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 "../../fs/io-wq.h"
20 #include "../smpboot.h"
24 #define CREATE_TRACE_POINTS
25 #include <trace/events/sched.h>
28 * Export tracepoints that act as a bare tracehook (ie: have no trace event
29 * associated with them) to allow external modules to probe them.
31 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
32 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
38 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
40 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_JUMP_LABEL)
42 * Debugging: various feature bits
44 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
45 * sysctl_sched_features, defined in sched.h, to allow constants propagation
46 * at compile time and compiler optimization based on features default.
48 #define SCHED_FEAT(name, enabled) \
49 (1UL << __SCHED_FEAT_##name) * enabled |
50 const_debug unsigned int sysctl_sched_features =
57 * Number of tasks to iterate in a single balance run.
58 * Limited because this is done with IRQs disabled.
60 const_debug unsigned int sysctl_sched_nr_migrate = 32;
63 * period over which we measure -rt task CPU usage in us.
66 unsigned int sysctl_sched_rt_period = 1000000;
68 __read_mostly int scheduler_running;
71 * part of the period that we allow rt tasks to run in us.
74 int sysctl_sched_rt_runtime = 950000;
77 * __task_rq_lock - lock the rq @p resides on.
79 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
84 lockdep_assert_held(&p->pi_lock);
88 raw_spin_lock(&rq->lock);
89 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
93 raw_spin_unlock(&rq->lock);
95 while (unlikely(task_on_rq_migrating(p)))
101 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
103 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
104 __acquires(p->pi_lock)
110 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
112 raw_spin_lock(&rq->lock);
114 * move_queued_task() task_rq_lock()
117 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
118 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
119 * [S] ->cpu = new_cpu [L] task_rq()
123 * If we observe the old CPU in task_rq_lock(), the acquire of
124 * the old rq->lock will fully serialize against the stores.
126 * If we observe the new CPU in task_rq_lock(), the address
127 * dependency headed by '[L] rq = task_rq()' and the acquire
128 * will pair with the WMB to ensure we then also see migrating.
130 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
134 raw_spin_unlock(&rq->lock);
135 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
137 while (unlikely(task_on_rq_migrating(p)))
143 * RQ-clock updating methods:
146 static void update_rq_clock_task(struct rq *rq, s64 delta)
149 * In theory, the compile should just see 0 here, and optimize out the call
150 * to sched_rt_avg_update. But I don't trust it...
152 s64 __maybe_unused steal = 0, irq_delta = 0;
154 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
155 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
158 * Since irq_time is only updated on {soft,}irq_exit, we might run into
159 * this case when a previous update_rq_clock() happened inside a
162 * When this happens, we stop ->clock_task and only update the
163 * prev_irq_time stamp to account for the part that fit, so that a next
164 * update will consume the rest. This ensures ->clock_task is
167 * It does however cause some slight miss-attribution of {soft,}irq
168 * time, a more accurate solution would be to update the irq_time using
169 * the current rq->clock timestamp, except that would require using
172 if (irq_delta > delta)
175 rq->prev_irq_time += irq_delta;
178 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
179 if (static_key_false((¶virt_steal_rq_enabled))) {
180 steal = paravirt_steal_clock(cpu_of(rq));
181 steal -= rq->prev_steal_time_rq;
183 if (unlikely(steal > delta))
186 rq->prev_steal_time_rq += steal;
191 rq->clock_task += delta;
193 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
194 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
195 update_irq_load_avg(rq, irq_delta + steal);
197 update_rq_clock_pelt(rq, delta);
200 void update_rq_clock(struct rq *rq)
204 lockdep_assert_held(&rq->lock);
206 if (rq->clock_update_flags & RQCF_ACT_SKIP)
209 #ifdef CONFIG_SCHED_DEBUG
210 if (sched_feat(WARN_DOUBLE_CLOCK))
211 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
212 rq->clock_update_flags |= RQCF_UPDATED;
215 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
219 update_rq_clock_task(rq, delta);
223 #ifdef CONFIG_SCHED_HRTICK
225 * Use HR-timers to deliver accurate preemption points.
228 static void hrtick_clear(struct rq *rq)
230 if (hrtimer_active(&rq->hrtick_timer))
231 hrtimer_cancel(&rq->hrtick_timer);
235 * High-resolution timer tick.
236 * Runs from hardirq context with interrupts disabled.
238 static enum hrtimer_restart hrtick(struct hrtimer *timer)
240 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
243 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
247 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
250 return HRTIMER_NORESTART;
255 static void __hrtick_restart(struct rq *rq)
257 struct hrtimer *timer = &rq->hrtick_timer;
259 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
263 * called from hardirq (IPI) context
265 static void __hrtick_start(void *arg)
271 __hrtick_restart(rq);
272 rq->hrtick_csd_pending = 0;
277 * Called to set the hrtick timer state.
279 * called with rq->lock held and irqs disabled
281 void hrtick_start(struct rq *rq, u64 delay)
283 struct hrtimer *timer = &rq->hrtick_timer;
288 * Don't schedule slices shorter than 10000ns, that just
289 * doesn't make sense and can cause timer DoS.
291 delta = max_t(s64, delay, 10000LL);
292 time = ktime_add_ns(timer->base->get_time(), delta);
294 hrtimer_set_expires(timer, time);
296 if (rq == this_rq()) {
297 __hrtick_restart(rq);
298 } else if (!rq->hrtick_csd_pending) {
299 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
300 rq->hrtick_csd_pending = 1;
306 * Called to set the hrtick timer state.
308 * called with rq->lock held and irqs disabled
310 void hrtick_start(struct rq *rq, u64 delay)
313 * Don't schedule slices shorter than 10000ns, that just
314 * doesn't make sense. Rely on vruntime for fairness.
316 delay = max_t(u64, delay, 10000LL);
317 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
318 HRTIMER_MODE_REL_PINNED_HARD);
320 #endif /* CONFIG_SMP */
322 static void hrtick_rq_init(struct rq *rq)
325 rq->hrtick_csd_pending = 0;
327 rq->hrtick_csd.flags = 0;
328 rq->hrtick_csd.func = __hrtick_start;
329 rq->hrtick_csd.info = rq;
332 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
333 rq->hrtick_timer.function = hrtick;
335 #else /* CONFIG_SCHED_HRTICK */
336 static inline void hrtick_clear(struct rq *rq)
340 static inline void hrtick_rq_init(struct rq *rq)
343 #endif /* CONFIG_SCHED_HRTICK */
346 * cmpxchg based fetch_or, macro so it works for different integer types
348 #define fetch_or(ptr, mask) \
350 typeof(ptr) _ptr = (ptr); \
351 typeof(mask) _mask = (mask); \
352 typeof(*_ptr) _old, _val = *_ptr; \
355 _old = cmpxchg(_ptr, _val, _val | _mask); \
363 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
365 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
366 * this avoids any races wrt polling state changes and thereby avoids
369 static bool set_nr_and_not_polling(struct task_struct *p)
371 struct thread_info *ti = task_thread_info(p);
372 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
376 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
378 * If this returns true, then the idle task promises to call
379 * sched_ttwu_pending() and reschedule soon.
381 static bool set_nr_if_polling(struct task_struct *p)
383 struct thread_info *ti = task_thread_info(p);
384 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
387 if (!(val & _TIF_POLLING_NRFLAG))
389 if (val & _TIF_NEED_RESCHED)
391 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
400 static bool set_nr_and_not_polling(struct task_struct *p)
402 set_tsk_need_resched(p);
407 static bool set_nr_if_polling(struct task_struct *p)
414 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
416 struct wake_q_node *node = &task->wake_q;
419 * Atomically grab the task, if ->wake_q is !nil already it means
420 * its already queued (either by us or someone else) and will get the
421 * wakeup due to that.
423 * In order to ensure that a pending wakeup will observe our pending
424 * state, even in the failed case, an explicit smp_mb() must be used.
426 smp_mb__before_atomic();
427 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
431 * The head is context local, there can be no concurrency.
434 head->lastp = &node->next;
439 * wake_q_add() - queue a wakeup for 'later' waking.
440 * @head: the wake_q_head to add @task to
441 * @task: the task to queue for 'later' wakeup
443 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
444 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
447 * This function must be used as-if it were wake_up_process(); IOW the task
448 * must be ready to be woken at this location.
450 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
452 if (__wake_q_add(head, task))
453 get_task_struct(task);
457 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
458 * @head: the wake_q_head to add @task to
459 * @task: the task to queue for 'later' wakeup
461 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
462 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
465 * This function must be used as-if it were wake_up_process(); IOW the task
466 * must be ready to be woken at this location.
468 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
469 * that already hold reference to @task can call the 'safe' version and trust
470 * wake_q to do the right thing depending whether or not the @task is already
473 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
475 if (!__wake_q_add(head, task))
476 put_task_struct(task);
479 void wake_up_q(struct wake_q_head *head)
481 struct wake_q_node *node = head->first;
483 while (node != WAKE_Q_TAIL) {
484 struct task_struct *task;
486 task = container_of(node, struct task_struct, wake_q);
488 /* Task can safely be re-inserted now: */
490 task->wake_q.next = NULL;
493 * wake_up_process() executes a full barrier, which pairs with
494 * the queueing in wake_q_add() so as not to miss wakeups.
496 wake_up_process(task);
497 put_task_struct(task);
502 * resched_curr - mark rq's current task 'to be rescheduled now'.
504 * On UP this means the setting of the need_resched flag, on SMP it
505 * might also involve a cross-CPU call to trigger the scheduler on
508 void resched_curr(struct rq *rq)
510 struct task_struct *curr = rq->curr;
513 lockdep_assert_held(&rq->lock);
515 if (test_tsk_need_resched(curr))
520 if (cpu == smp_processor_id()) {
521 set_tsk_need_resched(curr);
522 set_preempt_need_resched();
526 if (set_nr_and_not_polling(curr))
527 smp_send_reschedule(cpu);
529 trace_sched_wake_idle_without_ipi(cpu);
532 void resched_cpu(int cpu)
534 struct rq *rq = cpu_rq(cpu);
537 raw_spin_lock_irqsave(&rq->lock, flags);
538 if (cpu_online(cpu) || cpu == smp_processor_id())
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
544 #ifdef CONFIG_NO_HZ_COMMON
546 * In the semi idle case, use the nearest busy CPU for migrating timers
547 * from an idle CPU. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle CPU will add more delays to the timers than intended
551 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int i, cpu = smp_processor_id(), default_cpu = -1;
556 struct sched_domain *sd;
558 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
565 for_each_domain(cpu, sd) {
566 for_each_cpu_and(i, sched_domain_span(sd),
567 housekeeping_cpumask(HK_FLAG_TIMER)) {
578 if (default_cpu == -1)
579 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
587 * When add_timer_on() enqueues a timer into the timer wheel of an
588 * idle CPU then this timer might expire before the next timer event
589 * which is scheduled to wake up that CPU. In case of a completely
590 * idle system the next event might even be infinite time into the
591 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
592 * leaves the inner idle loop so the newly added timer is taken into
593 * account when the CPU goes back to idle and evaluates the timer
594 * wheel for the next timer event.
596 static void wake_up_idle_cpu(int cpu)
598 struct rq *rq = cpu_rq(cpu);
600 if (cpu == smp_processor_id())
603 if (set_nr_and_not_polling(rq->idle))
604 smp_send_reschedule(cpu);
606 trace_sched_wake_idle_without_ipi(cpu);
609 static bool wake_up_full_nohz_cpu(int cpu)
612 * We just need the target to call irq_exit() and re-evaluate
613 * the next tick. The nohz full kick at least implies that.
614 * If needed we can still optimize that later with an
617 if (cpu_is_offline(cpu))
618 return true; /* Don't try to wake offline CPUs. */
619 if (tick_nohz_full_cpu(cpu)) {
620 if (cpu != smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 tick_nohz_full_kick_cpu(cpu);
630 * Wake up the specified CPU. If the CPU is going offline, it is the
631 * caller's responsibility to deal with the lost wakeup, for example,
632 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
634 void wake_up_nohz_cpu(int cpu)
636 if (!wake_up_full_nohz_cpu(cpu))
637 wake_up_idle_cpu(cpu);
640 static inline bool got_nohz_idle_kick(void)
642 int cpu = smp_processor_id();
644 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
647 if (idle_cpu(cpu) && !need_resched())
651 * We can't run Idle Load Balance on this CPU for this time so we
652 * cancel it and clear NOHZ_BALANCE_KICK
654 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
658 #else /* CONFIG_NO_HZ_COMMON */
660 static inline bool got_nohz_idle_kick(void)
665 #endif /* CONFIG_NO_HZ_COMMON */
667 #ifdef CONFIG_NO_HZ_FULL
668 bool sched_can_stop_tick(struct rq *rq)
672 /* Deadline tasks, even if single, need the tick */
673 if (rq->dl.dl_nr_running)
677 * If there are more than one RR tasks, we need the tick to effect the
678 * actual RR behaviour.
680 if (rq->rt.rr_nr_running) {
681 if (rq->rt.rr_nr_running == 1)
688 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
689 * forced preemption between FIFO tasks.
691 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
696 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
697 * if there's more than one we need the tick for involuntary
700 if (rq->nr_running > 1)
705 #endif /* CONFIG_NO_HZ_FULL */
706 #endif /* CONFIG_SMP */
708 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
709 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
711 * Iterate task_group tree rooted at *from, calling @down when first entering a
712 * node and @up when leaving it for the final time.
714 * Caller must hold rcu_lock or sufficient equivalent.
716 int walk_tg_tree_from(struct task_group *from,
717 tg_visitor down, tg_visitor up, void *data)
719 struct task_group *parent, *child;
725 ret = (*down)(parent, data);
728 list_for_each_entry_rcu(child, &parent->children, siblings) {
735 ret = (*up)(parent, data);
736 if (ret || parent == from)
740 parent = parent->parent;
747 int tg_nop(struct task_group *tg, void *data)
753 static void set_load_weight(struct task_struct *p, bool update_load)
755 int prio = p->static_prio - MAX_RT_PRIO;
756 struct load_weight *load = &p->se.load;
759 * SCHED_IDLE tasks get minimal weight:
761 if (task_has_idle_policy(p)) {
762 load->weight = scale_load(WEIGHT_IDLEPRIO);
763 load->inv_weight = WMULT_IDLEPRIO;
768 * SCHED_OTHER tasks have to update their load when changing their
771 if (update_load && p->sched_class == &fair_sched_class) {
772 reweight_task(p, prio);
774 load->weight = scale_load(sched_prio_to_weight[prio]);
775 load->inv_weight = sched_prio_to_wmult[prio];
779 #ifdef CONFIG_UCLAMP_TASK
781 * Serializes updates of utilization clamp values
783 * The (slow-path) user-space triggers utilization clamp value updates which
784 * can require updates on (fast-path) scheduler's data structures used to
785 * support enqueue/dequeue operations.
786 * While the per-CPU rq lock protects fast-path update operations, user-space
787 * requests are serialized using a mutex to reduce the risk of conflicting
788 * updates or API abuses.
790 static DEFINE_MUTEX(uclamp_mutex);
792 /* Max allowed minimum utilization */
793 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
795 /* Max allowed maximum utilization */
796 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
798 /* All clamps are required to be less or equal than these values */
799 static struct uclamp_se uclamp_default[UCLAMP_CNT];
801 /* Integer rounded range for each bucket */
802 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
804 #define for_each_clamp_id(clamp_id) \
805 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
807 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
809 return clamp_value / UCLAMP_BUCKET_DELTA;
812 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
814 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
817 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
819 if (clamp_id == UCLAMP_MIN)
821 return SCHED_CAPACITY_SCALE;
824 static inline void uclamp_se_set(struct uclamp_se *uc_se,
825 unsigned int value, bool user_defined)
827 uc_se->value = value;
828 uc_se->bucket_id = uclamp_bucket_id(value);
829 uc_se->user_defined = user_defined;
832 static inline unsigned int
833 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
834 unsigned int clamp_value)
837 * Avoid blocked utilization pushing up the frequency when we go
838 * idle (which drops the max-clamp) by retaining the last known
841 if (clamp_id == UCLAMP_MAX) {
842 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
846 return uclamp_none(UCLAMP_MIN);
849 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
850 unsigned int clamp_value)
852 /* Reset max-clamp retention only on idle exit */
853 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
856 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
860 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
861 unsigned int clamp_value)
863 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
864 int bucket_id = UCLAMP_BUCKETS - 1;
867 * Since both min and max clamps are max aggregated, find the
868 * top most bucket with tasks in.
870 for ( ; bucket_id >= 0; bucket_id--) {
871 if (!bucket[bucket_id].tasks)
873 return bucket[bucket_id].value;
876 /* No tasks -- default clamp values */
877 return uclamp_idle_value(rq, clamp_id, clamp_value);
880 static inline struct uclamp_se
881 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
883 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
884 #ifdef CONFIG_UCLAMP_TASK_GROUP
885 struct uclamp_se uc_max;
888 * Tasks in autogroups or root task group will be
889 * restricted by system defaults.
891 if (task_group_is_autogroup(task_group(p)))
893 if (task_group(p) == &root_task_group)
896 uc_max = task_group(p)->uclamp[clamp_id];
897 if (uc_req.value > uc_max.value || !uc_req.user_defined)
905 * The effective clamp bucket index of a task depends on, by increasing
907 * - the task specific clamp value, when explicitly requested from userspace
908 * - the task group effective clamp value, for tasks not either in the root
909 * group or in an autogroup
910 * - the system default clamp value, defined by the sysadmin
912 static inline struct uclamp_se
913 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
915 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
916 struct uclamp_se uc_max = uclamp_default[clamp_id];
918 /* System default restrictions always apply */
919 if (unlikely(uc_req.value > uc_max.value))
925 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
927 struct uclamp_se uc_eff;
929 /* Task currently refcounted: use back-annotated (effective) value */
930 if (p->uclamp[clamp_id].active)
931 return (unsigned long)p->uclamp[clamp_id].value;
933 uc_eff = uclamp_eff_get(p, clamp_id);
935 return (unsigned long)uc_eff.value;
939 * When a task is enqueued on a rq, the clamp bucket currently defined by the
940 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
941 * updates the rq's clamp value if required.
943 * Tasks can have a task-specific value requested from user-space, track
944 * within each bucket the maximum value for tasks refcounted in it.
945 * This "local max aggregation" allows to track the exact "requested" value
946 * for each bucket when all its RUNNABLE tasks require the same clamp.
948 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
949 enum uclamp_id clamp_id)
951 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
952 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
953 struct uclamp_bucket *bucket;
955 lockdep_assert_held(&rq->lock);
957 /* Update task effective clamp */
958 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
960 bucket = &uc_rq->bucket[uc_se->bucket_id];
962 uc_se->active = true;
964 uclamp_idle_reset(rq, clamp_id, uc_se->value);
967 * Local max aggregation: rq buckets always track the max
968 * "requested" clamp value of its RUNNABLE tasks.
970 if (bucket->tasks == 1 || uc_se->value > bucket->value)
971 bucket->value = uc_se->value;
973 if (uc_se->value > READ_ONCE(uc_rq->value))
974 WRITE_ONCE(uc_rq->value, uc_se->value);
978 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
979 * is released. If this is the last task reference counting the rq's max
980 * active clamp value, then the rq's clamp value is updated.
982 * Both refcounted tasks and rq's cached clamp values are expected to be
983 * always valid. If it's detected they are not, as defensive programming,
984 * enforce the expected state and warn.
986 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
987 enum uclamp_id clamp_id)
989 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
990 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
991 struct uclamp_bucket *bucket;
992 unsigned int bkt_clamp;
993 unsigned int rq_clamp;
995 lockdep_assert_held(&rq->lock);
997 bucket = &uc_rq->bucket[uc_se->bucket_id];
998 SCHED_WARN_ON(!bucket->tasks);
999 if (likely(bucket->tasks))
1001 uc_se->active = false;
1004 * Keep "local max aggregation" simple and accept to (possibly)
1005 * overboost some RUNNABLE tasks in the same bucket.
1006 * The rq clamp bucket value is reset to its base value whenever
1007 * there are no more RUNNABLE tasks refcounting it.
1009 if (likely(bucket->tasks))
1012 rq_clamp = READ_ONCE(uc_rq->value);
1014 * Defensive programming: this should never happen. If it happens,
1015 * e.g. due to future modification, warn and fixup the expected value.
1017 SCHED_WARN_ON(bucket->value > rq_clamp);
1018 if (bucket->value >= rq_clamp) {
1019 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1020 WRITE_ONCE(uc_rq->value, bkt_clamp);
1024 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1026 enum uclamp_id clamp_id;
1028 if (unlikely(!p->sched_class->uclamp_enabled))
1031 for_each_clamp_id(clamp_id)
1032 uclamp_rq_inc_id(rq, p, clamp_id);
1034 /* Reset clamp idle holding when there is one RUNNABLE task */
1035 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1036 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1039 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1041 enum uclamp_id clamp_id;
1043 if (unlikely(!p->sched_class->uclamp_enabled))
1046 for_each_clamp_id(clamp_id)
1047 uclamp_rq_dec_id(rq, p, clamp_id);
1051 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1057 * Lock the task and the rq where the task is (or was) queued.
1059 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1060 * price to pay to safely serialize util_{min,max} updates with
1061 * enqueues, dequeues and migration operations.
1062 * This is the same locking schema used by __set_cpus_allowed_ptr().
1064 rq = task_rq_lock(p, &rf);
1067 * Setting the clamp bucket is serialized by task_rq_lock().
1068 * If the task is not yet RUNNABLE and its task_struct is not
1069 * affecting a valid clamp bucket, the next time it's enqueued,
1070 * it will already see the updated clamp bucket value.
