4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <uapi/linux/sched/types.h>
11 #include <linux/sched/loadavg.h>
12 #include <linux/sched/hotplug.h>
13 #include <linux/cpuset.h>
14 #include <linux/delayacct.h>
15 #include <linux/init_task.h>
16 #include <linux/context_tracking.h>
17 #include <linux/rcupdate_wait.h>
19 #include <linux/blkdev.h>
20 #include <linux/kprobes.h>
21 #include <linux/mmu_context.h>
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/prefetch.h>
25 #include <linux/profile.h>
26 #include <linux/security.h>
27 #include <linux/syscalls.h>
29 #include <asm/switch_to.h>
31 #ifdef CONFIG_PARAVIRT
32 #include <asm/paravirt.h>
36 #include "../workqueue_internal.h"
37 #include "../smpboot.h"
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/sched.h>
42 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
45 * Debugging: various feature bits
48 #define SCHED_FEAT(name, enabled) \
49 (1UL << __SCHED_FEAT_##name) * enabled |
51 const_debug unsigned int sysctl_sched_features =
58 * Number of tasks to iterate in a single balance run.
59 * Limited because this is done with IRQs disabled.
61 const_debug unsigned int sysctl_sched_nr_migrate = 32;
64 * period over which we average the RT time consumption, measured
69 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
72 * period over which we measure -rt task CPU usage in us.
75 unsigned int sysctl_sched_rt_period = 1000000;
77 __read_mostly int scheduler_running;
80 * part of the period that we allow rt tasks to run in us.
83 int sysctl_sched_rt_runtime = 950000;
85 /* CPUs with isolated domains */
86 cpumask_var_t cpu_isolated_map;
89 * __task_rq_lock - lock the rq @p resides on.
91 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
96 lockdep_assert_held(&p->pi_lock);
100 raw_spin_lock(&rq->lock);
101 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
105 raw_spin_unlock(&rq->lock);
107 while (unlikely(task_on_rq_migrating(p)))
113 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
115 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
116 __acquires(p->pi_lock)
122 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
124 raw_spin_lock(&rq->lock);
126 * move_queued_task() task_rq_lock()
129 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
130 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
131 * [S] ->cpu = new_cpu [L] task_rq()
135 * If we observe the old cpu in task_rq_lock, the acquire of
136 * the old rq->lock will fully serialize against the stores.
138 * If we observe the new CPU in task_rq_lock, the acquire will
139 * pair with the WMB to ensure we must then also see migrating.
141 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
145 raw_spin_unlock(&rq->lock);
146 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
148 while (unlikely(task_on_rq_migrating(p)))
154 * RQ-clock updating methods:
157 static void update_rq_clock_task(struct rq *rq, s64 delta)
160 * In theory, the compile should just see 0 here, and optimize out the call
161 * to sched_rt_avg_update. But I don't trust it...
163 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
164 s64 steal = 0, irq_delta = 0;
166 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
167 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
170 * Since irq_time is only updated on {soft,}irq_exit, we might run into
171 * this case when a previous update_rq_clock() happened inside a
174 * When this happens, we stop ->clock_task and only update the
175 * prev_irq_time stamp to account for the part that fit, so that a next
176 * update will consume the rest. This ensures ->clock_task is
179 * It does however cause some slight miss-attribution of {soft,}irq
180 * time, a more accurate solution would be to update the irq_time using
181 * the current rq->clock timestamp, except that would require using
184 if (irq_delta > delta)
187 rq->prev_irq_time += irq_delta;
190 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
191 if (static_key_false((¶virt_steal_rq_enabled))) {
192 steal = paravirt_steal_clock(cpu_of(rq));
193 steal -= rq->prev_steal_time_rq;
195 if (unlikely(steal > delta))
198 rq->prev_steal_time_rq += steal;
203 rq->clock_task += delta;
205 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
206 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
207 sched_rt_avg_update(rq, irq_delta + steal);
211 void update_rq_clock(struct rq *rq)
215 lockdep_assert_held(&rq->lock);
217 if (rq->clock_update_flags & RQCF_ACT_SKIP)
220 #ifdef CONFIG_SCHED_DEBUG
221 if (sched_feat(WARN_DOUBLE_CLOCK))
222 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
223 rq->clock_update_flags |= RQCF_UPDATED;
226 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
230 update_rq_clock_task(rq, delta);
234 #ifdef CONFIG_SCHED_HRTICK
236 * Use HR-timers to deliver accurate preemption points.
239 static void hrtick_clear(struct rq *rq)
241 if (hrtimer_active(&rq->hrtick_timer))
242 hrtimer_cancel(&rq->hrtick_timer);
246 * High-resolution timer tick.
247 * Runs from hardirq context with interrupts disabled.
249 static enum hrtimer_restart hrtick(struct hrtimer *timer)
251 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
254 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
258 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
261 return HRTIMER_NORESTART;
266 static void __hrtick_restart(struct rq *rq)
268 struct hrtimer *timer = &rq->hrtick_timer;
270 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
274 * called from hardirq (IPI) context
276 static void __hrtick_start(void *arg)
282 __hrtick_restart(rq);
283 rq->hrtick_csd_pending = 0;
288 * Called to set the hrtick timer state.
290 * called with rq->lock held and irqs disabled
292 void hrtick_start(struct rq *rq, u64 delay)
294 struct hrtimer *timer = &rq->hrtick_timer;
299 * Don't schedule slices shorter than 10000ns, that just
300 * doesn't make sense and can cause timer DoS.
302 delta = max_t(s64, delay, 10000LL);
303 time = ktime_add_ns(timer->base->get_time(), delta);
305 hrtimer_set_expires(timer, time);
307 if (rq == this_rq()) {
308 __hrtick_restart(rq);
309 } else if (!rq->hrtick_csd_pending) {
310 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
311 rq->hrtick_csd_pending = 1;
317 * Called to set the hrtick timer state.
319 * called with rq->lock held and irqs disabled
321 void hrtick_start(struct rq *rq, u64 delay)
324 * Don't schedule slices shorter than 10000ns, that just
325 * doesn't make sense. Rely on vruntime for fairness.
327 delay = max_t(u64, delay, 10000LL);
328 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
329 HRTIMER_MODE_REL_PINNED);
331 #endif /* CONFIG_SMP */
333 static void init_rq_hrtick(struct rq *rq)
336 rq->hrtick_csd_pending = 0;
338 rq->hrtick_csd.flags = 0;
339 rq->hrtick_csd.func = __hrtick_start;
340 rq->hrtick_csd.info = rq;
343 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
344 rq->hrtick_timer.function = hrtick;
346 #else /* CONFIG_SCHED_HRTICK */
347 static inline void hrtick_clear(struct rq *rq)
351 static inline void init_rq_hrtick(struct rq *rq)
354 #endif /* CONFIG_SCHED_HRTICK */
357 * cmpxchg based fetch_or, macro so it works for different integer types
359 #define fetch_or(ptr, mask) \
361 typeof(ptr) _ptr = (ptr); \
362 typeof(mask) _mask = (mask); \
363 typeof(*_ptr) _old, _val = *_ptr; \
366 _old = cmpxchg(_ptr, _val, _val | _mask); \
374 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
376 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
377 * this avoids any races wrt polling state changes and thereby avoids
380 static bool set_nr_and_not_polling(struct task_struct *p)
382 struct thread_info *ti = task_thread_info(p);
383 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
387 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
389 * If this returns true, then the idle task promises to call
390 * sched_ttwu_pending() and reschedule soon.
