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;
2506 activate_task(rq, p, en_flags);
2507 ttwu_do_wakeup(rq, p, wake_flags, rf);
2511 * Consider @p being inside a wait loop:
2514 * set_current_state(TASK_UNINTERRUPTIBLE);
2521 * __set_current_state(TASK_RUNNING);
2523 * between set_current_state() and schedule(). In this case @p is still
2524 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
2527 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
2528 * then schedule() must still happen and p->state can be changed to
2529 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
2530 * need to do a full wakeup with enqueue.
2532 * Returns: %true when the wakeup is done,
2535 static int ttwu_runnable(struct task_struct *p, int wake_flags)
2541 rq = __task_rq_lock(p, &rf);
2542 if (task_on_rq_queued(p)) {
2543 /* check_preempt_curr() may use rq clock */
2544 update_rq_clock(rq);
2545 ttwu_do_wakeup(rq, p, wake_flags, &rf);
2548 __task_rq_unlock(rq, &rf);
2554 void sched_ttwu_pending(void *arg)
2556 struct llist_node *llist = arg;
2557 struct rq *rq = this_rq();
2558 struct task_struct *p, *t;
2565 * rq::ttwu_pending racy indication of out-standing wakeups.
2566 * Races such that false-negatives are possible, since they
2567 * are shorter lived that false-positives would be.
2569 WRITE_ONCE(rq->ttwu_pending, 0);
2571 rq_lock_irqsave(rq, &rf);
2572 update_rq_clock(rq);
2574 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
2575 if (WARN_ON_ONCE(p->on_cpu))
2576 smp_cond_load_acquire(&p->on_cpu, !VAL);
2578 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
2579 set_task_cpu(p, cpu_of(rq));
2581 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
2584 rq_unlock_irqrestore(rq, &rf);
2587 void send_call_function_single_ipi(int cpu)
2589 struct rq *rq = cpu_rq(cpu);
2591 if (!set_nr_if_polling(rq->idle))
2592 arch_send_call_function_single_ipi(cpu);
2594 trace_sched_wake_idle_without_ipi(cpu);
2598 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
2599 * necessary. The wakee CPU on receipt of the IPI will queue the task
2600 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
2601 * of the wakeup instead of the waker.
2603 static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2605 struct rq *rq = cpu_rq(cpu);
2607 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
2609 WRITE_ONCE(rq->ttwu_pending, 1);
2610 __smp_call_single_queue(cpu, &p->wake_entry.llist);
2613 void wake_up_if_idle(int cpu)
2615 struct rq *rq = cpu_rq(cpu);
2620 if (!is_idle_task(rcu_dereference(rq->curr)))
2623 if (set_nr_if_polling(rq->idle)) {
2624 trace_sched_wake_idle_without_ipi(cpu);
2626 rq_lock_irqsave(rq, &rf);
2627 if (is_idle_task(rq->curr))
2628 smp_send_reschedule(cpu);
2629 /* Else CPU is not idle, do nothing here: */
2630 rq_unlock_irqrestore(rq, &rf);
2637 bool cpus_share_cache(int this_cpu, int that_cpu)
2639 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2642 static inline bool ttwu_queue_cond(int cpu, int wake_flags)
2645 * If the CPU does not share cache, then queue the task on the
2646 * remote rqs wakelist to avoid accessing remote data.
2648 if (!cpus_share_cache(smp_processor_id(), cpu))
2652 * If the task is descheduling and the only running task on the
2653 * CPU then use the wakelist to offload the task activation to
2654 * the soon-to-be-idle CPU as the current CPU is likely busy.
2655 * nr_running is checked to avoid unnecessary task stacking.
2657 if ((wake_flags & WF_ON_CPU) && cpu_rq(cpu)->nr_running <= 1)
2663 static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2665 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
2666 if (WARN_ON_ONCE(cpu == smp_processor_id()))
2669 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
2670 __ttwu_queue_wakelist(p, cpu, wake_flags);
2677 #else /* !CONFIG_SMP */
2679 static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
2684 #endif /* CONFIG_SMP */
2686 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
2688 struct rq *rq = cpu_rq(cpu);
2691 if (ttwu_queue_wakelist(p, cpu, wake_flags))
2695 update_rq_clock(rq);
2696 ttwu_do_activate(rq, p, wake_flags, &rf);
2701 * Notes on Program-Order guarantees on SMP systems.
2705 * The basic program-order guarantee on SMP systems is that when a task [t]
2706 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
2707 * execution on its new CPU [c1].
2709 * For migration (of runnable tasks) this is provided by the following means:
2711 * A) UNLOCK of the rq(c0)->lock scheduling out task t
2712 * B) migration for t is required to synchronize *both* rq(c0)->lock and
2713 * rq(c1)->lock (if not at the same time, then in that order).
2714 * C) LOCK of the rq(c1)->lock scheduling in task
2716 * Release/acquire chaining guarantees that B happens after A and C after B.
2717 * Note: the CPU doing B need not be c0 or c1
2726 * UNLOCK rq(0)->lock
2728 * LOCK rq(0)->lock // orders against CPU0
2730 * UNLOCK rq(0)->lock
2734 * UNLOCK rq(1)->lock
2736 * LOCK rq(1)->lock // orders against CPU2
2739 * UNLOCK rq(1)->lock
2742 * BLOCKING -- aka. SLEEP + WAKEUP
2744 * For blocking we (obviously) need to provide the same guarantee as for
2745 * migration. However the means are completely different as there is no lock
2746 * chain to provide order. Instead we do:
2748 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
2749 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
2753 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
2755 * LOCK rq(0)->lock LOCK X->pi_lock
2758 * smp_store_release(X->on_cpu, 0);
2760 * smp_cond_load_acquire(&X->on_cpu, !VAL);
2766 * X->state = RUNNING
2767 * UNLOCK rq(2)->lock
2769 * LOCK rq(2)->lock // orders against CPU1
2772 * UNLOCK rq(2)->lock
2775 * UNLOCK rq(0)->lock
2778 * However, for wakeups there is a second guarantee we must provide, namely we
2779 * must ensure that CONDITION=1 done by the caller can not be reordered with
2780 * accesses to the task state; see try_to_wake_up() and set_current_state().
2784 * try_to_wake_up - wake up a thread
2785 * @p: the thread to be awakened
2786 * @state: the mask of task states that can be woken
2787 * @wake_flags: wake modifier flags (WF_*)
2789 * Conceptually does:
2791 * If (@state & @p->state) @p->state = TASK_RUNNING.
2793 * If the task was not queued/runnable, also place it back on a runqueue.
2795 * This function is atomic against schedule() which would dequeue the task.
2797 * It issues a full memory barrier before accessing @p->state, see the comment
2798 * with set_current_state().
2800 * Uses p->pi_lock to serialize against concurrent wake-ups.
2802 * Relies on p->pi_lock stabilizing:
2805 * - p->sched_task_group
2806 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
2808 * Tries really hard to only take one task_rq(p)->lock for performance.
2809 * Takes rq->lock in:
2810 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
2811 * - ttwu_queue() -- new rq, for enqueue of the task;
2812 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
2814 * As a consequence we race really badly with just about everything. See the
2815 * many memory barriers and their comments for details.
2817 * Return: %true if @p->state changes (an actual wakeup was done),
2821 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2823 unsigned long flags;
2824 int cpu, success = 0;
2829 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
2830 * == smp_processor_id()'. Together this means we can special
2831 * case the whole 'p->on_rq && ttwu_runnable()' case below
2832 * without taking any locks.
2835 * - we rely on Program-Order guarantees for all the ordering,
2836 * - we're serialized against set_special_state() by virtue of
2837 * it disabling IRQs (this allows not taking ->pi_lock).
2839 if (!(p->state & state))
2843 trace_sched_waking(p);
2844 p->state = TASK_RUNNING;
2845 trace_sched_wakeup(p);
2850 * If we are going to wake up a thread waiting for CONDITION we
2851 * need to ensure that CONDITION=1 done by the caller can not be
2852 * reordered with p->state check below. This pairs with smp_store_mb()
2853 * in set_current_state() that the waiting thread does.
2855 raw_spin_lock_irqsave(&p->pi_lock, flags);
2856 smp_mb__after_spinlock();
2857 if (!(p->state & state))
2860 trace_sched_waking(p);
2862 /* We're going to change ->state: */
2866 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2867 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2868 * in smp_cond_load_acquire() below.
2870 * sched_ttwu_pending() try_to_wake_up()
2871 * STORE p->on_rq = 1 LOAD p->state
2874 * __schedule() (switch to task 'p')
2875 * LOCK rq->lock smp_rmb();
2876 * smp_mb__after_spinlock();
2880 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2882 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2883 * __schedule(). See the comment for smp_mb__after_spinlock().
2885 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
2888 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
2892 delayacct_blkio_end(p);
2893 atomic_dec(&task_rq(p)->nr_iowait);
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) {
2966 wake_flags |= WF_MIGRATED;
2967 psi_ttwu_dequeue(p);
2968 set_task_cpu(p, cpu);
2972 #endif /* CONFIG_SMP */
2974 ttwu_queue(p, cpu, wake_flags);
2976 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2979 ttwu_stat(p, task_cpu(p), wake_flags);
2986 * try_invoke_on_locked_down_task - Invoke a function on task in fixed state
2987 * @p: Process for which the function is to be invoked.
2988 * @func: Function to invoke.
2989 * @arg: Argument to function.
2991 * If the specified task can be quickly locked into a definite state
2992 * (either sleeping or on a given runqueue), arrange to keep it in that
2993 * state while invoking @func(@arg). This function can use ->on_rq and
2994 * task_curr() to work out what the state is, if required. Given that
2995 * @func can be invoked with a runqueue lock held, it had better be quite
2999 * @false if the task slipped out from under the locks.
3000 * @true if the task was locked onto a runqueue or is sleeping.
3001 * However, @func can override this by returning @false.
3003 bool try_invoke_on_locked_down_task(struct task_struct *p, bool (*func)(struct task_struct *t, void *arg), void *arg)
3009 lockdep_assert_irqs_enabled();
3010 raw_spin_lock_irq(&p->pi_lock);
3012 rq = __task_rq_lock(p, &rf);
3013 if (task_rq(p) == rq)
3022 smp_rmb(); // See smp_rmb() comment in try_to_wake_up().
3027 raw_spin_unlock_irq(&p->pi_lock);
3032 * wake_up_process - Wake up a specific process
3033 * @p: The process to be woken up.
3035 * Attempt to wake up the nominated process and move it to the set of runnable
3038 * Return: 1 if the process was woken up, 0 if it was already running.
3040 * This function executes a full memory barrier before accessing the task state.
3042 int wake_up_process(struct task_struct *p)
3044 return try_to_wake_up(p, TASK_NORMAL, 0);
3046 EXPORT_SYMBOL(wake_up_process);
3048 int wake_up_state(struct task_struct *p, unsigned int state)
3050 return try_to_wake_up(p, state, 0);
3054 * Perform scheduler related setup for a newly forked process p.
3055 * p is forked by current.
3057 * __sched_fork() is basic setup used by init_idle() too:
3059 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
3064 p->se.exec_start = 0;
3065 p->se.sum_exec_runtime = 0;
3066 p->se.prev_sum_exec_runtime = 0;
3067 p->se.nr_migrations = 0;
3069 INIT_LIST_HEAD(&p->se.group_node);
3071 #ifdef CONFIG_FAIR_GROUP_SCHED
3072 p->se.cfs_rq = NULL;
3075 #ifdef CONFIG_SCHEDSTATS
3076 /* Even if schedstat is disabled, there should not be garbage */
3077 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
3080 RB_CLEAR_NODE(&p->dl.rb_node);
3081 init_dl_task_timer(&p->dl);
3082 init_dl_inactive_task_timer(&p->dl);
3083 __dl_clear_params(p);
3085 INIT_LIST_HEAD(&p->rt.run_list);
3087 p->rt.time_slice = sched_rr_timeslice;
3091 #ifdef CONFIG_PREEMPT_NOTIFIERS
3092 INIT_HLIST_HEAD(&p->preempt_notifiers);
3095 #ifdef CONFIG_COMPACTION
3096 p->capture_control = NULL;
3098 init_numa_balancing(clone_flags, p);
3100 p->wake_entry.u_flags = CSD_TYPE_TTWU;
3104 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
3106 #ifdef CONFIG_NUMA_BALANCING
3108 void set_numabalancing_state(bool enabled)
3111 static_branch_enable(&sched_numa_balancing);
3113 static_branch_disable(&sched_numa_balancing);
3116 #ifdef CONFIG_PROC_SYSCTL
3117 int sysctl_numa_balancing(struct ctl_table *table, int write,
3118 void *buffer, size_t *lenp, loff_t *ppos)
3122 int state = static_branch_likely(&sched_numa_balancing);
3124 if (write && !capable(CAP_SYS_ADMIN))
3129 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3133 set_numabalancing_state(state);
3139 #ifdef CONFIG_SCHEDSTATS
3141 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
3142 static bool __initdata __sched_schedstats = false;
3144 static void set_schedstats(bool enabled)
3147 static_branch_enable(&sched_schedstats);
3149 static_branch_disable(&sched_schedstats);
3152 void force_schedstat_enabled(void)
3154 if (!schedstat_enabled()) {
3155 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
3156 static_branch_enable(&sched_schedstats);
3160 static int __init setup_schedstats(char *str)
3167 * This code is called before jump labels have been set up, so we can't
3168 * change the static branch directly just yet. Instead set a temporary
3169 * variable so init_schedstats() can do it later.
3171 if (!strcmp(str, "enable")) {
3172 __sched_schedstats = true;
3174 } else if (!strcmp(str, "disable")) {
3175 __sched_schedstats = false;
3180 pr_warn("Unable to parse schedstats=\n");
3184 __setup("schedstats=", setup_schedstats);
3186 static void __init init_schedstats(void)
3188 set_schedstats(__sched_schedstats);
3191 #ifdef CONFIG_PROC_SYSCTL
3192 int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
3193 size_t *lenp, loff_t *ppos)
3197 int state = static_branch_likely(&sched_schedstats);
3199 if (write && !capable(CAP_SYS_ADMIN))
3204 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
3208 set_schedstats(state);
3211 #endif /* CONFIG_PROC_SYSCTL */
3212 #else /* !CONFIG_SCHEDSTATS */
3213 static inline void init_schedstats(void) {}
3214 #endif /* CONFIG_SCHEDSTATS */
3217 * fork()/clone()-time setup:
3219 int sched_fork(unsigned long clone_flags, struct task_struct *p)
3221 unsigned long flags;
3223 __sched_fork(clone_flags, p);
3225 * We mark the process as NEW here. This guarantees that
3226 * nobody will actually run it, and a signal or other external
3227 * event cannot wake it up and insert it on the runqueue either.
3229 p->state = TASK_NEW;
3232 * Make sure we do not leak PI boosting priority to the child.
3234 p->prio = current->normal_prio;
3239 * Revert to default priority/policy on fork if requested.
3241 if (unlikely(p->sched_reset_on_fork)) {
3242 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3243 p->policy = SCHED_NORMAL;
3244 p->static_prio = NICE_TO_PRIO(0);
3246 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3247 p->static_prio = NICE_TO_PRIO(0);
3249 p->prio = p->normal_prio = __normal_prio(p);
3250 set_load_weight(p, false);
3253 * We don't need the reset flag anymore after the fork. It has
3254 * fulfilled its duty:
3256 p->sched_reset_on_fork = 0;
3259 if (dl_prio(p->prio))
3261 else if (rt_prio(p->prio))
3262 p->sched_class = &rt_sched_class;
3264 p->sched_class = &fair_sched_class;
3266 init_entity_runnable_average(&p->se);
3269 * The child is not yet in the pid-hash so no cgroup attach races,
3270 * and the cgroup is pinned to this child due to cgroup_fork()
3271 * is ran before sched_fork().
3273 * Silence PROVE_RCU.
3275 raw_spin_lock_irqsave(&p->pi_lock, flags);
3278 * We're setting the CPU for the first time, we don't migrate,
3279 * so use __set_task_cpu().
3281 __set_task_cpu(p, smp_processor_id());
3282 if (p->sched_class->task_fork)
3283 p->sched_class->task_fork(p);
3284 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3286 #ifdef CONFIG_SCHED_INFO
3287 if (likely(sched_info_on()))
3288 memset(&p->sched_info, 0, sizeof(p->sched_info));
3290 #if defined(CONFIG_SMP)
3293 init_task_preempt_count(p);
3295 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3296 RB_CLEAR_NODE(&p->pushable_dl_tasks);
3301 void sched_post_fork(struct task_struct *p)
3303 uclamp_post_fork(p);
3306 unsigned long to_ratio(u64 period, u64 runtime)
3308 if (runtime == RUNTIME_INF)
3312 * Doing this here saves a lot of checks in all
3313 * the calling paths, and returning zero seems
3314 * safe for them anyway.
3319 return div64_u64(runtime << BW_SHIFT, period);
3323 * wake_up_new_task - wake up a newly created task for the first time.
