set_cpus_allowed(p, cpu_possible_mask);
1965 if (state != cpuset) {
1967 * Don't tell them about moving exiting tasks or
1968 * kernel threads (both mm NULL), since they never
1971 if (p->mm && printk_ratelimit()) {
1972 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1973 task_pid_nr(p), p->comm, cpu);
1981 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
1984 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1986 lockdep_assert_held(&p->pi_lock);
1988 if (p->nr_cpus_allowed > 1)
1989 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1991 cpu = cpumask_any(p->cpus_ptr);
1994 * In order not to call set_task_cpu() on a blocking task we need
1995 * to rely on ttwu() to place the task on a valid ->cpus_ptr
1998 * Since this is common to all placement strategies, this lives here.
2000 * [ this allows ->select_task() to simply return task_cpu(p) and
2001 * not worry about this generic constraint ]
2003 if (unlikely(!is_cpu_allowed(p, cpu)))
2004 cpu = select_fallback_rq(task_cpu(p), p);
2009 static void update_avg(u64 *avg, u64 sample)
2011 s64 diff = sample - *avg;
2015 void sched_set_stop_task(int cpu, struct task_struct *stop)
2017 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2018 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2022 * Make it appear like a SCHED_FIFO task, its something
2023 * userspace knows about and won't get confused about.
2025 * Also, it will make PI more or less work without too
2026 * much confusion -- but then, stop work should not
2027 * rely on PI working anyway.
2029 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2031 stop->sched_class = &stop_sched_class;
2034 cpu_rq(cpu)->stop = stop;
2038 * Reset it back to a normal scheduling class so that
2039 * it can die in pieces.
2041 old_stop->sched_class = &rt_sched_class;
2047 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2048 const struct cpumask *new_mask, bool check)
2050 return set_cpus_allowed_ptr(p, new_mask);
2053 #endif /* CONFIG_SMP */
2056 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2060 if (!schedstat_enabled())
2066 if (cpu == rq->cpu) {
2067 __schedstat_inc(rq->ttwu_local);
2068 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2070 struct sched_domain *sd;
2072 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2074 for_each_domain(rq->cpu, sd) {
2075 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2076 __schedstat_inc(sd->ttwu_wake_remote);
2083 if (wake_flags & WF_MIGRATED)
2084 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2085 #endif /* CONFIG_SMP */
2087 __schedstat_inc(rq->ttwu_count);
2088 __schedstat_inc(p->se.statistics.nr_wakeups);
2090 if (wake_flags & WF_SYNC)
2091 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2095 * Mark the task runnable and perform wakeup-preemption.
2097 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2098 struct rq_flags *rf)
2100 check_preempt_curr(rq, p, wake_flags);
2101 p->state = TASK_RUNNING;
2102 trace_sched_wakeup(p);
2105 if (p->sched_class->task_woken) {
2107 * Our task @p is fully woken up and running; so its safe to
2108 * drop the rq->lock, hereafter rq is only used for statistics.
2110 rq_unpin_lock(rq, rf);
2111 p->sched_class->task_woken(rq, p);
2112 rq_repin_lock(rq, rf);
2115 if (rq->idle_stamp) {
2116 u64 delta = rq_clock(rq) - rq->idle_stamp;
2117 u64 max = 2*rq->max_idle_balance_cost;
2119 update_avg(&rq->avg_idle, delta);
2121 if (rq->avg_idle > max)
2130 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2131 struct rq_flags *rf)
2133 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2135 lockdep_assert_held(&rq->lock);
2138 if (p->sched_contributes_to_load)
2139 rq->nr_uninterruptible--;
2141 if (wake_flags & WF_MIGRATED)
2142 en_flags |= ENQUEUE_MIGRATED;
2145 activate_task(rq, p, en_flags);
2146 ttwu_do_wakeup(rq, p, wake_flags, rf);
2150 * Called in case the task @p isn't fully descheduled from its runqueue,
2151 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2152 * since all we need to do is flip p->state to TASK_RUNNING, since
2153 * the task is still ->on_rq.
2155 static int ttwu_remote(struct task_struct *p, int wake_flags)
2161 rq = __task_rq_lock(p, &rf);
2162 if (task_on_rq_queued(p)) {
2163 /* check_preempt_curr() may use rq clock */
2164 update_rq_clock(rq);
2165 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2168 __task_rq_unlock(rq, &rf);
2174 void sched_ttwu_pending(void)
2176 struct rq *rq = this_rq();
2177 struct llist_node *llist = llist_del_all(&rq->wake_list);
2178 struct task_struct *p, *t;
2184 rq_lock_irqsave(rq, &rf);
2185 update_rq_clock(rq);
2187 llist_for_each_entry_safe(p, t, llist, wake_entry)
2188 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2190 rq_unlock_irqrestore(rq, &rf);
2193 void scheduler_ipi(void)
2196 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2197 * TIF_NEED_RESCHED remotely (for the first time) will also send
2200 preempt_fold_need_resched();
2202 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2206 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2207 * traditionally all their work was done from the interrupt return
2208 * path. Now that we actually do some work, we need to make sure
2211 * Some archs already do call them, luckily irq_enter/exit nest
2214 * Arguably we should visit all archs and update all handlers,
2215 * however a fair share of IPIs are still resched only so this would
2216 * somewhat pessimize the simple resched case.
2219 sched_ttwu_pending();
2222 * Check if someone kicked us for doing the nohz idle load balance.
2224 if (unlikely(got_nohz_idle_kick())) {
2225 this_rq()->idle_balance = 1;
2226 raise_softirq_irqoff(SCHED_SOFTIRQ);
2231 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2233 struct rq *rq = cpu_rq(cpu);
2235 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2237 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2238 if (!set_nr_if_polling(rq->idle))
2239 smp_send_reschedule(cpu);
2241 trace_sched_wake_idle_without_ipi(cpu);
2245 void wake_up_if_idle(int cpu)
2247 struct rq *rq = cpu_rq(cpu);
2252 if (!is_idle_task(rcu_dereference(rq->curr)))
2255 if (set_nr_if_polling(rq->idle)) {
2256 trace_sched_wake_idle_without_ipi(cpu);
2258 rq_lock_irqsave(rq, &rf);
2259 if (is_idle_task(rq->curr))
2260 smp_send_reschedule(cpu);
2261 /* Else CPU is not idle, do nothing here: */
2262 rq_unlock_irqrestore(rq, &rf);
2269 bool cpus_share_cache(int this_cpu, int that_cpu)
2271 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2273 #endif /* CONFIG_SMP */
2275 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2277 struct rq *rq = cpu_rq(cpu);
2280 #if defined(CONFIG_SMP)
2281 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2282 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2283 ttwu_queue_remote(p, cpu, wake_flags);
2289 update_rq_clock(rq);
2290 ttwu_do_activate(rq, p, wake_flags, &rf);
2295 * Notes on Program-Order guarantees on SMP systems.
2299 * The basic program-order guarantee on SMP systems is that when a task [t]
2300 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2301 * execution on its new CPU [c1].
2303 * For migration (of runnable tasks) this is provided by the following means:
2305 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2306 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2307 * rq(c1)->lock (if not at the same time, then in that order).
2308 * C) LOCK of the rq(c1)->lock scheduling in task
2310 * Release/acquire chaining guarantees that B happens after A and C after B.
2311 * Note: the CPU doing B need not be c0 or c1
2320 * UNLOCK rq(0)->lock
2322 * LOCK rq(0)->lock // orders against CPU0
2324 * UNLOCK rq(0)->lock
2328 * UNLOCK rq(1)->lock
2330 * LOCK rq(1)->lock // orders against CPU2
2333 * UNLOCK rq(1)->lock
2336 * BLOCKING -- aka. SLEEP + WAKEUP
2338 * For blocking we (obviously) need to provide the same guarantee as for
2339 * migration. However the means are completely different as there is no lock
2340 * chain to provide order. Instead we do:
2342 * 1) smp_store_release(X->on_cpu, 0)
2343 * 2) smp_cond_load_acquire(!X->on_cpu)
2347 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2349 * LOCK rq(0)->lock LOCK X->pi_lock
2352 * smp_store_release(X->on_cpu, 0);
2354 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2360 * X->state = RUNNING
2361 * UNLOCK rq(2)->lock
2363 * LOCK rq(2)->lock // orders against CPU1
2366 * UNLOCK rq(2)->lock
2369 * UNLOCK rq(0)->lock
2372 * However, for wakeups there is a second guarantee we must provide, namely we
2373 * must ensure that CONDITION=1 done by the caller can not be reordered with
2374 * accesses to the task state; see try_to_wake_up() and set_current_state().
2378 * try_to_wake_up - wake up a thread
2379 * @p: the thread to be awakened
2380 * @state: the mask of task states that can be woken
2381 * @wake_flags: wake modifier flags (WF_*)
2383 * If (@state & @p->state) @p->state = TASK_RUNNING.
2385 * If the task was not queued/runnable, also place it back on a runqueue.
2387 * Atomic against schedule() which would dequeue a task, also see
2388 * set_current_state().
2390 * This function executes a full memory barrier before accessing the task
2391 * state; see set_current_state().
2393 * Return: %true if @p->state changes (an actual wakeup was done),
2397 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2399 unsigned long flags;
2400 int cpu, success = 0;
2405 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2406 * == smp_processor_id()'. Together this means we can special
2407 * case the whole 'p->on_rq && ttwu_remote()' case below
2408 * without taking any locks.
2411 * - we rely on Program-Order guarantees for all the ordering,
2412 * - we're serialized against set_special_state() by virtue of
2413 * it disabling IRQs (this allows not taking ->pi_lock).
2415 if (!(p->state & state))
2420 trace_sched_waking(p);
2421 p->state = TASK_RUNNING;
2422 trace_sched_wakeup(p);
2427 * If we are going to wake up a thread waiting for CONDITION we
2428 * need to ensure that CONDITION=1 done by the caller can not be
2429 * reordered with p->state check below. This pairs with mb() in
2430 * set_current_state() the waiting thread does.
2432 raw_spin_lock_irqsave(&p->pi_lock, flags);
2433 smp_mb__after_spinlock();
2434 if (!(p->state & state))
2437 trace_sched_waking(p);
2439 /* We're going to change ->state: */
2444 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2445 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2446 * in smp_cond_load_acquire() below.
2448 * sched_ttwu_pending() try_to_wake_up()
2449 * STORE p->on_rq = 1 LOAD p->state
2452 * __schedule() (switch to task 'p')
2453 * LOCK rq->lock smp_rmb();
2454 * smp_mb__after_spinlock();
2458 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2460 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2461 * __schedule(). See the comment for smp_mb__after_spinlock().
2464 if (p->on_rq && ttwu_remote(p, wake_flags))
2469 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2470 * possible to, falsely, observe p->on_cpu == 0.
2472 * One must be running (->on_cpu == 1) in order to remove oneself
2473 * from the runqueue.
2475 * __schedule() (switch to task 'p') try_to_wake_up()
2476 * STORE p->on_cpu = 1 LOAD p->on_rq
2479 * __schedule() (put 'p' to sleep)
2480 * LOCK rq->lock smp_rmb();
2481 * smp_mb__after_spinlock();
2482 * STORE p->on_rq = 0 LOAD p->on_cpu
2484 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2485 * __schedule(). See the comment for smp_mb__after_spinlock().
2490 * If the owning (remote) CPU is still in the middle of schedule() with
2491 * this task as prev, wait until its done referencing the task.
2493 * Pairs with the smp_store_release() in finish_task().
2495 * This ensures that tasks getting woken will be fully ordered against
2496 * their previous state and preserve Program Order.
2498 smp_cond_load_acquire(&p->on_cpu, !VAL);
2500 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2501 p->state = TASK_WAKING;
2504 delayacct_blkio_end(p);
2505 atomic_dec(&task_rq(p)->nr_iowait);
2508 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2509 if (task_cpu(p) != cpu) {
2510 wake_flags |= WF_MIGRATED;
2511 psi_ttwu_dequeue(p);
2512 set_task_cpu(p, cpu);
2515 #else /* CONFIG_SMP */
2518 delayacct_blkio_end(p);
2519 atomic_dec(&task_rq(p)->nr_iowait);
2522 #endif /* CONFIG_SMP */
2524 ttwu_queue(p, cpu, wake_flags);
2526 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2529 ttwu_stat(p, cpu, wake_flags);
2536 * wake_up_process - Wake up a specific process
2537 * @p: The process to be woken up.
2539 * Attempt to wake up the nominated process and move it to the set of runnable
2542 * Return: 1 if the process was woken up, 0 if it was already running.
2544 * This function executes a full memory barrier before accessing the task state.
2546 int wake_up_process(struct task_struct *p)
2548 return try_to_wake_up(p, TASK_NORMAL, 0);
2550 EXPORT_SYMBOL(wake_up_process);
2552 int wake_up_state(struct task_struct *p, unsigned int state)
2554 return try_to_wake_up(p, state, 0);
2558 * Perform scheduler related setup for a newly forked process p.
2559 * p is forked by current.
2561 * __sched_fork() is basic setup used by init_idle() too:
2563 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2568 p->se.exec_start = 0;
2569 p->se.sum_exec_runtime = 0;
2570 p->se.prev_sum_exec_runtime = 0;
2571 p->se.nr_migrations = 0;
2573 INIT_LIST_HEAD(&p->se.group_node);
2575 #ifdef CONFIG_FAIR_GROUP_SCHED
2576 p->se.cfs_rq = NULL;
2579 #ifdef CONFIG_SCHEDSTATS
2580 /* Even if schedstat is disabled, there should not be garbage */
2581 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2584 RB_CLEAR_NODE(&p->dl.rb_node);
2585 init_dl_task_timer(&p->dl);
2586 init_dl_inactive_task_timer(&p->dl);
2587 __dl_clear_params(p);
2589 INIT_LIST_HEAD(&p->rt.run_list);
2591 p->rt.time_slice = sched_rr_timeslice;
2595 #ifdef CONFIG_PREEMPT_NOTIFIERS
2596 INIT_HLIST_HEAD(&p->preempt_notifiers);
2599 #ifdef CONFIG_COMPACTION
2600 p->capture_control = NULL;
2602 init_numa_balancing(clone_flags, p);
2605 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2607 #ifdef CONFIG_NUMA_BALANCING
2609 void set_numabalancing_state(bool enabled)
2612 static_branch_enable(&sched_numa_balancing);
2614 static_branch_disable(&sched_numa_balancing);
2617 #ifdef CONFIG_PROC_SYSCTL
2618 int sysctl_numa_balancing(struct ctl_table *table, int write,
2619 void __user *buffer, size_t *lenp, loff_t *ppos)
2623 int state = static_branch_likely(&sched_numa_balancing);
2625 if (write && !capable(CAP_SYS_ADMIN))
2630 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2634 set_numabalancing_state(state);
2640 #ifdef CONFIG_SCHEDSTATS
2642 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2643 static bool __initdata __sched_schedstats = false;
2645 static void set_schedstats(bool enabled)
2648 static_branch_enable(&sched_schedstats);
2650 static_branch_disable(&sched_schedstats);
2653 void force_schedstat_enabled(void)
2655 if (!schedstat_enabled()) {
2656 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2657 static_branch_enable(&sched_schedstats);
2661 static int __init setup_schedstats(char *str)
2668 * This code is called before jump labels have been set up, so we can't
2669 * change the static branch directly just yet. Instead set a temporary
2670 * variable so init_schedstats() can do it later.
