1 // SPDX-License-Identifier: GPL-2.0-only
5 * Core kernel scheduler code and related syscalls
7 * Copyright (C) 1991-2002 Linus Torvalds
9 #define CREATE_TRACE_POINTS
10 #include <trace/events/sched.h>
11 #undef CREATE_TRACE_POINTS
15 #include <linux/nospec.h>
17 #include <linux/kcov.h>
18 #include <linux/scs.h>
20 #include <asm/switch_to.h>
23 #include "../workqueue_internal.h"
24 #include "../../fs/io-wq.h"
25 #include "../smpboot.h"
31 * Export tracepoints that act as a bare tracehook (ie: have no trace event
32 * associated with them) to allow external modules to probe them.
34 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
35 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
36 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
37 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
38 EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
39 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
40 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
43 EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #ifdef CONFIG_SCHED_DEBUG
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
64 * Number of tasks to iterate in a single balance run.
65 * Limited because this is done with IRQs disabled.
67 const_debug unsigned int sysctl_sched_nr_migrate = 32;
70 * period over which we measure -rt task CPU usage in us.
73 unsigned int sysctl_sched_rt_period = 1000000;
75 __read_mostly int scheduler_running;
78 * part of the period that we allow rt tasks to run in us.
81 int sysctl_sched_rt_runtime = 950000;
85 * Serialization rules:
91 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
94 * rq2->lock where: rq1 < rq2
98 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
99 * local CPU's rq->lock, it optionally removes the task from the runqueue and
100 * always looks at the local rq data structures to find the most eligible task
103 * Task enqueue is also under rq->lock, possibly taken from another CPU.
104 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
105 * the local CPU to avoid bouncing the runqueue state around [ see
106 * ttwu_queue_wakelist() ]
108 * Task wakeup, specifically wakeups that involve migration, are horribly
109 * complicated to avoid having to take two rq->locks.
113 * System-calls and anything external will use task_rq_lock() which acquires
114 * both p->pi_lock and rq->lock. As a consequence the state they change is
115 * stable while holding either lock:
117 * - sched_setaffinity()/
118 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
119 * - set_user_nice(): p->se.load, p->*prio
120 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
121 * p->se.load, p->rt_priority,
122 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
123 * - sched_setnuma(): p->numa_preferred_nid
124 * - sched_move_task()/
125 * cpu_cgroup_fork(): p->sched_task_group
126 * - uclamp_update_active() p->uclamp*
128 * p->state <- TASK_*:
130 * is changed locklessly using set_current_state(), __set_current_state() or
131 * set_special_state(), see their respective comments, or by
132 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
135 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
137 * is set by activate_task() and cleared by deactivate_task(), under
138 * rq->lock. Non-zero indicates the task is runnable, the special
139 * ON_RQ_MIGRATING state is used for migration without holding both
140 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
142 * p->on_cpu <- { 0, 1 }:
144 * is set by prepare_task() and cleared by finish_task() such that it will be
145 * set before p is scheduled-in and cleared after p is scheduled-out, both
146 * under rq->lock. Non-zero indicates the task is running on its CPU.
148 * [ The astute reader will observe that it is possible for two tasks on one
149 * CPU to have ->on_cpu = 1 at the same time. ]
151 * task_cpu(p): is changed by set_task_cpu(), the rules are:
153 * - Don't call set_task_cpu() on a blocked task:
155 * We don't care what CPU we're not running on, this simplifies hotplug,
156 * the CPU assignment of blocked tasks isn't required to be valid.
158 * - for try_to_wake_up(), called under p->pi_lock:
160 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
162 * - for migration called under rq->lock:
163 * [ see task_on_rq_migrating() in task_rq_lock() ]
165 * o move_queued_task()
168 * - for migration called under double_rq_lock():
170 * o __migrate_swap_task()
171 * o push_rt_task() / pull_rt_task()
172 * o push_dl_task() / pull_dl_task()
173 * o dl_task_offline_migration()
178 * __task_rq_lock - lock the rq @p resides on.
180 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
185 lockdep_assert_held(&p->pi_lock);
189 raw_spin_lock(&rq->lock);
190 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
194 raw_spin_unlock(&rq->lock);
196 while (unlikely(task_on_rq_migrating(p)))
202 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
204 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
205 __acquires(p->pi_lock)
211 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
213 raw_spin_lock(&rq->lock);
215 * move_queued_task() task_rq_lock()
218 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
219 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
220 * [S] ->cpu = new_cpu [L] task_rq()
224 * If we observe the old CPU in task_rq_lock(), the acquire of
225 * the old rq->lock will fully serialize against the stores.
227 * If we observe the new CPU in task_rq_lock(), the address
228 * dependency headed by '[L] rq = task_rq()' and the acquire
229 * will pair with the WMB to ensure we then also see migrating.
231 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
235 raw_spin_unlock(&rq->lock);
236 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
238 while (unlikely(task_on_rq_migrating(p)))
244 * RQ-clock updating methods:
247 static void update_rq_clock_task(struct rq *rq, s64 delta)
250 * In theory, the compile should just see 0 here, and optimize out the call
251 * to sched_rt_avg_update. But I don't trust it...
253 s64 __maybe_unused steal = 0, irq_delta = 0;
255 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
256 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
259 * Since irq_time is only updated on {soft,}irq_exit, we might run into
260 * this case when a previous update_rq_clock() happened inside a
263 * When this happens, we stop ->clock_task and only update the
264 * prev_irq_time stamp to account for the part that fit, so that a next
265 * update will consume the rest. This ensures ->clock_task is
268 * It does however cause some slight miss-attribution of {soft,}irq
269 * time, a more accurate solution would be to update the irq_time using
270 * the current rq->clock timestamp, except that would require using
273 if (irq_delta > delta)
276 rq->prev_irq_time += irq_delta;
279 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
280 if (static_key_false((¶virt_steal_rq_enabled))) {
281 steal = paravirt_steal_clock(cpu_of(rq));
282 steal -= rq->prev_steal_time_rq;
284 if (unlikely(steal > delta))
287 rq->prev_steal_time_rq += steal;
292 rq->clock_task += delta;
294 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
295 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
296 update_irq_load_avg(rq, irq_delta + steal);
298 update_rq_clock_pelt(rq, delta);
301 void update_rq_clock(struct rq *rq)
305 lockdep_assert_held(&rq->lock);
307 if (rq->clock_update_flags & RQCF_ACT_SKIP)
310 #ifdef CONFIG_SCHED_DEBUG
311 if (sched_feat(WARN_DOUBLE_CLOCK))
312 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
313 rq->clock_update_flags |= RQCF_UPDATED;
316 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
320 update_rq_clock_task(rq, delta);
323 #ifdef CONFIG_SCHED_HRTICK
325 * Use HR-timers to deliver accurate preemption points.
328 static void hrtick_clear(struct rq *rq)
330 if (hrtimer_active(&rq->hrtick_timer))
331 hrtimer_cancel(&rq->hrtick_timer);
335 * High-resolution timer tick.
336 * Runs from hardirq context with interrupts disabled.
338 static enum hrtimer_restart hrtick(struct hrtimer *timer)
340 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
343 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
347 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
350 return HRTIMER_NORESTART;
355 static void __hrtick_restart(struct rq *rq)
357 struct hrtimer *timer = &rq->hrtick_timer;
359 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
363 * called from hardirq (IPI) context
365 static void __hrtick_start(void *arg)
371 __hrtick_restart(rq);
376 * Called to set the hrtick timer state.
378 * called with rq->lock held and irqs disabled
380 void hrtick_start(struct rq *rq, u64 delay)
382 struct hrtimer *timer = &rq->hrtick_timer;
387 * Don't schedule slices shorter than 10000ns, that just
388 * doesn't make sense and can cause timer DoS.
390 delta = max_t(s64, delay, 10000LL);
391 time = ktime_add_ns(timer->base->get_time(), delta);
393 hrtimer_set_expires(timer, time);
396 __hrtick_restart(rq);
398 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
403 * Called to set the hrtick timer state.
405 * called with rq->lock held and irqs disabled
407 void hrtick_start(struct rq *rq, u64 delay)
410 * Don't schedule slices shorter than 10000ns, that just
411 * doesn't make sense. Rely on vruntime for fairness.
413 delay = max_t(u64, delay, 10000LL);
414 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
415 HRTIMER_MODE_REL_PINNED_HARD);
418 #endif /* CONFIG_SMP */
420 static void hrtick_rq_init(struct rq *rq)
423 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
425 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
426 rq->hrtick_timer.function = hrtick;
428 #else /* CONFIG_SCHED_HRTICK */
429 static inline void hrtick_clear(struct rq *rq)
433 static inline void hrtick_rq_init(struct rq *rq)
436 #endif /* CONFIG_SCHED_HRTICK */
439 * cmpxchg based fetch_or, macro so it works for different integer types
441 #define fetch_or(ptr, mask) \
443 typeof(ptr) _ptr = (ptr); \
444 typeof(mask) _mask = (mask); \
445 typeof(*_ptr) _old, _val = *_ptr; \
448 _old = cmpxchg(_ptr, _val, _val | _mask); \
456 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
458 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
459 * this avoids any races wrt polling state changes and thereby avoids
462 static bool set_nr_and_not_polling(struct task_struct *p)
464 struct thread_info *ti = task_thread_info(p);
465 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
469 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
471 * If this returns true, then the idle task promises to call
472 * sched_ttwu_pending() and reschedule soon.
474 static bool set_nr_if_polling(struct task_struct *p)
476 struct thread_info *ti = task_thread_info(p);
477 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
480 if (!(val & _TIF_POLLING_NRFLAG))
482 if (val & _TIF_NEED_RESCHED)
484 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
493 static bool set_nr_and_not_polling(struct task_struct *p)
495 set_tsk_need_resched(p);
500 static bool set_nr_if_polling(struct task_struct *p)
507 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
509 struct wake_q_node *node = &task->wake_q;
512 * Atomically grab the task, if ->wake_q is !nil already it means
513 * it's already queued (either by us or someone else) and will get the
514 * wakeup due to that.
516 * In order to ensure that a pending wakeup will observe our pending
517 * state, even in the failed case, an explicit smp_mb() must be used.
519 smp_mb__before_atomic();
520 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
524 * The head is context local, there can be no concurrency.
527 head->lastp = &node->next;
532 * wake_q_add() - queue a wakeup for 'later' waking.
533 * @head: the wake_q_head to add @task to
534 * @task: the task to queue for 'later' wakeup
536 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
537 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
540 * This function must be used as-if it were wake_up_process(); IOW the task
541 * must be ready to be woken at this location.
543 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
545 if (__wake_q_add(head, task))
546 get_task_struct(task);
550 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
551 * @head: the wake_q_head to add @task to
552 * @task: the task to queue for 'later' wakeup
554 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
555 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
558 * This function must be used as-if it were wake_up_process(); IOW the task
559 * must be ready to be woken at this location.
561 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
562 * that already hold reference to @task can call the 'safe' version and trust
563 * wake_q to do the right thing depending whether or not the @task is already
566 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
568 if (!__wake_q_add(head, task))
569 put_task_struct(task);
572 void wake_up_q(struct wake_q_head *head)
574 struct wake_q_node *node = head->first;
576 while (node != WAKE_Q_TAIL) {
577 struct task_struct *task;
579 task = container_of(node, struct task_struct, wake_q);
581 /* Task can safely be re-inserted now: */
583 task->wake_q.next = NULL;
586 * wake_up_process() executes a full barrier, which pairs with
587 * the queueing in wake_q_add() so as not to miss wakeups.
589 wake_up_process(task);
590 put_task_struct(task);
595 * resched_curr - mark rq's current task 'to be rescheduled now'.
597 * On UP this means the setting of the need_resched flag, on SMP it
598 * might also involve a cross-CPU call to trigger the scheduler on
601 void resched_curr(struct rq *rq)
603 struct task_struct *curr = rq->curr;
606 lockdep_assert_held(&rq->lock);
608 if (test_tsk_need_resched(curr))
613 if (cpu == smp_processor_id()) {
614 set_tsk_need_resched(curr);
615 set_preempt_need_resched();
619 if (set_nr_and_not_polling(curr))
620 smp_send_reschedule(cpu);
622 trace_sched_wake_idle_without_ipi(cpu);
625 void resched_cpu(int cpu)
627 struct rq *rq = cpu_rq(cpu);
630 raw_spin_lock_irqsave(&rq->lock, flags);
631 if (cpu_online(cpu) || cpu == smp_processor_id())
633 raw_spin_unlock_irqrestore(&rq->lock, flags);
637 #ifdef CONFIG_NO_HZ_COMMON
639 * In the semi idle case, use the nearest busy CPU for migrating timers
640 * from an idle CPU. This is good for power-savings.
642 * We don't do similar optimization for completely idle system, as
643 * selecting an idle CPU will add more delays to the timers than intended
644 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
646 int get_nohz_timer_target(void)
648 int i, cpu = smp_processor_id(), default_cpu = -1;
649 struct sched_domain *sd;
651 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
658 for_each_domain(cpu, sd) {
659 for_each_cpu_and(i, sched_domain_span(sd),
660 housekeeping_cpumask(HK_FLAG_TIMER)) {
671 if (default_cpu == -1)
672 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
680 * When add_timer_on() enqueues a timer into the timer wheel of an
681 * idle CPU then this timer might expire before the next timer event
682 * which is scheduled to wake up that CPU. In case of a completely
683 * idle system the next event might even be infinite time into the
684 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
685 * leaves the inner idle loop so the newly added timer is taken into
686 * account when the CPU goes back to idle and evaluates the timer
687 * wheel for the next timer event.
689 static void wake_up_idle_cpu(int cpu)
691 struct rq *rq = cpu_rq(cpu);
693 if (cpu == smp_processor_id())
696 if (set_nr_and_not_polling(rq->idle))
697 smp_send_reschedule(cpu);
699 trace_sched_wake_idle_without_ipi(cpu);
702 static bool wake_up_full_nohz_cpu(int cpu)
705 * We just need the target to call irq_exit() and re-evaluate
706 * the next tick. The nohz full kick at least implies that.
707 * If needed we can still optimize that later with an
710 if (cpu_is_offline(cpu))
711 return true; /* Don't try to wake offline CPUs. */
712 if (tick_nohz_full_cpu(cpu)) {
713 if (cpu != smp_processor_id() ||
714 tick_nohz_tick_stopped())
715 tick_nohz_full_kick_cpu(cpu);
723 * Wake up the specified CPU. If the CPU is going offline, it is the
724 * caller's responsibility to deal with the lost wakeup, for example,
725 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
727 void wake_up_nohz_cpu(int cpu)
729 if (!wake_up_full_nohz_cpu(cpu))
730 wake_up_idle_cpu(cpu);
733 static void nohz_csd_func(void *info)
735 struct rq *rq = info;
736 int cpu = cpu_of(rq);
740 * Release the rq::nohz_csd.
742 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
743 WARN_ON(!(flags & NOHZ_KICK_MASK));
745 rq->idle_balance = idle_cpu(cpu);
746 if (rq->idle_balance && !need_resched()) {
747 rq->nohz_idle_balance = flags;
748 raise_softirq_irqoff(SCHED_SOFTIRQ);
752 #endif /* CONFIG_NO_HZ_COMMON */
754 #ifdef CONFIG_NO_HZ_FULL
755 bool sched_can_stop_tick(struct rq *rq)
759 /* Deadline tasks, even if single, need the tick */
760 if (rq->dl.dl_nr_running)
764 * If there are more than one RR tasks, we need the tick to affect the
765 * actual RR behaviour.
767 if (rq->rt.rr_nr_running) {
768 if (rq->rt.rr_nr_running == 1)
775 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
776 * forced preemption between FIFO tasks.
778 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
783 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
784 * if there's more than one we need the tick for involuntary
787 if (rq->nr_running > 1)
792 #endif /* CONFIG_NO_HZ_FULL */
793 #endif /* CONFIG_SMP */
795 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
796 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
798 * Iterate task_group tree rooted at *from, calling @down when first entering a
799 * node and @up when leaving it for the final time.
801 * Caller must hold rcu_lock or sufficient equivalent.
803 int walk_tg_tree_from(struct task_group *from,
804 tg_visitor down, tg_visitor up, void *data)
806 struct task_group *parent, *child;
812 ret = (*down)(parent, data);
815 list_for_each_entry_rcu(child, &parent->children, siblings) {
822 ret = (*up)(parent, data);
823 if (ret || parent == from)
827 parent = parent->parent;
834 int tg_nop(struct task_group *tg, void *data)
840 static void set_load_weight(struct task_struct *p, bool update_load)
842 int prio = p->static_prio - MAX_RT_PRIO;
843 struct load_weight *load = &p->se.load;
846 * SCHED_IDLE tasks get minimal weight:
848 if (task_has_idle_policy(p)) {
849 load->weight = scale_load(WEIGHT_IDLEPRIO);
850 load->inv_weight = WMULT_IDLEPRIO;
855 * SCHED_OTHER tasks have to update their load when changing their
858 if (update_load && p->sched_class == &fair_sched_class) {
859 reweight_task(p, prio);
861 load->weight = scale_load(sched_prio_to_weight[prio]);
862 load->inv_weight = sched_prio_to_wmult[prio];
866 #ifdef CONFIG_UCLAMP_TASK
868 * Serializes updates of utilization clamp values
870 * The (slow-path) user-space triggers utilization clamp value updates which
871 * can require updates on (fast-path) scheduler's data structures used to
872 * support enqueue/dequeue operations.
873 * While the per-CPU rq lock protects fast-path update operations, user-space
874 * requests are serialized using a mutex to reduce the risk of conflicting
875 * updates or API abuses.
877 static DEFINE_MUTEX(uclamp_mutex);
879 /* Max allowed minimum utilization */
880 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
882 /* Max allowed maximum utilization */
883 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
886 * By default RT tasks run at the maximum performance point/capacity of the
887 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
888 * SCHED_CAPACITY_SCALE.
890 * This knob allows admins to change the default behavior when uclamp is being
891 * used. In battery powered devices, particularly, running at the maximum
892 * capacity and frequency will increase energy consumption and shorten the
895 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
897 * This knob will not override the system default sched_util_clamp_min defined
900 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
902 /* All clamps are required to be less or equal than these values */
903 static struct uclamp_se uclamp_default[UCLAMP_CNT];
906 * This static key is used to reduce the uclamp overhead in the fast path. It
907 * primarily disables the call to uclamp_rq_{inc, dec}() in
908 * enqueue/dequeue_task().
910 * This allows users to continue to enable uclamp in their kernel config with
911 * minimum uclamp overhead in the fast path.
913 * As soon as userspace modifies any of the uclamp knobs, the static key is
914 * enabled, since we have an actual users that make use of uclamp
917 * The knobs that would enable this static key are:
919 * * A task modifying its uclamp value with sched_setattr().
920 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
921 * * An admin modifying the cgroup cpu.uclamp.{min, max}
923 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
925 /* Integer rounded range for each bucket */
926 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
928 #define for_each_clamp_id(clamp_id) \
929 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
931 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
933 return clamp_value / UCLAMP_BUCKET_DELTA;
936 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
938 if (clamp_id == UCLAMP_MIN)
940 return SCHED_CAPACITY_SCALE;
943 static inline void uclamp_se_set(struct uclamp_se *uc_se,
944 unsigned int value, bool user_defined)
946 uc_se->value = value;
947 uc_se->bucket_id = uclamp_bucket_id(value);
948 uc_se->user_defined = user_defined;
951 static inline unsigned int
952 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
953 unsigned int clamp_value)
956 * Avoid blocked utilization pushing up the frequency when we go
957 * idle (which drops the max-clamp) by retaining the last known
960 if (clamp_id == UCLAMP_MAX) {
961 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
965 return uclamp_none(UCLAMP_MIN);
968 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
969 unsigned int clamp_value)
971 /* Reset max-clamp retention only on idle exit */
972 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
975 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
979 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
980 unsigned int clamp_value)
982 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
983 int bucket_id = UCLAMP_BUCKETS - 1;
986 * Since both min and max clamps are max aggregated, find the
987 * top most bucket with tasks in.
989 for ( ; bucket_id >= 0; bucket_id--) {
990 if (!bucket[bucket_id].tasks)
992 return bucket[bucket_id].value;
995 /* No tasks -- default clamp values */
996 return uclamp_idle_value(rq, clamp_id, clamp_value);
999 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1001 unsigned int default_util_min;
1002 struct uclamp_se *uc_se;
1004 lockdep_assert_held(&p->pi_lock);
1006 uc_se = &p->uclamp_req[UCLAMP_MIN];
1008 /* Only sync if user didn't override the default */
1009 if (uc_se->user_defined)
1012 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1013 uclamp_se_set(uc_se, default_util_min, false);
1016 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1024 /* Protect updates to p->uclamp_* */
1025 rq = task_rq_lock(p, &rf);
1026 __uclamp_update_util_min_rt_default(p);
1027 task_rq_unlock(rq, p, &rf);
1030 static void uclamp_sync_util_min_rt_default(void)
1032 struct task_struct *g, *p;
1035 * copy_process() sysctl_uclamp
1036 * uclamp_min_rt = X;
1037 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1038 * // link thread smp_mb__after_spinlock()
1039 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1040 * sched_post_fork() for_each_process_thread()
1041 * __uclamp_sync_rt() __uclamp_sync_rt()
1043 * Ensures that either sched_post_fork() will observe the new
1044 * uclamp_min_rt or for_each_process_thread() will observe the new
1047 read_lock(&tasklist_lock);
1048 smp_mb__after_spinlock();
1049 read_unlock(&tasklist_lock);
1052 for_each_process_thread(g, p)
1053 uclamp_update_util_min_rt_default(p);
1057 static inline struct uclamp_se
1058 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1060 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1061 #ifdef CONFIG_UCLAMP_TASK_GROUP
1062 struct uclamp_se uc_max;
1065 * Tasks in autogroups or root task group will be
1066 * restricted by system defaults.
1068 if (task_group_is_autogroup(task_group(p)))
1070 if (task_group(p) == &root_task_group)
1073 uc_max = task_group(p)->uclamp[clamp_id];
1074 if (uc_req.value > uc_max.value || !uc_req.user_defined)
1082 * The effective clamp bucket index of a task depends on, by increasing
1084 * - the task specific clamp value, when explicitly requested from userspace
1085 * - the task group effective clamp value, for tasks not either in the root
1086 * group or in an autogroup
1087 * - the system default clamp value, defined by the sysadmin
1089 static inline struct uclamp_se
1090 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1092 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1093 struct uclamp_se uc_max = uclamp_default[clamp_id];
1095 /* System default restrictions always apply */
1096 if (unlikely(uc_req.value > uc_max.value))
1102 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1104 struct uclamp_se uc_eff;
1106 /* Task currently refcounted: use back-annotated (effective) value */
1107 if (p->uclamp[clamp_id].active)
1108 return (unsigned long)p->uclamp[clamp_id].value;
1110 uc_eff = uclamp_eff_get(p, clamp_id);
1112 return (unsigned long)uc_eff.value;
1116 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1117 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1118 * updates the rq's clamp value if required.
1120 * Tasks can have a task-specific value requested from user-space, track
1121 * within each bucket the maximum value for tasks refcounted in it.
1122 * This "local max aggregation" allows to track the exact "requested" value
1123 * for each bucket when all its RUNNABLE tasks require the same clamp.
1125 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1126 enum uclamp_id clamp_id)
1128 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1129 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1130 struct uclamp_bucket *bucket;
1132 lockdep_assert_held(&rq->lock);
1134 /* Update task effective clamp */
1135 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1137 bucket = &uc_rq->bucket[uc_se->bucket_id];
1139 uc_se->active = true;
1141 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1144 * Local max aggregation: rq buckets always track the max
1145 * "requested" clamp value of its RUNNABLE tasks.
1147 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1148 bucket->value = uc_se->value;
1150 if (uc_se->value > READ_ONCE(uc_rq->value))
1151 WRITE_ONCE(uc_rq->value, uc_se->value);
1155 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1156 * is released. If this is the last task reference counting the rq's max
1157 * active clamp value, then the rq's clamp value is updated.
1159 * Both refcounted tasks and rq's cached clamp values are expected to be
1160 * always valid. If it's detected they are not, as defensive programming,
1161 * enforce the expected state and warn.
1163 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1164 enum uclamp_id clamp_id)
1166 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1167 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1168 struct uclamp_bucket *bucket;
1169 unsigned int bkt_clamp;
1170 unsigned int rq_clamp;
1172 lockdep_assert_held(&rq->lock);
1175 * If sched_uclamp_used was enabled after task @p was enqueued,
1176 * we could end up with unbalanced call to uclamp_rq_dec_id().
1178 * In this case the uc_se->active flag should be false since no uclamp
1179 * accounting was performed at enqueue time and we can just return
1182 * Need to be careful of the following enqueue/dequeue ordering
1186 * // sched_uclamp_used gets enabled
1189 * // Must not decrement bucket->tasks here
1192 * where we could end up with stale data in uc_se and
1193 * bucket[uc_se->bucket_id].
1195 * The following check here eliminates the possibility of such race.
1197 if (unlikely(!uc_se->active))
1200 bucket = &uc_rq->bucket[uc_se->bucket_id];
1202 SCHED_WARN_ON(!bucket->tasks);
1203 if (likely(bucket->tasks))
1206 uc_se->active = false;
1209 * Keep "local max aggregation" simple and accept to (possibly)
1210 * overboost some RUNNABLE tasks in the same bucket.
