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();
556 struct sched_domain *sd;
558 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
562 for_each_domain(cpu, sd) {
563 for_each_cpu(i, sched_domain_span(sd)) {
567 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
574 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
575 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
582 * When add_timer_on() enqueues a timer into the timer wheel of an
583 * idle CPU then this timer might expire before the next timer event
584 * which is scheduled to wake up that CPU. In case of a completely
585 * idle system the next event might even be infinite time into the
586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587 * leaves the inner idle loop so the newly added timer is taken into
588 * account when the CPU goes back to idle and evaluates the timer
589 * wheel for the next timer event.
591 static void wake_up_idle_cpu(int cpu)
593 struct rq *rq = cpu_rq(cpu);
595 if (cpu == smp_processor_id())
598 if (set_nr_and_not_polling(rq->idle))
599 smp_send_reschedule(cpu);
601 trace_sched_wake_idle_without_ipi(cpu);
604 static bool wake_up_full_nohz_cpu(int cpu)
607 * We just need the target to call irq_exit() and re-evaluate
608 * the next tick. The nohz full kick at least implies that.
609 * If needed we can still optimize that later with an
612 if (cpu_is_offline(cpu))
613 return true; /* Don't try to wake offline CPUs. */
614 if (tick_nohz_full_cpu(cpu)) {
615 if (cpu != smp_processor_id() ||
616 tick_nohz_tick_stopped())
617 tick_nohz_full_kick_cpu(cpu);
625 * Wake up the specified CPU. If the CPU is going offline, it is the
626 * caller's responsibility to deal with the lost wakeup, for example,
627 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
629 void wake_up_nohz_cpu(int cpu)
631 if (!wake_up_full_nohz_cpu(cpu))
632 wake_up_idle_cpu(cpu);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu = smp_processor_id();
639 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
642 if (idle_cpu(cpu) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(struct rq *rq)
667 /* Deadline tasks, even if single, need the tick */
668 if (rq->dl.dl_nr_running)
672 * If there are more than one RR tasks, we need the tick to effect the
673 * actual RR behaviour.
675 if (rq->rt.rr_nr_running) {
676 if (rq->rt.rr_nr_running == 1)
683 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
684 * forced preemption between FIFO tasks.
686 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
691 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
692 * if there's more than one we need the tick for involuntary
695 if (rq->nr_running > 1)
700 #endif /* CONFIG_NO_HZ_FULL */
701 #endif /* CONFIG_SMP */
703 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
704 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
706 * Iterate task_group tree rooted at *from, calling @down when first entering a
707 * node and @up when leaving it for the final time.
709 * Caller must hold rcu_lock or sufficient equivalent.
711 int walk_tg_tree_from(struct task_group *from,
712 tg_visitor down, tg_visitor up, void *data)
714 struct task_group *parent, *child;
720 ret = (*down)(parent, data);
723 list_for_each_entry_rcu(child, &parent->children, siblings) {
730 ret = (*up)(parent, data);
731 if (ret || parent == from)
735 parent = parent->parent;
742 int tg_nop(struct task_group *tg, void *data)
748 static void set_load_weight(struct task_struct *p, bool update_load)
750 int prio = p->static_prio - MAX_RT_PRIO;
751 struct load_weight *load = &p->se.load;
754 * SCHED_IDLE tasks get minimal weight:
756 if (task_has_idle_policy(p)) {
757 load->weight = scale_load(WEIGHT_IDLEPRIO);
758 load->inv_weight = WMULT_IDLEPRIO;
759 p->se.runnable_weight = load->weight;
764 * SCHED_OTHER tasks have to update their load when changing their
767 if (update_load && p->sched_class == &fair_sched_class) {
768 reweight_task(p, prio);
770 load->weight = scale_load(sched_prio_to_weight[prio]);
771 load->inv_weight = sched_prio_to_wmult[prio];
772 p->se.runnable_weight = load->weight;
776 #ifdef CONFIG_UCLAMP_TASK
778 * Serializes updates of utilization clamp values
780 * The (slow-path) user-space triggers utilization clamp value updates which
781 * can require updates on (fast-path) scheduler's data structures used to
782 * support enqueue/dequeue operations.
783 * While the per-CPU rq lock protects fast-path update operations, user-space
784 * requests are serialized using a mutex to reduce the risk of conflicting
785 * updates or API abuses.
787 static DEFINE_MUTEX(uclamp_mutex);
789 /* Max allowed minimum utilization */
790 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
792 /* Max allowed maximum utilization */
793 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
795 /* All clamps are required to be less or equal than these values */
796 static struct uclamp_se uclamp_default[UCLAMP_CNT];
798 /* Integer rounded range for each bucket */
799 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
801 #define for_each_clamp_id(clamp_id) \
802 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
804 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
806 return clamp_value / UCLAMP_BUCKET_DELTA;
809 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
811 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
814 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
816 if (clamp_id == UCLAMP_MIN)
818 return SCHED_CAPACITY_SCALE;
821 static inline void uclamp_se_set(struct uclamp_se *uc_se,
822 unsigned int value, bool user_defined)
824 uc_se->value = value;
825 uc_se->bucket_id = uclamp_bucket_id(value);
826 uc_se->user_defined = user_defined;
829 static inline unsigned int
830 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
831 unsigned int clamp_value)
834 * Avoid blocked utilization pushing up the frequency when we go
835 * idle (which drops the max-clamp) by retaining the last known
838 if (clamp_id == UCLAMP_MAX) {
839 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
843 return uclamp_none(UCLAMP_MIN);
846 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
847 unsigned int clamp_value)
849 /* Reset max-clamp retention only on idle exit */
850 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
853 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
857 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
858 unsigned int clamp_value)
860 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
861 int bucket_id = UCLAMP_BUCKETS - 1;
864 * Since both min and max clamps are max aggregated, find the
865 * top most bucket with tasks in.
867 for ( ; bucket_id >= 0; bucket_id--) {
868 if (!bucket[bucket_id].tasks)
870 return bucket[bucket_id].value;
873 /* No tasks -- default clamp values */
874 return uclamp_idle_value(rq, clamp_id, clamp_value);
877 static inline struct uclamp_se
878 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
880 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
881 #ifdef CONFIG_UCLAMP_TASK_GROUP
882 struct uclamp_se uc_max;
885 * Tasks in autogroups or root task group will be
886 * restricted by system defaults.
888 if (task_group_is_autogroup(task_group(p)))
890 if (task_group(p) == &root_task_group)
893 uc_max = task_group(p)->uclamp[clamp_id];
894 if (uc_req.value > uc_max.value || !uc_req.user_defined)
902 * The effective clamp bucket index of a task depends on, by increasing
904 * - the task specific clamp value, when explicitly requested from userspace
905 * - the task group effective clamp value, for tasks not either in the root
906 * group or in an autogroup
907 * - the system default clamp value, defined by the sysadmin
909 static inline struct uclamp_se
910 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
912 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
913 struct uclamp_se uc_max = uclamp_default[clamp_id];
915 /* System default restrictions always apply */
916 if (unlikely(uc_req.value > uc_max.value))
922 unsigned int uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
924 struct uclamp_se uc_eff;
926 /* Task currently refcounted: use back-annotated (effective) value */
927 if (p->uclamp[clamp_id].active)
928 return p->uclamp[clamp_id].value;
930 uc_eff = uclamp_eff_get(p, clamp_id);
936 * When a task is enqueued on a rq, the clamp bucket currently defined by the
937 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
938 * updates the rq's clamp value if required.
940 * Tasks can have a task-specific value requested from user-space, track
941 * within each bucket the maximum value for tasks refcounted in it.
942 * This "local max aggregation" allows to track the exact "requested" value
943 * for each bucket when all its RUNNABLE tasks require the same clamp.
945 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
946 enum uclamp_id clamp_id)
948 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
949 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
950 struct uclamp_bucket *bucket;
952 lockdep_assert_held(&rq->lock);
954 /* Update task effective clamp */
955 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
957 bucket = &uc_rq->bucket[uc_se->bucket_id];
959 uc_se->active = true;
961 uclamp_idle_reset(rq, clamp_id, uc_se->value);
964 * Local max aggregation: rq buckets always track the max
965 * "requested" clamp value of its RUNNABLE tasks.
967 if (bucket->tasks == 1 || uc_se->value > bucket->value)
968 bucket->value = uc_se->value;
970 if (uc_se->value > READ_ONCE(uc_rq->value))
971 WRITE_ONCE(uc_rq->value, uc_se->value);
975 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
976 * is released. If this is the last task reference counting the rq's max
977 * active clamp value, then the rq's clamp value is updated.
979 * Both refcounted tasks and rq's cached clamp values are expected to be
980 * always valid. If it's detected they are not, as defensive programming,
981 * enforce the expected state and warn.
983 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
984 enum uclamp_id clamp_id)
986 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
987 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
988 struct uclamp_bucket *bucket;
989 unsigned int bkt_clamp;
990 unsigned int rq_clamp;
992 lockdep_assert_held(&rq->lock);
994 bucket = &uc_rq->bucket[uc_se->bucket_id];
995 SCHED_WARN_ON(!bucket->tasks);
996 if (likely(bucket->tasks))
998 uc_se->active = false;
1001 * Keep "local max aggregation" simple and accept to (possibly)
1002 * overboost some RUNNABLE tasks in the same bucket.
1003 * The rq clamp bucket value is reset to its base value whenever
1004 * there are no more RUNNABLE tasks refcounting it.
1006 if (likely(bucket->tasks))
1009 rq_clamp = READ_ONCE(uc_rq->value);
1011 * Defensive programming: this should never happen. If it happens,
1012 * e.g. due to future modification, warn and fixup the expected value.
1014 SCHED_WARN_ON(bucket->value > rq_clamp);
1015 if (bucket->value >= rq_clamp) {
1016 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1017 WRITE_ONCE(uc_rq->value, bkt_clamp);
1021 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1023 enum uclamp_id clamp_id;
1025 if (unlikely(!p->sched_class->uclamp_enabled))
1028 for_each_clamp_id(clamp_id)
1029 uclamp_rq_inc_id(rq, p, clamp_id);
1031 /* Reset clamp idle holding when there is one RUNNABLE task */
1032 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1033 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1036 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1038 enum uclamp_id clamp_id;
1040 if (unlikely(!p->sched_class->uclamp_enabled))
1043 for_each_clamp_id(clamp_id)
1044 uclamp_rq_dec_id(rq, p, clamp_id);
1048 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1054 * Lock the task and the rq where the task is (or was) queued.
1056 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1057 * price to pay to safely serialize util_{min,max} updates with
1058 * enqueues, dequeues and migration operations.
1059 * This is the same locking schema used by __set_cpus_allowed_ptr().
1061 rq = task_rq_lock(p, &rf);
1064 * Setting the clamp bucket is serialized by task_rq_lock().
1065 * If the task is not yet RUNNABLE and its task_struct is not
1066 * affecting a valid clamp bucket, the next time it's enqueued,
1067 * it will already see the updated clamp bucket value.
1069 if (p->uclamp[clamp_id].active) {
1070 uclamp_rq_dec_id(rq, p, clamp_id);
1071 uclamp_rq_inc_id(rq, p, clamp_id);
1074 task_rq_unlock(rq, p, &rf);
1077 #ifdef CONFIG_UCLAMP_TASK_GROUP
1079 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1080 unsigned int clamps)
1082 enum uclamp_id clamp_id;
1083 struct css_task_iter it;
1084 struct task_struct *p;
1086 css_task_iter_start(css, 0, &it);
1087 while ((p = css_task_iter_next(&it))) {
1088 for_each_clamp_id(clamp_id) {
1089 if ((0x1 << clamp_id) & clamps)
1090 uclamp_update_active(p, clamp_id);
1093 css_task_iter_end(&it);
1096 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1097 static void uclamp_update_root_tg(void)
1099 struct task_group *tg = &root_task_group;
1101 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1102 sysctl_sched_uclamp_util_min, false);
1103 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1104 sysctl_sched_uclamp_util_max, false);
1107 cpu_util_update_eff(&root_task_group.css);
1111 static void uclamp_update_root_tg(void) { }
1114 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1115 void __user *buffer, size_t *lenp,
1118 bool update_root_tg = false;
1119 int old_min, old_max;
1122 mutex_lock(&uclamp_mutex);
1123 old_min = sysctl_sched_uclamp_util_min;
1124 old_max = sysctl_sched_uclamp_util_max;
1126 result = proc_dointvec(table, write, buffer, lenp, ppos);
1132 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1133 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1138 if (old_min != sysctl_sched_uclamp_util_min) {
1139 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1140 sysctl_sched_uclamp_util_min, false);
1141 update_root_tg = true;
1143 if (old_max != sysctl_sched_uclamp_util_max) {
1144 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1145 sysctl_sched_uclamp_util_max, false);
1146 update_root_tg = true;
1150 uclamp_update_root_tg();
1153 * We update all RUNNABLE tasks only when task groups are in use.
1154 * Otherwise, keep it simple and do just a lazy update at each next
1155 * task enqueue time.
1161 sysctl_sched_uclamp_util_min = old_min;
1162 sysctl_sched_uclamp_util_max = old_max;
1164 mutex_unlock(&uclamp_mutex);
1169 static int uclamp_validate(struct task_struct *p,
1170 const struct sched_attr *attr)
1172 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1173 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1175 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1176 lower_bound = attr->sched_util_min;
1177 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1178 upper_bound = attr->sched_util_max;
1180 if (lower_bound > upper_bound)
1182 if (upper_bound > SCHED_CAPACITY_SCALE)
1188 static void __setscheduler_uclamp(struct task_struct *p,
1189 const struct sched_attr *attr)
1191 enum uclamp_id clamp_id;
1194 * On scheduling class change, reset to default clamps for tasks
1195 * without a task-specific value.
1197 for_each_clamp_id(clamp_id) {
1198 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1199 unsigned int clamp_value = uclamp_none(clamp_id);
1201 /* Keep using defined clamps across class changes */
1202 if (uc_se->user_defined)
1205 /* By default, RT tasks always get 100% boost */
1206 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1207 clamp_value = uclamp_none(UCLAMP_MAX);
1209 uclamp_se_set(uc_se, clamp_value, false);
1212 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1215 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1216 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1217 attr->sched_util_min, true);
1220 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1221 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1222 attr->sched_util_max, true);
1226 static void uclamp_fork(struct task_struct *p)
1228 enum uclamp_id clamp_id;
1230 for_each_clamp_id(clamp_id)
1231 p->uclamp[clamp_id].active = false;
1233 if (likely(!p->sched_reset_on_fork))
1236 for_each_clamp_id(clamp_id) {
1237 unsigned int clamp_value = uclamp_none(clamp_id);
1239 /* By default, RT tasks always get 100% boost */
1240 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1241 clamp_value = uclamp_none(UCLAMP_MAX);
1243 uclamp_se_set(&p->uclamp_req[clamp_id], clamp_value, false);
1247 static void __init init_uclamp(void)
1249 struct uclamp_se uc_max = {};
1250 enum uclamp_id clamp_id;
1253 mutex_init(&uclamp_mutex);
1255 for_each_possible_cpu(cpu) {
1256 memset(&cpu_rq(cpu)->uclamp, 0, sizeof(struct uclamp_rq));
1257 cpu_rq(cpu)->uclamp_flags = 0;
1260 for_each_clamp_id(clamp_id) {
1261 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1262 uclamp_none(clamp_id), false);
1265 /* System defaults allow max clamp values for both indexes */
1266 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1267 for_each_clamp_id(clamp_id) {
1268 uclamp_default[clamp_id] = uc_max;
1269 #ifdef CONFIG_UCLAMP_TASK_GROUP
1270 root_task_group.uclamp_req[clamp_id] = uc_max;
1271 root_task_group.uclamp[clamp_id] = uc_max;
1276 #else /* CONFIG_UCLAMP_TASK */
1277 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1278 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1279 static inline int uclamp_validate(struct task_struct *p,
1280 const struct sched_attr *attr)
1284 static void __setscheduler_uclamp(struct task_struct *p,
1285 const struct sched_attr *attr) { }
1286 static inline void uclamp_fork(struct task_struct *p) { }
1287 static inline void init_uclamp(void) { }
1288 #endif /* CONFIG_UCLAMP_TASK */
1290 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1292 if (!(flags & ENQUEUE_NOCLOCK))
1293 update_rq_clock(rq);
1295 if (!(flags & ENQUEUE_RESTORE)) {
1296 sched_info_queued(rq, p);
1297 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1300 uclamp_rq_inc(rq, p);
1301 p->sched_class->enqueue_task(rq, p, flags);
1304 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1306 if (!(flags & DEQUEUE_NOCLOCK))
1307 update_rq_clock(rq);
1309 if (!(flags & DEQUEUE_SAVE)) {
1310 sched_info_dequeued(rq, p);
1311 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1314 uclamp_rq_dec(rq, p);
1315 p->sched_class->dequeue_task(rq, p, flags);
1318 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1320 if (task_contributes_to_load(p))
1321 rq->nr_uninterruptible--;
1323 enqueue_task(rq, p, flags);
1325 p->on_rq = TASK_ON_RQ_QUEUED;
1328 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1330 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1332 if (task_contributes_to_load(p))
1333 rq->nr_uninterruptible++;
1335 dequeue_task(rq, p, flags);
1339 * __normal_prio - return the priority that is based on the static prio
1341 static inline int __normal_prio(struct task_struct *p)
1343 return p->static_prio;
1347 * Calculate the expected normal priority: i.e. priority
1348 * without taking RT-inheritance into account. Might be
1349 * boosted by interactivity modifiers. Changes upon fork,
1350 * setprio syscalls, and whenever the interactivity
1351 * estimator recalculates.
1353 static inline int normal_prio(struct task_struct *p)
1357 if (task_has_dl_policy(p))
1358 prio = MAX_DL_PRIO-1;
1359 else if (task_has_rt_policy(p))
1360 prio = MAX_RT_PRIO-1 - p->rt_priority;
1362 prio = __normal_prio(p);
1367 * Calculate the current priority, i.e. the priority
1368 * taken into account by the scheduler. This value might
1369 * be boosted by RT tasks, or might be boosted by
1370 * interactivity modifiers. Will be RT if the task got
1371 * RT-boosted. If not then it returns p->normal_prio.
1373 static int effective_prio(struct task_struct *p)
1375 p->normal_prio = normal_prio(p);
1377 * If we are RT tasks or we were boosted to RT priority,
1378 * keep the priority unchanged. Otherwise, update priority
1379 * to the normal priority:
1381 if (!rt_prio(p->prio))
1382 return p->normal_prio;
1387 * task_curr - is this task currently executing on a CPU?
1388 * @p: the task in question.
1390 * Return: 1 if the task is currently executing. 0 otherwise.
1392 inline int task_curr(const struct task_struct *p)
1394 return cpu_curr(task_cpu(p)) == p;
1398 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1399 * use the balance_callback list if you want balancing.
1401 * this means any call to check_class_changed() must be followed by a call to
1402 * balance_callback().
1404 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1405 const struct sched_class *prev_class,
1408 if (prev_class != p->sched_class) {
1409 if (prev_class->switched_from)
1410 prev_class->switched_from(rq, p);
1412 p->sched_class->switched_to(rq, p);
1413 } else if (oldprio != p->prio || dl_task(p))
1414 p->sched_class->prio_changed(rq, p, oldprio);
1417 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1419 const struct sched_class *class;
1421 if (p->sched_class == rq->curr->sched_class) {
1422 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1424 for_each_class(class) {
1425 if (class == rq->curr->sched_class)
1427 if (class == p->sched_class) {
1435 * A queue event has occurred, and we're going to schedule. In
1436 * this case, we can save a useless back to back clock update.
1438 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1439 rq_clock_skip_update(rq);
1444 static inline bool is_per_cpu_kthread(struct task_struct *p)
1446 if (!(p->flags & PF_KTHREAD))
1449 if (p->nr_cpus_allowed != 1)
1456 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1457 * __set_cpus_allowed_ptr() and select_fallback_rq().
1459 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1461 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1464 if (is_per_cpu_kthread(p))
1465 return cpu_online(cpu);
1467 return cpu_active(cpu);
1471 * This is how migration works:
1473 * 1) we invoke migration_cpu_stop() on the target CPU using
1475 * 2) stopper starts to run (implicitly forcing the migrated thread
1477 * 3) it checks whether the migrated task is still in the wrong runqueue.
