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
777 * Serializes updates of utilization clamp values
779 * The (slow-path) user-space triggers utilization clamp value updates which
780 * can require updates on (fast-path) scheduler's data structures used to
781 * support enqueue/dequeue operations.
782 * While the per-CPU rq lock protects fast-path update operations, user-space
783 * requests are serialized using a mutex to reduce the risk of conflicting
784 * updates or API abuses.
786 static DEFINE_MUTEX(uclamp_mutex);
788 /* Max allowed minimum utilization */
789 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
791 /* Max allowed maximum utilization */
792 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
794 /* All clamps are required to be less or equal than these values */
795 static struct uclamp_se uclamp_default[UCLAMP_CNT];
797 /* Integer rounded range for each bucket */
798 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
800 #define for_each_clamp_id(clamp_id) \
801 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
803 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
805 return clamp_value / UCLAMP_BUCKET_DELTA;
808 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
810 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
813 static inline enum uclamp_id uclamp_none(enum uclamp_id clamp_id)
815 if (clamp_id == UCLAMP_MIN)
817 return SCHED_CAPACITY_SCALE;
820 static inline void uclamp_se_set(struct uclamp_se *uc_se,
821 unsigned int value, bool user_defined)
823 uc_se->value = value;
824 uc_se->bucket_id = uclamp_bucket_id(value);
825 uc_se->user_defined = user_defined;
828 static inline unsigned int
829 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
830 unsigned int clamp_value)
833 * Avoid blocked utilization pushing up the frequency when we go
834 * idle (which drops the max-clamp) by retaining the last known
837 if (clamp_id == UCLAMP_MAX) {
838 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
842 return uclamp_none(UCLAMP_MIN);
845 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
846 unsigned int clamp_value)
848 /* Reset max-clamp retention only on idle exit */
849 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
852 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
856 enum uclamp_id uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
857 unsigned int clamp_value)
859 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
860 int bucket_id = UCLAMP_BUCKETS - 1;
863 * Since both min and max clamps are max aggregated, find the
864 * top most bucket with tasks in.
866 for ( ; bucket_id >= 0; bucket_id--) {
867 if (!bucket[bucket_id].tasks)
869 return bucket[bucket_id].value;
872 /* No tasks -- default clamp values */
873 return uclamp_idle_value(rq, clamp_id, clamp_value);
876 static inline struct uclamp_se
877 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
879 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
880 #ifdef CONFIG_UCLAMP_TASK_GROUP
881 struct uclamp_se uc_max;
884 * Tasks in autogroups or root task group will be
885 * restricted by system defaults.
887 if (task_group_is_autogroup(task_group(p)))
889 if (task_group(p) == &root_task_group)
892 uc_max = task_group(p)->uclamp[clamp_id];
893 if (uc_req.value > uc_max.value || !uc_req.user_defined)
901 * The effective clamp bucket index of a task depends on, by increasing
903 * - the task specific clamp value, when explicitly requested from userspace
904 * - the task group effective clamp value, for tasks not either in the root
905 * group or in an autogroup
906 * - the system default clamp value, defined by the sysadmin
908 static inline struct uclamp_se
909 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
911 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
912 struct uclamp_se uc_max = uclamp_default[clamp_id];
914 /* System default restrictions always apply */
915 if (unlikely(uc_req.value > uc_max.value))
921 enum uclamp_id uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
923 struct uclamp_se uc_eff;
925 /* Task currently refcounted: use back-annotated (effective) value */
926 if (p->uclamp[clamp_id].active)
927 return p->uclamp[clamp_id].value;
929 uc_eff = uclamp_eff_get(p, clamp_id);
935 * When a task is enqueued on a rq, the clamp bucket currently defined by the
936 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
937 * updates the rq's clamp value if required.
939 * Tasks can have a task-specific value requested from user-space, track
940 * within each bucket the maximum value for tasks refcounted in it.
941 * This "local max aggregation" allows to track the exact "requested" value
942 * for each bucket when all its RUNNABLE tasks require the same clamp.
944 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
945 enum uclamp_id clamp_id)
947 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
948 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
949 struct uclamp_bucket *bucket;
951 lockdep_assert_held(&rq->lock);
953 /* Update task effective clamp */
954 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
956 bucket = &uc_rq->bucket[uc_se->bucket_id];
958 uc_se->active = true;
960 uclamp_idle_reset(rq, clamp_id, uc_se->value);
963 * Local max aggregation: rq buckets always track the max
964 * "requested" clamp value of its RUNNABLE tasks.
966 if (bucket->tasks == 1 || uc_se->value > bucket->value)
967 bucket->value = uc_se->value;
969 if (uc_se->value > READ_ONCE(uc_rq->value))
970 WRITE_ONCE(uc_rq->value, uc_se->value);
974 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
975 * is released. If this is the last task reference counting the rq's max
976 * active clamp value, then the rq's clamp value is updated.
978 * Both refcounted tasks and rq's cached clamp values are expected to be
979 * always valid. If it's detected they are not, as defensive programming,
980 * enforce the expected state and warn.
982 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
983 enum uclamp_id clamp_id)
985 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
986 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
987 struct uclamp_bucket *bucket;
988 unsigned int bkt_clamp;
989 unsigned int rq_clamp;
991 lockdep_assert_held(&rq->lock);
993 bucket = &uc_rq->bucket[uc_se->bucket_id];
994 SCHED_WARN_ON(!bucket->tasks);
995 if (likely(bucket->tasks))
997 uc_se->active = false;
1000 * Keep "local max aggregation" simple and accept to (possibly)
1001 * overboost some RUNNABLE tasks in the same bucket.
1002 * The rq clamp bucket value is reset to its base value whenever
1003 * there are no more RUNNABLE tasks refcounting it.
1005 if (likely(bucket->tasks))
1008 rq_clamp = READ_ONCE(uc_rq->value);
1010 * Defensive programming: this should never happen. If it happens,
1011 * e.g. due to future modification, warn and fixup the expected value.
1013 SCHED_WARN_ON(bucket->value > rq_clamp);
1014 if (bucket->value >= rq_clamp) {
1015 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1016 WRITE_ONCE(uc_rq->value, bkt_clamp);
1020 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1022 enum uclamp_id clamp_id;
1024 if (unlikely(!p->sched_class->uclamp_enabled))
1027 for_each_clamp_id(clamp_id)
1028 uclamp_rq_inc_id(rq, p, clamp_id);
1030 /* Reset clamp idle holding when there is one RUNNABLE task */
1031 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1032 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1035 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1037 enum uclamp_id clamp_id;
1039 if (unlikely(!p->sched_class->uclamp_enabled))
1042 for_each_clamp_id(clamp_id)
1043 uclamp_rq_dec_id(rq, p, clamp_id);
1047 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1053 * Lock the task and the rq where the task is (or was) queued.
1055 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1056 * price to pay to safely serialize util_{min,max} updates with
1057 * enqueues, dequeues and migration operations.
1058 * This is the same locking schema used by __set_cpus_allowed_ptr().
1060 rq = task_rq_lock(p, &rf);
1063 * Setting the clamp bucket is serialized by task_rq_lock().
1064 * If the task is not yet RUNNABLE and its task_struct is not
1065 * affecting a valid clamp bucket, the next time it's enqueued,
1066 * it will already see the updated clamp bucket value.
1068 if (!p->uclamp[clamp_id].active) {
1069 uclamp_rq_dec_id(rq, p, clamp_id);
1070 uclamp_rq_inc_id(rq, p, clamp_id);
1073 task_rq_unlock(rq, p, &rf);
1077 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1078 unsigned int clamps)
1080 enum uclamp_id clamp_id;
1081 struct css_task_iter it;
1082 struct task_struct *p;
1084 css_task_iter_start(css, 0, &it);
1085 while ((p = css_task_iter_next(&it))) {
1086 for_each_clamp_id(clamp_id) {
1087 if ((0x1 << clamp_id) & clamps)
1088 uclamp_update_active(p, clamp_id);
1091 css_task_iter_end(&it);
1094 #ifdef CONFIG_UCLAMP_TASK_GROUP
1095 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1096 static void uclamp_update_root_tg(void)
1098 struct task_group *tg = &root_task_group;
1100 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1101 sysctl_sched_uclamp_util_min, false);
1102 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1103 sysctl_sched_uclamp_util_max, false);
1106 cpu_util_update_eff(&root_task_group.css);
1110 static void uclamp_update_root_tg(void) { }
1113 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1114 void __user *buffer, size_t *lenp,
1117 bool update_root_tg = false;
1118 int old_min, old_max;
1121 mutex_lock(&uclamp_mutex);
1122 old_min = sysctl_sched_uclamp_util_min;
1123 old_max = sysctl_sched_uclamp_util_max;
1125 result = proc_dointvec(table, write, buffer, lenp, ppos);
1131 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1132 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1137 if (old_min != sysctl_sched_uclamp_util_min) {
1138 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1139 sysctl_sched_uclamp_util_min, false);
1140 update_root_tg = true;
1142 if (old_max != sysctl_sched_uclamp_util_max) {
1143 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1144 sysctl_sched_uclamp_util_max, false);
1145 update_root_tg = true;
1149 uclamp_update_root_tg();
1152 * We update all RUNNABLE tasks only when task groups are in use.
1153 * Otherwise, keep it simple and do just a lazy update at each next
1154 * task enqueue time.
1160 sysctl_sched_uclamp_util_min = old_min;
1161 sysctl_sched_uclamp_util_max = old_max;
1163 mutex_unlock(&uclamp_mutex);
1168 static int uclamp_validate(struct task_struct *p,
1169 const struct sched_attr *attr)
1171 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1172 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1174 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1175 lower_bound = attr->sched_util_min;
1176 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1177 upper_bound = attr->sched_util_max;
1179 if (lower_bound > upper_bound)
1181 if (upper_bound > SCHED_CAPACITY_SCALE)
1187 static void __setscheduler_uclamp(struct task_struct *p,
1188 const struct sched_attr *attr)
1190 enum uclamp_id clamp_id;
1193 * On scheduling class change, reset to default clamps for tasks
1194 * without a task-specific value.
1196 for_each_clamp_id(clamp_id) {
1197 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1198 unsigned int clamp_value = uclamp_none(clamp_id);
1200 /* Keep using defined clamps across class changes */
1201 if (uc_se->user_defined)
1204 /* By default, RT tasks always get 100% boost */
1205 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1206 clamp_value = uclamp_none(UCLAMP_MAX);
1208 uclamp_se_set(uc_se, clamp_value, false);
1211 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1214 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1215 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1216 attr->sched_util_min, true);
1219 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1220 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1221 attr->sched_util_max, true);
1225 static void uclamp_fork(struct task_struct *p)
1227 enum uclamp_id clamp_id;
1229 for_each_clamp_id(clamp_id)
1230 p->uclamp[clamp_id].active = false;
1232 if (likely(!p->sched_reset_on_fork))
1235 for_each_clamp_id(clamp_id) {
1236 unsigned int clamp_value = uclamp_none(clamp_id);
1238 /* By default, RT tasks always get 100% boost */
1239 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1240 clamp_value = uclamp_none(UCLAMP_MAX);
1242 uclamp_se_set(&p->uclamp_req[clamp_id], clamp_value, false);
1246 static void __init init_uclamp(void)
1248 struct uclamp_se uc_max = {};
1249 enum uclamp_id clamp_id;
1252 mutex_init(&uclamp_mutex);
1254 for_each_possible_cpu(cpu) {
1255 memset(&cpu_rq(cpu)->uclamp, 0, sizeof(struct uclamp_rq));
1256 cpu_rq(cpu)->uclamp_flags = 0;
1259 for_each_clamp_id(clamp_id) {
1260 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1261 uclamp_none(clamp_id), false);
1264 /* System defaults allow max clamp values for both indexes */
1265 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1266 for_each_clamp_id(clamp_id) {
1267 uclamp_default[clamp_id] = uc_max;
1268 #ifdef CONFIG_UCLAMP_TASK_GROUP
1269 root_task_group.uclamp_req[clamp_id] = uc_max;
1270 root_task_group.uclamp[clamp_id] = uc_max;
1275 #else /* CONFIG_UCLAMP_TASK */
1276 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1277 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1278 static inline int uclamp_validate(struct task_struct *p,
1279 const struct sched_attr *attr)
1283 static void __setscheduler_uclamp(struct task_struct *p,
1284 const struct sched_attr *attr) { }
1285 static inline void uclamp_fork(struct task_struct *p) { }
1286 static inline void init_uclamp(void) { }
1287 #endif /* CONFIG_UCLAMP_TASK */
1289 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1291 if (!(flags & ENQUEUE_NOCLOCK))
1292 update_rq_clock(rq);
1294 if (!(flags & ENQUEUE_RESTORE)) {
1295 sched_info_queued(rq, p);
1296 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1299 uclamp_rq_inc(rq, p);
1300 p->sched_class->enqueue_task(rq, p, flags);
1303 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1305 if (!(flags & DEQUEUE_NOCLOCK))
1306 update_rq_clock(rq);
1308 if (!(flags & DEQUEUE_SAVE)) {
1309 sched_info_dequeued(rq, p);
1310 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1313 uclamp_rq_dec(rq, p);
1314 p->sched_class->dequeue_task(rq, p, flags);
1317 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1319 if (task_contributes_to_load(p))
1320 rq->nr_uninterruptible--;
1322 enqueue_task(rq, p, flags);
1324 p->on_rq = TASK_ON_RQ_QUEUED;
1327 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1329 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1331 if (task_contributes_to_load(p))
1332 rq->nr_uninterruptible++;
1334 dequeue_task(rq, p, flags);
1338 * __normal_prio - return the priority that is based on the static prio
1340 static inline int __normal_prio(struct task_struct *p)
1342 return p->static_prio;
1346 * Calculate the expected normal priority: i.e. priority
1347 * without taking RT-inheritance into account. Might be
1348 * boosted by interactivity modifiers. Changes upon fork,
1349 * setprio syscalls, and whenever the interactivity
1350 * estimator recalculates.
1352 static inline int normal_prio(struct task_struct *p)
1356 if (task_has_dl_policy(p))
1357 prio = MAX_DL_PRIO-1;
1358 else if (task_has_rt_policy(p))
1359 prio = MAX_RT_PRIO-1 - p->rt_priority;
1361 prio = __normal_prio(p);
1366 * Calculate the current priority, i.e. the priority
1367 * taken into account by the scheduler. This value might
1368 * be boosted by RT tasks, or might be boosted by
1369 * interactivity modifiers. Will be RT if the task got
1370 * RT-boosted. If not then it returns p->normal_prio.
1372 static int effective_prio(struct task_struct *p)
1374 p->normal_prio = normal_prio(p);
1376 * If we are RT tasks or we were boosted to RT priority,
1377 * keep the priority unchanged. Otherwise, update priority
1378 * to the normal priority:
1380 if (!rt_prio(p->prio))
1381 return p->normal_prio;
1386 * task_curr - is this task currently executing on a CPU?
1387 * @p: the task in question.
1389 * Return: 1 if the task is currently executing. 0 otherwise.
1391 inline int task_curr(const struct task_struct *p)
1393 return cpu_curr(task_cpu(p)) == p;
1397 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1398 * use the balance_callback list if you want balancing.
1400 * this means any call to check_class_changed() must be followed by a call to
1401 * balance_callback().
1403 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1404 const struct sched_class *prev_class,
1407 if (prev_class != p->sched_class) {
1408 if (prev_class->switched_from)
1409 prev_class->switched_from(rq, p);
1411 p->sched_class->switched_to(rq, p);
1412 } else if (oldprio != p->prio || dl_task(p))
1413 p->sched_class->prio_changed(rq, p, oldprio);
1416 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1418 const struct sched_class *class;
1420 if (p->sched_class == rq->curr->sched_class) {
1421 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1423 for_each_class(class) {
1424 if (class == rq->curr->sched_class)
1426 if (class == p->sched_class) {
1434 * A queue event has occurred, and we're going to schedule. In
1435 * this case, we can save a useless back to back clock update.
1437 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1438 rq_clock_skip_update(rq);
1443 static inline bool is_per_cpu_kthread(struct task_struct *p)
1445 if (!(p->flags & PF_KTHREAD))
1448 if (p->nr_cpus_allowed != 1)
1455 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1456 * __set_cpus_allowed_ptr() and select_fallback_rq().
1458 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1460 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1463 if (is_per_cpu_kthread(p))
1464 return cpu_online(cpu);
1466 return cpu_active(cpu);
1470 * This is how migration works:
1472 * 1) we invoke migration_cpu_stop() on the target CPU using
1474 * 2) stopper starts to run (implicitly forcing the migrated thread
1476 * 3) it checks whether the migrated task is still in the wrong runqueue.
1477 * 4) if it's in the wrong runqueue then the migration thread removes
1478 * it and puts it into the right queue.
1479 * 5) stopper completes and stop_one_cpu() returns and the migration
1484 * move_queued_task - move a queued task to new rq.
1486 * Returns (locked) new rq. Old rq's lock is released.
1488 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1489 struct task_struct *p, int new_cpu)
1491 lockdep_assert_held(&rq->lock);
1493 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1494 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1495 set_task_cpu(p, new_cpu);
1498 rq = cpu_rq(new_cpu);
1501 BUG_ON(task_cpu(p) != new_cpu);
1502 enqueue_task(rq, p, 0);
1503 p->on_rq = TASK_ON_RQ_QUEUED;
1504 check_preempt_curr(rq, p, 0);
1509 struct migration_arg {
1510 struct task_struct *task;
1515 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1516 * this because either it can't run here any more (set_cpus_allowed()
1517 * away from this CPU, or CPU going down), or because we're
1518 * attempting to rebalance this task on exec (sched_exec).
1520 * So we race with normal scheduler movements, but that's OK, as long
1521 * as the task is no longer on this CPU.
1523 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1524 struct task_struct *p, int dest_cpu)
1526 /* Affinity changed (again). */
1527 if (!is_cpu_allowed(p, dest_cpu))
1530 update_rq_clock(rq);
1531 rq = move_queued_task(rq, rf, p, dest_cpu);
1537 * migration_cpu_stop - this will be executed by a highprio stopper thread
1538 * and performs thread migration by bumping thread off CPU then
1539 * 'pushing' onto another runqueue.
1541 static int migration_cpu_stop(void *data)
1543 struct migration_arg *arg = data;
1544 struct task_struct *p = arg->task;
1545 struct rq *rq = this_rq();
1549 * The original target CPU might have gone down and we might
1550 * be on another CPU but it doesn't matter.
1552 local_irq_disable();
1554 * We need to explicitly wake pending tasks before running
1555 * __migrate_task() such that we will not miss enforcing cpus_ptr
1556 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1558 sched_ttwu_pending();
1560 raw_spin_lock(&p->pi_lock);
1563 * If task_rq(p) != rq, it cannot be migrated here, because we're
1564 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1565 * we're holding p->pi_lock.
1567 if (task_rq(p) == rq) {
1568 if (task_on_rq_queued(p))
1569 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1571 p->wake_cpu = arg->dest_cpu;
1574 raw_spin_unlock(&p->pi_lock);
1581 * sched_class::set_cpus_allowed must do the below, but is not required to
1582 * actually call this function.
1584 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1586 cpumask_copy(&p->cpus_mask, new_mask);
1587 p->nr_cpus_allowed = cpumask_weight(new_mask);
1590 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1592 struct rq *rq = task_rq(p);
1593 bool queued, running;
1595 lockdep_assert_held(&p->pi_lock);
1597 queued = task_on_rq_queued(p);
1598 running = task_current(rq, p);
1602 * Because __kthread_bind() calls this on blocked tasks without
1605 lockdep_assert_held(&rq->lock);
1606 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1609 put_prev_task(rq, p);
1611 p->sched_class->set_cpus_allowed(p, new_mask);
1614 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1616 set_next_task(rq, p);
1620 * Change a given task's CPU affinity. Migrate the thread to a
1621 * proper CPU and schedule it away if the CPU it's executing on
1622 * is removed from the allowed bitmask.