1211 * The rq clamp bucket value is reset to its base value whenever
1212 * there are no more RUNNABLE tasks refcounting it.
1214 if (likely(bucket->tasks))
1217 rq_clamp = READ_ONCE(uc_rq->value);
1219 * Defensive programming: this should never happen. If it happens,
1220 * e.g. due to future modification, warn and fixup the expected value.
1222 SCHED_WARN_ON(bucket->value > rq_clamp);
1223 if (bucket->value >= rq_clamp) {
1224 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1225 WRITE_ONCE(uc_rq->value, bkt_clamp);
1229 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1231 enum uclamp_id clamp_id;
1234 * Avoid any overhead until uclamp is actually used by the userspace.
1236 * The condition is constructed such that a NOP is generated when
1237 * sched_uclamp_used is disabled.
1239 if (!static_branch_unlikely(&sched_uclamp_used))
1242 if (unlikely(!p->sched_class->uclamp_enabled))
1245 for_each_clamp_id(clamp_id)
1246 uclamp_rq_inc_id(rq, p, clamp_id);
1248 /* Reset clamp idle holding when there is one RUNNABLE task */
1249 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1250 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1253 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1255 enum uclamp_id clamp_id;
1258 * Avoid any overhead until uclamp is actually used by the userspace.
1260 * The condition is constructed such that a NOP is generated when
1261 * sched_uclamp_used is disabled.
1263 if (!static_branch_unlikely(&sched_uclamp_used))
1266 if (unlikely(!p->sched_class->uclamp_enabled))
1269 for_each_clamp_id(clamp_id)
1270 uclamp_rq_dec_id(rq, p, clamp_id);
1274 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1280 * Lock the task and the rq where the task is (or was) queued.
1282 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1283 * price to pay to safely serialize util_{min,max} updates with
1284 * enqueues, dequeues and migration operations.
1285 * This is the same locking schema used by __set_cpus_allowed_ptr().
1287 rq = task_rq_lock(p, &rf);
1290 * Setting the clamp bucket is serialized by task_rq_lock().
1291 * If the task is not yet RUNNABLE and its task_struct is not
1292 * affecting a valid clamp bucket, the next time it's enqueued,
1293 * it will already see the updated clamp bucket value.
1295 if (p->uclamp[clamp_id].active) {
1296 uclamp_rq_dec_id(rq, p, clamp_id);
1297 uclamp_rq_inc_id(rq, p, clamp_id);
1300 task_rq_unlock(rq, p, &rf);
1303 #ifdef CONFIG_UCLAMP_TASK_GROUP
1305 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1306 unsigned int clamps)
1308 enum uclamp_id clamp_id;
1309 struct css_task_iter it;
1310 struct task_struct *p;
1312 css_task_iter_start(css, 0, &it);
1313 while ((p = css_task_iter_next(&it))) {
1314 for_each_clamp_id(clamp_id) {
1315 if ((0x1 << clamp_id) & clamps)
1316 uclamp_update_active(p, clamp_id);
1319 css_task_iter_end(&it);
1322 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1323 static void uclamp_update_root_tg(void)
1325 struct task_group *tg = &root_task_group;
1327 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1328 sysctl_sched_uclamp_util_min, false);
1329 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1330 sysctl_sched_uclamp_util_max, false);
1333 cpu_util_update_eff(&root_task_group.css);
1337 static void uclamp_update_root_tg(void) { }
1340 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1341 void *buffer, size_t *lenp, loff_t *ppos)
1343 bool update_root_tg = false;
1344 int old_min, old_max, old_min_rt;
1347 mutex_lock(&uclamp_mutex);
1348 old_min = sysctl_sched_uclamp_util_min;
1349 old_max = sysctl_sched_uclamp_util_max;
1350 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1352 result = proc_dointvec(table, write, buffer, lenp, ppos);
1358 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1359 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1360 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1366 if (old_min != sysctl_sched_uclamp_util_min) {
1367 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1368 sysctl_sched_uclamp_util_min, false);
1369 update_root_tg = true;
1371 if (old_max != sysctl_sched_uclamp_util_max) {
1372 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1373 sysctl_sched_uclamp_util_max, false);
1374 update_root_tg = true;
1377 if (update_root_tg) {
1378 static_branch_enable(&sched_uclamp_used);
1379 uclamp_update_root_tg();
1382 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1383 static_branch_enable(&sched_uclamp_used);
1384 uclamp_sync_util_min_rt_default();
1388 * We update all RUNNABLE tasks only when task groups are in use.
1389 * Otherwise, keep it simple and do just a lazy update at each next
1390 * task enqueue time.
1396 sysctl_sched_uclamp_util_min = old_min;
1397 sysctl_sched_uclamp_util_max = old_max;
1398 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1400 mutex_unlock(&uclamp_mutex);
1405 static int uclamp_validate(struct task_struct *p,
1406 const struct sched_attr *attr)
1408 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1409 int util_max = p->uclamp_req[UCLAMP_MAX].value;
1411 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1412 util_min = attr->sched_util_min;
1414 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1418 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1419 util_max = attr->sched_util_max;
1421 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1425 if (util_min != -1 && util_max != -1 && util_min > util_max)
1429 * We have valid uclamp attributes; make sure uclamp is enabled.
1431 * We need to do that here, because enabling static branches is a
1432 * blocking operation which obviously cannot be done while holding
1435 static_branch_enable(&sched_uclamp_used);
1440 static bool uclamp_reset(const struct sched_attr *attr,
1441 enum uclamp_id clamp_id,
1442 struct uclamp_se *uc_se)
1444 /* Reset on sched class change for a non user-defined clamp value. */
1445 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1446 !uc_se->user_defined)
1449 /* Reset on sched_util_{min,max} == -1. */
1450 if (clamp_id == UCLAMP_MIN &&
1451 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1452 attr->sched_util_min == -1) {
1456 if (clamp_id == UCLAMP_MAX &&
1457 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1458 attr->sched_util_max == -1) {
1465 static void __setscheduler_uclamp(struct task_struct *p,
1466 const struct sched_attr *attr)
1468 enum uclamp_id clamp_id;
1470 for_each_clamp_id(clamp_id) {
1471 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1474 if (!uclamp_reset(attr, clamp_id, uc_se))
1478 * RT by default have a 100% boost value that could be modified
1481 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1482 value = sysctl_sched_uclamp_util_min_rt_default;
1484 value = uclamp_none(clamp_id);
1486 uclamp_se_set(uc_se, value, false);
1490 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1493 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1494 attr->sched_util_min != -1) {
1495 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1496 attr->sched_util_min, true);
1499 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1500 attr->sched_util_max != -1) {
1501 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1502 attr->sched_util_max, true);
1506 static void uclamp_fork(struct task_struct *p)
1508 enum uclamp_id clamp_id;
1511 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1512 * as the task is still at its early fork stages.
1514 for_each_clamp_id(clamp_id)
1515 p->uclamp[clamp_id].active = false;
1517 if (likely(!p->sched_reset_on_fork))
1520 for_each_clamp_id(clamp_id) {
1521 uclamp_se_set(&p->uclamp_req[clamp_id],
1522 uclamp_none(clamp_id), false);
1526 static void uclamp_post_fork(struct task_struct *p)
1528 uclamp_update_util_min_rt_default(p);
1531 static void __init init_uclamp_rq(struct rq *rq)
1533 enum uclamp_id clamp_id;
1534 struct uclamp_rq *uc_rq = rq->uclamp;
1536 for_each_clamp_id(clamp_id) {
1537 uc_rq[clamp_id] = (struct uclamp_rq) {
1538 .value = uclamp_none(clamp_id)
1542 rq->uclamp_flags = 0;
1545 static void __init init_uclamp(void)
1547 struct uclamp_se uc_max = {};
1548 enum uclamp_id clamp_id;
1551 for_each_possible_cpu(cpu)
1552 init_uclamp_rq(cpu_rq(cpu));
1554 for_each_clamp_id(clamp_id) {
1555 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1556 uclamp_none(clamp_id), false);
1559 /* System defaults allow max clamp values for both indexes */
1560 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1561 for_each_clamp_id(clamp_id) {
1562 uclamp_default[clamp_id] = uc_max;
1563 #ifdef CONFIG_UCLAMP_TASK_GROUP
1564 root_task_group.uclamp_req[clamp_id] = uc_max;
1565 root_task_group.uclamp[clamp_id] = uc_max;
1570 #else /* CONFIG_UCLAMP_TASK */
1571 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1572 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1573 static inline int uclamp_validate(struct task_struct *p,
1574 const struct sched_attr *attr)
1578 static void __setscheduler_uclamp(struct task_struct *p,
1579 const struct sched_attr *attr) { }
1580 static inline void uclamp_fork(struct task_struct *p) { }
1581 static inline void uclamp_post_fork(struct task_struct *p) { }
1582 static inline void init_uclamp(void) { }
1583 #endif /* CONFIG_UCLAMP_TASK */
1585 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1587 if (!(flags & ENQUEUE_NOCLOCK))
1588 update_rq_clock(rq);
1590 if (!(flags & ENQUEUE_RESTORE)) {
1591 sched_info_queued(rq, p);
1592 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1595 uclamp_rq_inc(rq, p);
1596 p->sched_class->enqueue_task(rq, p, flags);
1599 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1601 if (!(flags & DEQUEUE_NOCLOCK))
1602 update_rq_clock(rq);
1604 if (!(flags & DEQUEUE_SAVE)) {
1605 sched_info_dequeued(rq, p);
1606 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1609 uclamp_rq_dec(rq, p);
1610 p->sched_class->dequeue_task(rq, p, flags);
1613 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1615 enqueue_task(rq, p, flags);
1617 p->on_rq = TASK_ON_RQ_QUEUED;
1620 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1622 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1624 dequeue_task(rq, p, flags);
1628 * __normal_prio - return the priority that is based on the static prio
1630 static inline int __normal_prio(struct task_struct *p)
1632 return p->static_prio;
1636 * Calculate the expected normal priority: i.e. priority
1637 * without taking RT-inheritance into account. Might be
1638 * boosted by interactivity modifiers. Changes upon fork,
1639 * setprio syscalls, and whenever the interactivity
1640 * estimator recalculates.
1642 static inline int normal_prio(struct task_struct *p)
1646 if (task_has_dl_policy(p))
1647 prio = MAX_DL_PRIO-1;
1648 else if (task_has_rt_policy(p))
1649 prio = MAX_RT_PRIO-1 - p->rt_priority;
1651 prio = __normal_prio(p);
1656 * Calculate the current priority, i.e. the priority
1657 * taken into account by the scheduler. This value might
1658 * be boosted by RT tasks, or might be boosted by
1659 * interactivity modifiers. Will be RT if the task got
1660 * RT-boosted. If not then it returns p->normal_prio.
1662 static int effective_prio(struct task_struct *p)
1664 p->normal_prio = normal_prio(p);
1666 * If we are RT tasks or we were boosted to RT priority,
1667 * keep the priority unchanged. Otherwise, update priority
1668 * to the normal priority:
1670 if (!rt_prio(p->prio))
1671 return p->normal_prio;
1676 * task_curr - is this task currently executing on a CPU?
1677 * @p: the task in question.
1679 * Return: 1 if the task is currently executing. 0 otherwise.
1681 inline int task_curr(const struct task_struct *p)
1683 return cpu_curr(task_cpu(p)) == p;
1687 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1688 * use the balance_callback list if you want balancing.
1690 * this means any call to check_class_changed() must be followed by a call to
1691 * balance_callback().
1693 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1694 const struct sched_class *prev_class,
1697 if (prev_class != p->sched_class) {
1698 if (prev_class->switched_from)
1699 prev_class->switched_from(rq, p);
1701 p->sched_class->switched_to(rq, p);
1702 } else if (oldprio != p->prio || dl_task(p))
1703 p->sched_class->prio_changed(rq, p, oldprio);
1706 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1708 if (p->sched_class == rq->curr->sched_class)
1709 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1710 else if (p->sched_class > rq->curr->sched_class)
1714 * A queue event has occurred, and we're going to schedule. In
1715 * this case, we can save a useless back to back clock update.
1717 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1718 rq_clock_skip_update(rq);
1724 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
1726 static int __set_cpus_allowed_ptr(struct task_struct *p,
1727 const struct cpumask *new_mask,
1730 static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
1732 if (likely(!p->migration_disabled))
1735 if (p->cpus_ptr != &p->cpus_mask)
1739 * Violates locking rules! see comment in __do_set_cpus_allowed().
1741 __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
1744 void migrate_disable(void)
1746 struct task_struct *p = current;
1748 if (p->migration_disabled) {
1749 p->migration_disabled++;
1754 this_rq()->nr_pinned++;
1755 p->migration_disabled = 1;
1758 EXPORT_SYMBOL_GPL(migrate_disable);
1760 void migrate_enable(void)
1762 struct task_struct *p = current;
1764 if (p->migration_disabled > 1) {
1765 p->migration_disabled--;
1770 * Ensure stop_task runs either before or after this, and that
1771 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
1774 if (p->cpus_ptr != &p->cpus_mask)
1775 __set_cpus_allowed_ptr(p, &p->cpus_mask, SCA_MIGRATE_ENABLE);
1777 * Mustn't clear migration_disabled() until cpus_ptr points back at the
1778 * regular cpus_mask, otherwise things that race (eg.
1779 * select_fallback_rq) get confused.
1782 p->migration_disabled = 0;
1783 this_rq()->nr_pinned--;
1786 EXPORT_SYMBOL_GPL(migrate_enable);
1788 static inline bool rq_has_pinned_tasks(struct rq *rq)
1790 return rq->nr_pinned;
1794 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1795 * __set_cpus_allowed_ptr() and select_fallback_rq().
1797 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1799 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1802 if (is_per_cpu_kthread(p) || is_migration_disabled(p))
1803 return cpu_online(cpu);
1805 return cpu_active(cpu);
1809 * This is how migration works:
1811 * 1) we invoke migration_cpu_stop() on the target CPU using
1813 * 2) stopper starts to run (implicitly forcing the migrated thread
1815 * 3) it checks whether the migrated task is still in the wrong runqueue.
1816 * 4) if it's in the wrong runqueue then the migration thread removes
1817 * it and puts it into the right queue.
1818 * 5) stopper completes and stop_one_cpu() returns and the migration
1823 * move_queued_task - move a queued task to new rq.
1825 * Returns (locked) new rq. Old rq's lock is released.
1827 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1828 struct task_struct *p, int new_cpu)
1830 lockdep_assert_held(&rq->lock);
1832 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
1833 set_task_cpu(p, new_cpu);
1836 rq = cpu_rq(new_cpu);
1839 BUG_ON(task_cpu(p) != new_cpu);
1840 activate_task(rq, p, 0);
1841 check_preempt_curr(rq, p, 0);
1846 struct migration_arg {
1847 struct task_struct *task;
1849 struct set_affinity_pending *pending;
1852 struct set_affinity_pending {
1854 struct completion done;
1855 struct cpu_stop_work stop_work;
1856 struct migration_arg arg;
1860 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1861 * this because either it can't run here any more (set_cpus_allowed()
1862 * away from this CPU, or CPU going down), or because we're
1863 * attempting to rebalance this task on exec (sched_exec).
1865 * So we race with normal scheduler movements, but that's OK, as long
1866 * as the task is no longer on this CPU.
1868 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1869 struct task_struct *p, int dest_cpu)
1871 /* Affinity changed (again). */
1872 if (!is_cpu_allowed(p, dest_cpu))
1875 update_rq_clock(rq);
1876 rq = move_queued_task(rq, rf, p, dest_cpu);
1882 * migration_cpu_stop - this will be executed by a highprio stopper thread
1883 * and performs thread migration by bumping thread off CPU then
1884 * 'pushing' onto another runqueue.
1886 static int migration_cpu_stop(void *data)
1888 struct set_affinity_pending *pending;
1889 struct migration_arg *arg = data;
1890 struct task_struct *p = arg->task;
1891 int dest_cpu = arg->dest_cpu;
1892 struct rq *rq = this_rq();
1893 bool complete = false;
1897 * The original target CPU might have gone down and we might
1898 * be on another CPU but it doesn't matter.
1900 local_irq_save(rf.flags);
1902 * We need to explicitly wake pending tasks before running
1903 * __migrate_task() such that we will not miss enforcing cpus_ptr
1904 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1906 flush_smp_call_function_from_idle();
1908 raw_spin_lock(&p->pi_lock);
1911 pending = p->migration_pending;
1913 * If task_rq(p) != rq, it cannot be migrated here, because we're
1914 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1915 * we're holding p->pi_lock.
1917 if (task_rq(p) == rq) {
1918 if (is_migration_disabled(p))
1922 p->migration_pending = NULL;
1926 /* migrate_enable() -- we must not race against SCA */
1929 * When this was migrate_enable() but we no longer
1930 * have a @pending, a concurrent SCA 'fixed' things
1931 * and we should be valid again. Nothing to do.
1934 WARN_ON_ONCE(!cpumask_test_cpu(task_cpu(p), &p->cpus_mask));
1938 dest_cpu = cpumask_any_distribute(&p->cpus_mask);
1941 if (task_on_rq_queued(p))
1942 rq = __migrate_task(rq, &rf, p, dest_cpu);
1944 p->wake_cpu = dest_cpu;
1946 } else if (dest_cpu < 0 || pending) {
1948 * This happens when we get migrated between migrate_enable()'s
1949 * preempt_enable() and scheduling the stopper task. At that
1950 * point we're a regular task again and not current anymore.
1952 * A !PREEMPT kernel has a giant hole here, which makes it far
1957 * The task moved before the stopper got to run. We're holding
1958 * ->pi_lock, so the allowed mask is stable - if it got
1959 * somewhere allowed, we're done.
1961 if (pending && cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
1962 p->migration_pending = NULL;
1968 * When this was migrate_enable() but we no longer have an
1969 * @pending, a concurrent SCA 'fixed' things and we should be
1970 * valid again. Nothing to do.
1973 WARN_ON_ONCE(!cpumask_test_cpu(task_cpu(p), &p->cpus_mask));
1978 * When migrate_enable() hits a rq mis-match we can't reliably
1979 * determine is_migration_disabled() and so have to chase after
1982 task_rq_unlock(rq, p, &rf);
1983 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
1984 &pending->arg, &pending->stop_work);
1988 task_rq_unlock(rq, p, &rf);
1991 complete_all(&pending->done);
1993 /* For pending->{arg,stop_work} */
1994 pending = arg->pending;
1995 if (pending && refcount_dec_and_test(&pending->refs))
1996 wake_up_var(&pending->refs);
2001 int push_cpu_stop(void *arg)
2003 struct rq *lowest_rq = NULL, *rq = this_rq();
2004 struct task_struct *p = arg;
2006 raw_spin_lock_irq(&p->pi_lock);
2007 raw_spin_lock(&rq->lock);
2009 if (task_rq(p) != rq)
2012 if (is_migration_disabled(p)) {
2013 p->migration_flags |= MDF_PUSH;
2017 p->migration_flags &= ~MDF_PUSH;
2019 if (p->sched_class->find_lock_rq)
2020 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2025 // XXX validate p is still the highest prio task
2026 if (task_rq(p) == rq) {
2027 deactivate_task(rq, p, 0);
2028 set_task_cpu(p, lowest_rq->cpu);
2029 activate_task(lowest_rq, p, 0);
2030 resched_curr(lowest_rq);
2033 double_unlock_balance(rq, lowest_rq);
2036 rq->push_busy = false;
2037 raw_spin_unlock(&rq->lock);
2038 raw_spin_unlock_irq(&p->pi_lock);
2045 * sched_class::set_cpus_allowed must do the below, but is not required to
2046 * actually call this function.
2048 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2050 if (flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2051 p->cpus_ptr = new_mask;
2055 cpumask_copy(&p->cpus_mask, new_mask);
2056 p->nr_cpus_allowed = cpumask_weight(new_mask);
2060 __do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask, u32 flags)
2062 struct rq *rq = task_rq(p);
2063 bool queued, running;
2066 * This here violates the locking rules for affinity, since we're only
2067 * supposed to change these variables while holding both rq->lock and
2070 * HOWEVER, it magically works, because ttwu() is the only code that
2071 * accesses these variables under p->pi_lock and only does so after
2072 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2073 * before finish_task().
2075 * XXX do further audits, this smells like something putrid.
2077 if (flags & SCA_MIGRATE_DISABLE)
2078 SCHED_WARN_ON(!p->on_cpu);
2080 lockdep_assert_held(&p->pi_lock);
2082 queued = task_on_rq_queued(p);
2083 running = task_current(rq, p);
2087 * Because __kthread_bind() calls this on blocked tasks without
2090 lockdep_assert_held(&rq->lock);
2091 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
2094 put_prev_task(rq, p);
2096 p->sched_class->set_cpus_allowed(p, new_mask, flags);
2099 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
2101 set_next_task(rq, p);
2104 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2106 __do_set_cpus_allowed(p, new_mask, 0);
2110 * This function is wildly self concurrent; here be dragons.
2113 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2114 * designated task is enqueued on an allowed CPU. If that task is currently
2115 * running, we have to kick it out using the CPU stopper.
2117 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2120 * Initial conditions: P0->cpus_mask = [0, 1]
2124 * migrate_disable();
2126 * set_cpus_allowed_ptr(P0, [1]);
2128 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2129 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2130 * This means we need the following scheme:
2134 * migrate_disable();
2136 * set_cpus_allowed_ptr(P0, [1]);
2140 * __set_cpus_allowed_ptr();
2141 * <wakes local stopper>
2142 * `--> <woken on migration completion>
2144 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2145 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2146 * task p are serialized by p->pi_lock, which we can leverage: the one that
2147 * should come into effect at the end of the Migrate-Disable region is the last
2148 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2149 * but we still need to properly signal those waiting tasks at the appropriate
2152 * This is implemented using struct set_affinity_pending. The first
2153 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2154 * setup an instance of that struct and install it on the targeted task_struct.
2155 * Any and all further callers will reuse that instance. Those then wait for
2156 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2157 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2160 * (1) In the cases covered above. There is one more where the completion is
2161 * signaled within affine_move_task() itself: when a subsequent affinity request
2162 * cancels the need for an active migration. Consider:
2164 * Initial conditions: P0->cpus_mask = [0, 1]
2168 * migrate_disable();
2170 * set_cpus_allowed_ptr(P0, [1]);
2172 * set_cpus_allowed_ptr(P0, [0, 1]);
2173 * <signal completion>
2176 * Note that the above is safe vs a concurrent migrate_enable(), as any
2177 * pending affinity completion is preceded by an uninstallation of
2178 * p->migration_pending done with p->pi_lock held.
2180 static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2181 int dest_cpu, unsigned int flags)
2183 struct set_affinity_pending my_pending = { }, *pending = NULL;
2184 struct migration_arg arg = {
2186 .dest_cpu = dest_cpu,
2188 bool complete = false;
2190 /* Can the task run on the task's current CPU? If so, we're done */
2191 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
2192 struct task_struct *push_task = NULL;
2194 if ((flags & SCA_MIGRATE_ENABLE) &&
2195 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2196 rq->push_busy = true;
2197 push_task = get_task_struct(p);
2200 pending = p->migration_pending;
2202 refcount_inc(&pending->refs);
2203 p->migration_pending = NULL;
2206 task_rq_unlock(rq, p, rf);
2209 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2219 if (!(flags & SCA_MIGRATE_ENABLE)) {
2220 /* serialized by p->pi_lock */
2221 if (!p->migration_pending) {
2222 /* Install the request */
2223 refcount_set(&my_pending.refs, 1);
2224 init_completion(&my_pending.done);
2225 p->migration_pending = &my_pending;
2227 pending = p->migration_pending;
2228 refcount_inc(&pending->refs);
2231 pending = p->migration_pending;
2233 * - !MIGRATE_ENABLE:
2234 * we'll have installed a pending if there wasn't one already.
2237 * we're here because the current CPU isn't matching anymore,
2238 * the only way that can happen is because of a concurrent
2239 * set_cpus_allowed_ptr() call, which should then still be
2240 * pending completion.
2242 * Either way, we really should have a @pending here.
2244 if (WARN_ON_ONCE(!pending)) {
2245 task_rq_unlock(rq, p, rf);
2249 if (flags & SCA_MIGRATE_ENABLE) {
2251 refcount_inc(&pending->refs); /* pending->{arg,stop_work} */
2252 p->migration_flags &= ~MDF_PUSH;
2253 task_rq_unlock(rq, p, rf);
2255 pending->arg = (struct migration_arg) {
2261 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2262 &pending->arg, &pending->stop_work);
2267 if (task_running(rq, p) || p->state == TASK_WAKING) {
2269 * Lessen races (and headaches) by delegating
2270 * is_migration_disabled(p) checks to the stopper, which will
2271 * run on the same CPU as said p.
2273 task_rq_unlock(rq, p, rf);
2274 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
2278 if (!is_migration_disabled(p)) {
2279 if (task_on_rq_queued(p))
2280 rq = move_queued_task(rq, rf, p, dest_cpu);
2282 p->migration_pending = NULL;
2285 task_rq_unlock(rq, p, rf);
2289 complete_all(&pending->done);
2292 wait_for_completion(&pending->done);
2294 if (refcount_dec_and_test(&pending->refs))
2295 wake_up_var(&pending->refs);
2298 * Block the original owner of &pending until all subsequent callers
2299 * have seen the completion and decremented the refcount
2301 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2307 * Change a given task's CPU affinity. Migrate the thread to a
2308 * proper CPU and schedule it away if the CPU it's executing on
2309 * is removed from the allowed bitmask.