1478 * 4) if it's in the wrong runqueue then the migration thread removes
1479 * it and puts it into the right queue.
1480 * 5) stopper completes and stop_one_cpu() returns and the migration
1485 * move_queued_task - move a queued task to new rq.
1487 * Returns (locked) new rq. Old rq's lock is released.
1489 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1490 struct task_struct *p, int new_cpu)
1492 lockdep_assert_held(&rq->lock);
1494 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1495 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1496 set_task_cpu(p, new_cpu);
1499 rq = cpu_rq(new_cpu);
1502 BUG_ON(task_cpu(p) != new_cpu);
1503 enqueue_task(rq, p, 0);
1504 p->on_rq = TASK_ON_RQ_QUEUED;
1505 check_preempt_curr(rq, p, 0);
1510 struct migration_arg {
1511 struct task_struct *task;
1516 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1517 * this because either it can't run here any more (set_cpus_allowed()
1518 * away from this CPU, or CPU going down), or because we're
1519 * attempting to rebalance this task on exec (sched_exec).
1521 * So we race with normal scheduler movements, but that's OK, as long
1522 * as the task is no longer on this CPU.
1524 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1525 struct task_struct *p, int dest_cpu)
1527 /* Affinity changed (again). */
1528 if (!is_cpu_allowed(p, dest_cpu))
1531 update_rq_clock(rq);
1532 rq = move_queued_task(rq, rf, p, dest_cpu);
1538 * migration_cpu_stop - this will be executed by a highprio stopper thread
1539 * and performs thread migration by bumping thread off CPU then
1540 * 'pushing' onto another runqueue.
1542 static int migration_cpu_stop(void *data)
1544 struct migration_arg *arg = data;
1545 struct task_struct *p = arg->task;
1546 struct rq *rq = this_rq();
1550 * The original target CPU might have gone down and we might
1551 * be on another CPU but it doesn't matter.
1553 local_irq_disable();
1555 * We need to explicitly wake pending tasks before running
1556 * __migrate_task() such that we will not miss enforcing cpus_ptr
1557 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1559 sched_ttwu_pending();
1561 raw_spin_lock(&p->pi_lock);
1564 * If task_rq(p) != rq, it cannot be migrated here, because we're
1565 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1566 * we're holding p->pi_lock.
1568 if (task_rq(p) == rq) {
1569 if (task_on_rq_queued(p))
1570 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1572 p->wake_cpu = arg->dest_cpu;
1575 raw_spin_unlock(&p->pi_lock);
1582 * sched_class::set_cpus_allowed must do the below, but is not required to
1583 * actually call this function.
1585 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1587 cpumask_copy(&p->cpus_mask, new_mask);
1588 p->nr_cpus_allowed = cpumask_weight(new_mask);
1591 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1593 struct rq *rq = task_rq(p);
1594 bool queued, running;
1596 lockdep_assert_held(&p->pi_lock);
1598 queued = task_on_rq_queued(p);
1599 running = task_current(rq, p);
1603 * Because __kthread_bind() calls this on blocked tasks without
1606 lockdep_assert_held(&rq->lock);
1607 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1610 put_prev_task(rq, p);
1612 p->sched_class->set_cpus_allowed(p, new_mask);
1615 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1617 set_next_task(rq, p);
1621 * Change a given task's CPU affinity. Migrate the thread to a
1622 * proper CPU and schedule it away if the CPU it's executing on
1623 * is removed from the allowed bitmask.
1625 * NOTE: the caller must have a valid reference to the task, the
1626 * task must not exit() & deallocate itself prematurely. The
1627 * call is not atomic; no spinlocks may be held.
1629 static int __set_cpus_allowed_ptr(struct task_struct *p,
1630 const struct cpumask *new_mask, bool check)
1632 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1633 unsigned int dest_cpu;
1638 rq = task_rq_lock(p, &rf);
1639 update_rq_clock(rq);
1641 if (p->flags & PF_KTHREAD) {
1643 * Kernel threads are allowed on online && !active CPUs
1645 cpu_valid_mask = cpu_online_mask;
1649 * Must re-check here, to close a race against __kthread_bind(),
1650 * sched_setaffinity() is not guaranteed to observe the flag.
1652 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1657 if (cpumask_equal(p->cpus_ptr, new_mask))
1660 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1661 if (dest_cpu >= nr_cpu_ids) {
1666 do_set_cpus_allowed(p, new_mask);
1668 if (p->flags & PF_KTHREAD) {
1670 * For kernel threads that do indeed end up on online &&
1671 * !active we want to ensure they are strict per-CPU threads.
1673 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1674 !cpumask_intersects(new_mask, cpu_active_mask) &&
1675 p->nr_cpus_allowed != 1);
1678 /* Can the task run on the task's current CPU? If so, we're done */
1679 if (cpumask_test_cpu(task_cpu(p), new_mask))
1682 if (task_running(rq, p) || p->state == TASK_WAKING) {
1683 struct migration_arg arg = { p, dest_cpu };
1684 /* Need help from migration thread: drop lock and wait. */
1685 task_rq_unlock(rq, p, &rf);
1686 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1688 } else if (task_on_rq_queued(p)) {
1690 * OK, since we're going to drop the lock immediately
1691 * afterwards anyway.
1693 rq = move_queued_task(rq, &rf, p, dest_cpu);
1696 task_rq_unlock(rq, p, &rf);
1701 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1703 return __set_cpus_allowed_ptr(p, new_mask, false);
1705 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1707 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1709 #ifdef CONFIG_SCHED_DEBUG
1711 * We should never call set_task_cpu() on a blocked task,
1712 * ttwu() will sort out the placement.
1714 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1718 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1719 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1720 * time relying on p->on_rq.
1722 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1723 p->sched_class == &fair_sched_class &&
1724 (p->on_rq && !task_on_rq_migrating(p)));
1726 #ifdef CONFIG_LOCKDEP
1728 * The caller should hold either p->pi_lock or rq->lock, when changing
1729 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1731 * sched_move_task() holds both and thus holding either pins the cgroup,
1734 * Furthermore, all task_rq users should acquire both locks, see
1737 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1738 lockdep_is_held(&task_rq(p)->lock)));
1741 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1743 WARN_ON_ONCE(!cpu_online(new_cpu));
1746 trace_sched_migrate_task(p, new_cpu);
1748 if (task_cpu(p) != new_cpu) {
1749 if (p->sched_class->migrate_task_rq)
1750 p->sched_class->migrate_task_rq(p, new_cpu);
1751 p->se.nr_migrations++;
1753 perf_event_task_migrate(p);
1756 __set_task_cpu(p, new_cpu);
1759 #ifdef CONFIG_NUMA_BALANCING
1760 static void __migrate_swap_task(struct task_struct *p, int cpu)
1762 if (task_on_rq_queued(p)) {
1763 struct rq *src_rq, *dst_rq;
1764 struct rq_flags srf, drf;
1766 src_rq = task_rq(p);
1767 dst_rq = cpu_rq(cpu);
1769 rq_pin_lock(src_rq, &srf);
1770 rq_pin_lock(dst_rq, &drf);
1772 deactivate_task(src_rq, p, 0);
1773 set_task_cpu(p, cpu);
1774 activate_task(dst_rq, p, 0);
1775 check_preempt_curr(dst_rq, p, 0);
1777 rq_unpin_lock(dst_rq, &drf);
1778 rq_unpin_lock(src_rq, &srf);
1782 * Task isn't running anymore; make it appear like we migrated
1783 * it before it went to sleep. This means on wakeup we make the
1784 * previous CPU our target instead of where it really is.
1790 struct migration_swap_arg {
1791 struct task_struct *src_task, *dst_task;
1792 int src_cpu, dst_cpu;
1795 static int migrate_swap_stop(void *data)
1797 struct migration_swap_arg *arg = data;
1798 struct rq *src_rq, *dst_rq;
1801 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1804 src_rq = cpu_rq(arg->src_cpu);
1805 dst_rq = cpu_rq(arg->dst_cpu);
1807 double_raw_lock(&arg->src_task->pi_lock,
1808 &arg->dst_task->pi_lock);
1809 double_rq_lock(src_rq, dst_rq);
1811 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1814 if (task_cpu(arg->src_task) != arg->src_cpu)
1817 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1820 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1823 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1824 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1829 double_rq_unlock(src_rq, dst_rq);
1830 raw_spin_unlock(&arg->dst_task->pi_lock);
1831 raw_spin_unlock(&arg->src_task->pi_lock);
1837 * Cross migrate two tasks
1839 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1840 int target_cpu, int curr_cpu)
1842 struct migration_swap_arg arg;
1845 arg = (struct migration_swap_arg){
1847 .src_cpu = curr_cpu,
1849 .dst_cpu = target_cpu,
1852 if (arg.src_cpu == arg.dst_cpu)
1856 * These three tests are all lockless; this is OK since all of them
1857 * will be re-checked with proper locks held further down the line.
1859 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1862 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1865 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1868 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1869 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1874 #endif /* CONFIG_NUMA_BALANCING */
1877 * wait_task_inactive - wait for a thread to unschedule.
1879 * If @match_state is nonzero, it's the @p->state value just checked and
1880 * not expected to change. If it changes, i.e. @p might have woken up,
1881 * then return zero. When we succeed in waiting for @p to be off its CPU,
1882 * we return a positive number (its total switch count). If a second call
1883 * a short while later returns the same number, the caller can be sure that
1884 * @p has remained unscheduled the whole time.
1886 * The caller must ensure that the task *will* unschedule sometime soon,
1887 * else this function might spin for a *long* time. This function can't
1888 * be called with interrupts off, or it may introduce deadlock with
1889 * smp_call_function() if an IPI is sent by the same process we are
1890 * waiting to become inactive.
1892 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1894 int running, queued;
1901 * We do the initial early heuristics without holding
1902 * any task-queue locks at all. We'll only try to get
1903 * the runqueue lock when things look like they will
1909 * If the task is actively running on another CPU
1910 * still, just relax and busy-wait without holding
1913 * NOTE! Since we don't hold any locks, it's not
1914 * even sure that "rq" stays as the right runqueue!
1915 * But we don't care, since "task_running()" will
1916 * return false if the runqueue has changed and p
1917 * is actually now running somewhere else!
1919 while (task_running(rq, p)) {
1920 if (match_state && unlikely(p->state != match_state))
1926 * Ok, time to look more closely! We need the rq
1927 * lock now, to be *sure*. If we're wrong, we'll
1928 * just go back and repeat.
1930 rq = task_rq_lock(p, &rf);
1931 trace_sched_wait_task(p);
1932 running = task_running(rq, p);
1933 queued = task_on_rq_queued(p);
1935 if (!match_state || p->state == match_state)
1936 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1937 task_rq_unlock(rq, p, &rf);
1940 * If it changed from the expected state, bail out now.
1942 if (unlikely(!ncsw))
1946 * Was it really running after all now that we
1947 * checked with the proper locks actually held?
1949 * Oops. Go back and try again..
1951 if (unlikely(running)) {
1957 * It's not enough that it's not actively running,
1958 * it must be off the runqueue _entirely_, and not
1961 * So if it was still runnable (but just not actively
1962 * running right now), it's preempted, and we should
1963 * yield - it could be a while.
1965 if (unlikely(queued)) {
1966 ktime_t to = NSEC_PER_SEC / HZ;
1968 set_current_state(TASK_UNINTERRUPTIBLE);
1969 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1974 * Ahh, all good. It wasn't running, and it wasn't
1975 * runnable, which means that it will never become
1976 * running in the future either. We're all done!
1985 * kick_process - kick a running thread to enter/exit the kernel
1986 * @p: the to-be-kicked thread
1988 * Cause a process which is running on another CPU to enter
1989 * kernel-mode, without any delay. (to get signals handled.)
1991 * NOTE: this function doesn't have to take the runqueue lock,
1992 * because all it wants to ensure is that the remote task enters
1993 * the kernel. If the IPI races and the task has been migrated
1994 * to another CPU then no harm is done and the purpose has been
1997 void kick_process(struct task_struct *p)
2003 if ((cpu != smp_processor_id()) && task_curr(p))
2004 smp_send_reschedule(cpu);
2007 EXPORT_SYMBOL_GPL(kick_process);
2010 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2012 * A few notes on cpu_active vs cpu_online:
2014 * - cpu_active must be a subset of cpu_online
2016 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2017 * see __set_cpus_allowed_ptr(). At this point the newly online
2018 * CPU isn't yet part of the sched domains, and balancing will not
2021 * - on CPU-down we clear cpu_active() to mask the sched domains and
2022 * avoid the load balancer to place new tasks on the to be removed
2023 * CPU. Existing tasks will remain running there and will be taken
2026 * This means that fallback selection must not select !active CPUs.
2027 * And can assume that any active CPU must be online. Conversely
2028 * select_task_rq() below may allow selection of !active CPUs in order
2029 * to satisfy the above rules.
2031 static int select_fallback_rq(int cpu, struct task_struct *p)
2033 int nid = cpu_to_node(cpu);
2034 const struct cpumask *nodemask = NULL;
2035 enum { cpuset, possible, fail } state = cpuset;
2039 * If the node that the CPU is on has been offlined, cpu_to_node()
2040 * will return -1. There is no CPU on the node, and we should
2041 * select the CPU on the other node.
2044 nodemask = cpumask_of_node(nid);
2046 /* Look for allowed, online CPU in same node. */
2047 for_each_cpu(dest_cpu, nodemask) {
2048 if (!cpu_active(dest_cpu))
2050 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2056 /* Any allowed, online CPU? */
2057 for_each_cpu(dest_cpu, p->cpus_ptr) {
2058 if (!is_cpu_allowed(p, dest_cpu))
2064 /* No more Mr. Nice Guy. */
2067 if (IS_ENABLED(CONFIG_CPUSETS)) {
2068 cpuset_cpus_allowed_fallback(p);
2074 do_set_cpus_allowed(p, cpu_possible_mask);
2085 if (state != cpuset) {
2087 * Don't tell them about moving exiting tasks or
2088 * kernel threads (both mm NULL), since they never
2091 if (p->mm && printk_ratelimit()) {
2092 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2093 task_pid_nr(p), p->comm, cpu);
2101 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2104 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2106 lockdep_assert_held(&p->pi_lock);
2108 if (p->nr_cpus_allowed > 1)
2109 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2111 cpu = cpumask_any(p->cpus_ptr);
2114 * In order not to call set_task_cpu() on a blocking task we need
2115 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2118 * Since this is common to all placement strategies, this lives here.
2120 * [ this allows ->select_task() to simply return task_cpu(p) and
2121 * not worry about this generic constraint ]
2123 if (unlikely(!is_cpu_allowed(p, cpu)))
2124 cpu = select_fallback_rq(task_cpu(p), p);
2129 static void update_avg(u64 *avg, u64 sample)
2131 s64 diff = sample - *avg;
2135 void sched_set_stop_task(int cpu, struct task_struct *stop)
2137 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2138 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2142 * Make it appear like a SCHED_FIFO task, its something
2143 * userspace knows about and won't get confused about.
2145 * Also, it will make PI more or less work without too
2146 * much confusion -- but then, stop work should not
2147 * rely on PI working anyway.
2149 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2151 stop->sched_class = &stop_sched_class;
2154 cpu_rq(cpu)->stop = stop;
2158 * Reset it back to a normal scheduling class so that
2159 * it can die in pieces.
2161 old_stop->sched_class = &rt_sched_class;
2167 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2168 const struct cpumask *new_mask, bool check)
2170 return set_cpus_allowed_ptr(p, new_mask);
2173 #endif /* CONFIG_SMP */
2176 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2180 if (!schedstat_enabled())
2186 if (cpu == rq->cpu) {
2187 __schedstat_inc(rq->ttwu_local);
2188 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2190 struct sched_domain *sd;
2192 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2194 for_each_domain(rq->cpu, sd) {
2195 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2196 __schedstat_inc(sd->ttwu_wake_remote);
2203 if (wake_flags & WF_MIGRATED)
2204 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2205 #endif /* CONFIG_SMP */
2207 __schedstat_inc(rq->ttwu_count);
2208 __schedstat_inc(p->se.statistics.nr_wakeups);
2210 if (wake_flags & WF_SYNC)
2211 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2215 * Mark the task runnable and perform wakeup-preemption.
2217 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2218 struct rq_flags *rf)
2220 check_preempt_curr(rq, p, wake_flags);
2221 p->state = TASK_RUNNING;
2222 trace_sched_wakeup(p);
2225 if (p->sched_class->task_woken) {
2227 * Our task @p is fully woken up and running; so its safe to
2228 * drop the rq->lock, hereafter rq is only used for statistics.
2230 rq_unpin_lock(rq, rf);
2231 p->sched_class->task_woken(rq, p);
2232 rq_repin_lock(rq, rf);
2235 if (rq->idle_stamp) {
2236 u64 delta = rq_clock(rq) - rq->idle_stamp;
2237 u64 max = 2*rq->max_idle_balance_cost;
2239 update_avg(&rq->avg_idle, delta);
2241 if (rq->avg_idle > max)
2250 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2251 struct rq_flags *rf)
2253 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2255 lockdep_assert_held(&rq->lock);
2258 if (p->sched_contributes_to_load)
2259 rq->nr_uninterruptible--;
2261 if (wake_flags & WF_MIGRATED)
2262 en_flags |= ENQUEUE_MIGRATED;
2265 activate_task(rq, p, en_flags);
2266 ttwu_do_wakeup(rq, p, wake_flags, rf);
2270 * Called in case the task @p isn't fully descheduled from its runqueue,
2271 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2272 * since all we need to do is flip p->state to TASK_RUNNING, since
2273 * the task is still ->on_rq.
2275 static int ttwu_remote(struct task_struct *p, int wake_flags)
2281 rq = __task_rq_lock(p, &rf);
2282 if (task_on_rq_queued(p)) {
2283 /* check_preempt_curr() may use rq clock */
2284 update_rq_clock(rq);
2285 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2288 __task_rq_unlock(rq, &rf);
2294 void sched_ttwu_pending(void)
2296 struct rq *rq = this_rq();
2297 struct llist_node *llist = llist_del_all(&rq->wake_list);
2298 struct task_struct *p, *t;
2304 rq_lock_irqsave(rq, &rf);
2305 update_rq_clock(rq);
2307 llist_for_each_entry_safe(p, t, llist, wake_entry)
2308 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2310 rq_unlock_irqrestore(rq, &rf);
2313 void scheduler_ipi(void)
2316 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2317 * TIF_NEED_RESCHED remotely (for the first time) will also send
2320 preempt_fold_need_resched();
2322 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2326 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2327 * traditionally all their work was done from the interrupt return
2328 * path. Now that we actually do some work, we need to make sure
2331 * Some archs already do call them, luckily irq_enter/exit nest
2334 * Arguably we should visit all archs and update all handlers,
2335 * however a fair share of IPIs are still resched only so this would
2336 * somewhat pessimize the simple resched case.
2339 sched_ttwu_pending();
2342 * Check if someone kicked us for doing the nohz idle load balance.
2344 if (unlikely(got_nohz_idle_kick())) {
2345 this_rq()->idle_balance = 1;
2346 raise_softirq_irqoff(SCHED_SOFTIRQ);
2351 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2353 struct rq *rq = cpu_rq(cpu);
2355 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2357 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2358 if (!set_nr_if_polling(rq->idle))
2359 smp_send_reschedule(cpu);
2361 trace_sched_wake_idle_without_ipi(cpu);
2365 void wake_up_if_idle(int cpu)
2367 struct rq *rq = cpu_rq(cpu);
2372 if (!is_idle_task(rcu_dereference(rq->curr)))
2375 if (set_nr_if_polling(rq->idle)) {
2376 trace_sched_wake_idle_without_ipi(cpu);
2378 rq_lock_irqsave(rq, &rf);
2379 if (is_idle_task(rq->curr))
2380 smp_send_reschedule(cpu);
2381 /* Else CPU is not idle, do nothing here: */
2382 rq_unlock_irqrestore(rq, &rf);
2389 bool cpus_share_cache(int this_cpu, int that_cpu)
2391 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2393 #endif /* CONFIG_SMP */
2395 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2397 struct rq *rq = cpu_rq(cpu);
2400 #if defined(CONFIG_SMP)
2401 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2402 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2403 ttwu_queue_remote(p, cpu, wake_flags);
2409 update_rq_clock(rq);
2410 ttwu_do_activate(rq, p, wake_flags, &rf);
2415 * Notes on Program-Order guarantees on SMP systems.
