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 prepare_task_switch(rq, prev, next);
3220 * For paravirt, this is coupled with an exit in switch_to to
3221 * combine the page table reload and the switch backend into
3224 arch_start_context_switch(prev);
3227 * kernel -> kernel lazy + transfer active
3228 * user -> kernel lazy + mmgrab() active
3230 * kernel -> user switch + mmdrop() active
3231 * user -> user switch
3233 if (!next->mm) { // to kernel
3234 enter_lazy_tlb(prev->active_mm, next);
3236 next->active_mm = prev->active_mm;
3237 if (prev->mm) // from user
3238 mmgrab(prev->active_mm);
3240 prev->active_mm = NULL;
3243 * sys_membarrier() requires an smp_mb() between setting
3244 * rq->curr and returning to userspace.
3246 * The below provides this either through switch_mm(), or in
3247 * case 'prev->active_mm == next->mm' through
3248 * finish_task_switch()'s mmdrop().
3251 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3253 if (!prev->mm) { // from kernel
3254 /* will mmdrop() in finish_task_switch(). */
3255 rq->prev_mm = prev->active_mm;
3256 prev->active_mm = NULL;
3260 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3262 prepare_lock_switch(rq, next, rf);
3264 /* Here we just switch the register state and the stack. */
3265 switch_to(prev, next, prev);
3268 return finish_task_switch(prev);
3272 * nr_running and nr_context_switches:
3274 * externally visible scheduler statistics: current number of runnable
3275 * threads, total number of context switches performed since bootup.
3277 unsigned long nr_running(void)
3279 unsigned long i, sum = 0;
3281 for_each_online_cpu(i)
3282 sum += cpu_rq(i)->nr_running;
3288 * Check if only the current task is running on the CPU.
3290 * Caution: this function does not check that the caller has disabled
3291 * preemption, thus the result might have a time-of-check-to-time-of-use
3292 * race. The caller is responsible to use it correctly, for example:
3294 * - from a non-preemptible section (of course)
3296 * - from a thread that is bound to a single CPU
3298 * - in a loop with very short iterations (e.g. a polling loop)
3300 bool single_task_running(void)
3302 return raw_rq()->nr_running == 1;
3304 EXPORT_SYMBOL(single_task_running);
3306 unsigned long long nr_context_switches(void)
3309 unsigned long long sum = 0;
3311 for_each_possible_cpu(i)
3312 sum += cpu_rq(i)->nr_switches;
3318 * Consumers of these two interfaces, like for example the cpuidle menu
3319 * governor, are using nonsensical data. Preferring shallow idle state selection
3320 * for a CPU that has IO-wait which might not even end up running the task when
3321 * it does become runnable.
3324 unsigned long nr_iowait_cpu(int cpu)
3326 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3330 * IO-wait accounting, and how its mostly bollocks (on SMP).
3332 * The idea behind IO-wait account is to account the idle time that we could
3333 * have spend running if it were not for IO. That is, if we were to improve the
3334 * storage performance, we'd have a proportional reduction in IO-wait time.
3336 * This all works nicely on UP, where, when a task blocks on IO, we account
3337 * idle time as IO-wait, because if the storage were faster, it could've been
3338 * running and we'd not be idle.
3340 * This has been extended to SMP, by doing the same for each CPU. This however
3343 * Imagine for instance the case where two tasks block on one CPU, only the one
3344 * CPU will have IO-wait accounted, while the other has regular idle. Even
3345 * though, if the storage were faster, both could've ran at the same time,
3346 * utilising both CPUs.
3348 * This means, that when looking globally, the current IO-wait accounting on
3349 * SMP is a lower bound, by reason of under accounting.
3351 * Worse, since the numbers are provided per CPU, they are sometimes
3352 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3353 * associated with any one particular CPU, it can wake to another CPU than it
3354 * blocked on. This means the per CPU IO-wait number is meaningless.
3356 * Task CPU affinities can make all that even more 'interesting'.
3359 unsigned long nr_iowait(void)
3361 unsigned long i, sum = 0;
3363 for_each_possible_cpu(i)
3364 sum += nr_iowait_cpu(i);
3372 * sched_exec - execve() is a valuable balancing opportunity, because at
3373 * this point the task has the smallest effective memory and cache footprint.
3375 void sched_exec(void)
3377 struct task_struct *p = current;
3378 unsigned long flags;
3381 raw_spin_lock_irqsave(&p->pi_lock, flags);
3382 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3383 if (dest_cpu == smp_processor_id())
3386 if (likely(cpu_active(dest_cpu))) {
3387 struct migration_arg arg = { p, dest_cpu };
3389 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3390 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3394 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3399 DEFINE_PER_CPU(struct kernel_stat, kstat);
3400 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3402 EXPORT_PER_CPU_SYMBOL(kstat);
3403 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3406 * The function fair_sched_class.update_curr accesses the struct curr
3407 * and its field curr->exec_start; when called from task_sched_runtime(),
3408 * we observe a high rate of cache misses in practice.
3409 * Prefetching this data results in improved performance.
3411 static inline void prefetch_curr_exec_start(struct task_struct *p)
3413 #ifdef CONFIG_FAIR_GROUP_SCHED
3414 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3416 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3419 prefetch(&curr->exec_start);
3423 * Return accounted runtime for the task.
3424 * In case the task is currently running, return the runtime plus current's
3425 * pending runtime that have not been accounted yet.
3427 unsigned long long task_sched_runtime(struct task_struct *p)
3433 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3435 * 64-bit doesn't need locks to atomically read a 64-bit value.
3436 * So we have a optimization chance when the task's delta_exec is 0.
3437 * Reading ->on_cpu is racy, but this is ok.
3439 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3440 * If we race with it entering CPU, unaccounted time is 0. This is
3441 * indistinguishable from the read occurring a few cycles earlier.
3442 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3443 * been accounted, so we're correct here as well.
3445 if (!p->on_cpu || !task_on_rq_queued(p))
3446 return p->se.sum_exec_runtime;
3449 rq = task_rq_lock(p, &rf);
3451 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3452 * project cycles that may never be accounted to this
3453 * thread, breaking clock_gettime().
3455 if (task_current(rq, p) && task_on_rq_queued(p)) {
3456 prefetch_curr_exec_start(p);
3457 update_rq_clock(rq);
3458 p->sched_class->update_curr(rq);
3460 ns = p->se.sum_exec_runtime;
3461 task_rq_unlock(rq, p, &rf);
3467 * This function gets called by the timer code, with HZ frequency.
3468 * We call it with interrupts disabled.
3470 void scheduler_tick(void)
3472 int cpu = smp_processor_id();
3473 struct rq *rq = cpu_rq(cpu);
3474 struct task_struct *curr = rq->curr;
3481 update_rq_clock(rq);
3482 curr->sched_class->task_tick(rq, curr, 0);
3483 calc_global_load_tick(rq);
3488 perf_event_task_tick();
3491 rq->idle_balance = idle_cpu(cpu);
3492 trigger_load_balance(rq);
3496 #ifdef CONFIG_NO_HZ_FULL
3501 struct delayed_work work;
3503 /* Values for ->state, see diagram below. */
3504 #define TICK_SCHED_REMOTE_OFFLINE 0
3505 #define TICK_SCHED_REMOTE_OFFLINING 1
3506 #define TICK_SCHED_REMOTE_RUNNING 2
3509 * State diagram for ->state:
3512 * TICK_SCHED_REMOTE_OFFLINE
3515 * | | sched_tick_remote()
3518 * +--TICK_SCHED_REMOTE_OFFLINING
3521 * sched_tick_start() | | sched_tick_stop()
3524 * TICK_SCHED_REMOTE_RUNNING
3527 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3528 * and sched_tick_start() are happy to leave the state in RUNNING.
3531 static struct tick_work __percpu *tick_work_cpu;
3533 static void sched_tick_remote(struct work_struct *work)
3535 struct delayed_work *dwork = to_delayed_work(work);
3536 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3537 int cpu = twork->cpu;
3538 struct rq *rq = cpu_rq(cpu);
3539 struct task_struct *curr;
3545 * Handle the tick only if it appears the remote CPU is running in full
3546 * dynticks mode. The check is racy by nature, but missing a tick or
3547 * having one too much is no big deal because the scheduler tick updates
3548 * statistics and checks timeslices in a time-independent way, regardless
3549 * of when exactly it is running.
3551 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3554 rq_lock_irq(rq, &rf);
3556 if (is_idle_task(curr) || cpu_is_offline(cpu))
3559 update_rq_clock(rq);
3560 delta = rq_clock_task(rq) - curr->se.exec_start;
3563 * Make sure the next tick runs within a reasonable
3566 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3567 curr->sched_class->task_tick(rq, curr, 0);
3570 rq_unlock_irq(rq, &rf);
3574 * Run the remote tick once per second (1Hz). This arbitrary
3575 * frequency is large enough to avoid overload but short enough
3576 * to keep scheduler internal stats reasonably up to date. But
3577 * first update state to reflect hotplug activity if required.
3579 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3580 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3581 if (os == TICK_SCHED_REMOTE_RUNNING)
3582 queue_delayed_work(system_unbound_wq, dwork, HZ);
3585 static void sched_tick_start(int cpu)
3588 struct tick_work *twork;
3590 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3593 WARN_ON_ONCE(!tick_work_cpu);
3595 twork = per_cpu_ptr(tick_work_cpu, cpu);
3596 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3597 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3598 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3600 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3601 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3605 #ifdef CONFIG_HOTPLUG_CPU
3606 static void sched_tick_stop(int cpu)
3608 struct tick_work *twork;
3611 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3614 WARN_ON_ONCE(!tick_work_cpu);
3616 twork = per_cpu_ptr(tick_work_cpu, cpu);
3617 /* There cannot be competing actions, but don't rely on stop-machine. */
3618 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3619 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3620 /* Don't cancel, as this would mess up the state machine. */
3622 #endif /* CONFIG_HOTPLUG_CPU */
3624 int __init sched_tick_offload_init(void)
3626 tick_work_cpu = alloc_percpu(struct tick_work);
3627 BUG_ON(!tick_work_cpu);
3631 #else /* !CONFIG_NO_HZ_FULL */
3632 static inline void sched_tick_start(int cpu) { }
3633 static inline void sched_tick_stop(int cpu) { }
3636 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3637 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3639 * If the value passed in is equal to the current preempt count
3640 * then we just disabled preemption. Start timing the latency.
3642 static inline void preempt_latency_start(int val)
3644 if (preempt_count() == val) {
3645 unsigned long ip = get_lock_parent_ip();
3646 #ifdef CONFIG_DEBUG_PREEMPT
3647 current->preempt_disable_ip = ip;
3649 trace_preempt_off(CALLER_ADDR0, ip);
3653 void preempt_count_add(int val)
3655 #ifdef CONFIG_DEBUG_PREEMPT
3659 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3662 __preempt_count_add(val);
3663 #ifdef CONFIG_DEBUG_PREEMPT
3665 * Spinlock count overflowing soon?
3667 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3670 preempt_latency_start(val);
3672 EXPORT_SYMBOL(preempt_count_add);
3673 NOKPROBE_SYMBOL(preempt_count_add);
3676 * If the value passed in equals to the current preempt count
3677 * then we just enabled preemption. Stop timing the latency.
3679 static inline void preempt_latency_stop(int val)
3681 if (preempt_count() == val)
3682 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3685 void preempt_count_sub(int val)
3687 #ifdef CONFIG_DEBUG_PREEMPT
3691 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3694 * Is the spinlock portion underflowing?
