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 elegible 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);
324 rq_csd_init(struct rq *rq, call_single_data_t *csd, smp_call_func_t func)
331 #ifdef CONFIG_SCHED_HRTICK
333 * Use HR-timers to deliver accurate preemption points.
336 static void hrtick_clear(struct rq *rq)
338 if (hrtimer_active(&rq->hrtick_timer))
339 hrtimer_cancel(&rq->hrtick_timer);
343 * High-resolution timer tick.
344 * Runs from hardirq context with interrupts disabled.
346 static enum hrtimer_restart hrtick(struct hrtimer *timer)
348 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
351 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
355 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
358 return HRTIMER_NORESTART;
363 static void __hrtick_restart(struct rq *rq)
365 struct hrtimer *timer = &rq->hrtick_timer;
367 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
371 * called from hardirq (IPI) context
373 static void __hrtick_start(void *arg)
379 __hrtick_restart(rq);
384 * Called to set the hrtick timer state.
386 * called with rq->lock held and irqs disabled
388 void hrtick_start(struct rq *rq, u64 delay)
390 struct hrtimer *timer = &rq->hrtick_timer;
395 * Don't schedule slices shorter than 10000ns, that just
396 * doesn't make sense and can cause timer DoS.
398 delta = max_t(s64, delay, 10000LL);
399 time = ktime_add_ns(timer->base->get_time(), delta);
401 hrtimer_set_expires(timer, time);
404 __hrtick_restart(rq);
406 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq *rq, u64 delay)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay = max_t(u64, delay, 10000LL);
422 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
423 HRTIMER_MODE_REL_PINNED_HARD);
426 #endif /* CONFIG_SMP */
428 static void hrtick_rq_init(struct rq *rq)
431 rq_csd_init(rq, &rq->hrtick_csd, __hrtick_start);
433 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
434 rq->hrtick_timer.function = hrtick;
436 #else /* CONFIG_SCHED_HRTICK */
437 static inline void hrtick_clear(struct rq *rq)
441 static inline void hrtick_rq_init(struct rq *rq)
444 #endif /* CONFIG_SCHED_HRTICK */
447 * cmpxchg based fetch_or, macro so it works for different integer types
449 #define fetch_or(ptr, mask) \
451 typeof(ptr) _ptr = (ptr); \
452 typeof(mask) _mask = (mask); \
453 typeof(*_ptr) _old, _val = *_ptr; \
456 _old = cmpxchg(_ptr, _val, _val | _mask); \
464 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
466 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
467 * this avoids any races wrt polling state changes and thereby avoids
470 static bool set_nr_and_not_polling(struct task_struct *p)
472 struct thread_info *ti = task_thread_info(p);
473 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
477 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
479 * If this returns true, then the idle task promises to call
480 * sched_ttwu_pending() and reschedule soon.
482 static bool set_nr_if_polling(struct task_struct *p)
484 struct thread_info *ti = task_thread_info(p);
485 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
488 if (!(val & _TIF_POLLING_NRFLAG))
490 if (val & _TIF_NEED_RESCHED)
492 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
501 static bool set_nr_and_not_polling(struct task_struct *p)
503 set_tsk_need_resched(p);
508 static bool set_nr_if_polling(struct task_struct *p)
515 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
517 struct wake_q_node *node = &task->wake_q;
520 * Atomically grab the task, if ->wake_q is !nil already it means
521 * its already queued (either by us or someone else) and will get the
522 * wakeup due to that.
524 * In order to ensure that a pending wakeup will observe our pending
525 * state, even in the failed case, an explicit smp_mb() must be used.
527 smp_mb__before_atomic();
528 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
532 * The head is context local, there can be no concurrency.
535 head->lastp = &node->next;
540 * wake_q_add() - queue a wakeup for 'later' waking.
541 * @head: the wake_q_head to add @task to
542 * @task: the task to queue for 'later' wakeup
544 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
545 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
548 * This function must be used as-if it were wake_up_process(); IOW the task
549 * must be ready to be woken at this location.
551 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
553 if (__wake_q_add(head, task))
554 get_task_struct(task);
558 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
559 * @head: the wake_q_head to add @task to
560 * @task: the task to queue for 'later' wakeup
562 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
563 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
566 * This function must be used as-if it were wake_up_process(); IOW the task
567 * must be ready to be woken at this location.
569 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
570 * that already hold reference to @task can call the 'safe' version and trust
571 * wake_q to do the right thing depending whether or not the @task is already
574 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
576 if (!__wake_q_add(head, task))
577 put_task_struct(task);
580 void wake_up_q(struct wake_q_head *head)
582 struct wake_q_node *node = head->first;
584 while (node != WAKE_Q_TAIL) {
585 struct task_struct *task;
587 task = container_of(node, struct task_struct, wake_q);
589 /* Task can safely be re-inserted now: */
591 task->wake_q.next = NULL;
594 * wake_up_process() executes a full barrier, which pairs with
595 * the queueing in wake_q_add() so as not to miss wakeups.
597 wake_up_process(task);
598 put_task_struct(task);
603 * resched_curr - mark rq's current task 'to be rescheduled now'.
605 * On UP this means the setting of the need_resched flag, on SMP it
606 * might also involve a cross-CPU call to trigger the scheduler on
609 void resched_curr(struct rq *rq)
611 struct task_struct *curr = rq->curr;
614 lockdep_assert_held(&rq->lock);
616 if (test_tsk_need_resched(curr))
621 if (cpu == smp_processor_id()) {
622 set_tsk_need_resched(curr);
623 set_preempt_need_resched();
627 if (set_nr_and_not_polling(curr))
628 smp_send_reschedule(cpu);
630 trace_sched_wake_idle_without_ipi(cpu);
633 void resched_cpu(int cpu)
635 struct rq *rq = cpu_rq(cpu);
638 raw_spin_lock_irqsave(&rq->lock, flags);
639 if (cpu_online(cpu) || cpu == smp_processor_id())
641 raw_spin_unlock_irqrestore(&rq->lock, flags);
645 #ifdef CONFIG_NO_HZ_COMMON
647 * In the semi idle case, use the nearest busy CPU for migrating timers
648 * from an idle CPU. This is good for power-savings.
650 * We don't do similar optimization for completely idle system, as
651 * selecting an idle CPU will add more delays to the timers than intended
652 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
654 int get_nohz_timer_target(void)
656 int i, cpu = smp_processor_id(), default_cpu = -1;
657 struct sched_domain *sd;
659 if (housekeeping_cpu(cpu, HK_FLAG_TIMER)) {
666 for_each_domain(cpu, sd) {
667 for_each_cpu_and(i, sched_domain_span(sd),
668 housekeeping_cpumask(HK_FLAG_TIMER)) {
679 if (default_cpu == -1)
680 default_cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
688 * When add_timer_on() enqueues a timer into the timer wheel of an
689 * idle CPU then this timer might expire before the next timer event
690 * which is scheduled to wake up that CPU. In case of a completely
691 * idle system the next event might even be infinite time into the
692 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
693 * leaves the inner idle loop so the newly added timer is taken into
694 * account when the CPU goes back to idle and evaluates the timer
695 * wheel for the next timer event.
697 static void wake_up_idle_cpu(int cpu)
699 struct rq *rq = cpu_rq(cpu);
701 if (cpu == smp_processor_id())
704 if (set_nr_and_not_polling(rq->idle))
705 smp_send_reschedule(cpu);
707 trace_sched_wake_idle_without_ipi(cpu);
710 static bool wake_up_full_nohz_cpu(int cpu)
713 * We just need the target to call irq_exit() and re-evaluate
714 * the next tick. The nohz full kick at least implies that.
715 * If needed we can still optimize that later with an
718 if (cpu_is_offline(cpu))
719 return true; /* Don't try to wake offline CPUs. */
720 if (tick_nohz_full_cpu(cpu)) {
721 if (cpu != smp_processor_id() ||
722 tick_nohz_tick_stopped())
723 tick_nohz_full_kick_cpu(cpu);
731 * Wake up the specified CPU. If the CPU is going offline, it is the
732 * caller's responsibility to deal with the lost wakeup, for example,
733 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
735 void wake_up_nohz_cpu(int cpu)
737 if (!wake_up_full_nohz_cpu(cpu))
738 wake_up_idle_cpu(cpu);
741 static void nohz_csd_func(void *info)
743 struct rq *rq = info;
744 int cpu = cpu_of(rq);
748 * Release the rq::nohz_csd.
750 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
751 WARN_ON(!(flags & NOHZ_KICK_MASK));
753 rq->idle_balance = idle_cpu(cpu);
754 if (rq->idle_balance && !need_resched()) {
755 rq->nohz_idle_balance = flags;
756 raise_softirq_irqoff(SCHED_SOFTIRQ);
760 #endif /* CONFIG_NO_HZ_COMMON */
762 #ifdef CONFIG_NO_HZ_FULL
763 bool sched_can_stop_tick(struct rq *rq)
767 /* Deadline tasks, even if single, need the tick */
768 if (rq->dl.dl_nr_running)
772 * If there are more than one RR tasks, we need the tick to effect the
773 * actual RR behaviour.
775 if (rq->rt.rr_nr_running) {
776 if (rq->rt.rr_nr_running == 1)
783 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
784 * forced preemption between FIFO tasks.
786 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
791 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
792 * if there's more than one we need the tick for involuntary
795 if (rq->nr_running > 1)
800 #endif /* CONFIG_NO_HZ_FULL */
801 #endif /* CONFIG_SMP */
803 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
804 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
806 * Iterate task_group tree rooted at *from, calling @down when first entering a
807 * node and @up when leaving it for the final time.
809 * Caller must hold rcu_lock or sufficient equivalent.
811 int walk_tg_tree_from(struct task_group *from,
812 tg_visitor down, tg_visitor up, void *data)
814 struct task_group *parent, *child;
820 ret = (*down)(parent, data);
823 list_for_each_entry_rcu(child, &parent->children, siblings) {
830 ret = (*up)(parent, data);
831 if (ret || parent == from)
835 parent = parent->parent;
842 int tg_nop(struct task_group *tg, void *data)
848 static void set_load_weight(struct task_struct *p, bool update_load)
850 int prio = p->static_prio - MAX_RT_PRIO;
851 struct load_weight *load = &p->se.load;
854 * SCHED_IDLE tasks get minimal weight:
856 if (task_has_idle_policy(p)) {
857 load->weight = scale_load(WEIGHT_IDLEPRIO);
858 load->inv_weight = WMULT_IDLEPRIO;
863 * SCHED_OTHER tasks have to update their load when changing their
866 if (update_load && p->sched_class == &fair_sched_class) {
867 reweight_task(p, prio);
869 load->weight = scale_load(sched_prio_to_weight[prio]);
870 load->inv_weight = sched_prio_to_wmult[prio];
874 #ifdef CONFIG_UCLAMP_TASK
876 * Serializes updates of utilization clamp values
878 * The (slow-path) user-space triggers utilization clamp value updates which
879 * can require updates on (fast-path) scheduler's data structures used to
880 * support enqueue/dequeue operations.
881 * While the per-CPU rq lock protects fast-path update operations, user-space
882 * requests are serialized using a mutex to reduce the risk of conflicting
883 * updates or API abuses.
885 static DEFINE_MUTEX(uclamp_mutex);
887 /* Max allowed minimum utilization */
888 unsigned int sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
890 /* Max allowed maximum utilization */
891 unsigned int sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
894 * By default RT tasks run at the maximum performance point/capacity of the
895 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
896 * SCHED_CAPACITY_SCALE.
898 * This knob allows admins to change the default behavior when uclamp is being
899 * used. In battery powered devices, particularly, running at the maximum
900 * capacity and frequency will increase energy consumption and shorten the
903 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
905 * This knob will not override the system default sched_util_clamp_min defined
908 unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
910 /* All clamps are required to be less or equal than these values */
911 static struct uclamp_se uclamp_default[UCLAMP_CNT];
914 * This static key is used to reduce the uclamp overhead in the fast path. It
915 * primarily disables the call to uclamp_rq_{inc, dec}() in
916 * enqueue/dequeue_task().
918 * This allows users to continue to enable uclamp in their kernel config with
919 * minimum uclamp overhead in the fast path.
921 * As soon as userspace modifies any of the uclamp knobs, the static key is
922 * enabled, since we have an actual users that make use of uclamp
925 * The knobs that would enable this static key are:
927 * * A task modifying its uclamp value with sched_setattr().
928 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
929 * * An admin modifying the cgroup cpu.uclamp.{min, max}
931 DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
933 /* Integer rounded range for each bucket */
934 #define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
936 #define for_each_clamp_id(clamp_id) \
937 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
939 static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
941 return clamp_value / UCLAMP_BUCKET_DELTA;
944 static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
946 if (clamp_id == UCLAMP_MIN)
948 return SCHED_CAPACITY_SCALE;
951 static inline void uclamp_se_set(struct uclamp_se *uc_se,
952 unsigned int value, bool user_defined)
954 uc_se->value = value;
955 uc_se->bucket_id = uclamp_bucket_id(value);
956 uc_se->user_defined = user_defined;
959 static inline unsigned int
960 uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
961 unsigned int clamp_value)
964 * Avoid blocked utilization pushing up the frequency when we go
965 * idle (which drops the max-clamp) by retaining the last known
968 if (clamp_id == UCLAMP_MAX) {
969 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
973 return uclamp_none(UCLAMP_MIN);
976 static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
977 unsigned int clamp_value)
979 /* Reset max-clamp retention only on idle exit */
980 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
983 WRITE_ONCE(rq->uclamp[clamp_id].value, clamp_value);
987 unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
988 unsigned int clamp_value)
990 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
991 int bucket_id = UCLAMP_BUCKETS - 1;
994 * Since both min and max clamps are max aggregated, find the
995 * top most bucket with tasks in.
997 for ( ; bucket_id >= 0; bucket_id--) {
998 if (!bucket[bucket_id].tasks)
1000 return bucket[bucket_id].value;
1003 /* No tasks -- default clamp values */
1004 return uclamp_idle_value(rq, clamp_id, clamp_value);
1007 static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1009 unsigned int default_util_min;
1010 struct uclamp_se *uc_se;
1012 lockdep_assert_held(&p->pi_lock);
1014 uc_se = &p->uclamp_req[UCLAMP_MIN];
1016 /* Only sync if user didn't override the default */
1017 if (uc_se->user_defined)
1020 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1021 uclamp_se_set(uc_se, default_util_min, false);
1024 static void uclamp_update_util_min_rt_default(struct task_struct *p)
1032 /* Protect updates to p->uclamp_* */
1033 rq = task_rq_lock(p, &rf);
1034 __uclamp_update_util_min_rt_default(p);
1035 task_rq_unlock(rq, p, &rf);
1038 static void uclamp_sync_util_min_rt_default(void)
1040 struct task_struct *g, *p;
1043 * copy_process() sysctl_uclamp
1044 * uclamp_min_rt = X;
1045 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1046 * // link thread smp_mb__after_spinlock()
1047 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1048 * sched_post_fork() for_each_process_thread()
1049 * __uclamp_sync_rt() __uclamp_sync_rt()
1051 * Ensures that either sched_post_fork() will observe the new
1052 * uclamp_min_rt or for_each_process_thread() will observe the new
1055 read_lock(&tasklist_lock);
1056 smp_mb__after_spinlock();
1057 read_unlock(&tasklist_lock);
1060 for_each_process_thread(g, p)
1061 uclamp_update_util_min_rt_default(p);
1065 static inline struct uclamp_se
1066 uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
1068 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1069 #ifdef CONFIG_UCLAMP_TASK_GROUP
1070 struct uclamp_se uc_max;
1073 * Tasks in autogroups or root task group will be
1074 * restricted by system defaults.
1076 if (task_group_is_autogroup(task_group(p)))
1078 if (task_group(p) == &root_task_group)
1081 uc_max = task_group(p)->uclamp[clamp_id];
1082 if (uc_req.value > uc_max.value || !uc_req.user_defined)
1090 * The effective clamp bucket index of a task depends on, by increasing
1092 * - the task specific clamp value, when explicitly requested from userspace
1093 * - the task group effective clamp value, for tasks not either in the root
1094 * group or in an autogroup
1095 * - the system default clamp value, defined by the sysadmin
1097 static inline struct uclamp_se
1098 uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
1100 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
1101 struct uclamp_se uc_max = uclamp_default[clamp_id];
1103 /* System default restrictions always apply */
1104 if (unlikely(uc_req.value > uc_max.value))
1110 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
1112 struct uclamp_se uc_eff;
1114 /* Task currently refcounted: use back-annotated (effective) value */
1115 if (p->uclamp[clamp_id].active)
1116 return (unsigned long)p->uclamp[clamp_id].value;
1118 uc_eff = uclamp_eff_get(p, clamp_id);
1120 return (unsigned long)uc_eff.value;
1124 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1125 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1126 * updates the rq's clamp value if required.
1128 * Tasks can have a task-specific value requested from user-space, track
1129 * within each bucket the maximum value for tasks refcounted in it.
1130 * This "local max aggregation" allows to track the exact "requested" value
1131 * for each bucket when all its RUNNABLE tasks require the same clamp.
1133 static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
1134 enum uclamp_id clamp_id)
1136 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1137 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1138 struct uclamp_bucket *bucket;
1140 lockdep_assert_held(&rq->lock);
1142 /* Update task effective clamp */
1143 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1145 bucket = &uc_rq->bucket[uc_se->bucket_id];
1147 uc_se->active = true;
1149 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1152 * Local max aggregation: rq buckets always track the max
1153 * "requested" clamp value of its RUNNABLE tasks.
1155 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1156 bucket->value = uc_se->value;
1158 if (uc_se->value > READ_ONCE(uc_rq->value))
1159 WRITE_ONCE(uc_rq->value, uc_se->value);
1163 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1164 * is released. If this is the last task reference counting the rq's max
1165 * active clamp value, then the rq's clamp value is updated.
1167 * Both refcounted tasks and rq's cached clamp values are expected to be
1168 * always valid. If it's detected they are not, as defensive programming,
1169 * enforce the expected state and warn.
1171 static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
1172 enum uclamp_id clamp_id)
1174 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1175 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1176 struct uclamp_bucket *bucket;
1177 unsigned int bkt_clamp;
1178 unsigned int rq_clamp;
1180 lockdep_assert_held(&rq->lock);
1183 * If sched_uclamp_used was enabled after task @p was enqueued,
1184 * we could end up with unbalanced call to uclamp_rq_dec_id().
1186 * In this case the uc_se->active flag should be false since no uclamp
1187 * accounting was performed at enqueue time and we can just return
1190 * Need to be careful of the following enqeueue/dequeue ordering
1194 * // sched_uclamp_used gets enabled
1197 * // Must not decrement bukcet->tasks here
1200 * where we could end up with stale data in uc_se and
1201 * bucket[uc_se->bucket_id].
1203 * The following check here eliminates the possibility of such race.
1205 if (unlikely(!uc_se->active))
1208 bucket = &uc_rq->bucket[uc_se->bucket_id];
1210 SCHED_WARN_ON(!bucket->tasks);
1211 if (likely(bucket->tasks))
1214 uc_se->active = false;
1217 * Keep "local max aggregation" simple and accept to (possibly)
1218 * overboost some RUNNABLE tasks in the same bucket.
1219 * The rq clamp bucket value is reset to its base value whenever
1220 * there are no more RUNNABLE tasks refcounting it.
1222 if (likely(bucket->tasks))
1225 rq_clamp = READ_ONCE(uc_rq->value);
1227 * Defensive programming: this should never happen. If it happens,
1228 * e.g. due to future modification, warn and fixup the expected value.
1230 SCHED_WARN_ON(bucket->value > rq_clamp);
1231 if (bucket->value >= rq_clamp) {
1232 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
1233 WRITE_ONCE(uc_rq->value, bkt_clamp);
1237 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1239 enum uclamp_id clamp_id;
1242 * Avoid any overhead until uclamp is actually used by the userspace.
1244 * The condition is constructed such that a NOP is generated when
1245 * sched_uclamp_used is disabled.
1247 if (!static_branch_unlikely(&sched_uclamp_used))
1250 if (unlikely(!p->sched_class->uclamp_enabled))
1253 for_each_clamp_id(clamp_id)
1254 uclamp_rq_inc_id(rq, p, clamp_id);
1256 /* Reset clamp idle holding when there is one RUNNABLE task */
1257 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1258 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1261 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1263 enum uclamp_id clamp_id;
1266 * Avoid any overhead until uclamp is actually used by the userspace.
1268 * The condition is constructed such that a NOP is generated when
1269 * sched_uclamp_used is disabled.
1271 if (!static_branch_unlikely(&sched_uclamp_used))
1274 if (unlikely(!p->sched_class->uclamp_enabled))
1277 for_each_clamp_id(clamp_id)
1278 uclamp_rq_dec_id(rq, p, clamp_id);
1282 uclamp_update_active(struct task_struct *p, enum uclamp_id clamp_id)
1288 * Lock the task and the rq where the task is (or was) queued.
1290 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1291 * price to pay to safely serialize util_{min,max} updates with
1292 * enqueues, dequeues and migration operations.
1293 * This is the same locking schema used by __set_cpus_allowed_ptr().
1295 rq = task_rq_lock(p, &rf);
1298 * Setting the clamp bucket is serialized by task_rq_lock().
1299 * If the task is not yet RUNNABLE and its task_struct is not
1300 * affecting a valid clamp bucket, the next time it's enqueued,
1301 * it will already see the updated clamp bucket value.
1303 if (p->uclamp[clamp_id].active) {
1304 uclamp_rq_dec_id(rq, p, clamp_id);
1305 uclamp_rq_inc_id(rq, p, clamp_id);
1308 task_rq_unlock(rq, p, &rf);
1311 #ifdef CONFIG_UCLAMP_TASK_GROUP
1313 uclamp_update_active_tasks(struct cgroup_subsys_state *css,
1314 unsigned int clamps)
1316 enum uclamp_id clamp_id;
1317 struct css_task_iter it;
1318 struct task_struct *p;
1320 css_task_iter_start(css, 0, &it);
1321 while ((p = css_task_iter_next(&it))) {
1322 for_each_clamp_id(clamp_id) {
1323 if ((0x1 << clamp_id) & clamps)
1324 uclamp_update_active(p, clamp_id);
1327 css_task_iter_end(&it);
1330 static void cpu_util_update_eff(struct cgroup_subsys_state *css);
1331 static void uclamp_update_root_tg(void)
1333 struct task_group *tg = &root_task_group;
1335 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1336 sysctl_sched_uclamp_util_min, false);
1337 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1338 sysctl_sched_uclamp_util_max, false);
1341 cpu_util_update_eff(&root_task_group.css);
1345 static void uclamp_update_root_tg(void) { }
1348 int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
1349 void *buffer, size_t *lenp, loff_t *ppos)
1351 bool update_root_tg = false;
1352 int old_min, old_max, old_min_rt;
1355 mutex_lock(&uclamp_mutex);
1356 old_min = sysctl_sched_uclamp_util_min;
1357 old_max = sysctl_sched_uclamp_util_max;
1358 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
1360 result = proc_dointvec(table, write, buffer, lenp, ppos);
1366 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
1367 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1368 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1374 if (old_min != sysctl_sched_uclamp_util_min) {
1375 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
1376 sysctl_sched_uclamp_util_min, false);
1377 update_root_tg = true;
1379 if (old_max != sysctl_sched_uclamp_util_max) {
1380 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
1381 sysctl_sched_uclamp_util_max, false);
1382 update_root_tg = true;
1385 if (update_root_tg) {
1386 static_branch_enable(&sched_uclamp_used);
1387 uclamp_update_root_tg();
1390 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1391 static_branch_enable(&sched_uclamp_used);
1392 uclamp_sync_util_min_rt_default();
1396 * We update all RUNNABLE tasks only when task groups are in use.