2672 if (!strcmp(str, "enable")) {
2673 __sched_schedstats = true;
2675 } else if (!strcmp(str, "disable")) {
2676 __sched_schedstats = false;
2681 pr_warn("Unable to parse schedstats=\n");
2685 __setup("schedstats=", setup_schedstats);
2687 static void __init init_schedstats(void)
2689 set_schedstats(__sched_schedstats);
2692 #ifdef CONFIG_PROC_SYSCTL
2693 int sysctl_schedstats(struct ctl_table *table, int write,
2694 void __user *buffer, size_t *lenp, loff_t *ppos)
2698 int state = static_branch_likely(&sched_schedstats);
2700 if (write && !capable(CAP_SYS_ADMIN))
2705 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2709 set_schedstats(state);
2712 #endif /* CONFIG_PROC_SYSCTL */
2713 #else /* !CONFIG_SCHEDSTATS */
2714 static inline void init_schedstats(void) {}
2715 #endif /* CONFIG_SCHEDSTATS */
2718 * fork()/clone()-time setup:
2720 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2722 unsigned long flags;
2724 __sched_fork(clone_flags, p);
2726 * We mark the process as NEW here. This guarantees that
2727 * nobody will actually run it, and a signal or other external
2728 * event cannot wake it up and insert it on the runqueue either.
2730 p->state = TASK_NEW;
2733 * Make sure we do not leak PI boosting priority to the child.
2735 p->prio = current->normal_prio;
2740 * Revert to default priority/policy on fork if requested.
2742 if (unlikely(p->sched_reset_on_fork)) {
2743 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2744 p->policy = SCHED_NORMAL;
2745 p->static_prio = NICE_TO_PRIO(0);
2747 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2748 p->static_prio = NICE_TO_PRIO(0);
2750 p->prio = p->normal_prio = __normal_prio(p);
2751 set_load_weight(p, false);
2754 * We don't need the reset flag anymore after the fork. It has
2755 * fulfilled its duty:
2757 p->sched_reset_on_fork = 0;
2760 if (dl_prio(p->prio))
2762 else if (rt_prio(p->prio))
2763 p->sched_class = &rt_sched_class;
2765 p->sched_class = &fair_sched_class;
2767 init_entity_runnable_average(&p->se);
2770 * The child is not yet in the pid-hash so no cgroup attach races,
2771 * and the cgroup is pinned to this child due to cgroup_fork()
2772 * is ran before sched_fork().
2774 * Silence PROVE_RCU.
2776 raw_spin_lock_irqsave(&p->pi_lock, flags);
2778 * We're setting the CPU for the first time, we don't migrate,
2779 * so use __set_task_cpu().
2781 __set_task_cpu(p, smp_processor_id());
2782 if (p->sched_class->task_fork)
2783 p->sched_class->task_fork(p);
2784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2786 #ifdef CONFIG_SCHED_INFO
2787 if (likely(sched_info_on()))
2788 memset(&p->sched_info, 0, sizeof(p->sched_info));
2790 #if defined(CONFIG_SMP)
2793 init_task_preempt_count(p);
2795 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2796 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2801 unsigned long to_ratio(u64 period, u64 runtime)
2803 if (runtime == RUNTIME_INF)
2807 * Doing this here saves a lot of checks in all
2808 * the calling paths, and returning zero seems
2809 * safe for them anyway.
2814 return div64_u64(runtime << BW_SHIFT, period);
2818 * wake_up_new_task - wake up a newly created task for the first time.
2820 * This function will do some initial scheduler statistics housekeeping
2821 * that must be done for every newly created context, then puts the task
2822 * on the runqueue and wakes it.
2824 void wake_up_new_task(struct task_struct *p)
2829 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2830 p->state = TASK_RUNNING;
2833 * Fork balancing, do it here and not earlier because:
2834 * - cpus_ptr can change in the fork path
2835 * - any previously selected CPU might disappear through hotplug
2837 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2838 * as we're not fully set-up yet.
2840 p->recent_used_cpu = task_cpu(p);
2841 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2843 rq = __task_rq_lock(p, &rf);
2844 update_rq_clock(rq);
2845 post_init_entity_util_avg(p);
2847 activate_task(rq, p, ENQUEUE_NOCLOCK);
2848 trace_sched_wakeup_new(p);
2849 check_preempt_curr(rq, p, WF_FORK);
2851 if (p->sched_class->task_woken) {
2853 * Nothing relies on rq->lock after this, so its fine to
2856 rq_unpin_lock(rq, &rf);
2857 p->sched_class->task_woken(rq, p);
2858 rq_repin_lock(rq, &rf);
2861 task_rq_unlock(rq, p, &rf);
2864 #ifdef CONFIG_PREEMPT_NOTIFIERS
2866 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2868 void preempt_notifier_inc(void)
2870 static_branch_inc(&preempt_notifier_key);
2872 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2874 void preempt_notifier_dec(void)
2876 static_branch_dec(&preempt_notifier_key);
2878 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2881 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2882 * @notifier: notifier struct to register
2884 void preempt_notifier_register(struct preempt_notifier *notifier)
2886 if (!static_branch_unlikely(&preempt_notifier_key))
2887 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2889 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2891 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2894 * preempt_notifier_unregister - no longer interested in preemption notifications
2895 * @notifier: notifier struct to unregister
2897 * This is *not* safe to call from within a preemption notifier.
2899 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2901 hlist_del(¬ifier->link);
2903 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2905 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2907 struct preempt_notifier *notifier;
2909 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2910 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2913 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2915 if (static_branch_unlikely(&preempt_notifier_key))
2916 __fire_sched_in_preempt_notifiers(curr);
2920 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2921 struct task_struct *next)
2923 struct preempt_notifier *notifier;
2925 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2926 notifier->ops->sched_out(notifier, next);
2929 static __always_inline void
2930 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2931 struct task_struct *next)
2933 if (static_branch_unlikely(&preempt_notifier_key))
2934 __fire_sched_out_preempt_notifiers(curr, next);
2937 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2939 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2944 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2945 struct task_struct *next)
2949 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2951 static inline void prepare_task(struct task_struct *next)
2955 * Claim the task as running, we do this before switching to it
2956 * such that any running task will have this set.
2962 static inline void finish_task(struct task_struct *prev)
2966 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2967 * We must ensure this doesn't happen until the switch is completely
2970 * In particular, the load of prev->state in finish_task_switch() must
2971 * happen before this.
2973 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2975 smp_store_release(&prev->on_cpu, 0);
2980 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2983 * Since the runqueue lock will be released by the next
2984 * task (which is an invalid locking op but in the case
2985 * of the scheduler it's an obvious special-case), so we
2986 * do an early lockdep release here:
2988 rq_unpin_lock(rq, rf);
2989 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2990 #ifdef CONFIG_DEBUG_SPINLOCK
2991 /* this is a valid case when another task releases the spinlock */
2992 rq->lock.owner = next;
2996 static inline void finish_lock_switch(struct rq *rq)
2999 * If we are tracking spinlock dependencies then we have to
3000 * fix up the runqueue lock - which gets 'carried over' from
3001 * prev into current:
3003 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3004 raw_spin_unlock_irq(&rq->lock);
3008 * NOP if the arch has not defined these:
3011 #ifndef prepare_arch_switch
3012 # define prepare_arch_switch(next) do { } while (0)
3015 #ifndef finish_arch_post_lock_switch
3016 # define finish_arch_post_lock_switch() do { } while (0)
3020 * prepare_task_switch - prepare to switch tasks
3021 * @rq: the runqueue preparing to switch
3022 * @prev: the current task that is being switched out
3023 * @next: the task we are going to switch to.
3025 * This is called with the rq lock held and interrupts off. It must
3026 * be paired with a subsequent finish_task_switch after the context
3029 * prepare_task_switch sets up locking and calls architecture specific
3033 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3034 struct task_struct *next)
3036 kcov_prepare_switch(prev);
3037 sched_info_switch(rq, prev, next);
3038 perf_event_task_sched_out(prev, next);
3040 fire_sched_out_preempt_notifiers(prev, next);
3042 prepare_arch_switch(next);
3046 * finish_task_switch - clean up after a task-switch
3047 * @prev: the thread we just switched away from.
3049 * finish_task_switch must be called after the context switch, paired
3050 * with a prepare_task_switch call before the context switch.
3051 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3052 * and do any other architecture-specific cleanup actions.
3054 * Note that we may have delayed dropping an mm in context_switch(). If
3055 * so, we finish that here outside of the runqueue lock. (Doing it
3056 * with the lock held can cause deadlocks; see schedule() for
3059 * The context switch have flipped the stack from under us and restored the
3060 * local variables which were saved when this task called schedule() in the
3061 * past. prev == current is still correct but we need to recalculate this_rq
3062 * because prev may have moved to another CPU.
3064 static struct rq *finish_task_switch(struct task_struct *prev)
3065 __releases(rq->lock)
3067 struct rq *rq = this_rq();
3068 struct mm_struct *mm = rq->prev_mm;
3072 * The previous task will have left us with a preempt_count of 2
3073 * because it left us after:
3076 * preempt_disable(); // 1
3078 * raw_spin_lock_irq(&rq->lock) // 2
3080 * Also, see FORK_PREEMPT_COUNT.
3082 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3083 "corrupted preempt_count: %s/%d/0x%x\n",
3084 current->comm, current->pid, preempt_count()))
3085 preempt_count_set(FORK_PREEMPT_COUNT);
3090 * A task struct has one reference for the use as "current".
3091 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3092 * schedule one last time. The schedule call will never return, and
3093 * the scheduled task must drop that reference.
3095 * We must observe prev->state before clearing prev->on_cpu (in
3096 * finish_task), otherwise a concurrent wakeup can get prev
3097 * running on another CPU and we could rave with its RUNNING -> DEAD
3098 * transition, resulting in a double drop.
3100 prev_state = prev->state;
3101 vtime_task_switch(prev);
3102 perf_event_task_sched_in(prev, current);
3104 finish_lock_switch(rq);
3105 finish_arch_post_lock_switch();
3106 kcov_finish_switch(current);
3108 fire_sched_in_preempt_notifiers(current);
3110 * When switching through a kernel thread, the loop in
3111 * membarrier_{private,global}_expedited() may have observed that
3112 * kernel thread and not issued an IPI. It is therefore possible to
3113 * schedule between user->kernel->user threads without passing though
3114 * switch_mm(). Membarrier requires a barrier after storing to
3115 * rq->curr, before returning to userspace, so provide them here:
3117 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3118 * provided by mmdrop(),
3119 * - a sync_core for SYNC_CORE.
3122 membarrier_mm_sync_core_before_usermode(mm);
3125 if (unlikely(prev_state == TASK_DEAD)) {
3126 if (prev->sched_class->task_dead)
3127 prev->sched_class->task_dead(prev);
3130 * Remove function-return probe instances associated with this
3131 * task and put them back on the free list.
3133 kprobe_flush_task(prev);
3135 /* Task is done with its stack. */
3136 put_task_stack(prev);
3138 put_task_struct(prev);
3141 tick_nohz_task_switch();
3147 /* rq->lock is NOT held, but preemption is disabled */
3148 static void __balance_callback(struct rq *rq)
3150 struct callback_head *head, *next;
3151 void (*func)(struct rq *rq);
3152 unsigned long flags;
3154 raw_spin_lock_irqsave(&rq->lock, flags);
3155 head = rq->balance_callback;
3156 rq->balance_callback = NULL;
3158 func = (void (*)(struct rq *))head->func;
3165 raw_spin_unlock_irqrestore(&rq->lock, flags);
3168 static inline void balance_callback(struct rq *rq)
3170 if (unlikely(rq->balance_callback))
3171 __balance_callback(rq);
3176 static inline void balance_callback(struct rq *rq)
3183 * schedule_tail - first thing a freshly forked thread must call.
3184 * @prev: the thread we just switched away from.
3186 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3187 __releases(rq->lock)
3192 * New tasks start with FORK_PREEMPT_COUNT, see there and
3193 * finish_task_switch() for details.
3195 * finish_task_switch() will drop rq->lock() and lower preempt_count
3196 * and the preempt_enable() will end up enabling preemption (on
3197 * PREEMPT_COUNT kernels).
3200 rq = finish_task_switch(prev);
3201 balance_callback(rq);
3204 if (current->set_child_tid)
3205 put_user(task_pid_vnr(current), current->set_child_tid);
3207 calculate_sigpending();
3211 * context_switch - switch to the new MM and the new thread's register state.
3213 static __always_inline struct rq *
3214 context_switch(struct rq *rq, struct task_struct *prev,
3215 struct task_struct *next, struct rq_flags *rf)
3217 struct mm_struct *mm, *oldmm;
3219 prepare_task_switch(rq, prev, next);
3222 oldmm = prev->active_mm;
3224 * For paravirt, this is coupled with an exit in switch_to to
3225 * combine the page table reload and the switch backend into
3228 arch_start_context_switch(prev);
3231 * If mm is non-NULL, we pass through switch_mm(). If mm is
3232 * NULL, we will pass through mmdrop() in finish_task_switch().
3233 * Both of these contain the full memory barrier required by
3234 * membarrier after storing to rq->curr, before returning to
3238 next->active_mm = oldmm;
3240 enter_lazy_tlb(oldmm, next);
3242 switch_mm_irqs_off(oldmm, mm, next);
3245 prev->active_mm = NULL;
3246 rq->prev_mm = oldmm;
3249 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3251 prepare_lock_switch(rq, next, rf);
3253 /* Here we just switch the register state and the stack. */
3254 switch_to(prev, next, prev);
3257 return finish_task_switch(prev);
3261 * nr_running and nr_context_switches:
3263 * externally visible scheduler statistics: current number of runnable
3264 * threads, total number of context switches performed since bootup.
3266 unsigned long nr_running(void)
3268 unsigned long i, sum = 0;
3270 for_each_online_cpu(i)
3271 sum += cpu_rq(i)->nr_running;
3277 * Check if only the current task is running on the CPU.
3279 * Caution: this function does not check that the caller has disabled
3280 * preemption, thus the result might have a time-of-check-to-time-of-use
3281 * race. The caller is responsible to use it correctly, for example:
3283 * - from a non-preemptible section (of course)
3285 * - from a thread that is bound to a single CPU
3287 * - in a loop with very short iterations (e.g. a polling loop)
3289 bool single_task_running(void)
3291 return raw_rq()->nr_running == 1;
3293 EXPORT_SYMBOL(single_task_running);
3295 unsigned long long nr_context_switches(void)
3298 unsigned long long sum = 0;
3300 for_each_possible_cpu(i)
3301 sum += cpu_rq(i)->nr_switches;
3307 * Consumers of these two interfaces, like for example the cpuidle menu
3308 * governor, are using nonsensical data. Preferring shallow idle state selection
3309 * for a CPU that has IO-wait which might not even end up running the task when
3310 * it does become runnable.
3313 unsigned long nr_iowait_cpu(int cpu)
3315 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3319 * IO-wait accounting, and how its mostly bollocks (on SMP).
3321 * The idea behind IO-wait account is to account the idle time that we could
3322 * have spend running if it were not for IO. That is, if we were to improve the
3323 * storage performance, we'd have a proportional reduction in IO-wait time.