3696 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3697 !(preempt_count() & PREEMPT_MASK)))
3701 preempt_latency_stop(val);
3702 __preempt_count_sub(val);
3704 EXPORT_SYMBOL(preempt_count_sub);
3705 NOKPROBE_SYMBOL(preempt_count_sub);
3708 static inline void preempt_latency_start(int val) { }
3709 static inline void preempt_latency_stop(int val) { }
3712 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3714 #ifdef CONFIG_DEBUG_PREEMPT
3715 return p->preempt_disable_ip;
3722 * Print scheduling while atomic bug:
3724 static noinline void __schedule_bug(struct task_struct *prev)
3726 /* Save this before calling printk(), since that will clobber it */
3727 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3729 if (oops_in_progress)
3732 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3733 prev->comm, prev->pid, preempt_count());
3735 debug_show_held_locks(prev);
3737 if (irqs_disabled())
3738 print_irqtrace_events(prev);
3739 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3740 && in_atomic_preempt_off()) {
3741 pr_err("Preemption disabled at:");
3742 print_ip_sym(preempt_disable_ip);
3746 panic("scheduling while atomic\n");
3749 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3753 * Various schedule()-time debugging checks and statistics:
3755 static inline void schedule_debug(struct task_struct *prev)
3757 #ifdef CONFIG_SCHED_STACK_END_CHECK
3758 if (task_stack_end_corrupted(prev))
3759 panic("corrupted stack end detected inside scheduler\n");
3762 if (unlikely(in_atomic_preempt_off())) {
3763 __schedule_bug(prev);
3764 preempt_count_set(PREEMPT_DISABLED);
3768 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3770 schedstat_inc(this_rq()->sched_count);
3774 * Pick up the highest-prio task:
3776 static inline struct task_struct *
3777 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3779 const struct sched_class *class;
3780 struct task_struct *p;
3783 * Optimization: we know that if all tasks are in the fair class we can
3784 * call that function directly, but only if the @prev task wasn't of a
3785 * higher scheduling class, because otherwise those loose the
3786 * opportunity to pull in more work from other CPUs.
3788 if (likely((prev->sched_class == &idle_sched_class ||
3789 prev->sched_class == &fair_sched_class) &&
3790 rq->nr_running == rq->cfs.h_nr_running)) {
3792 p = fair_sched_class.pick_next_task(rq, prev, rf);
3793 if (unlikely(p == RETRY_TASK))
3796 /* Assumes fair_sched_class->next == idle_sched_class */
3798 p = idle_sched_class.pick_next_task(rq, prev, rf);
3804 for_each_class(class) {
3805 p = class->pick_next_task(rq, prev, rf);
3807 if (unlikely(p == RETRY_TASK))
3813 /* The idle class should always have a runnable task: */
3818 * __schedule() is the main scheduler function.
3820 * The main means of driving the scheduler and thus entering this function are:
3822 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3824 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3825 * paths. For example, see arch/x86/entry_64.S.
3827 * To drive preemption between tasks, the scheduler sets the flag in timer
3828 * interrupt handler scheduler_tick().
3830 * 3. Wakeups don't really cause entry into schedule(). They add a
3831 * task to the run-queue and that's it.
3833 * Now, if the new task added to the run-queue preempts the current
3834 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3835 * called on the nearest possible occasion:
3837 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3839 * - in syscall or exception context, at the next outmost
3840 * preempt_enable(). (this might be as soon as the wake_up()'s
3843 * - in IRQ context, return from interrupt-handler to
3844 * preemptible context
3846 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3849 * - cond_resched() call
3850 * - explicit schedule() call
3851 * - return from syscall or exception to user-space
3852 * - return from interrupt-handler to user-space
3854 * WARNING: must be called with preemption disabled!
3856 static void __sched notrace __schedule(bool preempt)
3858 struct task_struct *prev, *next;
3859 unsigned long *switch_count;
3864 cpu = smp_processor_id();
3868 schedule_debug(prev);
3870 if (sched_feat(HRTICK))
3873 local_irq_disable();
3874 rcu_note_context_switch(preempt);
3877 * Make sure that signal_pending_state()->signal_pending() below
3878 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3879 * done by the caller to avoid the race with signal_wake_up().
3881 * The membarrier system call requires a full memory barrier
3882 * after coming from user-space, before storing to rq->curr.
3885 smp_mb__after_spinlock();
3887 /* Promote REQ to ACT */
3888 rq->clock_update_flags <<= 1;
3889 update_rq_clock(rq);
3891 switch_count = &prev->nivcsw;
3892 if (!preempt && prev->state) {
3893 if (signal_pending_state(prev->state, prev)) {
3894 prev->state = TASK_RUNNING;
3896 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3898 if (prev->in_iowait) {
3899 atomic_inc(&rq->nr_iowait);
3900 delayacct_blkio_start();
3903 switch_count = &prev->nvcsw;
3906 next = pick_next_task(rq, prev, &rf);
3907 clear_tsk_need_resched(prev);
3908 clear_preempt_need_resched();
3910 if (likely(prev != next)) {
3914 * The membarrier system call requires each architecture
3915 * to have a full memory barrier after updating
3916 * rq->curr, before returning to user-space.
3918 * Here are the schemes providing that barrier on the
3919 * various architectures:
3920 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3921 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3922 * - finish_lock_switch() for weakly-ordered
3923 * architectures where spin_unlock is a full barrier,
3924 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3925 * is a RELEASE barrier),
3929 trace_sched_switch(preempt, prev, next);
3931 /* Also unlocks the rq: */
3932 rq = context_switch(rq, prev, next, &rf);
3934 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3935 rq_unlock_irq(rq, &rf);
3938 balance_callback(rq);
3941 void __noreturn do_task_dead(void)
3943 /* Causes final put_task_struct in finish_task_switch(): */
3944 set_special_state(TASK_DEAD);
3946 /* Tell freezer to ignore us: */
3947 current->flags |= PF_NOFREEZE;
3952 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3957 static inline void sched_submit_work(struct task_struct *tsk)
3959 if (!tsk->state || tsk_is_pi_blocked(tsk))
3963 * If a worker went to sleep, notify and ask workqueue whether
3964 * it wants to wake up a task to maintain concurrency.
3965 * As this function is called inside the schedule() context,
3966 * we disable preemption to avoid it calling schedule() again
3967 * in the possible wakeup of a kworker.
3969 if (tsk->flags & PF_WQ_WORKER) {
3971 wq_worker_sleeping(tsk);
3972 preempt_enable_no_resched();
3976 * If we are going to sleep and we have plugged IO queued,
3977 * make sure to submit it to avoid deadlocks.
3979 if (blk_needs_flush_plug(tsk))
3980 blk_schedule_flush_plug(tsk);
3983 static void sched_update_worker(struct task_struct *tsk)
3985 if (tsk->flags & PF_WQ_WORKER)
3986 wq_worker_running(tsk);
3989 asmlinkage __visible void __sched schedule(void)
3991 struct task_struct *tsk = current;
3993 sched_submit_work(tsk);
3997 sched_preempt_enable_no_resched();
3998 } while (need_resched());
3999 sched_update_worker(tsk);
4001 EXPORT_SYMBOL(schedule);
4004 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4005 * state (have scheduled out non-voluntarily) by making sure that all
4006 * tasks have either left the run queue or have gone into user space.
4007 * As idle tasks do not do either, they must not ever be preempted
4008 * (schedule out non-voluntarily).
4010 * schedule_idle() is similar to schedule_preempt_disable() except that it
4011 * never enables preemption because it does not call sched_submit_work().
4013 void __sched schedule_idle(void)
4016 * As this skips calling sched_submit_work(), which the idle task does
4017 * regardless because that function is a nop when the task is in a
4018 * TASK_RUNNING state, make sure this isn't used someplace that the
4019 * current task can be in any other state. Note, idle is always in the
4020 * TASK_RUNNING state.
4022 WARN_ON_ONCE(current->state);
4025 } while (need_resched());
4028 #ifdef CONFIG_CONTEXT_TRACKING
4029 asmlinkage __visible void __sched schedule_user(void)
4032 * If we come here after a random call to set_need_resched(),
4033 * or we have been woken up remotely but the IPI has not yet arrived,
4034 * we haven't yet exited the RCU idle mode. Do it here manually until
4035 * we find a better solution.
4037 * NB: There are buggy callers of this function. Ideally we
4038 * should warn if prev_state != CONTEXT_USER, but that will trigger
4039 * too frequently to make sense yet.
4041 enum ctx_state prev_state = exception_enter();
4043 exception_exit(prev_state);
4048 * schedule_preempt_disabled - called with preemption disabled
4050 * Returns with preemption disabled. Note: preempt_count must be 1
4052 void __sched schedule_preempt_disabled(void)
4054 sched_preempt_enable_no_resched();
4059 static void __sched notrace preempt_schedule_common(void)
4063 * Because the function tracer can trace preempt_count_sub()
4064 * and it also uses preempt_enable/disable_notrace(), if
4065 * NEED_RESCHED is set, the preempt_enable_notrace() called
4066 * by the function tracer will call this function again and
4067 * cause infinite recursion.
4069 * Preemption must be disabled here before the function
4070 * tracer can trace. Break up preempt_disable() into two
4071 * calls. One to disable preemption without fear of being
4072 * traced. The other to still record the preemption latency,
4073 * which can also be traced by the function tracer.
4075 preempt_disable_notrace();
4076 preempt_latency_start(1);
4078 preempt_latency_stop(1);
4079 preempt_enable_no_resched_notrace();
4082 * Check again in case we missed a preemption opportunity
4083 * between schedule and now.
4085 } while (need_resched());
4088 #ifdef CONFIG_PREEMPT
4090 * this is the entry point to schedule() from in-kernel preemption
4091 * off of preempt_enable. Kernel preemptions off return from interrupt
4092 * occur there and call schedule directly.
4094 asmlinkage __visible void __sched notrace preempt_schedule(void)
4097 * If there is a non-zero preempt_count or interrupts are disabled,
4098 * we do not want to preempt the current task. Just return..
4100 if (likely(!preemptible()))
4103 preempt_schedule_common();
4105 NOKPROBE_SYMBOL(preempt_schedule);
4106 EXPORT_SYMBOL(preempt_schedule);
4109 * preempt_schedule_notrace - preempt_schedule called by tracing
4111 * The tracing infrastructure uses preempt_enable_notrace to prevent
4112 * recursion and tracing preempt enabling caused by the tracing
4113 * infrastructure itself. But as tracing can happen in areas coming
4114 * from userspace or just about to enter userspace, a preempt enable
4115 * can occur before user_exit() is called. This will cause the scheduler
4116 * to be called when the system is still in usermode.
4118 * To prevent this, the preempt_enable_notrace will use this function
4119 * instead of preempt_schedule() to exit user context if needed before
4120 * calling the scheduler.
4122 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4124 enum ctx_state prev_ctx;
4126 if (likely(!preemptible()))
4131 * Because the function tracer can trace preempt_count_sub()
4132 * and it also uses preempt_enable/disable_notrace(), if
4133 * NEED_RESCHED is set, the preempt_enable_notrace() called
4134 * by the function tracer will call this function again and
4135 * cause infinite recursion.
4137 * Preemption must be disabled here before the function
4138 * tracer can trace. Break up preempt_disable() into two
4139 * calls. One to disable preemption without fear of being
4140 * traced. The other to still record the preemption latency,
4141 * which can also be traced by the function tracer.
4143 preempt_disable_notrace();
4144 preempt_latency_start(1);
4146 * Needs preempt disabled in case user_exit() is traced
4147 * and the tracer calls preempt_enable_notrace() causing
4148 * an infinite recursion.
4150 prev_ctx = exception_enter();
4152 exception_exit(prev_ctx);
4154 preempt_latency_stop(1);
4155 preempt_enable_no_resched_notrace();
4156 } while (need_resched());
4158 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4160 #endif /* CONFIG_PREEMPT */
4163 * this is the entry point to schedule() from kernel preemption
4164 * off of irq context.
4165 * Note, that this is called and return with irqs disabled. This will
4166 * protect us against recursive calling from irq.
