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_HARD);
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_HARD);
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_HARD);
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 unsigned int uclamp_none(enum uclamp_id clamp_id)
815 if (clamp_id == UCLAMP_MIN)
817 return SCHED_CAPACITY_SCALE;
820 static inline void uclamp_se_set(struct uclamp_se *uc_se,
821 unsigned int value, bool user_defined)
823 uc_se->value = value;
824 uc_se->bucket_id = uclamp_bucket_id(value);
825 uc_se->user_defined = user_defined;
828 static inline unsigned int
829 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
830 unsigned int clamp_value)
833 * Avoid blocked utilization pushing up the frequency when we go
834 * idle (which drops the max-clamp) by retaining the last known
837 if (clamp_id == UCLAMP_MAX) {
838 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
842 return uclamp_none(UCLAMP_MIN);
845 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
846 unsigned int clamp_value)
848 /* Reset max-clamp retention only on idle exit */
849 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
852 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
856 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
857 unsigned int clamp_value)
859 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
860 int bucket_id = UCLAMP_BUCKETS - 1;
863 * Since both min and max clamps are max aggregated, find the
864 * top most bucket with tasks in.
866 for ( ; bucket_id >= 0; bucket_id--) {
867 if (!bucket[bucket_id].tasks)
869 return bucket[bucket_id].value;
872 /* No tasks -- default clamp values */
873 return uclamp_idle_value(rq, clamp_id, clamp_value);
876 static inline struct uclamp_se
877 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
879 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
880 #ifdef CONFIG_UCLAMP_TASK_GROUP
881 struct uclamp_se uc_max;
884 * Tasks in autogroups or root task group will be
885 * restricted by system defaults.
887 if (task_group_is_autogroup(task_group(p)))
889 if (task_group(p) == &root_task_group)
892 uc_max = task_group(p)->uclamp[clamp_id];
893 if (uc_req.value > uc_max.value || !uc_req.user_defined)
901 * The effective clamp bucket index of a task depends on, by increasing
903 * - the task specific clamp value, when explicitly requested from userspace
904 * - the task group effective clamp value, for tasks not either in the root
905 * group or in an autogroup
906 * - the system default clamp value, defined by the sysadmin
908 static inline struct uclamp_se
909 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
911 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
912 struct uclamp_se uc_max = uclamp_default[clamp_id];
914 /* System default restrictions always apply */
915 if (unlikely(uc_req.value > uc_max.value))
921 unsigned int 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);
1076 #ifdef CONFIG_UCLAMP_TASK_GROUP
1078 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1079 unsigned int clamps)
1081 enum uclamp_id clamp_id;
1082 struct css_task_iter it;
1083 struct task_struct *p;
1085 css_task_iter_start(css, 0, &it);
1086 while ((p = css_task_iter_next(&it))) {
1087 for_each_clamp_id(clamp_id) {
1088 if ((0x1 << clamp_id) & clamps)
1089 uclamp_update_active(p, clamp_id);
1092 css_task_iter_end(&it);
1095 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1096 static void uclamp_update_root_tg(void)
1098 struct task_group *tg = &root_task_group;
1100 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1101 sysctl_sched_uclamp_util_min, false);
1102 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1103 sysctl_sched_uclamp_util_max, false);
1106 cpu_util_update_eff(&root_task_group.css);
1110 static void uclamp_update_root_tg(void) { }
1113 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1114 void __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 uclamp_se_set(&p->uclamp_req[clamp_id],
1237 uclamp_none(clamp_id), false);
1241 static void __init init_uclamp(void)
1243 struct uclamp_se uc_max = {};
1244 enum uclamp_id clamp_id;
1247 mutex_init(&uclamp_mutex);
1249 for_each_possible_cpu(cpu) {
1250 memset(&cpu_rq(cpu)->uclamp, 0,
1251 sizeof(struct uclamp_rq)*UCLAMP_CNT);
1252 cpu_rq(cpu)->uclamp_flags = 0;
1255 for_each_clamp_id(clamp_id) {
1256 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1257 uclamp_none(clamp_id), false);
1260 /* System defaults allow max clamp values for both indexes */
1261 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1262 for_each_clamp_id(clamp_id) {
1263 uclamp_default[clamp_id] = uc_max;
1264 #ifdef CONFIG_UCLAMP_TASK_GROUP
1265 root_task_group.uclamp_req[clamp_id] = uc_max;
1266 root_task_group.uclamp[clamp_id] = uc_max;
1271 #else /* CONFIG_UCLAMP_TASK */
1272 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1273 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1274 static inline int uclamp_validate(struct task_struct *p,
1275 const struct sched_attr *attr)
1279 static void __setscheduler_uclamp(struct task_struct *p,
1280 const struct sched_attr *attr) { }
1281 static inline void uclamp_fork(struct task_struct *p) { }
1282 static inline void init_uclamp(void) { }
1283 #endif /* CONFIG_UCLAMP_TASK */
1285 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1287 if (!(flags & ENQUEUE_NOCLOCK))
1288 update_rq_clock(rq);
1290 if (!(flags & ENQUEUE_RESTORE)) {
1291 sched_info_queued(rq, p);
1292 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1295 uclamp_rq_inc(rq, p);
1296 p->sched_class->enqueue_task(rq, p, flags);
1299 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1301 if (!(flags & DEQUEUE_NOCLOCK))
1302 update_rq_clock(rq);
1304 if (!(flags & DEQUEUE_SAVE)) {
1305 sched_info_dequeued(rq, p);
1306 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1309 uclamp_rq_dec(rq, p);
1310 p->sched_class->dequeue_task(rq, p, flags);
1313 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1315 if (task_contributes_to_load(p))
1316 rq->nr_uninterruptible--;
1318 enqueue_task(rq, p, flags);
1320 p->on_rq = TASK_ON_RQ_QUEUED;
1323 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1325 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1327 if (task_contributes_to_load(p))
1328 rq->nr_uninterruptible++;
1330 dequeue_task(rq, p, flags);
1334 * __normal_prio - return the priority that is based on the static prio
1336 static inline int __normal_prio(struct task_struct *p)
1338 return p->static_prio;
1342 * Calculate the expected normal priority: i.e. priority
1343 * without taking RT-inheritance into account. Might be
1344 * boosted by interactivity modifiers. Changes upon fork,
1345 * setprio syscalls, and whenever the interactivity
1346 * estimator recalculates.
1348 static inline int normal_prio(struct task_struct *p)
1352 if (task_has_dl_policy(p))
1353 prio = MAX_DL_PRIO-1;
1354 else if (task_has_rt_policy(p))
1355 prio = MAX_RT_PRIO-1 - p->rt_priority;
1357 prio = __normal_prio(p);
1362 * Calculate the current priority, i.e. the priority
1363 * taken into account by the scheduler. This value might
1364 * be boosted by RT tasks, or might be boosted by
1365 * interactivity modifiers. Will be RT if the task got
1366 * RT-boosted. If not then it returns p->normal_prio.
1368 static int effective_prio(struct task_struct *p)
1370 p->normal_prio = normal_prio(p);
1372 * If we are RT tasks or we were boosted to RT priority,
1373 * keep the priority unchanged. Otherwise, update priority
1374 * to the normal priority:
1376 if (!rt_prio(p->prio))
1377 return p->normal_prio;
1382 * task_curr - is this task currently executing on a CPU?
1383 * @p: the task in question.
1385 * Return: 1 if the task is currently executing. 0 otherwise.
1387 inline int task_curr(const struct task_struct *p)
1389 return cpu_curr(task_cpu(p)) == p;
1393 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1394 * use the balance_callback list if you want balancing.
1396 * this means any call to check_class_changed() must be followed by a call to
1397 * balance_callback().
1399 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1400 const struct sched_class *prev_class,
1403 if (prev_class != p->sched_class) {
1404 if (prev_class->switched_from)
1405 prev_class->switched_from(rq, p);
1407 p->sched_class->switched_to(rq, p);
1408 } else if (oldprio != p->prio || dl_task(p))
1409 p->sched_class->prio_changed(rq, p, oldprio);
1412 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1414 const struct sched_class *class;
1416 if (p->sched_class == rq->curr->sched_class) {
1417 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1419 for_each_class(class) {
1420 if (class == rq->curr->sched_class)
1422 if (class == p->sched_class) {
1430 * A queue event has occurred, and we're going to schedule. In
1431 * this case, we can save a useless back to back clock update.
1433 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1434 rq_clock_skip_update(rq);
1439 static inline bool is_per_cpu_kthread(struct task_struct *p)
1441 if (!(p->flags & PF_KTHREAD))
1444 if (p->nr_cpus_allowed != 1)
1451 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1452 * __set_cpus_allowed_ptr() and select_fallback_rq().
1454 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1456 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1459 if (is_per_cpu_kthread(p))
1460 return cpu_online(cpu);
1462 return cpu_active(cpu);
1466 * This is how migration works:
1468 * 1) we invoke migration_cpu_stop() on the target CPU using
1470 * 2) stopper starts to run (implicitly forcing the migrated thread
1472 * 3) it checks whether the migrated task is still in the wrong runqueue.
1473 * 4) if it's in the wrong runqueue then the migration thread removes
1474 * it and puts it into the right queue.
1475 * 5) stopper completes and stop_one_cpu() returns and the migration
1480 * move_queued_task - move a queued task to new rq.
1482 * Returns (locked) new rq. Old rq's lock is released.
1484 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1485 struct task_struct *p, int new_cpu)
1487 lockdep_assert_held(&rq->lock);
1489 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1490 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1491 set_task_cpu(p, new_cpu);
1494 rq = cpu_rq(new_cpu);
1497 BUG_ON(task_cpu(p) != new_cpu);
1498 enqueue_task(rq, p, 0);
1499 p->on_rq = TASK_ON_RQ_QUEUED;
1500 check_preempt_curr(rq, p, 0);
1505 struct migration_arg {
1506 struct task_struct *task;
1511 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1512 * this because either it can't run here any more (set_cpus_allowed()
1513 * away from this CPU, or CPU going down), or because we're
1514 * attempting to rebalance this task on exec (sched_exec).
1516 * So we race with normal scheduler movements, but that's OK, as long
1517 * as the task is no longer on this CPU.
1519 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1520 struct task_struct *p, int dest_cpu)
1522 /* Affinity changed (again). */
1523 if (!is_cpu_allowed(p, dest_cpu))
1526 update_rq_clock(rq);
1527 rq = move_queued_task(rq, rf, p, dest_cpu);
1533 * migration_cpu_stop - this will be executed by a highprio stopper thread
1534 * and performs thread migration by bumping thread off CPU then
1535 * 'pushing' onto another runqueue.
1537 static int migration_cpu_stop(void *data)
1539 struct migration_arg *arg = data;
1540 struct task_struct *p = arg->task;
1541 struct rq *rq = this_rq();
1545 * The original target CPU might have gone down and we might
1546 * be on another CPU but it doesn't matter.
1548 local_irq_disable();
1550 * We need to explicitly wake pending tasks before running
1551 * __migrate_task() such that we will not miss enforcing cpus_ptr
1552 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1554 sched_ttwu_pending();
1556 raw_spin_lock(&p->pi_lock);
1559 * If task_rq(p) != rq, it cannot be migrated here, because we're
1560 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1561 * we're holding p->pi_lock.
1563 if (task_rq(p) == rq) {
1564 if (task_on_rq_queued(p))
1565 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1567 p->wake_cpu = arg->dest_cpu;
1570 raw_spin_unlock(&p->pi_lock);
1577 * sched_class::set_cpus_allowed must do the below, but is not required to
1578 * actually call this function.
1580 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1582 cpumask_copy(&p->cpus_mask, new_mask);
1583 p->nr_cpus_allowed = cpumask_weight(new_mask);
1586 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1588 struct rq *rq = task_rq(p);
1589 bool queued, running;
1591 lockdep_assert_held(&p->pi_lock);
1593 queued = task_on_rq_queued(p);
1594 running = task_current(rq, p);
1598 * Because __kthread_bind() calls this on blocked tasks without
1601 lockdep_assert_held(&rq->lock);
1602 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1605 put_prev_task(rq, p);
1607 p->sched_class->set_cpus_allowed(p, new_mask);
1610 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1612 set_next_task(rq, p);
1616 * Change a given task's CPU affinity. Migrate the thread to a
1617 * proper CPU and schedule it away if the CPU it's executing on
1618 * is removed from the allowed bitmask.
1620 * NOTE: the caller must have a valid reference to the task, the
1621 * task must not exit() & deallocate itself prematurely. The
1622 * call is not atomic; no spinlocks may be held.
1624 static int __set_cpus_allowed_ptr(struct task_struct *p,
1625 const struct cpumask *new_mask, bool check)
1627 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1628 unsigned int dest_cpu;
1633 rq = task_rq_lock(p, &rf);
1634 update_rq_clock(rq);
1636 if (p->flags & PF_KTHREAD) {
1638 * Kernel threads are allowed on online && !active CPUs
1640 cpu_valid_mask = cpu_online_mask;
1644 * Must re-check here, to close a race against __kthread_bind(),
1645 * sched_setaffinity() is not guaranteed to observe the flag.
1647 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1652 if (cpumask_equal(p->cpus_ptr, new_mask))
1655 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1656 if (dest_cpu >= nr_cpu_ids) {
1661 do_set_cpus_allowed(p, new_mask);
1663 if (p->flags & PF_KTHREAD) {
1665 * For kernel threads that do indeed end up on online &&
1666 * !active we want to ensure they are strict per-CPU threads.
1668 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1669 !cpumask_intersects(new_mask, cpu_active_mask) &&
1670 p->nr_cpus_allowed != 1);
1673 /* Can the task run on the task's current CPU? If so, we're done */
1674 if (cpumask_test_cpu(task_cpu(p), new_mask))
1677 if (task_running(rq, p) || p->state == TASK_WAKING) {
1678 struct migration_arg arg = { p, dest_cpu };
1679 /* Need help from migration thread: drop lock and wait. */
1680 task_rq_unlock(rq, p, &rf);
1681 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1683 } else if (task_on_rq_queued(p)) {
1685 * OK, since we're going to drop the lock immediately
1686 * afterwards anyway.
1688 rq = move_queued_task(rq, &rf, p, dest_cpu);
1691 task_rq_unlock(rq, p, &rf);
1696 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1698 return __set_cpus_allowed_ptr(p, new_mask, false);
1700 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1702 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1704 #ifdef CONFIG_SCHED_DEBUG
1706 * We should never call set_task_cpu() on a blocked task,
1707 * ttwu() will sort out the placement.
1709 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1713 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1714 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1715 * time relying on p->on_rq.
1717 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1718 p->sched_class == &fair_sched_class &&
1719 (p->on_rq && !task_on_rq_migrating(p)));
1721 #ifdef CONFIG_LOCKDEP
1723 * The caller should hold either p->pi_lock or rq->lock, when changing
1724 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1726 * sched_move_task() holds both and thus holding either pins the cgroup,
1729 * Furthermore, all task_rq users should acquire both locks, see
1732 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1733 lockdep_is_held(&task_rq(p)->lock)));
1736 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1738 WARN_ON_ONCE(!cpu_online(new_cpu));
1741 trace_sched_migrate_task(p, new_cpu);
1743 if (task_cpu(p) != new_cpu) {
1744 if (p->sched_class->migrate_task_rq)
1745 p->sched_class->migrate_task_rq(p, new_cpu);
1746 p->se.nr_migrations++;
1748 perf_event_task_migrate(p);
1751 __set_task_cpu(p, new_cpu);
1754 #ifdef CONFIG_NUMA_BALANCING
1755 static void __migrate_swap_task(struct task_struct *p, int cpu)
1757 if (task_on_rq_queued(p)) {
1758 struct rq *src_rq, *dst_rq;
1759 struct rq_flags srf, drf;
1761 src_rq = task_rq(p);
1762 dst_rq = cpu_rq(cpu);
1764 rq_pin_lock(src_rq, &srf);
1765 rq_pin_lock(dst_rq, &drf);
1767 deactivate_task(src_rq, p, 0);
1768 set_task_cpu(p, cpu);
1769 activate_task(dst_rq, p, 0);
1770 check_preempt_curr(dst_rq, p, 0);
1772 rq_unpin_lock(dst_rq, &drf);
1773 rq_unpin_lock(src_rq, &srf);
1777 * Task isn't running anymore; make it appear like we migrated
1778 * it before it went to sleep. This means on wakeup we make the
1779 * previous CPU our target instead of where it really is.
1785 struct migration_swap_arg {
1786 struct task_struct *src_task, *dst_task;
1787 int src_cpu, dst_cpu;
1790 static int migrate_swap_stop(void *data)
1792 struct migration_swap_arg *arg = data;
1793 struct rq *src_rq, *dst_rq;
1796 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1799 src_rq = cpu_rq(arg->src_cpu);
1800 dst_rq = cpu_rq(arg->dst_cpu);
1802 double_raw_lock(&arg->src_task->pi_lock,
1803 &arg->dst_task->pi_lock);
1804 double_rq_lock(src_rq, dst_rq);
1806 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1809 if (task_cpu(arg->src_task) != arg->src_cpu)
1812 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1815 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1818 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1819 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1824 double_rq_unlock(src_rq, dst_rq);
1825 raw_spin_unlock(&arg->dst_task->pi_lock);
1826 raw_spin_unlock(&arg->src_task->pi_lock);
1832 * Cross migrate two tasks
1834 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1835 int target_cpu, int curr_cpu)
1837 struct migration_swap_arg arg;
1840 arg = (struct migration_swap_arg){
1842 .src_cpu = curr_cpu,
1844 .dst_cpu = target_cpu,
1847 if (arg.src_cpu == arg.dst_cpu)
1851 * These three tests are all lockless; this is OK since all of them
1852 * will be re-checked with proper locks held further down the line.
1854 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1857 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1860 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1863 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1864 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1869 #endif /* CONFIG_NUMA_BALANCING */
1872 * wait_task_inactive - wait for a thread to unschedule.
1874 * If @match_state is nonzero, it's the @p->state value just checked and
1875 * not expected to change. If it changes, i.e. @p might have woken up,
1876 * then return zero. When we succeed in waiting for @p to be off its CPU,
1877 * we return a positive number (its total switch count). If a second call
1878 * a short while later returns the same number, the caller can be sure that
1879 * @p has remained unscheduled the whole time.
1881 * The caller must ensure that the task *will* unschedule sometime soon,
1882 * else this function might spin for a *long* time. This function can't
1883 * be called with interrupts off, or it may introduce deadlock with
1884 * smp_call_function() if an IPI is sent by the same process we are
1885 * waiting to become inactive.
1887 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1889 int running, queued;
1896 * We do the initial early heuristics without holding
1897 * any task-queue locks at all. We'll only try to get
1898 * the runqueue lock when things look like they will
1904 * If the task is actively running on another CPU
1905 * still, just relax and busy-wait without holding
1908 * NOTE! Since we don't hold any locks, it's not
1909 * even sure that "rq" stays as the right runqueue!
