1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
11 #include <linux/nospec.h>
13 #include <linux/kcov.h>
15 #include <asm/switch_to.h>
18 #include "../workqueue_internal.h"
19 #include "../smpboot.h"
23 #define CREATE_TRACE_POINTS
24 #include <trace/events/sched.h>
27 * Export tracepoints that act as a bare tracehook (ie: have no trace event
28 * associated with them) to allow external modules to probe them.
30 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
31 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
32 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
33 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
37 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
39 #ifdef CONFIG_SCHED_DEBUG
41 * Debugging: various feature bits
43 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
44 * sysctl_sched_features, defined in sched.h, to allow constants propagation
45 * at compile time and compiler optimization based on features default.
47 #define SCHED_FEAT(name, enabled) \
48 (1UL << __SCHED_FEAT_##name) * enabled |
49 const_debug unsigned int sysctl_sched_features =
56 * Number of tasks to iterate in a single balance run.
57 * Limited because this is done with IRQs disabled.
59 const_debug unsigned int sysctl_sched_nr_migrate = 32;
62 * period over which we measure -rt task CPU usage in us.
65 unsigned int sysctl_sched_rt_period = 1000000;
67 __read_mostly int scheduler_running;
70 * part of the period that we allow rt tasks to run in us.
73 int sysctl_sched_rt_runtime = 950000;
76 * __task_rq_lock - lock the rq @p resides on.
78 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
83 lockdep_assert_held(&p->pi_lock);
87 raw_spin_lock(&rq->lock);
88 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
92 raw_spin_unlock(&rq->lock);
94 while (unlikely(task_on_rq_migrating(p)))
100 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
102 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
103 __acquires(p->pi_lock)
109 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
111 raw_spin_lock(&rq->lock);
113 * move_queued_task() task_rq_lock()
116 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
117 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
118 * [S] ->cpu = new_cpu [L] task_rq()
122 * If we observe the old CPU in task_rq_lock(), the acquire of
123 * the old rq->lock will fully serialize against the stores.
125 * If we observe the new CPU in task_rq_lock(), the address
126 * dependency headed by '[L] rq = task_rq()' and the acquire
127 * will pair with the WMB to ensure we then also see migrating.
129 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
133 raw_spin_unlock(&rq->lock);
134 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
136 while (unlikely(task_on_rq_migrating(p)))
142 * RQ-clock updating methods:
145 static void update_rq_clock_task(struct rq *rq, s64 delta)
148 * In theory, the compile should just see 0 here, and optimize out the call
149 * to sched_rt_avg_update. But I don't trust it...
151 s64 __maybe_unused steal = 0, irq_delta = 0;
153 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
154 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
157 * Since irq_time is only updated on {soft,}irq_exit, we might run into
158 * this case when a previous update_rq_clock() happened inside a
161 * When this happens, we stop ->clock_task and only update the
162 * prev_irq_time stamp to account for the part that fit, so that a next
163 * update will consume the rest. This ensures ->clock_task is
166 * It does however cause some slight miss-attribution of {soft,}irq
167 * time, a more accurate solution would be to update the irq_time using
168 * the current rq->clock timestamp, except that would require using
171 if (irq_delta > delta)
174 rq->prev_irq_time += irq_delta;
177 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
178 if (static_key_false((¶virt_steal_rq_enabled))) {
179 steal = paravirt_steal_clock(cpu_of(rq));
180 steal -= rq->prev_steal_time_rq;
182 if (unlikely(steal > delta))
185 rq->prev_steal_time_rq += steal;
190 rq->clock_task += delta;
192 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
193 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
194 update_irq_load_avg(rq, irq_delta + steal);
196 update_rq_clock_pelt(rq, delta);
199 void update_rq_clock(struct rq *rq)
203 lockdep_assert_held(&rq->lock);
205 if (rq->clock_update_flags & RQCF_ACT_SKIP)
208 #ifdef CONFIG_SCHED_DEBUG
209 if (sched_feat(WARN_DOUBLE_CLOCK))
210 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
211 rq->clock_update_flags |= RQCF_UPDATED;
214 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
218 update_rq_clock_task(rq, delta);
222 #ifdef CONFIG_SCHED_HRTICK
224 * Use HR-timers to deliver accurate preemption points.
227 static void hrtick_clear(struct rq *rq)
229 if (hrtimer_active(&rq->hrtick_timer))
230 hrtimer_cancel(&rq->hrtick_timer);
234 * High-resolution timer tick.
235 * Runs from hardirq context with interrupts disabled.
237 static enum hrtimer_restart hrtick(struct hrtimer *timer)
239 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
242 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
246 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
249 return HRTIMER_NORESTART;
254 static void __hrtick_restart(struct rq *rq)
256 struct hrtimer *timer = &rq->hrtick_timer;
257 ktime_t time = rq->hrtick_time;
259 hrtimer_start(timer, time, 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;
287 * Don't schedule slices shorter than 10000ns, that just
288 * doesn't make sense and can cause timer DoS.
290 delta = max_t(s64, delay, 10000LL);
291 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
293 if (rq == this_rq()) {
294 __hrtick_restart(rq);
295 } else if (!rq->hrtick_csd_pending) {
296 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
297 rq->hrtick_csd_pending = 1;
303 * Called to set the hrtick timer state.
305 * called with rq->lock held and irqs disabled
307 void hrtick_start(struct rq *rq, u64 delay)
310 * Don't schedule slices shorter than 10000ns, that just
311 * doesn't make sense. Rely on vruntime for fairness.
313 delay = max_t(u64, delay, 10000LL);
314 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
315 HRTIMER_MODE_REL_PINNED_HARD);
317 #endif /* CONFIG_SMP */
319 static void hrtick_rq_init(struct rq *rq)
322 rq->hrtick_csd_pending = 0;
324 rq->hrtick_csd.flags = 0;
325 rq->hrtick_csd.func = __hrtick_start;
326 rq->hrtick_csd.info = rq;
329 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
330 rq->hrtick_timer.function = hrtick;
332 #else /* CONFIG_SCHED_HRTICK */
333 static inline void hrtick_clear(struct rq *rq)
337 static inline void hrtick_rq_init(struct rq *rq)
340 #endif /* CONFIG_SCHED_HRTICK */
343 * cmpxchg based fetch_or, macro so it works for different integer types
345 #define fetch_or(ptr, mask) \
347 typeof(ptr) _ptr = (ptr); \
348 typeof(mask) _mask = (mask); \
349 typeof(*_ptr) _old, _val = *_ptr; \
352 _old = cmpxchg(_ptr, _val, _val | _mask); \
360 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
362 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
363 * this avoids any races wrt polling state changes and thereby avoids
366 static bool set_nr_and_not_polling(struct task_struct *p)
368 struct thread_info *ti = task_thread_info(p);
369 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
373 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
375 * If this returns true, then the idle task promises to call
376 * sched_ttwu_pending() and reschedule soon.
378 static bool set_nr_if_polling(struct task_struct *p)
380 struct thread_info *ti = task_thread_info(p);
381 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
384 if (!(val & _TIF_POLLING_NRFLAG))
386 if (val & _TIF_NEED_RESCHED)
388 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
397 static bool set_nr_and_not_polling(struct task_struct *p)
399 set_tsk_need_resched(p);
404 static bool set_nr_if_polling(struct task_struct *p)
411 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
413 struct wake_q_node *node = &task->wake_q;
416 * Atomically grab the task, if ->wake_q is !nil already it means
417 * its already queued (either by us or someone else) and will get the
418 * wakeup due to that.
420 * In order to ensure that a pending wakeup will observe our pending
421 * state, even in the failed case, an explicit smp_mb() must be used.
423 smp_mb__before_atomic();
424 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
428 * The head is context local, there can be no concurrency.
431 head->lastp = &node->next;
436 * wake_q_add() - queue a wakeup for 'later' waking.
437 * @head: the wake_q_head to add @task to
438 * @task: the task to queue for 'later' wakeup
440 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
441 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
444 * This function must be used as-if it were wake_up_process(); IOW the task
445 * must be ready to be woken at this location.
447 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
449 if (__wake_q_add(head, task))
450 get_task_struct(task);
454 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
455 * @head: the wake_q_head to add @task to
456 * @task: the task to queue for 'later' wakeup
458 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
459 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
462 * This function must be used as-if it were wake_up_process(); IOW the task
463 * must be ready to be woken at this location.
465 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
466 * that already hold reference to @task can call the 'safe' version and trust
467 * wake_q to do the right thing depending whether or not the @task is already
470 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
472 if (!__wake_q_add(head, task))
473 put_task_struct(task);
476 void wake_up_q(struct wake_q_head *head)
478 struct wake_q_node *node = head->first;
480 while (node != WAKE_Q_TAIL) {
481 struct task_struct *task;
483 task = container_of(node, struct task_struct, wake_q);
485 /* Task can safely be re-inserted now: */
487 task->wake_q.next = NULL;
490 * wake_up_process() executes a full barrier, which pairs with
491 * the queueing in wake_q_add() so as not to miss wakeups.
493 wake_up_process(task);
494 put_task_struct(task);
499 * resched_curr - mark rq's current task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
505 void resched_curr(struct rq *rq)
507 struct task_struct *curr = rq->curr;
510 lockdep_assert_held(&rq->lock);
512 if (test_tsk_need_resched(curr))
517 if (cpu == smp_processor_id()) {
518 set_tsk_need_resched(curr);
519 set_preempt_need_resched();
523 if (set_nr_and_not_polling(curr))
524 smp_send_reschedule(cpu);
526 trace_sched_wake_idle_without_ipi(cpu);
529 void resched_cpu(int cpu)
531 struct rq *rq = cpu_rq(cpu);
534 raw_spin_lock_irqsave(&rq->lock, flags);
535 if (cpu_online(cpu) || cpu == smp_processor_id())
537 raw_spin_unlock_irqrestore(&rq->lock, flags);
541 #ifdef CONFIG_NO_HZ_COMMON
543 * In the semi idle case, use the nearest busy CPU for migrating timers
544 * from an idle CPU. This is good for power-savings.
546 * We don't do similar optimization for completely idle system, as
547 * selecting an idle CPU will add more delays to the timers than intended
548 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
550 int get_nohz_timer_target(void)
552 int i, cpu = smp_processor_id();
553 struct sched_domain *sd;
555 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
564 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
571 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
572 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
579 * When add_timer_on() enqueues a timer into the timer wheel of an
580 * idle CPU then this timer might expire before the next timer event
581 * which is scheduled to wake up that CPU. In case of a completely
582 * idle system the next event might even be infinite time into the
583 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
584 * leaves the inner idle loop so the newly added timer is taken into
585 * account when the CPU goes back to idle and evaluates the timer
586 * wheel for the next timer event.
588 static void wake_up_idle_cpu(int cpu)
590 struct rq *rq = cpu_rq(cpu);
592 if (cpu == smp_processor_id())
595 if (set_nr_and_not_polling(rq->idle))
596 smp_send_reschedule(cpu);
598 trace_sched_wake_idle_without_ipi(cpu);
601 static bool wake_up_full_nohz_cpu(int cpu)
604 * We just need the target to call irq_exit() and re-evaluate
605 * the next tick. The nohz full kick at least implies that.
606 * If needed we can still optimize that later with an
609 if (cpu_is_offline(cpu))
610 return true; /* Don't try to wake offline CPUs. */
611 if (tick_nohz_full_cpu(cpu)) {
612 if (cpu != smp_processor_id() ||
613 tick_nohz_tick_stopped())
614 tick_nohz_full_kick_cpu(cpu);
622 * Wake up the specified CPU. If the CPU is going offline, it is the
623 * caller's responsibility to deal with the lost wakeup, for example,
624 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
626 void wake_up_nohz_cpu(int cpu)
628 if (!wake_up_full_nohz_cpu(cpu))
629 wake_up_idle_cpu(cpu);
632 static inline bool got_nohz_idle_kick(void)
634 int cpu = smp_processor_id();
636 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
639 if (idle_cpu(cpu) && !need_resched())
643 * We can't run Idle Load Balance on this CPU for this time so we
644 * cancel it and clear NOHZ_BALANCE_KICK
646 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
650 #else /* CONFIG_NO_HZ_COMMON */
652 static inline bool got_nohz_idle_kick(void)
657 #endif /* CONFIG_NO_HZ_COMMON */
659 #ifdef CONFIG_NO_HZ_FULL
660 bool sched_can_stop_tick(struct rq *rq)
664 /* Deadline tasks, even if single, need the tick */
665 if (rq->dl.dl_nr_running)
669 * If there are more than one RR tasks, we need the tick to effect the
670 * actual RR behaviour.
672 if (rq->rt.rr_nr_running) {
673 if (rq->rt.rr_nr_running == 1)
680 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
681 * forced preemption between FIFO tasks.
683 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
688 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
689 * if there's more than one we need the tick for involuntary
692 if (rq->nr_running > 1)
697 #endif /* CONFIG_NO_HZ_FULL */
698 #endif /* CONFIG_SMP */
700 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
703 * Iterate task_group tree rooted at *from, calling @down when first entering a
704 * node and @up when leaving it for the final time.
706 * Caller must hold rcu_lock or sufficient equivalent.
708 int walk_tg_tree_from(struct task_group *from,
709 tg_visitor down, tg_visitor up, void *data)
711 struct task_group *parent, *child;
717 ret = (*down)(parent, data);
720 list_for_each_entry_rcu(child, &parent->children, siblings) {
727 ret = (*up)(parent, data);
728 if (ret || parent == from)
732 parent = parent->parent;
739 int tg_nop(struct task_group *tg, void *data)
745 static void set_load_weight(struct task_struct *p, bool update_load)
747 int prio = p->static_prio - MAX_RT_PRIO;
748 struct load_weight *load = &p->se.load;
751 * SCHED_IDLE tasks get minimal weight:
753 if (task_has_idle_policy(p)) {
754 load->weight = scale_load(WEIGHT_IDLEPRIO);
755 load->inv_weight = WMULT_IDLEPRIO;
756 p->se.runnable_weight = load->weight;
761 * SCHED_OTHER tasks have to update their load when changing their
764 if (update_load && p->sched_class == &fair_sched_class) {
765 reweight_task(p, prio);
767 load->weight = scale_load(sched_prio_to_weight[prio]);
768 load->inv_weight = sched_prio_to_wmult[prio];
769 p->se.runnable_weight = load->weight;
773 #ifdef CONFIG_UCLAMP_TASK
775 * Serializes updates of utilization clamp values
777 * The (slow-path) user-space triggers utilization clamp value updates which
778 * can require updates on (fast-path) scheduler's data structures used to
779 * support enqueue/dequeue operations.
780 * While the per-CPU rq lock protects fast-path update operations, user-space
781 * requests are serialized using a mutex to reduce the risk of conflicting
782 * updates or API abuses.
784 static DEFINE_MUTEX(uclamp_mutex);
786 /* Max allowed minimum utilization */
787 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
789 /* Max allowed maximum utilization */
790 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
792 /* All clamps are required to be less or equal than these values */
793 static struct uclamp_se uclamp_default[UCLAMP_CNT];
796 * This static key is used to reduce the uclamp overhead in the fast path. It
797 * primarily disables the call to uclamp_rq_{inc, dec}() in
798 * enqueue/dequeue_task().
800 * This allows users to continue to enable uclamp in their kernel config with
801 * minimum uclamp overhead in the fast path.
803 * As soon as userspace modifies any of the uclamp knobs, the static key is
804 * enabled, since we have an actual users that make use of uclamp
807 * The knobs that would enable this static key are:
809 * * A task modifying its uclamp value with sched_setattr().
810 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
811 * * An admin modifying the cgroup cpu.uclamp.{min, max}
813 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
815 /* Integer rounded range for each bucket */
816 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
818 #define for_each_clamp_id(clamp_id) \
819 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
821 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
823 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
826 static inline unsigned int uclamp_bucket_base_value(unsigned int clamp_value)
828 return UCLAMP_BUCKET_DELTA * uclamp_bucket_id(clamp_value);
831 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
833 if (clamp_id == UCLAMP_MIN)
835 return SCHED_CAPACITY_SCALE;
838 static inline void uclamp_se_set(struct uclamp_se *uc_se,
839 unsigned int value, bool user_defined)
841 uc_se->value = value;
842 uc_se->bucket_id = uclamp_bucket_id(value);
843 uc_se->user_defined = user_defined;
846 static inline unsigned int
847 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
848 unsigned int clamp_value)
851 * Avoid blocked utilization pushing up the frequency when we go
852 * idle (which drops the max-clamp) by retaining the last known
855 if (clamp_id == UCLAMP_MAX) {
856 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
860 return uclamp_none(UCLAMP_MIN);
863 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
864 unsigned int clamp_value)
866 /* Reset max-clamp retention only on idle exit */
867 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
870 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
874 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
875 unsigned int clamp_value)
877 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
878 int bucket_id = UCLAMP_BUCKETS - 1;
881 * Since both min and max clamps are max aggregated, find the
882 * top most bucket with tasks in.
884 for ( ; bucket_id >= 0; bucket_id--) {
885 if (!bucket[bucket_id].tasks)
887 return bucket[bucket_id].value;
890 /* No tasks -- default clamp values */
891 return uclamp_idle_value(rq, clamp_id, clamp_value);
894 static inline struct uclamp_se
895 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
897 /* Copy by value as we could modify it */
898 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
899 #ifdef CONFIG_UCLAMP_TASK_GROUP
900 unsigned int tg_min, tg_max, value;
903 * Tasks in autogroups or root task group will be
904 * restricted by system defaults.
906 if (task_group_is_autogroup(task_group(p)))
908 if (task_group(p) == &root_task_group)
911 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
912 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
913 value = uc_req.value;
914 value = clamp(value, tg_min, tg_max);
915 uclamp_se_set(&uc_req, value, false);
922 * The effective clamp bucket index of a task depends on, by increasing
924 * - the task specific clamp value, when explicitly requested from userspace
925 * - the task group effective clamp value, for tasks not either in the root
926 * group or in an autogroup
927 * - the system default clamp value, defined by the sysadmin
929 static inline struct uclamp_se
930 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
932 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
933 struct uclamp_se uc_max = uclamp_default[clamp_id];
935 /* System default restrictions always apply */
936 if (unlikely(uc_req.value > uc_max.value))
942 unsigned int uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
944 struct uclamp_se uc_eff;
946 /* Task currently refcounted: use back-annotated (effective) value */
947 if (p->uclamp[clamp_id].active)
948 return p->uclamp[clamp_id].value;
950 uc_eff = uclamp_eff_get(p, clamp_id);
956 * When a task is enqueued on a rq, the clamp bucket currently defined by the
957 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
958 * updates the rq's clamp value if required.
960 * Tasks can have a task-specific value requested from user-space, track
961 * within each bucket the maximum value for tasks refcounted in it.
962 * This "local max aggregation" allows to track the exact "requested" value
963 * for each bucket when all its RUNNABLE tasks require the same clamp.
965 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
966 enum uclamp_id clamp_id)
968 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
969 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
970 struct uclamp_bucket *bucket;
972 lockdep_assert_held(&rq->lock);
974 /* Update task effective clamp */
975 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
977 bucket = &uc_rq->bucket[uc_se->bucket_id];
979 uc_se->active = true;
981 uclamp_idle_reset(rq, clamp_id, uc_se->value);
984 * Local max aggregation: rq buckets always track the max
985 * "requested" clamp value of its RUNNABLE tasks.
987 if (bucket->tasks == 1 || uc_se->value > bucket->value)
988 bucket->value = uc_se->value;
990 if (uc_se->value > READ_ONCE(uc_rq->value))
991 WRITE_ONCE(uc_rq->value, uc_se->value);
995 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
996 * is released. If this is the last task reference counting the rq's max
997 * active clamp value, then the rq's clamp value is updated.
999 * Both refcounted tasks and rq's cached clamp values are expected to be
1000 * always valid. If it's detected they are not, as defensive programming,
1001 * enforce the expected state and warn.
1003 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1004 enum uclamp_id clamp_id)
1006 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1007 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1008 struct uclamp_bucket *bucket;
1009 unsigned int bkt_clamp;
1010 unsigned int rq_clamp;
1012 lockdep_assert_held(&rq->lock);
1015 * If sched_uclamp_used was enabled after task @p was enqueued,
1016 * we could end up with unbalanced call to uclamp_rq_dec_id().
1018 * In this case the uc_se->active flag should be false since no uclamp
1019 * accounting was performed at enqueue time and we can just return
1022 * Need to be careful of the following enqeueue/dequeue ordering
1026 * // sched_uclamp_used gets enabled
1029 * // Must not decrement bukcet->tasks here
1032 * where we could end up with stale data in uc_se and
1033 * bucket[uc_se->bucket_id].
1035 * The following check here eliminates the possibility of such race.
1037 if (unlikely(!uc_se->active))
1040 bucket = &uc_rq->bucket[uc_se->bucket_id];
1042 SCHED_WARN_ON(!bucket->tasks);
1043 if (likely(bucket->tasks))
1046 uc_se->active = false;
1049 * Keep "local max aggregation" simple and accept to (possibly)
1050 * overboost some RUNNABLE tasks in the same bucket.
1051 * The rq clamp bucket value is reset to its base value whenever
1052 * there are no more RUNNABLE tasks refcounting it.
1054 if (likely(bucket->tasks))
1057 rq_clamp = READ_ONCE(uc_rq->value);
1059 * Defensive programming: this should never happen. If it happens,
1060 * e.g. due to future modification, warn and fixup the expected value.
1062 SCHED_WARN_ON(bucket->value > rq_clamp);
1063 if (bucket->value >= rq_clamp) {
1064 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1065 WRITE_ONCE(uc_rq->value, bkt_clamp);
1069 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1071 enum uclamp_id clamp_id;
1074 * Avoid any overhead until uclamp is actually used by the userspace.
1076 * The condition is constructed such that a NOP is generated when
1077 * sched_uclamp_used is disabled.