4168 asmlinkage __visible void __sched preempt_schedule_irq(void)
4170 enum ctx_state prev_state;
4172 /* Catch callers which need to be fixed */
4173 BUG_ON(preempt_count() || !irqs_disabled());
4175 prev_state = exception_enter();
4181 local_irq_disable();
4182 sched_preempt_enable_no_resched();
4183 } while (need_resched());
4185 exception_exit(prev_state);
4188 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4191 return try_to_wake_up(curr->private, mode, wake_flags);
4193 EXPORT_SYMBOL(default_wake_function);
4195 #ifdef CONFIG_RT_MUTEXES
4197 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4200 prio = min(prio, pi_task->prio);
4205 static inline int rt_effective_prio(struct task_struct *p, int prio)
4207 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4209 return __rt_effective_prio(pi_task, prio);
4213 * rt_mutex_setprio - set the current priority of a task
4215 * @pi_task: donor task
4217 * This function changes the 'effective' priority of a task. It does
4218 * not touch ->normal_prio like __setscheduler().
4220 * Used by the rt_mutex code to implement priority inheritance
4221 * logic. Call site only calls if the priority of the task changed.
4223 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4225 int prio, oldprio, queued, running, queue_flag =
4226 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4227 const struct sched_class *prev_class;
4231 /* XXX used to be waiter->prio, not waiter->task->prio */
4232 prio = __rt_effective_prio(pi_task, p->normal_prio);
4235 * If nothing changed; bail early.
4237 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4240 rq = __task_rq_lock(p, &rf);
4241 update_rq_clock(rq);
4243 * Set under pi_lock && rq->lock, such that the value can be used under
4246 * Note that there is loads of tricky to make this pointer cache work
4247 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4248 * ensure a task is de-boosted (pi_task is set to NULL) before the
4249 * task is allowed to run again (and can exit). This ensures the pointer
4250 * points to a blocked task -- which guaratees the task is present.
4252 p->pi_top_task = pi_task;
4255 * For FIFO/RR we only need to set prio, if that matches we're done.
4257 if (prio == p->prio && !dl_prio(prio))
4261 * Idle task boosting is a nono in general. There is one
4262 * exception, when PREEMPT_RT and NOHZ is active:
4264 * The idle task calls get_next_timer_interrupt() and holds
4265 * the timer wheel base->lock on the CPU and another CPU wants
4266 * to access the timer (probably to cancel it). We can safely
4267 * ignore the boosting request, as the idle CPU runs this code
4268 * with interrupts disabled and will complete the lock
4269 * protected section without being interrupted. So there is no
4270 * real need to boost.
4272 if (unlikely(p == rq->idle)) {
4273 WARN_ON(p != rq->curr);
4274 WARN_ON(p->pi_blocked_on);
4278 trace_sched_pi_setprio(p, pi_task);
4281 if (oldprio == prio)
4282 queue_flag &= ~DEQUEUE_MOVE;
4284 prev_class = p->sched_class;
4285 queued = task_on_rq_queued(p);
4286 running = task_current(rq, p);
4288 dequeue_task(rq, p, queue_flag);
4290 put_prev_task(rq, p);
4293 * Boosting condition are:
4294 * 1. -rt task is running and holds mutex A
4295 * --> -dl task blocks on mutex A
4297 * 2. -dl task is running and holds mutex A
4298 * --> -dl task blocks on mutex A and could preempt the
4301 if (dl_prio(prio)) {
4302 if (!dl_prio(p->normal_prio) ||
4303 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4304 p->dl.dl_boosted = 1;
4305 queue_flag |= ENQUEUE_REPLENISH;
4307 p->dl.dl_boosted = 0;
4308 p->sched_class = &dl_sched_class;
4309 } else if (rt_prio(prio)) {
4310 if (dl_prio(oldprio))
4311 p->dl.dl_boosted = 0;
4313 queue_flag |= ENQUEUE_HEAD;
4314 p->sched_class = &rt_sched_class;
4316 if (dl_prio(oldprio))
4317 p->dl.dl_boosted = 0;
4318 if (rt_prio(oldprio))
4320 p->sched_class = &fair_sched_class;
4326 enqueue_task(rq, p, queue_flag);
4328 set_curr_task(rq, p);
4330 check_class_changed(rq, p, prev_class, oldprio);
4332 /* Avoid rq from going away on us: */
4334 __task_rq_unlock(rq, &rf);
4336 balance_callback(rq);
4340 static inline int rt_effective_prio(struct task_struct *p, int prio)
4346 void set_user_nice(struct task_struct *p, long nice)
4348 bool queued, running;
4349 int old_prio, delta;
4353 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4356 * We have to be careful, if called from sys_setpriority(),
4357 * the task might be in the middle of scheduling on another CPU.
4359 rq = task_rq_lock(p, &rf);
4360 update_rq_clock(rq);
4363 * The RT priorities are set via sched_setscheduler(), but we still
4364 * allow the 'normal' nice value to be set - but as expected
4365 * it wont have any effect on scheduling until the task is
4366 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4368 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4369 p->static_prio = NICE_TO_PRIO(nice);
4372 queued = task_on_rq_queued(p);
4373 running = task_current(rq, p);
4375 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4377 put_prev_task(rq, p);
4379 p->static_prio = NICE_TO_PRIO(nice);
4380 set_load_weight(p, true);
4382 p->prio = effective_prio(p);
4383 delta = p->prio - old_prio;
4386 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4388 * If the task increased its priority or is running and
4389 * lowered its priority, then reschedule its CPU:
4391 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4395 set_curr_task(rq, p);
4397 task_rq_unlock(rq, p, &rf);
4399 EXPORT_SYMBOL(set_user_nice);
4402 * can_nice - check if a task can reduce its nice value
4406 int can_nice(const struct task_struct *p, const int nice)
4408 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4409 int nice_rlim = nice_to_rlimit(nice);
4411 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4412 capable(CAP_SYS_NICE));
4415 #ifdef __ARCH_WANT_SYS_NICE
4418 * sys_nice - change the priority of the current process.
4419 * @increment: priority increment
4421 * sys_setpriority is a more generic, but much slower function that
4422 * does similar things.
4424 SYSCALL_DEFINE1(nice, int, increment)
4429 * Setpriority might change our priority at the same moment.
4430 * We don't have to worry. Conceptually one call occurs first
4431 * and we have a single winner.
4433 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4434 nice = task_nice(current) + increment;
4436 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4437 if (increment < 0 && !can_nice(current, nice))
4440 retval = security_task_setnice(current, nice);
4444 set_user_nice(current, nice);
4451 * task_prio - return the priority value of a given task.
4452 * @p: the task in question.
4454 * Return: The priority value as seen by users in /proc.
4455 * RT tasks are offset by -200. Normal tasks are centered
4456 * around 0, value goes from -16 to +15.
4458 int task_prio(const struct task_struct *p)
4460 return p->prio - MAX_RT_PRIO;
4464 * idle_cpu - is a given CPU idle currently?
4465 * @cpu: the processor in question.
4467 * Return: 1 if the CPU is currently idle. 0 otherwise.
4469 int idle_cpu(int cpu)
4471 struct rq *rq = cpu_rq(cpu);
4473 if (rq->curr != rq->idle)
4480 if (!llist_empty(&rq->wake_list))
4488 * available_idle_cpu - is a given CPU idle for enqueuing work.
4489 * @cpu: the CPU in question.
4491 * Return: 1 if the CPU is currently idle. 0 otherwise.
4493 int available_idle_cpu(int cpu)
4498 if (vcpu_is_preempted(cpu))
4505 * idle_task - return the idle task for a given CPU.
4506 * @cpu: the processor in question.
4508 * Return: The idle task for the CPU @cpu.
4510 struct task_struct *idle_task(int cpu)
4512 return cpu_rq(cpu)->idle;
4516 * find_process_by_pid - find a process with a matching PID value.
4517 * @pid: the pid in question.
4519 * The task of @pid, if found. %NULL otherwise.
4521 static struct task_struct *find_process_by_pid(pid_t pid)
4523 return pid ? find_task_by_vpid(pid) : current;
4527 * sched_setparam() passes in -1 for its policy, to let the functions
4528 * it calls know not to change it.
4530 #define SETPARAM_POLICY -1
4532 static void __setscheduler_params(struct task_struct *p,
4533 const struct sched_attr *attr)
4535 int policy = attr->sched_policy;
4537 if (policy == SETPARAM_POLICY)
4542 if (dl_policy(policy))
4543 __setparam_dl(p, attr);
4544 else if (fair_policy(policy))
4545 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4548 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4549 * !rt_policy. Always setting this ensures that things like
4550 * getparam()/getattr() don't report silly values for !rt tasks.
4552 p->rt_priority = attr->sched_priority;
4553 p->normal_prio = normal_prio(p);
4554 set_load_weight(p, true);
4557 /* Actually do priority change: must hold pi & rq lock. */
4558 static void __setscheduler(struct rq *rq, struct task_struct *p,
4559 const struct sched_attr *attr, bool keep_boost)
4562 * If params can't change scheduling class changes aren't allowed
4565 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4568 __setscheduler_params(p, attr);
4571 * Keep a potential priority boosting if called from
4572 * sched_setscheduler().
4574 p->prio = normal_prio(p);
4576 p->prio = rt_effective_prio(p, p->prio);
4578 if (dl_prio(p->prio))
4579 p->sched_class = &dl_sched_class;
4580 else if (rt_prio(p->prio))
4581 p->sched_class = &rt_sched_class;
4583 p->sched_class = &fair_sched_class;
4587 * Check the target process has a UID that matches the current process's:
4589 static bool check_same_owner(struct task_struct *p)
4591 const struct cred *cred = current_cred(), *pcred;
4595 pcred = __task_cred(p);
4596 match = (uid_eq(cred->euid, pcred->euid) ||
4597 uid_eq(cred->euid, pcred->uid));
4602 static int __sched_setscheduler(struct task_struct *p,
4603 const struct sched_attr *attr,
4606 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4607 MAX_RT_PRIO - 1 - attr->sched_priority;
4608 int retval, oldprio, oldpolicy = -1, queued, running;
4609 int new_effective_prio, policy = attr->sched_policy;
4610 const struct sched_class *prev_class;
4613 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4616 /* The pi code expects interrupts enabled */
4617 BUG_ON(pi && in_interrupt());
4619 /* Double check policy once rq lock held: */
4621 reset_on_fork = p->sched_reset_on_fork;
4622 policy = oldpolicy = p->policy;
4624 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4626 if (!valid_policy(policy))
4630 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4634 * Valid priorities for SCHED_FIFO and SCHED_RR are
4635 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4636 * SCHED_BATCH and SCHED_IDLE is 0.
4638 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4639 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4641 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4642 (rt_policy(policy) != (attr->sched_priority != 0)))
4646 * Allow unprivileged RT tasks to decrease priority:
4648 if (user && !capable(CAP_SYS_NICE)) {
4649 if (fair_policy(policy)) {
4650 if (attr->sched_nice < task_nice(p) &&
4651 !can_nice(p, attr->sched_nice))
4655 if (rt_policy(policy)) {
4656 unsigned long rlim_rtprio =
4657 task_rlimit(p, RLIMIT_RTPRIO);
4659 /* Can't set/change the rt policy: */
4660 if (policy != p->policy && !rlim_rtprio)
4663 /* Can't increase priority: */
4664 if (attr->sched_priority > p->rt_priority &&
4665 attr->sched_priority > rlim_rtprio)
4670 * Can't set/change SCHED_DEADLINE policy at all for now
4671 * (safest behavior); in the future we would like to allow
4672 * unprivileged DL tasks to increase their relative deadline
4673 * or reduce their runtime (both ways reducing utilization)
4675 if (dl_policy(policy))
4679 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4680 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4682 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4683 if (!can_nice(p, task_nice(p)))
4687 /* Can't change other user's priorities: */
4688 if (!check_same_owner(p))
4691 /* Normal users shall not reset the sched_reset_on_fork flag: */
4692 if (p->sched_reset_on_fork && !reset_on_fork)
4697 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4700 retval = security_task_setscheduler(p);
4705 /* Update task specific "requested" clamps */
4706 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4707 retval = uclamp_validate(p, attr);
4716 * Make sure no PI-waiters arrive (or leave) while we are
4717 * changing the priority of the task:
4719 * To be able to change p->policy safely, the appropriate
4720 * runqueue lock must be held.