1072 if (p->uclamp[clamp_id].active) {
1073 uclamp_rq_dec_id(rq, p, clamp_id);
1074 uclamp_rq_inc_id(rq, p, clamp_id);
1077 task_rq_unlock(rq, p, &rf);
1080 #ifdef CONFIG_UCLAMP_TASK_GROUP
1082 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1083 unsigned int clamps)
1085 enum uclamp_id clamp_id;
1086 struct css_task_iter it;
1087 struct task_struct *p;
1089 css_task_iter_start(css, 0, &it);
1090 while ((p = css_task_iter_next(&it))) {
1091 for_each_clamp_id(clamp_id) {
1092 if ((0x1 << clamp_id) & clamps)
1093 uclamp_update_active(p, clamp_id);
1096 css_task_iter_end(&it);
1099 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1100 static void uclamp_update_root_tg(void)
1102 struct task_group *tg = &root_task_group;
1104 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1105 sysctl_sched_uclamp_util_min, false);
1106 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1107 sysctl_sched_uclamp_util_max, false);
1110 cpu_util_update_eff(&root_task_group.css);
1114 static void uclamp_update_root_tg(void) { }
1117 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1118 void __user *buffer, size_t *lenp,
1121 bool update_root_tg = false;
1122 int old_min, old_max;
1125 mutex_lock(&uclamp_mutex);
1126 old_min = sysctl_sched_uclamp_util_min;
1127 old_max = sysctl_sched_uclamp_util_max;
1129 result = proc_dointvec(table, write, buffer, lenp, ppos);
1135 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1136 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1141 if (old_min != sysctl_sched_uclamp_util_min) {
1142 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1143 sysctl_sched_uclamp_util_min, false);
1144 update_root_tg = true;
1146 if (old_max != sysctl_sched_uclamp_util_max) {
1147 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1148 sysctl_sched_uclamp_util_max, false);
1149 update_root_tg = true;
1153 uclamp_update_root_tg();
1156 * We update all RUNNABLE tasks only when task groups are in use.
1157 * Otherwise, keep it simple and do just a lazy update at each next
1158 * task enqueue time.
1164 sysctl_sched_uclamp_util_min = old_min;
1165 sysctl_sched_uclamp_util_max = old_max;
1167 mutex_unlock(&uclamp_mutex);
1172 static int uclamp_validate(struct task_struct *p,
1173 const struct sched_attr *attr)
1175 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1176 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1178 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1179 lower_bound = attr->sched_util_min;
1180 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1181 upper_bound = attr->sched_util_max;
1183 if (lower_bound > upper_bound)
1185 if (upper_bound > SCHED_CAPACITY_SCALE)
1191 static void __setscheduler_uclamp(struct task_struct *p,
1192 const struct sched_attr *attr)
1194 enum uclamp_id clamp_id;
1197 * On scheduling class change, reset to default clamps for tasks
1198 * without a task-specific value.
1200 for_each_clamp_id(clamp_id) {
1201 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1202 unsigned int clamp_value = uclamp_none(clamp_id);
1204 /* Keep using defined clamps across class changes */
1205 if (uc_se->user_defined)
1208 /* By default, RT tasks always get 100% boost */
1209 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1210 clamp_value = uclamp_none(UCLAMP_MAX);
1212 uclamp_se_set(uc_se, clamp_value, false);
1215 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1218 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1219 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1220 attr->sched_util_min, true);
1223 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1224 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1225 attr->sched_util_max, true);
1229 static void uclamp_fork(struct task_struct *p)
1231 enum uclamp_id clamp_id;
1233 for_each_clamp_id(clamp_id)
1234 p->uclamp[clamp_id].active = false;
1236 if (likely(!p->sched_reset_on_fork))
1239 for_each_clamp_id(clamp_id) {
1240 unsigned int clamp_value = uclamp_none(clamp_id);
1242 /* By default, RT tasks always get 100% boost */
1243 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1244 clamp_value = uclamp_none(UCLAMP_MAX);
1246 uclamp_se_set(&p->uclamp_req[clamp_id], clamp_value, false);
1250 static void __init init_uclamp(void)
1252 struct uclamp_se uc_max = {};
1253 enum uclamp_id clamp_id;
1256 mutex_init(&uclamp_mutex);
1258 for_each_possible_cpu(cpu) {
1259 memset(&cpu_rq(cpu)->uclamp, 0,
1260 sizeof(struct uclamp_rq)*UCLAMP_CNT);
1261 cpu_rq(cpu)->uclamp_flags = 0;
1264 for_each_clamp_id(clamp_id) {
1265 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1266 uclamp_none(clamp_id), false);
1269 /* System defaults allow max clamp values for both indexes */
1270 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1271 for_each_clamp_id(clamp_id) {
1272 uclamp_default[clamp_id] = uc_max;
1273 #ifdef CONFIG_UCLAMP_TASK_GROUP
1274 root_task_group.uclamp_req[clamp_id] = uc_max;
1275 root_task_group.uclamp[clamp_id] = uc_max;
1280 #else /* CONFIG_UCLAMP_TASK */
1281 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1282 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1283 static inline int uclamp_validate(struct task_struct *p,
1284 const struct sched_attr *attr)
1288 static void __setscheduler_uclamp(struct task_struct *p,
1289 const struct sched_attr *attr) { }
1290 static inline void uclamp_fork(struct task_struct *p) { }
1291 static inline void init_uclamp(void) { }
1292 #endif /* CONFIG_UCLAMP_TASK */
1294 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1296 if (!(flags & ENQUEUE_NOCLOCK))
1297 update_rq_clock(rq);
1299 if (!(flags & ENQUEUE_RESTORE)) {
1300 sched_info_queued(rq, p);
1301 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1304 uclamp_rq_inc(rq, p);
1305 p->sched_class->enqueue_task(rq, p, flags);
1308 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1310 if (!(flags & DEQUEUE_NOCLOCK))
1311 update_rq_clock(rq);
1313 if (!(flags & DEQUEUE_SAVE)) {
1314 sched_info_dequeued(rq, p);
1315 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1318 uclamp_rq_dec(rq, p);
1319 p->sched_class->dequeue_task(rq, p, flags);
1322 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1324 if (task_contributes_to_load(p))
1325 rq->nr_uninterruptible--;
1327 enqueue_task(rq, p, flags);
1329 p->on_rq = TASK_ON_RQ_QUEUED;
1332 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1334 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1336 if (task_contributes_to_load(p))
1337 rq->nr_uninterruptible++;
1339 dequeue_task(rq, p, flags);
1343 * __normal_prio - return the priority that is based on the static prio
1345 static inline int __normal_prio(struct task_struct *p)
1347 return p->static_prio;
1351 * Calculate the expected normal priority: i.e. priority
1352 * without taking RT-inheritance into account. Might be
1353 * boosted by interactivity modifiers. Changes upon fork,
1354 * setprio syscalls, and whenever the interactivity
1355 * estimator recalculates.
1357 static inline int normal_prio(struct task_struct *p)
1361 if (task_has_dl_policy(p))
1362 prio = MAX_DL_PRIO-1;
1363 else if (task_has_rt_policy(p))
1364 prio = MAX_RT_PRIO-1 - p->rt_priority;
1366 prio = __normal_prio(p);
1371 * Calculate the current priority, i.e. the priority
1372 * taken into account by the scheduler. This value might
1373 * be boosted by RT tasks, or might be boosted by
1374 * interactivity modifiers. Will be RT if the task got
1375 * RT-boosted. If not then it returns p->normal_prio.
1377 static int effective_prio(struct task_struct *p)
1379 p->normal_prio = normal_prio(p);
1381 * If we are RT tasks or we were boosted to RT priority,
1382 * keep the priority unchanged. Otherwise, update priority
1383 * to the normal priority:
1385 if (!rt_prio(p->prio))
1386 return p->normal_prio;
1391 * task_curr - is this task currently executing on a CPU?
1392 * @p: the task in question.
1394 * Return: 1 if the task is currently executing. 0 otherwise.
1396 inline int task_curr(const struct task_struct *p)
1398 return cpu_curr(task_cpu(p)) == p;
1402 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1403 * use the balance_callback list if you want balancing.
1405 * this means any call to check_class_changed() must be followed by a call to
1406 * balance_callback().
1408 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1409 const struct sched_class *prev_class,
1412 if (prev_class != p->sched_class) {
1413 if (prev_class->switched_from)
1414 prev_class->switched_from(rq, p);
1416 p->sched_class->switched_to(rq, p);
1417 } else if (oldprio != p->prio || dl_task(p))
1418 p->sched_class->prio_changed(rq, p, oldprio);
1421 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1423 const struct sched_class *class;
1425 if (p->sched_class == rq->curr->sched_class) {
1426 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1428 for_each_class(class) {
1429 if (class == rq->curr->sched_class)
1431 if (class == p->sched_class) {
1439 * A queue event has occurred, and we're going to schedule. In
1440 * this case, we can save a useless back to back clock update.
1442 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1443 rq_clock_skip_update(rq);
1449 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1450 * __set_cpus_allowed_ptr() and select_fallback_rq().
1452 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1454 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1457 if (is_per_cpu_kthread(p))
1458 return cpu_online(cpu);
1460 return cpu_active(cpu);
1464 * This is how migration works:
1466 * 1) we invoke migration_cpu_stop() on the target CPU using
1468 * 2) stopper starts to run (implicitly forcing the migrated thread
1470 * 3) it checks whether the migrated task is still in the wrong runqueue.
1471 * 4) if it's in the wrong runqueue then the migration thread removes
1472 * it and puts it into the right queue.
1473 * 5) stopper completes and stop_one_cpu() returns and the migration
1478 * move_queued_task - move a queued task to new rq.
1480 * Returns (locked) new rq. Old rq's lock is released.
1482 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1483 struct task_struct *p, int new_cpu)
1485 lockdep_assert_held(&rq->lock);
1487 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1488 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1489 set_task_cpu(p, new_cpu);
1492 rq = cpu_rq(new_cpu);
1495 BUG_ON(task_cpu(p) != new_cpu);
1496 enqueue_task(rq, p, 0);
1497 p->on_rq = TASK_ON_RQ_QUEUED;
1498 check_preempt_curr(rq, p, 0);
1503 struct migration_arg {
1504 struct task_struct *task;
1509 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1510 * this because either it can't run here any more (set_cpus_allowed()
1511 * away from this CPU, or CPU going down), or because we're
1512 * attempting to rebalance this task on exec (sched_exec).
1514 * So we race with normal scheduler movements, but that's OK, as long
1515 * as the task is no longer on this CPU.
1517 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1518 struct task_struct *p, int dest_cpu)
1520 /* Affinity changed (again). */
1521 if (!is_cpu_allowed(p, dest_cpu))
1524 update_rq_clock(rq);
1525 rq = move_queued_task(rq, rf, p, dest_cpu);
1531 * migration_cpu_stop - this will be executed by a highprio stopper thread
1532 * and performs thread migration by bumping thread off CPU then
1533 * 'pushing' onto another runqueue.
1535 static int migration_cpu_stop(void *data)
1537 struct migration_arg *arg = data;
1538 struct task_struct *p = arg->task;
1539 struct rq *rq = this_rq();
1543 * The original target CPU might have gone down and we might
1544 * be on another CPU but it doesn't matter.
1546 local_irq_disable();
1548 * We need to explicitly wake pending tasks before running
1549 * __migrate_task() such that we will not miss enforcing cpus_ptr
1550 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1552 sched_ttwu_pending();
1554 raw_spin_lock(&p->pi_lock);
1557 * If task_rq(p) != rq, it cannot be migrated here, because we're
1558 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1559 * we're holding p->pi_lock.
1561 if (task_rq(p) == rq) {
1562 if (task_on_rq_queued(p))
1563 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1565 p->wake_cpu = arg->dest_cpu;
1568 raw_spin_unlock(&p->pi_lock);
1575 * sched_class::set_cpus_allowed must do the below, but is not required to
1576 * actually call this function.
1578 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1580 cpumask_copy(&p->cpus_mask, new_mask);
1581 p->nr_cpus_allowed = cpumask_weight(new_mask);
1584 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1586 struct rq *rq = task_rq(p);
1587 bool queued, running;
1589 lockdep_assert_held(&p->pi_lock);
1591 queued = task_on_rq_queued(p);
1592 running = task_current(rq, p);
1596 * Because __kthread_bind() calls this on blocked tasks without
1599 lockdep_assert_held(&rq->lock);
1600 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1603 put_prev_task(rq, p);
1605 p->sched_class->set_cpus_allowed(p, new_mask);
1608 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1610 set_next_task(rq, p);
1614 * Change a given task's CPU affinity. Migrate the thread to a
1615 * proper CPU and schedule it away if the CPU it's executing on
1616 * is removed from the allowed bitmask.
1618 * NOTE: the caller must have a valid reference to the task, the
1619 * task must not exit() & deallocate itself prematurely. The
1620 * call is not atomic; no spinlocks may be held.
1622 static int __set_cpus_allowed_ptr(struct task_struct *p,
1623 const struct cpumask *new_mask, bool check)
1625 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1626 unsigned int dest_cpu;
1631 rq = task_rq_lock(p, &rf);
1632 update_rq_clock(rq);
1634 if (p->flags & PF_KTHREAD) {
1636 * Kernel threads are allowed on online && !active CPUs
1638 cpu_valid_mask = cpu_online_mask;
1642 * Must re-check here, to close a race against __kthread_bind(),
1643 * sched_setaffinity() is not guaranteed to observe the flag.
1645 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1650 if (cpumask_equal(p->cpus_ptr, new_mask))
1654 * Picking a ~random cpu helps in cases where we are changing affinity
1655 * for groups of tasks (ie. cpuset), so that load balancing is not
1656 * immediately required to distribute the tasks within their new mask.
1658 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
1659 if (dest_cpu >= nr_cpu_ids) {
1664 do_set_cpus_allowed(p, new_mask);
1666 if (p->flags & PF_KTHREAD) {
1668 * For kernel threads that do indeed end up on online &&
1669 * !active we want to ensure they are strict per-CPU threads.
1671 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1672 !cpumask_intersects(new_mask, cpu_active_mask) &&
1673 p->nr_cpus_allowed != 1);
1676 /* Can the task run on the task's current CPU? If so, we're done */
1677 if (cpumask_test_cpu(task_cpu(p), new_mask))
1680 if (task_running(rq, p) || p->state == TASK_WAKING) {
1681 struct migration_arg arg = { p, dest_cpu };
1682 /* Need help from migration thread: drop lock and wait. */
1683 task_rq_unlock(rq, p, &rf);
1684 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1686 } else if (task_on_rq_queued(p)) {
1688 * OK, since we're going to drop the lock immediately
1689 * afterwards anyway.
1691 rq = move_queued_task(rq, &rf, p, dest_cpu);
1694 task_rq_unlock(rq, p, &rf);
1699 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1701 return __set_cpus_allowed_ptr(p, new_mask, false);
1703 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1705 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1707 #ifdef CONFIG_SCHED_DEBUG
1709 * We should never call set_task_cpu() on a blocked task,
1710 * ttwu() will sort out the placement.
1712 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1716 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1717 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1718 * time relying on p->on_rq.
1720 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1721 p->sched_class == &fair_sched_class &&
1722 (p->on_rq && !task_on_rq_migrating(p)));
1724 #ifdef CONFIG_LOCKDEP
1726 * The caller should hold either p->pi_lock or rq->lock, when changing
1727 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1729 * sched_move_task() holds both and thus holding either pins the cgroup,
1732 * Furthermore, all task_rq users should acquire both locks, see
1735 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1736 lockdep_is_held(&task_rq(p)->lock)));
1739 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1741 WARN_ON_ONCE(!cpu_online(new_cpu));
1744 trace_sched_migrate_task(p, new_cpu);
1746 if (task_cpu(p) != new_cpu) {
1747 if (p->sched_class->migrate_task_rq)
1748 p->sched_class->migrate_task_rq(p, new_cpu);
1749 p->se.nr_migrations++;
1751 perf_event_task_migrate(p);
1754 __set_task_cpu(p, new_cpu);
1757 #ifdef CONFIG_NUMA_BALANCING
1758 static void __migrate_swap_task(struct task_struct *p, int cpu)
1760 if (task_on_rq_queued(p)) {
1761 struct rq *src_rq, *dst_rq;
1762 struct rq_flags srf, drf;
1764 src_rq = task_rq(p);
1765 dst_rq = cpu_rq(cpu);
1767 rq_pin_lock(src_rq, &srf);
1768 rq_pin_lock(dst_rq, &drf);
1770 deactivate_task(src_rq, p, 0);
1771 set_task_cpu(p, cpu);
1772 activate_task(dst_rq, p, 0);
1773 check_preempt_curr(dst_rq, p, 0);
1775 rq_unpin_lock(dst_rq, &drf);
1776 rq_unpin_lock(src_rq, &srf);
1780 * Task isn't running anymore; make it appear like we migrated
1781 * it before it went to sleep. This means on wakeup we make the
1782 * previous CPU our target instead of where it really is.
1788 struct migration_swap_arg {
1789 struct task_struct *src_task, *dst_task;
1790 int src_cpu, dst_cpu;
1793 static int migrate_swap_stop(void *data)
1795 struct migration_swap_arg *arg = data;
1796 struct rq *src_rq, *dst_rq;
1799 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1802 src_rq = cpu_rq(arg->src_cpu);
1803 dst_rq = cpu_rq(arg->dst_cpu);
1805 double_raw_lock(&arg->src_task->pi_lock,
1806 &arg->dst_task->pi_lock);
1807 double_rq_lock(src_rq, dst_rq);
1809 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1812 if (task_cpu(arg->src_task) != arg->src_cpu)
1815 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1818 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1821 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1822 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1827 double_rq_unlock(src_rq, dst_rq);
1828 raw_spin_unlock(&arg->dst_task->pi_lock);
1829 raw_spin_unlock(&arg->src_task->pi_lock);
1835 * Cross migrate two tasks
1837 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1838 int target_cpu, int curr_cpu)
1840 struct migration_swap_arg arg;
1843 arg = (struct migration_swap_arg){
1845 .src_cpu = curr_cpu,
1847 .dst_cpu = target_cpu,
1850 if (arg.src_cpu == arg.dst_cpu)
1854 * These three tests are all lockless; this is OK since all of them
1855 * will be re-checked with proper locks held further down the line.
1857 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1860 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1863 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1866 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1867 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1872 #endif /* CONFIG_NUMA_BALANCING */
1875 * wait_task_inactive - wait for a thread to unschedule.
1877 * If @match_state is nonzero, it's the @p->state value just checked and
1878 * not expected to change. If it changes, i.e. @p might have woken up,
1879 * then return zero. When we succeed in waiting for @p to be off its CPU,
1880 * we return a positive number (its total switch count). If a second call
1881 * a short while later returns the same number, the caller can be sure that
1882 * @p has remained unscheduled the whole time.
1884 * The caller must ensure that the task *will* unschedule sometime soon,
1885 * else this function might spin for a *long* time. This function can't
1886 * be called with interrupts off, or it may introduce deadlock with
1887 * smp_call_function() if an IPI is sent by the same process we are
1888 * waiting to become inactive.
1890 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1892 int running, queued;
1899 * We do the initial early heuristics without holding
1900 * any task-queue locks at all. We'll only try to get
1901 * the runqueue lock when things look like they will
1907 * If the task is actively running on another CPU
1908 * still, just relax and busy-wait without holding
1911 * NOTE! Since we don't hold any locks, it's not
1912 * even sure that "rq" stays as the right runqueue!
1913 * But we don't care, since "task_running()" will
1914 * return false if the runqueue has changed and p
1915 * is actually now running somewhere else!
1917 while (task_running(rq, p)) {
1918 if (match_state && unlikely(p->state != match_state))
1924 * Ok, time to look more closely! We need the rq
1925 * lock now, to be *sure*. If we're wrong, we'll
1926 * just go back and repeat.
1928 rq = task_rq_lock(p, &rf);
1929 trace_sched_wait_task(p);
1930 running = task_running(rq, p);
1931 queued = task_on_rq_queued(p);
1933 if (!match_state || p->state == match_state)
1934 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1935 task_rq_unlock(rq, p, &rf);
1938 * If it changed from the expected state, bail out now.
1940 if (unlikely(!ncsw))
1944 * Was it really running after all now that we
1945 * checked with the proper locks actually held?
1947 * Oops. Go back and try again..
1949 if (unlikely(running)) {
1955 * It's not enough that it's not actively running,
1956 * it must be off the runqueue _entirely_, and not
1959 * So if it was still runnable (but just not actively
1960 * running right now), it's preempted, and we should
1961 * yield - it could be a while.
1963 if (unlikely(queued)) {
1964 ktime_t to = NSEC_PER_SEC / HZ;
1966 set_current_state(TASK_UNINTERRUPTIBLE);
1967 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1972 * Ahh, all good. It wasn't running, and it wasn't
1973 * runnable, which means that it will never become
1974 * running in the future either. We're all done!
1983 * kick_process - kick a running thread to enter/exit the kernel
1984 * @p: the to-be-kicked thread
1986 * Cause a process which is running on another CPU to enter
1987 * kernel-mode, without any delay. (to get signals handled.)
1989 * NOTE: this function doesn't have to take the runqueue lock,
1990 * because all it wants to ensure is that the remote task enters
1991 * the kernel. If the IPI races and the task has been migrated
1992 * to another CPU then no harm is done and the purpose has been
1995 void kick_process(struct task_struct *p)
2001 if ((cpu != smp_processor_id()) && task_curr(p))
2002 smp_send_reschedule(cpu);
2005 EXPORT_SYMBOL_GPL(kick_process);
2008 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2010 * A few notes on cpu_active vs cpu_online:
2012 * - cpu_active must be a subset of cpu_online
2014 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2015 * see __set_cpus_allowed_ptr(). At this point the newly online
2016 * CPU isn't yet part of the sched domains, and balancing will not
2019 * - on CPU-down we clear cpu_active() to mask the sched domains and
2020 * avoid the load balancer to place new tasks on the to be removed
2021 * CPU. Existing tasks will remain running there and will be taken
2024 * This means that fallback selection must not select !active CPUs.