3325 * This function will do some initial scheduler statistics housekeeping
3326 * that must be done for every newly created context, then puts the task
3327 * on the runqueue and wakes it.
3329 void wake_up_new_task(struct task_struct *p)
3334 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
3335 p->state = TASK_RUNNING;
3338 * Fork balancing, do it here and not earlier because:
3339 * - cpus_ptr can change in the fork path
3340 * - any previously selected CPU might disappear through hotplug
3342 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
3343 * as we're not fully set-up yet.
3345 p->recent_used_cpu = task_cpu(p);
3347 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
3349 rq = __task_rq_lock(p, &rf);
3350 update_rq_clock(rq);
3351 post_init_entity_util_avg(p);
3353 activate_task(rq, p, ENQUEUE_NOCLOCK);
3354 trace_sched_wakeup_new(p);
3355 check_preempt_curr(rq, p, WF_FORK);
3357 if (p->sched_class->task_woken) {
3359 * Nothing relies on rq->lock after this, so its fine to
3362 rq_unpin_lock(rq, &rf);
3363 p->sched_class->task_woken(rq, p);
3364 rq_repin_lock(rq, &rf);
3367 task_rq_unlock(rq, p, &rf);
3370 #ifdef CONFIG_PREEMPT_NOTIFIERS
3372 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
3374 void preempt_notifier_inc(void)
3376 static_branch_inc(&preempt_notifier_key);
3378 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
3380 void preempt_notifier_dec(void)
3382 static_branch_dec(&preempt_notifier_key);
3384 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
3387 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3388 * @notifier: notifier struct to register
3390 void preempt_notifier_register(struct preempt_notifier *notifier)
3392 if (!static_branch_unlikely(&preempt_notifier_key))
3393 WARN(1, "registering preempt_notifier while notifiers disabled\n");
3395 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3397 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3400 * preempt_notifier_unregister - no longer interested in preemption notifications
3401 * @notifier: notifier struct to unregister
3403 * This is *not* safe to call from within a preemption notifier.
3405 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3407 hlist_del(¬ifier->link);
3409 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3411 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
3413 struct preempt_notifier *notifier;
3415 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3416 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3419 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3421 if (static_branch_unlikely(&preempt_notifier_key))
3422 __fire_sched_in_preempt_notifiers(curr);
3426 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
3427 struct task_struct *next)
3429 struct preempt_notifier *notifier;
3431 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
3432 notifier->ops->sched_out(notifier, next);
3435 static __always_inline void
3436 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3437 struct task_struct *next)
3439 if (static_branch_unlikely(&preempt_notifier_key))
3440 __fire_sched_out_preempt_notifiers(curr, next);
3443 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3445 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3450 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3451 struct task_struct *next)
3455 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3457 static inline void prepare_task(struct task_struct *next)
3461 * Claim the task as running, we do this before switching to it
3462 * such that any running task will have this set.
3464 * See the ttwu() WF_ON_CPU case and its ordering comment.
3466 WRITE_ONCE(next->on_cpu, 1);
3470 static inline void finish_task(struct task_struct *prev)
3474 * This must be the very last reference to @prev from this CPU. After
3475 * p->on_cpu is cleared, the task can be moved to a different CPU. We
3476 * must ensure this doesn't happen until the switch is completely
3479 * In particular, the load of prev->state in finish_task_switch() must
3480 * happen before this.
3482 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
3484 smp_store_release(&prev->on_cpu, 0);
3490 static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
3492 void (*func)(struct rq *rq);
3493 struct callback_head *next;
3495 lockdep_assert_held(&rq->lock);
3498 func = (void (*)(struct rq *))head->func;
3507 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
3509 struct callback_head *head = rq->balance_callback;
3511 lockdep_assert_held(&rq->lock);
3513 rq->balance_callback = NULL;
3514 rq->balance_flags &= ~BALANCE_WORK;
3520 static void __balance_callbacks(struct rq *rq)
3522 do_balance_callbacks(rq, splice_balance_callbacks(rq));
3525 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
3527 unsigned long flags;
3529 if (unlikely(head)) {
3530 raw_spin_lock_irqsave(&rq->lock, flags);
3531 do_balance_callbacks(rq, head);
3532 raw_spin_unlock_irqrestore(&rq->lock, flags);
3536 static void balance_push(struct rq *rq);
3538 static inline void balance_switch(struct rq *rq)
3540 if (likely(!rq->balance_flags))
3543 if (rq->balance_flags & BALANCE_PUSH) {
3548 __balance_callbacks(rq);
3553 static inline void __balance_callbacks(struct rq *rq)
3557 static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
3562 static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
3566 static inline void balance_switch(struct rq *rq)
3573 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
3576 * Since the runqueue lock will be released by the next
3577 * task (which is an invalid locking op but in the case
3578 * of the scheduler it's an obvious special-case), so we
3579 * do an early lockdep release here:
3581 rq_unpin_lock(rq, rf);
3582 spin_release(&rq->lock.dep_map, _THIS_IP_);
3583 #ifdef CONFIG_DEBUG_SPINLOCK
3584 /* this is a valid case when another task releases the spinlock */
3585 rq->lock.owner = next;
3589 static inline void finish_lock_switch(struct rq *rq)
3592 * If we are tracking spinlock dependencies then we have to
3593 * fix up the runqueue lock - which gets 'carried over' from
3594 * prev into current:
3596 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
3598 raw_spin_unlock_irq(&rq->lock);
3602 * NOP if the arch has not defined these:
3605 #ifndef prepare_arch_switch
3606 # define prepare_arch_switch(next) do { } while (0)
3609 #ifndef finish_arch_post_lock_switch
3610 # define finish_arch_post_lock_switch() do { } while (0)
3614 * prepare_task_switch - prepare to switch tasks
3615 * @rq: the runqueue preparing to switch
3616 * @prev: the current task that is being switched out
3617 * @next: the task we are going to switch to.
3619 * This is called with the rq lock held and interrupts off. It must
3620 * be paired with a subsequent finish_task_switch after the context
3623 * prepare_task_switch sets up locking and calls architecture specific
3627 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3628 struct task_struct *next)
3630 kcov_prepare_switch(prev);
3631 sched_info_switch(rq, prev, next);
3632 perf_event_task_sched_out(prev, next);
3634 fire_sched_out_preempt_notifiers(prev, next);
3636 prepare_arch_switch(next);
3640 * finish_task_switch - clean up after a task-switch
3641 * @prev: the thread we just switched away from.
3643 * finish_task_switch must be called after the context switch, paired
3644 * with a prepare_task_switch call before the context switch.
3645 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3646 * and do any other architecture-specific cleanup actions.
3648 * Note that we may have delayed dropping an mm in context_switch(). If
3649 * so, we finish that here outside of the runqueue lock. (Doing it
3650 * with the lock held can cause deadlocks; see schedule() for
3653 * The context switch have flipped the stack from under us and restored the
3654 * local variables which were saved when this task called schedule() in the
3655 * past. prev == current is still correct but we need to recalculate this_rq
3656 * because prev may have moved to another CPU.
3658 static struct rq *finish_task_switch(struct task_struct *prev)
3659 __releases(rq->lock)
3661 struct rq *rq = this_rq();
3662 struct mm_struct *mm = rq->prev_mm;
3666 * The previous task will have left us with a preempt_count of 2
3667 * because it left us after:
3670 * preempt_disable(); // 1
3672 * raw_spin_lock_irq(&rq->lock) // 2
3674 * Also, see FORK_PREEMPT_COUNT.
3676 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
3677 "corrupted preempt_count: %s/%d/0x%x\n",
3678 current->comm, current->pid, preempt_count()))
3679 preempt_count_set(FORK_PREEMPT_COUNT);
3684 * A task struct has one reference for the use as "current".
3685 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3686 * schedule one last time. The schedule call will never return, and
3687 * the scheduled task must drop that reference.
3689 * We must observe prev->state before clearing prev->on_cpu (in
3690 * finish_task), otherwise a concurrent wakeup can get prev
3691 * running on another CPU and we could rave with its RUNNING -> DEAD
3692 * transition, resulting in a double drop.
3694 prev_state = prev->state;
3695 vtime_task_switch(prev);
3696 perf_event_task_sched_in(prev, current);
3698 finish_lock_switch(rq);
3699 finish_arch_post_lock_switch();
3700 kcov_finish_switch(current);
3702 fire_sched_in_preempt_notifiers(current);
3704 * When switching through a kernel thread, the loop in
3705 * membarrier_{private,global}_expedited() may have observed that
3706 * kernel thread and not issued an IPI. It is therefore possible to
3707 * schedule between user->kernel->user threads without passing though
3708 * switch_mm(). Membarrier requires a barrier after storing to
3709 * rq->curr, before returning to userspace, so provide them here:
3711 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
3712 * provided by mmdrop(),
3713 * - a sync_core for SYNC_CORE.
3716 membarrier_mm_sync_core_before_usermode(mm);
3719 if (unlikely(prev_state == TASK_DEAD)) {
3720 if (prev->sched_class->task_dead)
3721 prev->sched_class->task_dead(prev);
3724 * Remove function-return probe instances associated with this
3725 * task and put them back on the free list.
3727 kprobe_flush_task(prev);
3729 /* Task is done with its stack. */
3730 put_task_stack(prev);
3732 put_task_struct_rcu_user(prev);
3735 tick_nohz_task_switch();
3740 * schedule_tail - first thing a freshly forked thread must call.
3741 * @prev: the thread we just switched away from.
3743 asmlinkage __visible void schedule_tail(struct task_struct *prev)
3744 __releases(rq->lock)
3749 * New tasks start with FORK_PREEMPT_COUNT, see there and
3750 * finish_task_switch() for details.
3752 * finish_task_switch() will drop rq->lock() and lower preempt_count
3753 * and the preempt_enable() will end up enabling preemption (on
3754 * PREEMPT_COUNT kernels).
3757 rq = finish_task_switch(prev);
3760 if (current->set_child_tid)
3761 put_user(task_pid_vnr(current), current->set_child_tid);
3763 calculate_sigpending();
3767 * context_switch - switch to the new MM and the new thread's register state.
3769 static __always_inline struct rq *
3770 context_switch(struct rq *rq, struct task_struct *prev,
3771 struct task_struct *next, struct rq_flags *rf)
3773 prepare_task_switch(rq, prev, next);
3776 * For paravirt, this is coupled with an exit in switch_to to
3777 * combine the page table reload and the switch backend into
3780 arch_start_context_switch(prev);
3783 * kernel -> kernel lazy + transfer active
3784 * user -> kernel lazy + mmgrab() active
3786 * kernel -> user switch + mmdrop() active
3787 * user -> user switch
3789 if (!next->mm) { // to kernel
3790 enter_lazy_tlb(prev->active_mm, next);
3792 next->active_mm = prev->active_mm;
3793 if (prev->mm) // from user
3794 mmgrab(prev->active_mm);
3796 prev->active_mm = NULL;
3798 membarrier_switch_mm(rq, prev->active_mm, next->mm);
3800 * sys_membarrier() requires an smp_mb() between setting
3801 * rq->curr / membarrier_switch_mm() and returning to userspace.
3803 * The below provides this either through switch_mm(), or in
3804 * case 'prev->active_mm == next->mm' through
3805 * finish_task_switch()'s mmdrop().
3807 switch_mm_irqs_off(prev->active_mm, next->mm, next);
3809 if (!prev->mm) { // from kernel
3810 /* will mmdrop() in finish_task_switch(). */
3811 rq->prev_mm = prev->active_mm;
3812 prev->active_mm = NULL;
3816 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3818 prepare_lock_switch(rq, next, rf);
3820 /* Here we just switch the register state and the stack. */
3821 switch_to(prev, next, prev);
3824 return finish_task_switch(prev);
3828 * nr_running and nr_context_switches:
3830 * externally visible scheduler statistics: current number of runnable
3831 * threads, total number of context switches performed since bootup.
3833 unsigned long nr_running(void)
3835 unsigned long i, sum = 0;
3837 for_each_online_cpu(i)
3838 sum += cpu_rq(i)->nr_running;
3844 * Check if only the current task is running on the CPU.
3846 * Caution: this function does not check that the caller has disabled
3847 * preemption, thus the result might have a time-of-check-to-time-of-use
3848 * race. The caller is responsible to use it correctly, for example:
3850 * - from a non-preemptible section (of course)
3852 * - from a thread that is bound to a single CPU
3854 * - in a loop with very short iterations (e.g. a polling loop)
3856 bool single_task_running(void)
3858 return raw_rq()->nr_running == 1;
3860 EXPORT_SYMBOL(single_task_running);
3862 unsigned long long nr_context_switches(void)
3865 unsigned long long sum = 0;
3867 for_each_possible_cpu(i)
3868 sum += cpu_rq(i)->nr_switches;
3874 * Consumers of these two interfaces, like for example the cpuidle menu
3875 * governor, are using nonsensical data. Preferring shallow idle state selection
3876 * for a CPU that has IO-wait which might not even end up running the task when
3877 * it does become runnable.
3880 unsigned long nr_iowait_cpu(int cpu)
3882 return atomic_read(&cpu_rq(cpu)->nr_iowait);
3886 * IO-wait accounting, and how its mostly bollocks (on SMP).
3888 * The idea behind IO-wait account is to account the idle time that we could
3889 * have spend running if it were not for IO. That is, if we were to improve the
3890 * storage performance, we'd have a proportional reduction in IO-wait time.
3892 * This all works nicely on UP, where, when a task blocks on IO, we account
3893 * idle time as IO-wait, because if the storage were faster, it could've been
3894 * running and we'd not be idle.
3896 * This has been extended to SMP, by doing the same for each CPU. This however
3899 * Imagine for instance the case where two tasks block on one CPU, only the one
3900 * CPU will have IO-wait accounted, while the other has regular idle. Even
3901 * though, if the storage were faster, both could've ran at the same time,
3902 * utilising both CPUs.
3904 * This means, that when looking globally, the current IO-wait accounting on
3905 * SMP is a lower bound, by reason of under accounting.
3907 * Worse, since the numbers are provided per CPU, they are sometimes
3908 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3909 * associated with any one particular CPU, it can wake to another CPU than it
3910 * blocked on. This means the per CPU IO-wait number is meaningless.
3912 * Task CPU affinities can make all that even more 'interesting'.
3915 unsigned long nr_iowait(void)
3917 unsigned long i, sum = 0;
3919 for_each_possible_cpu(i)
3920 sum += nr_iowait_cpu(i);
3928 * sched_exec - execve() is a valuable balancing opportunity, because at
3929 * this point the task has the smallest effective memory and cache footprint.
3931 void sched_exec(void)
3933 struct task_struct *p = current;
3934 unsigned long flags;
3937 raw_spin_lock_irqsave(&p->pi_lock, flags);
3938 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3939 if (dest_cpu == smp_processor_id())
3942 if (likely(cpu_active(dest_cpu))) {
3943 struct migration_arg arg = { p, dest_cpu };
3945 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3946 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3950 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3955 DEFINE_PER_CPU(struct kernel_stat, kstat);
3956 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3958 EXPORT_PER_CPU_SYMBOL(kstat);
3959 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3962 * The function fair_sched_class.update_curr accesses the struct curr
3963 * and its field curr->exec_start; when called from task_sched_runtime(),
3964 * we observe a high rate of cache misses in practice.
3965 * Prefetching this data results in improved performance.
3967 static inline void prefetch_curr_exec_start(struct task_struct *p)
3969 #ifdef CONFIG_FAIR_GROUP_SCHED
3970 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3972 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3975 prefetch(&curr->exec_start);
3979 * Return accounted runtime for the task.
3980 * In case the task is currently running, return the runtime plus current's
3981 * pending runtime that have not been accounted yet.
3983 unsigned long long task_sched_runtime(struct task_struct *p)
3989 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3991 * 64-bit doesn't need locks to atomically read a 64-bit value.
3992 * So we have a optimization chance when the task's delta_exec is 0.
3993 * Reading ->on_cpu is racy, but this is ok.
3995 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3996 * If we race with it entering CPU, unaccounted time is 0. This is
3997 * indistinguishable from the read occurring a few cycles earlier.
3998 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3999 * been accounted, so we're correct here as well.
4001 if (!p->on_cpu || !task_on_rq_queued(p))
4002 return p->se.sum_exec_runtime;
4005 rq = task_rq_lock(p, &rf);
4007 * Must be ->curr _and_ ->on_rq. If dequeued, we would
4008 * project cycles that may never be accounted to this
4009 * thread, breaking clock_gettime().
4011 if (task_current(rq, p) && task_on_rq_queued(p)) {
4012 prefetch_curr_exec_start(p);
4013 update_rq_clock(rq);
4014 p->sched_class->update_curr(rq);
4016 ns = p->se.sum_exec_runtime;
4017 task_rq_unlock(rq, p, &rf);
4023 * This function gets called by the timer code, with HZ frequency.
4024 * We call it with interrupts disabled.