4722 rq = task_rq_lock(p, &rf);
4723 update_rq_clock(rq);
4726 * Changing the policy of the stop threads its a very bad idea:
4728 if (p == rq->stop) {
4734 * If not changing anything there's no need to proceed further,
4735 * but store a possible modification of reset_on_fork.
4737 if (unlikely(policy == p->policy)) {
4738 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4740 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4742 if (dl_policy(policy) && dl_param_changed(p, attr))
4744 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4747 p->sched_reset_on_fork = reset_on_fork;
4754 #ifdef CONFIG_RT_GROUP_SCHED
4756 * Do not allow realtime tasks into groups that have no runtime
4759 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4760 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4761 !task_group_is_autogroup(task_group(p))) {
4767 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4768 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4769 cpumask_t *span = rq->rd->span;
4772 * Don't allow tasks with an affinity mask smaller than
4773 * the entire root_domain to become SCHED_DEADLINE. We
4774 * will also fail if there's no bandwidth available.
4776 if (!cpumask_subset(span, p->cpus_ptr) ||
4777 rq->rd->dl_bw.bw == 0) {
4785 /* Re-check policy now with rq lock held: */
4786 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4787 policy = oldpolicy = -1;
4788 task_rq_unlock(rq, p, &rf);
4790 cpuset_read_unlock();
4795 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4796 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4799 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4804 p->sched_reset_on_fork = reset_on_fork;
4809 * Take priority boosted tasks into account. If the new
4810 * effective priority is unchanged, we just store the new
4811 * normal parameters and do not touch the scheduler class and
4812 * the runqueue. This will be done when the task deboost
4815 new_effective_prio = rt_effective_prio(p, newprio);
4816 if (new_effective_prio == oldprio)
4817 queue_flags &= ~DEQUEUE_MOVE;
4820 queued = task_on_rq_queued(p);
4821 running = task_current(rq, p);
4823 dequeue_task(rq, p, queue_flags);
4825 put_prev_task(rq, p);
4827 prev_class = p->sched_class;
4829 __setscheduler(rq, p, attr, pi);
4830 __setscheduler_uclamp(p, attr);
4834 * We enqueue to tail when the priority of a task is
4835 * increased (user space view).
4837 if (oldprio < p->prio)
4838 queue_flags |= ENQUEUE_HEAD;
4840 enqueue_task(rq, p, queue_flags);
4843 set_curr_task(rq, p);
4845 check_class_changed(rq, p, prev_class, oldprio);
4847 /* Avoid rq from going away on us: */
4849 task_rq_unlock(rq, p, &rf);
4852 cpuset_read_unlock();
4853 rt_mutex_adjust_pi(p);
4856 /* Run balance callbacks after we've adjusted the PI chain: */
4857 balance_callback(rq);
4863 task_rq_unlock(rq, p, &rf);
4865 cpuset_read_unlock();
4869 static int _sched_setscheduler(struct task_struct *p, int policy,
4870 const struct sched_param *param, bool check)
4872 struct sched_attr attr = {
4873 .sched_policy = policy,
4874 .sched_priority = param->sched_priority,
4875 .sched_nice = PRIO_TO_NICE(p->static_prio),
4878 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4879 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4880 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4881 policy &= ~SCHED_RESET_ON_FORK;
4882 attr.sched_policy = policy;
4885 return __sched_setscheduler(p, &attr, check, true);
4888 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4889 * @p: the task in question.
4890 * @policy: new policy.
4891 * @param: structure containing the new RT priority.
4893 * Return: 0 on success. An error code otherwise.
4895 * NOTE that the task may be already dead.
4897 int sched_setscheduler(struct task_struct *p, int policy,
4898 const struct sched_param *param)
4900 return _sched_setscheduler(p, policy, param, true);
4902 EXPORT_SYMBOL_GPL(sched_setscheduler);
4904 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4906 return __sched_setscheduler(p, attr, true, true);
4908 EXPORT_SYMBOL_GPL(sched_setattr);
4910 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4912 return __sched_setscheduler(p, attr, false, true);
4916 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4917 * @p: the task in question.
4918 * @policy: new policy.
4919 * @param: structure containing the new RT priority.
4921 * Just like sched_setscheduler, only don't bother checking if the
4922 * current context has permission. For example, this is needed in
4923 * stop_machine(): we create temporary high priority worker threads,
4924 * but our caller might not have that capability.
4926 * Return: 0 on success. An error code otherwise.
4928 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4929 const struct sched_param *param)
4931 return _sched_setscheduler(p, policy, param, false);
4933 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4936 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4938 struct sched_param lparam;
4939 struct task_struct *p;
4942 if (!param || pid < 0)
4944 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4949 p = find_process_by_pid(pid);
4955 retval = sched_setscheduler(p, policy, &lparam);
4963 * Mimics kernel/events/core.c perf_copy_attr().
4965 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4970 if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
4973 /* Zero the full structure, so that a short copy will be nice: */
4974 memset(attr, 0, sizeof(*attr));
4976 ret = get_user(size, &uattr->size);
4980 /* Bail out on silly large: */
4981 if (size > PAGE_SIZE)
4984 /* ABI compatibility quirk: */
4986 size = SCHED_ATTR_SIZE_VER0;
4988 if (size < SCHED_ATTR_SIZE_VER0)
4992 * If we're handed a bigger struct than we know of,
4993 * ensure all the unknown bits are 0 - i.e. new
4994 * user-space does not rely on any kernel feature
4995 * extensions we dont know about yet.
4997 if (size > sizeof(*attr)) {
4998 unsigned char __user *addr;
4999 unsigned char __user *end;
5002 addr = (void __user *)uattr + sizeof(*attr);
5003 end = (void __user *)uattr + size;
5005 for (; addr < end; addr++) {
5006 ret = get_user(val, addr);
5012 size = sizeof(*attr);
5015 ret = copy_from_user(attr, uattr, size);
5019 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5020 size < SCHED_ATTR_SIZE_VER1)
5024 * XXX: Do we want to be lenient like existing syscalls; or do we want
5025 * to be strict and return an error on out-of-bounds values?
5027 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5032 put_user(sizeof(*attr), &uattr->size);
5037 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5038 * @pid: the pid in question.
5039 * @policy: new policy.
5040 * @param: structure containing the new RT priority.
5042 * Return: 0 on success. An error code otherwise.
5044 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5049 return do_sched_setscheduler(pid, policy, param);
5053 * sys_sched_setparam - set/change the RT priority of a thread
5054 * @pid: the pid in question.
5055 * @param: structure containing the new RT priority.
5057 * Return: 0 on success. An error code otherwise.
5059 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5061 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5065 * sys_sched_setattr - same as above, but with extended sched_attr
5066 * @pid: the pid in question.
5067 * @uattr: structure containing the extended parameters.
5068 * @flags: for future extension.
5070 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5071 unsigned int, flags)
5073 struct sched_attr attr;
5074 struct task_struct *p;
5077 if (!uattr || pid < 0 || flags)
5080 retval = sched_copy_attr(uattr, &attr);
5084 if ((int)attr.sched_policy < 0)
5086 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5087 attr.sched_policy = SETPARAM_POLICY;
5091 p = find_process_by_pid(pid);
5097 retval = sched_setattr(p, &attr);
5105 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5106 * @pid: the pid in question.
5108 * Return: On success, the policy of the thread. Otherwise, a negative error
5111 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5113 struct task_struct *p;
5121 p = find_process_by_pid(pid);
5123 retval = security_task_getscheduler(p);
5126 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5133 * sys_sched_getparam - get the RT priority of a thread
5134 * @pid: the pid in question.
5135 * @param: structure containing the RT priority.
5137 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5140 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5142 struct sched_param lp = { .sched_priority = 0 };
5143 struct task_struct *p;
5146 if (!param || pid < 0)
5150 p = find_process_by_pid(pid);
5155 retval = security_task_getscheduler(p);
5159 if (task_has_rt_policy(p))
5160 lp.sched_priority = p->rt_priority;
5164 * This one might sleep, we cannot do it with a spinlock held ...
5166 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5175 static int sched_read_attr(struct sched_attr __user *uattr,
5176 struct sched_attr *attr,
5181 if (!access_ok(uattr, usize))
5185 * If we're handed a smaller struct than we know of,
5186 * ensure all the unknown bits are 0 - i.e. old
5187 * user-space does not get uncomplete information.
5189 if (usize < sizeof(*attr)) {
5190 unsigned char *addr;
5193 addr = (void *)attr + usize;
5194 end = (void *)attr + sizeof(*attr);
5196 for (; addr < end; addr++) {
5204 ret = copy_to_user(uattr, attr, attr->size);
5212 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5213 * @pid: the pid in question.
5214 * @uattr: structure containing the extended parameters.
5215 * @size: sizeof(attr) for fwd/bwd comp.
5216 * @flags: for future extension.
5218 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5219 unsigned int, size, unsigned int, flags)
5221 struct sched_attr attr = {
5222 .size = sizeof(struct sched_attr),
5224 struct task_struct *p;
5227 if (!uattr || pid < 0 || size > PAGE_SIZE ||
5228 size < SCHED_ATTR_SIZE_VER0 || flags)
5232 p = find_process_by_pid(pid);
5237 retval = security_task_getscheduler(p);
5241 attr.sched_policy = p->policy;
5242 if (p->sched_reset_on_fork)
5243 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5244 if (task_has_dl_policy(p))
5245 __getparam_dl(p, &attr);
5246 else if (task_has_rt_policy(p))
5247 attr.sched_priority = p->rt_priority;
5249 attr.sched_nice = task_nice(p);
5251 #ifdef CONFIG_UCLAMP_TASK
5252 attr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5253 attr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5258 retval = sched_read_attr(uattr, &attr, size);
5266 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5268 cpumask_var_t cpus_allowed, new_mask;
5269 struct task_struct *p;
5274 p = find_process_by_pid(pid);
5280 /* Prevent p going away */
5284 if (p->flags & PF_NO_SETAFFINITY) {
5288 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5292 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5294 goto out_free_cpus_allowed;
5297 if (!check_same_owner(p)) {
5299 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5301 goto out_free_new_mask;
5306 retval = security_task_setscheduler(p);
5308 goto out_free_new_mask;
5311 cpuset_cpus_allowed(p, cpus_allowed);
5312 cpumask_and(new_mask, in_mask, cpus_allowed);
5315 * Since bandwidth control happens on root_domain basis,
5316 * if admission test is enabled, we only admit -deadline
5317 * tasks allowed to run on all the CPUs in the task's
5321 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5323 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5326 goto out_free_new_mask;
5332 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5335 cpuset_cpus_allowed(p, cpus_allowed);
5336 if (!cpumask_subset(new_mask, cpus_allowed)) {
5338 * We must have raced with a concurrent cpuset
5339 * update. Just reset the cpus_allowed to the
5340 * cpuset's cpus_allowed
5342 cpumask_copy(new_mask, cpus_allowed);
5347 free_cpumask_var(new_mask);
5348 out_free_cpus_allowed:
5349 free_cpumask_var(cpus_allowed);
5355 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5356 struct cpumask *new_mask)
5358 if (len < cpumask_size())
5359 cpumask_clear(new_mask);
5360 else if (len > cpumask_size())
5361 len = cpumask_size();
5363 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5367 * sys_sched_setaffinity - set the CPU affinity of a process
5368 * @pid: pid of the process
5369 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5370 * @user_mask_ptr: user-space pointer to the new CPU mask
5372 * Return: 0 on success. An error code otherwise.
5374 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5375 unsigned long __user *, user_mask_ptr)
5377 cpumask_var_t new_mask;
5380 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5383 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5385 retval = sched_setaffinity(pid, new_mask);
5386 free_cpumask_var(new_mask);
5390 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5392 struct task_struct *p;
5393 unsigned long flags;
5399 p = find_process_by_pid(pid);
5403 retval = security_task_getscheduler(p);
5407 raw_spin_lock_irqsave(&p->pi_lock, flags);
5408 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5409 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5418 * sys_sched_getaffinity - get the CPU affinity of a process
5419 * @pid: pid of the process
5420 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5421 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5423 * Return: size of CPU mask copied to user_mask_ptr on success. An
5424 * error code otherwise.