2311 * NOTE: the caller must have a valid reference to the task, the
2312 * task must not exit() & deallocate itself prematurely. The
2313 * call is not atomic; no spinlocks may be held.
2315 static int __set_cpus_allowed_ptr(struct task_struct *p,
2316 const struct cpumask *new_mask,
2319 const struct cpumask *cpu_valid_mask = cpu_active_mask;
2320 unsigned int dest_cpu;
2325 rq = task_rq_lock(p, &rf);
2326 update_rq_clock(rq);
2328 if (p->flags & PF_KTHREAD || is_migration_disabled(p)) {
2330 * Kernel threads are allowed on online && !active CPUs.
2332 * Specifically, migration_disabled() tasks must not fail the
2333 * cpumask_any_and_distribute() pick below, esp. so on
2334 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2335 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
2337 cpu_valid_mask = cpu_online_mask;
2341 * Must re-check here, to close a race against __kthread_bind(),
2342 * sched_setaffinity() is not guaranteed to observe the flag.
2344 if ((flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
2349 if (!(flags & SCA_MIGRATE_ENABLE)) {
2350 if (cpumask_equal(&p->cpus_mask, new_mask))
2353 if (WARN_ON_ONCE(p == current &&
2354 is_migration_disabled(p) &&
2355 !cpumask_test_cpu(task_cpu(p), new_mask))) {
2362 * Picking a ~random cpu helps in cases where we are changing affinity
2363 * for groups of tasks (ie. cpuset), so that load balancing is not
2364 * immediately required to distribute the tasks within their new mask.
2366 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
2367 if (dest_cpu >= nr_cpu_ids) {
2372 __do_set_cpus_allowed(p, new_mask, flags);
2374 if (p->flags & PF_KTHREAD) {
2376 * For kernel threads that do indeed end up on online &&
2377 * !active we want to ensure they are strict per-CPU threads.
2379 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
2380 !cpumask_intersects(new_mask, cpu_active_mask) &&
2381 p->nr_cpus_allowed != 1);
2384 return affine_move_task(rq, p, &rf, dest_cpu, flags);
2387 task_rq_unlock(rq, p, &rf);
2392 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
2394 return __set_cpus_allowed_ptr(p, new_mask, 0);
2396 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
2398 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2400 #ifdef CONFIG_SCHED_DEBUG
2402 * We should never call set_task_cpu() on a blocked task,
2403 * ttwu() will sort out the placement.
2405 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2409 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
2410 * because schedstat_wait_{start,end} rebase migrating task's wait_start
2411 * time relying on p->on_rq.
2413 WARN_ON_ONCE(p->state == TASK_RUNNING &&
2414 p->sched_class == &fair_sched_class &&
2415 (p->on_rq && !task_on_rq_migrating(p)));
2417 #ifdef CONFIG_LOCKDEP
2419 * The caller should hold either p->pi_lock or rq->lock, when changing
2420 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2422 * sched_move_task() holds both and thus holding either pins the cgroup,
2425 * Furthermore, all task_rq users should acquire both locks, see
2428 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2429 lockdep_is_held(&task_rq(p)->lock)));
2432 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
2434 WARN_ON_ONCE(!cpu_online(new_cpu));
2436 WARN_ON_ONCE(is_migration_disabled(p));
2439 trace_sched_migrate_task(p, new_cpu);
2441 if (task_cpu(p) != new_cpu) {
2442 if (p->sched_class->migrate_task_rq)
2443 p->sched_class->migrate_task_rq(p, new_cpu);
2444 p->se.nr_migrations++;
2446 perf_event_task_migrate(p);
2449 __set_task_cpu(p, new_cpu);
2452 #ifdef CONFIG_NUMA_BALANCING
2453 static void __migrate_swap_task(struct task_struct *p, int cpu)
2455 if (task_on_rq_queued(p)) {
2456 struct rq *src_rq, *dst_rq;
2457 struct rq_flags srf, drf;
2459 src_rq = task_rq(p);
2460 dst_rq = cpu_rq(cpu);
2462 rq_pin_lock(src_rq, &srf);
2463 rq_pin_lock(dst_rq, &drf);
2465 deactivate_task(src_rq, p, 0);
2466 set_task_cpu(p, cpu);
2467 activate_task(dst_rq, p, 0);
2468 check_preempt_curr(dst_rq, p, 0);
2470 rq_unpin_lock(dst_rq, &drf);
2471 rq_unpin_lock(src_rq, &srf);
2475 * Task isn't running anymore; make it appear like we migrated
2476 * it before it went to sleep. This means on wakeup we make the
2477 * previous CPU our target instead of where it really is.
2483 struct migration_swap_arg {
2484 struct task_struct *src_task, *dst_task;
2485 int src_cpu, dst_cpu;
2488 static int migrate_swap_stop(void *data)
2490 struct migration_swap_arg *arg = data;
2491 struct rq *src_rq, *dst_rq;
2494 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2497 src_rq = cpu_rq(arg->src_cpu);
2498 dst_rq = cpu_rq(arg->dst_cpu);
2500 double_raw_lock(&arg->src_task->pi_lock,
2501 &arg->dst_task->pi_lock);
2502 double_rq_lock(src_rq, dst_rq);
2504 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2507 if (task_cpu(arg->src_task) != arg->src_cpu)
2510 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2513 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2516 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2517 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2522 double_rq_unlock(src_rq, dst_rq);
2523 raw_spin_unlock(&arg->dst_task->pi_lock);
2524 raw_spin_unlock(&arg->src_task->pi_lock);
2530 * Cross migrate two tasks
2532 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2533 int target_cpu, int curr_cpu)
2535 struct migration_swap_arg arg;
2538 arg = (struct migration_swap_arg){
2540 .src_cpu = curr_cpu,
2542 .dst_cpu = target_cpu,
2545 if (arg.src_cpu == arg.dst_cpu)
2549 * These three tests are all lockless; this is OK since all of them
2550 * will be re-checked with proper locks held further down the line.
2552 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2555 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2558 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2561 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2562 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2567 #endif /* CONFIG_NUMA_BALANCING */
2570 * wait_task_inactive - wait for a thread to unschedule.
2572 * If @match_state is nonzero, it's the @p->state value just checked and
2573 * not expected to change. If it changes, i.e. @p might have woken up,
2574 * then return zero. When we succeed in waiting for @p to be off its CPU,
2575 * we return a positive number (its total switch count). If a second call
2576 * a short while later returns the same number, the caller can be sure that
2577 * @p has remained unscheduled the whole time.
2579 * The caller must ensure that the task *will* unschedule sometime soon,
2580 * else this function might spin for a *long* time. This function can't
2581 * be called with interrupts off, or it may introduce deadlock with
2582 * smp_call_function() if an IPI is sent by the same process we are
2583 * waiting to become inactive.
2585 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2587 int running, queued;
2594 * We do the initial early heuristics without holding
2595 * any task-queue locks at all. We'll only try to get
2596 * the runqueue lock when things look like they will
2602 * If the task is actively running on another CPU
2603 * still, just relax and busy-wait without holding
2606 * NOTE! Since we don't hold any locks, it's not
2607 * even sure that "rq" stays as the right runqueue!
2608 * But we don't care, since "task_running()" will
2609 * return false if the runqueue has changed and p
2610 * is actually now running somewhere else!
2612 while (task_running(rq, p)) {
2613 if (match_state && unlikely(p->state != match_state))
2619 * Ok, time to look more closely! We need the rq
2620 * lock now, to be *sure*. If we're wrong, we'll
2621 * just go back and repeat.
2623 rq = task_rq_lock(p, &rf);
2624 trace_sched_wait_task(p);
2625 running = task_running(rq, p);
2626 queued = task_on_rq_queued(p);
2628 if (!match_state || p->state == match_state)
2629 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2630 task_rq_unlock(rq, p, &rf);
2633 * If it changed from the expected state, bail out now.
2635 if (unlikely(!ncsw))
2639 * Was it really running after all now that we
2640 * checked with the proper locks actually held?
2642 * Oops. Go back and try again..
2644 if (unlikely(running)) {
2650 * It's not enough that it's not actively running,
2651 * it must be off the runqueue _entirely_, and not
2654 * So if it was still runnable (but just not actively
2655 * running right now), it's preempted, and we should
2656 * yield - it could be a while.
2658 if (unlikely(queued)) {
2659 ktime_t to = NSEC_PER_SEC / HZ;
2661 set_current_state(TASK_UNINTERRUPTIBLE);
2662 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2667 * Ahh, all good. It wasn't running, and it wasn't
2668 * runnable, which means that it will never become
2669 * running in the future either. We're all done!
2678 * kick_process - kick a running thread to enter/exit the kernel
2679 * @p: the to-be-kicked thread
2681 * Cause a process which is running on another CPU to enter
2682 * kernel-mode, without any delay. (to get signals handled.)
2684 * NOTE: this function doesn't have to take the runqueue lock,
2685 * because all it wants to ensure is that the remote task enters
2686 * the kernel. If the IPI races and the task has been migrated
2687 * to another CPU then no harm is done and the purpose has been
2690 void kick_process(struct task_struct *p)
2696 if ((cpu != smp_processor_id()) && task_curr(p))
2697 smp_send_reschedule(cpu);
2700 EXPORT_SYMBOL_GPL(kick_process);
2703 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2705 * A few notes on cpu_active vs cpu_online:
2707 * - cpu_active must be a subset of cpu_online
2709 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2710 * see __set_cpus_allowed_ptr(). At this point the newly online
2711 * CPU isn't yet part of the sched domains, and balancing will not
2714 * - on CPU-down we clear cpu_active() to mask the sched domains and
2715 * avoid the load balancer to place new tasks on the to be removed
2716 * CPU. Existing tasks will remain running there and will be taken
2719 * This means that fallback selection must not select !active CPUs.
2720 * And can assume that any active CPU must be online. Conversely
2721 * select_task_rq() below may allow selection of !active CPUs in order
2722 * to satisfy the above rules.
2724 static int select_fallback_rq(int cpu, struct task_struct *p)
2726 int nid = cpu_to_node(cpu);
2727 const struct cpumask *nodemask = NULL;
2728 enum { cpuset, possible, fail } state = cpuset;
2732 * If the node that the CPU is on has been offlined, cpu_to_node()
2733 * will return -1. There is no CPU on the node, and we should
2734 * select the CPU on the other node.
2737 nodemask = cpumask_of_node(nid);
2739 /* Look for allowed, online CPU in same node. */
2740 for_each_cpu(dest_cpu, nodemask) {
2741 if (!cpu_active(dest_cpu))
2743 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2749 /* Any allowed, online CPU? */
2750 for_each_cpu(dest_cpu, p->cpus_ptr) {
2751 if (!is_cpu_allowed(p, dest_cpu))
2757 /* No more Mr. Nice Guy. */
2760 if (IS_ENABLED(CONFIG_CPUSETS)) {
2761 cpuset_cpus_allowed_fallback(p);
2768 * XXX When called from select_task_rq() we only
2769 * hold p->pi_lock and again violate locking order.
2771 * More yuck to audit.
2773 do_set_cpus_allowed(p, cpu_possible_mask);
2784 if (state != cpuset) {
2786 * Don't tell them about moving exiting tasks or
2787 * kernel threads (both mm NULL), since they never
2790 if (p->mm && printk_ratelimit()) {
2791 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2792 task_pid_nr(p), p->comm, cpu);
2800 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2803 int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
2805 lockdep_assert_held(&p->pi_lock);
2807 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
2808 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
2810 cpu = cpumask_any(p->cpus_ptr);
2813 * In order not to call set_task_cpu() on a blocking task we need
2814 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2817 * Since this is common to all placement strategies, this lives here.
2819 * [ this allows ->select_task() to simply return task_cpu(p) and
2820 * not worry about this generic constraint ]
2822 if (unlikely(!is_cpu_allowed(p, cpu)))
2823 cpu = select_fallback_rq(task_cpu(p), p);
2828 void sched_set_stop_task(int cpu, struct task_struct *stop)
2830 static struct lock_class_key stop_pi_lock;
2831 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2832 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2836 * Make it appear like a SCHED_FIFO task, its something
2837 * userspace knows about and won't get confused about.
2839 * Also, it will make PI more or less work without too
2840 * much confusion -- but then, stop work should not
2841 * rely on PI working anyway.
2843 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2845 stop->sched_class = &stop_sched_class;
2848 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
2849 * adjust the effective priority of a task. As a result,
2850 * rt_mutex_setprio() can trigger (RT) balancing operations,
2851 * which can then trigger wakeups of the stop thread to push
2852 * around the current task.
2854 * The stop task itself will never be part of the PI-chain, it
2855 * never blocks, therefore that ->pi_lock recursion is safe.
2856 * Tell lockdep about this by placing the stop->pi_lock in its
2859 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
2862 cpu_rq(cpu)->stop = stop;
2866 * Reset it back to a normal scheduling class so that
2867 * it can die in pieces.
2869 old_stop->sched_class = &rt_sched_class;
2873 #else /* CONFIG_SMP */
2875 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2876 const struct cpumask *new_mask,
2879 return set_cpus_allowed_ptr(p, new_mask);
2882 static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
2884 static inline bool rq_has_pinned_tasks(struct rq *rq)
2889 #endif /* !CONFIG_SMP */
2892 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2896 if (!schedstat_enabled())
2902 if (cpu == rq->cpu) {
2903 __schedstat_inc(rq->ttwu_local);
2904 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2906 struct sched_domain *sd;
2908 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2910 for_each_domain(rq->cpu, sd) {
2911 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2912 __schedstat_inc(sd->ttwu_wake_remote);
2919 if (wake_flags & WF_MIGRATED)
2920 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2921 #endif /* CONFIG_SMP */
2923 __schedstat_inc(rq->ttwu_count);
2924 __schedstat_inc(p->se.statistics.nr_wakeups);
2926 if (wake_flags & WF_SYNC)
2927 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2931 * Mark the task runnable and perform wakeup-preemption.
2933 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2934 struct rq_flags *rf)
2936 check_preempt_curr(rq, p, wake_flags);
2937 p->state = TASK_RUNNING;
2938 trace_sched_wakeup(p);
2941 if (p->sched_class->task_woken) {
2943 * Our task @p is fully woken up and running; so it's safe to
2944 * drop the rq->lock, hereafter rq is only used for statistics.
2946 rq_unpin_lock(rq, rf);
2947 p->sched_class->task_woken(rq, p);
2948 rq_repin_lock(rq, rf);
2951 if (rq->idle_stamp) {
2952 u64 delta = rq_clock(rq) - rq->idle_stamp;
2953 u64 max = 2*rq->max_idle_balance_cost;
2955 update_avg(&rq->avg_idle, delta);
2957 if (rq->avg_idle > max)
2966 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2967 struct rq_flags *rf)
2969 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2971 lockdep_assert_held(&rq->lock);
2973 if (p->sched_contributes_to_load)
2974 rq->nr_uninterruptible--;
2977 if (wake_flags & WF_MIGRATED)
2978 en_flags |= ENQUEUE_MIGRATED;
2982 delayacct_blkio_end(p);
2983 atomic_dec(&task_rq(p)->nr_iowait);
2986 activate_task(rq, p, en_flags);
2987 ttwu_do_wakeup(rq, p, wake_flags, rf);
2991 * Consider @p being inside a wait loop:
2994 * set_current_state(TASK_UNINTERRUPTIBLE);
3001 * __set_current_state(TASK_RUNNING);
3003 * between set_current_state() and schedule(). In this case @p is still
3004 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3007 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3008 * then schedule() must still happen and p->state can be changed to
3009 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3010 * need to do a full wakeup with enqueue.
3012 * Returns: %true when the wakeup is done,
3015 static int ttwu_runnable(struct task_struct *p, int wake_flags)
3021 rq = __task_rq_lock(p, &rf);
3022 if (task_on_rq_queued(p)) {
3023 /* check_preempt_curr() may use rq clock */
3024 update_rq_clock(rq);
3025 ttwu_do_wakeup(rq, p, wake_flags, &rf);
3028 __task_rq_unlock(rq, &rf);
3034 void sched_ttwu_pending(void *arg)
3036 struct llist_node *llist = arg;
3037 struct rq *rq = this_rq();
3038 struct task_struct *p, *t;
3045 * rq::ttwu_pending racy indication of out-standing wakeups.
3046 * Races such that false-negatives are possible, since they
3047 * are shorter lived that false-positives would be.
3049 WRITE_ONCE(rq->ttwu_pending, 0);
3051 rq_lock_irqsave(rq, &rf);
3052 update_rq_clock(rq);
3054 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
3055 if (WARN_ON_ONCE(p->on_cpu))
3056 smp_cond_load_acquire(&p->on_cpu, !VAL);
3058 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3059 set_task_cpu(p, cpu_of(rq));
3061 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
3064 rq_unlock_irqrestore(rq, &rf);
3067 void send_call_function_single_ipi(int cpu)
3069 struct rq *rq = cpu_rq(cpu);
3071 if (!set_nr_if_polling(rq->idle))
3072 arch_send_call_function_single_ipi(cpu);
3074 trace_sched_wake_idle_without_ipi(cpu);
3078 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3079 * necessary. The wakee CPU on receipt of the IPI will queue the task
3080 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3081 * of the wakeup instead of the waker.
3083 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3085 struct rq *rq = cpu_rq(cpu);
3087 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3089 WRITE_ONCE(rq->ttwu_pending, 1);
3090 __smp_call_single_queue(cpu, &p->wake_entry.llist);
3093 void wake_up_if_idle(int cpu)
3095 struct rq *rq = cpu_rq(cpu);
3100 if (!is_idle_task(rcu_dereference(rq->curr)))
3103 if (set_nr_if_polling(rq->idle)) {
3104 trace_sched_wake_idle_without_ipi(cpu);
3106 rq_lock_irqsave(rq, &rf);
3107 if (is_idle_task(rq->curr))
3108 smp_send_reschedule(cpu);
3109 /* Else CPU is not idle, do nothing here: */
3110 rq_unlock_irqrestore(rq, &rf);
3117 bool cpus_share_cache(int this_cpu, int that_cpu)
3119 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3122 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
3125 * If the CPU does not share cache, then queue the task on the
3126 * remote rqs wakelist to avoid accessing remote data.
3128 if (!cpus_share_cache(smp_processor_id(), cpu))
3132 * If the task is descheduling and the only running task on the
3133 * CPU then use the wakelist to offload the task activation to
3134 * the soon-to-be-idle CPU as the current CPU is likely busy.
3135 * nr_running is checked to avoid unnecessary task stacking.
3137 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
3143 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3145 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
3146 if (WARN_ON_ONCE(cpu == smp_processor_id()))
3149 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
3150 __ttwu_queue_wakelist(p, cpu, wake_flags);
3157 #else /* !CONFIG_SMP */
3159 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3164 #endif /* CONFIG_SMP */
3166 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
3168 struct rq *rq = cpu_rq(cpu);
3171 if (ttwu_queue_wakelist(p, cpu, wake_flags))
3175 update_rq_clock(rq);
3176 ttwu_do_activate(rq, p, wake_flags, &rf);
3181 * Notes on Program-Order guarantees on SMP systems.
3185 * The basic program-order guarantee on SMP systems is that when a task [t]
3186 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
3187 * execution on its new CPU [c1].
3189 * For migration (of runnable tasks) this is provided by the following means:
3191 * A) UNLOCK of the rq(c0)->lock scheduling out task t
3192 * B) migration for t is required to synchronize *both* rq(c0)->lock and
3193 * rq(c1)->lock (if not at the same time, then in that order).
3194 * C) LOCK of the rq(c1)->lock scheduling in task
3196 * Release/acquire chaining guarantees that B happens after A and C after B.
3197 * Note: the CPU doing B need not be c0 or c1
3206 * UNLOCK rq(0)->lock
3208 * LOCK rq(0)->lock // orders against CPU0
3210 * UNLOCK rq(0)->lock
3214 * UNLOCK rq(1)->lock
3216 * LOCK rq(1)->lock // orders against CPU2
3219 * UNLOCK rq(1)->lock
3222 * BLOCKING -- aka. SLEEP + WAKEUP
3224 * For blocking we (obviously) need to provide the same guarantee as for
3225 * migration. However the means are completely different as there is no lock
3226 * chain to provide order. Instead we do:
3228 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
3229 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
3233 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
3235 * LOCK rq(0)->lock LOCK X->pi_lock
3238 * smp_store_release(X->on_cpu, 0);
3240 * smp_cond_load_acquire(&X->on_cpu, !VAL);
3246 * X->state = RUNNING
3247 * UNLOCK rq(2)->lock
3249 * LOCK rq(2)->lock // orders against CPU1
3252 * UNLOCK rq(2)->lock
3255 * UNLOCK rq(0)->lock
3258 * However, for wakeups there is a second guarantee we must provide, namely we
3259 * must ensure that CONDITION=1 done by the caller can not be reordered with
3260 * accesses to the task state; see try_to_wake_up() and set_current_state().
3264 * try_to_wake_up - wake up a thread
3265 * @p: the thread to be awakened
3266 * @state: the mask of task states that can be woken
3267 * @wake_flags: wake modifier flags (WF_*)
3269 * Conceptually does:
3271 * If (@state & @p->state) @p->state = TASK_RUNNING.
3273 * If the task was not queued/runnable, also place it back on a runqueue.
3275 * This function is atomic against schedule() which would dequeue the task.
3277 * It issues a full memory barrier before accessing @p->state, see the comment
3278 * with set_current_state().
3280 * Uses p->pi_lock to serialize against concurrent wake-ups.
3282 * Relies on p->pi_lock stabilizing:
3285 * - p->sched_task_group
3286 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
3288 * Tries really hard to only take one task_rq(p)->lock for performance.
3289 * Takes rq->lock in:
3290 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
3291 * - ttwu_queue() -- new rq, for enqueue of the task;
3292 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
3294 * As a consequence we race really badly with just about everything. See the
3295 * many memory barriers and their comments for details.
3297 * Return: %true if @p->state changes (an actual wakeup was done),
3301 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
3303 unsigned long flags;
3304 int cpu, success = 0;
3309 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
3310 * == smp_processor_id()'. Together this means we can special
3311 * case the whole 'p->on_rq && ttwu_runnable()' case below
3312 * without taking any locks.
3315 * - we rely on Program-Order guarantees for all the ordering,
3316 * - we're serialized against set_special_state() by virtue of
3317 * it disabling IRQs (this allows not taking ->pi_lock).
3319 if (!(p->state & state))
3323 trace_sched_waking(p);
3324 p->state = TASK_RUNNING;
3325 trace_sched_wakeup(p);
3330 * If we are going to wake up a thread waiting for CONDITION we
3331 * need to ensure that CONDITION=1 done by the caller can not be
3332 * reordered with p->state check below. This pairs with smp_store_mb()
3333 * in set_current_state() that the waiting thread does.
3335 raw_spin_lock_irqsave(&p->pi_lock, flags);
3336 smp_mb__after_spinlock();
3337 if (!(p->state & state))
3340 trace_sched_waking(p);
3342 /* We're going to change ->state: */
3346 * Ensure we load p->on_rq _after_ p->state, otherwise it would
3347 * be possible to, falsely, observe p->on_rq == 0 and get stuck
3348 * in smp_cond_load_acquire() below.
3350 * sched_ttwu_pending() try_to_wake_up()
3351 * STORE p->on_rq = 1 LOAD p->state
3354 * __schedule() (switch to task 'p')
3355 * LOCK rq->lock smp_rmb();
3356 * smp_mb__after_spinlock();
3360 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
3362 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3363 * __schedule(). See the comment for smp_mb__after_spinlock().
3365 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
3368 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
3373 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
3374 * possible to, falsely, observe p->on_cpu == 0.
3376 * One must be running (->on_cpu == 1) in order to remove oneself
3377 * from the runqueue.
3379 * __schedule() (switch to task 'p') try_to_wake_up()
3380 * STORE p->on_cpu = 1 LOAD p->on_rq
3383 * __schedule() (put 'p' to sleep)
3384 * LOCK rq->lock smp_rmb();
3385 * smp_mb__after_spinlock();
3386 * STORE p->on_rq = 0 LOAD p->on_cpu
3388 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
3389 * __schedule(). See the comment for smp_mb__after_spinlock().
3391 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
3392 * schedule()'s deactivate_task() has 'happened' and p will no longer
3393 * care about it's own p->state. See the comment in __schedule().
3395 smp_acquire__after_ctrl_dep();
3398 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
3399 * == 0), which means we need to do an enqueue, change p->state to
3400 * TASK_WAKING such that we can unlock p->pi_lock before doing the
3401 * enqueue, such as ttwu_queue_wakelist().
3403 p->state = TASK_WAKING;
3406 * If the owning (remote) CPU is still in the middle of schedule() with
3407 * this task as prev, considering queueing p on the remote CPUs wake_list
3408 * which potentially sends an IPI instead of spinning on p->on_cpu to
3409 * let the waker make forward progress. This is safe because IRQs are
3410 * disabled and the IPI will deliver after on_cpu is cleared.
3412 * Ensure we load task_cpu(p) after p->on_cpu:
3414 * set_task_cpu(p, cpu);
3415 * STORE p->cpu = @cpu
3416 * __schedule() (switch to task 'p')
3418 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
3419 * STORE p->on_cpu = 1 LOAD p->cpu
3421 * to ensure we observe the correct CPU on which the task is currently
3424 if (smp_load_acquire(&p->on_cpu) &&
3425 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
3429 * If the owning (remote) CPU is still in the middle of schedule() with
3430 * this task as prev, wait until it's done referencing the task.