3325 * This all works nicely on UP, where, when a task blocks on IO, we account
3326 * idle time as IO-wait, because if the storage were faster, it could've been
3327 * running and we'd not be idle.
3329 * This has been extended to SMP, by doing the same for each CPU. This however
3332 * Imagine for instance the case where two tasks block on one CPU, only the one
3333 * CPU will have IO-wait accounted, while the other has regular idle. Even
3334 * though, if the storage were faster, both could've ran at the same time,
3335 * utilising both CPUs.
3337 * This means, that when looking globally, the current IO-wait accounting on
3338 * SMP is a lower bound, by reason of under accounting.
3340 * Worse, since the numbers are provided per CPU, they are sometimes
3341 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3342 * associated with any one particular CPU, it can wake to another CPU than it
3343 * blocked on. This means the per CPU IO-wait number is meaningless.
3345 * Task CPU affinities can make all that even more 'interesting'.
3348 unsigned long nr_iowait(void)
3350 unsigned long i, sum = 0;
3352 for_each_possible_cpu(i)
3353 sum += nr_iowait_cpu(i);
3361 * sched_exec - execve() is a valuable balancing opportunity, because at
3362 * this point the task has the smallest effective memory and cache footprint.
3364 void sched_exec(void)
3366 struct task_struct *p = current;
3367 unsigned long flags;
3370 raw_spin_lock_irqsave(&p->pi_lock, flags);
3371 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3372 if (dest_cpu == smp_processor_id())
3375 if (likely(cpu_active(dest_cpu))) {
3376 struct migration_arg arg = { p, dest_cpu };
3378 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3379 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3383 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3388 DEFINE_PER_CPU(struct kernel_stat, kstat);
3389 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3391 EXPORT_PER_CPU_SYMBOL(kstat);
3392 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3395 * The function fair_sched_class.update_curr accesses the struct curr
3396 * and its field curr->exec_start; when called from task_sched_runtime(),
3397 * we observe a high rate of cache misses in practice.
3398 * Prefetching this data results in improved performance.
3400 static inline void prefetch_curr_exec_start(struct task_struct *p)
3402 #ifdef CONFIG_FAIR_GROUP_SCHED
3403 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3405 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3408 prefetch(&curr->exec_start);
3412 * Return accounted runtime for the task.
3413 * In case the task is currently running, return the runtime plus current's
3414 * pending runtime that have not been accounted yet.
3416 unsigned long long task_sched_runtime(struct task_struct *p)
3422 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3424 * 64-bit doesn't need locks to atomically read a 64-bit value.
3425 * So we have a optimization chance when the task's delta_exec is 0.
3426 * Reading ->on_cpu is racy, but this is ok.
3428 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3429 * If we race with it entering CPU, unaccounted time is 0. This is
3430 * indistinguishable from the read occurring a few cycles earlier.
3431 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3432 * been accounted, so we're correct here as well.
3434 if (!p->on_cpu || !task_on_rq_queued(p))
3435 return p->se.sum_exec_runtime;
3438 rq = task_rq_lock(p, &rf);
3440 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3441 * project cycles that may never be accounted to this
3442 * thread, breaking clock_gettime().
3444 if (task_current(rq, p) && task_on_rq_queued(p)) {
3445 prefetch_curr_exec_start(p);
3446 update_rq_clock(rq);
3447 p->sched_class->update_curr(rq);
3449 ns = p->se.sum_exec_runtime;
3450 task_rq_unlock(rq, p, &rf);
3456 * This function gets called by the timer code, with HZ frequency.
3457 * We call it with interrupts disabled.
3459 void scheduler_tick(void)
3461 int cpu = smp_processor_id();
3462 struct rq *rq = cpu_rq(cpu);
3463 struct task_struct *curr = rq->curr;
3470 update_rq_clock(rq);
3471 curr->sched_class->task_tick(rq, curr, 0);
3472 calc_global_load_tick(rq);
3477 perf_event_task_tick();
3480 rq->idle_balance = idle_cpu(cpu);
3481 trigger_load_balance(rq);
3485 #ifdef CONFIG_NO_HZ_FULL
3489 struct delayed_work work;
3492 static struct tick_work __percpu *tick_work_cpu;
3494 static void sched_tick_remote(struct work_struct *work)
3496 struct delayed_work *dwork = to_delayed_work(work);
3497 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3498 int cpu = twork->cpu;
3499 struct rq *rq = cpu_rq(cpu);
3500 struct task_struct *curr;
3505 * Handle the tick only if it appears the remote CPU is running in full
3506 * dynticks mode. The check is racy by nature, but missing a tick or
3507 * having one too much is no big deal because the scheduler tick updates
3508 * statistics and checks timeslices in a time-independent way, regardless
3509 * of when exactly it is running.
3511 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3514 rq_lock_irq(rq, &rf);
3516 if (is_idle_task(curr))
3519 update_rq_clock(rq);
3520 delta = rq_clock_task(rq) - curr->se.exec_start;
3523 * Make sure the next tick runs within a reasonable
3526 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3527 curr->sched_class->task_tick(rq, curr, 0);
3530 rq_unlock_irq(rq, &rf);
3534 * Run the remote tick once per second (1Hz). This arbitrary
3535 * frequency is large enough to avoid overload but short enough
3536 * to keep scheduler internal stats reasonably up to date.
3538 queue_delayed_work(system_unbound_wq, dwork, HZ);
3541 static void sched_tick_start(int cpu)
3543 struct tick_work *twork;
3545 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3548 WARN_ON_ONCE(!tick_work_cpu);
3550 twork = per_cpu_ptr(tick_work_cpu, cpu);
3552 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3553 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3556 #ifdef CONFIG_HOTPLUG_CPU
3557 static void sched_tick_stop(int cpu)
3559 struct tick_work *twork;
3561 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3564 WARN_ON_ONCE(!tick_work_cpu);
3566 twork = per_cpu_ptr(tick_work_cpu, cpu);
3567 cancel_delayed_work_sync(&twork->work);
3569 #endif /* CONFIG_HOTPLUG_CPU */
3571 int __init sched_tick_offload_init(void)
3573 tick_work_cpu = alloc_percpu(struct tick_work);
3574 BUG_ON(!tick_work_cpu);
3579 #else /* !CONFIG_NO_HZ_FULL */
3580 static inline void sched_tick_start(int cpu) { }
3581 static inline void sched_tick_stop(int cpu) { }
3584 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3585 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3587 * If the value passed in is equal to the current preempt count
3588 * then we just disabled preemption. Start timing the latency.
3590 static inline void preempt_latency_start(int val)
3592 if (preempt_count() == val) {
3593 unsigned long ip = get_lock_parent_ip();
3594 #ifdef CONFIG_DEBUG_PREEMPT
3595 current->preempt_disable_ip = ip;
3597 trace_preempt_off(CALLER_ADDR0, ip);
3601 void preempt_count_add(int val)
3603 #ifdef CONFIG_DEBUG_PREEMPT
3607 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3610 __preempt_count_add(val);
3611 #ifdef CONFIG_DEBUG_PREEMPT
3613 * Spinlock count overflowing soon?
3615 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3618 preempt_latency_start(val);
3620 EXPORT_SYMBOL(preempt_count_add);
3621 NOKPROBE_SYMBOL(preempt_count_add);
3624 * If the value passed in equals to the current preempt count
3625 * then we just enabled preemption. Stop timing the latency.
3627 static inline void preempt_latency_stop(int val)
3629 if (preempt_count() == val)
3630 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3633 void preempt_count_sub(int val)
3635 #ifdef CONFIG_DEBUG_PREEMPT
3639 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3642 * Is the spinlock portion underflowing?
3644 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3645 !(preempt_count() & PREEMPT_MASK)))
3649 preempt_latency_stop(val);
3650 __preempt_count_sub(val);
3652 EXPORT_SYMBOL(preempt_count_sub);
3653 NOKPROBE_SYMBOL(preempt_count_sub);
3656 static inline void preempt_latency_start(int val) { }
3657 static inline void preempt_latency_stop(int val) { }
3660 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3662 #ifdef CONFIG_DEBUG_PREEMPT
3663 return p->preempt_disable_ip;
3670 * Print scheduling while atomic bug:
3672 static noinline void __schedule_bug(struct task_struct *prev)
3674 /* Save this before calling printk(), since that will clobber it */
3675 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3677 if (oops_in_progress)
3680 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3681 prev->comm, prev->pid, preempt_count());
3683 debug_show_held_locks(prev);
3685 if (irqs_disabled())
3686 print_irqtrace_events(prev);
3687 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3688 && in_atomic_preempt_off()) {
3689 pr_err("Preemption disabled at:");
3690 print_ip_sym(preempt_disable_ip);
3694 panic("scheduling while atomic\n");
3697 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3701 * Various schedule()-time debugging checks and statistics:
3703 static inline void schedule_debug(struct task_struct *prev)
3705 #ifdef CONFIG_SCHED_STACK_END_CHECK
3706 if (task_stack_end_corrupted(prev))
3707 panic("corrupted stack end detected inside scheduler\n");
3710 if (unlikely(in_atomic_preempt_off())) {
3711 __schedule_bug(prev);
3712 preempt_count_set(PREEMPT_DISABLED);
3716 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3718 schedstat_inc(this_rq()->sched_count);
3722 * Pick up the highest-prio task:
3724 static inline struct task_struct *
3725 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3727 const struct sched_class *class;
3728 struct task_struct *p;
3731 * Optimization: we know that if all tasks are in the fair class we can
3732 * call that function directly, but only if the @prev task wasn't of a
3733 * higher scheduling class, because otherwise those loose the
3734 * opportunity to pull in more work from other CPUs.
3736 if (likely((prev->sched_class == &idle_sched_class ||
3737 prev->sched_class == &fair_sched_class) &&
3738 rq->nr_running == rq->cfs.h_nr_running)) {
3740 p = fair_sched_class.pick_next_task(rq, prev, rf);
3741 if (unlikely(p == RETRY_TASK))
3744 /* Assumes fair_sched_class->next == idle_sched_class */
3746 p = idle_sched_class.pick_next_task(rq, prev, rf);
3752 for_each_class(class) {
3753 p = class->pick_next_task(rq, prev, rf);
3755 if (unlikely(p == RETRY_TASK))
3761 /* The idle class should always have a runnable task: */
3766 * __schedule() is the main scheduler function.
3768 * The main means of driving the scheduler and thus entering this function are:
3770 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3772 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3773 * paths. For example, see arch/x86/entry_64.S.
3775 * To drive preemption between tasks, the scheduler sets the flag in timer
3776 * interrupt handler scheduler_tick().
3778 * 3. Wakeups don't really cause entry into schedule(). They add a
3779 * task to the run-queue and that's it.
3781 * Now, if the new task added to the run-queue preempts the current
3782 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3783 * called on the nearest possible occasion:
3785 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3787 * - in syscall or exception context, at the next outmost
3788 * preempt_enable(). (this might be as soon as the wake_up()'s
3791 * - in IRQ context, return from interrupt-handler to
3792 * preemptible context
3794 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3797 * - cond_resched() call
3798 * - explicit schedule() call
3799 * - return from syscall or exception to user-space
3800 * - return from interrupt-handler to user-space
3802 * WARNING: must be called with preemption disabled!
3804 static void __sched notrace __schedule(bool preempt)
3806 struct task_struct *prev, *next;
3807 unsigned long *switch_count;
3812 cpu = smp_processor_id();
3816 schedule_debug(prev);
3818 if (sched_feat(HRTICK))
3821 local_irq_disable();
3822 rcu_note_context_switch(preempt);
3825 * Make sure that signal_pending_state()->signal_pending() below
3826 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3827 * done by the caller to avoid the race with signal_wake_up().
3829 * The membarrier system call requires a full memory barrier
3830 * after coming from user-space, before storing to rq->curr.
3833 smp_mb__after_spinlock();
3835 /* Promote REQ to ACT */
3836 rq->clock_update_flags <<= 1;
3837 update_rq_clock(rq);
3839 switch_count = &prev->nivcsw;
3840 if (!preempt && prev->state) {
3841 if (signal_pending_state(prev->state, prev)) {
3842 prev->state = TASK_RUNNING;
3844 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3846 if (prev->in_iowait) {
3847 atomic_inc(&rq->nr_iowait);
3848 delayacct_blkio_start();
3851 switch_count = &prev->nvcsw;
3854 next = pick_next_task(rq, prev, &rf);
3855 clear_tsk_need_resched(prev);
3856 clear_preempt_need_resched();
3858 if (likely(prev != next)) {
3862 * The membarrier system call requires each architecture
3863 * to have a full memory barrier after updating
3864 * rq->curr, before returning to user-space.
3866 * Here are the schemes providing that barrier on the
3867 * various architectures:
3868 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3869 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3870 * - finish_lock_switch() for weakly-ordered
3871 * architectures where spin_unlock is a full barrier,
3872 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3873 * is a RELEASE barrier),
3877 trace_sched_switch(preempt, prev, next);
3879 /* Also unlocks the rq: */
3880 rq = context_switch(rq, prev, next, &rf);
3882 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3883 rq_unlock_irq(rq, &rf);
3886 balance_callback(rq);
3889 void __noreturn do_task_dead(void)
3891 /* Causes final put_task_struct in finish_task_switch(): */
3892 set_special_state(TASK_DEAD);
3894 /* Tell freezer to ignore us: */
3895 current->flags |= PF_NOFREEZE;
3900 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3905 static inline void sched_submit_work(struct task_struct *tsk)
3911 * If a worker went to sleep, notify and ask workqueue whether
3912 * it wants to wake up a task to maintain concurrency.
3913 * As this function is called inside the schedule() context,
3914 * we disable preemption to avoid it calling schedule() again
3915 * in the possible wakeup of a kworker.
3917 if (tsk->flags & PF_WQ_WORKER) {
3919 wq_worker_sleeping(tsk);
3920 preempt_enable_no_resched();
3923 if (tsk_is_pi_blocked(tsk))
3927 * If we are going to sleep and we have plugged IO queued,
3928 * make sure to submit it to avoid deadlocks.
3930 if (blk_needs_flush_plug(tsk))
3931 blk_schedule_flush_plug(tsk);
3934 static void sched_update_worker(struct task_struct *tsk)
3936 if (tsk->flags & PF_WQ_WORKER)
3937 wq_worker_running(tsk);
3940 asmlinkage __visible void __sched schedule(void)
3942 struct task_struct *tsk = current;
3944 sched_submit_work(tsk);
3948 sched_preempt_enable_no_resched();
3949 } while (need_resched());
3950 sched_update_worker(tsk);
3952 EXPORT_SYMBOL(schedule);
3955 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3956 * state (have scheduled out non-voluntarily) by making sure that all
3957 * tasks have either left the run queue or have gone into user space.
3958 * As idle tasks do not do either, they must not ever be preempted
3959 * (schedule out non-voluntarily).
3961 * schedule_idle() is similar to schedule_preempt_disable() except that it
3962 * never enables preemption because it does not call sched_submit_work().
3964 void __sched schedule_idle(void)
3967 * As this skips calling sched_submit_work(), which the idle task does
3968 * regardless because that function is a nop when the task is in a
3969 * TASK_RUNNING state, make sure this isn't used someplace that the
3970 * current task can be in any other state. Note, idle is always in the
3971 * TASK_RUNNING state.