5426 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5427 unsigned long __user *, user_mask_ptr)
5432 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5434 if (len & (sizeof(unsigned long)-1))
5437 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5440 ret = sched_getaffinity(pid, mask);
5442 unsigned int retlen = min(len, cpumask_size());
5444 if (copy_to_user(user_mask_ptr, mask, retlen))
5449 free_cpumask_var(mask);
5455 * sys_sched_yield - yield the current processor to other threads.
5457 * This function yields the current CPU to other tasks. If there are no
5458 * other threads running on this CPU then this function will return.
5462 static void do_sched_yield(void)
5467 rq = this_rq_lock_irq(&rf);
5469 schedstat_inc(rq->yld_count);
5470 current->sched_class->yield_task(rq);
5473 * Since we are going to call schedule() anyway, there's
5474 * no need to preempt or enable interrupts:
5478 sched_preempt_enable_no_resched();
5483 SYSCALL_DEFINE0(sched_yield)
5489 #ifndef CONFIG_PREEMPT
5490 int __sched _cond_resched(void)
5492 if (should_resched(0)) {
5493 preempt_schedule_common();
5499 EXPORT_SYMBOL(_cond_resched);
5503 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5504 * call schedule, and on return reacquire the lock.
5506 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5507 * operations here to prevent schedule() from being called twice (once via
5508 * spin_unlock(), once by hand).
5510 int __cond_resched_lock(spinlock_t *lock)
5512 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5515 lockdep_assert_held(lock);
5517 if (spin_needbreak(lock) || resched) {
5520 preempt_schedule_common();
5528 EXPORT_SYMBOL(__cond_resched_lock);
5531 * yield - yield the current processor to other threads.
5533 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5535 * The scheduler is at all times free to pick the calling task as the most
5536 * eligible task to run, if removing the yield() call from your code breaks
5537 * it, its already broken.
5539 * Typical broken usage is:
5544 * where one assumes that yield() will let 'the other' process run that will
5545 * make event true. If the current task is a SCHED_FIFO task that will never
5546 * happen. Never use yield() as a progress guarantee!!
5548 * If you want to use yield() to wait for something, use wait_event().
5549 * If you want to use yield() to be 'nice' for others, use cond_resched().
5550 * If you still want to use yield(), do not!
5552 void __sched yield(void)
5554 set_current_state(TASK_RUNNING);
5557 EXPORT_SYMBOL(yield);
5560 * yield_to - yield the current processor to another thread in
5561 * your thread group, or accelerate that thread toward the
5562 * processor it's on.
5564 * @preempt: whether task preemption is allowed or not
5566 * It's the caller's job to ensure that the target task struct
5567 * can't go away on us before we can do any checks.
5570 * true (>0) if we indeed boosted the target task.
5571 * false (0) if we failed to boost the target.
5572 * -ESRCH if there's no task to yield to.
5574 int __sched yield_to(struct task_struct *p, bool preempt)
5576 struct task_struct *curr = current;
5577 struct rq *rq, *p_rq;
5578 unsigned long flags;
5581 local_irq_save(flags);
5587 * If we're the only runnable task on the rq and target rq also
5588 * has only one task, there's absolutely no point in yielding.
5590 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5595 double_rq_lock(rq, p_rq);
5596 if (task_rq(p) != p_rq) {
5597 double_rq_unlock(rq, p_rq);
5601 if (!curr->sched_class->yield_to_task)
5604 if (curr->sched_class != p->sched_class)
5607 if (task_running(p_rq, p) || p->state)
5610 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5612 schedstat_inc(rq->yld_count);
5614 * Make p's CPU reschedule; pick_next_entity takes care of
5617 if (preempt && rq != p_rq)
5622 double_rq_unlock(rq, p_rq);
5624 local_irq_restore(flags);
5631 EXPORT_SYMBOL_GPL(yield_to);
5633 int io_schedule_prepare(void)
5635 int old_iowait = current->in_iowait;
5637 current->in_iowait = 1;
5638 blk_schedule_flush_plug(current);
5643 void io_schedule_finish(int token)
5645 current->in_iowait = token;
5649 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5650 * that process accounting knows that this is a task in IO wait state.
5652 long __sched io_schedule_timeout(long timeout)
5657 token = io_schedule_prepare();
5658 ret = schedule_timeout(timeout);
5659 io_schedule_finish(token);
5663 EXPORT_SYMBOL(io_schedule_timeout);
5665 void __sched io_schedule(void)
5669 token = io_schedule_prepare();
5671 io_schedule_finish(token);
5673 EXPORT_SYMBOL(io_schedule);
5676 * sys_sched_get_priority_max - return maximum RT priority.
5677 * @policy: scheduling class.
5679 * Return: On success, this syscall returns the maximum
5680 * rt_priority that can be used by a given scheduling class.
5681 * On failure, a negative error code is returned.
5683 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5690 ret = MAX_USER_RT_PRIO-1;
5692 case SCHED_DEADLINE:
5703 * sys_sched_get_priority_min - return minimum RT priority.
5704 * @policy: scheduling class.
5706 * Return: On success, this syscall returns the minimum
5707 * rt_priority that can be used by a given scheduling class.
5708 * On failure, a negative error code is returned.
5710 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5719 case SCHED_DEADLINE:
5728 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5730 struct task_struct *p;
5731 unsigned int time_slice;
5741 p = find_process_by_pid(pid);
5745 retval = security_task_getscheduler(p);
5749 rq = task_rq_lock(p, &rf);
5751 if (p->sched_class->get_rr_interval)
5752 time_slice = p->sched_class->get_rr_interval(rq, p);
5753 task_rq_unlock(rq, p, &rf);
5756 jiffies_to_timespec64(time_slice, t);
5765 * sys_sched_rr_get_interval - return the default timeslice of a process.
5766 * @pid: pid of the process.
5767 * @interval: userspace pointer to the timeslice value.
5769 * this syscall writes the default timeslice value of a given process
5770 * into the user-space timespec buffer. A value of '0' means infinity.
5772 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5775 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5776 struct __kernel_timespec __user *, interval)
5778 struct timespec64 t;
5779 int retval = sched_rr_get_interval(pid, &t);
5782 retval = put_timespec64(&t, interval);
5787 #ifdef CONFIG_COMPAT_32BIT_TIME
5788 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5789 struct old_timespec32 __user *, interval)
5791 struct timespec64 t;
5792 int retval = sched_rr_get_interval(pid, &t);
5795 retval = put_old_timespec32(&t, interval);
5800 void sched_show_task(struct task_struct *p)
5802 unsigned long free = 0;
5805 if (!try_get_task_stack(p))
5808 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5810 if (p->state == TASK_RUNNING)
5811 printk(KERN_CONT " running task ");
5812 #ifdef CONFIG_DEBUG_STACK_USAGE
5813 free = stack_not_used(p);
5818 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5820 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5821 task_pid_nr(p), ppid,
5822 (unsigned long)task_thread_info(p)->flags);
5824 print_worker_info(KERN_INFO, p);
5825 show_stack(p, NULL);
5828 EXPORT_SYMBOL_GPL(sched_show_task);
5831 state_filter_match(unsigned long state_filter, struct task_struct *p)
5833 /* no filter, everything matches */
5837 /* filter, but doesn't match */
5838 if (!(p->state & state_filter))
5842 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5845 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5852 void show_state_filter(unsigned long state_filter)
5854 struct task_struct *g, *p;
5856 #if BITS_PER_LONG == 32
5858 " task PC stack pid father\n");
5861 " task PC stack pid father\n");
5864 for_each_process_thread(g, p) {
5866 * reset the NMI-timeout, listing all files on a slow
5867 * console might take a lot of time:
5868 * Also, reset softlockup watchdogs on all CPUs, because
5869 * another CPU might be blocked waiting for us to process
5872 touch_nmi_watchdog();
5873 touch_all_softlockup_watchdogs();
5874 if (state_filter_match(state_filter, p))
5878 #ifdef CONFIG_SCHED_DEBUG
5880 sysrq_sched_debug_show();
5884 * Only show locks if all tasks are dumped:
5887 debug_show_all_locks();
5891 * init_idle - set up an idle thread for a given CPU
5892 * @idle: task in question
5893 * @cpu: CPU the idle task belongs to
5895 * NOTE: this function does not set the idle thread's NEED_RESCHED
5896 * flag, to make booting more robust.
5898 void init_idle(struct task_struct *idle, int cpu)
5900 struct rq *rq = cpu_rq(cpu);
5901 unsigned long flags;
5903 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5904 raw_spin_lock(&rq->lock);
5906 __sched_fork(0, idle);
5907 idle->state = TASK_RUNNING;
5908 idle->se.exec_start = sched_clock();
5909 idle->flags |= PF_IDLE;
5911 kasan_unpoison_task_stack(idle);
5915 * Its possible that init_idle() gets called multiple times on a task,
5916 * in that case do_set_cpus_allowed() will not do the right thing.
5918 * And since this is boot we can forgo the serialization.
5920 set_cpus_allowed_common(idle, cpumask_of(cpu));
5923 * We're having a chicken and egg problem, even though we are
5924 * holding rq->lock, the CPU isn't yet set to this CPU so the
5925 * lockdep check in task_group() will fail.
5927 * Similar case to sched_fork(). / Alternatively we could
5928 * use task_rq_lock() here and obtain the other rq->lock.
5933 __set_task_cpu(idle, cpu);
5936 rq->curr = rq->idle = idle;
5937 idle->on_rq = TASK_ON_RQ_QUEUED;
5941 raw_spin_unlock(&rq->lock);
5942 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5944 /* Set the preempt count _outside_ the spinlocks! */
5945 init_idle_preempt_count(idle, cpu);
5948 * The idle tasks have their own, simple scheduling class:
5950 idle->sched_class = &idle_sched_class;
5951 ftrace_graph_init_idle_task(idle, cpu);
5952 vtime_init_idle(idle, cpu);
5954 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5960 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5961 const struct cpumask *trial)
5965 if (!cpumask_weight(cur))
5968 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5973 int task_can_attach(struct task_struct *p,
5974 const struct cpumask *cs_cpus_allowed)
5979 * Kthreads which disallow setaffinity shouldn't be moved
5980 * to a new cpuset; we don't want to change their CPU
5981 * affinity and isolating such threads by their set of
5982 * allowed nodes is unnecessary. Thus, cpusets are not
5983 * applicable for such threads. This prevents checking for
5984 * success of set_cpus_allowed_ptr() on all attached tasks
5985 * before cpus_mask may be changed.
5987 if (p->flags & PF_NO_SETAFFINITY) {
5992 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5994 ret = dl_task_can_attach(p, cs_cpus_allowed);
6000 bool sched_smp_initialized __read_mostly;
6002 #ifdef CONFIG_NUMA_BALANCING
6003 /* Migrate current task p to target_cpu */
6004 int migrate_task_to(struct task_struct *p, int target_cpu)
6006 struct migration_arg arg = { p, target_cpu };
6007 int curr_cpu = task_cpu(p);
6009 if (curr_cpu == target_cpu)
6012 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6015 /* TODO: This is not properly updating schedstats */
6017 trace_sched_move_numa(p, curr_cpu, target_cpu);
6018 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6022 * Requeue a task on a given node and accurately track the number of NUMA
6023 * tasks on the runqueues
6025 void sched_setnuma(struct task_struct *p, int nid)
6027 bool queued, running;
6031 rq = task_rq_lock(p, &rf);
6032 queued = task_on_rq_queued(p);
6033 running = task_current(rq, p);
6036 dequeue_task(rq, p, DEQUEUE_SAVE);
6038 put_prev_task(rq, p);
6040 p->numa_preferred_nid = nid;
6043 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6045 set_curr_task(rq, p);
6046 task_rq_unlock(rq, p, &rf);
6048 #endif /* CONFIG_NUMA_BALANCING */
6050 #ifdef CONFIG_HOTPLUG_CPU
6052 * Ensure that the idle task is using init_mm right before its CPU goes
6055 void idle_task_exit(void)
6057 struct mm_struct *mm = current->active_mm;
6059 BUG_ON(cpu_online(smp_processor_id()));
6061 if (mm != &init_mm) {
6062 switch_mm(mm, &init_mm, current);
6063 current->active_mm = &init_mm;
6064 finish_arch_post_lock_switch();
6070 * Since this CPU is going 'away' for a while, fold any nr_active delta
6071 * we might have. Assumes we're called after migrate_tasks() so that the
6072 * nr_active count is stable. We need to take the teardown thread which
6073 * is calling this into account, so we hand in adjust = 1 to the load
6076 * Also see the comment "Global load-average calculations".