1397 * Otherwise, keep it simple and do just a lazy update at each next
1398 * task enqueue time.
1404 sysctl_sched_uclamp_util_min = old_min;
1405 sysctl_sched_uclamp_util_max = old_max;
1406 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
1408 mutex_unlock(&uclamp_mutex);
1413 static int uclamp_validate(struct task_struct *p,
1414 const struct sched_attr *attr)
1416 unsigned int lower_bound = p->uclamp_req[UCLAMP_MIN].value;
1417 unsigned int upper_bound = p->uclamp_req[UCLAMP_MAX].value;
1419 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN)
1420 lower_bound = attr->sched_util_min;
1421 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX)
1422 upper_bound = attr->sched_util_max;
1424 if (lower_bound > upper_bound)
1426 if (upper_bound > SCHED_CAPACITY_SCALE)
1430 * We have valid uclamp attributes; make sure uclamp is enabled.
1432 * We need to do that here, because enabling static branches is a
1433 * blocking operation which obviously cannot be done while holding
1436 static_branch_enable(&sched_uclamp_used);
1441 static void __setscheduler_uclamp(struct task_struct *p,
1442 const struct sched_attr *attr)
1444 enum uclamp_id clamp_id;
1447 * On scheduling class change, reset to default clamps for tasks
1448 * without a task-specific value.
1450 for_each_clamp_id(clamp_id) {
1451 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
1453 /* Keep using defined clamps across class changes */
1454 if (uc_se->user_defined)
1458 * RT by default have a 100% boost value that could be modified
1461 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
1462 __uclamp_update_util_min_rt_default(p);
1464 uclamp_se_set(uc_se, uclamp_none(clamp_id), false);
1468 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1471 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1472 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1473 attr->sched_util_min, true);
1476 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1477 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1478 attr->sched_util_max, true);
1482 static void uclamp_fork(struct task_struct *p)
1484 enum uclamp_id clamp_id;
1487 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1488 * as the task is still at its early fork stages.
1490 for_each_clamp_id(clamp_id)
1491 p->uclamp[clamp_id].active = false;
1493 if (likely(!p->sched_reset_on_fork))
1496 for_each_clamp_id(clamp_id) {
1497 uclamp_se_set(&p->uclamp_req[clamp_id],
1498 uclamp_none(clamp_id), false);
1502 static void uclamp_post_fork(struct task_struct *p)
1504 uclamp_update_util_min_rt_default(p);
1507 static void __init init_uclamp_rq(struct rq *rq)
1509 enum uclamp_id clamp_id;
1510 struct uclamp_rq *uc_rq = rq->uclamp;
1512 for_each_clamp_id(clamp_id) {
1513 uc_rq[clamp_id] = (struct uclamp_rq) {
1514 .value = uclamp_none(clamp_id)
1518 rq->uclamp_flags = 0;
1521 static void __init init_uclamp(void)
1523 struct uclamp_se uc_max = {};
1524 enum uclamp_id clamp_id;
1527 for_each_possible_cpu(cpu)
1528 init_uclamp_rq(cpu_rq(cpu));
1530 for_each_clamp_id(clamp_id) {
1531 uclamp_se_set(&init_task.uclamp_req[clamp_id],
1532 uclamp_none(clamp_id), false);
1535 /* System defaults allow max clamp values for both indexes */
1536 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
1537 for_each_clamp_id(clamp_id) {
1538 uclamp_default[clamp_id] = uc_max;
1539 #ifdef CONFIG_UCLAMP_TASK_GROUP
1540 root_task_group.uclamp_req[clamp_id] = uc_max;
1541 root_task_group.uclamp[clamp_id] = uc_max;
1546 #else /* CONFIG_UCLAMP_TASK */
1547 static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
1548 static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
1549 static inline int uclamp_validate(struct task_struct *p,
1550 const struct sched_attr *attr)
1554 static void __setscheduler_uclamp(struct task_struct *p,
1555 const struct sched_attr *attr) { }
1556 static inline void uclamp_fork(struct task_struct *p) { }
1557 static inline void uclamp_post_fork(struct task_struct *p) { }
1558 static inline void init_uclamp(void) { }
1559 #endif /* CONFIG_UCLAMP_TASK */
1561 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1563 if (!(flags & ENQUEUE_NOCLOCK))
1564 update_rq_clock(rq);
1566 if (!(flags & ENQUEUE_RESTORE)) {
1567 sched_info_queued(rq, p);
1568 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
1571 uclamp_rq_inc(rq, p);
1572 p->sched_class->enqueue_task(rq, p, flags);
1575 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1577 if (!(flags & DEQUEUE_NOCLOCK))
1578 update_rq_clock(rq);
1580 if (!(flags & DEQUEUE_SAVE)) {
1581 sched_info_dequeued(rq, p);
1582 psi_dequeue(p, flags & DEQUEUE_SLEEP);
1585 uclamp_rq_dec(rq, p);
1586 p->sched_class->dequeue_task(rq, p, flags);
1589 void activate_task(struct rq *rq, struct task_struct *p, int flags)
1591 enqueue_task(rq, p, flags);
1593 p->on_rq = TASK_ON_RQ_QUEUED;
1596 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1598 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
1600 dequeue_task(rq, p, flags);
1604 * __normal_prio - return the priority that is based on the static prio
1606 static inline int __normal_prio(struct task_struct *p)
1608 return p->static_prio;
1612 * Calculate the expected normal priority: i.e. priority
1613 * without taking RT-inheritance into account. Might be
1614 * boosted by interactivity modifiers. Changes upon fork,
1615 * setprio syscalls, and whenever the interactivity
1616 * estimator recalculates.
1618 static inline int normal_prio(struct task_struct *p)
1622 if (task_has_dl_policy(p))
1623 prio = MAX_DL_PRIO-1;
1624 else if (task_has_rt_policy(p))
1625 prio = MAX_RT_PRIO-1 - p->rt_priority;
1627 prio = __normal_prio(p);
1632 * Calculate the current priority, i.e. the priority
1633 * taken into account by the scheduler. This value might
1634 * be boosted by RT tasks, or might be boosted by
1635 * interactivity modifiers. Will be RT if the task got
1636 * RT-boosted. If not then it returns p->normal_prio.
1638 static int effective_prio(struct task_struct *p)
1640 p->normal_prio = normal_prio(p);
1642 * If we are RT tasks or we were boosted to RT priority,
1643 * keep the priority unchanged. Otherwise, update priority
1644 * to the normal priority:
1646 if (!rt_prio(p->prio))
1647 return p->normal_prio;
1652 * task_curr - is this task currently executing on a CPU?
1653 * @p: the task in question.
1655 * Return: 1 if the task is currently executing. 0 otherwise.
1657 inline int task_curr(const struct task_struct *p)
1659 return cpu_curr(task_cpu(p)) == p;
1663 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1664 * use the balance_callback list if you want balancing.
1666 * this means any call to check_class_changed() must be followed by a call to
1667 * balance_callback().
1669 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1670 const struct sched_class *prev_class,
1673 if (prev_class != p->sched_class) {
1674 if (prev_class->switched_from)
1675 prev_class->switched_from(rq, p);
1677 p->sched_class->switched_to(rq, p);
1678 } else if (oldprio != p->prio || dl_task(p))
1679 p->sched_class->prio_changed(rq, p, oldprio);
1682 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1684 if (p->sched_class == rq->curr->sched_class)
1685 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1686 else if (p->sched_class > rq->curr->sched_class)
1690 * A queue event has occurred, and we're going to schedule. In
1691 * this case, we can save a useless back to back clock update.
1693 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1694 rq_clock_skip_update(rq);
1700 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
1701 * __set_cpus_allowed_ptr() and select_fallback_rq().
1703 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
1705 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
1708 if (is_per_cpu_kthread(p))
1709 return cpu_online(cpu);
1711 return cpu_active(cpu);
1715 * This is how migration works:
1717 * 1) we invoke migration_cpu_stop() on the target CPU using
1719 * 2) stopper starts to run (implicitly forcing the migrated thread
1721 * 3) it checks whether the migrated task is still in the wrong runqueue.
1722 * 4) if it's in the wrong runqueue then the migration thread removes
1723 * it and puts it into the right queue.
1724 * 5) stopper completes and stop_one_cpu() returns and the migration
1729 * move_queued_task - move a queued task to new rq.
1731 * Returns (locked) new rq. Old rq's lock is released.
1733 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
1734 struct task_struct *p, int new_cpu)
1736 lockdep_assert_held(&rq->lock);
1738 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
1739 set_task_cpu(p, new_cpu);
1742 rq = cpu_rq(new_cpu);
1745 BUG_ON(task_cpu(p) != new_cpu);
1746 activate_task(rq, p, 0);
1747 check_preempt_curr(rq, p, 0);
1752 struct migration_arg {
1753 struct task_struct *task;
1758 * Move (not current) task off this CPU, onto the destination CPU. We're doing
1759 * this because either it can't run here any more (set_cpus_allowed()
1760 * away from this CPU, or CPU going down), or because we're
1761 * attempting to rebalance this task on exec (sched_exec).
1763 * So we race with normal scheduler movements, but that's OK, as long
1764 * as the task is no longer on this CPU.
1766 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
1767 struct task_struct *p, int dest_cpu)
1769 /* Affinity changed (again). */
1770 if (!is_cpu_allowed(p, dest_cpu))
1773 update_rq_clock(rq);
1774 rq = move_queued_task(rq, rf, p, dest_cpu);
1780 * migration_cpu_stop - this will be executed by a highprio stopper thread
1781 * and performs thread migration by bumping thread off CPU then
1782 * 'pushing' onto another runqueue.
1784 static int migration_cpu_stop(void *data)
1786 struct migration_arg *arg = data;
1787 struct task_struct *p = arg->task;
1788 struct rq *rq = this_rq();
1792 * The original target CPU might have gone down and we might
1793 * be on another CPU but it doesn't matter.
1795 local_irq_disable();
1797 * We need to explicitly wake pending tasks before running
1798 * __migrate_task() such that we will not miss enforcing cpus_ptr
1799 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1801 flush_smp_call_function_from_idle();
1803 raw_spin_lock(&p->pi_lock);
1806 * If task_rq(p) != rq, it cannot be migrated here, because we're
1807 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1808 * we're holding p->pi_lock.
1810 if (task_rq(p) == rq) {
1811 if (task_on_rq_queued(p))
1812 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1814 p->wake_cpu = arg->dest_cpu;
1817 raw_spin_unlock(&p->pi_lock);
1824 * sched_class::set_cpus_allowed must do the below, but is not required to
1825 * actually call this function.
1827 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1829 cpumask_copy(&p->cpus_mask, new_mask);
1830 p->nr_cpus_allowed = cpumask_weight(new_mask);
1833 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1835 struct rq *rq = task_rq(p);
1836 bool queued, running;
1838 lockdep_assert_held(&p->pi_lock);
1840 queued = task_on_rq_queued(p);
1841 running = task_current(rq, p);
1845 * Because __kthread_bind() calls this on blocked tasks without
1848 lockdep_assert_held(&rq->lock);
1849 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1852 put_prev_task(rq, p);
1854 p->sched_class->set_cpus_allowed(p, new_mask);
1857 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1859 set_next_task(rq, p);
1863 * Change a given task's CPU affinity. Migrate the thread to a
1864 * proper CPU and schedule it away if the CPU it's executing on
1865 * is removed from the allowed bitmask.
1867 * NOTE: the caller must have a valid reference to the task, the
1868 * task must not exit() & deallocate itself prematurely. The
1869 * call is not atomic; no spinlocks may be held.
1871 static int __set_cpus_allowed_ptr(struct task_struct *p,
1872 const struct cpumask *new_mask, bool check)
1874 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1875 unsigned int dest_cpu;
1880 rq = task_rq_lock(p, &rf);
1881 update_rq_clock(rq);
1883 if (p->flags & PF_KTHREAD) {
1885 * Kernel threads are allowed on online && !active CPUs
1887 cpu_valid_mask = cpu_online_mask;
1891 * Must re-check here, to close a race against __kthread_bind(),
1892 * sched_setaffinity() is not guaranteed to observe the flag.
1894 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1899 if (cpumask_equal(&p->cpus_mask, new_mask))
1903 * Picking a ~random cpu helps in cases where we are changing affinity
1904 * for groups of tasks (ie. cpuset), so that load balancing is not
1905 * immediately required to distribute the tasks within their new mask.
1907 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, new_mask);
1908 if (dest_cpu >= nr_cpu_ids) {
1913 do_set_cpus_allowed(p, new_mask);
1915 if (p->flags & PF_KTHREAD) {
1917 * For kernel threads that do indeed end up on online &&
1918 * !active we want to ensure they are strict per-CPU threads.
1920 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1921 !cpumask_intersects(new_mask, cpu_active_mask) &&
1922 p->nr_cpus_allowed != 1);
1925 /* Can the task run on the task's current CPU? If so, we're done */
1926 if (cpumask_test_cpu(task_cpu(p), new_mask))
1929 if (task_running(rq, p) || p->state == TASK_WAKING) {
1930 struct migration_arg arg = { p, dest_cpu };
1931 /* Need help from migration thread: drop lock and wait. */
1932 task_rq_unlock(rq, p, &rf);
1933 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1935 } else if (task_on_rq_queued(p)) {
1937 * OK, since we're going to drop the lock immediately
1938 * afterwards anyway.
1940 rq = move_queued_task(rq, &rf, p, dest_cpu);
1943 task_rq_unlock(rq, p, &rf);
1948 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1950 return __set_cpus_allowed_ptr(p, new_mask, false);
1952 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1954 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1956 #ifdef CONFIG_SCHED_DEBUG
1958 * We should never call set_task_cpu() on a blocked task,
1959 * ttwu() will sort out the placement.
1961 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1965 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1966 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1967 * time relying on p->on_rq.
1969 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1970 p->sched_class == &fair_sched_class &&
1971 (p->on_rq && !task_on_rq_migrating(p)));
1973 #ifdef CONFIG_LOCKDEP
1975 * The caller should hold either p->pi_lock or rq->lock, when changing
1976 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1978 * sched_move_task() holds both and thus holding either pins the cgroup,
1981 * Furthermore, all task_rq users should acquire both locks, see
1984 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1985 lockdep_is_held(&task_rq(p)->lock)));
1988 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1990 WARN_ON_ONCE(!cpu_online(new_cpu));
1993 trace_sched_migrate_task(p, new_cpu);
1995 if (task_cpu(p) != new_cpu) {
1996 if (p->sched_class->migrate_task_rq)
1997 p->sched_class->migrate_task_rq(p, new_cpu);
1998 p->se.nr_migrations++;
2000 perf_event_task_migrate(p);
2003 __set_task_cpu(p, new_cpu);
2006 #ifdef CONFIG_NUMA_BALANCING
2007 static void __migrate_swap_task(struct task_struct *p, int cpu)
2009 if (task_on_rq_queued(p)) {
2010 struct rq *src_rq, *dst_rq;
2011 struct rq_flags srf, drf;
2013 src_rq = task_rq(p);
2014 dst_rq = cpu_rq(cpu);
2016 rq_pin_lock(src_rq, &srf);
2017 rq_pin_lock(dst_rq, &drf);
2019 deactivate_task(src_rq, p, 0);
2020 set_task_cpu(p, cpu);
2021 activate_task(dst_rq, p, 0);
2022 check_preempt_curr(dst_rq, p, 0);
2024 rq_unpin_lock(dst_rq, &drf);
2025 rq_unpin_lock(src_rq, &srf);
2029 * Task isn't running anymore; make it appear like we migrated
2030 * it before it went to sleep. This means on wakeup we make the
2031 * previous CPU our target instead of where it really is.
2037 struct migration_swap_arg {
2038 struct task_struct *src_task, *dst_task;
2039 int src_cpu, dst_cpu;
2042 static int migrate_swap_stop(void *data)
2044 struct migration_swap_arg *arg = data;
2045 struct rq *src_rq, *dst_rq;
2048 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
2051 src_rq = cpu_rq(arg->src_cpu);
2052 dst_rq = cpu_rq(arg->dst_cpu);
2054 double_raw_lock(&arg->src_task->pi_lock,
2055 &arg->dst_task->pi_lock);
2056 double_rq_lock(src_rq, dst_rq);
2058 if (task_cpu(arg->dst_task) != arg->dst_cpu)
2061 if (task_cpu(arg->src_task) != arg->src_cpu)
2064 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
2067 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
2070 __migrate_swap_task(arg->src_task, arg->dst_cpu);
2071 __migrate_swap_task(arg->dst_task, arg->src_cpu);
2076 double_rq_unlock(src_rq, dst_rq);
2077 raw_spin_unlock(&arg->dst_task->pi_lock);
2078 raw_spin_unlock(&arg->src_task->pi_lock);
2084 * Cross migrate two tasks
2086 int migrate_swap(struct task_struct *cur, struct task_struct *p,
2087 int target_cpu, int curr_cpu)
2089 struct migration_swap_arg arg;
2092 arg = (struct migration_swap_arg){
2094 .src_cpu = curr_cpu,
2096 .dst_cpu = target_cpu,
2099 if (arg.src_cpu == arg.dst_cpu)
2103 * These three tests are all lockless; this is OK since all of them
2104 * will be re-checked with proper locks held further down the line.
2106 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
2109 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
2112 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
2115 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
2116 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
2121 #endif /* CONFIG_NUMA_BALANCING */
2124 * wait_task_inactive - wait for a thread to unschedule.
2126 * If @match_state is nonzero, it's the @p->state value just checked and
2127 * not expected to change. If it changes, i.e. @p might have woken up,
2128 * then return zero. When we succeed in waiting for @p to be off its CPU,
2129 * we return a positive number (its total switch count). If a second call
2130 * a short while later returns the same number, the caller can be sure that
2131 * @p has remained unscheduled the whole time.
2133 * The caller must ensure that the task *will* unschedule sometime soon,
2134 * else this function might spin for a *long* time. This function can't
2135 * be called with interrupts off, or it may introduce deadlock with
2136 * smp_call_function() if an IPI is sent by the same process we are
2137 * waiting to become inactive.
2139 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2141 int running, queued;
2148 * We do the initial early heuristics without holding
2149 * any task-queue locks at all. We'll only try to get
2150 * the runqueue lock when things look like they will
2156 * If the task is actively running on another CPU
2157 * still, just relax and busy-wait without holding
2160 * NOTE! Since we don't hold any locks, it's not
2161 * even sure that "rq" stays as the right runqueue!
2162 * But we don't care, since "task_running()" will
2163 * return false if the runqueue has changed and p
2164 * is actually now running somewhere else!
2166 while (task_running(rq, p)) {
2167 if (match_state && unlikely(p->state != match_state))
2173 * Ok, time to look more closely! We need the rq
2174 * lock now, to be *sure*. If we're wrong, we'll
2175 * just go back and repeat.
2177 rq = task_rq_lock(p, &rf);
2178 trace_sched_wait_task(p);
2179 running = task_running(rq, p);
2180 queued = task_on_rq_queued(p);
2182 if (!match_state || p->state == match_state)
2183 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2184 task_rq_unlock(rq, p, &rf);
2187 * If it changed from the expected state, bail out now.
2189 if (unlikely(!ncsw))
2193 * Was it really running after all now that we
2194 * checked with the proper locks actually held?
2196 * Oops. Go back and try again..
2198 if (unlikely(running)) {
2204 * It's not enough that it's not actively running,
2205 * it must be off the runqueue _entirely_, and not
2208 * So if it was still runnable (but just not actively
2209 * running right now), it's preempted, and we should
2210 * yield - it could be a while.
2212 if (unlikely(queued)) {
2213 ktime_t to = NSEC_PER_SEC / HZ;
2215 set_current_state(TASK_UNINTERRUPTIBLE);
2216 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2221 * Ahh, all good. It wasn't running, and it wasn't
2222 * runnable, which means that it will never become
2223 * running in the future either. We're all done!
2232 * kick_process - kick a running thread to enter/exit the kernel
2233 * @p: the to-be-kicked thread
2235 * Cause a process which is running on another CPU to enter
2236 * kernel-mode, without any delay. (to get signals handled.)
2238 * NOTE: this function doesn't have to take the runqueue lock,
2239 * because all it wants to ensure is that the remote task enters
2240 * the kernel. If the IPI races and the task has been migrated
2241 * to another CPU then no harm is done and the purpose has been
2244 void kick_process(struct task_struct *p)
2250 if ((cpu != smp_processor_id()) && task_curr(p))
2251 smp_send_reschedule(cpu);
2254 EXPORT_SYMBOL_GPL(kick_process);
2257 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
2259 * A few notes on cpu_active vs cpu_online:
2261 * - cpu_active must be a subset of cpu_online
2263 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
2264 * see __set_cpus_allowed_ptr(). At this point the newly online
2265 * CPU isn't yet part of the sched domains, and balancing will not
2268 * - on CPU-down we clear cpu_active() to mask the sched domains and
2269 * avoid the load balancer to place new tasks on the to be removed
2270 * CPU. Existing tasks will remain running there and will be taken
2273 * This means that fallback selection must not select !active CPUs.
2274 * And can assume that any active CPU must be online. Conversely
2275 * select_task_rq() below may allow selection of !active CPUs in order
2276 * to satisfy the above rules.
2278 static int select_fallback_rq(int cpu, struct task_struct *p)
2280 int nid = cpu_to_node(cpu);
2281 const struct cpumask *nodemask = NULL;
2282 enum { cpuset, possible, fail } state = cpuset;
2286 * If the node that the CPU is on has been offlined, cpu_to_node()
2287 * will return -1. There is no CPU on the node, and we should
2288 * select the CPU on the other node.