1624 * NOTE: the caller must have a valid reference to the task, the
1625 * task must not exit() & deallocate itself prematurely. The
1626 * call is not atomic; no spinlocks may be held.
1628 static int __set_cpus_allowed_ptr(struct task_struct *p,
1629 const struct cpumask *new_mask, bool check)
1631 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1632 unsigned int dest_cpu;
1637 rq = task_rq_lock(p, &rf);
1638 update_rq_clock(rq);
1640 if (p->flags & PF_KTHREAD) {
1642 * Kernel threads are allowed on online && !active CPUs
1644 cpu_valid_mask = cpu_online_mask;
1648 * Must re-check here, to close a race against __kthread_bind(),
1649 * sched_setaffinity() is not guaranteed to observe the flag.
1651 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1656 if (cpumask_equal(p->cpus_ptr, new_mask))
1659 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1660 if (dest_cpu >= nr_cpu_ids) {
1665 do_set_cpus_allowed(p, new_mask);
1667 if (p->flags & PF_KTHREAD) {
1669 * For kernel threads that do indeed end up on online &&
1670 * !active we want to ensure they are strict per-CPU threads.
1672 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1673 !cpumask_intersects(new_mask, cpu_active_mask) &&
1674 p->nr_cpus_allowed != 1);
1677 /* Can the task run on the task's current CPU? If so, we're done */
1678 if (cpumask_test_cpu(task_cpu(p), new_mask))
1681 if (task_running(rq, p) || p->state == TASK_WAKING) {
1682 struct migration_arg arg = { p, dest_cpu };
1683 /* Need help from migration thread: drop lock and wait. */
1684 task_rq_unlock(rq, p, &rf);
1685 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1687 } else if (task_on_rq_queued(p)) {
1689 * OK, since we're going to drop the lock immediately
1690 * afterwards anyway.
1692 rq = move_queued_task(rq, &rf, p, dest_cpu);
1695 task_rq_unlock(rq, p, &rf);
1700 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1702 return __set_cpus_allowed_ptr(p, new_mask, false);
1704 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1706 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1708 #ifdef CONFIG_SCHED_DEBUG
1710 * We should never call set_task_cpu() on a blocked task,
1711 * ttwu() will sort out the placement.
1713 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1717 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1718 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1719 * time relying on p->on_rq.
1721 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1722 p->sched_class == &fair_sched_class &&
1723 (p->on_rq && !task_on_rq_migrating(p)));
1725 #ifdef CONFIG_LOCKDEP
1727 * The caller should hold either p->pi_lock or rq->lock, when changing
1728 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1730 * sched_move_task() holds both and thus holding either pins the cgroup,
1733 * Furthermore, all task_rq users should acquire both locks, see
1736 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1737 lockdep_is_held(&task_rq(p)->lock)));
1740 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1742 WARN_ON_ONCE(!cpu_online(new_cpu));
1745 trace_sched_migrate_task(p, new_cpu);
1747 if (task_cpu(p) != new_cpu) {
1748 if (p->sched_class->migrate_task_rq)
1749 p->sched_class->migrate_task_rq(p, new_cpu);
1750 p->se.nr_migrations++;
1752 perf_event_task_migrate(p);
1755 __set_task_cpu(p, new_cpu);
1758 #ifdef CONFIG_NUMA_BALANCING
1759 static void __migrate_swap_task(struct task_struct *p, int cpu)
1761 if (task_on_rq_queued(p)) {
1762 struct rq *src_rq, *dst_rq;
1763 struct rq_flags srf, drf;
1765 src_rq = task_rq(p);
1766 dst_rq = cpu_rq(cpu);
1768 rq_pin_lock(src_rq, &srf);
1769 rq_pin_lock(dst_rq, &drf);
1771 deactivate_task(src_rq, p, 0);
1772 set_task_cpu(p, cpu);
1773 activate_task(dst_rq, p, 0);
1774 check_preempt_curr(dst_rq, p, 0);
1776 rq_unpin_lock(dst_rq, &drf);
1777 rq_unpin_lock(src_rq, &srf);
1781 * Task isn't running anymore; make it appear like we migrated
1782 * it before it went to sleep. This means on wakeup we make the
1783 * previous CPU our target instead of where it really is.
1789 struct migration_swap_arg {
1790 struct task_struct *src_task, *dst_task;
1791 int src_cpu, dst_cpu;
1794 static int migrate_swap_stop(void *data)
1796 struct migration_swap_arg *arg = data;
1797 struct rq *src_rq, *dst_rq;
1800 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1803 src_rq = cpu_rq(arg->src_cpu);
1804 dst_rq = cpu_rq(arg->dst_cpu);
1806 double_raw_lock(&arg->src_task->pi_lock,
1807 &arg->dst_task->pi_lock);
1808 double_rq_lock(src_rq, dst_rq);
1810 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1813 if (task_cpu(arg->src_task) != arg->src_cpu)
1816 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1819 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1822 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1823 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1828 double_rq_unlock(src_rq, dst_rq);
1829 raw_spin_unlock(&arg->dst_task->pi_lock);
1830 raw_spin_unlock(&arg->src_task->pi_lock);
1836 * Cross migrate two tasks
1838 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1839 int target_cpu, int curr_cpu)
1841 struct migration_swap_arg arg;
1844 arg = (struct migration_swap_arg){
1846 .src_cpu = curr_cpu,
1848 .dst_cpu = target_cpu,
1851 if (arg.src_cpu == arg.dst_cpu)
1855 * These three tests are all lockless; this is OK since all of them
1856 * will be re-checked with proper locks held further down the line.
1858 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1861 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1864 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1867 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1868 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1873 #endif /* CONFIG_NUMA_BALANCING */
1876 * wait_task_inactive - wait for a thread to unschedule.
1878 * If @match_state is nonzero, it's the @p->state value just checked and
1879 * not expected to change. If it changes, i.e. @p might have woken up,
1880 * then return zero. When we succeed in waiting for @p to be off its CPU,
1881 * we return a positive number (its total switch count). If a second call
1882 * a short while later returns the same number, the caller can be sure that
1883 * @p has remained unscheduled the whole time.
1885 * The caller must ensure that the task *will* unschedule sometime soon,
1886 * else this function might spin for a *long* time. This function can't
1887 * be called with interrupts off, or it may introduce deadlock with
1888 * smp_call_function() if an IPI is sent by the same process we are
1889 * waiting to become inactive.
1891 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1893 int running, queued;
1900 * We do the initial early heuristics without holding
1901 * any task-queue locks at all. We'll only try to get
1902 * the runqueue lock when things look like they will
1908 * If the task is actively running on another CPU
1909 * still, just relax and busy-wait without holding
1912 * NOTE! Since we don't hold any locks, it's not
1913 * even sure that "rq" stays as the right runqueue!
1914 * But we don't care, since "task_running()" will
1915 * return false if the runqueue has changed and p
1916 * is actually now running somewhere else!
1918 while (task_running(rq, p)) {
1919 if (match_state && unlikely(p->state != match_state))
1925 * Ok, time to look more closely! We need the rq
1926 * lock now, to be *sure*. If we're wrong, we'll
1927 * just go back and repeat.
1929 rq = task_rq_lock(p, &rf);
1930 trace_sched_wait_task(p);
1931 running = task_running(rq, p);
1932 queued = task_on_rq_queued(p);
1934 if (!match_state || p->state == match_state)
1935 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1936 task_rq_unlock(rq, p, &rf);
1939 * If it changed from the expected state, bail out now.
1941 if (unlikely(!ncsw))
1945 * Was it really running after all now that we
1946 * checked with the proper locks actually held?
1948 * Oops. Go back and try again..
1950 if (unlikely(running)) {
1956 * It's not enough that it's not actively running,
1957 * it must be off the runqueue _entirely_, and not
1960 * So if it was still runnable (but just not actively
1961 * running right now), it's preempted, and we should
1962 * yield - it could be a while.
1964 if (unlikely(queued)) {
1965 ktime_t to = NSEC_PER_SEC / HZ;
1967 set_current_state(TASK_UNINTERRUPTIBLE);
1968 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1973 * Ahh, all good. It wasn't running, and it wasn't
1974 * runnable, which means that it will never become
1975 * running in the future either. We're all done!
1984 * kick_process - kick a running thread to enter/exit the kernel
1985 * @p: the to-be-kicked thread
1987 * Cause a process which is running on another CPU to enter
1988 * kernel-mode, without any delay. (to get signals handled.)
1990 * NOTE: this function doesn't have to take the runqueue lock,
1991 * because all it wants to ensure is that the remote task enters
1992 * the kernel. If the IPI races and the task has been migrated
1993 * to another CPU then no harm is done and the purpose has been
1996 void kick_process(struct task_struct *p)
2002 if ((cpu != smp_processor_id()) && task_curr(p))
2003 smp_send_reschedule(cpu);
2006 EXPORT_SYMBOL_GPL(kick_process);
2009 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2011 * A few notes on cpu_active vs cpu_online:
2013 * - cpu_active must be a subset of cpu_online
2015 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2016 * see __set_cpus_allowed_ptr(). At this point the newly online
2017 * CPU isn't yet part of the sched domains, and balancing will not
2020 * - on CPU-down we clear cpu_active() to mask the sched domains and
2021 * avoid the load balancer to place new tasks on the to be removed
2022 * CPU. Existing tasks will remain running there and will be taken
2025 * This means that fallback selection must not select !active CPUs.
2026 * And can assume that any active CPU must be online. Conversely
2027 * select_task_rq() below may allow selection of !active CPUs in order
2028 * to satisfy the above rules.
2030 static int select_fallback_rq(int cpu, struct task_struct *p)
2032 int nid = cpu_to_node(cpu);
2033 const struct cpumask *nodemask = NULL;
2034 enum { cpuset, possible, fail } state = cpuset;
2038 * If the node that the CPU is on has been offlined, cpu_to_node()
2039 * will return -1. There is no CPU on the node, and we should
2040 * select the CPU on the other node.
2043 nodemask = cpumask_of_node(nid);
2045 /* Look for allowed, online CPU in same node. */
2046 for_each_cpu(dest_cpu, nodemask) {
2047 if (!cpu_active(dest_cpu))
2049 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2055 /* Any allowed, online CPU? */
2056 for_each_cpu(dest_cpu, p->cpus_ptr) {
2057 if (!is_cpu_allowed(p, dest_cpu))
2063 /* No more Mr. Nice Guy. */
2066 if (IS_ENABLED(CONFIG_CPUSETS)) {
2067 cpuset_cpus_allowed_fallback(p);
2073 do_set_cpus_allowed(p, cpu_possible_mask);
2084 if (state != cpuset) {
2086 * Don't tell them about moving exiting tasks or
2087 * kernel threads (both mm NULL), since they never
2090 if (p->mm && printk_ratelimit()) {
2091 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2092 task_pid_nr(p), p->comm, cpu);
2100 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2103 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2105 lockdep_assert_held(&p->pi_lock);
2107 if (p->nr_cpus_allowed > 1)
2108 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2110 cpu = cpumask_any(p->cpus_ptr);
2113 * In order not to call set_task_cpu() on a blocking task we need
2114 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2117 * Since this is common to all placement strategies, this lives here.
2119 * [ this allows ->select_task() to simply return task_cpu(p) and
2120 * not worry about this generic constraint ]
2122 if (unlikely(!is_cpu_allowed(p, cpu)))
2123 cpu = select_fallback_rq(task_cpu(p), p);
2128 static void update_avg(u64 *avg, u64 sample)
2130 s64 diff = sample - *avg;
2134 void sched_set_stop_task(int cpu, struct task_struct *stop)
2136 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2137 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2141 * Make it appear like a SCHED_FIFO task, its something
2142 * userspace knows about and won't get confused about.
2144 * Also, it will make PI more or less work without too
2145 * much confusion -- but then, stop work should not
2146 * rely on PI working anyway.
2148 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2150 stop->sched_class = &stop_sched_class;
2153 cpu_rq(cpu)->stop = stop;
2157 * Reset it back to a normal scheduling class so that
2158 * it can die in pieces.
2160 old_stop->sched_class = &rt_sched_class;
2166 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2167 const struct cpumask *new_mask, bool check)
2169 return set_cpus_allowed_ptr(p, new_mask);
2172 #endif /* CONFIG_SMP */
2175 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2179 if (!schedstat_enabled())
2185 if (cpu == rq->cpu) {
2186 __schedstat_inc(rq->ttwu_local);
2187 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2189 struct sched_domain *sd;
2191 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2193 for_each_domain(rq->cpu, sd) {
2194 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2195 __schedstat_inc(sd->ttwu_wake_remote);
2202 if (wake_flags & WF_MIGRATED)
2203 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2204 #endif /* CONFIG_SMP */
2206 __schedstat_inc(rq->ttwu_count);
2207 __schedstat_inc(p->se.statistics.nr_wakeups);
2209 if (wake_flags & WF_SYNC)
2210 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2214 * Mark the task runnable and perform wakeup-preemption.
2216 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2217 struct rq_flags *rf)
2219 check_preempt_curr(rq, p, wake_flags);
2220 p->state = TASK_RUNNING;
2221 trace_sched_wakeup(p);
2224 if (p->sched_class->task_woken) {
2226 * Our task @p is fully woken up and running; so its safe to
2227 * drop the rq->lock, hereafter rq is only used for statistics.
2229 rq_unpin_lock(rq, rf);
2230 p->sched_class->task_woken(rq, p);
2231 rq_repin_lock(rq, rf);
2234 if (rq->idle_stamp) {
2235 u64 delta = rq_clock(rq) - rq->idle_stamp;
2236 u64 max = 2*rq->max_idle_balance_cost;
2238 update_avg(&rq->avg_idle, delta);
2240 if (rq->avg_idle > max)
2249 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2250 struct rq_flags *rf)
2252 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2254 lockdep_assert_held(&rq->lock);
2257 if (p->sched_contributes_to_load)
2258 rq->nr_uninterruptible--;
2260 if (wake_flags & WF_MIGRATED)
2261 en_flags |= ENQUEUE_MIGRATED;
2264 activate_task(rq, p, en_flags);
2265 ttwu_do_wakeup(rq, p, wake_flags, rf);
2269 * Called in case the task @p isn't fully descheduled from its runqueue,
2270 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2271 * since all we need to do is flip p->state to TASK_RUNNING, since
2272 * the task is still ->on_rq.
2274 static int ttwu_remote(struct task_struct *p, int wake_flags)
2280 rq = __task_rq_lock(p, &rf);
2281 if (task_on_rq_queued(p)) {
2282 /* check_preempt_curr() may use rq clock */
2283 update_rq_clock(rq);
2284 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2287 __task_rq_unlock(rq, &rf);
2293 void sched_ttwu_pending(void)
2295 struct rq *rq = this_rq();
2296 struct llist_node *llist = llist_del_all(&rq->wake_list);
2297 struct task_struct *p, *t;
2303 rq_lock_irqsave(rq, &rf);
2304 update_rq_clock(rq);
2306 llist_for_each_entry_safe(p, t, llist, wake_entry)
2307 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2309 rq_unlock_irqrestore(rq, &rf);
2312 void scheduler_ipi(void)
2315 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2316 * TIF_NEED_RESCHED remotely (for the first time) will also send
2319 preempt_fold_need_resched();
2321 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2325 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2326 * traditionally all their work was done from the interrupt return
2327 * path. Now that we actually do some work, we need to make sure
2330 * Some archs already do call them, luckily irq_enter/exit nest
2333 * Arguably we should visit all archs and update all handlers,
2334 * however a fair share of IPIs are still resched only so this would
2335 * somewhat pessimize the simple resched case.
2338 sched_ttwu_pending();
2341 * Check if someone kicked us for doing the nohz idle load balance.
2343 if (unlikely(got_nohz_idle_kick())) {
2344 this_rq()->idle_balance = 1;
2345 raise_softirq_irqoff(SCHED_SOFTIRQ);
2350 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2352 struct rq *rq = cpu_rq(cpu);
2354 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2356 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2357 if (!set_nr_if_polling(rq->idle))
2358 smp_send_reschedule(cpu);
2360 trace_sched_wake_idle_without_ipi(cpu);
2364 void wake_up_if_idle(int cpu)
2366 struct rq *rq = cpu_rq(cpu);
2371 if (!is_idle_task(rcu_dereference(rq->curr)))
2374 if (set_nr_if_polling(rq->idle)) {
2375 trace_sched_wake_idle_without_ipi(cpu);
2377 rq_lock_irqsave(rq, &rf);
2378 if (is_idle_task(rq->curr))
2379 smp_send_reschedule(cpu);
2380 /* Else CPU is not idle, do nothing here: */
2381 rq_unlock_irqrestore(rq, &rf);
2388 bool cpus_share_cache(int this_cpu, int that_cpu)
2390 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2392 #endif /* CONFIG_SMP */
2394 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2396 struct rq *rq = cpu_rq(cpu);
2399 #if defined(CONFIG_SMP)
2400 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2401 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2402 ttwu_queue_remote(p, cpu, wake_flags);
2408 update_rq_clock(rq);
2409 ttwu_do_activate(rq, p, wake_flags, &rf);
2414 * Notes on Program-Order guarantees on SMP systems.
2418 * The basic program-order guarantee on SMP systems is that when a task [t]
2419 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2420 * execution on its new CPU [c1].
2422 * For migration (of runnable tasks) this is provided by the following means:
2424 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2425 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2426 * rq(c1)->lock (if not at the same time, then in that order).
2427 * C) LOCK of the rq(c1)->lock scheduling in task
2429 * Release/acquire chaining guarantees that B happens after A and C after B.
2430 * Note: the CPU doing B need not be c0 or c1
2439 * UNLOCK rq(0)->lock
2441 * LOCK rq(0)->lock // orders against CPU0
2443 * UNLOCK rq(0)->lock
2447 * UNLOCK rq(1)->lock
2449 * LOCK rq(1)->lock // orders against CPU2
2452 * UNLOCK rq(1)->lock
2455 * BLOCKING -- aka. SLEEP + WAKEUP
2457 * For blocking we (obviously) need to provide the same guarantee as for
2458 * migration. However the means are completely different as there is no lock
2459 * chain to provide order. Instead we do:
2461 * 1) smp_store_release(X->on_cpu, 0)
2462 * 2) smp_cond_load_acquire(!X->on_cpu)
2466 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2468 * LOCK rq(0)->lock LOCK X->pi_lock
2471 * smp_store_release(X->on_cpu, 0);
2473 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2479 * X->state = RUNNING
2480 * UNLOCK rq(2)->lock
2482 * LOCK rq(2)->lock // orders against CPU1
2485 * UNLOCK rq(2)->lock
2488 * UNLOCK rq(0)->lock
2491 * However, for wakeups there is a second guarantee we must provide, namely we
2492 * must ensure that CONDITION=1 done by the caller can not be reordered with
2493 * accesses to the task state; see try_to_wake_up() and set_current_state().
2497 * try_to_wake_up - wake up a thread
2498 * @p: the thread to be awakened
2499 * @state: the mask of task states that can be woken
2500 * @wake_flags: wake modifier flags (WF_*)
2502 * If (@state & @p->state) @p->state = TASK_RUNNING.
2504 * If the task was not queued/runnable, also place it back on a runqueue.
2506 * Atomic against schedule() which would dequeue a task, also see
2507 * set_current_state().
2509 * This function executes a full memory barrier before accessing the task
2510 * state; see set_current_state().
2512 * Return: %true if @p->state changes (an actual wakeup was done),
2516 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2518 unsigned long flags;
2519 int cpu, success = 0;
2524 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2525 * == smp_processor_id()'. Together this means we can special
2526 * case the whole 'p->on_rq && ttwu_remote()' case below
2527 * without taking any locks.
2530 * - we rely on Program-Order guarantees for all the ordering,
2531 * - we're serialized against set_special_state() by virtue of
2532 * it disabling IRQs (this allows not taking ->pi_lock).