3432 * Pairs with the smp_store_release() in finish_task().
3434 * This ensures that tasks getting woken will be fully ordered against
3435 * their previous state and preserve Program Order.
3437 smp_cond_load_acquire(&p->on_cpu, !VAL);
3439 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
3440 if (task_cpu(p) != cpu) {
3442 delayacct_blkio_end(p);
3443 atomic_dec(&task_rq(p)->nr_iowait);
3446 wake_flags |= WF_MIGRATED;
3447 psi_ttwu_dequeue(p);
3448 set_task_cpu(p, cpu);
3452 #endif /* CONFIG_SMP */
3454 ttwu_queue(p, cpu, wake_flags);
3456 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3459 ttwu_stat(p, task_cpu(p), wake_flags);
3466 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
3467 * @p: Process for which the function is to be invoked.
3468 * @func: Function to invoke.
3469 * @arg: Argument to function.
3471 * If the specified task can be quickly locked into a definite state
3472 * (either sleeping or on a given runqueue), arrange to keep it in that
3473 * state while invoking @func(@arg). This function can use ->on_rq and
3474 * task_curr() to work out what the state is, if required. Given that
3475 * @func can be invoked with a runqueue lock held, it had better be quite
3479 * @false if the task slipped out from under the locks.
3480 * @true if the task was locked onto a runqueue or is sleeping.
3481 * However, @func can override this by returning @false.
3483 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3489 lockdep_assert_irqs_enabled();
3490 raw_spin_lock_irq(&p->pi_lock);
3492 rq = __task_rq_lock(p, &rf);
3493 if (task_rq(p) == rq)
3502 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3507 raw_spin_unlock_irq(&p->pi_lock);
3512 * wake_up_process - Wake up a specific process
3513 * @p: The process to be woken up.
3515 * Attempt to wake up the nominated process and move it to the set of runnable
3518 * Return: 1 if the process was woken up, 0 if it was already running.
3520 * This function executes a full memory barrier before accessing the task state.
3522 int wake_up_process(struct task_struct *p)
3524 return try_to_wake_up(p, TASK_NORMAL, 0);
3526 EXPORT_SYMBOL(wake_up_process);
3528 int wake_up_state(struct task_struct *p, unsigned int state)
3530 return try_to_wake_up(p, state, 0);
3534 * Perform scheduler related setup for a newly forked process p.
3535 * p is forked by current.
3537 * __sched_fork() is basic setup used by init_idle() too:
3539 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3544 p->se.exec_start = 0;
3545 p->se.sum_exec_runtime = 0;
3546 p->se.prev_sum_exec_runtime = 0;
3547 p->se.nr_migrations = 0;
3549 INIT_LIST_HEAD(&p->se.group_node);
3551 #ifdef CONFIG_FAIR_GROUP_SCHED
3552 p->se.cfs_rq = NULL;
3555 #ifdef CONFIG_SCHEDSTATS
3556 /* Even if schedstat is disabled, there should not be garbage */
3557 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3560 RB_CLEAR_NODE(&p->dl.rb_node);
3561 init_dl_task_timer(&p->dl);
3562 init_dl_inactive_task_timer(&p->dl);
3563 __dl_clear_params(p);
3565 INIT_LIST_HEAD(&p->rt.run_list);
3567 p->rt.time_slice = sched_rr_timeslice;
3571 #ifdef CONFIG_PREEMPT_NOTIFIERS
3572 INIT_HLIST_HEAD(&p->preempt_notifiers);
3575 #ifdef CONFIG_COMPACTION
3576 p->capture_control = NULL;
3578 init_numa_balancing(clone_flags, p);
3580 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3581 p->migration_pending = NULL;
3585 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3587 #ifdef CONFIG_NUMA_BALANCING
3589 void set_numabalancing_state(bool enabled)
3592 static_branch_enable(&sched_numa_balancing);
3594 static_branch_disable(&sched_numa_balancing);
3597 #ifdef CONFIG_PROC_SYSCTL
3598 int sysctl_numa_balancing(struct ctl_table *table, int write,
3599 void *buffer, size_t *lenp, loff_t *ppos)
3603 int state = static_branch_likely(&sched_numa_balancing);
3605 if (write && !capable(CAP_SYS_ADMIN))
3610 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3614 set_numabalancing_state(state);
3620 #ifdef CONFIG_SCHEDSTATS
3622 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3623 static bool __initdata __sched_schedstats = false;
3625 static void set_schedstats(bool enabled)
3628 static_branch_enable(&sched_schedstats);
3630 static_branch_disable(&sched_schedstats);
3633 void force_schedstat_enabled(void)
3635 if (!schedstat_enabled()) {
3636 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3637 static_branch_enable(&sched_schedstats);
3641 static int __init setup_schedstats(char *str)
3648 * This code is called before jump labels have been set up, so we can't
3649 * change the static branch directly just yet. Instead set a temporary
3650 * variable so init_schedstats() can do it later.
3652 if (!strcmp(str, "enable")) {
3653 __sched_schedstats = true;
3655 } else if (!strcmp(str, "disable")) {
3656 __sched_schedstats = false;
3661 pr_warn("Unable to parse schedstats=\n");
3665 __setup("schedstats=", setup_schedstats);
3667 static void __init init_schedstats(void)
3669 set_schedstats(__sched_schedstats);
3672 #ifdef CONFIG_PROC_SYSCTL
3673 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
3674 size_t *lenp, loff_t *ppos)
3678 int state = static_branch_likely(&sched_schedstats);
3680 if (write && !capable(CAP_SYS_ADMIN))
3685 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3689 set_schedstats(state);
3692 #endif /* CONFIG_PROC_SYSCTL */
3693 #else /* !CONFIG_SCHEDSTATS */
3694 static inline void init_schedstats(void) {}
3695 #endif /* CONFIG_SCHEDSTATS */
3698 * fork()/clone()-time setup:
3700 int sched_fork(unsigned long clone_flags, struct task_struct *p)
3702 unsigned long flags;
3704 __sched_fork(clone_flags, p);
3706 * We mark the process as NEW here. This guarantees that
3707 * nobody will actually run it, and a signal or other external
3708 * event cannot wake it up and insert it on the runqueue either.
3710 p->state = TASK_NEW;
3713 * Make sure we do not leak PI boosting priority to the child.
3715 p->prio = current->normal_prio;
3720 * Revert to default priority/policy on fork if requested.
3722 if (unlikely(p->sched_reset_on_fork)) {
3723 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3724 p->policy = SCHED_NORMAL;
3725 p->static_prio = NICE_TO_PRIO(0);
3727 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3728 p->static_prio = NICE_TO_PRIO(0);
3730 p->prio = p->normal_prio = __normal_prio(p);
3731 set_load_weight(p, false);
3734 * We don't need the reset flag anymore after the fork. It has
3735 * fulfilled its duty:
3737 p->sched_reset_on_fork = 0;
3740 if (dl_prio(p->prio))
3742 else if (rt_prio(p->prio))
3743 p->sched_class = &rt_sched_class;
3745 p->sched_class = &fair_sched_class;
3747 init_entity_runnable_average(&p->se);
3750 * The child is not yet in the pid-hash so no cgroup attach races,
3751 * and the cgroup is pinned to this child due to cgroup_fork()
3752 * is ran before sched_fork().
3754 * Silence PROVE_RCU.
3756 raw_spin_lock_irqsave(&p->pi_lock, flags);
3759 * We're setting the CPU for the first time, we don't migrate,
3760 * so use __set_task_cpu().
3762 __set_task_cpu(p, smp_processor_id());
3763 if (p->sched_class->task_fork)
3764 p->sched_class->task_fork(p);
3765 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3767 #ifdef CONFIG_SCHED_INFO
3768 if (likely(sched_info_on()))
3769 memset(&p->sched_info, 0, sizeof(p->sched_info));
3771 #if defined(CONFIG_SMP)
3774 init_task_preempt_count(p);
3776 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3777 RB_CLEAR_NODE(&p->pushable_dl_tasks);
3782 void sched_post_fork(struct task_struct *p)
3784 uclamp_post_fork(p);
3787 unsigned long to_ratio(u64 period, u64 runtime)
3789 if (runtime == RUNTIME_INF)
3793 * Doing this here saves a lot of checks in all
3794 * the calling paths, and returning zero seems
3795 * safe for them anyway.
3800 return div64_u64(runtime << BW_SHIFT, period);
3804 * wake_up_new_task - wake up a newly created task for the first time.
3806 * This function will do some initial scheduler statistics housekeeping
3807 * that must be done for every newly created context, then puts the task
3808 * on the runqueue and wakes it.
3810 void wake_up_new_task(struct task_struct *p)
3815 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3816 p->state = TASK_RUNNING;
3819 * Fork balancing, do it here and not earlier because:
3820 * - cpus_ptr can change in the fork path
3821 * - any previously selected CPU might disappear through hotplug
3823 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3824 * as we're not fully set-up yet.
3826 p->recent_used_cpu = task_cpu(p);
3828 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
3830 rq = __task_rq_lock(p, &rf);
3831 update_rq_clock(rq);
3832 post_init_entity_util_avg(p);
3834 activate_task(rq, p, ENQUEUE_NOCLOCK);
3835 trace_sched_wakeup_new(p);
3836 check_preempt_curr(rq, p, WF_FORK);
3838 if (p->sched_class->task_woken) {
3840 * Nothing relies on rq->lock after this, so it's fine to
3843 rq_unpin_lock(rq, &rf);
3844 p->sched_class->task_woken(rq, p);
3845 rq_repin_lock(rq, &rf);
3848 task_rq_unlock(rq, p, &rf);
3851 #ifdef CONFIG_PREEMPT_NOTIFIERS
3853 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3855 void preempt_notifier_inc(void)
3857 static_branch_inc(&preempt_notifier_key);
3859 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3861 void preempt_notifier_dec(void)
3863 static_branch_dec(&preempt_notifier_key);
3865 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3868 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3869 * @notifier: notifier struct to register
3871 void preempt_notifier_register(struct preempt_notifier *notifier)
3873 if (!static_branch_unlikely(&preempt_notifier_key))
3874 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3876 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3878 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3881 * preempt_notifier_unregister - no longer interested in preemption notifications
3882 * @notifier: notifier struct to unregister
3884 * This is *not* safe to call from within a preemption notifier.
3886 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3888 hlist_del(¬ifier->link);
3890 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3892 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3894 struct preempt_notifier *notifier;
3896 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3897 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3900 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3902 if (static_branch_unlikely(&preempt_notifier_key))
3903 __fire_sched_in_preempt_notifiers(curr);
3907 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3908 struct task_struct *next)
3910 struct preempt_notifier *notifier;
3912 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3913 notifier->ops->sched_out(notifier, next);
3916 static __always_inline void
3917 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3918 struct task_struct *next)
3920 if (static_branch_unlikely(&preempt_notifier_key))
3921 __fire_sched_out_preempt_notifiers(curr, next);
3924 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3926 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3931 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3932 struct task_struct *next)
3936 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3938 static inline void prepare_task(struct task_struct *next)
3942 * Claim the task as running, we do this before switching to it
3943 * such that any running task will have this set.
3945 * See the ttwu() WF_ON_CPU case and its ordering comment.
3947 WRITE_ONCE(next->on_cpu, 1);
3951 static inline void finish_task(struct task_struct *prev)
3955 * This must be the very last reference to @prev from this CPU. After
3956 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3957 * must ensure this doesn't happen until the switch is completely
3960 * In particular, the load of prev->state in finish_task_switch() must
3961 * happen before this.
3963 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3965 smp_store_release(&prev->on_cpu, 0);
3971 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
3973 void (*func)(struct rq *rq);
3974 struct callback_head *next;
3976 lockdep_assert_held(&rq->lock);
3979 func = (void (*)(struct rq *))head->func;
3988 static void balance_push(struct rq *rq);
3990 struct callback_head balance_push_callback = {
3992 .func = (void (*)(struct callback_head *))balance_push,
3995 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
3997 struct callback_head *head = rq->balance_callback;
3999 lockdep_assert_held(&rq->lock);
4001 rq->balance_callback = NULL;
4006 static void __balance_callbacks(struct rq *rq)
4008 do_balance_callbacks(rq, splice_balance_callbacks(rq));
4011 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4013 unsigned long flags;
4015 if (unlikely(head)) {
4016 raw_spin_lock_irqsave(&rq->lock, flags);
4017 do_balance_callbacks(rq, head);
4018 raw_spin_unlock_irqrestore(&rq->lock, flags);
4024 static inline void __balance_callbacks(struct rq *rq)
4028 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
4033 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
4040 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
4043 * Since the runqueue lock will be released by the next
4044 * task (which is an invalid locking op but in the case
4045 * of the scheduler it's an obvious special-case), so we
4046 * do an early lockdep release here:
4048 rq_unpin_lock(rq, rf);
4049 spin_release(&rq->lock.dep_map, _THIS_IP_);
4050 #ifdef CONFIG_DEBUG_SPINLOCK
4051 /* this is a valid case when another task releases the spinlock */
4052 rq->lock.owner = next;
4056 static inline void finish_lock_switch(struct rq *rq)
4059 * If we are tracking spinlock dependencies then we have to
4060 * fix up the runqueue lock - which gets 'carried over' from
4061 * prev into current:
4063 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
4064 __balance_callbacks(rq);
4065 raw_spin_unlock_irq(&rq->lock);
4069 * NOP if the arch has not defined these:
4072 #ifndef prepare_arch_switch
4073 # define prepare_arch_switch(next) do { } while (0)
4076 #ifndef finish_arch_post_lock_switch
4077 # define finish_arch_post_lock_switch() do { } while (0)
4081 * prepare_task_switch - prepare to switch tasks
4082 * @rq: the runqueue preparing to switch
4083 * @prev: the current task that is being switched out
4084 * @next: the task we are going to switch to.
4086 * This is called with the rq lock held and interrupts off. It must
4087 * be paired with a subsequent finish_task_switch after the context
4090 * prepare_task_switch sets up locking and calls architecture specific
4094 prepare_task_switch(struct rq *rq, struct task_struct *prev,
4095 struct task_struct *next)
4097 kcov_prepare_switch(prev);
4098 sched_info_switch(rq, prev, next);
4099 perf_event_task_sched_out(prev, next);
4101 fire_sched_out_preempt_notifiers(prev, next);
4103 prepare_arch_switch(next);
4107 * finish_task_switch - clean up after a task-switch
4108 * @prev: the thread we just switched away from.
4110 * finish_task_switch must be called after the context switch, paired
4111 * with a prepare_task_switch call before the context switch.
4112 * finish_task_switch will reconcile locking set up by prepare_task_switch,
4113 * and do any other architecture-specific cleanup actions.
4115 * Note that we may have delayed dropping an mm in context_switch(). If
4116 * so, we finish that here outside of the runqueue lock. (Doing it
4117 * with the lock held can cause deadlocks; see schedule() for
4120 * The context switch have flipped the stack from under us and restored the
4121 * local variables which were saved when this task called schedule() in the
4122 * past. prev == current is still correct but we need to recalculate this_rq
4123 * because prev may have moved to another CPU.
4125 static struct rq *finish_task_switch(struct task_struct *prev)
4126 __releases(rq->lock)
4128 struct rq *rq = this_rq();
4129 struct mm_struct *mm = rq->prev_mm;
4133 * The previous task will have left us with a preempt_count of 2
4134 * because it left us after:
4137 * preempt_disable(); // 1
4139 * raw_spin_lock_irq(&rq->lock) // 2
4141 * Also, see FORK_PREEMPT_COUNT.
4143 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
4144 "corrupted preempt_count: %s/%d/0x%x\n",
4145 current->comm, current->pid, preempt_count()))
4146 preempt_count_set(FORK_PREEMPT_COUNT);
4151 * A task struct has one reference for the use as "current".
4152 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
4153 * schedule one last time. The schedule call will never return, and
4154 * the scheduled task must drop that reference.
4156 * We must observe prev->state before clearing prev->on_cpu (in
4157 * finish_task), otherwise a concurrent wakeup can get prev
4158 * running on another CPU and we could rave with its RUNNING -> DEAD
4159 * transition, resulting in a double drop.
4161 prev_state = prev->state;
4162 vtime_task_switch(prev);
4163 perf_event_task_sched_in(prev, current);
4165 finish_lock_switch(rq);
4166 finish_arch_post_lock_switch();
4167 kcov_finish_switch(current);
4169 fire_sched_in_preempt_notifiers(current);
4171 * When switching through a kernel thread, the loop in
4172 * membarrier_{private,global}_expedited() may have observed that
4173 * kernel thread and not issued an IPI. It is therefore possible to
4174 * schedule between user->kernel->user threads without passing though
4175 * switch_mm(). Membarrier requires a barrier after storing to
4176 * rq->curr, before returning to userspace, so provide them here:
4178 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
4179 * provided by mmdrop(),
4180 * - a sync_core for SYNC_CORE.
4183 membarrier_mm_sync_core_before_usermode(mm);
4186 if (unlikely(prev_state == TASK_DEAD)) {
4187 if (prev->sched_class->task_dead)
4188 prev->sched_class->task_dead(prev);
4191 * Remove function-return probe instances associated with this
4192 * task and put them back on the free list.
4194 kprobe_flush_task(prev);
4196 /* Task is done with its stack. */
4197 put_task_stack(prev);
4199 put_task_struct_rcu_user(prev);
4202 tick_nohz_task_switch();
4207 * schedule_tail - first thing a freshly forked thread must call.
4208 * @prev: the thread we just switched away from.
4210 asmlinkage __visible void schedule_tail(struct task_struct *prev)
4211 __releases(rq->lock)
4216 * New tasks start with FORK_PREEMPT_COUNT, see there and
4217 * finish_task_switch() for details.
4219 * finish_task_switch() will drop rq->lock() and lower preempt_count
4220 * and the preempt_enable() will end up enabling preemption (on
4221 * PREEMPT_COUNT kernels).
4224 rq = finish_task_switch(prev);
4227 if (current->set_child_tid)
4228 put_user(task_pid_vnr(current), current->set_child_tid);
4230 calculate_sigpending();
4234 * context_switch - switch to the new MM and the new thread's register state.
4236 static __always_inline struct rq *
4237 context_switch(struct rq *rq, struct task_struct *prev,
4238 struct task_struct *next, struct rq_flags *rf)
4240 prepare_task_switch(rq, prev, next);
4243 * For paravirt, this is coupled with an exit in switch_to to
4244 * combine the page table reload and the switch backend into
4247 arch_start_context_switch(prev);
4250 * kernel -> kernel lazy + transfer active
4251 * user -> kernel lazy + mmgrab() active
4253 * kernel -> user switch + mmdrop() active
4254 * user -> user switch
4256 if (!next->mm) { // to kernel
4257 enter_lazy_tlb(prev->active_mm, next);
4259 next->active_mm = prev->active_mm;
4260 if (prev->mm) // from user
4261 mmgrab(prev->active_mm);
4263 prev->active_mm = NULL;
4265 membarrier_switch_mm(rq, prev->active_mm, next->mm);
4267 * sys_membarrier() requires an smp_mb() between setting
4268 * rq->curr / membarrier_switch_mm() and returning to userspace.
4270 * The below provides this either through switch_mm(), or in
4271 * case 'prev->active_mm == next->mm' through
4272 * finish_task_switch()'s mmdrop().
4274 switch_mm_irqs_off(prev->active_mm, next->mm, next);
4276 if (!prev->mm) { // from kernel
4277 /* will mmdrop() in finish_task_switch(). */
4278 rq->prev_mm = prev->active_mm;
4279 prev->active_mm = NULL;
4283 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4285 prepare_lock_switch(rq, next, rf);
4287 /* Here we just switch the register state and the stack. */
4288 switch_to(prev, next, prev);
4291 return finish_task_switch(prev);
4295 * nr_running and nr_context_switches:
4297 * externally visible scheduler statistics: current number of runnable
4298 * threads, total number of context switches performed since bootup.
4300 unsigned long nr_running(void)
4302 unsigned long i, sum = 0;
4304 for_each_online_cpu(i)
4305 sum += cpu_rq(i)->nr_running;
4311 * Check if only the current task is running on the CPU.
4313 * Caution: this function does not check that the caller has disabled
4314 * preemption, thus the result might have a time-of-check-to-time-of-use
4315 * race. The caller is responsible to use it correctly, for example:
4317 * - from a non-preemptible section (of course)
4319 * - from a thread that is bound to a single CPU
4321 * - in a loop with very short iterations (e.g. a polling loop)
4323 bool single_task_running(void)
4325 return raw_rq()->nr_running == 1;
4327 EXPORT_SYMBOL(single_task_running);
4329 unsigned long long nr_context_switches(void)
4332 unsigned long long sum = 0;
4334 for_each_possible_cpu(i)
4335 sum += cpu_rq(i)->nr_switches;
4341 * Consumers of these two interfaces, like for example the cpuidle menu
4342 * governor, are using nonsensical data. Preferring shallow idle state selection
4343 * for a CPU that has IO-wait which might not even end up running the task when
4344 * it does become runnable.
4347 unsigned long nr_iowait_cpu(int cpu)
4349 return atomic_read(&cpu_rq(cpu)->nr_iowait);
4353 * IO-wait accounting, and how it's mostly bollocks (on SMP).
4355 * The idea behind IO-wait account is to account the idle time that we could
4356 * have spend running if it were not for IO. That is, if we were to improve the
4357 * storage performance, we'd have a proportional reduction in IO-wait time.
4359 * This all works nicely on UP, where, when a task blocks on IO, we account
4360 * idle time as IO-wait, because if the storage were faster, it could've been
4361 * running and we'd not be idle.
4363 * This has been extended to SMP, by doing the same for each CPU. This however
4366 * Imagine for instance the case where two tasks block on one CPU, only the one
4367 * CPU will have IO-wait accounted, while the other has regular idle. Even
4368 * though, if the storage were faster, both could've ran at the same time,
4369 * utilising both CPUs.
4371 * This means, that when looking globally, the current IO-wait accounting on
4372 * SMP is a lower bound, by reason of under accounting.
4374 * Worse, since the numbers are provided per CPU, they are sometimes
4375 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
4376 * associated with any one particular CPU, it can wake to another CPU than it
4377 * blocked on. This means the per CPU IO-wait number is meaningless.
4379 * Task CPU affinities can make all that even more 'interesting'.
4382 unsigned long nr_iowait(void)
4384 unsigned long i, sum = 0;
4386 for_each_possible_cpu(i)
4387 sum += nr_iowait_cpu(i);
4395 * sched_exec - execve() is a valuable balancing opportunity, because at
4396 * this point the task has the smallest effective memory and cache footprint.
4398 void sched_exec(void)
4400 struct task_struct *p = current;
4401 unsigned long flags;
4404 raw_spin_lock_irqsave(&p->pi_lock, flags);
4405 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
4406 if (dest_cpu == smp_processor_id())
4409 if (likely(cpu_active(dest_cpu))) {
4410 struct migration_arg arg = { p, dest_cpu };
4412 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4413 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4417 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4422 DEFINE_PER_CPU(struct kernel_stat, kstat);
4423 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
4425 EXPORT_PER_CPU_SYMBOL(kstat);
4426 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
4429 * The function fair_sched_class.update_curr accesses the struct curr
4430 * and its field curr->exec_start; when called from task_sched_runtime(),
4431 * we observe a high rate of cache misses in practice.
4432 * Prefetching this data results in improved performance.
4434 static inline void prefetch_curr_exec_start(struct task_struct *p)
4436 #ifdef CONFIG_FAIR_GROUP_SCHED
4437 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
4439 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
4442 prefetch(&curr->exec_start);
4446 * Return accounted runtime for the task.
4447 * In case the task is currently running, return the runtime plus current's
4448 * pending runtime that have not been accounted yet.
4450 unsigned long long task_sched_runtime(struct task_struct *p)
4456 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
4458 * 64-bit doesn't need locks to atomically read a 64-bit value.
4459 * So we have a optimization chance when the task's delta_exec is 0.
4460 * Reading ->on_cpu is racy, but this is ok.
4462 * If we race with it leaving CPU, we'll take a lock. So we're correct.
4463 * If we race with it entering CPU, unaccounted time is 0. This is
4464 * indistinguishable from the read occurring a few cycles earlier.
4465 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
4466 * been accounted, so we're correct here as well.
4468 if (!p->on_cpu || !task_on_rq_queued(p))
4469 return p->se.sum_exec_runtime;
4472 rq = task_rq_lock(p, &rf);
4474 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4475 * project cycles that may never be accounted to this
4476 * thread, breaking clock_gettime().
4478 if (task_current(rq, p) && task_on_rq_queued(p)) {
4479 prefetch_curr_exec_start(p);
4480 update_rq_clock(rq);
4481 p->sched_class->update_curr(rq);
4483 ns = p->se.sum_exec_runtime;
4484 task_rq_unlock(rq, p, &rf);
4490 * This function gets called by the timer code, with HZ frequency.
4491 * We call it with interrupts disabled.