6078 static void calc_load_migrate(struct rq *rq)
6080 long delta = calc_load_fold_active(rq, 1);
6082 atomic_long_add(delta, &calc_load_tasks);
6085 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
6089 static const struct sched_class fake_sched_class = {
6090 .put_prev_task = put_prev_task_fake,
6093 static struct task_struct fake_task = {
6095 * Avoid pull_{rt,dl}_task()
6097 .prio = MAX_PRIO + 1,
6098 .sched_class = &fake_sched_class,
6102 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6103 * try_to_wake_up()->select_task_rq().
6105 * Called with rq->lock held even though we'er in stop_machine() and
6106 * there's no concurrency possible, we hold the required locks anyway
6107 * because of lock validation efforts.
6109 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6111 struct rq *rq = dead_rq;
6112 struct task_struct *next, *stop = rq->stop;
6113 struct rq_flags orf = *rf;
6117 * Fudge the rq selection such that the below task selection loop
6118 * doesn't get stuck on the currently eligible stop task.
6120 * We're currently inside stop_machine() and the rq is either stuck
6121 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6122 * either way we should never end up calling schedule() until we're
6128 * put_prev_task() and pick_next_task() sched
6129 * class method both need to have an up-to-date
6130 * value of rq->clock[_task]
6132 update_rq_clock(rq);
6136 * There's this thread running, bail when that's the only
6139 if (rq->nr_running == 1)
6143 * pick_next_task() assumes pinned rq->lock:
6145 next = pick_next_task(rq, &fake_task, rf);
6147 put_prev_task(rq, next);
6150 * Rules for changing task_struct::cpus_mask are holding
6151 * both pi_lock and rq->lock, such that holding either
6152 * stabilizes the mask.
6154 * Drop rq->lock is not quite as disastrous as it usually is
6155 * because !cpu_active at this point, which means load-balance
6156 * will not interfere. Also, stop-machine.
6159 raw_spin_lock(&next->pi_lock);
6163 * Since we're inside stop-machine, _nothing_ should have
6164 * changed the task, WARN if weird stuff happened, because in
6165 * that case the above rq->lock drop is a fail too.
6167 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6168 raw_spin_unlock(&next->pi_lock);
6172 /* Find suitable destination for @next, with force if needed. */
6173 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6174 rq = __migrate_task(rq, rf, next, dest_cpu);
6175 if (rq != dead_rq) {
6181 raw_spin_unlock(&next->pi_lock);
6186 #endif /* CONFIG_HOTPLUG_CPU */
6188 void set_rq_online(struct rq *rq)
6191 const struct sched_class *class;
6193 cpumask_set_cpu(rq->cpu, rq->rd->online);
6196 for_each_class(class) {
6197 if (class->rq_online)
6198 class->rq_online(rq);
6203 void set_rq_offline(struct rq *rq)
6206 const struct sched_class *class;
6208 for_each_class(class) {
6209 if (class->rq_offline)
6210 class->rq_offline(rq);
6213 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6219 * used to mark begin/end of suspend/resume:
6221 static int num_cpus_frozen;
6224 * Update cpusets according to cpu_active mask. If cpusets are
6225 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6226 * around partition_sched_domains().
6228 * If we come here as part of a suspend/resume, don't touch cpusets because we
6229 * want to restore it back to its original state upon resume anyway.
6231 static void cpuset_cpu_active(void)
6233 if (cpuhp_tasks_frozen) {
6235 * num_cpus_frozen tracks how many CPUs are involved in suspend
6236 * resume sequence. As long as this is not the last online
6237 * operation in the resume sequence, just build a single sched
6238 * domain, ignoring cpusets.
6240 partition_sched_domains(1, NULL, NULL);
6241 if (--num_cpus_frozen)
6244 * This is the last CPU online operation. So fall through and
6245 * restore the original sched domains by considering the
6246 * cpuset configurations.
6248 cpuset_force_rebuild();
6250 cpuset_update_active_cpus();
6253 static int cpuset_cpu_inactive(unsigned int cpu)
6255 if (!cpuhp_tasks_frozen) {
6256 if (dl_cpu_busy(cpu))
6258 cpuset_update_active_cpus();
6261 partition_sched_domains(1, NULL, NULL);
6266 int sched_cpu_activate(unsigned int cpu)
6268 struct rq *rq = cpu_rq(cpu);
6271 #ifdef CONFIG_SCHED_SMT
6273 * When going up, increment the number of cores with SMT present.
6275 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6276 static_branch_inc_cpuslocked(&sched_smt_present);
6278 set_cpu_active(cpu, true);
6280 if (sched_smp_initialized) {
6281 sched_domains_numa_masks_set(cpu);
6282 cpuset_cpu_active();
6286 * Put the rq online, if not already. This happens:
6288 * 1) In the early boot process, because we build the real domains
6289 * after all CPUs have been brought up.
6291 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6294 rq_lock_irqsave(rq, &rf);
6296 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6299 rq_unlock_irqrestore(rq, &rf);
6301 update_max_interval();
6306 int sched_cpu_deactivate(unsigned int cpu)
6310 set_cpu_active(cpu, false);
6312 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6313 * users of this state to go away such that all new such users will
6316 * Do sync before park smpboot threads to take care the rcu boost case.
6320 #ifdef CONFIG_SCHED_SMT
6322 * When going down, decrement the number of cores with SMT present.
6324 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6325 static_branch_dec_cpuslocked(&sched_smt_present);
6328 if (!sched_smp_initialized)
6331 ret = cpuset_cpu_inactive(cpu);
6333 set_cpu_active(cpu, true);
6336 sched_domains_numa_masks_clear(cpu);
6340 static void sched_rq_cpu_starting(unsigned int cpu)
6342 struct rq *rq = cpu_rq(cpu);
6344 rq->calc_load_update = calc_load_update;
6345 update_max_interval();
6348 int sched_cpu_starting(unsigned int cpu)
6350 sched_rq_cpu_starting(cpu);
6351 sched_tick_start(cpu);
6355 #ifdef CONFIG_HOTPLUG_CPU
6356 int sched_cpu_dying(unsigned int cpu)
6358 struct rq *rq = cpu_rq(cpu);
6361 /* Handle pending wakeups and then migrate everything off */
6362 sched_ttwu_pending();
6363 sched_tick_stop(cpu);
6365 rq_lock_irqsave(rq, &rf);
6367 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6370 migrate_tasks(rq, &rf);
6371 BUG_ON(rq->nr_running != 1);
6372 rq_unlock_irqrestore(rq, &rf);
6374 calc_load_migrate(rq);
6375 update_max_interval();
6376 nohz_balance_exit_idle(rq);
6382 void __init sched_init_smp(void)
6387 * There's no userspace yet to cause hotplug operations; hence all the
6388 * CPU masks are stable and all blatant races in the below code cannot
6391 mutex_lock(&sched_domains_mutex);
6392 sched_init_domains(cpu_active_mask);
6393 mutex_unlock(&sched_domains_mutex);
6395 /* Move init over to a non-isolated CPU */
6396 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6398 sched_init_granularity();
6400 init_sched_rt_class();
6401 init_sched_dl_class();
6403 sched_smp_initialized = true;
6406 static int __init migration_init(void)
6408 sched_cpu_starting(smp_processor_id());
6411 early_initcall(migration_init);
6414 void __init sched_init_smp(void)
6416 sched_init_granularity();
6418 #endif /* CONFIG_SMP */
6420 int in_sched_functions(unsigned long addr)
6422 return in_lock_functions(addr) ||
6423 (addr >= (unsigned long)__sched_text_start
6424 && addr < (unsigned long)__sched_text_end);
6427 #ifdef CONFIG_CGROUP_SCHED
6429 * Default task group.
6430 * Every task in system belongs to this group at bootup.
6432 struct task_group root_task_group;
6433 LIST_HEAD(task_groups);
6435 /* Cacheline aligned slab cache for task_group */
6436 static struct kmem_cache *task_group_cache __read_mostly;
6439 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6440 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6442 void __init sched_init(void)
6444 unsigned long ptr = 0;
6449 #ifdef CONFIG_FAIR_GROUP_SCHED
6450 ptr += 2 * nr_cpu_ids * sizeof(void **);
6452 #ifdef CONFIG_RT_GROUP_SCHED
6453 ptr += 2 * nr_cpu_ids * sizeof(void **);
6456 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6458 #ifdef CONFIG_FAIR_GROUP_SCHED
6459 root_task_group.se = (struct sched_entity **)ptr;
6460 ptr += nr_cpu_ids * sizeof(void **);
6462 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6463 ptr += nr_cpu_ids * sizeof(void **);
6465 #endif /* CONFIG_FAIR_GROUP_SCHED */
6466 #ifdef CONFIG_RT_GROUP_SCHED
6467 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6468 ptr += nr_cpu_ids * sizeof(void **);
6470 root_task_group.rt_rq = (struct rt_rq **)ptr;
6471 ptr += nr_cpu_ids * sizeof(void **);
6473 #endif /* CONFIG_RT_GROUP_SCHED */
6475 #ifdef CONFIG_CPUMASK_OFFSTACK
6476 for_each_possible_cpu(i) {
6477 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6478 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6479 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6480 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6482 #endif /* CONFIG_CPUMASK_OFFSTACK */
6484 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6485 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6488 init_defrootdomain();
6491 #ifdef CONFIG_RT_GROUP_SCHED
6492 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6493 global_rt_period(), global_rt_runtime());
6494 #endif /* CONFIG_RT_GROUP_SCHED */
6496 #ifdef CONFIG_CGROUP_SCHED
6497 task_group_cache = KMEM_CACHE(task_group, 0);
6499 list_add(&root_task_group.list, &task_groups);
6500 INIT_LIST_HEAD(&root_task_group.children);
6501 INIT_LIST_HEAD(&root_task_group.siblings);
6502 autogroup_init(&init_task);
6503 #endif /* CONFIG_CGROUP_SCHED */
6505 for_each_possible_cpu(i) {
6509 raw_spin_lock_init(&rq->lock);
6511 rq->calc_load_active = 0;
6512 rq->calc_load_update = jiffies + LOAD_FREQ;
6513 init_cfs_rq(&rq->cfs);
6514 init_rt_rq(&rq->rt);
6515 init_dl_rq(&rq->dl);
6516 #ifdef CONFIG_FAIR_GROUP_SCHED
6517 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6518 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6519 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6521 * How much CPU bandwidth does root_task_group get?
6523 * In case of task-groups formed thr' the cgroup filesystem, it
6524 * gets 100% of the CPU resources in the system. This overall
6525 * system CPU resource is divided among the tasks of
6526 * root_task_group and its child task-groups in a fair manner,
6527 * based on each entity's (task or task-group's) weight
6528 * (se->load.weight).