2025 * And can assume that any active CPU must be online. Conversely
2026 * select_task_rq() below may allow selection of !active CPUs in order
2027 * to satisfy the above rules.
2029 static int select_fallback_rq(int cpu, struct task_struct *p)
2031 int nid = cpu_to_node(cpu);
2032 const struct cpumask *nodemask = NULL;
2033 enum { cpuset, possible, fail } state = cpuset;
2037 * If the node that the CPU is on has been offlined, cpu_to_node()
2038 * will return -1. There is no CPU on the node, and we should
2039 * select the CPU on the other node.
2042 nodemask = cpumask_of_node(nid);
2044 /* Look for allowed, online CPU in same node. */
2045 for_each_cpu(dest_cpu, nodemask) {
2046 if (!cpu_active(dest_cpu))
2048 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2054 /* Any allowed, online CPU? */
2055 for_each_cpu(dest_cpu, p->cpus_ptr) {
2056 if (!is_cpu_allowed(p, dest_cpu))
2062 /* No more Mr. Nice Guy. */
2065 if (IS_ENABLED(CONFIG_CPUSETS)) {
2066 cpuset_cpus_allowed_fallback(p);
2072 do_set_cpus_allowed(p, cpu_possible_mask);
2083 if (state != cpuset) {
2085 * Don't tell them about moving exiting tasks or
2086 * kernel threads (both mm NULL), since they never
2089 if (p->mm && printk_ratelimit()) {
2090 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2091 task_pid_nr(p), p->comm, cpu);
2099 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2102 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2104 lockdep_assert_held(&p->pi_lock);
2106 if (p->nr_cpus_allowed > 1)
2107 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2109 cpu = cpumask_any(p->cpus_ptr);
2112 * In order not to call set_task_cpu() on a blocking task we need
2113 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2116 * Since this is common to all placement strategies, this lives here.
2118 * [ this allows ->select_task() to simply return task_cpu(p) and
2119 * not worry about this generic constraint ]
2121 if (unlikely(!is_cpu_allowed(p, cpu)))
2122 cpu = select_fallback_rq(task_cpu(p), p);
2127 static void update_avg(u64 *avg, u64 sample)
2129 s64 diff = sample - *avg;
2133 void sched_set_stop_task(int cpu, struct task_struct *stop)
2135 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2136 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2140 * Make it appear like a SCHED_FIFO task, its something
2141 * userspace knows about and won't get confused about.
2143 * Also, it will make PI more or less work without too
2144 * much confusion -- but then, stop work should not
2145 * rely on PI working anyway.
2147 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2149 stop->sched_class = &stop_sched_class;
2152 cpu_rq(cpu)->stop = stop;
2156 * Reset it back to a normal scheduling class so that
2157 * it can die in pieces.
2159 old_stop->sched_class = &rt_sched_class;
2165 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2166 const struct cpumask *new_mask, bool check)
2168 return set_cpus_allowed_ptr(p, new_mask);
2171 #endif /* CONFIG_SMP */
2174 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2178 if (!schedstat_enabled())
2184 if (cpu == rq->cpu) {
2185 __schedstat_inc(rq->ttwu_local);
2186 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2188 struct sched_domain *sd;
2190 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2192 for_each_domain(rq->cpu, sd) {
2193 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2194 __schedstat_inc(sd->ttwu_wake_remote);
2201 if (wake_flags & WF_MIGRATED)
2202 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2203 #endif /* CONFIG_SMP */
2205 __schedstat_inc(rq->ttwu_count);
2206 __schedstat_inc(p->se.statistics.nr_wakeups);
2208 if (wake_flags & WF_SYNC)
2209 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2213 * Mark the task runnable and perform wakeup-preemption.
2215 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2216 struct rq_flags *rf)
2218 check_preempt_curr(rq, p, wake_flags);
2219 p->state = TASK_RUNNING;
2220 trace_sched_wakeup(p);
2223 if (p->sched_class->task_woken) {
2225 * Our task @p is fully woken up and running; so its safe to
2226 * drop the rq->lock, hereafter rq is only used for statistics.
2228 rq_unpin_lock(rq, rf);
2229 p->sched_class->task_woken(rq, p);
2230 rq_repin_lock(rq, rf);
2233 if (rq->idle_stamp) {
2234 u64 delta = rq_clock(rq) - rq->idle_stamp;
2235 u64 max = 2*rq->max_idle_balance_cost;
2237 update_avg(&rq->avg_idle, delta);
2239 if (rq->avg_idle > max)
2248 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2249 struct rq_flags *rf)
2251 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2253 lockdep_assert_held(&rq->lock);
2256 if (p->sched_contributes_to_load)
2257 rq->nr_uninterruptible--;
2259 if (wake_flags & WF_MIGRATED)
2260 en_flags |= ENQUEUE_MIGRATED;
2263 activate_task(rq, p, en_flags);
2264 ttwu_do_wakeup(rq, p, wake_flags, rf);
2268 * Called in case the task @p isn't fully descheduled from its runqueue,
2269 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2270 * since all we need to do is flip p->state to TASK_RUNNING, since
2271 * the task is still ->on_rq.
2273 static int ttwu_remote(struct task_struct *p, int wake_flags)
2279 rq = __task_rq_lock(p, &rf);
2280 if (task_on_rq_queued(p)) {
2281 /* check_preempt_curr() may use rq clock */
2282 update_rq_clock(rq);
2283 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2286 __task_rq_unlock(rq, &rf);
2292 void sched_ttwu_pending(void)
2294 struct rq *rq = this_rq();
2295 struct llist_node *llist = llist_del_all(&rq->wake_list);
2296 struct task_struct *p, *t;
2302 rq_lock_irqsave(rq, &rf);
2303 update_rq_clock(rq);
2305 llist_for_each_entry_safe(p, t, llist, wake_entry)
2306 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2308 rq_unlock_irqrestore(rq, &rf);
2311 void scheduler_ipi(void)
2314 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2315 * TIF_NEED_RESCHED remotely (for the first time) will also send
2318 preempt_fold_need_resched();
2320 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2324 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2325 * traditionally all their work was done from the interrupt return
2326 * path. Now that we actually do some work, we need to make sure
2329 * Some archs already do call them, luckily irq_enter/exit nest
2332 * Arguably we should visit all archs and update all handlers,
2333 * however a fair share of IPIs are still resched only so this would
2334 * somewhat pessimize the simple resched case.
2337 sched_ttwu_pending();
2340 * Check if someone kicked us for doing the nohz idle load balance.
2342 if (unlikely(got_nohz_idle_kick())) {
2343 this_rq()->idle_balance = 1;
2344 raise_softirq_irqoff(SCHED_SOFTIRQ);
2349 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2351 struct rq *rq = cpu_rq(cpu);
2353 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2355 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2356 if (!set_nr_if_polling(rq->idle))
2357 smp_send_reschedule(cpu);
2359 trace_sched_wake_idle_without_ipi(cpu);
2363 void wake_up_if_idle(int cpu)
2365 struct rq *rq = cpu_rq(cpu);
2370 if (!is_idle_task(rcu_dereference(rq->curr)))
2373 if (set_nr_if_polling(rq->idle)) {
2374 trace_sched_wake_idle_without_ipi(cpu);
2376 rq_lock_irqsave(rq, &rf);
2377 if (is_idle_task(rq->curr))
2378 smp_send_reschedule(cpu);
2379 /* Else CPU is not idle, do nothing here: */
2380 rq_unlock_irqrestore(rq, &rf);
2387 bool cpus_share_cache(int this_cpu, int that_cpu)
2389 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2391 #endif /* CONFIG_SMP */
2393 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2395 struct rq *rq = cpu_rq(cpu);
2398 #if defined(CONFIG_SMP)
2399 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2400 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2401 ttwu_queue_remote(p, cpu, wake_flags);
2407 update_rq_clock(rq);
2408 ttwu_do_activate(rq, p, wake_flags, &rf);
2413 * Notes on Program-Order guarantees on SMP systems.
2417 * The basic program-order guarantee on SMP systems is that when a task [t]
2418 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2419 * execution on its new CPU [c1].
2421 * For migration (of runnable tasks) this is provided by the following means:
2423 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2424 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2425 * rq(c1)->lock (if not at the same time, then in that order).
2426 * C) LOCK of the rq(c1)->lock scheduling in task
2428 * Release/acquire chaining guarantees that B happens after A and C after B.
2429 * Note: the CPU doing B need not be c0 or c1
2438 * UNLOCK rq(0)->lock
2440 * LOCK rq(0)->lock // orders against CPU0
2442 * UNLOCK rq(0)->lock
2446 * UNLOCK rq(1)->lock
2448 * LOCK rq(1)->lock // orders against CPU2
2451 * UNLOCK rq(1)->lock
2454 * BLOCKING -- aka. SLEEP + WAKEUP
2456 * For blocking we (obviously) need to provide the same guarantee as for
2457 * migration. However the means are completely different as there is no lock
2458 * chain to provide order. Instead we do:
2460 * 1) smp_store_release(X->on_cpu, 0)
2461 * 2) smp_cond_load_acquire(!X->on_cpu)
2465 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2467 * LOCK rq(0)->lock LOCK X->pi_lock
2470 * smp_store_release(X->on_cpu, 0);
2472 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2478 * X->state = RUNNING
2479 * UNLOCK rq(2)->lock
2481 * LOCK rq(2)->lock // orders against CPU1
2484 * UNLOCK rq(2)->lock
2487 * UNLOCK rq(0)->lock
2490 * However, for wakeups there is a second guarantee we must provide, namely we
2491 * must ensure that CONDITION=1 done by the caller can not be reordered with
2492 * accesses to the task state; see try_to_wake_up() and set_current_state().
2496 * try_to_wake_up - wake up a thread
2497 * @p: the thread to be awakened
2498 * @state: the mask of task states that can be woken
2499 * @wake_flags: wake modifier flags (WF_*)
2501 * If (@state & @p->state) @p->state = TASK_RUNNING.
2503 * If the task was not queued/runnable, also place it back on a runqueue.
2505 * Atomic against schedule() which would dequeue a task, also see
2506 * set_current_state().
2508 * This function executes a full memory barrier before accessing the task
2509 * state; see set_current_state().
2511 * Return: %true if @p->state changes (an actual wakeup was done),
2515 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2517 unsigned long flags;
2518 int cpu, success = 0;
2523 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2524 * == smp_processor_id()'. Together this means we can special
2525 * case the whole 'p->on_rq && ttwu_remote()' case below
2526 * without taking any locks.
2529 * - we rely on Program-Order guarantees for all the ordering,
2530 * - we're serialized against set_special_state() by virtue of
2531 * it disabling IRQs (this allows not taking ->pi_lock).
2533 if (!(p->state & state))
2538 trace_sched_waking(p);
2539 p->state = TASK_RUNNING;
2540 trace_sched_wakeup(p);
2545 * If we are going to wake up a thread waiting for CONDITION we
2546 * need to ensure that CONDITION=1 done by the caller can not be
2547 * reordered with p->state check below. This pairs with mb() in
2548 * set_current_state() the waiting thread does.
2550 raw_spin_lock_irqsave(&p->pi_lock, flags);
2551 smp_mb__after_spinlock();
2552 if (!(p->state & state))
2555 trace_sched_waking(p);
2557 /* We're going to change ->state: */
2562 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2563 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2564 * in smp_cond_load_acquire() below.
2566 * sched_ttwu_pending() try_to_wake_up()
2567 * STORE p->on_rq = 1 LOAD p->state
2570 * __schedule() (switch to task 'p')
2571 * LOCK rq->lock smp_rmb();
2572 * smp_mb__after_spinlock();
2576 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2578 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2579 * __schedule(). See the comment for smp_mb__after_spinlock().
2582 if (p->on_rq && ttwu_remote(p, wake_flags))
2587 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2588 * possible to, falsely, observe p->on_cpu == 0.
2590 * One must be running (->on_cpu == 1) in order to remove oneself
2591 * from the runqueue.
2593 * __schedule() (switch to task 'p') try_to_wake_up()
2594 * STORE p->on_cpu = 1 LOAD p->on_rq
2597 * __schedule() (put 'p' to sleep)
2598 * LOCK rq->lock smp_rmb();
2599 * smp_mb__after_spinlock();
2600 * STORE p->on_rq = 0 LOAD p->on_cpu
2602 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2603 * __schedule(). See the comment for smp_mb__after_spinlock().
2608 * If the owning (remote) CPU is still in the middle of schedule() with
2609 * this task as prev, wait until its done referencing the task.
2611 * Pairs with the smp_store_release() in finish_task().
2613 * This ensures that tasks getting woken will be fully ordered against
2614 * their previous state and preserve Program Order.
2616 smp_cond_load_acquire(&p->on_cpu, !VAL);
2618 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2619 p->state = TASK_WAKING;
2622 delayacct_blkio_end(p);
2623 atomic_dec(&task_rq(p)->nr_iowait);
2626 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2627 if (task_cpu(p) != cpu) {
2628 wake_flags |= WF_MIGRATED;
2629 psi_ttwu_dequeue(p);
2630 set_task_cpu(p, cpu);
2633 #else /* CONFIG_SMP */
2636 delayacct_blkio_end(p);
2637 atomic_dec(&task_rq(p)->nr_iowait);
2640 #endif /* CONFIG_SMP */
2642 ttwu_queue(p, cpu, wake_flags);
2644 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2647 ttwu_stat(p, cpu, wake_flags);
2654 * wake_up_process - Wake up a specific process
2655 * @p: The process to be woken up.
2657 * Attempt to wake up the nominated process and move it to the set of runnable
2660 * Return: 1 if the process was woken up, 0 if it was already running.
2662 * This function executes a full memory barrier before accessing the task state.
2664 int wake_up_process(struct task_struct *p)
2666 return try_to_wake_up(p, TASK_NORMAL, 0);
2668 EXPORT_SYMBOL(wake_up_process);
2670 int wake_up_state(struct task_struct *p, unsigned int state)
2672 return try_to_wake_up(p, state, 0);
2676 * Perform scheduler related setup for a newly forked process p.
2677 * p is forked by current.
2679 * __sched_fork() is basic setup used by init_idle() too:
2681 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2686 p->se.exec_start = 0;
2687 p->se.sum_exec_runtime = 0;
2688 p->se.prev_sum_exec_runtime = 0;
2689 p->se.nr_migrations = 0;
2691 INIT_LIST_HEAD(&p->se.group_node);
2693 #ifdef CONFIG_FAIR_GROUP_SCHED
2694 p->se.cfs_rq = NULL;
2697 #ifdef CONFIG_SCHEDSTATS
2698 /* Even if schedstat is disabled, there should not be garbage */
2699 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2702 RB_CLEAR_NODE(&p->dl.rb_node);
2703 init_dl_task_timer(&p->dl);
2704 init_dl_inactive_task_timer(&p->dl);
2705 __dl_clear_params(p);
2707 INIT_LIST_HEAD(&p->rt.run_list);
2709 p->rt.time_slice = sched_rr_timeslice;
2713 #ifdef CONFIG_PREEMPT_NOTIFIERS
2714 INIT_HLIST_HEAD(&p->preempt_notifiers);
2717 #ifdef CONFIG_COMPACTION
2718 p->capture_control = NULL;
2720 init_numa_balancing(clone_flags, p);
2723 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2725 #ifdef CONFIG_NUMA_BALANCING
2727 void set_numabalancing_state(bool enabled)
2730 static_branch_enable(&sched_numa_balancing);
2732 static_branch_disable(&sched_numa_balancing);
2735 #ifdef CONFIG_PROC_SYSCTL
2736 int sysctl_numa_balancing(struct ctl_table *table, int write,
2737 void __user *buffer, size_t *lenp, loff_t *ppos)
2741 int state = static_branch_likely(&sched_numa_balancing);
2743 if (write && !capable(CAP_SYS_ADMIN))
2748 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2752 set_numabalancing_state(state);
2758 #ifdef CONFIG_SCHEDSTATS
2760 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2761 static bool __initdata __sched_schedstats = false;
2763 static void set_schedstats(bool enabled)
2766 static_branch_enable(&sched_schedstats);
2768 static_branch_disable(&sched_schedstats);
2771 void force_schedstat_enabled(void)
2773 if (!schedstat_enabled()) {
2774 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2775 static_branch_enable(&sched_schedstats);
2779 static int __init setup_schedstats(char *str)
2786 * This code is called before jump labels have been set up, so we can't
2787 * change the static branch directly just yet. Instead set a temporary
2788 * variable so init_schedstats() can do it later.
2790 if (!strcmp(str, "enable")) {
2791 __sched_schedstats = true;
2793 } else if (!strcmp(str, "disable")) {
2794 __sched_schedstats = false;
2799 pr_warn("Unable to parse schedstats=\n");
2803 __setup("schedstats=", setup_schedstats);
2805 static void __init init_schedstats(void)
2807 set_schedstats(__sched_schedstats);
2810 #ifdef CONFIG_PROC_SYSCTL
2811 int sysctl_schedstats(struct ctl_table *table, int write,
2812 void __user *buffer, size_t *lenp, loff_t *ppos)
2816 int state = static_branch_likely(&sched_schedstats);
2818 if (write && !capable(CAP_SYS_ADMIN))
2823 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2827 set_schedstats(state);
2830 #endif /* CONFIG_PROC_SYSCTL */
2831 #else /* !CONFIG_SCHEDSTATS */
2832 static inline void init_schedstats(void) {}
2833 #endif /* CONFIG_SCHEDSTATS */
2836 * fork()/clone()-time setup:
2838 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2840 unsigned long flags;
2842 __sched_fork(clone_flags, p);
2844 * We mark the process as NEW here. This guarantees that
2845 * nobody will actually run it, and a signal or other external
2846 * event cannot wake it up and insert it on the runqueue either.
2848 p->state = TASK_NEW;
2851 * Make sure we do not leak PI boosting priority to the child.
2853 p->prio = current->normal_prio;
2858 * Revert to default priority/policy on fork if requested.
2860 if (unlikely(p->sched_reset_on_fork)) {
2861 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2862 p->policy = SCHED_NORMAL;
2863 p->static_prio = NICE_TO_PRIO(0);
2865 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2866 p->static_prio = NICE_TO_PRIO(0);
2868 p->prio = p->normal_prio = __normal_prio(p);
2869 set_load_weight(p, false);
2872 * We don't need the reset flag anymore after the fork. It has
2873 * fulfilled its duty:
2875 p->sched_reset_on_fork = 0;
2878 if (dl_prio(p->prio))
2880 else if (rt_prio(p->prio))
2881 p->sched_class = &rt_sched_class;
2883 p->sched_class = &fair_sched_class;
2885 init_entity_runnable_average(&p->se);
2888 * The child is not yet in the pid-hash so no cgroup attach races,
2889 * and the cgroup is pinned to this child due to cgroup_fork()
2890 * is ran before sched_fork().
2892 * Silence PROVE_RCU.
2894 raw_spin_lock_irqsave(&p->pi_lock, flags);
2896 * We're setting the CPU for the first time, we don't migrate,
2897 * so use __set_task_cpu().
2899 __set_task_cpu(p, smp_processor_id());
2900 if (p->sched_class->task_fork)
2901 p->sched_class->task_fork(p);
2902 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2904 #ifdef CONFIG_SCHED_INFO
2905 if (likely(sched_info_on()))
2906 memset(&p->sched_info, 0, sizeof(p->sched_info));
2908 #if defined(CONFIG_SMP)
2911 init_task_preempt_count(p);
2913 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2914 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2919 unsigned long to_ratio(u64 period, u64 runtime)
2921 if (runtime == RUNTIME_INF)
2925 * Doing this here saves a lot of checks in all
2926 * the calling paths, and returning zero seems
2927 * safe for them anyway.
2932 return div64_u64(runtime << BW_SHIFT, period);
2936 * wake_up_new_task - wake up a newly created task for the first time.
2938 * This function will do some initial scheduler statistics housekeeping
2939 * that must be done for every newly created context, then puts the task
2940 * on the runqueue and wakes it.
2942 void wake_up_new_task(struct task_struct *p)
2947 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2948 p->state = TASK_RUNNING;
2951 * Fork balancing, do it here and not earlier because:
2952 * - cpus_ptr can change in the fork path
2953 * - any previously selected CPU might disappear through hotplug
2955 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2956 * as we're not fully set-up yet.
2958 p->recent_used_cpu = task_cpu(p);
2959 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2961 rq = __task_rq_lock(p, &rf);
2962 update_rq_clock(rq);
2963 post_init_entity_util_avg(p);
2965 activate_task(rq, p, ENQUEUE_NOCLOCK);
2966 trace_sched_wakeup_new(p);
2967 check_preempt_curr(rq, p, WF_FORK);
2969 if (p->sched_class->task_woken) {
2971 * Nothing relies on rq->lock after this, so its fine to
2974 rq_unpin_lock(rq, &rf);
2975 p->sched_class->task_woken(rq, p);
2976 rq_repin_lock(rq, &rf);
2979 task_rq_unlock(rq, p, &rf);
2982 #ifdef CONFIG_PREEMPT_NOTIFIERS
2984 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2986 void preempt_notifier_inc(void)
2988 static_branch_inc(&preempt_notifier_key);
2990 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2992 void preempt_notifier_dec(void)
2994 static_branch_dec(&preempt_notifier_key);
2996 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2999 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3000 * @notifier: notifier struct to register
3002 void preempt_notifier_register(struct preempt_notifier *notifier)
3004 if (!static_branch_unlikely(&preempt_notifier_key))
3005 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3007 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3009 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3012 * preempt_notifier_unregister - no longer interested in preemption notifications
3013 * @notifier: notifier struct to unregister
3015 * This is *not* safe to call from within a preemption notifier.