1910 * But we don't care, since "task_running()" will
1911 * return false if the runqueue has changed and p
1912 * is actually now running somewhere else!
1914 while (task_running(rq, p)) {
1915 if (match_state && unlikely(p->state != match_state))
1921 * Ok, time to look more closely! We need the rq
1922 * lock now, to be *sure*. If we're wrong, we'll
1923 * just go back and repeat.
1925 rq = task_rq_lock(p, &rf);
1926 trace_sched_wait_task(p);
1927 running = task_running(rq, p);
1928 queued = task_on_rq_queued(p);
1930 if (!match_state || p->state == match_state)
1931 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1932 task_rq_unlock(rq, p, &rf);
1935 * If it changed from the expected state, bail out now.
1937 if (unlikely(!ncsw))
1941 * Was it really running after all now that we
1942 * checked with the proper locks actually held?
1944 * Oops. Go back and try again..
1946 if (unlikely(running)) {
1952 * It's not enough that it's not actively running,
1953 * it must be off the runqueue _entirely_, and not
1956 * So if it was still runnable (but just not actively
1957 * running right now), it's preempted, and we should
1958 * yield - it could be a while.
1960 if (unlikely(queued)) {
1961 ktime_t to = NSEC_PER_SEC / HZ;
1963 set_current_state(TASK_UNINTERRUPTIBLE);
1964 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1969 * Ahh, all good. It wasn't running, and it wasn't
1970 * runnable, which means that it will never become
1971 * running in the future either. We're all done!
1980 * kick_process - kick a running thread to enter/exit the kernel
1981 * @p: the to-be-kicked thread
1983 * Cause a process which is running on another CPU to enter
1984 * kernel-mode, without any delay. (to get signals handled.)
1986 * NOTE: this function doesn't have to take the runqueue lock,
1987 * because all it wants to ensure is that the remote task enters
1988 * the kernel. If the IPI races and the task has been migrated
1989 * to another CPU then no harm is done and the purpose has been
1992 void kick_process(struct task_struct *p)
1998 if ((cpu != smp_processor_id()) && task_curr(p))
1999 smp_send_reschedule(cpu);
2002 EXPORT_SYMBOL_GPL(kick_process);
2005 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2007 * A few notes on cpu_active vs cpu_online:
2009 * - cpu_active must be a subset of cpu_online
2011 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2012 * see __set_cpus_allowed_ptr(). At this point the newly online
2013 * CPU isn't yet part of the sched domains, and balancing will not
2016 * - on CPU-down we clear cpu_active() to mask the sched domains and
2017 * avoid the load balancer to place new tasks on the to be removed
2018 * CPU. Existing tasks will remain running there and will be taken
2021 * This means that fallback selection must not select !active CPUs.
2022 * And can assume that any active CPU must be online. Conversely
2023 * select_task_rq() below may allow selection of !active CPUs in order
2024 * to satisfy the above rules.
2026 static int select_fallback_rq(int cpu, struct task_struct *p)
2028 int nid = cpu_to_node(cpu);
2029 const struct cpumask *nodemask = NULL;
2030 enum { cpuset, possible, fail } state = cpuset;
2034 * If the node that the CPU is on has been offlined, cpu_to_node()
2035 * will return -1. There is no CPU on the node, and we should
2036 * select the CPU on the other node.
2039 nodemask = cpumask_of_node(nid);
2041 /* Look for allowed, online CPU in same node. */
2042 for_each_cpu(dest_cpu, nodemask) {
2043 if (!cpu_active(dest_cpu))
2045 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2051 /* Any allowed, online CPU? */
2052 for_each_cpu(dest_cpu, p->cpus_ptr) {
2053 if (!is_cpu_allowed(p, dest_cpu))
2059 /* No more Mr. Nice Guy. */
2062 if (IS_ENABLED(CONFIG_CPUSETS)) {
2063 cpuset_cpus_allowed_fallback(p);
2069 do_set_cpus_allowed(p, cpu_possible_mask);
2080 if (state != cpuset) {
2082 * Don't tell them about moving exiting tasks or
2083 * kernel threads (both mm NULL), since they never
2086 if (p->mm && printk_ratelimit()) {
2087 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2088 task_pid_nr(p), p->comm, cpu);
2096 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2099 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2101 lockdep_assert_held(&p->pi_lock);
2103 if (p->nr_cpus_allowed > 1)
2104 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2106 cpu = cpumask_any(p->cpus_ptr);
2109 * In order not to call set_task_cpu() on a blocking task we need
2110 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2113 * Since this is common to all placement strategies, this lives here.
2115 * [ this allows ->select_task() to simply return task_cpu(p) and
2116 * not worry about this generic constraint ]
2118 if (unlikely(!is_cpu_allowed(p, cpu)))
2119 cpu = select_fallback_rq(task_cpu(p), p);
2124 static void update_avg(u64 *avg, u64 sample)
2126 s64 diff = sample - *avg;
2130 void sched_set_stop_task(int cpu, struct task_struct *stop)
2132 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2133 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2137 * Make it appear like a SCHED_FIFO task, its something
2138 * userspace knows about and won't get confused about.
2140 * Also, it will make PI more or less work without too
2141 * much confusion -- but then, stop work should not
2142 * rely on PI working anyway.
2144 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2146 stop->sched_class = &stop_sched_class;
2149 cpu_rq(cpu)->stop = stop;
2153 * Reset it back to a normal scheduling class so that
2154 * it can die in pieces.
2156 old_stop->sched_class = &rt_sched_class;
2162 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2163 const struct cpumask *new_mask, bool check)
2165 return set_cpus_allowed_ptr(p, new_mask);
2168 #endif /* CONFIG_SMP */
2171 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2175 if (!schedstat_enabled())
2181 if (cpu == rq->cpu) {
2182 __schedstat_inc(rq->ttwu_local);
2183 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2185 struct sched_domain *sd;
2187 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2189 for_each_domain(rq->cpu, sd) {
2190 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2191 __schedstat_inc(sd->ttwu_wake_remote);
2198 if (wake_flags & WF_MIGRATED)
2199 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2200 #endif /* CONFIG_SMP */
2202 __schedstat_inc(rq->ttwu_count);
2203 __schedstat_inc(p->se.statistics.nr_wakeups);
2205 if (wake_flags & WF_SYNC)
2206 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2210 * Mark the task runnable and perform wakeup-preemption.
2212 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2213 struct rq_flags *rf)
2215 check_preempt_curr(rq, p, wake_flags);
2216 p->state = TASK_RUNNING;
2217 trace_sched_wakeup(p);
2220 if (p->sched_class->task_woken) {
2222 * Our task @p is fully woken up and running; so its safe to
2223 * drop the rq->lock, hereafter rq is only used for statistics.
2225 rq_unpin_lock(rq, rf);
2226 p->sched_class->task_woken(rq, p);
2227 rq_repin_lock(rq, rf);
2230 if (rq->idle_stamp) {
2231 u64 delta = rq_clock(rq) - rq->idle_stamp;
2232 u64 max = 2*rq->max_idle_balance_cost;
2234 update_avg(&rq->avg_idle, delta);
2236 if (rq->avg_idle > max)
2245 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2246 struct rq_flags *rf)
2248 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2250 lockdep_assert_held(&rq->lock);
2253 if (p->sched_contributes_to_load)
2254 rq->nr_uninterruptible--;
2256 if (wake_flags & WF_MIGRATED)
2257 en_flags |= ENQUEUE_MIGRATED;
2260 activate_task(rq, p, en_flags);
2261 ttwu_do_wakeup(rq, p, wake_flags, rf);
2265 * Called in case the task @p isn't fully descheduled from its runqueue,
2266 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2267 * since all we need to do is flip p->state to TASK_RUNNING, since
2268 * the task is still ->on_rq.
2270 static int ttwu_remote(struct task_struct *p, int wake_flags)
2276 rq = __task_rq_lock(p, &rf);
2277 if (task_on_rq_queued(p)) {
2278 /* check_preempt_curr() may use rq clock */
2279 update_rq_clock(rq);
2280 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2283 __task_rq_unlock(rq, &rf);
2289 void sched_ttwu_pending(void)
2291 struct rq *rq = this_rq();
2292 struct llist_node *llist = llist_del_all(&rq->wake_list);
2293 struct task_struct *p, *t;
2299 rq_lock_irqsave(rq, &rf);
2300 update_rq_clock(rq);
2302 llist_for_each_entry_safe(p, t, llist, wake_entry)
2303 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2305 rq_unlock_irqrestore(rq, &rf);
2308 void scheduler_ipi(void)
2311 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2312 * TIF_NEED_RESCHED remotely (for the first time) will also send
2315 preempt_fold_need_resched();
2317 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2321 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2322 * traditionally all their work was done from the interrupt return
2323 * path. Now that we actually do some work, we need to make sure
2326 * Some archs already do call them, luckily irq_enter/exit nest
2329 * Arguably we should visit all archs and update all handlers,
2330 * however a fair share of IPIs are still resched only so this would
2331 * somewhat pessimize the simple resched case.
2334 sched_ttwu_pending();
2337 * Check if someone kicked us for doing the nohz idle load balance.
2339 if (unlikely(got_nohz_idle_kick())) {
2340 this_rq()->idle_balance = 1;
2341 raise_softirq_irqoff(SCHED_SOFTIRQ);
2346 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2348 struct rq *rq = cpu_rq(cpu);
2350 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2352 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2353 if (!set_nr_if_polling(rq->idle))
2354 smp_send_reschedule(cpu);
2356 trace_sched_wake_idle_without_ipi(cpu);
2360 void wake_up_if_idle(int cpu)
2362 struct rq *rq = cpu_rq(cpu);
2367 if (!is_idle_task(rcu_dereference(rq->curr)))
2370 if (set_nr_if_polling(rq->idle)) {
2371 trace_sched_wake_idle_without_ipi(cpu);
2373 rq_lock_irqsave(rq, &rf);
2374 if (is_idle_task(rq->curr))
2375 smp_send_reschedule(cpu);
2376 /* Else CPU is not idle, do nothing here: */
2377 rq_unlock_irqrestore(rq, &rf);
2384 bool cpus_share_cache(int this_cpu, int that_cpu)
2386 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2388 #endif /* CONFIG_SMP */
2390 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2392 struct rq *rq = cpu_rq(cpu);
2395 #if defined(CONFIG_SMP)
2396 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2397 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2398 ttwu_queue_remote(p, cpu, wake_flags);
2404 update_rq_clock(rq);
2405 ttwu_do_activate(rq, p, wake_flags, &rf);
2410 * Notes on Program-Order guarantees on SMP systems.
2414 * The basic program-order guarantee on SMP systems is that when a task [t]
2415 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2416 * execution on its new CPU [c1].
2418 * For migration (of runnable tasks) this is provided by the following means:
2420 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2421 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2422 * rq(c1)->lock (if not at the same time, then in that order).
2423 * C) LOCK of the rq(c1)->lock scheduling in task
2425 * Release/acquire chaining guarantees that B happens after A and C after B.
2426 * Note: the CPU doing B need not be c0 or c1
2435 * UNLOCK rq(0)->lock
2437 * LOCK rq(0)->lock // orders against CPU0
2439 * UNLOCK rq(0)->lock
2443 * UNLOCK rq(1)->lock
2445 * LOCK rq(1)->lock // orders against CPU2
2448 * UNLOCK rq(1)->lock
2451 * BLOCKING -- aka. SLEEP + WAKEUP
2453 * For blocking we (obviously) need to provide the same guarantee as for
2454 * migration. However the means are completely different as there is no lock
2455 * chain to provide order. Instead we do:
2457 * 1) smp_store_release(X->on_cpu, 0)
2458 * 2) smp_cond_load_acquire(!X->on_cpu)
2462 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2464 * LOCK rq(0)->lock LOCK X->pi_lock
2467 * smp_store_release(X->on_cpu, 0);
2469 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2475 * X->state = RUNNING
2476 * UNLOCK rq(2)->lock
2478 * LOCK rq(2)->lock // orders against CPU1
2481 * UNLOCK rq(2)->lock
2484 * UNLOCK rq(0)->lock
2487 * However, for wakeups there is a second guarantee we must provide, namely we
2488 * must ensure that CONDITION=1 done by the caller can not be reordered with
2489 * accesses to the task state; see try_to_wake_up() and set_current_state().
2493 * try_to_wake_up - wake up a thread
2494 * @p: the thread to be awakened
2495 * @state: the mask of task states that can be woken
2496 * @wake_flags: wake modifier flags (WF_*)
2498 * If (@state & @p->state) @p->state = TASK_RUNNING.
2500 * If the task was not queued/runnable, also place it back on a runqueue.
2502 * Atomic against schedule() which would dequeue a task, also see
2503 * set_current_state().
2505 * This function executes a full memory barrier before accessing the task
2506 * state; see set_current_state().
2508 * Return: %true if @p->state changes (an actual wakeup was done),
2512 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2514 unsigned long flags;
2515 int cpu, success = 0;
2520 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2521 * == smp_processor_id()'. Together this means we can special
2522 * case the whole 'p->on_rq && ttwu_remote()' case below
2523 * without taking any locks.
2526 * - we rely on Program-Order guarantees for all the ordering,
2527 * - we're serialized against set_special_state() by virtue of
2528 * it disabling IRQs (this allows not taking ->pi_lock).
2530 if (!(p->state & state))
2535 trace_sched_waking(p);
2536 p->state = TASK_RUNNING;
2537 trace_sched_wakeup(p);
2542 * If we are going to wake up a thread waiting for CONDITION we
2543 * need to ensure that CONDITION=1 done by the caller can not be
2544 * reordered with p->state check below. This pairs with mb() in
2545 * set_current_state() the waiting thread does.
2547 raw_spin_lock_irqsave(&p->pi_lock, flags);
2548 smp_mb__after_spinlock();
2549 if (!(p->state & state))
2552 trace_sched_waking(p);
2554 /* We're going to change ->state: */
2559 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2560 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2561 * in smp_cond_load_acquire() below.
2563 * sched_ttwu_pending() try_to_wake_up()
2564 * STORE p->on_rq = 1 LOAD p->state
2567 * __schedule() (switch to task 'p')
2568 * LOCK rq->lock smp_rmb();
2569 * smp_mb__after_spinlock();
2573 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2575 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2576 * __schedule(). See the comment for smp_mb__after_spinlock().
2579 if (p->on_rq && ttwu_remote(p, wake_flags))
2584 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2585 * possible to, falsely, observe p->on_cpu == 0.
2587 * One must be running (->on_cpu == 1) in order to remove oneself
2588 * from the runqueue.
2590 * __schedule() (switch to task 'p') try_to_wake_up()
2591 * STORE p->on_cpu = 1 LOAD p->on_rq
2594 * __schedule() (put 'p' to sleep)
2595 * LOCK rq->lock smp_rmb();
2596 * smp_mb__after_spinlock();
2597 * STORE p->on_rq = 0 LOAD p->on_cpu
2599 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2600 * __schedule(). See the comment for smp_mb__after_spinlock().
2605 * If the owning (remote) CPU is still in the middle of schedule() with
2606 * this task as prev, wait until its done referencing the task.
2608 * Pairs with the smp_store_release() in finish_task().
2610 * This ensures that tasks getting woken will be fully ordered against
2611 * their previous state and preserve Program Order.
2613 smp_cond_load_acquire(&p->on_cpu, !VAL);
2615 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2616 p->state = TASK_WAKING;
2619 delayacct_blkio_end(p);
2620 atomic_dec(&task_rq(p)->nr_iowait);
2623 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2624 if (task_cpu(p) != cpu) {
2625 wake_flags |= WF_MIGRATED;
2626 psi_ttwu_dequeue(p);
2627 set_task_cpu(p, cpu);
2630 #else /* CONFIG_SMP */
2633 delayacct_blkio_end(p);
2634 atomic_dec(&task_rq(p)->nr_iowait);
2637 #endif /* CONFIG_SMP */
2639 ttwu_queue(p, cpu, wake_flags);
2641 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2644 ttwu_stat(p, cpu, wake_flags);
2651 * wake_up_process - Wake up a specific process
2652 * @p: The process to be woken up.
2654 * Attempt to wake up the nominated process and move it to the set of runnable
2657 * Return: 1 if the process was woken up, 0 if it was already running.
2659 * This function executes a full memory barrier before accessing the task state.
2661 int wake_up_process(struct task_struct *p)
2663 return try_to_wake_up(p, TASK_NORMAL, 0);
2665 EXPORT_SYMBOL(wake_up_process);
2667 int wake_up_state(struct task_struct *p, unsigned int state)
2669 return try_to_wake_up(p, state, 0);
2673 * Perform scheduler related setup for a newly forked process p.
2674 * p is forked by current.
2676 * __sched_fork() is basic setup used by init_idle() too:
2678 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2683 p->se.exec_start = 0;
2684 p->se.sum_exec_runtime = 0;
2685 p->se.prev_sum_exec_runtime = 0;
2686 p->se.nr_migrations = 0;
2688 INIT_LIST_HEAD(&p->se.group_node);
2690 #ifdef CONFIG_FAIR_GROUP_SCHED
2691 p->se.cfs_rq = NULL;
2694 #ifdef CONFIG_SCHEDSTATS
2695 /* Even if schedstat is disabled, there should not be garbage */
2696 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2699 RB_CLEAR_NODE(&p->dl.rb_node);
2700 init_dl_task_timer(&p->dl);
2701 init_dl_inactive_task_timer(&p->dl);
2702 __dl_clear_params(p);
2704 INIT_LIST_HEAD(&p->rt.run_list);
2706 p->rt.time_slice = sched_rr_timeslice;
2710 #ifdef CONFIG_PREEMPT_NOTIFIERS
2711 INIT_HLIST_HEAD(&p->preempt_notifiers);
2714 #ifdef CONFIG_COMPACTION
2715 p->capture_control = NULL;
2717 init_numa_balancing(clone_flags, p);
2720 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2722 #ifdef CONFIG_NUMA_BALANCING
2724 void set_numabalancing_state(bool enabled)
2727 static_branch_enable(&sched_numa_balancing);
2729 static_branch_disable(&sched_numa_balancing);
2732 #ifdef CONFIG_PROC_SYSCTL
2733 int sysctl_numa_balancing(struct ctl_table *table, int write,
2734 void __user *buffer, size_t *lenp, loff_t *ppos)
2738 int state = static_branch_likely(&sched_numa_balancing);
2740 if (write && !capable(CAP_SYS_ADMIN))
2745 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2749 set_numabalancing_state(state);
2755 #ifdef CONFIG_SCHEDSTATS
2757 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2758 static bool __initdata __sched_schedstats = false;
2760 static void set_schedstats(bool enabled)
2763 static_branch_enable(&sched_schedstats);
2765 static_branch_disable(&sched_schedstats);
2768 void force_schedstat_enabled(void)
2770 if (!schedstat_enabled()) {
2771 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2772 static_branch_enable(&sched_schedstats);
2776 static int __init setup_schedstats(char *str)
2783 * This code is called before jump labels have been set up, so we can't
2784 * change the static branch directly just yet. Instead set a temporary
2785 * variable so init_schedstats() can do it later.