3973 WARN_ON_ONCE(current->state);
3976 } while (need_resched());
3979 #ifdef CONFIG_CONTEXT_TRACKING
3980 asmlinkage __visible void __sched schedule_user(void)
3983 * If we come here after a random call to set_need_resched(),
3984 * or we have been woken up remotely but the IPI has not yet arrived,
3985 * we haven't yet exited the RCU idle mode. Do it here manually until
3986 * we find a better solution.
3988 * NB: There are buggy callers of this function. Ideally we
3989 * should warn if prev_state != CONTEXT_USER, but that will trigger
3990 * too frequently to make sense yet.
3992 enum ctx_state prev_state = exception_enter();
3994 exception_exit(prev_state);
3999 * schedule_preempt_disabled - called with preemption disabled
4001 * Returns with preemption disabled. Note: preempt_count must be 1
4003 void __sched schedule_preempt_disabled(void)
4005 sched_preempt_enable_no_resched();
4010 static void __sched notrace preempt_schedule_common(void)
4014 * Because the function tracer can trace preempt_count_sub()
4015 * and it also uses preempt_enable/disable_notrace(), if
4016 * NEED_RESCHED is set, the preempt_enable_notrace() called
4017 * by the function tracer will call this function again and
4018 * cause infinite recursion.
4020 * Preemption must be disabled here before the function
4021 * tracer can trace. Break up preempt_disable() into two
4022 * calls. One to disable preemption without fear of being
4023 * traced. The other to still record the preemption latency,
4024 * which can also be traced by the function tracer.
4026 preempt_disable_notrace();
4027 preempt_latency_start(1);
4029 preempt_latency_stop(1);
4030 preempt_enable_no_resched_notrace();
4033 * Check again in case we missed a preemption opportunity
4034 * between schedule and now.
4036 } while (need_resched());
4039 #ifdef CONFIG_PREEMPT
4041 * this is the entry point to schedule() from in-kernel preemption
4042 * off of preempt_enable. Kernel preemptions off return from interrupt
4043 * occur there and call schedule directly.
4045 asmlinkage __visible void __sched notrace preempt_schedule(void)
4048 * If there is a non-zero preempt_count or interrupts are disabled,
4049 * we do not want to preempt the current task. Just return..
4051 if (likely(!preemptible()))
4054 preempt_schedule_common();
4056 NOKPROBE_SYMBOL(preempt_schedule);
4057 EXPORT_SYMBOL(preempt_schedule);
4060 * preempt_schedule_notrace - preempt_schedule called by tracing
4062 * The tracing infrastructure uses preempt_enable_notrace to prevent
4063 * recursion and tracing preempt enabling caused by the tracing
4064 * infrastructure itself. But as tracing can happen in areas coming
4065 * from userspace or just about to enter userspace, a preempt enable
4066 * can occur before user_exit() is called. This will cause the scheduler
4067 * to be called when the system is still in usermode.
4069 * To prevent this, the preempt_enable_notrace will use this function
4070 * instead of preempt_schedule() to exit user context if needed before
4071 * calling the scheduler.
4073 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4075 enum ctx_state prev_ctx;
4077 if (likely(!preemptible()))
4082 * Because the function tracer can trace preempt_count_sub()
4083 * and it also uses preempt_enable/disable_notrace(), if
4084 * NEED_RESCHED is set, the preempt_enable_notrace() called
4085 * by the function tracer will call this function again and
4086 * cause infinite recursion.
4088 * Preemption must be disabled here before the function
4089 * tracer can trace. Break up preempt_disable() into two
4090 * calls. One to disable preemption without fear of being
4091 * traced. The other to still record the preemption latency,
4092 * which can also be traced by the function tracer.
4094 preempt_disable_notrace();
4095 preempt_latency_start(1);
4097 * Needs preempt disabled in case user_exit() is traced
4098 * and the tracer calls preempt_enable_notrace() causing
4099 * an infinite recursion.
4101 prev_ctx = exception_enter();
4103 exception_exit(prev_ctx);
4105 preempt_latency_stop(1);
4106 preempt_enable_no_resched_notrace();
4107 } while (need_resched());
4109 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4111 #endif /* CONFIG_PREEMPT */
4114 * this is the entry point to schedule() from kernel preemption
4115 * off of irq context.
4116 * Note, that this is called and return with irqs disabled. This will
4117 * protect us against recursive calling from irq.
4119 asmlinkage __visible void __sched preempt_schedule_irq(void)
4121 enum ctx_state prev_state;
4123 /* Catch callers which need to be fixed */
4124 BUG_ON(preempt_count() || !irqs_disabled());
4126 prev_state = exception_enter();
4132 local_irq_disable();
4133 sched_preempt_enable_no_resched();
4134 } while (need_resched());
4136 exception_exit(prev_state);
4139 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4142 return try_to_wake_up(curr->private, mode, wake_flags);
4144 EXPORT_SYMBOL(default_wake_function);
4146 #ifdef CONFIG_RT_MUTEXES
4148 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4151 prio = min(prio, pi_task->prio);
4156 static inline int rt_effective_prio(struct task_struct *p, int prio)
4158 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4160 return __rt_effective_prio(pi_task, prio);
4164 * rt_mutex_setprio - set the current priority of a task
4166 * @pi_task: donor task
4168 * This function changes the 'effective' priority of a task. It does
4169 * not touch ->normal_prio like __setscheduler().
4171 * Used by the rt_mutex code to implement priority inheritance
4172 * logic. Call site only calls if the priority of the task changed.
4174 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4176 int prio, oldprio, queued, running, queue_flag =
4177 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4178 const struct sched_class *prev_class;
4182 /* XXX used to be waiter->prio, not waiter->task->prio */
4183 prio = __rt_effective_prio(pi_task, p->normal_prio);
4186 * If nothing changed; bail early.
4188 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4191 rq = __task_rq_lock(p, &rf);
4192 update_rq_clock(rq);
4194 * Set under pi_lock && rq->lock, such that the value can be used under
4197 * Note that there is loads of tricky to make this pointer cache work
4198 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4199 * ensure a task is de-boosted (pi_task is set to NULL) before the
4200 * task is allowed to run again (and can exit). This ensures the pointer
4201 * points to a blocked task -- which guaratees the task is present.
4203 p->pi_top_task = pi_task;
4206 * For FIFO/RR we only need to set prio, if that matches we're done.
4208 if (prio == p->prio && !dl_prio(prio))
4212 * Idle task boosting is a nono in general. There is one
4213 * exception, when PREEMPT_RT and NOHZ is active:
4215 * The idle task calls get_next_timer_interrupt() and holds
4216 * the timer wheel base->lock on the CPU and another CPU wants
4217 * to access the timer (probably to cancel it). We can safely
4218 * ignore the boosting request, as the idle CPU runs this code
4219 * with interrupts disabled and will complete the lock
4220 * protected section without being interrupted. So there is no
4221 * real need to boost.
4223 if (unlikely(p == rq->idle)) {
4224 WARN_ON(p != rq->curr);
4225 WARN_ON(p->pi_blocked_on);
4229 trace_sched_pi_setprio(p, pi_task);
4232 if (oldprio == prio)
4233 queue_flag &= ~DEQUEUE_MOVE;
4235 prev_class = p->sched_class;
4236 queued = task_on_rq_queued(p);
4237 running = task_current(rq, p);
4239 dequeue_task(rq, p, queue_flag);
4241 put_prev_task(rq, p);
4244 * Boosting condition are:
4245 * 1. -rt task is running and holds mutex A
4246 * --> -dl task blocks on mutex A
4248 * 2. -dl task is running and holds mutex A
4249 * --> -dl task blocks on mutex A and could preempt the
4252 if (dl_prio(prio)) {
4253 if (!dl_prio(p->normal_prio) ||
4254 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4255 p->dl.dl_boosted = 1;
4256 queue_flag |= ENQUEUE_REPLENISH;
4258 p->dl.dl_boosted = 0;
4259 p->sched_class = &dl_sched_class;
4260 } else if (rt_prio(prio)) {
4261 if (dl_prio(oldprio))
4262 p->dl.dl_boosted = 0;
4264 queue_flag |= ENQUEUE_HEAD;
4265 p->sched_class = &rt_sched_class;
4267 if (dl_prio(oldprio))
4268 p->dl.dl_boosted = 0;
4269 if (rt_prio(oldprio))
4271 p->sched_class = &fair_sched_class;
4277 enqueue_task(rq, p, queue_flag);
4279 set_curr_task(rq, p);
4281 check_class_changed(rq, p, prev_class, oldprio);
4283 /* Avoid rq from going away on us: */
4285 __task_rq_unlock(rq, &rf);
4287 balance_callback(rq);
4291 static inline int rt_effective_prio(struct task_struct *p, int prio)
4297 void set_user_nice(struct task_struct *p, long nice)
4299 bool queued, running;
4300 int old_prio, delta;
4304 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4307 * We have to be careful, if called from sys_setpriority(),
4308 * the task might be in the middle of scheduling on another CPU.
4310 rq = task_rq_lock(p, &rf);
4311 update_rq_clock(rq);
4314 * The RT priorities are set via sched_setscheduler(), but we still
4315 * allow the 'normal' nice value to be set - but as expected
4316 * it wont have any effect on scheduling until the task is
4317 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4319 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4320 p->static_prio = NICE_TO_PRIO(nice);
4323 queued = task_on_rq_queued(p);
4324 running = task_current(rq, p);
4326 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4328 put_prev_task(rq, p);
4330 p->static_prio = NICE_TO_PRIO(nice);
4331 set_load_weight(p, true);
4333 p->prio = effective_prio(p);
4334 delta = p->prio - old_prio;
4337 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4339 * If the task increased its priority or is running and
4340 * lowered its priority, then reschedule its CPU:
4342 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4346 set_curr_task(rq, p);
4348 task_rq_unlock(rq, p, &rf);
4350 EXPORT_SYMBOL(set_user_nice);
4353 * can_nice - check if a task can reduce its nice value
4357 int can_nice(const struct task_struct *p, const int nice)
4359 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4360 int nice_rlim = nice_to_rlimit(nice);
4362 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4363 capable(CAP_SYS_NICE));
4366 #ifdef __ARCH_WANT_SYS_NICE
4369 * sys_nice - change the priority of the current process.
4370 * @increment: priority increment
4372 * sys_setpriority is a more generic, but much slower function that
4373 * does similar things.
4375 SYSCALL_DEFINE1(nice, int, increment)
4380 * Setpriority might change our priority at the same moment.
4381 * We don't have to worry. Conceptually one call occurs first
4382 * and we have a single winner.
4384 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4385 nice = task_nice(current) + increment;
4387 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4388 if (increment < 0 && !can_nice(current, nice))
4391 retval = security_task_setnice(current, nice);
4395 set_user_nice(current, nice);
4402 * task_prio - return the priority value of a given task.
4403 * @p: the task in question.
4405 * Return: The priority value as seen by users in /proc.
4406 * RT tasks are offset by -200. Normal tasks are centered
4407 * around 0, value goes from -16 to +15.
4409 int task_prio(const struct task_struct *p)
4411 return p->prio - MAX_RT_PRIO;
4415 * idle_cpu - is a given CPU idle currently?
4416 * @cpu: the processor in question.
4418 * Return: 1 if the CPU is currently idle. 0 otherwise.
4420 int idle_cpu(int cpu)
4422 struct rq *rq = cpu_rq(cpu);
4424 if (rq->curr != rq->idle)
4431 if (!llist_empty(&rq->wake_list))
4439 * available_idle_cpu - is a given CPU idle for enqueuing work.
4440 * @cpu: the CPU in question.
4442 * Return: 1 if the CPU is currently idle. 0 otherwise.
4444 int available_idle_cpu(int cpu)
4449 if (vcpu_is_preempted(cpu))
4456 * idle_task - return the idle task for a given CPU.
4457 * @cpu: the processor in question.
4459 * Return: The idle task for the CPU @cpu.
4461 struct task_struct *idle_task(int cpu)
4463 return cpu_rq(cpu)->idle;
4467 * find_process_by_pid - find a process with a matching PID value.
4468 * @pid: the pid in question.
4470 * The task of @pid, if found. %NULL otherwise.
4472 static struct task_struct *find_process_by_pid(pid_t pid)
4474 return pid ? find_task_by_vpid(pid) : current;
4478 * sched_setparam() passes in -1 for its policy, to let the functions
4479 * it calls know not to change it.
4481 #define SETPARAM_POLICY -1
4483 static void __setscheduler_params(struct task_struct *p,
4484 const struct sched_attr *attr)
4486 int policy = attr->sched_policy;
4488 if (policy == SETPARAM_POLICY)
4493 if (dl_policy(policy))
4494 __setparam_dl(p, attr);
4495 else if (fair_policy(policy))
4496 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4499 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4500 * !rt_policy. Always setting this ensures that things like
4501 * getparam()/getattr() don't report silly values for !rt tasks.
4503 p->rt_priority = attr->sched_priority;
4504 p->normal_prio = normal_prio(p);
4505 set_load_weight(p, true);
4508 /* Actually do priority change: must hold pi & rq lock. */
4509 static void __setscheduler(struct rq *rq, struct task_struct *p,
4510 const struct sched_attr *attr, bool keep_boost)
4513 * If params can't change scheduling class changes aren't allowed
4516 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4519 __setscheduler_params(p, attr);
4522 * Keep a potential priority boosting if called from
4523 * sched_setscheduler().
4525 p->prio = normal_prio(p);
4527 p->prio = rt_effective_prio(p, p->prio);
4529 if (dl_prio(p->prio))
4530 p->sched_class = &dl_sched_class;
4531 else if (rt_prio(p->prio))
4532 p->sched_class = &rt_sched_class;
4534 p->sched_class = &fair_sched_class;
4538 * Check the target process has a UID that matches the current process's:
4540 static bool check_same_owner(struct task_struct *p)
4542 const struct cred *cred = current_cred(), *pcred;
4546 pcred = __task_cred(p);
4547 match = (uid_eq(cred->euid, pcred->euid) ||
4548 uid_eq(cred->euid, pcred->uid));
4553 static int __sched_setscheduler(struct task_struct *p,
4554 const struct sched_attr *attr,
4557 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4558 MAX_RT_PRIO - 1 - attr->sched_priority;
4559 int retval, oldprio, oldpolicy = -1, queued, running;
4560 int new_effective_prio, policy = attr->sched_policy;
4561 const struct sched_class *prev_class;
4564 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4567 /* The pi code expects interrupts enabled */
4568 BUG_ON(pi && in_interrupt());
4570 /* Double check policy once rq lock held: */
4572 reset_on_fork = p->sched_reset_on_fork;
4573 policy = oldpolicy = p->policy;
4575 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4577 if (!valid_policy(policy))
4581 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4585 * Valid priorities for SCHED_FIFO and SCHED_RR are
4586 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4587 * SCHED_BATCH and SCHED_IDLE is 0.