3017 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3019 hlist_del(¬ifier->link);
3021 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3023 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3025 struct preempt_notifier *notifier;
3027 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3028 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3031 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3033 if (static_branch_unlikely(&preempt_notifier_key))
3034 __fire_sched_in_preempt_notifiers(curr);
3038 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3039 struct task_struct *next)
3041 struct preempt_notifier *notifier;
3043 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3044 notifier->ops->sched_out(notifier, next);
3047 static __always_inline void
3048 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3049 struct task_struct *next)
3051 if (static_branch_unlikely(&preempt_notifier_key))
3052 __fire_sched_out_preempt_notifiers(curr, next);
3055 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3057 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3062 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3063 struct task_struct *next)
3067 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3069 static inline void prepare_task(struct task_struct *next)
3073 * Claim the task as running, we do this before switching to it
3074 * such that any running task will have this set.
3080 static inline void finish_task(struct task_struct *prev)
3084 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3085 * We must ensure this doesn't happen until the switch is completely
3088 * In particular, the load of prev->state in finish_task_switch() must
3089 * happen before this.
3091 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3093 smp_store_release(&prev->on_cpu, 0);
3098 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3101 * Since the runqueue lock will be released by the next
3102 * task (which is an invalid locking op but in the case
3103 * of the scheduler it's an obvious special-case), so we
3104 * do an early lockdep release here:
3106 rq_unpin_lock(rq, rf);
3107 spin_release(&rq->lock.dep_map, _THIS_IP_);
3108 #ifdef CONFIG_DEBUG_SPINLOCK
3109 /* this is a valid case when another task releases the spinlock */
3110 rq->lock.owner = next;
3114 static inline void finish_lock_switch(struct rq *rq)
3117 * If we are tracking spinlock dependencies then we have to
3118 * fix up the runqueue lock - which gets 'carried over' from
3119 * prev into current:
3121 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3122 raw_spin_unlock_irq(&rq->lock);
3126 * NOP if the arch has not defined these:
3129 #ifndef prepare_arch_switch
3130 # define prepare_arch_switch(next) do { } while (0)
3133 #ifndef finish_arch_post_lock_switch
3134 # define finish_arch_post_lock_switch() do { } while (0)
3138 * prepare_task_switch - prepare to switch tasks
3139 * @rq: the runqueue preparing to switch
3140 * @prev: the current task that is being switched out
3141 * @next: the task we are going to switch to.
3143 * This is called with the rq lock held and interrupts off. It must
3144 * be paired with a subsequent finish_task_switch after the context
3147 * prepare_task_switch sets up locking and calls architecture specific
3151 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3152 struct task_struct *next)
3154 kcov_prepare_switch(prev);
3155 sched_info_switch(rq, prev, next);
3156 perf_event_task_sched_out(prev, next);
3158 fire_sched_out_preempt_notifiers(prev, next);
3160 prepare_arch_switch(next);
3164 * finish_task_switch - clean up after a task-switch
3165 * @prev: the thread we just switched away from.
3167 * finish_task_switch must be called after the context switch, paired
3168 * with a prepare_task_switch call before the context switch.
3169 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3170 * and do any other architecture-specific cleanup actions.
3172 * Note that we may have delayed dropping an mm in context_switch(). If
3173 * so, we finish that here outside of the runqueue lock. (Doing it
3174 * with the lock held can cause deadlocks; see schedule() for
3177 * The context switch have flipped the stack from under us and restored the
3178 * local variables which were saved when this task called schedule() in the
3179 * past. prev == current is still correct but we need to recalculate this_rq
3180 * because prev may have moved to another CPU.
3182 static struct rq *finish_task_switch(struct task_struct *prev)
3183 __releases(rq->lock)
3185 struct rq *rq = this_rq();
3186 struct mm_struct *mm = rq->prev_mm;
3190 * The previous task will have left us with a preempt_count of 2
3191 * because it left us after:
3194 * preempt_disable(); // 1
3196 * raw_spin_lock_irq(&rq->lock) // 2
3198 * Also, see FORK_PREEMPT_COUNT.
3200 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3201 "corrupted preempt_count: %s/%d/0x%x\n",
3202 current->comm, current->pid, preempt_count()))
3203 preempt_count_set(FORK_PREEMPT_COUNT);
3208 * A task struct has one reference for the use as "current".
3209 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3210 * schedule one last time. The schedule call will never return, and
3211 * the scheduled task must drop that reference.
3213 * We must observe prev->state before clearing prev->on_cpu (in
3214 * finish_task), otherwise a concurrent wakeup can get prev
3215 * running on another CPU and we could rave with its RUNNING -> DEAD
3216 * transition, resulting in a double drop.
3218 prev_state = prev->state;
3219 vtime_task_switch(prev);
3220 perf_event_task_sched_in(prev, current);
3222 finish_lock_switch(rq);
3223 finish_arch_post_lock_switch();
3224 kcov_finish_switch(current);
3226 fire_sched_in_preempt_notifiers(current);
3228 * When switching through a kernel thread, the loop in
3229 * membarrier_{private,global}_expedited() may have observed that
3230 * kernel thread and not issued an IPI. It is therefore possible to
3231 * schedule between user->kernel->user threads without passing though
3232 * switch_mm(). Membarrier requires a barrier after storing to
3233 * rq->curr, before returning to userspace, so provide them here:
3235 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3236 * provided by mmdrop(),
3237 * - a sync_core for SYNC_CORE.
3240 membarrier_mm_sync_core_before_usermode(mm);
3243 if (unlikely(prev_state == TASK_DEAD)) {
3244 if (prev->sched_class->task_dead)
3245 prev->sched_class->task_dead(prev);
3248 * Remove function-return probe instances associated with this
3249 * task and put them back on the free list.
3251 kprobe_flush_task(prev);
3253 /* Task is done with its stack. */
3254 put_task_stack(prev);
3256 put_task_struct_rcu_user(prev);
3259 tick_nohz_task_switch();
3265 /* rq->lock is NOT held, but preemption is disabled */
3266 static void __balance_callback(struct rq *rq)
3268 struct callback_head *head, *next;
3269 void (*func)(struct rq *rq);
3270 unsigned long flags;
3272 raw_spin_lock_irqsave(&rq->lock, flags);
3273 head = rq->balance_callback;
3274 rq->balance_callback = NULL;
3276 func = (void (*)(struct rq *))head->func;
3283 raw_spin_unlock_irqrestore(&rq->lock, flags);
3286 static inline void balance_callback(struct rq *rq)
3288 if (unlikely(rq->balance_callback))
3289 __balance_callback(rq);
3294 static inline void balance_callback(struct rq *rq)
3301 * schedule_tail - first thing a freshly forked thread must call.
3302 * @prev: the thread we just switched away from.
3304 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3305 __releases(rq->lock)
3310 * New tasks start with FORK_PREEMPT_COUNT, see there and
3311 * finish_task_switch() for details.
3313 * finish_task_switch() will drop rq->lock() and lower preempt_count
3314 * and the preempt_enable() will end up enabling preemption (on
3315 * PREEMPT_COUNT kernels).
3318 rq = finish_task_switch(prev);
3319 balance_callback(rq);
3322 if (current->set_child_tid)
3323 put_user(task_pid_vnr(current), current->set_child_tid);
3325 calculate_sigpending();
3329 * context_switch - switch to the new MM and the new thread's register state.
3331 static __always_inline struct rq *
3332 context_switch(struct rq *rq, struct task_struct *prev,
3333 struct task_struct *next, struct rq_flags *rf)
3335 prepare_task_switch(rq, prev, next);
3338 * For paravirt, this is coupled with an exit in switch_to to
3339 * combine the page table reload and the switch backend into
3342 arch_start_context_switch(prev);
3345 * kernel -> kernel lazy + transfer active
3346 * user -> kernel lazy + mmgrab() active
3348 * kernel -> user switch + mmdrop() active
3349 * user -> user switch
3351 if (!next->mm) { // to kernel
3352 enter_lazy_tlb(prev->active_mm, next);
3354 next->active_mm = prev->active_mm;
3355 if (prev->mm) // from user
3356 mmgrab(prev->active_mm);
3358 prev->active_mm = NULL;
3360 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3362 * sys_membarrier() requires an smp_mb() between setting
3363 * rq->curr / membarrier_switch_mm() and returning to userspace.
3365 * The below provides this either through switch_mm(), or in
3366 * case 'prev->active_mm == next->mm' through
3367 * finish_task_switch()'s mmdrop().
3369 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3371 if (!prev->mm) { // from kernel
3372 /* will mmdrop() in finish_task_switch(). */
3373 rq->prev_mm = prev->active_mm;
3374 prev->active_mm = NULL;
3378 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3380 prepare_lock_switch(rq, next, rf);
3382 /* Here we just switch the register state and the stack. */
3383 switch_to(prev, next, prev);
3386 return finish_task_switch(prev);
3390 * nr_running and nr_context_switches:
3392 * externally visible scheduler statistics: current number of runnable
3393 * threads, total number of context switches performed since bootup.
3395 unsigned long nr_running(void)
3397 unsigned long i, sum = 0;
3399 for_each_online_cpu(i)
3400 sum += cpu_rq(i)->nr_running;
3406 * Check if only the current task is running on the CPU.
3408 * Caution: this function does not check that the caller has disabled
3409 * preemption, thus the result might have a time-of-check-to-time-of-use
3410 * race. The caller is responsible to use it correctly, for example:
3412 * - from a non-preemptible section (of course)
3414 * - from a thread that is bound to a single CPU
3416 * - in a loop with very short iterations (e.g. a polling loop)
3418 bool single_task_running(void)
3420 return raw_rq()->nr_running == 1;
3422 EXPORT_SYMBOL(single_task_running);
3424 unsigned long long nr_context_switches(void)
3427 unsigned long long sum = 0;
3429 for_each_possible_cpu(i)
3430 sum += cpu_rq(i)->nr_switches;
3436 * Consumers of these two interfaces, like for example the cpuidle menu
3437 * governor, are using nonsensical data. Preferring shallow idle state selection
3438 * for a CPU that has IO-wait which might not even end up running the task when
3439 * it does become runnable.
3442 unsigned long nr_iowait_cpu(int cpu)
3444 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3448 * IO-wait accounting, and how its mostly bollocks (on SMP).
3450 * The idea behind IO-wait account is to account the idle time that we could
3451 * have spend running if it were not for IO. That is, if we were to improve the
3452 * storage performance, we'd have a proportional reduction in IO-wait time.
3454 * This all works nicely on UP, where, when a task blocks on IO, we account
3455 * idle time as IO-wait, because if the storage were faster, it could've been
3456 * running and we'd not be idle.
3458 * This has been extended to SMP, by doing the same for each CPU. This however
3461 * Imagine for instance the case where two tasks block on one CPU, only the one
3462 * CPU will have IO-wait accounted, while the other has regular idle. Even
3463 * though, if the storage were faster, both could've ran at the same time,
3464 * utilising both CPUs.
3466 * This means, that when looking globally, the current IO-wait accounting on
3467 * SMP is a lower bound, by reason of under accounting.
3469 * Worse, since the numbers are provided per CPU, they are sometimes
3470 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3471 * associated with any one particular CPU, it can wake to another CPU than it
3472 * blocked on. This means the per CPU IO-wait number is meaningless.
3474 * Task CPU affinities can make all that even more 'interesting'.
3477 unsigned long nr_iowait(void)
3479 unsigned long i, sum = 0;
3481 for_each_possible_cpu(i)
3482 sum += nr_iowait_cpu(i);
3490 * sched_exec - execve() is a valuable balancing opportunity, because at
3491 * this point the task has the smallest effective memory and cache footprint.
3493 void sched_exec(void)
3495 struct task_struct *p = current;
3496 unsigned long flags;
3499 raw_spin_lock_irqsave(&p->pi_lock, flags);
3500 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3501 if (dest_cpu == smp_processor_id())
3504 if (likely(cpu_active(dest_cpu))) {
3505 struct migration_arg arg = { p, dest_cpu };
3507 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3508 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3512 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3517 DEFINE_PER_CPU(struct kernel_stat, kstat);
3518 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3520 EXPORT_PER_CPU_SYMBOL(kstat);
3521 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3524 * The function fair_sched_class.update_curr accesses the struct curr
3525 * and its field curr->exec_start; when called from task_sched_runtime(),
3526 * we observe a high rate of cache misses in practice.
3527 * Prefetching this data results in improved performance.
3529 static inline void prefetch_curr_exec_start(struct task_struct *p)
3531 #ifdef CONFIG_FAIR_GROUP_SCHED
3532 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3534 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3537 prefetch(&curr->exec_start);
3541 * Return accounted runtime for the task.
3542 * In case the task is currently running, return the runtime plus current's
3543 * pending runtime that have not been accounted yet.
3545 unsigned long long task_sched_runtime(struct task_struct *p)
3551 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3553 * 64-bit doesn't need locks to atomically read a 64-bit value.
3554 * So we have a optimization chance when the task's delta_exec is 0.
3555 * Reading ->on_cpu is racy, but this is ok.
3557 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3558 * If we race with it entering CPU, unaccounted time is 0. This is
3559 * indistinguishable from the read occurring a few cycles earlier.
3560 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3561 * been accounted, so we're correct here as well.
3563 if (!p->on_cpu || !task_on_rq_queued(p))
3564 return p->se.sum_exec_runtime;
3567 rq = task_rq_lock(p, &rf);
3569 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3570 * project cycles that may never be accounted to this
3571 * thread, breaking clock_gettime().
3573 if (task_current(rq, p) && task_on_rq_queued(p)) {
3574 prefetch_curr_exec_start(p);
3575 update_rq_clock(rq);
3576 p->sched_class->update_curr(rq);
3578 ns = p->se.sum_exec_runtime;
3579 task_rq_unlock(rq, p, &rf);
3584 DEFINE_PER_CPU(unsigned long, thermal_pressure);
3586 void arch_set_thermal_pressure(struct cpumask *cpus,
3587 unsigned long th_pressure)
3591 for_each_cpu(cpu, cpus)
3592 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
3596 * This function gets called by the timer code, with HZ frequency.
3597 * We call it with interrupts disabled.
3599 void scheduler_tick(void)
3601 int cpu = smp_processor_id();
3602 struct rq *rq = cpu_rq(cpu);
3603 struct task_struct *curr = rq->curr;
3605 unsigned long thermal_pressure;
3607 arch_scale_freq_tick();
3612 update_rq_clock(rq);
3613 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
3614 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
3615 curr->sched_class->task_tick(rq, curr, 0);
3616 calc_global_load_tick(rq);
3621 perf_event_task_tick();
3624 rq->idle_balance = idle_cpu(cpu);
3625 trigger_load_balance(rq);
3629 #ifdef CONFIG_NO_HZ_FULL
3634 struct delayed_work work;
3636 /* Values for ->state, see diagram below. */
3637 #define TICK_SCHED_REMOTE_OFFLINE 0
3638 #define TICK_SCHED_REMOTE_OFFLINING 1
3639 #define TICK_SCHED_REMOTE_RUNNING 2
3642 * State diagram for ->state:
3645 * TICK_SCHED_REMOTE_OFFLINE
3648 * | | sched_tick_remote()
3651 * +--TICK_SCHED_REMOTE_OFFLINING
3654 * sched_tick_start() | | sched_tick_stop()
3657 * TICK_SCHED_REMOTE_RUNNING
3660 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3661 * and sched_tick_start() are happy to leave the state in RUNNING.
3664 static struct tick_work __percpu *tick_work_cpu;
3666 static void sched_tick_remote(struct work_struct *work)
3668 struct delayed_work *dwork = to_delayed_work(work);
3669 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3670 int cpu = twork->cpu;
3671 struct rq *rq = cpu_rq(cpu);
3672 struct task_struct *curr;
3678 * Handle the tick only if it appears the remote CPU is running in full
3679 * dynticks mode. The check is racy by nature, but missing a tick or
3680 * having one too much is no big deal because the scheduler tick updates
3681 * statistics and checks timeslices in a time-independent way, regardless
3682 * of when exactly it is running.
3684 if (!tick_nohz_tick_stopped_cpu(cpu))
3687 rq_lock_irq(rq, &rf);
3689 if (cpu_is_offline(cpu))
3692 update_rq_clock(rq);
3694 if (!is_idle_task(curr)) {
3696 * Make sure the next tick runs within a reasonable
3699 delta = rq_clock_task(rq) - curr->se.exec_start;
3700 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3702 curr->sched_class->task_tick(rq, curr, 0);
3704 calc_load_nohz_remote(rq);
3706 rq_unlock_irq(rq, &rf);
3710 * Run the remote tick once per second (1Hz). This arbitrary
3711 * frequency is large enough to avoid overload but short enough
3712 * to keep scheduler internal stats reasonably up to date. But
3713 * first update state to reflect hotplug activity if required.
3715 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3716 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3717 if (os == TICK_SCHED_REMOTE_RUNNING)
3718 queue_delayed_work(system_unbound_wq, dwork, HZ);
3721 static void sched_tick_start(int cpu)
3724 struct tick_work *twork;
3726 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3729 WARN_ON_ONCE(!tick_work_cpu);
3731 twork = per_cpu_ptr(tick_work_cpu, cpu);
3732 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3733 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3734 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3736 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3737 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3741 #ifdef CONFIG_HOTPLUG_CPU
3742 static void sched_tick_stop(int cpu)
3744 struct tick_work *twork;
3747 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3750 WARN_ON_ONCE(!tick_work_cpu);
3752 twork = per_cpu_ptr(tick_work_cpu, cpu);
3753 /* There cannot be competing actions, but don't rely on stop-machine. */
3754 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3755 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3756 /* Don't cancel, as this would mess up the state machine. */
3758 #endif /* CONFIG_HOTPLUG_CPU */
3760 int __init sched_tick_offload_init(void)
3762 tick_work_cpu = alloc_percpu(struct tick_work);
3763 BUG_ON(!tick_work_cpu);
3767 #else /* !CONFIG_NO_HZ_FULL */
3768 static inline void sched_tick_start(int cpu) { }
3769 static inline void sched_tick_stop(int cpu) { }
3772 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3773 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3775 * If the value passed in is equal to the current preempt count
3776 * then we just disabled preemption. Start timing the latency.
3778 static inline void preempt_latency_start(int val)
3780 if (preempt_count() == val) {
3781 unsigned long ip = get_lock_parent_ip();
3782 #ifdef CONFIG_DEBUG_PREEMPT
3783 current->preempt_disable_ip = ip;
3785 trace_preempt_off(CALLER_ADDR0, ip);
3789 void preempt_count_add(int val)
3791 #ifdef CONFIG_DEBUG_PREEMPT
3795 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3798 __preempt_count_add(val);
3799 #ifdef CONFIG_DEBUG_PREEMPT
3801 * Spinlock count overflowing soon?
3803 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3806 preempt_latency_start(val);
3808 EXPORT_SYMBOL(preempt_count_add);
3809 NOKPROBE_SYMBOL(preempt_count_add);
3812 * If the value passed in equals to the current preempt count
3813 * then we just enabled preemption. Stop timing the latency.
3815 static inline void preempt_latency_stop(int val)
3817 if (preempt_count() == val)
3818 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3821 void preempt_count_sub(int val)
3823 #ifdef CONFIG_DEBUG_PREEMPT
3827 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3830 * Is the spinlock portion underflowing?
3832 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3833 !(preempt_count() & PREEMPT_MASK)))
3837 preempt_latency_stop(val);
3838 __preempt_count_sub(val);
3840 EXPORT_SYMBOL(preempt_count_sub);
3841 NOKPROBE_SYMBOL(preempt_count_sub);
3844 static inline void preempt_latency_start(int val) { }
3845 static inline void preempt_latency_stop(int val) { }
3848 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3850 #ifdef CONFIG_DEBUG_PREEMPT
3851 return p->preempt_disable_ip;
3858 * Print scheduling while atomic bug:
3860 static noinline void __schedule_bug(struct task_struct *prev)
3862 /* Save this before calling printk(), since that will clobber it */
3863 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3865 if (oops_in_progress)
3868 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3869 prev->comm, prev->pid, preempt_count());
3871 debug_show_held_locks(prev);
3873 if (irqs_disabled())
3874 print_irqtrace_events(prev);
3875 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3876 && in_atomic_preempt_off()) {
3877 pr_err("Preemption disabled at:");
3878 print_ip_sym(preempt_disable_ip);
3882 panic("scheduling while atomic\n");
3885 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3889 * Various schedule()-time debugging checks and statistics:
3891 static inline void schedule_debug(struct task_struct *prev, bool preempt)
3893 #ifdef CONFIG_SCHED_STACK_END_CHECK
3894 if (task_stack_end_corrupted(prev))
3895 panic("corrupted stack end detected inside scheduler\n");
3898 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3899 if (!preempt && prev->state && prev->non_block_count) {
3900 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3901 prev->comm, prev->pid, prev->non_block_count);
3903 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3907 if (unlikely(in_atomic_preempt_off())) {
3908 __schedule_bug(prev);
3909 preempt_count_set(PREEMPT_DISABLED);
3913 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3915 schedstat_inc(this_rq()->sched_count);
3919 * Pick up the highest-prio task:
3921 static inline struct task_struct *
3922 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3924 const struct sched_class *class;
3925 struct task_struct *p;
3928 * Optimization: we know that if all tasks are in the fair class we can
3929 * call that function directly, but only if the @prev task wasn't of a
3930 * higher scheduling class, because otherwise those loose the
3931 * opportunity to pull in more work from other CPUs.