2419 * The basic program-order guarantee on SMP systems is that when a task [t]
2420 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2421 * execution on its new CPU [c1].
2423 * For migration (of runnable tasks) this is provided by the following means:
2425 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2426 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2427 * rq(c1)->lock (if not at the same time, then in that order).
2428 * C) LOCK of the rq(c1)->lock scheduling in task
2430 * Release/acquire chaining guarantees that B happens after A and C after B.
2431 * Note: the CPU doing B need not be c0 or c1
2440 * UNLOCK rq(0)->lock
2442 * LOCK rq(0)->lock // orders against CPU0
2444 * UNLOCK rq(0)->lock
2448 * UNLOCK rq(1)->lock
2450 * LOCK rq(1)->lock // orders against CPU2
2453 * UNLOCK rq(1)->lock
2456 * BLOCKING -- aka. SLEEP + WAKEUP
2458 * For blocking we (obviously) need to provide the same guarantee as for
2459 * migration. However the means are completely different as there is no lock
2460 * chain to provide order. Instead we do:
2462 * 1) smp_store_release(X->on_cpu, 0)
2463 * 2) smp_cond_load_acquire(!X->on_cpu)
2467 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2469 * LOCK rq(0)->lock LOCK X->pi_lock
2472 * smp_store_release(X->on_cpu, 0);
2474 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2480 * X->state = RUNNING
2481 * UNLOCK rq(2)->lock
2483 * LOCK rq(2)->lock // orders against CPU1
2486 * UNLOCK rq(2)->lock
2489 * UNLOCK rq(0)->lock
2492 * However, for wakeups there is a second guarantee we must provide, namely we
2493 * must ensure that CONDITION=1 done by the caller can not be reordered with
2494 * accesses to the task state; see try_to_wake_up() and set_current_state().
2498 * try_to_wake_up - wake up a thread
2499 * @p: the thread to be awakened
2500 * @state: the mask of task states that can be woken
2501 * @wake_flags: wake modifier flags (WF_*)
2503 * If (@state & @p->state) @p->state = TASK_RUNNING.
2505 * If the task was not queued/runnable, also place it back on a runqueue.
2507 * Atomic against schedule() which would dequeue a task, also see
2508 * set_current_state().
2510 * This function executes a full memory barrier before accessing the task
2511 * state; see set_current_state().
2513 * Return: %true if @p->state changes (an actual wakeup was done),
2517 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2519 unsigned long flags;
2520 int cpu, success = 0;
2525 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2526 * == smp_processor_id()'. Together this means we can special
2527 * case the whole 'p->on_rq && ttwu_remote()' case below
2528 * without taking any locks.
2531 * - we rely on Program-Order guarantees for all the ordering,
2532 * - we're serialized against set_special_state() by virtue of
2533 * it disabling IRQs (this allows not taking ->pi_lock).
2535 if (!(p->state & state))
2540 trace_sched_waking(p);
2541 p->state = TASK_RUNNING;
2542 trace_sched_wakeup(p);
2547 * If we are going to wake up a thread waiting for CONDITION we
2548 * need to ensure that CONDITION=1 done by the caller can not be
2549 * reordered with p->state check below. This pairs with mb() in
2550 * set_current_state() the waiting thread does.
2552 raw_spin_lock_irqsave(&p->pi_lock, flags);
2553 smp_mb__after_spinlock();
2554 if (!(p->state & state))
2557 trace_sched_waking(p);
2559 /* We're going to change ->state: */
2564 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2565 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2566 * in smp_cond_load_acquire() below.
2568 * sched_ttwu_pending() try_to_wake_up()
2569 * STORE p->on_rq = 1 LOAD p->state
2572 * __schedule() (switch to task 'p')
2573 * LOCK rq->lock smp_rmb();
2574 * smp_mb__after_spinlock();
2578 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2580 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2581 * __schedule(). See the comment for smp_mb__after_spinlock().
2584 if (p->on_rq && ttwu_remote(p, wake_flags))
2589 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2590 * possible to, falsely, observe p->on_cpu == 0.
2592 * One must be running (->on_cpu == 1) in order to remove oneself
2593 * from the runqueue.
2595 * __schedule() (switch to task 'p') try_to_wake_up()
2596 * STORE p->on_cpu = 1 LOAD p->on_rq
2599 * __schedule() (put 'p' to sleep)
2600 * LOCK rq->lock smp_rmb();
2601 * smp_mb__after_spinlock();
2602 * STORE p->on_rq = 0 LOAD p->on_cpu
2604 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2605 * __schedule(). See the comment for smp_mb__after_spinlock().
2610 * If the owning (remote) CPU is still in the middle of schedule() with
2611 * this task as prev, wait until its done referencing the task.
2613 * Pairs with the smp_store_release() in finish_task().
2615 * This ensures that tasks getting woken will be fully ordered against
2616 * their previous state and preserve Program Order.
2618 smp_cond_load_acquire(&p->on_cpu, !VAL);
2620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2621 p->state = TASK_WAKING;
2624 delayacct_blkio_end(p);
2625 atomic_dec(&task_rq(p)->nr_iowait);
2628 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2629 if (task_cpu(p) != cpu) {
2630 wake_flags |= WF_MIGRATED;
2631 psi_ttwu_dequeue(p);
2632 set_task_cpu(p, cpu);
2635 #else /* CONFIG_SMP */
2638 delayacct_blkio_end(p);
2639 atomic_dec(&task_rq(p)->nr_iowait);
2642 #endif /* CONFIG_SMP */
2644 ttwu_queue(p, cpu, wake_flags);
2646 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2649 ttwu_stat(p, cpu, wake_flags);
2656 * wake_up_process - Wake up a specific process
2657 * @p: The process to be woken up.
2659 * Attempt to wake up the nominated process and move it to the set of runnable
2662 * Return: 1 if the process was woken up, 0 if it was already running.
2664 * This function executes a full memory barrier before accessing the task state.
2666 int wake_up_process(struct task_struct *p)
2668 return try_to_wake_up(p, TASK_NORMAL, 0);
2670 EXPORT_SYMBOL(wake_up_process);
2672 int wake_up_state(struct task_struct *p, unsigned int state)
2674 return try_to_wake_up(p, state, 0);
2678 * Perform scheduler related setup for a newly forked process p.
2679 * p is forked by current.
2681 * __sched_fork() is basic setup used by init_idle() too:
2683 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2688 p->se.exec_start = 0;
2689 p->se.sum_exec_runtime = 0;
2690 p->se.prev_sum_exec_runtime = 0;
2691 p->se.nr_migrations = 0;
2693 INIT_LIST_HEAD(&p->se.group_node);
2695 #ifdef CONFIG_FAIR_GROUP_SCHED
2696 p->se.cfs_rq = NULL;
2699 #ifdef CONFIG_SCHEDSTATS
2700 /* Even if schedstat is disabled, there should not be garbage */
2701 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2704 RB_CLEAR_NODE(&p->dl.rb_node);
2705 init_dl_task_timer(&p->dl);
2706 init_dl_inactive_task_timer(&p->dl);
2707 __dl_clear_params(p);
2709 INIT_LIST_HEAD(&p->rt.run_list);
2711 p->rt.time_slice = sched_rr_timeslice;
2715 #ifdef CONFIG_PREEMPT_NOTIFIERS
2716 INIT_HLIST_HEAD(&p->preempt_notifiers);
2719 #ifdef CONFIG_COMPACTION
2720 p->capture_control = NULL;
2722 init_numa_balancing(clone_flags, p);
2725 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2727 #ifdef CONFIG_NUMA_BALANCING
2729 void set_numabalancing_state(bool enabled)
2732 static_branch_enable(&sched_numa_balancing);
2734 static_branch_disable(&sched_numa_balancing);
2737 #ifdef CONFIG_PROC_SYSCTL
2738 int sysctl_numa_balancing(struct ctl_table *table, int write,
2739 void __user *buffer, size_t *lenp, loff_t *ppos)
2743 int state = static_branch_likely(&sched_numa_balancing);
2745 if (write && !capable(CAP_SYS_ADMIN))
2750 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2754 set_numabalancing_state(state);
2760 #ifdef CONFIG_SCHEDSTATS
2762 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2763 static bool __initdata __sched_schedstats = false;
2765 static void set_schedstats(bool enabled)
2768 static_branch_enable(&sched_schedstats);
2770 static_branch_disable(&sched_schedstats);
2773 void force_schedstat_enabled(void)
2775 if (!schedstat_enabled()) {
2776 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2777 static_branch_enable(&sched_schedstats);
2781 static int __init setup_schedstats(char *str)
2788 * This code is called before jump labels have been set up, so we can't
2789 * change the static branch directly just yet. Instead set a temporary
2790 * variable so init_schedstats() can do it later.
2792 if (!strcmp(str, "enable")) {
2793 __sched_schedstats = true;
2795 } else if (!strcmp(str, "disable")) {
2796 __sched_schedstats = false;
2801 pr_warn("Unable to parse schedstats=\n");
2805 __setup("schedstats=", setup_schedstats);
2807 static void __init init_schedstats(void)
2809 set_schedstats(__sched_schedstats);
2812 #ifdef CONFIG_PROC_SYSCTL
2813 int sysctl_schedstats(struct ctl_table *table, int write,
2814 void __user *buffer, size_t *lenp, loff_t *ppos)
2818 int state = static_branch_likely(&sched_schedstats);
2820 if (write && !capable(CAP_SYS_ADMIN))
2825 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2829 set_schedstats(state);
2832 #endif /* CONFIG_PROC_SYSCTL */
2833 #else /* !CONFIG_SCHEDSTATS */
2834 static inline void init_schedstats(void) {}
2835 #endif /* CONFIG_SCHEDSTATS */
2838 * fork()/clone()-time setup:
2840 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2842 unsigned long flags;
2844 __sched_fork(clone_flags, p);
2846 * We mark the process as NEW here. This guarantees that
2847 * nobody will actually run it, and a signal or other external
2848 * event cannot wake it up and insert it on the runqueue either.
2850 p->state = TASK_NEW;
2853 * Make sure we do not leak PI boosting priority to the child.
2855 p->prio = current->normal_prio;
2860 * Revert to default priority/policy on fork if requested.
2862 if (unlikely(p->sched_reset_on_fork)) {
2863 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2864 p->policy = SCHED_NORMAL;
2865 p->static_prio = NICE_TO_PRIO(0);
2867 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2868 p->static_prio = NICE_TO_PRIO(0);
2870 p->prio = p->normal_prio = __normal_prio(p);
2871 set_load_weight(p, false);
2874 * We don't need the reset flag anymore after the fork. It has
2875 * fulfilled its duty:
2877 p->sched_reset_on_fork = 0;
2880 if (dl_prio(p->prio))
2882 else if (rt_prio(p->prio))
2883 p->sched_class = &rt_sched_class;
2885 p->sched_class = &fair_sched_class;
2887 init_entity_runnable_average(&p->se);
2890 * The child is not yet in the pid-hash so no cgroup attach races,
2891 * and the cgroup is pinned to this child due to cgroup_fork()
2892 * is ran before sched_fork().
2894 * Silence PROVE_RCU.
2896 raw_spin_lock_irqsave(&p->pi_lock, flags);
2898 * We're setting the CPU for the first time, we don't migrate,
2899 * so use __set_task_cpu().
2901 __set_task_cpu(p, smp_processor_id());
2902 if (p->sched_class->task_fork)
2903 p->sched_class->task_fork(p);
2904 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2906 #ifdef CONFIG_SCHED_INFO
2907 if (likely(sched_info_on()))
2908 memset(&p->sched_info, 0, sizeof(p->sched_info));
2910 #if defined(CONFIG_SMP)
2913 init_task_preempt_count(p);
2915 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2916 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2921 unsigned long to_ratio(u64 period, u64 runtime)
2923 if (runtime == RUNTIME_INF)
2927 * Doing this here saves a lot of checks in all
2928 * the calling paths, and returning zero seems
2929 * safe for them anyway.
2934 return div64_u64(runtime << BW_SHIFT, period);
2938 * wake_up_new_task - wake up a newly created task for the first time.
2940 * This function will do some initial scheduler statistics housekeeping
2941 * that must be done for every newly created context, then puts the task
2942 * on the runqueue and wakes it.
2944 void wake_up_new_task(struct task_struct *p)
2949 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2950 p->state = TASK_RUNNING;
2953 * Fork balancing, do it here and not earlier because:
2954 * - cpus_ptr can change in the fork path
2955 * - any previously selected CPU might disappear through hotplug
2957 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2958 * as we're not fully set-up yet.
2960 p->recent_used_cpu = task_cpu(p);
2961 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2963 rq = __task_rq_lock(p, &rf);
2964 update_rq_clock(rq);
2965 post_init_entity_util_avg(p);
2967 activate_task(rq, p, ENQUEUE_NOCLOCK);
2968 trace_sched_wakeup_new(p);
2969 check_preempt_curr(rq, p, WF_FORK);
2971 if (p->sched_class->task_woken) {
2973 * Nothing relies on rq->lock after this, so its fine to
2976 rq_unpin_lock(rq, &rf);
2977 p->sched_class->task_woken(rq, p);
2978 rq_repin_lock(rq, &rf);
2981 task_rq_unlock(rq, p, &rf);
2984 #ifdef CONFIG_PREEMPT_NOTIFIERS
2986 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2988 void preempt_notifier_inc(void)
2990 static_branch_inc(&preempt_notifier_key);
2992 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2994 void preempt_notifier_dec(void)
2996 static_branch_dec(&preempt_notifier_key);
2998 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3001 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3002 * @notifier: notifier struct to register
3004 void preempt_notifier_register(struct preempt_notifier *notifier)
3006 if (!static_branch_unlikely(&preempt_notifier_key))
3007 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3009 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3011 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3014 * preempt_notifier_unregister - no longer interested in preemption notifications
3015 * @notifier: notifier struct to unregister
3017 * This is *not* safe to call from within a preemption notifier.
3019 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3021 hlist_del(¬ifier->link);
3023 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3025 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3027 struct preempt_notifier *notifier;
3029 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3030 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3033 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3035 if (static_branch_unlikely(&preempt_notifier_key))
3036 __fire_sched_in_preempt_notifiers(curr);
3040 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3041 struct task_struct *next)
3043 struct preempt_notifier *notifier;
3045 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3046 notifier->ops->sched_out(notifier, next);
3049 static __always_inline void
3050 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3051 struct task_struct *next)
3053 if (static_branch_unlikely(&preempt_notifier_key))
3054 __fire_sched_out_preempt_notifiers(curr, next);
3057 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3059 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3064 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3065 struct task_struct *next)
3069 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3071 static inline void prepare_task(struct task_struct *next)
3075 * Claim the task as running, we do this before switching to it
3076 * such that any running task will have this set.
3082 static inline void finish_task(struct task_struct *prev)
3086 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3087 * We must ensure this doesn't happen until the switch is completely
3090 * In particular, the load of prev->state in finish_task_switch() must
3091 * happen before this.
3093 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3095 smp_store_release(&prev->on_cpu, 0);
3100 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3103 * Since the runqueue lock will be released by the next
3104 * task (which is an invalid locking op but in the case
3105 * of the scheduler it's an obvious special-case), so we
3106 * do an early lockdep release here:
3108 rq_unpin_lock(rq, rf);
3109 spin_release(&rq->lock.dep_map, _THIS_IP_);
3110 #ifdef CONFIG_DEBUG_SPINLOCK
3111 /* this is a valid case when another task releases the spinlock */
3112 rq->lock.owner = next;
3116 static inline void finish_lock_switch(struct rq *rq)
3119 * If we are tracking spinlock dependencies then we have to
3120 * fix up the runqueue lock - which gets 'carried over' from
3121 * prev into current:
3123 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3124 raw_spin_unlock_irq(&rq->lock);
3128 * NOP if the arch has not defined these:
3131 #ifndef prepare_arch_switch
3132 # define prepare_arch_switch(next) do { } while (0)
3135 #ifndef finish_arch_post_lock_switch
3136 # define finish_arch_post_lock_switch() do { } while (0)
3140 * prepare_task_switch - prepare to switch tasks
3141 * @rq: the runqueue preparing to switch
3142 * @prev: the current task that is being switched out
3143 * @next: the task we are going to switch to.
3145 * This is called with the rq lock held and interrupts off. It must
3146 * be paired with a subsequent finish_task_switch after the context
3149 * prepare_task_switch sets up locking and calls architecture specific
3153 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3154 struct task_struct *next)
3156 kcov_prepare_switch(prev);
3157 sched_info_switch(rq, prev, next);
3158 perf_event_task_sched_out(prev, next);
3160 fire_sched_out_preempt_notifiers(prev, next);
3162 prepare_arch_switch(next);
3166 * finish_task_switch - clean up after a task-switch
3167 * @prev: the thread we just switched away from.
3169 * finish_task_switch must be called after the context switch, paired
3170 * with a prepare_task_switch call before the context switch.
3171 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3172 * and do any other architecture-specific cleanup actions.
3174 * Note that we may have delayed dropping an mm in context_switch(). If
3175 * so, we finish that here outside of the runqueue lock. (Doing it
3176 * with the lock held can cause deadlocks; see schedule() for
3179 * The context switch have flipped the stack from under us and restored the
3180 * local variables which were saved when this task called schedule() in the
3181 * past. prev == current is still correct but we need to recalculate this_rq
3182 * because prev may have moved to another CPU.
3184 static struct rq *finish_task_switch(struct task_struct *prev)
3185 __releases(rq->lock)
3187 struct rq *rq = this_rq();
3188 struct mm_struct *mm = rq->prev_mm;
3192 * The previous task will have left us with a preempt_count of 2
3193 * because it left us after:
3196 * preempt_disable(); // 1
3198 * raw_spin_lock_irq(&rq->lock) // 2
3200 * Also, see FORK_PREEMPT_COUNT.
3202 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3203 "corrupted preempt_count: %s/%d/0x%x\n",
3204 current->comm, current->pid, preempt_count()))
3205 preempt_count_set(FORK_PREEMPT_COUNT);
3210 * A task struct has one reference for the use as "current".
3211 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3212 * schedule one last time. The schedule call will never return, and
3213 * the scheduled task must drop that reference.
3215 * We must observe prev->state before clearing prev->on_cpu (in
3216 * finish_task), otherwise a concurrent wakeup can get prev
3217 * running on another CPU and we could rave with its RUNNING -> DEAD
3218 * transition, resulting in a double drop.
3220 prev_state = prev->state;
3221 vtime_task_switch(prev);
3222 perf_event_task_sched_in(prev, current);
3224 finish_lock_switch(rq);
3225 finish_arch_post_lock_switch();
3226 kcov_finish_switch(current);
3228 fire_sched_in_preempt_notifiers(current);
3230 * When switching through a kernel thread, the loop in
3231 * membarrier_{private,global}_expedited() may have observed that
3232 * kernel thread and not issued an IPI. It is therefore possible to
3233 * schedule between user->kernel->user threads without passing though
3234 * switch_mm(). Membarrier requires a barrier after storing to
3235 * rq->curr, before returning to userspace, so provide them here:
3237 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3238 * provided by mmdrop(),
3239 * - a sync_core for SYNC_CORE.
3242 membarrier_mm_sync_core_before_usermode(mm);
3245 if (unlikely(prev_state == TASK_DEAD)) {
3246 if (prev->sched_class->task_dead)
3247 prev->sched_class->task_dead(prev);
3250 * Remove function-return probe instances associated with this
3251 * task and put them back on the free list.
3253 kprobe_flush_task(prev);
3255 /* Task is done with its stack. */
3256 put_task_stack(prev);
3258 put_task_struct_rcu_user(prev);
3261 tick_nohz_task_switch();
3267 /* rq->lock is NOT held, but preemption is disabled */
3268 static void __balance_callback(struct rq *rq)
3270 struct callback_head *head, *next;
3271 void (*func)(struct rq *rq);
3272 unsigned long flags;
3274 raw_spin_lock_irqsave(&rq->lock, flags);
3275 head = rq->balance_callback;
3276 rq->balance_callback = NULL;
3278 func = (void (*)(struct rq *))head->func;
3285 raw_spin_unlock_irqrestore(&rq->lock, flags);
3288 static inline void balance_callback(struct rq *rq)
3290 if (unlikely(rq->balance_callback))
3291 __balance_callback(rq);
3296 static inline void balance_callback(struct rq *rq)
3303 * schedule_tail - first thing a freshly forked thread must call.