4493 void scheduler_tick(void)
4495 int cpu = smp_processor_id();
4496 struct rq *rq = cpu_rq(cpu);
4497 struct task_struct *curr = rq->curr;
4499 unsigned long thermal_pressure;
4501 arch_scale_freq_tick();
4506 update_rq_clock(rq);
4507 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4508 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4509 curr->sched_class->task_tick(rq, curr, 0);
4510 calc_global_load_tick(rq);
4515 perf_event_task_tick();
4518 rq->idle_balance = idle_cpu(cpu);
4519 trigger_load_balance(rq);
4523 #ifdef CONFIG_NO_HZ_FULL
4528 struct delayed_work work;
4530 /* Values for ->state, see diagram below. */
4531 #define TICK_SCHED_REMOTE_OFFLINE 0
4532 #define TICK_SCHED_REMOTE_OFFLINING 1
4533 #define TICK_SCHED_REMOTE_RUNNING 2
4536 * State diagram for ->state:
4539 * TICK_SCHED_REMOTE_OFFLINE
4542 * | | sched_tick_remote()
4545 * +--TICK_SCHED_REMOTE_OFFLINING
4548 * sched_tick_start() | | sched_tick_stop()
4551 * TICK_SCHED_REMOTE_RUNNING
4554 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4555 * and sched_tick_start() are happy to leave the state in RUNNING.
4558 static struct tick_work __percpu *tick_work_cpu;
4560 static void sched_tick_remote(struct work_struct *work)
4562 struct delayed_work *dwork = to_delayed_work(work);
4563 struct tick_work *twork = container_of(dwork, struct tick_work, work);
4564 int cpu = twork->cpu;
4565 struct rq *rq = cpu_rq(cpu);
4566 struct task_struct *curr;
4572 * Handle the tick only if it appears the remote CPU is running in full
4573 * dynticks mode. The check is racy by nature, but missing a tick or
4574 * having one too much is no big deal because the scheduler tick updates
4575 * statistics and checks timeslices in a time-independent way, regardless
4576 * of when exactly it is running.
4578 if (!tick_nohz_tick_stopped_cpu(cpu))
4581 rq_lock_irq(rq, &rf);
4583 if (cpu_is_offline(cpu))
4586 update_rq_clock(rq);
4588 if (!is_idle_task(curr)) {
4590 * Make sure the next tick runs within a reasonable
4593 delta = rq_clock_task(rq) - curr->se.exec_start;
4594 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
4596 curr->sched_class->task_tick(rq, curr, 0);
4598 calc_load_nohz_remote(rq);
4600 rq_unlock_irq(rq, &rf);
4604 * Run the remote tick once per second (1Hz). This arbitrary
4605 * frequency is large enough to avoid overload but short enough
4606 * to keep scheduler internal stats reasonably up to date. But
4607 * first update state to reflect hotplug activity if required.
4609 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
4610 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
4611 if (os == TICK_SCHED_REMOTE_RUNNING)
4612 queue_delayed_work(system_unbound_wq, dwork, HZ);
4615 static void sched_tick_start(int cpu)
4618 struct tick_work *twork;
4620 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4623 WARN_ON_ONCE(!tick_work_cpu);
4625 twork = per_cpu_ptr(tick_work_cpu, cpu);
4626 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
4627 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
4628 if (os == TICK_SCHED_REMOTE_OFFLINE) {
4630 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
4631 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
4635 #ifdef CONFIG_HOTPLUG_CPU
4636 static void sched_tick_stop(int cpu)
4638 struct tick_work *twork;
4641 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4644 WARN_ON_ONCE(!tick_work_cpu);
4646 twork = per_cpu_ptr(tick_work_cpu, cpu);
4647 /* There cannot be competing actions, but don't rely on stop-machine. */
4648 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
4649 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
4650 /* Don't cancel, as this would mess up the state machine. */
4652 #endif /* CONFIG_HOTPLUG_CPU */
4654 int __init sched_tick_offload_init(void)
4656 tick_work_cpu = alloc_percpu(struct tick_work);
4657 BUG_ON(!tick_work_cpu);
4661 #else /* !CONFIG_NO_HZ_FULL */
4662 static inline void sched_tick_start(int cpu) { }
4663 static inline void sched_tick_stop(int cpu) { }
4666 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4667 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4669 * If the value passed in is equal to the current preempt count
4670 * then we just disabled preemption. Start timing the latency.
4672 static inline void preempt_latency_start(int val)
4674 if (preempt_count() == val) {
4675 unsigned long ip = get_lock_parent_ip();
4676 #ifdef CONFIG_DEBUG_PREEMPT
4677 current->preempt_disable_ip = ip;
4679 trace_preempt_off(CALLER_ADDR0, ip);
4683 void preempt_count_add(int val)
4685 #ifdef CONFIG_DEBUG_PREEMPT
4689 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4692 __preempt_count_add(val);
4693 #ifdef CONFIG_DEBUG_PREEMPT
4695 * Spinlock count overflowing soon?
4697 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4700 preempt_latency_start(val);
4702 EXPORT_SYMBOL(preempt_count_add);
4703 NOKPROBE_SYMBOL(preempt_count_add);
4706 * If the value passed in equals to the current preempt count
4707 * then we just enabled preemption. Stop timing the latency.
4709 static inline void preempt_latency_stop(int val)
4711 if (preempt_count() == val)
4712 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
4715 void preempt_count_sub(int val)
4717 #ifdef CONFIG_DEBUG_PREEMPT
4721 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4724 * Is the spinlock portion underflowing?
4726 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4727 !(preempt_count() & PREEMPT_MASK)))
4731 preempt_latency_stop(val);
4732 __preempt_count_sub(val);
4734 EXPORT_SYMBOL(preempt_count_sub);
4735 NOKPROBE_SYMBOL(preempt_count_sub);
4738 static inline void preempt_latency_start(int val) { }
4739 static inline void preempt_latency_stop(int val) { }
4742 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
4744 #ifdef CONFIG_DEBUG_PREEMPT
4745 return p->preempt_disable_ip;
4752 * Print scheduling while atomic bug:
4754 static noinline void __schedule_bug(struct task_struct *prev)
4756 /* Save this before calling printk(), since that will clobber it */
4757 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
4759 if (oops_in_progress)
4762 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4763 prev->comm, prev->pid, preempt_count());
4765 debug_show_held_locks(prev);
4767 if (irqs_disabled())
4768 print_irqtrace_events(prev);
4769 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
4770 && in_atomic_preempt_off()) {
4771 pr_err("Preemption disabled at:");
4772 print_ip_sym(KERN_ERR, preempt_disable_ip);
4775 panic("scheduling while atomic\n");
4778 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4782 * Various schedule()-time debugging checks and statistics:
4784 static inline void schedule_debug(struct task_struct *prev, bool preempt)
4786 #ifdef CONFIG_SCHED_STACK_END_CHECK
4787 if (task_stack_end_corrupted(prev))
4788 panic("corrupted stack end detected inside scheduler\n");
4790 if (task_scs_end_corrupted(prev))
4791 panic("corrupted shadow stack detected inside scheduler\n");
4794 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4795 if (!preempt && prev->state && prev->non_block_count) {
4796 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4797 prev->comm, prev->pid, prev->non_block_count);
4799 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4803 if (unlikely(in_atomic_preempt_off())) {
4804 __schedule_bug(prev);
4805 preempt_count_set(PREEMPT_DISABLED);
4809 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4811 schedstat_inc(this_rq()->sched_count);
4814 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
4815 struct rq_flags *rf)
4818 const struct sched_class *class;
4820 * We must do the balancing pass before put_prev_task(), such
4821 * that when we release the rq->lock the task is in the same
4822 * state as before we took rq->lock.
4824 * We can terminate the balance pass as soon as we know there is
4825 * a runnable task of @class priority or higher.
4827 for_class_range(class, prev->sched_class, &idle_sched_class) {
4828 if (class->balance(rq, prev, rf))
4833 put_prev_task(rq, prev);
4837 * Pick up the highest-prio task:
4839 static inline struct task_struct *
4840 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4842 const struct sched_class *class;
4843 struct task_struct *p;
4846 * Optimization: we know that if all tasks are in the fair class we can
4847 * call that function directly, but only if the @prev task wasn't of a
4848 * higher scheduling class, because otherwise those lose the
4849 * opportunity to pull in more work from other CPUs.
4851 if (likely(prev->sched_class <= &fair_sched_class &&
4852 rq->nr_running == rq->cfs.h_nr_running)) {
4854 p = pick_next_task_fair(rq, prev, rf);
4855 if (unlikely(p == RETRY_TASK))
4858 /* Assumes fair_sched_class->next == idle_sched_class */
4860 put_prev_task(rq, prev);
4861 p = pick_next_task_idle(rq);
4868 put_prev_task_balance(rq, prev, rf);
4870 for_each_class(class) {
4871 p = class->pick_next_task(rq);
4876 /* The idle class should always have a runnable task: */
4881 * __schedule() is the main scheduler function.
4883 * The main means of driving the scheduler and thus entering this function are:
4885 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4887 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4888 * paths. For example, see arch/x86/entry_64.S.
4890 * To drive preemption between tasks, the scheduler sets the flag in timer
4891 * interrupt handler scheduler_tick().
4893 * 3. Wakeups don't really cause entry into schedule(). They add a
4894 * task to the run-queue and that's it.
4896 * Now, if the new task added to the run-queue preempts the current
4897 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4898 * called on the nearest possible occasion:
4900 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4902 * - in syscall or exception context, at the next outmost
4903 * preempt_enable(). (this might be as soon as the wake_up()'s
4906 * - in IRQ context, return from interrupt-handler to
4907 * preemptible context
4909 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4912 * - cond_resched() call
4913 * - explicit schedule() call
4914 * - return from syscall or exception to user-space
4915 * - return from interrupt-handler to user-space
4917 * WARNING: must be called with preemption disabled!
4919 static void __sched notrace __schedule(bool preempt)
4921 struct task_struct *prev, *next;
4922 unsigned long *switch_count;
4923 unsigned long prev_state;
4928 cpu = smp_processor_id();
4932 schedule_debug(prev, preempt);
4934 if (sched_feat(HRTICK))
4937 local_irq_disable();
4938 rcu_note_context_switch(preempt);
4941 * Make sure that signal_pending_state()->signal_pending() below
4942 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4943 * done by the caller to avoid the race with signal_wake_up():
4945 * __set_current_state(@state) signal_wake_up()
4946 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4947 * wake_up_state(p, state)
4948 * LOCK rq->lock LOCK p->pi_state
4949 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4950 * if (signal_pending_state()) if (p->state & @state)
4952 * Also, the membarrier system call requires a full memory barrier
4953 * after coming from user-space, before storing to rq->curr.
4956 smp_mb__after_spinlock();
4958 /* Promote REQ to ACT */
4959 rq->clock_update_flags <<= 1;
4960 update_rq_clock(rq);
4962 switch_count = &prev->nivcsw;
4965 * We must load prev->state once (task_struct::state is volatile), such
4968 * - we form a control dependency vs deactivate_task() below.
4969 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
4971 prev_state = prev->state;
4972 if (!preempt && prev_state) {
4973 if (signal_pending_state(prev_state, prev)) {
4974 prev->state = TASK_RUNNING;
4976 prev->sched_contributes_to_load =
4977 (prev_state & TASK_UNINTERRUPTIBLE) &&
4978 !(prev_state & TASK_NOLOAD) &&
4979 !(prev->flags & PF_FROZEN);
4981 if (prev->sched_contributes_to_load)
4982 rq->nr_uninterruptible++;
4985 * __schedule() ttwu()
4986 * prev_state = prev->state; if (p->on_rq && ...)
4987 * if (prev_state) goto out;
4988 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
4989 * p->state = TASK_WAKING
4991 * Where __schedule() and ttwu() have matching control dependencies.
4993 * After this, schedule() must not care about p->state any more.
4995 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4997 if (prev->in_iowait) {
4998 atomic_inc(&rq->nr_iowait);
4999 delayacct_blkio_start();
5002 switch_count = &prev->nvcsw;
5005 next = pick_next_task(rq, prev, &rf);
5006 clear_tsk_need_resched(prev);
5007 clear_preempt_need_resched();
5009 if (likely(prev != next)) {
5012 * RCU users of rcu_dereference(rq->curr) may not see
5013 * changes to task_struct made by pick_next_task().
5015 RCU_INIT_POINTER(rq->curr, next);
5017 * The membarrier system call requires each architecture
5018 * to have a full memory barrier after updating
5019 * rq->curr, before returning to user-space.
5021 * Here are the schemes providing that barrier on the
5022 * various architectures:
5023 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
5024 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
5025 * - finish_lock_switch() for weakly-ordered
5026 * architectures where spin_unlock is a full barrier,
5027 * - switch_to() for arm64 (weakly-ordered, spin_unlock
5028 * is a RELEASE barrier),
5032 migrate_disable_switch(rq, prev);
5033 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
5035 trace_sched_switch(preempt, prev, next);
5037 /* Also unlocks the rq: */
5038 rq = context_switch(rq, prev, next, &rf);
5040 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
5042 rq_unpin_lock(rq, &rf);
5043 __balance_callbacks(rq);
5044 raw_spin_unlock_irq(&rq->lock);
5048 void __noreturn do_task_dead(void)
5050 /* Causes final put_task_struct in finish_task_switch(): */
5051 set_special_state(TASK_DEAD);
5053 /* Tell freezer to ignore us: */
5054 current->flags |= PF_NOFREEZE;
5059 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
5064 static inline void sched_submit_work(struct task_struct *tsk)
5066 unsigned int task_flags;
5071 task_flags = tsk->flags;
5073 * If a worker went to sleep, notify and ask workqueue whether
5074 * it wants to wake up a task to maintain concurrency.
5075 * As this function is called inside the schedule() context,
5076 * we disable preemption to avoid it calling schedule() again
5077 * in the possible wakeup of a kworker and because wq_worker_sleeping()
5080 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5082 if (task_flags & PF_WQ_WORKER)
5083 wq_worker_sleeping(tsk);
5085 io_wq_worker_sleeping(tsk);
5086 preempt_enable_no_resched();
5089 if (tsk_is_pi_blocked(tsk))
5093 * If we are going to sleep and we have plugged IO queued,
5094 * make sure to submit it to avoid deadlocks.
5096 if (blk_needs_flush_plug(tsk))
5097 blk_schedule_flush_plug(tsk);
5100 static void sched_update_worker(struct task_struct *tsk)
5102 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
5103 if (tsk->flags & PF_WQ_WORKER)
5104 wq_worker_running(tsk);
5106 io_wq_worker_running(tsk);
5110 asmlinkage __visible void __sched schedule(void)
5112 struct task_struct *tsk = current;
5114 sched_submit_work(tsk);
5118 sched_preempt_enable_no_resched();
5119 } while (need_resched());
5120 sched_update_worker(tsk);
5122 EXPORT_SYMBOL(schedule);
5125 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
5126 * state (have scheduled out non-voluntarily) by making sure that all
5127 * tasks have either left the run queue or have gone into user space.
5128 * As idle tasks do not do either, they must not ever be preempted
5129 * (schedule out non-voluntarily).
5131 * schedule_idle() is similar to schedule_preempt_disable() except that it
5132 * never enables preemption because it does not call sched_submit_work().
5134 void __sched schedule_idle(void)
5137 * As this skips calling sched_submit_work(), which the idle task does
5138 * regardless because that function is a nop when the task is in a
5139 * TASK_RUNNING state, make sure this isn't used someplace that the
5140 * current task can be in any other state. Note, idle is always in the
5141 * TASK_RUNNING state.
5143 WARN_ON_ONCE(current->state);
5146 } while (need_resched());
5149 #ifdef CONFIG_CONTEXT_TRACKING
5150 asmlinkage __visible void __sched schedule_user(void)
5153 * If we come here after a random call to set_need_resched(),
5154 * or we have been woken up remotely but the IPI has not yet arrived,
5155 * we haven't yet exited the RCU idle mode. Do it here manually until
5156 * we find a better solution.
5158 * NB: There are buggy callers of this function. Ideally we
5159 * should warn if prev_state != CONTEXT_USER, but that will trigger
5160 * too frequently to make sense yet.
5162 enum ctx_state prev_state = exception_enter();
5164 exception_exit(prev_state);
5169 * schedule_preempt_disabled - called with preemption disabled
5171 * Returns with preemption disabled. Note: preempt_count must be 1
5173 void __sched schedule_preempt_disabled(void)
5175 sched_preempt_enable_no_resched();
5180 static void __sched notrace preempt_schedule_common(void)
5184 * Because the function tracer can trace preempt_count_sub()
5185 * and it also uses preempt_enable/disable_notrace(), if
5186 * NEED_RESCHED is set, the preempt_enable_notrace() called
5187 * by the function tracer will call this function again and
5188 * cause infinite recursion.
5190 * Preemption must be disabled here before the function
5191 * tracer can trace. Break up preempt_disable() into two
5192 * calls. One to disable preemption without fear of being
5193 * traced. The other to still record the preemption latency,
5194 * which can also be traced by the function tracer.
5196 preempt_disable_notrace();
5197 preempt_latency_start(1);
5199 preempt_latency_stop(1);
5200 preempt_enable_no_resched_notrace();
5203 * Check again in case we missed a preemption opportunity
5204 * between schedule and now.
5206 } while (need_resched());
5209 #ifdef CONFIG_PREEMPTION
5211 * This is the entry point to schedule() from in-kernel preemption
5212 * off of preempt_enable.
5214 asmlinkage __visible void __sched notrace preempt_schedule(void)
5217 * If there is a non-zero preempt_count or interrupts are disabled,
5218 * we do not want to preempt the current task. Just return..
5220 if (likely(!preemptible()))
5223 preempt_schedule_common();
5225 NOKPROBE_SYMBOL(preempt_schedule);
5226 EXPORT_SYMBOL(preempt_schedule);
5229 * preempt_schedule_notrace - preempt_schedule called by tracing
5231 * The tracing infrastructure uses preempt_enable_notrace to prevent
5232 * recursion and tracing preempt enabling caused by the tracing
5233 * infrastructure itself. But as tracing can happen in areas coming
5234 * from userspace or just about to enter userspace, a preempt enable
5235 * can occur before user_exit() is called. This will cause the scheduler
5236 * to be called when the system is still in usermode.
5238 * To prevent this, the preempt_enable_notrace will use this function
5239 * instead of preempt_schedule() to exit user context if needed before
5240 * calling the scheduler.
5242 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
5244 enum ctx_state prev_ctx;
5246 if (likely(!preemptible()))
5251 * Because the function tracer can trace preempt_count_sub()
5252 * and it also uses preempt_enable/disable_notrace(), if
5253 * NEED_RESCHED is set, the preempt_enable_notrace() called
5254 * by the function tracer will call this function again and
5255 * cause infinite recursion.
5257 * Preemption must be disabled here before the function
5258 * tracer can trace. Break up preempt_disable() into two
5259 * calls. One to disable preemption without fear of being
5260 * traced. The other to still record the preemption latency,
5261 * which can also be traced by the function tracer.
5263 preempt_disable_notrace();
5264 preempt_latency_start(1);
5266 * Needs preempt disabled in case user_exit() is traced
5267 * and the tracer calls preempt_enable_notrace() causing
5268 * an infinite recursion.
5270 prev_ctx = exception_enter();
5272 exception_exit(prev_ctx);
5274 preempt_latency_stop(1);
5275 preempt_enable_no_resched_notrace();
5276 } while (need_resched());
5278 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
5280 #endif /* CONFIG_PREEMPTION */
5283 * This is the entry point to schedule() from kernel preemption
5284 * off of irq context.
5285 * Note, that this is called and return with irqs disabled. This will
5286 * protect us against recursive calling from irq.
5288 asmlinkage __visible void __sched preempt_schedule_irq(void)
5290 enum ctx_state prev_state;
5292 /* Catch callers which need to be fixed */
5293 BUG_ON(preempt_count() || !irqs_disabled());
5295 prev_state = exception_enter();
5301 local_irq_disable();
5302 sched_preempt_enable_no_resched();
5303 } while (need_resched());
5305 exception_exit(prev_state);
5308 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
5311 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
5312 return try_to_wake_up(curr->private, mode, wake_flags);
5314 EXPORT_SYMBOL(default_wake_function);
5316 #ifdef CONFIG_RT_MUTEXES
5318 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
5321 prio = min(prio, pi_task->prio);
5326 static inline int rt_effective_prio(struct task_struct *p, int prio)
5328 struct task_struct *pi_task = rt_mutex_get_top_task(p);
5330 return __rt_effective_prio(pi_task, prio);
5334 * rt_mutex_setprio - set the current priority of a task
5336 * @pi_task: donor task
5338 * This function changes the 'effective' priority of a task. It does
5339 * not touch ->normal_prio like __setscheduler().
5341 * Used by the rt_mutex code to implement priority inheritance
5342 * logic. Call site only calls if the priority of the task changed.
5344 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
5346 int prio, oldprio, queued, running, queue_flag =
5347 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5348 const struct sched_class *prev_class;
5352 /* XXX used to be waiter->prio, not waiter->task->prio */
5353 prio = __rt_effective_prio(pi_task, p->normal_prio);
5356 * If nothing changed; bail early.
5358 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
5361 rq = __task_rq_lock(p, &rf);
5362 update_rq_clock(rq);
5364 * Set under pi_lock && rq->lock, such that the value can be used under
5367 * Note that there is loads of tricky to make this pointer cache work
5368 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
5369 * ensure a task is de-boosted (pi_task is set to NULL) before the
5370 * task is allowed to run again (and can exit). This ensures the pointer
5371 * points to a blocked task -- which guarantees the task is present.
5373 p->pi_top_task = pi_task;
5376 * For FIFO/RR we only need to set prio, if that matches we're done.
5378 if (prio == p->prio && !dl_prio(prio))
5382 * Idle task boosting is a nono in general. There is one
5383 * exception, when PREEMPT_RT and NOHZ is active:
5385 * The idle task calls get_next_timer_interrupt() and holds
5386 * the timer wheel base->lock on the CPU and another CPU wants
5387 * to access the timer (probably to cancel it). We can safely
5388 * ignore the boosting request, as the idle CPU runs this code
5389 * with interrupts disabled and will complete the lock
5390 * protected section without being interrupted. So there is no
5391 * real need to boost.
5393 if (unlikely(p == rq->idle)) {
5394 WARN_ON(p != rq->curr);
5395 WARN_ON(p->pi_blocked_on);
5399 trace_sched_pi_setprio(p, pi_task);
5402 if (oldprio == prio)
5403 queue_flag &= ~DEQUEUE_MOVE;
5405 prev_class = p->sched_class;
5406 queued = task_on_rq_queued(p);
5407 running = task_current(rq, p);
5409 dequeue_task(rq, p, queue_flag);
5411 put_prev_task(rq, p);
5414 * Boosting condition are:
5415 * 1. -rt task is running and holds mutex A
5416 * --> -dl task blocks on mutex A
5418 * 2. -dl task is running and holds mutex A
5419 * --> -dl task blocks on mutex A and could preempt the
5422 if (dl_prio(prio)) {
5423 if (!dl_prio(p->normal_prio) ||
5424 (pi_task && dl_prio(pi_task->prio) &&
5425 dl_entity_preempt(&pi_task->dl, &p->dl))) {
5426 p->dl.pi_se = pi_task->dl.pi_se;
5427 queue_flag |= ENQUEUE_REPLENISH;
5429 p->dl.pi_se = &p->dl;
5431 p->sched_class = &dl_sched_class;
5432 } else if (rt_prio(prio)) {
5433 if (dl_prio(oldprio))
5434 p->dl.pi_se = &p->dl;
5436 queue_flag |= ENQUEUE_HEAD;
5437 p->sched_class = &rt_sched_class;
5439 if (dl_prio(oldprio))
5440 p->dl.pi_se = &p->dl;
5441 if (rt_prio(oldprio))
5443 p->sched_class = &fair_sched_class;
5449 enqueue_task(rq, p, queue_flag);
5451 set_next_task(rq, p);
5453 check_class_changed(rq, p, prev_class, oldprio);
5455 /* Avoid rq from going away on us: */
5458 rq_unpin_lock(rq, &rf);
5459 __balance_callbacks(rq);
5460 raw_spin_unlock(&rq->lock);
5465 static inline int rt_effective_prio(struct task_struct *p, int prio)
5471 void set_user_nice(struct task_struct *p, long nice)
5473 bool queued, running;
5478 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
5481 * We have to be careful, if called from sys_setpriority(),
5482 * the task might be in the middle of scheduling on another CPU.
5484 rq = task_rq_lock(p, &rf);
5485 update_rq_clock(rq);
5488 * The RT priorities are set via sched_setscheduler(), but we still
5489 * allow the 'normal' nice value to be set - but as expected
5490 * it won't have any effect on scheduling until the task is
5491 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5493 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
5494 p->static_prio = NICE_TO_PRIO(nice);
5497 queued = task_on_rq_queued(p);
5498 running = task_current(rq, p);
5500 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
5502 put_prev_task(rq, p);
5504 p->static_prio = NICE_TO_PRIO(nice);
5505 set_load_weight(p, true);
5507 p->prio = effective_prio(p);
5510 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5512 set_next_task(rq, p);
5515 * If the task increased its priority or is running and
5516 * lowered its priority, then reschedule its CPU:
5518 p->sched_class->prio_changed(rq, p, old_prio);
5521 task_rq_unlock(rq, p, &rf);
5523 EXPORT_SYMBOL(set_user_nice);
5526 * can_nice - check if a task can reduce its nice value
5530 int can_nice(const struct task_struct *p, const int nice)
5532 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5533 int nice_rlim = nice_to_rlimit(nice);
5535 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5536 capable(CAP_SYS_NICE));
5539 #ifdef __ARCH_WANT_SYS_NICE
5542 * sys_nice - change the priority of the current process.
5543 * @increment: priority increment
5545 * sys_setpriority is a more generic, but much slower function that
5546 * does similar things.
5548 SYSCALL_DEFINE1(nice, int, increment)
5553 * Setpriority might change our priority at the same moment.