3933 if (likely((prev->sched_class == &idle_sched_class ||
3934 prev->sched_class == &fair_sched_class) &&
3935 rq->nr_running == rq->cfs.h_nr_running)) {
3937 p = pick_next_task_fair(rq, prev, rf);
3938 if (unlikely(p == RETRY_TASK))
3941 /* Assumes fair_sched_class->next == idle_sched_class */
3943 put_prev_task(rq, prev);
3944 p = pick_next_task_idle(rq);
3953 * We must do the balancing pass before put_next_task(), such
3954 * that when we release the rq->lock the task is in the same
3955 * state as before we took rq->lock.
3957 * We can terminate the balance pass as soon as we know there is
3958 * a runnable task of @class priority or higher.
3960 for_class_range(class, prev->sched_class, &idle_sched_class) {
3961 if (class->balance(rq, prev, rf))
3966 put_prev_task(rq, prev);
3968 for_each_class(class) {
3969 p = class->pick_next_task(rq);
3974 /* The idle class should always have a runnable task: */
3979 * __schedule() is the main scheduler function.
3981 * The main means of driving the scheduler and thus entering this function are:
3983 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3985 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3986 * paths. For example, see arch/x86/entry_64.S.
3988 * To drive preemption between tasks, the scheduler sets the flag in timer
3989 * interrupt handler scheduler_tick().
3991 * 3. Wakeups don't really cause entry into schedule(). They add a
3992 * task to the run-queue and that's it.
3994 * Now, if the new task added to the run-queue preempts the current
3995 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3996 * called on the nearest possible occasion:
3998 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4000 * - in syscall or exception context, at the next outmost
4001 * preempt_enable(). (this might be as soon as the wake_up()'s
4004 * - in IRQ context, return from interrupt-handler to
4005 * preemptible context
4007 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4010 * - cond_resched() call
4011 * - explicit schedule() call
4012 * - return from syscall or exception to user-space
4013 * - return from interrupt-handler to user-space
4015 * WARNING: must be called with preemption disabled!
4017 static void __sched notrace __schedule(bool preempt)
4019 struct task_struct *prev, *next;
4020 unsigned long *switch_count;
4025 cpu = smp_processor_id();
4029 schedule_debug(prev, preempt);
4031 if (sched_feat(HRTICK))
4034 local_irq_disable();
4035 rcu_note_context_switch(preempt);
4038 * Make sure that signal_pending_state()->signal_pending() below
4039 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4040 * done by the caller to avoid the race with signal_wake_up().
4042 * The membarrier system call requires a full memory barrier
4043 * after coming from user-space, before storing to rq->curr.
4046 smp_mb__after_spinlock();
4048 /* Promote REQ to ACT */
4049 rq->clock_update_flags <<= 1;
4050 update_rq_clock(rq);
4052 switch_count = &prev->nivcsw;
4053 if (!preempt && prev->state) {
4054 if (signal_pending_state(prev->state, prev)) {
4055 prev->state = TASK_RUNNING;
4057 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4059 if (prev->in_iowait) {
4060 atomic_inc(&rq->nr_iowait);
4061 delayacct_blkio_start();
4064 switch_count = &prev->nvcsw;
4067 next = pick_next_task(rq, prev, &rf);
4068 clear_tsk_need_resched(prev);
4069 clear_preempt_need_resched();
4071 if (likely(prev != next)) {
4074 * RCU users of rcu_dereference(rq->curr) may not see
4075 * changes to task_struct made by pick_next_task().
4077 RCU_INIT_POINTER(rq->curr, next);
4079 * The membarrier system call requires each architecture
4080 * to have a full memory barrier after updating
4081 * rq->curr, before returning to user-space.
4083 * Here are the schemes providing that barrier on the
4084 * various architectures:
4085 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4086 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4087 * - finish_lock_switch() for weakly-ordered
4088 * architectures where spin_unlock is a full barrier,
4089 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4090 * is a RELEASE barrier),
4094 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
4096 trace_sched_switch(preempt, prev, next);
4098 /* Also unlocks the rq: */
4099 rq = context_switch(rq, prev, next, &rf);
4101 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4102 rq_unlock_irq(rq, &rf);
4105 balance_callback(rq);
4108 void __noreturn do_task_dead(void)
4110 /* Causes final put_task_struct in finish_task_switch(): */
4111 set_special_state(TASK_DEAD);
4113 /* Tell freezer to ignore us: */
4114 current->flags |= PF_NOFREEZE;
4119 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4124 static inline void sched_submit_work(struct task_struct *tsk)
4130 * If a worker went to sleep, notify and ask workqueue whether
4131 * it wants to wake up a task to maintain concurrency.
4132 * As this function is called inside the schedule() context,
4133 * we disable preemption to avoid it calling schedule() again
4134 * in the possible wakeup of a kworker.
4136 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4138 if (tsk->flags & PF_WQ_WORKER)
4139 wq_worker_sleeping(tsk);
4141 io_wq_worker_sleeping(tsk);
4142 preempt_enable_no_resched();
4145 if (tsk_is_pi_blocked(tsk))
4149 * If we are going to sleep and we have plugged IO queued,
4150 * make sure to submit it to avoid deadlocks.
4152 if (blk_needs_flush_plug(tsk))
4153 blk_schedule_flush_plug(tsk);
4156 static void sched_update_worker(struct task_struct *tsk)
4158 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4159 if (tsk->flags & PF_WQ_WORKER)
4160 wq_worker_running(tsk);
4162 io_wq_worker_running(tsk);
4166 asmlinkage __visible void __sched schedule(void)
4168 struct task_struct *tsk = current;
4170 sched_submit_work(tsk);
4174 sched_preempt_enable_no_resched();
4175 } while (need_resched());
4176 sched_update_worker(tsk);
4178 EXPORT_SYMBOL(schedule);
4181 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4182 * state (have scheduled out non-voluntarily) by making sure that all
4183 * tasks have either left the run queue or have gone into user space.
4184 * As idle tasks do not do either, they must not ever be preempted
4185 * (schedule out non-voluntarily).
4187 * schedule_idle() is similar to schedule_preempt_disable() except that it
4188 * never enables preemption because it does not call sched_submit_work().
4190 void __sched schedule_idle(void)
4193 * As this skips calling sched_submit_work(), which the idle task does
4194 * regardless because that function is a nop when the task is in a
4195 * TASK_RUNNING state, make sure this isn't used someplace that the
4196 * current task can be in any other state. Note, idle is always in the
4197 * TASK_RUNNING state.
4199 WARN_ON_ONCE(current->state);
4202 } while (need_resched());
4205 #ifdef CONFIG_CONTEXT_TRACKING
4206 asmlinkage __visible void __sched schedule_user(void)
4209 * If we come here after a random call to set_need_resched(),
4210 * or we have been woken up remotely but the IPI has not yet arrived,
4211 * we haven't yet exited the RCU idle mode. Do it here manually until
4212 * we find a better solution.
4214 * NB: There are buggy callers of this function. Ideally we
4215 * should warn if prev_state != CONTEXT_USER, but that will trigger
4216 * too frequently to make sense yet.
4218 enum ctx_state prev_state = exception_enter();
4220 exception_exit(prev_state);
4225 * schedule_preempt_disabled - called with preemption disabled
4227 * Returns with preemption disabled. Note: preempt_count must be 1
4229 void __sched schedule_preempt_disabled(void)
4231 sched_preempt_enable_no_resched();
4236 static void __sched notrace preempt_schedule_common(void)
4240 * Because the function tracer can trace preempt_count_sub()
4241 * and it also uses preempt_enable/disable_notrace(), if
4242 * NEED_RESCHED is set, the preempt_enable_notrace() called
4243 * by the function tracer will call this function again and
4244 * cause infinite recursion.
4246 * Preemption must be disabled here before the function
4247 * tracer can trace. Break up preempt_disable() into two
4248 * calls. One to disable preemption without fear of being
4249 * traced. The other to still record the preemption latency,
4250 * which can also be traced by the function tracer.
4252 preempt_disable_notrace();
4253 preempt_latency_start(1);
4255 preempt_latency_stop(1);
4256 preempt_enable_no_resched_notrace();
4259 * Check again in case we missed a preemption opportunity
4260 * between schedule and now.
4262 } while (need_resched());
4265 #ifdef CONFIG_PREEMPTION
4267 * This is the entry point to schedule() from in-kernel preemption
4268 * off of preempt_enable.
4270 asmlinkage __visible void __sched notrace preempt_schedule(void)
4273 * If there is a non-zero preempt_count or interrupts are disabled,
4274 * we do not want to preempt the current task. Just return..
4276 if (likely(!preemptible()))
4279 preempt_schedule_common();
4281 NOKPROBE_SYMBOL(preempt_schedule);
4282 EXPORT_SYMBOL(preempt_schedule);
4285 * preempt_schedule_notrace - preempt_schedule called by tracing
4287 * The tracing infrastructure uses preempt_enable_notrace to prevent
4288 * recursion and tracing preempt enabling caused by the tracing
4289 * infrastructure itself. But as tracing can happen in areas coming
4290 * from userspace or just about to enter userspace, a preempt enable
4291 * can occur before user_exit() is called. This will cause the scheduler
4292 * to be called when the system is still in usermode.
4294 * To prevent this, the preempt_enable_notrace will use this function
4295 * instead of preempt_schedule() to exit user context if needed before
4296 * calling the scheduler.
4298 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4300 enum ctx_state prev_ctx;
4302 if (likely(!preemptible()))
4307 * Because the function tracer can trace preempt_count_sub()
4308 * and it also uses preempt_enable/disable_notrace(), if
4309 * NEED_RESCHED is set, the preempt_enable_notrace() called
4310 * by the function tracer will call this function again and
4311 * cause infinite recursion.
4313 * Preemption must be disabled here before the function
4314 * tracer can trace. Break up preempt_disable() into two
4315 * calls. One to disable preemption without fear of being
4316 * traced. The other to still record the preemption latency,
4317 * which can also be traced by the function tracer.
4319 preempt_disable_notrace();
4320 preempt_latency_start(1);
4322 * Needs preempt disabled in case user_exit() is traced
4323 * and the tracer calls preempt_enable_notrace() causing
4324 * an infinite recursion.
4326 prev_ctx = exception_enter();
4328 exception_exit(prev_ctx);
4330 preempt_latency_stop(1);
4331 preempt_enable_no_resched_notrace();
4332 } while (need_resched());
4334 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4336 #endif /* CONFIG_PREEMPTION */
4339 * This is the entry point to schedule() from kernel preemption
4340 * off of irq context.
4341 * Note, that this is called and return with irqs disabled. This will
4342 * protect us against recursive calling from irq.
4344 asmlinkage __visible void __sched preempt_schedule_irq(void)
4346 enum ctx_state prev_state;
4348 /* Catch callers which need to be fixed */
4349 BUG_ON(preempt_count() || !irqs_disabled());
4351 prev_state = exception_enter();
4357 local_irq_disable();
4358 sched_preempt_enable_no_resched();
4359 } while (need_resched());
4361 exception_exit(prev_state);
4364 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4367 return try_to_wake_up(curr->private, mode, wake_flags);
4369 EXPORT_SYMBOL(default_wake_function);
4371 #ifdef CONFIG_RT_MUTEXES
4373 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4376 prio = min(prio, pi_task->prio);
4381 static inline int rt_effective_prio(struct task_struct *p, int prio)
4383 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4385 return __rt_effective_prio(pi_task, prio);
4389 * rt_mutex_setprio - set the current priority of a task
4391 * @pi_task: donor task
4393 * This function changes the 'effective' priority of a task. It does
4394 * not touch ->normal_prio like __setscheduler().
4396 * Used by the rt_mutex code to implement priority inheritance
4397 * logic. Call site only calls if the priority of the task changed.
4399 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4401 int prio, oldprio, queued, running, queue_flag =
4402 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4403 const struct sched_class *prev_class;
4407 /* XXX used to be waiter->prio, not waiter->task->prio */
4408 prio = __rt_effective_prio(pi_task, p->normal_prio);
4411 * If nothing changed; bail early.
4413 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4416 rq = __task_rq_lock(p, &rf);
4417 update_rq_clock(rq);
4419 * Set under pi_lock && rq->lock, such that the value can be used under
4422 * Note that there is loads of tricky to make this pointer cache work
4423 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4424 * ensure a task is de-boosted (pi_task is set to NULL) before the
4425 * task is allowed to run again (and can exit). This ensures the pointer
4426 * points to a blocked task -- which guaratees the task is present.
4428 p->pi_top_task = pi_task;
4431 * For FIFO/RR we only need to set prio, if that matches we're done.
4433 if (prio == p->prio && !dl_prio(prio))
4437 * Idle task boosting is a nono in general. There is one
4438 * exception, when PREEMPT_RT and NOHZ is active:
4440 * The idle task calls get_next_timer_interrupt() and holds
4441 * the timer wheel base->lock on the CPU and another CPU wants
4442 * to access the timer (probably to cancel it). We can safely
4443 * ignore the boosting request, as the idle CPU runs this code
4444 * with interrupts disabled and will complete the lock
4445 * protected section without being interrupted. So there is no
4446 * real need to boost.
4448 if (unlikely(p == rq->idle)) {
4449 WARN_ON(p != rq->curr);
4450 WARN_ON(p->pi_blocked_on);
4454 trace_sched_pi_setprio(p, pi_task);
4457 if (oldprio == prio)
4458 queue_flag &= ~DEQUEUE_MOVE;
4460 prev_class = p->sched_class;
4461 queued = task_on_rq_queued(p);
4462 running = task_current(rq, p);
4464 dequeue_task(rq, p, queue_flag);
4466 put_prev_task(rq, p);
4469 * Boosting condition are:
4470 * 1. -rt task is running and holds mutex A
4471 * --> -dl task blocks on mutex A
4473 * 2. -dl task is running and holds mutex A
4474 * --> -dl task blocks on mutex A and could preempt the
4477 if (dl_prio(prio)) {
4478 if (!dl_prio(p->normal_prio) ||
4479 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4480 p->dl.dl_boosted = 1;
4481 queue_flag |= ENQUEUE_REPLENISH;
4483 p->dl.dl_boosted = 0;
4484 p->sched_class = &dl_sched_class;
4485 } else if (rt_prio(prio)) {
4486 if (dl_prio(oldprio))
4487 p->dl.dl_boosted = 0;
4489 queue_flag |= ENQUEUE_HEAD;
4490 p->sched_class = &rt_sched_class;
4492 if (dl_prio(oldprio))
4493 p->dl.dl_boosted = 0;
4494 if (rt_prio(oldprio))
4496 p->sched_class = &fair_sched_class;
4502 enqueue_task(rq, p, queue_flag);
4504 set_next_task(rq, p);
4506 check_class_changed(rq, p, prev_class, oldprio);
4508 /* Avoid rq from going away on us: */
4510 __task_rq_unlock(rq, &rf);
4512 balance_callback(rq);
4516 static inline int rt_effective_prio(struct task_struct *p, int prio)
4522 void set_user_nice(struct task_struct *p, long nice)
4524 bool queued, running;
4529 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4532 * We have to be careful, if called from sys_setpriority(),
4533 * the task might be in the middle of scheduling on another CPU.
4535 rq = task_rq_lock(p, &rf);
4536 update_rq_clock(rq);
4539 * The RT priorities are set via sched_setscheduler(), but we still
4540 * allow the 'normal' nice value to be set - but as expected
4541 * it wont have any effect on scheduling until the task is
4542 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4544 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4545 p->static_prio = NICE_TO_PRIO(nice);
4548 queued = task_on_rq_queued(p);
4549 running = task_current(rq, p);
4551 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4553 put_prev_task(rq, p);
4555 p->static_prio = NICE_TO_PRIO(nice);
4556 set_load_weight(p, true);
4558 p->prio = effective_prio(p);
4561 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4563 set_next_task(rq, p);
4566 * If the task increased its priority or is running and
4567 * lowered its priority, then reschedule its CPU:
4569 p->sched_class->prio_changed(rq, p, old_prio);
4572 task_rq_unlock(rq, p, &rf);
4574 EXPORT_SYMBOL(set_user_nice);
4577 * can_nice - check if a task can reduce its nice value
4581 int can_nice(const struct task_struct *p, const int nice)
4583 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4584 int nice_rlim = nice_to_rlimit(nice);
4586 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4587 capable(CAP_SYS_NICE));
4590 #ifdef __ARCH_WANT_SYS_NICE
4593 * sys_nice - change the priority of the current process.
4594 * @increment: priority increment
4596 * sys_setpriority is a more generic, but much slower function that
4597 * does similar things.
4599 SYSCALL_DEFINE1(nice, int, increment)
4604 * Setpriority might change our priority at the same moment.
4605 * We don't have to worry. Conceptually one call occurs first
4606 * and we have a single winner.
4608 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4609 nice = task_nice(current) + increment;
4611 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4612 if (increment < 0 && !can_nice(current, nice))
4615 retval = security_task_setnice(current, nice);
4619 set_user_nice(current, nice);
4626 * task_prio - return the priority value of a given task.
4627 * @p: the task in question.
4629 * Return: The priority value as seen by users in /proc.
4630 * RT tasks are offset by -200. Normal tasks are centered
4631 * around 0, value goes from -16 to +15.
4633 int task_prio(const struct task_struct *p)
4635 return p->prio - MAX_RT_PRIO;
4639 * idle_cpu - is a given CPU idle currently?
4640 * @cpu: the processor in question.
4642 * Return: 1 if the CPU is currently idle. 0 otherwise.
4644 int idle_cpu(int cpu)
4646 struct rq *rq = cpu_rq(cpu);
4648 if (rq->curr != rq->idle)
4655 if (!llist_empty(&rq->wake_list))
4663 * available_idle_cpu - is a given CPU idle for enqueuing work.
4664 * @cpu: the CPU in question.
4666 * Return: 1 if the CPU is currently idle. 0 otherwise.
4668 int available_idle_cpu(int cpu)
4673 if (vcpu_is_preempted(cpu))
4680 * idle_task - return the idle task for a given CPU.
4681 * @cpu: the processor in question.
4683 * Return: The idle task for the CPU @cpu.
4685 struct task_struct *idle_task(int cpu)
4687 return cpu_rq(cpu)->idle;
4691 * find_process_by_pid - find a process with a matching PID value.
4692 * @pid: the pid in question.
4694 * The task of @pid, if found. %NULL otherwise.
4696 static struct task_struct *find_process_by_pid(pid_t pid)
4698 return pid ? find_task_by_vpid(pid) : current;
4702 * sched_setparam() passes in -1 for its policy, to let the functions
4703 * it calls know not to change it.
4705 #define SETPARAM_POLICY -1
4707 static void __setscheduler_params(struct task_struct *p,
4708 const struct sched_attr *attr)
4710 int policy = attr->sched_policy;
4712 if (policy == SETPARAM_POLICY)
4717 if (dl_policy(policy))
4718 __setparam_dl(p, attr);
4719 else if (fair_policy(policy))
4720 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4723 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4724 * !rt_policy. Always setting this ensures that things like
4725 * getparam()/getattr() don't report silly values for !rt tasks.
4727 p->rt_priority = attr->sched_priority;
4728 p->normal_prio = normal_prio(p);
4729 set_load_weight(p, true);
4732 /* Actually do priority change: must hold pi & rq lock. */
4733 static void __setscheduler(struct rq *rq, struct task_struct *p,
4734 const struct sched_attr *attr, bool keep_boost)
4737 * If params can't change scheduling class changes aren't allowed
4740 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4743 __setscheduler_params(p, attr);
4746 * Keep a potential priority boosting if called from
4747 * sched_setscheduler().
4749 p->prio = normal_prio(p);
4751 p->prio = rt_effective_prio(p, p->prio);
4753 if (dl_prio(p->prio))
4754 p->sched_class = &dl_sched_class;
4755 else if (rt_prio(p->prio))
4756 p->sched_class = &rt_sched_class;
4758 p->sched_class = &fair_sched_class;
4762 * Check the target process has a UID that matches the current process's:
4764 static bool check_same_owner(struct task_struct *p)
4766 const struct cred *cred = current_cred(), *pcred;
4770 pcred = __task_cred(p);
4771 match = (uid_eq(cred->euid, pcred->euid) ||
4772 uid_eq(cred->euid, pcred->uid));
4777 static int __sched_setscheduler(struct task_struct *p,
4778 const struct sched_attr *attr,
4781 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4782 MAX_RT_PRIO - 1 - attr->sched_priority;
4783 int retval, oldprio, oldpolicy = -1, queued, running;
4784 int new_effective_prio, policy = attr->sched_policy;
4785 const struct sched_class *prev_class;
4788 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4791 /* The pi code expects interrupts enabled */
4792 BUG_ON(pi && in_interrupt());
4794 /* Double check policy once rq lock held: */
4796 reset_on_fork = p->sched_reset_on_fork;
4797 policy = oldpolicy = p->policy;
4799 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4801 if (!valid_policy(policy))
4805 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4809 * Valid priorities for SCHED_FIFO and SCHED_RR are
4810 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4811 * SCHED_BATCH and SCHED_IDLE is 0.