4589 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4590 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4592 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4593 (rt_policy(policy) != (attr->sched_priority != 0)))
4597 * Allow unprivileged RT tasks to decrease priority:
4599 if (user && !capable(CAP_SYS_NICE)) {
4600 if (fair_policy(policy)) {
4601 if (attr->sched_nice < task_nice(p) &&
4602 !can_nice(p, attr->sched_nice))
4606 if (rt_policy(policy)) {
4607 unsigned long rlim_rtprio =
4608 task_rlimit(p, RLIMIT_RTPRIO);
4610 /* Can't set/change the rt policy: */
4611 if (policy != p->policy && !rlim_rtprio)
4614 /* Can't increase priority: */
4615 if (attr->sched_priority > p->rt_priority &&
4616 attr->sched_priority > rlim_rtprio)
4621 * Can't set/change SCHED_DEADLINE policy at all for now
4622 * (safest behavior); in the future we would like to allow
4623 * unprivileged DL tasks to increase their relative deadline
4624 * or reduce their runtime (both ways reducing utilization)
4626 if (dl_policy(policy))
4630 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4631 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4633 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4634 if (!can_nice(p, task_nice(p)))
4638 /* Can't change other user's priorities: */
4639 if (!check_same_owner(p))
4642 /* Normal users shall not reset the sched_reset_on_fork flag: */
4643 if (p->sched_reset_on_fork && !reset_on_fork)
4648 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4651 retval = security_task_setscheduler(p);
4656 /* Update task specific "requested" clamps */
4657 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4658 retval = uclamp_validate(p, attr);
4664 * Make sure no PI-waiters arrive (or leave) while we are
4665 * changing the priority of the task:
4667 * To be able to change p->policy safely, the appropriate
4668 * runqueue lock must be held.
4670 rq = task_rq_lock(p, &rf);
4671 update_rq_clock(rq);
4674 * Changing the policy of the stop threads its a very bad idea:
4676 if (p == rq->stop) {
4677 task_rq_unlock(rq, p, &rf);
4682 * If not changing anything there's no need to proceed further,
4683 * but store a possible modification of reset_on_fork.
4685 if (unlikely(policy == p->policy)) {
4686 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4688 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4690 if (dl_policy(policy) && dl_param_changed(p, attr))
4692 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4695 p->sched_reset_on_fork = reset_on_fork;
4696 task_rq_unlock(rq, p, &rf);
4702 #ifdef CONFIG_RT_GROUP_SCHED
4704 * Do not allow realtime tasks into groups that have no runtime
4707 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4708 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4709 !task_group_is_autogroup(task_group(p))) {
4710 task_rq_unlock(rq, p, &rf);
4715 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4716 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4717 cpumask_t *span = rq->rd->span;
4720 * Don't allow tasks with an affinity mask smaller than
4721 * the entire root_domain to become SCHED_DEADLINE. We
4722 * will also fail if there's no bandwidth available.
4724 if (!cpumask_subset(span, p->cpus_ptr) ||
4725 rq->rd->dl_bw.bw == 0) {
4726 task_rq_unlock(rq, p, &rf);
4733 /* Re-check policy now with rq lock held: */
4734 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4735 policy = oldpolicy = -1;
4736 task_rq_unlock(rq, p, &rf);
4741 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4742 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4745 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4746 task_rq_unlock(rq, p, &rf);
4750 p->sched_reset_on_fork = reset_on_fork;
4755 * Take priority boosted tasks into account. If the new
4756 * effective priority is unchanged, we just store the new
4757 * normal parameters and do not touch the scheduler class and
4758 * the runqueue. This will be done when the task deboost
4761 new_effective_prio = rt_effective_prio(p, newprio);
4762 if (new_effective_prio == oldprio)
4763 queue_flags &= ~DEQUEUE_MOVE;
4766 queued = task_on_rq_queued(p);
4767 running = task_current(rq, p);
4769 dequeue_task(rq, p, queue_flags);
4771 put_prev_task(rq, p);
4773 prev_class = p->sched_class;
4775 __setscheduler(rq, p, attr, pi);
4776 __setscheduler_uclamp(p, attr);
4780 * We enqueue to tail when the priority of a task is
4781 * increased (user space view).
4783 if (oldprio < p->prio)
4784 queue_flags |= ENQUEUE_HEAD;
4786 enqueue_task(rq, p, queue_flags);
4789 set_curr_task(rq, p);
4791 check_class_changed(rq, p, prev_class, oldprio);
4793 /* Avoid rq from going away on us: */
4795 task_rq_unlock(rq, p, &rf);
4798 rt_mutex_adjust_pi(p);
4800 /* Run balance callbacks after we've adjusted the PI chain: */
4801 balance_callback(rq);
4807 static int _sched_setscheduler(struct task_struct *p, int policy,
4808 const struct sched_param *param, bool check)
4810 struct sched_attr attr = {
4811 .sched_policy = policy,
4812 .sched_priority = param->sched_priority,
4813 .sched_nice = PRIO_TO_NICE(p->static_prio),
4816 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4817 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4818 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4819 policy &= ~SCHED_RESET_ON_FORK;
4820 attr.sched_policy = policy;
4823 return __sched_setscheduler(p, &attr, check, true);
4826 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4827 * @p: the task in question.
4828 * @policy: new policy.
4829 * @param: structure containing the new RT priority.
4831 * Return: 0 on success. An error code otherwise.
4833 * NOTE that the task may be already dead.
4835 int sched_setscheduler(struct task_struct *p, int policy,
4836 const struct sched_param *param)
4838 return _sched_setscheduler(p, policy, param, true);
4840 EXPORT_SYMBOL_GPL(sched_setscheduler);
4842 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4844 return __sched_setscheduler(p, attr, true, true);
4846 EXPORT_SYMBOL_GPL(sched_setattr);
4848 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4850 return __sched_setscheduler(p, attr, false, true);
4854 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4855 * @p: the task in question.
4856 * @policy: new policy.
4857 * @param: structure containing the new RT priority.
4859 * Just like sched_setscheduler, only don't bother checking if the
4860 * current context has permission. For example, this is needed in
4861 * stop_machine(): we create temporary high priority worker threads,
4862 * but our caller might not have that capability.
4864 * Return: 0 on success. An error code otherwise.
4866 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4867 const struct sched_param *param)
4869 return _sched_setscheduler(p, policy, param, false);
4871 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4874 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4876 struct sched_param lparam;
4877 struct task_struct *p;
4880 if (!param || pid < 0)
4882 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4887 p = find_process_by_pid(pid);
4889 retval = sched_setscheduler(p, policy, &lparam);
4896 * Mimics kernel/events/core.c perf_copy_attr().
4898 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4903 if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
4906 /* Zero the full structure, so that a short copy will be nice: */
4907 memset(attr, 0, sizeof(*attr));
4909 ret = get_user(size, &uattr->size);
4913 /* Bail out on silly large: */
4914 if (size > PAGE_SIZE)
4917 /* ABI compatibility quirk: */
4919 size = SCHED_ATTR_SIZE_VER0;
4921 if (size < SCHED_ATTR_SIZE_VER0)
4925 * If we're handed a bigger struct than we know of,
4926 * ensure all the unknown bits are 0 - i.e. new
4927 * user-space does not rely on any kernel feature
4928 * extensions we dont know about yet.
4930 if (size > sizeof(*attr)) {
4931 unsigned char __user *addr;
4932 unsigned char __user *end;
4935 addr = (void __user *)uattr + sizeof(*attr);
4936 end = (void __user *)uattr + size;
4938 for (; addr < end; addr++) {
4939 ret = get_user(val, addr);
4945 size = sizeof(*attr);
4948 ret = copy_from_user(attr, uattr, size);
4952 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
4953 size < SCHED_ATTR_SIZE_VER1)
4957 * XXX: Do we want to be lenient like existing syscalls; or do we want
4958 * to be strict and return an error on out-of-bounds values?
4960 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4965 put_user(sizeof(*attr), &uattr->size);
4970 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4971 * @pid: the pid in question.
4972 * @policy: new policy.
4973 * @param: structure containing the new RT priority.
4975 * Return: 0 on success. An error code otherwise.
4977 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4982 return do_sched_setscheduler(pid, policy, param);
4986 * sys_sched_setparam - set/change the RT priority of a thread
4987 * @pid: the pid in question.
4988 * @param: structure containing the new RT priority.
4990 * Return: 0 on success. An error code otherwise.
4992 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4994 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4998 * sys_sched_setattr - same as above, but with extended sched_attr
4999 * @pid: the pid in question.
5000 * @uattr: structure containing the extended parameters.
5001 * @flags: for future extension.
5003 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5004 unsigned int, flags)
5006 struct sched_attr attr;
5007 struct task_struct *p;
5010 if (!uattr || pid < 0 || flags)
5013 retval = sched_copy_attr(uattr, &attr);
5017 if ((int)attr.sched_policy < 0)
5019 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5020 attr.sched_policy = SETPARAM_POLICY;
5024 p = find_process_by_pid(pid);
5030 retval = sched_setattr(p, &attr);
5038 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5039 * @pid: the pid in question.
5041 * Return: On success, the policy of the thread. Otherwise, a negative error
5044 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5046 struct task_struct *p;
5054 p = find_process_by_pid(pid);
5056 retval = security_task_getscheduler(p);
5059 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5066 * sys_sched_getparam - get the RT priority of a thread
5067 * @pid: the pid in question.
5068 * @param: structure containing the RT priority.
5070 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5073 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5075 struct sched_param lp = { .sched_priority = 0 };
5076 struct task_struct *p;
5079 if (!param || pid < 0)
5083 p = find_process_by_pid(pid);
5088 retval = security_task_getscheduler(p);
5092 if (task_has_rt_policy(p))
5093 lp.sched_priority = p->rt_priority;
5097 * This one might sleep, we cannot do it with a spinlock held ...
5099 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5108 static int sched_read_attr(struct sched_attr __user *uattr,
5109 struct sched_attr *attr,
5114 if (!access_ok(uattr, usize))
5118 * If we're handed a smaller struct than we know of,
5119 * ensure all the unknown bits are 0 - i.e. old
5120 * user-space does not get uncomplete information.
5122 if (usize < sizeof(*attr)) {
5123 unsigned char *addr;
5126 addr = (void *)attr + usize;
5127 end = (void *)attr + sizeof(*attr);
5129 for (; addr < end; addr++) {
5137 ret = copy_to_user(uattr, attr, attr->size);
5145 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5146 * @pid: the pid in question.
5147 * @uattr: structure containing the extended parameters.
5148 * @size: sizeof(attr) for fwd/bwd comp.
5149 * @flags: for future extension.
5151 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5152 unsigned int, size, unsigned int, flags)
5154 struct sched_attr attr = {
5155 .size = sizeof(struct sched_attr),
5157 struct task_struct *p;
5160 if (!uattr || pid < 0 || size > PAGE_SIZE ||
5161 size < SCHED_ATTR_SIZE_VER0 || flags)
5165 p = find_process_by_pid(pid);
5170 retval = security_task_getscheduler(p);
5174 attr.sched_policy = p->policy;
5175 if (p->sched_reset_on_fork)
5176 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5177 if (task_has_dl_policy(p))
5178 __getparam_dl(p, &attr);
5179 else if (task_has_rt_policy(p))
5180 attr.sched_priority = p->rt_priority;
5182 attr.sched_nice = task_nice(p);
5184 #ifdef CONFIG_UCLAMP_TASK
5185 attr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5186 attr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5191 retval = sched_read_attr(uattr, &attr, size);
5199 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5201 cpumask_var_t cpus_allowed, new_mask;
5202 struct task_struct *p;
5207 p = find_process_by_pid(pid);
5213 /* Prevent p going away */
5217 if (p->flags & PF_NO_SETAFFINITY) {
5221 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5225 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5227 goto out_free_cpus_allowed;
5230 if (!check_same_owner(p)) {
5232 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5234 goto out_free_new_mask;
5239 retval = security_task_setscheduler(p);
5241 goto out_free_new_mask;
5244 cpuset_cpus_allowed(p, cpus_allowed);
5245 cpumask_and(new_mask, in_mask, cpus_allowed);
5248 * Since bandwidth control happens on root_domain basis,
5249 * if admission test is enabled, we only admit -deadline
5250 * tasks allowed to run on all the CPUs in the task's
5254 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5256 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5259 goto out_free_new_mask;
5265 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5268 cpuset_cpus_allowed(p, cpus_allowed);
5269 if (!cpumask_subset(new_mask, cpus_allowed)) {
5271 * We must have raced with a concurrent cpuset
5272 * update. Just reset the cpus_allowed to the
5273 * cpuset's cpus_allowed
5275 cpumask_copy(new_mask, cpus_allowed);
5280 free_cpumask_var(new_mask);
5281 out_free_cpus_allowed:
5282 free_cpumask_var(cpus_allowed);
5288 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5289 struct cpumask *new_mask)
5291 if (len < cpumask_size())
5292 cpumask_clear(new_mask);
5293 else if (len > cpumask_size())
5294 len = cpumask_size();
5296 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5300 * sys_sched_setaffinity - set the CPU affinity of a process
5301 * @pid: pid of the process
5302 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5303 * @user_mask_ptr: user-space pointer to the new CPU mask
5305 * Return: 0 on success. An error code otherwise.
5307 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5308 unsigned long __user *, user_mask_ptr)
5310 cpumask_var_t new_mask;
5313 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5316 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5318 retval = sched_setaffinity(pid, new_mask);
5319 free_cpumask_var(new_mask);
5323 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5325 struct task_struct *p;
5326 unsigned long flags;
5332 p = find_process_by_pid(pid);
5336 retval = security_task_getscheduler(p);
5340 raw_spin_lock_irqsave(&p->pi_lock, flags);
5341 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5342 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5351 * sys_sched_getaffinity - get the CPU affinity of a process
5352 * @pid: pid of the process
5353 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5354 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5356 * Return: size of CPU mask copied to user_mask_ptr on success. An
5357 * error code otherwise.
5359 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5360 unsigned long __user *, user_mask_ptr)
5365 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5367 if (len & (sizeof(unsigned long)-1))
5370 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5373 ret = sched_getaffinity(pid, mask);
5375 unsigned int retlen = min(len, cpumask_size());
5377 if (copy_to_user(user_mask_ptr, mask, retlen))
5382 free_cpumask_var(mask);
5388 * sys_sched_yield - yield the current processor to other threads.
5390 * This function yields the current CPU to other tasks. If there are no
5391 * other threads running on this CPU then this function will return.
5395 static void do_sched_yield(void)
5400 rq = this_rq_lock_irq(&rf);
5402 schedstat_inc(rq->yld_count);
5403 current->sched_class->yield_task(rq);
5406 * Since we are going to call schedule() anyway, there's
5407 * no need to preempt or enable interrupts:
5411 sched_preempt_enable_no_resched();
5416 SYSCALL_DEFINE0(sched_yield)
5422 #ifndef CONFIG_PREEMPT
5423 int __sched _cond_resched(void)
5425 if (should_resched(0)) {
5426 preempt_schedule_common();
5432 EXPORT_SYMBOL(_cond_resched);
5436 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5437 * call schedule, and on return reacquire the lock.
5439 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5440 * operations here to prevent schedule() from being called twice (once via
5441 * spin_unlock(), once by hand).
5443 int __cond_resched_lock(spinlock_t *lock)
5445 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5448 lockdep_assert_held(lock);
5450 if (spin_needbreak(lock) || resched) {
5453 preempt_schedule_common();
5461 EXPORT_SYMBOL(__cond_resched_lock);
5464 * yield - yield the current processor to other threads.
5466 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5468 * The scheduler is at all times free to pick the calling task as the most
5469 * eligible task to run, if removing the yield() call from your code breaks
5470 * it, its already broken.