4026 void scheduler_tick(void)
4028 int cpu = smp_processor_id();
4029 struct rq *rq = cpu_rq(cpu);
4030 struct task_struct *curr = rq->curr;
4032 unsigned long thermal_pressure;
4034 arch_scale_freq_tick();
4039 update_rq_clock(rq);
4040 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
4041 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
4042 curr->sched_class->task_tick(rq, curr, 0);
4043 calc_global_load_tick(rq);
4048 perf_event_task_tick();
4051 rq->idle_balance = idle_cpu(cpu);
4052 trigger_load_balance(rq);
4056 #ifdef CONFIG_NO_HZ_FULL
4061 struct delayed_work work;
4063 /* Values for ->state, see diagram below. */
4064 #define TICK_SCHED_REMOTE_OFFLINE 0
4065 #define TICK_SCHED_REMOTE_OFFLINING 1
4066 #define TICK_SCHED_REMOTE_RUNNING 2
4069 * State diagram for ->state:
4072 * TICK_SCHED_REMOTE_OFFLINE
4075 * | | sched_tick_remote()
4078 * +--TICK_SCHED_REMOTE_OFFLINING
4081 * sched_tick_start() | | sched_tick_stop()
4084 * TICK_SCHED_REMOTE_RUNNING
4087 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
4088 * and sched_tick_start() are happy to leave the state in RUNNING.
4091 static struct tick_work __percpu *tick_work_cpu;
4093 static void sched_tick_remote(struct work_struct *work)
4095 struct delayed_work *dwork = to_delayed_work(work);
4096 struct tick_work *twork = container_of(dwork, struct tick_work, work);
4097 int cpu = twork->cpu;
4098 struct rq *rq = cpu_rq(cpu);
4099 struct task_struct *curr;
4105 * Handle the tick only if it appears the remote CPU is running in full
4106 * dynticks mode. The check is racy by nature, but missing a tick or
4107 * having one too much is no big deal because the scheduler tick updates
4108 * statistics and checks timeslices in a time-independent way, regardless
4109 * of when exactly it is running.
4111 if (!tick_nohz_tick_stopped_cpu(cpu))
4114 rq_lock_irq(rq, &rf);
4116 if (cpu_is_offline(cpu))
4119 update_rq_clock(rq);
4121 if (!is_idle_task(curr)) {
4123 * Make sure the next tick runs within a reasonable
4126 delta = rq_clock_task(rq) - curr->se.exec_start;
4127 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
4129 curr->sched_class->task_tick(rq, curr, 0);
4131 calc_load_nohz_remote(rq);
4133 rq_unlock_irq(rq, &rf);
4137 * Run the remote tick once per second (1Hz). This arbitrary
4138 * frequency is large enough to avoid overload but short enough
4139 * to keep scheduler internal stats reasonably up to date. But
4140 * first update state to reflect hotplug activity if required.
4142 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
4143 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
4144 if (os == TICK_SCHED_REMOTE_RUNNING)
4145 queue_delayed_work(system_unbound_wq, dwork, HZ);
4148 static void sched_tick_start(int cpu)
4151 struct tick_work *twork;
4153 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4156 WARN_ON_ONCE(!tick_work_cpu);
4158 twork = per_cpu_ptr(tick_work_cpu, cpu);
4159 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
4160 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
4161 if (os == TICK_SCHED_REMOTE_OFFLINE) {
4163 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
4164 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
4168 #ifdef CONFIG_HOTPLUG_CPU
4169 static void sched_tick_stop(int cpu)
4171 struct tick_work *twork;
4174 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
4177 WARN_ON_ONCE(!tick_work_cpu);
4179 twork = per_cpu_ptr(tick_work_cpu, cpu);
4180 /* There cannot be competing actions, but don't rely on stop-machine. */
4181 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
4182 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
4183 /* Don't cancel, as this would mess up the state machine. */
4185 #endif /* CONFIG_HOTPLUG_CPU */
4187 int __init sched_tick_offload_init(void)
4189 tick_work_cpu = alloc_percpu(struct tick_work);
4190 BUG_ON(!tick_work_cpu);
4194 #else /* !CONFIG_NO_HZ_FULL */
4195 static inline void sched_tick_start(int cpu) { }
4196 static inline void sched_tick_stop(int cpu) { }
4199 #if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
4200 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
4202 * If the value passed in is equal to the current preempt count
4203 * then we just disabled preemption. Start timing the latency.
4205 static inline void preempt_latency_start(int val)
4207 if (preempt_count() == val) {
4208 unsigned long ip = get_lock_parent_ip();
4209 #ifdef CONFIG_DEBUG_PREEMPT
4210 current->preempt_disable_ip = ip;
4212 trace_preempt_off(CALLER_ADDR0, ip);
4216 void preempt_count_add(int val)
4218 #ifdef CONFIG_DEBUG_PREEMPT
4222 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4225 __preempt_count_add(val);
4226 #ifdef CONFIG_DEBUG_PREEMPT
4228 * Spinlock count overflowing soon?
4230 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4233 preempt_latency_start(val);
4235 EXPORT_SYMBOL(preempt_count_add);
4236 NOKPROBE_SYMBOL(preempt_count_add);
4239 * If the value passed in equals to the current preempt count
4240 * then we just enabled preemption. Stop timing the latency.
4242 static inline void preempt_latency_stop(int val)
4244 if (preempt_count() == val)
4245 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
4248 void preempt_count_sub(int val)
4250 #ifdef CONFIG_DEBUG_PREEMPT
4254 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4257 * Is the spinlock portion underflowing?
4259 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4260 !(preempt_count() & PREEMPT_MASK)))
4264 preempt_latency_stop(val);
4265 __preempt_count_sub(val);
4267 EXPORT_SYMBOL(preempt_count_sub);
4268 NOKPROBE_SYMBOL(preempt_count_sub);
4271 static inline void preempt_latency_start(int val) { }
4272 static inline void preempt_latency_stop(int val) { }
4275 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
4277 #ifdef CONFIG_DEBUG_PREEMPT
4278 return p->preempt_disable_ip;
4285 * Print scheduling while atomic bug:
4287 static noinline void __schedule_bug(struct task_struct *prev)
4289 /* Save this before calling printk(), since that will clobber it */
4290 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
4292 if (oops_in_progress)
4295 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4296 prev->comm, prev->pid, preempt_count());
4298 debug_show_held_locks(prev);
4300 if (irqs_disabled())
4301 print_irqtrace_events(prev);
4302 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
4303 && in_atomic_preempt_off()) {
4304 pr_err("Preemption disabled at:");
4305 print_ip_sym(KERN_ERR, preempt_disable_ip);
4308 panic("scheduling while atomic\n");
4311 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4315 * Various schedule()-time debugging checks and statistics:
4317 static inline void schedule_debug(struct task_struct *prev, bool preempt)
4319 #ifdef CONFIG_SCHED_STACK_END_CHECK
4320 if (task_stack_end_corrupted(prev))
4321 panic("corrupted stack end detected inside scheduler\n");
4323 if (task_scs_end_corrupted(prev))
4324 panic("corrupted shadow stack detected inside scheduler\n");
4327 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
4328 if (!preempt && prev->state && prev->non_block_count) {
4329 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
4330 prev->comm, prev->pid, prev->non_block_count);
4332 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
4336 if (unlikely(in_atomic_preempt_off())) {
4337 __schedule_bug(prev);
4338 preempt_count_set(PREEMPT_DISABLED);
4342 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4344 schedstat_inc(this_rq()->sched_count);
4347 static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
4348 struct rq_flags *rf)
4351 const struct sched_class *class;
4353 * We must do the balancing pass before put_prev_task(), such
4354 * that when we release the rq->lock the task is in the same
4355 * state as before we took rq->lock.
4357 * We can terminate the balance pass as soon as we know there is
4358 * a runnable task of @class priority or higher.
4360 for_class_range(class, prev->sched_class, &idle_sched_class) {
4361 if (class->balance(rq, prev, rf))
4366 put_prev_task(rq, prev);
4370 * Pick up the highest-prio task:
4372 static inline struct task_struct *
4373 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
4375 const struct sched_class *class;
4376 struct task_struct *p;
4379 * Optimization: we know that if all tasks are in the fair class we can
4380 * call that function directly, but only if the @prev task wasn't of a
4381 * higher scheduling class, because otherwise those loose the
4382 * opportunity to pull in more work from other CPUs.
4384 if (likely(prev->sched_class <= &fair_sched_class &&
4385 rq->nr_running == rq->cfs.h_nr_running)) {
4387 p = pick_next_task_fair(rq, prev, rf);
4388 if (unlikely(p == RETRY_TASK))
4391 /* Assumes fair_sched_class->next == idle_sched_class */
4393 put_prev_task(rq, prev);
4394 p = pick_next_task_idle(rq);
4401 put_prev_task_balance(rq, prev, rf);
4403 for_each_class(class) {
4404 p = class->pick_next_task(rq);
4409 /* The idle class should always have a runnable task: */
4414 * __schedule() is the main scheduler function.
4416 * The main means of driving the scheduler and thus entering this function are:
4418 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
4420 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
4421 * paths. For example, see arch/x86/entry_64.S.
4423 * To drive preemption between tasks, the scheduler sets the flag in timer
4424 * interrupt handler scheduler_tick().
4426 * 3. Wakeups don't really cause entry into schedule(). They add a
4427 * task to the run-queue and that's it.
4429 * Now, if the new task added to the run-queue preempts the current
4430 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
4431 * called on the nearest possible occasion:
4433 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
4435 * - in syscall or exception context, at the next outmost
4436 * preempt_enable(). (this might be as soon as the wake_up()'s
4439 * - in IRQ context, return from interrupt-handler to
4440 * preemptible context
4442 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
4445 * - cond_resched() call
4446 * - explicit schedule() call
4447 * - return from syscall or exception to user-space
4448 * - return from interrupt-handler to user-space
4450 * WARNING: must be called with preemption disabled!
4452 static void __sched notrace __schedule(bool preempt)
4454 struct task_struct *prev, *next;
4455 unsigned long *switch_count;
4456 unsigned long prev_state;
4461 cpu = smp_processor_id();
4465 schedule_debug(prev, preempt);
4467 if (sched_feat(HRTICK))
4470 local_irq_disable();
4471 rcu_note_context_switch(preempt);
4474 * Make sure that signal_pending_state()->signal_pending() below
4475 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
4476 * done by the caller to avoid the race with signal_wake_up():
4478 * __set_current_state(@state) signal_wake_up()
4479 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
4480 * wake_up_state(p, state)
4481 * LOCK rq->lock LOCK p->pi_state
4482 * smp_mb__after_spinlock() smp_mb__after_spinlock()
4483 * if (signal_pending_state()) if (p->state & @state)
4485 * Also, the membarrier system call requires a full memory barrier
4486 * after coming from user-space, before storing to rq->curr.
4489 smp_mb__after_spinlock();
4491 /* Promote REQ to ACT */
4492 rq->clock_update_flags <<= 1;
4493 update_rq_clock(rq);
4495 switch_count = &prev->nivcsw;
4498 * We must load prev->state once (task_struct::state is volatile), such
4501 * - we form a control dependency vs deactivate_task() below.
4502 * - ptrace_{,un}freeze_traced() can change ->state underneath us.
4504 prev_state = prev->state;
4505 if (!preempt && prev_state) {
4506 if (signal_pending_state(prev_state, prev)) {
4507 prev->state = TASK_RUNNING;
4509 prev->sched_contributes_to_load =
4510 (prev_state & TASK_UNINTERRUPTIBLE) &&
4511 !(prev_state & TASK_NOLOAD) &&
4512 !(prev->flags & PF_FROZEN);
4514 if (prev->sched_contributes_to_load)
4515 rq->nr_uninterruptible++;
4518 * __schedule() ttwu()
4519 * prev_state = prev->state; if (p->on_rq && ...)
4520 * if (prev_state) goto out;
4521 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
4522 * p->state = TASK_WAKING
4524 * Where __schedule() and ttwu() have matching control dependencies.
4526 * After this, schedule() must not care about p->state any more.
4528 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
4530 if (prev->in_iowait) {
4531 atomic_inc(&rq->nr_iowait);
4532 delayacct_blkio_start();
4535 switch_count = &prev->nvcsw;
4538 next = pick_next_task(rq, prev, &rf);
4539 clear_tsk_need_resched(prev);
4540 clear_preempt_need_resched();
4542 if (likely(prev != next)) {
4545 * RCU users of rcu_dereference(rq->curr) may not see
4546 * changes to task_struct made by pick_next_task().
4548 RCU_INIT_POINTER(rq->curr, next);
4550 * The membarrier system call requires each architecture
4551 * to have a full memory barrier after updating
4552 * rq->curr, before returning to user-space.
4554 * Here are the schemes providing that barrier on the
4555 * various architectures:
4556 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
4557 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
4558 * - finish_lock_switch() for weakly-ordered
4559 * architectures where spin_unlock is a full barrier,
4560 * - switch_to() for arm64 (weakly-ordered, spin_unlock
4561 * is a RELEASE barrier),
4565 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
4567 trace_sched_switch(preempt, prev, next);
4569 /* Also unlocks the rq: */
4570 rq = context_switch(rq, prev, next, &rf);
4572 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
4574 rq_unpin_lock(rq, &rf);
4575 __balance_callbacks(rq);
4576 raw_spin_unlock_irq(&rq->lock);
4580 void __noreturn do_task_dead(void)
4582 /* Causes final put_task_struct in finish_task_switch(): */
4583 set_special_state(TASK_DEAD);
4585 /* Tell freezer to ignore us: */
4586 current->flags |= PF_NOFREEZE;
4591 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
4596 static inline void sched_submit_work(struct task_struct *tsk)
4598 unsigned int task_flags;
4603 task_flags = tsk->flags;
4605 * If a worker went to sleep, notify and ask workqueue whether
4606 * it wants to wake up a task to maintain concurrency.
4607 * As this function is called inside the schedule() context,
4608 * we disable preemption to avoid it calling schedule() again
4609 * in the possible wakeup of a kworker and because wq_worker_sleeping()
4612 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4614 if (task_flags & PF_WQ_WORKER)
4615 wq_worker_sleeping(tsk);
4617 io_wq_worker_sleeping(tsk);
4618 preempt_enable_no_resched();
4621 if (tsk_is_pi_blocked(tsk))
4625 * If we are going to sleep and we have plugged IO queued,
4626 * make sure to submit it to avoid deadlocks.
4628 if (blk_needs_flush_plug(tsk))
4629 blk_schedule_flush_plug(tsk);
4632 static void sched_update_worker(struct task_struct *tsk)
4634 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
4635 if (tsk->flags & PF_WQ_WORKER)
4636 wq_worker_running(tsk);
4638 io_wq_worker_running(tsk);
4642 asmlinkage __visible void __sched schedule(void)
4644 struct task_struct *tsk = current;
4646 sched_submit_work(tsk);
4650 sched_preempt_enable_no_resched();
4651 } while (need_resched());
4652 sched_update_worker(tsk);
4654 EXPORT_SYMBOL(schedule);
4657 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
4658 * state (have scheduled out non-voluntarily) by making sure that all
4659 * tasks have either left the run queue or have gone into user space.
4660 * As idle tasks do not do either, they must not ever be preempted
4661 * (schedule out non-voluntarily).
4663 * schedule_idle() is similar to schedule_preempt_disable() except that it
4664 * never enables preemption because it does not call sched_submit_work().
4666 void __sched schedule_idle(void)
4669 * As this skips calling sched_submit_work(), which the idle task does
4670 * regardless because that function is a nop when the task is in a
4671 * TASK_RUNNING state, make sure this isn't used someplace that the
4672 * current task can be in any other state. Note, idle is always in the
4673 * TASK_RUNNING state.
4675 WARN_ON_ONCE(current->state);
4678 } while (need_resched());
4681 #ifdef CONFIG_CONTEXT_TRACKING
4682 asmlinkage __visible void __sched schedule_user(void)
4685 * If we come here after a random call to set_need_resched(),
4686 * or we have been woken up remotely but the IPI has not yet arrived,
4687 * we haven't yet exited the RCU idle mode. Do it here manually until
4688 * we find a better solution.
4690 * NB: There are buggy callers of this function. Ideally we
4691 * should warn if prev_state != CONTEXT_USER, but that will trigger
4692 * too frequently to make sense yet.
4694 enum ctx_state prev_state = exception_enter();
4696 exception_exit(prev_state);
4701 * schedule_preempt_disabled - called with preemption disabled
4703 * Returns with preemption disabled. Note: preempt_count must be 1
4705 void __sched schedule_preempt_disabled(void)
4707 sched_preempt_enable_no_resched();
4712 static void __sched notrace preempt_schedule_common(void)
4716 * Because the function tracer can trace preempt_count_sub()
4717 * and it also uses preempt_enable/disable_notrace(), if
4718 * NEED_RESCHED is set, the preempt_enable_notrace() called
4719 * by the function tracer will call this function again and
4720 * cause infinite recursion.
4722 * Preemption must be disabled here before the function
4723 * tracer can trace. Break up preempt_disable() into two
4724 * calls. One to disable preemption without fear of being
4725 * traced. The other to still record the preemption latency,
4726 * which can also be traced by the function tracer.