2534 if (!(p->state & state))
2539 trace_sched_waking(p);
2540 p->state = TASK_RUNNING;
2541 trace_sched_wakeup(p);
2546 * If we are going to wake up a thread waiting for CONDITION we
2547 * need to ensure that CONDITION=1 done by the caller can not be
2548 * reordered with p->state check below. This pairs with mb() in
2549 * set_current_state() the waiting thread does.
2551 raw_spin_lock_irqsave(&p->pi_lock, flags);
2552 smp_mb__after_spinlock();
2553 if (!(p->state & state))
2556 trace_sched_waking(p);
2558 /* We're going to change ->state: */
2563 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2564 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2565 * in smp_cond_load_acquire() below.
2567 * sched_ttwu_pending() try_to_wake_up()
2568 * STORE p->on_rq = 1 LOAD p->state
2571 * __schedule() (switch to task 'p')
2572 * LOCK rq->lock smp_rmb();
2573 * smp_mb__after_spinlock();
2577 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2579 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2580 * __schedule(). See the comment for smp_mb__after_spinlock().
2583 if (p->on_rq && ttwu_remote(p, wake_flags))
2588 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2589 * possible to, falsely, observe p->on_cpu == 0.
2591 * One must be running (->on_cpu == 1) in order to remove oneself
2592 * from the runqueue.
2594 * __schedule() (switch to task 'p') try_to_wake_up()
2595 * STORE p->on_cpu = 1 LOAD p->on_rq
2598 * __schedule() (put 'p' to sleep)
2599 * LOCK rq->lock smp_rmb();
2600 * smp_mb__after_spinlock();
2601 * STORE p->on_rq = 0 LOAD p->on_cpu
2603 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2604 * __schedule(). See the comment for smp_mb__after_spinlock().
2609 * If the owning (remote) CPU is still in the middle of schedule() with
2610 * this task as prev, wait until its done referencing the task.
2612 * Pairs with the smp_store_release() in finish_task().
2614 * This ensures that tasks getting woken will be fully ordered against
2615 * their previous state and preserve Program Order.
2617 smp_cond_load_acquire(&p->on_cpu, !VAL);
2619 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2620 p->state = TASK_WAKING;
2623 delayacct_blkio_end(p);
2624 atomic_dec(&task_rq(p)->nr_iowait);
2627 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2628 if (task_cpu(p) != cpu) {
2629 wake_flags |= WF_MIGRATED;
2630 psi_ttwu_dequeue(p);
2631 set_task_cpu(p, cpu);
2634 #else /* CONFIG_SMP */
2637 delayacct_blkio_end(p);
2638 atomic_dec(&task_rq(p)->nr_iowait);
2641 #endif /* CONFIG_SMP */
2643 ttwu_queue(p, cpu, wake_flags);
2645 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2648 ttwu_stat(p, cpu, wake_flags);
2655 * wake_up_process - Wake up a specific process
2656 * @p: The process to be woken up.
2658 * Attempt to wake up the nominated process and move it to the set of runnable
2661 * Return: 1 if the process was woken up, 0 if it was already running.
2663 * This function executes a full memory barrier before accessing the task state.
2665 int wake_up_process(struct task_struct *p)
2667 return try_to_wake_up(p, TASK_NORMAL, 0);
2669 EXPORT_SYMBOL(wake_up_process);
2671 int wake_up_state(struct task_struct *p, unsigned int state)
2673 return try_to_wake_up(p, state, 0);
2677 * Perform scheduler related setup for a newly forked process p.
2678 * p is forked by current.
2680 * __sched_fork() is basic setup used by init_idle() too:
2682 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2687 p->se.exec_start = 0;
2688 p->se.sum_exec_runtime = 0;
2689 p->se.prev_sum_exec_runtime = 0;
2690 p->se.nr_migrations = 0;
2692 INIT_LIST_HEAD(&p->se.group_node);
2694 #ifdef CONFIG_FAIR_GROUP_SCHED
2695 p->se.cfs_rq = NULL;
2698 #ifdef CONFIG_SCHEDSTATS
2699 /* Even if schedstat is disabled, there should not be garbage */
2700 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2703 RB_CLEAR_NODE(&p->dl.rb_node);
2704 init_dl_task_timer(&p->dl);
2705 init_dl_inactive_task_timer(&p->dl);
2706 __dl_clear_params(p);
2708 INIT_LIST_HEAD(&p->rt.run_list);
2710 p->rt.time_slice = sched_rr_timeslice;
2714 #ifdef CONFIG_PREEMPT_NOTIFIERS
2715 INIT_HLIST_HEAD(&p->preempt_notifiers);
2718 #ifdef CONFIG_COMPACTION
2719 p->capture_control = NULL;
2721 init_numa_balancing(clone_flags, p);
2724 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2726 #ifdef CONFIG_NUMA_BALANCING
2728 void set_numabalancing_state(bool enabled)
2731 static_branch_enable(&sched_numa_balancing);
2733 static_branch_disable(&sched_numa_balancing);
2736 #ifdef CONFIG_PROC_SYSCTL
2737 int sysctl_numa_balancing(struct ctl_table *table, int write,
2738 void __user *buffer, size_t *lenp, loff_t *ppos)
2742 int state = static_branch_likely(&sched_numa_balancing);
2744 if (write && !capable(CAP_SYS_ADMIN))
2749 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2753 set_numabalancing_state(state);
2759 #ifdef CONFIG_SCHEDSTATS
2761 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2762 static bool __initdata __sched_schedstats = false;
2764 static void set_schedstats(bool enabled)
2767 static_branch_enable(&sched_schedstats);
2769 static_branch_disable(&sched_schedstats);
2772 void force_schedstat_enabled(void)
2774 if (!schedstat_enabled()) {
2775 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2776 static_branch_enable(&sched_schedstats);
2780 static int __init setup_schedstats(char *str)
2787 * This code is called before jump labels have been set up, so we can't
2788 * change the static branch directly just yet. Instead set a temporary
2789 * variable so init_schedstats() can do it later.
2791 if (!strcmp(str, "enable")) {
2792 __sched_schedstats = true;
2794 } else if (!strcmp(str, "disable")) {
2795 __sched_schedstats = false;
2800 pr_warn("Unable to parse schedstats=\n");
2804 __setup("schedstats=", setup_schedstats);
2806 static void __init init_schedstats(void)
2808 set_schedstats(__sched_schedstats);
2811 #ifdef CONFIG_PROC_SYSCTL
2812 int sysctl_schedstats(struct ctl_table *table, int write,
2813 void __user *buffer, size_t *lenp, loff_t *ppos)
2817 int state = static_branch_likely(&sched_schedstats);
2819 if (write && !capable(CAP_SYS_ADMIN))
2824 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2828 set_schedstats(state);
2831 #endif /* CONFIG_PROC_SYSCTL */
2832 #else /* !CONFIG_SCHEDSTATS */
2833 static inline void init_schedstats(void) {}
2834 #endif /* CONFIG_SCHEDSTATS */
2837 * fork()/clone()-time setup:
2839 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2841 unsigned long flags;
2843 __sched_fork(clone_flags, p);
2845 * We mark the process as NEW here. This guarantees that
2846 * nobody will actually run it, and a signal or other external
2847 * event cannot wake it up and insert it on the runqueue either.
2849 p->state = TASK_NEW;
2852 * Make sure we do not leak PI boosting priority to the child.
2854 p->prio = current->normal_prio;
2859 * Revert to default priority/policy on fork if requested.
2861 if (unlikely(p->sched_reset_on_fork)) {
2862 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2863 p->policy = SCHED_NORMAL;
2864 p->static_prio = NICE_TO_PRIO(0);
2866 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2867 p->static_prio = NICE_TO_PRIO(0);
2869 p->prio = p->normal_prio = __normal_prio(p);
2870 set_load_weight(p, false);
2873 * We don't need the reset flag anymore after the fork. It has
2874 * fulfilled its duty:
2876 p->sched_reset_on_fork = 0;
2879 if (dl_prio(p->prio))
2881 else if (rt_prio(p->prio))
2882 p->sched_class = &rt_sched_class;
2884 p->sched_class = &fair_sched_class;
2886 init_entity_runnable_average(&p->se);
2889 * The child is not yet in the pid-hash so no cgroup attach races,
2890 * and the cgroup is pinned to this child due to cgroup_fork()
2891 * is ran before sched_fork().
2893 * Silence PROVE_RCU.
2895 raw_spin_lock_irqsave(&p->pi_lock, flags);
2897 * We're setting the CPU for the first time, we don't migrate,
2898 * so use __set_task_cpu().
2900 __set_task_cpu(p, smp_processor_id());
2901 if (p->sched_class->task_fork)
2902 p->sched_class->task_fork(p);
2903 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2905 #ifdef CONFIG_SCHED_INFO
2906 if (likely(sched_info_on()))
2907 memset(&p->sched_info, 0, sizeof(p->sched_info));
2909 #if defined(CONFIG_SMP)
2912 init_task_preempt_count(p);
2914 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2915 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2920 unsigned long to_ratio(u64 period, u64 runtime)
2922 if (runtime == RUNTIME_INF)
2926 * Doing this here saves a lot of checks in all
2927 * the calling paths, and returning zero seems
2928 * safe for them anyway.
2933 return div64_u64(runtime << BW_SHIFT, period);
2937 * wake_up_new_task - wake up a newly created task for the first time.
2939 * This function will do some initial scheduler statistics housekeeping
2940 * that must be done for every newly created context, then puts the task
2941 * on the runqueue and wakes it.
2943 void wake_up_new_task(struct task_struct *p)
2948 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2949 p->state = TASK_RUNNING;
2952 * Fork balancing, do it here and not earlier because:
2953 * - cpus_ptr can change in the fork path
2954 * - any previously selected CPU might disappear through hotplug
2956 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2957 * as we're not fully set-up yet.
2959 p->recent_used_cpu = task_cpu(p);
2960 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2962 rq = __task_rq_lock(p, &rf);
2963 update_rq_clock(rq);
2964 post_init_entity_util_avg(p);
2966 activate_task(rq, p, ENQUEUE_NOCLOCK);
2967 trace_sched_wakeup_new(p);
2968 check_preempt_curr(rq, p, WF_FORK);
2970 if (p->sched_class->task_woken) {
2972 * Nothing relies on rq->lock after this, so its fine to
2975 rq_unpin_lock(rq, &rf);
2976 p->sched_class->task_woken(rq, p);
2977 rq_repin_lock(rq, &rf);
2980 task_rq_unlock(rq, p, &rf);
2983 #ifdef CONFIG_PREEMPT_NOTIFIERS
2985 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2987 void preempt_notifier_inc(void)
2989 static_branch_inc(&preempt_notifier_key);
2991 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2993 void preempt_notifier_dec(void)
2995 static_branch_dec(&preempt_notifier_key);
2997 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3000 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3001 * @notifier: notifier struct to register
3003 void preempt_notifier_register(struct preempt_notifier *notifier)
3005 if (!static_branch_unlikely(&preempt_notifier_key))
3006 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3008 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3010 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3013 * preempt_notifier_unregister - no longer interested in preemption notifications
3014 * @notifier: notifier struct to unregister
3016 * This is *not* safe to call from within a preemption notifier.
3018 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3020 hlist_del(¬ifier->link);
3022 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3024 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3026 struct preempt_notifier *notifier;
3028 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3029 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3032 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3034 if (static_branch_unlikely(&preempt_notifier_key))
3035 __fire_sched_in_preempt_notifiers(curr);
3039 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3040 struct task_struct *next)
3042 struct preempt_notifier *notifier;
3044 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3045 notifier->ops->sched_out(notifier, next);
3048 static __always_inline void
3049 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3050 struct task_struct *next)
3052 if (static_branch_unlikely(&preempt_notifier_key))
3053 __fire_sched_out_preempt_notifiers(curr, next);
3056 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3058 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3063 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3064 struct task_struct *next)
3068 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3070 static inline void prepare_task(struct task_struct *next)
3074 * Claim the task as running, we do this before switching to it
3075 * such that any running task will have this set.
3081 static inline void finish_task(struct task_struct *prev)
3085 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3086 * We must ensure this doesn't happen until the switch is completely
3089 * In particular, the load of prev->state in finish_task_switch() must
3090 * happen before this.
3092 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3094 smp_store_release(&prev->on_cpu, 0);
3099 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3102 * Since the runqueue lock will be released by the next
3103 * task (which is an invalid locking op but in the case
3104 * of the scheduler it's an obvious special-case), so we
3105 * do an early lockdep release here:
3107 rq_unpin_lock(rq, rf);
3108 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3109 #ifdef CONFIG_DEBUG_SPINLOCK
3110 /* this is a valid case when another task releases the spinlock */
3111 rq->lock.owner = next;
3115 static inline void finish_lock_switch(struct rq *rq)
3118 * If we are tracking spinlock dependencies then we have to
3119 * fix up the runqueue lock - which gets 'carried over' from
3120 * prev into current:
3122 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3123 raw_spin_unlock_irq(&rq->lock);
3127 * NOP if the arch has not defined these:
3130 #ifndef prepare_arch_switch
3131 # define prepare_arch_switch(next) do { } while (0)
3134 #ifndef finish_arch_post_lock_switch
3135 # define finish_arch_post_lock_switch() do { } while (0)
3139 * prepare_task_switch - prepare to switch tasks
3140 * @rq: the runqueue preparing to switch
3141 * @prev: the current task that is being switched out
3142 * @next: the task we are going to switch to.
3144 * This is called with the rq lock held and interrupts off. It must
3145 * be paired with a subsequent finish_task_switch after the context
3148 * prepare_task_switch sets up locking and calls architecture specific
3152 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3153 struct task_struct *next)
3155 kcov_prepare_switch(prev);
3156 sched_info_switch(rq, prev, next);
3157 perf_event_task_sched_out(prev, next);
3159 fire_sched_out_preempt_notifiers(prev, next);
3161 prepare_arch_switch(next);
3165 * finish_task_switch - clean up after a task-switch
3166 * @prev: the thread we just switched away from.
3168 * finish_task_switch must be called after the context switch, paired
3169 * with a prepare_task_switch call before the context switch.
3170 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3171 * and do any other architecture-specific cleanup actions.
3173 * Note that we may have delayed dropping an mm in context_switch(). If
3174 * so, we finish that here outside of the runqueue lock. (Doing it
3175 * with the lock held can cause deadlocks; see schedule() for
3178 * The context switch have flipped the stack from under us and restored the
3179 * local variables which were saved when this task called schedule() in the
3180 * past. prev == current is still correct but we need to recalculate this_rq
3181 * because prev may have moved to another CPU.
3183 static struct rq *finish_task_switch(struct task_struct *prev)
3184 __releases(rq->lock)
3186 struct rq *rq = this_rq();
3187 struct mm_struct *mm = rq->prev_mm;
3191 * The previous task will have left us with a preempt_count of 2
3192 * because it left us after:
3195 * preempt_disable(); // 1
3197 * raw_spin_lock_irq(&rq->lock) // 2
3199 * Also, see FORK_PREEMPT_COUNT.
3201 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3202 "corrupted preempt_count: %s/%d/0x%x\n",
3203 current->comm, current->pid, preempt_count()))
3204 preempt_count_set(FORK_PREEMPT_COUNT);
3209 * A task struct has one reference for the use as "current".
3210 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3211 * schedule one last time. The schedule call will never return, and
3212 * the scheduled task must drop that reference.
3214 * We must observe prev->state before clearing prev->on_cpu (in
3215 * finish_task), otherwise a concurrent wakeup can get prev
3216 * running on another CPU and we could rave with its RUNNING -> DEAD
3217 * transition, resulting in a double drop.
3219 prev_state = prev->state;
3220 vtime_task_switch(prev);
3221 perf_event_task_sched_in(prev, current);
3223 finish_lock_switch(rq);
3224 finish_arch_post_lock_switch();
3225 kcov_finish_switch(current);
3227 fire_sched_in_preempt_notifiers(current);
3229 * When switching through a kernel thread, the loop in
3230 * membarrier_{private,global}_expedited() may have observed that
3231 * kernel thread and not issued an IPI. It is therefore possible to
3232 * schedule between user->kernel->user threads without passing though
3233 * switch_mm(). Membarrier requires a barrier after storing to
3234 * rq->curr, before returning to userspace, so provide them here:
3236 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3237 * provided by mmdrop(),
3238 * - a sync_core for SYNC_CORE.
3241 membarrier_mm_sync_core_before_usermode(mm);
3244 if (unlikely(prev_state == TASK_DEAD)) {
3245 if (prev->sched_class->task_dead)
3246 prev->sched_class->task_dead(prev);
3249 * Remove function-return probe instances associated with this
3250 * task and put them back on the free list.
3252 kprobe_flush_task(prev);
3254 /* Task is done with its stack. */
3255 put_task_stack(prev);
3257 put_task_struct_rcu_user(prev);
3260 tick_nohz_task_switch();
3266 /* rq->lock is NOT held, but preemption is disabled */
3267 static void __balance_callback(struct rq *rq)
3269 struct callback_head *head, *next;
3270 void (*func)(struct rq *rq);
3271 unsigned long flags;
3273 raw_spin_lock_irqsave(&rq->lock, flags);
3274 head = rq->balance_callback;
3275 rq->balance_callback = NULL;
3277 func = (void (*)(struct rq *))head->func;
3284 raw_spin_unlock_irqrestore(&rq->lock, flags);
3287 static inline void balance_callback(struct rq *rq)
3289 if (unlikely(rq->balance_callback))
3290 __balance_callback(rq);
3295 static inline void balance_callback(struct rq *rq)
3302 * schedule_tail - first thing a freshly forked thread must call.
3303 * @prev: the thread we just switched away from.
3305 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3306 __releases(rq->lock)
3311 * New tasks start with FORK_PREEMPT_COUNT, see there and
3312 * finish_task_switch() for details.
3314 * finish_task_switch() will drop rq->lock() and lower preempt_count
3315 * and the preempt_enable() will end up enabling preemption (on
3316 * PREEMPT_COUNT kernels).
3319 rq = finish_task_switch(prev);
3320 balance_callback(rq);
3323 if (current->set_child_tid)
3324 put_user(task_pid_vnr(current), current->set_child_tid);
3326 calculate_sigpending();
3330 * context_switch - switch to the new MM and the new thread's register state.
3332 static __always_inline struct rq *
3333 context_switch(struct rq *rq, struct task_struct *prev,
3334 struct task_struct *next, struct rq_flags *rf)
3336 prepare_task_switch(rq, prev, next);
3339 * For paravirt, this is coupled with an exit in switch_to to
3340 * combine the page table reload and the switch backend into
3343 arch_start_context_switch(prev);
3346 * kernel -> kernel lazy + transfer active
3347 * user -> kernel lazy + mmgrab() active
3349 * kernel -> user switch + mmdrop() active
3350 * user -> user switch
3352 if (!next->mm) { // to kernel
3353 enter_lazy_tlb(prev->active_mm, next);
3355 next->active_mm = prev->active_mm;
3356 if (prev->mm) // from user
3357 mmgrab(prev->active_mm);
3359 prev->active_mm = NULL;
3361 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3363 * sys_membarrier() requires an smp_mb() between setting
3364 * rq->curr / membarrier_switch_mm() and returning to userspace.
3366 * The below provides this either through switch_mm(), or in
3367 * case 'prev->active_mm == next->mm' through
3368 * finish_task_switch()'s mmdrop().
3370 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3372 if (!prev->mm) { // from kernel
3373 /* will mmdrop() in finish_task_switch(). */
3374 rq->prev_mm = prev->active_mm;
3375 prev->active_mm = NULL;
3379 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3381 prepare_lock_switch(rq, next, rf);
3383 /* Here we just switch the register state and the stack. */
3384 switch_to(prev, next, prev);
3387 return finish_task_switch(prev);
3391 * nr_running and nr_context_switches:
3393 * externally visible scheduler statistics: current number of runnable
3394 * threads, total number of context switches performed since bootup.