4813 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4814 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4816 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4817 (rt_policy(policy) != (attr->sched_priority != 0)))
4821 * Allow unprivileged RT tasks to decrease priority:
4823 if (user && !capable(CAP_SYS_NICE)) {
4824 if (fair_policy(policy)) {
4825 if (attr->sched_nice < task_nice(p) &&
4826 !can_nice(p, attr->sched_nice))
4830 if (rt_policy(policy)) {
4831 unsigned long rlim_rtprio =
4832 task_rlimit(p, RLIMIT_RTPRIO);
4834 /* Can't set/change the rt policy: */
4835 if (policy != p->policy && !rlim_rtprio)
4838 /* Can't increase priority: */
4839 if (attr->sched_priority > p->rt_priority &&
4840 attr->sched_priority > rlim_rtprio)
4845 * Can't set/change SCHED_DEADLINE policy at all for now
4846 * (safest behavior); in the future we would like to allow
4847 * unprivileged DL tasks to increase their relative deadline
4848 * or reduce their runtime (both ways reducing utilization)
4850 if (dl_policy(policy))
4854 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4855 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4857 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4858 if (!can_nice(p, task_nice(p)))
4862 /* Can't change other user's priorities: */
4863 if (!check_same_owner(p))
4866 /* Normal users shall not reset the sched_reset_on_fork flag: */
4867 if (p->sched_reset_on_fork && !reset_on_fork)
4872 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4875 retval = security_task_setscheduler(p);
4880 /* Update task specific "requested" clamps */
4881 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4882 retval = uclamp_validate(p, attr);
4891 * Make sure no PI-waiters arrive (or leave) while we are
4892 * changing the priority of the task:
4894 * To be able to change p->policy safely, the appropriate
4895 * runqueue lock must be held.
4897 rq = task_rq_lock(p, &rf);
4898 update_rq_clock(rq);
4901 * Changing the policy of the stop threads its a very bad idea:
4903 if (p == rq->stop) {
4909 * If not changing anything there's no need to proceed further,
4910 * but store a possible modification of reset_on_fork.
4912 if (unlikely(policy == p->policy)) {
4913 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4915 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4917 if (dl_policy(policy) && dl_param_changed(p, attr))
4919 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4922 p->sched_reset_on_fork = reset_on_fork;
4929 #ifdef CONFIG_RT_GROUP_SCHED
4931 * Do not allow realtime tasks into groups that have no runtime
4934 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4935 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4936 !task_group_is_autogroup(task_group(p))) {
4942 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4943 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4944 cpumask_t *span = rq->rd->span;
4947 * Don't allow tasks with an affinity mask smaller than
4948 * the entire root_domain to become SCHED_DEADLINE. We
4949 * will also fail if there's no bandwidth available.
4951 if (!cpumask_subset(span, p->cpus_ptr) ||
4952 rq->rd->dl_bw.bw == 0) {
4960 /* Re-check policy now with rq lock held: */
4961 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4962 policy = oldpolicy = -1;
4963 task_rq_unlock(rq, p, &rf);
4965 cpuset_read_unlock();
4970 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4971 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4974 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4979 p->sched_reset_on_fork = reset_on_fork;
4984 * Take priority boosted tasks into account. If the new
4985 * effective priority is unchanged, we just store the new
4986 * normal parameters and do not touch the scheduler class and
4987 * the runqueue. This will be done when the task deboost
4990 new_effective_prio = rt_effective_prio(p, newprio);
4991 if (new_effective_prio == oldprio)
4992 queue_flags &= ~DEQUEUE_MOVE;
4995 queued = task_on_rq_queued(p);
4996 running = task_current(rq, p);
4998 dequeue_task(rq, p, queue_flags);
5000 put_prev_task(rq, p);
5002 prev_class = p->sched_class;
5004 __setscheduler(rq, p, attr, pi);
5005 __setscheduler_uclamp(p, attr);
5009 * We enqueue to tail when the priority of a task is
5010 * increased (user space view).
5012 if (oldprio < p->prio)
5013 queue_flags |= ENQUEUE_HEAD;
5015 enqueue_task(rq, p, queue_flags);
5018 set_next_task(rq, p);
5020 check_class_changed(rq, p, prev_class, oldprio);
5022 /* Avoid rq from going away on us: */
5024 task_rq_unlock(rq, p, &rf);
5027 cpuset_read_unlock();
5028 rt_mutex_adjust_pi(p);
5031 /* Run balance callbacks after we've adjusted the PI chain: */
5032 balance_callback(rq);
5038 task_rq_unlock(rq, p, &rf);
5040 cpuset_read_unlock();
5044 static int _sched_setscheduler(struct task_struct *p, int policy,
5045 const struct sched_param *param, bool check)
5047 struct sched_attr attr = {
5048 .sched_policy = policy,
5049 .sched_priority = param->sched_priority,
5050 .sched_nice = PRIO_TO_NICE(p->static_prio),
5053 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5054 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5055 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5056 policy &= ~SCHED_RESET_ON_FORK;
5057 attr.sched_policy = policy;
5060 return __sched_setscheduler(p, &attr, check, true);
5063 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5064 * @p: the task in question.
5065 * @policy: new policy.
5066 * @param: structure containing the new RT priority.
5068 * Return: 0 on success. An error code otherwise.
5070 * NOTE that the task may be already dead.
5072 int sched_setscheduler(struct task_struct *p, int policy,
5073 const struct sched_param *param)
5075 return _sched_setscheduler(p, policy, param, true);
5077 EXPORT_SYMBOL_GPL(sched_setscheduler);
5079 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5081 return __sched_setscheduler(p, attr, true, true);
5083 EXPORT_SYMBOL_GPL(sched_setattr);
5085 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5087 return __sched_setscheduler(p, attr, false, true);
5091 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5092 * @p: the task in question.
5093 * @policy: new policy.
5094 * @param: structure containing the new RT priority.
5096 * Just like sched_setscheduler, only don't bother checking if the
5097 * current context has permission. For example, this is needed in
5098 * stop_machine(): we create temporary high priority worker threads,
5099 * but our caller might not have that capability.
5101 * Return: 0 on success. An error code otherwise.
5103 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5104 const struct sched_param *param)
5106 return _sched_setscheduler(p, policy, param, false);
5108 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5111 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5113 struct sched_param lparam;
5114 struct task_struct *p;
5117 if (!param || pid < 0)
5119 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5124 p = find_process_by_pid(pid);
5130 retval = sched_setscheduler(p, policy, &lparam);
5138 * Mimics kernel/events/core.c perf_copy_attr().
5140 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5145 /* Zero the full structure, so that a short copy will be nice: */
5146 memset(attr, 0, sizeof(*attr));
5148 ret = get_user(size, &uattr->size);
5152 /* ABI compatibility quirk: */
5154 size = SCHED_ATTR_SIZE_VER0;
5155 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5158 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5165 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5166 size < SCHED_ATTR_SIZE_VER1)
5170 * XXX: Do we want to be lenient like existing syscalls; or do we want
5171 * to be strict and return an error on out-of-bounds values?
5173 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5178 put_user(sizeof(*attr), &uattr->size);
5183 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5184 * @pid: the pid in question.
5185 * @policy: new policy.
5186 * @param: structure containing the new RT priority.
5188 * Return: 0 on success. An error code otherwise.
5190 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5195 return do_sched_setscheduler(pid, policy, param);
5199 * sys_sched_setparam - set/change the RT priority of a thread
5200 * @pid: the pid in question.
5201 * @param: structure containing the new RT priority.
5203 * Return: 0 on success. An error code otherwise.
5205 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5207 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5211 * sys_sched_setattr - same as above, but with extended sched_attr
5212 * @pid: the pid in question.
5213 * @uattr: structure containing the extended parameters.
5214 * @flags: for future extension.
5216 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5217 unsigned int, flags)
5219 struct sched_attr attr;
5220 struct task_struct *p;
5223 if (!uattr || pid < 0 || flags)
5226 retval = sched_copy_attr(uattr, &attr);
5230 if ((int)attr.sched_policy < 0)
5232 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5233 attr.sched_policy = SETPARAM_POLICY;
5237 p = find_process_by_pid(pid);
5243 retval = sched_setattr(p, &attr);
5251 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5252 * @pid: the pid in question.
5254 * Return: On success, the policy of the thread. Otherwise, a negative error
5257 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5259 struct task_struct *p;
5267 p = find_process_by_pid(pid);
5269 retval = security_task_getscheduler(p);
5272 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5279 * sys_sched_getparam - get the RT priority of a thread
5280 * @pid: the pid in question.
5281 * @param: structure containing the RT priority.
5283 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5286 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5288 struct sched_param lp = { .sched_priority = 0 };
5289 struct task_struct *p;
5292 if (!param || pid < 0)
5296 p = find_process_by_pid(pid);
5301 retval = security_task_getscheduler(p);
5305 if (task_has_rt_policy(p))
5306 lp.sched_priority = p->rt_priority;
5310 * This one might sleep, we cannot do it with a spinlock held ...
5312 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5322 * Copy the kernel size attribute structure (which might be larger
5323 * than what user-space knows about) to user-space.
5325 * Note that all cases are valid: user-space buffer can be larger or
5326 * smaller than the kernel-space buffer. The usual case is that both
5327 * have the same size.
5330 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5331 struct sched_attr *kattr,
5334 unsigned int ksize = sizeof(*kattr);
5336 if (!access_ok(uattr, usize))
5340 * sched_getattr() ABI forwards and backwards compatibility:
5342 * If usize == ksize then we just copy everything to user-space and all is good.
5344 * If usize < ksize then we only copy as much as user-space has space for,
5345 * this keeps ABI compatibility as well. We skip the rest.
5347 * If usize > ksize then user-space is using a newer version of the ABI,
5348 * which part the kernel doesn't know about. Just ignore it - tooling can
5349 * detect the kernel's knowledge of attributes from the attr->size value
5350 * which is set to ksize in this case.
5352 kattr->size = min(usize, ksize);
5354 if (copy_to_user(uattr, kattr, kattr->size))
5361 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5362 * @pid: the pid in question.
5363 * @uattr: structure containing the extended parameters.
5364 * @usize: sizeof(attr) for fwd/bwd comp.
5365 * @flags: for future extension.
5367 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5368 unsigned int, usize, unsigned int, flags)
5370 struct sched_attr kattr = { };
5371 struct task_struct *p;
5374 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5375 usize < SCHED_ATTR_SIZE_VER0 || flags)
5379 p = find_process_by_pid(pid);
5384 retval = security_task_getscheduler(p);
5388 kattr.sched_policy = p->policy;
5389 if (p->sched_reset_on_fork)
5390 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5391 if (task_has_dl_policy(p))
5392 __getparam_dl(p, &kattr);
5393 else if (task_has_rt_policy(p))
5394 kattr.sched_priority = p->rt_priority;
5396 kattr.sched_nice = task_nice(p);
5398 #ifdef CONFIG_UCLAMP_TASK
5399 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5400 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5405 return sched_attr_copy_to_user(uattr, &kattr, usize);
5412 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5414 cpumask_var_t cpus_allowed, new_mask;
5415 struct task_struct *p;
5420 p = find_process_by_pid(pid);
5426 /* Prevent p going away */
5430 if (p->flags & PF_NO_SETAFFINITY) {
5434 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5438 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5440 goto out_free_cpus_allowed;
5443 if (!check_same_owner(p)) {
5445 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5447 goto out_free_new_mask;
5452 retval = security_task_setscheduler(p);
5454 goto out_free_new_mask;
5457 cpuset_cpus_allowed(p, cpus_allowed);
5458 cpumask_and(new_mask, in_mask, cpus_allowed);
5461 * Since bandwidth control happens on root_domain basis,
5462 * if admission test is enabled, we only admit -deadline
5463 * tasks allowed to run on all the CPUs in the task's
5467 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5469 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5472 goto out_free_new_mask;
5478 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5481 cpuset_cpus_allowed(p, cpus_allowed);
5482 if (!cpumask_subset(new_mask, cpus_allowed)) {
5484 * We must have raced with a concurrent cpuset
5485 * update. Just reset the cpus_allowed to the
5486 * cpuset's cpus_allowed
5488 cpumask_copy(new_mask, cpus_allowed);
5493 free_cpumask_var(new_mask);
5494 out_free_cpus_allowed:
5495 free_cpumask_var(cpus_allowed);
5501 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5502 struct cpumask *new_mask)
5504 if (len < cpumask_size())
5505 cpumask_clear(new_mask);
5506 else if (len > cpumask_size())
5507 len = cpumask_size();
5509 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5513 * sys_sched_setaffinity - set the CPU affinity of a process
5514 * @pid: pid of the process
5515 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5516 * @user_mask_ptr: user-space pointer to the new CPU mask
5518 * Return: 0 on success. An error code otherwise.
5520 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5521 unsigned long __user *, user_mask_ptr)
5523 cpumask_var_t new_mask;
5526 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5529 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5531 retval = sched_setaffinity(pid, new_mask);
5532 free_cpumask_var(new_mask);
5536 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5538 struct task_struct *p;
5539 unsigned long flags;
5545 p = find_process_by_pid(pid);
5549 retval = security_task_getscheduler(p);
5553 raw_spin_lock_irqsave(&p->pi_lock, flags);
5554 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5555 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5564 * sys_sched_getaffinity - get the CPU affinity of a process
5565 * @pid: pid of the process
5566 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5567 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5569 * Return: size of CPU mask copied to user_mask_ptr on success. An
5570 * error code otherwise.
5572 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5573 unsigned long __user *, user_mask_ptr)
5578 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5580 if (len & (sizeof(unsigned long)-1))
5583 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5586 ret = sched_getaffinity(pid, mask);
5588 unsigned int retlen = min(len, cpumask_size());
5590 if (copy_to_user(user_mask_ptr, mask, retlen))
5595 free_cpumask_var(mask);
5601 * sys_sched_yield - yield the current processor to other threads.
5603 * This function yields the current CPU to other tasks. If there are no
5604 * other threads running on this CPU then this function will return.
5608 static void do_sched_yield(void)
5613 rq = this_rq_lock_irq(&rf);
5615 schedstat_inc(rq->yld_count);
5616 current->sched_class->yield_task(rq);
5619 * Since we are going to call schedule() anyway, there's
5620 * no need to preempt or enable interrupts:
5624 sched_preempt_enable_no_resched();
5629 SYSCALL_DEFINE0(sched_yield)
5635 #ifndef CONFIG_PREEMPTION
5636 int __sched _cond_resched(void)
5638 if (should_resched(0)) {
5639 preempt_schedule_common();
5645 EXPORT_SYMBOL(_cond_resched);
5649 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5650 * call schedule, and on return reacquire the lock.
5652 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5653 * operations here to prevent schedule() from being called twice (once via
5654 * spin_unlock(), once by hand).
5656 int __cond_resched_lock(spinlock_t *lock)
5658 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5661 lockdep_assert_held(lock);
5663 if (spin_needbreak(lock) || resched) {
5666 preempt_schedule_common();
5674 EXPORT_SYMBOL(__cond_resched_lock);
5677 * yield - yield the current processor to other threads.
5679 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5681 * The scheduler is at all times free to pick the calling task as the most
5682 * eligible task to run, if removing the yield() call from your code breaks
5683 * it, its already broken.
5685 * Typical broken usage is:
5690 * where one assumes that yield() will let 'the other' process run that will
5691 * make event true. If the current task is a SCHED_FIFO task that will never
5692 * happen. Never use yield() as a progress guarantee!!
5694 * If you want to use yield() to wait for something, use wait_event().
5695 * If you want to use yield() to be 'nice' for others, use cond_resched().
5696 * If you still want to use yield(), do not!
5698 void __sched yield(void)
5700 set_current_state(TASK_RUNNING);
5703 EXPORT_SYMBOL(yield);
5706 * yield_to - yield the current processor to another thread in
5707 * your thread group, or accelerate that thread toward the
5708 * processor it's on.
5710 * @preempt: whether task preemption is allowed or not
5712 * It's the caller's job to ensure that the target task struct
5713 * can't go away on us before we can do any checks.
5716 * true (>0) if we indeed boosted the target task.
5717 * false (0) if we failed to boost the target.
5718 * -ESRCH if there's no task to yield to.
5720 int __sched yield_to(struct task_struct *p, bool preempt)
5722 struct task_struct *curr = current;
5723 struct rq *rq, *p_rq;
5724 unsigned long flags;
5727 local_irq_save(flags);
5733 * If we're the only runnable task on the rq and target rq also
5734 * has only one task, there's absolutely no point in yielding.
5736 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5741 double_rq_lock(rq, p_rq);
5742 if (task_rq(p) != p_rq) {
5743 double_rq_unlock(rq, p_rq);
5747 if (!curr->sched_class->yield_to_task)
5750 if (curr->sched_class != p->sched_class)
5753 if (task_running(p_rq, p) || p->state)
5756 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5758 schedstat_inc(rq->yld_count);
5760 * Make p's CPU reschedule; pick_next_entity takes care of
5763 if (preempt && rq != p_rq)
5768 double_rq_unlock(rq, p_rq);
5770 local_irq_restore(flags);
5777 EXPORT_SYMBOL_GPL(yield_to);
5779 int io_schedule_prepare(void)
5781 int old_iowait = current->in_iowait;
5783 current->in_iowait = 1;
5784 blk_schedule_flush_plug(current);
5789 void io_schedule_finish(int token)
5791 current->in_iowait = token;
5795 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5796 * that process accounting knows that this is a task in IO wait state.
5798 long __sched io_schedule_timeout(long timeout)
5803 token = io_schedule_prepare();
5804 ret = schedule_timeout(timeout);
5805 io_schedule_finish(token);
5809 EXPORT_SYMBOL(io_schedule_timeout);
5811 void __sched io_schedule(void)
5815 token = io_schedule_prepare();
5817 io_schedule_finish(token);
5819 EXPORT_SYMBOL(io_schedule);
5822 * sys_sched_get_priority_max - return maximum RT priority.
5823 * @policy: scheduling class.
5825 * Return: On success, this syscall returns the maximum
5826 * rt_priority that can be used by a given scheduling class.
5827 * On failure, a negative error code is returned.
5829 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5836 ret = MAX_USER_RT_PRIO-1;
5838 case SCHED_DEADLINE:
5849 * sys_sched_get_priority_min - return minimum RT priority.
5850 * @policy: scheduling class.
5852 * Return: On success, this syscall returns the minimum
5853 * rt_priority that can be used by a given scheduling class.
5854 * On failure, a negative error code is returned.
5856 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5865 case SCHED_DEADLINE:
5874 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5876 struct task_struct *p;
5877 unsigned int time_slice;
5887 p = find_process_by_pid(pid);
5891 retval = security_task_getscheduler(p);
5895 rq = task_rq_lock(p, &rf);
5897 if (p->sched_class->get_rr_interval)
5898 time_slice = p->sched_class->get_rr_interval(rq, p);
5899 task_rq_unlock(rq, p, &rf);
5902 jiffies_to_timespec64(time_slice, t);
5911 * sys_sched_rr_get_interval - return the default timeslice of a process.
5912 * @pid: pid of the process.
5913 * @interval: userspace pointer to the timeslice value.
5915 * this syscall writes the default timeslice value of a given process
5916 * into the user-space timespec buffer. A value of '0' means infinity.
5918 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5921 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5922 struct __kernel_timespec __user *, interval)
5924 struct timespec64 t;
5925 int retval = sched_rr_get_interval(pid, &t);
5928 retval = put_timespec64(&t, interval);
5933 #ifdef CONFIG_COMPAT_32BIT_TIME
5934 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5935 struct old_timespec32 __user *, interval)
5937 struct timespec64 t;
5938 int retval = sched_rr_get_interval(pid, &t);
5941 retval = put_old_timespec32(&t, interval);
5946 void sched_show_task(struct task_struct *p)
5948 unsigned long free = 0;
5951 if (!try_get_task_stack(p))
5954 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5956 if (p->state == TASK_RUNNING)
5957 printk(KERN_CONT " running task ");
5958 #ifdef CONFIG_DEBUG_STACK_USAGE
5959 free = stack_not_used(p);
5964 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5966 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5967 task_pid_nr(p), ppid,
5968 (unsigned long)task_thread_info(p)->flags);
5970 print_worker_info(KERN_INFO, p);
5971 show_stack(p, NULL);
5974 EXPORT_SYMBOL_GPL(sched_show_task);
5977 state_filter_match(unsigned long state_filter, struct task_struct *p)
5979 /* no filter, everything matches */
5983 /* filter, but doesn't match */
5984 if (!(p->state & state_filter))
5988 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5991 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5998 void show_state_filter(unsigned long state_filter)
6000 struct task_struct *g, *p;
6002 #if BITS_PER_LONG == 32
6004 " task PC stack pid father\n");
6007 " task PC stack pid father\n");
6010 for_each_process_thread(g, p) {
6012 * reset the NMI-timeout, listing all files on a slow
6013 * console might take a lot of time:
6014 * Also, reset softlockup watchdogs on all CPUs, because
6015 * another CPU might be blocked waiting for us to process
6018 touch_nmi_watchdog();
6019 touch_all_softlockup_watchdogs();
6020 if (state_filter_match(state_filter, p))
6024 #ifdef CONFIG_SCHED_DEBUG
6026 sysrq_sched_debug_show();
6030 * Only show locks if all tasks are dumped:
6033 debug_show_all_locks();
6037 * init_idle - set up an idle thread for a given CPU
6038 * @idle: task in question
6039 * @cpu: CPU the idle task belongs to
6041 * NOTE: this function does not set the idle thread's NEED_RESCHED
6042 * flag, to make booting more robust.
6044 void init_idle(struct task_struct *idle, int cpu)
6046 struct rq *rq = cpu_rq(cpu);
6047 unsigned long flags;
6049 __sched_fork(0, idle);
6051 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6052 raw_spin_lock(&rq->lock);
6054 idle->state = TASK_RUNNING;
6055 idle->se.exec_start = sched_clock();
6056 idle->flags |= PF_IDLE;
6058 kasan_unpoison_task_stack(idle);
6062 * Its possible that init_idle() gets called multiple times on a task,
6063 * in that case do_set_cpus_allowed() will not do the right thing.
6065 * And since this is boot we can forgo the serialization.
6067 set_cpus_allowed_common(idle, cpumask_of(cpu));
6070 * We're having a chicken and egg problem, even though we are
6071 * holding rq->lock, the CPU isn't yet set to this CPU so the
6072 * lockdep check in task_group() will fail.
6074 * Similar case to sched_fork(). / Alternatively we could
6075 * use task_rq_lock() here and obtain the other rq->lock.