2291 nodemask = cpumask_of_node(nid);
2293 /* Look for allowed, online CPU in same node. */
2294 for_each_cpu(dest_cpu, nodemask) {
2295 if (!cpu_active(dest_cpu))
2297 if (cpumask_test_cpu(dest_cpu, p->cpus_ptr))
2303 /* Any allowed, online CPU? */
2304 for_each_cpu(dest_cpu, p->cpus_ptr) {
2305 if (!is_cpu_allowed(p, dest_cpu))
2311 /* No more Mr. Nice Guy. */
2314 if (IS_ENABLED(CONFIG_CPUSETS)) {
2315 cpuset_cpus_allowed_fallback(p);
2321 do_set_cpus_allowed(p, cpu_possible_mask);
2332 if (state != cpuset) {
2334 * Don't tell them about moving exiting tasks or
2335 * kernel threads (both mm NULL), since they never
2338 if (p->mm && printk_ratelimit()) {
2339 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
2340 task_pid_nr(p), p->comm, cpu);
2348 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
2351 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
2353 lockdep_assert_held(&p->pi_lock);
2355 if (p->nr_cpus_allowed > 1)
2356 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
2358 cpu = cpumask_any(p->cpus_ptr);
2361 * In order not to call set_task_cpu() on a blocking task we need
2362 * to rely on ttwu() to place the task on a valid ->cpus_ptr
2365 * Since this is common to all placement strategies, this lives here.
2367 * [ this allows ->select_task() to simply return task_cpu(p) and
2368 * not worry about this generic constraint ]
2370 if (unlikely(!is_cpu_allowed(p, cpu)))
2371 cpu = select_fallback_rq(task_cpu(p), p);
2376 void sched_set_stop_task(int cpu, struct task_struct *stop)
2378 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2379 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2383 * Make it appear like a SCHED_FIFO task, its something
2384 * userspace knows about and won't get confused about.
2386 * Also, it will make PI more or less work without too
2387 * much confusion -- but then, stop work should not
2388 * rely on PI working anyway.
2390 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2392 stop->sched_class = &stop_sched_class;
2395 cpu_rq(cpu)->stop = stop;
2399 * Reset it back to a normal scheduling class so that
2400 * it can die in pieces.
2402 old_stop->sched_class = &rt_sched_class;
2408 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
2409 const struct cpumask *new_mask, bool check)
2411 return set_cpus_allowed_ptr(p, new_mask);
2414 #endif /* CONFIG_SMP */
2417 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2421 if (!schedstat_enabled())
2427 if (cpu == rq->cpu) {
2428 __schedstat_inc(rq->ttwu_local);
2429 __schedstat_inc(p->se.statistics.nr_wakeups_local);
2431 struct sched_domain *sd;
2433 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
2435 for_each_domain(rq->cpu, sd) {
2436 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2437 __schedstat_inc(sd->ttwu_wake_remote);
2444 if (wake_flags & WF_MIGRATED)
2445 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
2446 #endif /* CONFIG_SMP */
2448 __schedstat_inc(rq->ttwu_count);
2449 __schedstat_inc(p->se.statistics.nr_wakeups);
2451 if (wake_flags & WF_SYNC)
2452 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
2456 * Mark the task runnable and perform wakeup-preemption.
2458 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
2459 struct rq_flags *rf)
2461 check_preempt_curr(rq, p, wake_flags);
2462 p->state = TASK_RUNNING;
2463 trace_sched_wakeup(p);
2466 if (p->sched_class->task_woken) {
2468 * Our task @p is fully woken up and running; so its safe to
2469 * drop the rq->lock, hereafter rq is only used for statistics.
2471 rq_unpin_lock(rq, rf);
2472 p->sched_class->task_woken(rq, p);
2473 rq_repin_lock(rq, rf);
2476 if (rq->idle_stamp) {
2477 u64 delta = rq_clock(rq) - rq->idle_stamp;
2478 u64 max = 2*rq->max_idle_balance_cost;
2480 update_avg(&rq->avg_idle, delta);
2482 if (rq->avg_idle > max)
2491 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
2492 struct rq_flags *rf)
2494 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
2496 lockdep_assert_held(&rq->lock);
2498 if (p->sched_contributes_to_load)
2499 rq->nr_uninterruptible--;
2502 if (wake_flags & WF_MIGRATED)
2503 en_flags |= ENQUEUE_MIGRATED;
2507 delayacct_blkio_end(p);
2508 atomic_dec(&task_rq(p)->nr_iowait);
2511 activate_task(rq, p, en_flags);
2512 ttwu_do_wakeup(rq, p, wake_flags, rf);
2516 * Consider @p being inside a wait loop:
2519 * set_current_state(TASK_UNINTERRUPTIBLE);
2526 * __set_current_state(TASK_RUNNING);
2528 * between set_current_state() and schedule(). In this case @p is still
2529 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
2532 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
2533 * then schedule() must still happen and p->state can be changed to
2534 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
2535 * need to do a full wakeup with enqueue.
2537 * Returns: %true when the wakeup is done,
2540 static int ttwu_runnable(struct task_struct *p, int wake_flags)
2546 rq = __task_rq_lock(p, &rf);
2547 if (task_on_rq_queued(p)) {
2548 /* check_preempt_curr() may use rq clock */
2549 update_rq_clock(rq);
2550 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2553 __task_rq_unlock(rq, &rf);
2559 void sched_ttwu_pending(void *arg)
2561 struct llist_node *llist = arg;
2562 struct rq *rq = this_rq();
2563 struct task_struct *p, *t;
2570 * rq::ttwu_pending racy indication of out-standing wakeups.
2571 * Races such that false-negatives are possible, since they
2572 * are shorter lived that false-positives would be.
2574 WRITE_ONCE(rq->ttwu_pending, 0);
2576 rq_lock_irqsave(rq, &rf);
2577 update_rq_clock(rq);
2579 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
2580 if (WARN_ON_ONCE(p->on_cpu))
2581 smp_cond_load_acquire(&p->on_cpu, !VAL);
2583 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
2584 set_task_cpu(p, cpu_of(rq));
2586 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2589 rq_unlock_irqrestore(rq, &rf);
2592 void send_call_function_single_ipi(int cpu)
2594 struct rq *rq = cpu_rq(cpu);
2596 if (!set_nr_if_polling(rq->idle))
2597 arch_send_call_function_single_ipi(cpu);
2599 trace_sched_wake_idle_without_ipi(cpu);
2603 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2604 * necessary. The wakee CPU on receipt of the IPI will queue the task
2605 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2606 * of the wakeup instead of the waker.
2608 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2610 struct rq *rq = cpu_rq(cpu);
2612 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2614 WRITE_ONCE(rq->ttwu_pending, 1);
2615 __smp_call_single_queue(cpu, &p->wake_entry.llist);
2618 void wake_up_if_idle(int cpu)
2620 struct rq *rq = cpu_rq(cpu);
2625 if (!is_idle_task(rcu_dereference(rq->curr)))
2628 if (set_nr_if_polling(rq->idle)) {
2629 trace_sched_wake_idle_without_ipi(cpu);
2631 rq_lock_irqsave(rq, &rf);
2632 if (is_idle_task(rq->curr))
2633 smp_send_reschedule(cpu);
2634 /* Else CPU is not idle, do nothing here: */
2635 rq_unlock_irqrestore(rq, &rf);
2642 bool cpus_share_cache(int this_cpu, int that_cpu)
2644 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2647 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
2650 * If the CPU does not share cache, then queue the task on the
2651 * remote rqs wakelist to avoid accessing remote data.
2653 if (!cpus_share_cache(smp_processor_id(), cpu))
2657 * If the task is descheduling and the only running task on the
2658 * CPU then use the wakelist to offload the task activation to
2659 * the soon-to-be-idle CPU as the current CPU is likely busy.
2660 * nr_running is checked to avoid unnecessary task stacking.
2662 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
2668 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2670 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
2671 if (WARN_ON_ONCE(cpu == smp_processor_id()))
2674 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2675 __ttwu_queue_wakelist(p, cpu, wake_flags);
2682 #else /* !CONFIG_SMP */
2684 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2689 #endif /* CONFIG_SMP */
2691 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2693 struct rq *rq = cpu_rq(cpu);
2696 if (ttwu_queue_wakelist(p, cpu, wake_flags))
2700 update_rq_clock(rq);
2701 ttwu_do_activate(rq, p, wake_flags, &rf);
2706 * Notes on Program-Order guarantees on SMP systems.
2710 * The basic program-order guarantee on SMP systems is that when a task [t]
2711 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2712 * execution on its new CPU [c1].
2714 * For migration (of runnable tasks) this is provided by the following means:
2716 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2717 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2718 * rq(c1)->lock (if not at the same time, then in that order).
2719 * C) LOCK of the rq(c1)->lock scheduling in task
2721 * Release/acquire chaining guarantees that B happens after A and C after B.
2722 * Note: the CPU doing B need not be c0 or c1
2731 * UNLOCK rq(0)->lock
2733 * LOCK rq(0)->lock // orders against CPU0
2735 * UNLOCK rq(0)->lock
2739 * UNLOCK rq(1)->lock
2741 * LOCK rq(1)->lock // orders against CPU2
2744 * UNLOCK rq(1)->lock
2747 * BLOCKING -- aka. SLEEP + WAKEUP
2749 * For blocking we (obviously) need to provide the same guarantee as for
2750 * migration. However the means are completely different as there is no lock
2751 * chain to provide order. Instead we do:
2753 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
2754 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
2758 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2760 * LOCK rq(0)->lock LOCK X->pi_lock
2763 * smp_store_release(X->on_cpu, 0);
2765 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2771 * X->state = RUNNING
2772 * UNLOCK rq(2)->lock
2774 * LOCK rq(2)->lock // orders against CPU1
2777 * UNLOCK rq(2)->lock
2780 * UNLOCK rq(0)->lock
2783 * However, for wakeups there is a second guarantee we must provide, namely we
2784 * must ensure that CONDITION=1 done by the caller can not be reordered with
2785 * accesses to the task state; see try_to_wake_up() and set_current_state().
2789 * try_to_wake_up - wake up a thread
2790 * @p: the thread to be awakened
2791 * @state: the mask of task states that can be woken
2792 * @wake_flags: wake modifier flags (WF_*)
2794 * Conceptually does:
2796 * If (@state & @p->state) @p->state = TASK_RUNNING.
2798 * If the task was not queued/runnable, also place it back on a runqueue.
2800 * This function is atomic against schedule() which would dequeue the task.
2802 * It issues a full memory barrier before accessing @p->state, see the comment
2803 * with set_current_state().
2805 * Uses p->pi_lock to serialize against concurrent wake-ups.
2807 * Relies on p->pi_lock stabilizing:
2810 * - p->sched_task_group
2811 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
2813 * Tries really hard to only take one task_rq(p)->lock for performance.
2814 * Takes rq->lock in:
2815 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
2816 * - ttwu_queue() -- new rq, for enqueue of the task;
2817 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
2819 * As a consequence we race really badly with just about everything. See the
2820 * many memory barriers and their comments for details.
2822 * Return: %true if @p->state changes (an actual wakeup was done),
2826 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2828 unsigned long flags;
2829 int cpu, success = 0;
2834 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2835 * == smp_processor_id()'. Together this means we can special
2836 * case the whole 'p->on_rq && ttwu_runnable()' case below
2837 * without taking any locks.
2840 * - we rely on Program-Order guarantees for all the ordering,
2841 * - we're serialized against set_special_state() by virtue of
2842 * it disabling IRQs (this allows not taking ->pi_lock).
2844 if (!(p->state & state))
2848 trace_sched_waking(p);
2849 p->state = TASK_RUNNING;
2850 trace_sched_wakeup(p);
2855 * If we are going to wake up a thread waiting for CONDITION we
2856 * need to ensure that CONDITION=1 done by the caller can not be
2857 * reordered with p->state check below. This pairs with smp_store_mb()
2858 * in set_current_state() that the waiting thread does.
2860 raw_spin_lock_irqsave(&p->pi_lock, flags);
2861 smp_mb__after_spinlock();
2862 if (!(p->state & state))
2865 trace_sched_waking(p);
2867 /* We're going to change ->state: */
2871 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2872 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2873 * in smp_cond_load_acquire() below.
2875 * sched_ttwu_pending() try_to_wake_up()
2876 * STORE p->on_rq = 1 LOAD p->state
2879 * __schedule() (switch to task 'p')
2880 * LOCK rq->lock smp_rmb();
2881 * smp_mb__after_spinlock();
2885 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2887 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2888 * __schedule(). See the comment for smp_mb__after_spinlock().
2890 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2893 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
2898 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2899 * possible to, falsely, observe p->on_cpu == 0.
2901 * One must be running (->on_cpu == 1) in order to remove oneself
2902 * from the runqueue.
2904 * __schedule() (switch to task 'p') try_to_wake_up()
2905 * STORE p->on_cpu = 1 LOAD p->on_rq
2908 * __schedule() (put 'p' to sleep)
2909 * LOCK rq->lock smp_rmb();
2910 * smp_mb__after_spinlock();
2911 * STORE p->on_rq = 0 LOAD p->on_cpu
2913 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2914 * __schedule(). See the comment for smp_mb__after_spinlock().
2916 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
2917 * schedule()'s deactivate_task() has 'happened' and p will no longer
2918 * care about it's own p->state. See the comment in __schedule().
2920 smp_acquire__after_ctrl_dep();
2923 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
2924 * == 0), which means we need to do an enqueue, change p->state to
2925 * TASK_WAKING such that we can unlock p->pi_lock before doing the
2926 * enqueue, such as ttwu_queue_wakelist().
2928 p->state = TASK_WAKING;
2931 * If the owning (remote) CPU is still in the middle of schedule() with
2932 * this task as prev, considering queueing p on the remote CPUs wake_list
2933 * which potentially sends an IPI instead of spinning on p->on_cpu to
2934 * let the waker make forward progress. This is safe because IRQs are
2935 * disabled and the IPI will deliver after on_cpu is cleared.
2937 * Ensure we load task_cpu(p) after p->on_cpu:
2939 * set_task_cpu(p, cpu);
2940 * STORE p->cpu = @cpu
2941 * __schedule() (switch to task 'p')
2943 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
2944 * STORE p->on_cpu = 1 LOAD p->cpu
2946 * to ensure we observe the correct CPU on which the task is currently
2949 if (smp_load_acquire(&p->on_cpu) &&
2950 ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_CPU))
2954 * If the owning (remote) CPU is still in the middle of schedule() with
2955 * this task as prev, wait until its done referencing the task.
2957 * Pairs with the smp_store_release() in finish_task().
2959 * This ensures that tasks getting woken will be fully ordered against
2960 * their previous state and preserve Program Order.
2962 smp_cond_load_acquire(&p->on_cpu, !VAL);
2964 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2965 if (task_cpu(p) != cpu) {
2967 delayacct_blkio_end(p);
2968 atomic_dec(&task_rq(p)->nr_iowait);
2971 wake_flags |= WF_MIGRATED;
2972 psi_ttwu_dequeue(p);
2973 set_task_cpu(p, cpu);
2977 #endif /* CONFIG_SMP */
2979 ttwu_queue(p, cpu, wake_flags);
2981 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2984 ttwu_stat(p, task_cpu(p), wake_flags);
2991 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2992 * @p: Process for which the function is to be invoked.
2993 * @func: Function to invoke.
2994 * @arg: Argument to function.
2996 * If the specified task can be quickly locked into a definite state
2997 * (either sleeping or on a given runqueue), arrange to keep it in that
2998 * state while invoking @func(@arg). This function can use ->on_rq and
2999 * task_curr() to work out what the state is, if required. Given that
3000 * @func can be invoked with a runqueue lock held, it had better be quite
3004 * @false if the task slipped out from under the locks.
3005 * @true if the task was locked onto a runqueue or is sleeping.
3006 * However, @func can override this by returning @false.
3008 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3014 lockdep_assert_irqs_enabled();
3015 raw_spin_lock_irq(&p->pi_lock);
3017 rq = __task_rq_lock(p, &rf);
3018 if (task_rq(p) == rq)
3027 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3032 raw_spin_unlock_irq(&p->pi_lock);
3037 * wake_up_process - Wake up a specific process
3038 * @p: The process to be woken up.
3040 * Attempt to wake up the nominated process and move it to the set of runnable
3043 * Return: 1 if the process was woken up, 0 if it was already running.
3045 * This function executes a full memory barrier before accessing the task state.
3047 int wake_up_process(struct task_struct *p)
3049 return try_to_wake_up(p, TASK_NORMAL, 0);
3051 EXPORT_SYMBOL(wake_up_process);
3053 int wake_up_state(struct task_struct *p, unsigned int state)
3055 return try_to_wake_up(p, state, 0);
3059 * Perform scheduler related setup for a newly forked process p.
3060 * p is forked by current.
3062 * __sched_fork() is basic setup used by init_idle() too:
3064 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3069 p->se.exec_start = 0;
3070 p->se.sum_exec_runtime = 0;
3071 p->se.prev_sum_exec_runtime = 0;
3072 p->se.nr_migrations = 0;
3074 INIT_LIST_HEAD(&p->se.group_node);
3076 #ifdef CONFIG_FAIR_GROUP_SCHED
3077 p->se.cfs_rq = NULL;
3080 #ifdef CONFIG_SCHEDSTATS
3081 /* Even if schedstat is disabled, there should not be garbage */
3082 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3085 RB_CLEAR_NODE(&p->dl.rb_node);
3086 init_dl_task_timer(&p->dl);
3087 init_dl_inactive_task_timer(&p->dl);
3088 __dl_clear_params(p);
3090 INIT_LIST_HEAD(&p->rt.run_list);
3092 p->rt.time_slice = sched_rr_timeslice;
3096 #ifdef CONFIG_PREEMPT_NOTIFIERS
3097 INIT_HLIST_HEAD(&p->preempt_notifiers);
3100 #ifdef CONFIG_COMPACTION
3101 p->capture_control = NULL;
3103 init_numa_balancing(clone_flags, p);
3105 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3109 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3111 #ifdef CONFIG_NUMA_BALANCING
3113 void set_numabalancing_state(bool enabled)
3116 static_branch_enable(&sched_numa_balancing);
3118 static_branch_disable(&sched_numa_balancing);
3121 #ifdef CONFIG_PROC_SYSCTL
3122 int sysctl_numa_balancing(struct ctl_table *table, int write,
3123 void *buffer, size_t *lenp, loff_t *ppos)
3127 int state = static_branch_likely(&sched_numa_balancing);
3129 if (write && !capable(CAP_SYS_ADMIN))
3134 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3138 set_numabalancing_state(state);
3144 #ifdef CONFIG_SCHEDSTATS
3146 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3147 static bool __initdata __sched_schedstats = false;
3149 static void set_schedstats(bool enabled)
3152 static_branch_enable(&sched_schedstats);
3154 static_branch_disable(&sched_schedstats);
3157 void force_schedstat_enabled(void)
3159 if (!schedstat_enabled()) {
3160 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3161 static_branch_enable(&sched_schedstats);
3165 static int __init setup_schedstats(char *str)
3172 * This code is called before jump labels have been set up, so we can't
3173 * change the static branch directly just yet. Instead set a temporary
3174 * variable so init_schedstats() can do it later.
3176 if (!strcmp(str, "enable")) {
3177 __sched_schedstats = true;
3179 } else if (!strcmp(str, "disable")) {
3180 __sched_schedstats = false;
3185 pr_warn("Unable to parse schedstats=\n");
3189 __setup("schedstats=", setup_schedstats);
3191 static void __init init_schedstats(void)
3193 set_schedstats(__sched_schedstats);
3196 #ifdef CONFIG_PROC_SYSCTL
3197 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
3198 size_t *lenp, loff_t *ppos)
3202 int state = static_branch_likely(&sched_schedstats);
3204 if (write && !capable(CAP_SYS_ADMIN))
3209 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3213 set_schedstats(state);
3216 #endif /* CONFIG_PROC_SYSCTL */
3217 #else /* !CONFIG_SCHEDSTATS */
3218 static inline void init_schedstats(void) {}
3219 #endif /* CONFIG_SCHEDSTATS */
3222 * fork()/clone()-time setup:
3224 int sched_fork(unsigned long clone_flags, struct task_struct *p)
3226 unsigned long flags;
3228 __sched_fork(clone_flags, p);
3230 * We mark the process as NEW here. This guarantees that
3231 * nobody will actually run it, and a signal or other external
3232 * event cannot wake it up and insert it on the runqueue either.
3234 p->state = TASK_NEW;
3237 * Make sure we do not leak PI boosting priority to the child.
3239 p->prio = current->normal_prio;
3244 * Revert to default priority/policy on fork if requested.
3246 if (unlikely(p->sched_reset_on_fork)) {
3247 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3248 p->policy = SCHED_NORMAL;
3249 p->static_prio = NICE_TO_PRIO(0);
3251 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3252 p->static_prio = NICE_TO_PRIO(0);
3254 p->prio = p->normal_prio = __normal_prio(p);
3255 set_load_weight(p, false);
3258 * We don't need the reset flag anymore after the fork. It has
3259 * fulfilled its duty:
3261 p->sched_reset_on_fork = 0;
3264 if (dl_prio(p->prio))
3266 else if (rt_prio(p->prio))
3267 p->sched_class = &rt_sched_class;
3269 p->sched_class = &fair_sched_class;
3271 init_entity_runnable_average(&p->se);
3274 * The child is not yet in the pid-hash so no cgroup attach races,
3275 * and the cgroup is pinned to this child due to cgroup_fork()
3276 * is ran before sched_fork().
3278 * Silence PROVE_RCU.
3280 raw_spin_lock_irqsave(&p->pi_lock, flags);
3283 * We're setting the CPU for the first time, we don't migrate,
3284 * so use __set_task_cpu().
3286 __set_task_cpu(p, smp_processor_id());
3287 if (p->sched_class->task_fork)
3288 p->sched_class->task_fork(p);
3289 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3291 #ifdef CONFIG_SCHED_INFO
3292 if (likely(sched_info_on()))
3293 memset(&p->sched_info, 0, sizeof(p->sched_info));
3295 #if defined(CONFIG_SMP)
3298 init_task_preempt_count(p);
3300 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3301 RB_CLEAR_NODE(&p->pushable_dl_tasks);
3306 void sched_post_fork(struct task_struct *p)
3308 uclamp_post_fork(p);
3311 unsigned long to_ratio(u64 period, u64 runtime)
3313 if (runtime == RUNTIME_INF)
3317 * Doing this here saves a lot of checks in all
3318 * the calling paths, and returning zero seems
3319 * safe for them anyway.
3324 return div64_u64(runtime << BW_SHIFT, period);
3328 * wake_up_new_task - wake up a newly created task for the first time.
3330 * This function will do some initial scheduler statistics housekeeping
3331 * that must be done for every newly created context, then puts the task
3332 * on the runqueue and wakes it.
3334 void wake_up_new_task(struct task_struct *p)
3339 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3340 p->state = TASK_RUNNING;
3343 * Fork balancing, do it here and not earlier because:
3344 * - cpus_ptr can change in the fork path
3345 * - any previously selected CPU might disappear through hotplug
3347 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3348 * as we're not fully set-up yet.