3396 unsigned long nr_running(void)
3398 unsigned long i, sum = 0;
3400 for_each_online_cpu(i)
3401 sum += cpu_rq(i)->nr_running;
3407 * Check if only the current task is running on the CPU.
3409 * Caution: this function does not check that the caller has disabled
3410 * preemption, thus the result might have a time-of-check-to-time-of-use
3411 * race. The caller is responsible to use it correctly, for example:
3413 * - from a non-preemptible section (of course)
3415 * - from a thread that is bound to a single CPU
3417 * - in a loop with very short iterations (e.g. a polling loop)
3419 bool single_task_running(void)
3421 return raw_rq()->nr_running == 1;
3423 EXPORT_SYMBOL(single_task_running);
3425 unsigned long long nr_context_switches(void)
3428 unsigned long long sum = 0;
3430 for_each_possible_cpu(i)
3431 sum += cpu_rq(i)->nr_switches;
3437 * Consumers of these two interfaces, like for example the cpuidle menu
3438 * governor, are using nonsensical data. Preferring shallow idle state selection
3439 * for a CPU that has IO-wait which might not even end up running the task when
3440 * it does become runnable.
3443 unsigned long nr_iowait_cpu(int cpu)
3445 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3449 * IO-wait accounting, and how its mostly bollocks (on SMP).
3451 * The idea behind IO-wait account is to account the idle time that we could
3452 * have spend running if it were not for IO. That is, if we were to improve the
3453 * storage performance, we'd have a proportional reduction in IO-wait time.
3455 * This all works nicely on UP, where, when a task blocks on IO, we account
3456 * idle time as IO-wait, because if the storage were faster, it could've been
3457 * running and we'd not be idle.
3459 * This has been extended to SMP, by doing the same for each CPU. This however
3462 * Imagine for instance the case where two tasks block on one CPU, only the one
3463 * CPU will have IO-wait accounted, while the other has regular idle. Even
3464 * though, if the storage were faster, both could've ran at the same time,
3465 * utilising both CPUs.
3467 * This means, that when looking globally, the current IO-wait accounting on
3468 * SMP is a lower bound, by reason of under accounting.
3470 * Worse, since the numbers are provided per CPU, they are sometimes
3471 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3472 * associated with any one particular CPU, it can wake to another CPU than it
3473 * blocked on. This means the per CPU IO-wait number is meaningless.
3475 * Task CPU affinities can make all that even more 'interesting'.
3478 unsigned long nr_iowait(void)
3480 unsigned long i, sum = 0;
3482 for_each_possible_cpu(i)
3483 sum += nr_iowait_cpu(i);
3491 * sched_exec - execve() is a valuable balancing opportunity, because at
3492 * this point the task has the smallest effective memory and cache footprint.
3494 void sched_exec(void)
3496 struct task_struct *p = current;
3497 unsigned long flags;
3500 raw_spin_lock_irqsave(&p->pi_lock, flags);
3501 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3502 if (dest_cpu == smp_processor_id())
3505 if (likely(cpu_active(dest_cpu))) {
3506 struct migration_arg arg = { p, dest_cpu };
3508 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3509 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3513 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3518 DEFINE_PER_CPU(struct kernel_stat, kstat);
3519 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3521 EXPORT_PER_CPU_SYMBOL(kstat);
3522 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3525 * The function fair_sched_class.update_curr accesses the struct curr
3526 * and its field curr->exec_start; when called from task_sched_runtime(),
3527 * we observe a high rate of cache misses in practice.
3528 * Prefetching this data results in improved performance.
3530 static inline void prefetch_curr_exec_start(struct task_struct *p)
3532 #ifdef CONFIG_FAIR_GROUP_SCHED
3533 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3535 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3538 prefetch(&curr->exec_start);
3542 * Return accounted runtime for the task.
3543 * In case the task is currently running, return the runtime plus current's
3544 * pending runtime that have not been accounted yet.
3546 unsigned long long task_sched_runtime(struct task_struct *p)
3552 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3554 * 64-bit doesn't need locks to atomically read a 64-bit value.
3555 * So we have a optimization chance when the task's delta_exec is 0.
3556 * Reading ->on_cpu is racy, but this is ok.
3558 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3559 * If we race with it entering CPU, unaccounted time is 0. This is
3560 * indistinguishable from the read occurring a few cycles earlier.
3561 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3562 * been accounted, so we're correct here as well.
3564 if (!p->on_cpu || !task_on_rq_queued(p))
3565 return p->se.sum_exec_runtime;
3568 rq = task_rq_lock(p, &rf);
3570 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3571 * project cycles that may never be accounted to this
3572 * thread, breaking clock_gettime().
3574 if (task_current(rq, p) && task_on_rq_queued(p)) {
3575 prefetch_curr_exec_start(p);
3576 update_rq_clock(rq);
3577 p->sched_class->update_curr(rq);
3579 ns = p->se.sum_exec_runtime;
3580 task_rq_unlock(rq, p, &rf);
3586 * This function gets called by the timer code, with HZ frequency.
3587 * We call it with interrupts disabled.
3589 void scheduler_tick(void)
3591 int cpu = smp_processor_id();
3592 struct rq *rq = cpu_rq(cpu);
3593 struct task_struct *curr = rq->curr;
3600 update_rq_clock(rq);
3601 curr->sched_class->task_tick(rq, curr, 0);
3602 calc_global_load_tick(rq);
3607 perf_event_task_tick();
3610 rq->idle_balance = idle_cpu(cpu);
3611 trigger_load_balance(rq);
3615 #ifdef CONFIG_NO_HZ_FULL
3620 struct delayed_work work;
3622 /* Values for ->state, see diagram below. */
3623 #define TICK_SCHED_REMOTE_OFFLINE 0
3624 #define TICK_SCHED_REMOTE_OFFLINING 1
3625 #define TICK_SCHED_REMOTE_RUNNING 2
3628 * State diagram for ->state:
3631 * TICK_SCHED_REMOTE_OFFLINE
3634 * | | sched_tick_remote()
3637 * +--TICK_SCHED_REMOTE_OFFLINING
3640 * sched_tick_start() | | sched_tick_stop()
3643 * TICK_SCHED_REMOTE_RUNNING
3646 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3647 * and sched_tick_start() are happy to leave the state in RUNNING.
3650 static struct tick_work __percpu *tick_work_cpu;
3652 static void sched_tick_remote(struct work_struct *work)
3654 struct delayed_work *dwork = to_delayed_work(work);
3655 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3656 int cpu = twork->cpu;
3657 struct rq *rq = cpu_rq(cpu);
3658 struct task_struct *curr;
3664 * Handle the tick only if it appears the remote CPU is running in full
3665 * dynticks mode. The check is racy by nature, but missing a tick or
3666 * having one too much is no big deal because the scheduler tick updates
3667 * statistics and checks timeslices in a time-independent way, regardless
3668 * of when exactly it is running.
3670 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3673 rq_lock_irq(rq, &rf);
3675 if (is_idle_task(curr) || cpu_is_offline(cpu))
3678 update_rq_clock(rq);
3679 delta = rq_clock_task(rq) - curr->se.exec_start;
3682 * Make sure the next tick runs within a reasonable
3685 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3686 curr->sched_class->task_tick(rq, curr, 0);
3689 rq_unlock_irq(rq, &rf);
3693 * Run the remote tick once per second (1Hz). This arbitrary
3694 * frequency is large enough to avoid overload but short enough
3695 * to keep scheduler internal stats reasonably up to date. But
3696 * first update state to reflect hotplug activity if required.
3698 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3699 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3700 if (os == TICK_SCHED_REMOTE_RUNNING)
3701 queue_delayed_work(system_unbound_wq, dwork, HZ);
3704 static void sched_tick_start(int cpu)
3707 struct tick_work *twork;
3709 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3712 WARN_ON_ONCE(!tick_work_cpu);
3714 twork = per_cpu_ptr(tick_work_cpu, cpu);
3715 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3716 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3717 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3719 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3720 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3724 #ifdef CONFIG_HOTPLUG_CPU
3725 static void sched_tick_stop(int cpu)
3727 struct tick_work *twork;
3730 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3733 WARN_ON_ONCE(!tick_work_cpu);
3735 twork = per_cpu_ptr(tick_work_cpu, cpu);
3736 /* There cannot be competing actions, but don't rely on stop-machine. */
3737 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3738 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3739 /* Don't cancel, as this would mess up the state machine. */
3741 #endif /* CONFIG_HOTPLUG_CPU */
3743 int __init sched_tick_offload_init(void)
3745 tick_work_cpu = alloc_percpu(struct tick_work);
3746 BUG_ON(!tick_work_cpu);
3750 #else /* !CONFIG_NO_HZ_FULL */
3751 static inline void sched_tick_start(int cpu) { }
3752 static inline void sched_tick_stop(int cpu) { }
3755 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3756 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3758 * If the value passed in is equal to the current preempt count
3759 * then we just disabled preemption. Start timing the latency.
3761 static inline void preempt_latency_start(int val)
3763 if (preempt_count() == val) {
3764 unsigned long ip = get_lock_parent_ip();
3765 #ifdef CONFIG_DEBUG_PREEMPT
3766 current->preempt_disable_ip = ip;
3768 trace_preempt_off(CALLER_ADDR0, ip);
3772 void preempt_count_add(int val)
3774 #ifdef CONFIG_DEBUG_PREEMPT
3778 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3781 __preempt_count_add(val);
3782 #ifdef CONFIG_DEBUG_PREEMPT
3784 * Spinlock count overflowing soon?
3786 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3789 preempt_latency_start(val);
3791 EXPORT_SYMBOL(preempt_count_add);
3792 NOKPROBE_SYMBOL(preempt_count_add);
3795 * If the value passed in equals to the current preempt count
3796 * then we just enabled preemption. Stop timing the latency.
3798 static inline void preempt_latency_stop(int val)
3800 if (preempt_count() == val)
3801 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3804 void preempt_count_sub(int val)
3806 #ifdef CONFIG_DEBUG_PREEMPT
3810 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3813 * Is the spinlock portion underflowing?
3815 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3816 !(preempt_count() & PREEMPT_MASK)))
3820 preempt_latency_stop(val);
3821 __preempt_count_sub(val);
3823 EXPORT_SYMBOL(preempt_count_sub);
3824 NOKPROBE_SYMBOL(preempt_count_sub);
3827 static inline void preempt_latency_start(int val) { }
3828 static inline void preempt_latency_stop(int val) { }
3831 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3833 #ifdef CONFIG_DEBUG_PREEMPT
3834 return p->preempt_disable_ip;
3841 * Print scheduling while atomic bug:
3843 static noinline void __schedule_bug(struct task_struct *prev)
3845 /* Save this before calling printk(), since that will clobber it */
3846 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3848 if (oops_in_progress)
3851 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3852 prev->comm, prev->pid, preempt_count());
3854 debug_show_held_locks(prev);
3856 if (irqs_disabled())
3857 print_irqtrace_events(prev);
3858 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3859 && in_atomic_preempt_off()) {
3860 pr_err("Preemption disabled at:");
3861 print_ip_sym(preempt_disable_ip);
3865 panic("scheduling while atomic\n");
3868 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3872 * Various schedule()-time debugging checks and statistics:
3874 static inline void schedule_debug(struct task_struct *prev)
3876 #ifdef CONFIG_SCHED_STACK_END_CHECK
3877 if (task_stack_end_corrupted(prev))
3878 panic("corrupted stack end detected inside scheduler\n");
3881 if (unlikely(in_atomic_preempt_off())) {
3882 __schedule_bug(prev);
3883 preempt_count_set(PREEMPT_DISABLED);
3887 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3889 schedstat_inc(this_rq()->sched_count);
3893 * Pick up the highest-prio task:
3895 static inline struct task_struct *
3896 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3898 const struct sched_class *class;
3899 struct task_struct *p;
3902 * Optimization: we know that if all tasks are in the fair class we can
3903 * call that function directly, but only if the @prev task wasn't of a
3904 * higher scheduling class, because otherwise those loose the
3905 * opportunity to pull in more work from other CPUs.
3907 if (likely((prev->sched_class == &idle_sched_class ||
3908 prev->sched_class == &fair_sched_class) &&
3909 rq->nr_running == rq->cfs.h_nr_running)) {
3911 p = fair_sched_class.pick_next_task(rq, prev, rf);
3912 if (unlikely(p == RETRY_TASK))
3915 /* Assumes fair_sched_class->next == idle_sched_class */
3917 p = idle_sched_class.pick_next_task(rq, prev, rf);
3924 * Ensure that we put DL/RT tasks before the pick loop, such that they
3925 * can PULL higher prio tasks when we lower the RQ 'priority'.
3927 prev->sched_class->put_prev_task(rq, prev, rf);
3928 if (!rq->nr_running)
3929 newidle_balance(rq, rf);
3931 for_each_class(class) {
3932 p = class->pick_next_task(rq, NULL, NULL);
3937 /* The idle class should always have a runnable task: */
3942 * __schedule() is the main scheduler function.
3944 * The main means of driving the scheduler and thus entering this function are:
3946 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3948 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3949 * paths. For example, see arch/x86/entry_64.S.
3951 * To drive preemption between tasks, the scheduler sets the flag in timer
3952 * interrupt handler scheduler_tick().
3954 * 3. Wakeups don't really cause entry into schedule(). They add a
3955 * task to the run-queue and that's it.
3957 * Now, if the new task added to the run-queue preempts the current
3958 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3959 * called on the nearest possible occasion:
3961 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
3963 * - in syscall or exception context, at the next outmost
3964 * preempt_enable(). (this might be as soon as the wake_up()'s
3967 * - in IRQ context, return from interrupt-handler to
3968 * preemptible context
3970 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
3973 * - cond_resched() call
3974 * - explicit schedule() call
3975 * - return from syscall or exception to user-space
3976 * - return from interrupt-handler to user-space
3978 * WARNING: must be called with preemption disabled!
3980 static void __sched notrace __schedule(bool preempt)
3982 struct task_struct *prev, *next;
3983 unsigned long *switch_count;
3988 cpu = smp_processor_id();
3992 schedule_debug(prev);
3994 if (sched_feat(HRTICK))
3997 local_irq_disable();
3998 rcu_note_context_switch(preempt);
4001 * Make sure that signal_pending_state()->signal_pending() below
4002 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4003 * done by the caller to avoid the race with signal_wake_up().
4005 * The membarrier system call requires a full memory barrier
4006 * after coming from user-space, before storing to rq->curr.
4009 smp_mb__after_spinlock();
4011 /* Promote REQ to ACT */
4012 rq->clock_update_flags <<= 1;
4013 update_rq_clock(rq);
4015 switch_count = &prev->nivcsw;
4016 if (!preempt && prev->state) {
4017 if (signal_pending_state(prev->state, prev)) {
4018 prev->state = TASK_RUNNING;
4020 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4022 if (prev->in_iowait) {
4023 atomic_inc(&rq->nr_iowait);
4024 delayacct_blkio_start();
4027 switch_count = &prev->nvcsw;
4030 next = pick_next_task(rq, prev, &rf);
4031 clear_tsk_need_resched(prev);
4032 clear_preempt_need_resched();
4034 if (likely(prev != next)) {
4037 * RCU users of rcu_dereference(rq->curr) may not see
4038 * changes to task_struct made by pick_next_task().
4040 RCU_INIT_POINTER(rq->curr, next);
4042 * The membarrier system call requires each architecture
4043 * to have a full memory barrier after updating
4044 * rq->curr, before returning to user-space.
4046 * Here are the schemes providing that barrier on the
4047 * various architectures:
4048 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4049 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4050 * - finish_lock_switch() for weakly-ordered
4051 * architectures where spin_unlock is a full barrier,
4052 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4053 * is a RELEASE barrier),
4057 trace_sched_switch(preempt, prev, next);
4059 /* Also unlocks the rq: */
4060 rq = context_switch(rq, prev, next, &rf);
4062 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4063 rq_unlock_irq(rq, &rf);
4066 balance_callback(rq);
4069 void __noreturn do_task_dead(void)
4071 /* Causes final put_task_struct in finish_task_switch(): */
4072 set_special_state(TASK_DEAD);
4074 /* Tell freezer to ignore us: */
4075 current->flags |= PF_NOFREEZE;
4080 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4085 static inline void sched_submit_work(struct task_struct *tsk)
4091 * If a worker went to sleep, notify and ask workqueue whether
4092 * it wants to wake up a task to maintain concurrency.
4093 * As this function is called inside the schedule() context,
4094 * we disable preemption to avoid it calling schedule() again
4095 * in the possible wakeup of a kworker.
4097 if (tsk->flags & PF_WQ_WORKER) {
4099 wq_worker_sleeping(tsk);
4100 preempt_enable_no_resched();
4103 if (tsk_is_pi_blocked(tsk))
4107 * If we are going to sleep and we have plugged IO queued,
4108 * make sure to submit it to avoid deadlocks.
4110 if (blk_needs_flush_plug(tsk))
4111 blk_schedule_flush_plug(tsk);
4114 static void sched_update_worker(struct task_struct *tsk)
4116 if (tsk->flags & PF_WQ_WORKER)
4117 wq_worker_running(tsk);
4120 asmlinkage __visible void __sched schedule(void)
4122 struct task_struct *tsk = current;
4124 sched_submit_work(tsk);
4128 sched_preempt_enable_no_resched();
4129 } while (need_resched());
4130 sched_update_worker(tsk);
4132 EXPORT_SYMBOL(schedule);
4135 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4136 * state (have scheduled out non-voluntarily) by making sure that all
4137 * tasks have either left the run queue or have gone into user space.
4138 * As idle tasks do not do either, they must not ever be preempted
4139 * (schedule out non-voluntarily).
4141 * schedule_idle() is similar to schedule_preempt_disable() except that it
4142 * never enables preemption because it does not call sched_submit_work().
4144 void __sched schedule_idle(void)
4147 * As this skips calling sched_submit_work(), which the idle task does
4148 * regardless because that function is a nop when the task is in a
4149 * TASK_RUNNING state, make sure this isn't used someplace that the
4150 * current task can be in any other state. Note, idle is always in the
4151 * TASK_RUNNING state.
4153 WARN_ON_ONCE(current->state);
4156 } while (need_resched());
4159 #ifdef CONFIG_CONTEXT_TRACKING
4160 asmlinkage __visible void __sched schedule_user(void)
4163 * If we come here after a random call to set_need_resched(),
4164 * or we have been woken up remotely but the IPI has not yet arrived,
4165 * we haven't yet exited the RCU idle mode. Do it here manually until
4166 * we find a better solution.
4168 * NB: There are buggy callers of this function. Ideally we
4169 * should warn if prev_state != CONTEXT_USER, but that will trigger
4170 * too frequently to make sense yet.
4172 enum ctx_state prev_state = exception_enter();
4174 exception_exit(prev_state);
4179 * schedule_preempt_disabled - called with preemption disabled
4181 * Returns with preemption disabled. Note: preempt_count must be 1
4183 void __sched schedule_preempt_disabled(void)
4185 sched_preempt_enable_no_resched();
4190 static void __sched notrace preempt_schedule_common(void)
4194 * Because the function tracer can trace preempt_count_sub()
4195 * and it also uses preempt_enable/disable_notrace(), if
4196 * NEED_RESCHED is set, the preempt_enable_notrace() called
4197 * by the function tracer will call this function again and
4198 * cause infinite recursion.
4200 * Preemption must be disabled here before the function
4201 * tracer can trace. Break up preempt_disable() into two
4202 * calls. One to disable preemption without fear of being
4203 * traced. The other to still record the preemption latency,
4204 * which can also be traced by the function tracer.
4206 preempt_disable_notrace();
4207 preempt_latency_start(1);
4209 preempt_latency_stop(1);
4210 preempt_enable_no_resched_notrace();
4213 * Check again in case we missed a preemption opportunity
4214 * between schedule and now.