5554 * We don't have to worry. Conceptually one call occurs first
5555 * and we have a single winner.
5557 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
5558 nice = task_nice(current) + increment;
5560 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
5561 if (increment < 0 && !can_nice(current, nice))
5564 retval = security_task_setnice(current, nice);
5568 set_user_nice(current, nice);
5575 * task_prio - return the priority value of a given task.
5576 * @p: the task in question.
5578 * Return: The priority value as seen by users in /proc.
5579 * RT tasks are offset by -200. Normal tasks are centered
5580 * around 0, value goes from -16 to +15.
5582 int task_prio(const struct task_struct *p)
5584 return p->prio - MAX_RT_PRIO;
5588 * idle_cpu - is a given CPU idle currently?
5589 * @cpu: the processor in question.
5591 * Return: 1 if the CPU is currently idle. 0 otherwise.
5593 int idle_cpu(int cpu)
5595 struct rq *rq = cpu_rq(cpu);
5597 if (rq->curr != rq->idle)
5604 if (rq->ttwu_pending)
5612 * available_idle_cpu - is a given CPU idle for enqueuing work.
5613 * @cpu: the CPU in question.
5615 * Return: 1 if the CPU is currently idle. 0 otherwise.
5617 int available_idle_cpu(int cpu)
5622 if (vcpu_is_preempted(cpu))
5629 * idle_task - return the idle task for a given CPU.
5630 * @cpu: the processor in question.
5632 * Return: The idle task for the CPU @cpu.
5634 struct task_struct *idle_task(int cpu)
5636 return cpu_rq(cpu)->idle;
5640 * find_process_by_pid - find a process with a matching PID value.
5641 * @pid: the pid in question.
5643 * The task of @pid, if found. %NULL otherwise.
5645 static struct task_struct *find_process_by_pid(pid_t pid)
5647 return pid ? find_task_by_vpid(pid) : current;
5651 * sched_setparam() passes in -1 for its policy, to let the functions
5652 * it calls know not to change it.
5654 #define SETPARAM_POLICY -1
5656 static void __setscheduler_params(struct task_struct *p,
5657 const struct sched_attr *attr)
5659 int policy = attr->sched_policy;
5661 if (policy == SETPARAM_POLICY)
5666 if (dl_policy(policy))
5667 __setparam_dl(p, attr);
5668 else if (fair_policy(policy))
5669 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
5672 * __sched_setscheduler() ensures attr->sched_priority == 0 when
5673 * !rt_policy. Always setting this ensures that things like
5674 * getparam()/getattr() don't report silly values for !rt tasks.
5676 p->rt_priority = attr->sched_priority;
5677 p->normal_prio = normal_prio(p);
5678 set_load_weight(p, true);
5681 /* Actually do priority change: must hold pi & rq lock. */
5682 static void __setscheduler(struct rq *rq, struct task_struct *p,
5683 const struct sched_attr *attr, bool keep_boost)
5686 * If params can't change scheduling class changes aren't allowed
5689 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
5692 __setscheduler_params(p, attr);
5695 * Keep a potential priority boosting if called from
5696 * sched_setscheduler().
5698 p->prio = normal_prio(p);
5700 p->prio = rt_effective_prio(p, p->prio);
5702 if (dl_prio(p->prio))
5703 p->sched_class = &dl_sched_class;
5704 else if (rt_prio(p->prio))
5705 p->sched_class = &rt_sched_class;
5707 p->sched_class = &fair_sched_class;
5711 * Check the target process has a UID that matches the current process's:
5713 static bool check_same_owner(struct task_struct *p)
5715 const struct cred *cred = current_cred(), *pcred;
5719 pcred = __task_cred(p);
5720 match = (uid_eq(cred->euid, pcred->euid) ||
5721 uid_eq(cred->euid, pcred->uid));
5726 static int __sched_setscheduler(struct task_struct *p,
5727 const struct sched_attr *attr,
5730 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
5731 MAX_RT_PRIO - 1 - attr->sched_priority;
5732 int retval, oldprio, oldpolicy = -1, queued, running;
5733 int new_effective_prio, policy = attr->sched_policy;
5734 const struct sched_class *prev_class;
5735 struct callback_head *head;
5738 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5741 /* The pi code expects interrupts enabled */
5742 BUG_ON(pi && in_interrupt());
5744 /* Double check policy once rq lock held: */
5746 reset_on_fork = p->sched_reset_on_fork;
5747 policy = oldpolicy = p->policy;
5749 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
5751 if (!valid_policy(policy))
5755 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
5759 * Valid priorities for SCHED_FIFO and SCHED_RR are
5760 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5761 * SCHED_BATCH and SCHED_IDLE is 0.
5763 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
5764 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
5766 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
5767 (rt_policy(policy) != (attr->sched_priority != 0)))
5771 * Allow unprivileged RT tasks to decrease priority:
5773 if (user && !capable(CAP_SYS_NICE)) {
5774 if (fair_policy(policy)) {
5775 if (attr->sched_nice < task_nice(p) &&
5776 !can_nice(p, attr->sched_nice))
5780 if (rt_policy(policy)) {
5781 unsigned long rlim_rtprio =
5782 task_rlimit(p, RLIMIT_RTPRIO);
5784 /* Can't set/change the rt policy: */
5785 if (policy != p->policy && !rlim_rtprio)
5788 /* Can't increase priority: */
5789 if (attr->sched_priority > p->rt_priority &&
5790 attr->sched_priority > rlim_rtprio)
5795 * Can't set/change SCHED_DEADLINE policy at all for now
5796 * (safest behavior); in the future we would like to allow
5797 * unprivileged DL tasks to increase their relative deadline
5798 * or reduce their runtime (both ways reducing utilization)
5800 if (dl_policy(policy))
5804 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5805 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5807 if (task_has_idle_policy(p) && !idle_policy(policy)) {
5808 if (!can_nice(p, task_nice(p)))
5812 /* Can't change other user's priorities: */
5813 if (!check_same_owner(p))
5816 /* Normal users shall not reset the sched_reset_on_fork flag: */
5817 if (p->sched_reset_on_fork && !reset_on_fork)
5822 if (attr->sched_flags & SCHED_FLAG_SUGOV)
5825 retval = security_task_setscheduler(p);
5830 /* Update task specific "requested" clamps */
5831 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
5832 retval = uclamp_validate(p, attr);
5841 * Make sure no PI-waiters arrive (or leave) while we are
5842 * changing the priority of the task:
5844 * To be able to change p->policy safely, the appropriate
5845 * runqueue lock must be held.
5847 rq = task_rq_lock(p, &rf);
5848 update_rq_clock(rq);
5851 * Changing the policy of the stop threads its a very bad idea:
5853 if (p == rq->stop) {
5859 * If not changing anything there's no need to proceed further,
5860 * but store a possible modification of reset_on_fork.
5862 if (unlikely(policy == p->policy)) {
5863 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
5865 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
5867 if (dl_policy(policy) && dl_param_changed(p, attr))
5869 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
5872 p->sched_reset_on_fork = reset_on_fork;
5879 #ifdef CONFIG_RT_GROUP_SCHED
5881 * Do not allow realtime tasks into groups that have no runtime
5884 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5885 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5886 !task_group_is_autogroup(task_group(p))) {
5892 if (dl_bandwidth_enabled() && dl_policy(policy) &&
5893 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
5894 cpumask_t *span = rq->rd->span;
5897 * Don't allow tasks with an affinity mask smaller than
5898 * the entire root_domain to become SCHED_DEADLINE. We
5899 * will also fail if there's no bandwidth available.
5901 if (!cpumask_subset(span, p->cpus_ptr) ||
5902 rq->rd->dl_bw.bw == 0) {
5910 /* Re-check policy now with rq lock held: */
5911 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5912 policy = oldpolicy = -1;
5913 task_rq_unlock(rq, p, &rf);
5915 cpuset_read_unlock();
5920 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5921 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5924 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
5929 p->sched_reset_on_fork = reset_on_fork;
5934 * Take priority boosted tasks into account. If the new
5935 * effective priority is unchanged, we just store the new
5936 * normal parameters and do not touch the scheduler class and
5937 * the runqueue. This will be done when the task deboost
5940 new_effective_prio = rt_effective_prio(p, newprio);
5941 if (new_effective_prio == oldprio)
5942 queue_flags &= ~DEQUEUE_MOVE;
5945 queued = task_on_rq_queued(p);
5946 running = task_current(rq, p);
5948 dequeue_task(rq, p, queue_flags);
5950 put_prev_task(rq, p);
5952 prev_class = p->sched_class;
5954 __setscheduler(rq, p, attr, pi);
5955 __setscheduler_uclamp(p, attr);
5959 * We enqueue to tail when the priority of a task is
5960 * increased (user space view).
5962 if (oldprio < p->prio)
5963 queue_flags |= ENQUEUE_HEAD;
5965 enqueue_task(rq, p, queue_flags);
5968 set_next_task(rq, p);
5970 check_class_changed(rq, p, prev_class, oldprio);
5972 /* Avoid rq from going away on us: */
5974 head = splice_balance_callbacks(rq);
5975 task_rq_unlock(rq, p, &rf);
5978 cpuset_read_unlock();
5979 rt_mutex_adjust_pi(p);
5982 /* Run balance callbacks after we've adjusted the PI chain: */
5983 balance_callbacks(rq, head);
5989 task_rq_unlock(rq, p, &rf);
5991 cpuset_read_unlock();
5995 static int _sched_setscheduler(struct task_struct *p, int policy,
5996 const struct sched_param *param, bool check)
5998 struct sched_attr attr = {
5999 .sched_policy = policy,
6000 .sched_priority = param->sched_priority,
6001 .sched_nice = PRIO_TO_NICE(p->static_prio),
6004 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
6005 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
6006 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
6007 policy &= ~SCHED_RESET_ON_FORK;
6008 attr.sched_policy = policy;
6011 return __sched_setscheduler(p, &attr, check, true);
6014 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6015 * @p: the task in question.
6016 * @policy: new policy.
6017 * @param: structure containing the new RT priority.
6019 * Use sched_set_fifo(), read its comment.
6021 * Return: 0 on success. An error code otherwise.
6023 * NOTE that the task may be already dead.
6025 int sched_setscheduler(struct task_struct *p, int policy,
6026 const struct sched_param *param)
6028 return _sched_setscheduler(p, policy, param, true);
6031 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
6033 return __sched_setscheduler(p, attr, true, true);
6036 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
6038 return __sched_setscheduler(p, attr, false, true);
6042 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6043 * @p: the task in question.
6044 * @policy: new policy.
6045 * @param: structure containing the new RT priority.
6047 * Just like sched_setscheduler, only don't bother checking if the
6048 * current context has permission. For example, this is needed in
6049 * stop_machine(): we create temporary high priority worker threads,
6050 * but our caller might not have that capability.
6052 * Return: 0 on success. An error code otherwise.
6054 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6055 const struct sched_param *param)
6057 return _sched_setscheduler(p, policy, param, false);
6061 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
6062 * incapable of resource management, which is the one thing an OS really should
6065 * This is of course the reason it is limited to privileged users only.
6067 * Worse still; it is fundamentally impossible to compose static priority
6068 * workloads. You cannot take two correctly working static prio workloads
6069 * and smash them together and still expect them to work.
6071 * For this reason 'all' FIFO tasks the kernel creates are basically at:
6075 * The administrator _MUST_ configure the system, the kernel simply doesn't
6076 * know enough information to make a sensible choice.
6078 void sched_set_fifo(struct task_struct *p)
6080 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
6081 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
6083 EXPORT_SYMBOL_GPL(sched_set_fifo);
6086 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
6088 void sched_set_fifo_low(struct task_struct *p)
6090 struct sched_param sp = { .sched_priority = 1 };
6091 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
6093 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
6095 void sched_set_normal(struct task_struct *p, int nice)
6097 struct sched_attr attr = {
6098 .sched_policy = SCHED_NORMAL,
6101 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
6103 EXPORT_SYMBOL_GPL(sched_set_normal);
6106 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6108 struct sched_param lparam;
6109 struct task_struct *p;
6112 if (!param || pid < 0)
6114 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6119 p = find_process_by_pid(pid);
6125 retval = sched_setscheduler(p, policy, &lparam);
6133 * Mimics kernel/events/core.c perf_copy_attr().
6135 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
6140 /* Zero the full structure, so that a short copy will be nice: */
6141 memset(attr, 0, sizeof(*attr));
6143 ret = get_user(size, &uattr->size);
6147 /* ABI compatibility quirk: */
6149 size = SCHED_ATTR_SIZE_VER0;
6150 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
6153 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
6160 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
6161 size < SCHED_ATTR_SIZE_VER1)
6165 * XXX: Do we want to be lenient like existing syscalls; or do we want
6166 * to be strict and return an error on out-of-bounds values?
6168 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
6173 put_user(sizeof(*attr), &uattr->size);
6178 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6179 * @pid: the pid in question.
6180 * @policy: new policy.
6181 * @param: structure containing the new RT priority.
6183 * Return: 0 on success. An error code otherwise.
6185 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
6190 return do_sched_setscheduler(pid, policy, param);
6194 * sys_sched_setparam - set/change the RT priority of a thread
6195 * @pid: the pid in question.
6196 * @param: structure containing the new RT priority.
6198 * Return: 0 on success. An error code otherwise.
6200 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6202 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
6206 * sys_sched_setattr - same as above, but with extended sched_attr
6207 * @pid: the pid in question.
6208 * @uattr: structure containing the extended parameters.
6209 * @flags: for future extension.
6211 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
6212 unsigned int, flags)
6214 struct sched_attr attr;
6215 struct task_struct *p;
6218 if (!uattr || pid < 0 || flags)
6221 retval = sched_copy_attr(uattr, &attr);
6225 if ((int)attr.sched_policy < 0)
6227 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
6228 attr.sched_policy = SETPARAM_POLICY;
6232 p = find_process_by_pid(pid);
6238 retval = sched_setattr(p, &attr);
6246 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6247 * @pid: the pid in question.
6249 * Return: On success, the policy of the thread. Otherwise, a negative error
6252 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6254 struct task_struct *p;
6262 p = find_process_by_pid(pid);
6264 retval = security_task_getscheduler(p);
6267 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6274 * sys_sched_getparam - get the RT priority of a thread
6275 * @pid: the pid in question.
6276 * @param: structure containing the RT priority.
6278 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
6281 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6283 struct sched_param lp = { .sched_priority = 0 };
6284 struct task_struct *p;
6287 if (!param || pid < 0)
6291 p = find_process_by_pid(pid);
6296 retval = security_task_getscheduler(p);
6300 if (task_has_rt_policy(p))
6301 lp.sched_priority = p->rt_priority;
6305 * This one might sleep, we cannot do it with a spinlock held ...
6307 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6317 * Copy the kernel size attribute structure (which might be larger
6318 * than what user-space knows about) to user-space.
6320 * Note that all cases are valid: user-space buffer can be larger or
6321 * smaller than the kernel-space buffer. The usual case is that both
6322 * have the same size.
6325 sched_attr_copy_to_user(struct sched_attr __user *uattr,
6326 struct sched_attr *kattr,
6329 unsigned int ksize = sizeof(*kattr);
6331 if (!access_ok(uattr, usize))
6335 * sched_getattr() ABI forwards and backwards compatibility:
6337 * If usize == ksize then we just copy everything to user-space and all is good.
6339 * If usize < ksize then we only copy as much as user-space has space for,
6340 * this keeps ABI compatibility as well. We skip the rest.
6342 * If usize > ksize then user-space is using a newer version of the ABI,
6343 * which part the kernel doesn't know about. Just ignore it - tooling can
6344 * detect the kernel's knowledge of attributes from the attr->size value
6345 * which is set to ksize in this case.
6347 kattr->size = min(usize, ksize);
6349 if (copy_to_user(uattr, kattr, kattr->size))
6356 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
6357 * @pid: the pid in question.
6358 * @uattr: structure containing the extended parameters.
6359 * @usize: sizeof(attr) for fwd/bwd comp.
6360 * @flags: for future extension.
6362 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
6363 unsigned int, usize, unsigned int, flags)
6365 struct sched_attr kattr = { };
6366 struct task_struct *p;
6369 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
6370 usize < SCHED_ATTR_SIZE_VER0 || flags)
6374 p = find_process_by_pid(pid);
6379 retval = security_task_getscheduler(p);
6383 kattr.sched_policy = p->policy;
6384 if (p->sched_reset_on_fork)
6385 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
6386 if (task_has_dl_policy(p))
6387 __getparam_dl(p, &kattr);
6388 else if (task_has_rt_policy(p))
6389 kattr.sched_priority = p->rt_priority;
6391 kattr.sched_nice = task_nice(p);
6393 #ifdef CONFIG_UCLAMP_TASK
6395 * This could race with another potential updater, but this is fine
6396 * because it'll correctly read the old or the new value. We don't need
6397 * to guarantee who wins the race as long as it doesn't return garbage.
6399 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
6400 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
6405 return sched_attr_copy_to_user(uattr, &kattr, usize);
6412 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6414 cpumask_var_t cpus_allowed, new_mask;
6415 struct task_struct *p;
6420 p = find_process_by_pid(pid);
6426 /* Prevent p going away */
6430 if (p->flags & PF_NO_SETAFFINITY) {
6434 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6438 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6440 goto out_free_cpus_allowed;
6443 if (!check_same_owner(p)) {
6445 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
6447 goto out_free_new_mask;
6452 retval = security_task_setscheduler(p);
6454 goto out_free_new_mask;
6457 cpuset_cpus_allowed(p, cpus_allowed);
6458 cpumask_and(new_mask, in_mask, cpus_allowed);
6461 * Since bandwidth control happens on root_domain basis,
6462 * if admission test is enabled, we only admit -deadline
6463 * tasks allowed to run on all the CPUs in the task's
6467 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
6469 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
6472 goto out_free_new_mask;
6478 retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
6481 cpuset_cpus_allowed(p, cpus_allowed);
6482 if (!cpumask_subset(new_mask, cpus_allowed)) {
6484 * We must have raced with a concurrent cpuset
6485 * update. Just reset the cpus_allowed to the
6486 * cpuset's cpus_allowed
6488 cpumask_copy(new_mask, cpus_allowed);
6493 free_cpumask_var(new_mask);
6494 out_free_cpus_allowed:
6495 free_cpumask_var(cpus_allowed);
6501 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6502 struct cpumask *new_mask)
6504 if (len < cpumask_size())
6505 cpumask_clear(new_mask);
6506 else if (len > cpumask_size())
6507 len = cpumask_size();
6509 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6513 * sys_sched_setaffinity - set the CPU affinity of a process
6514 * @pid: pid of the process
6515 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6516 * @user_mask_ptr: user-space pointer to the new CPU mask
6518 * Return: 0 on success. An error code otherwise.
6520 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6521 unsigned long __user *, user_mask_ptr)
6523 cpumask_var_t new_mask;
6526 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6529 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6531 retval = sched_setaffinity(pid, new_mask);
6532 free_cpumask_var(new_mask);
6536 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6538 struct task_struct *p;
6539 unsigned long flags;
6545 p = find_process_by_pid(pid);
6549 retval = security_task_getscheduler(p);
6553 raw_spin_lock_irqsave(&p->pi_lock, flags);
6554 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
6555 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6564 * sys_sched_getaffinity - get the CPU affinity of a process
6565 * @pid: pid of the process
6566 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6567 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6569 * Return: size of CPU mask copied to user_mask_ptr on success. An
6570 * error code otherwise.
6572 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6573 unsigned long __user *, user_mask_ptr)
6578 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6580 if (len & (sizeof(unsigned long)-1))
6583 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6586 ret = sched_getaffinity(pid, mask);
6588 unsigned int retlen = min(len, cpumask_size());
6590 if (copy_to_user(user_mask_ptr, mask, retlen))
6595 free_cpumask_var(mask);
6600 static void do_sched_yield(void)
6605 rq = this_rq_lock_irq(&rf);
6607 schedstat_inc(rq->yld_count);
6608 current->sched_class->yield_task(rq);
6611 rq_unlock_irq(rq, &rf);
6612 sched_preempt_enable_no_resched();
6618 * sys_sched_yield - yield the current processor to other threads.
6620 * This function yields the current CPU to other tasks. If there are no
6621 * other threads running on this CPU then this function will return.
6625 SYSCALL_DEFINE0(sched_yield)
6631 #ifndef CONFIG_PREEMPTION
6632 int __sched _cond_resched(void)
6634 if (should_resched(0)) {
6635 preempt_schedule_common();
6641 EXPORT_SYMBOL(_cond_resched);
6645 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6646 * call schedule, and on return reacquire the lock.
6648 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
6649 * operations here to prevent schedule() from being called twice (once via
6650 * spin_unlock(), once by hand).
6652 int __cond_resched_lock(spinlock_t *lock)
6654 int resched = should_resched(PREEMPT_LOCK_OFFSET);
6657 lockdep_assert_held(lock);
6659 if (spin_needbreak(lock) || resched) {
6662 preempt_schedule_common();
6670 EXPORT_SYMBOL(__cond_resched_lock);
6673 * yield - yield the current processor to other threads.
6675 * Do not ever use this function, there's a 99% chance you're doing it wrong.
6677 * The scheduler is at all times free to pick the calling task as the most
6678 * eligible task to run, if removing the yield() call from your code breaks
6679 * it, it's already broken.
6681 * Typical broken usage is:
6686 * where one assumes that yield() will let 'the other' process run that will
6687 * make event true. If the current task is a SCHED_FIFO task that will never
6688 * happen. Never use yield() as a progress guarantee!!
6690 * If you want to use yield() to wait for something, use wait_event().
6691 * If you want to use yield() to be 'nice' for others, use cond_resched().
6692 * If you still want to use yield(), do not!
6694 void __sched yield(void)
6696 set_current_state(TASK_RUNNING);
6699 EXPORT_SYMBOL(yield);
6702 * yield_to - yield the current processor to another thread in
6703 * your thread group, or accelerate that thread toward the
6704 * processor it's on.
6706 * @preempt: whether task preemption is allowed or not
6708 * It's the caller's job to ensure that the target task struct
6709 * can't go away on us before we can do any checks.
6712 * true (>0) if we indeed boosted the target task.
6713 * false (0) if we failed to boost the target.
6714 * -ESRCH if there's no task to yield to.
6716 int __sched yield_to(struct task_struct *p, bool preempt)
6718 struct task_struct *curr = current;
6719 struct rq *rq, *p_rq;
6720 unsigned long flags;
6723 local_irq_save(flags);
6729 * If we're the only runnable task on the rq and target rq also
6730 * has only one task, there's absolutely no point in yielding.
6732 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
6737 double_rq_lock(rq, p_rq);
6738 if (task_rq(p) != p_rq) {
6739 double_rq_unlock(rq, p_rq);
6743 if (!curr->sched_class->yield_to_task)
6746 if (curr->sched_class != p->sched_class)
6749 if (task_running(p_rq, p) || p->state)
6752 yielded = curr->sched_class->yield_to_task(rq, p);
6754 schedstat_inc(rq->yld_count);
6756 * Make p's CPU reschedule; pick_next_entity takes care of
6759 if (preempt && rq != p_rq)
6764 double_rq_unlock(rq, p_rq);
6766 local_irq_restore(flags);
6773 EXPORT_SYMBOL_GPL(yield_to);
6775 int io_schedule_prepare(void)
6777 int old_iowait = current->in_iowait;
6779 current->in_iowait = 1;
6780 blk_schedule_flush_plug(current);
6785 void io_schedule_finish(int token)
6787 current->in_iowait = token;
6791 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6792 * that process accounting knows that this is a task in IO wait state.
6794 long __sched io_schedule_timeout(long timeout)
6799 token = io_schedule_prepare();
6800 ret = schedule_timeout(timeout);
6801 io_schedule_finish(token);
6805 EXPORT_SYMBOL(io_schedule_timeout);
6807 void __sched io_schedule(void)
6811 token = io_schedule_prepare();
6813 io_schedule_finish(token);
6815 EXPORT_SYMBOL(io_schedule);
6818 * sys_sched_get_priority_max - return maximum RT priority.
6819 * @policy: scheduling class.
6821 * Return: On success, this syscall returns the maximum
6822 * rt_priority that can be used by a given scheduling class.
6823 * On failure, a negative error code is returned.
6825 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6832 ret = MAX_USER_RT_PRIO-1;
6834 case SCHED_DEADLINE:
6845 * sys_sched_get_priority_min - return minimum RT priority.
6846 * @policy: scheduling class.
6848 * Return: On success, this syscall returns the minimum
6849 * rt_priority that can be used by a given scheduling class.
6850 * On failure, a negative error code is returned.
6852 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6861 case SCHED_DEADLINE:
6870 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
6872 struct task_struct *p;
6873 unsigned int time_slice;
6883 p = find_process_by_pid(pid);
6887 retval = security_task_getscheduler(p);
6891 rq = task_rq_lock(p, &rf);
6893 if (p->sched_class->get_rr_interval)
6894 time_slice = p->sched_class->get_rr_interval(rq, p);
6895 task_rq_unlock(rq, p, &rf);
6898 jiffies_to_timespec64(time_slice, t);
6907 * sys_sched_rr_get_interval - return the default timeslice of a process.
6908 * @pid: pid of the process.
6909 * @interval: userspace pointer to the timeslice value.
6911 * this syscall writes the default timeslice value of a given process
6912 * into the user-space timespec buffer. A value of '0' means infinity.