392 static bool set_nr_if_polling(struct task_struct *p)
394 struct thread_info *ti = task_thread_info(p);
395 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
398 if (!(val & _TIF_POLLING_NRFLAG))
400 if (val & _TIF_NEED_RESCHED)
402 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
411 static bool set_nr_and_not_polling(struct task_struct *p)
413 set_tsk_need_resched(p);
418 static bool set_nr_if_polling(struct task_struct *p)
425 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
427 struct wake_q_node *node = &task->wake_q;
430 * Atomically grab the task, if ->wake_q is !nil already it means
431 * its already queued (either by us or someone else) and will get the
432 * wakeup due to that.
434 * This cmpxchg() implies a full barrier, which pairs with the write
435 * barrier implied by the wakeup in wake_up_q().
437 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
440 get_task_struct(task);
443 * The head is context local, there can be no concurrency.
446 head->lastp = &node->next;
449 void wake_up_q(struct wake_q_head *head)
451 struct wake_q_node *node = head->first;
453 while (node != WAKE_Q_TAIL) {
454 struct task_struct *task;
456 task = container_of(node, struct task_struct, wake_q);
458 /* Task can safely be re-inserted now: */
460 task->wake_q.next = NULL;
463 * wake_up_process() implies a wmb() to pair with the queueing
464 * in wake_q_add() so as not to miss wakeups.
466 wake_up_process(task);
467 put_task_struct(task);
472 * resched_curr - mark rq's current task 'to be rescheduled now'.
474 * On UP this means the setting of the need_resched flag, on SMP it
475 * might also involve a cross-CPU call to trigger the scheduler on
478 void resched_curr(struct rq *rq)
480 struct task_struct *curr = rq->curr;
483 lockdep_assert_held(&rq->lock);
485 if (test_tsk_need_resched(curr))
490 if (cpu == smp_processor_id()) {
491 set_tsk_need_resched(curr);
492 set_preempt_need_resched();
496 if (set_nr_and_not_polling(curr))
497 smp_send_reschedule(cpu);
499 trace_sched_wake_idle_without_ipi(cpu);
502 void resched_cpu(int cpu)
504 struct rq *rq = cpu_rq(cpu);
507 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
510 raw_spin_unlock_irqrestore(&rq->lock, flags);
514 #ifdef CONFIG_NO_HZ_COMMON
516 * In the semi idle case, use the nearest busy CPU for migrating timers
517 * from an idle CPU. This is good for power-savings.
519 * We don't do similar optimization for completely idle system, as
520 * selecting an idle CPU will add more delays to the timers than intended
521 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
523 int get_nohz_timer_target(void)
525 int i, cpu = smp_processor_id();
526 struct sched_domain *sd;
528 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
532 for_each_domain(cpu, sd) {
533 for_each_cpu(i, sched_domain_span(sd)) {
537 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
544 if (!is_housekeeping_cpu(cpu))
545 cpu = housekeeping_any_cpu();
552 * When add_timer_on() enqueues a timer into the timer wheel of an
553 * idle CPU then this timer might expire before the next timer event
554 * which is scheduled to wake up that CPU. In case of a completely
555 * idle system the next event might even be infinite time into the
556 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
557 * leaves the inner idle loop so the newly added timer is taken into
558 * account when the CPU goes back to idle and evaluates the timer
559 * wheel for the next timer event.
561 static void wake_up_idle_cpu(int cpu)
563 struct rq *rq = cpu_rq(cpu);
565 if (cpu == smp_processor_id())
568 if (set_nr_and_not_polling(rq->idle))
569 smp_send_reschedule(cpu);
571 trace_sched_wake_idle_without_ipi(cpu);
574 static bool wake_up_full_nohz_cpu(int cpu)
577 * We just need the target to call irq_exit() and re-evaluate
578 * the next tick. The nohz full kick at least implies that.
579 * If needed we can still optimize that later with an
582 if (cpu_is_offline(cpu))
583 return true; /* Don't try to wake offline CPUs. */
584 if (tick_nohz_full_cpu(cpu)) {
585 if (cpu != smp_processor_id() ||
586 tick_nohz_tick_stopped())
587 tick_nohz_full_kick_cpu(cpu);
595 * Wake up the specified CPU. If the CPU is going offline, it is the
596 * caller's responsibility to deal with the lost wakeup, for example,
597 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
599 void wake_up_nohz_cpu(int cpu)
601 if (!wake_up_full_nohz_cpu(cpu))
602 wake_up_idle_cpu(cpu);
605 static inline bool got_nohz_idle_kick(void)
607 int cpu = smp_processor_id();
609 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
612 if (idle_cpu(cpu) && !need_resched())
616 * We can't run Idle Load Balance on this CPU for this time so we
617 * cancel it and clear NOHZ_BALANCE_KICK
619 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
623 #else /* CONFIG_NO_HZ_COMMON */
625 static inline bool got_nohz_idle_kick(void)
630 #endif /* CONFIG_NO_HZ_COMMON */
632 #ifdef CONFIG_NO_HZ_FULL
633 bool sched_can_stop_tick(struct rq *rq)
637 /* Deadline tasks, even if single, need the tick */
638 if (rq->dl.dl_nr_running)
642 * If there are more than one RR tasks, we need the tick to effect the
643 * actual RR behaviour.
645 if (rq->rt.rr_nr_running) {
646 if (rq->rt.rr_nr_running == 1)
653 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
654 * forced preemption between FIFO tasks.
656 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
661 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
662 * if there's more than one we need the tick for involuntary
665 if (rq->nr_running > 1)
670 #endif /* CONFIG_NO_HZ_FULL */
672 void sched_avg_update(struct rq *rq)
674 s64 period = sched_avg_period();
676 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
678 * Inline assembly required to prevent the compiler
679 * optimising this loop into a divmod call.
680 * See __iter_div_u64_rem() for another example of this.
682 asm("" : "+rm" (rq->age_stamp));
683 rq->age_stamp += period;
688 #endif /* CONFIG_SMP */
690 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
691 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
693 * Iterate task_group tree rooted at *from, calling @down when first entering a
694 * node and @up when leaving it for the final time.