3304 * @prev: the thread we just switched away from.
3306 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3307 __releases(rq->lock)
3312 * New tasks start with FORK_PREEMPT_COUNT, see there and
3313 * finish_task_switch() for details.
3315 * finish_task_switch() will drop rq->lock() and lower preempt_count
3316 * and the preempt_enable() will end up enabling preemption (on
3317 * PREEMPT_COUNT kernels).
3320 rq = finish_task_switch(prev);
3321 balance_callback(rq);
3324 if (current->set_child_tid)
3325 put_user(task_pid_vnr(current), current->set_child_tid);
3327 calculate_sigpending();
3331 * context_switch - switch to the new MM and the new thread's register state.
3333 static __always_inline struct rq *
3334 context_switch(struct rq *rq, struct task_struct *prev,
3335 struct task_struct *next, struct rq_flags *rf)
3337 prepare_task_switch(rq, prev, next);
3340 * For paravirt, this is coupled with an exit in switch_to to
3341 * combine the page table reload and the switch backend into
3344 arch_start_context_switch(prev);
3347 * kernel -> kernel lazy + transfer active
3348 * user -> kernel lazy + mmgrab() active
3350 * kernel -> user switch + mmdrop() active
3351 * user -> user switch
3353 if (!next->mm) { // to kernel
3354 enter_lazy_tlb(prev->active_mm, next);
3356 next->active_mm = prev->active_mm;
3357 if (prev->mm) // from user
3358 mmgrab(prev->active_mm);
3360 prev->active_mm = NULL;
3362 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3364 * sys_membarrier() requires an smp_mb() between setting
3365 * rq->curr / membarrier_switch_mm() and returning to userspace.
3367 * The below provides this either through switch_mm(), or in
3368 * case 'prev->active_mm == next->mm' through
3369 * finish_task_switch()'s mmdrop().
3371 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3373 if (!prev->mm) { // from kernel
3374 /* will mmdrop() in finish_task_switch(). */
3375 rq->prev_mm = prev->active_mm;
3376 prev->active_mm = NULL;
3380 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3382 prepare_lock_switch(rq, next, rf);
3384 /* Here we just switch the register state and the stack. */
3385 switch_to(prev, next, prev);
3388 return finish_task_switch(prev);
3392 * nr_running and nr_context_switches:
3394 * externally visible scheduler statistics: current number of runnable
3395 * threads, total number of context switches performed since bootup.
3397 unsigned long nr_running(void)
3399 unsigned long i, sum = 0;
3401 for_each_online_cpu(i)
3402 sum += cpu_rq(i)->nr_running;
3408 * Check if only the current task is running on the CPU.
3410 * Caution: this function does not check that the caller has disabled
3411 * preemption, thus the result might have a time-of-check-to-time-of-use
3412 * race. The caller is responsible to use it correctly, for example:
3414 * - from a non-preemptible section (of course)
3416 * - from a thread that is bound to a single CPU
3418 * - in a loop with very short iterations (e.g. a polling loop)
3420 bool single_task_running(void)
3422 return raw_rq()->nr_running == 1;
3424 EXPORT_SYMBOL(single_task_running);
3426 unsigned long long nr_context_switches(void)
3429 unsigned long long sum = 0;
3431 for_each_possible_cpu(i)
3432 sum += cpu_rq(i)->nr_switches;
3438 * Consumers of these two interfaces, like for example the cpuidle menu
3439 * governor, are using nonsensical data. Preferring shallow idle state selection
3440 * for a CPU that has IO-wait which might not even end up running the task when
3441 * it does become runnable.
3444 unsigned long nr_iowait_cpu(int cpu)
3446 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3450 * IO-wait accounting, and how its mostly bollocks (on SMP).
3452 * The idea behind IO-wait account is to account the idle time that we could
3453 * have spend running if it were not for IO. That is, if we were to improve the
3454 * storage performance, we'd have a proportional reduction in IO-wait time.
3456 * This all works nicely on UP, where, when a task blocks on IO, we account
3457 * idle time as IO-wait, because if the storage were faster, it could've been
3458 * running and we'd not be idle.
3460 * This has been extended to SMP, by doing the same for each CPU. This however
3463 * Imagine for instance the case where two tasks block on one CPU, only the one
3464 * CPU will have IO-wait accounted, while the other has regular idle. Even
3465 * though, if the storage were faster, both could've ran at the same time,
3466 * utilising both CPUs.
3468 * This means, that when looking globally, the current IO-wait accounting on
3469 * SMP is a lower bound, by reason of under accounting.
3471 * Worse, since the numbers are provided per CPU, they are sometimes
3472 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3473 * associated with any one particular CPU, it can wake to another CPU than it
3474 * blocked on. This means the per CPU IO-wait number is meaningless.
3476 * Task CPU affinities can make all that even more 'interesting'.
3479 unsigned long nr_iowait(void)
3481 unsigned long i, sum = 0;
3483 for_each_possible_cpu(i)
3484 sum += nr_iowait_cpu(i);
3492 * sched_exec - execve() is a valuable balancing opportunity, because at
3493 * this point the task has the smallest effective memory and cache footprint.
3495 void sched_exec(void)
3497 struct task_struct *p = current;
3498 unsigned long flags;
3501 raw_spin_lock_irqsave(&p->pi_lock, flags);
3502 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3503 if (dest_cpu == smp_processor_id())
3506 if (likely(cpu_active(dest_cpu))) {
3507 struct migration_arg arg = { p, dest_cpu };
3509 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3510 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3514 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3519 DEFINE_PER_CPU(struct kernel_stat, kstat);
3520 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3522 EXPORT_PER_CPU_SYMBOL(kstat);
3523 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3526 * The function fair_sched_class.update_curr accesses the struct curr
3527 * and its field curr->exec_start; when called from task_sched_runtime(),
3528 * we observe a high rate of cache misses in practice.
3529 * Prefetching this data results in improved performance.
3531 static inline void prefetch_curr_exec_start(struct task_struct *p)
3533 #ifdef CONFIG_FAIR_GROUP_SCHED
3534 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3536 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3539 prefetch(&curr->exec_start);
3543 * Return accounted runtime for the task.
3544 * In case the task is currently running, return the runtime plus current's
3545 * pending runtime that have not been accounted yet.
3547 unsigned long long task_sched_runtime(struct task_struct *p)
3553 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3555 * 64-bit doesn't need locks to atomically read a 64-bit value.
3556 * So we have a optimization chance when the task's delta_exec is 0.
3557 * Reading ->on_cpu is racy, but this is ok.
3559 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3560 * If we race with it entering CPU, unaccounted time is 0. This is
3561 * indistinguishable from the read occurring a few cycles earlier.
3562 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3563 * been accounted, so we're correct here as well.
3565 if (!p->on_cpu || !task_on_rq_queued(p))
3566 return p->se.sum_exec_runtime;
3569 rq = task_rq_lock(p, &rf);
3571 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3572 * project cycles that may never be accounted to this
3573 * thread, breaking clock_gettime().
3575 if (task_current(rq, p) && task_on_rq_queued(p)) {
3576 prefetch_curr_exec_start(p);
3577 update_rq_clock(rq);
3578 p->sched_class->update_curr(rq);
3580 ns = p->se.sum_exec_runtime;
3581 task_rq_unlock(rq, p, &rf);
3587 * This function gets called by the timer code, with HZ frequency.
3588 * We call it with interrupts disabled.
3590 void scheduler_tick(void)
3592 int cpu = smp_processor_id();
3593 struct rq *rq = cpu_rq(cpu);
3594 struct task_struct *curr = rq->curr;
3601 update_rq_clock(rq);
3602 curr->sched_class->task_tick(rq, curr, 0);
3603 calc_global_load_tick(rq);
3608 perf_event_task_tick();
3611 rq->idle_balance = idle_cpu(cpu);
3612 trigger_load_balance(rq);
3616 #ifdef CONFIG_NO_HZ_FULL
3621 struct delayed_work work;
3623 /* Values for ->state, see diagram below. */
3624 #define TICK_SCHED_REMOTE_OFFLINE 0
3625 #define TICK_SCHED_REMOTE_OFFLINING 1
3626 #define TICK_SCHED_REMOTE_RUNNING 2
3629 * State diagram for ->state:
3632 * TICK_SCHED_REMOTE_OFFLINE
3635 * | | sched_tick_remote()
3638 * +--TICK_SCHED_REMOTE_OFFLINING
3641 * sched_tick_start() | | sched_tick_stop()
3644 * TICK_SCHED_REMOTE_RUNNING
3647 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3648 * and sched_tick_start() are happy to leave the state in RUNNING.
3651 static struct tick_work __percpu *tick_work_cpu;
3653 static void sched_tick_remote(struct work_struct *work)
3655 struct delayed_work *dwork = to_delayed_work(work);
3656 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3657 int cpu = twork->cpu;
3658 struct rq *rq = cpu_rq(cpu);
3659 struct task_struct *curr;
3665 * Handle the tick only if it appears the remote CPU is running in full
3666 * dynticks mode. The check is racy by nature, but missing a tick or
3667 * having one too much is no big deal because the scheduler tick updates
3668 * statistics and checks timeslices in a time-independent way, regardless
3669 * of when exactly it is running.
3671 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3674 rq_lock_irq(rq, &rf);
3676 if (is_idle_task(curr) || cpu_is_offline(cpu))
3679 update_rq_clock(rq);
3680 delta = rq_clock_task(rq) - curr->se.exec_start;
3683 * Make sure the next tick runs within a reasonable
3686 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3687 curr->sched_class->task_tick(rq, curr, 0);
3690 rq_unlock_irq(rq, &rf);
3694 * Run the remote tick once per second (1Hz). This arbitrary
3695 * frequency is large enough to avoid overload but short enough
3696 * to keep scheduler internal stats reasonably up to date. But
3697 * first update state to reflect hotplug activity if required.
3699 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3700 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3701 if (os == TICK_SCHED_REMOTE_RUNNING)
3702 queue_delayed_work(system_unbound_wq, dwork, HZ);
3705 static void sched_tick_start(int cpu)
3708 struct tick_work *twork;
3710 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3713 WARN_ON_ONCE(!tick_work_cpu);
3715 twork = per_cpu_ptr(tick_work_cpu, cpu);
3716 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3717 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3718 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3720 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3721 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3725 #ifdef CONFIG_HOTPLUG_CPU
3726 static void sched_tick_stop(int cpu)
3728 struct tick_work *twork;
3731 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3734 WARN_ON_ONCE(!tick_work_cpu);
3736 twork = per_cpu_ptr(tick_work_cpu, cpu);
3737 /* There cannot be competing actions, but don't rely on stop-machine. */
3738 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3739 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3740 /* Don't cancel, as this would mess up the state machine. */
3742 #endif /* CONFIG_HOTPLUG_CPU */
3744 int __init sched_tick_offload_init(void)
3746 tick_work_cpu = alloc_percpu(struct tick_work);
3747 BUG_ON(!tick_work_cpu);
3751 #else /* !CONFIG_NO_HZ_FULL */
3752 static inline void sched_tick_start(int cpu) { }
3753 static inline void sched_tick_stop(int cpu) { }
3756 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3757 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3759 * If the value passed in is equal to the current preempt count
3760 * then we just disabled preemption. Start timing the latency.
3762 static inline void preempt_latency_start(int val)
3764 if (preempt_count() == val) {
3765 unsigned long ip = get_lock_parent_ip();
3766 #ifdef CONFIG_DEBUG_PREEMPT
3767 current->preempt_disable_ip = ip;
3769 trace_preempt_off(CALLER_ADDR0, ip);
3773 void preempt_count_add(int val)
3775 #ifdef CONFIG_DEBUG_PREEMPT
3779 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3782 __preempt_count_add(val);
3783 #ifdef CONFIG_DEBUG_PREEMPT
3785 * Spinlock count overflowing soon?
3787 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3790 preempt_latency_start(val);
3792 EXPORT_SYMBOL(preempt_count_add);
3793 NOKPROBE_SYMBOL(preempt_count_add);
3796 * If the value passed in equals to the current preempt count
3797 * then we just enabled preemption. Stop timing the latency.
3799 static inline void preempt_latency_stop(int val)
3801 if (preempt_count() == val)
3802 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3805 void preempt_count_sub(int val)
3807 #ifdef CONFIG_DEBUG_PREEMPT
3811 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3814 * Is the spinlock portion underflowing?
3816 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3817 !(preempt_count() & PREEMPT_MASK)))
3821 preempt_latency_stop(val);
3822 __preempt_count_sub(val);
3824 EXPORT_SYMBOL(preempt_count_sub);
3825 NOKPROBE_SYMBOL(preempt_count_sub);
3828 static inline void preempt_latency_start(int val) { }
3829 static inline void preempt_latency_stop(int val) { }
3832 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3834 #ifdef CONFIG_DEBUG_PREEMPT
3835 return p->preempt_disable_ip;
3842 * Print scheduling while atomic bug:
3844 static noinline void __schedule_bug(struct task_struct *prev)
3846 /* Save this before calling printk(), since that will clobber it */
3847 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3849 if (oops_in_progress)
3852 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3853 prev->comm, prev->pid, preempt_count());
3855 debug_show_held_locks(prev);
3857 if (irqs_disabled())
3858 print_irqtrace_events(prev);
3859 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3860 && in_atomic_preempt_off()) {
3861 pr_err("Preemption disabled at:");
3862 print_ip_sym(preempt_disable_ip);
3866 panic("scheduling while atomic\n");
3869 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3873 * Various schedule()-time debugging checks and statistics:
3875 static inline void schedule_debug(struct task_struct *prev, bool preempt)
3877 #ifdef CONFIG_SCHED_STACK_END_CHECK
3878 if (task_stack_end_corrupted(prev))
3879 panic("corrupted stack end detected inside scheduler\n");
3882 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3883 if (!preempt && prev->state && prev->non_block_count) {
3884 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3885 prev->comm, prev->pid, prev->non_block_count);
3887 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3891 if (unlikely(in_atomic_preempt_off())) {
3892 __schedule_bug(prev);
3893 preempt_count_set(PREEMPT_DISABLED);
3897 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3899 schedstat_inc(this_rq()->sched_count);
3903 * Pick up the highest-prio task:
3905 static inline struct task_struct *
3906 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3908 const struct sched_class *class;
3909 struct task_struct *p;
3912 * Optimization: we know that if all tasks are in the fair class we can
3913 * call that function directly, but only if the @prev task wasn't of a
3914 * higher scheduling class, because otherwise those loose the
3915 * opportunity to pull in more work from other CPUs.
3917 if (likely((prev->sched_class == &idle_sched_class ||
3918 prev->sched_class == &fair_sched_class) &&
3919 rq->nr_running == rq->cfs.h_nr_running)) {
3921 p = pick_next_task_fair(rq, prev, rf);
3922 if (unlikely(p == RETRY_TASK))
3925 /* Assumes fair_sched_class->next == idle_sched_class */
3927 put_prev_task(rq, prev);
3928 p = pick_next_task_idle(rq);
3937 * We must do the balancing pass before put_next_task(), such
3938 * that when we release the rq->lock the task is in the same
3939 * state as before we took rq->lock.
3941 * We can terminate the balance pass as soon as we know there is
3942 * a runnable task of @class priority or higher.
3944 for_class_range(class, prev->sched_class, &idle_sched_class) {
3945 if (class->balance(rq, prev, rf))
3950 put_prev_task(rq, prev);
3952 for_each_class(class) {
3953 p = class->pick_next_task(rq);
3958 /* The idle class should always have a runnable task: */
3963 * __schedule() is the main scheduler function.
3965 * The main means of driving the scheduler and thus entering this function are:
3967 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3969 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3970 * paths. For example, see arch/x86/entry_64.S.
3972 * To drive preemption between tasks, the scheduler sets the flag in timer
3973 * interrupt handler scheduler_tick().
3975 * 3. Wakeups don't really cause entry into schedule(). They add a
3976 * task to the run-queue and that's it.
3978 * Now, if the new task added to the run-queue preempts the current
3979 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3980 * called on the nearest possible occasion:
3982 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
3984 * - in syscall or exception context, at the next outmost
3985 * preempt_enable(). (this might be as soon as the wake_up()'s
3988 * - in IRQ context, return from interrupt-handler to
3989 * preemptible context
3991 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
3994 * - cond_resched() call
3995 * - explicit schedule() call
3996 * - return from syscall or exception to user-space
3997 * - return from interrupt-handler to user-space
3999 * WARNING: must be called with preemption disabled!
4001 static void __sched notrace __schedule(bool preempt)
4003 struct task_struct *prev, *next;
4004 unsigned long *switch_count;
4009 cpu = smp_processor_id();
4013 schedule_debug(prev, preempt);
4015 if (sched_feat(HRTICK))
4018 local_irq_disable();
4019 rcu_note_context_switch(preempt);
4022 * Make sure that signal_pending_state()->signal_pending() below
4023 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4024 * done by the caller to avoid the race with signal_wake_up().
4026 * The membarrier system call requires a full memory barrier
4027 * after coming from user-space, before storing to rq->curr.
4030 smp_mb__after_spinlock();
4032 /* Promote REQ to ACT */
4033 rq->clock_update_flags <<= 1;
4034 update_rq_clock(rq);
4036 switch_count = &prev->nivcsw;
4037 if (!preempt && prev->state) {
4038 if (signal_pending_state(prev->state, prev)) {
4039 prev->state = TASK_RUNNING;
4041 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4043 if (prev->in_iowait) {
4044 atomic_inc(&rq->nr_iowait);
4045 delayacct_blkio_start();
4048 switch_count = &prev->nvcsw;
4051 next = pick_next_task(rq, prev, &rf);
4052 clear_tsk_need_resched(prev);
4053 clear_preempt_need_resched();
4055 if (likely(prev != next)) {
4058 * RCU users of rcu_dereference(rq->curr) may not see
4059 * changes to task_struct made by pick_next_task().
4061 RCU_INIT_POINTER(rq->curr, next);
4063 * The membarrier system call requires each architecture
4064 * to have a full memory barrier after updating
4065 * rq->curr, before returning to user-space.
4067 * Here are the schemes providing that barrier on the
4068 * various architectures:
4069 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4070 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4071 * - finish_lock_switch() for weakly-ordered
4072 * architectures where spin_unlock is a full barrier,
4073 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4074 * is a RELEASE barrier),
4078 trace_sched_switch(preempt, prev, next);
4080 /* Also unlocks the rq: */
4081 rq = context_switch(rq, prev, next, &rf);
4083 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4084 rq_unlock_irq(rq, &rf);
4087 balance_callback(rq);
4090 void __noreturn do_task_dead(void)
4092 /* Causes final put_task_struct in finish_task_switch(): */
4093 set_special_state(TASK_DEAD);
4095 /* Tell freezer to ignore us: */
4096 current->flags |= PF_NOFREEZE;
4101 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4106 static inline void sched_submit_work(struct task_struct *tsk)
4112 * If a worker went to sleep, notify and ask workqueue whether
4113 * it wants to wake up a task to maintain concurrency.
4114 * As this function is called inside the schedule() context,
4115 * we disable preemption to avoid it calling schedule() again
4116 * in the possible wakeup of a kworker.
4118 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4120 if (tsk->flags & PF_WQ_WORKER)
4121 wq_worker_sleeping(tsk);
4123 io_wq_worker_sleeping(tsk);
4124 preempt_enable_no_resched();
4127 if (tsk_is_pi_blocked(tsk))
4131 * If we are going to sleep and we have plugged IO queued,
4132 * make sure to submit it to avoid deadlocks.
4134 if (blk_needs_flush_plug(tsk))
4135 blk_schedule_flush_plug(tsk);
4138 static void sched_update_worker(struct task_struct *tsk)
4140 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4141 if (tsk->flags & PF_WQ_WORKER)
4142 wq_worker_running(tsk);
4144 io_wq_worker_running(tsk);
4148 asmlinkage __visible void __sched schedule(void)
4150 struct task_struct *tsk = current;
4152 sched_submit_work(tsk);
4156 sched_preempt_enable_no_resched();
4157 } while (need_resched());
4158 sched_update_worker(tsk);
4160 EXPORT_SYMBOL(schedule);
4163 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4164 * state (have scheduled out non-voluntarily) by making sure that all
4165 * tasks have either left the run queue or have gone into user space.