6914 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6917 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6918 struct __kernel_timespec __user *, interval)
6920 struct timespec64 t;
6921 int retval = sched_rr_get_interval(pid, &t);
6924 retval = put_timespec64(&t, interval);
6929 #ifdef CONFIG_COMPAT_32BIT_TIME
6930 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
6931 struct old_timespec32 __user *, interval)
6933 struct timespec64 t;
6934 int retval = sched_rr_get_interval(pid, &t);
6937 retval = put_old_timespec32(&t, interval);
6942 void sched_show_task(struct task_struct *p)
6944 unsigned long free = 0;
6947 if (!try_get_task_stack(p))
6950 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
6952 if (p->state == TASK_RUNNING)
6953 pr_cont(" running task ");
6954 #ifdef CONFIG_DEBUG_STACK_USAGE
6955 free = stack_not_used(p);
6960 ppid = task_pid_nr(rcu_dereference(p->real_parent));
6962 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
6963 free, task_pid_nr(p), ppid,
6964 (unsigned long)task_thread_info(p)->flags);
6966 print_worker_info(KERN_INFO, p);
6967 print_stop_info(KERN_INFO, p);
6968 show_stack(p, NULL, KERN_INFO);
6971 EXPORT_SYMBOL_GPL(sched_show_task);
6974 state_filter_match(unsigned long state_filter, struct task_struct *p)
6976 /* no filter, everything matches */
6980 /* filter, but doesn't match */
6981 if (!(p->state & state_filter))
6985 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6988 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
6995 void show_state_filter(unsigned long state_filter)
6997 struct task_struct *g, *p;
7000 for_each_process_thread(g, p) {
7002 * reset the NMI-timeout, listing all files on a slow
7003 * console might take a lot of time:
7004 * Also, reset softlockup watchdogs on all CPUs, because
7005 * another CPU might be blocked waiting for us to process
7008 touch_nmi_watchdog();
7009 touch_all_softlockup_watchdogs();
7010 if (state_filter_match(state_filter, p))
7014 #ifdef CONFIG_SCHED_DEBUG
7016 sysrq_sched_debug_show();
7020 * Only show locks if all tasks are dumped:
7023 debug_show_all_locks();
7027 * init_idle - set up an idle thread for a given CPU
7028 * @idle: task in question
7029 * @cpu: CPU the idle task belongs to
7031 * NOTE: this function does not set the idle thread's NEED_RESCHED
7032 * flag, to make booting more robust.
7034 void init_idle(struct task_struct *idle, int cpu)
7036 struct rq *rq = cpu_rq(cpu);
7037 unsigned long flags;
7039 __sched_fork(0, idle);
7041 raw_spin_lock_irqsave(&idle->pi_lock, flags);
7042 raw_spin_lock(&rq->lock);
7044 idle->state = TASK_RUNNING;
7045 idle->se.exec_start = sched_clock();
7046 idle->flags |= PF_IDLE;
7048 scs_task_reset(idle);
7049 kasan_unpoison_task_stack(idle);
7053 * It's possible that init_idle() gets called multiple times on a task,
7054 * in that case do_set_cpus_allowed() will not do the right thing.
7056 * And since this is boot we can forgo the serialization.
7058 set_cpus_allowed_common(idle, cpumask_of(cpu), 0);
7061 * We're having a chicken and egg problem, even though we are
7062 * holding rq->lock, the CPU isn't yet set to this CPU so the
7063 * lockdep check in task_group() will fail.
7065 * Similar case to sched_fork(). / Alternatively we could
7066 * use task_rq_lock() here and obtain the other rq->lock.
7071 __set_task_cpu(idle, cpu);
7075 rcu_assign_pointer(rq->curr, idle);
7076 idle->on_rq = TASK_ON_RQ_QUEUED;
7080 raw_spin_unlock(&rq->lock);
7081 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
7083 /* Set the preempt count _outside_ the spinlocks! */
7084 init_idle_preempt_count(idle, cpu);
7087 * The idle tasks have their own, simple scheduling class:
7089 idle->sched_class = &idle_sched_class;
7090 ftrace_graph_init_idle_task(idle, cpu);
7091 vtime_init_idle(idle, cpu);
7093 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
7099 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
7100 const struct cpumask *trial)
7104 if (!cpumask_weight(cur))
7107 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
7112 int task_can_attach(struct task_struct *p,
7113 const struct cpumask *cs_cpus_allowed)
7118 * Kthreads which disallow setaffinity shouldn't be moved
7119 * to a new cpuset; we don't want to change their CPU
7120 * affinity and isolating such threads by their set of
7121 * allowed nodes is unnecessary. Thus, cpusets are not
7122 * applicable for such threads. This prevents checking for
7123 * success of set_cpus_allowed_ptr() on all attached tasks
7124 * before cpus_mask may be changed.
7126 if (p->flags & PF_NO_SETAFFINITY) {
7131 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
7133 ret = dl_task_can_attach(p, cs_cpus_allowed);
7139 bool sched_smp_initialized __read_mostly;
7141 #ifdef CONFIG_NUMA_BALANCING
7142 /* Migrate current task p to target_cpu */
7143 int migrate_task_to(struct task_struct *p, int target_cpu)
7145 struct migration_arg arg = { p, target_cpu };
7146 int curr_cpu = task_cpu(p);
7148 if (curr_cpu == target_cpu)
7151 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
7154 /* TODO: This is not properly updating schedstats */
7156 trace_sched_move_numa(p, curr_cpu, target_cpu);
7157 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
7161 * Requeue a task on a given node and accurately track the number of NUMA
7162 * tasks on the runqueues
7164 void sched_setnuma(struct task_struct *p, int nid)
7166 bool queued, running;
7170 rq = task_rq_lock(p, &rf);
7171 queued = task_on_rq_queued(p);
7172 running = task_current(rq, p);
7175 dequeue_task(rq, p, DEQUEUE_SAVE);
7177 put_prev_task(rq, p);
7179 p->numa_preferred_nid = nid;
7182 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
7184 set_next_task(rq, p);
7185 task_rq_unlock(rq, p, &rf);
7187 #endif /* CONFIG_NUMA_BALANCING */
7189 #ifdef CONFIG_HOTPLUG_CPU
7191 * Ensure that the idle task is using init_mm right before its CPU goes
7194 void idle_task_exit(void)
7196 struct mm_struct *mm = current->active_mm;
7198 BUG_ON(cpu_online(smp_processor_id()));
7199 BUG_ON(current != this_rq()->idle);
7201 if (mm != &init_mm) {
7202 switch_mm(mm, &init_mm, current);
7203 finish_arch_post_lock_switch();
7206 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
7209 static int __balance_push_cpu_stop(void *arg)
7211 struct task_struct *p = arg;
7212 struct rq *rq = this_rq();
7216 raw_spin_lock_irq(&p->pi_lock);
7219 update_rq_clock(rq);
7221 if (task_rq(p) == rq && task_on_rq_queued(p)) {
7222 cpu = select_fallback_rq(rq->cpu, p);
7223 rq = __migrate_task(rq, &rf, p, cpu);
7227 raw_spin_unlock_irq(&p->pi_lock);
7234 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
7237 * Ensure we only run per-cpu kthreads once the CPU goes !active.
7239 static void balance_push(struct rq *rq)
7241 struct task_struct *push_task = rq->curr;
7243 lockdep_assert_held(&rq->lock);
7244 SCHED_WARN_ON(rq->cpu != smp_processor_id());
7246 * Ensure the thing is persistent until balance_push_set(.on = false);
7248 rq->balance_callback = &balance_push_callback;
7251 * Both the cpu-hotplug and stop task are in this case and are
7252 * required to complete the hotplug process.
7254 if (is_per_cpu_kthread(push_task) || is_migration_disabled(push_task)) {
7256 * If this is the idle task on the outgoing CPU try to wake
7257 * up the hotplug control thread which might wait for the
7258 * last task to vanish. The rcuwait_active() check is
7259 * accurate here because the waiter is pinned on this CPU
7260 * and can't obviously be running in parallel.
7262 * On RT kernels this also has to check whether there are
7263 * pinned and scheduled out tasks on the runqueue. They
7264 * need to leave the migrate disabled section first.
7266 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
7267 rcuwait_active(&rq->hotplug_wait)) {
7268 raw_spin_unlock(&rq->lock);
7269 rcuwait_wake_up(&rq->hotplug_wait);
7270 raw_spin_lock(&rq->lock);
7275 get_task_struct(push_task);
7277 * Temporarily drop rq->lock such that we can wake-up the stop task.
7278 * Both preemption and IRQs are still disabled.
7280 raw_spin_unlock(&rq->lock);
7281 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
7282 this_cpu_ptr(&push_work));
7284 * At this point need_resched() is true and we'll take the loop in
7285 * schedule(). The next pick is obviously going to be the stop task
7286 * which is_per_cpu_kthread() and will push this task away.
7288 raw_spin_lock(&rq->lock);
7291 static void balance_push_set(int cpu, bool on)
7293 struct rq *rq = cpu_rq(cpu);
7296 rq_lock_irqsave(rq, &rf);
7298 rq->balance_callback = &balance_push_callback;
7300 rq->balance_callback = NULL;
7301 rq_unlock_irqrestore(rq, &rf);
7305 * Invoked from a CPUs hotplug control thread after the CPU has been marked
7306 * inactive. All tasks which are not per CPU kernel threads are either
7307 * pushed off this CPU now via balance_push() or placed on a different CPU
7308 * during wakeup. Wait until the CPU is quiescent.
7310 static void balance_hotplug_wait(void)
7312 struct rq *rq = this_rq();
7314 rcuwait_wait_event(&rq->hotplug_wait,
7315 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
7316 TASK_UNINTERRUPTIBLE);
7321 static inline void balance_push(struct rq *rq)
7325 static inline void balance_push_set(int cpu, bool on)
7329 static inline void balance_hotplug_wait(void)
7333 #endif /* CONFIG_HOTPLUG_CPU */
7335 void set_rq_online(struct rq *rq)
7338 const struct sched_class *class;
7340 cpumask_set_cpu(rq->cpu, rq->rd->online);
7343 for_each_class(class) {
7344 if (class->rq_online)
7345 class->rq_online(rq);
7350 void set_rq_offline(struct rq *rq)
7353 const struct sched_class *class;
7355 for_each_class(class) {
7356 if (class->rq_offline)
7357 class->rq_offline(rq);
7360 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7366 * used to mark begin/end of suspend/resume:
7368 static int num_cpus_frozen;
7371 * Update cpusets according to cpu_active mask. If cpusets are
7372 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7373 * around partition_sched_domains().
7375 * If we come here as part of a suspend/resume, don't touch cpusets because we
7376 * want to restore it back to its original state upon resume anyway.
7378 static void cpuset_cpu_active(void)
7380 if (cpuhp_tasks_frozen) {
7382 * num_cpus_frozen tracks how many CPUs are involved in suspend
7383 * resume sequence. As long as this is not the last online
7384 * operation in the resume sequence, just build a single sched
7385 * domain, ignoring cpusets.
7387 partition_sched_domains(1, NULL, NULL);
7388 if (--num_cpus_frozen)
7391 * This is the last CPU online operation. So fall through and
7392 * restore the original sched domains by considering the
7393 * cpuset configurations.
7395 cpuset_force_rebuild();
7397 cpuset_update_active_cpus();
7400 static int cpuset_cpu_inactive(unsigned int cpu)
7402 if (!cpuhp_tasks_frozen) {
7403 if (dl_cpu_busy(cpu))
7405 cpuset_update_active_cpus();
7408 partition_sched_domains(1, NULL, NULL);
7413 int sched_cpu_activate(unsigned int cpu)
7415 struct rq *rq = cpu_rq(cpu);
7418 balance_push_set(cpu, false);
7420 #ifdef CONFIG_SCHED_SMT
7422 * When going up, increment the number of cores with SMT present.
7424 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7425 static_branch_inc_cpuslocked(&sched_smt_present);
7427 set_cpu_active(cpu, true);
7429 if (sched_smp_initialized) {
7430 sched_domains_numa_masks_set(cpu);
7431 cpuset_cpu_active();
7435 * Put the rq online, if not already. This happens:
7437 * 1) In the early boot process, because we build the real domains
7438 * after all CPUs have been brought up.
7440 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7443 rq_lock_irqsave(rq, &rf);
7445 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7448 rq_unlock_irqrestore(rq, &rf);
7453 int sched_cpu_deactivate(unsigned int cpu)
7455 struct rq *rq = cpu_rq(cpu);
7459 set_cpu_active(cpu, false);
7461 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7462 * users of this state to go away such that all new such users will
7465 * Do sync before park smpboot threads to take care the rcu boost case.
7469 balance_push_set(cpu, true);
7471 rq_lock_irqsave(rq, &rf);
7473 update_rq_clock(rq);
7474 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7477 rq_unlock_irqrestore(rq, &rf);
7479 #ifdef CONFIG_SCHED_SMT
7481 * When going down, decrement the number of cores with SMT present.
7483 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7484 static_branch_dec_cpuslocked(&sched_smt_present);
7487 if (!sched_smp_initialized)
7490 ret = cpuset_cpu_inactive(cpu);
7492 balance_push_set(cpu, false);
7493 set_cpu_active(cpu, true);
7496 sched_domains_numa_masks_clear(cpu);
7500 static void sched_rq_cpu_starting(unsigned int cpu)
7502 struct rq *rq = cpu_rq(cpu);
7504 rq->calc_load_update = calc_load_update;
7505 update_max_interval();
7508 int sched_cpu_starting(unsigned int cpu)
7510 sched_rq_cpu_starting(cpu);
7511 sched_tick_start(cpu);
7515 #ifdef CONFIG_HOTPLUG_CPU
7518 * Invoked immediately before the stopper thread is invoked to bring the
7519 * CPU down completely. At this point all per CPU kthreads except the
7520 * hotplug thread (current) and the stopper thread (inactive) have been
7521 * either parked or have been unbound from the outgoing CPU. Ensure that
7522 * any of those which might be on the way out are gone.
7524 * If after this point a bound task is being woken on this CPU then the
7525 * responsible hotplug callback has failed to do it's job.
7526 * sched_cpu_dying() will catch it with the appropriate fireworks.
7528 int sched_cpu_wait_empty(unsigned int cpu)
7530 balance_hotplug_wait();
7535 * Since this CPU is going 'away' for a while, fold any nr_active delta we
7536 * might have. Called from the CPU stopper task after ensuring that the
7537 * stopper is the last running task on the CPU, so nr_active count is
7538 * stable. We need to take the teardown thread which is calling this into
7539 * account, so we hand in adjust = 1 to the load calculation.
7541 * Also see the comment "Global load-average calculations".
7543 static void calc_load_migrate(struct rq *rq)
7545 long delta = calc_load_fold_active(rq, 1);
7548 atomic_long_add(delta, &calc_load_tasks);
7551 int sched_cpu_dying(unsigned int cpu)
7553 struct rq *rq = cpu_rq(cpu);
7556 /* Handle pending wakeups and then migrate everything off */
7557 sched_tick_stop(cpu);
7559 rq_lock_irqsave(rq, &rf);
7560 BUG_ON(rq->nr_running != 1 || rq_has_pinned_tasks(rq));
7561 rq_unlock_irqrestore(rq, &rf);
7563 calc_load_migrate(rq);
7564 update_max_interval();
7565 nohz_balance_exit_idle(rq);
7571 void __init sched_init_smp(void)
7576 * There's no userspace yet to cause hotplug operations; hence all the
7577 * CPU masks are stable and all blatant races in the below code cannot
7580 mutex_lock(&sched_domains_mutex);
7581 sched_init_domains(cpu_active_mask);
7582 mutex_unlock(&sched_domains_mutex);
7584 /* Move init over to a non-isolated CPU */
7585 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
7587 sched_init_granularity();
7589 init_sched_rt_class();
7590 init_sched_dl_class();
7592 sched_smp_initialized = true;
7595 static int __init migration_init(void)
7597 sched_cpu_starting(smp_processor_id());
7600 early_initcall(migration_init);
7603 void __init sched_init_smp(void)
7605 sched_init_granularity();
7607 #endif /* CONFIG_SMP */
7609 int in_sched_functions(unsigned long addr)
7611 return in_lock_functions(addr) ||
7612 (addr >= (unsigned long)__sched_text_start
7613 && addr < (unsigned long)__sched_text_end);
7616 #ifdef CONFIG_CGROUP_SCHED
7618 * Default task group.
7619 * Every task in system belongs to this group at bootup.
7621 struct task_group root_task_group;
7622 LIST_HEAD(task_groups);
7624 /* Cacheline aligned slab cache for task_group */
7625 static struct kmem_cache *task_group_cache __read_mostly;
7628 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7629 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7631 void __init sched_init(void)
7633 unsigned long ptr = 0;
7636 /* Make sure the linker didn't screw up */
7637 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
7638 &fair_sched_class + 1 != &rt_sched_class ||
7639 &rt_sched_class + 1 != &dl_sched_class);
7641 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
7646 #ifdef CONFIG_FAIR_GROUP_SCHED
7647 ptr += 2 * nr_cpu_ids * sizeof(void **);
7649 #ifdef CONFIG_RT_GROUP_SCHED
7650 ptr += 2 * nr_cpu_ids * sizeof(void **);
7653 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 root_task_group.se = (struct sched_entity **)ptr;
7657 ptr += nr_cpu_ids * sizeof(void **);
7659 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7660 ptr += nr_cpu_ids * sizeof(void **);
7662 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7663 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7664 #endif /* CONFIG_FAIR_GROUP_SCHED */
7665 #ifdef CONFIG_RT_GROUP_SCHED
7666 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7667 ptr += nr_cpu_ids * sizeof(void **);
7669 root_task_group.rt_rq = (struct rt_rq **)ptr;
7670 ptr += nr_cpu_ids * sizeof(void **);
7672 #endif /* CONFIG_RT_GROUP_SCHED */
7674 #ifdef CONFIG_CPUMASK_OFFSTACK
7675 for_each_possible_cpu(i) {
7676 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7677 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7678 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7679 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7681 #endif /* CONFIG_CPUMASK_OFFSTACK */
7683 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
7684 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
7687 init_defrootdomain();
7690 #ifdef CONFIG_RT_GROUP_SCHED
7691 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7692 global_rt_period(), global_rt_runtime());
7693 #endif /* CONFIG_RT_GROUP_SCHED */
7695 #ifdef CONFIG_CGROUP_SCHED
7696 task_group_cache = KMEM_CACHE(task_group, 0);
7698 list_add(&root_task_group.list, &task_groups);
7699 INIT_LIST_HEAD(&root_task_group.children);
7700 INIT_LIST_HEAD(&root_task_group.siblings);
7701 autogroup_init(&init_task);
7702 #endif /* CONFIG_CGROUP_SCHED */
7704 for_each_possible_cpu(i) {
7708 raw_spin_lock_init(&rq->lock);
7710 rq->calc_load_active = 0;
7711 rq->calc_load_update = jiffies + LOAD_FREQ;
7712 init_cfs_rq(&rq->cfs);
7713 init_rt_rq(&rq->rt);
7714 init_dl_rq(&rq->dl);
7715 #ifdef CONFIG_FAIR_GROUP_SCHED
7716 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7717 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7719 * How much CPU bandwidth does root_task_group get?
7721 * In case of task-groups formed thr' the cgroup filesystem, it
7722 * gets 100% of the CPU resources in the system. This overall
7723 * system CPU resource is divided among the tasks of
7724 * root_task_group and its child task-groups in a fair manner,
7725 * based on each entity's (task or task-group's) weight
7726 * (se->load.weight).
7728 * In other words, if root_task_group has 10 tasks of weight
7729 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7730 * then A0's share of the CPU resource is:
7732 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7734 * We achieve this by letting root_task_group's tasks sit
7735 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7737 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7738 #endif /* CONFIG_FAIR_GROUP_SCHED */
7740 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7741 #ifdef CONFIG_RT_GROUP_SCHED
7742 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7747 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7748 rq->balance_callback = NULL;
7749 rq->active_balance = 0;
7750 rq->next_balance = jiffies;
7755 rq->avg_idle = 2*sysctl_sched_migration_cost;
7756 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7758 INIT_LIST_HEAD(&rq->cfs_tasks);
7760 rq_attach_root(rq, &def_root_domain);
7761 #ifdef CONFIG_NO_HZ_COMMON
7762 rq->last_blocked_load_update_tick = jiffies;
7763 atomic_set(&rq->nohz_flags, 0);
7765 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
7767 #ifdef CONFIG_HOTPLUG_CPU
7768 rcuwait_init(&rq->hotplug_wait);
7770 #endif /* CONFIG_SMP */
7772 atomic_set(&rq->nr_iowait, 0);
7775 set_load_weight(&init_task, false);
7778 * The boot idle thread does lazy MMU switching as well:
7781 enter_lazy_tlb(&init_mm, current);
7784 * Make us the idle thread. Technically, schedule() should not be
7785 * called from this thread, however somewhere below it might be,
7786 * but because we are the idle thread, we just pick up running again
7787 * when this runqueue becomes "idle".
7789 init_idle(current, smp_processor_id());
7791 calc_load_update = jiffies + LOAD_FREQ;
7794 idle_thread_set_boot_cpu();
7796 init_sched_fair_class();
7804 scheduler_running = 1;
7807 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7808 static inline int preempt_count_equals(int preempt_offset)
7810 int nested = preempt_count() + rcu_preempt_depth();
7812 return (nested == preempt_offset);
7815 void __might_sleep(const char *file, int line, int preempt_offset)
7818 * Blocking primitives will set (and therefore destroy) current->state,
7819 * since we will exit with TASK_RUNNING make sure we enter with it,
7820 * otherwise we will destroy state.
7822 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7823 "do not call blocking ops when !TASK_RUNNING; "
7824 "state=%lx set at [<%p>] %pS\n",
7826 (void *)current->task_state_change,
7827 (void *)current->task_state_change);
7829 ___might_sleep(file, line, preempt_offset);
7831 EXPORT_SYMBOL(__might_sleep);
7833 void ___might_sleep(const char *file, int line, int preempt_offset)
7835 /* Ratelimiting timestamp: */
7836 static unsigned long prev_jiffy;
7838 unsigned long preempt_disable_ip;
7840 /* WARN_ON_ONCE() by default, no rate limit required: */
7843 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7844 !is_idle_task(current) && !current->non_block_count) ||
7845 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
7849 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7851 prev_jiffy = jiffies;
7853 /* Save this before calling printk(), since that will clobber it: */
7854 preempt_disable_ip = get_preempt_disable_ip(current);
7857 "BUG: sleeping function called from invalid context at %s:%d\n",
7860 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
7861 in_atomic(), irqs_disabled(), current->non_block_count,
7862 current->pid, current->comm);
7864 if (task_stack_end_corrupted(current))
7865 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7867 debug_show_held_locks(current);
7868 if (irqs_disabled())
7869 print_irqtrace_events(current);
7870 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7871 && !preempt_count_equals(preempt_offset)) {
7872 pr_err("Preemption disabled at:");
7873 print_ip_sym(KERN_ERR, preempt_disable_ip);
7876 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7878 EXPORT_SYMBOL(___might_sleep);
7880 void __cant_sleep(const char *file, int line, int preempt_offset)
7882 static unsigned long prev_jiffy;
7884 if (irqs_disabled())
7887 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
7890 if (preempt_count() > preempt_offset)
7893 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7895 prev_jiffy = jiffies;
7897 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
7898 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7899 in_atomic(), irqs_disabled(),
7900 current->pid, current->comm);
7902 debug_show_held_locks(current);
7904 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7906 EXPORT_SYMBOL_GPL(__cant_sleep);
7909 void __cant_migrate(const char *file, int line)
7911 static unsigned long prev_jiffy;
7913 if (irqs_disabled())
7916 if (is_migration_disabled(current))
7919 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
7922 if (preempt_count() > 0)
7925 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7927 prev_jiffy = jiffies;
7929 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
7930 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
7931 in_atomic(), irqs_disabled(), is_migration_disabled(current),
7932 current->pid, current->comm);
7934 debug_show_held_locks(current);
7936 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7938 EXPORT_SYMBOL_GPL(__cant_migrate);
7942 #ifdef CONFIG_MAGIC_SYSRQ
7943 void normalize_rt_tasks(void)
7945 struct task_struct *g, *p;
7946 struct sched_attr attr = {
7947 .sched_policy = SCHED_NORMAL,
7950 read_lock(&tasklist_lock);
7951 for_each_process_thread(g, p) {
7953 * Only normalize user tasks:
7955 if (p->flags & PF_KTHREAD)
7958 p->se.exec_start = 0;
7959 schedstat_set(p->se.statistics.wait_start, 0);
7960 schedstat_set(p->se.statistics.sleep_start, 0);
7961 schedstat_set(p->se.statistics.block_start, 0);
7963 if (!dl_task(p) && !rt_task(p)) {
7965 * Renice negative nice level userspace
7968 if (task_nice(p) < 0)
7969 set_user_nice(p, 0);
7973 __sched_setscheduler(p, &attr, false, false);
7975 read_unlock(&tasklist_lock);
7978 #endif /* CONFIG_MAGIC_SYSRQ */
7980 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7982 * These functions are only useful for the IA64 MCA handling, or kdb.
7984 * They can only be called when the whole system has been
7985 * stopped - every CPU needs to be quiescent, and no scheduling
7986 * activity can take place. Using them for anything else would
7987 * be a serious bug, and as a result, they aren't even visible
7988 * under any other configuration.
7992 * curr_task - return the current task for a given CPU.
7993 * @cpu: the processor in question.