1079 if (!static_branch_unlikely(&sched_uclamp_used))
1082 if (unlikely(!p->sched_class->uclamp_enabled))
1085 for_each_clamp_id(clamp_id)
1086 uclamp_rq_inc_id(rq, p, clamp_id);
1088 /* Reset clamp idle holding when there is one RUNNABLE task */
1089 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1090 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1093 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1095 enum uclamp_id clamp_id;
1098 * Avoid any overhead until uclamp is actually used by the userspace.
1100 * The condition is constructed such that a NOP is generated when
1101 * sched_uclamp_used is disabled.
1103 if (!static_branch_unlikely(&sched_uclamp_used))
1106 if (unlikely(!p->sched_class->uclamp_enabled))
1109 for_each_clamp_id(clamp_id)
1110 uclamp_rq_dec_id(rq, p, clamp_id);
1113 static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1114 enum uclamp_id clamp_id)
1116 if (!p->uclamp[clamp_id].active)
1119 uclamp_rq_dec_id(rq, p, clamp_id);
1120 uclamp_rq_inc_id(rq, p, clamp_id);
1123 * Make sure to clear the idle flag if we've transiently reached 0
1124 * active tasks on rq.
1126 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1127 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1131 uclamp_update_active(struct task_struct *p)
1133 enum uclamp_id clamp_id;
1138 * Lock the task and the rq where the task is (or was) queued.
1140 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1141 * price to pay to safely serialize util_{min,max} updates with
1142 * enqueues, dequeues and migration operations.
1143 * This is the same locking schema used by __set_cpus_allowed_ptr().
1145 rq = task_rq_lock(p, &rf);
1148 * Setting the clamp bucket is serialized by task_rq_lock().
1149 * If the task is not yet RUNNABLE and its task_struct is not
1150 * affecting a valid clamp bucket, the next time it's enqueued,
1151 * it will already see the updated clamp bucket value.
1153 for_each_clamp_id(clamp_id)
1154 uclamp_rq_reinc_id(rq, p, clamp_id);
1156 task_rq_unlock(rq, p, &rf);
1159 #ifdef CONFIG_UCLAMP_TASK_GROUP
1161 uclamp_update_active_tasks(struct cgroup_subsys_state *css)
1163 struct css_task_iter it;
1164 struct task_struct *p;
1166 css_task_iter_start(css, 0, &it);
1167 while ((p = css_task_iter_next(&it)))
1168 uclamp_update_active(p);
1169 css_task_iter_end(&it);
1172 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1173 static void uclamp_update_root_tg(void)
1175 struct task_group *tg = &root_task_group;
1177 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1178 sysctl_sched_uclamp_util_min, false);
1179 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1180 sysctl_sched_uclamp_util_max, false);
1183 cpu_util_update_eff(&root_task_group.css);
1187 static void uclamp_update_root_tg(void) { }
1190 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1191 void __user *buffer, size_t *lenp,
1194 bool update_root_tg = false;
1195 int old_min, old_max;
1198 mutex_lock(&uclamp_mutex);
1199 old_min = sysctl_sched_uclamp_util_min;
1200 old_max = sysctl_sched_uclamp_util_max;
1202 result = proc_dointvec(table, write, buffer, lenp, ppos);
1208 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1209 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE) {
1214 if (old_min != sysctl_sched_uclamp_util_min) {
1215 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1216 sysctl_sched_uclamp_util_min, false);
1217 update_root_tg = true;
1219 if (old_max != sysctl_sched_uclamp_util_max) {
1220 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1221 sysctl_sched_uclamp_util_max, false);
1222 update_root_tg = true;
1225 if (update_root_tg) {
1226 static_branch_enable(&sched_uclamp_used);
1227 uclamp_update_root_tg();
1231 * We update all RUNNABLE tasks only when task groups are in use.
1232 * Otherwise, keep it simple and do just a lazy update at each next
1233 * task enqueue time.
1239 sysctl_sched_uclamp_util_min = old_min;
1240 sysctl_sched_uclamp_util_max = old_max;
1242 mutex_unlock(&uclamp_mutex);
1247 static int uclamp_validate(struct task_struct *p,
1248 const struct sched_attr *attr)
1250 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1251 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1253 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1254 lower_bound = attr->sched_util_min;
1255 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1256 upper_bound = attr->sched_util_max;
1258 if (lower_bound > upper_bound)
1260 if (upper_bound > SCHED_CAPACITY_SCALE)
1264 * We have valid uclamp attributes; make sure uclamp is enabled.
1266 * We need to do that here, because enabling static branches is a
1267 * blocking operation which obviously cannot be done while holding
1270 static_branch_enable(&sched_uclamp_used);
1275 static void __setscheduler_uclamp(struct task_struct *p,
1276 const struct sched_attr *attr)
1278 enum uclamp_id clamp_id;
1281 * On scheduling class change, reset to default clamps for tasks
1282 * without a task-specific value.
1284 for_each_clamp_id(clamp_id) {
1285 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1286 unsigned int clamp_value = uclamp_none(clamp_id);
1288 /* Keep using defined clamps across class changes */
1289 if (uc_se->user_defined)
1292 /* By default, RT tasks always get 100% boost */
1293 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1294 clamp_value = uclamp_none(UCLAMP_MAX);
1296 uclamp_se_set(uc_se, clamp_value, false);
1299 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1302 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1303 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1304 attr->sched_util_min, true);
1307 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1308 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1309 attr->sched_util_max, true);
1313 static void uclamp_fork(struct task_struct *p)
1315 enum uclamp_id clamp_id;
1317 for_each_clamp_id(clamp_id)
1318 p->uclamp[clamp_id].active = false;
1320 if (likely(!p->sched_reset_on_fork))
1323 for_each_clamp_id(clamp_id) {
1324 uclamp_se_set(&p->uclamp_req[clamp_id],
1325 uclamp_none(clamp_id), false);
1329 static void __init init_uclamp_rq(struct rq *rq)
1331 enum uclamp_id clamp_id;
1332 struct uclamp_rq *uc_rq = rq->uclamp;
1334 for_each_clamp_id(clamp_id) {
1335 uc_rq[clamp_id] = (struct uclamp_rq) {
1336 .value = uclamp_none(clamp_id)
1340 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
1343 static void __init init_uclamp(void)
1345 struct uclamp_se uc_max = {};
1346 enum uclamp_id clamp_id;
1349 mutex_init(&uclamp_mutex);
1351 for_each_possible_cpu(cpu)
1352 init_uclamp_rq(cpu_rq(cpu));
1354 for_each_clamp_id(clamp_id) {
1355 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1356 uclamp_none(clamp_id), false);
1359 /* System defaults allow max clamp values for both indexes */
1360 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1361 for_each_clamp_id(clamp_id) {
1362 uclamp_default[clamp_id] = uc_max;
1363 #ifdef CONFIG_UCLAMP_TASK_GROUP
1364 root_task_group.uclamp_req[clamp_id] = uc_max;
1365 root_task_group.uclamp[clamp_id] = uc_max;
1370 #else /* CONFIG_UCLAMP_TASK */
1371 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1372 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1373 static inline int uclamp_validate(struct task_struct *p,
1374 const struct sched_attr *attr)
1378 static void __setscheduler_uclamp(struct task_struct *p,
1379 const struct sched_attr *attr) { }
1380 static inline void uclamp_fork(struct task_struct *p) { }
1381 static inline void init_uclamp(void) { }
1382 #endif /* CONFIG_UCLAMP_TASK */
1384 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1386 if (!(flags & ENQUEUE_NOCLOCK))
1387 update_rq_clock(rq);
1389 if (!(flags & ENQUEUE_RESTORE)) {
1390 sched_info_queued(rq, p);
1391 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1394 uclamp_rq_inc(rq, p);
1395 p->sched_class->enqueue_task(rq, p, flags);
1398 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1400 if (!(flags & DEQUEUE_NOCLOCK))
1401 update_rq_clock(rq);
1403 if (!(flags & DEQUEUE_SAVE)) {
1404 sched_info_dequeued(rq, p);
1405 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1408 uclamp_rq_dec(rq, p);
1409 p->sched_class->dequeue_task(rq, p, flags);
1412 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1414 if (task_contributes_to_load(p))
1415 rq->nr_uninterruptible--;
1417 enqueue_task(rq, p, flags);
1419 p->on_rq = TASK_ON_RQ_QUEUED;
1422 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1424 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1426 if (task_contributes_to_load(p))
1427 rq->nr_uninterruptible++;
1429 dequeue_task(rq, p, flags);
1433 * __normal_prio - return the priority that is based on the static prio
1435 static inline int __normal_prio(struct task_struct *p)
1437 return p->static_prio;
1441 * Calculate the expected normal priority: i.e. priority
1442 * without taking RT-inheritance into account. Might be
1443 * boosted by interactivity modifiers. Changes upon fork,
1444 * setprio syscalls, and whenever the interactivity
1445 * estimator recalculates.
1447 static inline int normal_prio(struct task_struct *p)
1451 if (task_has_dl_policy(p))
1452 prio = MAX_DL_PRIO-1;
1453 else if (task_has_rt_policy(p))
1454 prio = MAX_RT_PRIO-1 - p->rt_priority;
1456 prio = __normal_prio(p);
1461 * Calculate the current priority, i.e. the priority
1462 * taken into account by the scheduler. This value might
1463 * be boosted by RT tasks, or might be boosted by
1464 * interactivity modifiers. Will be RT if the task got
1465 * RT-boosted. If not then it returns p->normal_prio.
1467 static int effective_prio(struct task_struct *p)
1469 p->normal_prio = normal_prio(p);
1471 * If we are RT tasks or we were boosted to RT priority,
1472 * keep the priority unchanged. Otherwise, update priority
1473 * to the normal priority:
1475 if (!rt_prio(p->prio))
1476 return p->normal_prio;
1481 * task_curr - is this task currently executing on a CPU?
1482 * @p: the task in question.
1484 * Return: 1 if the task is currently executing. 0 otherwise.
1486 inline int task_curr(const struct task_struct *p)
1488 return cpu_curr(task_cpu(p)) == p;
1492 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1493 * use the balance_callback list if you want balancing.
1495 * this means any call to check_class_changed() must be followed by a call to
1496 * balance_callback().
1498 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1499 const struct sched_class *prev_class,
1502 if (prev_class != p->sched_class) {
1503 if (prev_class->switched_from)
1504 prev_class->switched_from(rq, p);
1506 p->sched_class->switched_to(rq, p);
1507 } else if (oldprio != p->prio || dl_task(p))
1508 p->sched_class->prio_changed(rq, p, oldprio);
1511 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1513 const struct sched_class *class;
1515 if (p->sched_class == rq->curr->sched_class) {
1516 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1518 for_each_class(class) {
1519 if (class == rq->curr->sched_class)
1521 if (class == p->sched_class) {
1529 * A queue event has occurred, and we're going to schedule. In
1530 * this case, we can save a useless back to back clock update.
1532 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1533 rq_clock_skip_update(rq);
1538 static inline bool is_per_cpu_kthread(struct task_struct *p)
1540 if (!(p->flags & PF_KTHREAD))
1543 if (p->nr_cpus_allowed != 1)
1550 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1551 * __set_cpus_allowed_ptr() and select_fallback_rq().
1553 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1555 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1558 if (is_per_cpu_kthread(p))
1559 return cpu_online(cpu);
1561 return cpu_active(cpu);
1565 * This is how migration works:
1567 * 1) we invoke migration_cpu_stop() on the target CPU using
1569 * 2) stopper starts to run (implicitly forcing the migrated thread
1571 * 3) it checks whether the migrated task is still in the wrong runqueue.
1572 * 4) if it's in the wrong runqueue then the migration thread removes
1573 * it and puts it into the right queue.
1574 * 5) stopper completes and stop_one_cpu() returns and the migration
1579 * move_queued_task - move a queued task to new rq.
1581 * Returns (locked) new rq. Old rq's lock is released.
1583 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1584 struct task_struct *p, int new_cpu)
1586 lockdep_assert_held(&rq->lock);
1588 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
1589 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
1590 set_task_cpu(p, new_cpu);
1593 rq = cpu_rq(new_cpu);
1596 BUG_ON(task_cpu(p) != new_cpu);
1597 enqueue_task(rq, p, 0);
1598 p->on_rq = TASK_ON_RQ_QUEUED;
1599 check_preempt_curr(rq, p, 0);
1604 struct migration_arg {
1605 struct task_struct *task;
1610 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1611 * this because either it can't run here any more (set_cpus_allowed()
1612 * away from this CPU, or CPU going down), or because we're
1613 * attempting to rebalance this task on exec (sched_exec).
1615 * So we race with normal scheduler movements, but that's OK, as long
1616 * as the task is no longer on this CPU.
1618 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1619 struct task_struct *p, int dest_cpu)
1621 /* Affinity changed (again). */
1622 if (!is_cpu_allowed(p, dest_cpu))
1625 update_rq_clock(rq);
1626 rq = move_queued_task(rq, rf, p, dest_cpu);
1632 * migration_cpu_stop - this will be executed by a highprio stopper thread
1633 * and performs thread migration by bumping thread off CPU then
1634 * 'pushing' onto another runqueue.
1636 static int migration_cpu_stop(void *data)
1638 struct migration_arg *arg = data;
1639 struct task_struct *p = arg->task;
1640 struct rq *rq = this_rq();
1644 * The original target CPU might have gone down and we might
1645 * be on another CPU but it doesn't matter.
1647 local_irq_disable();
1649 * We need to explicitly wake pending tasks before running
1650 * __migrate_task() such that we will not miss enforcing cpus_ptr
1651 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1653 sched_ttwu_pending();
1655 raw_spin_lock(&p->pi_lock);
1658 * If task_rq(p) != rq, it cannot be migrated here, because we're
1659 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1660 * we're holding p->pi_lock.
1662 if (task_rq(p) == rq) {
1663 if (task_on_rq_queued(p))
1664 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1666 p->wake_cpu = arg->dest_cpu;
1669 raw_spin_unlock(&p->pi_lock);
1676 * sched_class::set_cpus_allowed must do the below, but is not required to
1677 * actually call this function.
1679 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1681 cpumask_copy(&p->cpus_mask, new_mask);
1682 p->nr_cpus_allowed = cpumask_weight(new_mask);
1685 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1687 struct rq *rq = task_rq(p);
1688 bool queued, running;
1690 lockdep_assert_held(&p->pi_lock);
1692 queued = task_on_rq_queued(p);
1693 running = task_current(rq, p);
1697 * Because __kthread_bind() calls this on blocked tasks without
1700 lockdep_assert_held(&rq->lock);
1701 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1704 put_prev_task(rq, p);
1706 p->sched_class->set_cpus_allowed(p, new_mask);
1709 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1711 set_next_task(rq, p);
1715 * Change a given task's CPU affinity. Migrate the thread to a
1716 * proper CPU and schedule it away if the CPU it's executing on
1717 * is removed from the allowed bitmask.
1719 * NOTE: the caller must have a valid reference to the task, the
1720 * task must not exit() & deallocate itself prematurely. The
1721 * call is not atomic; no spinlocks may be held.
1723 static int __set_cpus_allowed_ptr(struct task_struct *p,
1724 const struct cpumask *new_mask, bool check)
1726 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1727 unsigned int dest_cpu;
1732 rq = task_rq_lock(p, &rf);
1733 update_rq_clock(rq);
1735 if (p->flags & PF_KTHREAD) {
1737 * Kernel threads are allowed on online && !active CPUs
1739 cpu_valid_mask = cpu_online_mask;
1743 * Must re-check here, to close a race against __kthread_bind(),
1744 * sched_setaffinity() is not guaranteed to observe the flag.
1746 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1751 if (cpumask_equal(&p->cpus_mask, new_mask))
1754 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1755 if (dest_cpu >= nr_cpu_ids) {
1760 do_set_cpus_allowed(p, new_mask);
1762 if (p->flags & PF_KTHREAD) {
1764 * For kernel threads that do indeed end up on online &&
1765 * !active we want to ensure they are strict per-CPU threads.
1767 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1768 !cpumask_intersects(new_mask, cpu_active_mask) &&
1769 p->nr_cpus_allowed != 1);
1772 /* Can the task run on the task's current CPU? If so, we're done */
1773 if (cpumask_test_cpu(task_cpu(p), new_mask))
1776 if (task_running(rq, p) || p->state == TASK_WAKING) {
1777 struct migration_arg arg = { p, dest_cpu };
1778 /* Need help from migration thread: drop lock and wait. */
1779 task_rq_unlock(rq, p, &rf);
1780 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1782 } else if (task_on_rq_queued(p)) {
1784 * OK, since we're going to drop the lock immediately
1785 * afterwards anyway.
1787 rq = move_queued_task(rq, &rf, p, dest_cpu);
1790 task_rq_unlock(rq, p, &rf);
1795 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1797 return __set_cpus_allowed_ptr(p, new_mask, false);
1799 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1801 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1803 #ifdef CONFIG_SCHED_DEBUG
1805 * We should never call set_task_cpu() on a blocked task,
1806 * ttwu() will sort out the placement.
1808 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1812 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1813 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1814 * time relying on p->on_rq.
1816 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1817 p->sched_class == &fair_sched_class &&
1818 (p->on_rq && !task_on_rq_migrating(p)));
1820 #ifdef CONFIG_LOCKDEP
1822 * The caller should hold either p->pi_lock or rq->lock, when changing
1823 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1825 * sched_move_task() holds both and thus holding either pins the cgroup,
1828 * Furthermore, all task_rq users should acquire both locks, see
1831 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1832 lockdep_is_held(&task_rq(p)->lock)));
1835 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1837 WARN_ON_ONCE(!cpu_online(new_cpu));
1840 trace_sched_migrate_task(p, new_cpu);
1842 if (task_cpu(p) != new_cpu) {
1843 if (p->sched_class->migrate_task_rq)
1844 p->sched_class->migrate_task_rq(p, new_cpu);
1845 p->se.nr_migrations++;
1847 perf_event_task_migrate(p);
1850 __set_task_cpu(p, new_cpu);
1853 #ifdef CONFIG_NUMA_BALANCING
1854 static void __migrate_swap_task(struct task_struct *p, int cpu)
1856 if (task_on_rq_queued(p)) {
1857 struct rq *src_rq, *dst_rq;
1858 struct rq_flags srf, drf;
1860 src_rq = task_rq(p);
1861 dst_rq = cpu_rq(cpu);
1863 rq_pin_lock(src_rq, &srf);
1864 rq_pin_lock(dst_rq, &drf);
1866 deactivate_task(src_rq, p, 0);
1867 set_task_cpu(p, cpu);
1868 activate_task(dst_rq, p, 0);
1869 check_preempt_curr(dst_rq, p, 0);
1871 rq_unpin_lock(dst_rq, &drf);
1872 rq_unpin_lock(src_rq, &srf);
1876 * Task isn't running anymore; make it appear like we migrated
1877 * it before it went to sleep. This means on wakeup we make the
1878 * previous CPU our target instead of where it really is.
1884 struct migration_swap_arg {
1885 struct task_struct *src_task, *dst_task;
1886 int src_cpu, dst_cpu;
1889 static int migrate_swap_stop(void *data)
1891 struct migration_swap_arg *arg = data;
1892 struct rq *src_rq, *dst_rq;
1895 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1898 src_rq = cpu_rq(arg->src_cpu);
1899 dst_rq = cpu_rq(arg->dst_cpu);
1901 double_raw_lock(&arg->src_task->pi_lock,
1902 &arg->dst_task->pi_lock);
1903 double_rq_lock(src_rq, dst_rq);
1905 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1908 if (task_cpu(arg->src_task) != arg->src_cpu)
1911 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
1914 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
1917 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1918 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1923 double_rq_unlock(src_rq, dst_rq);
1924 raw_spin_unlock(&arg->dst_task->pi_lock);
1925 raw_spin_unlock(&arg->src_task->pi_lock);
1931 * Cross migrate two tasks
1933 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1934 int target_cpu, int curr_cpu)
1936 struct migration_swap_arg arg;
1939 arg = (struct migration_swap_arg){
1941 .src_cpu = curr_cpu,
1943 .dst_cpu = target_cpu,
1946 if (arg.src_cpu == arg.dst_cpu)
1950 * These three tests are all lockless; this is OK since all of them
1951 * will be re-checked with proper locks held further down the line.
1953 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1956 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
1959 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
1962 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1963 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1968 #endif /* CONFIG_NUMA_BALANCING */
1971 * wait_task_inactive - wait for a thread to unschedule.
1973 * If @match_state is nonzero, it's the @p->state value just checked and
1974 * not expected to change. If it changes, i.e. @p might have woken up,
1975 * then return zero. When we succeed in waiting for @p to be off its CPU,
1976 * we return a positive number (its total switch count). If a second call
1977 * a short while later returns the same number, the caller can be sure that
1978 * @p has remained unscheduled the whole time.
1980 * The caller must ensure that the task *will* unschedule sometime soon,
1981 * else this function might spin for a *long* time. This function can't
1982 * be called with interrupts off, or it may introduce deadlock with
1983 * smp_call_function() if an IPI is sent by the same process we are
1984 * waiting to become inactive.
1986 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1988 int running, queued;
1995 * We do the initial early heuristics without holding
1996 * any task-queue locks at all. We'll only try to get
1997 * the runqueue lock when things look like they will
2003 * If the task is actively running on another CPU
2004 * still, just relax and busy-wait without holding
2007 * NOTE! Since we don't hold any locks, it's not
2008 * even sure that "rq" stays as the right runqueue!
2009 * But we don't care, since "task_running()" will
2010 * return false if the runqueue has changed and p
2011 * is actually now running somewhere else!
2013 while (task_running(rq, p)) {
2014 if (match_state && unlikely(p->state != match_state))
2020 * Ok, time to look more closely! We need the rq
2021 * lock now, to be *sure*. If we're wrong, we'll
2022 * just go back and repeat.
2024 rq = task_rq_lock(p, &rf);
2025 trace_sched_wait_task(p);
2026 running = task_running(rq, p);
2027 queued = task_on_rq_queued(p);
2029 if (!match_state || p->state == match_state)
2030 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2031 task_rq_unlock(rq, p, &rf);
2034 * If it changed from the expected state, bail out now.
2036 if (unlikely(!ncsw))
2040 * Was it really running after all now that we
2041 * checked with the proper locks actually held?
2043 * Oops. Go back and try again..