2787 if (!strcmp(str, "enable")) {
2788 __sched_schedstats = true;
2790 } else if (!strcmp(str, "disable")) {
2791 __sched_schedstats = false;
2796 pr_warn("Unable to parse schedstats=\n");
2800 __setup("schedstats=", setup_schedstats);
2802 static void __init init_schedstats(void)
2804 set_schedstats(__sched_schedstats);
2807 #ifdef CONFIG_PROC_SYSCTL
2808 int sysctl_schedstats(struct ctl_table *table, int write,
2809 void __user *buffer, size_t *lenp, loff_t *ppos)
2813 int state = static_branch_likely(&sched_schedstats);
2815 if (write && !capable(CAP_SYS_ADMIN))
2820 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2824 set_schedstats(state);
2827 #endif /* CONFIG_PROC_SYSCTL */
2828 #else /* !CONFIG_SCHEDSTATS */
2829 static inline void init_schedstats(void) {}
2830 #endif /* CONFIG_SCHEDSTATS */
2833 * fork()/clone()-time setup:
2835 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2837 unsigned long flags;
2839 __sched_fork(clone_flags, p);
2841 * We mark the process as NEW here. This guarantees that
2842 * nobody will actually run it, and a signal or other external
2843 * event cannot wake it up and insert it on the runqueue either.
2845 p->state = TASK_NEW;
2848 * Make sure we do not leak PI boosting priority to the child.
2850 p->prio = current->normal_prio;
2855 * Revert to default priority/policy on fork if requested.
2857 if (unlikely(p->sched_reset_on_fork)) {
2858 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2859 p->policy = SCHED_NORMAL;
2860 p->static_prio = NICE_TO_PRIO(0);
2862 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2863 p->static_prio = NICE_TO_PRIO(0);
2865 p->prio = p->normal_prio = __normal_prio(p);
2866 set_load_weight(p, false);
2869 * We don't need the reset flag anymore after the fork. It has
2870 * fulfilled its duty:
2872 p->sched_reset_on_fork = 0;
2875 if (dl_prio(p->prio))
2877 else if (rt_prio(p->prio))
2878 p->sched_class = &rt_sched_class;
2880 p->sched_class = &fair_sched_class;
2882 init_entity_runnable_average(&p->se);
2885 * The child is not yet in the pid-hash so no cgroup attach races,
2886 * and the cgroup is pinned to this child due to cgroup_fork()
2887 * is ran before sched_fork().
2889 * Silence PROVE_RCU.
2891 raw_spin_lock_irqsave(&p->pi_lock, flags);
2893 * We're setting the CPU for the first time, we don't migrate,
2894 * so use __set_task_cpu().
2896 __set_task_cpu(p, smp_processor_id());
2897 if (p->sched_class->task_fork)
2898 p->sched_class->task_fork(p);
2899 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2901 #ifdef CONFIG_SCHED_INFO
2902 if (likely(sched_info_on()))
2903 memset(&p->sched_info, 0, sizeof(p->sched_info));
2905 #if defined(CONFIG_SMP)
2908 init_task_preempt_count(p);
2910 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2911 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2916 unsigned long to_ratio(u64 period, u64 runtime)
2918 if (runtime == RUNTIME_INF)
2922 * Doing this here saves a lot of checks in all
2923 * the calling paths, and returning zero seems
2924 * safe for them anyway.
2929 return div64_u64(runtime << BW_SHIFT, period);
2933 * wake_up_new_task - wake up a newly created task for the first time.
2935 * This function will do some initial scheduler statistics housekeeping
2936 * that must be done for every newly created context, then puts the task
2937 * on the runqueue and wakes it.
2939 void wake_up_new_task(struct task_struct *p)
2944 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2945 p->state = TASK_RUNNING;
2948 * Fork balancing, do it here and not earlier because:
2949 * - cpus_ptr can change in the fork path
2950 * - any previously selected CPU might disappear through hotplug
2952 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2953 * as we're not fully set-up yet.
2955 p->recent_used_cpu = task_cpu(p);
2956 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2958 rq = __task_rq_lock(p, &rf);
2959 update_rq_clock(rq);
2960 post_init_entity_util_avg(p);
2962 activate_task(rq, p, ENQUEUE_NOCLOCK);
2963 trace_sched_wakeup_new(p);
2964 check_preempt_curr(rq, p, WF_FORK);
2966 if (p->sched_class->task_woken) {
2968 * Nothing relies on rq->lock after this, so its fine to
2971 rq_unpin_lock(rq, &rf);
2972 p->sched_class->task_woken(rq, p);
2973 rq_repin_lock(rq, &rf);
2976 task_rq_unlock(rq, p, &rf);
2979 #ifdef CONFIG_PREEMPT_NOTIFIERS
2981 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2983 void preempt_notifier_inc(void)
2985 static_branch_inc(&preempt_notifier_key);
2987 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2989 void preempt_notifier_dec(void)
2991 static_branch_dec(&preempt_notifier_key);
2993 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2996 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2997 * @notifier: notifier struct to register
2999 void preempt_notifier_register(struct preempt_notifier *notifier)
3001 if (!static_branch_unlikely(&preempt_notifier_key))
3002 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3004 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3006 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3009 * preempt_notifier_unregister - no longer interested in preemption notifications
3010 * @notifier: notifier struct to unregister
3012 * This is *not* safe to call from within a preemption notifier.
3014 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3016 hlist_del(¬ifier->link);
3018 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3020 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3022 struct preempt_notifier *notifier;
3024 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3025 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3028 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3030 if (static_branch_unlikely(&preempt_notifier_key))
3031 __fire_sched_in_preempt_notifiers(curr);
3035 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3036 struct task_struct *next)
3038 struct preempt_notifier *notifier;
3040 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3041 notifier->ops->sched_out(notifier, next);
3044 static __always_inline void
3045 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3046 struct task_struct *next)
3048 if (static_branch_unlikely(&preempt_notifier_key))
3049 __fire_sched_out_preempt_notifiers(curr, next);
3052 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3054 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3059 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3060 struct task_struct *next)
3064 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3066 static inline void prepare_task(struct task_struct *next)
3070 * Claim the task as running, we do this before switching to it
3071 * such that any running task will have this set.
3077 static inline void finish_task(struct task_struct *prev)
3081 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3082 * We must ensure this doesn't happen until the switch is completely
3085 * In particular, the load of prev->state in finish_task_switch() must
3086 * happen before this.
3088 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3090 smp_store_release(&prev->on_cpu, 0);
3095 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3098 * Since the runqueue lock will be released by the next
3099 * task (which is an invalid locking op but in the case
3100 * of the scheduler it's an obvious special-case), so we
3101 * do an early lockdep release here:
3103 rq_unpin_lock(rq, rf);
3104 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3105 #ifdef CONFIG_DEBUG_SPINLOCK
3106 /* this is a valid case when another task releases the spinlock */
3107 rq->lock.owner = next;
3111 static inline void finish_lock_switch(struct rq *rq)
3114 * If we are tracking spinlock dependencies then we have to
3115 * fix up the runqueue lock - which gets 'carried over' from
3116 * prev into current:
3118 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3119 raw_spin_unlock_irq(&rq->lock);
3123 * NOP if the arch has not defined these:
3126 #ifndef prepare_arch_switch
3127 # define prepare_arch_switch(next) do { } while (0)
3130 #ifndef finish_arch_post_lock_switch
3131 # define finish_arch_post_lock_switch() do { } while (0)
3135 * prepare_task_switch - prepare to switch tasks
3136 * @rq: the runqueue preparing to switch
3137 * @prev: the current task that is being switched out
3138 * @next: the task we are going to switch to.
3140 * This is called with the rq lock held and interrupts off. It must
3141 * be paired with a subsequent finish_task_switch after the context
3144 * prepare_task_switch sets up locking and calls architecture specific
3148 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3149 struct task_struct *next)
3151 kcov_prepare_switch(prev);
3152 sched_info_switch(rq, prev, next);
3153 perf_event_task_sched_out(prev, next);
3155 fire_sched_out_preempt_notifiers(prev, next);
3157 prepare_arch_switch(next);
3161 * finish_task_switch - clean up after a task-switch
3162 * @prev: the thread we just switched away from.
3164 * finish_task_switch must be called after the context switch, paired
3165 * with a prepare_task_switch call before the context switch.
3166 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3167 * and do any other architecture-specific cleanup actions.
3169 * Note that we may have delayed dropping an mm in context_switch(). If
3170 * so, we finish that here outside of the runqueue lock. (Doing it
3171 * with the lock held can cause deadlocks; see schedule() for
3174 * The context switch have flipped the stack from under us and restored the
3175 * local variables which were saved when this task called schedule() in the
3176 * past. prev == current is still correct but we need to recalculate this_rq
3177 * because prev may have moved to another CPU.
3179 static struct rq *finish_task_switch(struct task_struct *prev)
3180 __releases(rq->lock)
3182 struct rq *rq = this_rq();
3183 struct mm_struct *mm = rq->prev_mm;
3187 * The previous task will have left us with a preempt_count of 2
3188 * because it left us after:
3191 * preempt_disable(); // 1
3193 * raw_spin_lock_irq(&rq->lock) // 2
3195 * Also, see FORK_PREEMPT_COUNT.
3197 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3198 "corrupted preempt_count: %s/%d/0x%x\n",
3199 current->comm, current->pid, preempt_count()))
3200 preempt_count_set(FORK_PREEMPT_COUNT);
3205 * A task struct has one reference for the use as "current".
3206 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3207 * schedule one last time. The schedule call will never return, and
3208 * the scheduled task must drop that reference.
3210 * We must observe prev->state before clearing prev->on_cpu (in
3211 * finish_task), otherwise a concurrent wakeup can get prev
3212 * running on another CPU and we could rave with its RUNNING -> DEAD
3213 * transition, resulting in a double drop.
3215 prev_state = prev->state;
3216 vtime_task_switch(prev);
3217 perf_event_task_sched_in(prev, current);
3219 finish_lock_switch(rq);
3220 finish_arch_post_lock_switch();
3221 kcov_finish_switch(current);
3223 fire_sched_in_preempt_notifiers(current);
3225 * When switching through a kernel thread, the loop in
3226 * membarrier_{private,global}_expedited() may have observed that
3227 * kernel thread and not issued an IPI. It is therefore possible to
3228 * schedule between user->kernel->user threads without passing though
3229 * switch_mm(). Membarrier requires a barrier after storing to
3230 * rq->curr, before returning to userspace, so provide them here:
3232 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3233 * provided by mmdrop(),
3234 * - a sync_core for SYNC_CORE.
3237 membarrier_mm_sync_core_before_usermode(mm);
3240 if (unlikely(prev_state == TASK_DEAD)) {
3241 if (prev->sched_class->task_dead)
3242 prev->sched_class->task_dead(prev);
3245 * Remove function-return probe instances associated with this
3246 * task and put them back on the free list.
3248 kprobe_flush_task(prev);
3250 /* Task is done with its stack. */
3251 put_task_stack(prev);
3253 put_task_struct_rcu_user(prev);
3256 tick_nohz_task_switch();
3262 /* rq->lock is NOT held, but preemption is disabled */
3263 static void __balance_callback(struct rq *rq)
3265 struct callback_head *head, *next;
3266 void (*func)(struct rq *rq);
3267 unsigned long flags;
3269 raw_spin_lock_irqsave(&rq->lock, flags);
3270 head = rq->balance_callback;
3271 rq->balance_callback = NULL;
3273 func = (void (*)(struct rq *))head->func;
3280 raw_spin_unlock_irqrestore(&rq->lock, flags);
3283 static inline void balance_callback(struct rq *rq)
3285 if (unlikely(rq->balance_callback))
3286 __balance_callback(rq);
3291 static inline void balance_callback(struct rq *rq)
3298 * schedule_tail - first thing a freshly forked thread must call.
3299 * @prev: the thread we just switched away from.
3301 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3302 __releases(rq->lock)
3307 * New tasks start with FORK_PREEMPT_COUNT, see there and
3308 * finish_task_switch() for details.
3310 * finish_task_switch() will drop rq->lock() and lower preempt_count
3311 * and the preempt_enable() will end up enabling preemption (on
3312 * PREEMPT_COUNT kernels).
3315 rq = finish_task_switch(prev);
3316 balance_callback(rq);
3319 if (current->set_child_tid)
3320 put_user(task_pid_vnr(current), current->set_child_tid);
3322 calculate_sigpending();
3326 * context_switch - switch to the new MM and the new thread's register state.
3328 static __always_inline struct rq *
3329 context_switch(struct rq *rq, struct task_struct *prev,
3330 struct task_struct *next, struct rq_flags *rf)
3332 prepare_task_switch(rq, prev, next);
3335 * For paravirt, this is coupled with an exit in switch_to to
3336 * combine the page table reload and the switch backend into
3339 arch_start_context_switch(prev);
3342 * kernel -> kernel lazy + transfer active
3343 * user -> kernel lazy + mmgrab() active
3345 * kernel -> user switch + mmdrop() active
3346 * user -> user switch
3348 if (!next->mm) { // to kernel
3349 enter_lazy_tlb(prev->active_mm, next);
3351 next->active_mm = prev->active_mm;
3352 if (prev->mm) // from user
3353 mmgrab(prev->active_mm);
3355 prev->active_mm = NULL;
3357 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3359 * sys_membarrier() requires an smp_mb() between setting
3360 * rq->curr / membarrier_switch_mm() and returning to userspace.
3362 * The below provides this either through switch_mm(), or in
3363 * case 'prev->active_mm == next->mm' through
3364 * finish_task_switch()'s mmdrop().
3366 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3368 if (!prev->mm) { // from kernel
3369 /* will mmdrop() in finish_task_switch(). */
3370 rq->prev_mm = prev->active_mm;
3371 prev->active_mm = NULL;
3375 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3377 prepare_lock_switch(rq, next, rf);
3379 /* Here we just switch the register state and the stack. */
3380 switch_to(prev, next, prev);
3383 return finish_task_switch(prev);
3387 * nr_running and nr_context_switches:
3389 * externally visible scheduler statistics: current number of runnable
3390 * threads, total number of context switches performed since bootup.
3392 unsigned long nr_running(void)
3394 unsigned long i, sum = 0;
3396 for_each_online_cpu(i)
3397 sum += cpu_rq(i)->nr_running;
3403 * Check if only the current task is running on the CPU.
3405 * Caution: this function does not check that the caller has disabled
3406 * preemption, thus the result might have a time-of-check-to-time-of-use
3407 * race. The caller is responsible to use it correctly, for example:
3409 * - from a non-preemptible section (of course)
3411 * - from a thread that is bound to a single CPU
3413 * - in a loop with very short iterations (e.g. a polling loop)
3415 bool single_task_running(void)
3417 return raw_rq()->nr_running == 1;
3419 EXPORT_SYMBOL(single_task_running);
3421 unsigned long long nr_context_switches(void)
3424 unsigned long long sum = 0;
3426 for_each_possible_cpu(i)
3427 sum += cpu_rq(i)->nr_switches;
3433 * Consumers of these two interfaces, like for example the cpuidle menu
3434 * governor, are using nonsensical data. Preferring shallow idle state selection
3435 * for a CPU that has IO-wait which might not even end up running the task when
3436 * it does become runnable.
3439 unsigned long nr_iowait_cpu(int cpu)
3441 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3445 * IO-wait accounting, and how its mostly bollocks (on SMP).
3447 * The idea behind IO-wait account is to account the idle time that we could
3448 * have spend running if it were not for IO. That is, if we were to improve the
3449 * storage performance, we'd have a proportional reduction in IO-wait time.
3451 * This all works nicely on UP, where, when a task blocks on IO, we account
3452 * idle time as IO-wait, because if the storage were faster, it could've been
3453 * running and we'd not be idle.
3455 * This has been extended to SMP, by doing the same for each CPU. This however
3458 * Imagine for instance the case where two tasks block on one CPU, only the one
3459 * CPU will have IO-wait accounted, while the other has regular idle. Even
3460 * though, if the storage were faster, both could've ran at the same time,
3461 * utilising both CPUs.
3463 * This means, that when looking globally, the current IO-wait accounting on
3464 * SMP is a lower bound, by reason of under accounting.
3466 * Worse, since the numbers are provided per CPU, they are sometimes
3467 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3468 * associated with any one particular CPU, it can wake to another CPU than it
3469 * blocked on. This means the per CPU IO-wait number is meaningless.
3471 * Task CPU affinities can make all that even more 'interesting'.
3474 unsigned long nr_iowait(void)
3476 unsigned long i, sum = 0;
3478 for_each_possible_cpu(i)
3479 sum += nr_iowait_cpu(i);
3487 * sched_exec - execve() is a valuable balancing opportunity, because at
3488 * this point the task has the smallest effective memory and cache footprint.
3490 void sched_exec(void)
3492 struct task_struct *p = current;
3493 unsigned long flags;
3496 raw_spin_lock_irqsave(&p->pi_lock, flags);
3497 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3498 if (dest_cpu == smp_processor_id())
3501 if (likely(cpu_active(dest_cpu))) {
3502 struct migration_arg arg = { p, dest_cpu };
3504 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3505 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3509 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3514 DEFINE_PER_CPU(struct kernel_stat, kstat);
3515 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3517 EXPORT_PER_CPU_SYMBOL(kstat);
3518 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3521 * The function fair_sched_class.update_curr accesses the struct curr
3522 * and its field curr->exec_start; when called from task_sched_runtime(),
3523 * we observe a high rate of cache misses in practice.
3524 * Prefetching this data results in improved performance.
3526 static inline void prefetch_curr_exec_start(struct task_struct *p)
3528 #ifdef CONFIG_FAIR_GROUP_SCHED
3529 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3531 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3534 prefetch(&curr->exec_start);
3538 * Return accounted runtime for the task.
3539 * In case the task is currently running, return the runtime plus current's
3540 * pending runtime that have not been accounted yet.
3542 unsigned long long task_sched_runtime(struct task_struct *p)
3548 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3550 * 64-bit doesn't need locks to atomically read a 64-bit value.
3551 * So we have a optimization chance when the task's delta_exec is 0.
3552 * Reading ->on_cpu is racy, but this is ok.
3554 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3555 * If we race with it entering CPU, unaccounted time is 0. This is
3556 * indistinguishable from the read occurring a few cycles earlier.
3557 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3558 * been accounted, so we're correct here as well.
3560 if (!p->on_cpu || !task_on_rq_queued(p))
3561 return p->se.sum_exec_runtime;
3564 rq = task_rq_lock(p, &rf);
3566 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3567 * project cycles that may never be accounted to this
3568 * thread, breaking clock_gettime().
3570 if (task_current(rq, p) && task_on_rq_queued(p)) {
3571 prefetch_curr_exec_start(p);
3572 update_rq_clock(rq);
3573 p->sched_class->update_curr(rq);
3575 ns = p->se.sum_exec_runtime;
3576 task_rq_unlock(rq, p, &rf);
3582 * This function gets called by the timer code, with HZ frequency.