3350 p->recent_used_cpu = task_cpu(p);
3352 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
3354 rq = __task_rq_lock(p, &rf);
3355 update_rq_clock(rq);
3356 post_init_entity_util_avg(p);
3358 activate_task(rq, p, ENQUEUE_NOCLOCK);
3359 trace_sched_wakeup_new(p);
3360 check_preempt_curr(rq, p, WF_FORK);
3362 if (p->sched_class->task_woken) {
3364 * Nothing relies on rq->lock after this, so its fine to
3367 rq_unpin_lock(rq, &rf);
3368 p->sched_class->task_woken(rq, p);
3369 rq_repin_lock(rq, &rf);
3372 task_rq_unlock(rq, p, &rf);
3375 #ifdef CONFIG_PREEMPT_NOTIFIERS
3377 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3379 void preempt_notifier_inc(void)
3381 static_branch_inc(&preempt_notifier_key);
3383 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3385 void preempt_notifier_dec(void)
3387 static_branch_dec(&preempt_notifier_key);
3389 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3392 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3393 * @notifier: notifier struct to register
3395 void preempt_notifier_register(struct preempt_notifier *notifier)
3397 if (!static_branch_unlikely(&preempt_notifier_key))
3398 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3400 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3402 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3405 * preempt_notifier_unregister - no longer interested in preemption notifications
3406 * @notifier: notifier struct to unregister
3408 * This is *not* safe to call from within a preemption notifier.
3410 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3412 hlist_del(¬ifier->link);
3414 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3416 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3418 struct preempt_notifier *notifier;
3420 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3421 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3424 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3426 if (static_branch_unlikely(&preempt_notifier_key))
3427 __fire_sched_in_preempt_notifiers(curr);
3431 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3432 struct task_struct *next)
3434 struct preempt_notifier *notifier;
3436 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3437 notifier->ops->sched_out(notifier, next);
3440 static __always_inline void
3441 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3442 struct task_struct *next)
3444 if (static_branch_unlikely(&preempt_notifier_key))
3445 __fire_sched_out_preempt_notifiers(curr, next);
3448 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3450 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3455 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3456 struct task_struct *next)
3460 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3462 static inline void prepare_task(struct task_struct *next)
3466 * Claim the task as running, we do this before switching to it
3467 * such that any running task will have this set.
3469 * See the ttwu() WF_ON_CPU case and its ordering comment.
3471 WRITE_ONCE(next->on_cpu, 1);
3475 static inline void finish_task(struct task_struct *prev)
3479 * This must be the very last reference to @prev from this CPU. After
3480 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3481 * must ensure this doesn't happen until the switch is completely
3484 * In particular, the load of prev->state in finish_task_switch() must
3485 * happen before this.
3487 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3489 smp_store_release(&prev->on_cpu, 0);
3494 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3497 * Since the runqueue lock will be released by the next
3498 * task (which is an invalid locking op but in the case
3499 * of the scheduler it's an obvious special-case), so we
3500 * do an early lockdep release here:
3502 rq_unpin_lock(rq, rf);
3503 spin_release(&rq->lock.dep_map, _THIS_IP_);
3504 #ifdef CONFIG_DEBUG_SPINLOCK
3505 /* this is a valid case when another task releases the spinlock */
3506 rq->lock.owner = next;
3510 static inline void finish_lock_switch(struct rq *rq)
3513 * If we are tracking spinlock dependencies then we have to
3514 * fix up the runqueue lock - which gets 'carried over' from
3515 * prev into current:
3517 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3518 raw_spin_unlock_irq(&rq->lock);
3522 * NOP if the arch has not defined these:
3525 #ifndef prepare_arch_switch
3526 # define prepare_arch_switch(next) do { } while (0)
3529 #ifndef finish_arch_post_lock_switch
3530 # define finish_arch_post_lock_switch() do { } while (0)
3534 * prepare_task_switch - prepare to switch tasks
3535 * @rq: the runqueue preparing to switch
3536 * @prev: the current task that is being switched out
3537 * @next: the task we are going to switch to.
3539 * This is called with the rq lock held and interrupts off. It must
3540 * be paired with a subsequent finish_task_switch after the context
3543 * prepare_task_switch sets up locking and calls architecture specific
3547 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3548 struct task_struct *next)
3550 kcov_prepare_switch(prev);
3551 sched_info_switch(rq, prev, next);
3552 perf_event_task_sched_out(prev, next);
3554 fire_sched_out_preempt_notifiers(prev, next);
3556 prepare_arch_switch(next);
3560 * finish_task_switch - clean up after a task-switch
3561 * @prev: the thread we just switched away from.
3563 * finish_task_switch must be called after the context switch, paired
3564 * with a prepare_task_switch call before the context switch.
3565 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3566 * and do any other architecture-specific cleanup actions.
3568 * Note that we may have delayed dropping an mm in context_switch(). If
3569 * so, we finish that here outside of the runqueue lock. (Doing it
3570 * with the lock held can cause deadlocks; see schedule() for
3573 * The context switch have flipped the stack from under us and restored the
3574 * local variables which were saved when this task called schedule() in the
3575 * past. prev == current is still correct but we need to recalculate this_rq
3576 * because prev may have moved to another CPU.
3578 static struct rq *finish_task_switch(struct task_struct *prev)
3579 __releases(rq->lock)
3581 struct rq *rq = this_rq();
3582 struct mm_struct *mm = rq->prev_mm;
3586 * The previous task will have left us with a preempt_count of 2
3587 * because it left us after:
3590 * preempt_disable(); // 1
3592 * raw_spin_lock_irq(&rq->lock) // 2
3594 * Also, see FORK_PREEMPT_COUNT.
3596 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3597 "corrupted preempt_count: %s/%d/0x%x\n",
3598 current->comm, current->pid, preempt_count()))
3599 preempt_count_set(FORK_PREEMPT_COUNT);
3604 * A task struct has one reference for the use as "current".
3605 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3606 * schedule one last time. The schedule call will never return, and
3607 * the scheduled task must drop that reference.
3609 * We must observe prev->state before clearing prev->on_cpu (in
3610 * finish_task), otherwise a concurrent wakeup can get prev
3611 * running on another CPU and we could rave with its RUNNING -> DEAD
3612 * transition, resulting in a double drop.
3614 prev_state = prev->state;
3615 vtime_task_switch(prev);
3616 perf_event_task_sched_in(prev, current);
3618 finish_lock_switch(rq);
3619 finish_arch_post_lock_switch();
3620 kcov_finish_switch(current);
3622 fire_sched_in_preempt_notifiers(current);
3624 * When switching through a kernel thread, the loop in
3625 * membarrier_{private,global}_expedited() may have observed that
3626 * kernel thread and not issued an IPI. It is therefore possible to
3627 * schedule between user->kernel->user threads without passing though
3628 * switch_mm(). Membarrier requires a barrier after storing to
3629 * rq->curr, before returning to userspace, so provide them here:
3631 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3632 * provided by mmdrop(),
3633 * - a sync_core for SYNC_CORE.
3636 membarrier_mm_sync_core_before_usermode(mm);
3639 if (unlikely(prev_state == TASK_DEAD)) {
3640 if (prev->sched_class->task_dead)
3641 prev->sched_class->task_dead(prev);
3644 * Remove function-return probe instances associated with this
3645 * task and put them back on the free list.
3647 kprobe_flush_task(prev);
3649 /* Task is done with its stack. */
3650 put_task_stack(prev);
3652 put_task_struct_rcu_user(prev);
3655 tick_nohz_task_switch();
3661 /* rq->lock is NOT held, but preemption is disabled */
3662 static void __balance_callback(struct rq *rq)
3664 struct callback_head *head, *next;
3665 void (*func)(struct rq *rq);
3666 unsigned long flags;
3668 raw_spin_lock_irqsave(&rq->lock, flags);
3669 head = rq->balance_callback;
3670 rq->balance_callback = NULL;
3672 func = (void (*)(struct rq *))head->func;
3679 raw_spin_unlock_irqrestore(&rq->lock, flags);
3682 static inline void balance_callback(struct rq *rq)
3684 if (unlikely(rq->balance_callback))
3685 __balance_callback(rq);
3690 static inline void balance_callback(struct rq *rq)
3697 * schedule_tail - first thing a freshly forked thread must call.
3698 * @prev: the thread we just switched away from.
3700 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3701 __releases(rq->lock)
3706 * New tasks start with FORK_PREEMPT_COUNT, see there and
3707 * finish_task_switch() for details.
3709 * finish_task_switch() will drop rq->lock() and lower preempt_count
3710 * and the preempt_enable() will end up enabling preemption (on
3711 * PREEMPT_COUNT kernels).
3714 rq = finish_task_switch(prev);
3715 balance_callback(rq);
3718 if (current->set_child_tid)
3719 put_user(task_pid_vnr(current), current->set_child_tid);
3721 calculate_sigpending();
3725 * context_switch - switch to the new MM and the new thread's register state.
3727 static __always_inline struct rq *
3728 context_switch(struct rq *rq, struct task_struct *prev,
3729 struct task_struct *next, struct rq_flags *rf)
3731 prepare_task_switch(rq, prev, next);
3734 * For paravirt, this is coupled with an exit in switch_to to
3735 * combine the page table reload and the switch backend into
3738 arch_start_context_switch(prev);
3741 * kernel -> kernel lazy + transfer active
3742 * user -> kernel lazy + mmgrab() active
3744 * kernel -> user switch + mmdrop() active
3745 * user -> user switch
3747 if (!next->mm) { // to kernel
3748 enter_lazy_tlb(prev->active_mm, next);
3750 next->active_mm = prev->active_mm;
3751 if (prev->mm) // from user
3752 mmgrab(prev->active_mm);
3754 prev->active_mm = NULL;
3756 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3758 * sys_membarrier() requires an smp_mb() between setting
3759 * rq->curr / membarrier_switch_mm() and returning to userspace.
3761 * The below provides this either through switch_mm(), or in
3762 * case 'prev->active_mm == next->mm' through
3763 * finish_task_switch()'s mmdrop().
3765 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3767 if (!prev->mm) { // from kernel
3768 /* will mmdrop() in finish_task_switch(). */
3769 rq->prev_mm = prev->active_mm;
3770 prev->active_mm = NULL;
3774 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3776 prepare_lock_switch(rq, next, rf);
3778 /* Here we just switch the register state and the stack. */
3779 switch_to(prev, next, prev);
3782 return finish_task_switch(prev);
3786 * nr_running and nr_context_switches:
3788 * externally visible scheduler statistics: current number of runnable
3789 * threads, total number of context switches performed since bootup.
3791 unsigned long nr_running(void)
3793 unsigned long i, sum = 0;
3795 for_each_online_cpu(i)
3796 sum += cpu_rq(i)->nr_running;
3802 * Check if only the current task is running on the CPU.
3804 * Caution: this function does not check that the caller has disabled
3805 * preemption, thus the result might have a time-of-check-to-time-of-use
3806 * race. The caller is responsible to use it correctly, for example:
3808 * - from a non-preemptible section (of course)
3810 * - from a thread that is bound to a single CPU
3812 * - in a loop with very short iterations (e.g. a polling loop)
3814 bool single_task_running(void)
3816 return raw_rq()->nr_running == 1;
3818 EXPORT_SYMBOL(single_task_running);
3820 unsigned long long nr_context_switches(void)
3823 unsigned long long sum = 0;
3825 for_each_possible_cpu(i)
3826 sum += cpu_rq(i)->nr_switches;
3832 * Consumers of these two interfaces, like for example the cpuidle menu
3833 * governor, are using nonsensical data. Preferring shallow idle state selection
3834 * for a CPU that has IO-wait which might not even end up running the task when
3835 * it does become runnable.
3838 unsigned long nr_iowait_cpu(int cpu)
3840 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3844 * IO-wait accounting, and how its mostly bollocks (on SMP).
3846 * The idea behind IO-wait account is to account the idle time that we could
3847 * have spend running if it were not for IO. That is, if we were to improve the
3848 * storage performance, we'd have a proportional reduction in IO-wait time.
3850 * This all works nicely on UP, where, when a task blocks on IO, we account
3851 * idle time as IO-wait, because if the storage were faster, it could've been
3852 * running and we'd not be idle.
3854 * This has been extended to SMP, by doing the same for each CPU. This however
3857 * Imagine for instance the case where two tasks block on one CPU, only the one
3858 * CPU will have IO-wait accounted, while the other has regular idle. Even
3859 * though, if the storage were faster, both could've ran at the same time,
3860 * utilising both CPUs.
3862 * This means, that when looking globally, the current IO-wait accounting on
3863 * SMP is a lower bound, by reason of under accounting.
3865 * Worse, since the numbers are provided per CPU, they are sometimes
3866 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3867 * associated with any one particular CPU, it can wake to another CPU than it
3868 * blocked on. This means the per CPU IO-wait number is meaningless.
3870 * Task CPU affinities can make all that even more 'interesting'.
3873 unsigned long nr_iowait(void)
3875 unsigned long i, sum = 0;
3877 for_each_possible_cpu(i)
3878 sum += nr_iowait_cpu(i);
3886 * sched_exec - execve() is a valuable balancing opportunity, because at
3887 * this point the task has the smallest effective memory and cache footprint.
3889 void sched_exec(void)
3891 struct task_struct *p = current;
3892 unsigned long flags;
3895 raw_spin_lock_irqsave(&p->pi_lock, flags);
3896 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3897 if (dest_cpu == smp_processor_id())
3900 if (likely(cpu_active(dest_cpu))) {
3901 struct migration_arg arg = { p, dest_cpu };
3903 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3904 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3908 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3913 DEFINE_PER_CPU(struct kernel_stat, kstat);
3914 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3916 EXPORT_PER_CPU_SYMBOL(kstat);
3917 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3920 * The function fair_sched_class.update_curr accesses the struct curr
3921 * and its field curr->exec_start; when called from task_sched_runtime(),
3922 * we observe a high rate of cache misses in practice.
3923 * Prefetching this data results in improved performance.
3925 static inline void prefetch_curr_exec_start(struct task_struct *p)
3927 #ifdef CONFIG_FAIR_GROUP_SCHED
3928 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3930 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3933 prefetch(&curr->exec_start);
3937 * Return accounted runtime for the task.
3938 * In case the task is currently running, return the runtime plus current's
3939 * pending runtime that have not been accounted yet.
3941 unsigned long long task_sched_runtime(struct task_struct *p)
3947 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3949 * 64-bit doesn't need locks to atomically read a 64-bit value.
3950 * So we have a optimization chance when the task's delta_exec is 0.
3951 * Reading ->on_cpu is racy, but this is ok.
3953 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3954 * If we race with it entering CPU, unaccounted time is 0. This is
3955 * indistinguishable from the read occurring a few cycles earlier.
3956 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3957 * been accounted, so we're correct here as well.
3959 if (!p->on_cpu || !task_on_rq_queued(p))
3960 return p->se.sum_exec_runtime;
3963 rq = task_rq_lock(p, &rf);
3965 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3966 * project cycles that may never be accounted to this
3967 * thread, breaking clock_gettime().
3969 if (task_current(rq, p) && task_on_rq_queued(p)) {
3970 prefetch_curr_exec_start(p);
3971 update_rq_clock(rq);
3972 p->sched_class->update_curr(rq);
3974 ns = p->se.sum_exec_runtime;
3975 task_rq_unlock(rq, p, &rf);
3981 * This function gets called by the timer code, with HZ frequency.
3982 * We call it with interrupts disabled.
3984 void scheduler_tick(void)
3986 int cpu = smp_processor_id();
3987 struct rq *rq = cpu_rq(cpu);
3988 struct task_struct *curr = rq->curr;
3990 unsigned long thermal_pressure;
3992 arch_scale_freq_tick();
3997 update_rq_clock(rq);
3998 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
3999 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4000 curr->sched_class->task_tick(rq, curr, 0);
4001 calc_global_load_tick(rq);
4006 perf_event_task_tick();
4009 rq->idle_balance = idle_cpu(cpu);
4010 trigger_load_balance(rq);
4014 #ifdef CONFIG_NO_HZ_FULL
4019 struct delayed_work work;
4021 /* Values for ->state, see diagram below. */
4022 #define TICK_SCHED_REMOTE_OFFLINE 0
4023 #define TICK_SCHED_REMOTE_OFFLINING 1
4024 #define TICK_SCHED_REMOTE_RUNNING 2
4027 * State diagram for ->state:
4030 * TICK_SCHED_REMOTE_OFFLINE
4033 * | | sched_tick_remote()
4036 * +--TICK_SCHED_REMOTE_OFFLINING
4039 * sched_tick_start() | | sched_tick_stop()
4042 * TICK_SCHED_REMOTE_RUNNING
4045 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4046 * and sched_tick_start() are happy to leave the state in RUNNING.
4049 static struct tick_work __percpu *tick_work_cpu;
4051 static void sched_tick_remote(struct work_struct *work)
4053 struct delayed_work *dwork = to_delayed_work(work);
4054 struct tick_work *twork = container_of(dwork, struct tick_work, work);
4055 int cpu = twork->cpu;
4056 struct rq *rq = cpu_rq(cpu);
4057 struct task_struct *curr;
4063 * Handle the tick only if it appears the remote CPU is running in full
4064 * dynticks mode. The check is racy by nature, but missing a tick or
4065 * having one too much is no big deal because the scheduler tick updates
4066 * statistics and checks timeslices in a time-independent way, regardless
4067 * of when exactly it is running.
4069 if (!tick_nohz_tick_stopped_cpu(cpu))
4072 rq_lock_irq(rq, &rf);
4074 if (cpu_is_offline(cpu))
4077 update_rq_clock(rq);
4079 if (!is_idle_task(curr)) {
4081 * Make sure the next tick runs within a reasonable
4084 delta = rq_clock_task(rq) - curr->se.exec_start;
4085 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
4087 curr->sched_class->task_tick(rq, curr, 0);
4089 calc_load_nohz_remote(rq);
4091 rq_unlock_irq(rq, &rf);
4095 * Run the remote tick once per second (1Hz). This arbitrary
4096 * frequency is large enough to avoid overload but short enough
4097 * to keep scheduler internal stats reasonably up to date. But
4098 * first update state to reflect hotplug activity if required.
4100 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
4101 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
4102 if (os == TICK_SCHED_REMOTE_RUNNING)
4103 queue_delayed_work(system_unbound_wq, dwork, HZ);
4106 static void sched_tick_start(int cpu)
4109 struct tick_work *twork;
4111 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4114 WARN_ON_ONCE(!tick_work_cpu);
4116 twork = per_cpu_ptr(tick_work_cpu, cpu);
4117 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
4118 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
4119 if (os == TICK_SCHED_REMOTE_OFFLINE) {
4121 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
4122 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
4126 #ifdef CONFIG_HOTPLUG_CPU
4127 static void sched_tick_stop(int cpu)
4129 struct tick_work *twork;
4132 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4135 WARN_ON_ONCE(!tick_work_cpu);
4137 twork = per_cpu_ptr(tick_work_cpu, cpu);
4138 /* There cannot be competing actions, but don't rely on stop-machine. */
4139 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
4140 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
4141 /* Don't cancel, as this would mess up the state machine. */
4143 #endif /* CONFIG_HOTPLUG_CPU */
4145 int __init sched_tick_offload_init(void)
4147 tick_work_cpu = alloc_percpu(struct tick_work);
4148 BUG_ON(!tick_work_cpu);
4152 #else /* !CONFIG_NO_HZ_FULL */
4153 static inline void sched_tick_start(int cpu) { }
4154 static inline void sched_tick_stop(int cpu) { }
4157 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4158 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4160 * If the value passed in is equal to the current preempt count
4161 * then we just disabled preemption. Start timing the latency.
4163 static inline void preempt_latency_start(int val)
4165 if (preempt_count() == val) {
4166 unsigned long ip = get_lock_parent_ip();
4167 #ifdef CONFIG_DEBUG_PREEMPT
4168 current->preempt_disable_ip = ip;
4170 trace_preempt_off(CALLER_ADDR0, ip);
4174 void preempt_count_add(int val)
4176 #ifdef CONFIG_DEBUG_PREEMPT
4180 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4183 __preempt_count_add(val);
4184 #ifdef CONFIG_DEBUG_PREEMPT
4186 * Spinlock count overflowing soon?
4188 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4191 preempt_latency_start(val);
4193 EXPORT_SYMBOL(preempt_count_add);
4194 NOKPROBE_SYMBOL(preempt_count_add);
4197 * If the value passed in equals to the current preempt count
4198 * then we just enabled preemption. Stop timing the latency.
4200 static inline void preempt_latency_stop(int val)
4202 if (preempt_count() == val)
4203 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
4206 void preempt_count_sub(int val)
4208 #ifdef CONFIG_DEBUG_PREEMPT
4212 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4215 * Is the spinlock portion underflowing?
4217 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4218 !(preempt_count() & PREEMPT_MASK)))
4222 preempt_latency_stop(val);
4223 __preempt_count_sub(val);
4225 EXPORT_SYMBOL(preempt_count_sub);
4226 NOKPROBE_SYMBOL(preempt_count_sub);
4229 static inline void preempt_latency_start(int val) { }
4230 static inline void preempt_latency_stop(int val) { }
4233 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
4235 #ifdef CONFIG_DEBUG_PREEMPT
4236 return p->preempt_disable_ip;
4243 * Print scheduling while atomic bug:
4245 static noinline void __schedule_bug(struct task_struct *prev)
4247 /* Save this before calling printk(), since that will clobber it */
4248 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
4250 if (oops_in_progress)
4253 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4254 prev->comm, prev->pid, preempt_count());
4256 debug_show_held_locks(prev);
4258 if (irqs_disabled())
4259 print_irqtrace_events(prev);
4260 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
4261 && in_atomic_preempt_off()) {
4262 pr_err("Preemption disabled at:");
4263 print_ip_sym(KERN_ERR, preempt_disable_ip);
4266 panic("scheduling while atomic\n");
4269 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4273 * Various schedule()-time debugging checks and statistics:
4275 static inline void schedule_debug(struct task_struct *prev, bool preempt)
4277 #ifdef CONFIG_SCHED_STACK_END_CHECK
4278 if (task_stack_end_corrupted(prev))
4279 panic("corrupted stack end detected inside scheduler\n");
4281 if (task_scs_end_corrupted(prev))
4282 panic("corrupted shadow stack detected inside scheduler\n");
4285 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4286 if (!preempt && prev->state && prev->non_block_count) {
4287 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4288 prev->comm, prev->pid, prev->non_block_count);
4290 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4294 if (unlikely(in_atomic_preempt_off())) {
4295 __schedule_bug(prev);
4296 preempt_count_set(PREEMPT_DISABLED);
4300 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4302 schedstat_inc(this_rq()->sched_count);
4305 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
4306 struct rq_flags *rf)
4309 const struct sched_class *class;
4311 * We must do the balancing pass before put_prev_task(), such
4312 * that when we release the rq->lock the task is in the same
4313 * state as before we took rq->lock.