4728 preempt_disable_notrace();
4729 preempt_latency_start(1);
4731 preempt_latency_stop(1);
4732 preempt_enable_no_resched_notrace();
4735 * Check again in case we missed a preemption opportunity
4736 * between schedule and now.
4738 } while (need_resched());
4741 #ifdef CONFIG_PREEMPTION
4743 * This is the entry point to schedule() from in-kernel preemption
4744 * off of preempt_enable.
4746 asmlinkage __visible void __sched notrace preempt_schedule(void)
4749 * If there is a non-zero preempt_count or interrupts are disabled,
4750 * we do not want to preempt the current task. Just return..
4752 if (likely(!preemptible()))
4755 preempt_schedule_common();
4757 NOKPROBE_SYMBOL(preempt_schedule);
4758 EXPORT_SYMBOL(preempt_schedule);
4761 * preempt_schedule_notrace - preempt_schedule called by tracing
4763 * The tracing infrastructure uses preempt_enable_notrace to prevent
4764 * recursion and tracing preempt enabling caused by the tracing
4765 * infrastructure itself. But as tracing can happen in areas coming
4766 * from userspace or just about to enter userspace, a preempt enable
4767 * can occur before user_exit() is called. This will cause the scheduler
4768 * to be called when the system is still in usermode.
4770 * To prevent this, the preempt_enable_notrace will use this function
4771 * instead of preempt_schedule() to exit user context if needed before
4772 * calling the scheduler.
4774 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
4776 enum ctx_state prev_ctx;
4778 if (likely(!preemptible()))
4783 * Because the function tracer can trace preempt_count_sub()
4784 * and it also uses preempt_enable/disable_notrace(), if
4785 * NEED_RESCHED is set, the preempt_enable_notrace() called
4786 * by the function tracer will call this function again and
4787 * cause infinite recursion.
4789 * Preemption must be disabled here before the function
4790 * tracer can trace. Break up preempt_disable() into two
4791 * calls. One to disable preemption without fear of being
4792 * traced. The other to still record the preemption latency,
4793 * which can also be traced by the function tracer.
4795 preempt_disable_notrace();
4796 preempt_latency_start(1);
4798 * Needs preempt disabled in case user_exit() is traced
4799 * and the tracer calls preempt_enable_notrace() causing
4800 * an infinite recursion.
4802 prev_ctx = exception_enter();
4804 exception_exit(prev_ctx);
4806 preempt_latency_stop(1);
4807 preempt_enable_no_resched_notrace();
4808 } while (need_resched());
4810 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
4812 #endif /* CONFIG_PREEMPTION */
4815 * This is the entry point to schedule() from kernel preemption
4816 * off of irq context.
4817 * Note, that this is called and return with irqs disabled. This will
4818 * protect us against recursive calling from irq.
4820 asmlinkage __visible void __sched preempt_schedule_irq(void)
4822 enum ctx_state prev_state;
4824 /* Catch callers which need to be fixed */
4825 BUG_ON(preempt_count() || !irqs_disabled());
4827 prev_state = exception_enter();
4833 local_irq_disable();
4834 sched_preempt_enable_no_resched();
4835 } while (need_resched());
4837 exception_exit(prev_state);
4840 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
4843 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
4844 return try_to_wake_up(curr->private, mode, wake_flags);
4846 EXPORT_SYMBOL(default_wake_function);
4848 #ifdef CONFIG_RT_MUTEXES
4850 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
4853 prio = min(prio, pi_task->prio);
4858 static inline int rt_effective_prio(struct task_struct *p, int prio)
4860 struct task_struct *pi_task = rt_mutex_get_top_task(p);
4862 return __rt_effective_prio(pi_task, prio);
4866 * rt_mutex_setprio - set the current priority of a task
4868 * @pi_task: donor task
4870 * This function changes the 'effective' priority of a task. It does
4871 * not touch ->normal_prio like __setscheduler().
4873 * Used by the rt_mutex code to implement priority inheritance
4874 * logic. Call site only calls if the priority of the task changed.
4876 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
4878 int prio, oldprio, queued, running, queue_flag =
4879 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4880 const struct sched_class *prev_class;
4884 /* XXX used to be waiter->prio, not waiter->task->prio */
4885 prio = __rt_effective_prio(pi_task, p->normal_prio);
4888 * If nothing changed; bail early.
4890 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
4893 rq = __task_rq_lock(p, &rf);
4894 update_rq_clock(rq);
4896 * Set under pi_lock && rq->lock, such that the value can be used under
4899 * Note that there is loads of tricky to make this pointer cache work
4900 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
4901 * ensure a task is de-boosted (pi_task is set to NULL) before the
4902 * task is allowed to run again (and can exit). This ensures the pointer
4903 * points to a blocked task -- which guaratees the task is present.
4905 p->pi_top_task = pi_task;
4908 * For FIFO/RR we only need to set prio, if that matches we're done.
4910 if (prio == p->prio && !dl_prio(prio))
4914 * Idle task boosting is a nono in general. There is one
4915 * exception, when PREEMPT_RT and NOHZ is active:
4917 * The idle task calls get_next_timer_interrupt() and holds
4918 * the timer wheel base->lock on the CPU and another CPU wants
4919 * to access the timer (probably to cancel it). We can safely
4920 * ignore the boosting request, as the idle CPU runs this code
4921 * with interrupts disabled and will complete the lock
4922 * protected section without being interrupted. So there is no
4923 * real need to boost.
4925 if (unlikely(p == rq->idle)) {
4926 WARN_ON(p != rq->curr);
4927 WARN_ON(p->pi_blocked_on);
4931 trace_sched_pi_setprio(p, pi_task);
4934 if (oldprio == prio)
4935 queue_flag &= ~DEQUEUE_MOVE;
4937 prev_class = p->sched_class;
4938 queued = task_on_rq_queued(p);
4939 running = task_current(rq, p);
4941 dequeue_task(rq, p, queue_flag);
4943 put_prev_task(rq, p);
4946 * Boosting condition are:
4947 * 1. -rt task is running and holds mutex A
4948 * --> -dl task blocks on mutex A
4950 * 2. -dl task is running and holds mutex A
4951 * --> -dl task blocks on mutex A and could preempt the
4954 if (dl_prio(prio)) {
4955 if (!dl_prio(p->normal_prio) ||
4956 (pi_task && dl_prio(pi_task->prio) &&
4957 dl_entity_preempt(&pi_task->dl, &p->dl))) {
4958 p->dl.dl_boosted = 1;
4959 queue_flag |= ENQUEUE_REPLENISH;
4961 p->dl.dl_boosted = 0;
4962 p->sched_class = &dl_sched_class;
4963 } else if (rt_prio(prio)) {
4964 if (dl_prio(oldprio))
4965 p->dl.dl_boosted = 0;
4967 queue_flag |= ENQUEUE_HEAD;
4968 p->sched_class = &rt_sched_class;
4970 if (dl_prio(oldprio))
4971 p->dl.dl_boosted = 0;
4972 if (rt_prio(oldprio))
4974 p->sched_class = &fair_sched_class;
4980 enqueue_task(rq, p, queue_flag);
4982 set_next_task(rq, p);
4984 check_class_changed(rq, p, prev_class, oldprio);
4986 /* Avoid rq from going away on us: */
4989 rq_unpin_lock(rq, &rf);
4990 __balance_callbacks(rq);
4991 raw_spin_unlock(&rq->lock);
4996 static inline int rt_effective_prio(struct task_struct *p, int prio)
5002 void set_user_nice(struct task_struct *p, long nice)
5004 bool queued, running;
5009 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
5012 * We have to be careful, if called from sys_setpriority(),
5013 * the task might be in the middle of scheduling on another CPU.
5015 rq = task_rq_lock(p, &rf);
5016 update_rq_clock(rq);
5019 * The RT priorities are set via sched_setscheduler(), but we still
5020 * allow the 'normal' nice value to be set - but as expected
5021 * it wont have any effect on scheduling until the task is
5022 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
5024 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
5025 p->static_prio = NICE_TO_PRIO(nice);
5028 queued = task_on_rq_queued(p);
5029 running = task_current(rq, p);
5031 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
5033 put_prev_task(rq, p);
5035 p->static_prio = NICE_TO_PRIO(nice);
5036 set_load_weight(p, true);
5038 p->prio = effective_prio(p);
5041 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5043 set_next_task(rq, p);
5046 * If the task increased its priority or is running and
5047 * lowered its priority, then reschedule its CPU:
5049 p->sched_class->prio_changed(rq, p, old_prio);
5052 task_rq_unlock(rq, p, &rf);
5054 EXPORT_SYMBOL(set_user_nice);
5057 * can_nice - check if a task can reduce its nice value
5061 int can_nice(const struct task_struct *p, const int nice)
5063 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
5064 int nice_rlim = nice_to_rlimit(nice);
5066 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5067 capable(CAP_SYS_NICE));
5070 #ifdef __ARCH_WANT_SYS_NICE
5073 * sys_nice - change the priority of the current process.
5074 * @increment: priority increment
5076 * sys_setpriority is a more generic, but much slower function that
5077 * does similar things.
5079 SYSCALL_DEFINE1(nice, int, increment)
5084 * Setpriority might change our priority at the same moment.
5085 * We don't have to worry. Conceptually one call occurs first
5086 * and we have a single winner.
5088 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
5089 nice = task_nice(current) + increment;
5091 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
5092 if (increment < 0 && !can_nice(current, nice))
5095 retval = security_task_setnice(current, nice);
5099 set_user_nice(current, nice);
5106 * task_prio - return the priority value of a given task.
5107 * @p: the task in question.
5109 * Return: The priority value as seen by users in /proc.
5110 * RT tasks are offset by -200. Normal tasks are centered
5111 * around 0, value goes from -16 to +15.
5113 int task_prio(const struct task_struct *p)
5115 return p->prio - MAX_RT_PRIO;
5119 * idle_cpu - is a given CPU idle currently?
5120 * @cpu: the processor in question.
5122 * Return: 1 if the CPU is currently idle. 0 otherwise.
5124 int idle_cpu(int cpu)
5126 struct rq *rq = cpu_rq(cpu);
5128 if (rq->curr != rq->idle)
5135 if (rq->ttwu_pending)
5143 * available_idle_cpu - is a given CPU idle for enqueuing work.
5144 * @cpu: the CPU in question.
5146 * Return: 1 if the CPU is currently idle. 0 otherwise.
5148 int available_idle_cpu(int cpu)
5153 if (vcpu_is_preempted(cpu))
5160 * idle_task - return the idle task for a given CPU.
5161 * @cpu: the processor in question.
5163 * Return: The idle task for the CPU @cpu.
5165 struct task_struct *idle_task(int cpu)
5167 return cpu_rq(cpu)->idle;
5171 * find_process_by_pid - find a process with a matching PID value.
5172 * @pid: the pid in question.
5174 * The task of @pid, if found. %NULL otherwise.
5176 static struct task_struct *find_process_by_pid(pid_t pid)
5178 return pid ? find_task_by_vpid(pid) : current;
5182 * sched_setparam() passes in -1 for its policy, to let the functions
5183 * it calls know not to change it.
5185 #define SETPARAM_POLICY -1
5187 static void __setscheduler_params(struct task_struct *p,
5188 const struct sched_attr *attr)
5190 int policy = attr->sched_policy;
5192 if (policy == SETPARAM_POLICY)
5197 if (dl_policy(policy))
5198 __setparam_dl(p, attr);
5199 else if (fair_policy(policy))
5200 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
5203 * __sched_setscheduler() ensures attr->sched_priority == 0 when
5204 * !rt_policy. Always setting this ensures that things like
5205 * getparam()/getattr() don't report silly values for !rt tasks.
5207 p->rt_priority = attr->sched_priority;
5208 p->normal_prio = normal_prio(p);
5209 set_load_weight(p, true);
5212 /* Actually do priority change: must hold pi & rq lock. */
5213 static void __setscheduler(struct rq *rq, struct task_struct *p,
5214 const struct sched_attr *attr, bool keep_boost)
5217 * If params can't change scheduling class changes aren't allowed
5220 if (attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)
5223 __setscheduler_params(p, attr);
5226 * Keep a potential priority boosting if called from
5227 * sched_setscheduler().
5229 p->prio = normal_prio(p);
5231 p->prio = rt_effective_prio(p, p->prio);
5233 if (dl_prio(p->prio))
5234 p->sched_class = &dl_sched_class;
5235 else if (rt_prio(p->prio))
5236 p->sched_class = &rt_sched_class;
5238 p->sched_class = &fair_sched_class;
5242 * Check the target process has a UID that matches the current process's:
5244 static bool check_same_owner(struct task_struct *p)
5246 const struct cred *cred = current_cred(), *pcred;
5250 pcred = __task_cred(p);
5251 match = (uid_eq(cred->euid, pcred->euid) ||
5252 uid_eq(cred->euid, pcred->uid));
5257 static int __sched_setscheduler(struct task_struct *p,
5258 const struct sched_attr *attr,
5261 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
5262 MAX_RT_PRIO - 1 - attr->sched_priority;
5263 int retval, oldprio, oldpolicy = -1, queued, running;
5264 int new_effective_prio, policy = attr->sched_policy;
5265 const struct sched_class *prev_class;
5266 struct callback_head *head;
5269 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
5272 /* The pi code expects interrupts enabled */
5273 BUG_ON(pi && in_interrupt());
5275 /* Double check policy once rq lock held: */
5277 reset_on_fork = p->sched_reset_on_fork;
5278 policy = oldpolicy = p->policy;
5280 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
5282 if (!valid_policy(policy))
5286 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
5290 * Valid priorities for SCHED_FIFO and SCHED_RR are
5291 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5292 * SCHED_BATCH and SCHED_IDLE is 0.
5294 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
5295 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
5297 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
5298 (rt_policy(policy) != (attr->sched_priority != 0)))
5302 * Allow unprivileged RT tasks to decrease priority:
5304 if (user && !capable(CAP_SYS_NICE)) {
5305 if (fair_policy(policy)) {
5306 if (attr->sched_nice < task_nice(p) &&
5307 !can_nice(p, attr->sched_nice))
5311 if (rt_policy(policy)) {
5312 unsigned long rlim_rtprio =
5313 task_rlimit(p, RLIMIT_RTPRIO);
5315 /* Can't set/change the rt policy: */
5316 if (policy != p->policy && !rlim_rtprio)
5319 /* Can't increase priority: */
5320 if (attr->sched_priority > p->rt_priority &&
5321 attr->sched_priority > rlim_rtprio)
5326 * Can't set/change SCHED_DEADLINE policy at all for now
5327 * (safest behavior); in the future we would like to allow
5328 * unprivileged DL tasks to increase their relative deadline
5329 * or reduce their runtime (both ways reducing utilization)
5331 if (dl_policy(policy))
5335 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5336 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5338 if (task_has_idle_policy(p) && !idle_policy(policy)) {
5339 if (!can_nice(p, task_nice(p)))
5343 /* Can't change other user's priorities: */
5344 if (!check_same_owner(p))
5347 /* Normal users shall not reset the sched_reset_on_fork flag: */
5348 if (p->sched_reset_on_fork && !reset_on_fork)
5353 if (attr->sched_flags & SCHED_FLAG_SUGOV)
5356 retval = security_task_setscheduler(p);
5361 /* Update task specific "requested" clamps */
5362 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
5363 retval = uclamp_validate(p, attr);
5372 * Make sure no PI-waiters arrive (or leave) while we are
5373 * changing the priority of the task:
5375 * To be able to change p->policy safely, the appropriate
5376 * runqueue lock must be held.
5378 rq = task_rq_lock(p, &rf);
5379 update_rq_clock(rq);
5382 * Changing the policy of the stop threads its a very bad idea:
5384 if (p == rq->stop) {
5390 * If not changing anything there's no need to proceed further,
5391 * but store a possible modification of reset_on_fork.
5393 if (unlikely(policy == p->policy)) {
5394 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
5396 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
5398 if (dl_policy(policy) && dl_param_changed(p, attr))
5400 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
5403 p->sched_reset_on_fork = reset_on_fork;
5410 #ifdef CONFIG_RT_GROUP_SCHED
5412 * Do not allow realtime tasks into groups that have no runtime
5415 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5416 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5417 !task_group_is_autogroup(task_group(p))) {
5423 if (dl_bandwidth_enabled() && dl_policy(policy) &&
5424 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
5425 cpumask_t *span = rq->rd->span;
5428 * Don't allow tasks with an affinity mask smaller than
5429 * the entire root_domain to become SCHED_DEADLINE. We
5430 * will also fail if there's no bandwidth available.