696 * Caller must hold rcu_lock or sufficient equivalent.
698 int walk_tg_tree_from(struct task_group *from,
699 tg_visitor down, tg_visitor up, void *data)
701 struct task_group *parent, *child;
707 ret = (*down)(parent, data);
710 list_for_each_entry_rcu(child, &parent->children, siblings) {
717 ret = (*up)(parent, data);
718 if (ret || parent == from)
722 parent = parent->parent;
729 int tg_nop(struct task_group *tg, void *data)
735 static void set_load_weight(struct task_struct *p)
737 int prio = p->static_prio - MAX_RT_PRIO;
738 struct load_weight *load = &p->se.load;
741 * SCHED_IDLE tasks get minimal weight:
743 if (idle_policy(p->policy)) {
744 load->weight = scale_load(WEIGHT_IDLEPRIO);
745 load->inv_weight = WMULT_IDLEPRIO;
749 load->weight = scale_load(sched_prio_to_weight[prio]);
750 load->inv_weight = sched_prio_to_wmult[prio];
753 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
756 if (!(flags & ENQUEUE_RESTORE))
757 sched_info_queued(rq, p);
758 p->sched_class->enqueue_task(rq, p, flags);
761 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
764 if (!(flags & DEQUEUE_SAVE))
765 sched_info_dequeued(rq, p);
766 p->sched_class->dequeue_task(rq, p, flags);
769 void activate_task(struct rq *rq, struct task_struct *p, int flags)
771 if (task_contributes_to_load(p))
772 rq->nr_uninterruptible--;
774 enqueue_task(rq, p, flags);
777 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
779 if (task_contributes_to_load(p))
780 rq->nr_uninterruptible++;
782 dequeue_task(rq, p, flags);
785 void sched_set_stop_task(int cpu, struct task_struct *stop)
787 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
788 struct task_struct *old_stop = cpu_rq(cpu)->stop;
792 * Make it appear like a SCHED_FIFO task, its something
793 * userspace knows about and won't get confused about.
795 * Also, it will make PI more or less work without too
796 * much confusion -- but then, stop work should not
797 * rely on PI working anyway.
799 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
801 stop->sched_class = &stop_sched_class;
804 cpu_rq(cpu)->stop = stop;
808 * Reset it back to a normal scheduling class so that
809 * it can die in pieces.
811 old_stop->sched_class = &rt_sched_class;
816 * __normal_prio - return the priority that is based on the static prio
818 static inline int __normal_prio(struct task_struct *p)
820 return p->static_prio;
824 * Calculate the expected normal priority: i.e. priority
825 * without taking RT-inheritance into account. Might be
826 * boosted by interactivity modifiers. Changes upon fork,
827 * setprio syscalls, and whenever the interactivity
828 * estimator recalculates.
830 static inline int normal_prio(struct task_struct *p)
834 if (task_has_dl_policy(p))
835 prio = MAX_DL_PRIO-1;
836 else if (task_has_rt_policy(p))
837 prio = MAX_RT_PRIO-1 - p->rt_priority;
839 prio = __normal_prio(p);
844 * Calculate the current priority, i.e. the priority
845 * taken into account by the scheduler. This value might
846 * be boosted by RT tasks, or might be boosted by
847 * interactivity modifiers. Will be RT if the task got
848 * RT-boosted. If not then it returns p->normal_prio.
850 static int effective_prio(struct task_struct *p)
852 p->normal_prio = normal_prio(p);
854 * If we are RT tasks or we were boosted to RT priority,
855 * keep the priority unchanged. Otherwise, update priority
856 * to the normal priority:
858 if (!rt_prio(p->prio))
859 return p->normal_prio;
864 * task_curr - is this task currently executing on a CPU?
865 * @p: the task in question.
867 * Return: 1 if the task is currently executing. 0 otherwise.
869 inline int task_curr(const struct task_struct *p)
871 return cpu_curr(task_cpu(p)) == p;
875 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
876 * use the balance_callback list if you want balancing.
878 * this means any call to check_class_changed() must be followed by a call to
879 * balance_callback().
881 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
882 const struct sched_class *prev_class,
885 if (prev_class != p->sched_class) {
886 if (prev_class->switched_from)
887 prev_class->switched_from(rq, p);
889 p->sched_class->switched_to(rq, p);
890 } else if (oldprio != p->prio || dl_task(p))
891 p->sched_class->prio_changed(rq, p, oldprio);
894 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
896 const struct sched_class *class;
898 if (p->sched_class == rq->curr->sched_class) {
899 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
901 for_each_class(class) {
902 if (class == rq->curr->sched_class)
904 if (class == p->sched_class) {
912 * A queue event has occurred, and we're going to schedule. In
913 * this case, we can save a useless back to back clock update.
915 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
916 rq_clock_skip_update(rq, true);
921 * This is how migration works:
923 * 1) we invoke migration_cpu_stop() on the target CPU using
925 * 2) stopper starts to run (implicitly forcing the migrated thread
927 * 3) it checks whether the migrated task is still in the wrong runqueue.
928 * 4) if it's in the wrong runqueue then the migration thread removes
929 * it and puts it into the right queue.
930 * 5) stopper completes and stop_one_cpu() returns and the migration
935 * move_queued_task - move a queued task to new rq.
937 * Returns (locked) new rq. Old rq's lock is released.
939 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
940 struct task_struct *p, int new_cpu)
942 lockdep_assert_held(&rq->lock);
944 p->on_rq = TASK_ON_RQ_MIGRATING;
945 dequeue_task(rq, p, 0);
946 set_task_cpu(p, new_cpu);
949 rq = cpu_rq(new_cpu);
952 BUG_ON(task_cpu(p) != new_cpu);
953 enqueue_task(rq, p, 0);
954 p->on_rq = TASK_ON_RQ_QUEUED;
955 check_preempt_curr(rq, p, 0);
960 struct migration_arg {
961 struct task_struct *task;
966 * Move (not current) task off this CPU, onto the destination CPU. We're doing
967 * this because either it can't run here any more (set_cpus_allowed()
968 * away from this CPU, or CPU going down), or because we're
969 * attempting to rebalance this task on exec (sched_exec).
971 * So we race with normal scheduler movements, but that's OK, as long
972 * as the task is no longer on this CPU.
974 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
975 struct task_struct *p, int dest_cpu)
977 if (unlikely(!cpu_active(dest_cpu)))
980 /* Affinity changed (again). */
981 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
984 rq = move_queued_task(rq, rf, p, dest_cpu);
990 * migration_cpu_stop - this will be executed by a highprio stopper thread
991 * and performs thread migration by bumping thread off CPU then
992 * 'pushing' onto another runqueue.
994 static int migration_cpu_stop(void *data)
996 struct migration_arg *arg = data;
997 struct task_struct *p = arg->task;
998 struct rq *rq = this_rq();
1002 * The original target CPU might have gone down and we might
1003 * be on another CPU but it doesn't matter.