4315 * We can terminate the balance pass as soon as we know there is
4316 * a runnable task of @class priority or higher.
4318 for_class_range(class, prev->sched_class, &idle_sched_class) {
4319 if (class->balance(rq, prev, rf))
4324 put_prev_task(rq, prev);
4328 * Pick up the highest-prio task:
4330 static inline struct task_struct *
4331 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4333 const struct sched_class *class;
4334 struct task_struct *p;
4337 * Optimization: we know that if all tasks are in the fair class we can
4338 * call that function directly, but only if the @prev task wasn't of a
4339 * higher scheduling class, because otherwise those loose the
4340 * opportunity to pull in more work from other CPUs.
4342 if (likely(prev->sched_class <= &fair_sched_class &&
4343 rq->nr_running == rq->cfs.h_nr_running)) {
4345 p = pick_next_task_fair(rq, prev, rf);
4346 if (unlikely(p == RETRY_TASK))
4349 /* Assumes fair_sched_class->next == idle_sched_class */
4351 put_prev_task(rq, prev);
4352 p = pick_next_task_idle(rq);
4359 put_prev_task_balance(rq, prev, rf);
4361 for_each_class(class) {
4362 p = class->pick_next_task(rq);
4367 /* The idle class should always have a runnable task: */
4372 * __schedule() is the main scheduler function.
4374 * The main means of driving the scheduler and thus entering this function are:
4376 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4378 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4379 * paths. For example, see arch/x86/entry_64.S.
4381 * To drive preemption between tasks, the scheduler sets the flag in timer
4382 * interrupt handler scheduler_tick().
4384 * 3. Wakeups don't really cause entry into schedule(). They add a
4385 * task to the run-queue and that's it.
4387 * Now, if the new task added to the run-queue preempts the current
4388 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4389 * called on the nearest possible occasion:
4391 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4393 * - in syscall or exception context, at the next outmost
4394 * preempt_enable(). (this might be as soon as the wake_up()'s
4397 * - in IRQ context, return from interrupt-handler to
4398 * preemptible context
4400 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4403 * - cond_resched() call
4404 * - explicit schedule() call
4405 * - return from syscall or exception to user-space
4406 * - return from interrupt-handler to user-space
4408 * WARNING: must be called with preemption disabled!
4410 static void __sched notrace __schedule(bool preempt)
4412 struct task_struct *prev, *next;
4413 unsigned long *switch_count;
4414 unsigned long prev_state;
4419 cpu = smp_processor_id();
4423 schedule_debug(prev, preempt);
4425 if (sched_feat(HRTICK))
4428 local_irq_disable();
4429 rcu_note_context_switch(preempt);
4432 * Make sure that signal_pending_state()->signal_pending() below
4433 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4434 * done by the caller to avoid the race with signal_wake_up():
4436 * __set_current_state(@state) signal_wake_up()
4437 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4438 * wake_up_state(p, state)
4439 * LOCK rq->lock LOCK p->pi_state
4440 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4441 * if (signal_pending_state()) if (p->state & @state)
4443 * Also, the membarrier system call requires a full memory barrier
4444 * after coming from user-space, before storing to rq->curr.
4447 smp_mb__after_spinlock();
4449 /* Promote REQ to ACT */
4450 rq->clock_update_flags <<= 1;
4451 update_rq_clock(rq);
4453 switch_count = &prev->nivcsw;
4456 * We must load prev->state once (task_struct::state is volatile), such
4459 * - we form a control dependency vs deactivate_task() below.
4460 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
4462 prev_state = prev->state;
4463 if (!preempt && prev_state) {
4464 if (signal_pending_state(prev_state, prev)) {
4465 prev->state = TASK_RUNNING;
4467 prev->sched_contributes_to_load =
4468 (prev_state & TASK_UNINTERRUPTIBLE) &&
4469 !(prev_state & TASK_NOLOAD) &&
4470 !(prev->flags & PF_FROZEN);
4472 if (prev->sched_contributes_to_load)
4473 rq->nr_uninterruptible++;
4476 * __schedule() ttwu()
4477 * prev_state = prev->state; if (p->on_rq && ...)
4478 * if (prev_state) goto out;
4479 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
4480 * p->state = TASK_WAKING
4482 * Where __schedule() and ttwu() have matching control dependencies.
4484 * After this, schedule() must not care about p->state any more.
4486 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4488 if (prev->in_iowait) {
4489 atomic_inc(&rq->nr_iowait);
4490 delayacct_blkio_start();
4493 switch_count = &prev->nvcsw;
4496 next = pick_next_task(rq, prev, &rf);
4497 clear_tsk_need_resched(prev);
4498 clear_preempt_need_resched();
4500 if (likely(prev != next)) {
4503 * RCU users of rcu_dereference(rq->curr) may not see
4504 * changes to task_struct made by pick_next_task().
4506 RCU_INIT_POINTER(rq->curr, next);
4508 * The membarrier system call requires each architecture
4509 * to have a full memory barrier after updating
4510 * rq->curr, before returning to user-space.
4512 * Here are the schemes providing that barrier on the
4513 * various architectures:
4514 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4515 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4516 * - finish_lock_switch() for weakly-ordered
4517 * architectures where spin_unlock is a full barrier,
4518 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4519 * is a RELEASE barrier),
4523 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
4525 trace_sched_switch(preempt, prev, next);
4527 /* Also unlocks the rq: */
4528 rq = context_switch(rq, prev, next, &rf);
4530 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4531 rq_unlock_irq(rq, &rf);
4534 balance_callback(rq);
4537 void __noreturn do_task_dead(void)
4539 /* Causes final put_task_struct in finish_task_switch(): */
4540 set_special_state(TASK_DEAD);
4542 /* Tell freezer to ignore us: */
4543 current->flags |= PF_NOFREEZE;
4548 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4553 static inline void sched_submit_work(struct task_struct *tsk)
4555 unsigned int task_flags;
4560 task_flags = tsk->flags;
4562 * If a worker went to sleep, notify and ask workqueue whether
4563 * it wants to wake up a task to maintain concurrency.
4564 * As this function is called inside the schedule() context,
4565 * we disable preemption to avoid it calling schedule() again
4566 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4569 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4571 if (task_flags & PF_WQ_WORKER)
4572 wq_worker_sleeping(tsk);
4574 io_wq_worker_sleeping(tsk);
4575 preempt_enable_no_resched();
4578 if (tsk_is_pi_blocked(tsk))
4582 * If we are going to sleep and we have plugged IO queued,
4583 * make sure to submit it to avoid deadlocks.
4585 if (blk_needs_flush_plug(tsk))
4586 blk_schedule_flush_plug(tsk);
4589 static void sched_update_worker(struct task_struct *tsk)
4591 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4592 if (tsk->flags & PF_WQ_WORKER)
4593 wq_worker_running(tsk);
4595 io_wq_worker_running(tsk);
4599 asmlinkage __visible void __sched schedule(void)
4601 struct task_struct *tsk = current;
4603 sched_submit_work(tsk);
4607 sched_preempt_enable_no_resched();
4608 } while (need_resched());
4609 sched_update_worker(tsk);
4611 EXPORT_SYMBOL(schedule);
4614 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4615 * state (have scheduled out non-voluntarily) by making sure that all
4616 * tasks have either left the run queue or have gone into user space.
4617 * As idle tasks do not do either, they must not ever be preempted
4618 * (schedule out non-voluntarily).
4620 * schedule_idle() is similar to schedule_preempt_disable() except that it
4621 * never enables preemption because it does not call sched_submit_work().
4623 void __sched schedule_idle(void)
4626 * As this skips calling sched_submit_work(), which the idle task does
4627 * regardless because that function is a nop when the task is in a
4628 * TASK_RUNNING state, make sure this isn't used someplace that the
4629 * current task can be in any other state. Note, idle is always in the
4630 * TASK_RUNNING state.
4632 WARN_ON_ONCE(current->state);
4635 } while (need_resched());
4638 #ifdef CONFIG_CONTEXT_TRACKING
4639 asmlinkage __visible void __sched schedule_user(void)
4642 * If we come here after a random call to set_need_resched(),
4643 * or we have been woken up remotely but the IPI has not yet arrived,
4644 * we haven't yet exited the RCU idle mode. Do it here manually until
4645 * we find a better solution.
4647 * NB: There are buggy callers of this function. Ideally we
4648 * should warn if prev_state != CONTEXT_USER, but that will trigger
4649 * too frequently to make sense yet.
4651 enum ctx_state prev_state = exception_enter();
4653 exception_exit(prev_state);
4658 * schedule_preempt_disabled - called with preemption disabled
4660 * Returns with preemption disabled. Note: preempt_count must be 1
4662 void __sched schedule_preempt_disabled(void)
4664 sched_preempt_enable_no_resched();
4669 static void __sched notrace preempt_schedule_common(void)
4673 * Because the function tracer can trace preempt_count_sub()
4674 * and it also uses preempt_enable/disable_notrace(), if
4675 * NEED_RESCHED is set, the preempt_enable_notrace() called
4676 * by the function tracer will call this function again and
4677 * cause infinite recursion.
4679 * Preemption must be disabled here before the function
4680 * tracer can trace. Break up preempt_disable() into two
4681 * calls. One to disable preemption without fear of being
4682 * traced. The other to still record the preemption latency,
4683 * which can also be traced by the function tracer.
4685 preempt_disable_notrace();
4686 preempt_latency_start(1);
4688 preempt_latency_stop(1);
4689 preempt_enable_no_resched_notrace();
4692 * Check again in case we missed a preemption opportunity
4693 * between schedule and now.
4695 } while (need_resched());
4698 #ifdef CONFIG_PREEMPTION
4700 * This is the entry point to schedule() from in-kernel preemption
4701 * off of preempt_enable.
4703 asmlinkage __visible void __sched notrace preempt_schedule(void)
4706 * If there is a non-zero preempt_count or interrupts are disabled,
4707 * we do not want to preempt the current task. Just return..
4709 if (likely(!preemptible()))
4712 preempt_schedule_common();
4714 NOKPROBE_SYMBOL(preempt_schedule);
4715 EXPORT_SYMBOL(preempt_schedule);
4718 * preempt_schedule_notrace - preempt_schedule called by tracing
4720 * The tracing infrastructure uses preempt_enable_notrace to prevent
4721 * recursion and tracing preempt enabling caused by the tracing
4722 * infrastructure itself. But as tracing can happen in areas coming
4723 * from userspace or just about to enter userspace, a preempt enable
4724 * can occur before user_exit() is called. This will cause the scheduler
4725 * to be called when the system is still in usermode.
4727 * To prevent this, the preempt_enable_notrace will use this function
4728 * instead of preempt_schedule() to exit user context if needed before
4729 * calling the scheduler.
4731 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4733 enum ctx_state prev_ctx;
4735 if (likely(!preemptible()))
4740 * Because the function tracer can trace preempt_count_sub()
4741 * and it also uses preempt_enable/disable_notrace(), if
4742 * NEED_RESCHED is set, the preempt_enable_notrace() called
4743 * by the function tracer will call this function again and
4744 * cause infinite recursion.
4746 * Preemption must be disabled here before the function
4747 * tracer can trace. Break up preempt_disable() into two
4748 * calls. One to disable preemption without fear of being
4749 * traced. The other to still record the preemption latency,
4750 * which can also be traced by the function tracer.
4752 preempt_disable_notrace();
4753 preempt_latency_start(1);
4755 * Needs preempt disabled in case user_exit() is traced
4756 * and the tracer calls preempt_enable_notrace() causing
4757 * an infinite recursion.
4759 prev_ctx = exception_enter();
4761 exception_exit(prev_ctx);
4763 preempt_latency_stop(1);
4764 preempt_enable_no_resched_notrace();
4765 } while (need_resched());
4767 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4769 #endif /* CONFIG_PREEMPTION */
4772 * This is the entry point to schedule() from kernel preemption
4773 * off of irq context.
4774 * Note, that this is called and return with irqs disabled. This will
4775 * protect us against recursive calling from irq.
4777 asmlinkage __visible void __sched preempt_schedule_irq(void)
4779 enum ctx_state prev_state;
4781 /* Catch callers which need to be fixed */
4782 BUG_ON(preempt_count() || !irqs_disabled());
4784 prev_state = exception_enter();
4790 local_irq_disable();
4791 sched_preempt_enable_no_resched();
4792 } while (need_resched());
4794 exception_exit(prev_state);
4797 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4800 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
4801 return try_to_wake_up(curr->private, mode, wake_flags);
4803 EXPORT_SYMBOL(default_wake_function);
4805 #ifdef CONFIG_RT_MUTEXES
4807 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4810 prio = min(prio, pi_task->prio);
4815 static inline int rt_effective_prio(struct task_struct *p, int prio)
4817 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4819 return __rt_effective_prio(pi_task, prio);
4823 * rt_mutex_setprio - set the current priority of a task
4825 * @pi_task: donor task
4827 * This function changes the 'effective' priority of a task. It does
4828 * not touch ->normal_prio like __setscheduler().
4830 * Used by the rt_mutex code to implement priority inheritance
4831 * logic. Call site only calls if the priority of the task changed.
4833 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4835 int prio, oldprio, queued, running, queue_flag =
4836 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4837 const struct sched_class *prev_class;
4841 /* XXX used to be waiter->prio, not waiter->task->prio */
4842 prio = __rt_effective_prio(pi_task, p->normal_prio);
4845 * If nothing changed; bail early.
4847 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4850 rq = __task_rq_lock(p, &rf);
4851 update_rq_clock(rq);
4853 * Set under pi_lock && rq->lock, such that the value can be used under
4856 * Note that there is loads of tricky to make this pointer cache work
4857 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4858 * ensure a task is de-boosted (pi_task is set to NULL) before the
4859 * task is allowed to run again (and can exit). This ensures the pointer
4860 * points to a blocked task -- which guaratees the task is present.
4862 p->pi_top_task = pi_task;
4865 * For FIFO/RR we only need to set prio, if that matches we're done.
4867 if (prio == p->prio && !dl_prio(prio))
4871 * Idle task boosting is a nono in general. There is one
4872 * exception, when PREEMPT_RT and NOHZ is active:
4874 * The idle task calls get_next_timer_interrupt() and holds
4875 * the timer wheel base->lock on the CPU and another CPU wants
4876 * to access the timer (probably to cancel it). We can safely
4877 * ignore the boosting request, as the idle CPU runs this code
4878 * with interrupts disabled and will complete the lock
4879 * protected section without being interrupted. So there is no
4880 * real need to boost.
4882 if (unlikely(p == rq->idle)) {
4883 WARN_ON(p != rq->curr);
4884 WARN_ON(p->pi_blocked_on);
4888 trace_sched_pi_setprio(p, pi_task);
4891 if (oldprio == prio)
4892 queue_flag &= ~DEQUEUE_MOVE;
4894 prev_class = p->sched_class;
4895 queued = task_on_rq_queued(p);
4896 running = task_current(rq, p);
4898 dequeue_task(rq, p, queue_flag);
4900 put_prev_task(rq, p);
4903 * Boosting condition are:
4904 * 1. -rt task is running and holds mutex A
4905 * --> -dl task blocks on mutex A
4907 * 2. -dl task is running and holds mutex A
4908 * --> -dl task blocks on mutex A and could preempt the
4911 if (dl_prio(prio)) {
4912 if (!dl_prio(p->normal_prio) ||
4913 (pi_task && dl_prio(pi_task->prio) &&
4914 dl_entity_preempt(&pi_task->dl, &p->dl))) {
4915 p->dl.pi_se = pi_task->dl.pi_se;
4916 queue_flag |= ENQUEUE_REPLENISH;
4918 p->dl.pi_se = &p->dl;
4920 p->sched_class = &dl_sched_class;
4921 } else if (rt_prio(prio)) {
4922 if (dl_prio(oldprio))
4923 p->dl.pi_se = &p->dl;
4925 queue_flag |= ENQUEUE_HEAD;
4926 p->sched_class = &rt_sched_class;
4928 if (dl_prio(oldprio))
4929 p->dl.pi_se = &p->dl;
4930 if (rt_prio(oldprio))
4932 p->sched_class = &fair_sched_class;
4938 enqueue_task(rq, p, queue_flag);
4940 set_next_task(rq, p);
4942 check_class_changed(rq, p, prev_class, oldprio);
4944 /* Avoid rq from going away on us: */
4946 __task_rq_unlock(rq, &rf);
4948 balance_callback(rq);
4952 static inline int rt_effective_prio(struct task_struct *p, int prio)
4958 void set_user_nice(struct task_struct *p, long nice)
4960 bool queued, running;
4965 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
4968 * We have to be careful, if called from sys_setpriority(),
4969 * the task might be in the middle of scheduling on another CPU.
4971 rq = task_rq_lock(p, &rf);
4972 update_rq_clock(rq);
4975 * The RT priorities are set via sched_setscheduler(), but we still
4976 * allow the 'normal' nice value to be set - but as expected
4977 * it wont have any effect on scheduling until the task is
4978 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
4980 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
4981 p->static_prio = NICE_TO_PRIO(nice);
4984 queued = task_on_rq_queued(p);
4985 running = task_current(rq, p);
4987 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
4989 put_prev_task(rq, p);
4991 p->static_prio = NICE_TO_PRIO(nice);
4992 set_load_weight(p, true);
4994 p->prio = effective_prio(p);
4997 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
4999 set_next_task(rq, p);
5002 * If the task increased its priority or is running and
5003 * lowered its priority, then reschedule its CPU:
5005 p->sched_class->prio_changed(rq, p, old_prio);
5008 task_rq_unlock(rq, p, &rf);
5010 EXPORT_SYMBOL(set_user_nice);
5013 * can_nice - check if a task can reduce its nice value
5017 int can_nice(const struct task_struct *p, const int nice)
5019 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5020 int nice_rlim = nice_to_rlimit(nice);
5022 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5023 capable(CAP_SYS_NICE));
5026 #ifdef __ARCH_WANT_SYS_NICE
5029 * sys_nice - change the priority of the current process.
5030 * @increment: priority increment
5032 * sys_setpriority is a more generic, but much slower function that
5033 * does similar things.
5035 SYSCALL_DEFINE1(nice, int, increment)
5040 * Setpriority might change our priority at the same moment.
5041 * We don't have to worry. Conceptually one call occurs first
5042 * and we have a single winner.
5044 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
5045 nice = task_nice(current) + increment;
5047 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
5048 if (increment < 0 && !can_nice(current, nice))
5051 retval = security_task_setnice(current, nice);
5055 set_user_nice(current, nice);
5062 * task_prio - return the priority value of a given task.
5063 * @p: the task in question.
5065 * Return: The priority value as seen by users in /proc.
5066 * RT tasks are offset by -200. Normal tasks are centered
5067 * around 0, value goes from -16 to +15.
5069 int task_prio(const struct task_struct *p)
5071 return p->prio - MAX_RT_PRIO;
5075 * idle_cpu - is a given CPU idle currently?
5076 * @cpu: the processor in question.
5078 * Return: 1 if the CPU is currently idle. 0 otherwise.
5080 int idle_cpu(int cpu)
5082 struct rq *rq = cpu_rq(cpu);
5084 if (rq->curr != rq->idle)
5091 if (rq->ttwu_pending)
5099 * available_idle_cpu - is a given CPU idle for enqueuing work.
5100 * @cpu: the CPU in question.
5102 * Return: 1 if the CPU is currently idle. 0 otherwise.
5104 int available_idle_cpu(int cpu)
5109 if (vcpu_is_preempted(cpu))
5116 * idle_task - return the idle task for a given CPU.
5117 * @cpu: the processor in question.
5119 * Return: The idle task for the CPU @cpu.
5121 struct task_struct *idle_task(int cpu)
5123 return cpu_rq(cpu)->idle;
5127 * find_process_by_pid - find a process with a matching PID value.
5128 * @pid: the pid in question.
5130 * The task of @pid, if found. %NULL otherwise.
5132 static struct task_struct *find_process_by_pid(pid_t pid)
5134 return pid ? find_task_by_vpid(pid) : current;
5138 * sched_setparam() passes in -1 for its policy, to let the functions
5139 * it calls know not to change it.
5141 #define SETPARAM_POLICY -1
5143 static void __setscheduler_params(struct task_struct *p,
5144 const struct sched_attr *attr)
5146 int policy = attr->sched_policy;
5148 if (policy == SETPARAM_POLICY)
5153 if (dl_policy(policy))
5154 __setparam_dl(p, attr);
5155 else if (fair_policy(policy))
5156 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
5159 * __sched_setscheduler() ensures attr->sched_priority == 0 when
5160 * !rt_policy. Always setting this ensures that things like
5161 * getparam()/getattr() don't report silly values for !rt tasks.
5163 p->rt_priority = attr->sched_priority;
5164 p->normal_prio = normal_prio(p);
5165 set_load_weight(p, true);
5168 /* Actually do priority change: must hold pi & rq lock. */
5169 static void __setscheduler(struct rq *rq, struct task_struct *p,
5170 const struct sched_attr *attr, bool keep_boost)
5173 * If params can't change scheduling class changes aren't allowed
5176 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
5179 __setscheduler_params(p, attr);
5182 * Keep a potential priority boosting if called from
5183 * sched_setscheduler().
5185 p->prio = normal_prio(p);
5187 p->prio = rt_effective_prio(p, p->prio);
5189 if (dl_prio(p->prio))
5190 p->sched_class = &dl_sched_class;
5191 else if (rt_prio(p->prio))
5192 p->sched_class = &rt_sched_class;
5194 p->sched_class = &fair_sched_class;
5198 * Check the target process has a UID that matches the current process's:
5200 static bool check_same_owner(struct task_struct *p)
5202 const struct cred *cred = current_cred(), *pcred;
5206 pcred = __task_cred(p);
5207 match = (uid_eq(cred->euid, pcred->euid) ||
5208 uid_eq(cred->euid, pcred->uid));
5213 static int __sched_setscheduler(struct task_struct *p,
5214 const struct sched_attr *attr,
5217 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
5218 MAX_RT_PRIO - 1 - attr->sched_priority;
5219 int retval, oldprio, oldpolicy = -1, queued, running;
5220 int new_effective_prio, policy = attr->sched_policy;
5221 const struct sched_class *prev_class;
5224 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5227 /* The pi code expects interrupts enabled */
5228 BUG_ON(pi && in_interrupt());
5230 /* Double check policy once rq lock held: */
5232 reset_on_fork = p->sched_reset_on_fork;
5233 policy = oldpolicy = p->policy;
5235 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
5237 if (!valid_policy(policy))
5241 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
5245 * Valid priorities for SCHED_FIFO and SCHED_RR are
5246 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5247 * SCHED_BATCH and SCHED_IDLE is 0.