6530 * In other words, if root_task_group has 10 tasks of weight
6531 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6532 * then A0's share of the CPU resource is:
6534 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6536 * We achieve this by letting root_task_group's tasks sit
6537 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6539 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6540 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6541 #endif /* CONFIG_FAIR_GROUP_SCHED */
6543 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6544 #ifdef CONFIG_RT_GROUP_SCHED
6545 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6550 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6551 rq->balance_callback = NULL;
6552 rq->active_balance = 0;
6553 rq->next_balance = jiffies;
6558 rq->avg_idle = 2*sysctl_sched_migration_cost;
6559 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6561 INIT_LIST_HEAD(&rq->cfs_tasks);
6563 rq_attach_root(rq, &def_root_domain);
6564 #ifdef CONFIG_NO_HZ_COMMON
6565 rq->last_load_update_tick = jiffies;
6566 rq->last_blocked_load_update_tick = jiffies;
6567 atomic_set(&rq->nohz_flags, 0);
6569 #endif /* CONFIG_SMP */
6571 atomic_set(&rq->nr_iowait, 0);
6574 set_load_weight(&init_task, false);
6577 * The boot idle thread does lazy MMU switching as well:
6580 enter_lazy_tlb(&init_mm, current);
6583 * Make us the idle thread. Technically, schedule() should not be
6584 * called from this thread, however somewhere below it might be,
6585 * but because we are the idle thread, we just pick up running again
6586 * when this runqueue becomes "idle".
6588 init_idle(current, smp_processor_id());
6590 calc_load_update = jiffies + LOAD_FREQ;
6593 idle_thread_set_boot_cpu();
6595 init_sched_fair_class();
6603 scheduler_running = 1;
6606 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6607 static inline int preempt_count_equals(int preempt_offset)
6609 int nested = preempt_count() + rcu_preempt_depth();
6611 return (nested == preempt_offset);
6614 void __might_sleep(const char *file, int line, int preempt_offset)
6617 * Blocking primitives will set (and therefore destroy) current->state,
6618 * since we will exit with TASK_RUNNING make sure we enter with it,
6619 * otherwise we will destroy state.
6621 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6622 "do not call blocking ops when !TASK_RUNNING; "
6623 "state=%lx set at [<%p>] %pS\n",
6625 (void *)current->task_state_change,
6626 (void *)current->task_state_change);
6628 ___might_sleep(file, line, preempt_offset);
6630 EXPORT_SYMBOL(__might_sleep);
6632 void ___might_sleep(const char *file, int line, int preempt_offset)
6634 /* Ratelimiting timestamp: */
6635 static unsigned long prev_jiffy;
6637 unsigned long preempt_disable_ip;
6639 /* WARN_ON_ONCE() by default, no rate limit required: */
6642 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6643 !is_idle_task(current)) ||
6644 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6648 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6650 prev_jiffy = jiffies;
6652 /* Save this before calling printk(), since that will clobber it: */
6653 preempt_disable_ip = get_preempt_disable_ip(current);
6656 "BUG: sleeping function called from invalid context at %s:%d\n",
6659 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6660 in_atomic(), irqs_disabled(),
6661 current->pid, current->comm);
6663 if (task_stack_end_corrupted(current))
6664 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6666 debug_show_held_locks(current);
6667 if (irqs_disabled())
6668 print_irqtrace_events(current);
6669 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6670 && !preempt_count_equals(preempt_offset)) {
6671 pr_err("Preemption disabled at:");
6672 print_ip_sym(preempt_disable_ip);
6676 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6678 EXPORT_SYMBOL(___might_sleep);
6680 void __cant_sleep(const char *file, int line, int preempt_offset)
6682 static unsigned long prev_jiffy;
6684 if (irqs_disabled())
6687 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6690 if (preempt_count() > preempt_offset)
6693 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6695 prev_jiffy = jiffies;
6697 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6698 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6699 in_atomic(), irqs_disabled(),
6700 current->pid, current->comm);
6702 debug_show_held_locks(current);
6704 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6706 EXPORT_SYMBOL_GPL(__cant_sleep);
6709 #ifdef CONFIG_MAGIC_SYSRQ
6710 void normalize_rt_tasks(void)
6712 struct task_struct *g, *p;
6713 struct sched_attr attr = {
6714 .sched_policy = SCHED_NORMAL,
6717 read_lock(&tasklist_lock);
6718 for_each_process_thread(g, p) {
6720 * Only normalize user tasks:
6722 if (p->flags & PF_KTHREAD)
6725 p->se.exec_start = 0;
6726 schedstat_set(p->se.statistics.wait_start, 0);
6727 schedstat_set(p->se.statistics.sleep_start, 0);
6728 schedstat_set(p->se.statistics.block_start, 0);
6730 if (!dl_task(p) && !rt_task(p)) {
6732 * Renice negative nice level userspace
6735 if (task_nice(p) < 0)
6736 set_user_nice(p, 0);
6740 __sched_setscheduler(p, &attr, false, false);
6742 read_unlock(&tasklist_lock);
6745 #endif /* CONFIG_MAGIC_SYSRQ */
6747 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6749 * These functions are only useful for the IA64 MCA handling, or kdb.
6751 * They can only be called when the whole system has been
6752 * stopped - every CPU needs to be quiescent, and no scheduling
6753 * activity can take place. Using them for anything else would
6754 * be a serious bug, and as a result, they aren't even visible
6755 * under any other configuration.
6759 * curr_task - return the current task for a given CPU.
6760 * @cpu: the processor in question.
6762 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6764 * Return: The current task for @cpu.
6766 struct task_struct *curr_task(int cpu)
6768 return cpu_curr(cpu);
6771 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6775 * ia64_set_curr_task - set the current task for a given CPU.
6776 * @cpu: the processor in question.
6777 * @p: the task pointer to set.
6779 * Description: This function must only be used when non-maskable interrupts
6780 * are serviced on a separate stack. It allows the architecture to switch the
6781 * notion of the current task on a CPU in a non-blocking manner. This function
6782 * must be called with all CPU's synchronized, and interrupts disabled, the
6783 * and caller must save the original value of the current task (see
6784 * curr_task() above) and restore that value before reenabling interrupts and
6785 * re-starting the system.
6787 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6789 void ia64_set_curr_task(int cpu, struct task_struct *p)
6796 #ifdef CONFIG_CGROUP_SCHED
6797 /* task_group_lock serializes the addition/removal of task groups */
6798 static DEFINE_SPINLOCK(task_group_lock);
6800 static void sched_free_group(struct task_group *tg)
6802 free_fair_sched_group(tg);
6803 free_rt_sched_group(tg);
6805 kmem_cache_free(task_group_cache, tg);
6808 /* allocate runqueue etc for a new task group */
6809 struct task_group *sched_create_group(struct task_group *parent)
6811 struct task_group *tg;
6813 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6815 return ERR_PTR(-ENOMEM);
6817 if (!alloc_fair_sched_group(tg, parent))
6820 if (!alloc_rt_sched_group(tg, parent))
6826 sched_free_group(tg);
6827 return ERR_PTR(-ENOMEM);
6830 void sched_online_group(struct task_group *tg, struct task_group *parent)
6832 unsigned long flags;
6834 spin_lock_irqsave(&task_group_lock, flags);
6835 list_add_rcu(&tg->list, &task_groups);
6837 /* Root should already exist: */
6840 tg->parent = parent;
6841 INIT_LIST_HEAD(&tg->children);
6842 list_add_rcu(&tg->siblings, &parent->children);
6843 spin_unlock_irqrestore(&task_group_lock, flags);
6845 online_fair_sched_group(tg);
6848 /* rcu callback to free various structures associated with a task group */
6849 static void sched_free_group_rcu(struct rcu_head *rhp)
6851 /* Now it should be safe to free those cfs_rqs: */
6852 sched_free_group(container_of(rhp, struct task_group, rcu));
6855 void sched_destroy_group(struct task_group *tg)
6857 /* Wait for possible concurrent references to cfs_rqs complete: */
6858 call_rcu(&tg->rcu, sched_free_group_rcu);
6861 void sched_offline_group(struct task_group *tg)
6863 unsigned long flags;
6865 /* End participation in shares distribution: */
6866 unregister_fair_sched_group(tg);
6868 spin_lock_irqsave(&task_group_lock, flags);
6869 list_del_rcu(&tg->list);
6870 list_del_rcu(&tg->siblings);
6871 spin_unlock_irqrestore(&task_group_lock, flags);
6874 static void sched_change_group(struct task_struct *tsk, int type)
6876 struct task_group *tg;
6879 * All callers are synchronized by task_rq_lock(); we do not use RCU
6880 * which is pointless here. Thus, we pass "true" to task_css_check()
6881 * to prevent lockdep warnings.
6883 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6884 struct task_group, css);
6885 tg = autogroup_task_group(tsk, tg);
6886 tsk->sched_task_group = tg;
6888 #ifdef CONFIG_FAIR_GROUP_SCHED
6889 if (tsk->sched_class->task_change_group)
6890 tsk->sched_class->task_change_group(tsk, type);
6893 set_task_rq(tsk, task_cpu(tsk));
6897 * Change task's runqueue when it moves between groups.
6899 * The caller of this function should have put the task in its new group by
6900 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6903 void sched_move_task(struct task_struct *tsk)
6905 int queued, running, queue_flags =
6906 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6910 rq = task_rq_lock(tsk, &rf);
6911 update_rq_clock(rq);
6913 running = task_current(rq, tsk);
6914 queued = task_on_rq_queued(tsk);
6917 dequeue_task(rq, tsk, queue_flags);
6919 put_prev_task(rq, tsk);
6921 sched_change_group(tsk, TASK_MOVE_GROUP);
6924 enqueue_task(rq, tsk, queue_flags);
6926 set_curr_task(rq, tsk);
6928 task_rq_unlock(rq, tsk, &rf);
6931 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6933 return css ? container_of(css, struct task_group, css) : NULL;
6936 static struct cgroup_subsys_state *
6937 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6939 struct task_group *parent = css_tg(parent_css);
6940 struct task_group *tg;
6943 /* This is early initialization for the top cgroup */
6944 return &root_task_group.css;
6947 tg = sched_create_group(parent);
6949 return ERR_PTR(-ENOMEM);
6954 /* Expose task group only after completing cgroup initialization */
6955 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6957 struct task_group *tg = css_tg(css);
6958 struct task_group *parent = css_tg(css->parent);
6961 sched_online_group(tg, parent);
6965 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6967 struct task_group *tg = css_tg(css);
6969 sched_offline_group(tg);
6972 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6974 struct task_group *tg = css_tg(css);
6977 * Relies on the RCU grace period between css_released() and this.
6979 sched_free_group(tg);
6983 * This is called before wake_up_new_task(), therefore we really only
6984 * have to set its group bits, all the other stuff does not apply.
6986 static void cpu_cgroup_fork(struct task_struct *task)
6991 rq = task_rq_lock(task, &rf);
6993 update_rq_clock(rq);
6994 sched_change_group(task, TASK_SET_GROUP);
6996 task_rq_unlock(rq, task, &rf);
6999 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7001 struct task_struct *task;
7002 struct cgroup_subsys_state *css;
7005 cgroup_taskset_for_each(task, css, tset) {
7006 #ifdef CONFIG_RT_GROUP_SCHED
7007 if (!sched_rt_can_attach(css_tg(css), task))
7011 * Serialize against wake_up_new_task() such that if its
7012 * running, we're sure to observe its full state.
7014 raw_spin_lock_irq(&task->pi_lock);
7016 * Avoid calling sched_move_task() before wake_up_new_task()
7017 * has happened. This would lead to problems with PELT, due to
7018 * move wanting to detach+attach while we're not attached yet.
7020 if (task->state == TASK_NEW)
7022 raw_spin_unlock_irq(&task->pi_lock);
7030 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7032 struct task_struct *task;
7033 struct cgroup_subsys_state *css;
7035 cgroup_taskset_for_each(task, css, tset)
7036 sched_move_task(task);
7039 #ifdef CONFIG_FAIR_GROUP_SCHED
7040 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7041 struct cftype *cftype, u64 shareval)
7043 if (shareval > scale_load_down(ULONG_MAX))
7044 shareval = MAX_SHARES;
7045 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7048 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7051 struct task_group *tg = css_tg(css);
7053 return (u64) scale_load_down(tg->shares);
7056 #ifdef CONFIG_CFS_BANDWIDTH
7057 static DEFINE_MUTEX(cfs_constraints_mutex);
7059 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7060 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7062 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7064 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7066 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7067 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7069 if (tg == &root_task_group)
7073 * Ensure we have at some amount of bandwidth every period. This is
7074 * to prevent reaching a state of large arrears when throttled via
7075 * entity_tick() resulting in prolonged exit starvation.