1005 local_irq_disable();
1007 * We need to explicitly wake pending tasks before running
1008 * __migrate_task() such that we will not miss enforcing cpus_allowed
1009 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1011 sched_ttwu_pending();
1013 raw_spin_lock(&p->pi_lock);
1016 * If task_rq(p) != rq, it cannot be migrated here, because we're
1017 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1018 * we're holding p->pi_lock.
1020 if (task_rq(p) == rq) {
1021 if (task_on_rq_queued(p))
1022 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1024 p->wake_cpu = arg->dest_cpu;
1027 raw_spin_unlock(&p->pi_lock);
1034 * sched_class::set_cpus_allowed must do the below, but is not required to
1035 * actually call this function.
1037 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1039 cpumask_copy(&p->cpus_allowed, new_mask);
1040 p->nr_cpus_allowed = cpumask_weight(new_mask);
1043 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1045 struct rq *rq = task_rq(p);
1046 bool queued, running;
1048 lockdep_assert_held(&p->pi_lock);
1050 queued = task_on_rq_queued(p);
1051 running = task_current(rq, p);
1055 * Because __kthread_bind() calls this on blocked tasks without
1058 lockdep_assert_held(&rq->lock);
1059 dequeue_task(rq, p, DEQUEUE_SAVE);
1062 put_prev_task(rq, p);
1064 p->sched_class->set_cpus_allowed(p, new_mask);
1067 enqueue_task(rq, p, ENQUEUE_RESTORE);
1069 set_curr_task(rq, p);
1073 * Change a given task's CPU affinity. Migrate the thread to a
1074 * proper CPU and schedule it away if the CPU it's executing on
1075 * is removed from the allowed bitmask.
1077 * NOTE: the caller must have a valid reference to the task, the
1078 * task must not exit() & deallocate itself prematurely. The
1079 * call is not atomic; no spinlocks may be held.
1081 static int __set_cpus_allowed_ptr(struct task_struct *p,
1082 const struct cpumask *new_mask, bool check)
1084 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1085 unsigned int dest_cpu;
1090 rq = task_rq_lock(p, &rf);
1091 update_rq_clock(rq);
1093 if (p->flags & PF_KTHREAD) {
1095 * Kernel threads are allowed on online && !active CPUs
1097 cpu_valid_mask = cpu_online_mask;
1101 * Must re-check here, to close a race against __kthread_bind(),
1102 * sched_setaffinity() is not guaranteed to observe the flag.
1104 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1109 if (cpumask_equal(&p->cpus_allowed, new_mask))
1112 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1117 do_set_cpus_allowed(p, new_mask);
1119 if (p->flags & PF_KTHREAD) {
1121 * For kernel threads that do indeed end up on online &&
1122 * !active we want to ensure they are strict per-CPU threads.
1124 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1125 !cpumask_intersects(new_mask, cpu_active_mask) &&
1126 p->nr_cpus_allowed != 1);
1129 /* Can the task run on the task's current CPU? If so, we're done */
1130 if (cpumask_test_cpu(task_cpu(p), new_mask))
1133 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1134 if (task_running(rq, p) || p->state == TASK_WAKING) {
1135 struct migration_arg arg = { p, dest_cpu };
1136 /* Need help from migration thread: drop lock and wait. */
1137 task_rq_unlock(rq, p, &rf);
1138 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1139 tlb_migrate_finish(p->mm);
1141 } else if (task_on_rq_queued(p)) {
1143 * OK, since we're going to drop the lock immediately
1144 * afterwards anyway.
1146 rq = move_queued_task(rq, &rf, p, dest_cpu);
1149 task_rq_unlock(rq, p, &rf);
1154 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1156 return __set_cpus_allowed_ptr(p, new_mask, false);
1158 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1160 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1162 #ifdef CONFIG_SCHED_DEBUG
1164 * We should never call set_task_cpu() on a blocked task,
1165 * ttwu() will sort out the placement.
1167 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1171 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1172 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1173 * time relying on p->on_rq.
1175 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1176 p->sched_class == &fair_sched_class &&
1177 (p->on_rq && !task_on_rq_migrating(p)));
1179 #ifdef CONFIG_LOCKDEP
1181 * The caller should hold either p->pi_lock or rq->lock, when changing
1182 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1184 * sched_move_task() holds both and thus holding either pins the cgroup,
1187 * Furthermore, all task_rq users should acquire both locks, see
1190 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1191 lockdep_is_held(&task_rq(p)->lock)));
1195 trace_sched_migrate_task(p, new_cpu);
1197 if (task_cpu(p) != new_cpu) {
1198 if (p->sched_class->migrate_task_rq)
1199 p->sched_class->migrate_task_rq(p);
1200 p->se.nr_migrations++;
1201 perf_event_task_migrate(p);
1204 __set_task_cpu(p, new_cpu);
1207 static void __migrate_swap_task(struct task_struct *p, int cpu)
1209 if (task_on_rq_queued(p)) {
1210 struct rq *src_rq, *dst_rq;
1211 struct rq_flags srf, drf;
1213 src_rq = task_rq(p);
1214 dst_rq = cpu_rq(cpu);
1216 rq_pin_lock(src_rq, &srf);
1217 rq_pin_lock(dst_rq, &drf);
1219 p->on_rq = TASK_ON_RQ_MIGRATING;
1220 deactivate_task(src_rq, p, 0);
1221 set_task_cpu(p, cpu);
1222 activate_task(dst_rq, p, 0);
1223 p->on_rq = TASK_ON_RQ_QUEUED;
1224 check_preempt_curr(dst_rq, p, 0);
1226 rq_unpin_lock(dst_rq, &drf);
1227 rq_unpin_lock(src_rq, &srf);
1231 * Task isn't running anymore; make it appear like we migrated
1232 * it before it went to sleep. This means on wakeup we make the
1233 * previous CPU our target instead of where it really is.
1239 struct migration_swap_arg {
1240 struct task_struct *src_task, *dst_task;
1241 int src_cpu, dst_cpu;
1244 static int migrate_swap_stop(void *data)
1246 struct migration_swap_arg *arg = data;
1247 struct rq *src_rq, *dst_rq;
1250 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1253 src_rq = cpu_rq(arg->src_cpu);
1254 dst_rq = cpu_rq(arg->dst_cpu);
1256 double_raw_lock(&arg->src_task->pi_lock,
1257 &arg->dst_task->pi_lock);
1258 double_rq_lock(src_rq, dst_rq);
1260 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1263 if (task_cpu(arg->src_task) != arg->src_cpu)
1266 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1269 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1272 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1273 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1278 double_rq_unlock(src_rq, dst_rq);
1279 raw_spin_unlock(&arg->dst_task->pi_lock);
1280 raw_spin_unlock(&arg->src_task->pi_lock);
1286 * Cross migrate two tasks
1288 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1290 struct migration_swap_arg arg;
1293 arg = (struct migration_swap_arg){
1295 .src_cpu = task_cpu(cur),
1297 .dst_cpu = task_cpu(p),
1300 if (arg.src_cpu == arg.dst_cpu)
1304 * These three tests are all lockless; this is OK since all of them
1305 * will be re-checked with proper locks held further down the line.