5432 if (!cpumask_subset(span, p->cpus_ptr) ||
5433 rq->rd->dl_bw.bw == 0) {
5441 /* Re-check policy now with rq lock held: */
5442 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5443 policy = oldpolicy = -1;
5444 task_rq_unlock(rq, p, &rf);
5446 cpuset_read_unlock();
5451 * If setscheduling to SCHED_DEADLINE (or changing the parameters
5452 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
5455 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
5460 p->sched_reset_on_fork = reset_on_fork;
5465 * Take priority boosted tasks into account. If the new
5466 * effective priority is unchanged, we just store the new
5467 * normal parameters and do not touch the scheduler class and
5468 * the runqueue. This will be done when the task deboost
5471 new_effective_prio = rt_effective_prio(p, newprio);
5472 if (new_effective_prio == oldprio)
5473 queue_flags &= ~DEQUEUE_MOVE;
5476 queued = task_on_rq_queued(p);
5477 running = task_current(rq, p);
5479 dequeue_task(rq, p, queue_flags);
5481 put_prev_task(rq, p);
5483 prev_class = p->sched_class;
5485 __setscheduler(rq, p, attr, pi);
5486 __setscheduler_uclamp(p, attr);
5490 * We enqueue to tail when the priority of a task is
5491 * increased (user space view).
5493 if (oldprio < p->prio)
5494 queue_flags |= ENQUEUE_HEAD;
5496 enqueue_task(rq, p, queue_flags);
5499 set_next_task(rq, p);
5501 check_class_changed(rq, p, prev_class, oldprio);
5503 /* Avoid rq from going away on us: */
5505 head = splice_balance_callbacks(rq);
5506 task_rq_unlock(rq, p, &rf);
5509 cpuset_read_unlock();
5510 rt_mutex_adjust_pi(p);
5513 /* Run balance callbacks after we've adjusted the PI chain: */
5514 balance_callbacks(rq, head);
5520 task_rq_unlock(rq, p, &rf);
5522 cpuset_read_unlock();
5526 static int _sched_setscheduler(struct task_struct *p, int policy,
5527 const struct sched_param *param, bool check)
5529 struct sched_attr attr = {
5530 .sched_policy = policy,
5531 .sched_priority = param->sched_priority,
5532 .sched_nice = PRIO_TO_NICE(p->static_prio),
5535 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
5536 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
5537 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5538 policy &= ~SCHED_RESET_ON_FORK;
5539 attr.sched_policy = policy;
5542 return __sched_setscheduler(p, &attr, check, true);
5545 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5546 * @p: the task in question.
5547 * @policy: new policy.
5548 * @param: structure containing the new RT priority.
5550 * Use sched_set_fifo(), read its comment.
5552 * Return: 0 on success. An error code otherwise.
5554 * NOTE that the task may be already dead.
5556 int sched_setscheduler(struct task_struct *p, int policy,
5557 const struct sched_param *param)
5559 return _sched_setscheduler(p, policy, param, true);
5562 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
5564 return __sched_setscheduler(p, attr, true, true);
5567 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
5569 return __sched_setscheduler(p, attr, false, true);
5573 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5574 * @p: the task in question.
5575 * @policy: new policy.
5576 * @param: structure containing the new RT priority.
5578 * Just like sched_setscheduler, only don't bother checking if the
5579 * current context has permission. For example, this is needed in
5580 * stop_machine(): we create temporary high priority worker threads,
5581 * but our caller might not have that capability.
5583 * Return: 0 on success. An error code otherwise.
5585 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5586 const struct sched_param *param)
5588 return _sched_setscheduler(p, policy, param, false);
5592 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
5593 * incapable of resource management, which is the one thing an OS really should
5596 * This is of course the reason it is limited to privileged users only.
5598 * Worse still; it is fundamentally impossible to compose static priority
5599 * workloads. You cannot take two correctly working static prio workloads
5600 * and smash them together and still expect them to work.
5602 * For this reason 'all' FIFO tasks the kernel creates are basically at:
5606 * The administrator _MUST_ configure the system, the kernel simply doesn't
5607 * know enough information to make a sensible choice.
5609 void sched_set_fifo(struct task_struct *p)
5611 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
5612 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
5614 EXPORT_SYMBOL_GPL(sched_set_fifo);
5617 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
5619 void sched_set_fifo_low(struct task_struct *p)
5621 struct sched_param sp = { .sched_priority = 1 };
5622 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
5624 EXPORT_SYMBOL_GPL(sched_set_fifo_low);
5626 void sched_set_normal(struct task_struct *p, int nice)
5628 struct sched_attr attr = {
5629 .sched_policy = SCHED_NORMAL,
5632 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
5634 EXPORT_SYMBOL_GPL(sched_set_normal);
5637 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5639 struct sched_param lparam;
5640 struct task_struct *p;
5643 if (!param || pid < 0)
5645 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5650 p = find_process_by_pid(pid);
5656 retval = sched_setscheduler(p, policy, &lparam);
5664 * Mimics kernel/events/core.c perf_copy_attr().
5666 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
5671 /* Zero the full structure, so that a short copy will be nice: */
5672 memset(attr, 0, sizeof(*attr));
5674 ret = get_user(size, &uattr->size);
5678 /* ABI compatibility quirk: */
5680 size = SCHED_ATTR_SIZE_VER0;
5681 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
5684 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
5691 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
5692 size < SCHED_ATTR_SIZE_VER1)
5696 * XXX: Do we want to be lenient like existing syscalls; or do we want
5697 * to be strict and return an error on out-of-bounds values?
5699 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
5704 put_user(sizeof(*attr), &uattr->size);
5709 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5710 * @pid: the pid in question.
5711 * @policy: new policy.
5712 * @param: structure containing the new RT priority.
5714 * Return: 0 on success. An error code otherwise.
5716 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
5721 return do_sched_setscheduler(pid, policy, param);
5725 * sys_sched_setparam - set/change the RT priority of a thread
5726 * @pid: the pid in question.
5727 * @param: structure containing the new RT priority.
5729 * Return: 0 on success. An error code otherwise.
5731 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5733 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
5737 * sys_sched_setattr - same as above, but with extended sched_attr
5738 * @pid: the pid in question.
5739 * @uattr: structure containing the extended parameters.
5740 * @flags: for future extension.
5742 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
5743 unsigned int, flags)
5745 struct sched_attr attr;
5746 struct task_struct *p;
5749 if (!uattr || pid < 0 || flags)
5752 retval = sched_copy_attr(uattr, &attr);
5756 if ((int)attr.sched_policy < 0)
5758 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
5759 attr.sched_policy = SETPARAM_POLICY;
5763 p = find_process_by_pid(pid);
5769 retval = sched_setattr(p, &attr);
5777 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5778 * @pid: the pid in question.
5780 * Return: On success, the policy of the thread. Otherwise, a negative error
5783 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5785 struct task_struct *p;
5793 p = find_process_by_pid(pid);
5795 retval = security_task_getscheduler(p);
5798 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5805 * sys_sched_getparam - get the RT priority of a thread
5806 * @pid: the pid in question.
5807 * @param: structure containing the RT priority.
5809 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
5812 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5814 struct sched_param lp = { .sched_priority = 0 };
5815 struct task_struct *p;
5818 if (!param || pid < 0)
5822 p = find_process_by_pid(pid);
5827 retval = security_task_getscheduler(p);
5831 if (task_has_rt_policy(p))
5832 lp.sched_priority = p->rt_priority;
5836 * This one might sleep, we cannot do it with a spinlock held ...
5838 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5848 * Copy the kernel size attribute structure (which might be larger
5849 * than what user-space knows about) to user-space.
5851 * Note that all cases are valid: user-space buffer can be larger or
5852 * smaller than the kernel-space buffer. The usual case is that both
5853 * have the same size.
5856 sched_attr_copy_to_user(struct sched_attr __user *uattr,
5857 struct sched_attr *kattr,
5860 unsigned int ksize = sizeof(*kattr);
5862 if (!access_ok(uattr, usize))
5866 * sched_getattr() ABI forwards and backwards compatibility:
5868 * If usize == ksize then we just copy everything to user-space and all is good.
5870 * If usize < ksize then we only copy as much as user-space has space for,
5871 * this keeps ABI compatibility as well. We skip the rest.
5873 * If usize > ksize then user-space is using a newer version of the ABI,
5874 * which part the kernel doesn't know about. Just ignore it - tooling can
5875 * detect the kernel's knowledge of attributes from the attr->size value
5876 * which is set to ksize in this case.
5878 kattr->size = min(usize, ksize);
5880 if (copy_to_user(uattr, kattr, kattr->size))
5887 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
5888 * @pid: the pid in question.
5889 * @uattr: structure containing the extended parameters.
5890 * @usize: sizeof(attr) for fwd/bwd comp.
5891 * @flags: for future extension.
5893 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
5894 unsigned int, usize, unsigned int, flags)
5896 struct sched_attr kattr = { };
5897 struct task_struct *p;
5900 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
5901 usize < SCHED_ATTR_SIZE_VER0 || flags)
5905 p = find_process_by_pid(pid);
5910 retval = security_task_getscheduler(p);
5914 kattr.sched_policy = p->policy;
5915 if (p->sched_reset_on_fork)
5916 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
5917 if (task_has_dl_policy(p))
5918 __getparam_dl(p, &kattr);
5919 else if (task_has_rt_policy(p))
5920 kattr.sched_priority = p->rt_priority;
5922 kattr.sched_nice = task_nice(p);
5924 #ifdef CONFIG_UCLAMP_TASK
5926 * This could race with another potential updater, but this is fine
5927 * because it'll correctly read the old or the new value. We don't need
5928 * to guarantee who wins the race as long as it doesn't return garbage.
5930 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
5931 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
5936 return sched_attr_copy_to_user(uattr, &kattr, usize);
5943 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5945 cpumask_var_t cpus_allowed, new_mask;
5946 struct task_struct *p;
5951 p = find_process_by_pid(pid);
5957 /* Prevent p going away */
5961 if (p->flags & PF_NO_SETAFFINITY) {
5965 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5969 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5971 goto out_free_cpus_allowed;
5974 if (!check_same_owner(p)) {
5976 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
5978 goto out_free_new_mask;
5983 retval = security_task_setscheduler(p);
5985 goto out_free_new_mask;
5988 cpuset_cpus_allowed(p, cpus_allowed);
5989 cpumask_and(new_mask, in_mask, cpus_allowed);
5992 * Since bandwidth control happens on root_domain basis,
5993 * if admission test is enabled, we only admit -deadline
5994 * tasks allowed to run on all the CPUs in the task's
5998 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
6000 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
6003 goto out_free_new_mask;
6009 retval = __set_cpus_allowed_ptr(p, new_mask, true);
6012 cpuset_cpus_allowed(p, cpus_allowed);
6013 if (!cpumask_subset(new_mask, cpus_allowed)) {
6015 * We must have raced with a concurrent cpuset
6016 * update. Just reset the cpus_allowed to the
6017 * cpuset's cpus_allowed
6019 cpumask_copy(new_mask, cpus_allowed);
6024 free_cpumask_var(new_mask);
6025 out_free_cpus_allowed:
6026 free_cpumask_var(cpus_allowed);
6032 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6033 struct cpumask *new_mask)
6035 if (len < cpumask_size())
6036 cpumask_clear(new_mask);
6037 else if (len > cpumask_size())
6038 len = cpumask_size();
6040 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6044 * sys_sched_setaffinity - set the CPU affinity of a process
6045 * @pid: pid of the process
6046 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6047 * @user_mask_ptr: user-space pointer to the new CPU mask
6049 * Return: 0 on success. An error code otherwise.
6051 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6052 unsigned long __user *, user_mask_ptr)
6054 cpumask_var_t new_mask;
6057 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6060 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6062 retval = sched_setaffinity(pid, new_mask);
6063 free_cpumask_var(new_mask);
6067 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6069 struct task_struct *p;
6070 unsigned long flags;
6076 p = find_process_by_pid(pid);
6080 retval = security_task_getscheduler(p);
6084 raw_spin_lock_irqsave(&p->pi_lock, flags);
6085 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
6086 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6095 * sys_sched_getaffinity - get the CPU affinity of a process
6096 * @pid: pid of the process
6097 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6098 * @user_mask_ptr: user-space pointer to hold the current CPU mask
6100 * Return: size of CPU mask copied to user_mask_ptr on success. An
6101 * error code otherwise.
6103 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6104 unsigned long __user *, user_mask_ptr)
6109 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6111 if (len & (sizeof(unsigned long)-1))
6114 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6117 ret = sched_getaffinity(pid, mask);
6119 unsigned int retlen = min(len, cpumask_size());
6121 if (copy_to_user(user_mask_ptr, mask, retlen))
6126 free_cpumask_var(mask);
6132 * sys_sched_yield - yield the current processor to other threads.
6134 * This function yields the current CPU to other tasks. If there are no
6135 * other threads running on this CPU then this function will return.
6139 static void do_sched_yield(void)
6144 rq = this_rq_lock_irq(&rf);
6146 schedstat_inc(rq->yld_count);
6147 current->sched_class->yield_task(rq);
6150 * Since we are going to call schedule() anyway, there's
6151 * no need to preempt or enable interrupts:
6155 sched_preempt_enable_no_resched();
6160 SYSCALL_DEFINE0(sched_yield)
6166 #ifndef CONFIG_PREEMPTION
6167 int __sched _cond_resched(void)
6169 if (should_resched(0)) {
6170 preempt_schedule_common();
6176 EXPORT_SYMBOL(_cond_resched);
6180 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6181 * call schedule, and on return reacquire the lock.
6183 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
6184 * operations here to prevent schedule() from being called twice (once via
6185 * spin_unlock(), once by hand).
6187 int __cond_resched_lock(spinlock_t *lock)
6189 int resched = should_resched(PREEMPT_LOCK_OFFSET);
6192 lockdep_assert_held(lock);
6194 if (spin_needbreak(lock) || resched) {
6197 preempt_schedule_common();
6205 EXPORT_SYMBOL(__cond_resched_lock);
6208 * yield - yield the current processor to other threads.
6210 * Do not ever use this function, there's a 99% chance you're doing it wrong.
6212 * The scheduler is at all times free to pick the calling task as the most
6213 * eligible task to run, if removing the yield() call from your code breaks
6214 * it, its already broken.
6216 * Typical broken usage is:
6221 * where one assumes that yield() will let 'the other' process run that will
6222 * make event true. If the current task is a SCHED_FIFO task that will never
6223 * happen. Never use yield() as a progress guarantee!!
6225 * If you want to use yield() to wait for something, use wait_event().
6226 * If you want to use yield() to be 'nice' for others, use cond_resched().
6227 * If you still want to use yield(), do not!
6229 void __sched yield(void)
6231 set_current_state(TASK_RUNNING);
6234 EXPORT_SYMBOL(yield);
6237 * yield_to - yield the current processor to another thread in
6238 * your thread group, or accelerate that thread toward the
6239 * processor it's on.
6241 * @preempt: whether task preemption is allowed or not
6243 * It's the caller's job to ensure that the target task struct
6244 * can't go away on us before we can do any checks.
6247 * true (>0) if we indeed boosted the target task.
6248 * false (0) if we failed to boost the target.
6249 * -ESRCH if there's no task to yield to.
6251 int __sched yield_to(struct task_struct *p, bool preempt)
6253 struct task_struct *curr = current;
6254 struct rq *rq, *p_rq;
6255 unsigned long flags;
6258 local_irq_save(flags);
6264 * If we're the only runnable task on the rq and target rq also
6265 * has only one task, there's absolutely no point in yielding.
6267 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
6272 double_rq_lock(rq, p_rq);
6273 if (task_rq(p) != p_rq) {
6274 double_rq_unlock(rq, p_rq);
6278 if (!curr->sched_class->yield_to_task)
6281 if (curr->sched_class != p->sched_class)
6284 if (task_running(p_rq, p) || p->state)
6287 yielded = curr->sched_class->yield_to_task(rq, p);
6289 schedstat_inc(rq->yld_count);
6291 * Make p's CPU reschedule; pick_next_entity takes care of
6294 if (preempt && rq != p_rq)
6299 double_rq_unlock(rq, p_rq);
6301 local_irq_restore(flags);
6308 EXPORT_SYMBOL_GPL(yield_to);
6310 int io_schedule_prepare(void)
6312 int old_iowait = current->in_iowait;
6314 current->in_iowait = 1;
6315 blk_schedule_flush_plug(current);
6320 void io_schedule_finish(int token)
6322 current->in_iowait = token;
6326 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6327 * that process accounting knows that this is a task in IO wait state.
6329 long __sched io_schedule_timeout(long timeout)
6334 token = io_schedule_prepare();
6335 ret = schedule_timeout(timeout);
6336 io_schedule_finish(token);
6340 EXPORT_SYMBOL(io_schedule_timeout);
6342 void __sched io_schedule(void)
6346 token = io_schedule_prepare();
6348 io_schedule_finish(token);
6350 EXPORT_SYMBOL(io_schedule);
6353 * sys_sched_get_priority_max - return maximum RT priority.
6354 * @policy: scheduling class.
6356 * Return: On success, this syscall returns the maximum
6357 * rt_priority that can be used by a given scheduling class.
6358 * On failure, a negative error code is returned.
6360 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6367 ret = MAX_USER_RT_PRIO-1;
6369 case SCHED_DEADLINE:
6380 * sys_sched_get_priority_min - return minimum RT priority.
6381 * @policy: scheduling class.
6383 * Return: On success, this syscall returns the minimum
6384 * rt_priority that can be used by a given scheduling class.