4216 } while (need_resched());
4219 #ifdef CONFIG_PREEMPTION
4221 * This is the entry point to schedule() from in-kernel preemption
4222 * off of preempt_enable.
4224 asmlinkage __visible void __sched notrace preempt_schedule(void)
4227 * If there is a non-zero preempt_count or interrupts are disabled,
4228 * we do not want to preempt the current task. Just return..
4230 if (likely(!preemptible()))
4233 preempt_schedule_common();
4235 NOKPROBE_SYMBOL(preempt_schedule);
4236 EXPORT_SYMBOL(preempt_schedule);
4239 * preempt_schedule_notrace - preempt_schedule called by tracing
4241 * The tracing infrastructure uses preempt_enable_notrace to prevent
4242 * recursion and tracing preempt enabling caused by the tracing
4243 * infrastructure itself. But as tracing can happen in areas coming
4244 * from userspace or just about to enter userspace, a preempt enable
4245 * can occur before user_exit() is called. This will cause the scheduler
4246 * to be called when the system is still in usermode.
4248 * To prevent this, the preempt_enable_notrace will use this function
4249 * instead of preempt_schedule() to exit user context if needed before
4250 * calling the scheduler.
4252 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4254 enum ctx_state prev_ctx;
4256 if (likely(!preemptible()))
4261 * Because the function tracer can trace preempt_count_sub()
4262 * and it also uses preempt_enable/disable_notrace(), if
4263 * NEED_RESCHED is set, the preempt_enable_notrace() called
4264 * by the function tracer will call this function again and
4265 * cause infinite recursion.
4267 * Preemption must be disabled here before the function
4268 * tracer can trace. Break up preempt_disable() into two
4269 * calls. One to disable preemption without fear of being
4270 * traced. The other to still record the preemption latency,
4271 * which can also be traced by the function tracer.
4273 preempt_disable_notrace();
4274 preempt_latency_start(1);
4276 * Needs preempt disabled in case user_exit() is traced
4277 * and the tracer calls preempt_enable_notrace() causing
4278 * an infinite recursion.
4280 prev_ctx = exception_enter();
4282 exception_exit(prev_ctx);
4284 preempt_latency_stop(1);
4285 preempt_enable_no_resched_notrace();
4286 } while (need_resched());
4288 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4290 #endif /* CONFIG_PREEMPTION */
4293 * This is the entry point to schedule() from kernel preemption
4294 * off of irq context.
4295 * Note, that this is called and return with irqs disabled. This will
4296 * protect us against recursive calling from irq.
4298 asmlinkage __visible void __sched preempt_schedule_irq(void)
4300 enum ctx_state prev_state;
4302 /* Catch callers which need to be fixed */
4303 BUG_ON(preempt_count() || !irqs_disabled());
4305 prev_state = exception_enter();
4311 local_irq_disable();
4312 sched_preempt_enable_no_resched();
4313 } while (need_resched());
4315 exception_exit(prev_state);
4318 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4321 return try_to_wake_up(curr->private, mode, wake_flags);
4323 EXPORT_SYMBOL(default_wake_function);
4325 #ifdef CONFIG_RT_MUTEXES
4327 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4330 prio = min(prio, pi_task->prio);
4335 static inline int rt_effective_prio(struct task_struct *p, int prio)
4337 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4339 return __rt_effective_prio(pi_task, prio);
4343 * rt_mutex_setprio - set the current priority of a task
4345 * @pi_task: donor task
4347 * This function changes the 'effective' priority of a task. It does
4348 * not touch ->normal_prio like __setscheduler().
4350 * Used by the rt_mutex code to implement priority inheritance
4351 * logic. Call site only calls if the priority of the task changed.
4353 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4355 int prio, oldprio, queued, running, queue_flag =
4356 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4357 const struct sched_class *prev_class;
4361 /* XXX used to be waiter->prio, not waiter->task->prio */
4362 prio = __rt_effective_prio(pi_task, p->normal_prio);
4365 * If nothing changed; bail early.
4367 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4370 rq = __task_rq_lock(p, &rf);
4371 update_rq_clock(rq);
4373 * Set under pi_lock && rq->lock, such that the value can be used under
4376 * Note that there is loads of tricky to make this pointer cache work
4377 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4378 * ensure a task is de-boosted (pi_task is set to NULL) before the
4379 * task is allowed to run again (and can exit). This ensures the pointer
4380 * points to a blocked task -- which guaratees the task is present.
4382 p->pi_top_task = pi_task;
4385 * For FIFO/RR we only need to set prio, if that matches we're done.
4387 if (prio == p->prio && !dl_prio(prio))
4391 * Idle task boosting is a nono in general. There is one
4392 * exception, when PREEMPT_RT and NOHZ is active:
4394 * The idle task calls get_next_timer_interrupt() and holds
4395 * the timer wheel base->lock on the CPU and another CPU wants
4396 * to access the timer (probably to cancel it). We can safely
4397 * ignore the boosting request, as the idle CPU runs this code
4398 * with interrupts disabled and will complete the lock
4399 * protected section without being interrupted. So there is no
4400 * real need to boost.
4402 if (unlikely(p == rq->idle)) {
4403 WARN_ON(p != rq->curr);
4404 WARN_ON(p->pi_blocked_on);
4408 trace_sched_pi_setprio(p, pi_task);
4411 if (oldprio == prio)
4412 queue_flag &= ~DEQUEUE_MOVE;
4414 prev_class = p->sched_class;
4415 queued = task_on_rq_queued(p);
4416 running = task_current(rq, p);
4418 dequeue_task(rq, p, queue_flag);
4420 put_prev_task(rq, p);
4423 * Boosting condition are:
4424 * 1. -rt task is running and holds mutex A
4425 * --> -dl task blocks on mutex A
4427 * 2. -dl task is running and holds mutex A
4428 * --> -dl task blocks on mutex A and could preempt the
4431 if (dl_prio(prio)) {
4432 if (!dl_prio(p->normal_prio) ||
4433 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4434 p->dl.dl_boosted = 1;
4435 queue_flag |= ENQUEUE_REPLENISH;
4437 p->dl.dl_boosted = 0;
4438 p->sched_class = &dl_sched_class;
4439 } else if (rt_prio(prio)) {
4440 if (dl_prio(oldprio))
4441 p->dl.dl_boosted = 0;
4443 queue_flag |= ENQUEUE_HEAD;
4444 p->sched_class = &rt_sched_class;
4446 if (dl_prio(oldprio))
4447 p->dl.dl_boosted = 0;
4448 if (rt_prio(oldprio))
4450 p->sched_class = &fair_sched_class;
4456 enqueue_task(rq, p, queue_flag);
4458 set_next_task(rq, p);
4460 check_class_changed(rq, p, prev_class, oldprio);
4462 /* Avoid rq from going away on us: */
4464 __task_rq_unlock(rq, &rf);
4466 balance_callback(rq);
4470 static inline int rt_effective_prio(struct task_struct *p, int prio)
4476 void set_user_nice(struct task_struct *p, long nice)
4478 bool queued, running;
4479 int old_prio, delta;
4483 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4486 * We have to be careful, if called from sys_setpriority(),
4487 * the task might be in the middle of scheduling on another CPU.
4489 rq = task_rq_lock(p, &rf);
4490 update_rq_clock(rq);
4493 * The RT priorities are set via sched_setscheduler(), but we still
4494 * allow the 'normal' nice value to be set - but as expected
4495 * it wont have any effect on scheduling until the task is
4496 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4498 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4499 p->static_prio = NICE_TO_PRIO(nice);
4502 queued = task_on_rq_queued(p);
4503 running = task_current(rq, p);
4505 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4507 put_prev_task(rq, p);
4509 p->static_prio = NICE_TO_PRIO(nice);
4510 set_load_weight(p, true);
4512 p->prio = effective_prio(p);
4513 delta = p->prio - old_prio;
4516 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4518 * If the task increased its priority or is running and
4519 * lowered its priority, then reschedule its CPU:
4521 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4525 set_next_task(rq, p);
4527 task_rq_unlock(rq, p, &rf);
4529 EXPORT_SYMBOL(set_user_nice);
4532 * can_nice - check if a task can reduce its nice value
4536 int can_nice(const struct task_struct *p, const int nice)
4538 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4539 int nice_rlim = nice_to_rlimit(nice);
4541 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4542 capable(CAP_SYS_NICE));
4545 #ifdef __ARCH_WANT_SYS_NICE
4548 * sys_nice - change the priority of the current process.
4549 * @increment: priority increment
4551 * sys_setpriority is a more generic, but much slower function that
4552 * does similar things.
4554 SYSCALL_DEFINE1(nice, int, increment)
4559 * Setpriority might change our priority at the same moment.
4560 * We don't have to worry. Conceptually one call occurs first
4561 * and we have a single winner.
4563 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4564 nice = task_nice(current) + increment;
4566 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4567 if (increment < 0 && !can_nice(current, nice))
4570 retval = security_task_setnice(current, nice);
4574 set_user_nice(current, nice);
4581 * task_prio - return the priority value of a given task.
4582 * @p: the task in question.
4584 * Return: The priority value as seen by users in /proc.
4585 * RT tasks are offset by -200. Normal tasks are centered
4586 * around 0, value goes from -16 to +15.
4588 int task_prio(const struct task_struct *p)
4590 return p->prio - MAX_RT_PRIO;
4594 * idle_cpu - is a given CPU idle currently?
4595 * @cpu: the processor in question.
4597 * Return: 1 if the CPU is currently idle. 0 otherwise.
4599 int idle_cpu(int cpu)
4601 struct rq *rq = cpu_rq(cpu);
4603 if (rq->curr != rq->idle)
4610 if (!llist_empty(&rq->wake_list))
4618 * available_idle_cpu - is a given CPU idle for enqueuing work.
4619 * @cpu: the CPU in question.
4621 * Return: 1 if the CPU is currently idle. 0 otherwise.
4623 int available_idle_cpu(int cpu)
4628 if (vcpu_is_preempted(cpu))
4635 * idle_task - return the idle task for a given CPU.
4636 * @cpu: the processor in question.
4638 * Return: The idle task for the CPU @cpu.
4640 struct task_struct *idle_task(int cpu)
4642 return cpu_rq(cpu)->idle;
4646 * find_process_by_pid - find a process with a matching PID value.
4647 * @pid: the pid in question.
4649 * The task of @pid, if found. %NULL otherwise.
4651 static struct task_struct *find_process_by_pid(pid_t pid)
4653 return pid ? find_task_by_vpid(pid) : current;
4657 * sched_setparam() passes in -1 for its policy, to let the functions
4658 * it calls know not to change it.
4660 #define SETPARAM_POLICY -1
4662 static void __setscheduler_params(struct task_struct *p,
4663 const struct sched_attr *attr)
4665 int policy = attr->sched_policy;
4667 if (policy == SETPARAM_POLICY)
4672 if (dl_policy(policy))
4673 __setparam_dl(p, attr);
4674 else if (fair_policy(policy))
4675 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4678 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4679 * !rt_policy. Always setting this ensures that things like
4680 * getparam()/getattr() don't report silly values for !rt tasks.
4682 p->rt_priority = attr->sched_priority;
4683 p->normal_prio = normal_prio(p);
4684 set_load_weight(p, true);
4687 /* Actually do priority change: must hold pi & rq lock. */
4688 static void __setscheduler(struct rq *rq, struct task_struct *p,
4689 const struct sched_attr *attr, bool keep_boost)
4692 * If params can't change scheduling class changes aren't allowed
4695 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4698 __setscheduler_params(p, attr);
4701 * Keep a potential priority boosting if called from
4702 * sched_setscheduler().
4704 p->prio = normal_prio(p);
4706 p->prio = rt_effective_prio(p, p->prio);
4708 if (dl_prio(p->prio))
4709 p->sched_class = &dl_sched_class;
4710 else if (rt_prio(p->prio))
4711 p->sched_class = &rt_sched_class;
4713 p->sched_class = &fair_sched_class;
4717 * Check the target process has a UID that matches the current process's:
4719 static bool check_same_owner(struct task_struct *p)
4721 const struct cred *cred = current_cred(), *pcred;
4725 pcred = __task_cred(p);
4726 match = (uid_eq(cred->euid, pcred->euid) ||
4727 uid_eq(cred->euid, pcred->uid));
4732 static int __sched_setscheduler(struct task_struct *p,
4733 const struct sched_attr *attr,
4736 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4737 MAX_RT_PRIO - 1 - attr->sched_priority;
4738 int retval, oldprio, oldpolicy = -1, queued, running;
4739 int new_effective_prio, policy = attr->sched_policy;
4740 const struct sched_class *prev_class;
4743 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4746 /* The pi code expects interrupts enabled */
4747 BUG_ON(pi && in_interrupt());
4749 /* Double check policy once rq lock held: */
4751 reset_on_fork = p->sched_reset_on_fork;
4752 policy = oldpolicy = p->policy;
4754 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4756 if (!valid_policy(policy))
4760 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4764 * Valid priorities for SCHED_FIFO and SCHED_RR are
4765 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4766 * SCHED_BATCH and SCHED_IDLE is 0.
4768 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4769 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4771 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4772 (rt_policy(policy) != (attr->sched_priority != 0)))
4776 * Allow unprivileged RT tasks to decrease priority:
4778 if (user && !capable(CAP_SYS_NICE)) {
4779 if (fair_policy(policy)) {
4780 if (attr->sched_nice < task_nice(p) &&
4781 !can_nice(p, attr->sched_nice))
4785 if (rt_policy(policy)) {
4786 unsigned long rlim_rtprio =
4787 task_rlimit(p, RLIMIT_RTPRIO);
4789 /* Can't set/change the rt policy: */
4790 if (policy != p->policy && !rlim_rtprio)
4793 /* Can't increase priority: */
4794 if (attr->sched_priority > p->rt_priority &&
4795 attr->sched_priority > rlim_rtprio)
4800 * Can't set/change SCHED_DEADLINE policy at all for now
4801 * (safest behavior); in the future we would like to allow
4802 * unprivileged DL tasks to increase their relative deadline
4803 * or reduce their runtime (both ways reducing utilization)
4805 if (dl_policy(policy))
4809 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4810 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4812 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4813 if (!can_nice(p, task_nice(p)))
4817 /* Can't change other user's priorities: */
4818 if (!check_same_owner(p))
4821 /* Normal users shall not reset the sched_reset_on_fork flag: */
4822 if (p->sched_reset_on_fork && !reset_on_fork)
4827 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4830 retval = security_task_setscheduler(p);
4835 /* Update task specific "requested" clamps */
4836 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4837 retval = uclamp_validate(p, attr);
4846 * Make sure no PI-waiters arrive (or leave) while we are
4847 * changing the priority of the task:
4849 * To be able to change p->policy safely, the appropriate
4850 * runqueue lock must be held.
4852 rq = task_rq_lock(p, &rf);
4853 update_rq_clock(rq);
4856 * Changing the policy of the stop threads its a very bad idea:
4858 if (p == rq->stop) {
4864 * If not changing anything there's no need to proceed further,
4865 * but store a possible modification of reset_on_fork.
4867 if (unlikely(policy == p->policy)) {
4868 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4870 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4872 if (dl_policy(policy) && dl_param_changed(p, attr))
4874 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4877 p->sched_reset_on_fork = reset_on_fork;
4884 #ifdef CONFIG_RT_GROUP_SCHED
4886 * Do not allow realtime tasks into groups that have no runtime
4889 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4890 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4891 !task_group_is_autogroup(task_group(p))) {
4897 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4898 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4899 cpumask_t *span = rq->rd->span;
4902 * Don't allow tasks with an affinity mask smaller than
4903 * the entire root_domain to become SCHED_DEADLINE. We
4904 * will also fail if there's no bandwidth available.
4906 if (!cpumask_subset(span, p->cpus_ptr) ||
4907 rq->rd->dl_bw.bw == 0) {
4915 /* Re-check policy now with rq lock held: */
4916 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4917 policy = oldpolicy = -1;
4918 task_rq_unlock(rq, p, &rf);
4920 cpuset_read_unlock();
4925 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4926 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4929 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4934 p->sched_reset_on_fork = reset_on_fork;
4939 * Take priority boosted tasks into account. If the new
4940 * effective priority is unchanged, we just store the new
4941 * normal parameters and do not touch the scheduler class and
4942 * the runqueue. This will be done when the task deboost
4945 new_effective_prio = rt_effective_prio(p, newprio);
4946 if (new_effective_prio == oldprio)
4947 queue_flags &= ~DEQUEUE_MOVE;
4950 queued = task_on_rq_queued(p);
4951 running = task_current(rq, p);
4953 dequeue_task(rq, p, queue_flags);
4955 put_prev_task(rq, p);
4957 prev_class = p->sched_class;
4959 __setscheduler(rq, p, attr, pi);
4960 __setscheduler_uclamp(p, attr);
4964 * We enqueue to tail when the priority of a task is
4965 * increased (user space view).
4967 if (oldprio < p->prio)
4968 queue_flags |= ENQUEUE_HEAD;
4970 enqueue_task(rq, p, queue_flags);
4973 set_next_task(rq, p);
4975 check_class_changed(rq, p, prev_class, oldprio);
4977 /* Avoid rq from going away on us: */
4979 task_rq_unlock(rq, p, &rf);
4982 cpuset_read_unlock();
4983 rt_mutex_adjust_pi(p);
4986 /* Run balance callbacks after we've adjusted the PI chain: */
4987 balance_callback(rq);
4993 task_rq_unlock(rq, p, &rf);
4995 cpuset_read_unlock();
4999 static int _sched_setscheduler(struct task_struct *p, int policy,
5000 const struct sched_param *param, bool check)
5002 struct sched_attr attr = {
5003 .sched_policy = policy,
5004 .sched_priority = param->sched_priority,
5005 .sched_nice = PRIO_TO_NICE(p->static_prio),
5008 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5009 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5010 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5011 policy &= ~SCHED_RESET_ON_FORK;
5012 attr.sched_policy = policy;
5015 return __sched_setscheduler(p, &attr, check, true);
5018 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5019 * @p: the task in question.
5020 * @policy: new policy.
5021 * @param: structure containing the new RT priority.
5023 * Return: 0 on success. An error code otherwise.
5025 * NOTE that the task may be already dead.
5027 int sched_setscheduler(struct task_struct *p, int policy,
5028 const struct sched_param *param)
5030 return _sched_setscheduler(p, policy, param, true);
5032 EXPORT_SYMBOL_GPL(sched_setscheduler);
5034 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5036 return __sched_setscheduler(p, attr, true, true);
5038 EXPORT_SYMBOL_GPL(sched_setattr);
5040 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5042 return __sched_setscheduler(p, attr, false, true);
5046 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5047 * @p: the task in question.
5048 * @policy: new policy.
5049 * @param: structure containing the new RT priority.
5051 * Just like sched_setscheduler, only don't bother checking if the
5052 * current context has permission. For example, this is needed in
5053 * stop_machine(): we create temporary high priority worker threads,
5054 * but our caller might not have that capability.
5056 * Return: 0 on success. An error code otherwise.
5058 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5059 const struct sched_param *param)
5061 return _sched_setscheduler(p, policy, param, false);
5063 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5066 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5068 struct sched_param lparam;
5069 struct task_struct *p;
5072 if (!param || pid < 0)
5074 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5079 p = find_process_by_pid(pid);
5085 retval = sched_setscheduler(p, policy, &lparam);
5093 * Mimics kernel/events/core.c perf_copy_attr().
5095 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5100 if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
5103 /* Zero the full structure, so that a short copy will be nice: */
5104 memset(attr, 0, sizeof(*attr));
5106 ret = get_user(size, &uattr->size);
5110 /* Bail out on silly large: */
5111 if (size > PAGE_SIZE)
5114 /* ABI compatibility quirk: */
5116 size = SCHED_ATTR_SIZE_VER0;
5118 if (size < SCHED_ATTR_SIZE_VER0)
5122 * If we're handed a bigger struct than we know of,
5123 * ensure all the unknown bits are 0 - i.e. new
5124 * user-space does not rely on any kernel feature
5125 * extensions we dont know about yet.