4166 * As idle tasks do not do either, they must not ever be preempted
4167 * (schedule out non-voluntarily).
4169 * schedule_idle() is similar to schedule_preempt_disable() except that it
4170 * never enables preemption because it does not call sched_submit_work().
4172 void __sched schedule_idle(void)
4175 * As this skips calling sched_submit_work(), which the idle task does
4176 * regardless because that function is a nop when the task is in a
4177 * TASK_RUNNING state, make sure this isn't used someplace that the
4178 * current task can be in any other state. Note, idle is always in the
4179 * TASK_RUNNING state.
4181 WARN_ON_ONCE(current->state);
4184 } while (need_resched());
4187 #ifdef CONFIG_CONTEXT_TRACKING
4188 asmlinkage __visible void __sched schedule_user(void)
4191 * If we come here after a random call to set_need_resched(),
4192 * or we have been woken up remotely but the IPI has not yet arrived,
4193 * we haven't yet exited the RCU idle mode. Do it here manually until
4194 * we find a better solution.
4196 * NB: There are buggy callers of this function. Ideally we
4197 * should warn if prev_state != CONTEXT_USER, but that will trigger
4198 * too frequently to make sense yet.
4200 enum ctx_state prev_state = exception_enter();
4202 exception_exit(prev_state);
4207 * schedule_preempt_disabled - called with preemption disabled
4209 * Returns with preemption disabled. Note: preempt_count must be 1
4211 void __sched schedule_preempt_disabled(void)
4213 sched_preempt_enable_no_resched();
4218 static void __sched notrace preempt_schedule_common(void)
4222 * Because the function tracer can trace preempt_count_sub()
4223 * and it also uses preempt_enable/disable_notrace(), if
4224 * NEED_RESCHED is set, the preempt_enable_notrace() called
4225 * by the function tracer will call this function again and
4226 * cause infinite recursion.
4228 * Preemption must be disabled here before the function
4229 * tracer can trace. Break up preempt_disable() into two
4230 * calls. One to disable preemption without fear of being
4231 * traced. The other to still record the preemption latency,
4232 * which can also be traced by the function tracer.
4234 preempt_disable_notrace();
4235 preempt_latency_start(1);
4237 preempt_latency_stop(1);
4238 preempt_enable_no_resched_notrace();
4241 * Check again in case we missed a preemption opportunity
4242 * between schedule and now.
4244 } while (need_resched());
4247 #ifdef CONFIG_PREEMPTION
4249 * This is the entry point to schedule() from in-kernel preemption
4250 * off of preempt_enable.
4252 asmlinkage __visible void __sched notrace preempt_schedule(void)
4255 * If there is a non-zero preempt_count or interrupts are disabled,
4256 * we do not want to preempt the current task. Just return..
4258 if (likely(!preemptible()))
4261 preempt_schedule_common();
4263 NOKPROBE_SYMBOL(preempt_schedule);
4264 EXPORT_SYMBOL(preempt_schedule);
4267 * preempt_schedule_notrace - preempt_schedule called by tracing
4269 * The tracing infrastructure uses preempt_enable_notrace to prevent
4270 * recursion and tracing preempt enabling caused by the tracing
4271 * infrastructure itself. But as tracing can happen in areas coming
4272 * from userspace or just about to enter userspace, a preempt enable
4273 * can occur before user_exit() is called. This will cause the scheduler
4274 * to be called when the system is still in usermode.
4276 * To prevent this, the preempt_enable_notrace will use this function
4277 * instead of preempt_schedule() to exit user context if needed before
4278 * calling the scheduler.
4280 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4282 enum ctx_state prev_ctx;
4284 if (likely(!preemptible()))
4289 * Because the function tracer can trace preempt_count_sub()
4290 * and it also uses preempt_enable/disable_notrace(), if
4291 * NEED_RESCHED is set, the preempt_enable_notrace() called
4292 * by the function tracer will call this function again and
4293 * cause infinite recursion.
4295 * Preemption must be disabled here before the function
4296 * tracer can trace. Break up preempt_disable() into two
4297 * calls. One to disable preemption without fear of being
4298 * traced. The other to still record the preemption latency,
4299 * which can also be traced by the function tracer.
4301 preempt_disable_notrace();
4302 preempt_latency_start(1);
4304 * Needs preempt disabled in case user_exit() is traced
4305 * and the tracer calls preempt_enable_notrace() causing
4306 * an infinite recursion.
4308 prev_ctx = exception_enter();
4310 exception_exit(prev_ctx);
4312 preempt_latency_stop(1);
4313 preempt_enable_no_resched_notrace();
4314 } while (need_resched());
4316 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4318 #endif /* CONFIG_PREEMPTION */
4321 * This is the entry point to schedule() from kernel preemption
4322 * off of irq context.
4323 * Note, that this is called and return with irqs disabled. This will
4324 * protect us against recursive calling from irq.
4326 asmlinkage __visible void __sched preempt_schedule_irq(void)
4328 enum ctx_state prev_state;
4330 /* Catch callers which need to be fixed */
4331 BUG_ON(preempt_count() || !irqs_disabled());
4333 prev_state = exception_enter();
4339 local_irq_disable();
4340 sched_preempt_enable_no_resched();
4341 } while (need_resched());
4343 exception_exit(prev_state);
4346 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4349 return try_to_wake_up(curr->private, mode, wake_flags);
4351 EXPORT_SYMBOL(default_wake_function);
4353 #ifdef CONFIG_RT_MUTEXES
4355 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4358 prio = min(prio, pi_task->prio);
4363 static inline int rt_effective_prio(struct task_struct *p, int prio)
4365 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4367 return __rt_effective_prio(pi_task, prio);
4371 * rt_mutex_setprio - set the current priority of a task
4373 * @pi_task: donor task
4375 * This function changes the 'effective' priority of a task. It does
4376 * not touch ->normal_prio like __setscheduler().
4378 * Used by the rt_mutex code to implement priority inheritance
4379 * logic. Call site only calls if the priority of the task changed.
4381 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4383 int prio, oldprio, queued, running, queue_flag =
4384 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4385 const struct sched_class *prev_class;
4389 /* XXX used to be waiter->prio, not waiter->task->prio */
4390 prio = __rt_effective_prio(pi_task, p->normal_prio);
4393 * If nothing changed; bail early.
4395 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4398 rq = __task_rq_lock(p, &rf);
4399 update_rq_clock(rq);
4401 * Set under pi_lock && rq->lock, such that the value can be used under
4404 * Note that there is loads of tricky to make this pointer cache work
4405 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4406 * ensure a task is de-boosted (pi_task is set to NULL) before the
4407 * task is allowed to run again (and can exit). This ensures the pointer
4408 * points to a blocked task -- which guaratees the task is present.
4410 p->pi_top_task = pi_task;
4413 * For FIFO/RR we only need to set prio, if that matches we're done.
4415 if (prio == p->prio && !dl_prio(prio))
4419 * Idle task boosting is a nono in general. There is one
4420 * exception, when PREEMPT_RT and NOHZ is active:
4422 * The idle task calls get_next_timer_interrupt() and holds
4423 * the timer wheel base->lock on the CPU and another CPU wants
4424 * to access the timer (probably to cancel it). We can safely
4425 * ignore the boosting request, as the idle CPU runs this code
4426 * with interrupts disabled and will complete the lock
4427 * protected section without being interrupted. So there is no
4428 * real need to boost.
4430 if (unlikely(p == rq->idle)) {
4431 WARN_ON(p != rq->curr);
4432 WARN_ON(p->pi_blocked_on);
4436 trace_sched_pi_setprio(p, pi_task);
4439 if (oldprio == prio)
4440 queue_flag &= ~DEQUEUE_MOVE;
4442 prev_class = p->sched_class;
4443 queued = task_on_rq_queued(p);
4444 running = task_current(rq, p);
4446 dequeue_task(rq, p, queue_flag);
4448 put_prev_task(rq, p);
4451 * Boosting condition are:
4452 * 1. -rt task is running and holds mutex A
4453 * --> -dl task blocks on mutex A
4455 * 2. -dl task is running and holds mutex A
4456 * --> -dl task blocks on mutex A and could preempt the
4459 if (dl_prio(prio)) {
4460 if (!dl_prio(p->normal_prio) ||
4461 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
4462 p->dl.dl_boosted = 1;
4463 queue_flag |= ENQUEUE_REPLENISH;
4465 p->dl.dl_boosted = 0;
4466 p->sched_class = &dl_sched_class;
4467 } else if (rt_prio(prio)) {
4468 if (dl_prio(oldprio))
4469 p->dl.dl_boosted = 0;
4471 queue_flag |= ENQUEUE_HEAD;
4472 p->sched_class = &rt_sched_class;
4474 if (dl_prio(oldprio))
4475 p->dl.dl_boosted = 0;
4476 if (rt_prio(oldprio))
4478 p->sched_class = &fair_sched_class;
4484 enqueue_task(rq, p, queue_flag);
4486 set_next_task(rq, p);
4488 check_class_changed(rq, p, prev_class, oldprio);
4490 /* Avoid rq from going away on us: */
4492 __task_rq_unlock(rq, &rf);
4494 balance_callback(rq);
4498 static inline int rt_effective_prio(struct task_struct *p, int prio)
4504 void set_user_nice(struct task_struct *p, long nice)
4506 bool queued, running;
4507 int old_prio, delta;
4511 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4514 * We have to be careful, if called from sys_setpriority(),
4515 * the task might be in the middle of scheduling on another CPU.
4517 rq = task_rq_lock(p, &rf);
4518 update_rq_clock(rq);
4521 * The RT priorities are set via sched_setscheduler(), but we still
4522 * allow the 'normal' nice value to be set - but as expected
4523 * it wont have any effect on scheduling until the task is
4524 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4526 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4527 p->static_prio = NICE_TO_PRIO(nice);
4530 queued = task_on_rq_queued(p);
4531 running = task_current(rq, p);
4533 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4535 put_prev_task(rq, p);
4537 p->static_prio = NICE_TO_PRIO(nice);
4538 set_load_weight(p, true);
4540 p->prio = effective_prio(p);
4541 delta = p->prio - old_prio;
4544 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4546 * If the task increased its priority or is running and
4547 * lowered its priority, then reschedule its CPU:
4549 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4553 set_next_task(rq, p);
4555 task_rq_unlock(rq, p, &rf);
4557 EXPORT_SYMBOL(set_user_nice);
4560 * can_nice - check if a task can reduce its nice value
4564 int can_nice(const struct task_struct *p, const int nice)
4566 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4567 int nice_rlim = nice_to_rlimit(nice);
4569 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4570 capable(CAP_SYS_NICE));
4573 #ifdef __ARCH_WANT_SYS_NICE
4576 * sys_nice - change the priority of the current process.
4577 * @increment: priority increment
4579 * sys_setpriority is a more generic, but much slower function that
4580 * does similar things.
4582 SYSCALL_DEFINE1(nice, int, increment)
4587 * Setpriority might change our priority at the same moment.
4588 * We don't have to worry. Conceptually one call occurs first
4589 * and we have a single winner.
4591 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4592 nice = task_nice(current) + increment;
4594 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4595 if (increment < 0 && !can_nice(current, nice))
4598 retval = security_task_setnice(current, nice);
4602 set_user_nice(current, nice);
4609 * task_prio - return the priority value of a given task.
4610 * @p: the task in question.
4612 * Return: The priority value as seen by users in /proc.
4613 * RT tasks are offset by -200. Normal tasks are centered
4614 * around 0, value goes from -16 to +15.
4616 int task_prio(const struct task_struct *p)
4618 return p->prio - MAX_RT_PRIO;
4622 * idle_cpu - is a given CPU idle currently?
4623 * @cpu: the processor in question.
4625 * Return: 1 if the CPU is currently idle. 0 otherwise.
4627 int idle_cpu(int cpu)
4629 struct rq *rq = cpu_rq(cpu);
4631 if (rq->curr != rq->idle)
4638 if (!llist_empty(&rq->wake_list))
4646 * available_idle_cpu - is a given CPU idle for enqueuing work.
4647 * @cpu: the CPU in question.
4649 * Return: 1 if the CPU is currently idle. 0 otherwise.
4651 int available_idle_cpu(int cpu)
4656 if (vcpu_is_preempted(cpu))
4663 * idle_task - return the idle task for a given CPU.
4664 * @cpu: the processor in question.
4666 * Return: The idle task for the CPU @cpu.
4668 struct task_struct *idle_task(int cpu)
4670 return cpu_rq(cpu)->idle;
4674 * find_process_by_pid - find a process with a matching PID value.
4675 * @pid: the pid in question.
4677 * The task of @pid, if found. %NULL otherwise.
4679 static struct task_struct *find_process_by_pid(pid_t pid)
4681 return pid ? find_task_by_vpid(pid) : current;
4685 * sched_setparam() passes in -1 for its policy, to let the functions
4686 * it calls know not to change it.
4688 #define SETPARAM_POLICY -1
4690 static void __setscheduler_params(struct task_struct *p,
4691 const struct sched_attr *attr)
4693 int policy = attr->sched_policy;
4695 if (policy == SETPARAM_POLICY)
4700 if (dl_policy(policy))
4701 __setparam_dl(p, attr);
4702 else if (fair_policy(policy))
4703 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4706 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4707 * !rt_policy. Always setting this ensures that things like
4708 * getparam()/getattr() don't report silly values for !rt tasks.
4710 p->rt_priority = attr->sched_priority;
4711 p->normal_prio = normal_prio(p);
4712 set_load_weight(p, true);
4715 /* Actually do priority change: must hold pi & rq lock. */
4716 static void __setscheduler(struct rq *rq, struct task_struct *p,
4717 const struct sched_attr *attr, bool keep_boost)
4720 * If params can't change scheduling class changes aren't allowed
4723 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4726 __setscheduler_params(p, attr);
4729 * Keep a potential priority boosting if called from
4730 * sched_setscheduler().
4732 p->prio = normal_prio(p);
4734 p->prio = rt_effective_prio(p, p->prio);
4736 if (dl_prio(p->prio))
4737 p->sched_class = &dl_sched_class;
4738 else if (rt_prio(p->prio))
4739 p->sched_class = &rt_sched_class;
4741 p->sched_class = &fair_sched_class;
4745 * Check the target process has a UID that matches the current process's:
4747 static bool check_same_owner(struct task_struct *p)
4749 const struct cred *cred = current_cred(), *pcred;
4753 pcred = __task_cred(p);
4754 match = (uid_eq(cred->euid, pcred->euid) ||
4755 uid_eq(cred->euid, pcred->uid));
4760 static int __sched_setscheduler(struct task_struct *p,
4761 const struct sched_attr *attr,
4764 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4765 MAX_RT_PRIO - 1 - attr->sched_priority;
4766 int retval, oldprio, oldpolicy = -1, queued, running;
4767 int new_effective_prio, policy = attr->sched_policy;
4768 const struct sched_class *prev_class;
4771 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4774 /* The pi code expects interrupts enabled */
4775 BUG_ON(pi && in_interrupt());
4777 /* Double check policy once rq lock held: */
4779 reset_on_fork = p->sched_reset_on_fork;
4780 policy = oldpolicy = p->policy;
4782 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4784 if (!valid_policy(policy))
4788 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4792 * Valid priorities for SCHED_FIFO and SCHED_RR are
4793 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4794 * SCHED_BATCH and SCHED_IDLE is 0.
4796 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4797 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4799 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4800 (rt_policy(policy) != (attr->sched_priority != 0)))
4804 * Allow unprivileged RT tasks to decrease priority:
4806 if (user && !capable(CAP_SYS_NICE)) {
4807 if (fair_policy(policy)) {
4808 if (attr->sched_nice < task_nice(p) &&
4809 !can_nice(p, attr->sched_nice))
4813 if (rt_policy(policy)) {
4814 unsigned long rlim_rtprio =
4815 task_rlimit(p, RLIMIT_RTPRIO);
4817 /* Can't set/change the rt policy: */
4818 if (policy != p->policy && !rlim_rtprio)
4821 /* Can't increase priority: */
4822 if (attr->sched_priority > p->rt_priority &&
4823 attr->sched_priority > rlim_rtprio)
4828 * Can't set/change SCHED_DEADLINE policy at all for now
4829 * (safest behavior); in the future we would like to allow
4830 * unprivileged DL tasks to increase their relative deadline
4831 * or reduce their runtime (both ways reducing utilization)
4833 if (dl_policy(policy))
4837 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4838 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4840 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4841 if (!can_nice(p, task_nice(p)))
4845 /* Can't change other user's priorities: */
4846 if (!check_same_owner(p))
4849 /* Normal users shall not reset the sched_reset_on_fork flag: */
4850 if (p->sched_reset_on_fork && !reset_on_fork)
4855 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4858 retval = security_task_setscheduler(p);
4863 /* Update task specific "requested" clamps */
4864 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4865 retval = uclamp_validate(p, attr);
4874 * Make sure no PI-waiters arrive (or leave) while we are
4875 * changing the priority of the task:
4877 * To be able to change p->policy safely, the appropriate
4878 * runqueue lock must be held.
4880 rq = task_rq_lock(p, &rf);
4881 update_rq_clock(rq);
4884 * Changing the policy of the stop threads its a very bad idea:
4886 if (p == rq->stop) {
4892 * If not changing anything there's no need to proceed further,
4893 * but store a possible modification of reset_on_fork.
4895 if (unlikely(policy == p->policy)) {
4896 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4898 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4900 if (dl_policy(policy) && dl_param_changed(p, attr))
4902 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
4905 p->sched_reset_on_fork = reset_on_fork;
4912 #ifdef CONFIG_RT_GROUP_SCHED
4914 * Do not allow realtime tasks into groups that have no runtime
4917 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4918 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4919 !task_group_is_autogroup(task_group(p))) {
4925 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4926 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4927 cpumask_t *span = rq->rd->span;
4930 * Don't allow tasks with an affinity mask smaller than
4931 * the entire root_domain to become SCHED_DEADLINE. We
4932 * will also fail if there's no bandwidth available.
4934 if (!cpumask_subset(span, p->cpus_ptr) ||
4935 rq->rd->dl_bw.bw == 0) {
4943 /* Re-check policy now with rq lock held: */
4944 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4945 policy = oldpolicy = -1;
4946 task_rq_unlock(rq, p, &rf);
4948 cpuset_read_unlock();
4953 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4954 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4957 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4962 p->sched_reset_on_fork = reset_on_fork;
4967 * Take priority boosted tasks into account. If the new
4968 * effective priority is unchanged, we just store the new
4969 * normal parameters and do not touch the scheduler class and
4970 * the runqueue. This will be done when the task deboost
4973 new_effective_prio = rt_effective_prio(p, newprio);
4974 if (new_effective_prio == oldprio)
4975 queue_flags &= ~DEQUEUE_MOVE;
4978 queued = task_on_rq_queued(p);
4979 running = task_current(rq, p);
4981 dequeue_task(rq, p, queue_flags);
4983 put_prev_task(rq, p);
4985 prev_class = p->sched_class;
4987 __setscheduler(rq, p, attr, pi);
4988 __setscheduler_uclamp(p, attr);
4992 * We enqueue to tail when the priority of a task is
4993 * increased (user space view).
4995 if (oldprio < p->prio)
4996 queue_flags |= ENQUEUE_HEAD;
4998 enqueue_task(rq, p, queue_flags);
5001 set_next_task(rq, p);
5003 check_class_changed(rq, p, prev_class, oldprio);
5005 /* Avoid rq from going away on us: */
5007 task_rq_unlock(rq, p, &rf);
5010 cpuset_read_unlock();
5011 rt_mutex_adjust_pi(p);
5014 /* Run balance callbacks after we've adjusted the PI chain: */
5015 balance_callback(rq);
5021 task_rq_unlock(rq, p, &rf);
5023 cpuset_read_unlock();
5027 static int _sched_setscheduler(struct task_struct *p, int policy,
5028 const struct sched_param *param, bool check)
5030 struct sched_attr attr = {
5031 .sched_policy = policy,
5032 .sched_priority = param->sched_priority,
5033 .sched_nice = PRIO_TO_NICE(p->static_prio),
5036 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5037 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5038 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5039 policy &= ~SCHED_RESET_ON_FORK;
5040 attr.sched_policy = policy;
5043 return __sched_setscheduler(p, &attr, check, true);
5046 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5047 * @p: the task in question.