1307 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1310 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1313 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1316 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1317 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1324 * wait_task_inactive - wait for a thread to unschedule.
1326 * If @match_state is nonzero, it's the @p->state value just checked and
1327 * not expected to change. If it changes, i.e. @p might have woken up,
1328 * then return zero. When we succeed in waiting for @p to be off its CPU,
1329 * we return a positive number (its total switch count). If a second call
1330 * a short while later returns the same number, the caller can be sure that
1331 * @p has remained unscheduled the whole time.
1333 * The caller must ensure that the task *will* unschedule sometime soon,
1334 * else this function might spin for a *long* time. This function can't
1335 * be called with interrupts off, or it may introduce deadlock with
1336 * smp_call_function() if an IPI is sent by the same process we are
1337 * waiting to become inactive.
1339 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1341 int running, queued;
1348 * We do the initial early heuristics without holding
1349 * any task-queue locks at all. We'll only try to get
1350 * the runqueue lock when things look like they will
1356 * If the task is actively running on another CPU
1357 * still, just relax and busy-wait without holding
1360 * NOTE! Since we don't hold any locks, it's not
1361 * even sure that "rq" stays as the right runqueue!
1362 * But we don't care, since "task_running()" will
1363 * return false if the runqueue has changed and p
1364 * is actually now running somewhere else!
1366 while (task_running(rq, p)) {
1367 if (match_state && unlikely(p->state != match_state))
1373 * Ok, time to look more closely! We need the rq
1374 * lock now, to be *sure*. If we're wrong, we'll
1375 * just go back and repeat.
1377 rq = task_rq_lock(p, &rf);
1378 trace_sched_wait_task(p);
1379 running = task_running(rq, p);
1380 queued = task_on_rq_queued(p);
1382 if (!match_state || p->state == match_state)
1383 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1384 task_rq_unlock(rq, p, &rf);
1387 * If it changed from the expected state, bail out now.
1389 if (unlikely(!ncsw))
1393 * Was it really running after all now that we
1394 * checked with the proper locks actually held?
1396 * Oops. Go back and try again..
1398 if (unlikely(running)) {
1404 * It's not enough that it's not actively running,
1405 * it must be off the runqueue _entirely_, and not
1408 * So if it was still runnable (but just not actively
1409 * running right now), it's preempted, and we should
1410 * yield - it could be a while.
1412 if (unlikely(queued)) {
1413 ktime_t to = NSEC_PER_SEC / HZ;
1415 set_current_state(TASK_UNINTERRUPTIBLE);
1416 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1421 * Ahh, all good. It wasn't running, and it wasn't
1422 * runnable, which means that it will never become
1423 * running in the future either. We're all done!
1432 * kick_process - kick a running thread to enter/exit the kernel
1433 * @p: the to-be-kicked thread
1435 * Cause a process which is running on another CPU to enter
1436 * kernel-mode, without any delay. (to get signals handled.)
1438 * NOTE: this function doesn't have to take the runqueue lock,
1439 * because all it wants to ensure is that the remote task enters
1440 * the kernel. If the IPI races and the task has been migrated
1441 * to another CPU then no harm is done and the purpose has been
1444 void kick_process(struct task_struct *p)
1450 if ((cpu != smp_processor_id()) && task_curr(p))
1451 smp_send_reschedule(cpu);
1454 EXPORT_SYMBOL_GPL(kick_process);
1457 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1459 * A few notes on cpu_active vs cpu_online:
1461 * - cpu_active must be a subset of cpu_online
1463 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1464 * see __set_cpus_allowed_ptr(). At this point the newly online
1465 * CPU isn't yet part of the sched domains, and balancing will not
1468 * - on CPU-down we clear cpu_active() to mask the sched domains and
1469 * avoid the load balancer to place new tasks on the to be removed
1470 * CPU. Existing tasks will remain running there and will be taken
1473 * This means that fallback selection must not select !active CPUs.
1474 * And can assume that any active CPU must be online. Conversely
1475 * select_task_rq() below may allow selection of !active CPUs in order
1476 * to satisfy the above rules.
1478 static int select_fallback_rq(int cpu, struct task_struct *p)
1480 int nid = cpu_to_node(cpu);
1481 const struct cpumask *nodemask = NULL;
1482 enum { cpuset, possible, fail } state = cpuset;
1486 * If the node that the CPU is on has been offlined, cpu_to_node()
1487 * will return -1. There is no CPU on the node, and we should
1488 * select the CPU on the other node.
1491 nodemask = cpumask_of_node(nid);
1493 /* Look for allowed, online CPU in same node. */
1494 for_each_cpu(dest_cpu, nodemask) {
1495 if (!cpu_active(dest_cpu))
1497 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1503 /* Any allowed, online CPU? */
1504 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1505 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1507 if (!cpu_online(dest_cpu))
1512 /* No more Mr. Nice Guy. */
1515 if (IS_ENABLED(CONFIG_CPUSETS)) {
1516 cpuset_cpus_allowed_fallback(p);
1522 do_set_cpus_allowed(p, cpu_possible_mask);
1533 if (state != cpuset) {
1535 * Don't tell them about moving exiting tasks or
1536 * kernel threads (both mm NULL), since they never
1539 if (p->mm && printk_ratelimit()) {
1540 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1541 task_pid_nr(p), p->comm, cpu);
1549 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1552 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1554 lockdep_assert_held(&p->pi_lock);
1556 if (p->nr_cpus_allowed > 1)
1557 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1559 cpu = cpumask_any(&p->cpus_allowed);
1562 * In order not to call set_task_cpu() on a blocking task we need
1563 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1566 * Since this is common to all placement strategies, this lives here.