2045 if (unlikely(running)) {
2051 * It's not enough that it's not actively running,
2052 * it must be off the runqueue _entirely_, and not
2055 * So if it was still runnable (but just not actively
2056 * running right now), it's preempted, and we should
2057 * yield - it could be a while.
2059 if (unlikely(queued)) {
2060 ktime_t to = NSEC_PER_SEC / HZ;
2062 set_current_state(TASK_UNINTERRUPTIBLE);
2063 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2068 * Ahh, all good. It wasn't running, and it wasn't
2069 * runnable, which means that it will never become
2070 * running in the future either. We're all done!
2079 * kick_process - kick a running thread to enter/exit the kernel
2080 * @p: the to-be-kicked thread
2082 * Cause a process which is running on another CPU to enter
2083 * kernel-mode, without any delay. (to get signals handled.)
2085 * NOTE: this function doesn't have to take the runqueue lock,
2086 * because all it wants to ensure is that the remote task enters
2087 * the kernel. If the IPI races and the task has been migrated
2088 * to another CPU then no harm is done and the purpose has been
2091 void kick_process(struct task_struct *p)
2097 if ((cpu != smp_processor_id()) && task_curr(p))
2098 smp_send_reschedule(cpu);
2101 EXPORT_SYMBOL_GPL(kick_process);
2104 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2106 * A few notes on cpu_active vs cpu_online:
2108 * - cpu_active must be a subset of cpu_online
2110 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2111 * see __set_cpus_allowed_ptr(). At this point the newly online
2112 * CPU isn't yet part of the sched domains, and balancing will not
2115 * - on CPU-down we clear cpu_active() to mask the sched domains and
2116 * avoid the load balancer to place new tasks on the to be removed
2117 * CPU. Existing tasks will remain running there and will be taken
2120 * This means that fallback selection must not select !active CPUs.
2121 * And can assume that any active CPU must be online. Conversely
2122 * select_task_rq() below may allow selection of !active CPUs in order
2123 * to satisfy the above rules.
2125 static int select_fallback_rq(int cpu, struct task_struct *p)
2127 int nid = cpu_to_node(cpu);
2128 const struct cpumask *nodemask = NULL;
2129 enum { cpuset, possible, fail } state = cpuset;
2133 * If the node that the CPU is on has been offlined, cpu_to_node()
2134 * will return -1. There is no CPU on the node, and we should
2135 * select the CPU on the other node.
2138 nodemask = cpumask_of_node(nid);
2140 /* Look for allowed, online CPU in same node. */
2141 for_each_cpu(dest_cpu, nodemask) {
2142 if (!cpu_active(dest_cpu))
2144 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2150 /* Any allowed, online CPU? */
2151 for_each_cpu(dest_cpu, p->cpus_ptr) {
2152 if (!is_cpu_allowed(p, dest_cpu))
2158 /* No more Mr. Nice Guy. */
2161 if (IS_ENABLED(CONFIG_CPUSETS)) {
2162 cpuset_cpus_allowed_fallback(p);
2168 do_set_cpus_allowed(p, cpu_possible_mask);
2179 if (state != cpuset) {
2181 * Don't tell them about moving exiting tasks or
2182 * kernel threads (both mm NULL), since they never
2185 if (p->mm && printk_ratelimit()) {
2186 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2187 task_pid_nr(p), p->comm, cpu);
2195 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2198 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2200 lockdep_assert_held(&p->pi_lock);
2202 if (p->nr_cpus_allowed > 1)
2203 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2205 cpu = cpumask_any(p->cpus_ptr);
2208 * In order not to call set_task_cpu() on a blocking task we need
2209 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2212 * Since this is common to all placement strategies, this lives here.
2214 * [ this allows ->select_task() to simply return task_cpu(p) and
2215 * not worry about this generic constraint ]
2217 if (unlikely(!is_cpu_allowed(p, cpu)))
2218 cpu = select_fallback_rq(task_cpu(p), p);
2223 static void update_avg(u64 *avg, u64 sample)
2225 s64 diff = sample - *avg;
2229 void sched_set_stop_task(int cpu, struct task_struct *stop)
2231 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2232 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2236 * Make it appear like a SCHED_FIFO task, its something
2237 * userspace knows about and won't get confused about.
2239 * Also, it will make PI more or less work without too
2240 * much confusion -- but then, stop work should not
2241 * rely on PI working anyway.
2243 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2245 stop->sched_class = &stop_sched_class;
2248 cpu_rq(cpu)->stop = stop;
2252 * Reset it back to a normal scheduling class so that
2253 * it can die in pieces.
2255 old_stop->sched_class = &rt_sched_class;
2261 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2262 const struct cpumask *new_mask, bool check)
2264 return set_cpus_allowed_ptr(p, new_mask);
2267 #endif /* CONFIG_SMP */
2270 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2274 if (!schedstat_enabled())
2280 if (cpu == rq->cpu) {
2281 __schedstat_inc(rq->ttwu_local);
2282 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2284 struct sched_domain *sd;
2286 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2288 for_each_domain(rq->cpu, sd) {
2289 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2290 __schedstat_inc(sd->ttwu_wake_remote);
2297 if (wake_flags & WF_MIGRATED)
2298 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2299 #endif /* CONFIG_SMP */
2301 __schedstat_inc(rq->ttwu_count);
2302 __schedstat_inc(p->se.statistics.nr_wakeups);
2304 if (wake_flags & WF_SYNC)
2305 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2309 * Mark the task runnable and perform wakeup-preemption.
2311 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2312 struct rq_flags *rf)
2314 check_preempt_curr(rq, p, wake_flags);
2315 p->state = TASK_RUNNING;
2316 trace_sched_wakeup(p);
2319 if (p->sched_class->task_woken) {
2321 * Our task @p is fully woken up and running; so its safe to
2322 * drop the rq->lock, hereafter rq is only used for statistics.
2324 rq_unpin_lock(rq, rf);
2325 p->sched_class->task_woken(rq, p);
2326 rq_repin_lock(rq, rf);
2329 if (rq->idle_stamp) {
2330 u64 delta = rq_clock(rq) - rq->idle_stamp;
2331 u64 max = 2*rq->max_idle_balance_cost;
2333 update_avg(&rq->avg_idle, delta);
2335 if (rq->avg_idle > max)
2344 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2345 struct rq_flags *rf)
2347 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2349 lockdep_assert_held(&rq->lock);
2352 if (p->sched_contributes_to_load)
2353 rq->nr_uninterruptible--;
2355 if (wake_flags & WF_MIGRATED)
2356 en_flags |= ENQUEUE_MIGRATED;
2359 activate_task(rq, p, en_flags);
2360 ttwu_do_wakeup(rq, p, wake_flags, rf);
2364 * Called in case the task @p isn't fully descheduled from its runqueue,
2365 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2366 * since all we need to do is flip p->state to TASK_RUNNING, since
2367 * the task is still ->on_rq.
2369 static int ttwu_remote(struct task_struct *p, int wake_flags)
2375 rq = __task_rq_lock(p, &rf);
2376 if (task_on_rq_queued(p)) {
2377 /* check_preempt_curr() may use rq clock */
2378 update_rq_clock(rq);
2379 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2382 __task_rq_unlock(rq, &rf);
2388 void sched_ttwu_pending(void)
2390 struct rq *rq = this_rq();
2391 struct llist_node *llist = llist_del_all(&rq->wake_list);
2392 struct task_struct *p, *t;
2398 rq_lock_irqsave(rq, &rf);
2399 update_rq_clock(rq);
2401 llist_for_each_entry_safe(p, t, llist, wake_entry)
2402 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2404 rq_unlock_irqrestore(rq, &rf);
2407 void scheduler_ipi(void)
2410 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
2411 * TIF_NEED_RESCHED remotely (for the first time) will also send
2414 preempt_fold_need_resched();
2416 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2420 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2421 * traditionally all their work was done from the interrupt return
2422 * path. Now that we actually do some work, we need to make sure
2425 * Some archs already do call them, luckily irq_enter/exit nest
2428 * Arguably we should visit all archs and update all handlers,
2429 * however a fair share of IPIs are still resched only so this would
2430 * somewhat pessimize the simple resched case.
2433 sched_ttwu_pending();
2436 * Check if someone kicked us for doing the nohz idle load balance.
2438 if (unlikely(got_nohz_idle_kick())) {
2439 this_rq()->idle_balance = 1;
2440 raise_softirq_irqoff(SCHED_SOFTIRQ);
2445 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
2447 struct rq *rq = cpu_rq(cpu);
2449 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2451 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
2452 if (!set_nr_if_polling(rq->idle))
2453 smp_send_reschedule(cpu);
2455 trace_sched_wake_idle_without_ipi(cpu);
2459 void wake_up_if_idle(int cpu)
2461 struct rq *rq = cpu_rq(cpu);
2466 if (!is_idle_task(rcu_dereference(rq->curr)))
2469 if (set_nr_if_polling(rq->idle)) {
2470 trace_sched_wake_idle_without_ipi(cpu);
2472 rq_lock_irqsave(rq, &rf);
2473 if (is_idle_task(rq->curr))
2474 smp_send_reschedule(cpu);
2475 /* Else CPU is not idle, do nothing here: */
2476 rq_unlock_irqrestore(rq, &rf);
2483 bool cpus_share_cache(int this_cpu, int that_cpu)
2485 if (this_cpu == that_cpu)
2488 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2490 #endif /* CONFIG_SMP */
2492 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2494 struct rq *rq = cpu_rq(cpu);
2497 #if defined(CONFIG_SMP)
2498 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2499 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2500 ttwu_queue_remote(p, cpu, wake_flags);
2506 update_rq_clock(rq);
2507 ttwu_do_activate(rq, p, wake_flags, &rf);
2512 * Notes on Program-Order guarantees on SMP systems.
2516 * The basic program-order guarantee on SMP systems is that when a task [t]
2517 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2518 * execution on its new CPU [c1].
2520 * For migration (of runnable tasks) this is provided by the following means:
2522 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2523 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2524 * rq(c1)->lock (if not at the same time, then in that order).
2525 * C) LOCK of the rq(c1)->lock scheduling in task
2527 * Release/acquire chaining guarantees that B happens after A and C after B.
2528 * Note: the CPU doing B need not be c0 or c1
2537 * UNLOCK rq(0)->lock
2539 * LOCK rq(0)->lock // orders against CPU0
2541 * UNLOCK rq(0)->lock
2545 * UNLOCK rq(1)->lock
2547 * LOCK rq(1)->lock // orders against CPU2
2550 * UNLOCK rq(1)->lock
2553 * BLOCKING -- aka. SLEEP + WAKEUP
2555 * For blocking we (obviously) need to provide the same guarantee as for
2556 * migration. However the means are completely different as there is no lock
2557 * chain to provide order. Instead we do:
2559 * 1) smp_store_release(X->on_cpu, 0)
2560 * 2) smp_cond_load_acquire(!X->on_cpu)
2564 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2566 * LOCK rq(0)->lock LOCK X->pi_lock
2569 * smp_store_release(X->on_cpu, 0);
2571 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2577 * X->state = RUNNING
2578 * UNLOCK rq(2)->lock
2580 * LOCK rq(2)->lock // orders against CPU1
2583 * UNLOCK rq(2)->lock
2586 * UNLOCK rq(0)->lock
2589 * However, for wakeups there is a second guarantee we must provide, namely we
2590 * must ensure that CONDITION=1 done by the caller can not be reordered with
2591 * accesses to the task state; see try_to_wake_up() and set_current_state().
2595 * try_to_wake_up - wake up a thread
2596 * @p: the thread to be awakened
2597 * @state: the mask of task states that can be woken
2598 * @wake_flags: wake modifier flags (WF_*)
2600 * If (@state & @p->state) @p->state = TASK_RUNNING.
2602 * If the task was not queued/runnable, also place it back on a runqueue.
2604 * Atomic against schedule() which would dequeue a task, also see
2605 * set_current_state().
2607 * This function executes a full memory barrier before accessing the task
2608 * state; see set_current_state().
2610 * Return: %true if @p->state changes (an actual wakeup was done),
2614 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2616 unsigned long flags;
2617 int cpu, success = 0;
2622 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2623 * == smp_processor_id()'. Together this means we can special
2624 * case the whole 'p->on_rq && ttwu_remote()' case below
2625 * without taking any locks.
2628 * - we rely on Program-Order guarantees for all the ordering,
2629 * - we're serialized against set_special_state() by virtue of
2630 * it disabling IRQs (this allows not taking ->pi_lock).
2632 if (!(p->state & state))
2637 trace_sched_waking(p);
2638 p->state = TASK_RUNNING;
2639 trace_sched_wakeup(p);
2644 * If we are going to wake up a thread waiting for CONDITION we
2645 * need to ensure that CONDITION=1 done by the caller can not be
2646 * reordered with p->state check below. This pairs with mb() in
2647 * set_current_state() the waiting thread does.
2649 raw_spin_lock_irqsave(&p->pi_lock, flags);
2650 smp_mb__after_spinlock();
2651 if (!(p->state & state))
2654 trace_sched_waking(p);
2656 /* We're going to change ->state: */
2661 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2662 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2663 * in smp_cond_load_acquire() below.
2665 * sched_ttwu_pending() try_to_wake_up()
2666 * STORE p->on_rq = 1 LOAD p->state
2669 * __schedule() (switch to task 'p')
2670 * LOCK rq->lock smp_rmb();
2671 * smp_mb__after_spinlock();
2675 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2677 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2678 * __schedule(). See the comment for smp_mb__after_spinlock().
2681 if (p->on_rq && ttwu_remote(p, wake_flags))
2686 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2687 * possible to, falsely, observe p->on_cpu == 0.
2689 * One must be running (->on_cpu == 1) in order to remove oneself
2690 * from the runqueue.
2692 * __schedule() (switch to task 'p') try_to_wake_up()
2693 * STORE p->on_cpu = 1 LOAD p->on_rq
2696 * __schedule() (put 'p' to sleep)
2697 * LOCK rq->lock smp_rmb();
2698 * smp_mb__after_spinlock();
2699 * STORE p->on_rq = 0 LOAD p->on_cpu
2701 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2702 * __schedule(). See the comment for smp_mb__after_spinlock().
2707 * If the owning (remote) CPU is still in the middle of schedule() with
2708 * this task as prev, wait until its done referencing the task.
2710 * Pairs with the smp_store_release() in finish_task().
2712 * This ensures that tasks getting woken will be fully ordered against
2713 * their previous state and preserve Program Order.
2715 smp_cond_load_acquire(&p->on_cpu, !VAL);
2717 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2718 p->state = TASK_WAKING;
2721 delayacct_blkio_end(p);
2722 atomic_dec(&task_rq(p)->nr_iowait);
2725 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2726 if (task_cpu(p) != cpu) {
2727 wake_flags |= WF_MIGRATED;
2728 psi_ttwu_dequeue(p);
2729 set_task_cpu(p, cpu);
2732 #else /* CONFIG_SMP */
2735 delayacct_blkio_end(p);
2736 atomic_dec(&task_rq(p)->nr_iowait);
2739 #endif /* CONFIG_SMP */
2741 ttwu_queue(p, cpu, wake_flags);
2743 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 ttwu_stat(p, cpu, wake_flags);
2753 * wake_up_process - Wake up a specific process
2754 * @p: The process to be woken up.
2756 * Attempt to wake up the nominated process and move it to the set of runnable
2759 * Return: 1 if the process was woken up, 0 if it was already running.
2761 * This function executes a full memory barrier before accessing the task state.
2763 int wake_up_process(struct task_struct *p)
2765 return try_to_wake_up(p, TASK_NORMAL, 0);
2767 EXPORT_SYMBOL(wake_up_process);
2769 int wake_up_state(struct task_struct *p, unsigned int state)
2771 return try_to_wake_up(p, state, 0);
2775 * Perform scheduler related setup for a newly forked process p.
2776 * p is forked by current.
2778 * __sched_fork() is basic setup used by init_idle() too:
2780 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2785 p->se.exec_start = 0;
2786 p->se.sum_exec_runtime = 0;
2787 p->se.prev_sum_exec_runtime = 0;
2788 p->se.nr_migrations = 0;
2790 INIT_LIST_HEAD(&p->se.group_node);
2792 #ifdef CONFIG_FAIR_GROUP_SCHED
2793 p->se.cfs_rq = NULL;
2796 #ifdef CONFIG_SCHEDSTATS
2797 /* Even if schedstat is disabled, there should not be garbage */
2798 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2801 RB_CLEAR_NODE(&p->dl.rb_node);
2802 init_dl_task_timer(&p->dl);
2803 init_dl_inactive_task_timer(&p->dl);
2804 __dl_clear_params(p);
2806 INIT_LIST_HEAD(&p->rt.run_list);
2808 p->rt.time_slice = sched_rr_timeslice;
2812 #ifdef CONFIG_PREEMPT_NOTIFIERS
2813 INIT_HLIST_HEAD(&p->preempt_notifiers);
2816 #ifdef CONFIG_COMPACTION
2817 p->capture_control = NULL;
2819 init_numa_balancing(clone_flags, p);
2822 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2824 #ifdef CONFIG_NUMA_BALANCING
2826 void set_numabalancing_state(bool enabled)
2829 static_branch_enable(&sched_numa_balancing);
2831 static_branch_disable(&sched_numa_balancing);
2834 #ifdef CONFIG_PROC_SYSCTL
2835 int sysctl_numa_balancing(struct ctl_table *table, int write,
2836 void __user *buffer, size_t *lenp, loff_t *ppos)
2840 int state = static_branch_likely(&sched_numa_balancing);
2842 if (write && !capable(CAP_SYS_ADMIN))
2847 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2851 set_numabalancing_state(state);
2857 #ifdef CONFIG_SCHEDSTATS
2859 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2860 static bool __initdata __sched_schedstats = false;
2862 static void set_schedstats(bool enabled)
2865 static_branch_enable(&sched_schedstats);
2867 static_branch_disable(&sched_schedstats);
2870 void force_schedstat_enabled(void)
2872 if (!schedstat_enabled()) {
2873 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2874 static_branch_enable(&sched_schedstats);
2878 static int __init setup_schedstats(char *str)
2885 * This code is called before jump labels have been set up, so we can't
2886 * change the static branch directly just yet. Instead set a temporary
2887 * variable so init_schedstats() can do it later.
2889 if (!strcmp(str, "enable")) {
2890 __sched_schedstats = true;
2892 } else if (!strcmp(str, "disable")) {
2893 __sched_schedstats = false;
2898 pr_warn("Unable to parse schedstats=\n");
2902 __setup("schedstats=", setup_schedstats);
2904 static void __init init_schedstats(void)
2906 set_schedstats(__sched_schedstats);
2909 #ifdef CONFIG_PROC_SYSCTL
2910 int sysctl_schedstats(struct ctl_table *table, int write,
2911 void __user *buffer, size_t *lenp, loff_t *ppos)
2915 int state = static_branch_likely(&sched_schedstats);
2917 if (write && !capable(CAP_SYS_ADMIN))
2922 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2926 set_schedstats(state);
2929 #endif /* CONFIG_PROC_SYSCTL */
2930 #else /* !CONFIG_SCHEDSTATS */
2931 static inline void init_schedstats(void) {}
2932 #endif /* CONFIG_SCHEDSTATS */
2935 * fork()/clone()-time setup:
2937 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2939 unsigned long flags;
2941 __sched_fork(clone_flags, p);
2943 * We mark the process as NEW here. This guarantees that
2944 * nobody will actually run it, and a signal or other external
2945 * event cannot wake it up and insert it on the runqueue either.
2947 p->state = TASK_NEW;
2950 * Make sure we do not leak PI boosting priority to the child.
2952 p->prio = current->normal_prio;
2957 * Revert to default priority/policy on fork if requested.
2959 if (unlikely(p->sched_reset_on_fork)) {
2960 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2961 p->policy = SCHED_NORMAL;
2962 p->static_prio = NICE_TO_PRIO(0);
2964 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2965 p->static_prio = NICE_TO_PRIO(0);
2967 p->prio = p->normal_prio = __normal_prio(p);
2968 set_load_weight(p, false);
2971 * We don't need the reset flag anymore after the fork. It has
2972 * fulfilled its duty:
2974 p->sched_reset_on_fork = 0;
2977 if (dl_prio(p->prio))
2979 else if (rt_prio(p->prio))
2980 p->sched_class = &rt_sched_class;
2982 p->sched_class = &fair_sched_class;
2984 init_entity_runnable_average(&p->se);
2987 * The child is not yet in the pid-hash so no cgroup attach races,
2988 * and the cgroup is pinned to this child due to cgroup_fork()
2989 * is ran before sched_fork().
2991 * Silence PROVE_RCU.
2993 raw_spin_lock_irqsave(&p->pi_lock, flags);
2996 * We're setting the CPU for the first time, we don't migrate,
2997 * so use __set_task_cpu().
2999 __set_task_cpu(p, smp_processor_id());
3000 if (p->sched_class->task_fork)
3001 p->sched_class->task_fork(p);
3002 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3004 #ifdef CONFIG_SCHED_INFO
3005 if (likely(sched_info_on()))
3006 memset(&p->sched_info, 0, sizeof(p->sched_info));
3008 #if defined(CONFIG_SMP)
3011 init_task_preempt_count(p);
3013 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3014 RB_CLEAR_NODE(&p->pushable_dl_tasks);
3019 unsigned long to_ratio(u64 period, u64 runtime)
3021 if (runtime == RUNTIME_INF)
3025 * Doing this here saves a lot of checks in all
3026 * the calling paths, and returning zero seems
3027 * safe for them anyway.
3032 return div64_u64(runtime << BW_SHIFT, period);
3036 * wake_up_new_task - wake up a newly created task for the first time.
3038 * This function will do some initial scheduler statistics housekeeping
3039 * that must be done for every newly created context, then puts the task
3040 * on the runqueue and wakes it.
3042 void wake_up_new_task(struct task_struct *p)
3047 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3048 p->state = TASK_RUNNING;
3051 * Fork balancing, do it here and not earlier because:
3052 * - cpus_ptr can change in the fork path
3053 * - any previously selected CPU might disappear through hotplug
3055 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3056 * as we're not fully set-up yet.