5048 * @policy: new policy.
5049 * @param: structure containing the new RT priority.
5051 * Return: 0 on success. An error code otherwise.
5053 * NOTE that the task may be already dead.
5055 int sched_setscheduler(struct task_struct *p, int policy,
5056 const struct sched_param *param)
5058 return _sched_setscheduler(p, policy, param, true);
5060 EXPORT_SYMBOL_GPL(sched_setscheduler);
5062 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5064 return __sched_setscheduler(p, attr, true, true);
5066 EXPORT_SYMBOL_GPL(sched_setattr);
5068 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5070 return __sched_setscheduler(p, attr, false, true);
5074 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5075 * @p: the task in question.
5076 * @policy: new policy.
5077 * @param: structure containing the new RT priority.
5079 * Just like sched_setscheduler, only don't bother checking if the
5080 * current context has permission. For example, this is needed in
5081 * stop_machine(): we create temporary high priority worker threads,
5082 * but our caller might not have that capability.
5084 * Return: 0 on success. An error code otherwise.
5086 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5087 const struct sched_param *param)
5089 return _sched_setscheduler(p, policy, param, false);
5091 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5094 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5096 struct sched_param lparam;
5097 struct task_struct *p;
5100 if (!param || pid < 0)
5102 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5107 p = find_process_by_pid(pid);
5113 retval = sched_setscheduler(p, policy, &lparam);
5121 * Mimics kernel/events/core.c perf_copy_attr().
5123 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5128 /* Zero the full structure, so that a short copy will be nice: */
5129 memset(attr, 0, sizeof(*attr));
5131 ret = get_user(size, &uattr->size);
5135 /* ABI compatibility quirk: */
5137 size = SCHED_ATTR_SIZE_VER0;
5138 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5141 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5148 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5149 size < SCHED_ATTR_SIZE_VER1)
5153 * XXX: Do we want to be lenient like existing syscalls; or do we want
5154 * to be strict and return an error on out-of-bounds values?
5156 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5161 put_user(sizeof(*attr), &uattr->size);
5166 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5167 * @pid: the pid in question.
5168 * @policy: new policy.
5169 * @param: structure containing the new RT priority.
5171 * Return: 0 on success. An error code otherwise.
5173 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5178 return do_sched_setscheduler(pid, policy, param);
5182 * sys_sched_setparam - set/change the RT priority of a thread
5183 * @pid: the pid in question.
5184 * @param: structure containing the new RT priority.
5186 * Return: 0 on success. An error code otherwise.
5188 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5190 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5194 * sys_sched_setattr - same as above, but with extended sched_attr
5195 * @pid: the pid in question.
5196 * @uattr: structure containing the extended parameters.
5197 * @flags: for future extension.
5199 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5200 unsigned int, flags)
5202 struct sched_attr attr;
5203 struct task_struct *p;
5206 if (!uattr || pid < 0 || flags)
5209 retval = sched_copy_attr(uattr, &attr);
5213 if ((int)attr.sched_policy < 0)
5215 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5216 attr.sched_policy = SETPARAM_POLICY;
5220 p = find_process_by_pid(pid);
5226 retval = sched_setattr(p, &attr);
5234 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5235 * @pid: the pid in question.
5237 * Return: On success, the policy of the thread. Otherwise, a negative error
5240 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5242 struct task_struct *p;
5250 p = find_process_by_pid(pid);
5252 retval = security_task_getscheduler(p);
5255 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5262 * sys_sched_getparam - get the RT priority of a thread
5263 * @pid: the pid in question.
5264 * @param: structure containing the RT priority.
5266 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5269 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5271 struct sched_param lp = { .sched_priority = 0 };
5272 struct task_struct *p;
5275 if (!param || pid < 0)
5279 p = find_process_by_pid(pid);
5284 retval = security_task_getscheduler(p);
5288 if (task_has_rt_policy(p))
5289 lp.sched_priority = p->rt_priority;
5293 * This one might sleep, we cannot do it with a spinlock held ...
5295 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5305 * Copy the kernel size attribute structure (which might be larger
5306 * than what user-space knows about) to user-space.
5308 * Note that all cases are valid: user-space buffer can be larger or
5309 * smaller than the kernel-space buffer. The usual case is that both
5310 * have the same size.
5313 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5314 struct sched_attr *kattr,
5317 unsigned int ksize = sizeof(*kattr);
5319 if (!access_ok(uattr, usize))
5323 * sched_getattr() ABI forwards and backwards compatibility:
5325 * If usize == ksize then we just copy everything to user-space and all is good.
5327 * If usize < ksize then we only copy as much as user-space has space for,
5328 * this keeps ABI compatibility as well. We skip the rest.
5330 * If usize > ksize then user-space is using a newer version of the ABI,
5331 * which part the kernel doesn't know about. Just ignore it - tooling can
5332 * detect the kernel's knowledge of attributes from the attr->size value
5333 * which is set to ksize in this case.
5335 kattr->size = min(usize, ksize);
5337 if (copy_to_user(uattr, kattr, kattr->size))
5344 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5345 * @pid: the pid in question.
5346 * @uattr: structure containing the extended parameters.
5347 * @usize: sizeof(attr) for fwd/bwd comp.
5348 * @flags: for future extension.
5350 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5351 unsigned int, usize, unsigned int, flags)
5353 struct sched_attr kattr = { };
5354 struct task_struct *p;
5357 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5358 usize < SCHED_ATTR_SIZE_VER0 || flags)
5362 p = find_process_by_pid(pid);
5367 retval = security_task_getscheduler(p);
5371 kattr.sched_policy = p->policy;
5372 if (p->sched_reset_on_fork)
5373 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5374 if (task_has_dl_policy(p))
5375 __getparam_dl(p, &kattr);
5376 else if (task_has_rt_policy(p))
5377 kattr.sched_priority = p->rt_priority;
5379 kattr.sched_nice = task_nice(p);
5381 #ifdef CONFIG_UCLAMP_TASK
5382 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5383 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5388 return sched_attr_copy_to_user(uattr, &kattr, usize);
5395 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5397 cpumask_var_t cpus_allowed, new_mask;
5398 struct task_struct *p;
5403 p = find_process_by_pid(pid);
5409 /* Prevent p going away */
5413 if (p->flags & PF_NO_SETAFFINITY) {
5417 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5421 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5423 goto out_free_cpus_allowed;
5426 if (!check_same_owner(p)) {
5428 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5430 goto out_free_new_mask;
5435 retval = security_task_setscheduler(p);
5437 goto out_free_new_mask;
5440 cpuset_cpus_allowed(p, cpus_allowed);
5441 cpumask_and(new_mask, in_mask, cpus_allowed);
5444 * Since bandwidth control happens on root_domain basis,
5445 * if admission test is enabled, we only admit -deadline
5446 * tasks allowed to run on all the CPUs in the task's
5450 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5452 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5455 goto out_free_new_mask;
5461 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5464 cpuset_cpus_allowed(p, cpus_allowed);
5465 if (!cpumask_subset(new_mask, cpus_allowed)) {
5467 * We must have raced with a concurrent cpuset
5468 * update. Just reset the cpus_allowed to the
5469 * cpuset's cpus_allowed
5471 cpumask_copy(new_mask, cpus_allowed);
5476 free_cpumask_var(new_mask);
5477 out_free_cpus_allowed:
5478 free_cpumask_var(cpus_allowed);
5484 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5485 struct cpumask *new_mask)
5487 if (len < cpumask_size())
5488 cpumask_clear(new_mask);
5489 else if (len > cpumask_size())
5490 len = cpumask_size();
5492 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5496 * sys_sched_setaffinity - set the CPU affinity of a process
5497 * @pid: pid of the process
5498 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5499 * @user_mask_ptr: user-space pointer to the new CPU mask
5501 * Return: 0 on success. An error code otherwise.
5503 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5504 unsigned long __user *, user_mask_ptr)
5506 cpumask_var_t new_mask;
5509 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5512 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5514 retval = sched_setaffinity(pid, new_mask);
5515 free_cpumask_var(new_mask);
5519 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5521 struct task_struct *p;
5522 unsigned long flags;
5528 p = find_process_by_pid(pid);
5532 retval = security_task_getscheduler(p);
5536 raw_spin_lock_irqsave(&p->pi_lock, flags);
5537 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5538 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5547 * sys_sched_getaffinity - get the CPU affinity of a process
5548 * @pid: pid of the process
5549 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5550 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5552 * Return: size of CPU mask copied to user_mask_ptr on success. An
5553 * error code otherwise.
5555 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5556 unsigned long __user *, user_mask_ptr)
5561 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5563 if (len & (sizeof(unsigned long)-1))
5566 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5569 ret = sched_getaffinity(pid, mask);
5571 unsigned int retlen = min(len, cpumask_size());
5573 if (copy_to_user(user_mask_ptr, mask, retlen))
5578 free_cpumask_var(mask);
5584 * sys_sched_yield - yield the current processor to other threads.
5586 * This function yields the current CPU to other tasks. If there are no
5587 * other threads running on this CPU then this function will return.
5591 static void do_sched_yield(void)
5596 rq = this_rq_lock_irq(&rf);
5598 schedstat_inc(rq->yld_count);
5599 current->sched_class->yield_task(rq);
5602 * Since we are going to call schedule() anyway, there's
5603 * no need to preempt or enable interrupts:
5607 sched_preempt_enable_no_resched();
5612 SYSCALL_DEFINE0(sched_yield)
5618 #ifndef CONFIG_PREEMPTION
5619 int __sched _cond_resched(void)
5621 if (should_resched(0)) {
5622 preempt_schedule_common();
5628 EXPORT_SYMBOL(_cond_resched);
5632 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5633 * call schedule, and on return reacquire the lock.
5635 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5636 * operations here to prevent schedule() from being called twice (once via
5637 * spin_unlock(), once by hand).
5639 int __cond_resched_lock(spinlock_t *lock)
5641 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5644 lockdep_assert_held(lock);
5646 if (spin_needbreak(lock) || resched) {
5649 preempt_schedule_common();
5657 EXPORT_SYMBOL(__cond_resched_lock);
5660 * yield - yield the current processor to other threads.
5662 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5664 * The scheduler is at all times free to pick the calling task as the most
5665 * eligible task to run, if removing the yield() call from your code breaks
5666 * it, its already broken.
5668 * Typical broken usage is:
5673 * where one assumes that yield() will let 'the other' process run that will
5674 * make event true. If the current task is a SCHED_FIFO task that will never
5675 * happen. Never use yield() as a progress guarantee!!
5677 * If you want to use yield() to wait for something, use wait_event().
5678 * If you want to use yield() to be 'nice' for others, use cond_resched().
5679 * If you still want to use yield(), do not!
5681 void __sched yield(void)
5683 set_current_state(TASK_RUNNING);
5686 EXPORT_SYMBOL(yield);
5689 * yield_to - yield the current processor to another thread in
5690 * your thread group, or accelerate that thread toward the
5691 * processor it's on.
5693 * @preempt: whether task preemption is allowed or not
5695 * It's the caller's job to ensure that the target task struct
5696 * can't go away on us before we can do any checks.
5699 * true (>0) if we indeed boosted the target task.
5700 * false (0) if we failed to boost the target.
5701 * -ESRCH if there's no task to yield to.
5703 int __sched yield_to(struct task_struct *p, bool preempt)
5705 struct task_struct *curr = current;
5706 struct rq *rq, *p_rq;
5707 unsigned long flags;
5710 local_irq_save(flags);
5716 * If we're the only runnable task on the rq and target rq also
5717 * has only one task, there's absolutely no point in yielding.
5719 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5724 double_rq_lock(rq, p_rq);
5725 if (task_rq(p) != p_rq) {
5726 double_rq_unlock(rq, p_rq);
5730 if (!curr->sched_class->yield_to_task)
5733 if (curr->sched_class != p->sched_class)
5736 if (task_running(p_rq, p) || p->state)
5739 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5741 schedstat_inc(rq->yld_count);
5743 * Make p's CPU reschedule; pick_next_entity takes care of
5746 if (preempt && rq != p_rq)
5751 double_rq_unlock(rq, p_rq);
5753 local_irq_restore(flags);
5760 EXPORT_SYMBOL_GPL(yield_to);
5762 int io_schedule_prepare(void)
5764 int old_iowait = current->in_iowait;
5766 current->in_iowait = 1;
5767 blk_schedule_flush_plug(current);
5772 void io_schedule_finish(int token)
5774 current->in_iowait = token;
5778 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5779 * that process accounting knows that this is a task in IO wait state.
5781 long __sched io_schedule_timeout(long timeout)
5786 token = io_schedule_prepare();
5787 ret = schedule_timeout(timeout);
5788 io_schedule_finish(token);
5792 EXPORT_SYMBOL(io_schedule_timeout);
5794 void __sched io_schedule(void)
5798 token = io_schedule_prepare();
5800 io_schedule_finish(token);
5802 EXPORT_SYMBOL(io_schedule);
5805 * sys_sched_get_priority_max - return maximum RT priority.
5806 * @policy: scheduling class.
5808 * Return: On success, this syscall returns the maximum
5809 * rt_priority that can be used by a given scheduling class.
5810 * On failure, a negative error code is returned.
5812 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5819 ret = MAX_USER_RT_PRIO-1;
5821 case SCHED_DEADLINE:
5832 * sys_sched_get_priority_min - return minimum RT priority.
5833 * @policy: scheduling class.
5835 * Return: On success, this syscall returns the minimum
5836 * rt_priority that can be used by a given scheduling class.
5837 * On failure, a negative error code is returned.
5839 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5848 case SCHED_DEADLINE:
5857 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5859 struct task_struct *p;
5860 unsigned int time_slice;
5870 p = find_process_by_pid(pid);
5874 retval = security_task_getscheduler(p);
5878 rq = task_rq_lock(p, &rf);
5880 if (p->sched_class->get_rr_interval)
5881 time_slice = p->sched_class->get_rr_interval(rq, p);
5882 task_rq_unlock(rq, p, &rf);
5885 jiffies_to_timespec64(time_slice, t);
5894 * sys_sched_rr_get_interval - return the default timeslice of a process.
5895 * @pid: pid of the process.
5896 * @interval: userspace pointer to the timeslice value.
5898 * this syscall writes the default timeslice value of a given process
5899 * into the user-space timespec buffer. A value of '0' means infinity.
5901 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5904 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5905 struct __kernel_timespec __user *, interval)
5907 struct timespec64 t;
5908 int retval = sched_rr_get_interval(pid, &t);
5911 retval = put_timespec64(&t, interval);
5916 #ifdef CONFIG_COMPAT_32BIT_TIME
5917 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5918 struct old_timespec32 __user *, interval)
5920 struct timespec64 t;
5921 int retval = sched_rr_get_interval(pid, &t);
5924 retval = put_old_timespec32(&t, interval);
5929 void sched_show_task(struct task_struct *p)
5931 unsigned long free = 0;
5934 if (!try_get_task_stack(p))
5937 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5939 if (p->state == TASK_RUNNING)
5940 printk(KERN_CONT " running task ");
5941 #ifdef CONFIG_DEBUG_STACK_USAGE
5942 free = stack_not_used(p);
5947 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5949 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5950 task_pid_nr(p), ppid,
5951 (unsigned long)task_thread_info(p)->flags);
5953 print_worker_info(KERN_INFO, p);
5954 show_stack(p, NULL);
5957 EXPORT_SYMBOL_GPL(sched_show_task);
5960 state_filter_match(unsigned long state_filter, struct task_struct *p)
5962 /* no filter, everything matches */
5966 /* filter, but doesn't match */
5967 if (!(p->state & state_filter))
5971 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5974 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5981 void show_state_filter(unsigned long state_filter)
5983 struct task_struct *g, *p;
5985 #if BITS_PER_LONG == 32
5987 " task PC stack pid father\n");
5990 " task PC stack pid father\n");
5993 for_each_process_thread(g, p) {
5995 * reset the NMI-timeout, listing all files on a slow
5996 * console might take a lot of time:
5997 * Also, reset softlockup watchdogs on all CPUs, because
5998 * another CPU might be blocked waiting for us to process
6001 touch_nmi_watchdog();
6002 touch_all_softlockup_watchdogs();
6003 if (state_filter_match(state_filter, p))
6007 #ifdef CONFIG_SCHED_DEBUG
6009 sysrq_sched_debug_show();
6013 * Only show locks if all tasks are dumped:
6016 debug_show_all_locks();
6020 * init_idle - set up an idle thread for a given CPU
6021 * @idle: task in question
6022 * @cpu: CPU the idle task belongs to
6024 * NOTE: this function does not set the idle thread's NEED_RESCHED
6025 * flag, to make booting more robust.
6027 void init_idle(struct task_struct *idle, int cpu)
6029 struct rq *rq = cpu_rq(cpu);
6030 unsigned long flags;
6032 __sched_fork(0, idle);
6034 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6035 raw_spin_lock(&rq->lock);
6037 idle->state = TASK_RUNNING;
6038 idle->se.exec_start = sched_clock();
6039 idle->flags |= PF_IDLE;
6041 kasan_unpoison_task_stack(idle);
6045 * Its possible that init_idle() gets called multiple times on a task,
6046 * in that case do_set_cpus_allowed() will not do the right thing.
6048 * And since this is boot we can forgo the serialization.
6050 set_cpus_allowed_common(idle, cpumask_of(cpu));
6053 * We're having a chicken and egg problem, even though we are
6054 * holding rq->lock, the CPU isn't yet set to this CPU so the
6055 * lockdep check in task_group() will fail.
6057 * Similar case to sched_fork(). / Alternatively we could
6058 * use task_rq_lock() here and obtain the other rq->lock.
6063 __set_task_cpu(idle, cpu);
6067 rcu_assign_pointer(rq->curr, idle);
6068 idle->on_rq = TASK_ON_RQ_QUEUED;
6072 raw_spin_unlock(&rq->lock);
6073 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6075 /* Set the preempt count _outside_ the spinlocks! */
6076 init_idle_preempt_count(idle, cpu);
6079 * The idle tasks have their own, simple scheduling class:
6081 idle->sched_class = &idle_sched_class;
6082 ftrace_graph_init_idle_task(idle, cpu);
6083 vtime_init_idle(idle, cpu);
6085 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6091 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6092 const struct cpumask *trial)
6096 if (!cpumask_weight(cur))
6099 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6104 int task_can_attach(struct task_struct *p,
6105 const struct cpumask *cs_cpus_allowed)
6110 * Kthreads which disallow setaffinity shouldn't be moved
6111 * to a new cpuset; we don't want to change their CPU
6112 * affinity and isolating such threads by their set of
6113 * allowed nodes is unnecessary. Thus, cpusets are not
6114 * applicable for such threads. This prevents checking for
6115 * success of set_cpus_allowed_ptr() on all attached tasks
6116 * before cpus_mask may be changed.
6118 if (p->flags & PF_NO_SETAFFINITY) {
6123 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6125 ret = dl_task_can_attach(p, cs_cpus_allowed);
6131 bool sched_smp_initialized __read_mostly;
6133 #ifdef CONFIG_NUMA_BALANCING
6134 /* Migrate current task p to target_cpu */
6135 int migrate_task_to(struct task_struct *p, int target_cpu)
6137 struct migration_arg arg = { p, target_cpu };
6138 int curr_cpu = task_cpu(p);
6140 if (curr_cpu == target_cpu)
6143 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6146 /* TODO: This is not properly updating schedstats */
6148 trace_sched_move_numa(p, curr_cpu, target_cpu);
6149 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6153 * Requeue a task on a given node and accurately track the number of NUMA
6154 * tasks on the runqueues
6156 void sched_setnuma(struct task_struct *p, int nid)
6158 bool queued, running;
6162 rq = task_rq_lock(p, &rf);
6163 queued = task_on_rq_queued(p);
6164 running = task_current(rq, p);
6167 dequeue_task(rq, p, DEQUEUE_SAVE);
6169 put_prev_task(rq, p);
6171 p->numa_preferred_nid = nid;
6174 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6176 set_next_task(rq, p);
6177 task_rq_unlock(rq, p, &rf);
6179 #endif /* CONFIG_NUMA_BALANCING */
6181 #ifdef CONFIG_HOTPLUG_CPU
6183 * Ensure that the idle task is using init_mm right before its CPU goes
6186 void idle_task_exit(void)
6188 struct mm_struct *mm = current->active_mm;
6190 BUG_ON(cpu_online(smp_processor_id()));
6192 if (mm != &init_mm) {
6193 switch_mm(mm, &init_mm, current);
6194 current->active_mm = &init_mm;
6195 finish_arch_post_lock_switch();
6201 * Since this CPU is going 'away' for a while, fold any nr_active delta
6202 * we might have. Assumes we're called after migrate_tasks() so that the
6203 * nr_active count is stable. We need to take the teardown thread which
6204 * is calling this into account, so we hand in adjust = 1 to the load
6207 * Also see the comment "Global load-average calculations".