3583 * We call it with interrupts disabled.
3585 void scheduler_tick(void)
3587 int cpu = smp_processor_id();
3588 struct rq *rq = cpu_rq(cpu);
3589 struct task_struct *curr = rq->curr;
3596 update_rq_clock(rq);
3597 curr->sched_class->task_tick(rq, curr, 0);
3598 calc_global_load_tick(rq);
3603 perf_event_task_tick();
3606 rq->idle_balance = idle_cpu(cpu);
3607 trigger_load_balance(rq);
3611 #ifdef CONFIG_NO_HZ_FULL
3616 struct delayed_work work;
3618 /* Values for ->state, see diagram below. */
3619 #define TICK_SCHED_REMOTE_OFFLINE 0
3620 #define TICK_SCHED_REMOTE_OFFLINING 1
3621 #define TICK_SCHED_REMOTE_RUNNING 2
3624 * State diagram for ->state:
3627 * TICK_SCHED_REMOTE_OFFLINE
3630 * | | sched_tick_remote()
3633 * +--TICK_SCHED_REMOTE_OFFLINING
3636 * sched_tick_start() | | sched_tick_stop()
3639 * TICK_SCHED_REMOTE_RUNNING
3642 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3643 * and sched_tick_start() are happy to leave the state in RUNNING.
3646 static struct tick_work __percpu *tick_work_cpu;
3648 static void sched_tick_remote(struct work_struct *work)
3650 struct delayed_work *dwork = to_delayed_work(work);
3651 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3652 int cpu = twork->cpu;
3653 struct rq *rq = cpu_rq(cpu);
3654 struct task_struct *curr;
3660 * Handle the tick only if it appears the remote CPU is running in full
3661 * dynticks mode. The check is racy by nature, but missing a tick or
3662 * having one too much is no big deal because the scheduler tick updates
3663 * statistics and checks timeslices in a time-independent way, regardless
3664 * of when exactly it is running.
3666 if (!tick_nohz_tick_stopped_cpu(cpu))
3669 rq_lock_irq(rq, &rf);
3671 if (cpu_is_offline(cpu))
3674 update_rq_clock(rq);
3676 if (!is_idle_task(curr)) {
3678 * Make sure the next tick runs within a reasonable
3681 delta = rq_clock_task(rq) - curr->se.exec_start;
3682 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3684 curr->sched_class->task_tick(rq, curr, 0);
3686 calc_load_nohz_remote(rq);
3688 rq_unlock_irq(rq, &rf);
3692 * Run the remote tick once per second (1Hz). This arbitrary
3693 * frequency is large enough to avoid overload but short enough
3694 * to keep scheduler internal stats reasonably up to date. But
3695 * first update state to reflect hotplug activity if required.
3697 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3698 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3699 if (os == TICK_SCHED_REMOTE_RUNNING)
3700 queue_delayed_work(system_unbound_wq, dwork, HZ);
3703 static void sched_tick_start(int cpu)
3706 struct tick_work *twork;
3708 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3711 WARN_ON_ONCE(!tick_work_cpu);
3713 twork = per_cpu_ptr(tick_work_cpu, cpu);
3714 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3715 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3716 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3718 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3719 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3723 #ifdef CONFIG_HOTPLUG_CPU
3724 static void sched_tick_stop(int cpu)
3726 struct tick_work *twork;
3729 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3732 WARN_ON_ONCE(!tick_work_cpu);
3734 twork = per_cpu_ptr(tick_work_cpu, cpu);
3735 /* There cannot be competing actions, but don't rely on stop-machine. */
3736 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3737 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3738 /* Don't cancel, as this would mess up the state machine. */
3740 #endif /* CONFIG_HOTPLUG_CPU */
3742 int __init sched_tick_offload_init(void)
3744 tick_work_cpu = alloc_percpu(struct tick_work);
3745 BUG_ON(!tick_work_cpu);
3749 #else /* !CONFIG_NO_HZ_FULL */
3750 static inline void sched_tick_start(int cpu) { }
3751 static inline void sched_tick_stop(int cpu) { }
3754 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3755 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3757 * If the value passed in is equal to the current preempt count
3758 * then we just disabled preemption. Start timing the latency.
3760 static inline void preempt_latency_start(int val)
3762 if (preempt_count() == val) {
3763 unsigned long ip = get_lock_parent_ip();
3764 #ifdef CONFIG_DEBUG_PREEMPT
3765 current->preempt_disable_ip = ip;
3767 trace_preempt_off(CALLER_ADDR0, ip);
3771 void preempt_count_add(int val)
3773 #ifdef CONFIG_DEBUG_PREEMPT
3777 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3780 __preempt_count_add(val);
3781 #ifdef CONFIG_DEBUG_PREEMPT
3783 * Spinlock count overflowing soon?
3785 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3788 preempt_latency_start(val);
3790 EXPORT_SYMBOL(preempt_count_add);
3791 NOKPROBE_SYMBOL(preempt_count_add);
3794 * If the value passed in equals to the current preempt count
3795 * then we just enabled preemption. Stop timing the latency.
3797 static inline void preempt_latency_stop(int val)
3799 if (preempt_count() == val)
3800 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3803 void preempt_count_sub(int val)
3805 #ifdef CONFIG_DEBUG_PREEMPT
3809 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3812 * Is the spinlock portion underflowing?
3814 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3815 !(preempt_count() & PREEMPT_MASK)))
3819 preempt_latency_stop(val);
3820 __preempt_count_sub(val);
3822 EXPORT_SYMBOL(preempt_count_sub);
3823 NOKPROBE_SYMBOL(preempt_count_sub);
3826 static inline void preempt_latency_start(int val) { }
3827 static inline void preempt_latency_stop(int val) { }
3830 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3832 #ifdef CONFIG_DEBUG_PREEMPT
3833 return p->preempt_disable_ip;
3840 * Print scheduling while atomic bug:
3842 static noinline void __schedule_bug(struct task_struct *prev)
3844 /* Save this before calling printk(), since that will clobber it */
3845 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3847 if (oops_in_progress)
3850 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3851 prev->comm, prev->pid, preempt_count());
3853 debug_show_held_locks(prev);
3855 if (irqs_disabled())
3856 print_irqtrace_events(prev);
3857 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3858 && in_atomic_preempt_off()) {
3859 pr_err("Preemption disabled at:");
3860 print_ip_sym(preempt_disable_ip);
3864 panic("scheduling while atomic\n");
3867 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3871 * Various schedule()-time debugging checks and statistics:
3873 static inline void schedule_debug(struct task_struct *prev, bool preempt)
3875 #ifdef CONFIG_SCHED_STACK_END_CHECK
3876 if (task_stack_end_corrupted(prev))
3877 panic("corrupted stack end detected inside scheduler\n");
3880 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3881 if (!preempt && prev->state && prev->non_block_count) {
3882 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3883 prev->comm, prev->pid, prev->non_block_count);
3885 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3889 if (unlikely(in_atomic_preempt_off())) {
3890 __schedule_bug(prev);
3891 preempt_count_set(PREEMPT_DISABLED);
3895 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3897 schedstat_inc(this_rq()->sched_count);
3901 * Pick up the highest-prio task:
3903 static inline struct task_struct *
3904 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3906 const struct sched_class *class;
3907 struct task_struct *p;
3910 * Optimization: we know that if all tasks are in the fair class we can
3911 * call that function directly, but only if the @prev task wasn't of a
3912 * higher scheduling class, because otherwise those loose the
3913 * opportunity to pull in more work from other CPUs.
3915 if (likely((prev->sched_class == &idle_sched_class ||
3916 prev->sched_class == &fair_sched_class) &&
3917 rq->nr_running == rq->cfs.h_nr_running)) {
3919 p = fair_sched_class.pick_next_task(rq, prev, rf);
3920 if (unlikely(p == RETRY_TASK))
3923 /* Assumes fair_sched_class->next == idle_sched_class */
3925 p = idle_sched_class.pick_next_task(rq, prev, rf);
3933 * We must do the balancing pass before put_next_task(), such
3934 * that when we release the rq->lock the task is in the same
3935 * state as before we took rq->lock.
3937 * We can terminate the balance pass as soon as we know there is
3938 * a runnable task of @class priority or higher.
3940 for_class_range(class, prev->sched_class, &idle_sched_class) {
3941 if (class->balance(rq, prev, rf))
3946 put_prev_task(rq, prev);
3948 for_each_class(class) {
3949 p = class->pick_next_task(rq, NULL, NULL);
3954 /* The idle class should always have a runnable task: */
3959 * __schedule() is the main scheduler function.
3961 * The main means of driving the scheduler and thus entering this function are:
3963 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3965 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3966 * paths. For example, see arch/x86/entry_64.S.
3968 * To drive preemption between tasks, the scheduler sets the flag in timer
3969 * interrupt handler scheduler_tick().
3971 * 3. Wakeups don't really cause entry into schedule(). They add a
3972 * task to the run-queue and that's it.
3974 * Now, if the new task added to the run-queue preempts the current
3975 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3976 * called on the nearest possible occasion:
3978 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
3980 * - in syscall or exception context, at the next outmost
3981 * preempt_enable(). (this might be as soon as the wake_up()'s
3984 * - in IRQ context, return from interrupt-handler to
3985 * preemptible context
3987 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
3990 * - cond_resched() call
3991 * - explicit schedule() call
3992 * - return from syscall or exception to user-space
3993 * - return from interrupt-handler to user-space
3995 * WARNING: must be called with preemption disabled!
3997 static void __sched notrace __schedule(bool preempt)
3999 struct task_struct *prev, *next;
4000 unsigned long *switch_count;
4005 cpu = smp_processor_id();
4009 schedule_debug(prev, preempt);
4011 if (sched_feat(HRTICK))
4014 local_irq_disable();
4015 rcu_note_context_switch(preempt);
4018 * Make sure that signal_pending_state()->signal_pending() below
4019 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4020 * done by the caller to avoid the race with signal_wake_up().
4022 * The membarrier system call requires a full memory barrier
4023 * after coming from user-space, before storing to rq->curr.
4026 smp_mb__after_spinlock();
4028 /* Promote REQ to ACT */
4029 rq->clock_update_flags <<= 1;
4030 update_rq_clock(rq);
4032 switch_count = &prev->nivcsw;
4033 if (!preempt && prev->state) {
4034 if (signal_pending_state(prev->state, prev)) {
4035 prev->state = TASK_RUNNING;
4037 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4039 if (prev->in_iowait) {
4040 atomic_inc(&rq->nr_iowait);
4041 delayacct_blkio_start();
4044 switch_count = &prev->nvcsw;
4047 next = pick_next_task(rq, prev, &rf);
4048 clear_tsk_need_resched(prev);
4049 clear_preempt_need_resched();
4051 if (likely(prev != next)) {
4054 * RCU users of rcu_dereference(rq->curr) may not see
4055 * changes to task_struct made by pick_next_task().
4057 RCU_INIT_POINTER(rq->curr, next);
4059 * The membarrier system call requires each architecture
4060 * to have a full memory barrier after updating
4061 * rq->curr, before returning to user-space.
4063 * Here are the schemes providing that barrier on the
4064 * various architectures:
4065 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4066 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4067 * - finish_lock_switch() for weakly-ordered
4068 * architectures where spin_unlock is a full barrier,
4069 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4070 * is a RELEASE barrier),
4074 trace_sched_switch(preempt, prev, next);
4076 /* Also unlocks the rq: */
4077 rq = context_switch(rq, prev, next, &rf);
4079 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4080 rq_unlock_irq(rq, &rf);
4083 balance_callback(rq);
4086 void __noreturn do_task_dead(void)
4088 /* Causes final put_task_struct in finish_task_switch(): */
4089 set_special_state(TASK_DEAD);
4091 /* Tell freezer to ignore us: */
4092 current->flags |= PF_NOFREEZE;
4097 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4102 static inline void sched_submit_work(struct task_struct *tsk)
4108 * If a worker went to sleep, notify and ask workqueue whether
4109 * it wants to wake up a task to maintain concurrency.
4110 * As this function is called inside the schedule() context,
4111 * we disable preemption to avoid it calling schedule() again
4112 * in the possible wakeup of a kworker.
4114 if (tsk->flags & PF_WQ_WORKER) {
4116 wq_worker_sleeping(tsk);
4117 preempt_enable_no_resched();
4120 if (tsk_is_pi_blocked(tsk))
4124 * If we are going to sleep and we have plugged IO queued,
4125 * make sure to submit it to avoid deadlocks.
4127 if (blk_needs_flush_plug(tsk))
4128 blk_schedule_flush_plug(tsk);
4131 static void sched_update_worker(struct task_struct *tsk)
4133 if (tsk->flags & PF_WQ_WORKER)
4134 wq_worker_running(tsk);
4137 asmlinkage __visible void __sched schedule(void)
4139 struct task_struct *tsk = current;
4141 sched_submit_work(tsk);
4145 sched_preempt_enable_no_resched();
4146 } while (need_resched());
4147 sched_update_worker(tsk);
4149 EXPORT_SYMBOL(schedule);
4152 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4153 * state (have scheduled out non-voluntarily) by making sure that all
4154 * tasks have either left the run queue or have gone into user space.
4155 * As idle tasks do not do either, they must not ever be preempted
4156 * (schedule out non-voluntarily).
4158 * schedule_idle() is similar to schedule_preempt_disable() except that it
4159 * never enables preemption because it does not call sched_submit_work().
4161 void __sched schedule_idle(void)
4164 * As this skips calling sched_submit_work(), which the idle task does
4165 * regardless because that function is a nop when the task is in a
4166 * TASK_RUNNING state, make sure this isn't used someplace that the
4167 * current task can be in any other state. Note, idle is always in the
4168 * TASK_RUNNING state.
4170 WARN_ON_ONCE(current->state);
4173 } while (need_resched());
4176 #ifdef CONFIG_CONTEXT_TRACKING
4177 asmlinkage __visible void __sched schedule_user(void)
4180 * If we come here after a random call to set_need_resched(),
4181 * or we have been woken up remotely but the IPI has not yet arrived,
4182 * we haven't yet exited the RCU idle mode. Do it here manually until
4183 * we find a better solution.
4185 * NB: There are buggy callers of this function. Ideally we
4186 * should warn if prev_state != CONTEXT_USER, but that will trigger
4187 * too frequently to make sense yet.
4189 enum ctx_state prev_state = exception_enter();
4191 exception_exit(prev_state);
4196 * schedule_preempt_disabled - called with preemption disabled
4198 * Returns with preemption disabled. Note: preempt_count must be 1
4200 void __sched schedule_preempt_disabled(void)
4202 sched_preempt_enable_no_resched();
4207 static void __sched notrace preempt_schedule_common(void)
4211 * Because the function tracer can trace preempt_count_sub()
4212 * and it also uses preempt_enable/disable_notrace(), if
4213 * NEED_RESCHED is set, the preempt_enable_notrace() called
4214 * by the function tracer will call this function again and
4215 * cause infinite recursion.
4217 * Preemption must be disabled here before the function
4218 * tracer can trace. Break up preempt_disable() into two
4219 * calls. One to disable preemption without fear of being
4220 * traced. The other to still record the preemption latency,
4221 * which can also be traced by the function tracer.
4223 preempt_disable_notrace();
4224 preempt_latency_start(1);
4226 preempt_latency_stop(1);
4227 preempt_enable_no_resched_notrace();
4230 * Check again in case we missed a preemption opportunity
4231 * between schedule and now.
4233 } while (need_resched());
4236 #ifdef CONFIG_PREEMPTION
4238 * This is the entry point to schedule() from in-kernel preemption
4239 * off of preempt_enable.
4241 asmlinkage __visible void __sched notrace preempt_schedule(void)
4244 * If there is a non-zero preempt_count or interrupts are disabled,
4245 * we do not want to preempt the current task. Just return..
4247 if (likely(!preemptible()))
4250 preempt_schedule_common();
4252 NOKPROBE_SYMBOL(preempt_schedule);
4253 EXPORT_SYMBOL(preempt_schedule);
4256 * preempt_schedule_notrace - preempt_schedule called by tracing
4258 * The tracing infrastructure uses preempt_enable_notrace to prevent
4259 * recursion and tracing preempt enabling caused by the tracing
4260 * infrastructure itself. But as tracing can happen in areas coming
4261 * from userspace or just about to enter userspace, a preempt enable
4262 * can occur before user_exit() is called. This will cause the scheduler
4263 * to be called when the system is still in usermode.
4265 * To prevent this, the preempt_enable_notrace will use this function
4266 * instead of preempt_schedule() to exit user context if needed before
4267 * calling the scheduler.
4269 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4271 enum ctx_state prev_ctx;
4273 if (likely(!preemptible()))
4278 * Because the function tracer can trace preempt_count_sub()
4279 * and it also uses preempt_enable/disable_notrace(), if
4280 * NEED_RESCHED is set, the preempt_enable_notrace() called
4281 * by the function tracer will call this function again and
4282 * cause infinite recursion.
4284 * Preemption must be disabled here before the function
4285 * tracer can trace. Break up preempt_disable() into two
4286 * calls. One to disable preemption without fear of being
4287 * traced. The other to still record the preemption latency,
4288 * which can also be traced by the function tracer.
4290 preempt_disable_notrace();
4291 preempt_latency_start(1);
4293 * Needs preempt disabled in case user_exit() is traced
4294 * and the tracer calls preempt_enable_notrace() causing
4295 * an infinite recursion.
4297 prev_ctx = exception_enter();
4299 exception_exit(prev_ctx);
4301 preempt_latency_stop(1);
4302 preempt_enable_no_resched_notrace();
4303 } while (need_resched());
4305 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4307 #endif /* CONFIG_PREEMPTION */
4310 * This is the entry point to schedule() from kernel preemption
4311 * off of irq context.
4312 * Note, that this is called and return with irqs disabled. This will
4313 * protect us against recursive calling from irq.
4315 asmlinkage __visible void __sched preempt_schedule_irq(void)
4317 enum ctx_state prev_state;
4319 /* Catch callers which need to be fixed */
4320 BUG_ON(preempt_count() || !irqs_disabled());
4322 prev_state = exception_enter();
4328 local_irq_disable();
4329 sched_preempt_enable_no_resched();
4330 } while (need_resched());
4332 exception_exit(prev_state);
4335 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4338 return try_to_wake_up(curr->private, mode, wake_flags);
4340 EXPORT_SYMBOL(default_wake_function);
4342 #ifdef CONFIG_RT_MUTEXES
4344 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4347 prio = min(prio, pi_task->prio);
4352 static inline int rt_effective_prio(struct task_struct *p, int prio)
4354 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4356 return __rt_effective_prio(pi_task, prio);
4360 * rt_mutex_setprio - set the current priority of a task
4362 * @pi_task: donor task
4364 * This function changes the 'effective' priority of a task. It does
4365 * not touch ->normal_prio like __setscheduler().
4367 * Used by the rt_mutex code to implement priority inheritance
4368 * logic. Call site only calls if the priority of the task changed.