5249 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
5250 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
5252 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
5253 (rt_policy(policy) != (attr->sched_priority != 0)))
5257 * Allow unprivileged RT tasks to decrease priority:
5259 if (user && !capable(CAP_SYS_NICE)) {
5260 if (fair_policy(policy)) {
5261 if (attr->sched_nice < task_nice(p) &&
5262 !can_nice(p, attr->sched_nice))
5266 if (rt_policy(policy)) {
5267 unsigned long rlim_rtprio =
5268 task_rlimit(p, RLIMIT_RTPRIO);
5270 /* Can't set/change the rt policy: */
5271 if (policy != p->policy && !rlim_rtprio)
5274 /* Can't increase priority: */
5275 if (attr->sched_priority > p->rt_priority &&
5276 attr->sched_priority > rlim_rtprio)
5281 * Can't set/change SCHED_DEADLINE policy at all for now
5282 * (safest behavior); in the future we would like to allow
5283 * unprivileged DL tasks to increase their relative deadline
5284 * or reduce their runtime (both ways reducing utilization)
5286 if (dl_policy(policy))
5290 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5291 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5293 if (task_has_idle_policy(p) && !idle_policy(policy)) {
5294 if (!can_nice(p, task_nice(p)))
5298 /* Can't change other user's priorities: */
5299 if (!check_same_owner(p))
5302 /* Normal users shall not reset the sched_reset_on_fork flag: */
5303 if (p->sched_reset_on_fork && !reset_on_fork)
5308 if (attr->sched_flags & SCHED_FLAG_SUGOV)
5311 retval = security_task_setscheduler(p);
5316 /* Update task specific "requested" clamps */
5317 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
5318 retval = uclamp_validate(p, attr);
5327 * Make sure no PI-waiters arrive (or leave) while we are
5328 * changing the priority of the task:
5330 * To be able to change p->policy safely, the appropriate
5331 * runqueue lock must be held.
5333 rq = task_rq_lock(p, &rf);
5334 update_rq_clock(rq);
5337 * Changing the policy of the stop threads its a very bad idea:
5339 if (p == rq->stop) {
5345 * If not changing anything there's no need to proceed further,
5346 * but store a possible modification of reset_on_fork.
5348 if (unlikely(policy == p->policy)) {
5349 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
5351 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
5353 if (dl_policy(policy) && dl_param_changed(p, attr))
5355 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
5358 p->sched_reset_on_fork = reset_on_fork;
5365 #ifdef CONFIG_RT_GROUP_SCHED
5367 * Do not allow realtime tasks into groups that have no runtime
5370 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5371 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5372 !task_group_is_autogroup(task_group(p))) {
5378 if (dl_bandwidth_enabled() && dl_policy(policy) &&
5379 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
5380 cpumask_t *span = rq->rd->span;
5383 * Don't allow tasks with an affinity mask smaller than
5384 * the entire root_domain to become SCHED_DEADLINE. We
5385 * will also fail if there's no bandwidth available.
5387 if (!cpumask_subset(span, p->cpus_ptr) ||
5388 rq->rd->dl_bw.bw == 0) {
5396 /* Re-check policy now with rq lock held: */
5397 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5398 policy = oldpolicy = -1;
5399 task_rq_unlock(rq, p, &rf);
5401 cpuset_read_unlock();
5406 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5407 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5410 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
5415 p->sched_reset_on_fork = reset_on_fork;
5420 * Take priority boosted tasks into account. If the new
5421 * effective priority is unchanged, we just store the new
5422 * normal parameters and do not touch the scheduler class and
5423 * the runqueue. This will be done when the task deboost
5426 new_effective_prio = rt_effective_prio(p, newprio);
5427 if (new_effective_prio == oldprio)
5428 queue_flags &= ~DEQUEUE_MOVE;
5431 queued = task_on_rq_queued(p);
5432 running = task_current(rq, p);
5434 dequeue_task(rq, p, queue_flags);
5436 put_prev_task(rq, p);
5438 prev_class = p->sched_class;
5440 __setscheduler(rq, p, attr, pi);
5441 __setscheduler_uclamp(p, attr);
5445 * We enqueue to tail when the priority of a task is
5446 * increased (user space view).
5448 if (oldprio < p->prio)
5449 queue_flags |= ENQUEUE_HEAD;
5451 enqueue_task(rq, p, queue_flags);
5454 set_next_task(rq, p);
5456 check_class_changed(rq, p, prev_class, oldprio);
5458 /* Avoid rq from going away on us: */
5460 task_rq_unlock(rq, p, &rf);
5463 cpuset_read_unlock();
5464 rt_mutex_adjust_pi(p);
5467 /* Run balance callbacks after we've adjusted the PI chain: */
5468 balance_callback(rq);
5474 task_rq_unlock(rq, p, &rf);
5476 cpuset_read_unlock();
5480 static int _sched_setscheduler(struct task_struct *p, int policy,
5481 const struct sched_param *param, bool check)
5483 struct sched_attr attr = {
5484 .sched_policy = policy,
5485 .sched_priority = param->sched_priority,
5486 .sched_nice = PRIO_TO_NICE(p->static_prio),
5489 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5490 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5491 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5492 policy &= ~SCHED_RESET_ON_FORK;
5493 attr.sched_policy = policy;
5496 return __sched_setscheduler(p, &attr, check, true);
5499 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5500 * @p: the task in question.
5501 * @policy: new policy.
5502 * @param: structure containing the new RT priority.
5504 * Use sched_set_fifo(), read its comment.
5506 * Return: 0 on success. An error code otherwise.
5508 * NOTE that the task may be already dead.
5510 int sched_setscheduler(struct task_struct *p, int policy,
5511 const struct sched_param *param)
5513 return _sched_setscheduler(p, policy, param, true);
5516 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5518 return __sched_setscheduler(p, attr, true, true);
5521 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5523 return __sched_setscheduler(p, attr, false, true);
5527 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5528 * @p: the task in question.
5529 * @policy: new policy.
5530 * @param: structure containing the new RT priority.
5532 * Just like sched_setscheduler, only don't bother checking if the
5533 * current context has permission. For example, this is needed in
5534 * stop_machine(): we create temporary high priority worker threads,
5535 * but our caller might not have that capability.
5537 * Return: 0 on success. An error code otherwise.
5539 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5540 const struct sched_param *param)
5542 return _sched_setscheduler(p, policy, param, false);
5546 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
5547 * incapable of resource management, which is the one thing an OS really should
5550 * This is of course the reason it is limited to privileged users only.
5552 * Worse still; it is fundamentally impossible to compose static priority
5553 * workloads. You cannot take two correctly working static prio workloads
5554 * and smash them together and still expect them to work.
5556 * For this reason 'all' FIFO tasks the kernel creates are basically at:
5560 * The administrator _MUST_ configure the system, the kernel simply doesn't
5561 * know enough information to make a sensible choice.
5563 void sched_set_fifo(struct task_struct *p)
5565 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
5566 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
5568 EXPORT_SYMBOL_GPL(sched_set_fifo);
5571 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
5573 void sched_set_fifo_low(struct task_struct *p)
5575 struct sched_param sp = { .sched_priority = 1 };
5576 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
5578 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
5580 void sched_set_normal(struct task_struct *p, int nice)
5582 struct sched_attr attr = {
5583 .sched_policy = SCHED_NORMAL,
5586 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
5588 EXPORT_SYMBOL_GPL(sched_set_normal);
5591 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5593 struct sched_param lparam;
5594 struct task_struct *p;
5597 if (!param || pid < 0)
5599 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5604 p = find_process_by_pid(pid);
5610 retval = sched_setscheduler(p, policy, &lparam);
5618 * Mimics kernel/events/core.c perf_copy_attr().
5620 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5625 /* Zero the full structure, so that a short copy will be nice: */
5626 memset(attr, 0, sizeof(*attr));
5628 ret = get_user(size, &uattr->size);
5632 /* ABI compatibility quirk: */
5634 size = SCHED_ATTR_SIZE_VER0;
5635 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5638 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5645 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5646 size < SCHED_ATTR_SIZE_VER1)
5650 * XXX: Do we want to be lenient like existing syscalls; or do we want
5651 * to be strict and return an error on out-of-bounds values?
5653 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5658 put_user(sizeof(*attr), &uattr->size);
5663 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5664 * @pid: the pid in question.
5665 * @policy: new policy.
5666 * @param: structure containing the new RT priority.
5668 * Return: 0 on success. An error code otherwise.
5670 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5675 return do_sched_setscheduler(pid, policy, param);
5679 * sys_sched_setparam - set/change the RT priority of a thread
5680 * @pid: the pid in question.
5681 * @param: structure containing the new RT priority.
5683 * Return: 0 on success. An error code otherwise.
5685 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5687 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5691 * sys_sched_setattr - same as above, but with extended sched_attr
5692 * @pid: the pid in question.
5693 * @uattr: structure containing the extended parameters.
5694 * @flags: for future extension.
5696 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5697 unsigned int, flags)
5699 struct sched_attr attr;
5700 struct task_struct *p;
5703 if (!uattr || pid < 0 || flags)
5706 retval = sched_copy_attr(uattr, &attr);
5710 if ((int)attr.sched_policy < 0)
5712 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5713 attr.sched_policy = SETPARAM_POLICY;
5717 p = find_process_by_pid(pid);
5723 retval = sched_setattr(p, &attr);
5731 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5732 * @pid: the pid in question.
5734 * Return: On success, the policy of the thread. Otherwise, a negative error
5737 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5739 struct task_struct *p;
5747 p = find_process_by_pid(pid);
5749 retval = security_task_getscheduler(p);
5752 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5759 * sys_sched_getparam - get the RT priority of a thread
5760 * @pid: the pid in question.
5761 * @param: structure containing the RT priority.
5763 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5766 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5768 struct sched_param lp = { .sched_priority = 0 };
5769 struct task_struct *p;
5772 if (!param || pid < 0)
5776 p = find_process_by_pid(pid);
5781 retval = security_task_getscheduler(p);
5785 if (task_has_rt_policy(p))
5786 lp.sched_priority = p->rt_priority;
5790 * This one might sleep, we cannot do it with a spinlock held ...
5792 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5802 * Copy the kernel size attribute structure (which might be larger
5803 * than what user-space knows about) to user-space.
5805 * Note that all cases are valid: user-space buffer can be larger or
5806 * smaller than the kernel-space buffer. The usual case is that both
5807 * have the same size.
5810 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5811 struct sched_attr *kattr,
5814 unsigned int ksize = sizeof(*kattr);
5816 if (!access_ok(uattr, usize))
5820 * sched_getattr() ABI forwards and backwards compatibility:
5822 * If usize == ksize then we just copy everything to user-space and all is good.
5824 * If usize < ksize then we only copy as much as user-space has space for,
5825 * this keeps ABI compatibility as well. We skip the rest.
5827 * If usize > ksize then user-space is using a newer version of the ABI,
5828 * which part the kernel doesn't know about. Just ignore it - tooling can
5829 * detect the kernel's knowledge of attributes from the attr->size value
5830 * which is set to ksize in this case.
5832 kattr->size = min(usize, ksize);
5834 if (copy_to_user(uattr, kattr, kattr->size))
5841 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5842 * @pid: the pid in question.
5843 * @uattr: structure containing the extended parameters.
5844 * @usize: sizeof(attr) for fwd/bwd comp.
5845 * @flags: for future extension.
5847 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5848 unsigned int, usize, unsigned int, flags)
5850 struct sched_attr kattr = { };
5851 struct task_struct *p;
5854 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5855 usize < SCHED_ATTR_SIZE_VER0 || flags)
5859 p = find_process_by_pid(pid);
5864 retval = security_task_getscheduler(p);
5868 kattr.sched_policy = p->policy;
5869 if (p->sched_reset_on_fork)
5870 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5871 if (task_has_dl_policy(p))
5872 __getparam_dl(p, &kattr);
5873 else if (task_has_rt_policy(p))
5874 kattr.sched_priority = p->rt_priority;
5876 kattr.sched_nice = task_nice(p);
5878 #ifdef CONFIG_UCLAMP_TASK
5880 * This could race with another potential updater, but this is fine
5881 * because it'll correctly read the old or the new value. We don't need
5882 * to guarantee who wins the race as long as it doesn't return garbage.
5884 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5885 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5890 return sched_attr_copy_to_user(uattr, &kattr, usize);
5897 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5899 cpumask_var_t cpus_allowed, new_mask;
5900 struct task_struct *p;
5905 p = find_process_by_pid(pid);
5911 /* Prevent p going away */
5915 if (p->flags & PF_NO_SETAFFINITY) {
5919 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5923 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5925 goto out_free_cpus_allowed;
5928 if (!check_same_owner(p)) {
5930 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5932 goto out_free_new_mask;
5937 retval = security_task_setscheduler(p);
5939 goto out_free_new_mask;
5942 cpuset_cpus_allowed(p, cpus_allowed);
5943 cpumask_and(new_mask, in_mask, cpus_allowed);
5946 * Since bandwidth control happens on root_domain basis,
5947 * if admission test is enabled, we only admit -deadline
5948 * tasks allowed to run on all the CPUs in the task's
5952 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
5954 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
5957 goto out_free_new_mask;
5963 retval = __set_cpus_allowed_ptr(p, new_mask, true);
5966 cpuset_cpus_allowed(p, cpus_allowed);
5967 if (!cpumask_subset(new_mask, cpus_allowed)) {
5969 * We must have raced with a concurrent cpuset
5970 * update. Just reset the cpus_allowed to the
5971 * cpuset's cpus_allowed
5973 cpumask_copy(new_mask, cpus_allowed);
5978 free_cpumask_var(new_mask);
5979 out_free_cpus_allowed:
5980 free_cpumask_var(cpus_allowed);
5986 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5987 struct cpumask *new_mask)
5989 if (len < cpumask_size())
5990 cpumask_clear(new_mask);
5991 else if (len > cpumask_size())
5992 len = cpumask_size();
5994 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5998 * sys_sched_setaffinity - set the CPU affinity of a process
5999 * @pid: pid of the process
6000 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6001 * @user_mask_ptr: user-space pointer to the new CPU mask
6003 * Return: 0 on success. An error code otherwise.
6005 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6006 unsigned long __user *, user_mask_ptr)
6008 cpumask_var_t new_mask;
6011 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6014 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6016 retval = sched_setaffinity(pid, new_mask);
6017 free_cpumask_var(new_mask);
6021 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6023 struct task_struct *p;
6024 unsigned long flags;
6030 p = find_process_by_pid(pid);
6034 retval = security_task_getscheduler(p);
6038 raw_spin_lock_irqsave(&p->pi_lock, flags);
6039 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
6040 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6049 * sys_sched_getaffinity - get the CPU affinity of a process
6050 * @pid: pid of the process
6051 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6052 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6054 * Return: size of CPU mask copied to user_mask_ptr on success. An
6055 * error code otherwise.
6057 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6058 unsigned long __user *, user_mask_ptr)
6063 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6065 if (len & (sizeof(unsigned long)-1))
6068 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6071 ret = sched_getaffinity(pid, mask);
6073 unsigned int retlen = min(len, cpumask_size());
6075 if (copy_to_user(user_mask_ptr, mask, retlen))
6080 free_cpumask_var(mask);
6086 * sys_sched_yield - yield the current processor to other threads.
6088 * This function yields the current CPU to other tasks. If there are no
6089 * other threads running on this CPU then this function will return.
6093 static void do_sched_yield(void)
6098 rq = this_rq_lock_irq(&rf);
6100 schedstat_inc(rq->yld_count);
6101 current->sched_class->yield_task(rq);
6104 * Since we are going to call schedule() anyway, there's
6105 * no need to preempt or enable interrupts:
6109 sched_preempt_enable_no_resched();
6114 SYSCALL_DEFINE0(sched_yield)
6120 #ifndef CONFIG_PREEMPTION
6121 int __sched _cond_resched(void)
6123 if (should_resched(0)) {
6124 preempt_schedule_common();
6130 EXPORT_SYMBOL(_cond_resched);
6134 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6135 * call schedule, and on return reacquire the lock.
6137 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
6138 * operations here to prevent schedule() from being called twice (once via
6139 * spin_unlock(), once by hand).
6141 int __cond_resched_lock(spinlock_t *lock)
6143 int resched = should_resched(PREEMPT_LOCK_OFFSET);
6146 lockdep_assert_held(lock);
6148 if (spin_needbreak(lock) || resched) {
6151 preempt_schedule_common();
6159 EXPORT_SYMBOL(__cond_resched_lock);
6162 * yield - yield the current processor to other threads.
6164 * Do not ever use this function, there's a 99% chance you're doing it wrong.
6166 * The scheduler is at all times free to pick the calling task as the most
6167 * eligible task to run, if removing the yield() call from your code breaks
6168 * it, its already broken.
6170 * Typical broken usage is:
6175 * where one assumes that yield() will let 'the other' process run that will
6176 * make event true. If the current task is a SCHED_FIFO task that will never
6177 * happen. Never use yield() as a progress guarantee!!
6179 * If you want to use yield() to wait for something, use wait_event().
6180 * If you want to use yield() to be 'nice' for others, use cond_resched().
6181 * If you still want to use yield(), do not!
6183 void __sched yield(void)
6185 set_current_state(TASK_RUNNING);
6188 EXPORT_SYMBOL(yield);
6191 * yield_to - yield the current processor to another thread in
6192 * your thread group, or accelerate that thread toward the
6193 * processor it's on.
6195 * @preempt: whether task preemption is allowed or not
6197 * It's the caller's job to ensure that the target task struct
6198 * can't go away on us before we can do any checks.
6201 * true (>0) if we indeed boosted the target task.
6202 * false (0) if we failed to boost the target.
6203 * -ESRCH if there's no task to yield to.
6205 int __sched yield_to(struct task_struct *p, bool preempt)
6207 struct task_struct *curr = current;
6208 struct rq *rq, *p_rq;
6209 unsigned long flags;
6212 local_irq_save(flags);
6218 * If we're the only runnable task on the rq and target rq also
6219 * has only one task, there's absolutely no point in yielding.
6221 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
6226 double_rq_lock(rq, p_rq);
6227 if (task_rq(p) != p_rq) {
6228 double_rq_unlock(rq, p_rq);
6232 if (!curr->sched_class->yield_to_task)
6235 if (curr->sched_class != p->sched_class)
6238 if (task_running(p_rq, p) || p->state)
6241 yielded = curr->sched_class->yield_to_task(rq, p);
6243 schedstat_inc(rq->yld_count);
6245 * Make p's CPU reschedule; pick_next_entity takes care of
6248 if (preempt && rq != p_rq)
6253 double_rq_unlock(rq, p_rq);
6255 local_irq_restore(flags);
6262 EXPORT_SYMBOL_GPL(yield_to);
6264 int io_schedule_prepare(void)
6266 int old_iowait = current->in_iowait;
6268 current->in_iowait = 1;
6269 blk_schedule_flush_plug(current);
6274 void io_schedule_finish(int token)
6276 current->in_iowait = token;
6280 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6281 * that process accounting knows that this is a task in IO wait state.
6283 long __sched io_schedule_timeout(long timeout)
6288 token = io_schedule_prepare();
6289 ret = schedule_timeout(timeout);
6290 io_schedule_finish(token);
6294 EXPORT_SYMBOL(io_schedule_timeout);
6296 void __sched io_schedule(void)
6300 token = io_schedule_prepare();
6302 io_schedule_finish(token);
6304 EXPORT_SYMBOL(io_schedule);
6307 * sys_sched_get_priority_max - return maximum RT priority.
6308 * @policy: scheduling class.
6310 * Return: On success, this syscall returns the maximum
6311 * rt_priority that can be used by a given scheduling class.
6312 * On failure, a negative error code is returned.
6314 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6321 ret = MAX_USER_RT_PRIO-1;
6323 case SCHED_DEADLINE:
6334 * sys_sched_get_priority_min - return minimum RT priority.
6335 * @policy: scheduling class.
6337 * Return: On success, this syscall returns the minimum
6338 * rt_priority that can be used by a given scheduling class.
6339 * On failure, a negative error code is returned.
6341 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6350 case SCHED_DEADLINE:
6359 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
6361 struct task_struct *p;
6362 unsigned int time_slice;
6372 p = find_process_by_pid(pid);
6376 retval = security_task_getscheduler(p);
6380 rq = task_rq_lock(p, &rf);
6382 if (p->sched_class->get_rr_interval)
6383 time_slice = p->sched_class->get_rr_interval(rq, p);
6384 task_rq_unlock(rq, p, &rf);
6387 jiffies_to_timespec64(time_slice, t);
6396 * sys_sched_rr_get_interval - return the default timeslice of a process.
6397 * @pid: pid of the process.
6398 * @interval: userspace pointer to the timeslice value.
6400 * this syscall writes the default timeslice value of a given process
6401 * into the user-space timespec buffer. A value of '0' means infinity.
6403 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6406 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6407 struct __kernel_timespec __user *, interval)
6409 struct timespec64 t;
6410 int retval = sched_rr_get_interval(pid, &t);
6413 retval = put_timespec64(&t, interval);
6418 #ifdef CONFIG_COMPAT_32BIT_TIME
6419 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
6420 struct old_timespec32 __user *, interval)
6422 struct timespec64 t;
6423 int retval = sched_rr_get_interval(pid, &t);
6426 retval = put_old_timespec32(&t, interval);
6431 void sched_show_task(struct task_struct *p)
6433 unsigned long free = 0;
6436 if (!try_get_task_stack(p))
6439 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
6441 if (p->state == TASK_RUNNING)
6442 pr_cont(" running task ");
6443 #ifdef CONFIG_DEBUG_STACK_USAGE
6444 free = stack_not_used(p);
6449 ppid = task_pid_nr(rcu_dereference(p->real_parent));
6451 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
6452 free, task_pid_nr(p), ppid,
6453 (unsigned long)task_thread_info(p)->flags);
6455 print_worker_info(KERN_INFO, p);
6456 show_stack(p, NULL, KERN_INFO);
6459 EXPORT_SYMBOL_GPL(sched_show_task);
6462 state_filter_match(unsigned long state_filter, struct task_struct *p)
6464 /* no filter, everything matches */
6468 /* filter, but doesn't match */
6469 if (!(p->state & state_filter))
6473 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6476 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
6483 void show_state_filter(unsigned long state_filter)
6485 struct task_struct *g, *p;
6488 for_each_process_thread(g, p) {
6490 * reset the NMI-timeout, listing all files on a slow
6491 * console might take a lot of time:
6492 * Also, reset softlockup watchdogs on all CPUs, because
6493 * another CPU might be blocked waiting for us to process
6496 touch_nmi_watchdog();
6497 touch_all_softlockup_watchdogs();
6498 if (state_filter_match(state_filter, p))
6502 #ifdef CONFIG_SCHED_DEBUG
6504 sysrq_sched_debug_show();
6508 * Only show locks if all tasks are dumped:
6511 debug_show_all_locks();
6515 * init_idle - set up an idle thread for a given CPU
6516 * @idle: task in question
6517 * @cpu: CPU the idle task belongs to
6519 * NOTE: this function does not set the idle thread's NEED_RESCHED
6520 * flag, to make booting more robust.