5472 * Typical broken usage is:
5477 * where one assumes that yield() will let 'the other' process run that will
5478 * make event true. If the current task is a SCHED_FIFO task that will never
5479 * happen. Never use yield() as a progress guarantee!!
5481 * If you want to use yield() to wait for something, use wait_event().
5482 * If you want to use yield() to be 'nice' for others, use cond_resched().
5483 * If you still want to use yield(), do not!
5485 void __sched yield(void)
5487 set_current_state(TASK_RUNNING);
5490 EXPORT_SYMBOL(yield);
5493 * yield_to - yield the current processor to another thread in
5494 * your thread group, or accelerate that thread toward the
5495 * processor it's on.
5497 * @preempt: whether task preemption is allowed or not
5499 * It's the caller's job to ensure that the target task struct
5500 * can't go away on us before we can do any checks.
5503 * true (>0) if we indeed boosted the target task.
5504 * false (0) if we failed to boost the target.
5505 * -ESRCH if there's no task to yield to.
5507 int __sched yield_to(struct task_struct *p, bool preempt)
5509 struct task_struct *curr = current;
5510 struct rq *rq, *p_rq;
5511 unsigned long flags;
5514 local_irq_save(flags);
5520 * If we're the only runnable task on the rq and target rq also
5521 * has only one task, there's absolutely no point in yielding.
5523 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5528 double_rq_lock(rq, p_rq);
5529 if (task_rq(p) != p_rq) {
5530 double_rq_unlock(rq, p_rq);
5534 if (!curr->sched_class->yield_to_task)
5537 if (curr->sched_class != p->sched_class)
5540 if (task_running(p_rq, p) || p->state)
5543 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5545 schedstat_inc(rq->yld_count);
5547 * Make p's CPU reschedule; pick_next_entity takes care of
5550 if (preempt && rq != p_rq)
5555 double_rq_unlock(rq, p_rq);
5557 local_irq_restore(flags);
5564 EXPORT_SYMBOL_GPL(yield_to);
5566 int io_schedule_prepare(void)
5568 int old_iowait = current->in_iowait;
5570 current->in_iowait = 1;
5571 blk_schedule_flush_plug(current);
5576 void io_schedule_finish(int token)
5578 current->in_iowait = token;
5582 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5583 * that process accounting knows that this is a task in IO wait state.
5585 long __sched io_schedule_timeout(long timeout)
5590 token = io_schedule_prepare();
5591 ret = schedule_timeout(timeout);
5592 io_schedule_finish(token);
5596 EXPORT_SYMBOL(io_schedule_timeout);
5598 void __sched io_schedule(void)
5602 token = io_schedule_prepare();
5604 io_schedule_finish(token);
5606 EXPORT_SYMBOL(io_schedule);
5609 * sys_sched_get_priority_max - return maximum RT priority.
5610 * @policy: scheduling class.
5612 * Return: On success, this syscall returns the maximum
5613 * rt_priority that can be used by a given scheduling class.
5614 * On failure, a negative error code is returned.
5616 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5623 ret = MAX_USER_RT_PRIO-1;
5625 case SCHED_DEADLINE:
5636 * sys_sched_get_priority_min - return minimum RT priority.
5637 * @policy: scheduling class.
5639 * Return: On success, this syscall returns the minimum
5640 * rt_priority that can be used by a given scheduling class.
5641 * On failure, a negative error code is returned.
5643 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5652 case SCHED_DEADLINE:
5661 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5663 struct task_struct *p;
5664 unsigned int time_slice;
5674 p = find_process_by_pid(pid);
5678 retval = security_task_getscheduler(p);
5682 rq = task_rq_lock(p, &rf);
5684 if (p->sched_class->get_rr_interval)
5685 time_slice = p->sched_class->get_rr_interval(rq, p);
5686 task_rq_unlock(rq, p, &rf);
5689 jiffies_to_timespec64(time_slice, t);
5698 * sys_sched_rr_get_interval - return the default timeslice of a process.
5699 * @pid: pid of the process.
5700 * @interval: userspace pointer to the timeslice value.
5702 * this syscall writes the default timeslice value of a given process
5703 * into the user-space timespec buffer. A value of '0' means infinity.
5705 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5708 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5709 struct __kernel_timespec __user *, interval)
5711 struct timespec64 t;
5712 int retval = sched_rr_get_interval(pid, &t);
5715 retval = put_timespec64(&t, interval);
5720 #ifdef CONFIG_COMPAT_32BIT_TIME
5721 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5722 struct old_timespec32 __user *, interval)
5724 struct timespec64 t;
5725 int retval = sched_rr_get_interval(pid, &t);
5728 retval = put_old_timespec32(&t, interval);
5733 void sched_show_task(struct task_struct *p)
5735 unsigned long free = 0;
5738 if (!try_get_task_stack(p))
5741 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5743 if (p->state == TASK_RUNNING)
5744 printk(KERN_CONT " running task ");
5745 #ifdef CONFIG_DEBUG_STACK_USAGE
5746 free = stack_not_used(p);
5751 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5753 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5754 task_pid_nr(p), ppid,
5755 (unsigned long)task_thread_info(p)->flags);
5757 print_worker_info(KERN_INFO, p);
5758 show_stack(p, NULL);
5761 EXPORT_SYMBOL_GPL(sched_show_task);
5764 state_filter_match(unsigned long state_filter, struct task_struct *p)
5766 /* no filter, everything matches */
5770 /* filter, but doesn't match */
5771 if (!(p->state & state_filter))
5775 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5778 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5785 void show_state_filter(unsigned long state_filter)
5787 struct task_struct *g, *p;
5789 #if BITS_PER_LONG == 32
5791 " task PC stack pid father\n");
5794 " task PC stack pid father\n");
5797 for_each_process_thread(g, p) {
5799 * reset the NMI-timeout, listing all files on a slow
5800 * console might take a lot of time:
5801 * Also, reset softlockup watchdogs on all CPUs, because
5802 * another CPU might be blocked waiting for us to process
5805 touch_nmi_watchdog();
5806 touch_all_softlockup_watchdogs();
5807 if (state_filter_match(state_filter, p))
5811 #ifdef CONFIG_SCHED_DEBUG
5813 sysrq_sched_debug_show();
5817 * Only show locks if all tasks are dumped:
5820 debug_show_all_locks();
5824 * init_idle - set up an idle thread for a given CPU
5825 * @idle: task in question
5826 * @cpu: CPU the idle task belongs to
5828 * NOTE: this function does not set the idle thread's NEED_RESCHED
5829 * flag, to make booting more robust.
5831 void init_idle(struct task_struct *idle, int cpu)
5833 struct rq *rq = cpu_rq(cpu);
5834 unsigned long flags;
5836 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5837 raw_spin_lock(&rq->lock);
5839 __sched_fork(0, idle);
5840 idle->state = TASK_RUNNING;
5841 idle->se.exec_start = sched_clock();
5842 idle->flags |= PF_IDLE;
5844 kasan_unpoison_task_stack(idle);
5848 * Its possible that init_idle() gets called multiple times on a task,
5849 * in that case do_set_cpus_allowed() will not do the right thing.
5851 * And since this is boot we can forgo the serialization.
5853 set_cpus_allowed_common(idle, cpumask_of(cpu));
5856 * We're having a chicken and egg problem, even though we are
5857 * holding rq->lock, the CPU isn't yet set to this CPU so the
5858 * lockdep check in task_group() will fail.
5860 * Similar case to sched_fork(). / Alternatively we could
5861 * use task_rq_lock() here and obtain the other rq->lock.
5866 __set_task_cpu(idle, cpu);
5869 rq->curr = rq->idle = idle;
5870 idle->on_rq = TASK_ON_RQ_QUEUED;
5874 raw_spin_unlock(&rq->lock);
5875 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5877 /* Set the preempt count _outside_ the spinlocks! */
5878 init_idle_preempt_count(idle, cpu);
5881 * The idle tasks have their own, simple scheduling class:
5883 idle->sched_class = &idle_sched_class;
5884 ftrace_graph_init_idle_task(idle, cpu);
5885 vtime_init_idle(idle, cpu);
5887 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5893 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5894 const struct cpumask *trial)
5898 if (!cpumask_weight(cur))
5901 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5906 int task_can_attach(struct task_struct *p,
5907 const struct cpumask *cs_cpus_allowed)
5912 * Kthreads which disallow setaffinity shouldn't be moved
5913 * to a new cpuset; we don't want to change their CPU
5914 * affinity and isolating such threads by their set of
5915 * allowed nodes is unnecessary. Thus, cpusets are not
5916 * applicable for such threads. This prevents checking for
5917 * success of set_cpus_allowed_ptr() on all attached tasks
5918 * before cpus_mask may be changed.
5920 if (p->flags & PF_NO_SETAFFINITY) {
5925 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5927 ret = dl_task_can_attach(p, cs_cpus_allowed);
5933 bool sched_smp_initialized __read_mostly;
5935 #ifdef CONFIG_NUMA_BALANCING
5936 /* Migrate current task p to target_cpu */
5937 int migrate_task_to(struct task_struct *p, int target_cpu)
5939 struct migration_arg arg = { p, target_cpu };
5940 int curr_cpu = task_cpu(p);
5942 if (curr_cpu == target_cpu)
5945 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
5948 /* TODO: This is not properly updating schedstats */
5950 trace_sched_move_numa(p, curr_cpu, target_cpu);
5951 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5955 * Requeue a task on a given node and accurately track the number of NUMA
5956 * tasks on the runqueues
5958 void sched_setnuma(struct task_struct *p, int nid)
5960 bool queued, running;
5964 rq = task_rq_lock(p, &rf);
5965 queued = task_on_rq_queued(p);
5966 running = task_current(rq, p);
5969 dequeue_task(rq, p, DEQUEUE_SAVE);
5971 put_prev_task(rq, p);
5973 p->numa_preferred_nid = nid;
5976 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5978 set_curr_task(rq, p);
5979 task_rq_unlock(rq, p, &rf);
5981 #endif /* CONFIG_NUMA_BALANCING */
5983 #ifdef CONFIG_HOTPLUG_CPU
5985 * Ensure that the idle task is using init_mm right before its CPU goes
5988 void idle_task_exit(void)
5990 struct mm_struct *mm = current->active_mm;
5992 BUG_ON(cpu_online(smp_processor_id()));
5994 if (mm != &init_mm) {
5995 switch_mm(mm, &init_mm, current);
5996 current->active_mm = &init_mm;
5997 finish_arch_post_lock_switch();
6003 * Since this CPU is going 'away' for a while, fold any nr_active delta
6004 * we might have. Assumes we're called after migrate_tasks() so that the
6005 * nr_active count is stable. We need to take the teardown thread which
6006 * is calling this into account, so we hand in adjust = 1 to the load
6009 * Also see the comment "Global load-average calculations".
6011 static void calc_load_migrate(struct rq *rq)
6013 long delta = calc_load_fold_active(rq, 1);
6015 atomic_long_add(delta, &calc_load_tasks);
6018 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
6022 static const struct sched_class fake_sched_class = {
6023 .put_prev_task = put_prev_task_fake,
6026 static struct task_struct fake_task = {
6028 * Avoid pull_{rt,dl}_task()
6030 .prio = MAX_PRIO + 1,
6031 .sched_class = &fake_sched_class,
6035 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6036 * try_to_wake_up()->select_task_rq().
6038 * Called with rq->lock held even though we'er in stop_machine() and
6039 * there's no concurrency possible, we hold the required locks anyway
6040 * because of lock validation efforts.
6042 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6044 struct rq *rq = dead_rq;
6045 struct task_struct *next, *stop = rq->stop;
6046 struct rq_flags orf = *rf;
6050 * Fudge the rq selection such that the below task selection loop
6051 * doesn't get stuck on the currently eligible stop task.
6053 * We're currently inside stop_machine() and the rq is either stuck
6054 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6055 * either way we should never end up calling schedule() until we're
6061 * put_prev_task() and pick_next_task() sched
6062 * class method both need to have an up-to-date
6063 * value of rq->clock[_task]
6065 update_rq_clock(rq);
6069 * There's this thread running, bail when that's the only
6072 if (rq->nr_running == 1)
6076 * pick_next_task() assumes pinned rq->lock:
6078 next = pick_next_task(rq, &fake_task, rf);
6080 put_prev_task(rq, next);
6083 * Rules for changing task_struct::cpus_mask are holding
6084 * both pi_lock and rq->lock, such that holding either
6085 * stabilizes the mask.
6087 * Drop rq->lock is not quite as disastrous as it usually is
6088 * because !cpu_active at this point, which means load-balance
6089 * will not interfere. Also, stop-machine.
6092 raw_spin_lock(&next->pi_lock);
6096 * Since we're inside stop-machine, _nothing_ should have
6097 * changed the task, WARN if weird stuff happened, because in
6098 * that case the above rq->lock drop is a fail too.
6100 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6101 raw_spin_unlock(&next->pi_lock);
6105 /* Find suitable destination for @next, with force if needed. */
6106 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6107 rq = __migrate_task(rq, rf, next, dest_cpu);
6108 if (rq != dead_rq) {
6114 raw_spin_unlock(&next->pi_lock);
6119 #endif /* CONFIG_HOTPLUG_CPU */
6121 void set_rq_online(struct rq *rq)
6124 const struct sched_class *class;
6126 cpumask_set_cpu(rq->cpu, rq->rd->online);
6129 for_each_class(class) {
6130 if (class->rq_online)
6131 class->rq_online(rq);
6136 void set_rq_offline(struct rq *rq)
6139 const struct sched_class *class;
6141 for_each_class(class) {
6142 if (class->rq_offline)
6143 class->rq_offline(rq);
6146 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6152 * used to mark begin/end of suspend/resume:
6154 static int num_cpus_frozen;
6157 * Update cpusets according to cpu_active mask. If cpusets are
6158 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6159 * around partition_sched_domains().
6161 * If we come here as part of a suspend/resume, don't touch cpusets because we
6162 * want to restore it back to its original state upon resume anyway.
6164 static void cpuset_cpu_active(void)
6166 if (cpuhp_tasks_frozen) {
6168 * num_cpus_frozen tracks how many CPUs are involved in suspend
6169 * resume sequence. As long as this is not the last online
6170 * operation in the resume sequence, just build a single sched
6171 * domain, ignoring cpusets.
6173 partition_sched_domains(1, NULL, NULL);
6174 if (--num_cpus_frozen)
6177 * This is the last CPU online operation. So fall through and
6178 * restore the original sched domains by considering the
6179 * cpuset configurations.
6181 cpuset_force_rebuild();
6183 cpuset_update_active_cpus();
6186 static int cpuset_cpu_inactive(unsigned int cpu)
6188 if (!cpuhp_tasks_frozen) {
6189 if (dl_cpu_busy(cpu))
6191 cpuset_update_active_cpus();
6194 partition_sched_domains(1, NULL, NULL);
6199 int sched_cpu_activate(unsigned int cpu)
6201 struct rq *rq = cpu_rq(cpu);
6204 #ifdef CONFIG_SCHED_SMT
6206 * When going up, increment the number of cores with SMT present.
6208 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6209 static_branch_inc_cpuslocked(&sched_smt_present);
6211 set_cpu_active(cpu, true);
6213 if (sched_smp_initialized) {
6214 sched_domains_numa_masks_set(cpu);
6215 cpuset_cpu_active();
6219 * Put the rq online, if not already. This happens:
6221 * 1) In the early boot process, because we build the real domains
6222 * after all CPUs have been brought up.