4370 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4372 int prio, oldprio, queued, running, queue_flag =
4373 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4374 const struct sched_class *prev_class;
4378 /* XXX used to be waiter->prio, not waiter->task->prio */
4379 prio = __rt_effective_prio(pi_task, p->normal_prio);
4382 * If nothing changed; bail early.
4384 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4387 rq = __task_rq_lock(p, &rf);
4388 update_rq_clock(rq);
4390 * Set under pi_lock && rq->lock, such that the value can be used under
4393 * Note that there is loads of tricky to make this pointer cache work
4394 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4395 * ensure a task is de-boosted (pi_task is set to NULL) before the
4396 * task is allowed to run again (and can exit). This ensures the pointer
4397 * points to a blocked task -- which guaratees the task is present.
4399 p->pi_top_task = pi_task;
4402 * For FIFO/RR we only need to set prio, if that matches we're done.
4404 if (prio == p->prio && !dl_prio(prio))
4408 * Idle task boosting is a nono in general. There is one
4409 * exception, when PREEMPT_RT and NOHZ is active:
4411 * The idle task calls get_next_timer_interrupt() and holds
4412 * the timer wheel base->lock on the CPU and another CPU wants
4413 * to access the timer (probably to cancel it). We can safely
4414 * ignore the boosting request, as the idle CPU runs this code
4415 * with interrupts disabled and will complete the lock
4416 * protected section without being interrupted. So there is no
4417 * real need to boost.
4419 if (unlikely(p == rq->idle)) {
4420 WARN_ON(p != rq->curr);
4421 WARN_ON(p->pi_blocked_on);
4425 trace_sched_pi_setprio(p, pi_task);
4428 if (oldprio == prio)
4429 queue_flag &= ~DEQUEUE_MOVE;
4431 prev_class = p->sched_class;
4432 queued = task_on_rq_queued(p);
4433 running = task_current(rq, p);
4435 dequeue_task(rq, p, queue_flag);
4437 put_prev_task(rq, p);
4440 * Boosting condition are:
4441 * 1. -rt task is running and holds mutex A
4442 * --> -dl task blocks on mutex A
4444 * 2. -dl task is running and holds mutex A
4445 * --> -dl task blocks on mutex A and could preempt the
4448 if (dl_prio(prio)) {
4449 if (!dl_prio(p->normal_prio) ||
4450 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4451 p->dl.dl_boosted = 1;
4452 queue_flag |= ENQUEUE_REPLENISH;
4454 p->dl.dl_boosted = 0;
4455 p->sched_class = &dl_sched_class;
4456 } else if (rt_prio(prio)) {
4457 if (dl_prio(oldprio))
4458 p->dl.dl_boosted = 0;
4460 queue_flag |= ENQUEUE_HEAD;
4461 p->sched_class = &rt_sched_class;
4463 if (dl_prio(oldprio))
4464 p->dl.dl_boosted = 0;
4465 if (rt_prio(oldprio))
4467 p->sched_class = &fair_sched_class;
4473 enqueue_task(rq, p, queue_flag);
4475 set_next_task(rq, p);
4477 check_class_changed(rq, p, prev_class, oldprio);
4479 /* Avoid rq from going away on us: */
4481 __task_rq_unlock(rq, &rf);
4483 balance_callback(rq);
4487 static inline int rt_effective_prio(struct task_struct *p, int prio)
4493 void set_user_nice(struct task_struct *p, long nice)
4495 bool queued, running;
4496 int old_prio, delta;
4500 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4503 * We have to be careful, if called from sys_setpriority(),
4504 * the task might be in the middle of scheduling on another CPU.
4506 rq = task_rq_lock(p, &rf);
4507 update_rq_clock(rq);
4510 * The RT priorities are set via sched_setscheduler(), but we still
4511 * allow the 'normal' nice value to be set - but as expected
4512 * it wont have any effect on scheduling until the task is
4513 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4515 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4516 p->static_prio = NICE_TO_PRIO(nice);
4519 queued = task_on_rq_queued(p);
4520 running = task_current(rq, p);
4522 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4524 put_prev_task(rq, p);
4526 p->static_prio = NICE_TO_PRIO(nice);
4527 set_load_weight(p, true);
4529 p->prio = effective_prio(p);
4530 delta = p->prio - old_prio;
4533 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4535 * If the task increased its priority or is running and
4536 * lowered its priority, then reschedule its CPU:
4538 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4542 set_next_task(rq, p);
4544 task_rq_unlock(rq, p, &rf);
4546 EXPORT_SYMBOL(set_user_nice);
4549 * can_nice - check if a task can reduce its nice value
4553 int can_nice(const struct task_struct *p, const int nice)
4555 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4556 int nice_rlim = nice_to_rlimit(nice);
4558 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4559 capable(CAP_SYS_NICE));
4562 #ifdef __ARCH_WANT_SYS_NICE
4565 * sys_nice - change the priority of the current process.
4566 * @increment: priority increment
4568 * sys_setpriority is a more generic, but much slower function that
4569 * does similar things.
4571 SYSCALL_DEFINE1(nice, int, increment)
4576 * Setpriority might change our priority at the same moment.
4577 * We don't have to worry. Conceptually one call occurs first
4578 * and we have a single winner.
4580 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4581 nice = task_nice(current) + increment;
4583 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4584 if (increment < 0 && !can_nice(current, nice))
4587 retval = security_task_setnice(current, nice);
4591 set_user_nice(current, nice);
4598 * task_prio - return the priority value of a given task.
4599 * @p: the task in question.
4601 * Return: The priority value as seen by users in /proc.
4602 * RT tasks are offset by -200. Normal tasks are centered
4603 * around 0, value goes from -16 to +15.
4605 int task_prio(const struct task_struct *p)
4607 return p->prio - MAX_RT_PRIO;
4611 * idle_cpu - is a given CPU idle currently?
4612 * @cpu: the processor in question.
4614 * Return: 1 if the CPU is currently idle. 0 otherwise.
4616 int idle_cpu(int cpu)
4618 struct rq *rq = cpu_rq(cpu);
4620 if (rq->curr != rq->idle)
4627 if (!llist_empty(&rq->wake_list))
4635 * available_idle_cpu - is a given CPU idle for enqueuing work.
4636 * @cpu: the CPU in question.
4638 * Return: 1 if the CPU is currently idle. 0 otherwise.
4640 int available_idle_cpu(int cpu)
4645 if (vcpu_is_preempted(cpu))
4652 * idle_task - return the idle task for a given CPU.
4653 * @cpu: the processor in question.
4655 * Return: The idle task for the CPU @cpu.
4657 struct task_struct *idle_task(int cpu)
4659 return cpu_rq(cpu)->idle;
4663 * find_process_by_pid - find a process with a matching PID value.
4664 * @pid: the pid in question.
4666 * The task of @pid, if found. %NULL otherwise.
4668 static struct task_struct *find_process_by_pid(pid_t pid)
4670 return pid ? find_task_by_vpid(pid) : current;
4674 * sched_setparam() passes in -1 for its policy, to let the functions
4675 * it calls know not to change it.
4677 #define SETPARAM_POLICY -1
4679 static void __setscheduler_params(struct task_struct *p,
4680 const struct sched_attr *attr)
4682 int policy = attr->sched_policy;
4684 if (policy == SETPARAM_POLICY)
4689 if (dl_policy(policy))
4690 __setparam_dl(p, attr);
4691 else if (fair_policy(policy))
4692 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4695 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4696 * !rt_policy. Always setting this ensures that things like
4697 * getparam()/getattr() don't report silly values for !rt tasks.
4699 p->rt_priority = attr->sched_priority;
4700 p->normal_prio = normal_prio(p);
4701 set_load_weight(p, true);
4704 /* Actually do priority change: must hold pi & rq lock. */
4705 static void __setscheduler(struct rq *rq, struct task_struct *p,
4706 const struct sched_attr *attr, bool keep_boost)
4709 * If params can't change scheduling class changes aren't allowed
4712 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4715 __setscheduler_params(p, attr);
4718 * Keep a potential priority boosting if called from
4719 * sched_setscheduler().
4721 p->prio = normal_prio(p);
4723 p->prio = rt_effective_prio(p, p->prio);
4725 if (dl_prio(p->prio))
4726 p->sched_class = &dl_sched_class;
4727 else if (rt_prio(p->prio))
4728 p->sched_class = &rt_sched_class;
4730 p->sched_class = &fair_sched_class;
4734 * Check the target process has a UID that matches the current process's:
4736 static bool check_same_owner(struct task_struct *p)
4738 const struct cred *cred = current_cred(), *pcred;
4742 pcred = __task_cred(p);
4743 match = (uid_eq(cred->euid, pcred->euid) ||
4744 uid_eq(cred->euid, pcred->uid));
4749 static int __sched_setscheduler(struct task_struct *p,
4750 const struct sched_attr *attr,
4753 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4754 MAX_RT_PRIO - 1 - attr->sched_priority;
4755 int retval, oldprio, oldpolicy = -1, queued, running;
4756 int new_effective_prio, policy = attr->sched_policy;
4757 const struct sched_class *prev_class;
4760 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4763 /* The pi code expects interrupts enabled */
4764 BUG_ON(pi && in_interrupt());
4766 /* Double check policy once rq lock held: */
4768 reset_on_fork = p->sched_reset_on_fork;
4769 policy = oldpolicy = p->policy;
4771 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4773 if (!valid_policy(policy))
4777 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4781 * Valid priorities for SCHED_FIFO and SCHED_RR are
4782 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4783 * SCHED_BATCH and SCHED_IDLE is 0.
4785 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4786 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4788 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4789 (rt_policy(policy) != (attr->sched_priority != 0)))
4793 * Allow unprivileged RT tasks to decrease priority:
4795 if (user && !capable(CAP_SYS_NICE)) {
4796 if (fair_policy(policy)) {
4797 if (attr->sched_nice < task_nice(p) &&
4798 !can_nice(p, attr->sched_nice))
4802 if (rt_policy(policy)) {
4803 unsigned long rlim_rtprio =
4804 task_rlimit(p, RLIMIT_RTPRIO);
4806 /* Can't set/change the rt policy: */
4807 if (policy != p->policy && !rlim_rtprio)
4810 /* Can't increase priority: */
4811 if (attr->sched_priority > p->rt_priority &&
4812 attr->sched_priority > rlim_rtprio)
4817 * Can't set/change SCHED_DEADLINE policy at all for now
4818 * (safest behavior); in the future we would like to allow
4819 * unprivileged DL tasks to increase their relative deadline
4820 * or reduce their runtime (both ways reducing utilization)
4822 if (dl_policy(policy))
4826 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4827 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4829 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4830 if (!can_nice(p, task_nice(p)))
4834 /* Can't change other user's priorities: */
4835 if (!check_same_owner(p))
4838 /* Normal users shall not reset the sched_reset_on_fork flag: */
4839 if (p->sched_reset_on_fork && !reset_on_fork)
4844 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4847 retval = security_task_setscheduler(p);
4852 /* Update task specific "requested" clamps */
4853 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4854 retval = uclamp_validate(p, attr);
4863 * Make sure no PI-waiters arrive (or leave) while we are
4864 * changing the priority of the task:
4866 * To be able to change p->policy safely, the appropriate
4867 * runqueue lock must be held.
4869 rq = task_rq_lock(p, &rf);
4870 update_rq_clock(rq);
4873 * Changing the policy of the stop threads its a very bad idea:
4875 if (p == rq->stop) {
4881 * If not changing anything there's no need to proceed further,
4882 * but store a possible modification of reset_on_fork.
4884 if (unlikely(policy == p->policy)) {
4885 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4887 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4889 if (dl_policy(policy) && dl_param_changed(p, attr))
4891 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4894 p->sched_reset_on_fork = reset_on_fork;
4901 #ifdef CONFIG_RT_GROUP_SCHED
4903 * Do not allow realtime tasks into groups that have no runtime
4906 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4907 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4908 !task_group_is_autogroup(task_group(p))) {
4914 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4915 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4916 cpumask_t *span = rq->rd->span;
4919 * Don't allow tasks with an affinity mask smaller than
4920 * the entire root_domain to become SCHED_DEADLINE. We
4921 * will also fail if there's no bandwidth available.
4923 if (!cpumask_subset(span, p->cpus_ptr) ||
4924 rq->rd->dl_bw.bw == 0) {
4932 /* Re-check policy now with rq lock held: */
4933 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4934 policy = oldpolicy = -1;
4935 task_rq_unlock(rq, p, &rf);
4937 cpuset_read_unlock();
4942 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4943 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4946 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4951 p->sched_reset_on_fork = reset_on_fork;
4956 * Take priority boosted tasks into account. If the new
4957 * effective priority is unchanged, we just store the new
4958 * normal parameters and do not touch the scheduler class and
4959 * the runqueue. This will be done when the task deboost
4962 new_effective_prio = rt_effective_prio(p, newprio);
4963 if (new_effective_prio == oldprio)
4964 queue_flags &= ~DEQUEUE_MOVE;
4967 queued = task_on_rq_queued(p);
4968 running = task_current(rq, p);
4970 dequeue_task(rq, p, queue_flags);
4972 put_prev_task(rq, p);
4974 prev_class = p->sched_class;
4976 __setscheduler(rq, p, attr, pi);
4977 __setscheduler_uclamp(p, attr);
4981 * We enqueue to tail when the priority of a task is
4982 * increased (user space view).
4984 if (oldprio < p->prio)
4985 queue_flags |= ENQUEUE_HEAD;
4987 enqueue_task(rq, p, queue_flags);
4990 set_next_task(rq, p);
4992 check_class_changed(rq, p, prev_class, oldprio);
4994 /* Avoid rq from going away on us: */
4996 task_rq_unlock(rq, p, &rf);
4999 cpuset_read_unlock();
5000 rt_mutex_adjust_pi(p);
5003 /* Run balance callbacks after we've adjusted the PI chain: */
5004 balance_callback(rq);
5010 task_rq_unlock(rq, p, &rf);
5012 cpuset_read_unlock();
5016 static int _sched_setscheduler(struct task_struct *p, int policy,
5017 const struct sched_param *param, bool check)
5019 struct sched_attr attr = {
5020 .sched_policy = policy,
5021 .sched_priority = param->sched_priority,
5022 .sched_nice = PRIO_TO_NICE(p->static_prio),
5025 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5026 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5027 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5028 policy &= ~SCHED_RESET_ON_FORK;
5029 attr.sched_policy = policy;
5032 return __sched_setscheduler(p, &attr, check, true);
5035 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5036 * @p: the task in question.
5037 * @policy: new policy.
5038 * @param: structure containing the new RT priority.
5040 * Return: 0 on success. An error code otherwise.
5042 * NOTE that the task may be already dead.
5044 int sched_setscheduler(struct task_struct *p, int policy,
5045 const struct sched_param *param)
5047 return _sched_setscheduler(p, policy, param, true);
5049 EXPORT_SYMBOL_GPL(sched_setscheduler);
5051 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5053 return __sched_setscheduler(p, attr, true, true);
5055 EXPORT_SYMBOL_GPL(sched_setattr);
5057 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5059 return __sched_setscheduler(p, attr, false, true);
5063 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5064 * @p: the task in question.
5065 * @policy: new policy.
5066 * @param: structure containing the new RT priority.
5068 * Just like sched_setscheduler, only don't bother checking if the
5069 * current context has permission. For example, this is needed in
5070 * stop_machine(): we create temporary high priority worker threads,
5071 * but our caller might not have that capability.
5073 * Return: 0 on success. An error code otherwise.
5075 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5076 const struct sched_param *param)
5078 return _sched_setscheduler(p, policy, param, false);
5080 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5083 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5085 struct sched_param lparam;
5086 struct task_struct *p;
5089 if (!param || pid < 0)
5091 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5096 p = find_process_by_pid(pid);
5102 retval = sched_setscheduler(p, policy, &lparam);
5110 * Mimics kernel/events/core.c perf_copy_attr().
5112 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5117 /* Zero the full structure, so that a short copy will be nice: */
5118 memset(attr, 0, sizeof(*attr));
5120 ret = get_user(size, &uattr->size);
5124 /* ABI compatibility quirk: */
5126 size = SCHED_ATTR_SIZE_VER0;
5127 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5130 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5137 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5138 size < SCHED_ATTR_SIZE_VER1)
5142 * XXX: Do we want to be lenient like existing syscalls; or do we want
5143 * to be strict and return an error on out-of-bounds values?
5145 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5150 put_user(sizeof(*attr), &uattr->size);
5155 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5156 * @pid: the pid in question.
5157 * @policy: new policy.
5158 * @param: structure containing the new RT priority.
5160 * Return: 0 on success. An error code otherwise.
5162 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5167 return do_sched_setscheduler(pid, policy, param);
5171 * sys_sched_setparam - set/change the RT priority of a thread
5172 * @pid: the pid in question.
5173 * @param: structure containing the new RT priority.
5175 * Return: 0 on success. An error code otherwise.
5177 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5179 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5183 * sys_sched_setattr - same as above, but with extended sched_attr
5184 * @pid: the pid in question.
5185 * @uattr: structure containing the extended parameters.
5186 * @flags: for future extension.
5188 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5189 unsigned int, flags)
5191 struct sched_attr attr;
5192 struct task_struct *p;
5195 if (!uattr || pid < 0 || flags)
5198 retval = sched_copy_attr(uattr, &attr);
5202 if ((int)attr.sched_policy < 0)
5204 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5205 attr.sched_policy = SETPARAM_POLICY;
5209 p = find_process_by_pid(pid);
5215 retval = sched_setattr(p, &attr);
5223 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5224 * @pid: the pid in question.
5226 * Return: On success, the policy of the thread. Otherwise, a negative error
5229 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5231 struct task_struct *p;
5239 p = find_process_by_pid(pid);
5241 retval = security_task_getscheduler(p);
5244 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5251 * sys_sched_getparam - get the RT priority of a thread
5252 * @pid: the pid in question.
5253 * @param: structure containing the RT priority.
5255 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5258 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5260 struct sched_param lp = { .sched_priority = 0 };
5261 struct task_struct *p;
5264 if (!param || pid < 0)
5268 p = find_process_by_pid(pid);
5273 retval = security_task_getscheduler(p);
5277 if (task_has_rt_policy(p))
5278 lp.sched_priority = p->rt_priority;
5282 * This one might sleep, we cannot do it with a spinlock held ...
5284 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5294 * Copy the kernel size attribute structure (which might be larger
5295 * than what user-space knows about) to user-space.
5297 * Note that all cases are valid: user-space buffer can be larger or
5298 * smaller than the kernel-space buffer. The usual case is that both
5299 * have the same size.
5302 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5303 struct sched_attr *kattr,
5306 unsigned int ksize = sizeof(*kattr);
5308 if (!access_ok(uattr, usize))
5312 * sched_getattr() ABI forwards and backwards compatibility:
5314 * If usize == ksize then we just copy everything to user-space and all is good.
5316 * If usize < ksize then we only copy as much as user-space has space for,
5317 * this keeps ABI compatibility as well. We skip the rest.