6522 void init_idle(struct task_struct *idle, int cpu)
6524 struct rq *rq = cpu_rq(cpu);
6525 unsigned long flags;
6527 __sched_fork(0, idle);
6529 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6530 raw_spin_lock(&rq->lock);
6532 idle->state = TASK_RUNNING;
6533 idle->se.exec_start = sched_clock();
6534 idle->flags |= PF_IDLE;
6536 scs_task_reset(idle);
6537 kasan_unpoison_task_stack(idle);
6541 * Its possible that init_idle() gets called multiple times on a task,
6542 * in that case do_set_cpus_allowed() will not do the right thing.
6544 * And since this is boot we can forgo the serialization.
6546 set_cpus_allowed_common(idle, cpumask_of(cpu));
6549 * We're having a chicken and egg problem, even though we are
6550 * holding rq->lock, the CPU isn't yet set to this CPU so the
6551 * lockdep check in task_group() will fail.
6553 * Similar case to sched_fork(). / Alternatively we could
6554 * use task_rq_lock() here and obtain the other rq->lock.
6559 __set_task_cpu(idle, cpu);
6563 rcu_assign_pointer(rq->curr, idle);
6564 idle->on_rq = TASK_ON_RQ_QUEUED;
6568 raw_spin_unlock(&rq->lock);
6569 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6571 /* Set the preempt count _outside_ the spinlocks! */
6572 init_idle_preempt_count(idle, cpu);
6575 * The idle tasks have their own, simple scheduling class:
6577 idle->sched_class = &idle_sched_class;
6578 ftrace_graph_init_idle_task(idle, cpu);
6579 vtime_init_idle(idle, cpu);
6581 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6587 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6588 const struct cpumask *trial)
6592 if (!cpumask_weight(cur))
6595 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6600 int task_can_attach(struct task_struct *p,
6601 const struct cpumask *cs_cpus_allowed)
6606 * Kthreads which disallow setaffinity shouldn't be moved
6607 * to a new cpuset; we don't want to change their CPU
6608 * affinity and isolating such threads by their set of
6609 * allowed nodes is unnecessary. Thus, cpusets are not
6610 * applicable for such threads. This prevents checking for
6611 * success of set_cpus_allowed_ptr() on all attached tasks
6612 * before cpus_mask may be changed.
6614 if (p->flags & PF_NO_SETAFFINITY) {
6619 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6621 ret = dl_task_can_attach(p, cs_cpus_allowed);
6627 bool sched_smp_initialized __read_mostly;
6629 #ifdef CONFIG_NUMA_BALANCING
6630 /* Migrate current task p to target_cpu */
6631 int migrate_task_to(struct task_struct *p, int target_cpu)
6633 struct migration_arg arg = { p, target_cpu };
6634 int curr_cpu = task_cpu(p);
6636 if (curr_cpu == target_cpu)
6639 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6642 /* TODO: This is not properly updating schedstats */
6644 trace_sched_move_numa(p, curr_cpu, target_cpu);
6645 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6649 * Requeue a task on a given node and accurately track the number of NUMA
6650 * tasks on the runqueues
6652 void sched_setnuma(struct task_struct *p, int nid)
6654 bool queued, running;
6658 rq = task_rq_lock(p, &rf);
6659 queued = task_on_rq_queued(p);
6660 running = task_current(rq, p);
6663 dequeue_task(rq, p, DEQUEUE_SAVE);
6665 put_prev_task(rq, p);
6667 p->numa_preferred_nid = nid;
6670 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6672 set_next_task(rq, p);
6673 task_rq_unlock(rq, p, &rf);
6675 #endif /* CONFIG_NUMA_BALANCING */
6677 #ifdef CONFIG_HOTPLUG_CPU
6679 * Ensure that the idle task is using init_mm right before its CPU goes
6682 void idle_task_exit(void)
6684 struct mm_struct *mm = current->active_mm;
6686 BUG_ON(cpu_online(smp_processor_id()));
6687 BUG_ON(current != this_rq()->idle);
6689 if (mm != &init_mm) {
6690 switch_mm(mm, &init_mm, current);
6691 finish_arch_post_lock_switch();
6694 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6698 * Since this CPU is going 'away' for a while, fold any nr_active delta
6699 * we might have. Assumes we're called after migrate_tasks() so that the
6700 * nr_active count is stable. We need to take the teardown thread which
6701 * is calling this into account, so we hand in adjust = 1 to the load
6704 * Also see the comment "Global load-average calculations".
6706 static void calc_load_migrate(struct rq *rq)
6708 long delta = calc_load_fold_active(rq, 1);
6710 atomic_long_add(delta, &calc_load_tasks);
6713 static struct task_struct *__pick_migrate_task(struct rq *rq)
6715 const struct sched_class *class;
6716 struct task_struct *next;
6718 for_each_class(class) {
6719 next = class->pick_next_task(rq);
6721 next->sched_class->put_prev_task(rq, next);
6726 /* The idle class should always have a runnable task */
6731 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6732 * try_to_wake_up()->select_task_rq().
6734 * Called with rq->lock held even though we'er in stop_machine() and
6735 * there's no concurrency possible, we hold the required locks anyway
6736 * because of lock validation efforts.
6738 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
6740 struct rq *rq = dead_rq;
6741 struct task_struct *next, *stop = rq->stop;
6742 struct rq_flags orf = *rf;
6746 * Fudge the rq selection such that the below task selection loop
6747 * doesn't get stuck on the currently eligible stop task.
6749 * We're currently inside stop_machine() and the rq is either stuck
6750 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6751 * either way we should never end up calling schedule() until we're
6757 * put_prev_task() and pick_next_task() sched
6758 * class method both need to have an up-to-date
6759 * value of rq->clock[_task]
6761 update_rq_clock(rq);
6765 * There's this thread running, bail when that's the only
6768 if (rq->nr_running == 1)
6771 next = __pick_migrate_task(rq);
6774 * Rules for changing task_struct::cpus_mask are holding
6775 * both pi_lock and rq->lock, such that holding either
6776 * stabilizes the mask.
6778 * Drop rq->lock is not quite as disastrous as it usually is
6779 * because !cpu_active at this point, which means load-balance
6780 * will not interfere. Also, stop-machine.
6783 raw_spin_lock(&next->pi_lock);
6787 * Since we're inside stop-machine, _nothing_ should have
6788 * changed the task, WARN if weird stuff happened, because in
6789 * that case the above rq->lock drop is a fail too.
6791 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
6792 raw_spin_unlock(&next->pi_lock);
6796 /* Find suitable destination for @next, with force if needed. */
6797 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
6798 rq = __migrate_task(rq, rf, next, dest_cpu);
6799 if (rq != dead_rq) {
6805 raw_spin_unlock(&next->pi_lock);
6810 #endif /* CONFIG_HOTPLUG_CPU */
6812 void set_rq_online(struct rq *rq)
6815 const struct sched_class *class;
6817 cpumask_set_cpu(rq->cpu, rq->rd->online);
6820 for_each_class(class) {
6821 if (class->rq_online)
6822 class->rq_online(rq);
6827 void set_rq_offline(struct rq *rq)
6830 const struct sched_class *class;
6832 for_each_class(class) {
6833 if (class->rq_offline)
6834 class->rq_offline(rq);
6837 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6843 * used to mark begin/end of suspend/resume:
6845 static int num_cpus_frozen;
6848 * Update cpusets according to cpu_active mask. If cpusets are
6849 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6850 * around partition_sched_domains().
6852 * If we come here as part of a suspend/resume, don't touch cpusets because we
6853 * want to restore it back to its original state upon resume anyway.
6855 static void cpuset_cpu_active(void)
6857 if (cpuhp_tasks_frozen) {
6859 * num_cpus_frozen tracks how many CPUs are involved in suspend
6860 * resume sequence. As long as this is not the last online
6861 * operation in the resume sequence, just build a single sched
6862 * domain, ignoring cpusets.
6864 partition_sched_domains(1, NULL, NULL);
6865 if (--num_cpus_frozen)
6868 * This is the last CPU online operation. So fall through and
6869 * restore the original sched domains by considering the
6870 * cpuset configurations.
6872 cpuset_force_rebuild();
6874 cpuset_update_active_cpus();
6877 static int cpuset_cpu_inactive(unsigned int cpu)
6879 if (!cpuhp_tasks_frozen) {
6880 if (dl_cpu_busy(cpu))
6882 cpuset_update_active_cpus();
6885 partition_sched_domains(1, NULL, NULL);
6890 int sched_cpu_activate(unsigned int cpu)
6892 struct rq *rq = cpu_rq(cpu);
6895 #ifdef CONFIG_SCHED_SMT
6897 * When going up, increment the number of cores with SMT present.
6899 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6900 static_branch_inc_cpuslocked(&sched_smt_present);
6902 set_cpu_active(cpu, true);
6904 if (sched_smp_initialized) {
6905 sched_domains_numa_masks_set(cpu);
6906 cpuset_cpu_active();
6910 * Put the rq online, if not already. This happens:
6912 * 1) In the early boot process, because we build the real domains
6913 * after all CPUs have been brought up.
6915 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6918 rq_lock_irqsave(rq, &rf);
6920 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6923 rq_unlock_irqrestore(rq, &rf);
6928 int sched_cpu_deactivate(unsigned int cpu)
6932 set_cpu_active(cpu, false);
6934 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6935 * users of this state to go away such that all new such users will
6938 * Do sync before park smpboot threads to take care the rcu boost case.
6942 #ifdef CONFIG_SCHED_SMT
6944 * When going down, decrement the number of cores with SMT present.
6946 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6947 static_branch_dec_cpuslocked(&sched_smt_present);
6950 if (!sched_smp_initialized)
6953 ret = cpuset_cpu_inactive(cpu);
6955 set_cpu_active(cpu, true);
6958 sched_domains_numa_masks_clear(cpu);
6962 static void sched_rq_cpu_starting(unsigned int cpu)
6964 struct rq *rq = cpu_rq(cpu);
6966 rq->calc_load_update = calc_load_update;
6967 update_max_interval();
6970 int sched_cpu_starting(unsigned int cpu)
6972 sched_rq_cpu_starting(cpu);
6973 sched_tick_start(cpu);
6977 #ifdef CONFIG_HOTPLUG_CPU
6978 int sched_cpu_dying(unsigned int cpu)
6980 struct rq *rq = cpu_rq(cpu);
6983 /* Handle pending wakeups and then migrate everything off */
6984 sched_tick_stop(cpu);
6986 rq_lock_irqsave(rq, &rf);
6988 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6991 migrate_tasks(rq, &rf);
6992 BUG_ON(rq->nr_running != 1);
6993 rq_unlock_irqrestore(rq, &rf);
6995 calc_load_migrate(rq);
6996 update_max_interval();
6997 nohz_balance_exit_idle(rq);
7003 void __init sched_init_smp(void)
7008 * There's no userspace yet to cause hotplug operations; hence all the
7009 * CPU masks are stable and all blatant races in the below code cannot
7012 mutex_lock(&sched_domains_mutex);
7013 sched_init_domains(cpu_active_mask);
7014 mutex_unlock(&sched_domains_mutex);
7016 /* Move init over to a non-isolated CPU */
7017 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
7019 sched_init_granularity();
7021 init_sched_rt_class();
7022 init_sched_dl_class();
7024 sched_smp_initialized = true;
7027 static int __init migration_init(void)
7029 sched_cpu_starting(smp_processor_id());
7032 early_initcall(migration_init);
7035 void __init sched_init_smp(void)
7037 sched_init_granularity();
7039 #endif /* CONFIG_SMP */
7041 int in_sched_functions(unsigned long addr)
7043 return in_lock_functions(addr) ||
7044 (addr >= (unsigned long)__sched_text_start
7045 && addr < (unsigned long)__sched_text_end);
7048 #ifdef CONFIG_CGROUP_SCHED
7050 * Default task group.
7051 * Every task in system belongs to this group at bootup.
7053 struct task_group root_task_group;
7054 LIST_HEAD(task_groups);
7056 /* Cacheline aligned slab cache for task_group */
7057 static struct kmem_cache *task_group_cache __read_mostly;
7060 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7061 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7063 void __init sched_init(void)
7065 unsigned long ptr = 0;
7068 /* Make sure the linker didn't screw up */
7069 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
7070 &fair_sched_class + 1 != &rt_sched_class ||
7071 &rt_sched_class + 1 != &dl_sched_class);
7073 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
7078 #ifdef CONFIG_FAIR_GROUP_SCHED
7079 ptr += 2 * nr_cpu_ids * sizeof(void **);
7081 #ifdef CONFIG_RT_GROUP_SCHED
7082 ptr += 2 * nr_cpu_ids * sizeof(void **);
7085 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
7087 #ifdef CONFIG_FAIR_GROUP_SCHED
7088 root_task_group.se = (struct sched_entity **)ptr;
7089 ptr += nr_cpu_ids * sizeof(void **);
7091 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7092 ptr += nr_cpu_ids * sizeof(void **);
7094 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7095 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7096 #endif /* CONFIG_FAIR_GROUP_SCHED */
7097 #ifdef CONFIG_RT_GROUP_SCHED
7098 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7099 ptr += nr_cpu_ids * sizeof(void **);
7101 root_task_group.rt_rq = (struct rt_rq **)ptr;
7102 ptr += nr_cpu_ids * sizeof(void **);
7104 #endif /* CONFIG_RT_GROUP_SCHED */
7106 #ifdef CONFIG_CPUMASK_OFFSTACK
7107 for_each_possible_cpu(i) {
7108 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7109 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7110 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7111 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7113 #endif /* CONFIG_CPUMASK_OFFSTACK */
7115 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
7116 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
7119 init_defrootdomain();
7122 #ifdef CONFIG_RT_GROUP_SCHED
7123 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7124 global_rt_period(), global_rt_runtime());
7125 #endif /* CONFIG_RT_GROUP_SCHED */
7127 #ifdef CONFIG_CGROUP_SCHED
7128 task_group_cache = KMEM_CACHE(task_group, 0);
7130 list_add(&root_task_group.list, &task_groups);
7131 INIT_LIST_HEAD(&root_task_group.children);
7132 INIT_LIST_HEAD(&root_task_group.siblings);
7133 autogroup_init(&init_task);
7134 #endif /* CONFIG_CGROUP_SCHED */
7136 for_each_possible_cpu(i) {
7140 raw_spin_lock_init(&rq->lock);
7142 rq->calc_load_active = 0;
7143 rq->calc_load_update = jiffies + LOAD_FREQ;
7144 init_cfs_rq(&rq->cfs);
7145 init_rt_rq(&rq->rt);
7146 init_dl_rq(&rq->dl);
7147 #ifdef CONFIG_FAIR_GROUP_SCHED
7148 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7149 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7151 * How much CPU bandwidth does root_task_group get?
7153 * In case of task-groups formed thr' the cgroup filesystem, it
7154 * gets 100% of the CPU resources in the system. This overall
7155 * system CPU resource is divided among the tasks of
7156 * root_task_group and its child task-groups in a fair manner,
7157 * based on each entity's (task or task-group's) weight
7158 * (se->load.weight).
7160 * In other words, if root_task_group has 10 tasks of weight
7161 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7162 * then A0's share of the CPU resource is:
7164 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7166 * We achieve this by letting root_task_group's tasks sit
7167 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7169 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7170 #endif /* CONFIG_FAIR_GROUP_SCHED */
7172 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7173 #ifdef CONFIG_RT_GROUP_SCHED
7174 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7179 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7180 rq->balance_callback = NULL;
7181 rq->active_balance = 0;
7182 rq->next_balance = jiffies;
7187 rq->avg_idle = 2*sysctl_sched_migration_cost;
7188 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7190 INIT_LIST_HEAD(&rq->cfs_tasks);
7192 rq_attach_root(rq, &def_root_domain);
7193 #ifdef CONFIG_NO_HZ_COMMON
7194 rq->last_blocked_load_update_tick = jiffies;
7195 atomic_set(&rq->nohz_flags, 0);
7197 rq_csd_init(rq, &rq->nohz_csd, nohz_csd_func);
7199 #endif /* CONFIG_SMP */
7201 atomic_set(&rq->nr_iowait, 0);
7204 set_load_weight(&init_task, false);
7207 * The boot idle thread does lazy MMU switching as well:
7210 enter_lazy_tlb(&init_mm, current);
7213 * Make us the idle thread. Technically, schedule() should not be
7214 * called from this thread, however somewhere below it might be,
7215 * but because we are the idle thread, we just pick up running again
7216 * when this runqueue becomes "idle".
7218 init_idle(current, smp_processor_id());
7220 calc_load_update = jiffies + LOAD_FREQ;
7223 idle_thread_set_boot_cpu();
7225 init_sched_fair_class();
7233 scheduler_running = 1;
7236 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7237 static inline int preempt_count_equals(int preempt_offset)
7239 int nested = preempt_count() + rcu_preempt_depth();
7241 return (nested == preempt_offset);
7244 void __might_sleep(const char *file, int line, int preempt_offset)
7247 * Blocking primitives will set (and therefore destroy) current->state,
7248 * since we will exit with TASK_RUNNING make sure we enter with it,
7249 * otherwise we will destroy state.
7251 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7252 "do not call blocking ops when !TASK_RUNNING; "
7253 "state=%lx set at [<%p>] %pS\n",
7255 (void *)current->task_state_change,
7256 (void *)current->task_state_change);
7258 ___might_sleep(file, line, preempt_offset);
7260 EXPORT_SYMBOL(__might_sleep);
7262 void ___might_sleep(const char *file, int line, int preempt_offset)
7264 /* Ratelimiting timestamp: */
7265 static unsigned long prev_jiffy;
7267 unsigned long preempt_disable_ip;
7269 /* WARN_ON_ONCE() by default, no rate limit required: */
7272 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7273 !is_idle_task(current) && !current->non_block_count) ||
7274 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
7278 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7280 prev_jiffy = jiffies;
7282 /* Save this before calling printk(), since that will clobber it: */
7283 preempt_disable_ip = get_preempt_disable_ip(current);
7286 "BUG: sleeping function called from invalid context at %s:%d\n",
7289 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
7290 in_atomic(), irqs_disabled(), current->non_block_count,
7291 current->pid, current->comm);
7293 if (task_stack_end_corrupted(current))
7294 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7296 debug_show_held_locks(current);
7297 if (irqs_disabled())
7298 print_irqtrace_events(current);
7299 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7300 && !preempt_count_equals(preempt_offset)) {
7301 pr_err("Preemption disabled at:");
7302 print_ip_sym(KERN_ERR, preempt_disable_ip);
7305 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7307 EXPORT_SYMBOL(___might_sleep);
7309 void __cant_sleep(const char *file, int line, int preempt_offset)
7311 static unsigned long prev_jiffy;
7313 if (irqs_disabled())
7316 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
7319 if (preempt_count() > preempt_offset)
7322 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7324 prev_jiffy = jiffies;
7326 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
7327 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7328 in_atomic(), irqs_disabled(),
7329 current->pid, current->comm);
7331 debug_show_held_locks(current);
7333 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7335 EXPORT_SYMBOL_GPL(__cant_sleep);
7338 #ifdef CONFIG_MAGIC_SYSRQ
7339 void normalize_rt_tasks(void)
7341 struct task_struct *g, *p;
7342 struct sched_attr attr = {
7343 .sched_policy = SCHED_NORMAL,
7346 read_lock(&tasklist_lock);
7347 for_each_process_thread(g, p) {
7349 * Only normalize user tasks:
7351 if (p->flags & PF_KTHREAD)
7354 p->se.exec_start = 0;
7355 schedstat_set(p->se.statistics.wait_start, 0);
7356 schedstat_set(p->se.statistics.sleep_start, 0);
7357 schedstat_set(p->se.statistics.block_start, 0);
7359 if (!dl_task(p) && !rt_task(p)) {
7361 * Renice negative nice level userspace
7364 if (task_nice(p) < 0)
7365 set_user_nice(p, 0);
7369 __sched_setscheduler(p, &attr, false, false);
7371 read_unlock(&tasklist_lock);
7374 #endif /* CONFIG_MAGIC_SYSRQ */
7376 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7378 * These functions are only useful for the IA64 MCA handling, or kdb.
7380 * They can only be called when the whole system has been
7381 * stopped - every CPU needs to be quiescent, and no scheduling
7382 * activity can take place. Using them for anything else would
7383 * be a serious bug, and as a result, they aren't even visible
7384 * under any other configuration.
7388 * curr_task - return the current task for a given CPU.
7389 * @cpu: the processor in question.
7391 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7393 * Return: The current task for @cpu.
7395 struct task_struct *curr_task(int cpu)
7397 return cpu_curr(cpu);
7400 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7404 * ia64_set_curr_task - set the current task for a given CPU.
7405 * @cpu: the processor in question.
7406 * @p: the task pointer to set.
7408 * Description: This function must only be used when non-maskable interrupts
7409 * are serviced on a separate stack. It allows the architecture to switch the
7410 * notion of the current task on a CPU in a non-blocking manner. This function
7411 * must be called with all CPU's synchronized, and interrupts disabled, the
7412 * and caller must save the original value of the current task (see
7413 * curr_task() above) and restore that value before reenabling interrupts and
7414 * re-starting the system.