6224 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6227 rq_lock_irqsave(rq, &rf);
6229 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6232 rq_unlock_irqrestore(rq, &rf);
6234 update_max_interval();
6239 int sched_cpu_deactivate(unsigned int cpu)
6243 set_cpu_active(cpu, false);
6245 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6246 * users of this state to go away such that all new such users will
6249 * Do sync before park smpboot threads to take care the rcu boost case.
6253 #ifdef CONFIG_SCHED_SMT
6255 * When going down, decrement the number of cores with SMT present.
6257 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6258 static_branch_dec_cpuslocked(&sched_smt_present);
6261 if (!sched_smp_initialized)
6264 ret = cpuset_cpu_inactive(cpu);
6266 set_cpu_active(cpu, true);
6269 sched_domains_numa_masks_clear(cpu);
6273 static void sched_rq_cpu_starting(unsigned int cpu)
6275 struct rq *rq = cpu_rq(cpu);
6277 rq->calc_load_update = calc_load_update;
6278 update_max_interval();
6281 int sched_cpu_starting(unsigned int cpu)
6283 sched_rq_cpu_starting(cpu);
6284 sched_tick_start(cpu);
6288 #ifdef CONFIG_HOTPLUG_CPU
6289 int sched_cpu_dying(unsigned int cpu)
6291 struct rq *rq = cpu_rq(cpu);
6294 /* Handle pending wakeups and then migrate everything off */
6295 sched_ttwu_pending();
6296 sched_tick_stop(cpu);
6298 rq_lock_irqsave(rq, &rf);
6300 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6303 migrate_tasks(rq, &rf);
6304 BUG_ON(rq->nr_running != 1);
6305 rq_unlock_irqrestore(rq, &rf);
6307 calc_load_migrate(rq);
6308 update_max_interval();
6309 nohz_balance_exit_idle(rq);
6315 void __init sched_init_smp(void)
6320 * There's no userspace yet to cause hotplug operations; hence all the
6321 * CPU masks are stable and all blatant races in the below code cannot
6324 mutex_lock(&sched_domains_mutex);
6325 sched_init_domains(cpu_active_mask);
6326 mutex_unlock(&sched_domains_mutex);
6328 /* Move init over to a non-isolated CPU */
6329 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6331 sched_init_granularity();
6333 init_sched_rt_class();
6334 init_sched_dl_class();
6336 sched_smp_initialized = true;
6339 static int __init migration_init(void)
6341 sched_cpu_starting(smp_processor_id());
6344 early_initcall(migration_init);
6347 void __init sched_init_smp(void)
6349 sched_init_granularity();
6351 #endif /* CONFIG_SMP */
6353 int in_sched_functions(unsigned long addr)
6355 return in_lock_functions(addr) ||
6356 (addr >= (unsigned long)__sched_text_start
6357 && addr < (unsigned long)__sched_text_end);
6360 #ifdef CONFIG_CGROUP_SCHED
6362 * Default task group.
6363 * Every task in system belongs to this group at bootup.
6365 struct task_group root_task_group;
6366 LIST_HEAD(task_groups);
6368 /* Cacheline aligned slab cache for task_group */
6369 static struct kmem_cache *task_group_cache __read_mostly;
6372 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6373 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6375 void __init sched_init(void)
6377 unsigned long alloc_size = 0, ptr;
6382 #ifdef CONFIG_FAIR_GROUP_SCHED
6383 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6385 #ifdef CONFIG_RT_GROUP_SCHED
6386 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6389 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6391 #ifdef CONFIG_FAIR_GROUP_SCHED
6392 root_task_group.se = (struct sched_entity **)ptr;
6393 ptr += nr_cpu_ids * sizeof(void **);
6395 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6396 ptr += nr_cpu_ids * sizeof(void **);
6398 #endif /* CONFIG_FAIR_GROUP_SCHED */
6399 #ifdef CONFIG_RT_GROUP_SCHED
6400 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6401 ptr += nr_cpu_ids * sizeof(void **);
6403 root_task_group.rt_rq = (struct rt_rq **)ptr;
6404 ptr += nr_cpu_ids * sizeof(void **);
6406 #endif /* CONFIG_RT_GROUP_SCHED */
6408 #ifdef CONFIG_CPUMASK_OFFSTACK
6409 for_each_possible_cpu(i) {
6410 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6411 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6412 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6413 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6415 #endif /* CONFIG_CPUMASK_OFFSTACK */
6417 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6418 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6421 init_defrootdomain();
6424 #ifdef CONFIG_RT_GROUP_SCHED
6425 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6426 global_rt_period(), global_rt_runtime());
6427 #endif /* CONFIG_RT_GROUP_SCHED */
6429 #ifdef CONFIG_CGROUP_SCHED
6430 task_group_cache = KMEM_CACHE(task_group, 0);
6432 list_add(&root_task_group.list, &task_groups);
6433 INIT_LIST_HEAD(&root_task_group.children);
6434 INIT_LIST_HEAD(&root_task_group.siblings);
6435 autogroup_init(&init_task);
6436 #endif /* CONFIG_CGROUP_SCHED */
6438 for_each_possible_cpu(i) {
6442 raw_spin_lock_init(&rq->lock);
6444 rq->calc_load_active = 0;
6445 rq->calc_load_update = jiffies + LOAD_FREQ;
6446 init_cfs_rq(&rq->cfs);
6447 init_rt_rq(&rq->rt);
6448 init_dl_rq(&rq->dl);
6449 #ifdef CONFIG_FAIR_GROUP_SCHED
6450 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6451 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6452 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6454 * How much CPU bandwidth does root_task_group get?
6456 * In case of task-groups formed thr' the cgroup filesystem, it
6457 * gets 100% of the CPU resources in the system. This overall
6458 * system CPU resource is divided among the tasks of
6459 * root_task_group and its child task-groups in a fair manner,
6460 * based on each entity's (task or task-group's) weight
6461 * (se->load.weight).
6463 * In other words, if root_task_group has 10 tasks of weight
6464 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6465 * then A0's share of the CPU resource is:
6467 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6469 * We achieve this by letting root_task_group's tasks sit
6470 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6472 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6473 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6474 #endif /* CONFIG_FAIR_GROUP_SCHED */
6476 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6477 #ifdef CONFIG_RT_GROUP_SCHED
6478 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6483 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6484 rq->balance_callback = NULL;
6485 rq->active_balance = 0;
6486 rq->next_balance = jiffies;
6491 rq->avg_idle = 2*sysctl_sched_migration_cost;
6492 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6494 INIT_LIST_HEAD(&rq->cfs_tasks);
6496 rq_attach_root(rq, &def_root_domain);
6497 #ifdef CONFIG_NO_HZ_COMMON
6498 rq->last_load_update_tick = jiffies;
6499 rq->last_blocked_load_update_tick = jiffies;
6500 atomic_set(&rq->nohz_flags, 0);
6502 #endif /* CONFIG_SMP */
6504 atomic_set(&rq->nr_iowait, 0);
6507 set_load_weight(&init_task, false);
6510 * The boot idle thread does lazy MMU switching as well:
6513 enter_lazy_tlb(&init_mm, current);
6516 * Make us the idle thread. Technically, schedule() should not be
6517 * called from this thread, however somewhere below it might be,
6518 * but because we are the idle thread, we just pick up running again
6519 * when this runqueue becomes "idle".
6521 init_idle(current, smp_processor_id());
6523 calc_load_update = jiffies + LOAD_FREQ;
6526 idle_thread_set_boot_cpu();
6528 init_sched_fair_class();
6536 scheduler_running = 1;
6539 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6540 static inline int preempt_count_equals(int preempt_offset)
6542 int nested = preempt_count() + rcu_preempt_depth();
6544 return (nested == preempt_offset);
6547 void __might_sleep(const char *file, int line, int preempt_offset)
6550 * Blocking primitives will set (and therefore destroy) current->state,
6551 * since we will exit with TASK_RUNNING make sure we enter with it,
6552 * otherwise we will destroy state.
6554 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6555 "do not call blocking ops when !TASK_RUNNING; "
6556 "state=%lx set at [<%p>] %pS\n",
6558 (void *)current->task_state_change,
6559 (void *)current->task_state_change);
6561 ___might_sleep(file, line, preempt_offset);
6563 EXPORT_SYMBOL(__might_sleep);
6565 void ___might_sleep(const char *file, int line, int preempt_offset)
6567 /* Ratelimiting timestamp: */
6568 static unsigned long prev_jiffy;
6570 unsigned long preempt_disable_ip;
6572 /* WARN_ON_ONCE() by default, no rate limit required: */
6575 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6576 !is_idle_task(current)) ||
6577 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6581 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6583 prev_jiffy = jiffies;
6585 /* Save this before calling printk(), since that will clobber it: */
6586 preempt_disable_ip = get_preempt_disable_ip(current);
6589 "BUG: sleeping function called from invalid context at %s:%d\n",
6592 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6593 in_atomic(), irqs_disabled(),
6594 current->pid, current->comm);
6596 if (task_stack_end_corrupted(current))
6597 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6599 debug_show_held_locks(current);
6600 if (irqs_disabled())
6601 print_irqtrace_events(current);
6602 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6603 && !preempt_count_equals(preempt_offset)) {
6604 pr_err("Preemption disabled at:");
6605 print_ip_sym(preempt_disable_ip);
6609 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6611 EXPORT_SYMBOL(___might_sleep);
6613 void __cant_sleep(const char *file, int line, int preempt_offset)
6615 static unsigned long prev_jiffy;
6617 if (irqs_disabled())
6620 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6623 if (preempt_count() > preempt_offset)
6626 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6628 prev_jiffy = jiffies;
6630 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6631 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6632 in_atomic(), irqs_disabled(),
6633 current->pid, current->comm);
6635 debug_show_held_locks(current);
6637 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6639 EXPORT_SYMBOL_GPL(__cant_sleep);
6642 #ifdef CONFIG_MAGIC_SYSRQ
6643 void normalize_rt_tasks(void)
6645 struct task_struct *g, *p;
6646 struct sched_attr attr = {
6647 .sched_policy = SCHED_NORMAL,
6650 read_lock(&tasklist_lock);
6651 for_each_process_thread(g, p) {
6653 * Only normalize user tasks:
6655 if (p->flags & PF_KTHREAD)
6658 p->se.exec_start = 0;
6659 schedstat_set(p->se.statistics.wait_start, 0);
6660 schedstat_set(p->se.statistics.sleep_start, 0);
6661 schedstat_set(p->se.statistics.block_start, 0);
6663 if (!dl_task(p) && !rt_task(p)) {
6665 * Renice negative nice level userspace
6668 if (task_nice(p) < 0)
6669 set_user_nice(p, 0);
6673 __sched_setscheduler(p, &attr, false, false);
6675 read_unlock(&tasklist_lock);
6678 #endif /* CONFIG_MAGIC_SYSRQ */
6680 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6682 * These functions are only useful for the IA64 MCA handling, or kdb.
6684 * They can only be called when the whole system has been
6685 * stopped - every CPU needs to be quiescent, and no scheduling
6686 * activity can take place. Using them for anything else would
6687 * be a serious bug, and as a result, they aren't even visible
6688 * under any other configuration.
6692 * curr_task - return the current task for a given CPU.
6693 * @cpu: the processor in question.
6695 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6697 * Return: The current task for @cpu.
6699 struct task_struct *curr_task(int cpu)
6701 return cpu_curr(cpu);
6704 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6708 * set_curr_task - set the current task for a given CPU.
6709 * @cpu: the processor in question.
6710 * @p: the task pointer to set.
6712 * Description: This function must only be used when non-maskable interrupts
6713 * are serviced on a separate stack. It allows the architecture to switch the
6714 * notion of the current task on a CPU in a non-blocking manner. This function
6715 * must be called with all CPU's synchronized, and interrupts disabled, the
6716 * and caller must save the original value of the current task (see
6717 * curr_task() above) and restore that value before reenabling interrupts and
6718 * re-starting the system.
6720 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6722 void ia64_set_curr_task(int cpu, struct task_struct *p)
6729 #ifdef CONFIG_CGROUP_SCHED
6730 /* task_group_lock serializes the addition/removal of task groups */
6731 static DEFINE_SPINLOCK(task_group_lock);
6733 static void sched_free_group(struct task_group *tg)
6735 free_fair_sched_group(tg);
6736 free_rt_sched_group(tg);
6738 kmem_cache_free(task_group_cache, tg);
6741 /* allocate runqueue etc for a new task group */
6742 struct task_group *sched_create_group(struct task_group *parent)
6744 struct task_group *tg;
6746 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6748 return ERR_PTR(-ENOMEM);
6750 if (!alloc_fair_sched_group(tg, parent))
6753 if (!alloc_rt_sched_group(tg, parent))
6759 sched_free_group(tg);
6760 return ERR_PTR(-ENOMEM);
6763 void sched_online_group(struct task_group *tg, struct task_group *parent)
6765 unsigned long flags;
6767 spin_lock_irqsave(&task_group_lock, flags);
6768 list_add_rcu(&tg->list, &task_groups);
6770 /* Root should already exist: */
6773 tg->parent = parent;
6774 INIT_LIST_HEAD(&tg->children);
6775 list_add_rcu(&tg->siblings, &parent->children);
6776 spin_unlock_irqrestore(&task_group_lock, flags);
6778 online_fair_sched_group(tg);
6781 /* rcu callback to free various structures associated with a task group */
6782 static void sched_free_group_rcu(struct rcu_head *rhp)
6784 /* Now it should be safe to free those cfs_rqs: */
6785 sched_free_group(container_of(rhp, struct task_group, rcu));
6788 void sched_destroy_group(struct task_group *tg)
6790 /* Wait for possible concurrent references to cfs_rqs complete: */
6791 call_rcu(&tg->rcu, sched_free_group_rcu);
6794 void sched_offline_group(struct task_group *tg)
6796 unsigned long flags;
6798 /* End participation in shares distribution: */
6799 unregister_fair_sched_group(tg);
6801 spin_lock_irqsave(&task_group_lock, flags);
6802 list_del_rcu(&tg->list);
6803 list_del_rcu(&tg->siblings);
6804 spin_unlock_irqrestore(&task_group_lock, flags);
6807 static void sched_change_group(struct task_struct *tsk, int type)
6809 struct task_group *tg;
6812 * All callers are synchronized by task_rq_lock(); we do not use RCU
6813 * which is pointless here. Thus, we pass "true" to task_css_check()
6814 * to prevent lockdep warnings.
6816 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6817 struct task_group, css);
6818 tg = autogroup_task_group(tsk, tg);
6819 tsk->sched_task_group = tg;
6821 #ifdef CONFIG_FAIR_GROUP_SCHED
6822 if (tsk->sched_class->task_change_group)
6823 tsk->sched_class->task_change_group(tsk, type);
6826 set_task_rq(tsk, task_cpu(tsk));
6830 * Change task's runqueue when it moves between groups.