1568 * [ this allows ->select_task() to simply return task_cpu(p) and
1569 * not worry about this generic constraint ]
1571 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
1573 cpu = select_fallback_rq(task_cpu(p), p);
1578 static void update_avg(u64 *avg, u64 sample)
1580 s64 diff = sample - *avg;
1586 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1587 const struct cpumask *new_mask, bool check)
1589 return set_cpus_allowed_ptr(p, new_mask);
1592 #endif /* CONFIG_SMP */
1595 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1599 if (!schedstat_enabled())
1605 if (cpu == rq->cpu) {
1606 schedstat_inc(rq->ttwu_local);
1607 schedstat_inc(p->se.statistics.nr_wakeups_local);
1609 struct sched_domain *sd;
1611 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1613 for_each_domain(rq->cpu, sd) {
1614 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1615 schedstat_inc(sd->ttwu_wake_remote);
1622 if (wake_flags & WF_MIGRATED)
1623 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1624 #endif /* CONFIG_SMP */
1626 schedstat_inc(rq->ttwu_count);
1627 schedstat_inc(p->se.statistics.nr_wakeups);
1629 if (wake_flags & WF_SYNC)
1630 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1633 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1635 activate_task(rq, p, en_flags);
1636 p->on_rq = TASK_ON_RQ_QUEUED;
1638 /* If a worker is waking up, notify the workqueue: */
1639 if (p->flags & PF_WQ_WORKER)
1640 wq_worker_waking_up(p, cpu_of(rq));
1644 * Mark the task runnable and perform wakeup-preemption.
1646 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1647 struct rq_flags *rf)
1649 check_preempt_curr(rq, p, wake_flags);
1650 p->state = TASK_RUNNING;
1651 trace_sched_wakeup(p);
1654 if (p->sched_class->task_woken) {
1656 * Our task @p is fully woken up and running; so its safe to
1657 * drop the rq->lock, hereafter rq is only used for statistics.
1659 rq_unpin_lock(rq, rf);
1660 p->sched_class->task_woken(rq, p);
1661 rq_repin_lock(rq, rf);
1664 if (rq->idle_stamp) {
1665 u64 delta = rq_clock(rq) - rq->idle_stamp;
1666 u64 max = 2*rq->max_idle_balance_cost;
1668 update_avg(&rq->avg_idle, delta);
1670 if (rq->avg_idle > max)
1679 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1680 struct rq_flags *rf)
1682 int en_flags = ENQUEUE_WAKEUP;
1684 lockdep_assert_held(&rq->lock);
1687 if (p->sched_contributes_to_load)
1688 rq->nr_uninterruptible--;
1690 if (wake_flags & WF_MIGRATED)
1691 en_flags |= ENQUEUE_MIGRATED;
1694 ttwu_activate(rq, p, en_flags);
1695 ttwu_do_wakeup(rq, p, wake_flags, rf);
1699 * Called in case the task @p isn't fully descheduled from its runqueue,
1700 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1701 * since all we need to do is flip p->state to TASK_RUNNING, since
1702 * the task is still ->on_rq.
1704 static int ttwu_remote(struct task_struct *p, int wake_flags)
1710 rq = __task_rq_lock(p, &rf);
1711 if (task_on_rq_queued(p)) {
1712 /* check_preempt_curr() may use rq clock */
1713 update_rq_clock(rq);
1714 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1717 __task_rq_unlock(rq, &rf);
1723 void sched_ttwu_pending(void)
1725 struct rq *rq = this_rq();
1726 struct llist_node *llist = llist_del_all(&rq->wake_list);
1727 struct task_struct *p;
1733 rq_lock_irqsave(rq, &rf);
1738 p = llist_entry(llist, struct task_struct, wake_entry);
1739 llist = llist_next(llist);
1741 if (p->sched_remote_wakeup)
1742 wake_flags = WF_MIGRATED;
1744 ttwu_do_activate(rq, p, wake_flags, &rf);
1747 rq_unlock_irqrestore(rq, &rf);
1750 void scheduler_ipi(void)
1753 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1754 * TIF_NEED_RESCHED remotely (for the first time) will also send
1757 preempt_fold_need_resched();
1759 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1763 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1764 * traditionally all their work was done from the interrupt return
1765 * path. Now that we actually do some work, we need to make sure
1768 * Some archs already do call them, luckily irq_enter/exit nest
1771 * Arguably we should visit all archs and update all handlers,
1772 * however a fair share of IPIs are still resched only so this would
1773 * somewhat pessimize the simple resched case.
1776 sched_ttwu_pending();
1779 * Check if someone kicked us for doing the nohz idle load balance.
1781 if (unlikely(got_nohz_idle_kick())) {
1782 this_rq()->idle_balance = 1;
1783 raise_softirq_irqoff(SCHED_SOFTIRQ);
1788 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1790 struct rq *rq = cpu_rq(cpu);
1792 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1794 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1795 if (!set_nr_if_polling(rq->idle))
1796 smp_send_reschedule(cpu);
1798 trace_sched_wake_idle_without_ipi(cpu);
1802 void wake_up_if_idle(int cpu)
1804 struct rq *rq = cpu_rq(cpu);
1809 if (!is_idle_task(rcu_dereference(rq->curr)))
1812 if (set_nr_if_polling(rq->idle)) {
1813 trace_sched_wake_idle_without_ipi(cpu);
1815 rq_lock_irqsave(rq, &rf);
1816 if (is_idle_task(rq->curr))
1817 smp_send_reschedule(cpu);
1818 /* Else CPU is not idle, do nothing here: */
1819 rq_unlock_irqrestore(rq, &rf);
1826 bool cpus_share_cache(int this_cpu, int that_cpu)
1828 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1830 #endif /* CONFIG_SMP */
1832 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1834 struct rq *rq = cpu_rq(cpu);
1837 #if defined(CONFIG_SMP)
1838 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1839 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1840 ttwu_queue_remote(p, cpu, wake_flags);
1846 ttwu_do_activate(rq, p, wake_flags, &rf);
1851 * Notes on Program-Order guarantees on SMP systems.
1855 * The basic program-order guarantee on SMP systems is that when a task [t]
1856 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1857 * execution on its new CPU [c1].
1859 * For migration (of runnable tasks) this is provided by the following means:
1861 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1862 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1863 * rq(c1)->lock (if not at the same time, then in that order).
1864 * C) LOCK of the rq(c1)->lock scheduling in task
1866 * Transitivity guarantees that B happens after A and C after B.
1867 * Note: we only require RCpc transitivity.
1868 * Note: the CPU doing B need not be c0 or c1
1877 * UNLOCK rq(0)->lock
1879 * LOCK rq(0)->lock // orders against CPU0
1881 * UNLOCK rq(0)->lock
1885 * UNLOCK rq(1)->lock
1887 * LOCK rq(1)->lock // orders against CPU2
1890 * UNLOCK rq(1)->lock
1893 * BLOCKING -- aka. SLEEP + WAKEUP
1895 * For blocking we (obviously) need to provide the same guarantee as for
1896 * migration. However the means are completely different as there is no lock
1897 * chain to provide order. Instead we do:
1899 * 1) smp_store_release(X->on_cpu, 0)
1900 * 2) smp_cond_load_acquire(!X->on_cpu)
1904 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1906 * LOCK rq(0)->lock LOCK X->pi_lock
1909 * smp_store_release(X->on_cpu, 0);
1911 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1917 * X->state = RUNNING
1918 * UNLOCK rq(2)->lock
1920 * LOCK rq(2)->lock // orders against CPU1
1923 * UNLOCK rq(2)->lock
1926 * UNLOCK rq(0)->lock
1929 * However; for wakeups there is a second guarantee we must provide, namely we
1930 * must observe the state that lead to our wakeup. That is, not only must our
1931 * task observe its own prior state, it must also observe the stores prior to
1934 * This means that any means of doing remote wakeups must order the CPU doing
1935 * the wakeup against the CPU the task is going to end up running on. This,
1936 * however, is already required for the regular Program-Order guarantee above,
1937 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1942 * try_to_wake_up - wake up a thread
1943 * @p: the thread to be awakened
1944 * @state: the mask of task states that can be woken
1945 * @wake_flags: wake modifier flags (WF_*)
1947 * If (@state & @p->state) @p->state = TASK_RUNNING.