5319 * If usize > ksize then user-space is using a newer version of the ABI,
5320 * which part the kernel doesn't know about. Just ignore it - tooling can
5321 * detect the kernel's knowledge of attributes from the attr->size value
5322 * which is set to ksize in this case.
5324 kattr->size = min(usize, ksize);
5326 if (copy_to_user(uattr, kattr, kattr->size))
5333 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5334 * @pid: the pid in question.
5335 * @uattr: structure containing the extended parameters.
5336 * @usize: sizeof(attr) for fwd/bwd comp.
5337 * @flags: for future extension.
5339 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5340 unsigned int, usize, unsigned int, flags)
5342 struct sched_attr kattr = { };
5343 struct task_struct *p;
5346 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5347 usize < SCHED_ATTR_SIZE_VER0 || flags)
5351 p = find_process_by_pid(pid);
5356 retval = security_task_getscheduler(p);
5360 kattr.sched_policy = p->policy;
5361 if (p->sched_reset_on_fork)
5362 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5363 if (task_has_dl_policy(p))
5364 __getparam_dl(p, &kattr);
5365 else if (task_has_rt_policy(p))
5366 kattr.sched_priority = p->rt_priority;
5368 kattr.sched_nice = task_nice(p);
5370 #ifdef CONFIG_UCLAMP_TASK
5371 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5372 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5377 return sched_attr_copy_to_user(uattr, &kattr, usize);
5384 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5386 cpumask_var_t cpus_allowed, new_mask;
5387 struct task_struct *p;
5392 p = find_process_by_pid(pid);
5398 /* Prevent p going away */
5402 if (p->flags & PF_NO_SETAFFINITY) {
5406 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5410 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5412 goto out_free_cpus_allowed;
5415 if (!check_same_owner(p)) {
5417 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5419 goto out_free_new_mask;
5424 retval = security_task_setscheduler(p);
5426 goto out_free_new_mask;
5429 cpuset_cpus_allowed(p, cpus_allowed);
5430 cpumask_and(new_mask, in_mask, cpus_allowed);
5433 * Since bandwidth control happens on root_domain basis,
5434 * if admission test is enabled, we only admit -deadline
5435 * tasks allowed to run on all the CPUs in the task's
5439 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5441 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5444 goto out_free_new_mask;
5450 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5453 cpuset_cpus_allowed(p, cpus_allowed);
5454 if (!cpumask_subset(new_mask, cpus_allowed)) {
5456 * We must have raced with a concurrent cpuset
5457 * update. Just reset the cpus_allowed to the
5458 * cpuset's cpus_allowed
5460 cpumask_copy(new_mask, cpus_allowed);
5465 free_cpumask_var(new_mask);
5466 out_free_cpus_allowed:
5467 free_cpumask_var(cpus_allowed);
5473 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5474 struct cpumask *new_mask)
5476 if (len < cpumask_size())
5477 cpumask_clear(new_mask);
5478 else if (len > cpumask_size())
5479 len = cpumask_size();
5481 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5485 * sys_sched_setaffinity - set the CPU affinity of a process
5486 * @pid: pid of the process
5487 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5488 * @user_mask_ptr: user-space pointer to the new CPU mask
5490 * Return: 0 on success. An error code otherwise.
5492 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5493 unsigned long __user *, user_mask_ptr)
5495 cpumask_var_t new_mask;
5498 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5501 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5503 retval = sched_setaffinity(pid, new_mask);
5504 free_cpumask_var(new_mask);
5508 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5510 struct task_struct *p;
5511 unsigned long flags;
5517 p = find_process_by_pid(pid);
5521 retval = security_task_getscheduler(p);
5525 raw_spin_lock_irqsave(&p->pi_lock, flags);
5526 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5527 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5536 * sys_sched_getaffinity - get the CPU affinity of a process
5537 * @pid: pid of the process
5538 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5539 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5541 * Return: size of CPU mask copied to user_mask_ptr on success. An
5542 * error code otherwise.
5544 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5545 unsigned long __user *, user_mask_ptr)
5550 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5552 if (len & (sizeof(unsigned long)-1))
5555 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5558 ret = sched_getaffinity(pid, mask);
5560 unsigned int retlen = min(len, cpumask_size());
5562 if (copy_to_user(user_mask_ptr, mask, retlen))
5567 free_cpumask_var(mask);
5573 * sys_sched_yield - yield the current processor to other threads.
5575 * This function yields the current CPU to other tasks. If there are no
5576 * other threads running on this CPU then this function will return.
5580 static void do_sched_yield(void)
5585 rq = this_rq_lock_irq(&rf);
5587 schedstat_inc(rq->yld_count);
5588 current->sched_class->yield_task(rq);
5591 * Since we are going to call schedule() anyway, there's
5592 * no need to preempt or enable interrupts:
5596 sched_preempt_enable_no_resched();
5601 SYSCALL_DEFINE0(sched_yield)
5607 #ifndef CONFIG_PREEMPTION
5608 int __sched _cond_resched(void)
5610 if (should_resched(0)) {
5611 preempt_schedule_common();
5617 EXPORT_SYMBOL(_cond_resched);
5621 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5622 * call schedule, and on return reacquire the lock.
5624 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5625 * operations here to prevent schedule() from being called twice (once via
5626 * spin_unlock(), once by hand).
5628 int __cond_resched_lock(spinlock_t *lock)
5630 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5633 lockdep_assert_held(lock);
5635 if (spin_needbreak(lock) || resched) {
5638 preempt_schedule_common();
5646 EXPORT_SYMBOL(__cond_resched_lock);
5649 * yield - yield the current processor to other threads.
5651 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5653 * The scheduler is at all times free to pick the calling task as the most
5654 * eligible task to run, if removing the yield() call from your code breaks
5655 * it, its already broken.
5657 * Typical broken usage is:
5662 * where one assumes that yield() will let 'the other' process run that will
5663 * make event true. If the current task is a SCHED_FIFO task that will never
5664 * happen. Never use yield() as a progress guarantee!!
5666 * If you want to use yield() to wait for something, use wait_event().
5667 * If you want to use yield() to be 'nice' for others, use cond_resched().
5668 * If you still want to use yield(), do not!
5670 void __sched yield(void)
5672 set_current_state(TASK_RUNNING);
5675 EXPORT_SYMBOL(yield);
5678 * yield_to - yield the current processor to another thread in
5679 * your thread group, or accelerate that thread toward the
5680 * processor it's on.
5682 * @preempt: whether task preemption is allowed or not
5684 * It's the caller's job to ensure that the target task struct
5685 * can't go away on us before we can do any checks.
5688 * true (>0) if we indeed boosted the target task.
5689 * false (0) if we failed to boost the target.
5690 * -ESRCH if there's no task to yield to.
5692 int __sched yield_to(struct task_struct *p, bool preempt)
5694 struct task_struct *curr = current;
5695 struct rq *rq, *p_rq;
5696 unsigned long flags;
5699 local_irq_save(flags);
5705 * If we're the only runnable task on the rq and target rq also
5706 * has only one task, there's absolutely no point in yielding.
5708 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5713 double_rq_lock(rq, p_rq);
5714 if (task_rq(p) != p_rq) {
5715 double_rq_unlock(rq, p_rq);
5719 if (!curr->sched_class->yield_to_task)
5722 if (curr->sched_class != p->sched_class)
5725 if (task_running(p_rq, p) || p->state)
5728 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5730 schedstat_inc(rq->yld_count);
5732 * Make p's CPU reschedule; pick_next_entity takes care of
5735 if (preempt && rq != p_rq)
5740 double_rq_unlock(rq, p_rq);
5742 local_irq_restore(flags);
5749 EXPORT_SYMBOL_GPL(yield_to);
5751 int io_schedule_prepare(void)
5753 int old_iowait = current->in_iowait;
5755 current->in_iowait = 1;
5756 blk_schedule_flush_plug(current);
5761 void io_schedule_finish(int token)
5763 current->in_iowait = token;
5767 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5768 * that process accounting knows that this is a task in IO wait state.
5770 long __sched io_schedule_timeout(long timeout)
5775 token = io_schedule_prepare();
5776 ret = schedule_timeout(timeout);
5777 io_schedule_finish(token);
5781 EXPORT_SYMBOL(io_schedule_timeout);
5783 void __sched io_schedule(void)
5787 token = io_schedule_prepare();
5789 io_schedule_finish(token);
5791 EXPORT_SYMBOL(io_schedule);
5794 * sys_sched_get_priority_max - return maximum RT priority.
5795 * @policy: scheduling class.
5797 * Return: On success, this syscall returns the maximum
5798 * rt_priority that can be used by a given scheduling class.
5799 * On failure, a negative error code is returned.
5801 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5808 ret = MAX_USER_RT_PRIO-1;
5810 case SCHED_DEADLINE:
5821 * sys_sched_get_priority_min - return minimum RT priority.
5822 * @policy: scheduling class.
5824 * Return: On success, this syscall returns the minimum
5825 * rt_priority that can be used by a given scheduling class.
5826 * On failure, a negative error code is returned.
5828 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5837 case SCHED_DEADLINE:
5846 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5848 struct task_struct *p;
5849 unsigned int time_slice;
5859 p = find_process_by_pid(pid);
5863 retval = security_task_getscheduler(p);
5867 rq = task_rq_lock(p, &rf);
5869 if (p->sched_class->get_rr_interval)
5870 time_slice = p->sched_class->get_rr_interval(rq, p);
5871 task_rq_unlock(rq, p, &rf);
5874 jiffies_to_timespec64(time_slice, t);
5883 * sys_sched_rr_get_interval - return the default timeslice of a process.
5884 * @pid: pid of the process.
5885 * @interval: userspace pointer to the timeslice value.
5887 * this syscall writes the default timeslice value of a given process
5888 * into the user-space timespec buffer. A value of '0' means infinity.
5890 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5893 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5894 struct __kernel_timespec __user *, interval)
5896 struct timespec64 t;
5897 int retval = sched_rr_get_interval(pid, &t);
5900 retval = put_timespec64(&t, interval);
5905 #ifdef CONFIG_COMPAT_32BIT_TIME
5906 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5907 struct old_timespec32 __user *, interval)
5909 struct timespec64 t;
5910 int retval = sched_rr_get_interval(pid, &t);
5913 retval = put_old_timespec32(&t, interval);
5918 void sched_show_task(struct task_struct *p)
5920 unsigned long free = 0;
5923 if (!try_get_task_stack(p))
5926 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5928 if (p->state == TASK_RUNNING)
5929 printk(KERN_CONT " running task ");
5930 #ifdef CONFIG_DEBUG_STACK_USAGE
5931 free = stack_not_used(p);
5936 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5938 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5939 task_pid_nr(p), ppid,
5940 (unsigned long)task_thread_info(p)->flags);
5942 print_worker_info(KERN_INFO, p);
5943 show_stack(p, NULL);
5946 EXPORT_SYMBOL_GPL(sched_show_task);
5949 state_filter_match(unsigned long state_filter, struct task_struct *p)
5951 /* no filter, everything matches */
5955 /* filter, but doesn't match */
5956 if (!(p->state & state_filter))
5960 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5963 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5970 void show_state_filter(unsigned long state_filter)
5972 struct task_struct *g, *p;
5974 #if BITS_PER_LONG == 32
5976 " task PC stack pid father\n");
5979 " task PC stack pid father\n");
5982 for_each_process_thread(g, p) {
5984 * reset the NMI-timeout, listing all files on a slow
5985 * console might take a lot of time:
5986 * Also, reset softlockup watchdogs on all CPUs, because
5987 * another CPU might be blocked waiting for us to process
5990 touch_nmi_watchdog();
5991 touch_all_softlockup_watchdogs();
5992 if (state_filter_match(state_filter, p))
5996 #ifdef CONFIG_SCHED_DEBUG
5998 sysrq_sched_debug_show();
6002 * Only show locks if all tasks are dumped:
6005 debug_show_all_locks();
6009 * init_idle - set up an idle thread for a given CPU
6010 * @idle: task in question
6011 * @cpu: CPU the idle task belongs to
6013 * NOTE: this function does not set the idle thread's NEED_RESCHED
6014 * flag, to make booting more robust.
6016 void init_idle(struct task_struct *idle, int cpu)
6018 struct rq *rq = cpu_rq(cpu);
6019 unsigned long flags;
6021 __sched_fork(0, idle);
6023 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6024 raw_spin_lock(&rq->lock);
6026 idle->state = TASK_RUNNING;
6027 idle->se.exec_start = sched_clock();
6028 idle->flags |= PF_IDLE;
6030 kasan_unpoison_task_stack(idle);
6034 * Its possible that init_idle() gets called multiple times on a task,
6035 * in that case do_set_cpus_allowed() will not do the right thing.
6037 * And since this is boot we can forgo the serialization.
6039 set_cpus_allowed_common(idle, cpumask_of(cpu));
6042 * We're having a chicken and egg problem, even though we are
6043 * holding rq->lock, the CPU isn't yet set to this CPU so the
6044 * lockdep check in task_group() will fail.
6046 * Similar case to sched_fork(). / Alternatively we could
6047 * use task_rq_lock() here and obtain the other rq->lock.
6052 __set_task_cpu(idle, cpu);
6056 rcu_assign_pointer(rq->curr, idle);
6057 idle->on_rq = TASK_ON_RQ_QUEUED;
6061 raw_spin_unlock(&rq->lock);
6062 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6064 /* Set the preempt count _outside_ the spinlocks! */
6065 init_idle_preempt_count(idle, cpu);
6068 * The idle tasks have their own, simple scheduling class:
6070 idle->sched_class = &idle_sched_class;
6071 ftrace_graph_init_idle_task(idle, cpu);
6072 vtime_init_idle(idle, cpu);
6074 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6080 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6081 const struct cpumask *trial)
6085 if (!cpumask_weight(cur))
6088 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6093 int task_can_attach(struct task_struct *p,
6094 const struct cpumask *cs_cpus_allowed)
6099 * Kthreads which disallow setaffinity shouldn't be moved
6100 * to a new cpuset; we don't want to change their CPU
6101 * affinity and isolating such threads by their set of
6102 * allowed nodes is unnecessary. Thus, cpusets are not
6103 * applicable for such threads. This prevents checking for
6104 * success of set_cpus_allowed_ptr() on all attached tasks
6105 * before cpus_mask may be changed.
6107 if (p->flags & PF_NO_SETAFFINITY) {
6112 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6114 ret = dl_task_can_attach(p, cs_cpus_allowed);
6120 bool sched_smp_initialized __read_mostly;
6122 #ifdef CONFIG_NUMA_BALANCING
6123 /* Migrate current task p to target_cpu */
6124 int migrate_task_to(struct task_struct *p, int target_cpu)
6126 struct migration_arg arg = { p, target_cpu };
6127 int curr_cpu = task_cpu(p);
6129 if (curr_cpu == target_cpu)
6132 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6135 /* TODO: This is not properly updating schedstats */
6137 trace_sched_move_numa(p, curr_cpu, target_cpu);
6138 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6142 * Requeue a task on a given node and accurately track the number of NUMA
6143 * tasks on the runqueues
6145 void sched_setnuma(struct task_struct *p, int nid)
6147 bool queued, running;
6151 rq = task_rq_lock(p, &rf);
6152 queued = task_on_rq_queued(p);
6153 running = task_current(rq, p);
6156 dequeue_task(rq, p, DEQUEUE_SAVE);
6158 put_prev_task(rq, p);
6160 p->numa_preferred_nid = nid;
6163 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6165 set_next_task(rq, p);
6166 task_rq_unlock(rq, p, &rf);
6168 #endif /* CONFIG_NUMA_BALANCING */
6170 #ifdef CONFIG_HOTPLUG_CPU
6172 * Ensure that the idle task is using init_mm right before its CPU goes
6175 void idle_task_exit(void)
6177 struct mm_struct *mm = current->active_mm;
6179 BUG_ON(cpu_online(smp_processor_id()));
6180 BUG_ON(current != this_rq()->idle);
6182 if (mm != &init_mm) {
6183 switch_mm(mm, &init_mm, current);
6184 finish_arch_post_lock_switch();
6187 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6191 * Since this CPU is going 'away' for a while, fold any nr_active delta
6192 * we might have. Assumes we're called after migrate_tasks() so that the
6193 * nr_active count is stable. We need to take the teardown thread which
6194 * is calling this into account, so we hand in adjust = 1 to the load
6197 * Also see the comment "Global load-average calculations".
6199 static void calc_load_migrate(struct rq *rq)
6201 long delta = calc_load_fold_active(rq, 1);
6203 atomic_long_add(delta, &calc_load_tasks);
6206 static struct task_struct *__pick_migrate_task(struct rq *rq)
6208 const struct sched_class *class;
6209 struct task_struct *next;
6211 for_each_class(class) {
6212 next = class->pick_next_task(rq, NULL, NULL);
6214 next->sched_class->put_prev_task(rq, next);
6219 /* The idle class should always have a runnable task */
6224 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6225 * try_to_wake_up()->select_task_rq().
6227 * Called with rq->lock held even though we'er in stop_machine() and
6228 * there's no concurrency possible, we hold the required locks anyway
6229 * because of lock validation efforts.
6231 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6233 struct rq *rq = dead_rq;
6234 struct task_struct *next, *stop = rq->stop;
6235 struct rq_flags orf = *rf;
6239 * Fudge the rq selection such that the below task selection loop
6240 * doesn't get stuck on the currently eligible stop task.
6242 * We're currently inside stop_machine() and the rq is either stuck
6243 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6244 * either way we should never end up calling schedule() until we're
6250 * put_prev_task() and pick_next_task() sched
6251 * class method both need to have an up-to-date
6252 * value of rq->clock[_task]
6254 update_rq_clock(rq);
6258 * There's this thread running, bail when that's the only
6261 if (rq->nr_running == 1)
6264 next = __pick_migrate_task(rq);
6267 * Rules for changing task_struct::cpus_mask are holding
6268 * both pi_lock and rq->lock, such that holding either
6269 * stabilizes the mask.
6271 * Drop rq->lock is not quite as disastrous as it usually is
6272 * because !cpu_active at this point, which means load-balance
6273 * will not interfere. Also, stop-machine.
6276 raw_spin_lock(&next->pi_lock);
6280 * Since we're inside stop-machine, _nothing_ should have
6281 * changed the task, WARN if weird stuff happened, because in
6282 * that case the above rq->lock drop is a fail too.
6284 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6285 raw_spin_unlock(&next->pi_lock);
6289 /* Find suitable destination for @next, with force if needed. */
6290 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6291 rq = __migrate_task(rq, rf, next, dest_cpu);
6292 if (rq != dead_rq) {
6298 raw_spin_unlock(&next->pi_lock);
6303 #endif /* CONFIG_HOTPLUG_CPU */
6305 void set_rq_online(struct rq *rq)
6308 const struct sched_class *class;
6310 cpumask_set_cpu(rq->cpu, rq->rd->online);
6313 for_each_class(class) {
6314 if (class->rq_online)
6315 class->rq_online(rq);
6320 void set_rq_offline(struct rq *rq)
6323 const struct sched_class *class;
6325 for_each_class(class) {
6326 if (class->rq_offline)
6327 class->rq_offline(rq);
6330 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6336 * used to mark begin/end of suspend/resume:
6338 static int num_cpus_frozen;
6341 * Update cpusets according to cpu_active mask. If cpusets are
6342 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6343 * around partition_sched_domains().