5127 if (size > sizeof(*attr)) {
5128 unsigned char __user *addr;
5129 unsigned char __user *end;
5132 addr = (void __user *)uattr + sizeof(*attr);
5133 end = (void __user *)uattr + size;
5135 for (; addr < end; addr++) {
5136 ret = get_user(val, addr);
5142 size = sizeof(*attr);
5145 ret = copy_from_user(attr, uattr, size);
5149 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5150 size < SCHED_ATTR_SIZE_VER1)
5154 * XXX: Do we want to be lenient like existing syscalls; or do we want
5155 * to be strict and return an error on out-of-bounds values?
5157 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5162 put_user(sizeof(*attr), &uattr->size);
5167 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5168 * @pid: the pid in question.
5169 * @policy: new policy.
5170 * @param: structure containing the new RT priority.
5172 * Return: 0 on success. An error code otherwise.
5174 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5179 return do_sched_setscheduler(pid, policy, param);
5183 * sys_sched_setparam - set/change the RT priority of a thread
5184 * @pid: the pid in question.
5185 * @param: structure containing the new RT priority.
5187 * Return: 0 on success. An error code otherwise.
5189 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5191 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5195 * sys_sched_setattr - same as above, but with extended sched_attr
5196 * @pid: the pid in question.
5197 * @uattr: structure containing the extended parameters.
5198 * @flags: for future extension.
5200 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5201 unsigned int, flags)
5203 struct sched_attr attr;
5204 struct task_struct *p;
5207 if (!uattr || pid < 0 || flags)
5210 retval = sched_copy_attr(uattr, &attr);
5214 if ((int)attr.sched_policy < 0)
5216 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5217 attr.sched_policy = SETPARAM_POLICY;
5221 p = find_process_by_pid(pid);
5227 retval = sched_setattr(p, &attr);
5235 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5236 * @pid: the pid in question.
5238 * Return: On success, the policy of the thread. Otherwise, a negative error
5241 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5243 struct task_struct *p;
5251 p = find_process_by_pid(pid);
5253 retval = security_task_getscheduler(p);
5256 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5263 * sys_sched_getparam - get the RT priority of a thread
5264 * @pid: the pid in question.
5265 * @param: structure containing the RT priority.
5267 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5270 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5272 struct sched_param lp = { .sched_priority = 0 };
5273 struct task_struct *p;
5276 if (!param || pid < 0)
5280 p = find_process_by_pid(pid);
5285 retval = security_task_getscheduler(p);
5289 if (task_has_rt_policy(p))
5290 lp.sched_priority = p->rt_priority;
5294 * This one might sleep, we cannot do it with a spinlock held ...
5296 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5306 * Copy the kernel size attribute structure (which might be larger
5307 * than what user-space knows about) to user-space.
5309 * Note that all cases are valid: user-space buffer can be larger or
5310 * smaller than the kernel-space buffer. The usual case is that both
5311 * have the same size.
5314 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5315 struct sched_attr *kattr,
5318 unsigned int ksize = sizeof(*kattr);
5320 if (!access_ok(uattr, usize))
5324 * sched_getattr() ABI forwards and backwards compatibility:
5326 * If usize == ksize then we just copy everything to user-space and all is good.
5328 * If usize < ksize then we only copy as much as user-space has space for,
5329 * this keeps ABI compatibility as well. We skip the rest.
5331 * If usize > ksize then user-space is using a newer version of the ABI,
5332 * which part the kernel doesn't know about. Just ignore it - tooling can
5333 * detect the kernel's knowledge of attributes from the attr->size value
5334 * which is set to ksize in this case.
5336 kattr->size = min(usize, ksize);
5338 if (copy_to_user(uattr, kattr, kattr->size))
5345 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5346 * @pid: the pid in question.
5347 * @uattr: structure containing the extended parameters.
5348 * @usize: sizeof(attr) that user-space knows about, for forwards and backwards compatibility.
5349 * @flags: for future extension.
5351 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5352 unsigned int, usize, unsigned int, flags)
5354 struct sched_attr kattr = { };
5355 struct task_struct *p;
5358 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5359 usize < SCHED_ATTR_SIZE_VER0 || flags)
5363 p = find_process_by_pid(pid);
5368 retval = security_task_getscheduler(p);
5372 kattr.sched_policy = p->policy;
5373 if (p->sched_reset_on_fork)
5374 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5375 if (task_has_dl_policy(p))
5376 __getparam_dl(p, &kattr);
5377 else if (task_has_rt_policy(p))
5378 kattr.sched_priority = p->rt_priority;
5380 kattr.sched_nice = task_nice(p);
5382 #ifdef CONFIG_UCLAMP_TASK
5383 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5384 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5389 return sched_attr_copy_to_user(uattr, &kattr, usize);
5396 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5398 cpumask_var_t cpus_allowed, new_mask;
5399 struct task_struct *p;
5404 p = find_process_by_pid(pid);
5410 /* Prevent p going away */
5414 if (p->flags & PF_NO_SETAFFINITY) {
5418 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5422 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5424 goto out_free_cpus_allowed;
5427 if (!check_same_owner(p)) {
5429 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5431 goto out_free_new_mask;
5436 retval = security_task_setscheduler(p);
5438 goto out_free_new_mask;
5441 cpuset_cpus_allowed(p, cpus_allowed);
5442 cpumask_and(new_mask, in_mask, cpus_allowed);
5445 * Since bandwidth control happens on root_domain basis,
5446 * if admission test is enabled, we only admit -deadline
5447 * tasks allowed to run on all the CPUs in the task's
5451 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5453 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5456 goto out_free_new_mask;
5462 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5465 cpuset_cpus_allowed(p, cpus_allowed);
5466 if (!cpumask_subset(new_mask, cpus_allowed)) {
5468 * We must have raced with a concurrent cpuset
5469 * update. Just reset the cpus_allowed to the
5470 * cpuset's cpus_allowed
5472 cpumask_copy(new_mask, cpus_allowed);
5477 free_cpumask_var(new_mask);
5478 out_free_cpus_allowed:
5479 free_cpumask_var(cpus_allowed);
5485 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5486 struct cpumask *new_mask)
5488 if (len < cpumask_size())
5489 cpumask_clear(new_mask);
5490 else if (len > cpumask_size())
5491 len = cpumask_size();
5493 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5497 * sys_sched_setaffinity - set the CPU affinity of a process
5498 * @pid: pid of the process
5499 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5500 * @user_mask_ptr: user-space pointer to the new CPU mask
5502 * Return: 0 on success. An error code otherwise.
5504 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5505 unsigned long __user *, user_mask_ptr)
5507 cpumask_var_t new_mask;
5510 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5513 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5515 retval = sched_setaffinity(pid, new_mask);
5516 free_cpumask_var(new_mask);
5520 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5522 struct task_struct *p;
5523 unsigned long flags;
5529 p = find_process_by_pid(pid);
5533 retval = security_task_getscheduler(p);
5537 raw_spin_lock_irqsave(&p->pi_lock, flags);
5538 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5539 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5548 * sys_sched_getaffinity - get the CPU affinity of a process
5549 * @pid: pid of the process
5550 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5551 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5553 * Return: size of CPU mask copied to user_mask_ptr on success. An
5554 * error code otherwise.
5556 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5557 unsigned long __user *, user_mask_ptr)
5562 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5564 if (len & (sizeof(unsigned long)-1))
5567 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5570 ret = sched_getaffinity(pid, mask);
5572 unsigned int retlen = min(len, cpumask_size());
5574 if (copy_to_user(user_mask_ptr, mask, retlen))
5579 free_cpumask_var(mask);
5585 * sys_sched_yield - yield the current processor to other threads.
5587 * This function yields the current CPU to other tasks. If there are no
5588 * other threads running on this CPU then this function will return.
5592 static void do_sched_yield(void)
5597 rq = this_rq_lock_irq(&rf);
5599 schedstat_inc(rq->yld_count);
5600 current->sched_class->yield_task(rq);
5603 * Since we are going to call schedule() anyway, there's
5604 * no need to preempt or enable interrupts:
5608 sched_preempt_enable_no_resched();
5613 SYSCALL_DEFINE0(sched_yield)
5619 #ifndef CONFIG_PREEMPTION
5620 int __sched _cond_resched(void)
5622 if (should_resched(0)) {
5623 preempt_schedule_common();
5629 EXPORT_SYMBOL(_cond_resched);
5633 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5634 * call schedule, and on return reacquire the lock.
5636 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5637 * operations here to prevent schedule() from being called twice (once via
5638 * spin_unlock(), once by hand).
5640 int __cond_resched_lock(spinlock_t *lock)
5642 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5645 lockdep_assert_held(lock);
5647 if (spin_needbreak(lock) || resched) {
5650 preempt_schedule_common();
5658 EXPORT_SYMBOL(__cond_resched_lock);
5661 * yield - yield the current processor to other threads.
5663 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5665 * The scheduler is at all times free to pick the calling task as the most
5666 * eligible task to run, if removing the yield() call from your code breaks
5667 * it, its already broken.
5669 * Typical broken usage is:
5674 * where one assumes that yield() will let 'the other' process run that will
5675 * make event true. If the current task is a SCHED_FIFO task that will never
5676 * happen. Never use yield() as a progress guarantee!!
5678 * If you want to use yield() to wait for something, use wait_event().
5679 * If you want to use yield() to be 'nice' for others, use cond_resched().
5680 * If you still want to use yield(), do not!
5682 void __sched yield(void)
5684 set_current_state(TASK_RUNNING);
5687 EXPORT_SYMBOL(yield);
5690 * yield_to - yield the current processor to another thread in
5691 * your thread group, or accelerate that thread toward the
5692 * processor it's on.
5694 * @preempt: whether task preemption is allowed or not
5696 * It's the caller's job to ensure that the target task struct
5697 * can't go away on us before we can do any checks.
5700 * true (>0) if we indeed boosted the target task.
5701 * false (0) if we failed to boost the target.
5702 * -ESRCH if there's no task to yield to.
5704 int __sched yield_to(struct task_struct *p, bool preempt)
5706 struct task_struct *curr = current;
5707 struct rq *rq, *p_rq;
5708 unsigned long flags;
5711 local_irq_save(flags);
5717 * If we're the only runnable task on the rq and target rq also
5718 * has only one task, there's absolutely no point in yielding.
5720 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5725 double_rq_lock(rq, p_rq);
5726 if (task_rq(p) != p_rq) {
5727 double_rq_unlock(rq, p_rq);
5731 if (!curr->sched_class->yield_to_task)
5734 if (curr->sched_class != p->sched_class)
5737 if (task_running(p_rq, p) || p->state)
5740 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5742 schedstat_inc(rq->yld_count);
5744 * Make p's CPU reschedule; pick_next_entity takes care of
5747 if (preempt && rq != p_rq)
5752 double_rq_unlock(rq, p_rq);
5754 local_irq_restore(flags);
5761 EXPORT_SYMBOL_GPL(yield_to);
5763 int io_schedule_prepare(void)
5765 int old_iowait = current->in_iowait;
5767 current->in_iowait = 1;
5768 blk_schedule_flush_plug(current);
5773 void io_schedule_finish(int token)
5775 current->in_iowait = token;
5779 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5780 * that process accounting knows that this is a task in IO wait state.
5782 long __sched io_schedule_timeout(long timeout)
5787 token = io_schedule_prepare();
5788 ret = schedule_timeout(timeout);
5789 io_schedule_finish(token);
5793 EXPORT_SYMBOL(io_schedule_timeout);
5795 void __sched io_schedule(void)
5799 token = io_schedule_prepare();
5801 io_schedule_finish(token);
5803 EXPORT_SYMBOL(io_schedule);
5806 * sys_sched_get_priority_max - return maximum RT priority.
5807 * @policy: scheduling class.
5809 * Return: On success, this syscall returns the maximum
5810 * rt_priority that can be used by a given scheduling class.
5811 * On failure, a negative error code is returned.
5813 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5820 ret = MAX_USER_RT_PRIO-1;
5822 case SCHED_DEADLINE:
5833 * sys_sched_get_priority_min - return minimum RT priority.
5834 * @policy: scheduling class.
5836 * Return: On success, this syscall returns the minimum
5837 * rt_priority that can be used by a given scheduling class.
5838 * On failure, a negative error code is returned.
5840 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5849 case SCHED_DEADLINE:
5858 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5860 struct task_struct *p;
5861 unsigned int time_slice;
5871 p = find_process_by_pid(pid);
5875 retval = security_task_getscheduler(p);
5879 rq = task_rq_lock(p, &rf);
5881 if (p->sched_class->get_rr_interval)
5882 time_slice = p->sched_class->get_rr_interval(rq, p);
5883 task_rq_unlock(rq, p, &rf);
5886 jiffies_to_timespec64(time_slice, t);
5895 * sys_sched_rr_get_interval - return the default timeslice of a process.
5896 * @pid: pid of the process.
5897 * @interval: userspace pointer to the timeslice value.
5899 * this syscall writes the default timeslice value of a given process
5900 * into the user-space timespec buffer. A value of '0' means infinity.
5902 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5905 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5906 struct __kernel_timespec __user *, interval)
5908 struct timespec64 t;
5909 int retval = sched_rr_get_interval(pid, &t);
5912 retval = put_timespec64(&t, interval);
5917 #ifdef CONFIG_COMPAT_32BIT_TIME
5918 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5919 struct old_timespec32 __user *, interval)
5921 struct timespec64 t;
5922 int retval = sched_rr_get_interval(pid, &t);
5925 retval = put_old_timespec32(&t, interval);
5930 void sched_show_task(struct task_struct *p)
5932 unsigned long free = 0;
5935 if (!try_get_task_stack(p))
5938 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5940 if (p->state == TASK_RUNNING)
5941 printk(KERN_CONT " running task ");
5942 #ifdef CONFIG_DEBUG_STACK_USAGE
5943 free = stack_not_used(p);
5948 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5950 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5951 task_pid_nr(p), ppid,
5952 (unsigned long)task_thread_info(p)->flags);
5954 print_worker_info(KERN_INFO, p);
5955 show_stack(p, NULL);
5958 EXPORT_SYMBOL_GPL(sched_show_task);
5961 state_filter_match(unsigned long state_filter, struct task_struct *p)
5963 /* no filter, everything matches */
5967 /* filter, but doesn't match */
5968 if (!(p->state & state_filter))
5972 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5975 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5982 void show_state_filter(unsigned long state_filter)
5984 struct task_struct *g, *p;
5986 #if BITS_PER_LONG == 32
5988 " task PC stack pid father\n");
5991 " task PC stack pid father\n");
5994 for_each_process_thread(g, p) {
5996 * reset the NMI-timeout, listing all files on a slow
5997 * console might take a lot of time:
5998 * Also, reset softlockup watchdogs on all CPUs, because
5999 * another CPU might be blocked waiting for us to process
6002 touch_nmi_watchdog();
6003 touch_all_softlockup_watchdogs();
6004 if (state_filter_match(state_filter, p))
6008 #ifdef CONFIG_SCHED_DEBUG
6010 sysrq_sched_debug_show();
6014 * Only show locks if all tasks are dumped:
6017 debug_show_all_locks();
6021 * init_idle - set up an idle thread for a given CPU
6022 * @idle: task in question
6023 * @cpu: CPU the idle task belongs to
6025 * NOTE: this function does not set the idle thread's NEED_RESCHED
6026 * flag, to make booting more robust.
6028 void init_idle(struct task_struct *idle, int cpu)
6030 struct rq *rq = cpu_rq(cpu);
6031 unsigned long flags;
6033 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6034 raw_spin_lock(&rq->lock);
6036 __sched_fork(0, idle);
6037 idle->state = TASK_RUNNING;
6038 idle->se.exec_start = sched_clock();
6039 idle->flags |= PF_IDLE;
6041 kasan_unpoison_task_stack(idle);
6045 * Its possible that init_idle() gets called multiple times on a task,
6046 * in that case do_set_cpus_allowed() will not do the right thing.
6048 * And since this is boot we can forgo the serialization.
6050 set_cpus_allowed_common(idle, cpumask_of(cpu));
6053 * We're having a chicken and egg problem, even though we are
6054 * holding rq->lock, the CPU isn't yet set to this CPU so the
6055 * lockdep check in task_group() will fail.
6057 * Similar case to sched_fork(). / Alternatively we could
6058 * use task_rq_lock() here and obtain the other rq->lock.
6063 __set_task_cpu(idle, cpu);
6067 rcu_assign_pointer(rq->curr, idle);
6068 idle->on_rq = TASK_ON_RQ_QUEUED;
6072 raw_spin_unlock(&rq->lock);
6073 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6075 /* Set the preempt count _outside_ the spinlocks! */
6076 init_idle_preempt_count(idle, cpu);
6079 * The idle tasks have their own, simple scheduling class:
6081 idle->sched_class = &idle_sched_class;
6082 ftrace_graph_init_idle_task(idle, cpu);
6083 vtime_init_idle(idle, cpu);
6085 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6091 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6092 const struct cpumask *trial)
6096 if (!cpumask_weight(cur))
6099 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6104 int task_can_attach(struct task_struct *p,
6105 const struct cpumask *cs_cpus_allowed)
6110 * Kthreads which disallow setaffinity shouldn't be moved
6111 * to a new cpuset; we don't want to change their CPU
6112 * affinity and isolating such threads by their set of
6113 * allowed nodes is unnecessary. Thus, cpusets are not
6114 * applicable for such threads. This prevents checking for
6115 * success of set_cpus_allowed_ptr() on all attached tasks
6116 * before cpus_mask may be changed.
6118 if (p->flags & PF_NO_SETAFFINITY) {
6123 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6125 ret = dl_task_can_attach(p, cs_cpus_allowed);
6131 bool sched_smp_initialized __read_mostly;
6133 #ifdef CONFIG_NUMA_BALANCING
6134 /* Migrate current task p to target_cpu */
6135 int migrate_task_to(struct task_struct *p, int target_cpu)
6137 struct migration_arg arg = { p, target_cpu };
6138 int curr_cpu = task_cpu(p);
6140 if (curr_cpu == target_cpu)
6143 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6146 /* TODO: This is not properly updating schedstats */
6148 trace_sched_move_numa(p, curr_cpu, target_cpu);
6149 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6153 * Requeue a task on a given node and accurately track the number of NUMA
6154 * tasks on the runqueues
6156 void sched_setnuma(struct task_struct *p, int nid)
6158 bool queued, running;
6162 rq = task_rq_lock(p, &rf);
6163 queued = task_on_rq_queued(p);
6164 running = task_current(rq, p);
6167 dequeue_task(rq, p, DEQUEUE_SAVE);
6169 put_prev_task(rq, p);
6171 p->numa_preferred_nid = nid;
6174 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6176 set_next_task(rq, p);
6177 task_rq_unlock(rq, p, &rf);
6179 #endif /* CONFIG_NUMA_BALANCING */
6181 #ifdef CONFIG_HOTPLUG_CPU
6183 * Ensure that the idle task is using init_mm right before its CPU goes
6186 void idle_task_exit(void)
6188 struct mm_struct *mm = current->active_mm;
6190 BUG_ON(cpu_online(smp_processor_id()));
6192 if (mm != &init_mm) {
6193 switch_mm(mm, &init_mm, current);
6194 current->active_mm = &init_mm;
6195 finish_arch_post_lock_switch();
6201 * Since this CPU is going 'away' for a while, fold any nr_active delta
6202 * we might have. Assumes we're called after migrate_tasks() so that the
6203 * nr_active count is stable. We need to take the teardown thread which
6204 * is calling this into account, so we hand in adjust = 1 to the load
6207 * Also see the comment "Global load-average calculations".