1949 * If the task was not queued/runnable, also place it back on a runqueue.
1951 * Atomic against schedule() which would dequeue a task, also see
1952 * set_current_state().
1954 * Return: %true if @p->state changes (an actual wakeup was done),
1958 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1960 unsigned long flags;
1961 int cpu, success = 0;
1964 * If we are going to wake up a thread waiting for CONDITION we
1965 * need to ensure that CONDITION=1 done by the caller can not be
1966 * reordered with p->state check below. This pairs with mb() in
1967 * set_current_state() the waiting thread does.
1969 smp_mb__before_spinlock();
1970 raw_spin_lock_irqsave(&p->pi_lock, flags);
1971 if (!(p->state & state))
1974 trace_sched_waking(p);
1976 /* We're going to change ->state: */
1981 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1982 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1983 * in smp_cond_load_acquire() below.
1985 * sched_ttwu_pending() try_to_wake_up()
1986 * [S] p->on_rq = 1; [L] P->state
1987 * UNLOCK rq->lock -----.
1991 * LOCK rq->lock -----'
1995 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1997 * Pairs with the UNLOCK+LOCK on rq->lock from the
1998 * last wakeup of our task and the schedule that got our task
2002 if (p->on_rq && ttwu_remote(p, wake_flags))
2007 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2008 * possible to, falsely, observe p->on_cpu == 0.
2010 * One must be running (->on_cpu == 1) in order to remove oneself
2011 * from the runqueue.
2013 * [S] ->on_cpu = 1; [L] ->on_rq
2017 * [S] ->on_rq = 0; [L] ->on_cpu
2019 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2020 * from the consecutive calls to schedule(); the first switching to our
2021 * task, the second putting it to sleep.
2026 * If the owning (remote) CPU is still in the middle of schedule() with
2027 * this task as prev, wait until its done referencing the task.
2029 * Pairs with the smp_store_release() in finish_lock_switch().
2031 * This ensures that tasks getting woken will be fully ordered against
2032 * their previous state and preserve Program Order.
2034 smp_cond_load_acquire(&p->on_cpu, !VAL);
2036 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2037 p->state = TASK_WAKING;
2040 delayacct_blkio_end();
2041 atomic_dec(&task_rq(p)->nr_iowait);
2044 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2045 if (task_cpu(p) != cpu) {
2046 wake_flags |= WF_MIGRATED;
2047 set_task_cpu(p, cpu);
2050 #else /* CONFIG_SMP */
2053 delayacct_blkio_end();
2054 atomic_dec(&task_rq(p)->nr_iowait);
2057 #endif /* CONFIG_SMP */
2059 ttwu_queue(p, cpu, wake_flags);
2061 ttwu_stat(p, cpu, wake_flags);
2063 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2069 * try_to_wake_up_local - try to wake up a local task with rq lock held
2070 * @p: the thread to be awakened
2071 * @cookie: context's cookie for pinning
2073 * Put @p on the run-queue if it's not already there. The caller must
2074 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2077 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2079 struct rq *rq = task_rq(p);
2081 if (WARN_ON_ONCE(rq != this_rq()) ||
2082 WARN_ON_ONCE(p == current))
2085 lockdep_assert_held(&rq->lock);
2087 if (!raw_spin_trylock(&p->pi_lock)) {
2089 * This is OK, because current is on_cpu, which avoids it being
2090 * picked for load-balance and preemption/IRQs are still
2091 * disabled avoiding further scheduler activity on it and we've
2092 * not yet picked a replacement task.
2095 raw_spin_lock(&p->pi_lock);
2099 if (!(p->state & TASK_NORMAL))
2102 trace_sched_waking(p);
2104 if (!task_on_rq_queued(p)) {
2106 delayacct_blkio_end();
2107 atomic_dec(&rq->nr_iowait);
2109 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2112 ttwu_do_wakeup(rq, p, 0, rf);
2113 ttwu_stat(p, smp_processor_id(), 0);
2115 raw_spin_unlock(&p->pi_lock);
2119 * wake_up_process - Wake up a specific process
2120 * @p: The process to be woken up.
2122 * Attempt to wake up the nominated process and move it to the set of runnable
2125 * Return: 1 if the process was woken up, 0 if it was already running.
2127 * It may be assumed that this function implies a write memory barrier before
2128 * changing the task state if and only if any tasks are woken up.
2130 int wake_up_process(struct task_struct *p)
2132 return try_to_wake_up(p, TASK_NORMAL, 0);
2134 EXPORT_SYMBOL(wake_up_process);
2136 int wake_up_state(struct task_struct *p, unsigned int state)
2138 return try_to_wake_up(p, state, 0);
2142 * This function clears the sched_dl_entity static params.
2144 void __dl_clear_params(struct task_struct *p)
2146 struct sched_dl_entity *dl_se = &p->dl;
2148 dl_se->dl_runtime = 0;
2149 dl_se->dl_deadline = 0;
2150 dl_se->dl_period = 0;
2154 dl_se->dl_throttled = 0;
2155 dl_se->dl_yielded = 0;
2159 * Perform scheduler related setup for a newly forked process p.
2160 * p is forked by current.