6345 * If we come here as part of a suspend/resume, don't touch cpusets because we
6346 * want to restore it back to its original state upon resume anyway.
6348 static void cpuset_cpu_active(void)
6350 if (cpuhp_tasks_frozen) {
6352 * num_cpus_frozen tracks how many CPUs are involved in suspend
6353 * resume sequence. As long as this is not the last online
6354 * operation in the resume sequence, just build a single sched
6355 * domain, ignoring cpusets.
6357 partition_sched_domains(1, NULL, NULL);
6358 if (--num_cpus_frozen)
6361 * This is the last CPU online operation. So fall through and
6362 * restore the original sched domains by considering the
6363 * cpuset configurations.
6365 cpuset_force_rebuild();
6367 cpuset_update_active_cpus();
6370 static int cpuset_cpu_inactive(unsigned int cpu)
6372 if (!cpuhp_tasks_frozen) {
6373 if (dl_cpu_busy(cpu))
6375 cpuset_update_active_cpus();
6378 partition_sched_domains(1, NULL, NULL);
6383 int sched_cpu_activate(unsigned int cpu)
6385 struct rq *rq = cpu_rq(cpu);
6388 #ifdef CONFIG_SCHED_SMT
6390 * When going up, increment the number of cores with SMT present.
6392 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6393 static_branch_inc_cpuslocked(&sched_smt_present);
6395 set_cpu_active(cpu, true);
6397 if (sched_smp_initialized) {
6398 sched_domains_numa_masks_set(cpu);
6399 cpuset_cpu_active();
6403 * Put the rq online, if not already. This happens:
6405 * 1) In the early boot process, because we build the real domains
6406 * after all CPUs have been brought up.
6408 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6411 rq_lock_irqsave(rq, &rf);
6413 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6416 rq_unlock_irqrestore(rq, &rf);
6421 int sched_cpu_deactivate(unsigned int cpu)
6425 set_cpu_active(cpu, false);
6427 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6428 * users of this state to go away such that all new such users will
6431 * Do sync before park smpboot threads to take care the rcu boost case.
6435 #ifdef CONFIG_SCHED_SMT
6437 * When going down, decrement the number of cores with SMT present.
6439 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6440 static_branch_dec_cpuslocked(&sched_smt_present);
6443 if (!sched_smp_initialized)
6446 ret = cpuset_cpu_inactive(cpu);
6448 set_cpu_active(cpu, true);
6451 sched_domains_numa_masks_clear(cpu);
6455 static void sched_rq_cpu_starting(unsigned int cpu)
6457 struct rq *rq = cpu_rq(cpu);
6459 rq->calc_load_update = calc_load_update;
6460 update_max_interval();
6463 int sched_cpu_starting(unsigned int cpu)
6465 sched_rq_cpu_starting(cpu);
6466 sched_tick_start(cpu);
6470 #ifdef CONFIG_HOTPLUG_CPU
6471 int sched_cpu_dying(unsigned int cpu)
6473 struct rq *rq = cpu_rq(cpu);
6476 /* Handle pending wakeups and then migrate everything off */
6477 sched_ttwu_pending();
6478 sched_tick_stop(cpu);
6480 rq_lock_irqsave(rq, &rf);
6482 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6485 migrate_tasks(rq, &rf);
6486 BUG_ON(rq->nr_running != 1);
6487 rq_unlock_irqrestore(rq, &rf);
6489 calc_load_migrate(rq);
6490 update_max_interval();
6491 nohz_balance_exit_idle(rq);
6497 void __init sched_init_smp(void)
6502 * There's no userspace yet to cause hotplug operations; hence all the
6503 * CPU masks are stable and all blatant races in the below code cannot
6506 mutex_lock(&sched_domains_mutex);
6507 sched_init_domains(cpu_active_mask);
6508 mutex_unlock(&sched_domains_mutex);
6510 /* Move init over to a non-isolated CPU */
6511 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6513 sched_init_granularity();
6515 init_sched_rt_class();
6516 init_sched_dl_class();
6518 sched_smp_initialized = true;
6521 static int __init migration_init(void)
6523 sched_cpu_starting(smp_processor_id());
6526 early_initcall(migration_init);
6529 void __init sched_init_smp(void)
6531 sched_init_granularity();
6533 #endif /* CONFIG_SMP */
6535 int in_sched_functions(unsigned long addr)
6537 return in_lock_functions(addr) ||
6538 (addr >= (unsigned long)__sched_text_start
6539 && addr < (unsigned long)__sched_text_end);
6542 #ifdef CONFIG_CGROUP_SCHED
6544 * Default task group.
6545 * Every task in system belongs to this group at bootup.
6547 struct task_group root_task_group;
6548 LIST_HEAD(task_groups);
6550 /* Cacheline aligned slab cache for task_group */
6551 static struct kmem_cache *task_group_cache __read_mostly;
6554 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6555 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6557 void __init sched_init(void)
6559 unsigned long ptr = 0;
6564 #ifdef CONFIG_FAIR_GROUP_SCHED
6565 ptr += 2 * nr_cpu_ids * sizeof(void **);
6567 #ifdef CONFIG_RT_GROUP_SCHED
6568 ptr += 2 * nr_cpu_ids * sizeof(void **);
6571 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6573 #ifdef CONFIG_FAIR_GROUP_SCHED
6574 root_task_group.se = (struct sched_entity **)ptr;
6575 ptr += nr_cpu_ids * sizeof(void **);
6577 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6578 ptr += nr_cpu_ids * sizeof(void **);
6580 #endif /* CONFIG_FAIR_GROUP_SCHED */
6581 #ifdef CONFIG_RT_GROUP_SCHED
6582 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6583 ptr += nr_cpu_ids * sizeof(void **);
6585 root_task_group.rt_rq = (struct rt_rq **)ptr;
6586 ptr += nr_cpu_ids * sizeof(void **);
6588 #endif /* CONFIG_RT_GROUP_SCHED */
6590 #ifdef CONFIG_CPUMASK_OFFSTACK
6591 for_each_possible_cpu(i) {
6592 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6593 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6594 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6595 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6597 #endif /* CONFIG_CPUMASK_OFFSTACK */
6599 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6600 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6603 init_defrootdomain();
6606 #ifdef CONFIG_RT_GROUP_SCHED
6607 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6608 global_rt_period(), global_rt_runtime());
6609 #endif /* CONFIG_RT_GROUP_SCHED */
6611 #ifdef CONFIG_CGROUP_SCHED
6612 task_group_cache = KMEM_CACHE(task_group, 0);
6614 list_add(&root_task_group.list, &task_groups);
6615 INIT_LIST_HEAD(&root_task_group.children);
6616 INIT_LIST_HEAD(&root_task_group.siblings);
6617 autogroup_init(&init_task);
6618 #endif /* CONFIG_CGROUP_SCHED */
6620 for_each_possible_cpu(i) {
6624 raw_spin_lock_init(&rq->lock);
6626 rq->calc_load_active = 0;
6627 rq->calc_load_update = jiffies + LOAD_FREQ;
6628 init_cfs_rq(&rq->cfs);
6629 init_rt_rq(&rq->rt);
6630 init_dl_rq(&rq->dl);
6631 #ifdef CONFIG_FAIR_GROUP_SCHED
6632 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6633 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6634 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6636 * How much CPU bandwidth does root_task_group get?
6638 * In case of task-groups formed thr' the cgroup filesystem, it
6639 * gets 100% of the CPU resources in the system. This overall
6640 * system CPU resource is divided among the tasks of
6641 * root_task_group and its child task-groups in a fair manner,
6642 * based on each entity's (task or task-group's) weight
6643 * (se->load.weight).
6645 * In other words, if root_task_group has 10 tasks of weight
6646 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6647 * then A0's share of the CPU resource is:
6649 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6651 * We achieve this by letting root_task_group's tasks sit
6652 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6654 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6655 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6656 #endif /* CONFIG_FAIR_GROUP_SCHED */
6658 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6659 #ifdef CONFIG_RT_GROUP_SCHED
6660 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6665 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6666 rq->balance_callback = NULL;
6667 rq->active_balance = 0;
6668 rq->next_balance = jiffies;
6673 rq->avg_idle = 2*sysctl_sched_migration_cost;
6674 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6676 INIT_LIST_HEAD(&rq->cfs_tasks);
6678 rq_attach_root(rq, &def_root_domain);
6679 #ifdef CONFIG_NO_HZ_COMMON
6680 rq->last_load_update_tick = jiffies;
6681 rq->last_blocked_load_update_tick = jiffies;
6682 atomic_set(&rq->nohz_flags, 0);
6684 #endif /* CONFIG_SMP */
6686 atomic_set(&rq->nr_iowait, 0);
6689 set_load_weight(&init_task, false);
6692 * The boot idle thread does lazy MMU switching as well:
6695 enter_lazy_tlb(&init_mm, current);
6698 * Make us the idle thread. Technically, schedule() should not be
6699 * called from this thread, however somewhere below it might be,
6700 * but because we are the idle thread, we just pick up running again
6701 * when this runqueue becomes "idle".
6703 init_idle(current, smp_processor_id());
6705 calc_load_update = jiffies + LOAD_FREQ;
6708 idle_thread_set_boot_cpu();
6710 init_sched_fair_class();
6718 scheduler_running = 1;
6721 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6722 static inline int preempt_count_equals(int preempt_offset)
6724 int nested = preempt_count() + rcu_preempt_depth();
6726 return (nested == preempt_offset);
6729 void __might_sleep(const char *file, int line, int preempt_offset)
6732 * Blocking primitives will set (and therefore destroy) current->state,
6733 * since we will exit with TASK_RUNNING make sure we enter with it,
6734 * otherwise we will destroy state.
6736 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6737 "do not call blocking ops when !TASK_RUNNING; "
6738 "state=%lx set at [<%p>] %pS\n",
6740 (void *)current->task_state_change,
6741 (void *)current->task_state_change);
6743 ___might_sleep(file, line, preempt_offset);
6745 EXPORT_SYMBOL(__might_sleep);
6747 void ___might_sleep(const char *file, int line, int preempt_offset)
6749 /* Ratelimiting timestamp: */
6750 static unsigned long prev_jiffy;
6752 unsigned long preempt_disable_ip;
6754 /* WARN_ON_ONCE() by default, no rate limit required: */
6757 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6758 !is_idle_task(current) && !current->non_block_count) ||
6759 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6763 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6765 prev_jiffy = jiffies;
6767 /* Save this before calling printk(), since that will clobber it: */
6768 preempt_disable_ip = get_preempt_disable_ip(current);
6771 "BUG: sleeping function called from invalid context at %s:%d\n",
6774 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6775 in_atomic(), irqs_disabled(), current->non_block_count,
6776 current->pid, current->comm);
6778 if (task_stack_end_corrupted(current))
6779 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6781 debug_show_held_locks(current);
6782 if (irqs_disabled())
6783 print_irqtrace_events(current);
6784 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6785 && !preempt_count_equals(preempt_offset)) {
6786 pr_err("Preemption disabled at:");
6787 print_ip_sym(preempt_disable_ip);
6791 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6793 EXPORT_SYMBOL(___might_sleep);
6795 void __cant_sleep(const char *file, int line, int preempt_offset)
6797 static unsigned long prev_jiffy;
6799 if (irqs_disabled())
6802 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6805 if (preempt_count() > preempt_offset)
6808 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6810 prev_jiffy = jiffies;
6812 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6813 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6814 in_atomic(), irqs_disabled(),
6815 current->pid, current->comm);
6817 debug_show_held_locks(current);
6819 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6821 EXPORT_SYMBOL_GPL(__cant_sleep);
6824 #ifdef CONFIG_MAGIC_SYSRQ
6825 void normalize_rt_tasks(void)
6827 struct task_struct *g, *p;
6828 struct sched_attr attr = {
6829 .sched_policy = SCHED_NORMAL,
6832 read_lock(&tasklist_lock);
6833 for_each_process_thread(g, p) {
6835 * Only normalize user tasks:
6837 if (p->flags & PF_KTHREAD)
6840 p->se.exec_start = 0;
6841 schedstat_set(p->se.statistics.wait_start, 0);
6842 schedstat_set(p->se.statistics.sleep_start, 0);
6843 schedstat_set(p->se.statistics.block_start, 0);
6845 if (!dl_task(p) && !rt_task(p)) {
6847 * Renice negative nice level userspace
6850 if (task_nice(p) < 0)
6851 set_user_nice(p, 0);
6855 __sched_setscheduler(p, &attr, false, false);
6857 read_unlock(&tasklist_lock);
6860 #endif /* CONFIG_MAGIC_SYSRQ */
6862 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6864 * These functions are only useful for the IA64 MCA handling, or kdb.
6866 * They can only be called when the whole system has been
6867 * stopped - every CPU needs to be quiescent, and no scheduling
6868 * activity can take place. Using them for anything else would
6869 * be a serious bug, and as a result, they aren't even visible
6870 * under any other configuration.
6874 * curr_task - return the current task for a given CPU.
6875 * @cpu: the processor in question.
6877 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6879 * Return: The current task for @cpu.
6881 struct task_struct *curr_task(int cpu)
6883 return cpu_curr(cpu);
6886 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6890 * ia64_set_curr_task - set the current task for a given CPU.
6891 * @cpu: the processor in question.
6892 * @p: the task pointer to set.
6894 * Description: This function must only be used when non-maskable interrupts
6895 * are serviced on a separate stack. It allows the architecture to switch the
6896 * notion of the current task on a CPU in a non-blocking manner. This function
6897 * must be called with all CPU's synchronized, and interrupts disabled, the
6898 * and caller must save the original value of the current task (see
6899 * curr_task() above) and restore that value before reenabling interrupts and
6900 * re-starting the system.
6902 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6904 void ia64_set_curr_task(int cpu, struct task_struct *p)
6911 #ifdef CONFIG_CGROUP_SCHED
6912 /* task_group_lock serializes the addition/removal of task groups */
6913 static DEFINE_SPINLOCK(task_group_lock);
6915 static inline void alloc_uclamp_sched_group(struct task_group *tg,
6916 struct task_group *parent)
6918 #ifdef CONFIG_UCLAMP_TASK_GROUP
6919 enum uclamp_id clamp_id;
6921 for_each_clamp_id(clamp_id) {
6922 uclamp_se_set(&tg->uclamp_req[clamp_id],
6923 uclamp_none(clamp_id), false);
6924 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
6929 static void sched_free_group(struct task_group *tg)
6931 free_fair_sched_group(tg);
6932 free_rt_sched_group(tg);
6934 kmem_cache_free(task_group_cache, tg);
6937 /* allocate runqueue etc for a new task group */
6938 struct task_group *sched_create_group(struct task_group *parent)
6940 struct task_group *tg;
6942 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6944 return ERR_PTR(-ENOMEM);
6946 if (!alloc_fair_sched_group(tg, parent))
6949 if (!alloc_rt_sched_group(tg, parent))
6952 alloc_uclamp_sched_group(tg, parent);
6957 sched_free_group(tg);
6958 return ERR_PTR(-ENOMEM);
6961 void sched_online_group(struct task_group *tg, struct task_group *parent)
6963 unsigned long flags;
6965 spin_lock_irqsave(&task_group_lock, flags);
6966 list_add_rcu(&tg->list, &task_groups);
6968 /* Root should already exist: */
6971 tg->parent = parent;
6972 INIT_LIST_HEAD(&tg->children);
6973 list_add_rcu(&tg->siblings, &parent->children);
6974 spin_unlock_irqrestore(&task_group_lock, flags);
6976 online_fair_sched_group(tg);
6979 /* rcu callback to free various structures associated with a task group */
6980 static void sched_free_group_rcu(struct rcu_head *rhp)
6982 /* Now it should be safe to free those cfs_rqs: */
6983 sched_free_group(container_of(rhp, struct task_group, rcu));
6986 void sched_destroy_group(struct task_group *tg)
6988 /* Wait for possible concurrent references to cfs_rqs complete: */
6989 call_rcu(&tg->rcu, sched_free_group_rcu);
6992 void sched_offline_group(struct task_group *tg)
6994 unsigned long flags;
6996 /* End participation in shares distribution: */
6997 unregister_fair_sched_group(tg);
6999 spin_lock_irqsave(&task_group_lock, flags);
7000 list_del_rcu(&tg->list);
7001 list_del_rcu(&tg->siblings);
7002 spin_unlock_irqrestore(&task_group_lock, flags);
7005 static void sched_change_group(struct task_struct *tsk, int type)
7007 struct task_group *tg;
7010 * All callers are synchronized by task_rq_lock(); we do not use RCU
7011 * which is pointless here. Thus, we pass "true" to task_css_check()
7012 * to prevent lockdep warnings.
7014 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7015 struct task_group, css);
7016 tg = autogroup_task_group(tsk, tg);
7017 tsk->sched_task_group = tg;
7019 #ifdef CONFIG_FAIR_GROUP_SCHED
7020 if (tsk->sched_class->task_change_group)
7021 tsk->sched_class->task_change_group(tsk, type);
7024 set_task_rq(tsk, task_cpu(tsk));
7028 * Change task's runqueue when it moves between groups.
7030 * The caller of this function should have put the task in its new group by
7031 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7034 void sched_move_task(struct task_struct *tsk)
7036 int queued, running, queue_flags =
7037 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7041 rq = task_rq_lock(tsk, &rf);
7042 update_rq_clock(rq);
7044 running = task_current(rq, tsk);
7045 queued = task_on_rq_queued(tsk);
7048 dequeue_task(rq, tsk, queue_flags);
7050 put_prev_task(rq, tsk);
7052 sched_change_group(tsk, TASK_MOVE_GROUP);
7055 enqueue_task(rq, tsk, queue_flags);
7057 set_next_task(rq, tsk);
7059 * After changing group, the running task may have joined a
7060 * throttled one but it's still the running task. Trigger a
7061 * resched to make sure that task can still run.
7066 task_rq_unlock(rq, tsk, &rf);
7069 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7071 return css ? container_of(css, struct task_group, css) : NULL;
7074 static struct cgroup_subsys_state *
7075 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7077 struct task_group *parent = css_tg(parent_css);
7078 struct task_group *tg;
7081 /* This is early initialization for the top cgroup */
7082 return &root_task_group.css;
7085 tg = sched_create_group(parent);
7087 return ERR_PTR(-ENOMEM);
7092 /* Expose task group only after completing cgroup initialization */
7093 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7095 struct task_group *tg = css_tg(css);
7096 struct task_group *parent = css_tg(css->parent);
7099 sched_online_group(tg, parent);
7101 #ifdef CONFIG_UCLAMP_TASK_GROUP
7102 /* Propagate the effective uclamp value for the new group */
7103 cpu_util_update_eff(css);
7109 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7111 struct task_group *tg = css_tg(css);
7113 sched_offline_group(tg);
7116 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7118 struct task_group *tg = css_tg(css);
7121 * Relies on the RCU grace period between css_released() and this.