6209 static void calc_load_migrate(struct rq *rq)
6211 long delta = calc_load_fold_active(rq, 1);
6213 atomic_long_add(delta, &calc_load_tasks);
6216 static struct task_struct *__pick_migrate_task(struct rq *rq)
6218 const struct sched_class *class;
6219 struct task_struct *next;
6221 for_each_class(class) {
6222 next = class->pick_next_task(rq, NULL, NULL);
6224 next->sched_class->put_prev_task(rq, next, NULL);
6229 /* The idle class should always have a runnable task */
6234 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6235 * try_to_wake_up()->select_task_rq().
6237 * Called with rq->lock held even though we'er in stop_machine() and
6238 * there's no concurrency possible, we hold the required locks anyway
6239 * because of lock validation efforts.
6241 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6243 struct rq *rq = dead_rq;
6244 struct task_struct *next, *stop = rq->stop;
6245 struct rq_flags orf = *rf;
6249 * Fudge the rq selection such that the below task selection loop
6250 * doesn't get stuck on the currently eligible stop task.
6252 * We're currently inside stop_machine() and the rq is either stuck
6253 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6254 * either way we should never end up calling schedule() until we're
6260 * put_prev_task() and pick_next_task() sched
6261 * class method both need to have an up-to-date
6262 * value of rq->clock[_task]
6264 update_rq_clock(rq);
6268 * There's this thread running, bail when that's the only
6271 if (rq->nr_running == 1)
6274 next = __pick_migrate_task(rq);
6277 * Rules for changing task_struct::cpus_mask are holding
6278 * both pi_lock and rq->lock, such that holding either
6279 * stabilizes the mask.
6281 * Drop rq->lock is not quite as disastrous as it usually is
6282 * because !cpu_active at this point, which means load-balance
6283 * will not interfere. Also, stop-machine.
6286 raw_spin_lock(&next->pi_lock);
6290 * Since we're inside stop-machine, _nothing_ should have
6291 * changed the task, WARN if weird stuff happened, because in
6292 * that case the above rq->lock drop is a fail too.
6294 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6295 raw_spin_unlock(&next->pi_lock);
6299 /* Find suitable destination for @next, with force if needed. */
6300 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6301 rq = __migrate_task(rq, rf, next, dest_cpu);
6302 if (rq != dead_rq) {
6308 raw_spin_unlock(&next->pi_lock);
6313 #endif /* CONFIG_HOTPLUG_CPU */
6315 void set_rq_online(struct rq *rq)
6318 const struct sched_class *class;
6320 cpumask_set_cpu(rq->cpu, rq->rd->online);
6323 for_each_class(class) {
6324 if (class->rq_online)
6325 class->rq_online(rq);
6330 void set_rq_offline(struct rq *rq)
6333 const struct sched_class *class;
6335 for_each_class(class) {
6336 if (class->rq_offline)
6337 class->rq_offline(rq);
6340 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6346 * used to mark begin/end of suspend/resume:
6348 static int num_cpus_frozen;
6351 * Update cpusets according to cpu_active mask. If cpusets are
6352 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6353 * around partition_sched_domains().
6355 * If we come here as part of a suspend/resume, don't touch cpusets because we
6356 * want to restore it back to its original state upon resume anyway.
6358 static void cpuset_cpu_active(void)
6360 if (cpuhp_tasks_frozen) {
6362 * num_cpus_frozen tracks how many CPUs are involved in suspend
6363 * resume sequence. As long as this is not the last online
6364 * operation in the resume sequence, just build a single sched
6365 * domain, ignoring cpusets.
6367 partition_sched_domains(1, NULL, NULL);
6368 if (--num_cpus_frozen)
6371 * This is the last CPU online operation. So fall through and
6372 * restore the original sched domains by considering the
6373 * cpuset configurations.
6375 cpuset_force_rebuild();
6377 cpuset_update_active_cpus();
6380 static int cpuset_cpu_inactive(unsigned int cpu)
6382 if (!cpuhp_tasks_frozen) {
6383 if (dl_cpu_busy(cpu))
6385 cpuset_update_active_cpus();
6388 partition_sched_domains(1, NULL, NULL);
6393 int sched_cpu_activate(unsigned int cpu)
6395 struct rq *rq = cpu_rq(cpu);
6398 #ifdef CONFIG_SCHED_SMT
6400 * When going up, increment the number of cores with SMT present.
6402 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6403 static_branch_inc_cpuslocked(&sched_smt_present);
6405 set_cpu_active(cpu, true);
6407 if (sched_smp_initialized) {
6408 sched_domains_numa_masks_set(cpu);
6409 cpuset_cpu_active();
6413 * Put the rq online, if not already. This happens:
6415 * 1) In the early boot process, because we build the real domains
6416 * after all CPUs have been brought up.
6418 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6421 rq_lock_irqsave(rq, &rf);
6423 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6426 rq_unlock_irqrestore(rq, &rf);
6428 update_max_interval();
6433 int sched_cpu_deactivate(unsigned int cpu)
6437 set_cpu_active(cpu, false);
6439 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6440 * users of this state to go away such that all new such users will
6443 * Do sync before park smpboot threads to take care the rcu boost case.
6447 #ifdef CONFIG_SCHED_SMT
6449 * When going down, decrement the number of cores with SMT present.
6451 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6452 static_branch_dec_cpuslocked(&sched_smt_present);
6455 if (!sched_smp_initialized)
6458 ret = cpuset_cpu_inactive(cpu);
6460 set_cpu_active(cpu, true);
6463 sched_domains_numa_masks_clear(cpu);
6467 static void sched_rq_cpu_starting(unsigned int cpu)
6469 struct rq *rq = cpu_rq(cpu);
6471 rq->calc_load_update = calc_load_update;
6472 update_max_interval();
6475 int sched_cpu_starting(unsigned int cpu)
6477 sched_rq_cpu_starting(cpu);
6478 sched_tick_start(cpu);
6482 #ifdef CONFIG_HOTPLUG_CPU
6483 int sched_cpu_dying(unsigned int cpu)
6485 struct rq *rq = cpu_rq(cpu);
6488 /* Handle pending wakeups and then migrate everything off */
6489 sched_ttwu_pending();
6490 sched_tick_stop(cpu);
6492 rq_lock_irqsave(rq, &rf);
6494 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6497 migrate_tasks(rq, &rf);
6498 BUG_ON(rq->nr_running != 1);
6499 rq_unlock_irqrestore(rq, &rf);
6501 calc_load_migrate(rq);
6502 update_max_interval();
6503 nohz_balance_exit_idle(rq);
6509 void __init sched_init_smp(void)
6514 * There's no userspace yet to cause hotplug operations; hence all the
6515 * CPU masks are stable and all blatant races in the below code cannot
6518 mutex_lock(&sched_domains_mutex);
6519 sched_init_domains(cpu_active_mask);
6520 mutex_unlock(&sched_domains_mutex);
6522 /* Move init over to a non-isolated CPU */
6523 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6525 sched_init_granularity();
6527 init_sched_rt_class();
6528 init_sched_dl_class();
6530 sched_smp_initialized = true;
6533 static int __init migration_init(void)
6535 sched_cpu_starting(smp_processor_id());
6538 early_initcall(migration_init);
6541 void __init sched_init_smp(void)
6543 sched_init_granularity();
6545 #endif /* CONFIG_SMP */
6547 int in_sched_functions(unsigned long addr)
6549 return in_lock_functions(addr) ||
6550 (addr >= (unsigned long)__sched_text_start
6551 && addr < (unsigned long)__sched_text_end);
6554 #ifdef CONFIG_CGROUP_SCHED
6556 * Default task group.
6557 * Every task in system belongs to this group at bootup.
6559 struct task_group root_task_group;
6560 LIST_HEAD(task_groups);
6562 /* Cacheline aligned slab cache for task_group */
6563 static struct kmem_cache *task_group_cache __read_mostly;
6566 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6567 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6569 void __init sched_init(void)
6571 unsigned long ptr = 0;
6576 #ifdef CONFIG_FAIR_GROUP_SCHED
6577 ptr += 2 * nr_cpu_ids * sizeof(void **);
6579 #ifdef CONFIG_RT_GROUP_SCHED
6580 ptr += 2 * nr_cpu_ids * sizeof(void **);
6583 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6585 #ifdef CONFIG_FAIR_GROUP_SCHED
6586 root_task_group.se = (struct sched_entity **)ptr;
6587 ptr += nr_cpu_ids * sizeof(void **);
6589 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6590 ptr += nr_cpu_ids * sizeof(void **);
6592 #endif /* CONFIG_FAIR_GROUP_SCHED */
6593 #ifdef CONFIG_RT_GROUP_SCHED
6594 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6595 ptr += nr_cpu_ids * sizeof(void **);
6597 root_task_group.rt_rq = (struct rt_rq **)ptr;
6598 ptr += nr_cpu_ids * sizeof(void **);
6600 #endif /* CONFIG_RT_GROUP_SCHED */
6602 #ifdef CONFIG_CPUMASK_OFFSTACK
6603 for_each_possible_cpu(i) {
6604 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6605 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6606 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6607 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6609 #endif /* CONFIG_CPUMASK_OFFSTACK */
6611 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6612 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6615 init_defrootdomain();
6618 #ifdef CONFIG_RT_GROUP_SCHED
6619 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6620 global_rt_period(), global_rt_runtime());
6621 #endif /* CONFIG_RT_GROUP_SCHED */
6623 #ifdef CONFIG_CGROUP_SCHED
6624 task_group_cache = KMEM_CACHE(task_group, 0);
6626 list_add(&root_task_group.list, &task_groups);
6627 INIT_LIST_HEAD(&root_task_group.children);
6628 INIT_LIST_HEAD(&root_task_group.siblings);
6629 autogroup_init(&init_task);
6630 #endif /* CONFIG_CGROUP_SCHED */
6632 for_each_possible_cpu(i) {
6636 raw_spin_lock_init(&rq->lock);
6638 rq->calc_load_active = 0;
6639 rq->calc_load_update = jiffies + LOAD_FREQ;
6640 init_cfs_rq(&rq->cfs);
6641 init_rt_rq(&rq->rt);
6642 init_dl_rq(&rq->dl);
6643 #ifdef CONFIG_FAIR_GROUP_SCHED
6644 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6645 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6646 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6648 * How much CPU bandwidth does root_task_group get?
6650 * In case of task-groups formed thr' the cgroup filesystem, it
6651 * gets 100% of the CPU resources in the system. This overall
6652 * system CPU resource is divided among the tasks of
6653 * root_task_group and its child task-groups in a fair manner,
6654 * based on each entity's (task or task-group's) weight
6655 * (se->load.weight).
6657 * In other words, if root_task_group has 10 tasks of weight
6658 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6659 * then A0's share of the CPU resource is:
6661 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6663 * We achieve this by letting root_task_group's tasks sit
6664 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6666 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6667 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6668 #endif /* CONFIG_FAIR_GROUP_SCHED */
6670 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6671 #ifdef CONFIG_RT_GROUP_SCHED
6672 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6677 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6678 rq->balance_callback = NULL;
6679 rq->active_balance = 0;
6680 rq->next_balance = jiffies;
6685 rq->avg_idle = 2*sysctl_sched_migration_cost;
6686 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6688 INIT_LIST_HEAD(&rq->cfs_tasks);
6690 rq_attach_root(rq, &def_root_domain);
6691 #ifdef CONFIG_NO_HZ_COMMON
6692 rq->last_load_update_tick = jiffies;
6693 rq->last_blocked_load_update_tick = jiffies;
6694 atomic_set(&rq->nohz_flags, 0);
6696 #endif /* CONFIG_SMP */
6698 atomic_set(&rq->nr_iowait, 0);
6701 set_load_weight(&init_task, false);
6704 * The boot idle thread does lazy MMU switching as well:
6707 enter_lazy_tlb(&init_mm, current);
6710 * Make us the idle thread. Technically, schedule() should not be
6711 * called from this thread, however somewhere below it might be,
6712 * but because we are the idle thread, we just pick up running again
6713 * when this runqueue becomes "idle".
6715 init_idle(current, smp_processor_id());
6717 calc_load_update = jiffies + LOAD_FREQ;
6720 idle_thread_set_boot_cpu();
6722 init_sched_fair_class();
6730 scheduler_running = 1;
6733 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6734 static inline int preempt_count_equals(int preempt_offset)
6736 int nested = preempt_count() + rcu_preempt_depth();
6738 return (nested == preempt_offset);
6741 void __might_sleep(const char *file, int line, int preempt_offset)
6744 * Blocking primitives will set (and therefore destroy) current->state,
6745 * since we will exit with TASK_RUNNING make sure we enter with it,
6746 * otherwise we will destroy state.
6748 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6749 "do not call blocking ops when !TASK_RUNNING; "
6750 "state=%lx set at [<%p>] %pS\n",
6752 (void *)current->task_state_change,
6753 (void *)current->task_state_change);
6755 ___might_sleep(file, line, preempt_offset);
6757 EXPORT_SYMBOL(__might_sleep);
6759 void ___might_sleep(const char *file, int line, int preempt_offset)
6761 /* Ratelimiting timestamp: */
6762 static unsigned long prev_jiffy;
6764 unsigned long preempt_disable_ip;
6766 /* WARN_ON_ONCE() by default, no rate limit required: */
6769 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6770 !is_idle_task(current)) ||
6771 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6775 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6777 prev_jiffy = jiffies;
6779 /* Save this before calling printk(), since that will clobber it: */
6780 preempt_disable_ip = get_preempt_disable_ip(current);
6783 "BUG: sleeping function called from invalid context at %s:%d\n",
6786 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6787 in_atomic(), irqs_disabled(),
6788 current->pid, current->comm);
6790 if (task_stack_end_corrupted(current))
6791 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6793 debug_show_held_locks(current);
6794 if (irqs_disabled())
6795 print_irqtrace_events(current);
6796 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6797 && !preempt_count_equals(preempt_offset)) {
6798 pr_err("Preemption disabled at:");
6799 print_ip_sym(preempt_disable_ip);
6803 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6805 EXPORT_SYMBOL(___might_sleep);
6807 void __cant_sleep(const char *file, int line, int preempt_offset)
6809 static unsigned long prev_jiffy;
6811 if (irqs_disabled())
6814 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6817 if (preempt_count() > preempt_offset)
6820 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6822 prev_jiffy = jiffies;
6824 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6825 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6826 in_atomic(), irqs_disabled(),
6827 current->pid, current->comm);
6829 debug_show_held_locks(current);
6831 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6833 EXPORT_SYMBOL_GPL(__cant_sleep);
6836 #ifdef CONFIG_MAGIC_SYSRQ
6837 void normalize_rt_tasks(void)
6839 struct task_struct *g, *p;
6840 struct sched_attr attr = {
6841 .sched_policy = SCHED_NORMAL,
6844 read_lock(&tasklist_lock);
6845 for_each_process_thread(g, p) {
6847 * Only normalize user tasks:
6849 if (p->flags & PF_KTHREAD)
6852 p->se.exec_start = 0;
6853 schedstat_set(p->se.statistics.wait_start, 0);
6854 schedstat_set(p->se.statistics.sleep_start, 0);
6855 schedstat_set(p->se.statistics.block_start, 0);
6857 if (!dl_task(p) && !rt_task(p)) {
6859 * Renice negative nice level userspace
6862 if (task_nice(p) < 0)
6863 set_user_nice(p, 0);
6867 __sched_setscheduler(p, &attr, false, false);
6869 read_unlock(&tasklist_lock);
6872 #endif /* CONFIG_MAGIC_SYSRQ */
6874 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6876 * These functions are only useful for the IA64 MCA handling, or kdb.
6878 * They can only be called when the whole system has been
6879 * stopped - every CPU needs to be quiescent, and no scheduling
6880 * activity can take place. Using them for anything else would
6881 * be a serious bug, and as a result, they aren't even visible
6882 * under any other configuration.
6886 * curr_task - return the current task for a given CPU.
6887 * @cpu: the processor in question.
6889 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6891 * Return: The current task for @cpu.
6893 struct task_struct *curr_task(int cpu)
6895 return cpu_curr(cpu);
6898 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6902 * ia64_set_curr_task - set the current task for a given CPU.
6903 * @cpu: the processor in question.
6904 * @p: the task pointer to set.
6906 * Description: This function must only be used when non-maskable interrupts
6907 * are serviced on a separate stack. It allows the architecture to switch the
6908 * notion of the current task on a CPU in a non-blocking manner. This function
6909 * must be called with all CPU's synchronized, and interrupts disabled, the
6910 * and caller must save the original value of the current task (see
6911 * curr_task() above) and restore that value before reenabling interrupts and
6912 * re-starting the system.
6914 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6916 void ia64_set_curr_task(int cpu, struct task_struct *p)
6923 #ifdef CONFIG_CGROUP_SCHED
6924 /* task_group_lock serializes the addition/removal of task groups */
6925 static DEFINE_SPINLOCK(task_group_lock);
6927 static inline void alloc_uclamp_sched_group(struct task_group *tg,
6928 struct task_group *parent)
6930 #ifdef CONFIG_UCLAMP_TASK_GROUP
6931 enum uclamp_id clamp_id;
6933 for_each_clamp_id(clamp_id) {
6934 uclamp_se_set(&tg->uclamp_req[clamp_id],
6935 uclamp_none(clamp_id), false);
6936 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
6941 static void sched_free_group(struct task_group *tg)
6943 free_fair_sched_group(tg);
6944 free_rt_sched_group(tg);
6946 kmem_cache_free(task_group_cache, tg);
6949 /* allocate runqueue etc for a new task group */
6950 struct task_group *sched_create_group(struct task_group *parent)
6952 struct task_group *tg;
6954 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6956 return ERR_PTR(-ENOMEM);
6958 if (!alloc_fair_sched_group(tg, parent))
6961 if (!alloc_rt_sched_group(tg, parent))
6964 alloc_uclamp_sched_group(tg, parent);
6969 sched_free_group(tg);
6970 return ERR_PTR(-ENOMEM);
6973 void sched_online_group(struct task_group *tg, struct task_group *parent)
6975 unsigned long flags;
6977 spin_lock_irqsave(&task_group_lock, flags);
6978 list_add_rcu(&tg->list, &task_groups);
6980 /* Root should already exist: */
6983 tg->parent = parent;
6984 INIT_LIST_HEAD(&tg->children);
6985 list_add_rcu(&tg->siblings, &parent->children);
6986 spin_unlock_irqrestore(&task_group_lock, flags);
6988 online_fair_sched_group(tg);
6991 /* rcu callback to free various structures associated with a task group */
6992 static void sched_free_group_rcu(struct rcu_head *rhp)
6994 /* Now it should be safe to free those cfs_rqs: */
6995 sched_free_group(container_of(rhp, struct task_group, rcu));
6998 void sched_destroy_group(struct task_group *tg)
7000 /* Wait for possible concurrent references to cfs_rqs complete: */
7001 call_rcu(&tg->rcu, sched_free_group_rcu);
7004 void sched_offline_group(struct task_group *tg)
7006 unsigned long flags;
7008 /* End participation in shares distribution: */
7009 unregister_fair_sched_group(tg);
7011 spin_lock_irqsave(&task_group_lock, flags);
7012 list_del_rcu(&tg->list);
7013 list_del_rcu(&tg->siblings);
7014 spin_unlock_irqrestore(&task_group_lock, flags);
7017 static void sched_change_group(struct task_struct *tsk, int type)
7019 struct task_group *tg;
7022 * All callers are synchronized by task_rq_lock(); we do not use RCU
7023 * which is pointless here. Thus, we pass "true" to task_css_check()
7024 * to prevent lockdep warnings.
7026 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7027 struct task_group, css);
7028 tg = autogroup_task_group(tsk, tg);
7029 tsk->sched_task_group = tg;
7031 #ifdef CONFIG_FAIR_GROUP_SCHED
7032 if (tsk->sched_class->task_change_group)
7033 tsk->sched_class->task_change_group(tsk, type);
7036 set_task_rq(tsk, task_cpu(tsk));
7040 * Change task's runqueue when it moves between groups.