3058 p->recent_used_cpu = task_cpu(p);
3060 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
3062 rq = __task_rq_lock(p, &rf);
3063 update_rq_clock(rq);
3064 post_init_entity_util_avg(p);
3066 activate_task(rq, p, ENQUEUE_NOCLOCK);
3067 trace_sched_wakeup_new(p);
3068 check_preempt_curr(rq, p, WF_FORK);
3070 if (p->sched_class->task_woken) {
3072 * Nothing relies on rq->lock after this, so its fine to
3075 rq_unpin_lock(rq, &rf);
3076 p->sched_class->task_woken(rq, p);
3077 rq_repin_lock(rq, &rf);
3080 task_rq_unlock(rq, p, &rf);
3083 #ifdef CONFIG_PREEMPT_NOTIFIERS
3085 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3087 void preempt_notifier_inc(void)
3089 static_branch_inc(&preempt_notifier_key);
3091 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3093 void preempt_notifier_dec(void)
3095 static_branch_dec(&preempt_notifier_key);
3097 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3100 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3101 * @notifier: notifier struct to register
3103 void preempt_notifier_register(struct preempt_notifier *notifier)
3105 if (!static_branch_unlikely(&preempt_notifier_key))
3106 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3108 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3110 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3113 * preempt_notifier_unregister - no longer interested in preemption notifications
3114 * @notifier: notifier struct to unregister
3116 * This is *not* safe to call from within a preemption notifier.
3118 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3120 hlist_del(¬ifier->link);
3122 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3124 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3126 struct preempt_notifier *notifier;
3128 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3129 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3132 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3134 if (static_branch_unlikely(&preempt_notifier_key))
3135 __fire_sched_in_preempt_notifiers(curr);
3139 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3140 struct task_struct *next)
3142 struct preempt_notifier *notifier;
3144 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3145 notifier->ops->sched_out(notifier, next);
3148 static __always_inline void
3149 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3150 struct task_struct *next)
3152 if (static_branch_unlikely(&preempt_notifier_key))
3153 __fire_sched_out_preempt_notifiers(curr, next);
3156 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3158 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3163 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3164 struct task_struct *next)
3168 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3170 static inline void prepare_task(struct task_struct *next)
3174 * Claim the task as running, we do this before switching to it
3175 * such that any running task will have this set.
3181 static inline void finish_task(struct task_struct *prev)
3185 * After ->on_cpu is cleared, the task can be moved to a different CPU.
3186 * We must ensure this doesn't happen until the switch is completely
3189 * In particular, the load of prev->state in finish_task_switch() must
3190 * happen before this.
3192 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3194 smp_store_release(&prev->on_cpu, 0);
3199 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3202 * Since the runqueue lock will be released by the next
3203 * task (which is an invalid locking op but in the case
3204 * of the scheduler it's an obvious special-case), so we
3205 * do an early lockdep release here:
3207 rq_unpin_lock(rq, rf);
3208 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3209 #ifdef CONFIG_DEBUG_SPINLOCK
3210 /* this is a valid case when another task releases the spinlock */
3211 rq->lock.owner = next;
3215 static inline void finish_lock_switch(struct rq *rq)
3218 * If we are tracking spinlock dependencies then we have to
3219 * fix up the runqueue lock - which gets 'carried over' from
3220 * prev into current:
3222 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3223 raw_spin_unlock_irq(&rq->lock);
3227 * NOP if the arch has not defined these:
3230 #ifndef prepare_arch_switch
3231 # define prepare_arch_switch(next) do { } while (0)
3234 #ifndef finish_arch_post_lock_switch
3235 # define finish_arch_post_lock_switch() do { } while (0)
3239 * prepare_task_switch - prepare to switch tasks
3240 * @rq: the runqueue preparing to switch
3241 * @prev: the current task that is being switched out
3242 * @next: the task we are going to switch to.
3244 * This is called with the rq lock held and interrupts off. It must
3245 * be paired with a subsequent finish_task_switch after the context
3248 * prepare_task_switch sets up locking and calls architecture specific
3252 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3253 struct task_struct *next)
3255 kcov_prepare_switch(prev);
3256 sched_info_switch(rq, prev, next);
3257 perf_event_task_sched_out(prev, next);
3259 fire_sched_out_preempt_notifiers(prev, next);
3261 prepare_arch_switch(next);
3265 * finish_task_switch - clean up after a task-switch
3266 * @prev: the thread we just switched away from.
3268 * finish_task_switch must be called after the context switch, paired
3269 * with a prepare_task_switch call before the context switch.
3270 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3271 * and do any other architecture-specific cleanup actions.
3273 * Note that we may have delayed dropping an mm in context_switch(). If
3274 * so, we finish that here outside of the runqueue lock. (Doing it
3275 * with the lock held can cause deadlocks; see schedule() for
3278 * The context switch have flipped the stack from under us and restored the
3279 * local variables which were saved when this task called schedule() in the
3280 * past. prev == current is still correct but we need to recalculate this_rq
3281 * because prev may have moved to another CPU.
3283 static struct rq *finish_task_switch(struct task_struct *prev)
3284 __releases(rq->lock)
3286 struct rq *rq = this_rq();
3287 struct mm_struct *mm = rq->prev_mm;
3291 * The previous task will have left us with a preempt_count of 2
3292 * because it left us after:
3295 * preempt_disable(); // 1
3297 * raw_spin_lock_irq(&rq->lock) // 2
3299 * Also, see FORK_PREEMPT_COUNT.
3301 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3302 "corrupted preempt_count: %s/%d/0x%x\n",
3303 current->comm, current->pid, preempt_count()))
3304 preempt_count_set(FORK_PREEMPT_COUNT);
3309 * A task struct has one reference for the use as "current".
3310 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3311 * schedule one last time. The schedule call will never return, and
3312 * the scheduled task must drop that reference.
3314 * We must observe prev->state before clearing prev->on_cpu (in
3315 * finish_task), otherwise a concurrent wakeup can get prev
3316 * running on another CPU and we could rave with its RUNNING -> DEAD
3317 * transition, resulting in a double drop.
3319 prev_state = prev->state;
3320 vtime_task_switch(prev);
3321 perf_event_task_sched_in(prev, current);
3323 finish_lock_switch(rq);
3324 finish_arch_post_lock_switch();
3325 kcov_finish_switch(current);
3327 fire_sched_in_preempt_notifiers(current);
3329 * When switching through a kernel thread, the loop in
3330 * membarrier_{private,global}_expedited() may have observed that
3331 * kernel thread and not issued an IPI. It is therefore possible to
3332 * schedule between user->kernel->user threads without passing though
3333 * switch_mm(). Membarrier requires a barrier after storing to
3334 * rq->curr, before returning to userspace, so provide them here:
3336 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3337 * provided by mmdrop(),
3338 * - a sync_core for SYNC_CORE.
3341 membarrier_mm_sync_core_before_usermode(mm);
3344 if (unlikely(prev_state == TASK_DEAD)) {
3345 if (prev->sched_class->task_dead)
3346 prev->sched_class->task_dead(prev);
3349 * Remove function-return probe instances associated with this
3350 * task and put them back on the free list.
3352 kprobe_flush_task(prev);
3354 /* Task is done with its stack. */
3355 put_task_stack(prev);
3357 put_task_struct_rcu_user(prev);
3360 tick_nohz_task_switch();
3366 /* rq->lock is NOT held, but preemption is disabled */
3367 static void __balance_callback(struct rq *rq)
3369 struct callback_head *head, *next;
3370 void (*func)(struct rq *rq);
3371 unsigned long flags;
3373 raw_spin_lock_irqsave(&rq->lock, flags);
3374 head = rq->balance_callback;
3375 rq->balance_callback = NULL;
3377 func = (void (*)(struct rq *))head->func;
3384 raw_spin_unlock_irqrestore(&rq->lock, flags);
3387 static inline void balance_callback(struct rq *rq)
3389 if (unlikely(rq->balance_callback))
3390 __balance_callback(rq);
3395 static inline void balance_callback(struct rq *rq)
3402 * schedule_tail - first thing a freshly forked thread must call.
3403 * @prev: the thread we just switched away from.
3405 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3406 __releases(rq->lock)
3411 * New tasks start with FORK_PREEMPT_COUNT, see there and
3412 * finish_task_switch() for details.
3414 * finish_task_switch() will drop rq->lock() and lower preempt_count
3415 * and the preempt_enable() will end up enabling preemption (on
3416 * PREEMPT_COUNT kernels).
3419 rq = finish_task_switch(prev);
3420 balance_callback(rq);
3423 if (current->set_child_tid)
3424 put_user(task_pid_vnr(current), current->set_child_tid);
3426 calculate_sigpending();
3430 * context_switch - switch to the new MM and the new thread's register state.
3432 static __always_inline struct rq *
3433 context_switch(struct rq *rq, struct task_struct *prev,
3434 struct task_struct *next, struct rq_flags *rf)
3436 prepare_task_switch(rq, prev, next);
3439 * For paravirt, this is coupled with an exit in switch_to to
3440 * combine the page table reload and the switch backend into
3443 arch_start_context_switch(prev);
3446 * kernel -> kernel lazy + transfer active
3447 * user -> kernel lazy + mmgrab() active
3449 * kernel -> user switch + mmdrop() active
3450 * user -> user switch
3452 if (!next->mm) { // to kernel
3453 enter_lazy_tlb(prev->active_mm, next);
3455 next->active_mm = prev->active_mm;
3456 if (prev->mm) // from user
3457 mmgrab(prev->active_mm);
3459 prev->active_mm = NULL;
3461 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3463 * sys_membarrier() requires an smp_mb() between setting
3464 * rq->curr / membarrier_switch_mm() and returning to userspace.
3466 * The below provides this either through switch_mm(), or in
3467 * case 'prev->active_mm == next->mm' through
3468 * finish_task_switch()'s mmdrop().
3470 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3472 if (!prev->mm) { // from kernel
3473 /* will mmdrop() in finish_task_switch(). */
3474 rq->prev_mm = prev->active_mm;
3475 prev->active_mm = NULL;
3479 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3481 prepare_lock_switch(rq, next, rf);
3483 /* Here we just switch the register state and the stack. */
3484 switch_to(prev, next, prev);
3487 return finish_task_switch(prev);
3491 * nr_running and nr_context_switches:
3493 * externally visible scheduler statistics: current number of runnable
3494 * threads, total number of context switches performed since bootup.
3496 unsigned long nr_running(void)
3498 unsigned long i, sum = 0;
3500 for_each_online_cpu(i)
3501 sum += cpu_rq(i)->nr_running;
3507 * Check if only the current task is running on the CPU.
3509 * Caution: this function does not check that the caller has disabled
3510 * preemption, thus the result might have a time-of-check-to-time-of-use
3511 * race. The caller is responsible to use it correctly, for example:
3513 * - from a non-preemptible section (of course)
3515 * - from a thread that is bound to a single CPU
3517 * - in a loop with very short iterations (e.g. a polling loop)
3519 bool single_task_running(void)
3521 return raw_rq()->nr_running == 1;
3523 EXPORT_SYMBOL(single_task_running);
3525 unsigned long long nr_context_switches(void)
3528 unsigned long long sum = 0;
3530 for_each_possible_cpu(i)
3531 sum += cpu_rq(i)->nr_switches;
3537 * Consumers of these two interfaces, like for example the cpuidle menu
3538 * governor, are using nonsensical data. Preferring shallow idle state selection
3539 * for a CPU that has IO-wait which might not even end up running the task when
3540 * it does become runnable.
3543 unsigned long nr_iowait_cpu(int cpu)
3545 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3549 * IO-wait accounting, and how its mostly bollocks (on SMP).
3551 * The idea behind IO-wait account is to account the idle time that we could
3552 * have spend running if it were not for IO. That is, if we were to improve the
3553 * storage performance, we'd have a proportional reduction in IO-wait time.
3555 * This all works nicely on UP, where, when a task blocks on IO, we account
3556 * idle time as IO-wait, because if the storage were faster, it could've been
3557 * running and we'd not be idle.
3559 * This has been extended to SMP, by doing the same for each CPU. This however
3562 * Imagine for instance the case where two tasks block on one CPU, only the one
3563 * CPU will have IO-wait accounted, while the other has regular idle. Even
3564 * though, if the storage were faster, both could've ran at the same time,
3565 * utilising both CPUs.
3567 * This means, that when looking globally, the current IO-wait accounting on
3568 * SMP is a lower bound, by reason of under accounting.
3570 * Worse, since the numbers are provided per CPU, they are sometimes
3571 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3572 * associated with any one particular CPU, it can wake to another CPU than it
3573 * blocked on. This means the per CPU IO-wait number is meaningless.
3575 * Task CPU affinities can make all that even more 'interesting'.
3578 unsigned long nr_iowait(void)
3580 unsigned long i, sum = 0;
3582 for_each_possible_cpu(i)
3583 sum += nr_iowait_cpu(i);
3591 * sched_exec - execve() is a valuable balancing opportunity, because at
3592 * this point the task has the smallest effective memory and cache footprint.
3594 void sched_exec(void)
3596 struct task_struct *p = current;
3597 unsigned long flags;
3600 raw_spin_lock_irqsave(&p->pi_lock, flags);
3601 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3602 if (dest_cpu == smp_processor_id())
3605 if (likely(cpu_active(dest_cpu))) {
3606 struct migration_arg arg = { p, dest_cpu };
3608 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3609 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3613 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3618 DEFINE_PER_CPU(struct kernel_stat, kstat);
3619 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3621 EXPORT_PER_CPU_SYMBOL(kstat);
3622 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3625 * The function fair_sched_class.update_curr accesses the struct curr
3626 * and its field curr->exec_start; when called from task_sched_runtime(),
3627 * we observe a high rate of cache misses in practice.
3628 * Prefetching this data results in improved performance.
3630 static inline void prefetch_curr_exec_start(struct task_struct *p)
3632 #ifdef CONFIG_FAIR_GROUP_SCHED
3633 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3635 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3638 prefetch(&curr->exec_start);
3642 * Return accounted runtime for the task.
3643 * In case the task is currently running, return the runtime plus current's
3644 * pending runtime that have not been accounted yet.
3646 unsigned long long task_sched_runtime(struct task_struct *p)
3652 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3654 * 64-bit doesn't need locks to atomically read a 64-bit value.
3655 * So we have a optimization chance when the task's delta_exec is 0.
3656 * Reading ->on_cpu is racy, but this is ok.
3658 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3659 * If we race with it entering CPU, unaccounted time is 0. This is
3660 * indistinguishable from the read occurring a few cycles earlier.
3661 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3662 * been accounted, so we're correct here as well.
3664 if (!p->on_cpu || !task_on_rq_queued(p))
3665 return p->se.sum_exec_runtime;
3668 rq = task_rq_lock(p, &rf);
3670 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3671 * project cycles that may never be accounted to this
3672 * thread, breaking clock_gettime().
3674 if (task_current(rq, p) && task_on_rq_queued(p)) {
3675 prefetch_curr_exec_start(p);
3676 update_rq_clock(rq);
3677 p->sched_class->update_curr(rq);
3679 ns = p->se.sum_exec_runtime;
3680 task_rq_unlock(rq, p, &rf);
3686 * This function gets called by the timer code, with HZ frequency.
3687 * We call it with interrupts disabled.
3689 void scheduler_tick(void)
3691 int cpu = smp_processor_id();
3692 struct rq *rq = cpu_rq(cpu);
3693 struct task_struct *curr = rq->curr;
3700 update_rq_clock(rq);
3701 curr->sched_class->task_tick(rq, curr, 0);
3702 calc_global_load_tick(rq);
3707 perf_event_task_tick();
3710 rq->idle_balance = idle_cpu(cpu);
3711 trigger_load_balance(rq);
3715 #ifdef CONFIG_NO_HZ_FULL
3720 struct delayed_work work;
3722 /* Values for ->state, see diagram below. */
3723 #define TICK_SCHED_REMOTE_OFFLINE 0
3724 #define TICK_SCHED_REMOTE_OFFLINING 1
3725 #define TICK_SCHED_REMOTE_RUNNING 2
3728 * State diagram for ->state:
3731 * TICK_SCHED_REMOTE_OFFLINE
3734 * | | sched_tick_remote()
3737 * +--TICK_SCHED_REMOTE_OFFLINING
3740 * sched_tick_start() | | sched_tick_stop()
3743 * TICK_SCHED_REMOTE_RUNNING
3746 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3747 * and sched_tick_start() are happy to leave the state in RUNNING.
3750 static struct tick_work __percpu *tick_work_cpu;
3752 static void sched_tick_remote(struct work_struct *work)
3754 struct delayed_work *dwork = to_delayed_work(work);
3755 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3756 int cpu = twork->cpu;
3757 struct rq *rq = cpu_rq(cpu);
3758 struct task_struct *curr;
3764 * Handle the tick only if it appears the remote CPU is running in full
3765 * dynticks mode. The check is racy by nature, but missing a tick or
3766 * having one too much is no big deal because the scheduler tick updates
3767 * statistics and checks timeslices in a time-independent way, regardless
3768 * of when exactly it is running.
3770 if (!tick_nohz_tick_stopped_cpu(cpu))
3773 rq_lock_irq(rq, &rf);
3775 if (cpu_is_offline(cpu))
3778 update_rq_clock(rq);
3780 if (!is_idle_task(curr)) {
3782 * Make sure the next tick runs within a reasonable
3785 delta = rq_clock_task(rq) - curr->se.exec_start;
3786 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3788 curr->sched_class->task_tick(rq, curr, 0);
3790 calc_load_nohz_remote(rq);
3792 rq_unlock_irq(rq, &rf);
3796 * Run the remote tick once per second (1Hz). This arbitrary
3797 * frequency is large enough to avoid overload but short enough
3798 * to keep scheduler internal stats reasonably up to date. But
3799 * first update state to reflect hotplug activity if required.
3801 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3802 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3803 if (os == TICK_SCHED_REMOTE_RUNNING)
3804 queue_delayed_work(system_unbound_wq, dwork, HZ);
3807 static void sched_tick_start(int cpu)
3810 struct tick_work *twork;
3812 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3815 WARN_ON_ONCE(!tick_work_cpu);
3817 twork = per_cpu_ptr(tick_work_cpu, cpu);
3818 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3819 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3820 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3822 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3823 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3827 #ifdef CONFIG_HOTPLUG_CPU
3828 static void sched_tick_stop(int cpu)
3830 struct tick_work *twork;
3833 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3836 WARN_ON_ONCE(!tick_work_cpu);
3838 twork = per_cpu_ptr(tick_work_cpu, cpu);
3839 /* There cannot be competing actions, but don't rely on stop-machine. */
3840 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3841 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3842 /* Don't cancel, as this would mess up the state machine. */
3844 #endif /* CONFIG_HOTPLUG_CPU */
3846 int __init sched_tick_offload_init(void)
3848 tick_work_cpu = alloc_percpu(struct tick_work);
3849 BUG_ON(!tick_work_cpu);
3853 #else /* !CONFIG_NO_HZ_FULL */
3854 static inline void sched_tick_start(int cpu) { }
3855 static inline void sched_tick_stop(int cpu) { }
3858 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
3859 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3861 * If the value passed in is equal to the current preempt count
3862 * then we just disabled preemption. Start timing the latency.
3864 static inline void preempt_latency_start(int val)
3866 if (preempt_count() == val) {
3867 unsigned long ip = get_lock_parent_ip();
3868 #ifdef CONFIG_DEBUG_PREEMPT
3869 current->preempt_disable_ip = ip;
3871 trace_preempt_off(CALLER_ADDR0, ip);
3875 void preempt_count_add(int val)
3877 #ifdef CONFIG_DEBUG_PREEMPT
3881 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3884 __preempt_count_add(val);
3885 #ifdef CONFIG_DEBUG_PREEMPT
3887 * Spinlock count overflowing soon?
3889 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3892 preempt_latency_start(val);
3894 EXPORT_SYMBOL(preempt_count_add);
3895 NOKPROBE_SYMBOL(preempt_count_add);
3898 * If the value passed in equals to the current preempt count
3899 * then we just enabled preemption. Stop timing the latency.
3901 static inline void preempt_latency_stop(int val)
3903 if (preempt_count() == val)
3904 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3907 void preempt_count_sub(int val)
3909 #ifdef CONFIG_DEBUG_PREEMPT
3913 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3916 * Is the spinlock portion underflowing?
3918 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3919 !(preempt_count() & PREEMPT_MASK)))
3923 preempt_latency_stop(val);
3924 __preempt_count_sub(val);
3926 EXPORT_SYMBOL(preempt_count_sub);
3927 NOKPROBE_SYMBOL(preempt_count_sub);
3930 static inline void preempt_latency_start(int val) { }
3931 static inline void preempt_latency_stop(int val) { }
3934 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3936 #ifdef CONFIG_DEBUG_PREEMPT
3937 return p->preempt_disable_ip;
3944 * Print scheduling while atomic bug:
3946 static noinline void __schedule_bug(struct task_struct *prev)
3948 /* Save this before calling printk(), since that will clobber it */
3949 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3951 if (oops_in_progress)
3954 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3955 prev->comm, prev->pid, preempt_count());
3957 debug_show_held_locks(prev);
3959 if (irqs_disabled())
3960 print_irqtrace_events(prev);
3961 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3962 && in_atomic_preempt_off()) {
3963 pr_err("Preemption disabled at:");
3964 print_ip_sym(preempt_disable_ip);
3968 panic("scheduling while atomic\n");
3971 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3975 * Various schedule()-time debugging checks and statistics:
3977 static inline void schedule_debug(struct task_struct *prev, bool preempt)
3979 #ifdef CONFIG_SCHED_STACK_END_CHECK
3980 if (task_stack_end_corrupted(prev))
3981 panic("corrupted stack end detected inside scheduler\n");
3984 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
3985 if (!preempt && prev->state && prev->non_block_count) {
3986 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
3987 prev->comm, prev->pid, prev->non_block_count);
3989 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3993 if (unlikely(in_atomic_preempt_off())) {
3994 __schedule_bug(prev);
3995 preempt_count_set(PREEMPT_DISABLED);
3999 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4001 schedstat_inc(this_rq()->sched_count);
4005 * Pick up the highest-prio task:
4007 static inline struct task_struct *
4008 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4010 const struct sched_class *class;
4011 struct task_struct *p;
4014 * Optimization: we know that if all tasks are in the fair class we can
4015 * call that function directly, but only if the @prev task wasn't of a
4016 * higher scheduling class, because otherwise those loose the
4017 * opportunity to pull in more work from other CPUs.