7995 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7997 * Return: The current task for @cpu.
7999 struct task_struct *curr_task(int cpu)
8001 return cpu_curr(cpu);
8004 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8008 * ia64_set_curr_task - set the current task for a given CPU.
8009 * @cpu: the processor in question.
8010 * @p: the task pointer to set.
8012 * Description: This function must only be used when non-maskable interrupts
8013 * are serviced on a separate stack. It allows the architecture to switch the
8014 * notion of the current task on a CPU in a non-blocking manner. This function
8015 * must be called with all CPU's synchronized, and interrupts disabled, the
8016 * and caller must save the original value of the current task (see
8017 * curr_task() above) and restore that value before reenabling interrupts and
8018 * re-starting the system.
8020 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8022 void ia64_set_curr_task(int cpu, struct task_struct *p)
8029 #ifdef CONFIG_CGROUP_SCHED
8030 /* task_group_lock serializes the addition/removal of task groups */
8031 static DEFINE_SPINLOCK(task_group_lock);
8033 static inline void alloc_uclamp_sched_group(struct task_group *tg,
8034 struct task_group *parent)
8036 #ifdef CONFIG_UCLAMP_TASK_GROUP
8037 enum uclamp_id clamp_id;
8039 for_each_clamp_id(clamp_id) {
8040 uclamp_se_set(&tg->uclamp_req[clamp_id],
8041 uclamp_none(clamp_id), false);
8042 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
8047 static void sched_free_group(struct task_group *tg)
8049 free_fair_sched_group(tg);
8050 free_rt_sched_group(tg);
8052 kmem_cache_free(task_group_cache, tg);
8055 /* allocate runqueue etc for a new task group */
8056 struct task_group *sched_create_group(struct task_group *parent)
8058 struct task_group *tg;
8060 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
8062 return ERR_PTR(-ENOMEM);
8064 if (!alloc_fair_sched_group(tg, parent))
8067 if (!alloc_rt_sched_group(tg, parent))
8070 alloc_uclamp_sched_group(tg, parent);
8075 sched_free_group(tg);
8076 return ERR_PTR(-ENOMEM);
8079 void sched_online_group(struct task_group *tg, struct task_group *parent)
8081 unsigned long flags;
8083 spin_lock_irqsave(&task_group_lock, flags);
8084 list_add_rcu(&tg->list, &task_groups);
8086 /* Root should already exist: */
8089 tg->parent = parent;
8090 INIT_LIST_HEAD(&tg->children);
8091 list_add_rcu(&tg->siblings, &parent->children);
8092 spin_unlock_irqrestore(&task_group_lock, flags);
8094 online_fair_sched_group(tg);
8097 /* rcu callback to free various structures associated with a task group */
8098 static void sched_free_group_rcu(struct rcu_head *rhp)
8100 /* Now it should be safe to free those cfs_rqs: */
8101 sched_free_group(container_of(rhp, struct task_group, rcu));
8104 void sched_destroy_group(struct task_group *tg)
8106 /* Wait for possible concurrent references to cfs_rqs complete: */
8107 call_rcu(&tg->rcu, sched_free_group_rcu);
8110 void sched_offline_group(struct task_group *tg)
8112 unsigned long flags;
8114 /* End participation in shares distribution: */
8115 unregister_fair_sched_group(tg);
8117 spin_lock_irqsave(&task_group_lock, flags);
8118 list_del_rcu(&tg->list);
8119 list_del_rcu(&tg->siblings);
8120 spin_unlock_irqrestore(&task_group_lock, flags);
8123 static void sched_change_group(struct task_struct *tsk, int type)
8125 struct task_group *tg;
8128 * All callers are synchronized by task_rq_lock(); we do not use RCU
8129 * which is pointless here. Thus, we pass "true" to task_css_check()
8130 * to prevent lockdep warnings.
8132 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8133 struct task_group, css);
8134 tg = autogroup_task_group(tsk, tg);
8135 tsk->sched_task_group = tg;
8137 #ifdef CONFIG_FAIR_GROUP_SCHED
8138 if (tsk->sched_class->task_change_group)
8139 tsk->sched_class->task_change_group(tsk, type);
8142 set_task_rq(tsk, task_cpu(tsk));
8146 * Change task's runqueue when it moves between groups.
8148 * The caller of this function should have put the task in its new group by
8149 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8152 void sched_move_task(struct task_struct *tsk)
8154 int queued, running, queue_flags =
8155 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
8159 rq = task_rq_lock(tsk, &rf);
8160 update_rq_clock(rq);
8162 running = task_current(rq, tsk);
8163 queued = task_on_rq_queued(tsk);
8166 dequeue_task(rq, tsk, queue_flags);
8168 put_prev_task(rq, tsk);
8170 sched_change_group(tsk, TASK_MOVE_GROUP);
8173 enqueue_task(rq, tsk, queue_flags);
8175 set_next_task(rq, tsk);
8177 * After changing group, the running task may have joined a
8178 * throttled one but it's still the running task. Trigger a
8179 * resched to make sure that task can still run.
8184 task_rq_unlock(rq, tsk, &rf);
8187 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8189 return css ? container_of(css, struct task_group, css) : NULL;
8192 static struct cgroup_subsys_state *
8193 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8195 struct task_group *parent = css_tg(parent_css);
8196 struct task_group *tg;
8199 /* This is early initialization for the top cgroup */
8200 return &root_task_group.css;
8203 tg = sched_create_group(parent);
8205 return ERR_PTR(-ENOMEM);
8210 /* Expose task group only after completing cgroup initialization */
8211 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8213 struct task_group *tg = css_tg(css);
8214 struct task_group *parent = css_tg(css->parent);
8217 sched_online_group(tg, parent);
8219 #ifdef CONFIG_UCLAMP_TASK_GROUP
8220 /* Propagate the effective uclamp value for the new group */
8221 cpu_util_update_eff(css);
8227 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8229 struct task_group *tg = css_tg(css);
8231 sched_offline_group(tg);
8234 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8236 struct task_group *tg = css_tg(css);
8239 * Relies on the RCU grace period between css_released() and this.
8241 sched_free_group(tg);
8245 * This is called before wake_up_new_task(), therefore we really only
8246 * have to set its group bits, all the other stuff does not apply.
8248 static void cpu_cgroup_fork(struct task_struct *task)
8253 rq = task_rq_lock(task, &rf);
8255 update_rq_clock(rq);
8256 sched_change_group(task, TASK_SET_GROUP);
8258 task_rq_unlock(rq, task, &rf);
8261 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8263 struct task_struct *task;
8264 struct cgroup_subsys_state *css;
8267 cgroup_taskset_for_each(task, css, tset) {
8268 #ifdef CONFIG_RT_GROUP_SCHED
8269 if (!sched_rt_can_attach(css_tg(css), task))
8273 * Serialize against wake_up_new_task() such that if it's
8274 * running, we're sure to observe its full state.
8276 raw_spin_lock_irq(&task->pi_lock);
8278 * Avoid calling sched_move_task() before wake_up_new_task()
8279 * has happened. This would lead to problems with PELT, due to
8280 * move wanting to detach+attach while we're not attached yet.
8282 if (task->state == TASK_NEW)
8284 raw_spin_unlock_irq(&task->pi_lock);
8292 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8294 struct task_struct *task;
8295 struct cgroup_subsys_state *css;
8297 cgroup_taskset_for_each(task, css, tset)
8298 sched_move_task(task);
8301 #ifdef CONFIG_UCLAMP_TASK_GROUP
8302 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
8304 struct cgroup_subsys_state *top_css = css;
8305 struct uclamp_se *uc_parent = NULL;
8306 struct uclamp_se *uc_se = NULL;
8307 unsigned int eff[UCLAMP_CNT];
8308 enum uclamp_id clamp_id;
8309 unsigned int clamps;
8311 css_for_each_descendant_pre(css, top_css) {
8312 uc_parent = css_tg(css)->parent
8313 ? css_tg(css)->parent->uclamp : NULL;
8315 for_each_clamp_id(clamp_id) {
8316 /* Assume effective clamps matches requested clamps */
8317 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
8318 /* Cap effective clamps with parent's effective clamps */
8320 eff[clamp_id] > uc_parent[clamp_id].value) {
8321 eff[clamp_id] = uc_parent[clamp_id].value;
8324 /* Ensure protection is always capped by limit */
8325 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
8327 /* Propagate most restrictive effective clamps */
8329 uc_se = css_tg(css)->uclamp;
8330 for_each_clamp_id(clamp_id) {
8331 if (eff[clamp_id] == uc_se[clamp_id].value)
8333 uc_se[clamp_id].value = eff[clamp_id];
8334 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
8335 clamps |= (0x1 << clamp_id);
8338 css = css_rightmost_descendant(css);
8342 /* Immediately update descendants RUNNABLE tasks */
8343 uclamp_update_active_tasks(css, clamps);
8348 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
8349 * C expression. Since there is no way to convert a macro argument (N) into a
8350 * character constant, use two levels of macros.
8352 #define _POW10(exp) ((unsigned int)1e##exp)
8353 #define POW10(exp) _POW10(exp)
8355 struct uclamp_request {
8356 #define UCLAMP_PERCENT_SHIFT 2
8357 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
8363 static inline struct uclamp_request
8364 capacity_from_percent(char *buf)
8366 struct uclamp_request req = {
8367 .percent = UCLAMP_PERCENT_SCALE,
8368 .util = SCHED_CAPACITY_SCALE,
8373 if (strcmp(buf, "max")) {
8374 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
8378 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
8383 req.util = req.percent << SCHED_CAPACITY_SHIFT;
8384 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
8390 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
8391 size_t nbytes, loff_t off,
8392 enum uclamp_id clamp_id)
8394 struct uclamp_request req;
8395 struct task_group *tg;
8397 req = capacity_from_percent(buf);
8401 static_branch_enable(&sched_uclamp_used);
8403 mutex_lock(&uclamp_mutex);
8406 tg = css_tg(of_css(of));
8407 if (tg->uclamp_req[clamp_id].value != req.util)
8408 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
8411 * Because of not recoverable conversion rounding we keep track of the
8412 * exact requested value
8414 tg->uclamp_pct[clamp_id] = req.percent;
8416 /* Update effective clamps to track the most restrictive value */
8417 cpu_util_update_eff(of_css(of));
8420 mutex_unlock(&uclamp_mutex);
8425 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
8426 char *buf, size_t nbytes,
8429 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
8432 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
8433 char *buf, size_t nbytes,
8436 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
8439 static inline void cpu_uclamp_print(struct seq_file *sf,
8440 enum uclamp_id clamp_id)
8442 struct task_group *tg;
8448 tg = css_tg(seq_css(sf));
8449 util_clamp = tg->uclamp_req[clamp_id].value;
8452 if (util_clamp == SCHED_CAPACITY_SCALE) {
8453 seq_puts(sf, "max\n");
8457 percent = tg->uclamp_pct[clamp_id];
8458 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
8459 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
8462 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
8464 cpu_uclamp_print(sf, UCLAMP_MIN);
8468 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
8470 cpu_uclamp_print(sf, UCLAMP_MAX);
8473 #endif /* CONFIG_UCLAMP_TASK_GROUP */
8475 #ifdef CONFIG_FAIR_GROUP_SCHED
8476 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8477 struct cftype *cftype, u64 shareval)
8479 if (shareval > scale_load_down(ULONG_MAX))
8480 shareval = MAX_SHARES;
8481 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8484 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8487 struct task_group *tg = css_tg(css);
8489 return (u64) scale_load_down(tg->shares);
8492 #ifdef CONFIG_CFS_BANDWIDTH
8493 static DEFINE_MUTEX(cfs_constraints_mutex);
8495 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8496 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8497 /* More than 203 days if BW_SHIFT equals 20. */
8498 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
8500 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8502 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8504 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8505 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8507 if (tg == &root_task_group)
8511 * Ensure we have at some amount of bandwidth every period. This is
8512 * to prevent reaching a state of large arrears when throttled via
8513 * entity_tick() resulting in prolonged exit starvation.
8515 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8519 * Likewise, bound things on the otherside by preventing insane quota
8520 * periods. This also allows us to normalize in computing quota
8523 if (period > max_cfs_quota_period)
8527 * Bound quota to defend quota against overflow during bandwidth shift.
8529 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
8533 * Prevent race between setting of cfs_rq->runtime_enabled and
8534 * unthrottle_offline_cfs_rqs().
8537 mutex_lock(&cfs_constraints_mutex);
8538 ret = __cfs_schedulable(tg, period, quota);
8542 runtime_enabled = quota != RUNTIME_INF;
8543 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8545 * If we need to toggle cfs_bandwidth_used, off->on must occur
8546 * before making related changes, and on->off must occur afterwards
8548 if (runtime_enabled && !runtime_was_enabled)
8549 cfs_bandwidth_usage_inc();
8550 raw_spin_lock_irq(&cfs_b->lock);
8551 cfs_b->period = ns_to_ktime(period);
8552 cfs_b->quota = quota;
8554 __refill_cfs_bandwidth_runtime(cfs_b);
8556 /* Restart the period timer (if active) to handle new period expiry: */
8557 if (runtime_enabled)
8558 start_cfs_bandwidth(cfs_b);
8560 raw_spin_unlock_irq(&cfs_b->lock);
8562 for_each_online_cpu(i) {
8563 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8564 struct rq *rq = cfs_rq->rq;
8567 rq_lock_irq(rq, &rf);
8568 cfs_rq->runtime_enabled = runtime_enabled;
8569 cfs_rq->runtime_remaining = 0;
8571 if (cfs_rq->throttled)
8572 unthrottle_cfs_rq(cfs_rq);
8573 rq_unlock_irq(rq, &rf);
8575 if (runtime_was_enabled && !runtime_enabled)
8576 cfs_bandwidth_usage_dec();
8578 mutex_unlock(&cfs_constraints_mutex);
8584 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8588 period = ktime_to_ns(tg->cfs_bandwidth.period);
8589 if (cfs_quota_us < 0)
8590 quota = RUNTIME_INF;
8591 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
8592 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8596 return tg_set_cfs_bandwidth(tg, period, quota);
8599 static long tg_get_cfs_quota(struct task_group *tg)
8603 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8606 quota_us = tg->cfs_bandwidth.quota;
8607 do_div(quota_us, NSEC_PER_USEC);
8612 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8616 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
8619 period = (u64)cfs_period_us * NSEC_PER_USEC;
8620 quota = tg->cfs_bandwidth.quota;
8622 return tg_set_cfs_bandwidth(tg, period, quota);
8625 static long tg_get_cfs_period(struct task_group *tg)
8629 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8630 do_div(cfs_period_us, NSEC_PER_USEC);
8632 return cfs_period_us;
8635 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8638 return tg_get_cfs_quota(css_tg(css));
8641 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8642 struct cftype *cftype, s64 cfs_quota_us)
8644 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8647 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8650 return tg_get_cfs_period(css_tg(css));
8653 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8654 struct cftype *cftype, u64 cfs_period_us)
8656 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8659 struct cfs_schedulable_data {
8660 struct task_group *tg;
8665 * normalize group quota/period to be quota/max_period
8666 * note: units are usecs
8668 static u64 normalize_cfs_quota(struct task_group *tg,
8669 struct cfs_schedulable_data *d)
8677 period = tg_get_cfs_period(tg);
8678 quota = tg_get_cfs_quota(tg);
8681 /* note: these should typically be equivalent */
8682 if (quota == RUNTIME_INF || quota == -1)
8685 return to_ratio(period, quota);
8688 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8690 struct cfs_schedulable_data *d = data;
8691 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8692 s64 quota = 0, parent_quota = -1;
8695 quota = RUNTIME_INF;
8697 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8699 quota = normalize_cfs_quota(tg, d);
8700 parent_quota = parent_b->hierarchical_quota;
8703 * Ensure max(child_quota) <= parent_quota. On cgroup2,
8704 * always take the min. On cgroup1, only inherit when no
8707 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
8708 quota = min(quota, parent_quota);
8710 if (quota == RUNTIME_INF)
8711 quota = parent_quota;
8712 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8716 cfs_b->hierarchical_quota = quota;
8721 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8724 struct cfs_schedulable_data data = {
8730 if (quota != RUNTIME_INF) {
8731 do_div(data.period, NSEC_PER_USEC);
8732 do_div(data.quota, NSEC_PER_USEC);
8736 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8742 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
8744 struct task_group *tg = css_tg(seq_css(sf));
8745 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8747 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8748 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8749 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8751 if (schedstat_enabled() && tg != &root_task_group) {
8755 for_each_possible_cpu(i)
8756 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
8758 seq_printf(sf, "wait_sum %llu\n", ws);
8763 #endif /* CONFIG_CFS_BANDWIDTH */
8764 #endif /* CONFIG_FAIR_GROUP_SCHED */
8766 #ifdef CONFIG_RT_GROUP_SCHED
8767 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8768 struct cftype *cft, s64 val)
8770 return sched_group_set_rt_runtime(css_tg(css), val);
8773 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8776 return sched_group_rt_runtime(css_tg(css));
8779 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8780 struct cftype *cftype, u64 rt_period_us)
8782 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8785 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8788 return sched_group_rt_period(css_tg(css));
8790 #endif /* CONFIG_RT_GROUP_SCHED */
8792 static struct cftype cpu_legacy_files[] = {
8793 #ifdef CONFIG_FAIR_GROUP_SCHED
8796 .read_u64 = cpu_shares_read_u64,
8797 .write_u64 = cpu_shares_write_u64,
8800 #ifdef CONFIG_CFS_BANDWIDTH
8802 .name = "cfs_quota_us",
8803 .read_s64 = cpu_cfs_quota_read_s64,
8804 .write_s64 = cpu_cfs_quota_write_s64,
8807 .name = "cfs_period_us",
8808 .read_u64 = cpu_cfs_period_read_u64,
8809 .write_u64 = cpu_cfs_period_write_u64,
8813 .seq_show = cpu_cfs_stat_show,
8816 #ifdef CONFIG_RT_GROUP_SCHED
8818 .name = "rt_runtime_us",
8819 .read_s64 = cpu_rt_runtime_read,
8820 .write_s64 = cpu_rt_runtime_write,
8823 .name = "rt_period_us",
8824 .read_u64 = cpu_rt_period_read_uint,
8825 .write_u64 = cpu_rt_period_write_uint,
8828 #ifdef CONFIG_UCLAMP_TASK_GROUP
8830 .name = "uclamp.min",
8831 .flags = CFTYPE_NOT_ON_ROOT,
8832 .seq_show = cpu_uclamp_min_show,
8833 .write = cpu_uclamp_min_write,
8836 .name = "uclamp.max",
8837 .flags = CFTYPE_NOT_ON_ROOT,
8838 .seq_show = cpu_uclamp_max_show,
8839 .write = cpu_uclamp_max_write,
8845 static int cpu_extra_stat_show(struct seq_file *sf,
8846 struct cgroup_subsys_state *css)
8848 #ifdef CONFIG_CFS_BANDWIDTH
8850 struct task_group *tg = css_tg(css);
8851 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8854 throttled_usec = cfs_b->throttled_time;
8855 do_div(throttled_usec, NSEC_PER_USEC);
8857 seq_printf(sf, "nr_periods %d\n"
8859 "throttled_usec %llu\n",
8860 cfs_b->nr_periods, cfs_b->nr_throttled,
8867 #ifdef CONFIG_FAIR_GROUP_SCHED
8868 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
8871 struct task_group *tg = css_tg(css);
8872 u64 weight = scale_load_down(tg->shares);
8874 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
8877 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
8878 struct cftype *cft, u64 weight)
8881 * cgroup weight knobs should use the common MIN, DFL and MAX
8882 * values which are 1, 100 and 10000 respectively. While it loses
8883 * a bit of range on both ends, it maps pretty well onto the shares
8884 * value used by scheduler and the round-trip conversions preserve
8885 * the original value over the entire range.
8887 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
8890 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
8892 return sched_group_set_shares(css_tg(css), scale_load(weight));
8895 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
8898 unsigned long weight = scale_load_down(css_tg(css)->shares);
8899 int last_delta = INT_MAX;
8902 /* find the closest nice value to the current weight */
8903 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
8904 delta = abs(sched_prio_to_weight[prio] - weight);
8905 if (delta >= last_delta)
8910 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
8913 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
8914 struct cftype *cft, s64 nice)
8916 unsigned long weight;
8919 if (nice < MIN_NICE || nice > MAX_NICE)
8922 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
8923 idx = array_index_nospec(idx, 40);
8924 weight = sched_prio_to_weight[idx];
8926 return sched_group_set_shares(css_tg(css), scale_load(weight));
8930 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
8931 long period, long quota)
8934 seq_puts(sf, "max");
8936 seq_printf(sf, "%ld", quota);
8938 seq_printf(sf, " %ld\n", period);
8941 /* caller should put the current value in *@periodp before calling */
8942 static int __maybe_unused cpu_period_quota_parse(char *buf,
8943 u64 *periodp, u64 *quotap)
8945 char tok[21]; /* U64_MAX */
8947 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
8950 *periodp *= NSEC_PER_USEC;
8952 if (sscanf(tok, "%llu", quotap))
8953 *quotap *= NSEC_PER_USEC;
8954 else if (!strcmp(tok, "max"))
8955 *quotap = RUNTIME_INF;
8962 #ifdef CONFIG_CFS_BANDWIDTH
8963 static int cpu_max_show(struct seq_file *sf, void *v)
8965 struct task_group *tg = css_tg(seq_css(sf));
8967 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
8971 static ssize_t cpu_max_write(struct kernfs_open_file *of,
8972 char *buf, size_t nbytes, loff_t off)
8974 struct task_group *tg = css_tg(of_css(of));
8975 u64 period = tg_get_cfs_period(tg);
8979 ret = cpu_period_quota_parse(buf, &period, "a);
8981 ret = tg_set_cfs_bandwidth(tg, period, quota);
8982 return ret ?: nbytes;
8986 static struct cftype cpu_files[] = {
8987 #ifdef CONFIG_FAIR_GROUP_SCHED
8990 .flags = CFTYPE_NOT_ON_ROOT,
8991 .read_u64 = cpu_weight_read_u64,
8992 .write_u64 = cpu_weight_write_u64,
8995 .name = "weight.nice",
8996 .flags = CFTYPE_NOT_ON_ROOT,
8997 .read_s64 = cpu_weight_nice_read_s64,
8998 .write_s64 = cpu_weight_nice_write_s64,
9001 #ifdef CONFIG_CFS_BANDWIDTH
9004 .flags = CFTYPE_NOT_ON_ROOT,
9005 .seq_show = cpu_max_show,
9006 .write = cpu_max_write,
9009 #ifdef CONFIG_UCLAMP_TASK_GROUP
9011 .name = "uclamp.min",
9012 .flags = CFTYPE_NOT_ON_ROOT,
9013 .seq_show = cpu_uclamp_min_show,
9014 .write = cpu_uclamp_min_write,
9017 .name = "uclamp.max",
9018 .flags = CFTYPE_NOT_ON_ROOT,
9019 .seq_show = cpu_uclamp_max_show,
9020 .write = cpu_uclamp_max_write,
9026 struct cgroup_subsys cpu_cgrp_subsys = {
9027 .css_alloc = cpu_cgroup_css_alloc,
9028 .css_online = cpu_cgroup_css_online,
9029 .css_released = cpu_cgroup_css_released,
9030 .css_free = cpu_cgroup_css_free,
9031 .css_extra_stat_show = cpu_extra_stat_show,
9032 .fork = cpu_cgroup_fork,
9033 .can_attach = cpu_cgroup_can_attach,
9034 .attach = cpu_cgroup_attach,
9035 .legacy_cftypes = cpu_legacy_files,
9036 .dfl_cftypes = cpu_files,
9041 #endif /* CONFIG_CGROUP_SCHED */
9043 void dump_cpu_task(int cpu)
9045 pr_info("Task dump for CPU %d:\n", cpu);
9046 sched_show_task(cpu_curr(cpu));
9050 * Nice levels are multiplicative, with a gentle 10% change for every
9051 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
9052 * nice 1, it will get ~10% less CPU time than another CPU-bound task
9053 * that remained on nice 0.
9055 * The "10% effect" is relative and cumulative: from _any_ nice level,
9056 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
9057 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
9058 * If a task goes up by ~10% and another task goes down by ~10% then
9059 * the relative distance between them is ~25%.)
9061 const int sched_prio_to_weight[40] = {
9062 /* -20 */ 88761, 71755, 56483, 46273, 36291,
9063 /* -15 */ 29154, 23254, 18705, 14949, 11916,
9064 /* -10 */ 9548, 7620, 6100, 4904, 3906,
9065 /* -5 */ 3121, 2501, 1991, 1586, 1277,
9066 /* 0 */ 1024, 820, 655, 526, 423,
9067 /* 5 */ 335, 272, 215, 172, 137,
9068 /* 10 */ 110, 87, 70, 56, 45,
9069 /* 15 */ 36, 29, 23, 18, 15,
9073 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
9075 * In cases where the weight does not change often, we can use the
9076 * precalculated inverse to speed up arithmetics by turning divisions
9077 * into multiplications:
9079 const u32 sched_prio_to_wmult[40] = {
9080 /* -20 */ 48388, 59856, 76040, 92818, 118348,
9081 /* -15 */ 147320, 184698, 229616, 287308, 360437,
9082 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
9083 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
9084 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
9085 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
9086 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
9087 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
9090 void call_trace_sched_update_nr_running(struct rq *rq, int count)
9092 trace_sched_update_nr_running_tp(rq, count);