6209 static void calc_load_migrate(struct rq *rq)
6211 long delta = calc_load_fold_active(rq, 1);
6213 atomic_long_add(delta, &calc_load_tasks);
6216 static struct task_struct *__pick_migrate_task(struct rq *rq)
6218 const struct sched_class *class;
6219 struct task_struct *next;
6221 for_each_class(class) {
6222 next = class->pick_next_task(rq);
6224 next->sched_class->put_prev_task(rq, next);
6229 /* The idle class should always have a runnable task */
6234 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6235 * try_to_wake_up()->select_task_rq().
6237 * Called with rq->lock held even though we'er in stop_machine() and
6238 * there's no concurrency possible, we hold the required locks anyway
6239 * because of lock validation efforts.
6241 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6243 struct rq *rq = dead_rq;
6244 struct task_struct *next, *stop = rq->stop;
6245 struct rq_flags orf = *rf;
6249 * Fudge the rq selection such that the below task selection loop
6250 * doesn't get stuck on the currently eligible stop task.
6252 * We're currently inside stop_machine() and the rq is either stuck
6253 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6254 * either way we should never end up calling schedule() until we're
6260 * put_prev_task() and pick_next_task() sched
6261 * class method both need to have an up-to-date
6262 * value of rq->clock[_task]
6264 update_rq_clock(rq);
6268 * There's this thread running, bail when that's the only
6271 if (rq->nr_running == 1)
6274 next = __pick_migrate_task(rq);
6277 * Rules for changing task_struct::cpus_mask are holding
6278 * both pi_lock and rq->lock, such that holding either
6279 * stabilizes the mask.
6281 * Drop rq->lock is not quite as disastrous as it usually is
6282 * because !cpu_active at this point, which means load-balance
6283 * will not interfere. Also, stop-machine.
6286 raw_spin_lock(&next->pi_lock);
6290 * Since we're inside stop-machine, _nothing_ should have
6291 * changed the task, WARN if weird stuff happened, because in
6292 * that case the above rq->lock drop is a fail too.
6294 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6295 raw_spin_unlock(&next->pi_lock);
6299 /* Find suitable destination for @next, with force if needed. */
6300 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6301 rq = __migrate_task(rq, rf, next, dest_cpu);
6302 if (rq != dead_rq) {
6308 raw_spin_unlock(&next->pi_lock);
6313 #endif /* CONFIG_HOTPLUG_CPU */
6315 void set_rq_online(struct rq *rq)
6318 const struct sched_class *class;
6320 cpumask_set_cpu(rq->cpu, rq->rd->online);
6323 for_each_class(class) {
6324 if (class->rq_online)
6325 class->rq_online(rq);
6330 void set_rq_offline(struct rq *rq)
6333 const struct sched_class *class;
6335 for_each_class(class) {
6336 if (class->rq_offline)
6337 class->rq_offline(rq);
6340 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6346 * used to mark begin/end of suspend/resume:
6348 static int num_cpus_frozen;
6351 * Update cpusets according to cpu_active mask. If cpusets are
6352 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6353 * around partition_sched_domains().
6355 * If we come here as part of a suspend/resume, don't touch cpusets because we
6356 * want to restore it back to its original state upon resume anyway.
6358 static void cpuset_cpu_active(void)
6360 if (cpuhp_tasks_frozen) {
6362 * num_cpus_frozen tracks how many CPUs are involved in suspend
6363 * resume sequence. As long as this is not the last online
6364 * operation in the resume sequence, just build a single sched
6365 * domain, ignoring cpusets.
6367 partition_sched_domains(1, NULL, NULL);
6368 if (--num_cpus_frozen)
6371 * This is the last CPU online operation. So fall through and
6372 * restore the original sched domains by considering the
6373 * cpuset configurations.
6375 cpuset_force_rebuild();
6377 cpuset_update_active_cpus();
6380 static int cpuset_cpu_inactive(unsigned int cpu)
6382 if (!cpuhp_tasks_frozen) {
6383 if (dl_cpu_busy(cpu))
6385 cpuset_update_active_cpus();
6388 partition_sched_domains(1, NULL, NULL);
6393 int sched_cpu_activate(unsigned int cpu)
6395 struct rq *rq = cpu_rq(cpu);
6398 #ifdef CONFIG_SCHED_SMT
6400 * When going up, increment the number of cores with SMT present.
6402 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6403 static_branch_inc_cpuslocked(&sched_smt_present);
6405 set_cpu_active(cpu, true);
6407 if (sched_smp_initialized) {
6408 sched_domains_numa_masks_set(cpu);
6409 cpuset_cpu_active();
6413 * Put the rq online, if not already. This happens:
6415 * 1) In the early boot process, because we build the real domains
6416 * after all CPUs have been brought up.
6418 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6421 rq_lock_irqsave(rq, &rf);
6423 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6426 rq_unlock_irqrestore(rq, &rf);
6431 int sched_cpu_deactivate(unsigned int cpu)
6435 set_cpu_active(cpu, false);
6437 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6438 * users of this state to go away such that all new such users will
6441 * Do sync before park smpboot threads to take care the rcu boost case.
6445 #ifdef CONFIG_SCHED_SMT
6447 * When going down, decrement the number of cores with SMT present.
6449 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6450 static_branch_dec_cpuslocked(&sched_smt_present);
6453 if (!sched_smp_initialized)
6456 ret = cpuset_cpu_inactive(cpu);
6458 set_cpu_active(cpu, true);
6461 sched_domains_numa_masks_clear(cpu);
6465 static void sched_rq_cpu_starting(unsigned int cpu)
6467 struct rq *rq = cpu_rq(cpu);
6469 rq->calc_load_update = calc_load_update;
6470 update_max_interval();
6473 int sched_cpu_starting(unsigned int cpu)
6475 sched_rq_cpu_starting(cpu);
6476 sched_tick_start(cpu);
6480 #ifdef CONFIG_HOTPLUG_CPU
6481 int sched_cpu_dying(unsigned int cpu)
6483 struct rq *rq = cpu_rq(cpu);
6486 /* Handle pending wakeups and then migrate everything off */
6487 sched_ttwu_pending();
6488 sched_tick_stop(cpu);
6490 rq_lock_irqsave(rq, &rf);
6492 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6495 migrate_tasks(rq, &rf);
6496 BUG_ON(rq->nr_running != 1);
6497 rq_unlock_irqrestore(rq, &rf);
6499 calc_load_migrate(rq);
6500 update_max_interval();
6501 nohz_balance_exit_idle(rq);
6507 void __init sched_init_smp(void)
6512 * There's no userspace yet to cause hotplug operations; hence all the
6513 * CPU masks are stable and all blatant races in the below code cannot
6516 mutex_lock(&sched_domains_mutex);
6517 sched_init_domains(cpu_active_mask);
6518 mutex_unlock(&sched_domains_mutex);
6520 /* Move init over to a non-isolated CPU */
6521 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6523 sched_init_granularity();
6525 init_sched_rt_class();
6526 init_sched_dl_class();
6528 sched_smp_initialized = true;
6531 static int __init migration_init(void)
6533 sched_cpu_starting(smp_processor_id());
6536 early_initcall(migration_init);
6539 void __init sched_init_smp(void)
6541 sched_init_granularity();
6543 #endif /* CONFIG_SMP */
6545 int in_sched_functions(unsigned long addr)
6547 return in_lock_functions(addr) ||
6548 (addr >= (unsigned long)__sched_text_start
6549 && addr < (unsigned long)__sched_text_end);
6552 #ifdef CONFIG_CGROUP_SCHED
6554 * Default task group.
6555 * Every task in system belongs to this group at bootup.
6557 struct task_group root_task_group;
6558 LIST_HEAD(task_groups);
6560 /* Cacheline aligned slab cache for task_group */
6561 static struct kmem_cache *task_group_cache __read_mostly;
6564 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6565 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6567 void __init sched_init(void)
6569 unsigned long ptr = 0;
6574 #ifdef CONFIG_FAIR_GROUP_SCHED
6575 ptr += 2 * nr_cpu_ids * sizeof(void **);
6577 #ifdef CONFIG_RT_GROUP_SCHED
6578 ptr += 2 * nr_cpu_ids * sizeof(void **);
6581 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6583 #ifdef CONFIG_FAIR_GROUP_SCHED
6584 root_task_group.se = (struct sched_entity **)ptr;
6585 ptr += nr_cpu_ids * sizeof(void **);
6587 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6588 ptr += nr_cpu_ids * sizeof(void **);
6590 #endif /* CONFIG_FAIR_GROUP_SCHED */
6591 #ifdef CONFIG_RT_GROUP_SCHED
6592 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6593 ptr += nr_cpu_ids * sizeof(void **);
6595 root_task_group.rt_rq = (struct rt_rq **)ptr;
6596 ptr += nr_cpu_ids * sizeof(void **);
6598 #endif /* CONFIG_RT_GROUP_SCHED */
6600 #ifdef CONFIG_CPUMASK_OFFSTACK
6601 for_each_possible_cpu(i) {
6602 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6603 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6604 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6605 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6607 #endif /* CONFIG_CPUMASK_OFFSTACK */
6609 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6610 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6613 init_defrootdomain();
6616 #ifdef CONFIG_RT_GROUP_SCHED
6617 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6618 global_rt_period(), global_rt_runtime());
6619 #endif /* CONFIG_RT_GROUP_SCHED */
6621 #ifdef CONFIG_CGROUP_SCHED
6622 task_group_cache = KMEM_CACHE(task_group, 0);
6624 list_add(&root_task_group.list, &task_groups);
6625 INIT_LIST_HEAD(&root_task_group.children);
6626 INIT_LIST_HEAD(&root_task_group.siblings);
6627 autogroup_init(&init_task);
6628 #endif /* CONFIG_CGROUP_SCHED */
6630 for_each_possible_cpu(i) {
6634 raw_spin_lock_init(&rq->lock);
6636 rq->calc_load_active = 0;
6637 rq->calc_load_update = jiffies + LOAD_FREQ;
6638 init_cfs_rq(&rq->cfs);
6639 init_rt_rq(&rq->rt);
6640 init_dl_rq(&rq->dl);
6641 #ifdef CONFIG_FAIR_GROUP_SCHED
6642 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6643 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6644 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6646 * How much CPU bandwidth does root_task_group get?
6648 * In case of task-groups formed thr' the cgroup filesystem, it
6649 * gets 100% of the CPU resources in the system. This overall
6650 * system CPU resource is divided among the tasks of
6651 * root_task_group and its child task-groups in a fair manner,
6652 * based on each entity's (task or task-group's) weight
6653 * (se->load.weight).
6655 * In other words, if root_task_group has 10 tasks of weight
6656 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6657 * then A0's share of the CPU resource is:
6659 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6661 * We achieve this by letting root_task_group's tasks sit
6662 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6664 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6665 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6666 #endif /* CONFIG_FAIR_GROUP_SCHED */
6668 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6669 #ifdef CONFIG_RT_GROUP_SCHED
6670 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6675 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6676 rq->balance_callback = NULL;
6677 rq->active_balance = 0;
6678 rq->next_balance = jiffies;
6683 rq->avg_idle = 2*sysctl_sched_migration_cost;
6684 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6686 INIT_LIST_HEAD(&rq->cfs_tasks);
6688 rq_attach_root(rq, &def_root_domain);
6689 #ifdef CONFIG_NO_HZ_COMMON
6690 rq->last_load_update_tick = jiffies;
6691 rq->last_blocked_load_update_tick = jiffies;
6692 atomic_set(&rq->nohz_flags, 0);
6694 #endif /* CONFIG_SMP */
6696 atomic_set(&rq->nr_iowait, 0);
6699 set_load_weight(&init_task, false);
6702 * The boot idle thread does lazy MMU switching as well:
6705 enter_lazy_tlb(&init_mm, current);
6708 * Make us the idle thread. Technically, schedule() should not be
6709 * called from this thread, however somewhere below it might be,
6710 * but because we are the idle thread, we just pick up running again
6711 * when this runqueue becomes "idle".
6713 init_idle(current, smp_processor_id());
6715 calc_load_update = jiffies + LOAD_FREQ;
6718 idle_thread_set_boot_cpu();
6720 init_sched_fair_class();
6728 scheduler_running = 1;
6731 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6732 static inline int preempt_count_equals(int preempt_offset)
6734 int nested = preempt_count() + rcu_preempt_depth();
6736 return (nested == preempt_offset);
6739 void __might_sleep(const char *file, int line, int preempt_offset)
6742 * Blocking primitives will set (and therefore destroy) current->state,
6743 * since we will exit with TASK_RUNNING make sure we enter with it,
6744 * otherwise we will destroy state.
6746 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6747 "do not call blocking ops when !TASK_RUNNING; "
6748 "state=%lx set at [<%p>] %pS\n",
6750 (void *)current->task_state_change,
6751 (void *)current->task_state_change);
6753 ___might_sleep(file, line, preempt_offset);
6755 EXPORT_SYMBOL(__might_sleep);
6757 void ___might_sleep(const char *file, int line, int preempt_offset)
6759 /* Ratelimiting timestamp: */
6760 static unsigned long prev_jiffy;
6762 unsigned long preempt_disable_ip;
6764 /* WARN_ON_ONCE() by default, no rate limit required: */
6767 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6768 !is_idle_task(current) && !current->non_block_count) ||
6769 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6773 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6775 prev_jiffy = jiffies;
6777 /* Save this before calling printk(), since that will clobber it: */
6778 preempt_disable_ip = get_preempt_disable_ip(current);
6781 "BUG: sleeping function called from invalid context at %s:%d\n",
6784 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6785 in_atomic(), irqs_disabled(), current->non_block_count,
6786 current->pid, current->comm);
6788 if (task_stack_end_corrupted(current))
6789 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6791 debug_show_held_locks(current);
6792 if (irqs_disabled())
6793 print_irqtrace_events(current);
6794 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6795 && !preempt_count_equals(preempt_offset)) {
6796 pr_err("Preemption disabled at:");
6797 print_ip_sym(preempt_disable_ip);
6801 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6803 EXPORT_SYMBOL(___might_sleep);
6805 void __cant_sleep(const char *file, int line, int preempt_offset)
6807 static unsigned long prev_jiffy;
6809 if (irqs_disabled())
6812 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6815 if (preempt_count() > preempt_offset)
6818 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6820 prev_jiffy = jiffies;
6822 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6823 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6824 in_atomic(), irqs_disabled(),
6825 current->pid, current->comm);
6827 debug_show_held_locks(current);
6829 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6831 EXPORT_SYMBOL_GPL(__cant_sleep);
6834 #ifdef CONFIG_MAGIC_SYSRQ
6835 void normalize_rt_tasks(void)
6837 struct task_struct *g, *p;
6838 struct sched_attr attr = {
6839 .sched_policy = SCHED_NORMAL,
6842 read_lock(&tasklist_lock);
6843 for_each_process_thread(g, p) {
6845 * Only normalize user tasks:
6847 if (p->flags & PF_KTHREAD)
6850 p->se.exec_start = 0;
6851 schedstat_set(p->se.statistics.wait_start, 0);
6852 schedstat_set(p->se.statistics.sleep_start, 0);
6853 schedstat_set(p->se.statistics.block_start, 0);
6855 if (!dl_task(p) && !rt_task(p)) {
6857 * Renice negative nice level userspace
6860 if (task_nice(p) < 0)
6861 set_user_nice(p, 0);
6865 __sched_setscheduler(p, &attr, false, false);
6867 read_unlock(&tasklist_lock);
6870 #endif /* CONFIG_MAGIC_SYSRQ */
6872 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6874 * These functions are only useful for the IA64 MCA handling, or kdb.
6876 * They can only be called when the whole system has been
6877 * stopped - every CPU needs to be quiescent, and no scheduling
6878 * activity can take place. Using them for anything else would
6879 * be a serious bug, and as a result, they aren't even visible
6880 * under any other configuration.
6884 * curr_task - return the current task for a given CPU.
6885 * @cpu: the processor in question.
6887 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6889 * Return: The current task for @cpu.
6891 struct task_struct *curr_task(int cpu)
6893 return cpu_curr(cpu);
6896 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6900 * ia64_set_curr_task - set the current task for a given CPU.
6901 * @cpu: the processor in question.
6902 * @p: the task pointer to set.
6904 * Description: This function must only be used when non-maskable interrupts
6905 * are serviced on a separate stack. It allows the architecture to switch the
6906 * notion of the current task on a CPU in a non-blocking manner. This function
6907 * must be called with all CPU's synchronized, and interrupts disabled, the
6908 * and caller must save the original value of the current task (see
6909 * curr_task() above) and restore that value before reenabling interrupts and
6910 * re-starting the system.
6912 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6914 void ia64_set_curr_task(int cpu, struct task_struct *p)
6921 #ifdef CONFIG_CGROUP_SCHED
6922 /* task_group_lock serializes the addition/removal of task groups */
6923 static DEFINE_SPINLOCK(task_group_lock);
6925 static inline void alloc_uclamp_sched_group(struct task_group *tg,
6926 struct task_group *parent)
6928 #ifdef CONFIG_UCLAMP_TASK_GROUP
6929 enum uclamp_id clamp_id;
6931 for_each_clamp_id(clamp_id) {
6932 uclamp_se_set(&tg->uclamp_req[clamp_id],
6933 uclamp_none(clamp_id), false);
6934 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
6939 static void sched_free_group(struct task_group *tg)
6941 free_fair_sched_group(tg);
6942 free_rt_sched_group(tg);
6944 kmem_cache_free(task_group_cache, tg);
6947 /* allocate runqueue etc for a new task group */
6948 struct task_group *sched_create_group(struct task_group *parent)
6950 struct task_group *tg;
6952 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6954 return ERR_PTR(-ENOMEM);
6956 if (!alloc_fair_sched_group(tg, parent))
6959 if (!alloc_rt_sched_group(tg, parent))
6962 alloc_uclamp_sched_group(tg, parent);
6967 sched_free_group(tg);
6968 return ERR_PTR(-ENOMEM);
6971 void sched_online_group(struct task_group *tg, struct task_group *parent)
6973 unsigned long flags;
6975 spin_lock_irqsave(&task_group_lock, flags);
6976 list_add_rcu(&tg->list, &task_groups);
6978 /* Root should already exist: */
6981 tg->parent = parent;
6982 INIT_LIST_HEAD(&tg->children);
6983 list_add_rcu(&tg->siblings, &parent->children);
6984 spin_unlock_irqrestore(&task_group_lock, flags);
6986 online_fair_sched_group(tg);
6989 /* rcu callback to free various structures associated with a task group */
6990 static void sched_free_group_rcu(struct rcu_head *rhp)
6992 /* Now it should be safe to free those cfs_rqs: */
6993 sched_free_group(container_of(rhp, struct task_group, rcu));
6996 void sched_destroy_group(struct task_group *tg)
6998 /* Wait for possible concurrent references to cfs_rqs complete: */
6999 call_rcu(&tg->rcu, sched_free_group_rcu);
7002 void sched_offline_group(struct task_group *tg)
7004 unsigned long flags;
7006 /* End participation in shares distribution: */
7007 unregister_fair_sched_group(tg);
7009 spin_lock_irqsave(&task_group_lock, flags);
7010 list_del_rcu(&tg->list);
7011 list_del_rcu(&tg->siblings);
7012 spin_unlock_irqrestore(&task_group_lock, flags);
7015 static void sched_change_group(struct task_struct *tsk, int type)
7017 struct task_group *tg;
7020 * All callers are synchronized by task_rq_lock(); we do not use RCU
7021 * which is pointless here. Thus, we pass "true" to task_css_check()
7022 * to prevent lockdep warnings.