4019 if (likely((prev->sched_class == &idle_sched_class ||
4020 prev->sched_class == &fair_sched_class) &&
4021 rq->nr_running == rq->cfs.h_nr_running)) {
4023 p = fair_sched_class.pick_next_task(rq, prev, rf);
4024 if (unlikely(p == RETRY_TASK))
4027 /* Assumes fair_sched_class->next == idle_sched_class */
4029 p = idle_sched_class.pick_next_task(rq, prev, rf);
4037 * We must do the balancing pass before put_next_task(), such
4038 * that when we release the rq->lock the task is in the same
4039 * state as before we took rq->lock.
4041 * We can terminate the balance pass as soon as we know there is
4042 * a runnable task of @class priority or higher.
4044 for_class_range(class, prev->sched_class, &idle_sched_class) {
4045 if (class->balance(rq, prev, rf))
4050 put_prev_task(rq, prev);
4052 for_each_class(class) {
4053 p = class->pick_next_task(rq, NULL, NULL);
4058 /* The idle class should always have a runnable task: */
4063 * __schedule() is the main scheduler function.
4065 * The main means of driving the scheduler and thus entering this function are:
4067 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4069 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4070 * paths. For example, see arch/x86/entry_64.S.
4072 * To drive preemption between tasks, the scheduler sets the flag in timer
4073 * interrupt handler scheduler_tick().
4075 * 3. Wakeups don't really cause entry into schedule(). They add a
4076 * task to the run-queue and that's it.
4078 * Now, if the new task added to the run-queue preempts the current
4079 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4080 * called on the nearest possible occasion:
4082 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4084 * - in syscall or exception context, at the next outmost
4085 * preempt_enable(). (this might be as soon as the wake_up()'s
4088 * - in IRQ context, return from interrupt-handler to
4089 * preemptible context
4091 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4094 * - cond_resched() call
4095 * - explicit schedule() call
4096 * - return from syscall or exception to user-space
4097 * - return from interrupt-handler to user-space
4099 * WARNING: must be called with preemption disabled!
4101 static void __sched notrace __schedule(bool preempt)
4103 struct task_struct *prev, *next;
4104 unsigned long *switch_count;
4109 cpu = smp_processor_id();
4113 schedule_debug(prev, preempt);
4115 if (sched_feat(HRTICK))
4118 local_irq_disable();
4119 rcu_note_context_switch(preempt);
4122 * Make sure that signal_pending_state()->signal_pending() below
4123 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4124 * done by the caller to avoid the race with signal_wake_up().
4126 * The membarrier system call requires a full memory barrier
4127 * after coming from user-space, before storing to rq->curr.
4130 smp_mb__after_spinlock();
4132 /* Promote REQ to ACT */
4133 rq->clock_update_flags <<= 1;
4134 update_rq_clock(rq);
4136 switch_count = &prev->nivcsw;
4137 if (!preempt && prev->state) {
4138 if (signal_pending_state(prev->state, prev)) {
4139 prev->state = TASK_RUNNING;
4141 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4143 if (prev->in_iowait) {
4144 atomic_inc(&rq->nr_iowait);
4145 delayacct_blkio_start();
4148 switch_count = &prev->nvcsw;
4151 next = pick_next_task(rq, prev, &rf);
4152 clear_tsk_need_resched(prev);
4153 clear_preempt_need_resched();
4155 if (likely(prev != next)) {
4158 * RCU users of rcu_dereference(rq->curr) may not see
4159 * changes to task_struct made by pick_next_task().
4161 RCU_INIT_POINTER(rq->curr, next);
4163 * The membarrier system call requires each architecture
4164 * to have a full memory barrier after updating
4165 * rq->curr, before returning to user-space.
4167 * Here are the schemes providing that barrier on the
4168 * various architectures:
4169 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4170 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4171 * - finish_lock_switch() for weakly-ordered
4172 * architectures where spin_unlock is a full barrier,
4173 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4174 * is a RELEASE barrier),
4178 trace_sched_switch(preempt, prev, next);
4180 /* Also unlocks the rq: */
4181 rq = context_switch(rq, prev, next, &rf);
4183 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4184 rq_unlock_irq(rq, &rf);
4187 balance_callback(rq);
4190 void __noreturn do_task_dead(void)
4192 /* Causes final put_task_struct in finish_task_switch(): */
4193 set_special_state(TASK_DEAD);
4195 /* Tell freezer to ignore us: */
4196 current->flags |= PF_NOFREEZE;
4201 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4206 static inline void sched_submit_work(struct task_struct *tsk)
4212 * If a worker went to sleep, notify and ask workqueue whether
4213 * it wants to wake up a task to maintain concurrency.
4214 * As this function is called inside the schedule() context,
4215 * we disable preemption to avoid it calling schedule() again
4216 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4219 if (tsk->flags & PF_WQ_WORKER) {
4221 wq_worker_sleeping(tsk);
4222 preempt_enable_no_resched();
4225 if (tsk_is_pi_blocked(tsk))
4229 * If we are going to sleep and we have plugged IO queued,
4230 * make sure to submit it to avoid deadlocks.
4232 if (blk_needs_flush_plug(tsk))
4233 blk_schedule_flush_plug(tsk);
4236 static void sched_update_worker(struct task_struct *tsk)
4238 if (tsk->flags & PF_WQ_WORKER)
4239 wq_worker_running(tsk);
4242 asmlinkage __visible void __sched schedule(void)
4244 struct task_struct *tsk = current;
4246 sched_submit_work(tsk);
4250 sched_preempt_enable_no_resched();
4251 } while (need_resched());
4252 sched_update_worker(tsk);
4254 EXPORT_SYMBOL(schedule);
4257 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4258 * state (have scheduled out non-voluntarily) by making sure that all
4259 * tasks have either left the run queue or have gone into user space.
4260 * As idle tasks do not do either, they must not ever be preempted
4261 * (schedule out non-voluntarily).
4263 * schedule_idle() is similar to schedule_preempt_disable() except that it
4264 * never enables preemption because it does not call sched_submit_work().
4266 void __sched schedule_idle(void)
4269 * As this skips calling sched_submit_work(), which the idle task does
4270 * regardless because that function is a nop when the task is in a
4271 * TASK_RUNNING state, make sure this isn't used someplace that the
4272 * current task can be in any other state. Note, idle is always in the
4273 * TASK_RUNNING state.
4275 WARN_ON_ONCE(current->state);
4278 } while (need_resched());
4281 #ifdef CONFIG_CONTEXT_TRACKING
4282 asmlinkage __visible void __sched schedule_user(void)
4285 * If we come here after a random call to set_need_resched(),
4286 * or we have been woken up remotely but the IPI has not yet arrived,
4287 * we haven't yet exited the RCU idle mode. Do it here manually until
4288 * we find a better solution.
4290 * NB: There are buggy callers of this function. Ideally we
4291 * should warn if prev_state != CONTEXT_USER, but that will trigger
4292 * too frequently to make sense yet.
4294 enum ctx_state prev_state = exception_enter();
4296 exception_exit(prev_state);
4301 * schedule_preempt_disabled - called with preemption disabled
4303 * Returns with preemption disabled. Note: preempt_count must be 1
4305 void __sched schedule_preempt_disabled(void)
4307 sched_preempt_enable_no_resched();
4312 static void __sched notrace preempt_schedule_common(void)
4316 * Because the function tracer can trace preempt_count_sub()
4317 * and it also uses preempt_enable/disable_notrace(), if
4318 * NEED_RESCHED is set, the preempt_enable_notrace() called
4319 * by the function tracer will call this function again and
4320 * cause infinite recursion.
4322 * Preemption must be disabled here before the function
4323 * tracer can trace. Break up preempt_disable() into two
4324 * calls. One to disable preemption without fear of being
4325 * traced. The other to still record the preemption latency,
4326 * which can also be traced by the function tracer.
4328 preempt_disable_notrace();
4329 preempt_latency_start(1);
4331 preempt_latency_stop(1);
4332 preempt_enable_no_resched_notrace();
4335 * Check again in case we missed a preemption opportunity
4336 * between schedule and now.
4338 } while (need_resched());
4341 #ifdef CONFIG_PREEMPTION
4343 * This is the entry point to schedule() from in-kernel preemption
4344 * off of preempt_enable.
4346 asmlinkage __visible void __sched notrace preempt_schedule(void)
4349 * If there is a non-zero preempt_count or interrupts are disabled,
4350 * we do not want to preempt the current task. Just return..
4352 if (likely(!preemptible()))
4355 preempt_schedule_common();
4357 NOKPROBE_SYMBOL(preempt_schedule);
4358 EXPORT_SYMBOL(preempt_schedule);
4361 * preempt_schedule_notrace - preempt_schedule called by tracing
4363 * The tracing infrastructure uses preempt_enable_notrace to prevent
4364 * recursion and tracing preempt enabling caused by the tracing
4365 * infrastructure itself. But as tracing can happen in areas coming
4366 * from userspace or just about to enter userspace, a preempt enable
4367 * can occur before user_exit() is called. This will cause the scheduler
4368 * to be called when the system is still in usermode.
4370 * To prevent this, the preempt_enable_notrace will use this function
4371 * instead of preempt_schedule() to exit user context if needed before
4372 * calling the scheduler.
4374 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4376 enum ctx_state prev_ctx;
4378 if (likely(!preemptible()))
4383 * Because the function tracer can trace preempt_count_sub()
4384 * and it also uses preempt_enable/disable_notrace(), if
4385 * NEED_RESCHED is set, the preempt_enable_notrace() called
4386 * by the function tracer will call this function again and
4387 * cause infinite recursion.
4389 * Preemption must be disabled here before the function
4390 * tracer can trace. Break up preempt_disable() into two
4391 * calls. One to disable preemption without fear of being
4392 * traced. The other to still record the preemption latency,
4393 * which can also be traced by the function tracer.
4395 preempt_disable_notrace();
4396 preempt_latency_start(1);
4398 * Needs preempt disabled in case user_exit() is traced
4399 * and the tracer calls preempt_enable_notrace() causing
4400 * an infinite recursion.
4402 prev_ctx = exception_enter();
4404 exception_exit(prev_ctx);
4406 preempt_latency_stop(1);
4407 preempt_enable_no_resched_notrace();
4408 } while (need_resched());
4410 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4412 #endif /* CONFIG_PREEMPTION */
4415 * This is the entry point to schedule() from kernel preemption
4416 * off of irq context.
4417 * Note, that this is called and return with irqs disabled. This will
4418 * protect us against recursive calling from irq.
4420 asmlinkage __visible void __sched preempt_schedule_irq(void)
4422 enum ctx_state prev_state;
4424 /* Catch callers which need to be fixed */
4425 BUG_ON(preempt_count() || !irqs_disabled());
4427 prev_state = exception_enter();
4433 local_irq_disable();
4434 sched_preempt_enable_no_resched();
4435 } while (need_resched());
4437 exception_exit(prev_state);
4440 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4443 return try_to_wake_up(curr->private, mode, wake_flags);
4445 EXPORT_SYMBOL(default_wake_function);
4447 #ifdef CONFIG_RT_MUTEXES
4449 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4452 prio = min(prio, pi_task->prio);
4457 static inline int rt_effective_prio(struct task_struct *p, int prio)
4459 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4461 return __rt_effective_prio(pi_task, prio);
4465 * rt_mutex_setprio - set the current priority of a task
4467 * @pi_task: donor task
4469 * This function changes the 'effective' priority of a task. It does
4470 * not touch ->normal_prio like __setscheduler().
4472 * Used by the rt_mutex code to implement priority inheritance
4473 * logic. Call site only calls if the priority of the task changed.
4475 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4477 int prio, oldprio, queued, running, queue_flag =
4478 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4479 const struct sched_class *prev_class;
4483 /* XXX used to be waiter->prio, not waiter->task->prio */
4484 prio = __rt_effective_prio(pi_task, p->normal_prio);
4487 * If nothing changed; bail early.
4489 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4492 rq = __task_rq_lock(p, &rf);
4493 update_rq_clock(rq);
4495 * Set under pi_lock && rq->lock, such that the value can be used under
4498 * Note that there is loads of tricky to make this pointer cache work
4499 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4500 * ensure a task is de-boosted (pi_task is set to NULL) before the
4501 * task is allowed to run again (and can exit). This ensures the pointer
4502 * points to a blocked task -- which guaratees the task is present.
4504 p->pi_top_task = pi_task;
4507 * For FIFO/RR we only need to set prio, if that matches we're done.
4509 if (prio == p->prio && !dl_prio(prio))
4513 * Idle task boosting is a nono in general. There is one
4514 * exception, when PREEMPT_RT and NOHZ is active:
4516 * The idle task calls get_next_timer_interrupt() and holds
4517 * the timer wheel base->lock on the CPU and another CPU wants
4518 * to access the timer (probably to cancel it). We can safely
4519 * ignore the boosting request, as the idle CPU runs this code
4520 * with interrupts disabled and will complete the lock
4521 * protected section without being interrupted. So there is no
4522 * real need to boost.
4524 if (unlikely(p == rq->idle)) {
4525 WARN_ON(p != rq->curr);
4526 WARN_ON(p->pi_blocked_on);
4530 trace_sched_pi_setprio(p, pi_task);
4533 if (oldprio == prio)
4534 queue_flag &= ~DEQUEUE_MOVE;
4536 prev_class = p->sched_class;
4537 queued = task_on_rq_queued(p);
4538 running = task_current(rq, p);
4540 dequeue_task(rq, p, queue_flag);
4542 put_prev_task(rq, p);
4545 * Boosting condition are:
4546 * 1. -rt task is running and holds mutex A
4547 * --> -dl task blocks on mutex A
4549 * 2. -dl task is running and holds mutex A
4550 * --> -dl task blocks on mutex A and could preempt the
4553 if (dl_prio(prio)) {
4554 if (!dl_prio(p->normal_prio) ||
4555 (pi_task && dl_prio(pi_task->prio) &&
4556 dl_entity_preempt(&pi_task->dl, &p->dl))) {
4557 p->dl.dl_boosted = 1;
4558 queue_flag |= ENQUEUE_REPLENISH;
4560 p->dl.dl_boosted = 0;
4561 p->sched_class = &dl_sched_class;
4562 } else if (rt_prio(prio)) {
4563 if (dl_prio(oldprio))
4564 p->dl.dl_boosted = 0;
4566 queue_flag |= ENQUEUE_HEAD;
4567 p->sched_class = &rt_sched_class;
4569 if (dl_prio(oldprio))
4570 p->dl.dl_boosted = 0;
4571 if (rt_prio(oldprio))
4573 p->sched_class = &fair_sched_class;
4579 enqueue_task(rq, p, queue_flag);
4581 set_next_task(rq, p);
4583 check_class_changed(rq, p, prev_class, oldprio);
4585 /* Avoid rq from going away on us: */
4587 __task_rq_unlock(rq, &rf);
4589 balance_callback(rq);
4593 static inline int rt_effective_prio(struct task_struct *p, int prio)
4599 void set_user_nice(struct task_struct *p, long nice)
4601 bool queued, running;
4602 int old_prio, delta;
4606 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4609 * We have to be careful, if called from sys_setpriority(),
4610 * the task might be in the middle of scheduling on another CPU.
4612 rq = task_rq_lock(p, &rf);
4613 update_rq_clock(rq);
4616 * The RT priorities are set via sched_setscheduler(), but we still
4617 * allow the 'normal' nice value to be set - but as expected
4618 * it wont have any effect on scheduling until the task is
4619 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4621 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4622 p->static_prio = NICE_TO_PRIO(nice);
4625 queued = task_on_rq_queued(p);
4626 running = task_current(rq, p);
4628 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4630 put_prev_task(rq, p);
4632 p->static_prio = NICE_TO_PRIO(nice);
4633 set_load_weight(p, true);
4635 p->prio = effective_prio(p);
4636 delta = p->prio - old_prio;
4639 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4641 * If the task increased its priority or is running and
4642 * lowered its priority, then reschedule its CPU:
4644 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4648 set_next_task(rq, p);
4650 task_rq_unlock(rq, p, &rf);
4652 EXPORT_SYMBOL(set_user_nice);
4655 * can_nice - check if a task can reduce its nice value
4659 int can_nice(const struct task_struct *p, const int nice)
4661 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
4662 int nice_rlim = nice_to_rlimit(nice);
4664 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4665 capable(CAP_SYS_NICE));
4668 #ifdef __ARCH_WANT_SYS_NICE
4671 * sys_nice - change the priority of the current process.
4672 * @increment: priority increment
4674 * sys_setpriority is a more generic, but much slower function that
4675 * does similar things.
4677 SYSCALL_DEFINE1(nice, int, increment)
4682 * Setpriority might change our priority at the same moment.
4683 * We don't have to worry. Conceptually one call occurs first
4684 * and we have a single winner.
4686 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4687 nice = task_nice(current) + increment;
4689 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4690 if (increment < 0 && !can_nice(current, nice))
4693 retval = security_task_setnice(current, nice);
4697 set_user_nice(current, nice);
4704 * task_prio - return the priority value of a given task.
4705 * @p: the task in question.
4707 * Return: The priority value as seen by users in /proc.
4708 * RT tasks are offset by -200. Normal tasks are centered
4709 * around 0, value goes from -16 to +15.
4711 int task_prio(const struct task_struct *p)
4713 return p->prio - MAX_RT_PRIO;
4717 * idle_cpu - is a given CPU idle currently?
4718 * @cpu: the processor in question.
4720 * Return: 1 if the CPU is currently idle. 0 otherwise.
4722 int idle_cpu(int cpu)
4724 struct rq *rq = cpu_rq(cpu);
4726 if (rq->curr != rq->idle)
4733 if (!llist_empty(&rq->wake_list))
4741 * available_idle_cpu - is a given CPU idle for enqueuing work.
4742 * @cpu: the CPU in question.
4744 * Return: 1 if the CPU is currently idle. 0 otherwise.
4746 int available_idle_cpu(int cpu)
4751 if (vcpu_is_preempted(cpu))
4758 * idle_task - return the idle task for a given CPU.
4759 * @cpu: the processor in question.
4761 * Return: The idle task for the CPU @cpu.
4763 struct task_struct *idle_task(int cpu)
4765 return cpu_rq(cpu)->idle;
4769 * find_process_by_pid - find a process with a matching PID value.
4770 * @pid: the pid in question.
4772 * The task of @pid, if found. %NULL otherwise.
4774 static struct task_struct *find_process_by_pid(pid_t pid)
4776 return pid ? find_task_by_vpid(pid) : current;
4780 * sched_setparam() passes in -1 for its policy, to let the functions
4781 * it calls know not to change it.
4783 #define SETPARAM_POLICY -1
4785 static void __setscheduler_params(struct task_struct *p,
4786 const struct sched_attr *attr)
4788 int policy = attr->sched_policy;
4790 if (policy == SETPARAM_POLICY)
4795 if (dl_policy(policy))
4796 __setparam_dl(p, attr);
4797 else if (fair_policy(policy))
4798 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4801 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4802 * !rt_policy. Always setting this ensures that things like
4803 * getparam()/getattr() don't report silly values for !rt tasks.
4805 p->rt_priority = attr->sched_priority;
4806 p->normal_prio = normal_prio(p);
4807 set_load_weight(p, true);
4810 /* Actually do priority change: must hold pi & rq lock. */
4811 static void __setscheduler(struct rq *rq, struct task_struct *p,
4812 const struct sched_attr *attr, bool keep_boost)
4815 * If params can't change scheduling class changes aren't allowed
4818 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
4821 __setscheduler_params(p, attr);
4824 * Keep a potential priority boosting if called from
4825 * sched_setscheduler().
4827 p->prio = normal_prio(p);
4829 p->prio = rt_effective_prio(p, p->prio);
4831 if (dl_prio(p->prio))
4832 p->sched_class = &dl_sched_class;
4833 else if (rt_prio(p->prio))
4834 p->sched_class = &rt_sched_class;
4836 p->sched_class = &fair_sched_class;
4840 * Check the target process has a UID that matches the current process's:
4842 static bool check_same_owner(struct task_struct *p)
4844 const struct cred *cred = current_cred(), *pcred;
4848 pcred = __task_cred(p);
4849 match = (uid_eq(cred->euid, pcred->euid) ||
4850 uid_eq(cred->euid, pcred->uid));
4855 static int __sched_setscheduler(struct task_struct *p,
4856 const struct sched_attr *attr,
4859 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4860 MAX_RT_PRIO - 1 - attr->sched_priority;
4861 int retval, oldprio, oldpolicy = -1, queued, running;
4862 int new_effective_prio, policy = attr->sched_policy;
4863 const struct sched_class *prev_class;
4866 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4869 /* The pi code expects interrupts enabled */
4870 BUG_ON(pi && in_interrupt());
4872 /* Double check policy once rq lock held: */
4874 reset_on_fork = p->sched_reset_on_fork;
4875 policy = oldpolicy = p->policy;
4877 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4879 if (!valid_policy(policy))
4883 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4887 * Valid priorities for SCHED_FIFO and SCHED_RR are
4888 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4889 * SCHED_BATCH and SCHED_IDLE is 0.