6385 * On failure, a negative error code is returned.
6387 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6396 case SCHED_DEADLINE:
6405 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
6407 struct task_struct *p;
6408 unsigned int time_slice;
6418 p = find_process_by_pid(pid);
6422 retval = security_task_getscheduler(p);
6426 rq = task_rq_lock(p, &rf);
6428 if (p->sched_class->get_rr_interval)
6429 time_slice = p->sched_class->get_rr_interval(rq, p);
6430 task_rq_unlock(rq, p, &rf);
6433 jiffies_to_timespec64(time_slice, t);
6442 * sys_sched_rr_get_interval - return the default timeslice of a process.
6443 * @pid: pid of the process.
6444 * @interval: userspace pointer to the timeslice value.
6446 * this syscall writes the default timeslice value of a given process
6447 * into the user-space timespec buffer. A value of '0' means infinity.
6449 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
6452 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6453 struct __kernel_timespec __user *, interval)
6455 struct timespec64 t;
6456 int retval = sched_rr_get_interval(pid, &t);
6459 retval = put_timespec64(&t, interval);
6464 #ifdef CONFIG_COMPAT_32BIT_TIME
6465 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
6466 struct old_timespec32 __user *, interval)
6468 struct timespec64 t;
6469 int retval = sched_rr_get_interval(pid, &t);
6472 retval = put_old_timespec32(&t, interval);
6477 void sched_show_task(struct task_struct *p)
6479 unsigned long free = 0;
6482 if (!try_get_task_stack(p))
6485 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
6487 if (p->state == TASK_RUNNING)
6488 pr_cont(" running task ");
6489 #ifdef CONFIG_DEBUG_STACK_USAGE
6490 free = stack_not_used(p);
6495 ppid = task_pid_nr(rcu_dereference(p->real_parent));
6497 pr_cont(" stack:%5lu pid:%5d ppid:%6d flags:0x%08lx\n",
6498 free, task_pid_nr(p), ppid,
6499 (unsigned long)task_thread_info(p)->flags);
6501 print_worker_info(KERN_INFO, p);
6502 print_stop_info(KERN_INFO, p);
6503 show_stack(p, NULL, KERN_INFO);
6506 EXPORT_SYMBOL_GPL(sched_show_task);
6509 state_filter_match(unsigned long state_filter, struct task_struct *p)
6511 /* no filter, everything matches */
6515 /* filter, but doesn't match */
6516 if (!(p->state & state_filter))
6520 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
6523 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
6530 void show_state_filter(unsigned long state_filter)
6532 struct task_struct *g, *p;
6535 for_each_process_thread(g, p) {
6537 * reset the NMI-timeout, listing all files on a slow
6538 * console might take a lot of time:
6539 * Also, reset softlockup watchdogs on all CPUs, because
6540 * another CPU might be blocked waiting for us to process
6543 touch_nmi_watchdog();
6544 touch_all_softlockup_watchdogs();
6545 if (state_filter_match(state_filter, p))
6549 #ifdef CONFIG_SCHED_DEBUG
6551 sysrq_sched_debug_show();
6555 * Only show locks if all tasks are dumped:
6558 debug_show_all_locks();
6562 * init_idle - set up an idle thread for a given CPU
6563 * @idle: task in question
6564 * @cpu: CPU the idle task belongs to
6566 * NOTE: this function does not set the idle thread's NEED_RESCHED
6567 * flag, to make booting more robust.
6569 void init_idle(struct task_struct *idle, int cpu)
6571 struct rq *rq = cpu_rq(cpu);
6572 unsigned long flags;
6574 __sched_fork(0, idle);
6576 raw_spin_lock_irqsave(&idle->pi_lock, flags);
6577 raw_spin_lock(&rq->lock);
6579 idle->state = TASK_RUNNING;
6580 idle->se.exec_start = sched_clock();
6581 idle->flags |= PF_IDLE;
6583 scs_task_reset(idle);
6584 kasan_unpoison_task_stack(idle);
6588 * Its possible that init_idle() gets called multiple times on a task,
6589 * in that case do_set_cpus_allowed() will not do the right thing.
6591 * And since this is boot we can forgo the serialization.
6593 set_cpus_allowed_common(idle, cpumask_of(cpu));
6596 * We're having a chicken and egg problem, even though we are
6597 * holding rq->lock, the CPU isn't yet set to this CPU so the
6598 * lockdep check in task_group() will fail.
6600 * Similar case to sched_fork(). / Alternatively we could
6601 * use task_rq_lock() here and obtain the other rq->lock.
6606 __set_task_cpu(idle, cpu);
6610 rcu_assign_pointer(rq->curr, idle);
6611 idle->on_rq = TASK_ON_RQ_QUEUED;
6615 raw_spin_unlock(&rq->lock);
6616 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
6618 /* Set the preempt count _outside_ the spinlocks! */
6619 init_idle_preempt_count(idle, cpu);
6622 * The idle tasks have their own, simple scheduling class:
6624 idle->sched_class = &idle_sched_class;
6625 ftrace_graph_init_idle_task(idle, cpu);
6626 vtime_init_idle(idle, cpu);
6628 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6634 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
6635 const struct cpumask *trial)
6639 if (!cpumask_weight(cur))
6642 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
6647 int task_can_attach(struct task_struct *p,
6648 const struct cpumask *cs_cpus_allowed)
6653 * Kthreads which disallow setaffinity shouldn't be moved
6654 * to a new cpuset; we don't want to change their CPU
6655 * affinity and isolating such threads by their set of
6656 * allowed nodes is unnecessary. Thus, cpusets are not
6657 * applicable for such threads. This prevents checking for
6658 * success of set_cpus_allowed_ptr() on all attached tasks
6659 * before cpus_mask may be changed.
6661 if (p->flags & PF_NO_SETAFFINITY) {
6666 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
6668 ret = dl_task_can_attach(p, cs_cpus_allowed);
6674 bool sched_smp_initialized __read_mostly;
6676 #ifdef CONFIG_NUMA_BALANCING
6677 /* Migrate current task p to target_cpu */
6678 int migrate_task_to(struct task_struct *p, int target_cpu)
6680 struct migration_arg arg = { p, target_cpu };
6681 int curr_cpu = task_cpu(p);
6683 if (curr_cpu == target_cpu)
6686 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
6689 /* TODO: This is not properly updating schedstats */
6691 trace_sched_move_numa(p, curr_cpu, target_cpu);
6692 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
6696 * Requeue a task on a given node and accurately track the number of NUMA
6697 * tasks on the runqueues
6699 void sched_setnuma(struct task_struct *p, int nid)
6701 bool queued, running;
6705 rq = task_rq_lock(p, &rf);
6706 queued = task_on_rq_queued(p);
6707 running = task_current(rq, p);
6710 dequeue_task(rq, p, DEQUEUE_SAVE);
6712 put_prev_task(rq, p);
6714 p->numa_preferred_nid = nid;
6717 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
6719 set_next_task(rq, p);
6720 task_rq_unlock(rq, p, &rf);
6722 #endif /* CONFIG_NUMA_BALANCING */
6724 #ifdef CONFIG_HOTPLUG_CPU
6726 * Ensure that the idle task is using init_mm right before its CPU goes
6729 void idle_task_exit(void)
6731 struct mm_struct *mm = current->active_mm;
6733 BUG_ON(cpu_online(smp_processor_id()));
6734 BUG_ON(current != this_rq()->idle);
6736 if (mm != &init_mm) {
6737 switch_mm(mm, &init_mm, current);
6738 finish_arch_post_lock_switch();
6741 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
6744 static int __balance_push_cpu_stop(void *arg)
6746 struct task_struct *p = arg;
6747 struct rq *rq = this_rq();
6751 raw_spin_lock_irq(&p->pi_lock);
6754 update_rq_clock(rq);
6756 if (task_rq(p) == rq && task_on_rq_queued(p)) {
6757 cpu = select_fallback_rq(rq->cpu, p);
6758 rq = __migrate_task(rq, &rf, p, cpu);
6762 raw_spin_unlock_irq(&p->pi_lock);
6769 static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
6772 * Ensure we only run per-cpu kthreads once the CPU goes !active.
6774 static void balance_push(struct rq *rq)
6776 struct task_struct *push_task = rq->curr;
6778 lockdep_assert_held(&rq->lock);
6779 SCHED_WARN_ON(rq->cpu != smp_processor_id());
6782 * Both the cpu-hotplug and stop task are in this case and are
6783 * required to complete the hotplug process.
6785 if (is_per_cpu_kthread(push_task)) {
6787 * If this is the idle task on the outgoing CPU try to wake
6788 * up the hotplug control thread which might wait for the
6789 * last task to vanish. The rcuwait_active() check is
6790 * accurate here because the waiter is pinned on this CPU
6791 * and can't obviously be running in parallel.
6793 if (!rq->nr_running && rcuwait_active(&rq->hotplug_wait)) {
6794 raw_spin_unlock(&rq->lock);
6795 rcuwait_wake_up(&rq->hotplug_wait);
6796 raw_spin_lock(&rq->lock);
6801 get_task_struct(push_task);
6803 * Temporarily drop rq->lock such that we can wake-up the stop task.
6804 * Both preemption and IRQs are still disabled.
6806 raw_spin_unlock(&rq->lock);
6807 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
6808 this_cpu_ptr(&push_work));
6810 * At this point need_resched() is true and we'll take the loop in
6811 * schedule(). The next pick is obviously going to be the stop task
6812 * which is_per_cpu_kthread() and will push this task away.
6814 raw_spin_lock(&rq->lock);
6817 static void balance_push_set(int cpu, bool on)
6819 struct rq *rq = cpu_rq(cpu);
6822 rq_lock_irqsave(rq, &rf);
6824 rq->balance_flags |= BALANCE_PUSH;
6826 rq->balance_flags &= ~BALANCE_PUSH;
6827 rq_unlock_irqrestore(rq, &rf);
6831 * Invoked from a CPUs hotplug control thread after the CPU has been marked
6832 * inactive. All tasks which are not per CPU kernel threads are either
6833 * pushed off this CPU now via balance_push() or placed on a different CPU
6834 * during wakeup. Wait until the CPU is quiescent.
6836 static void balance_hotplug_wait(void)
6838 struct rq *rq = this_rq();
6840 rcuwait_wait_event(&rq->hotplug_wait, rq->nr_running == 1,
6841 TASK_UNINTERRUPTIBLE);
6846 static inline void balance_push(struct rq *rq)
6850 static inline void balance_push_set(int cpu, bool on)
6854 static inline void balance_hotplug_wait(void)
6858 #endif /* CONFIG_HOTPLUG_CPU */
6860 void set_rq_online(struct rq *rq)
6863 const struct sched_class *class;
6865 cpumask_set_cpu(rq->cpu, rq->rd->online);
6868 for_each_class(class) {
6869 if (class->rq_online)
6870 class->rq_online(rq);
6875 void set_rq_offline(struct rq *rq)
6878 const struct sched_class *class;
6880 for_each_class(class) {
6881 if (class->rq_offline)
6882 class->rq_offline(rq);
6885 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6891 * used to mark begin/end of suspend/resume:
6893 static int num_cpus_frozen;
6896 * Update cpusets according to cpu_active mask. If cpusets are
6897 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6898 * around partition_sched_domains().
6900 * If we come here as part of a suspend/resume, don't touch cpusets because we
6901 * want to restore it back to its original state upon resume anyway.
6903 static void cpuset_cpu_active(void)
6905 if (cpuhp_tasks_frozen) {
6907 * num_cpus_frozen tracks how many CPUs are involved in suspend
6908 * resume sequence. As long as this is not the last online
6909 * operation in the resume sequence, just build a single sched
6910 * domain, ignoring cpusets.
6912 partition_sched_domains(1, NULL, NULL);
6913 if (--num_cpus_frozen)
6916 * This is the last CPU online operation. So fall through and
6917 * restore the original sched domains by considering the
6918 * cpuset configurations.
6920 cpuset_force_rebuild();
6922 cpuset_update_active_cpus();
6925 static int cpuset_cpu_inactive(unsigned int cpu)
6927 if (!cpuhp_tasks_frozen) {
6928 if (dl_cpu_busy(cpu))
6930 cpuset_update_active_cpus();
6933 partition_sched_domains(1, NULL, NULL);
6938 int sched_cpu_activate(unsigned int cpu)
6940 struct rq *rq = cpu_rq(cpu);
6943 balance_push_set(cpu, false);
6945 #ifdef CONFIG_SCHED_SMT
6947 * When going up, increment the number of cores with SMT present.
6949 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6950 static_branch_inc_cpuslocked(&sched_smt_present);
6952 set_cpu_active(cpu, true);
6954 if (sched_smp_initialized) {
6955 sched_domains_numa_masks_set(cpu);
6956 cpuset_cpu_active();
6960 * Put the rq online, if not already. This happens:
6962 * 1) In the early boot process, because we build the real domains
6963 * after all CPUs have been brought up.
6965 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
6968 rq_lock_irqsave(rq, &rf);
6970 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6973 rq_unlock_irqrestore(rq, &rf);
6978 int sched_cpu_deactivate(unsigned int cpu)
6982 set_cpu_active(cpu, false);
6984 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
6985 * users of this state to go away such that all new such users will
6988 * Do sync before park smpboot threads to take care the rcu boost case.
6992 balance_push_set(cpu, true);
6994 #ifdef CONFIG_SCHED_SMT
6996 * When going down, decrement the number of cores with SMT present.
6998 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
6999 static_branch_dec_cpuslocked(&sched_smt_present);
7002 if (!sched_smp_initialized)
7005 ret = cpuset_cpu_inactive(cpu);
7007 balance_push_set(cpu, false);
7008 set_cpu_active(cpu, true);
7011 sched_domains_numa_masks_clear(cpu);
7015 static void sched_rq_cpu_starting(unsigned int cpu)
7017 struct rq *rq = cpu_rq(cpu);
7019 rq->calc_load_update = calc_load_update;
7020 update_max_interval();
7023 int sched_cpu_starting(unsigned int cpu)
7025 sched_rq_cpu_starting(cpu);
7026 sched_tick_start(cpu);
7030 #ifdef CONFIG_HOTPLUG_CPU
7033 * Invoked immediately before the stopper thread is invoked to bring the
7034 * CPU down completely. At this point all per CPU kthreads except the
7035 * hotplug thread (current) and the stopper thread (inactive) have been
7036 * either parked or have been unbound from the outgoing CPU. Ensure that
7037 * any of those which might be on the way out are gone.
7039 * If after this point a bound task is being woken on this CPU then the
7040 * responsible hotplug callback has failed to do it's job.
7041 * sched_cpu_dying() will catch it with the appropriate fireworks.
7043 int sched_cpu_wait_empty(unsigned int cpu)
7045 balance_hotplug_wait();
7050 * Since this CPU is going 'away' for a while, fold any nr_active delta we
7051 * might have. Called from the CPU stopper task after ensuring that the
7052 * stopper is the last running task on the CPU, so nr_active count is
7053 * stable. We need to take the teardown thread which is calling this into
7054 * account, so we hand in adjust = 1 to the load calculation.
7056 * Also see the comment "Global load-average calculations".
7058 static void calc_load_migrate(struct rq *rq)
7060 long delta = calc_load_fold_active(rq, 1);
7063 atomic_long_add(delta, &calc_load_tasks);
7066 int sched_cpu_dying(unsigned int cpu)
7068 struct rq *rq = cpu_rq(cpu);
7071 /* Handle pending wakeups and then migrate everything off */
7072 sched_tick_stop(cpu);
7074 rq_lock_irqsave(rq, &rf);
7076 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7079 BUG_ON(rq->nr_running != 1);
7080 rq_unlock_irqrestore(rq, &rf);
7082 calc_load_migrate(rq);
7083 update_max_interval();
7084 nohz_balance_exit_idle(rq);
7090 void __init sched_init_smp(void)
7095 * There's no userspace yet to cause hotplug operations; hence all the
7096 * CPU masks are stable and all blatant races in the below code cannot
7099 mutex_lock(&sched_domains_mutex);
7100 sched_init_domains(cpu_active_mask);
7101 mutex_unlock(&sched_domains_mutex);
7103 /* Move init over to a non-isolated CPU */
7104 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
7106 sched_init_granularity();
7108 init_sched_rt_class();
7109 init_sched_dl_class();
7111 sched_smp_initialized = true;
7114 static int __init migration_init(void)
7116 sched_cpu_starting(smp_processor_id());
7119 early_initcall(migration_init);
7122 void __init sched_init_smp(void)
7124 sched_init_granularity();
7126 #endif /* CONFIG_SMP */
7128 int in_sched_functions(unsigned long addr)
7130 return in_lock_functions(addr) ||
7131 (addr >= (unsigned long)__sched_text_start
7132 && addr < (unsigned long)__sched_text_end);
7135 #ifdef CONFIG_CGROUP_SCHED
7137 * Default task group.
7138 * Every task in system belongs to this group at bootup.