7077 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7081 * Likewise, bound things on the otherside by preventing insane quota
7082 * periods. This also allows us to normalize in computing quota
7085 if (period > max_cfs_quota_period)
7089 * Prevent race between setting of cfs_rq->runtime_enabled and
7090 * unthrottle_offline_cfs_rqs().
7093 mutex_lock(&cfs_constraints_mutex);
7094 ret = __cfs_schedulable(tg, period, quota);
7098 runtime_enabled = quota != RUNTIME_INF;
7099 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7101 * If we need to toggle cfs_bandwidth_used, off->on must occur
7102 * before making related changes, and on->off must occur afterwards
7104 if (runtime_enabled && !runtime_was_enabled)
7105 cfs_bandwidth_usage_inc();
7106 raw_spin_lock_irq(&cfs_b->lock);
7107 cfs_b->period = ns_to_ktime(period);
7108 cfs_b->quota = quota;
7110 __refill_cfs_bandwidth_runtime(cfs_b);
7112 /* Restart the period timer (if active) to handle new period expiry: */
7113 if (runtime_enabled)
7114 start_cfs_bandwidth(cfs_b);
7116 raw_spin_unlock_irq(&cfs_b->lock);
7118 for_each_online_cpu(i) {
7119 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7120 struct rq *rq = cfs_rq->rq;
7123 rq_lock_irq(rq, &rf);
7124 cfs_rq->runtime_enabled = runtime_enabled;
7125 cfs_rq->runtime_remaining = 0;
7127 if (cfs_rq->throttled)
7128 unthrottle_cfs_rq(cfs_rq);
7129 rq_unlock_irq(rq, &rf);
7131 if (runtime_was_enabled && !runtime_enabled)
7132 cfs_bandwidth_usage_dec();
7134 mutex_unlock(&cfs_constraints_mutex);
7140 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7144 period = ktime_to_ns(tg->cfs_bandwidth.period);
7145 if (cfs_quota_us < 0)
7146 quota = RUNTIME_INF;
7147 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7148 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7152 return tg_set_cfs_bandwidth(tg, period, quota);
7155 static long tg_get_cfs_quota(struct task_group *tg)
7159 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7162 quota_us = tg->cfs_bandwidth.quota;
7163 do_div(quota_us, NSEC_PER_USEC);
7168 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7172 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7175 period = (u64)cfs_period_us * NSEC_PER_USEC;
7176 quota = tg->cfs_bandwidth.quota;
7178 return tg_set_cfs_bandwidth(tg, period, quota);
7181 static long tg_get_cfs_period(struct task_group *tg)
7185 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7186 do_div(cfs_period_us, NSEC_PER_USEC);
7188 return cfs_period_us;
7191 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7194 return tg_get_cfs_quota(css_tg(css));
7197 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7198 struct cftype *cftype, s64 cfs_quota_us)
7200 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7203 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7206 return tg_get_cfs_period(css_tg(css));
7209 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7210 struct cftype *cftype, u64 cfs_period_us)
7212 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7215 struct cfs_schedulable_data {
7216 struct task_group *tg;
7221 * normalize group quota/period to be quota/max_period
7222 * note: units are usecs
7224 static u64 normalize_cfs_quota(struct task_group *tg,
7225 struct cfs_schedulable_data *d)
7233 period = tg_get_cfs_period(tg);
7234 quota = tg_get_cfs_quota(tg);
7237 /* note: these should typically be equivalent */
7238 if (quota == RUNTIME_INF || quota == -1)
7241 return to_ratio(period, quota);
7244 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7246 struct cfs_schedulable_data *d = data;
7247 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7248 s64 quota = 0, parent_quota = -1;
7251 quota = RUNTIME_INF;
7253 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7255 quota = normalize_cfs_quota(tg, d);
7256 parent_quota = parent_b->hierarchical_quota;
7259 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7260 * always take the min. On cgroup1, only inherit when no
7263 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7264 quota = min(quota, parent_quota);
7266 if (quota == RUNTIME_INF)
7267 quota = parent_quota;
7268 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7272 cfs_b->hierarchical_quota = quota;
7277 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7280 struct cfs_schedulable_data data = {
7286 if (quota != RUNTIME_INF) {
7287 do_div(data.period, NSEC_PER_USEC);
7288 do_div(data.quota, NSEC_PER_USEC);
7292 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7298 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7300 struct task_group *tg = css_tg(seq_css(sf));
7301 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7303 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7304 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7305 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7307 if (schedstat_enabled() && tg != &root_task_group) {
7311 for_each_possible_cpu(i)
7312 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7314 seq_printf(sf, "wait_sum %llu\n", ws);
7319 #endif /* CONFIG_CFS_BANDWIDTH */
7320 #endif /* CONFIG_FAIR_GROUP_SCHED */
7322 #ifdef CONFIG_RT_GROUP_SCHED
7323 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7324 struct cftype *cft, s64 val)
7326 return sched_group_set_rt_runtime(css_tg(css), val);
7329 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7332 return sched_group_rt_runtime(css_tg(css));
7335 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7336 struct cftype *cftype, u64 rt_period_us)
7338 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7341 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7344 return sched_group_rt_period(css_tg(css));
7346 #endif /* CONFIG_RT_GROUP_SCHED */
7348 static struct cftype cpu_legacy_files[] = {
7349 #ifdef CONFIG_FAIR_GROUP_SCHED
7352 .read_u64 = cpu_shares_read_u64,
7353 .write_u64 = cpu_shares_write_u64,
7356 #ifdef CONFIG_CFS_BANDWIDTH
7358 .name = "cfs_quota_us",
7359 .read_s64 = cpu_cfs_quota_read_s64,
7360 .write_s64 = cpu_cfs_quota_write_s64,
7363 .name = "cfs_period_us",
7364 .read_u64 = cpu_cfs_period_read_u64,
7365 .write_u64 = cpu_cfs_period_write_u64,
7369 .seq_show = cpu_cfs_stat_show,
7372 #ifdef CONFIG_RT_GROUP_SCHED
7374 .name = "rt_runtime_us",
7375 .read_s64 = cpu_rt_runtime_read,
7376 .write_s64 = cpu_rt_runtime_write,
7379 .name = "rt_period_us",
7380 .read_u64 = cpu_rt_period_read_uint,
7381 .write_u64 = cpu_rt_period_write_uint,
7387 static int cpu_extra_stat_show(struct seq_file *sf,
7388 struct cgroup_subsys_state *css)
7390 #ifdef CONFIG_CFS_BANDWIDTH
7392 struct task_group *tg = css_tg(css);
7393 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7396 throttled_usec = cfs_b->throttled_time;
7397 do_div(throttled_usec, NSEC_PER_USEC);
7399 seq_printf(sf, "nr_periods %d\n"
7401 "throttled_usec %llu\n",
7402 cfs_b->nr_periods, cfs_b->nr_throttled,
7409 #ifdef CONFIG_FAIR_GROUP_SCHED
7410 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7413 struct task_group *tg = css_tg(css);
7414 u64 weight = scale_load_down(tg->shares);
7416 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7419 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7420 struct cftype *cft, u64 weight)
7423 * cgroup weight knobs should use the common MIN, DFL and MAX
7424 * values which are 1, 100 and 10000 respectively. While it loses
7425 * a bit of range on both ends, it maps pretty well onto the shares
7426 * value used by scheduler and the round-trip conversions preserve
7427 * the original value over the entire range.
7429 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7432 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7434 return sched_group_set_shares(css_tg(css), scale_load(weight));
7437 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7440 unsigned long weight = scale_load_down(css_tg(css)->shares);
7441 int last_delta = INT_MAX;
7444 /* find the closest nice value to the current weight */
7445 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7446 delta = abs(sched_prio_to_weight[prio] - weight);
7447 if (delta >= last_delta)
7452 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7455 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7456 struct cftype *cft, s64 nice)
7458 unsigned long weight;
7461 if (nice < MIN_NICE || nice > MAX_NICE)
7464 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7465 idx = array_index_nospec(idx, 40);
7466 weight = sched_prio_to_weight[idx];
7468 return sched_group_set_shares(css_tg(css), scale_load(weight));
7472 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7473 long period, long quota)
7476 seq_puts(sf, "max");
7478 seq_printf(sf, "%ld", quota);
7480 seq_printf(sf, " %ld\n", period);
7483 /* caller should put the current value in *@periodp before calling */
7484 static int __maybe_unused cpu_period_quota_parse(char *buf,
7485 u64 *periodp, u64 *quotap)
7487 char tok[21]; /* U64_MAX */
7489 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7492 *periodp *= NSEC_PER_USEC;
7494 if (sscanf(tok, "%llu", quotap))
7495 *quotap *= NSEC_PER_USEC;
7496 else if (!strcmp(tok, "max"))
7497 *quotap = RUNTIME_INF;
7504 #ifdef CONFIG_CFS_BANDWIDTH
7505 static int cpu_max_show(struct seq_file *sf, void *v)
7507 struct task_group *tg = css_tg(seq_css(sf));
7509 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7513 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7514 char *buf, size_t nbytes, loff_t off)
7516 struct task_group *tg = css_tg(of_css(of));
7517 u64 period = tg_get_cfs_period(tg);
7521 ret = cpu_period_quota_parse(buf, &period, "a);
7523 ret = tg_set_cfs_bandwidth(tg, period, quota);
7524 return ret ?: nbytes;
7528 static struct cftype cpu_files[] = {
7529 #ifdef CONFIG_FAIR_GROUP_SCHED
7532 .flags = CFTYPE_NOT_ON_ROOT,
7533 .read_u64 = cpu_weight_read_u64,
7534 .write_u64 = cpu_weight_write_u64,
7537 .name = "weight.nice",
7538 .flags = CFTYPE_NOT_ON_ROOT,
7539 .read_s64 = cpu_weight_nice_read_s64,
7540 .write_s64 = cpu_weight_nice_write_s64,
7543 #ifdef CONFIG_CFS_BANDWIDTH
7546 .flags = CFTYPE_NOT_ON_ROOT,
7547 .seq_show = cpu_max_show,
7548 .write = cpu_max_write,
7554 struct cgroup_subsys cpu_cgrp_subsys = {
7555 .css_alloc = cpu_cgroup_css_alloc,
7556 .css_online = cpu_cgroup_css_online,
7557 .css_released = cpu_cgroup_css_released,
7558 .css_free = cpu_cgroup_css_free,
7559 .css_extra_stat_show = cpu_extra_stat_show,
7560 .fork = cpu_cgroup_fork,
7561 .can_attach = cpu_cgroup_can_attach,
7562 .attach = cpu_cgroup_attach,
7563 .legacy_cftypes = cpu_legacy_files,
7564 .dfl_cftypes = cpu_files,
7569 #endif /* CONFIG_CGROUP_SCHED */
7571 void dump_cpu_task(int cpu)
7573 pr_info("Task dump for CPU %d:\n", cpu);
7574 sched_show_task(cpu_curr(cpu));
7578 * Nice levels are multiplicative, with a gentle 10% change for every
7579 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7580 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7581 * that remained on nice 0.
7583 * The "10% effect" is relative and cumulative: from _any_ nice level,
7584 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7585 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7586 * If a task goes up by ~10% and another task goes down by ~10% then
7587 * the relative distance between them is ~25%.)
7589 const int sched_prio_to_weight[40] = {
7590 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7591 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7592 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7593 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7594 /* 0 */ 1024, 820, 655, 526, 423,
7595 /* 5 */ 335, 272, 215, 172, 137,
7596 /* 10 */ 110, 87, 70, 56, 45,
7597 /* 15 */ 36, 29, 23, 18, 15,
7601 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7603 * In cases where the weight does not change often, we can use the
7604 * precalculated inverse to speed up arithmetics by turning divisions
7605 * into multiplications:
7607 const u32 sched_prio_to_wmult[40] = {
7608 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7609 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7610 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7611 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7612 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7613 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7614 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7615 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7618 #undef CREATE_TRACE_POINTS