6080 __set_task_cpu(idle, cpu);
6084 rcu_assign_pointer(rq->curr, idle);
6085 idle->on_rq = TASK_ON_RQ_QUEUED;
6089 raw_spin_unlock(&rq->lock);
6090 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6092 /* Set the preempt count _outside_ the spinlocks! */
6093 init_idle_preempt_count(idle, cpu);
6096 * The idle tasks have their own, simple scheduling class:
6098 idle->sched_class = &idle_sched_class;
6099 ftrace_graph_init_idle_task(idle, cpu);
6100 vtime_init_idle(idle, cpu);
6102 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6108 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6109 const struct cpumask *trial)
6113 if (!cpumask_weight(cur))
6116 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6121 int task_can_attach(struct task_struct *p,
6122 const struct cpumask *cs_cpus_allowed)
6127 * Kthreads which disallow setaffinity shouldn't be moved
6128 * to a new cpuset; we don't want to change their CPU
6129 * affinity and isolating such threads by their set of
6130 * allowed nodes is unnecessary. Thus, cpusets are not
6131 * applicable for such threads. This prevents checking for
6132 * success of set_cpus_allowed_ptr() on all attached tasks
6133 * before cpus_mask may be changed.
6135 if (p->flags & PF_NO_SETAFFINITY) {
6140 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6142 ret = dl_task_can_attach(p, cs_cpus_allowed);
6148 bool sched_smp_initialized __read_mostly;
6150 #ifdef CONFIG_NUMA_BALANCING
6151 /* Migrate current task p to target_cpu */
6152 int migrate_task_to(struct task_struct *p, int target_cpu)
6154 struct migration_arg arg = { p, target_cpu };
6155 int curr_cpu = task_cpu(p);
6157 if (curr_cpu == target_cpu)
6160 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6163 /* TODO: This is not properly updating schedstats */
6165 trace_sched_move_numa(p, curr_cpu, target_cpu);
6166 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6170 * Requeue a task on a given node and accurately track the number of NUMA
6171 * tasks on the runqueues
6173 void sched_setnuma(struct task_struct *p, int nid)
6175 bool queued, running;
6179 rq = task_rq_lock(p, &rf);
6180 queued = task_on_rq_queued(p);
6181 running = task_current(rq, p);
6184 dequeue_task(rq, p, DEQUEUE_SAVE);
6186 put_prev_task(rq, p);
6188 p->numa_preferred_nid = nid;
6191 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6193 set_next_task(rq, p);
6194 task_rq_unlock(rq, p, &rf);
6196 #endif /* CONFIG_NUMA_BALANCING */
6198 #ifdef CONFIG_HOTPLUG_CPU
6200 * Ensure that the idle task is using init_mm right before its CPU goes
6203 void idle_task_exit(void)
6205 struct mm_struct *mm = current->active_mm;
6207 BUG_ON(cpu_online(smp_processor_id()));
6209 if (mm != &init_mm) {
6210 switch_mm(mm, &init_mm, current);
6211 current->active_mm = &init_mm;
6212 finish_arch_post_lock_switch();
6218 * Since this CPU is going 'away' for a while, fold any nr_active delta
6219 * we might have. Assumes we're called after migrate_tasks() so that the
6220 * nr_active count is stable. We need to take the teardown thread which
6221 * is calling this into account, so we hand in adjust = 1 to the load
6224 * Also see the comment "Global load-average calculations".
6226 static void calc_load_migrate(struct rq *rq)
6228 long delta = calc_load_fold_active(rq, 1);
6230 atomic_long_add(delta, &calc_load_tasks);
6233 static struct task_struct *__pick_migrate_task(struct rq *rq)
6235 const struct sched_class *class;
6236 struct task_struct *next;
6238 for_each_class(class) {
6239 next = class->pick_next_task(rq);
6241 next->sched_class->put_prev_task(rq, next);
6246 /* The idle class should always have a runnable task */
6251 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6252 * try_to_wake_up()->select_task_rq().
6254 * Called with rq->lock held even though we'er in stop_machine() and
6255 * there's no concurrency possible, we hold the required locks anyway
6256 * because of lock validation efforts.
6258 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6260 struct rq *rq = dead_rq;
6261 struct task_struct *next, *stop = rq->stop;
6262 struct rq_flags orf = *rf;
6266 * Fudge the rq selection such that the below task selection loop
6267 * doesn't get stuck on the currently eligible stop task.
6269 * We're currently inside stop_machine() and the rq is either stuck
6270 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6271 * either way we should never end up calling schedule() until we're
6277 * put_prev_task() and pick_next_task() sched
6278 * class method both need to have an up-to-date
6279 * value of rq->clock[_task]
6281 update_rq_clock(rq);
6285 * There's this thread running, bail when that's the only
6288 if (rq->nr_running == 1)
6291 next = __pick_migrate_task(rq);
6294 * Rules for changing task_struct::cpus_mask are holding
6295 * both pi_lock and rq->lock, such that holding either
6296 * stabilizes the mask.
6298 * Drop rq->lock is not quite as disastrous as it usually is
6299 * because !cpu_active at this point, which means load-balance
6300 * will not interfere. Also, stop-machine.
6303 raw_spin_lock(&next->pi_lock);
6307 * Since we're inside stop-machine, _nothing_ should have
6308 * changed the task, WARN if weird stuff happened, because in
6309 * that case the above rq->lock drop is a fail too.
6311 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6312 raw_spin_unlock(&next->pi_lock);
6316 /* Find suitable destination for @next, with force if needed. */
6317 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6318 rq = __migrate_task(rq, rf, next, dest_cpu);
6319 if (rq != dead_rq) {
6325 raw_spin_unlock(&next->pi_lock);
6330 #endif /* CONFIG_HOTPLUG_CPU */
6332 void set_rq_online(struct rq *rq)
6335 const struct sched_class *class;
6337 cpumask_set_cpu(rq->cpu, rq->rd->online);
6340 for_each_class(class) {
6341 if (class->rq_online)
6342 class->rq_online(rq);
6347 void set_rq_offline(struct rq *rq)
6350 const struct sched_class *class;
6352 for_each_class(class) {
6353 if (class->rq_offline)
6354 class->rq_offline(rq);
6357 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6363 * used to mark begin/end of suspend/resume:
6365 static int num_cpus_frozen;
6368 * Update cpusets according to cpu_active mask. If cpusets are
6369 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6370 * around partition_sched_domains().
6372 * If we come here as part of a suspend/resume, don't touch cpusets because we
6373 * want to restore it back to its original state upon resume anyway.
6375 static void cpuset_cpu_active(void)
6377 if (cpuhp_tasks_frozen) {
6379 * num_cpus_frozen tracks how many CPUs are involved in suspend
6380 * resume sequence. As long as this is not the last online
6381 * operation in the resume sequence, just build a single sched
6382 * domain, ignoring cpusets.
6384 partition_sched_domains(1, NULL, NULL);
6385 if (--num_cpus_frozen)
6388 * This is the last CPU online operation. So fall through and
6389 * restore the original sched domains by considering the
6390 * cpuset configurations.
6392 cpuset_force_rebuild();
6394 cpuset_update_active_cpus();
6397 static int cpuset_cpu_inactive(unsigned int cpu)
6399 if (!cpuhp_tasks_frozen) {
6400 if (dl_cpu_busy(cpu))
6402 cpuset_update_active_cpus();
6405 partition_sched_domains(1, NULL, NULL);
6410 int sched_cpu_activate(unsigned int cpu)
6412 struct rq *rq = cpu_rq(cpu);
6415 #ifdef CONFIG_SCHED_SMT
6417 * When going up, increment the number of cores with SMT present.
6419 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6420 static_branch_inc_cpuslocked(&sched_smt_present);
6422 set_cpu_active(cpu, true);
6424 if (sched_smp_initialized) {
6425 sched_domains_numa_masks_set(cpu);
6426 cpuset_cpu_active();
6430 * Put the rq online, if not already. This happens:
6432 * 1) In the early boot process, because we build the real domains
6433 * after all CPUs have been brought up.
6435 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6438 rq_lock_irqsave(rq, &rf);
6440 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6443 rq_unlock_irqrestore(rq, &rf);
6448 int sched_cpu_deactivate(unsigned int cpu)
6452 set_cpu_active(cpu, false);
6454 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6455 * users of this state to go away such that all new such users will
6458 * Do sync before park smpboot threads to take care the rcu boost case.
6462 #ifdef CONFIG_SCHED_SMT
6464 * When going down, decrement the number of cores with SMT present.
6466 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6467 static_branch_dec_cpuslocked(&sched_smt_present);
6470 if (!sched_smp_initialized)
6473 ret = cpuset_cpu_inactive(cpu);
6475 set_cpu_active(cpu, true);
6478 sched_domains_numa_masks_clear(cpu);
6482 static void sched_rq_cpu_starting(unsigned int cpu)
6484 struct rq *rq = cpu_rq(cpu);
6486 rq->calc_load_update = calc_load_update;
6487 update_max_interval();
6490 int sched_cpu_starting(unsigned int cpu)
6492 sched_rq_cpu_starting(cpu);
6493 sched_tick_start(cpu);
6497 #ifdef CONFIG_HOTPLUG_CPU
6498 int sched_cpu_dying(unsigned int cpu)
6500 struct rq *rq = cpu_rq(cpu);
6503 /* Handle pending wakeups and then migrate everything off */
6504 sched_ttwu_pending();
6505 sched_tick_stop(cpu);
6507 rq_lock_irqsave(rq, &rf);
6509 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6512 migrate_tasks(rq, &rf);
6513 BUG_ON(rq->nr_running != 1);
6514 rq_unlock_irqrestore(rq, &rf);
6516 calc_load_migrate(rq);
6517 update_max_interval();
6518 nohz_balance_exit_idle(rq);
6524 void __init sched_init_smp(void)
6529 * There's no userspace yet to cause hotplug operations; hence all the
6530 * CPU masks are stable and all blatant races in the below code cannot
6533 mutex_lock(&sched_domains_mutex);
6534 sched_init_domains(cpu_active_mask);
6535 mutex_unlock(&sched_domains_mutex);
6537 /* Move init over to a non-isolated CPU */
6538 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6540 sched_init_granularity();
6542 init_sched_rt_class();
6543 init_sched_dl_class();
6545 sched_smp_initialized = true;
6548 static int __init migration_init(void)
6550 sched_cpu_starting(smp_processor_id());
6553 early_initcall(migration_init);
6556 void __init sched_init_smp(void)
6558 sched_init_granularity();
6560 #endif /* CONFIG_SMP */
6562 int in_sched_functions(unsigned long addr)
6564 return in_lock_functions(addr) ||
6565 (addr >= (unsigned long)__sched_text_start
6566 && addr < (unsigned long)__sched_text_end);
6569 #ifdef CONFIG_CGROUP_SCHED
6571 * Default task group.
6572 * Every task in system belongs to this group at bootup.
6574 struct task_group root_task_group;
6575 LIST_HEAD(task_groups);
6577 /* Cacheline aligned slab cache for task_group */
6578 static struct kmem_cache *task_group_cache __read_mostly;
6581 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6582 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6584 void __init sched_init(void)
6586 unsigned long ptr = 0;
6591 #ifdef CONFIG_FAIR_GROUP_SCHED
6592 ptr += 2 * nr_cpu_ids * sizeof(void **);
6594 #ifdef CONFIG_RT_GROUP_SCHED
6595 ptr += 2 * nr_cpu_ids * sizeof(void **);
6598 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6600 #ifdef CONFIG_FAIR_GROUP_SCHED
6601 root_task_group.se = (struct sched_entity **)ptr;
6602 ptr += nr_cpu_ids * sizeof(void **);
6604 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6605 ptr += nr_cpu_ids * sizeof(void **);
6607 #endif /* CONFIG_FAIR_GROUP_SCHED */
6608 #ifdef CONFIG_RT_GROUP_SCHED
6609 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6610 ptr += nr_cpu_ids * sizeof(void **);
6612 root_task_group.rt_rq = (struct rt_rq **)ptr;
6613 ptr += nr_cpu_ids * sizeof(void **);
6615 #endif /* CONFIG_RT_GROUP_SCHED */
6617 #ifdef CONFIG_CPUMASK_OFFSTACK
6618 for_each_possible_cpu(i) {
6619 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6620 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6621 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6622 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6624 #endif /* CONFIG_CPUMASK_OFFSTACK */
6626 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6627 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6630 init_defrootdomain();
6633 #ifdef CONFIG_RT_GROUP_SCHED
6634 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6635 global_rt_period(), global_rt_runtime());
6636 #endif /* CONFIG_RT_GROUP_SCHED */
6638 #ifdef CONFIG_CGROUP_SCHED
6639 task_group_cache = KMEM_CACHE(task_group, 0);
6641 list_add(&root_task_group.list, &task_groups);
6642 INIT_LIST_HEAD(&root_task_group.children);
6643 INIT_LIST_HEAD(&root_task_group.siblings);
6644 autogroup_init(&init_task);
6645 #endif /* CONFIG_CGROUP_SCHED */
6647 for_each_possible_cpu(i) {
6651 raw_spin_lock_init(&rq->lock);
6653 rq->calc_load_active = 0;
6654 rq->calc_load_update = jiffies + LOAD_FREQ;
6655 init_cfs_rq(&rq->cfs);
6656 init_rt_rq(&rq->rt);
6657 init_dl_rq(&rq->dl);
6658 #ifdef CONFIG_FAIR_GROUP_SCHED
6659 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6660 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6661 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6663 * How much CPU bandwidth does root_task_group get?
6665 * In case of task-groups formed thr' the cgroup filesystem, it
6666 * gets 100% of the CPU resources in the system. This overall
6667 * system CPU resource is divided among the tasks of
6668 * root_task_group and its child task-groups in a fair manner,
6669 * based on each entity's (task or task-group's) weight
6670 * (se->load.weight).
6672 * In other words, if root_task_group has 10 tasks of weight
6673 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6674 * then A0's share of the CPU resource is:
6676 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6678 * We achieve this by letting root_task_group's tasks sit
6679 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6681 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6682 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6683 #endif /* CONFIG_FAIR_GROUP_SCHED */
6685 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6686 #ifdef CONFIG_RT_GROUP_SCHED
6687 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6692 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6693 rq->balance_callback = NULL;
6694 rq->active_balance = 0;
6695 rq->next_balance = jiffies;
6700 rq->avg_idle = 2*sysctl_sched_migration_cost;
6701 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6703 INIT_LIST_HEAD(&rq->cfs_tasks);
6705 rq_attach_root(rq, &def_root_domain);
6706 #ifdef CONFIG_NO_HZ_COMMON
6707 rq->last_load_update_tick = jiffies;
6708 rq->last_blocked_load_update_tick = jiffies;
6709 atomic_set(&rq->nohz_flags, 0);
6711 #endif /* CONFIG_SMP */
6713 atomic_set(&rq->nr_iowait, 0);
6716 set_load_weight(&init_task, false);
6719 * The boot idle thread does lazy MMU switching as well:
6722 enter_lazy_tlb(&init_mm, current);
6725 * Make us the idle thread. Technically, schedule() should not be
6726 * called from this thread, however somewhere below it might be,
6727 * but because we are the idle thread, we just pick up running again
6728 * when this runqueue becomes "idle".
6730 init_idle(current, smp_processor_id());
6732 calc_load_update = jiffies + LOAD_FREQ;
6735 idle_thread_set_boot_cpu();
6737 init_sched_fair_class();
6745 scheduler_running = 1;
6748 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6749 static inline int preempt_count_equals(int preempt_offset)
6751 int nested = preempt_count() + rcu_preempt_depth();
6753 return (nested == preempt_offset);
6756 void __might_sleep(const char *file, int line, int preempt_offset)
6759 * Blocking primitives will set (and therefore destroy) current->state,
6760 * since we will exit with TASK_RUNNING make sure we enter with it,
6761 * otherwise we will destroy state.
6763 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6764 "do not call blocking ops when !TASK_RUNNING; "
6765 "state=%lx set at [<%p>] %pS\n",
6767 (void *)current->task_state_change,
6768 (void *)current->task_state_change);
6770 ___might_sleep(file, line, preempt_offset);
6772 EXPORT_SYMBOL(__might_sleep);
6774 void ___might_sleep(const char *file, int line, int preempt_offset)
6776 /* Ratelimiting timestamp: */
6777 static unsigned long prev_jiffy;
6779 unsigned long preempt_disable_ip;
6781 /* WARN_ON_ONCE() by default, no rate limit required: */
6784 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6785 !is_idle_task(current) && !current->non_block_count) ||
6786 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6790 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6792 prev_jiffy = jiffies;
6794 /* Save this before calling printk(), since that will clobber it: */
6795 preempt_disable_ip = get_preempt_disable_ip(current);
6798 "BUG: sleeping function called from invalid context at %s:%d\n",
6801 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6802 in_atomic(), irqs_disabled(), current->non_block_count,
6803 current->pid, current->comm);
6805 if (task_stack_end_corrupted(current))
6806 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6808 debug_show_held_locks(current);
6809 if (irqs_disabled())
6810 print_irqtrace_events(current);
6811 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6812 && !preempt_count_equals(preempt_offset)) {
6813 pr_err("Preemption disabled at:");
6814 print_ip_sym(preempt_disable_ip);
6818 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6820 EXPORT_SYMBOL(___might_sleep);
6822 void __cant_sleep(const char *file, int line, int preempt_offset)
6824 static unsigned long prev_jiffy;
6826 if (irqs_disabled())
6829 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6832 if (preempt_count() > preempt_offset)
6835 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6837 prev_jiffy = jiffies;
6839 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6840 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6841 in_atomic(), irqs_disabled(),
6842 current->pid, current->comm);
6844 debug_show_held_locks(current);
6846 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6848 EXPORT_SYMBOL_GPL(__cant_sleep);
6851 #ifdef CONFIG_MAGIC_SYSRQ
6852 void normalize_rt_tasks(void)
6854 struct task_struct *g, *p;
6855 struct sched_attr attr = {
6856 .sched_policy = SCHED_NORMAL,
6859 read_lock(&tasklist_lock);
6860 for_each_process_thread(g, p) {
6862 * Only normalize user tasks:
6864 if (p->flags & PF_KTHREAD)
6867 p->se.exec_start = 0;
6868 schedstat_set(p->se.statistics.wait_start, 0);
6869 schedstat_set(p->se.statistics.sleep_start, 0);
6870 schedstat_set(p->se.statistics.block_start, 0);
6872 if (!dl_task(p) && !rt_task(p)) {
6874 * Renice negative nice level userspace
6877 if (task_nice(p) < 0)
6878 set_user_nice(p, 0);
6882 __sched_setscheduler(p, &attr, false, false);
6884 read_unlock(&tasklist_lock);
6887 #endif /* CONFIG_MAGIC_SYSRQ */
6889 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6891 * These functions are only useful for the IA64 MCA handling, or kdb.
6893 * They can only be called when the whole system has been
6894 * stopped - every CPU needs to be quiescent, and no scheduling
6895 * activity can take place. Using them for anything else would
6896 * be a serious bug, and as a result, they aren't even visible
6897 * under any other configuration.
6901 * curr_task - return the current task for a given CPU.
6902 * @cpu: the processor in question.
6904 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6906 * Return: The current task for @cpu.
6908 struct task_struct *curr_task(int cpu)
6910 return cpu_curr(cpu);
6913 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6917 * ia64_set_curr_task - set the current task for a given CPU.
6918 * @cpu: the processor in question.
6919 * @p: the task pointer to set.
6921 * Description: This function must only be used when non-maskable interrupts
6922 * are serviced on a separate stack. It allows the architecture to switch the
6923 * notion of the current task on a CPU in a non-blocking manner. This function
6924 * must be called with all CPU's synchronized, and interrupts disabled, the
6925 * and caller must save the original value of the current task (see
6926 * curr_task() above) and restore that value before reenabling interrupts and
6927 * re-starting the system.
6929 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6931 void ia64_set_curr_task(int cpu, struct task_struct *p)
6938 #ifdef CONFIG_CGROUP_SCHED
6939 /* task_group_lock serializes the addition/removal of task groups */
6940 static DEFINE_SPINLOCK(task_group_lock);
6942 static inline void alloc_uclamp_sched_group(struct task_group *tg,
6943 struct task_group *parent)
6945 #ifdef CONFIG_UCLAMP_TASK_GROUP
6946 enum uclamp_id clamp_id;
6948 for_each_clamp_id(clamp_id) {
6949 uclamp_se_set(&tg->uclamp_req[clamp_id],
6950 uclamp_none(clamp_id), false);
6951 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
6956 static void sched_free_group(struct task_group *tg)
6958 free_fair_sched_group(tg);
6959 free_rt_sched_group(tg);
6961 kmem_cache_free(task_group_cache, tg);
6964 /* allocate runqueue etc for a new task group */
6965 struct task_group *sched_create_group(struct task_group *parent)
6967 struct task_group *tg;
6969 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6971 return ERR_PTR(-ENOMEM);
6973 if (!alloc_fair_sched_group(tg, parent))
6976 if (!alloc_rt_sched_group(tg, parent))
6979 alloc_uclamp_sched_group(tg, parent);
6984 sched_free_group(tg);
6985 return ERR_PTR(-ENOMEM);
6988 void sched_online_group(struct task_group *tg, struct task_group *parent)
6990 unsigned long flags;
6992 spin_lock_irqsave(&task_group_lock, flags);
6993 list_add_rcu(&tg->list, &task_groups);
6995 /* Root should already exist: */
6998 tg->parent = parent;
6999 INIT_LIST_HEAD(&tg->children);
7000 list_add_rcu(&tg->siblings, &parent->children);
7001 spin_unlock_irqrestore(&task_group_lock, flags);
7003 online_fair_sched_group(tg);
7006 /* rcu callback to free various structures associated with a task group */
7007 static void sched_free_group_rcu(struct rcu_head *rhp)
7009 /* Now it should be safe to free those cfs_rqs: */
7010 sched_free_group(container_of(rhp, struct task_group, rcu));
7013 void sched_destroy_group(struct task_group *tg)
7015 /* Wait for possible concurrent references to cfs_rqs complete: */
7016 call_rcu(&tg->rcu, sched_free_group_rcu);
7019 void sched_offline_group(struct task_group *tg)
7021 unsigned long flags;
7023 /* End participation in shares distribution: */
7024 unregister_fair_sched_group(tg);
7026 spin_lock_irqsave(&task_group_lock, flags);
7027 list_del_rcu(&tg->list);
7028 list_del_rcu(&tg->siblings);
7029 spin_unlock_irqrestore(&task_group_lock, flags);
7032 static void sched_change_group(struct task_struct *tsk, int type)
7034 struct task_group *tg;
7037 * All callers are synchronized by task_rq_lock(); we do not use RCU
7038 * which is pointless here. Thus, we pass "true" to task_css_check()
7039 * to prevent lockdep warnings.