7024 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7025 struct task_group, css);
7026 tg = autogroup_task_group(tsk, tg);
7027 tsk->sched_task_group = tg;
7029 #ifdef CONFIG_FAIR_GROUP_SCHED
7030 if (tsk->sched_class->task_change_group)
7031 tsk->sched_class->task_change_group(tsk, type);
7034 set_task_rq(tsk, task_cpu(tsk));
7038 * Change task's runqueue when it moves between groups.
7040 * The caller of this function should have put the task in its new group by
7041 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7044 void sched_move_task(struct task_struct *tsk)
7046 int queued, running, queue_flags =
7047 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7051 rq = task_rq_lock(tsk, &rf);
7052 update_rq_clock(rq);
7054 running = task_current(rq, tsk);
7055 queued = task_on_rq_queued(tsk);
7058 dequeue_task(rq, tsk, queue_flags);
7060 put_prev_task(rq, tsk);
7062 sched_change_group(tsk, TASK_MOVE_GROUP);
7065 enqueue_task(rq, tsk, queue_flags);
7067 set_next_task(rq, tsk);
7069 task_rq_unlock(rq, tsk, &rf);
7072 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7074 return css ? container_of(css, struct task_group, css) : NULL;
7077 static struct cgroup_subsys_state *
7078 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7080 struct task_group *parent = css_tg(parent_css);
7081 struct task_group *tg;
7084 /* This is early initialization for the top cgroup */
7085 return &root_task_group.css;
7088 tg = sched_create_group(parent);
7090 return ERR_PTR(-ENOMEM);
7095 /* Expose task group only after completing cgroup initialization */
7096 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7098 struct task_group *tg = css_tg(css);
7099 struct task_group *parent = css_tg(css->parent);
7102 sched_online_group(tg, parent);
7106 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7108 struct task_group *tg = css_tg(css);
7110 sched_offline_group(tg);
7113 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7115 struct task_group *tg = css_tg(css);
7118 * Relies on the RCU grace period between css_released() and this.
7120 sched_free_group(tg);
7124 * This is called before wake_up_new_task(), therefore we really only
7125 * have to set its group bits, all the other stuff does not apply.
7127 static void cpu_cgroup_fork(struct task_struct *task)
7132 rq = task_rq_lock(task, &rf);
7134 update_rq_clock(rq);
7135 sched_change_group(task, TASK_SET_GROUP);
7137 task_rq_unlock(rq, task, &rf);
7140 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7142 struct task_struct *task;
7143 struct cgroup_subsys_state *css;
7146 cgroup_taskset_for_each(task, css, tset) {
7147 #ifdef CONFIG_RT_GROUP_SCHED
7148 if (!sched_rt_can_attach(css_tg(css), task))
7152 * Serialize against wake_up_new_task() such that if its
7153 * running, we're sure to observe its full state.
7155 raw_spin_lock_irq(&task->pi_lock);
7157 * Avoid calling sched_move_task() before wake_up_new_task()
7158 * has happened. This would lead to problems with PELT, due to
7159 * move wanting to detach+attach while we're not attached yet.
7161 if (task->state == TASK_NEW)
7163 raw_spin_unlock_irq(&task->pi_lock);
7171 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7173 struct task_struct *task;
7174 struct cgroup_subsys_state *css;
7176 cgroup_taskset_for_each(task, css, tset)
7177 sched_move_task(task);
7180 #ifdef CONFIG_UCLAMP_TASK_GROUP
7181 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7183 struct cgroup_subsys_state *top_css = css;
7184 struct uclamp_se *uc_parent = NULL;
7185 struct uclamp_se *uc_se = NULL;
7186 unsigned int eff[UCLAMP_CNT];
7187 enum uclamp_id clamp_id;
7188 unsigned int clamps;
7190 css_for_each_descendant_pre(css, top_css) {
7191 uc_parent = css_tg(css)->parent
7192 ? css_tg(css)->parent->uclamp : NULL;
7194 for_each_clamp_id(clamp_id) {
7195 /* Assume effective clamps matches requested clamps */
7196 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7197 /* Cap effective clamps with parent's effective clamps */
7199 eff[clamp_id] > uc_parent[clamp_id].value) {
7200 eff[clamp_id] = uc_parent[clamp_id].value;
7203 /* Ensure protection is always capped by limit */
7204 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7206 /* Propagate most restrictive effective clamps */
7208 uc_se = css_tg(css)->uclamp;
7209 for_each_clamp_id(clamp_id) {
7210 if (eff[clamp_id] == uc_se[clamp_id].value)
7212 uc_se[clamp_id].value = eff[clamp_id];
7213 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7214 clamps |= (0x1 << clamp_id);
7217 css = css_rightmost_descendant(css);
7221 /* Immediately update descendants RUNNABLE tasks */
7222 uclamp_update_active_tasks(css, clamps);
7227 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7228 * C expression. Since there is no way to convert a macro argument (N) into a
7229 * character constant, use two levels of macros.
7231 #define _POW10(exp) ((unsigned int)1e##exp)
7232 #define POW10(exp) _POW10(exp)
7234 struct uclamp_request {
7235 #define UCLAMP_PERCENT_SHIFT 2
7236 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7242 static inline struct uclamp_request
7243 capacity_from_percent(char *buf)
7245 struct uclamp_request req = {
7246 .percent = UCLAMP_PERCENT_SCALE,
7247 .util = SCHED_CAPACITY_SCALE,
7252 if (strcmp(buf, "max")) {
7253 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7257 if (req.percent > UCLAMP_PERCENT_SCALE) {
7262 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7263 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7269 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7270 size_t nbytes, loff_t off,
7271 enum uclamp_id clamp_id)
7273 struct uclamp_request req;
7274 struct task_group *tg;
7276 req = capacity_from_percent(buf);
7280 mutex_lock(&uclamp_mutex);
7283 tg = css_tg(of_css(of));
7284 if (tg->uclamp_req[clamp_id].value != req.util)
7285 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7288 * Because of not recoverable conversion rounding we keep track of the
7289 * exact requested value
7291 tg->uclamp_pct[clamp_id] = req.percent;
7293 /* Update effective clamps to track the most restrictive value */
7294 cpu_util_update_eff(of_css(of));
7297 mutex_unlock(&uclamp_mutex);
7302 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7303 char *buf, size_t nbytes,
7306 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7309 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7310 char *buf, size_t nbytes,
7313 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7316 static inline void cpu_uclamp_print(struct seq_file *sf,
7317 enum uclamp_id clamp_id)
7319 struct task_group *tg;
7325 tg = css_tg(seq_css(sf));
7326 util_clamp = tg->uclamp_req[clamp_id].value;
7329 if (util_clamp == SCHED_CAPACITY_SCALE) {
7330 seq_puts(sf, "max\n");
7334 percent = tg->uclamp_pct[clamp_id];
7335 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7336 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7339 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7341 cpu_uclamp_print(sf, UCLAMP_MIN);
7345 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7347 cpu_uclamp_print(sf, UCLAMP_MAX);
7350 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7352 #ifdef CONFIG_FAIR_GROUP_SCHED
7353 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7354 struct cftype *cftype, u64 shareval)
7356 if (shareval > scale_load_down(ULONG_MAX))
7357 shareval = MAX_SHARES;
7358 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7361 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7364 struct task_group *tg = css_tg(css);
7366 return (u64) scale_load_down(tg->shares);
7369 #ifdef CONFIG_CFS_BANDWIDTH
7370 static DEFINE_MUTEX(cfs_constraints_mutex);
7372 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7373 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7375 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7377 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7379 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7380 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7382 if (tg == &root_task_group)
7386 * Ensure we have at some amount of bandwidth every period. This is
7387 * to prevent reaching a state of large arrears when throttled via
7388 * entity_tick() resulting in prolonged exit starvation.
7390 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7394 * Likewise, bound things on the otherside by preventing insane quota
7395 * periods. This also allows us to normalize in computing quota
7398 if (period > max_cfs_quota_period)
7402 * Prevent race between setting of cfs_rq->runtime_enabled and
7403 * unthrottle_offline_cfs_rqs().
7406 mutex_lock(&cfs_constraints_mutex);
7407 ret = __cfs_schedulable(tg, period, quota);
7411 runtime_enabled = quota != RUNTIME_INF;
7412 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7414 * If we need to toggle cfs_bandwidth_used, off->on must occur
7415 * before making related changes, and on->off must occur afterwards
7417 if (runtime_enabled && !runtime_was_enabled)
7418 cfs_bandwidth_usage_inc();
7419 raw_spin_lock_irq(&cfs_b->lock);
7420 cfs_b->period = ns_to_ktime(period);
7421 cfs_b->quota = quota;
7423 __refill_cfs_bandwidth_runtime(cfs_b);
7425 /* Restart the period timer (if active) to handle new period expiry: */
7426 if (runtime_enabled)
7427 start_cfs_bandwidth(cfs_b);
7429 raw_spin_unlock_irq(&cfs_b->lock);
7431 for_each_online_cpu(i) {
7432 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7433 struct rq *rq = cfs_rq->rq;
7436 rq_lock_irq(rq, &rf);
7437 cfs_rq->runtime_enabled = runtime_enabled;
7438 cfs_rq->runtime_remaining = 0;
7440 if (cfs_rq->throttled)
7441 unthrottle_cfs_rq(cfs_rq);
7442 rq_unlock_irq(rq, &rf);
7444 if (runtime_was_enabled && !runtime_enabled)
7445 cfs_bandwidth_usage_dec();
7447 mutex_unlock(&cfs_constraints_mutex);
7453 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7457 period = ktime_to_ns(tg->cfs_bandwidth.period);
7458 if (cfs_quota_us < 0)
7459 quota = RUNTIME_INF;
7460 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7461 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7465 return tg_set_cfs_bandwidth(tg, period, quota);
7468 static long tg_get_cfs_quota(struct task_group *tg)
7472 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7475 quota_us = tg->cfs_bandwidth.quota;
7476 do_div(quota_us, NSEC_PER_USEC);
7481 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7485 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7488 period = (u64)cfs_period_us * NSEC_PER_USEC;
7489 quota = tg->cfs_bandwidth.quota;
7491 return tg_set_cfs_bandwidth(tg, period, quota);
7494 static long tg_get_cfs_period(struct task_group *tg)
7498 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7499 do_div(cfs_period_us, NSEC_PER_USEC);
7501 return cfs_period_us;
7504 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7507 return tg_get_cfs_quota(css_tg(css));
7510 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7511 struct cftype *cftype, s64 cfs_quota_us)
7513 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7516 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7519 return tg_get_cfs_period(css_tg(css));
7522 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7523 struct cftype *cftype, u64 cfs_period_us)
7525 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7528 struct cfs_schedulable_data {
7529 struct task_group *tg;
7534 * normalize group quota/period to be quota/max_period
7535 * note: units are usecs
7537 static u64 normalize_cfs_quota(struct task_group *tg,
7538 struct cfs_schedulable_data *d)
7546 period = tg_get_cfs_period(tg);
7547 quota = tg_get_cfs_quota(tg);
7550 /* note: these should typically be equivalent */
7551 if (quota == RUNTIME_INF || quota == -1)
7554 return to_ratio(period, quota);
7557 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7559 struct cfs_schedulable_data *d = data;
7560 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7561 s64 quota = 0, parent_quota = -1;
7564 quota = RUNTIME_INF;
7566 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7568 quota = normalize_cfs_quota(tg, d);
7569 parent_quota = parent_b->hierarchical_quota;
7572 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7573 * always take the min. On cgroup1, only inherit when no
7576 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7577 quota = min(quota, parent_quota);
7579 if (quota == RUNTIME_INF)
7580 quota = parent_quota;
7581 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7585 cfs_b->hierarchical_quota = quota;
7590 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7593 struct cfs_schedulable_data data = {
7599 if (quota != RUNTIME_INF) {
7600 do_div(data.period, NSEC_PER_USEC);
7601 do_div(data.quota, NSEC_PER_USEC);
7605 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7611 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7613 struct task_group *tg = css_tg(seq_css(sf));
7614 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7616 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7617 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7618 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7620 if (schedstat_enabled() && tg != &root_task_group) {
7624 for_each_possible_cpu(i)
7625 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7627 seq_printf(sf, "wait_sum %llu\n", ws);
7632 #endif /* CONFIG_CFS_BANDWIDTH */
7633 #endif /* CONFIG_FAIR_GROUP_SCHED */
7635 #ifdef CONFIG_RT_GROUP_SCHED
7636 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7637 struct cftype *cft, s64 val)
7639 return sched_group_set_rt_runtime(css_tg(css), val);
7642 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7645 return sched_group_rt_runtime(css_tg(css));
7648 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7649 struct cftype *cftype, u64 rt_period_us)
7651 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7654 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7657 return sched_group_rt_period(css_tg(css));
7659 #endif /* CONFIG_RT_GROUP_SCHED */
7661 static struct cftype cpu_legacy_files[] = {
7662 #ifdef CONFIG_FAIR_GROUP_SCHED
7665 .read_u64 = cpu_shares_read_u64,
7666 .write_u64 = cpu_shares_write_u64,
7669 #ifdef CONFIG_CFS_BANDWIDTH
7671 .name = "cfs_quota_us",
7672 .read_s64 = cpu_cfs_quota_read_s64,
7673 .write_s64 = cpu_cfs_quota_write_s64,
7676 .name = "cfs_period_us",
7677 .read_u64 = cpu_cfs_period_read_u64,
7678 .write_u64 = cpu_cfs_period_write_u64,
7682 .seq_show = cpu_cfs_stat_show,
7685 #ifdef CONFIG_RT_GROUP_SCHED
7687 .name = "rt_runtime_us",
7688 .read_s64 = cpu_rt_runtime_read,
7689 .write_s64 = cpu_rt_runtime_write,
7692 .name = "rt_period_us",
7693 .read_u64 = cpu_rt_period_read_uint,
7694 .write_u64 = cpu_rt_period_write_uint,
7697 #ifdef CONFIG_UCLAMP_TASK_GROUP
7699 .name = "uclamp.min",
7700 .flags = CFTYPE_NOT_ON_ROOT,
7701 .seq_show = cpu_uclamp_min_show,
7702 .write = cpu_uclamp_min_write,
7705 .name = "uclamp.max",
7706 .flags = CFTYPE_NOT_ON_ROOT,
7707 .seq_show = cpu_uclamp_max_show,
7708 .write = cpu_uclamp_max_write,
7714 static int cpu_extra_stat_show(struct seq_file *sf,
7715 struct cgroup_subsys_state *css)
7717 #ifdef CONFIG_CFS_BANDWIDTH
7719 struct task_group *tg = css_tg(css);
7720 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7723 throttled_usec = cfs_b->throttled_time;
7724 do_div(throttled_usec, NSEC_PER_USEC);
7726 seq_printf(sf, "nr_periods %d\n"
7728 "throttled_usec %llu\n",
7729 cfs_b->nr_periods, cfs_b->nr_throttled,
7736 #ifdef CONFIG_FAIR_GROUP_SCHED
7737 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7740 struct task_group *tg = css_tg(css);
7741 u64 weight = scale_load_down(tg->shares);
7743 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7746 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7747 struct cftype *cft, u64 weight)
7750 * cgroup weight knobs should use the common MIN, DFL and MAX
7751 * values which are 1, 100 and 10000 respectively. While it loses
7752 * a bit of range on both ends, it maps pretty well onto the shares
7753 * value used by scheduler and the round-trip conversions preserve
7754 * the original value over the entire range.
7756 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7759 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7761 return sched_group_set_shares(css_tg(css), scale_load(weight));
7764 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7767 unsigned long weight = scale_load_down(css_tg(css)->shares);
7768 int last_delta = INT_MAX;
7771 /* find the closest nice value to the current weight */
7772 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7773 delta = abs(sched_prio_to_weight[prio] - weight);
7774 if (delta >= last_delta)
7779 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7782 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7783 struct cftype *cft, s64 nice)
7785 unsigned long weight;
7788 if (nice < MIN_NICE || nice > MAX_NICE)
7791 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7792 idx = array_index_nospec(idx, 40);
7793 weight = sched_prio_to_weight[idx];
7795 return sched_group_set_shares(css_tg(css), scale_load(weight));
7799 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7800 long period, long quota)
7803 seq_puts(sf, "max");
7805 seq_printf(sf, "%ld", quota);
7807 seq_printf(sf, " %ld\n", period);
7810 /* caller should put the current value in *@periodp before calling */
7811 static int __maybe_unused cpu_period_quota_parse(char *buf,
7812 u64 *periodp, u64 *quotap)
7814 char tok[21]; /* U64_MAX */
7816 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7819 *periodp *= NSEC_PER_USEC;
7821 if (sscanf(tok, "%llu", quotap))
7822 *quotap *= NSEC_PER_USEC;
7823 else if (!strcmp(tok, "max"))
7824 *quotap = RUNTIME_INF;
7831 #ifdef CONFIG_CFS_BANDWIDTH
7832 static int cpu_max_show(struct seq_file *sf, void *v)
7834 struct task_group *tg = css_tg(seq_css(sf));
7836 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7840 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7841 char *buf, size_t nbytes, loff_t off)
7843 struct task_group *tg = css_tg(of_css(of));
7844 u64 period = tg_get_cfs_period(tg);
7848 ret = cpu_period_quota_parse(buf, &period, "a);
7850 ret = tg_set_cfs_bandwidth(tg, period, quota);
7851 return ret ?: nbytes;
7855 static struct cftype cpu_files[] = {
7856 #ifdef CONFIG_FAIR_GROUP_SCHED
7859 .flags = CFTYPE_NOT_ON_ROOT,
7860 .read_u64 = cpu_weight_read_u64,
7861 .write_u64 = cpu_weight_write_u64,
7864 .name = "weight.nice",
7865 .flags = CFTYPE_NOT_ON_ROOT,
7866 .read_s64 = cpu_weight_nice_read_s64,
7867 .write_s64 = cpu_weight_nice_write_s64,
7870 #ifdef CONFIG_CFS_BANDWIDTH
7873 .flags = CFTYPE_NOT_ON_ROOT,
7874 .seq_show = cpu_max_show,
7875 .write = cpu_max_write,
7878 #ifdef CONFIG_UCLAMP_TASK_GROUP
7880 .name = "uclamp.min",
7881 .flags = CFTYPE_NOT_ON_ROOT,
7882 .seq_show = cpu_uclamp_min_show,
7883 .write = cpu_uclamp_min_write,
7886 .name = "uclamp.max",
7887 .flags = CFTYPE_NOT_ON_ROOT,
7888 .seq_show = cpu_uclamp_max_show,
7889 .write = cpu_uclamp_max_write,
7895 struct cgroup_subsys cpu_cgrp_subsys = {
7896 .css_alloc = cpu_cgroup_css_alloc,
7897 .css_online = cpu_cgroup_css_online,
7898 .css_released = cpu_cgroup_css_released,
7899 .css_free = cpu_cgroup_css_free,
7900 .css_extra_stat_show = cpu_extra_stat_show,
7901 .fork = cpu_cgroup_fork,
7902 .can_attach = cpu_cgroup_can_attach,
7903 .attach = cpu_cgroup_attach,
7904 .legacy_cftypes = cpu_legacy_files,
7905 .dfl_cftypes = cpu_files,
7910 #endif /* CONFIG_CGROUP_SCHED */
7912 void dump_cpu_task(int cpu)
7914 pr_info("Task dump for CPU %d:\n", cpu);
7915 sched_show_task(cpu_curr(cpu));
7919 * Nice levels are multiplicative, with a gentle 10% change for every
7920 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7921 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7922 * that remained on nice 0.
7924 * The "10% effect" is relative and cumulative: from _any_ nice level,
7925 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7926 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7927 * If a task goes up by ~10% and another task goes down by ~10% then
7928 * the relative distance between them is ~25%.)
7930 const int sched_prio_to_weight[40] = {
7931 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7932 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7933 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7934 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7935 /* 0 */ 1024, 820, 655, 526, 423,
7936 /* 5 */ 335, 272, 215, 172, 137,
7937 /* 10 */ 110, 87, 70, 56, 45,
7938 /* 15 */ 36, 29, 23, 18, 15,
7942 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7944 * In cases where the weight does not change often, we can use the
7945 * precalculated inverse to speed up arithmetics by turning divisions
7946 * into multiplications:
7948 const u32 sched_prio_to_wmult[40] = {
7949 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7950 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7951 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7952 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7953 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7954 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7955 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7956 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7959 #undef CREATE_TRACE_POINTS