4891 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4892 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4894 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4895 (rt_policy(policy) != (attr->sched_priority != 0)))
4899 * Allow unprivileged RT tasks to decrease priority:
4901 if (user && !capable(CAP_SYS_NICE)) {
4902 if (fair_policy(policy)) {
4903 if (attr->sched_nice < task_nice(p) &&
4904 !can_nice(p, attr->sched_nice))
4908 if (rt_policy(policy)) {
4909 unsigned long rlim_rtprio =
4910 task_rlimit(p, RLIMIT_RTPRIO);
4912 /* Can't set/change the rt policy: */
4913 if (policy != p->policy && !rlim_rtprio)
4916 /* Can't increase priority: */
4917 if (attr->sched_priority > p->rt_priority &&
4918 attr->sched_priority > rlim_rtprio)
4923 * Can't set/change SCHED_DEADLINE policy at all for now
4924 * (safest behavior); in the future we would like to allow
4925 * unprivileged DL tasks to increase their relative deadline
4926 * or reduce their runtime (both ways reducing utilization)
4928 if (dl_policy(policy))
4932 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4933 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4935 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4936 if (!can_nice(p, task_nice(p)))
4940 /* Can't change other user's priorities: */
4941 if (!check_same_owner(p))
4944 /* Normal users shall not reset the sched_reset_on_fork flag: */
4945 if (p->sched_reset_on_fork && !reset_on_fork)
4950 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4953 retval = security_task_setscheduler(p);
4958 /* Update task specific "requested" clamps */
4959 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
4960 retval = uclamp_validate(p, attr);
4969 * Make sure no PI-waiters arrive (or leave) while we are
4970 * changing the priority of the task:
4972 * To be able to change p->policy safely, the appropriate
4973 * runqueue lock must be held.
4975 rq = task_rq_lock(p, &rf);
4976 update_rq_clock(rq);
4979 * Changing the policy of the stop threads its a very bad idea:
4981 if (p == rq->stop) {
4987 * If not changing anything there's no need to proceed further,
4988 * but store a possible modification of reset_on_fork.
4990 if (unlikely(policy == p->policy)) {
4991 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4993 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4995 if (dl_policy(policy) && dl_param_changed(p, attr))
4997 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
5000 p->sched_reset_on_fork = reset_on_fork;
5007 #ifdef CONFIG_RT_GROUP_SCHED
5009 * Do not allow realtime tasks into groups that have no runtime
5012 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5013 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5014 !task_group_is_autogroup(task_group(p))) {
5020 if (dl_bandwidth_enabled() && dl_policy(policy) &&
5021 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
5022 cpumask_t *span = rq->rd->span;
5025 * Don't allow tasks with an affinity mask smaller than
5026 * the entire root_domain to become SCHED_DEADLINE. We
5027 * will also fail if there's no bandwidth available.
5029 if (!cpumask_subset(span, p->cpus_ptr) ||
5030 rq->rd->dl_bw.bw == 0) {
5038 /* Re-check policy now with rq lock held: */
5039 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5040 policy = oldpolicy = -1;
5041 task_rq_unlock(rq, p, &rf);
5043 cpuset_read_unlock();
5048 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5049 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5052 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
5057 p->sched_reset_on_fork = reset_on_fork;
5062 * Take priority boosted tasks into account. If the new
5063 * effective priority is unchanged, we just store the new
5064 * normal parameters and do not touch the scheduler class and
5065 * the runqueue. This will be done when the task deboost
5068 new_effective_prio = rt_effective_prio(p, newprio);
5069 if (new_effective_prio == oldprio)
5070 queue_flags &= ~DEQUEUE_MOVE;
5073 queued = task_on_rq_queued(p);
5074 running = task_current(rq, p);
5076 dequeue_task(rq, p, queue_flags);
5078 put_prev_task(rq, p);
5080 prev_class = p->sched_class;
5082 __setscheduler(rq, p, attr, pi);
5083 __setscheduler_uclamp(p, attr);
5087 * We enqueue to tail when the priority of a task is
5088 * increased (user space view).
5090 if (oldprio < p->prio)
5091 queue_flags |= ENQUEUE_HEAD;
5093 enqueue_task(rq, p, queue_flags);
5096 set_next_task(rq, p);
5098 check_class_changed(rq, p, prev_class, oldprio);
5100 /* Avoid rq from going away on us: */
5102 task_rq_unlock(rq, p, &rf);
5105 cpuset_read_unlock();
5106 rt_mutex_adjust_pi(p);
5109 /* Run balance callbacks after we've adjusted the PI chain: */
5110 balance_callback(rq);
5116 task_rq_unlock(rq, p, &rf);
5118 cpuset_read_unlock();
5122 static int _sched_setscheduler(struct task_struct *p, int policy,
5123 const struct sched_param *param, bool check)
5125 struct sched_attr attr = {
5126 .sched_policy = policy,
5127 .sched_priority = param->sched_priority,
5128 .sched_nice = PRIO_TO_NICE(p->static_prio),
5131 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5132 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5133 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5134 policy &= ~SCHED_RESET_ON_FORK;
5135 attr.sched_policy = policy;
5138 return __sched_setscheduler(p, &attr, check, true);
5141 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5142 * @p: the task in question.
5143 * @policy: new policy.
5144 * @param: structure containing the new RT priority.
5146 * Return: 0 on success. An error code otherwise.
5148 * NOTE that the task may be already dead.
5150 int sched_setscheduler(struct task_struct *p, int policy,
5151 const struct sched_param *param)
5153 return _sched_setscheduler(p, policy, param, true);
5155 EXPORT_SYMBOL_GPL(sched_setscheduler);
5157 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5159 return __sched_setscheduler(p, attr, true, true);
5161 EXPORT_SYMBOL_GPL(sched_setattr);
5163 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5165 return __sched_setscheduler(p, attr, false, true);
5169 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5170 * @p: the task in question.
5171 * @policy: new policy.
5172 * @param: structure containing the new RT priority.
5174 * Just like sched_setscheduler, only don't bother checking if the
5175 * current context has permission. For example, this is needed in
5176 * stop_machine(): we create temporary high priority worker threads,
5177 * but our caller might not have that capability.
5179 * Return: 0 on success. An error code otherwise.
5181 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5182 const struct sched_param *param)
5184 return _sched_setscheduler(p, policy, param, false);
5186 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
5189 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5191 struct sched_param lparam;
5192 struct task_struct *p;
5195 if (!param || pid < 0)
5197 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5202 p = find_process_by_pid(pid);
5208 retval = sched_setscheduler(p, policy, &lparam);
5216 * Mimics kernel/events/core.c perf_copy_attr().
5218 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5223 /* Zero the full structure, so that a short copy will be nice: */
5224 memset(attr, 0, sizeof(*attr));
5226 ret = get_user(size, &uattr->size);
5230 /* ABI compatibility quirk: */
5232 size = SCHED_ATTR_SIZE_VER0;
5233 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5236 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5243 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5244 size < SCHED_ATTR_SIZE_VER1)
5248 * XXX: Do we want to be lenient like existing syscalls; or do we want
5249 * to be strict and return an error on out-of-bounds values?
5251 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5256 put_user(sizeof(*attr), &uattr->size);
5261 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5262 * @pid: the pid in question.
5263 * @policy: new policy.
5264 * @param: structure containing the new RT priority.
5266 * Return: 0 on success. An error code otherwise.
5268 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5273 return do_sched_setscheduler(pid, policy, param);
5277 * sys_sched_setparam - set/change the RT priority of a thread
5278 * @pid: the pid in question.
5279 * @param: structure containing the new RT priority.
5281 * Return: 0 on success. An error code otherwise.
5283 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5285 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5289 * sys_sched_setattr - same as above, but with extended sched_attr
5290 * @pid: the pid in question.
5291 * @uattr: structure containing the extended parameters.
5292 * @flags: for future extension.
5294 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5295 unsigned int, flags)
5297 struct sched_attr attr;
5298 struct task_struct *p;
5301 if (!uattr || pid < 0 || flags)
5304 retval = sched_copy_attr(uattr, &attr);
5308 if ((int)attr.sched_policy < 0)
5310 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5311 attr.sched_policy = SETPARAM_POLICY;
5315 p = find_process_by_pid(pid);
5321 retval = sched_setattr(p, &attr);
5329 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5330 * @pid: the pid in question.
5332 * Return: On success, the policy of the thread. Otherwise, a negative error
5335 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5337 struct task_struct *p;
5345 p = find_process_by_pid(pid);
5347 retval = security_task_getscheduler(p);
5350 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5357 * sys_sched_getparam - get the RT priority of a thread
5358 * @pid: the pid in question.
5359 * @param: structure containing the RT priority.
5361 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5364 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5366 struct sched_param lp = { .sched_priority = 0 };
5367 struct task_struct *p;
5370 if (!param || pid < 0)
5374 p = find_process_by_pid(pid);
5379 retval = security_task_getscheduler(p);
5383 if (task_has_rt_policy(p))
5384 lp.sched_priority = p->rt_priority;
5388 * This one might sleep, we cannot do it with a spinlock held ...
5390 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5400 * Copy the kernel size attribute structure (which might be larger
5401 * than what user-space knows about) to user-space.
5403 * Note that all cases are valid: user-space buffer can be larger or
5404 * smaller than the kernel-space buffer. The usual case is that both
5405 * have the same size.
5408 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5409 struct sched_attr *kattr,
5412 unsigned int ksize = sizeof(*kattr);
5414 if (!access_ok(uattr, usize))
5418 * sched_getattr() ABI forwards and backwards compatibility:
5420 * If usize == ksize then we just copy everything to user-space and all is good.
5422 * If usize < ksize then we only copy as much as user-space has space for,
5423 * this keeps ABI compatibility as well. We skip the rest.
5425 * If usize > ksize then user-space is using a newer version of the ABI,
5426 * which part the kernel doesn't know about. Just ignore it - tooling can
5427 * detect the kernel's knowledge of attributes from the attr->size value
5428 * which is set to ksize in this case.
5430 kattr->size = min(usize, ksize);
5432 if (copy_to_user(uattr, kattr, kattr->size))
5439 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5440 * @pid: the pid in question.
5441 * @uattr: structure containing the extended parameters.
5442 * @usize: sizeof(attr) for fwd/bwd comp.
5443 * @flags: for future extension.
5445 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5446 unsigned int, usize, unsigned int, flags)
5448 struct sched_attr kattr = { };
5449 struct task_struct *p;
5452 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5453 usize < SCHED_ATTR_SIZE_VER0 || flags)
5457 p = find_process_by_pid(pid);
5462 retval = security_task_getscheduler(p);
5466 kattr.sched_policy = p->policy;
5467 if (p->sched_reset_on_fork)
5468 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5469 if (task_has_dl_policy(p))
5470 __getparam_dl(p, &kattr);
5471 else if (task_has_rt_policy(p))
5472 kattr.sched_priority = p->rt_priority;
5474 kattr.sched_nice = task_nice(p);
5476 #ifdef CONFIG_UCLAMP_TASK
5477 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5478 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5483 return sched_attr_copy_to_user(uattr, &kattr, usize);
5490 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5492 cpumask_var_t cpus_allowed, new_mask;
5493 struct task_struct *p;
5498 p = find_process_by_pid(pid);
5504 /* Prevent p going away */
5508 if (p->flags & PF_NO_SETAFFINITY) {
5512 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5516 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5518 goto out_free_cpus_allowed;
5521 if (!check_same_owner(p)) {
5523 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5525 goto out_free_new_mask;
5530 retval = security_task_setscheduler(p);
5532 goto out_free_new_mask;
5535 cpuset_cpus_allowed(p, cpus_allowed);
5536 cpumask_and(new_mask, in_mask, cpus_allowed);
5539 * Since bandwidth control happens on root_domain basis,
5540 * if admission test is enabled, we only admit -deadline
5541 * tasks allowed to run on all the CPUs in the task's
5545 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5547 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5550 goto out_free_new_mask;
5556 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5559 cpuset_cpus_allowed(p, cpus_allowed);
5560 if (!cpumask_subset(new_mask, cpus_allowed)) {
5562 * We must have raced with a concurrent cpuset
5563 * update. Just reset the cpus_allowed to the
5564 * cpuset's cpus_allowed
5566 cpumask_copy(new_mask, cpus_allowed);
5571 free_cpumask_var(new_mask);
5572 out_free_cpus_allowed:
5573 free_cpumask_var(cpus_allowed);
5579 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5580 struct cpumask *new_mask)
5582 if (len < cpumask_size())
5583 cpumask_clear(new_mask);
5584 else if (len > cpumask_size())
5585 len = cpumask_size();
5587 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5591 * sys_sched_setaffinity - set the CPU affinity of a process
5592 * @pid: pid of the process
5593 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5594 * @user_mask_ptr: user-space pointer to the new CPU mask
5596 * Return: 0 on success. An error code otherwise.
5598 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5599 unsigned long __user *, user_mask_ptr)
5601 cpumask_var_t new_mask;
5604 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5607 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5609 retval = sched_setaffinity(pid, new_mask);
5610 free_cpumask_var(new_mask);
5614 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5616 struct task_struct *p;
5617 unsigned long flags;
5623 p = find_process_by_pid(pid);
5627 retval = security_task_getscheduler(p);
5631 raw_spin_lock_irqsave(&p->pi_lock, flags);
5632 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
5633 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5642 * sys_sched_getaffinity - get the CPU affinity of a process
5643 * @pid: pid of the process
5644 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5645 * @user_mask_ptr: user-space pointer to hold the current CPU mask
5647 * Return: size of CPU mask copied to user_mask_ptr on success. An
5648 * error code otherwise.
5650 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5651 unsigned long __user *, user_mask_ptr)
5656 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5658 if (len & (sizeof(unsigned long)-1))
5661 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5664 ret = sched_getaffinity(pid, mask);
5666 unsigned int retlen = min(len, cpumask_size());
5668 if (copy_to_user(user_mask_ptr, mask, retlen))
5673 free_cpumask_var(mask);
5679 * sys_sched_yield - yield the current processor to other threads.
5681 * This function yields the current CPU to other tasks. If there are no
5682 * other threads running on this CPU then this function will return.
5686 static void do_sched_yield(void)
5691 rq = this_rq_lock_irq(&rf);
5693 schedstat_inc(rq->yld_count);
5694 current->sched_class->yield_task(rq);
5697 rq_unlock_irq(rq, &rf);
5698 sched_preempt_enable_no_resched();
5703 SYSCALL_DEFINE0(sched_yield)
5709 #ifndef CONFIG_PREEMPTION
5710 int __sched _cond_resched(void)
5712 if (should_resched(0)) {
5713 preempt_schedule_common();
5719 EXPORT_SYMBOL(_cond_resched);
5723 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5724 * call schedule, and on return reacquire the lock.
5726 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
5727 * operations here to prevent schedule() from being called twice (once via
5728 * spin_unlock(), once by hand).
5730 int __cond_resched_lock(spinlock_t *lock)
5732 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5735 lockdep_assert_held(lock);
5737 if (spin_needbreak(lock) || resched) {
5740 preempt_schedule_common();
5748 EXPORT_SYMBOL(__cond_resched_lock);
5751 * yield - yield the current processor to other threads.
5753 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5755 * The scheduler is at all times free to pick the calling task as the most
5756 * eligible task to run, if removing the yield() call from your code breaks
5757 * it, its already broken.
5759 * Typical broken usage is:
5764 * where one assumes that yield() will let 'the other' process run that will
5765 * make event true. If the current task is a SCHED_FIFO task that will never
5766 * happen. Never use yield() as a progress guarantee!!
5768 * If you want to use yield() to wait for something, use wait_event().
5769 * If you want to use yield() to be 'nice' for others, use cond_resched().
5770 * If you still want to use yield(), do not!
5772 void __sched yield(void)
5774 set_current_state(TASK_RUNNING);
5777 EXPORT_SYMBOL(yield);
5780 * yield_to - yield the current processor to another thread in
5781 * your thread group, or accelerate that thread toward the
5782 * processor it's on.
5784 * @preempt: whether task preemption is allowed or not
5786 * It's the caller's job to ensure that the target task struct
5787 * can't go away on us before we can do any checks.
5790 * true (>0) if we indeed boosted the target task.
5791 * false (0) if we failed to boost the target.
5792 * -ESRCH if there's no task to yield to.
5794 int __sched yield_to(struct task_struct *p, bool preempt)
5796 struct task_struct *curr = current;
5797 struct rq *rq, *p_rq;
5798 unsigned long flags;
5801 local_irq_save(flags);
5807 * If we're the only runnable task on the rq and target rq also
5808 * has only one task, there's absolutely no point in yielding.
5810 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5815 double_rq_lock(rq, p_rq);
5816 if (task_rq(p) != p_rq) {
5817 double_rq_unlock(rq, p_rq);
5821 if (!curr->sched_class->yield_to_task)
5824 if (curr->sched_class != p->sched_class)
5827 if (task_running(p_rq, p) || p->state)
5830 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5832 schedstat_inc(rq->yld_count);
5834 * Make p's CPU reschedule; pick_next_entity takes care of
5837 if (preempt && rq != p_rq)
5842 double_rq_unlock(rq, p_rq);
5844 local_irq_restore(flags);
5851 EXPORT_SYMBOL_GPL(yield_to);
5853 int io_schedule_prepare(void)
5855 int old_iowait = current->in_iowait;
5857 current->in_iowait = 1;
5858 blk_schedule_flush_plug(current);
5863 void io_schedule_finish(int token)
5865 current->in_iowait = token;
5869 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5870 * that process accounting knows that this is a task in IO wait state.
5872 long __sched io_schedule_timeout(long timeout)
5877 token = io_schedule_prepare();
5878 ret = schedule_timeout(timeout);
5879 io_schedule_finish(token);
5883 EXPORT_SYMBOL(io_schedule_timeout);
5885 void __sched io_schedule(void)
5889 token = io_schedule_prepare();
5891 io_schedule_finish(token);
5893 EXPORT_SYMBOL(io_schedule);
5896 * sys_sched_get_priority_max - return maximum RT priority.
5897 * @policy: scheduling class.
5899 * Return: On success, this syscall returns the maximum
5900 * rt_priority that can be used by a given scheduling class.
5901 * On failure, a negative error code is returned.
5903 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5910 ret = MAX_USER_RT_PRIO-1;
5912 case SCHED_DEADLINE:
5923 * sys_sched_get_priority_min - return minimum RT priority.
5924 * @policy: scheduling class.
5926 * Return: On success, this syscall returns the minimum
5927 * rt_priority that can be used by a given scheduling class.
5928 * On failure, a negative error code is returned.
5930 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5939 case SCHED_DEADLINE:
5948 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5950 struct task_struct *p;
5951 unsigned int time_slice;
5961 p = find_process_by_pid(pid);
5965 retval = security_task_getscheduler(p);
5969 rq = task_rq_lock(p, &rf);
5971 if (p->sched_class->get_rr_interval)
5972 time_slice = p->sched_class->get_rr_interval(rq, p);
5973 task_rq_unlock(rq, p, &rf);
5976 jiffies_to_timespec64(time_slice, t);
5985 * sys_sched_rr_get_interval - return the default timeslice of a process.
5986 * @pid: pid of the process.
5987 * @interval: userspace pointer to the timeslice value.
5989 * this syscall writes the default timeslice value of a given process
5990 * into the user-space timespec buffer. A value of '0' means infinity.
5992 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5995 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5996 struct __kernel_timespec __user *, interval)
5998 struct timespec64 t;
5999 int retval = sched_rr_get_interval(pid, &t);
6002 retval = put_timespec64(&t, interval);
6007 #ifdef CONFIG_COMPAT_32BIT_TIME
6008 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
6009 struct old_timespec32 __user *, interval)
6011 struct timespec64 t;
6012 int retval = sched_rr_get_interval(pid, &t);
6015 retval = put_old_timespec32(&t, interval);
6020 void sched_show_task(struct task_struct *p)
6022 unsigned long free = 0;
6025 if (!try_get_task_stack(p))
6028 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
6030 if (p->state == TASK_RUNNING)
6031 printk(KERN_CONT " running task ");
6032 #ifdef CONFIG_DEBUG_STACK_USAGE
6033 free = stack_not_used(p);
6038 ppid = task_pid_nr(rcu_dereference(p->real_parent));
6040 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6041 task_pid_nr(p), ppid,
6042 (unsigned long)task_thread_info(p)->flags);
6044 print_worker_info(KERN_INFO, p);
6045 show_stack(p, NULL);
6048 EXPORT_SYMBOL_GPL(sched_show_task);
6051 state_filter_match(unsigned long state_filter, struct task_struct *p)
6053 /* no filter, everything matches */
6057 /* filter, but doesn't match */
6058 if (!(p->state & state_filter))
6062 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6065 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
6072 void show_state_filter(unsigned long state_filter)
6074 struct task_struct *g, *p;
6076 #if BITS_PER_LONG == 32
6078 " task PC stack pid father\n");
6081 " task PC stack pid father\n");
6084 for_each_process_thread(g, p) {
6086 * reset the NMI-timeout, listing all files on a slow
6087 * console might take a lot of time:
6088 * Also, reset softlockup watchdogs on all CPUs, because
6089 * another CPU might be blocked waiting for us to process
6092 touch_nmi_watchdog();
6093 touch_all_softlockup_watchdogs();
6094 if (state_filter_match(state_filter, p))
6098 #ifdef CONFIG_SCHED_DEBUG
6100 sysrq_sched_debug_show();
6104 * Only show locks if all tasks are dumped:
6107 debug_show_all_locks();
6111 * init_idle - set up an idle thread for a given CPU
6112 * @idle: task in question
6113 * @cpu: CPU the idle task belongs to
6115 * NOTE: this function does not set the idle thread's NEED_RESCHED
6116 * flag, to make booting more robust.
6118 void init_idle(struct task_struct *idle, int cpu)
6120 struct rq *rq = cpu_rq(cpu);
6121 unsigned long flags;
6123 __sched_fork(0, idle);
6125 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6126 raw_spin_lock(&rq->lock);
6128 idle->state = TASK_RUNNING;
6129 idle->se.exec_start = sched_clock();
6130 idle->flags |= PF_IDLE;
6132 kasan_unpoison_task_stack(idle);
6136 * Its possible that init_idle() gets called multiple times on a task,
6137 * in that case do_set_cpus_allowed() will not do the right thing.
6139 * And since this is boot we can forgo the serialization.
6141 set_cpus_allowed_common(idle, cpumask_of(cpu));
6144 * We're having a chicken and egg problem, even though we are
6145 * holding rq->lock, the CPU isn't yet set to this CPU so the
6146 * lockdep check in task_group() will fail.