7041 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7042 struct task_group, css);
7043 tg = autogroup_task_group(tsk, tg);
7044 tsk->sched_task_group = tg;
7046 #ifdef CONFIG_FAIR_GROUP_SCHED
7047 if (tsk->sched_class->task_change_group)
7048 tsk->sched_class->task_change_group(tsk, type);
7051 set_task_rq(tsk, task_cpu(tsk));
7055 * Change task's runqueue when it moves between groups.
7057 * The caller of this function should have put the task in its new group by
7058 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7061 void sched_move_task(struct task_struct *tsk)
7063 int queued, running, queue_flags =
7064 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7068 rq = task_rq_lock(tsk, &rf);
7069 update_rq_clock(rq);
7071 running = task_current(rq, tsk);
7072 queued = task_on_rq_queued(tsk);
7075 dequeue_task(rq, tsk, queue_flags);
7077 put_prev_task(rq, tsk);
7079 sched_change_group(tsk, TASK_MOVE_GROUP);
7082 enqueue_task(rq, tsk, queue_flags);
7084 set_next_task(rq, tsk);
7086 * After changing group, the running task may have joined a
7087 * throttled one but it's still the running task. Trigger a
7088 * resched to make sure that task can still run.
7093 task_rq_unlock(rq, tsk, &rf);
7096 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7098 return css ? container_of(css, struct task_group, css) : NULL;
7101 static struct cgroup_subsys_state *
7102 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7104 struct task_group *parent = css_tg(parent_css);
7105 struct task_group *tg;
7108 /* This is early initialization for the top cgroup */
7109 return &root_task_group.css;
7112 tg = sched_create_group(parent);
7114 return ERR_PTR(-ENOMEM);
7119 /* Expose task group only after completing cgroup initialization */
7120 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7122 struct task_group *tg = css_tg(css);
7123 struct task_group *parent = css_tg(css->parent);
7126 sched_online_group(tg, parent);
7128 #ifdef CONFIG_UCLAMP_TASK_GROUP
7129 /* Propagate the effective uclamp value for the new group */
7130 cpu_util_update_eff(css);
7136 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7138 struct task_group *tg = css_tg(css);
7140 sched_offline_group(tg);
7143 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7145 struct task_group *tg = css_tg(css);
7148 * Relies on the RCU grace period between css_released() and this.
7150 sched_free_group(tg);
7154 * This is called before wake_up_new_task(), therefore we really only
7155 * have to set its group bits, all the other stuff does not apply.
7157 static void cpu_cgroup_fork(struct task_struct *task)
7162 rq = task_rq_lock(task, &rf);
7164 update_rq_clock(rq);
7165 sched_change_group(task, TASK_SET_GROUP);
7167 task_rq_unlock(rq, task, &rf);
7170 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7172 struct task_struct *task;
7173 struct cgroup_subsys_state *css;
7176 cgroup_taskset_for_each(task, css, tset) {
7177 #ifdef CONFIG_RT_GROUP_SCHED
7178 if (!sched_rt_can_attach(css_tg(css), task))
7182 * Serialize against wake_up_new_task() such that if its
7183 * running, we're sure to observe its full state.
7185 raw_spin_lock_irq(&task->pi_lock);
7187 * Avoid calling sched_move_task() before wake_up_new_task()
7188 * has happened. This would lead to problems with PELT, due to
7189 * move wanting to detach+attach while we're not attached yet.
7191 if (task->state == TASK_NEW)
7193 raw_spin_unlock_irq(&task->pi_lock);
7201 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7203 struct task_struct *task;
7204 struct cgroup_subsys_state *css;
7206 cgroup_taskset_for_each(task, css, tset)
7207 sched_move_task(task);
7210 #ifdef CONFIG_UCLAMP_TASK_GROUP
7211 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7213 struct cgroup_subsys_state *top_css = css;
7214 struct uclamp_se *uc_parent = NULL;
7215 struct uclamp_se *uc_se = NULL;
7216 unsigned int eff[UCLAMP_CNT];
7217 enum uclamp_id clamp_id;
7218 unsigned int clamps;
7220 css_for_each_descendant_pre(css, top_css) {
7221 uc_parent = css_tg(css)->parent
7222 ? css_tg(css)->parent->uclamp : NULL;
7224 for_each_clamp_id(clamp_id) {
7225 /* Assume effective clamps matches requested clamps */
7226 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7227 /* Cap effective clamps with parent's effective clamps */
7229 eff[clamp_id] > uc_parent[clamp_id].value) {
7230 eff[clamp_id] = uc_parent[clamp_id].value;
7233 /* Ensure protection is always capped by limit */
7234 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7236 /* Propagate most restrictive effective clamps */
7238 uc_se = css_tg(css)->uclamp;
7239 for_each_clamp_id(clamp_id) {
7240 if (eff[clamp_id] == uc_se[clamp_id].value)
7242 uc_se[clamp_id].value = eff[clamp_id];
7243 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7244 clamps |= (0x1 << clamp_id);
7247 css = css_rightmost_descendant(css);
7251 /* Immediately update descendants RUNNABLE tasks */
7252 uclamp_update_active_tasks(css, clamps);
7257 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7258 * C expression. Since there is no way to convert a macro argument (N) into a
7259 * character constant, use two levels of macros.
7261 #define _POW10(exp) ((unsigned int)1e##exp)
7262 #define POW10(exp) _POW10(exp)
7264 struct uclamp_request {
7265 #define UCLAMP_PERCENT_SHIFT 2
7266 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7272 static inline struct uclamp_request
7273 capacity_from_percent(char *buf)
7275 struct uclamp_request req = {
7276 .percent = UCLAMP_PERCENT_SCALE,
7277 .util = SCHED_CAPACITY_SCALE,
7282 if (strcmp(buf, "max")) {
7283 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7287 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7292 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7293 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7299 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7300 size_t nbytes, loff_t off,
7301 enum uclamp_id clamp_id)
7303 struct uclamp_request req;
7304 struct task_group *tg;
7306 req = capacity_from_percent(buf);
7310 mutex_lock(&uclamp_mutex);
7313 tg = css_tg(of_css(of));
7314 if (tg->uclamp_req[clamp_id].value != req.util)
7315 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7318 * Because of not recoverable conversion rounding we keep track of the
7319 * exact requested value
7321 tg->uclamp_pct[clamp_id] = req.percent;
7323 /* Update effective clamps to track the most restrictive value */
7324 cpu_util_update_eff(of_css(of));
7327 mutex_unlock(&uclamp_mutex);
7332 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7333 char *buf, size_t nbytes,
7336 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7339 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7340 char *buf, size_t nbytes,
7343 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7346 static inline void cpu_uclamp_print(struct seq_file *sf,
7347 enum uclamp_id clamp_id)
7349 struct task_group *tg;
7355 tg = css_tg(seq_css(sf));
7356 util_clamp = tg->uclamp_req[clamp_id].value;
7359 if (util_clamp == SCHED_CAPACITY_SCALE) {
7360 seq_puts(sf, "max\n");
7364 percent = tg->uclamp_pct[clamp_id];
7365 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7366 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7369 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7371 cpu_uclamp_print(sf, UCLAMP_MIN);
7375 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7377 cpu_uclamp_print(sf, UCLAMP_MAX);
7380 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7382 #ifdef CONFIG_FAIR_GROUP_SCHED
7383 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7384 struct cftype *cftype, u64 shareval)
7386 if (shareval > scale_load_down(ULONG_MAX))
7387 shareval = MAX_SHARES;
7388 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7391 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7394 struct task_group *tg = css_tg(css);
7396 return (u64) scale_load_down(tg->shares);
7399 #ifdef CONFIG_CFS_BANDWIDTH
7400 static DEFINE_MUTEX(cfs_constraints_mutex);
7402 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7403 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7405 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7407 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7409 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7410 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7412 if (tg == &root_task_group)
7416 * Ensure we have at some amount of bandwidth every period. This is
7417 * to prevent reaching a state of large arrears when throttled via
7418 * entity_tick() resulting in prolonged exit starvation.
7420 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7424 * Likewise, bound things on the otherside by preventing insane quota
7425 * periods. This also allows us to normalize in computing quota
7428 if (period > max_cfs_quota_period)
7432 * Prevent race between setting of cfs_rq->runtime_enabled and
7433 * unthrottle_offline_cfs_rqs().
7436 mutex_lock(&cfs_constraints_mutex);
7437 ret = __cfs_schedulable(tg, period, quota);
7441 runtime_enabled = quota != RUNTIME_INF;
7442 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7444 * If we need to toggle cfs_bandwidth_used, off->on must occur
7445 * before making related changes, and on->off must occur afterwards
7447 if (runtime_enabled && !runtime_was_enabled)
7448 cfs_bandwidth_usage_inc();
7449 raw_spin_lock_irq(&cfs_b->lock);
7450 cfs_b->period = ns_to_ktime(period);
7451 cfs_b->quota = quota;
7453 __refill_cfs_bandwidth_runtime(cfs_b);
7455 /* Restart the period timer (if active) to handle new period expiry: */
7456 if (runtime_enabled)
7457 start_cfs_bandwidth(cfs_b);
7459 raw_spin_unlock_irq(&cfs_b->lock);
7461 for_each_online_cpu(i) {
7462 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7463 struct rq *rq = cfs_rq->rq;
7466 rq_lock_irq(rq, &rf);
7467 cfs_rq->runtime_enabled = runtime_enabled;
7468 cfs_rq->runtime_remaining = 0;
7470 if (cfs_rq->throttled)
7471 unthrottle_cfs_rq(cfs_rq);
7472 rq_unlock_irq(rq, &rf);
7474 if (runtime_was_enabled && !runtime_enabled)
7475 cfs_bandwidth_usage_dec();
7477 mutex_unlock(&cfs_constraints_mutex);
7483 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7487 period = ktime_to_ns(tg->cfs_bandwidth.period);
7488 if (cfs_quota_us < 0)
7489 quota = RUNTIME_INF;
7490 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7491 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7495 return tg_set_cfs_bandwidth(tg, period, quota);
7498 static long tg_get_cfs_quota(struct task_group *tg)
7502 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7505 quota_us = tg->cfs_bandwidth.quota;
7506 do_div(quota_us, NSEC_PER_USEC);
7511 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7515 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7518 period = (u64)cfs_period_us * NSEC_PER_USEC;
7519 quota = tg->cfs_bandwidth.quota;
7521 return tg_set_cfs_bandwidth(tg, period, quota);
7524 static long tg_get_cfs_period(struct task_group *tg)
7528 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7529 do_div(cfs_period_us, NSEC_PER_USEC);
7531 return cfs_period_us;
7534 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7537 return tg_get_cfs_quota(css_tg(css));
7540 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7541 struct cftype *cftype, s64 cfs_quota_us)
7543 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7546 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7549 return tg_get_cfs_period(css_tg(css));
7552 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7553 struct cftype *cftype, u64 cfs_period_us)
7555 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7558 struct cfs_schedulable_data {
7559 struct task_group *tg;
7564 * normalize group quota/period to be quota/max_period
7565 * note: units are usecs
7567 static u64 normalize_cfs_quota(struct task_group *tg,
7568 struct cfs_schedulable_data *d)
7576 period = tg_get_cfs_period(tg);
7577 quota = tg_get_cfs_quota(tg);
7580 /* note: these should typically be equivalent */
7581 if (quota == RUNTIME_INF || quota == -1)
7584 return to_ratio(period, quota);
7587 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7589 struct cfs_schedulable_data *d = data;
7590 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7591 s64 quota = 0, parent_quota = -1;
7594 quota = RUNTIME_INF;
7596 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7598 quota = normalize_cfs_quota(tg, d);
7599 parent_quota = parent_b->hierarchical_quota;
7602 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7603 * always take the min. On cgroup1, only inherit when no
7606 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7607 quota = min(quota, parent_quota);
7609 if (quota == RUNTIME_INF)
7610 quota = parent_quota;
7611 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7615 cfs_b->hierarchical_quota = quota;
7620 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7623 struct cfs_schedulable_data data = {
7629 if (quota != RUNTIME_INF) {
7630 do_div(data.period, NSEC_PER_USEC);
7631 do_div(data.quota, NSEC_PER_USEC);
7635 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7641 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7643 struct task_group *tg = css_tg(seq_css(sf));
7644 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7646 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7647 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7648 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7650 if (schedstat_enabled() && tg != &root_task_group) {
7654 for_each_possible_cpu(i)
7655 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7657 seq_printf(sf, "wait_sum %llu\n", ws);
7662 #endif /* CONFIG_CFS_BANDWIDTH */
7663 #endif /* CONFIG_FAIR_GROUP_SCHED */
7665 #ifdef CONFIG_RT_GROUP_SCHED
7666 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7667 struct cftype *cft, s64 val)
7669 return sched_group_set_rt_runtime(css_tg(css), val);
7672 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7675 return sched_group_rt_runtime(css_tg(css));
7678 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7679 struct cftype *cftype, u64 rt_period_us)
7681 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7684 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7687 return sched_group_rt_period(css_tg(css));
7689 #endif /* CONFIG_RT_GROUP_SCHED */
7691 static struct cftype cpu_legacy_files[] = {
7692 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 .read_u64 = cpu_shares_read_u64,
7696 .write_u64 = cpu_shares_write_u64,
7699 #ifdef CONFIG_CFS_BANDWIDTH
7701 .name = "cfs_quota_us",
7702 .read_s64 = cpu_cfs_quota_read_s64,
7703 .write_s64 = cpu_cfs_quota_write_s64,
7706 .name = "cfs_period_us",
7707 .read_u64 = cpu_cfs_period_read_u64,
7708 .write_u64 = cpu_cfs_period_write_u64,
7712 .seq_show = cpu_cfs_stat_show,
7715 #ifdef CONFIG_RT_GROUP_SCHED
7717 .name = "rt_runtime_us",
7718 .read_s64 = cpu_rt_runtime_read,
7719 .write_s64 = cpu_rt_runtime_write,
7722 .name = "rt_period_us",
7723 .read_u64 = cpu_rt_period_read_uint,
7724 .write_u64 = cpu_rt_period_write_uint,
7727 #ifdef CONFIG_UCLAMP_TASK_GROUP
7729 .name = "uclamp.min",
7730 .flags = CFTYPE_NOT_ON_ROOT,
7731 .seq_show = cpu_uclamp_min_show,
7732 .write = cpu_uclamp_min_write,
7735 .name = "uclamp.max",
7736 .flags = CFTYPE_NOT_ON_ROOT,
7737 .seq_show = cpu_uclamp_max_show,
7738 .write = cpu_uclamp_max_write,
7744 static int cpu_extra_stat_show(struct seq_file *sf,
7745 struct cgroup_subsys_state *css)
7747 #ifdef CONFIG_CFS_BANDWIDTH
7749 struct task_group *tg = css_tg(css);
7750 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7753 throttled_usec = cfs_b->throttled_time;
7754 do_div(throttled_usec, NSEC_PER_USEC);
7756 seq_printf(sf, "nr_periods %d\n"
7758 "throttled_usec %llu\n",
7759 cfs_b->nr_periods, cfs_b->nr_throttled,
7766 #ifdef CONFIG_FAIR_GROUP_SCHED
7767 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7770 struct task_group *tg = css_tg(css);
7771 u64 weight = scale_load_down(tg->shares);
7773 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7776 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7777 struct cftype *cft, u64 weight)
7780 * cgroup weight knobs should use the common MIN, DFL and MAX
7781 * values which are 1, 100 and 10000 respectively. While it loses
7782 * a bit of range on both ends, it maps pretty well onto the shares
7783 * value used by scheduler and the round-trip conversions preserve
7784 * the original value over the entire range.
7786 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7789 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7791 return sched_group_set_shares(css_tg(css), scale_load(weight));
7794 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7797 unsigned long weight = scale_load_down(css_tg(css)->shares);
7798 int last_delta = INT_MAX;
7801 /* find the closest nice value to the current weight */
7802 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7803 delta = abs(sched_prio_to_weight[prio] - weight);
7804 if (delta >= last_delta)
7809 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7812 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7813 struct cftype *cft, s64 nice)
7815 unsigned long weight;
7818 if (nice < MIN_NICE || nice > MAX_NICE)
7821 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7822 idx = array_index_nospec(idx, 40);
7823 weight = sched_prio_to_weight[idx];
7825 return sched_group_set_shares(css_tg(css), scale_load(weight));
7829 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7830 long period, long quota)
7833 seq_puts(sf, "max");
7835 seq_printf(sf, "%ld", quota);
7837 seq_printf(sf, " %ld\n", period);
7840 /* caller should put the current value in *@periodp before calling */
7841 static int __maybe_unused cpu_period_quota_parse(char *buf,
7842 u64 *periodp, u64 *quotap)
7844 char tok[21]; /* U64_MAX */
7846 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7849 *periodp *= NSEC_PER_USEC;
7851 if (sscanf(tok, "%llu", quotap))
7852 *quotap *= NSEC_PER_USEC;
7853 else if (!strcmp(tok, "max"))
7854 *quotap = RUNTIME_INF;
7861 #ifdef CONFIG_CFS_BANDWIDTH
7862 static int cpu_max_show(struct seq_file *sf, void *v)
7864 struct task_group *tg = css_tg(seq_css(sf));
7866 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7870 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7871 char *buf, size_t nbytes, loff_t off)
7873 struct task_group *tg = css_tg(of_css(of));
7874 u64 period = tg_get_cfs_period(tg);
7878 ret = cpu_period_quota_parse(buf, &period, "a);
7880 ret = tg_set_cfs_bandwidth(tg, period, quota);
7881 return ret ?: nbytes;
7885 static struct cftype cpu_files[] = {
7886 #ifdef CONFIG_FAIR_GROUP_SCHED
7889 .flags = CFTYPE_NOT_ON_ROOT,
7890 .read_u64 = cpu_weight_read_u64,
7891 .write_u64 = cpu_weight_write_u64,
7894 .name = "weight.nice",
7895 .flags = CFTYPE_NOT_ON_ROOT,
7896 .read_s64 = cpu_weight_nice_read_s64,
7897 .write_s64 = cpu_weight_nice_write_s64,
7900 #ifdef CONFIG_CFS_BANDWIDTH
7903 .flags = CFTYPE_NOT_ON_ROOT,
7904 .seq_show = cpu_max_show,
7905 .write = cpu_max_write,
7908 #ifdef CONFIG_UCLAMP_TASK_GROUP
7910 .name = "uclamp.min",
7911 .flags = CFTYPE_NOT_ON_ROOT,
7912 .seq_show = cpu_uclamp_min_show,
7913 .write = cpu_uclamp_min_write,
7916 .name = "uclamp.max",
7917 .flags = CFTYPE_NOT_ON_ROOT,
7918 .seq_show = cpu_uclamp_max_show,
7919 .write = cpu_uclamp_max_write,
7925 struct cgroup_subsys cpu_cgrp_subsys = {
7926 .css_alloc = cpu_cgroup_css_alloc,
7927 .css_online = cpu_cgroup_css_online,
7928 .css_released = cpu_cgroup_css_released,
7929 .css_free = cpu_cgroup_css_free,
7930 .css_extra_stat_show = cpu_extra_stat_show,
7931 .fork = cpu_cgroup_fork,
7932 .can_attach = cpu_cgroup_can_attach,
7933 .attach = cpu_cgroup_attach,
7934 .legacy_cftypes = cpu_legacy_files,
7935 .dfl_cftypes = cpu_files,
7940 #endif /* CONFIG_CGROUP_SCHED */
7942 void dump_cpu_task(int cpu)
7944 pr_info("Task dump for CPU %d:\n", cpu);
7945 sched_show_task(cpu_curr(cpu));
7949 * Nice levels are multiplicative, with a gentle 10% change for every
7950 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7951 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7952 * that remained on nice 0.
7954 * The "10% effect" is relative and cumulative: from _any_ nice level,
7955 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7956 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7957 * If a task goes up by ~10% and another task goes down by ~10% then
7958 * the relative distance between them is ~25%.)
7960 const int sched_prio_to_weight[40] = {
7961 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7962 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7963 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7964 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7965 /* 0 */ 1024, 820, 655, 526, 423,
7966 /* 5 */ 335, 272, 215, 172, 137,
7967 /* 10 */ 110, 87, 70, 56, 45,
7968 /* 15 */ 36, 29, 23, 18, 15,
7972 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7974 * In cases where the weight does not change often, we can use the
7975 * precalculated inverse to speed up arithmetics by turning divisions
7976 * into multiplications:
7978 const u32 sched_prio_to_wmult[40] = {
7979 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7980 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7981 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7982 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7983 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7984 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7985 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7986 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7989 #undef CREATE_TRACE_POINTS