7140 struct task_group root_task_group;
7141 LIST_HEAD(task_groups);
7143 /* Cacheline aligned slab cache for task_group */
7144 static struct kmem_cache *task_group_cache __read_mostly;
7147 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7148 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7150 void __init sched_init(void)
7152 unsigned long ptr = 0;
7155 /* Make sure the linker didn't screw up */
7156 BUG_ON(&idle_sched_class + 1 != &fair_sched_class ||
7157 &fair_sched_class + 1 != &rt_sched_class ||
7158 &rt_sched_class + 1 != &dl_sched_class);
7160 BUG_ON(&dl_sched_class + 1 != &stop_sched_class);
7165 #ifdef CONFIG_FAIR_GROUP_SCHED
7166 ptr += 2 * nr_cpu_ids * sizeof(void **);
7168 #ifdef CONFIG_RT_GROUP_SCHED
7169 ptr += 2 * nr_cpu_ids * sizeof(void **);
7172 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
7174 #ifdef CONFIG_FAIR_GROUP_SCHED
7175 root_task_group.se = (struct sched_entity **)ptr;
7176 ptr += nr_cpu_ids * sizeof(void **);
7178 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7179 ptr += nr_cpu_ids * sizeof(void **);
7181 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7182 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7183 #endif /* CONFIG_FAIR_GROUP_SCHED */
7184 #ifdef CONFIG_RT_GROUP_SCHED
7185 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7186 ptr += nr_cpu_ids * sizeof(void **);
7188 root_task_group.rt_rq = (struct rt_rq **)ptr;
7189 ptr += nr_cpu_ids * sizeof(void **);
7191 #endif /* CONFIG_RT_GROUP_SCHED */
7193 #ifdef CONFIG_CPUMASK_OFFSTACK
7194 for_each_possible_cpu(i) {
7195 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7196 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7197 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7198 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7200 #endif /* CONFIG_CPUMASK_OFFSTACK */
7202 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
7203 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
7206 init_defrootdomain();
7209 #ifdef CONFIG_RT_GROUP_SCHED
7210 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7211 global_rt_period(), global_rt_runtime());
7212 #endif /* CONFIG_RT_GROUP_SCHED */
7214 #ifdef CONFIG_CGROUP_SCHED
7215 task_group_cache = KMEM_CACHE(task_group, 0);
7217 list_add(&root_task_group.list, &task_groups);
7218 INIT_LIST_HEAD(&root_task_group.children);
7219 INIT_LIST_HEAD(&root_task_group.siblings);
7220 autogroup_init(&init_task);
7221 #endif /* CONFIG_CGROUP_SCHED */
7223 for_each_possible_cpu(i) {
7227 raw_spin_lock_init(&rq->lock);
7229 rq->calc_load_active = 0;
7230 rq->calc_load_update = jiffies + LOAD_FREQ;
7231 init_cfs_rq(&rq->cfs);
7232 init_rt_rq(&rq->rt);
7233 init_dl_rq(&rq->dl);
7234 #ifdef CONFIG_FAIR_GROUP_SCHED
7235 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7236 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7238 * How much CPU bandwidth does root_task_group get?
7240 * In case of task-groups formed thr' the cgroup filesystem, it
7241 * gets 100% of the CPU resources in the system. This overall
7242 * system CPU resource is divided among the tasks of
7243 * root_task_group and its child task-groups in a fair manner,
7244 * based on each entity's (task or task-group's) weight
7245 * (se->load.weight).
7247 * In other words, if root_task_group has 10 tasks of weight
7248 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7249 * then A0's share of the CPU resource is:
7251 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7253 * We achieve this by letting root_task_group's tasks sit
7254 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7256 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7257 #endif /* CONFIG_FAIR_GROUP_SCHED */
7259 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7260 #ifdef CONFIG_RT_GROUP_SCHED
7261 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7266 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7267 rq->balance_callback = NULL;
7268 rq->active_balance = 0;
7269 rq->next_balance = jiffies;
7274 rq->avg_idle = 2*sysctl_sched_migration_cost;
7275 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7277 INIT_LIST_HEAD(&rq->cfs_tasks);
7279 rq_attach_root(rq, &def_root_domain);
7280 #ifdef CONFIG_NO_HZ_COMMON
7281 rq->last_blocked_load_update_tick = jiffies;
7282 atomic_set(&rq->nohz_flags, 0);
7284 rq_csd_init(rq, &rq->nohz_csd, nohz_csd_func);
7286 #ifdef CONFIG_HOTPLUG_CPU
7287 rcuwait_init(&rq->hotplug_wait);
7289 #endif /* CONFIG_SMP */
7291 atomic_set(&rq->nr_iowait, 0);
7294 set_load_weight(&init_task, false);
7297 * The boot idle thread does lazy MMU switching as well:
7300 enter_lazy_tlb(&init_mm, current);
7303 * Make us the idle thread. Technically, schedule() should not be
7304 * called from this thread, however somewhere below it might be,
7305 * but because we are the idle thread, we just pick up running again
7306 * when this runqueue becomes "idle".
7308 init_idle(current, smp_processor_id());
7310 calc_load_update = jiffies + LOAD_FREQ;
7313 idle_thread_set_boot_cpu();
7315 init_sched_fair_class();
7323 scheduler_running = 1;
7326 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7327 static inline int preempt_count_equals(int preempt_offset)
7329 int nested = preempt_count() + rcu_preempt_depth();
7331 return (nested == preempt_offset);
7334 void __might_sleep(const char *file, int line, int preempt_offset)
7337 * Blocking primitives will set (and therefore destroy) current->state,
7338 * since we will exit with TASK_RUNNING make sure we enter with it,
7339 * otherwise we will destroy state.
7341 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7342 "do not call blocking ops when !TASK_RUNNING; "
7343 "state=%lx set at [<%p>] %pS\n",
7345 (void *)current->task_state_change,
7346 (void *)current->task_state_change);
7348 ___might_sleep(file, line, preempt_offset);
7350 EXPORT_SYMBOL(__might_sleep);
7352 void ___might_sleep(const char *file, int line, int preempt_offset)
7354 /* Ratelimiting timestamp: */
7355 static unsigned long prev_jiffy;
7357 unsigned long preempt_disable_ip;
7359 /* WARN_ON_ONCE() by default, no rate limit required: */
7362 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7363 !is_idle_task(current) && !current->non_block_count) ||
7364 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
7368 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7370 prev_jiffy = jiffies;
7372 /* Save this before calling printk(), since that will clobber it: */
7373 preempt_disable_ip = get_preempt_disable_ip(current);
7376 "BUG: sleeping function called from invalid context at %s:%d\n",
7379 "in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
7380 in_atomic(), irqs_disabled(), current->non_block_count,
7381 current->pid, current->comm);
7383 if (task_stack_end_corrupted(current))
7384 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7386 debug_show_held_locks(current);
7387 if (irqs_disabled())
7388 print_irqtrace_events(current);
7389 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7390 && !preempt_count_equals(preempt_offset)) {
7391 pr_err("Preemption disabled at:");
7392 print_ip_sym(KERN_ERR, preempt_disable_ip);
7395 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7397 EXPORT_SYMBOL(___might_sleep);
7399 void __cant_sleep(const char *file, int line, int preempt_offset)
7401 static unsigned long prev_jiffy;
7403 if (irqs_disabled())
7406 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
7409 if (preempt_count() > preempt_offset)
7412 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7414 prev_jiffy = jiffies;
7416 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
7417 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7418 in_atomic(), irqs_disabled(),
7419 current->pid, current->comm);
7421 debug_show_held_locks(current);
7423 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7425 EXPORT_SYMBOL_GPL(__cant_sleep);
7428 #ifdef CONFIG_MAGIC_SYSRQ
7429 void normalize_rt_tasks(void)
7431 struct task_struct *g, *p;
7432 struct sched_attr attr = {
7433 .sched_policy = SCHED_NORMAL,
7436 read_lock(&tasklist_lock);
7437 for_each_process_thread(g, p) {
7439 * Only normalize user tasks:
7441 if (p->flags & PF_KTHREAD)
7444 p->se.exec_start = 0;
7445 schedstat_set(p->se.statistics.wait_start, 0);
7446 schedstat_set(p->se.statistics.sleep_start, 0);
7447 schedstat_set(p->se.statistics.block_start, 0);
7449 if (!dl_task(p) && !rt_task(p)) {
7451 * Renice negative nice level userspace
7454 if (task_nice(p) < 0)
7455 set_user_nice(p, 0);
7459 __sched_setscheduler(p, &attr, false, false);
7461 read_unlock(&tasklist_lock);
7464 #endif /* CONFIG_MAGIC_SYSRQ */
7466 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7468 * These functions are only useful for the IA64 MCA handling, or kdb.
7470 * They can only be called when the whole system has been
7471 * stopped - every CPU needs to be quiescent, and no scheduling
7472 * activity can take place. Using them for anything else would
7473 * be a serious bug, and as a result, they aren't even visible
7474 * under any other configuration.
7478 * curr_task - return the current task for a given CPU.
7479 * @cpu: the processor in question.
7481 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7483 * Return: The current task for @cpu.
7485 struct task_struct *curr_task(int cpu)
7487 return cpu_curr(cpu);
7490 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7494 * ia64_set_curr_task - set the current task for a given CPU.
7495 * @cpu: the processor in question.
7496 * @p: the task pointer to set.
7498 * Description: This function must only be used when non-maskable interrupts
7499 * are serviced on a separate stack. It allows the architecture to switch the
7500 * notion of the current task on a CPU in a non-blocking manner. This function
7501 * must be called with all CPU's synchronized, and interrupts disabled, the
7502 * and caller must save the original value of the current task (see
7503 * curr_task() above) and restore that value before reenabling interrupts and
7504 * re-starting the system.
7506 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7508 void ia64_set_curr_task(int cpu, struct task_struct *p)
7515 #ifdef CONFIG_CGROUP_SCHED
7516 /* task_group_lock serializes the addition/removal of task groups */
7517 static DEFINE_SPINLOCK(task_group_lock);
7519 static inline void alloc_uclamp_sched_group(struct task_group *tg,
7520 struct task_group *parent)
7522 #ifdef CONFIG_UCLAMP_TASK_GROUP
7523 enum uclamp_id clamp_id;
7525 for_each_clamp_id(clamp_id) {
7526 uclamp_se_set(&tg->uclamp_req[clamp_id],
7527 uclamp_none(clamp_id), false);
7528 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
7533 static void sched_free_group(struct task_group *tg)
7535 free_fair_sched_group(tg);
7536 free_rt_sched_group(tg);
7538 kmem_cache_free(task_group_cache, tg);
7541 /* allocate runqueue etc for a new task group */
7542 struct task_group *sched_create_group(struct task_group *parent)
7544 struct task_group *tg;
7546 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7548 return ERR_PTR(-ENOMEM);
7550 if (!alloc_fair_sched_group(tg, parent))
7553 if (!alloc_rt_sched_group(tg, parent))
7556 alloc_uclamp_sched_group(tg, parent);
7561 sched_free_group(tg);
7562 return ERR_PTR(-ENOMEM);
7565 void sched_online_group(struct task_group *tg, struct task_group *parent)
7567 unsigned long flags;
7569 spin_lock_irqsave(&task_group_lock, flags);
7570 list_add_rcu(&tg->list, &task_groups);
7572 /* Root should already exist: */
7575 tg->parent = parent;
7576 INIT_LIST_HEAD(&tg->children);
7577 list_add_rcu(&tg->siblings, &parent->children);
7578 spin_unlock_irqrestore(&task_group_lock, flags);
7580 online_fair_sched_group(tg);
7583 /* rcu callback to free various structures associated with a task group */
7584 static void sched_free_group_rcu(struct rcu_head *rhp)
7586 /* Now it should be safe to free those cfs_rqs: */
7587 sched_free_group(container_of(rhp, struct task_group, rcu));
7590 void sched_destroy_group(struct task_group *tg)
7592 /* Wait for possible concurrent references to cfs_rqs complete: */
7593 call_rcu(&tg->rcu, sched_free_group_rcu);
7596 void sched_offline_group(struct task_group *tg)
7598 unsigned long flags;
7600 /* End participation in shares distribution: */
7601 unregister_fair_sched_group(tg);
7603 spin_lock_irqsave(&task_group_lock, flags);
7604 list_del_rcu(&tg->list);
7605 list_del_rcu(&tg->siblings);
7606 spin_unlock_irqrestore(&task_group_lock, flags);
7609 static void sched_change_group(struct task_struct *tsk, int type)
7611 struct task_group *tg;
7614 * All callers are synchronized by task_rq_lock(); we do not use RCU
7615 * which is pointless here. Thus, we pass "true" to task_css_check()
7616 * to prevent lockdep warnings.
7618 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7619 struct task_group, css);
7620 tg = autogroup_task_group(tsk, tg);
7621 tsk->sched_task_group = tg;
7623 #ifdef CONFIG_FAIR_GROUP_SCHED
7624 if (tsk->sched_class->task_change_group)
7625 tsk->sched_class->task_change_group(tsk, type);
7628 set_task_rq(tsk, task_cpu(tsk));
7632 * Change task's runqueue when it moves between groups.
7634 * The caller of this function should have put the task in its new group by
7635 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7638 void sched_move_task(struct task_struct *tsk)
7640 int queued, running, queue_flags =
7641 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
7645 rq = task_rq_lock(tsk, &rf);
7646 update_rq_clock(rq);
7648 running = task_current(rq, tsk);
7649 queued = task_on_rq_queued(tsk);
7652 dequeue_task(rq, tsk, queue_flags);
7654 put_prev_task(rq, tsk);
7656 sched_change_group(tsk, TASK_MOVE_GROUP);
7659 enqueue_task(rq, tsk, queue_flags);
7661 set_next_task(rq, tsk);
7663 * After changing group, the running task may have joined a
7664 * throttled one but it's still the running task. Trigger a
7665 * resched to make sure that task can still run.
7670 task_rq_unlock(rq, tsk, &rf);
7673 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7675 return css ? container_of(css, struct task_group, css) : NULL;
7678 static struct cgroup_subsys_state *
7679 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7681 struct task_group *parent = css_tg(parent_css);
7682 struct task_group *tg;
7685 /* This is early initialization for the top cgroup */
7686 return &root_task_group.css;
7689 tg = sched_create_group(parent);
7691 return ERR_PTR(-ENOMEM);
7696 /* Expose task group only after completing cgroup initialization */
7697 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7699 struct task_group *tg = css_tg(css);
7700 struct task_group *parent = css_tg(css->parent);
7703 sched_online_group(tg, parent);
7705 #ifdef CONFIG_UCLAMP_TASK_GROUP
7706 /* Propagate the effective uclamp value for the new group */
7707 cpu_util_update_eff(css);
7713 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
7715 struct task_group *tg = css_tg(css);
7717 sched_offline_group(tg);
7720 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7722 struct task_group *tg = css_tg(css);
7725 * Relies on the RCU grace period between css_released() and this.
7727 sched_free_group(tg);
7731 * This is called before wake_up_new_task(), therefore we really only
7732 * have to set its group bits, all the other stuff does not apply.
7734 static void cpu_cgroup_fork(struct task_struct *task)
7739 rq = task_rq_lock(task, &rf);
7741 update_rq_clock(rq);
7742 sched_change_group(task, TASK_SET_GROUP);
7744 task_rq_unlock(rq, task, &rf);
7747 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
7749 struct task_struct *task;
7750 struct cgroup_subsys_state *css;
7753 cgroup_taskset_for_each(task, css, tset) {
7754 #ifdef CONFIG_RT_GROUP_SCHED
7755 if (!sched_rt_can_attach(css_tg(css), task))
7759 * Serialize against wake_up_new_task() such that if its
7760 * running, we're sure to observe its full state.
7762 raw_spin_lock_irq(&task->pi_lock);
7764 * Avoid calling sched_move_task() before wake_up_new_task()
7765 * has happened. This would lead to problems with PELT, due to
7766 * move wanting to detach+attach while we're not attached yet.
7768 if (task->state == TASK_NEW)
7770 raw_spin_unlock_irq(&task->pi_lock);
7778 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
7780 struct task_struct *task;
7781 struct cgroup_subsys_state *css;
7783 cgroup_taskset_for_each(task, css, tset)
7784 sched_move_task(task);
7787 #ifdef CONFIG_UCLAMP_TASK_GROUP
7788 static void cpu_util_update_eff(struct cgroup_subsys_state *css)
7790 struct cgroup_subsys_state *top_css = css;
7791 struct uclamp_se *uc_parent = NULL;
7792 struct uclamp_se *uc_se = NULL;
7793 unsigned int eff[UCLAMP_CNT];
7794 enum uclamp_id clamp_id;
7795 unsigned int clamps;
7797 css_for_each_descendant_pre(css, top_css) {
7798 uc_parent = css_tg(css)->parent
7799 ? css_tg(css)->parent->uclamp : NULL;
7801 for_each_clamp_id(clamp_id) {
7802 /* Assume effective clamps matches requested clamps */
7803 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
7804 /* Cap effective clamps with parent's effective clamps */
7806 eff[clamp_id] > uc_parent[clamp_id].value) {
7807 eff[clamp_id] = uc_parent[clamp_id].value;
7810 /* Ensure protection is always capped by limit */
7811 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
7813 /* Propagate most restrictive effective clamps */
7815 uc_se = css_tg(css)->uclamp;
7816 for_each_clamp_id(clamp_id) {
7817 if (eff[clamp_id] == uc_se[clamp_id].value)
7819 uc_se[clamp_id].value = eff[clamp_id];
7820 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
7821 clamps |= (0x1 << clamp_id);
7824 css = css_rightmost_descendant(css);
7828 /* Immediately update descendants RUNNABLE tasks */
7829 uclamp_update_active_tasks(css, clamps);
7834 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
7835 * C expression. Since there is no way to convert a macro argument (N) into a
7836 * character constant, use two levels of macros.