7123 sched_free_group(tg);
7127 * This is called before wake_up_new_task(), therefore we really only
7128 * have to set its group bits, all the other stuff does not apply.
7130 static void cpu_cgroup_fork(struct task_struct *task)
7135 rq = task_rq_lock(task, &rf);
7137 update_rq_clock(rq);
7138 sched_change_group(task, TASK_SET_GROUP);
7140 task_rq_unlock(rq, task, &rf);
7143 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7145 struct task_struct *task;
7146 struct cgroup_subsys_state *css;
7149 cgroup_taskset_for_each(task, css, tset) {
7150 #ifdef CONFIG_RT_GROUP_SCHED
7151 if (!sched_rt_can_attach(css_tg(css), task))
7155 * Serialize against wake_up_new_task() such that if its
7156 * running, we're sure to observe its full state.
7158 raw_spin_lock_irq(&task->pi_lock);
7160 * Avoid calling sched_move_task() before wake_up_new_task()
7161 * has happened. This would lead to problems with PELT, due to
7162 * move wanting to detach+attach while we're not attached yet.
7164 if (task->state == TASK_NEW)
7166 raw_spin_unlock_irq(&task->pi_lock);
7174 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7176 struct task_struct *task;
7177 struct cgroup_subsys_state *css;
7179 cgroup_taskset_for_each(task, css, tset)
7180 sched_move_task(task);
7183 #ifdef CONFIG_UCLAMP_TASK_GROUP
7184 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7186 struct cgroup_subsys_state *top_css = css;
7187 struct uclamp_se *uc_parent = NULL;
7188 struct uclamp_se *uc_se = NULL;
7189 unsigned int eff[UCLAMP_CNT];
7190 enum uclamp_id clamp_id;
7191 unsigned int clamps;
7193 css_for_each_descendant_pre(css, top_css) {
7194 uc_parent = css_tg(css)->parent
7195 ? css_tg(css)->parent->uclamp : NULL;
7197 for_each_clamp_id(clamp_id) {
7198 /* Assume effective clamps matches requested clamps */
7199 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7200 /* Cap effective clamps with parent's effective clamps */
7202 eff[clamp_id] > uc_parent[clamp_id].value) {
7203 eff[clamp_id] = uc_parent[clamp_id].value;
7206 /* Ensure protection is always capped by limit */
7207 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7209 /* Propagate most restrictive effective clamps */
7211 uc_se = css_tg(css)->uclamp;
7212 for_each_clamp_id(clamp_id) {
7213 if (eff[clamp_id] == uc_se[clamp_id].value)
7215 uc_se[clamp_id].value = eff[clamp_id];
7216 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7217 clamps |= (0x1 << clamp_id);
7220 css = css_rightmost_descendant(css);
7224 /* Immediately update descendants RUNNABLE tasks */
7225 uclamp_update_active_tasks(css, clamps);
7230 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7231 * C expression. Since there is no way to convert a macro argument (N) into a
7232 * character constant, use two levels of macros.
7234 #define _POW10(exp) ((unsigned int)1e##exp)
7235 #define POW10(exp) _POW10(exp)
7237 struct uclamp_request {
7238 #define UCLAMP_PERCENT_SHIFT 2
7239 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7245 static inline struct uclamp_request
7246 capacity_from_percent(char *buf)
7248 struct uclamp_request req = {
7249 .percent = UCLAMP_PERCENT_SCALE,
7250 .util = SCHED_CAPACITY_SCALE,
7255 if (strcmp(buf, "max")) {
7256 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7260 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7265 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7266 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7272 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7273 size_t nbytes, loff_t off,
7274 enum uclamp_id clamp_id)
7276 struct uclamp_request req;
7277 struct task_group *tg;
7279 req = capacity_from_percent(buf);
7283 mutex_lock(&uclamp_mutex);
7286 tg = css_tg(of_css(of));
7287 if (tg->uclamp_req[clamp_id].value != req.util)
7288 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7291 * Because of not recoverable conversion rounding we keep track of the
7292 * exact requested value
7294 tg->uclamp_pct[clamp_id] = req.percent;
7296 /* Update effective clamps to track the most restrictive value */
7297 cpu_util_update_eff(of_css(of));
7300 mutex_unlock(&uclamp_mutex);
7305 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7306 char *buf, size_t nbytes,
7309 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7312 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7313 char *buf, size_t nbytes,
7316 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7319 static inline void cpu_uclamp_print(struct seq_file *sf,
7320 enum uclamp_id clamp_id)
7322 struct task_group *tg;
7328 tg = css_tg(seq_css(sf));
7329 util_clamp = tg->uclamp_req[clamp_id].value;
7332 if (util_clamp == SCHED_CAPACITY_SCALE) {
7333 seq_puts(sf, "max\n");
7337 percent = tg->uclamp_pct[clamp_id];
7338 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7339 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7342 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7344 cpu_uclamp_print(sf, UCLAMP_MIN);
7348 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7350 cpu_uclamp_print(sf, UCLAMP_MAX);
7353 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7355 #ifdef CONFIG_FAIR_GROUP_SCHED
7356 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7357 struct cftype *cftype, u64 shareval)
7359 if (shareval > scale_load_down(ULONG_MAX))
7360 shareval = MAX_SHARES;
7361 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7364 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7367 struct task_group *tg = css_tg(css);
7369 return (u64) scale_load_down(tg->shares);
7372 #ifdef CONFIG_CFS_BANDWIDTH
7373 static DEFINE_MUTEX(cfs_constraints_mutex);
7375 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7376 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7377 /* More than 203 days if BW_SHIFT equals 20. */
7378 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
7380 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7382 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7384 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7385 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7387 if (tg == &root_task_group)
7391 * Ensure we have at some amount of bandwidth every period. This is
7392 * to prevent reaching a state of large arrears when throttled via
7393 * entity_tick() resulting in prolonged exit starvation.
7395 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7399 * Likewise, bound things on the otherside by preventing insane quota
7400 * periods. This also allows us to normalize in computing quota
7403 if (period > max_cfs_quota_period)
7407 * Bound quota to defend quota against overflow during bandwidth shift.
7409 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
7413 * Prevent race between setting of cfs_rq->runtime_enabled and
7414 * unthrottle_offline_cfs_rqs().
7417 mutex_lock(&cfs_constraints_mutex);
7418 ret = __cfs_schedulable(tg, period, quota);
7422 runtime_enabled = quota != RUNTIME_INF;
7423 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7425 * If we need to toggle cfs_bandwidth_used, off->on must occur
7426 * before making related changes, and on->off must occur afterwards
7428 if (runtime_enabled && !runtime_was_enabled)
7429 cfs_bandwidth_usage_inc();
7430 raw_spin_lock_irq(&cfs_b->lock);
7431 cfs_b->period = ns_to_ktime(period);
7432 cfs_b->quota = quota;
7434 __refill_cfs_bandwidth_runtime(cfs_b);
7436 /* Restart the period timer (if active) to handle new period expiry: */
7437 if (runtime_enabled)
7438 start_cfs_bandwidth(cfs_b);
7440 raw_spin_unlock_irq(&cfs_b->lock);
7442 for_each_online_cpu(i) {
7443 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7444 struct rq *rq = cfs_rq->rq;
7447 rq_lock_irq(rq, &rf);
7448 cfs_rq->runtime_enabled = runtime_enabled;
7449 cfs_rq->runtime_remaining = 0;
7451 if (cfs_rq->throttled)
7452 unthrottle_cfs_rq(cfs_rq);
7453 rq_unlock_irq(rq, &rf);
7455 if (runtime_was_enabled && !runtime_enabled)
7456 cfs_bandwidth_usage_dec();
7458 mutex_unlock(&cfs_constraints_mutex);
7464 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7468 period = ktime_to_ns(tg->cfs_bandwidth.period);
7469 if (cfs_quota_us < 0)
7470 quota = RUNTIME_INF;
7471 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7472 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7476 return tg_set_cfs_bandwidth(tg, period, quota);
7479 static long tg_get_cfs_quota(struct task_group *tg)
7483 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7486 quota_us = tg->cfs_bandwidth.quota;
7487 do_div(quota_us, NSEC_PER_USEC);
7492 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7496 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7499 period = (u64)cfs_period_us * NSEC_PER_USEC;
7500 quota = tg->cfs_bandwidth.quota;
7502 return tg_set_cfs_bandwidth(tg, period, quota);
7505 static long tg_get_cfs_period(struct task_group *tg)
7509 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7510 do_div(cfs_period_us, NSEC_PER_USEC);
7512 return cfs_period_us;
7515 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7518 return tg_get_cfs_quota(css_tg(css));
7521 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7522 struct cftype *cftype, s64 cfs_quota_us)
7524 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7527 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7530 return tg_get_cfs_period(css_tg(css));
7533 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7534 struct cftype *cftype, u64 cfs_period_us)
7536 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7539 struct cfs_schedulable_data {
7540 struct task_group *tg;
7545 * normalize group quota/period to be quota/max_period
7546 * note: units are usecs
7548 static u64 normalize_cfs_quota(struct task_group *tg,
7549 struct cfs_schedulable_data *d)
7557 period = tg_get_cfs_period(tg);
7558 quota = tg_get_cfs_quota(tg);
7561 /* note: these should typically be equivalent */
7562 if (quota == RUNTIME_INF || quota == -1)
7565 return to_ratio(period, quota);
7568 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7570 struct cfs_schedulable_data *d = data;
7571 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7572 s64 quota = 0, parent_quota = -1;
7575 quota = RUNTIME_INF;
7577 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7579 quota = normalize_cfs_quota(tg, d);
7580 parent_quota = parent_b->hierarchical_quota;
7583 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7584 * always take the min. On cgroup1, only inherit when no
7587 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7588 quota = min(quota, parent_quota);
7590 if (quota == RUNTIME_INF)
7591 quota = parent_quota;
7592 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7596 cfs_b->hierarchical_quota = quota;
7601 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7604 struct cfs_schedulable_data data = {
7610 if (quota != RUNTIME_INF) {
7611 do_div(data.period, NSEC_PER_USEC);
7612 do_div(data.quota, NSEC_PER_USEC);
7616 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7622 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7624 struct task_group *tg = css_tg(seq_css(sf));
7625 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7627 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7628 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7629 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7631 if (schedstat_enabled() && tg != &root_task_group) {
7635 for_each_possible_cpu(i)
7636 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7638 seq_printf(sf, "wait_sum %llu\n", ws);
7643 #endif /* CONFIG_CFS_BANDWIDTH */
7644 #endif /* CONFIG_FAIR_GROUP_SCHED */
7646 #ifdef CONFIG_RT_GROUP_SCHED
7647 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7648 struct cftype *cft, s64 val)
7650 return sched_group_set_rt_runtime(css_tg(css), val);
7653 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7656 return sched_group_rt_runtime(css_tg(css));
7659 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7660 struct cftype *cftype, u64 rt_period_us)
7662 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7665 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7668 return sched_group_rt_period(css_tg(css));
7670 #endif /* CONFIG_RT_GROUP_SCHED */
7672 static struct cftype cpu_legacy_files[] = {
7673 #ifdef CONFIG_FAIR_GROUP_SCHED
7676 .read_u64 = cpu_shares_read_u64,
7677 .write_u64 = cpu_shares_write_u64,
7680 #ifdef CONFIG_CFS_BANDWIDTH
7682 .name = "cfs_quota_us",
7683 .read_s64 = cpu_cfs_quota_read_s64,
7684 .write_s64 = cpu_cfs_quota_write_s64,
7687 .name = "cfs_period_us",
7688 .read_u64 = cpu_cfs_period_read_u64,
7689 .write_u64 = cpu_cfs_period_write_u64,
7693 .seq_show = cpu_cfs_stat_show,
7696 #ifdef CONFIG_RT_GROUP_SCHED
7698 .name = "rt_runtime_us",
7699 .read_s64 = cpu_rt_runtime_read,
7700 .write_s64 = cpu_rt_runtime_write,
7703 .name = "rt_period_us",
7704 .read_u64 = cpu_rt_period_read_uint,
7705 .write_u64 = cpu_rt_period_write_uint,
7708 #ifdef CONFIG_UCLAMP_TASK_GROUP
7710 .name = "uclamp.min",
7711 .flags = CFTYPE_NOT_ON_ROOT,
7712 .seq_show = cpu_uclamp_min_show,
7713 .write = cpu_uclamp_min_write,
7716 .name = "uclamp.max",
7717 .flags = CFTYPE_NOT_ON_ROOT,
7718 .seq_show = cpu_uclamp_max_show,
7719 .write = cpu_uclamp_max_write,
7725 static int cpu_extra_stat_show(struct seq_file *sf,
7726 struct cgroup_subsys_state *css)
7728 #ifdef CONFIG_CFS_BANDWIDTH
7730 struct task_group *tg = css_tg(css);
7731 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7734 throttled_usec = cfs_b->throttled_time;
7735 do_div(throttled_usec, NSEC_PER_USEC);
7737 seq_printf(sf, "nr_periods %d\n"
7739 "throttled_usec %llu\n",
7740 cfs_b->nr_periods, cfs_b->nr_throttled,
7747 #ifdef CONFIG_FAIR_GROUP_SCHED
7748 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7751 struct task_group *tg = css_tg(css);
7752 u64 weight = scale_load_down(tg->shares);
7754 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7757 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7758 struct cftype *cft, u64 weight)
7761 * cgroup weight knobs should use the common MIN, DFL and MAX
7762 * values which are 1, 100 and 10000 respectively. While it loses
7763 * a bit of range on both ends, it maps pretty well onto the shares
7764 * value used by scheduler and the round-trip conversions preserve
7765 * the original value over the entire range.
7767 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7770 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7772 return sched_group_set_shares(css_tg(css), scale_load(weight));
7775 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7778 unsigned long weight = scale_load_down(css_tg(css)->shares);
7779 int last_delta = INT_MAX;
7782 /* find the closest nice value to the current weight */
7783 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7784 delta = abs(sched_prio_to_weight[prio] - weight);
7785 if (delta >= last_delta)
7790 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7793 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7794 struct cftype *cft, s64 nice)
7796 unsigned long weight;
7799 if (nice < MIN_NICE || nice > MAX_NICE)
7802 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7803 idx = array_index_nospec(idx, 40);
7804 weight = sched_prio_to_weight[idx];
7806 return sched_group_set_shares(css_tg(css), scale_load(weight));
7810 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7811 long period, long quota)
7814 seq_puts(sf, "max");
7816 seq_printf(sf, "%ld", quota);
7818 seq_printf(sf, " %ld\n", period);
7821 /* caller should put the current value in *@periodp before calling */
7822 static int __maybe_unused cpu_period_quota_parse(char *buf,
7823 u64 *periodp, u64 *quotap)
7825 char tok[21]; /* U64_MAX */
7827 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7830 *periodp *= NSEC_PER_USEC;
7832 if (sscanf(tok, "%llu", quotap))
7833 *quotap *= NSEC_PER_USEC;
7834 else if (!strcmp(tok, "max"))
7835 *quotap = RUNTIME_INF;
7842 #ifdef CONFIG_CFS_BANDWIDTH
7843 static int cpu_max_show(struct seq_file *sf, void *v)
7845 struct task_group *tg = css_tg(seq_css(sf));
7847 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7851 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7852 char *buf, size_t nbytes, loff_t off)
7854 struct task_group *tg = css_tg(of_css(of));
7855 u64 period = tg_get_cfs_period(tg);
7859 ret = cpu_period_quota_parse(buf, &period, "a);
7861 ret = tg_set_cfs_bandwidth(tg, period, quota);
7862 return ret ?: nbytes;
7866 static struct cftype cpu_files[] = {
7867 #ifdef CONFIG_FAIR_GROUP_SCHED
7870 .flags = CFTYPE_NOT_ON_ROOT,
7871 .read_u64 = cpu_weight_read_u64,
7872 .write_u64 = cpu_weight_write_u64,
7875 .name = "weight.nice",
7876 .flags = CFTYPE_NOT_ON_ROOT,
7877 .read_s64 = cpu_weight_nice_read_s64,
7878 .write_s64 = cpu_weight_nice_write_s64,
7881 #ifdef CONFIG_CFS_BANDWIDTH
7884 .flags = CFTYPE_NOT_ON_ROOT,
7885 .seq_show = cpu_max_show,
7886 .write = cpu_max_write,
7889 #ifdef CONFIG_UCLAMP_TASK_GROUP
7891 .name = "uclamp.min",
7892 .flags = CFTYPE_NOT_ON_ROOT,
7893 .seq_show = cpu_uclamp_min_show,
7894 .write = cpu_uclamp_min_write,
7897 .name = "uclamp.max",
7898 .flags = CFTYPE_NOT_ON_ROOT,
7899 .seq_show = cpu_uclamp_max_show,
7900 .write = cpu_uclamp_max_write,
7906 struct cgroup_subsys cpu_cgrp_subsys = {
7907 .css_alloc = cpu_cgroup_css_alloc,
7908 .css_online = cpu_cgroup_css_online,
7909 .css_released = cpu_cgroup_css_released,
7910 .css_free = cpu_cgroup_css_free,
7911 .css_extra_stat_show = cpu_extra_stat_show,
7912 .fork = cpu_cgroup_fork,
7913 .can_attach = cpu_cgroup_can_attach,
7914 .attach = cpu_cgroup_attach,
7915 .legacy_cftypes = cpu_legacy_files,
7916 .dfl_cftypes = cpu_files,
7921 #endif /* CONFIG_CGROUP_SCHED */
7923 void dump_cpu_task(int cpu)
7925 pr_info("Task dump for CPU %d:\n", cpu);
7926 sched_show_task(cpu_curr(cpu));
7930 * Nice levels are multiplicative, with a gentle 10% change for every
7931 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7932 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7933 * that remained on nice 0.
7935 * The "10% effect" is relative and cumulative: from _any_ nice level,
7936 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7937 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7938 * If a task goes up by ~10% and another task goes down by ~10% then
7939 * the relative distance between them is ~25%.)
7941 const int sched_prio_to_weight[40] = {
7942 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7943 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7944 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7945 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7946 /* 0 */ 1024, 820, 655, 526, 423,
7947 /* 5 */ 335, 272, 215, 172, 137,
7948 /* 10 */ 110, 87, 70, 56, 45,
7949 /* 15 */ 36, 29, 23, 18, 15,
7953 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7955 * In cases where the weight does not change often, we can use the
7956 * precalculated inverse to speed up arithmetics by turning divisions
7957 * into multiplications:
7959 const u32 sched_prio_to_wmult[40] = {
7960 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7961 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7962 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7963 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7964 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7965 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7966 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7967 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7970 #undef CREATE_TRACE_POINTS