7416 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7418 void ia64_set_curr_task(int cpu, struct task_struct *p)
7425 #ifdef CONFIG_CGROUP_SCHED
7426 /* task_group_lock serializes the addition/removal of task groups */
7427 static DEFINE_SPINLOCK(task_group_lock);
7429 static inline void alloc_uclamp_sched_group(struct task_group *tg,
7430 struct task_group *parent)
7432 #ifdef CONFIG_UCLAMP_TASK_GROUP
7433 enum uclamp_id clamp_id;
7435 for_each_clamp_id(clamp_id) {
7436 uclamp_se_set(&tg->uclamp_req[clamp_id],
7437 uclamp_none(clamp_id), false);
7438 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
7443 static void sched_free_group(struct task_group *tg)
7445 free_fair_sched_group(tg);
7446 free_rt_sched_group(tg);
7448 kmem_cache_free(task_group_cache, tg);
7451 /* allocate runqueue etc for a new task group */
7452 struct task_group *sched_create_group(struct task_group *parent)
7454 struct task_group *tg;
7456 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7458 return ERR_PTR(-ENOMEM);
7460 if (!alloc_fair_sched_group(tg, parent))
7463 if (!alloc_rt_sched_group(tg, parent))
7466 alloc_uclamp_sched_group(tg, parent);
7471 sched_free_group(tg);
7472 return ERR_PTR(-ENOMEM);
7475 void sched_online_group(struct task_group *tg, struct task_group *parent)
7477 unsigned long flags;
7479 spin_lock_irqsave(&task_group_lock, flags);
7480 list_add_rcu(&tg->list, &task_groups);
7482 /* Root should already exist: */
7485 tg->parent = parent;
7486 INIT_LIST_HEAD(&tg->children);
7487 list_add_rcu(&tg->siblings, &parent->children);
7488 spin_unlock_irqrestore(&task_group_lock, flags);
7490 online_fair_sched_group(tg);
7493 /* rcu callback to free various structures associated with a task group */
7494 static void sched_free_group_rcu(struct rcu_head *rhp)
7496 /* Now it should be safe to free those cfs_rqs: */
7497 sched_free_group(container_of(rhp, struct task_group, rcu));
7500 void sched_destroy_group(struct task_group *tg)
7502 /* Wait for possible concurrent references to cfs_rqs complete: */
7503 call_rcu(&tg->rcu, sched_free_group_rcu);
7506 void sched_offline_group(struct task_group *tg)
7508 unsigned long flags;
7510 /* End participation in shares distribution: */
7511 unregister_fair_sched_group(tg);
7513 spin_lock_irqsave(&task_group_lock, flags);
7514 list_del_rcu(&tg->list);
7515 list_del_rcu(&tg->siblings);
7516 spin_unlock_irqrestore(&task_group_lock, flags);
7519 static void sched_change_group(struct task_struct *tsk, int type)
7521 struct task_group *tg;
7524 * All callers are synchronized by task_rq_lock(); we do not use RCU
7525 * which is pointless here. Thus, we pass "true" to task_css_check()
7526 * to prevent lockdep warnings.
7528 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7529 struct task_group, css);
7530 tg = autogroup_task_group(tsk, tg);
7531 tsk->sched_task_group = tg;
7533 #ifdef CONFIG_FAIR_GROUP_SCHED
7534 if (tsk->sched_class->task_change_group)
7535 tsk->sched_class->task_change_group(tsk, type);
7538 set_task_rq(tsk, task_cpu(tsk));
7542 * Change task's runqueue when it moves between groups.
7544 * The caller of this function should have put the task in its new group by
7545 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7548 void sched_move_task(struct task_struct *tsk)
7550 int queued, running, queue_flags =
7551 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7555 rq = task_rq_lock(tsk, &rf);
7556 update_rq_clock(rq);
7558 running = task_current(rq, tsk);
7559 queued = task_on_rq_queued(tsk);
7562 dequeue_task(rq, tsk, queue_flags);
7564 put_prev_task(rq, tsk);
7566 sched_change_group(tsk, TASK_MOVE_GROUP);
7569 enqueue_task(rq, tsk, queue_flags);
7571 set_next_task(rq, tsk);
7573 * After changing group, the running task may have joined a
7574 * throttled one but it's still the running task. Trigger a
7575 * resched to make sure that task can still run.
7580 task_rq_unlock(rq, tsk, &rf);
7583 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7585 return css ? container_of(css, struct task_group, css) : NULL;
7588 static struct cgroup_subsys_state *
7589 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7591 struct task_group *parent = css_tg(parent_css);
7592 struct task_group *tg;
7595 /* This is early initialization for the top cgroup */
7596 return &root_task_group.css;
7599 tg = sched_create_group(parent);
7601 return ERR_PTR(-ENOMEM);
7606 /* Expose task group only after completing cgroup initialization */
7607 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7609 struct task_group *tg = css_tg(css);
7610 struct task_group *parent = css_tg(css->parent);
7613 sched_online_group(tg, parent);
7615 #ifdef CONFIG_UCLAMP_TASK_GROUP
7616 /* Propagate the effective uclamp value for the new group */
7617 cpu_util_update_eff(css);
7623 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7625 struct task_group *tg = css_tg(css);
7627 sched_offline_group(tg);
7630 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7632 struct task_group *tg = css_tg(css);
7635 * Relies on the RCU grace period between css_released() and this.
7637 sched_free_group(tg);
7641 * This is called before wake_up_new_task(), therefore we really only
7642 * have to set its group bits, all the other stuff does not apply.
7644 static void cpu_cgroup_fork(struct task_struct *task)
7649 rq = task_rq_lock(task, &rf);
7651 update_rq_clock(rq);
7652 sched_change_group(task, TASK_SET_GROUP);
7654 task_rq_unlock(rq, task, &rf);
7657 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7659 struct task_struct *task;
7660 struct cgroup_subsys_state *css;
7663 cgroup_taskset_for_each(task, css, tset) {
7664 #ifdef CONFIG_RT_GROUP_SCHED
7665 if (!sched_rt_can_attach(css_tg(css), task))
7669 * Serialize against wake_up_new_task() such that if its
7670 * running, we're sure to observe its full state.
7672 raw_spin_lock_irq(&task->pi_lock);
7674 * Avoid calling sched_move_task() before wake_up_new_task()
7675 * has happened. This would lead to problems with PELT, due to
7676 * move wanting to detach+attach while we're not attached yet.
7678 if (task->state == TASK_NEW)
7680 raw_spin_unlock_irq(&task->pi_lock);
7688 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7690 struct task_struct *task;
7691 struct cgroup_subsys_state *css;
7693 cgroup_taskset_for_each(task, css, tset)
7694 sched_move_task(task);
7697 #ifdef CONFIG_UCLAMP_TASK_GROUP
7698 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7700 struct cgroup_subsys_state *top_css = css;
7701 struct uclamp_se *uc_parent = NULL;
7702 struct uclamp_se *uc_se = NULL;
7703 unsigned int eff[UCLAMP_CNT];
7704 enum uclamp_id clamp_id;
7705 unsigned int clamps;
7707 css_for_each_descendant_pre(css, top_css) {
7708 uc_parent = css_tg(css)->parent
7709 ? css_tg(css)->parent->uclamp : NULL;
7711 for_each_clamp_id(clamp_id) {
7712 /* Assume effective clamps matches requested clamps */
7713 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7714 /* Cap effective clamps with parent's effective clamps */
7716 eff[clamp_id] > uc_parent[clamp_id].value) {
7717 eff[clamp_id] = uc_parent[clamp_id].value;
7720 /* Ensure protection is always capped by limit */
7721 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7723 /* Propagate most restrictive effective clamps */
7725 uc_se = css_tg(css)->uclamp;
7726 for_each_clamp_id(clamp_id) {
7727 if (eff[clamp_id] == uc_se[clamp_id].value)
7729 uc_se[clamp_id].value = eff[clamp_id];
7730 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7731 clamps |= (0x1 << clamp_id);
7734 css = css_rightmost_descendant(css);
7738 /* Immediately update descendants RUNNABLE tasks */
7739 uclamp_update_active_tasks(css, clamps);
7744 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7745 * C expression. Since there is no way to convert a macro argument (N) into a
7746 * character constant, use two levels of macros.
7748 #define _POW10(exp) ((unsigned int)1e##exp)
7749 #define POW10(exp) _POW10(exp)
7751 struct uclamp_request {
7752 #define UCLAMP_PERCENT_SHIFT 2
7753 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7759 static inline struct uclamp_request
7760 capacity_from_percent(char *buf)
7762 struct uclamp_request req = {
7763 .percent = UCLAMP_PERCENT_SCALE,
7764 .util = SCHED_CAPACITY_SCALE,
7769 if (strcmp(buf, "max")) {
7770 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7774 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7779 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7780 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7786 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7787 size_t nbytes, loff_t off,
7788 enum uclamp_id clamp_id)
7790 struct uclamp_request req;
7791 struct task_group *tg;
7793 req = capacity_from_percent(buf);
7797 static_branch_enable(&sched_uclamp_used);
7799 mutex_lock(&uclamp_mutex);
7802 tg = css_tg(of_css(of));
7803 if (tg->uclamp_req[clamp_id].value != req.util)
7804 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7807 * Because of not recoverable conversion rounding we keep track of the
7808 * exact requested value
7810 tg->uclamp_pct[clamp_id] = req.percent;
7812 /* Update effective clamps to track the most restrictive value */
7813 cpu_util_update_eff(of_css(of));
7816 mutex_unlock(&uclamp_mutex);
7821 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7822 char *buf, size_t nbytes,
7825 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7828 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7829 char *buf, size_t nbytes,
7832 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7835 static inline void cpu_uclamp_print(struct seq_file *sf,
7836 enum uclamp_id clamp_id)
7838 struct task_group *tg;
7844 tg = css_tg(seq_css(sf));
7845 util_clamp = tg->uclamp_req[clamp_id].value;
7848 if (util_clamp == SCHED_CAPACITY_SCALE) {
7849 seq_puts(sf, "max\n");
7853 percent = tg->uclamp_pct[clamp_id];
7854 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7855 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7858 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7860 cpu_uclamp_print(sf, UCLAMP_MIN);
7864 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7866 cpu_uclamp_print(sf, UCLAMP_MAX);
7869 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7871 #ifdef CONFIG_FAIR_GROUP_SCHED
7872 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7873 struct cftype *cftype, u64 shareval)
7875 if (shareval > scale_load_down(ULONG_MAX))
7876 shareval = MAX_SHARES;
7877 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7880 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7883 struct task_group *tg = css_tg(css);
7885 return (u64) scale_load_down(tg->shares);
7888 #ifdef CONFIG_CFS_BANDWIDTH
7889 static DEFINE_MUTEX(cfs_constraints_mutex);
7891 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7892 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7893 /* More than 203 days if BW_SHIFT equals 20. */
7894 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
7896 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7898 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7900 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7901 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7903 if (tg == &root_task_group)
7907 * Ensure we have at some amount of bandwidth every period. This is
7908 * to prevent reaching a state of large arrears when throttled via
7909 * entity_tick() resulting in prolonged exit starvation.
7911 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7915 * Likewise, bound things on the otherside by preventing insane quota
7916 * periods. This also allows us to normalize in computing quota
7919 if (period > max_cfs_quota_period)
7923 * Bound quota to defend quota against overflow during bandwidth shift.
7925 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
7929 * Prevent race between setting of cfs_rq->runtime_enabled and
7930 * unthrottle_offline_cfs_rqs().
7933 mutex_lock(&cfs_constraints_mutex);
7934 ret = __cfs_schedulable(tg, period, quota);
7938 runtime_enabled = quota != RUNTIME_INF;
7939 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7941 * If we need to toggle cfs_bandwidth_used, off->on must occur
7942 * before making related changes, and on->off must occur afterwards
7944 if (runtime_enabled && !runtime_was_enabled)
7945 cfs_bandwidth_usage_inc();
7946 raw_spin_lock_irq(&cfs_b->lock);
7947 cfs_b->period = ns_to_ktime(period);
7948 cfs_b->quota = quota;
7950 __refill_cfs_bandwidth_runtime(cfs_b);
7952 /* Restart the period timer (if active) to handle new period expiry: */
7953 if (runtime_enabled)
7954 start_cfs_bandwidth(cfs_b);
7956 raw_spin_unlock_irq(&cfs_b->lock);
7958 for_each_online_cpu(i) {
7959 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7960 struct rq *rq = cfs_rq->rq;
7963 rq_lock_irq(rq, &rf);
7964 cfs_rq->runtime_enabled = runtime_enabled;
7965 cfs_rq->runtime_remaining = 0;
7967 if (cfs_rq->throttled)
7968 unthrottle_cfs_rq(cfs_rq);
7969 rq_unlock_irq(rq, &rf);
7971 if (runtime_was_enabled && !runtime_enabled)
7972 cfs_bandwidth_usage_dec();
7974 mutex_unlock(&cfs_constraints_mutex);
7980 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7984 period = ktime_to_ns(tg->cfs_bandwidth.period);
7985 if (cfs_quota_us < 0)
7986 quota = RUNTIME_INF;
7987 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
7988 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7992 return tg_set_cfs_bandwidth(tg, period, quota);
7995 static long tg_get_cfs_quota(struct task_group *tg)
7999 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8002 quota_us = tg->cfs_bandwidth.quota;
8003 do_div(quota_us, NSEC_PER_USEC);
8008 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8012 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
8015 period = (u64)cfs_period_us * NSEC_PER_USEC;
8016 quota = tg->cfs_bandwidth.quota;
8018 return tg_set_cfs_bandwidth(tg, period, quota);
8021 static long tg_get_cfs_period(struct task_group *tg)
8025 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8026 do_div(cfs_period_us, NSEC_PER_USEC);
8028 return cfs_period_us;
8031 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8034 return tg_get_cfs_quota(css_tg(css));
8037 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8038 struct cftype *cftype, s64 cfs_quota_us)
8040 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8043 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8046 return tg_get_cfs_period(css_tg(css));
8049 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8050 struct cftype *cftype, u64 cfs_period_us)
8052 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8055 struct cfs_schedulable_data {
8056 struct task_group *tg;
8061 * normalize group quota/period to be quota/max_period
8062 * note: units are usecs
8064 static u64 normalize_cfs_quota(struct task_group *tg,
8065 struct cfs_schedulable_data *d)
8073 period = tg_get_cfs_period(tg);
8074 quota = tg_get_cfs_quota(tg);
8077 /* note: these should typically be equivalent */
8078 if (quota == RUNTIME_INF || quota == -1)
8081 return to_ratio(period, quota);
8084 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8086 struct cfs_schedulable_data *d = data;
8087 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8088 s64 quota = 0, parent_quota = -1;
8091 quota = RUNTIME_INF;
8093 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8095 quota = normalize_cfs_quota(tg, d);
8096 parent_quota = parent_b->hierarchical_quota;
8099 * Ensure max(child_quota) <= parent_quota. On cgroup2,
8100 * always take the min. On cgroup1, only inherit when no
8103 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
8104 quota = min(quota, parent_quota);
8106 if (quota == RUNTIME_INF)
8107 quota = parent_quota;
8108 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8112 cfs_b->hierarchical_quota = quota;
8117 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8120 struct cfs_schedulable_data data = {
8126 if (quota != RUNTIME_INF) {
8127 do_div(data.period, NSEC_PER_USEC);
8128 do_div(data.quota, NSEC_PER_USEC);
8132 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8138 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
8140 struct task_group *tg = css_tg(seq_css(sf));
8141 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8143 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8144 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8145 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8147 if (schedstat_enabled() && tg != &root_task_group) {
8151 for_each_possible_cpu(i)
8152 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
8154 seq_printf(sf, "wait_sum %llu\n", ws);
8159 #endif /* CONFIG_CFS_BANDWIDTH */
8160 #endif /* CONFIG_FAIR_GROUP_SCHED */
8162 #ifdef CONFIG_RT_GROUP_SCHED
8163 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8164 struct cftype *cft, s64 val)
8166 return sched_group_set_rt_runtime(css_tg(css), val);
8169 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8172 return sched_group_rt_runtime(css_tg(css));
8175 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8176 struct cftype *cftype, u64 rt_period_us)
8178 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8181 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8184 return sched_group_rt_period(css_tg(css));
8186 #endif /* CONFIG_RT_GROUP_SCHED */
8188 static struct cftype cpu_legacy_files[] = {
8189 #ifdef CONFIG_FAIR_GROUP_SCHED
8192 .read_u64 = cpu_shares_read_u64,
8193 .write_u64 = cpu_shares_write_u64,
8196 #ifdef CONFIG_CFS_BANDWIDTH
8198 .name = "cfs_quota_us",
8199 .read_s64 = cpu_cfs_quota_read_s64,
8200 .write_s64 = cpu_cfs_quota_write_s64,
8203 .name = "cfs_period_us",
8204 .read_u64 = cpu_cfs_period_read_u64,
8205 .write_u64 = cpu_cfs_period_write_u64,
8209 .seq_show = cpu_cfs_stat_show,
8212 #ifdef CONFIG_RT_GROUP_SCHED
8214 .name = "rt_runtime_us",
8215 .read_s64 = cpu_rt_runtime_read,
8216 .write_s64 = cpu_rt_runtime_write,
8219 .name = "rt_period_us",
8220 .read_u64 = cpu_rt_period_read_uint,
8221 .write_u64 = cpu_rt_period_write_uint,
8224 #ifdef CONFIG_UCLAMP_TASK_GROUP
8226 .name = "uclamp.min",
8227 .flags = CFTYPE_NOT_ON_ROOT,
8228 .seq_show = cpu_uclamp_min_show,
8229 .write = cpu_uclamp_min_write,
8232 .name = "uclamp.max",
8233 .flags = CFTYPE_NOT_ON_ROOT,
8234 .seq_show = cpu_uclamp_max_show,
8235 .write = cpu_uclamp_max_write,
8241 static int cpu_extra_stat_show(struct seq_file *sf,
8242 struct cgroup_subsys_state *css)
8244 #ifdef CONFIG_CFS_BANDWIDTH
8246 struct task_group *tg = css_tg(css);
8247 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8250 throttled_usec = cfs_b->throttled_time;
8251 do_div(throttled_usec, NSEC_PER_USEC);
8253 seq_printf(sf, "nr_periods %d\n"
8255 "throttled_usec %llu\n",
8256 cfs_b->nr_periods, cfs_b->nr_throttled,
8263 #ifdef CONFIG_FAIR_GROUP_SCHED
8264 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
8267 struct task_group *tg = css_tg(css);
8268 u64 weight = scale_load_down(tg->shares);
8270 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
8273 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
8274 struct cftype *cft, u64 weight)
8277 * cgroup weight knobs should use the common MIN, DFL and MAX
8278 * values which are 1, 100 and 10000 respectively. While it loses
8279 * a bit of range on both ends, it maps pretty well onto the shares
8280 * value used by scheduler and the round-trip conversions preserve
8281 * the original value over the entire range.
8283 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
8286 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
8288 return sched_group_set_shares(css_tg(css), scale_load(weight));
8291 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
8294 unsigned long weight = scale_load_down(css_tg(css)->shares);
8295 int last_delta = INT_MAX;
8298 /* find the closest nice value to the current weight */
8299 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
8300 delta = abs(sched_prio_to_weight[prio] - weight);
8301 if (delta >= last_delta)
8306 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
8309 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
8310 struct cftype *cft, s64 nice)
8312 unsigned long weight;
8315 if (nice < MIN_NICE || nice > MAX_NICE)
8318 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
8319 idx = array_index_nospec(idx, 40);
8320 weight = sched_prio_to_weight[idx];
8322 return sched_group_set_shares(css_tg(css), scale_load(weight));
8326 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
8327 long period, long quota)
8330 seq_puts(sf, "max");
8332 seq_printf(sf, "%ld", quota);
8334 seq_printf(sf, " %ld\n", period);
8337 /* caller should put the current value in *@periodp before calling */
8338 static int __maybe_unused cpu_period_quota_parse(char *buf,
8339 u64 *periodp, u64 *quotap)
8341 char tok[21]; /* U64_MAX */
8343 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
8346 *periodp *= NSEC_PER_USEC;
8348 if (sscanf(tok, "%llu", quotap))
8349 *quotap *= NSEC_PER_USEC;
8350 else if (!strcmp(tok, "max"))
8351 *quotap = RUNTIME_INF;
8358 #ifdef CONFIG_CFS_BANDWIDTH
8359 static int cpu_max_show(struct seq_file *sf, void *v)
8361 struct task_group *tg = css_tg(seq_css(sf));
8363 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
8367 static ssize_t cpu_max_write(struct kernfs_open_file *of,
8368 char *buf, size_t nbytes, loff_t off)
8370 struct task_group *tg = css_tg(of_css(of));
8371 u64 period = tg_get_cfs_period(tg);
8375 ret = cpu_period_quota_parse(buf, &period, "a);
8377 ret = tg_set_cfs_bandwidth(tg, period, quota);
8378 return ret ?: nbytes;
8382 static struct cftype cpu_files[] = {
8383 #ifdef CONFIG_FAIR_GROUP_SCHED
8386 .flags = CFTYPE_NOT_ON_ROOT,
8387 .read_u64 = cpu_weight_read_u64,
8388 .write_u64 = cpu_weight_write_u64,
8391 .name = "weight.nice",
8392 .flags = CFTYPE_NOT_ON_ROOT,
8393 .read_s64 = cpu_weight_nice_read_s64,
8394 .write_s64 = cpu_weight_nice_write_s64,
8397 #ifdef CONFIG_CFS_BANDWIDTH
8400 .flags = CFTYPE_NOT_ON_ROOT,
8401 .seq_show = cpu_max_show,
8402 .write = cpu_max_write,
8405 #ifdef CONFIG_UCLAMP_TASK_GROUP
8407 .name = "uclamp.min",
8408 .flags = CFTYPE_NOT_ON_ROOT,
8409 .seq_show = cpu_uclamp_min_show,
8410 .write = cpu_uclamp_min_write,
8413 .name = "uclamp.max",
8414 .flags = CFTYPE_NOT_ON_ROOT,
8415 .seq_show = cpu_uclamp_max_show,
8416 .write = cpu_uclamp_max_write,
8422 struct cgroup_subsys cpu_cgrp_subsys = {
8423 .css_alloc = cpu_cgroup_css_alloc,
8424 .css_online = cpu_cgroup_css_online,
8425 .css_released = cpu_cgroup_css_released,
8426 .css_free = cpu_cgroup_css_free,
8427 .css_extra_stat_show = cpu_extra_stat_show,
8428 .fork = cpu_cgroup_fork,
8429 .can_attach = cpu_cgroup_can_attach,
8430 .attach = cpu_cgroup_attach,
8431 .legacy_cftypes = cpu_legacy_files,
8432 .dfl_cftypes = cpu_files,
8437 #endif /* CONFIG_CGROUP_SCHED */
8439 void dump_cpu_task(int cpu)
8441 pr_info("Task dump for CPU %d:\n", cpu);
8442 sched_show_task(cpu_curr(cpu));
8446 * Nice levels are multiplicative, with a gentle 10% change for every
8447 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8448 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8449 * that remained on nice 0.
8451 * The "10% effect" is relative and cumulative: from _any_ nice level,
8452 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8453 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8454 * If a task goes up by ~10% and another task goes down by ~10% then
8455 * the relative distance between them is ~25%.)
8457 const int sched_prio_to_weight[40] = {
8458 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8459 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8460 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8461 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8462 /* 0 */ 1024, 820, 655, 526, 423,
8463 /* 5 */ 335, 272, 215, 172, 137,
8464 /* 10 */ 110, 87, 70, 56, 45,
8465 /* 15 */ 36, 29, 23, 18, 15,
8469 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8471 * In cases where the weight does not change often, we can use the
8472 * precalculated inverse to speed up arithmetics by turning divisions
8473 * into multiplications:
8475 const u32 sched_prio_to_wmult[40] = {
8476 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8477 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8478 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8479 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8480 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8481 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8482 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8483 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8486 void call_trace_sched_update_nr_running(struct rq *rq, int count)
8488 trace_sched_update_nr_running_tp(rq, count);