6148 * Similar case to sched_fork(). / Alternatively we could
6149 * use task_rq_lock() here and obtain the other rq->lock.
6154 __set_task_cpu(idle, cpu);
6158 rcu_assign_pointer(rq->curr, idle);
6159 idle->on_rq = TASK_ON_RQ_QUEUED;
6163 raw_spin_unlock(&rq->lock);
6164 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6166 /* Set the preempt count _outside_ the spinlocks! */
6167 init_idle_preempt_count(idle, cpu);
6170 * The idle tasks have their own, simple scheduling class:
6172 idle->sched_class = &idle_sched_class;
6173 ftrace_graph_init_idle_task(idle, cpu);
6174 vtime_init_idle(idle, cpu);
6176 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6182 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6183 const struct cpumask *trial)
6187 if (!cpumask_weight(cur))
6190 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6195 int task_can_attach(struct task_struct *p,
6196 const struct cpumask *cs_cpus_allowed)
6201 * Kthreads which disallow setaffinity shouldn't be moved
6202 * to a new cpuset; we don't want to change their CPU
6203 * affinity and isolating such threads by their set of
6204 * allowed nodes is unnecessary. Thus, cpusets are not
6205 * applicable for such threads. This prevents checking for
6206 * success of set_cpus_allowed_ptr() on all attached tasks
6207 * before cpus_mask may be changed.
6209 if (p->flags & PF_NO_SETAFFINITY) {
6214 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6216 ret = dl_task_can_attach(p, cs_cpus_allowed);
6222 bool sched_smp_initialized __read_mostly;
6224 #ifdef CONFIG_NUMA_BALANCING
6225 /* Migrate current task p to target_cpu */
6226 int migrate_task_to(struct task_struct *p, int target_cpu)
6228 struct migration_arg arg = { p, target_cpu };
6229 int curr_cpu = task_cpu(p);
6231 if (curr_cpu == target_cpu)
6234 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6237 /* TODO: This is not properly updating schedstats */
6239 trace_sched_move_numa(p, curr_cpu, target_cpu);
6240 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6244 * Requeue a task on a given node and accurately track the number of NUMA
6245 * tasks on the runqueues
6247 void sched_setnuma(struct task_struct *p, int nid)
6249 bool queued, running;
6253 rq = task_rq_lock(p, &rf);
6254 queued = task_on_rq_queued(p);
6255 running = task_current(rq, p);
6258 dequeue_task(rq, p, DEQUEUE_SAVE);
6260 put_prev_task(rq, p);
6262 p->numa_preferred_nid = nid;
6265 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6267 set_next_task(rq, p);
6268 task_rq_unlock(rq, p, &rf);
6270 #endif /* CONFIG_NUMA_BALANCING */
6272 #ifdef CONFIG_HOTPLUG_CPU
6274 * Ensure that the idle task is using init_mm right before its CPU goes
6277 void idle_task_exit(void)
6279 struct mm_struct *mm = current->active_mm;
6281 BUG_ON(cpu_online(smp_processor_id()));
6282 BUG_ON(current != this_rq()->idle);
6284 if (mm != &init_mm) {
6285 switch_mm(mm, &init_mm, current);
6286 finish_arch_post_lock_switch();
6289 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6293 * Since this CPU is going 'away' for a while, fold any nr_active delta
6294 * we might have. Assumes we're called after migrate_tasks() so that the
6295 * nr_active count is stable. We need to take the teardown thread which
6296 * is calling this into account, so we hand in adjust = 1 to the load
6299 * Also see the comment "Global load-average calculations".
6301 static void calc_load_migrate(struct rq *rq)
6303 long delta = calc_load_fold_active(rq, 1);
6305 atomic_long_add(delta, &calc_load_tasks);
6308 static struct task_struct *__pick_migrate_task(struct rq *rq)
6310 const struct sched_class *class;
6311 struct task_struct *next;
6313 for_each_class(class) {
6314 next = class->pick_next_task(rq, NULL, NULL);
6316 next->sched_class->put_prev_task(rq, next);
6321 /* The idle class should always have a runnable task */
6326 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6327 * try_to_wake_up()->select_task_rq().
6329 * Called with rq->lock held even though we'er in stop_machine() and
6330 * there's no concurrency possible, we hold the required locks anyway
6331 * because of lock validation efforts.
6333 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6335 struct rq *rq = dead_rq;
6336 struct task_struct *next, *stop = rq->stop;
6337 struct rq_flags orf = *rf;
6341 * Fudge the rq selection such that the below task selection loop
6342 * doesn't get stuck on the currently eligible stop task.
6344 * We're currently inside stop_machine() and the rq is either stuck
6345 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6346 * either way we should never end up calling schedule() until we're
6352 * put_prev_task() and pick_next_task() sched
6353 * class method both need to have an up-to-date
6354 * value of rq->clock[_task]
6356 update_rq_clock(rq);
6360 * There's this thread running, bail when that's the only
6363 if (rq->nr_running == 1)
6366 next = __pick_migrate_task(rq);
6369 * Rules for changing task_struct::cpus_mask are holding
6370 * both pi_lock and rq->lock, such that holding either
6371 * stabilizes the mask.
6373 * Drop rq->lock is not quite as disastrous as it usually is
6374 * because !cpu_active at this point, which means load-balance
6375 * will not interfere. Also, stop-machine.
6378 raw_spin_lock(&next->pi_lock);
6382 * Since we're inside stop-machine, _nothing_ should have
6383 * changed the task, WARN if weird stuff happened, because in
6384 * that case the above rq->lock drop is a fail too.
6386 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6387 raw_spin_unlock(&next->pi_lock);
6391 /* Find suitable destination for @next, with force if needed. */
6392 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6393 rq = __migrate_task(rq, rf, next, dest_cpu);
6394 if (rq != dead_rq) {
6400 raw_spin_unlock(&next->pi_lock);
6405 #endif /* CONFIG_HOTPLUG_CPU */
6407 void set_rq_online(struct rq *rq)
6410 const struct sched_class *class;
6412 cpumask_set_cpu(rq->cpu, rq->rd->online);
6415 for_each_class(class) {
6416 if (class->rq_online)
6417 class->rq_online(rq);
6422 void set_rq_offline(struct rq *rq)
6425 const struct sched_class *class;
6427 for_each_class(class) {
6428 if (class->rq_offline)
6429 class->rq_offline(rq);
6432 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6438 * used to mark begin/end of suspend/resume:
6440 static int num_cpus_frozen;
6443 * Update cpusets according to cpu_active mask. If cpusets are
6444 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6445 * around partition_sched_domains().
6447 * If we come here as part of a suspend/resume, don't touch cpusets because we
6448 * want to restore it back to its original state upon resume anyway.
6450 static void cpuset_cpu_active(void)
6452 if (cpuhp_tasks_frozen) {
6454 * num_cpus_frozen tracks how many CPUs are involved in suspend
6455 * resume sequence. As long as this is not the last online
6456 * operation in the resume sequence, just build a single sched
6457 * domain, ignoring cpusets.
6459 partition_sched_domains(1, NULL, NULL);
6460 if (--num_cpus_frozen)
6463 * This is the last CPU online operation. So fall through and
6464 * restore the original sched domains by considering the
6465 * cpuset configurations.
6467 cpuset_force_rebuild();
6469 cpuset_update_active_cpus();
6472 static int cpuset_cpu_inactive(unsigned int cpu)
6474 if (!cpuhp_tasks_frozen) {
6475 if (dl_cpu_busy(cpu))
6477 cpuset_update_active_cpus();
6480 partition_sched_domains(1, NULL, NULL);
6485 int sched_cpu_activate(unsigned int cpu)
6487 struct rq *rq = cpu_rq(cpu);
6490 #ifdef CONFIG_SCHED_SMT
6492 * When going up, increment the number of cores with SMT present.
6494 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6495 static_branch_inc_cpuslocked(&sched_smt_present);
6497 set_cpu_active(cpu, true);
6499 if (sched_smp_initialized) {
6500 sched_domains_numa_masks_set(cpu);
6501 cpuset_cpu_active();
6505 * Put the rq online, if not already. This happens:
6507 * 1) In the early boot process, because we build the real domains
6508 * after all CPUs have been brought up.
6510 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6513 rq_lock_irqsave(rq, &rf);
6515 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6518 rq_unlock_irqrestore(rq, &rf);
6523 int sched_cpu_deactivate(unsigned int cpu)
6527 set_cpu_active(cpu, false);
6529 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6530 * users of this state to go away such that all new such users will
6533 * Do sync before park smpboot threads to take care the rcu boost case.
6537 #ifdef CONFIG_SCHED_SMT
6539 * When going down, decrement the number of cores with SMT present.
6541 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6542 static_branch_dec_cpuslocked(&sched_smt_present);
6545 if (!sched_smp_initialized)
6548 ret = cpuset_cpu_inactive(cpu);
6550 set_cpu_active(cpu, true);
6553 sched_domains_numa_masks_clear(cpu);
6557 static void sched_rq_cpu_starting(unsigned int cpu)
6559 struct rq *rq = cpu_rq(cpu);
6561 rq->calc_load_update = calc_load_update;
6562 update_max_interval();
6565 int sched_cpu_starting(unsigned int cpu)
6567 sched_rq_cpu_starting(cpu);
6568 sched_tick_start(cpu);
6572 #ifdef CONFIG_HOTPLUG_CPU
6573 int sched_cpu_dying(unsigned int cpu)
6575 struct rq *rq = cpu_rq(cpu);
6578 /* Handle pending wakeups and then migrate everything off */
6579 sched_ttwu_pending();
6580 sched_tick_stop(cpu);
6582 rq_lock_irqsave(rq, &rf);
6584 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6587 migrate_tasks(rq, &rf);
6588 BUG_ON(rq->nr_running != 1);
6589 rq_unlock_irqrestore(rq, &rf);
6591 calc_load_migrate(rq);
6592 update_max_interval();
6593 nohz_balance_exit_idle(rq);
6599 void __init sched_init_smp(void)
6604 * There's no userspace yet to cause hotplug operations; hence all the
6605 * CPU masks are stable and all blatant races in the below code cannot
6608 mutex_lock(&sched_domains_mutex);
6609 sched_init_domains(cpu_active_mask);
6610 mutex_unlock(&sched_domains_mutex);
6612 /* Move init over to a non-isolated CPU */
6613 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
6615 sched_init_granularity();
6617 init_sched_rt_class();
6618 init_sched_dl_class();
6620 sched_smp_initialized = true;
6623 static int __init migration_init(void)
6625 sched_cpu_starting(smp_processor_id());
6628 early_initcall(migration_init);
6631 void __init sched_init_smp(void)
6633 sched_init_granularity();
6635 #endif /* CONFIG_SMP */
6637 int in_sched_functions(unsigned long addr)
6639 return in_lock_functions(addr) ||
6640 (addr >= (unsigned long)__sched_text_start
6641 && addr < (unsigned long)__sched_text_end);
6644 #ifdef CONFIG_CGROUP_SCHED
6646 * Default task group.
6647 * Every task in system belongs to this group at bootup.
6649 struct task_group root_task_group;
6650 LIST_HEAD(task_groups);
6652 /* Cacheline aligned slab cache for task_group */
6653 static struct kmem_cache *task_group_cache __read_mostly;
6656 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6657 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
6659 void __init sched_init(void)
6661 unsigned long ptr = 0;
6666 #ifdef CONFIG_FAIR_GROUP_SCHED
6667 ptr += 2 * nr_cpu_ids * sizeof(void **);
6669 #ifdef CONFIG_RT_GROUP_SCHED
6670 ptr += 2 * nr_cpu_ids * sizeof(void **);
6673 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
6675 #ifdef CONFIG_FAIR_GROUP_SCHED
6676 root_task_group.se = (struct sched_entity **)ptr;
6677 ptr += nr_cpu_ids * sizeof(void **);
6679 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6680 ptr += nr_cpu_ids * sizeof(void **);
6682 #endif /* CONFIG_FAIR_GROUP_SCHED */
6683 #ifdef CONFIG_RT_GROUP_SCHED
6684 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6685 ptr += nr_cpu_ids * sizeof(void **);
6687 root_task_group.rt_rq = (struct rt_rq **)ptr;
6688 ptr += nr_cpu_ids * sizeof(void **);
6690 #endif /* CONFIG_RT_GROUP_SCHED */
6692 #ifdef CONFIG_CPUMASK_OFFSTACK
6693 for_each_possible_cpu(i) {
6694 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6695 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6696 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6697 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6699 #endif /* CONFIG_CPUMASK_OFFSTACK */
6701 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6702 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6705 init_defrootdomain();
6708 #ifdef CONFIG_RT_GROUP_SCHED
6709 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6710 global_rt_period(), global_rt_runtime());
6711 #endif /* CONFIG_RT_GROUP_SCHED */
6713 #ifdef CONFIG_CGROUP_SCHED
6714 task_group_cache = KMEM_CACHE(task_group, 0);
6716 list_add(&root_task_group.list, &task_groups);
6717 INIT_LIST_HEAD(&root_task_group.children);
6718 INIT_LIST_HEAD(&root_task_group.siblings);
6719 autogroup_init(&init_task);
6720 #endif /* CONFIG_CGROUP_SCHED */
6722 for_each_possible_cpu(i) {
6726 raw_spin_lock_init(&rq->lock);
6728 rq->calc_load_active = 0;
6729 rq->calc_load_update = jiffies + LOAD_FREQ;
6730 init_cfs_rq(&rq->cfs);
6731 init_rt_rq(&rq->rt);
6732 init_dl_rq(&rq->dl);
6733 #ifdef CONFIG_FAIR_GROUP_SCHED
6734 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6735 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6736 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6738 * How much CPU bandwidth does root_task_group get?
6740 * In case of task-groups formed thr' the cgroup filesystem, it
6741 * gets 100% of the CPU resources in the system. This overall
6742 * system CPU resource is divided among the tasks of
6743 * root_task_group and its child task-groups in a fair manner,
6744 * based on each entity's (task or task-group's) weight
6745 * (se->load.weight).
6747 * In other words, if root_task_group has 10 tasks of weight
6748 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6749 * then A0's share of the CPU resource is:
6751 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6753 * We achieve this by letting root_task_group's tasks sit
6754 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6756 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6757 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6758 #endif /* CONFIG_FAIR_GROUP_SCHED */
6760 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6761 #ifdef CONFIG_RT_GROUP_SCHED
6762 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6767 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6768 rq->balance_callback = NULL;
6769 rq->active_balance = 0;
6770 rq->next_balance = jiffies;
6775 rq->avg_idle = 2*sysctl_sched_migration_cost;
6776 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6778 INIT_LIST_HEAD(&rq->cfs_tasks);
6780 rq_attach_root(rq, &def_root_domain);
6781 #ifdef CONFIG_NO_HZ_COMMON
6782 rq->last_load_update_tick = jiffies;
6783 rq->last_blocked_load_update_tick = jiffies;
6784 atomic_set(&rq->nohz_flags, 0);
6786 #endif /* CONFIG_SMP */
6788 atomic_set(&rq->nr_iowait, 0);
6791 set_load_weight(&init_task, false);
6794 * The boot idle thread does lazy MMU switching as well:
6797 enter_lazy_tlb(&init_mm, current);
6800 * Make us the idle thread. Technically, schedule() should not be
6801 * called from this thread, however somewhere below it might be,
6802 * but because we are the idle thread, we just pick up running again
6803 * when this runqueue becomes "idle".
6805 init_idle(current, smp_processor_id());
6807 calc_load_update = jiffies + LOAD_FREQ;
6810 idle_thread_set_boot_cpu();
6812 init_sched_fair_class();
6820 scheduler_running = 1;
6823 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6824 static inline int preempt_count_equals(int preempt_offset)
6826 int nested = preempt_count() + rcu_preempt_depth();
6828 return (nested == preempt_offset);
6831 void __might_sleep(const char *file, int line, int preempt_offset)
6834 * Blocking primitives will set (and therefore destroy) current->state,
6835 * since we will exit with TASK_RUNNING make sure we enter with it,
6836 * otherwise we will destroy state.
6838 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6839 "do not call blocking ops when !TASK_RUNNING; "
6840 "state=%lx set at [<%p>] %pS\n",
6842 (void *)current->task_state_change,
6843 (void *)current->task_state_change);
6845 ___might_sleep(file, line, preempt_offset);
6847 EXPORT_SYMBOL(__might_sleep);
6849 void ___might_sleep(const char *file, int line, int preempt_offset)
6851 /* Ratelimiting timestamp: */
6852 static unsigned long prev_jiffy;
6854 unsigned long preempt_disable_ip;
6856 /* WARN_ON_ONCE() by default, no rate limit required: */
6859 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6860 !is_idle_task(current) && !current->non_block_count) ||
6861 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6865 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6867 prev_jiffy = jiffies;
6869 /* Save this before calling printk(), since that will clobber it: */
6870 preempt_disable_ip = get_preempt_disable_ip(current);
6873 "BUG: sleeping function called from invalid context at %s:%d\n",
6876 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
6877 in_atomic(), irqs_disabled(), current->non_block_count,
6878 current->pid, current->comm);
6880 if (task_stack_end_corrupted(current))
6881 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6883 debug_show_held_locks(current);
6884 if (irqs_disabled())
6885 print_irqtrace_events(current);
6886 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6887 && !preempt_count_equals(preempt_offset)) {
6888 pr_err("Preemption disabled at:");
6889 print_ip_sym(preempt_disable_ip);
6893 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6895 EXPORT_SYMBOL(___might_sleep);
6897 void __cant_sleep(const char *file, int line, int preempt_offset)
6899 static unsigned long prev_jiffy;
6901 if (irqs_disabled())
6904 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6907 if (preempt_count() > preempt_offset)
6910 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6912 prev_jiffy = jiffies;
6914 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6915 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6916 in_atomic(), irqs_disabled(),
6917 current->pid, current->comm);
6919 debug_show_held_locks(current);
6921 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6923 EXPORT_SYMBOL_GPL(__cant_sleep);
6926 #ifdef CONFIG_MAGIC_SYSRQ
6927 void normalize_rt_tasks(void)
6929 struct task_struct *g, *p;
6930 struct sched_attr attr = {
6931 .sched_policy = SCHED_NORMAL,
6934 read_lock(&tasklist_lock);
6935 for_each_process_thread(g, p) {
6937 * Only normalize user tasks:
6939 if (p->flags & PF_KTHREAD)
6942 p->se.exec_start = 0;
6943 schedstat_set(p->se.statistics.wait_start, 0);
6944 schedstat_set(p->se.statistics.sleep_start, 0);
6945 schedstat_set(p->se.statistics.block_start, 0);
6947 if (!dl_task(p) && !rt_task(p)) {
6949 * Renice negative nice level userspace
6952 if (task_nice(p) < 0)
6953 set_user_nice(p, 0);
6957 __sched_setscheduler(p, &attr, false, false);
6959 read_unlock(&tasklist_lock);
6962 #endif /* CONFIG_MAGIC_SYSRQ */
6964 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6966 * These functions are only useful for the IA64 MCA handling, or kdb.
6968 * They can only be called when the whole system has been
6969 * stopped - every CPU needs to be quiescent, and no scheduling
6970 * activity can take place. Using them for anything else would
6971 * be a serious bug, and as a result, they aren't even visible
6972 * under any other configuration.
6976 * curr_task - return the current task for a given CPU.
6977 * @cpu: the processor in question.
6979 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6981 * Return: The current task for @cpu.
6983 struct task_struct *curr_task(int cpu)
6985 return cpu_curr(cpu);
6988 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6992 * ia64_set_curr_task - set the current task for a given CPU.
6993 * @cpu: the processor in question.
6994 * @p: the task pointer to set.
6996 * Description: This function must only be used when non-maskable interrupts
6997 * are serviced on a separate stack. It allows the architecture to switch the
6998 * notion of the current task on a CPU in a non-blocking manner. This function
6999 * must be called with all CPU's synchronized, and interrupts disabled, the
7000 * and caller must save the original value of the current task (see
7001 * curr_task() above) and restore that value before reenabling interrupts and
7002 * re-starting the system.
7004 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7006 void ia64_set_curr_task(int cpu, struct task_struct *p)
7013 #ifdef CONFIG_CGROUP_SCHED
7014 /* task_group_lock serializes the addition/removal of task groups */
7015 static DEFINE_SPINLOCK(task_group_lock);
7017 static inline void alloc_uclamp_sched_group(struct task_group *tg,
7018 struct task_group *parent)
7020 #ifdef CONFIG_UCLAMP_TASK_GROUP
7021 enum uclamp_id clamp_id;
7023 for_each_clamp_id(clamp_id) {
7024 uclamp_se_set(&tg->uclamp_req[clamp_id],
7025 uclamp_none(clamp_id), false);
7026 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
7031 static void sched_free_group(struct task_group *tg)
7033 free_fair_sched_group(tg);
7034 free_rt_sched_group(tg);
7036 kmem_cache_free(task_group_cache, tg);
7039 /* allocate runqueue etc for a new task group */
7040 struct task_group *sched_create_group(struct task_group *parent)
7042 struct task_group *tg;
7044 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7046 return ERR_PTR(-ENOMEM);
7048 if (!alloc_fair_sched_group(tg, parent))
7051 if (!alloc_rt_sched_group(tg, parent))
7054 alloc_uclamp_sched_group(tg, parent);
7059 sched_free_group(tg);
7060 return ERR_PTR(-ENOMEM);
7063 void sched_online_group(struct task_group *tg, struct task_group *parent)
7065 unsigned long flags;
7067 spin_lock_irqsave(&task_group_lock, flags);
7068 list_add_rcu(&tg->list, &task_groups);
7070 /* Root should already exist: */
7073 tg->parent = parent;
7074 INIT_LIST_HEAD(&tg->children);
7075 list_add_rcu(&tg->siblings, &parent->children);
7076 spin_unlock_irqrestore(&task_group_lock, flags);
7078 online_fair_sched_group(tg);
7081 /* rcu callback to free various structures associated with a task group */
7082 static void sched_free_group_rcu(struct rcu_head *rhp)
7084 /* Now it should be safe to free those cfs_rqs: */
7085 sched_free_group(container_of(rhp, struct task_group, rcu));
7088 void sched_destroy_group(struct task_group *tg)
7090 /* Wait for possible concurrent references to cfs_rqs complete: */
7091 call_rcu(&tg->rcu, sched_free_group_rcu);
7094 void sched_offline_group(struct task_group *tg)
7096 unsigned long flags;
7098 /* End participation in shares distribution: */
7099 unregister_fair_sched_group(tg);
7101 spin_lock_irqsave(&task_group_lock, flags);
7102 list_del_rcu(&tg->list);
7103 list_del_rcu(&tg->siblings);
7104 spin_unlock_irqrestore(&task_group_lock, flags);
7107 static void sched_change_group(struct task_struct *tsk, int type)
7109 struct task_group *tg;
7112 * All callers are synchronized by task_rq_lock(); we do not use RCU
7113 * which is pointless here. Thus, we pass "true" to task_css_check()
7114 * to prevent lockdep warnings.