7838 #define _POW10(exp) ((unsigned int)1e##exp)
7839 #define POW10(exp) _POW10(exp)
7841 struct uclamp_request {
7842 #define UCLAMP_PERCENT_SHIFT 2
7843 #define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
7849 static inline struct uclamp_request
7850 capacity_from_percent(char *buf)
7852 struct uclamp_request req = {
7853 .percent = UCLAMP_PERCENT_SCALE,
7854 .util = SCHED_CAPACITY_SCALE,
7859 if (strcmp(buf, "max")) {
7860 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
7864 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
7869 req.util = req.percent << SCHED_CAPACITY_SHIFT;
7870 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
7876 static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
7877 size_t nbytes, loff_t off,
7878 enum uclamp_id clamp_id)
7880 struct uclamp_request req;
7881 struct task_group *tg;
7883 req = capacity_from_percent(buf);
7887 static_branch_enable(&sched_uclamp_used);
7889 mutex_lock(&uclamp_mutex);
7892 tg = css_tg(of_css(of));
7893 if (tg->uclamp_req[clamp_id].value != req.util)
7894 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
7897 * Because of not recoverable conversion rounding we keep track of the
7898 * exact requested value
7900 tg->uclamp_pct[clamp_id] = req.percent;
7902 /* Update effective clamps to track the most restrictive value */
7903 cpu_util_update_eff(of_css(of));
7906 mutex_unlock(&uclamp_mutex);
7911 static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
7912 char *buf, size_t nbytes,
7915 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
7918 static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
7919 char *buf, size_t nbytes,
7922 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
7925 static inline void cpu_uclamp_print(struct seq_file *sf,
7926 enum uclamp_id clamp_id)
7928 struct task_group *tg;
7934 tg = css_tg(seq_css(sf));
7935 util_clamp = tg->uclamp_req[clamp_id].value;
7938 if (util_clamp == SCHED_CAPACITY_SCALE) {
7939 seq_puts(sf, "max\n");
7943 percent = tg->uclamp_pct[clamp_id];
7944 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
7945 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
7948 static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
7950 cpu_uclamp_print(sf, UCLAMP_MIN);
7954 static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
7956 cpu_uclamp_print(sf, UCLAMP_MAX);
7959 #endif /* CONFIG_UCLAMP_TASK_GROUP */
7961 #ifdef CONFIG_FAIR_GROUP_SCHED
7962 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7963 struct cftype *cftype, u64 shareval)
7965 if (shareval > scale_load_down(ULONG_MAX))
7966 shareval = MAX_SHARES;
7967 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7970 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7973 struct task_group *tg = css_tg(css);
7975 return (u64) scale_load_down(tg->shares);
7978 #ifdef CONFIG_CFS_BANDWIDTH
7979 static DEFINE_MUTEX(cfs_constraints_mutex);
7981 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7982 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7983 /* More than 203 days if BW_SHIFT equals 20. */
7984 static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
7986 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7988 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7990 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7991 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7993 if (tg == &root_task_group)
7997 * Ensure we have at some amount of bandwidth every period. This is
7998 * to prevent reaching a state of large arrears when throttled via
7999 * entity_tick() resulting in prolonged exit starvation.
8001 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8005 * Likewise, bound things on the otherside by preventing insane quota
8006 * periods. This also allows us to normalize in computing quota
8009 if (period > max_cfs_quota_period)
8013 * Bound quota to defend quota against overflow during bandwidth shift.
8015 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
8019 * Prevent race between setting of cfs_rq->runtime_enabled and
8020 * unthrottle_offline_cfs_rqs().
8023 mutex_lock(&cfs_constraints_mutex);
8024 ret = __cfs_schedulable(tg, period, quota);
8028 runtime_enabled = quota != RUNTIME_INF;
8029 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8031 * If we need to toggle cfs_bandwidth_used, off->on must occur
8032 * before making related changes, and on->off must occur afterwards
8034 if (runtime_enabled && !runtime_was_enabled)
8035 cfs_bandwidth_usage_inc();
8036 raw_spin_lock_irq(&cfs_b->lock);
8037 cfs_b->period = ns_to_ktime(period);
8038 cfs_b->quota = quota;
8040 __refill_cfs_bandwidth_runtime(cfs_b);
8042 /* Restart the period timer (if active) to handle new period expiry: */
8043 if (runtime_enabled)
8044 start_cfs_bandwidth(cfs_b);
8046 raw_spin_unlock_irq(&cfs_b->lock);
8048 for_each_online_cpu(i) {
8049 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8050 struct rq *rq = cfs_rq->rq;
8053 rq_lock_irq(rq, &rf);
8054 cfs_rq->runtime_enabled = runtime_enabled;
8055 cfs_rq->runtime_remaining = 0;
8057 if (cfs_rq->throttled)
8058 unthrottle_cfs_rq(cfs_rq);
8059 rq_unlock_irq(rq, &rf);
8061 if (runtime_was_enabled && !runtime_enabled)
8062 cfs_bandwidth_usage_dec();
8064 mutex_unlock(&cfs_constraints_mutex);
8070 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8074 period = ktime_to_ns(tg->cfs_bandwidth.period);
8075 if (cfs_quota_us < 0)
8076 quota = RUNTIME_INF;
8077 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
8078 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8082 return tg_set_cfs_bandwidth(tg, period, quota);
8085 static long tg_get_cfs_quota(struct task_group *tg)
8089 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8092 quota_us = tg->cfs_bandwidth.quota;
8093 do_div(quota_us, NSEC_PER_USEC);
8098 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8102 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
8105 period = (u64)cfs_period_us * NSEC_PER_USEC;
8106 quota = tg->cfs_bandwidth.quota;
8108 return tg_set_cfs_bandwidth(tg, period, quota);
8111 static long tg_get_cfs_period(struct task_group *tg)
8115 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8116 do_div(cfs_period_us, NSEC_PER_USEC);
8118 return cfs_period_us;
8121 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8124 return tg_get_cfs_quota(css_tg(css));
8127 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8128 struct cftype *cftype, s64 cfs_quota_us)
8130 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8133 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8136 return tg_get_cfs_period(css_tg(css));
8139 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8140 struct cftype *cftype, u64 cfs_period_us)
8142 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8145 struct cfs_schedulable_data {
8146 struct task_group *tg;
8151 * normalize group quota/period to be quota/max_period
8152 * note: units are usecs
8154 static u64 normalize_cfs_quota(struct task_group *tg,
8155 struct cfs_schedulable_data *d)
8163 period = tg_get_cfs_period(tg);
8164 quota = tg_get_cfs_quota(tg);
8167 /* note: these should typically be equivalent */
8168 if (quota == RUNTIME_INF || quota == -1)
8171 return to_ratio(period, quota);
8174 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8176 struct cfs_schedulable_data *d = data;
8177 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8178 s64 quota = 0, parent_quota = -1;
8181 quota = RUNTIME_INF;
8183 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8185 quota = normalize_cfs_quota(tg, d);
8186 parent_quota = parent_b->hierarchical_quota;
8189 * Ensure max(child_quota) <= parent_quota. On cgroup2,
8190 * always take the min. On cgroup1, only inherit when no
8193 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
8194 quota = min(quota, parent_quota);
8196 if (quota == RUNTIME_INF)
8197 quota = parent_quota;
8198 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8202 cfs_b->hierarchical_quota = quota;
8207 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8210 struct cfs_schedulable_data data = {
8216 if (quota != RUNTIME_INF) {
8217 do_div(data.period, NSEC_PER_USEC);
8218 do_div(data.quota, NSEC_PER_USEC);
8222 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8228 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
8230 struct task_group *tg = css_tg(seq_css(sf));
8231 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8233 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8234 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8235 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8237 if (schedstat_enabled() && tg != &root_task_group) {
8241 for_each_possible_cpu(i)
8242 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
8244 seq_printf(sf, "wait_sum %llu\n", ws);
8249 #endif /* CONFIG_CFS_BANDWIDTH */
8250 #endif /* CONFIG_FAIR_GROUP_SCHED */
8252 #ifdef CONFIG_RT_GROUP_SCHED
8253 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8254 struct cftype *cft, s64 val)
8256 return sched_group_set_rt_runtime(css_tg(css), val);
8259 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8262 return sched_group_rt_runtime(css_tg(css));
8265 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8266 struct cftype *cftype, u64 rt_period_us)
8268 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8271 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8274 return sched_group_rt_period(css_tg(css));
8276 #endif /* CONFIG_RT_GROUP_SCHED */
8278 static struct cftype cpu_legacy_files[] = {
8279 #ifdef CONFIG_FAIR_GROUP_SCHED
8282 .read_u64 = cpu_shares_read_u64,
8283 .write_u64 = cpu_shares_write_u64,
8286 #ifdef CONFIG_CFS_BANDWIDTH
8288 .name = "cfs_quota_us",
8289 .read_s64 = cpu_cfs_quota_read_s64,
8290 .write_s64 = cpu_cfs_quota_write_s64,
8293 .name = "cfs_period_us",
8294 .read_u64 = cpu_cfs_period_read_u64,
8295 .write_u64 = cpu_cfs_period_write_u64,
8299 .seq_show = cpu_cfs_stat_show,
8302 #ifdef CONFIG_RT_GROUP_SCHED
8304 .name = "rt_runtime_us",
8305 .read_s64 = cpu_rt_runtime_read,
8306 .write_s64 = cpu_rt_runtime_write,
8309 .name = "rt_period_us",
8310 .read_u64 = cpu_rt_period_read_uint,
8311 .write_u64 = cpu_rt_period_write_uint,
8314 #ifdef CONFIG_UCLAMP_TASK_GROUP
8316 .name = "uclamp.min",
8317 .flags = CFTYPE_NOT_ON_ROOT,
8318 .seq_show = cpu_uclamp_min_show,
8319 .write = cpu_uclamp_min_write,
8322 .name = "uclamp.max",
8323 .flags = CFTYPE_NOT_ON_ROOT,
8324 .seq_show = cpu_uclamp_max_show,
8325 .write = cpu_uclamp_max_write,
8331 static int cpu_extra_stat_show(struct seq_file *sf,
8332 struct cgroup_subsys_state *css)
8334 #ifdef CONFIG_CFS_BANDWIDTH
8336 struct task_group *tg = css_tg(css);
8337 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8340 throttled_usec = cfs_b->throttled_time;
8341 do_div(throttled_usec, NSEC_PER_USEC);
8343 seq_printf(sf, "nr_periods %d\n"
8345 "throttled_usec %llu\n",
8346 cfs_b->nr_periods, cfs_b->nr_throttled,
8353 #ifdef CONFIG_FAIR_GROUP_SCHED
8354 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
8357 struct task_group *tg = css_tg(css);
8358 u64 weight = scale_load_down(tg->shares);
8360 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
8363 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
8364 struct cftype *cft, u64 weight)
8367 * cgroup weight knobs should use the common MIN, DFL and MAX
8368 * values which are 1, 100 and 10000 respectively. While it loses
8369 * a bit of range on both ends, it maps pretty well onto the shares
8370 * value used by scheduler and the round-trip conversions preserve
8371 * the original value over the entire range.
8373 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
8376 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
8378 return sched_group_set_shares(css_tg(css), scale_load(weight));
8381 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
8384 unsigned long weight = scale_load_down(css_tg(css)->shares);
8385 int last_delta = INT_MAX;
8388 /* find the closest nice value to the current weight */
8389 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
8390 delta = abs(sched_prio_to_weight[prio] - weight);
8391 if (delta >= last_delta)
8396 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
8399 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
8400 struct cftype *cft, s64 nice)
8402 unsigned long weight;
8405 if (nice < MIN_NICE || nice > MAX_NICE)
8408 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
8409 idx = array_index_nospec(idx, 40);
8410 weight = sched_prio_to_weight[idx];
8412 return sched_group_set_shares(css_tg(css), scale_load(weight));
8416 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
8417 long period, long quota)
8420 seq_puts(sf, "max");
8422 seq_printf(sf, "%ld", quota);
8424 seq_printf(sf, " %ld\n", period);
8427 /* caller should put the current value in *@periodp before calling */
8428 static int __maybe_unused cpu_period_quota_parse(char *buf,
8429 u64 *periodp, u64 *quotap)
8431 char tok[21]; /* U64_MAX */
8433 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
8436 *periodp *= NSEC_PER_USEC;
8438 if (sscanf(tok, "%llu", quotap))
8439 *quotap *= NSEC_PER_USEC;
8440 else if (!strcmp(tok, "max"))
8441 *quotap = RUNTIME_INF;
8448 #ifdef CONFIG_CFS_BANDWIDTH
8449 static int cpu_max_show(struct seq_file *sf, void *v)
8451 struct task_group *tg = css_tg(seq_css(sf));
8453 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
8457 static ssize_t cpu_max_write(struct kernfs_open_file *of,
8458 char *buf, size_t nbytes, loff_t off)
8460 struct task_group *tg = css_tg(of_css(of));
8461 u64 period = tg_get_cfs_period(tg);
8465 ret = cpu_period_quota_parse(buf, &period, "a);
8467 ret = tg_set_cfs_bandwidth(tg, period, quota);
8468 return ret ?: nbytes;
8472 static struct cftype cpu_files[] = {
8473 #ifdef CONFIG_FAIR_GROUP_SCHED
8476 .flags = CFTYPE_NOT_ON_ROOT,
8477 .read_u64 = cpu_weight_read_u64,
8478 .write_u64 = cpu_weight_write_u64,
8481 .name = "weight.nice",
8482 .flags = CFTYPE_NOT_ON_ROOT,
8483 .read_s64 = cpu_weight_nice_read_s64,
8484 .write_s64 = cpu_weight_nice_write_s64,
8487 #ifdef CONFIG_CFS_BANDWIDTH
8490 .flags = CFTYPE_NOT_ON_ROOT,
8491 .seq_show = cpu_max_show,
8492 .write = cpu_max_write,
8495 #ifdef CONFIG_UCLAMP_TASK_GROUP
8497 .name = "uclamp.min",
8498 .flags = CFTYPE_NOT_ON_ROOT,
8499 .seq_show = cpu_uclamp_min_show,
8500 .write = cpu_uclamp_min_write,
8503 .name = "uclamp.max",
8504 .flags = CFTYPE_NOT_ON_ROOT,
8505 .seq_show = cpu_uclamp_max_show,
8506 .write = cpu_uclamp_max_write,
8512 struct cgroup_subsys cpu_cgrp_subsys = {
8513 .css_alloc = cpu_cgroup_css_alloc,
8514 .css_online = cpu_cgroup_css_online,
8515 .css_released = cpu_cgroup_css_released,
8516 .css_free = cpu_cgroup_css_free,
8517 .css_extra_stat_show = cpu_extra_stat_show,
8518 .fork = cpu_cgroup_fork,
8519 .can_attach = cpu_cgroup_can_attach,
8520 .attach = cpu_cgroup_attach,
8521 .legacy_cftypes = cpu_legacy_files,
8522 .dfl_cftypes = cpu_files,
8527 #endif /* CONFIG_CGROUP_SCHED */
8529 void dump_cpu_task(int cpu)
8531 pr_info("Task dump for CPU %d:\n", cpu);
8532 sched_show_task(cpu_curr(cpu));
8536 * Nice levels are multiplicative, with a gentle 10% change for every
8537 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8538 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8539 * that remained on nice 0.
8541 * The "10% effect" is relative and cumulative: from _any_ nice level,
8542 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8543 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8544 * If a task goes up by ~10% and another task goes down by ~10% then
8545 * the relative distance between them is ~25%.)
8547 const int sched_prio_to_weight[40] = {
8548 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8549 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8550 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8551 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8552 /* 0 */ 1024, 820, 655, 526, 423,
8553 /* 5 */ 335, 272, 215, 172, 137,
8554 /* 10 */ 110, 87, 70, 56, 45,
8555 /* 15 */ 36, 29, 23, 18, 15,
8559 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8561 * In cases where the weight does not change often, we can use the
8562 * precalculated inverse to speed up arithmetics by turning divisions
8563 * into multiplications:
8565 const u32 sched_prio_to_wmult[40] = {
8566 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8567 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8568 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8569 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8570 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8571 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8572 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8573 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
8576 void call_trace_sched_update_nr_running(struct rq *rq, int count)
8578 trace_sched_update_nr_running_tp(rq, count);