6832 * The caller of this function should have put the task in its new group by
6833 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6836 void sched_move_task(struct task_struct *tsk)
6838 int queued, running, queue_flags =
6839 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6843 rq = task_rq_lock(tsk, &rf);
6844 update_rq_clock(rq);
6846 running = task_current(rq, tsk);
6847 queued = task_on_rq_queued(tsk);
6850 dequeue_task(rq, tsk, queue_flags);
6852 put_prev_task(rq, tsk);
6854 sched_change_group(tsk, TASK_MOVE_GROUP);
6857 enqueue_task(rq, tsk, queue_flags);
6859 set_curr_task(rq, tsk);
6861 task_rq_unlock(rq, tsk, &rf);
6864 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6866 return css ? container_of(css, struct task_group, css) : NULL;
6869 static struct cgroup_subsys_state *
6870 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6872 struct task_group *parent = css_tg(parent_css);
6873 struct task_group *tg;
6876 /* This is early initialization for the top cgroup */
6877 return &root_task_group.css;
6880 tg = sched_create_group(parent);
6882 return ERR_PTR(-ENOMEM);
6887 /* Expose task group only after completing cgroup initialization */
6888 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6890 struct task_group *tg = css_tg(css);
6891 struct task_group *parent = css_tg(css->parent);
6894 sched_online_group(tg, parent);
6898 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6900 struct task_group *tg = css_tg(css);
6902 sched_offline_group(tg);
6905 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6907 struct task_group *tg = css_tg(css);
6910 * Relies on the RCU grace period between css_released() and this.
6912 sched_free_group(tg);
6916 * This is called before wake_up_new_task(), therefore we really only
6917 * have to set its group bits, all the other stuff does not apply.
6919 static void cpu_cgroup_fork(struct task_struct *task)
6924 rq = task_rq_lock(task, &rf);
6926 update_rq_clock(rq);
6927 sched_change_group(task, TASK_SET_GROUP);
6929 task_rq_unlock(rq, task, &rf);
6932 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6934 struct task_struct *task;
6935 struct cgroup_subsys_state *css;
6938 cgroup_taskset_for_each(task, css, tset) {
6939 #ifdef CONFIG_RT_GROUP_SCHED
6940 if (!sched_rt_can_attach(css_tg(css), task))
6943 /* We don't support RT-tasks being in separate groups */
6944 if (task->sched_class != &fair_sched_class)
6948 * Serialize against wake_up_new_task() such that if its
6949 * running, we're sure to observe its full state.
6951 raw_spin_lock_irq(&task->pi_lock);
6953 * Avoid calling sched_move_task() before wake_up_new_task()
6954 * has happened. This would lead to problems with PELT, due to
6955 * move wanting to detach+attach while we're not attached yet.
6957 if (task->state == TASK_NEW)
6959 raw_spin_unlock_irq(&task->pi_lock);
6967 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6969 struct task_struct *task;
6970 struct cgroup_subsys_state *css;
6972 cgroup_taskset_for_each(task, css, tset)
6973 sched_move_task(task);
6976 #ifdef CONFIG_FAIR_GROUP_SCHED
6977 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6978 struct cftype *cftype, u64 shareval)
6980 if (shareval > scale_load_down(ULONG_MAX))
6981 shareval = MAX_SHARES;
6982 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6985 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6988 struct task_group *tg = css_tg(css);
6990 return (u64) scale_load_down(tg->shares);
6993 #ifdef CONFIG_CFS_BANDWIDTH
6994 static DEFINE_MUTEX(cfs_constraints_mutex);
6996 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6997 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6999 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7001 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7003 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7004 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7006 if (tg == &root_task_group)
7010 * Ensure we have at some amount of bandwidth every period. This is
7011 * to prevent reaching a state of large arrears when throttled via
7012 * entity_tick() resulting in prolonged exit starvation.
7014 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7018 * Likewise, bound things on the otherside by preventing insane quota
7019 * periods. This also allows us to normalize in computing quota
7022 if (period > max_cfs_quota_period)
7026 * Prevent race between setting of cfs_rq->runtime_enabled and
7027 * unthrottle_offline_cfs_rqs().
7030 mutex_lock(&cfs_constraints_mutex);
7031 ret = __cfs_schedulable(tg, period, quota);
7035 runtime_enabled = quota != RUNTIME_INF;
7036 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7038 * If we need to toggle cfs_bandwidth_used, off->on must occur
7039 * before making related changes, and on->off must occur afterwards
7041 if (runtime_enabled && !runtime_was_enabled)
7042 cfs_bandwidth_usage_inc();
7043 raw_spin_lock_irq(&cfs_b->lock);
7044 cfs_b->period = ns_to_ktime(period);
7045 cfs_b->quota = quota;
7047 __refill_cfs_bandwidth_runtime(cfs_b);
7049 /* Restart the period timer (if active) to handle new period expiry: */
7050 if (runtime_enabled)
7051 start_cfs_bandwidth(cfs_b);
7053 raw_spin_unlock_irq(&cfs_b->lock);
7055 for_each_online_cpu(i) {
7056 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7057 struct rq *rq = cfs_rq->rq;
7060 rq_lock_irq(rq, &rf);
7061 cfs_rq->runtime_enabled = runtime_enabled;
7062 cfs_rq->runtime_remaining = 0;
7064 if (cfs_rq->throttled)
7065 unthrottle_cfs_rq(cfs_rq);
7066 rq_unlock_irq(rq, &rf);
7068 if (runtime_was_enabled && !runtime_enabled)
7069 cfs_bandwidth_usage_dec();
7071 mutex_unlock(&cfs_constraints_mutex);
7077 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7081 period = ktime_to_ns(tg->cfs_bandwidth.period);
7082 if (cfs_quota_us < 0)
7083 quota = RUNTIME_INF;
7084 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7085 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7089 return tg_set_cfs_bandwidth(tg, period, quota);
7092 static long tg_get_cfs_quota(struct task_group *tg)
7096 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7099 quota_us = tg->cfs_bandwidth.quota;
7100 do_div(quota_us, NSEC_PER_USEC);
7105 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7109 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7112 period = (u64)cfs_period_us * NSEC_PER_USEC;
7113 quota = tg->cfs_bandwidth.quota;
7115 return tg_set_cfs_bandwidth(tg, period, quota);
7118 static long tg_get_cfs_period(struct task_group *tg)
7122 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7123 do_div(cfs_period_us, NSEC_PER_USEC);
7125 return cfs_period_us;
7128 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7131 return tg_get_cfs_quota(css_tg(css));
7134 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7135 struct cftype *cftype, s64 cfs_quota_us)
7137 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7140 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7143 return tg_get_cfs_period(css_tg(css));
7146 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7147 struct cftype *cftype, u64 cfs_period_us)
7149 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7152 struct cfs_schedulable_data {
7153 struct task_group *tg;
7158 * normalize group quota/period to be quota/max_period
7159 * note: units are usecs
7161 static u64 normalize_cfs_quota(struct task_group *tg,
7162 struct cfs_schedulable_data *d)
7170 period = tg_get_cfs_period(tg);
7171 quota = tg_get_cfs_quota(tg);
7174 /* note: these should typically be equivalent */
7175 if (quota == RUNTIME_INF || quota == -1)
7178 return to_ratio(period, quota);
7181 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7183 struct cfs_schedulable_data *d = data;
7184 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7185 s64 quota = 0, parent_quota = -1;
7188 quota = RUNTIME_INF;
7190 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7192 quota = normalize_cfs_quota(tg, d);
7193 parent_quota = parent_b->hierarchical_quota;
7196 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7197 * always take the min. On cgroup1, only inherit when no
7200 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7201 quota = min(quota, parent_quota);
7203 if (quota == RUNTIME_INF)
7204 quota = parent_quota;
7205 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7209 cfs_b->hierarchical_quota = quota;
7214 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7217 struct cfs_schedulable_data data = {
7223 if (quota != RUNTIME_INF) {
7224 do_div(data.period, NSEC_PER_USEC);
7225 do_div(data.quota, NSEC_PER_USEC);
7229 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7235 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7237 struct task_group *tg = css_tg(seq_css(sf));
7238 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7240 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7241 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7242 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7244 if (schedstat_enabled() && tg != &root_task_group) {
7248 for_each_possible_cpu(i)
7249 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7251 seq_printf(sf, "wait_sum %llu\n", ws);
7256 #endif /* CONFIG_CFS_BANDWIDTH */
7257 #endif /* CONFIG_FAIR_GROUP_SCHED */
7259 #ifdef CONFIG_RT_GROUP_SCHED
7260 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7261 struct cftype *cft, s64 val)
7263 return sched_group_set_rt_runtime(css_tg(css), val);
7266 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7269 return sched_group_rt_runtime(css_tg(css));
7272 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7273 struct cftype *cftype, u64 rt_period_us)
7275 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7278 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7281 return sched_group_rt_period(css_tg(css));
7283 #endif /* CONFIG_RT_GROUP_SCHED */
7285 static struct cftype cpu_legacy_files[] = {
7286 #ifdef CONFIG_FAIR_GROUP_SCHED
7289 .read_u64 = cpu_shares_read_u64,
7290 .write_u64 = cpu_shares_write_u64,
7293 #ifdef CONFIG_CFS_BANDWIDTH
7295 .name = "cfs_quota_us",
7296 .read_s64 = cpu_cfs_quota_read_s64,
7297 .write_s64 = cpu_cfs_quota_write_s64,
7300 .name = "cfs_period_us",
7301 .read_u64 = cpu_cfs_period_read_u64,
7302 .write_u64 = cpu_cfs_period_write_u64,
7306 .seq_show = cpu_cfs_stat_show,
7309 #ifdef CONFIG_RT_GROUP_SCHED
7311 .name = "rt_runtime_us",
7312 .read_s64 = cpu_rt_runtime_read,
7313 .write_s64 = cpu_rt_runtime_write,
7316 .name = "rt_period_us",
7317 .read_u64 = cpu_rt_period_read_uint,
7318 .write_u64 = cpu_rt_period_write_uint,
7324 static int cpu_extra_stat_show(struct seq_file *sf,
7325 struct cgroup_subsys_state *css)
7327 #ifdef CONFIG_CFS_BANDWIDTH
7329 struct task_group *tg = css_tg(css);
7330 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7333 throttled_usec = cfs_b->throttled_time;
7334 do_div(throttled_usec, NSEC_PER_USEC);
7336 seq_printf(sf, "nr_periods %d\n"
7338 "throttled_usec %llu\n",
7339 cfs_b->nr_periods, cfs_b->nr_throttled,
7346 #ifdef CONFIG_FAIR_GROUP_SCHED
7347 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7350 struct task_group *tg = css_tg(css);
7351 u64 weight = scale_load_down(tg->shares);
7353 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7356 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7357 struct cftype *cft, u64 weight)
7360 * cgroup weight knobs should use the common MIN, DFL and MAX
7361 * values which are 1, 100 and 10000 respectively. While it loses
7362 * a bit of range on both ends, it maps pretty well onto the shares
7363 * value used by scheduler and the round-trip conversions preserve
7364 * the original value over the entire range.
7366 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7369 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7371 return sched_group_set_shares(css_tg(css), scale_load(weight));
7374 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7377 unsigned long weight = scale_load_down(css_tg(css)->shares);
7378 int last_delta = INT_MAX;
7381 /* find the closest nice value to the current weight */
7382 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7383 delta = abs(sched_prio_to_weight[prio] - weight);
7384 if (delta >= last_delta)
7389 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7392 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7393 struct cftype *cft, s64 nice)
7395 unsigned long weight;
7398 if (nice < MIN_NICE || nice > MAX_NICE)
7401 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7402 idx = array_index_nospec(idx, 40);
7403 weight = sched_prio_to_weight[idx];
7405 return sched_group_set_shares(css_tg(css), scale_load(weight));
7409 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7410 long period, long quota)
7413 seq_puts(sf, "max");
7415 seq_printf(sf, "%ld", quota);
7417 seq_printf(sf, " %ld\n", period);
7420 /* caller should put the current value in *@periodp before calling */
7421 static int __maybe_unused cpu_period_quota_parse(char *buf,
7422 u64 *periodp, u64 *quotap)
7424 char tok[21]; /* U64_MAX */
7426 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7429 *periodp *= NSEC_PER_USEC;
7431 if (sscanf(tok, "%llu", quotap))
7432 *quotap *= NSEC_PER_USEC;
7433 else if (!strcmp(tok, "max"))
7434 *quotap = RUNTIME_INF;
7441 #ifdef CONFIG_CFS_BANDWIDTH
7442 static int cpu_max_show(struct seq_file *sf, void *v)
7444 struct task_group *tg = css_tg(seq_css(sf));
7446 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7450 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7451 char *buf, size_t nbytes, loff_t off)
7453 struct task_group *tg = css_tg(of_css(of));
7454 u64 period = tg_get_cfs_period(tg);
7458 ret = cpu_period_quota_parse(buf, &period, "a);
7460 ret = tg_set_cfs_bandwidth(tg, period, quota);
7461 return ret ?: nbytes;
7465 static struct cftype cpu_files[] = {
7466 #ifdef CONFIG_FAIR_GROUP_SCHED
7469 .flags = CFTYPE_NOT_ON_ROOT,
7470 .read_u64 = cpu_weight_read_u64,
7471 .write_u64 = cpu_weight_write_u64,
7474 .name = "weight.nice",
7475 .flags = CFTYPE_NOT_ON_ROOT,
7476 .read_s64 = cpu_weight_nice_read_s64,
7477 .write_s64 = cpu_weight_nice_write_s64,
7480 #ifdef CONFIG_CFS_BANDWIDTH
7483 .flags = CFTYPE_NOT_ON_ROOT,
7484 .seq_show = cpu_max_show,
7485 .write = cpu_max_write,
7491 struct cgroup_subsys cpu_cgrp_subsys = {
7492 .css_alloc = cpu_cgroup_css_alloc,
7493 .css_online = cpu_cgroup_css_online,
7494 .css_released = cpu_cgroup_css_released,
7495 .css_free = cpu_cgroup_css_free,
7496 .css_extra_stat_show = cpu_extra_stat_show,
7497 .fork = cpu_cgroup_fork,
7498 .can_attach = cpu_cgroup_can_attach,
7499 .attach = cpu_cgroup_attach,
7500 .legacy_cftypes = cpu_legacy_files,
7501 .dfl_cftypes = cpu_files,
7506 #endif /* CONFIG_CGROUP_SCHED */
7508 void dump_cpu_task(int cpu)
7510 pr_info("Task dump for CPU %d:\n", cpu);
7511 sched_show_task(cpu_curr(cpu));
7515 * Nice levels are multiplicative, with a gentle 10% change for every
7516 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7517 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7518 * that remained on nice 0.
7520 * The "10% effect" is relative and cumulative: from _any_ nice level,
7521 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7522 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7523 * If a task goes up by ~10% and another task goes down by ~10% then
7524 * the relative distance between them is ~25%.)
7526 const int sched_prio_to_weight[40] = {
7527 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7528 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7529 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7530 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7531 /* 0 */ 1024, 820, 655, 526, 423,
7532 /* 5 */ 335, 272, 215, 172, 137,
7533 /* 10 */ 110, 87, 70, 56, 45,
7534 /* 15 */ 36, 29, 23, 18, 15,
7538 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7540 * In cases where the weight does not change often, we can use the
7541 * precalculated inverse to speed up arithmetics by turning divisions
7542 * into multiplications:
7544 const u32 sched_prio_to_wmult[40] = {
7545 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7546 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7547 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7548 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7549 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7550 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7551 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7552 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7555 #undef CREATE_TRACE_POINTS