7116 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7117 struct task_group, css);
7118 tg = autogroup_task_group(tsk, tg);
7119 tsk->sched_task_group = tg;
7121 #ifdef CONFIG_FAIR_GROUP_SCHED
7122 if (tsk->sched_class->task_change_group)
7123 tsk->sched_class->task_change_group(tsk, type);
7126 set_task_rq(tsk, task_cpu(tsk));
7130 * Change task's runqueue when it moves between groups.
7132 * The caller of this function should have put the task in its new group by
7133 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7136 void sched_move_task(struct task_struct *tsk)
7138 int queued, running, queue_flags =
7139 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7143 rq = task_rq_lock(tsk, &rf);
7144 update_rq_clock(rq);
7146 running = task_current(rq, tsk);
7147 queued = task_on_rq_queued(tsk);
7150 dequeue_task(rq, tsk, queue_flags);
7152 put_prev_task(rq, tsk);
7154 sched_change_group(tsk, TASK_MOVE_GROUP);
7157 enqueue_task(rq, tsk, queue_flags);
7159 set_next_task(rq, tsk);
7161 * After changing group, the running task may have joined a
7162 * throttled one but it's still the running task. Trigger a
7163 * resched to make sure that task can still run.
7168 task_rq_unlock(rq, tsk, &rf);
7171 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7173 return css ? container_of(css, struct task_group, css) : NULL;
7176 static struct cgroup_subsys_state *
7177 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7179 struct task_group *parent = css_tg(parent_css);
7180 struct task_group *tg;
7183 /* This is early initialization for the top cgroup */
7184 return &root_task_group.css;
7187 tg = sched_create_group(parent);
7189 return ERR_PTR(-ENOMEM);
7194 /* Expose task group only after completing cgroup initialization */
7195 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7197 struct task_group *tg = css_tg(css);
7198 struct task_group *parent = css_tg(css->parent);
7201 sched_online_group(tg, parent);
7203 #ifdef CONFIG_UCLAMP_TASK_GROUP
7204 /* Propagate the effective uclamp value for the new group */
7205 mutex_lock(&uclamp_mutex);
7207 cpu_util_update_eff(css);
7209 mutex_unlock(&uclamp_mutex);
7215 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7217 struct task_group *tg = css_tg(css);
7219 sched_offline_group(tg);
7222 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7224 struct task_group *tg = css_tg(css);
7227 * Relies on the RCU grace period between css_released() and this.
7229 sched_free_group(tg);
7233 * This is called before wake_up_new_task(), therefore we really only
7234 * have to set its group bits, all the other stuff does not apply.
7236 static void cpu_cgroup_fork(struct task_struct *task)
7241 rq = task_rq_lock(task, &rf);
7243 update_rq_clock(rq);
7244 sched_change_group(task, TASK_SET_GROUP);
7246 task_rq_unlock(rq, task, &rf);
7249 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7251 struct task_struct *task;
7252 struct cgroup_subsys_state *css;
7255 cgroup_taskset_for_each(task, css, tset) {
7256 #ifdef CONFIG_RT_GROUP_SCHED
7257 if (!sched_rt_can_attach(css_tg(css), task))
7261 * Serialize against wake_up_new_task() such that if its
7262 * running, we're sure to observe its full state.
7264 raw_spin_lock_irq(&task->pi_lock);
7266 * Avoid calling sched_move_task() before wake_up_new_task()
7267 * has happened. This would lead to problems with PELT, due to
7268 * move wanting to detach+attach while we're not attached yet.
7270 if (task->state == TASK_NEW)
7272 raw_spin_unlock_irq(&task->pi_lock);
7280 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7282 struct task_struct *task;
7283 struct cgroup_subsys_state *css;
7285 cgroup_taskset_for_each(task, css, tset)
7286 sched_move_task(task);
7289 #ifdef CONFIG_UCLAMP_TASK_GROUP
7290 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7292 struct cgroup_subsys_state *top_css = css;
7293 struct uclamp_se *uc_parent = NULL;
7294 struct uclamp_se *uc_se = NULL;
7295 unsigned int eff[UCLAMP_CNT];
7296 enum uclamp_id clamp_id;
7297 unsigned int clamps;
7299 lockdep_assert_held(&uclamp_mutex);
7300 SCHED_WARN_ON(!rcu_read_lock_held());
7302 css_for_each_descendant_pre(css, top_css) {
7303 uc_parent = css_tg(css)->parent
7304 ? css_tg(css)->parent->uclamp : NULL;
7306 for_each_clamp_id(clamp_id) {
7307 /* Assume effective clamps matches requested clamps */
7308 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7309 /* Cap effective clamps with parent's effective clamps */
7311 eff[clamp_id] > uc_parent[clamp_id].value) {
7312 eff[clamp_id] = uc_parent[clamp_id].value;
7315 /* Ensure protection is always capped by limit */
7316 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7318 /* Propagate most restrictive effective clamps */
7320 uc_se = css_tg(css)->uclamp;
7321 for_each_clamp_id(clamp_id) {
7322 if (eff[clamp_id] == uc_se[clamp_id].value)
7324 uc_se[clamp_id].value = eff[clamp_id];
7325 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7326 clamps |= (0x1 << clamp_id);
7329 css = css_rightmost_descendant(css);
7333 /* Immediately update descendants RUNNABLE tasks */
7334 uclamp_update_active_tasks(css);
7339 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7340 * C expression. Since there is no way to convert a macro argument (N) into a
7341 * character constant, use two levels of macros.
7343 #define _POW10(exp) ((unsigned int)1e##exp)
7344 #define POW10(exp) _POW10(exp)
7346 struct uclamp_request {
7347 #define UCLAMP_PERCENT_SHIFT 2
7348 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7354 static inline struct uclamp_request
7355 capacity_from_percent(char *buf)
7357 struct uclamp_request req = {
7358 .percent = UCLAMP_PERCENT_SCALE,
7359 .util = SCHED_CAPACITY_SCALE,
7364 if (strcmp(buf, "max")) {
7365 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7369 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7374 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7375 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7381 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7382 size_t nbytes, loff_t off,
7383 enum uclamp_id clamp_id)
7385 struct uclamp_request req;
7386 struct task_group *tg;
7388 req = capacity_from_percent(buf);
7392 static_branch_enable(&sched_uclamp_used);
7394 mutex_lock(&uclamp_mutex);
7397 tg = css_tg(of_css(of));
7398 if (tg->uclamp_req[clamp_id].value != req.util)
7399 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7402 * Because of not recoverable conversion rounding we keep track of the
7403 * exact requested value
7405 tg->uclamp_pct[clamp_id] = req.percent;
7407 /* Update effective clamps to track the most restrictive value */
7408 cpu_util_update_eff(of_css(of));
7411 mutex_unlock(&uclamp_mutex);
7416 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7417 char *buf, size_t nbytes,
7420 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7423 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7424 char *buf, size_t nbytes,
7427 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7430 static inline void cpu_uclamp_print(struct seq_file *sf,
7431 enum uclamp_id clamp_id)
7433 struct task_group *tg;
7439 tg = css_tg(seq_css(sf));
7440 util_clamp = tg->uclamp_req[clamp_id].value;
7443 if (util_clamp == SCHED_CAPACITY_SCALE) {
7444 seq_puts(sf, "max\n");
7448 percent = tg->uclamp_pct[clamp_id];
7449 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7450 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7453 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7455 cpu_uclamp_print(sf, UCLAMP_MIN);
7459 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7461 cpu_uclamp_print(sf, UCLAMP_MAX);
7464 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7466 #ifdef CONFIG_FAIR_GROUP_SCHED
7467 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7468 struct cftype *cftype, u64 shareval)
7470 if (shareval > scale_load_down(ULONG_MAX))
7471 shareval = MAX_SHARES;
7472 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7475 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7478 struct task_group *tg = css_tg(css);
7480 return (u64) scale_load_down(tg->shares);
7483 #ifdef CONFIG_CFS_BANDWIDTH
7484 static DEFINE_MUTEX(cfs_constraints_mutex);
7486 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7487 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7488 /* More than 203 days if BW_SHIFT equals 20. */
7489 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
7491 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7493 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7495 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7496 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7498 if (tg == &root_task_group)
7502 * Ensure we have at some amount of bandwidth every period. This is
7503 * to prevent reaching a state of large arrears when throttled via
7504 * entity_tick() resulting in prolonged exit starvation.
7506 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7510 * Likewise, bound things on the otherside by preventing insane quota
7511 * periods. This also allows us to normalize in computing quota
7514 if (period > max_cfs_quota_period)
7518 * Bound quota to defend quota against overflow during bandwidth shift.
7520 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
7524 * Prevent race between setting of cfs_rq->runtime_enabled and
7525 * unthrottle_offline_cfs_rqs().
7528 mutex_lock(&cfs_constraints_mutex);
7529 ret = __cfs_schedulable(tg, period, quota);
7533 runtime_enabled = quota != RUNTIME_INF;
7534 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7536 * If we need to toggle cfs_bandwidth_used, off->on must occur
7537 * before making related changes, and on->off must occur afterwards
7539 if (runtime_enabled && !runtime_was_enabled)
7540 cfs_bandwidth_usage_inc();
7541 raw_spin_lock_irq(&cfs_b->lock);
7542 cfs_b->period = ns_to_ktime(period);
7543 cfs_b->quota = quota;
7545 __refill_cfs_bandwidth_runtime(cfs_b);
7547 /* Restart the period timer (if active) to handle new period expiry: */
7548 if (runtime_enabled)
7549 start_cfs_bandwidth(cfs_b);
7551 raw_spin_unlock_irq(&cfs_b->lock);
7553 for_each_online_cpu(i) {
7554 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7555 struct rq *rq = cfs_rq->rq;
7558 rq_lock_irq(rq, &rf);
7559 cfs_rq->runtime_enabled = runtime_enabled;
7560 cfs_rq->runtime_remaining = 0;
7562 if (cfs_rq->throttled)
7563 unthrottle_cfs_rq(cfs_rq);
7564 rq_unlock_irq(rq, &rf);
7566 if (runtime_was_enabled && !runtime_enabled)
7567 cfs_bandwidth_usage_dec();
7569 mutex_unlock(&cfs_constraints_mutex);
7575 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7579 period = ktime_to_ns(tg->cfs_bandwidth.period);
7580 if (cfs_quota_us < 0)
7581 quota = RUNTIME_INF;
7582 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7583 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7587 return tg_set_cfs_bandwidth(tg, period, quota);
7590 static long tg_get_cfs_quota(struct task_group *tg)
7594 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7597 quota_us = tg->cfs_bandwidth.quota;
7598 do_div(quota_us, NSEC_PER_USEC);
7603 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7607 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
7610 period = (u64)cfs_period_us * NSEC_PER_USEC;
7611 quota = tg->cfs_bandwidth.quota;
7613 return tg_set_cfs_bandwidth(tg, period, quota);
7616 static long tg_get_cfs_period(struct task_group *tg)
7620 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7621 do_div(cfs_period_us, NSEC_PER_USEC);
7623 return cfs_period_us;
7626 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7629 return tg_get_cfs_quota(css_tg(css));
7632 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7633 struct cftype *cftype, s64 cfs_quota_us)
7635 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7638 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7641 return tg_get_cfs_period(css_tg(css));
7644 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7645 struct cftype *cftype, u64 cfs_period_us)
7647 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7650 struct cfs_schedulable_data {
7651 struct task_group *tg;
7656 * normalize group quota/period to be quota/max_period
7657 * note: units are usecs
7659 static u64 normalize_cfs_quota(struct task_group *tg,
7660 struct cfs_schedulable_data *d)
7668 period = tg_get_cfs_period(tg);
7669 quota = tg_get_cfs_quota(tg);
7672 /* note: these should typically be equivalent */
7673 if (quota == RUNTIME_INF || quota == -1)
7676 return to_ratio(period, quota);
7679 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7681 struct cfs_schedulable_data *d = data;
7682 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7683 s64 quota = 0, parent_quota = -1;
7686 quota = RUNTIME_INF;
7688 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7690 quota = normalize_cfs_quota(tg, d);
7691 parent_quota = parent_b->hierarchical_quota;
7694 * Ensure max(child_quota) <= parent_quota. On cgroup2,
7695 * always take the min. On cgroup1, only inherit when no
7698 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
7699 quota = min(quota, parent_quota);
7701 if (quota == RUNTIME_INF)
7702 quota = parent_quota;
7703 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7707 cfs_b->hierarchical_quota = quota;
7712 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7715 struct cfs_schedulable_data data = {
7721 if (quota != RUNTIME_INF) {
7722 do_div(data.period, NSEC_PER_USEC);
7723 do_div(data.quota, NSEC_PER_USEC);
7727 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7733 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
7735 struct task_group *tg = css_tg(seq_css(sf));
7736 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7738 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7739 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7740 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7742 if (schedstat_enabled() && tg != &root_task_group) {
7746 for_each_possible_cpu(i)
7747 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
7749 seq_printf(sf, "wait_sum %llu\n", ws);
7754 #endif /* CONFIG_CFS_BANDWIDTH */
7755 #endif /* CONFIG_FAIR_GROUP_SCHED */
7757 #ifdef CONFIG_RT_GROUP_SCHED
7758 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7759 struct cftype *cft, s64 val)
7761 return sched_group_set_rt_runtime(css_tg(css), val);
7764 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7767 return sched_group_rt_runtime(css_tg(css));
7770 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7771 struct cftype *cftype, u64 rt_period_us)
7773 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7776 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7779 return sched_group_rt_period(css_tg(css));
7781 #endif /* CONFIG_RT_GROUP_SCHED */
7783 static struct cftype cpu_legacy_files[] = {
7784 #ifdef CONFIG_FAIR_GROUP_SCHED
7787 .read_u64 = cpu_shares_read_u64,
7788 .write_u64 = cpu_shares_write_u64,
7791 #ifdef CONFIG_CFS_BANDWIDTH
7793 .name = "cfs_quota_us",
7794 .read_s64 = cpu_cfs_quota_read_s64,
7795 .write_s64 = cpu_cfs_quota_write_s64,
7798 .name = "cfs_period_us",
7799 .read_u64 = cpu_cfs_period_read_u64,
7800 .write_u64 = cpu_cfs_period_write_u64,
7804 .seq_show = cpu_cfs_stat_show,
7807 #ifdef CONFIG_RT_GROUP_SCHED
7809 .name = "rt_runtime_us",
7810 .read_s64 = cpu_rt_runtime_read,
7811 .write_s64 = cpu_rt_runtime_write,
7814 .name = "rt_period_us",
7815 .read_u64 = cpu_rt_period_read_uint,
7816 .write_u64 = cpu_rt_period_write_uint,
7819 #ifdef CONFIG_UCLAMP_TASK_GROUP
7821 .name = "uclamp.min",
7822 .flags = CFTYPE_NOT_ON_ROOT,
7823 .seq_show = cpu_uclamp_min_show,
7824 .write = cpu_uclamp_min_write,
7827 .name = "uclamp.max",
7828 .flags = CFTYPE_NOT_ON_ROOT,
7829 .seq_show = cpu_uclamp_max_show,
7830 .write = cpu_uclamp_max_write,
7836 static int cpu_extra_stat_show(struct seq_file *sf,
7837 struct cgroup_subsys_state *css)
7839 #ifdef CONFIG_CFS_BANDWIDTH
7841 struct task_group *tg = css_tg(css);
7842 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7845 throttled_usec = cfs_b->throttled_time;
7846 do_div(throttled_usec, NSEC_PER_USEC);
7848 seq_printf(sf, "nr_periods %d\n"
7850 "throttled_usec %llu\n",
7851 cfs_b->nr_periods, cfs_b->nr_throttled,
7858 #ifdef CONFIG_FAIR_GROUP_SCHED
7859 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
7862 struct task_group *tg = css_tg(css);
7863 u64 weight = scale_load_down(tg->shares);
7865 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
7868 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
7869 struct cftype *cft, u64 weight)
7872 * cgroup weight knobs should use the common MIN, DFL and MAX
7873 * values which are 1, 100 and 10000 respectively. While it loses
7874 * a bit of range on both ends, it maps pretty well onto the shares
7875 * value used by scheduler and the round-trip conversions preserve
7876 * the original value over the entire range.
7878 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
7881 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
7883 return sched_group_set_shares(css_tg(css), scale_load(weight));
7886 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
7889 unsigned long weight = scale_load_down(css_tg(css)->shares);
7890 int last_delta = INT_MAX;
7893 /* find the closest nice value to the current weight */
7894 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
7895 delta = abs(sched_prio_to_weight[prio] - weight);
7896 if (delta >= last_delta)
7901 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
7904 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
7905 struct cftype *cft, s64 nice)
7907 unsigned long weight;
7910 if (nice < MIN_NICE || nice > MAX_NICE)
7913 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
7914 idx = array_index_nospec(idx, 40);
7915 weight = sched_prio_to_weight[idx];
7917 return sched_group_set_shares(css_tg(css), scale_load(weight));
7921 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
7922 long period, long quota)
7925 seq_puts(sf, "max");
7927 seq_printf(sf, "%ld", quota);
7929 seq_printf(sf, " %ld\n", period);
7932 /* caller should put the current value in *@periodp before calling */
7933 static int __maybe_unused cpu_period_quota_parse(char *buf,
7934 u64 *periodp, u64 *quotap)
7936 char tok[21]; /* U64_MAX */
7938 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
7941 *periodp *= NSEC_PER_USEC;
7943 if (sscanf(tok, "%llu", quotap))
7944 *quotap *= NSEC_PER_USEC;
7945 else if (!strcmp(tok, "max"))
7946 *quotap = RUNTIME_INF;
7953 #ifdef CONFIG_CFS_BANDWIDTH
7954 static int cpu_max_show(struct seq_file *sf, void *v)
7956 struct task_group *tg = css_tg(seq_css(sf));
7958 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7962 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7963 char *buf, size_t nbytes, loff_t off)
7965 struct task_group *tg = css_tg(of_css(of));
7966 u64 period = tg_get_cfs_period(tg);
7970 ret = cpu_period_quota_parse(buf, &period, "a);
7972 ret = tg_set_cfs_bandwidth(tg, period, quota);
7973 return ret ?: nbytes;
7977 static struct cftype cpu_files[] = {
7978 #ifdef CONFIG_FAIR_GROUP_SCHED
7981 .flags = CFTYPE_NOT_ON_ROOT,
7982 .read_u64 = cpu_weight_read_u64,
7983 .write_u64 = cpu_weight_write_u64,
7986 .name = "weight.nice",
7987 .flags = CFTYPE_NOT_ON_ROOT,
7988 .read_s64 = cpu_weight_nice_read_s64,
7989 .write_s64 = cpu_weight_nice_write_s64,
7992 #ifdef CONFIG_CFS_BANDWIDTH
7995 .flags = CFTYPE_NOT_ON_ROOT,
7996 .seq_show = cpu_max_show,
7997 .write = cpu_max_write,
8000 #ifdef CONFIG_UCLAMP_TASK_GROUP
8002 .name = "uclamp.min",
8003 .flags = CFTYPE_NOT_ON_ROOT,
8004 .seq_show = cpu_uclamp_min_show,
8005 .write = cpu_uclamp_min_write,
8008 .name = "uclamp.max",
8009 .flags = CFTYPE_NOT_ON_ROOT,
8010 .seq_show = cpu_uclamp_max_show,
8011 .write = cpu_uclamp_max_write,
8017 struct cgroup_subsys cpu_cgrp_subsys = {
8018 .css_alloc = cpu_cgroup_css_alloc,
8019 .css_online = cpu_cgroup_css_online,
8020 .css_released = cpu_cgroup_css_released,
8021 .css_free = cpu_cgroup_css_free,
8022 .css_extra_stat_show = cpu_extra_stat_show,
8023 .fork = cpu_cgroup_fork,
8024 .can_attach = cpu_cgroup_can_attach,
8025 .attach = cpu_cgroup_attach,
8026 .legacy_cftypes = cpu_legacy_files,
8027 .dfl_cftypes = cpu_files,
8032 #endif /* CONFIG_CGROUP_SCHED */
8034 void dump_cpu_task(int cpu)
8036 pr_info("Task dump for CPU %d:\n", cpu);
8037 sched_show_task(cpu_curr(cpu));
8041 * Nice levels are multiplicative, with a gentle 10% change for every
8042 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8043 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8044 * that remained on nice 0.
8046 * The "10% effect" is relative and cumulative: from _any_ nice level,
8047 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8048 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8049 * If a task goes up by ~10% and another task goes down by ~10% then
8050 * the relative distance between them is ~25%.)
8052 const int sched_prio_to_weight[40] = {
8053 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8054 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8055 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8056 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8057 /* 0 */ 1024, 820, 655, 526, 423,
8058 /* 5 */ 335, 272, 215, 172, 137,
8059 /* 10 */ 110, 87, 70, 56, 45,
8060 /* 15 */ 36, 29, 23, 18, 15,
8064 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8066 * In cases where the weight does not change often, we can use the
8067 * precalculated inverse to speed up arithmetics by turning divisions
8068 * into multiplications:
8070 const u32 sched_prio_to_wmult[40] = {
8071 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8072 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8073 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8074 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8075 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8076 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8077 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8078 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8081 #undef CREATE_TRACE_POINTS