7042 * The caller of this function should have put the task in its new group by
7043 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7046 void sched_move_task(struct task_struct *tsk)
7048 int queued, running, queue_flags =
7049 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7053 rq = task_rq_lock(tsk, &rf);
7054 update_rq_clock(rq);
7056 running = task_current(rq, tsk);
7057 queued = task_on_rq_queued(tsk);
7060 dequeue_task(rq, tsk, queue_flags);
7062 put_prev_task(rq, tsk);
7064 sched_change_group(tsk, TASK_MOVE_GROUP);
7067 enqueue_task(rq, tsk, queue_flags);
7069 set_next_task(rq, tsk);
7071 task_rq_unlock(rq, tsk, &rf);
7074 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7076 return css ? container_of(css, struct task_group, css) : NULL;
7079 static struct cgroup_subsys_state *
7080 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7082 struct task_group *parent = css_tg(parent_css);
7083 struct task_group *tg;
7086 /* This is early initialization for the top cgroup */
7087 return &root_task_group.css;
7090 tg = sched_create_group(parent);
7092 return ERR_PTR(-ENOMEM);
7097 /* Expose task group only after completing cgroup initialization */
7098 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7100 struct task_group *tg = css_tg(css);
7101 struct task_group *parent = css_tg(css->parent);
7104 sched_online_group(tg, parent);
7108 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7110 struct task_group *tg = css_tg(css);
7112 sched_offline_group(tg);
7115 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7117 struct task_group *tg = css_tg(css);
7120 * Relies on the RCU grace period between css_released() and this.
7122 sched_free_group(tg);
7126 * This is called before wake_up_new_task(), therefore we really only
7127 * have to set its group bits, all the other stuff does not apply.
7129 static void cpu_cgroup_fork(struct task_struct *task)
7134 rq = task_rq_lock(task, &rf);
7136 update_rq_clock(rq);
7137 sched_change_group(task, TASK_SET_GROUP);
7139 task_rq_unlock(rq, task, &rf);
7142 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7144 struct task_struct *task;
7145 struct cgroup_subsys_state *css;
7148 cgroup_taskset_for_each(task, css, tset) {
7149 #ifdef CONFIG_RT_GROUP_SCHED
7150 if (!sched_rt_can_attach(css_tg(css), task))
7154 * Serialize against wake_up_new_task() such that if its
7155 * running, we're sure to observe its full state.
7157 raw_spin_lock_irq(&task->pi_lock);
7159 * Avoid calling sched_move_task() before wake_up_new_task()
7160 * has happened. This would lead to problems with PELT, due to
7161 * move wanting to detach+attach while we're not attached yet.
7163 if (task->state == TASK_NEW)
7165 raw_spin_unlock_irq(&task->pi_lock);
7173 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7175 struct task_struct *task;
7176 struct cgroup_subsys_state *css;
7178 cgroup_taskset_for_each(task, css, tset)
7179 sched_move_task(task);
7182 #ifdef CONFIG_UCLAMP_TASK_GROUP
7183 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7185 struct cgroup_subsys_state *top_css = css;
7186 struct uclamp_se *uc_parent = NULL;
7187 struct uclamp_se *uc_se = NULL;
7188 unsigned int eff[UCLAMP_CNT];
7189 enum uclamp_id clamp_id;
7190 unsigned int clamps;
7192 css_for_each_descendant_pre(css, top_css) {
7193 uc_parent = css_tg(css)->parent
7194 ? css_tg(css)->parent->uclamp : NULL;
7196 for_each_clamp_id(clamp_id) {
7197 /* Assume effective clamps matches requested clamps */
7198 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7199 /* Cap effective clamps with parent's effective clamps */
7201 eff[clamp_id] > uc_parent[clamp_id].value) {
7202 eff[clamp_id] = uc_parent[clamp_id].value;
7205 /* Ensure protection is always capped by limit */
7206 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7208 /* Propagate most restrictive effective clamps */
7210 uc_se = css_tg(css)->uclamp;
7211 for_each_clamp_id(clamp_id) {
7212 if (eff[clamp_id] == uc_se[clamp_id].value)
7214 uc_se[clamp_id].value = eff[clamp_id];
7215 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7216 clamps |= (0x1 << clamp_id);
7219 css = css_rightmost_descendant(css);
7223 /* Immediately update descendants RUNNABLE tasks */
7224 uclamp_update_active_tasks(css, clamps);
7229 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7230 * C expression. Since there is no way to convert a macro argument (N) into a
7231 * character constant, use two levels of macros.
7233 #define _POW10(exp) ((unsigned int)1e##exp)
7234 #define POW10(exp) _POW10(exp)
7236 struct uclamp_request {
7237 #define UCLAMP_PERCENT_SHIFT 2
7238 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7244 static inline struct uclamp_request
7245 capacity_from_percent(char *buf)
7247 struct uclamp_request req = {
7248 .percent = UCLAMP_PERCENT_SCALE,
7249 .util = SCHED_CAPACITY_SCALE,
7254 if (strcmp(buf, "max")) {
7255 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7259 if (req.percent > UCLAMP_PERCENT_SCALE) {
7264 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7265 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7271 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7272 size_t nbytes, loff_t off,
7273 enum uclamp_id clamp_id)
7275 struct uclamp_request req;
7276 struct task_group *tg;
7278 req = capacity_from_percent(buf);
7282 mutex_lock(&uclamp_mutex);
7285 tg = css_tg(of_css(of));
7286 if (tg->uclamp_req[clamp_id].value != req.util)
7287 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7290 * Because of not recoverable conversion rounding we keep track of the
7291 * exact requested value
7293 tg->uclamp_pct[clamp_id] = req.percent;
7295 /* Update effective clamps to track the most restrictive value */
7296 cpu_util_update_eff(of_css(of));
7299 mutex_unlock(&uclamp_mutex);
7304 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7305 char *buf, size_t nbytes,
7308 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7311 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7312 char *buf, size_t nbytes,
7315 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7318 static inline void cpu_uclamp_print(struct seq_file *sf,
7319 enum uclamp_id clamp_id)
7321 struct task_group *tg;
7327 tg = css_tg(seq_css(sf));
7328 util_clamp = tg->uclamp_req[clamp_id].value;
7331 if (util_clamp == SCHED_CAPACITY_SCALE) {
7332 seq_puts(sf, "max\n");
7336 percent = tg->uclamp_pct[clamp_id];
7337 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7338 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7341 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7343 cpu_uclamp_print(sf, UCLAMP_MIN);
7347 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7349 cpu_uclamp_print(sf, UCLAMP_MAX);
7352 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7354 #ifdef CONFIG_FAIR_GROUP_SCHED
7355 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7356 struct cftype *cftype, u64 shareval)
7358 if (shareval > scale_load_down(ULONG_MAX))
7359 shareval = MAX_SHARES;
7360 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7363 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7366 struct task_group *tg = css_tg(css);
7368 return (u64) scale_load_down(tg->shares);
7371 #ifdef CONFIG_CFS_BANDWIDTH
7372 static DEFINE_MUTEX(cfs_constraints_mutex);
7374 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7375 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7377 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7379 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7381 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7382 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7384 if (tg == &root_task_group)
7388 * Ensure we have at some amount of bandwidth every period. This is
7389 * to prevent reaching a state of large arrears when throttled via
7390 * entity_tick() resulting in prolonged exit starvation.
7392 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7396 * Likewise, bound things on the otherside by preventing insane quota
7397 * periods. This also allows us to normalize in computing quota
7400 if (period > max_cfs_quota_period)
7404 * Prevent race between setting of cfs_rq->runtime_enabled and
7405 * unthrottle_offline_cfs_rqs().
7408 mutex_lock(&cfs_constraints_mutex);
7409 ret = __cfs_schedulable(tg, period, quota);
7413 runtime_enabled = quota != RUNTIME_INF;
7414 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7416 * If we need to toggle cfs_bandwidth_used, off->on must occur
7417 * before making related changes, and on->off must occur afterwards
7419 if (runtime_enabled && !runtime_was_enabled)
7420 cfs_bandwidth_usage_inc();
7421 raw_spin_lock_irq(&cfs_b->lock);
7422 cfs_b->period = ns_to_ktime(period);
7423 cfs_b->quota = quota;
7425 __refill_cfs_bandwidth_runtime(cfs_b);
7427 /* Restart the period timer (if active) to handle new period expiry: */
7428 if (runtime_enabled)
7429 start_cfs_bandwidth(cfs_b);
7431 raw_spin_unlock_irq(&cfs_b->lock);
7433 for_each_online_cpu(i) {
7434 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7435 struct rq *rq = cfs_rq->rq;
7438 rq_lock_irq(rq, &rf);
7439 cfs_rq->runtime_enabled = runtime_enabled;
7440 cfs_rq->runtime_remaining = 0;
7442 if (cfs_rq->throttled)
7443 unthrottle_cfs_rq(cfs_rq);
7444 rq_unlock_irq(rq, &rf);
7446 if (runtime_was_enabled && !runtime_enabled)
7447 cfs_bandwidth_usage_dec();
7449 mutex_unlock(&cfs_constraints_mutex);
7455 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7459 period = ktime_to_ns(tg->cfs_bandwidth.period);
7460 if (cfs_quota_us < 0)
7461 quota = RUNTIME_INF;
7462 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7463 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7467 return tg_set_cfs_bandwidth(tg, period, quota);
7470 static long tg_get_cfs_quota(struct task_group *tg)
7474 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7477 quota_us = tg->cfs_bandwidth.quota;
7478 do_div(quota_us, NSEC_PER_USEC);
7483 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7487 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7490 period = (u64)cfs_period_us * NSEC_PER_USEC;
7491 quota = tg->cfs_bandwidth.quota;
7493 return tg_set_cfs_bandwidth(tg, period, quota);
7496 static long tg_get_cfs_period(struct task_group *tg)
7500 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7501 do_div(cfs_period_us, NSEC_PER_USEC);
7503 return cfs_period_us;
7506 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7509 return tg_get_cfs_quota(css_tg(css));
7512 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7513 struct cftype *cftype, s64 cfs_quota_us)
7515 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7518 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7521 return tg_get_cfs_period(css_tg(css));
7524 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7525 struct cftype *cftype, u64 cfs_period_us)
7527 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7530 struct cfs_schedulable_data {
7531 struct task_group *tg;
7536 * normalize group quota/period to be quota/max_period
7537 * note: units are usecs
7539 static u64 normalize_cfs_quota(struct task_group *tg,
7540 struct cfs_schedulable_data *d)
7548 period = tg_get_cfs_period(tg);
7549 quota = tg_get_cfs_quota(tg);
7552 /* note: these should typically be equivalent */
7553 if (quota == RUNTIME_INF || quota == -1)
7556 return to_ratio(period, quota);
7559 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7561 struct cfs_schedulable_data *d = data;
7562 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7563 s64 quota = 0, parent_quota = -1;
7566 quota = RUNTIME_INF;
7568 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7570 quota = normalize_cfs_quota(tg, d);
7571 parent_quota = parent_b->hierarchical_quota;
7574 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7575 * always take the min. On cgroup1, only inherit when no
7578 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7579 quota = min(quota, parent_quota);
7581 if (quota == RUNTIME_INF)
7582 quota = parent_quota;
7583 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7587 cfs_b->hierarchical_quota = quota;
7592 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7595 struct cfs_schedulable_data data = {
7601 if (quota != RUNTIME_INF) {
7602 do_div(data.period, NSEC_PER_USEC);
7603 do_div(data.quota, NSEC_PER_USEC);
7607 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7613 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7615 struct task_group *tg = css_tg(seq_css(sf));
7616 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7618 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7619 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7620 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7622 if (schedstat_enabled() && tg != &root_task_group) {
7626 for_each_possible_cpu(i)
7627 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7629 seq_printf(sf, "wait_sum %llu\n", ws);
7634 #endif /* CONFIG_CFS_BANDWIDTH */
7635 #endif /* CONFIG_FAIR_GROUP_SCHED */
7637 #ifdef CONFIG_RT_GROUP_SCHED
7638 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7639 struct cftype *cft, s64 val)
7641 return sched_group_set_rt_runtime(css_tg(css), val);
7644 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7647 return sched_group_rt_runtime(css_tg(css));
7650 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7651 struct cftype *cftype, u64 rt_period_us)
7653 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7656 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7659 return sched_group_rt_period(css_tg(css));
7661 #endif /* CONFIG_RT_GROUP_SCHED */
7663 static struct cftype cpu_legacy_files[] = {
7664 #ifdef CONFIG_FAIR_GROUP_SCHED
7667 .read_u64 = cpu_shares_read_u64,
7668 .write_u64 = cpu_shares_write_u64,
7671 #ifdef CONFIG_CFS_BANDWIDTH
7673 .name = "cfs_quota_us",
7674 .read_s64 = cpu_cfs_quota_read_s64,
7675 .write_s64 = cpu_cfs_quota_write_s64,
7678 .name = "cfs_period_us",
7679 .read_u64 = cpu_cfs_period_read_u64,
7680 .write_u64 = cpu_cfs_period_write_u64,
7684 .seq_show = cpu_cfs_stat_show,
7687 #ifdef CONFIG_RT_GROUP_SCHED
7689 .name = "rt_runtime_us",
7690 .read_s64 = cpu_rt_runtime_read,
7691 .write_s64 = cpu_rt_runtime_write,
7694 .name = "rt_period_us",
7695 .read_u64 = cpu_rt_period_read_uint,
7696 .write_u64 = cpu_rt_period_write_uint,
7699 #ifdef CONFIG_UCLAMP_TASK_GROUP
7701 .name = "uclamp.min",
7702 .flags = CFTYPE_NOT_ON_ROOT,
7703 .seq_show = cpu_uclamp_min_show,
7704 .write = cpu_uclamp_min_write,
7707 .name = "uclamp.max",
7708 .flags = CFTYPE_NOT_ON_ROOT,
7709 .seq_show = cpu_uclamp_max_show,
7710 .write = cpu_uclamp_max_write,
7716 static int cpu_extra_stat_show(struct seq_file *sf,
7717 struct cgroup_subsys_state *css)
7719 #ifdef CONFIG_CFS_BANDWIDTH
7721 struct task_group *tg = css_tg(css);
7722 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7725 throttled_usec = cfs_b->throttled_time;
7726 do_div(throttled_usec, NSEC_PER_USEC);
7728 seq_printf(sf, "nr_periods %d\n"
7730 "throttled_usec %llu\n",
7731 cfs_b->nr_periods, cfs_b->nr_throttled,
7738 #ifdef CONFIG_FAIR_GROUP_SCHED
7739 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7742 struct task_group *tg = css_tg(css);
7743 u64 weight = scale_load_down(tg->shares);
7745 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7748 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7749 struct cftype *cft, u64 weight)
7752 * cgroup weight knobs should use the common MIN, DFL and MAX
7753 * values which are 1, 100 and 10000 respectively. While it loses
7754 * a bit of range on both ends, it maps pretty well onto the shares
7755 * value used by scheduler and the round-trip conversions preserve
7756 * the original value over the entire range.
7758 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7761 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7763 return sched_group_set_shares(css_tg(css), scale_load(weight));
7766 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7769 unsigned long weight = scale_load_down(css_tg(css)->shares);
7770 int last_delta = INT_MAX;
7773 /* find the closest nice value to the current weight */
7774 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7775 delta = abs(sched_prio_to_weight[prio] - weight);
7776 if (delta >= last_delta)
7781 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7784 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7785 struct cftype *cft, s64 nice)
7787 unsigned long weight;
7790 if (nice < MIN_NICE || nice > MAX_NICE)
7793 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7794 idx = array_index_nospec(idx, 40);
7795 weight = sched_prio_to_weight[idx];
7797 return sched_group_set_shares(css_tg(css), scale_load(weight));
7801 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7802 long period, long quota)
7805 seq_puts(sf, "max");
7807 seq_printf(sf, "%ld", quota);
7809 seq_printf(sf, " %ld\n", period);
7812 /* caller should put the current value in *@periodp before calling */
7813 static int __maybe_unused cpu_period_quota_parse(char *buf,
7814 u64 *periodp, u64 *quotap)
7816 char tok[21]; /* U64_MAX */
7818 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7821 *periodp *= NSEC_PER_USEC;
7823 if (sscanf(tok, "%llu", quotap))
7824 *quotap *= NSEC_PER_USEC;
7825 else if (!strcmp(tok, "max"))
7826 *quotap = RUNTIME_INF;
7833 #ifdef CONFIG_CFS_BANDWIDTH
7834 static int cpu_max_show(struct seq_file *sf, void *v)
7836 struct task_group *tg = css_tg(seq_css(sf));
7838 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7842 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7843 char *buf, size_t nbytes, loff_t off)
7845 struct task_group *tg = css_tg(of_css(of));
7846 u64 period = tg_get_cfs_period(tg);
7850 ret = cpu_period_quota_parse(buf, &period, "a);
7852 ret = tg_set_cfs_bandwidth(tg, period, quota);
7853 return ret ?: nbytes;
7857 static struct cftype cpu_files[] = {
7858 #ifdef CONFIG_FAIR_GROUP_SCHED
7861 .flags = CFTYPE_NOT_ON_ROOT,
7862 .read_u64 = cpu_weight_read_u64,
7863 .write_u64 = cpu_weight_write_u64,
7866 .name = "weight.nice",
7867 .flags = CFTYPE_NOT_ON_ROOT,
7868 .read_s64 = cpu_weight_nice_read_s64,
7869 .write_s64 = cpu_weight_nice_write_s64,
7872 #ifdef CONFIG_CFS_BANDWIDTH
7875 .flags = CFTYPE_NOT_ON_ROOT,
7876 .seq_show = cpu_max_show,
7877 .write = cpu_max_write,
7880 #ifdef CONFIG_UCLAMP_TASK_GROUP
7882 .name = "uclamp.min",
7883 .flags = CFTYPE_NOT_ON_ROOT,
7884 .seq_show = cpu_uclamp_min_show,
7885 .write = cpu_uclamp_min_write,
7888 .name = "uclamp.max",
7889 .flags = CFTYPE_NOT_ON_ROOT,
7890 .seq_show = cpu_uclamp_max_show,
7891 .write = cpu_uclamp_max_write,
7897 struct cgroup_subsys cpu_cgrp_subsys = {
7898 .css_alloc = cpu_cgroup_css_alloc,
7899 .css_online = cpu_cgroup_css_online,
7900 .css_released = cpu_cgroup_css_released,
7901 .css_free = cpu_cgroup_css_free,
7902 .css_extra_stat_show = cpu_extra_stat_show,
7903 .fork = cpu_cgroup_fork,
7904 .can_attach = cpu_cgroup_can_attach,
7905 .attach = cpu_cgroup_attach,
7906 .legacy_cftypes = cpu_legacy_files,
7907 .dfl_cftypes = cpu_files,
7912 #endif /* CONFIG_CGROUP_SCHED */
7914 void dump_cpu_task(int cpu)
7916 pr_info("Task dump for CPU %d:\n", cpu);
7917 sched_show_task(cpu_curr(cpu));
7921 * Nice levels are multiplicative, with a gentle 10% change for every
7922 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7923 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7924 * that remained on nice 0.
7926 * The "10% effect" is relative and cumulative: from _any_ nice level,
7927 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7928 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7929 * If a task goes up by ~10% and another task goes down by ~10% then
7930 * the relative distance between them is ~25%.)
7932 const int sched_prio_to_weight[40] = {
7933 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7934 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7935 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7936 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7937 /* 0 */ 1024, 820, 655, 526, 423,
7938 /* 5 */ 335, 272, 215, 172, 137,
7939 /* 10 */ 110, 87, 70, 56, 45,
7940 /* 15 */ 36, 29, 23, 18, 15,
7944 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7946 * In cases where the weight does not change often, we can use the
7947 * precalculated inverse to speed up arithmetics by turning divisions
7948 * into multiplications:
7950 const u32 sched_prio_to_wmult[40] = {
7951 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7952 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7953 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7954 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7955 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7956 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7957 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7958 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7961 #undef CREATE_TRACE_POINTS