2162 * __sched_fork() is basic setup used by init_idle() too:
2164 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2169 p->se.exec_start = 0;
2170 p->se.sum_exec_runtime = 0;
2171 p->se.prev_sum_exec_runtime = 0;
2172 p->se.nr_migrations = 0;
2174 INIT_LIST_HEAD(&p->se.group_node);
2176 #ifdef CONFIG_FAIR_GROUP_SCHED
2177 p->se.cfs_rq = NULL;
2180 #ifdef CONFIG_SCHEDSTATS
2181 /* Even if schedstat is disabled, there should not be garbage */
2182 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2185 RB_CLEAR_NODE(&p->dl.rb_node);
2186 init_dl_task_timer(&p->dl);
2187 __dl_clear_params(p);
2189 INIT_LIST_HEAD(&p->rt.run_list);
2191 p->rt.time_slice = sched_rr_timeslice;
2195 #ifdef CONFIG_PREEMPT_NOTIFIERS
2196 INIT_HLIST_HEAD(&p->preempt_notifiers);
2199 #ifdef CONFIG_NUMA_BALANCING
2200 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2201 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2202 p->mm->numa_scan_seq = 0;
2205 if (clone_flags & CLONE_VM)
2206 p->numa_preferred_nid = current->numa_preferred_nid;
2208 p->numa_preferred_nid = -1;
2210 p->node_stamp = 0ULL;
2211 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2212 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2213 p->numa_work.next = &p->numa_work;
2214 p->numa_faults = NULL;
2215 p->last_task_numa_placement = 0;
2216 p->last_sum_exec_runtime = 0;
2218 p->numa_group = NULL;
2219 #endif /* CONFIG_NUMA_BALANCING */
2222 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2224 #ifdef CONFIG_NUMA_BALANCING
2226 void set_numabalancing_state(bool enabled)
2229 static_branch_enable(&sched_numa_balancing);
2231 static_branch_disable(&sched_numa_balancing);
2234 #ifdef CONFIG_PROC_SYSCTL
2235 int sysctl_numa_balancing(struct ctl_table *table, int write,
2236 void __user *buffer, size_t *lenp, loff_t *ppos)
2240 int state = static_branch_likely(&sched_numa_balancing);
2242 if (write && !capable(CAP_SYS_ADMIN))
2247 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2251 set_numabalancing_state(state);
2257 #ifdef CONFIG_SCHEDSTATS
2259 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2260 static bool __initdata __sched_schedstats = false;
2262 static void set_schedstats(bool enabled)
2265 static_branch_enable(&sched_schedstats);
2267 static_branch_disable(&sched_schedstats);
2270 void force_schedstat_enabled(void)
2272 if (!schedstat_enabled()) {
2273 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2274 static_branch_enable(&sched_schedstats);
2278 static int __init setup_schedstats(char *str)
2285 * This code is called before jump labels have been set up, so we can't
2286 * change the static branch directly just yet. Instead set a temporary
2287 * variable so init_schedstats() can do it later.
2289 if (!strcmp(str, "enable")) {
2290 __sched_schedstats = true;
2292 } else if (!strcmp(str, "disable")) {
2293 __sched_schedstats = false;
2298 pr_warn("Unable to parse schedstats=\n");
2302 __setup("schedstats=", setup_schedstats);
2304 static void __init init_schedstats(void)
2306 set_schedstats(__sched_schedstats);
2309 #ifdef CONFIG_PROC_SYSCTL
2310 int sysctl_schedstats(struct ctl_table *table, int write,
2311 void __user *buffer, size_t *lenp, loff_t *ppos)
2315 int state = static_branch_likely(&sched_schedstats);
2317 if (write && !capable(CAP_SYS_ADMIN))
2322 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2326 set_schedstats(state);
2329 #endif /* CONFIG_PROC_SYSCTL */
2330 #else /* !CONFIG_SCHEDSTATS */
2331 static inline void init_schedstats(void) {}
2332 #endif /* CONFIG_SCHEDSTATS */
2335 * fork()/clone()-time setup:
2337 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2339 unsigned long flags;
2340 int cpu = get_cpu();
2342 __sched_fork(clone_flags, p);
2344 * We mark the process as NEW here. This guarantees that
2345 * nobody will actually run it, and a signal or other external
2346 * event cannot wake it up and insert it on the runqueue either.
2348 p->state = TASK_NEW;
2351 * Make sure we do not leak PI boosting priority to the child.
2353 p->prio = current->normal_prio;
2356 * Revert to default priority/policy on fork if requested.
2358 if (unlikely(p->sched_reset_on_fork)) {
2359 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2360 p->policy = SCHED_NORMAL;
2361 p->static_prio = NICE_TO_PRIO(0);
2363 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2364 p->static_prio = NICE_TO_PRIO(0);
2366 p->prio = p->normal_prio = __normal_prio(p);
2370 * We don't need the reset flag anymore after the fork. It has
2371 * fulfilled its duty:
2373 p->sched_reset_on_fork = 0;
2376 if (dl_prio(p->prio)) {
2379 } else if (rt_prio(p->prio)) {
2380 p->sched_class = &rt_sched_class;
2382 p->sched_class = &fair_sched_class;
2385 init_entity_runnable_average(&p->se);
2388 * The child is not yet in the pid-hash so no cgroup attach races,
2389 * and the cgroup is pinned to this child due to cgroup_fork()
2390 * is ran before sched_fork().
2392 * Silence PROVE_RCU.
2394 raw_spin_lock_irqsave(&p->pi_lock, flags);
2396 * We're setting the CPU for the first time, we don't migrate,
2397 * so use __set_task_cpu().
2399 __set_task_cpu(p, cpu);
2400 if (p->sched_class->task_fork)
2401 p->sched_class->task_fork(p);
2402 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2404 #ifdef CONFIG_SCHED_INFO
2405 if (likely(sched_info_on()))
2406 memset(&p->sched_info, 0, sizeof(p->sched_info));
2408 #if defined(CONFIG_SMP)
2411 init_task_preempt_count(p);
2413 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2414 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2421 unsigned long to_ratio(u64 period, u64 runtime)
2423 if (runtime == RUNTIME_INF)
2427 * Doing this here saves a lot of checks in all
2428 * the calling paths, and returning zero seems
2429 * safe for them anyway.
2434 return div64_u64(runtime << 20, period);
2438 inline struct dl_bw *dl_bw_of(int i)
2440 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2441 "sched RCU must be held");
2442 return &cpu_rq(i)->rd->dl_bw;
2445 static inline int dl_bw_cpus(int i)
2447 struct root_domain *rd = cpu_rq(i)->rd;
2450 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2451 "sched RCU must be held");
2452 for_each_cpu_and(i, rd->span, cpu_active_mask)
2458 inline struct dl_bw *dl_bw_of(int i)
2460 return &cpu_rq(i)->dl.dl_bw;
2463 static inline int dl_bw_cpus(int i)
2470 * We must be sure that accepting a new task (or allowing changing the
2471 * parameters of an existing one) is consistent with the bandwidth
2472 * constraints. If yes, this function also accordingly updates the currently
2473 * allocated bandwidth to reflect the new situation.
2475 * This function is called while holding p's rq->lock.
2477 * XXX we should delay bw change until the task's 0-lag point, see
2480 static int dl_overflow(struct task_struct *p, int policy,
2481 const struct sched_attr *attr)
2484 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2485 u64 period = attr->sched_period ?: attr->sched_deadline;
2486 u64 runtime = attr->sched_runtime;
2487 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2490 /* !deadline task may carry old deadline bandwidth */
2491 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2495 * Either if a task, enters, leave, or stays -deadline but changes
2496 * its parameters, we may need to update accordingly the total
2497 * allocated bandwidth of the container.
2499 raw_spin_lock(&dl_b->lock);
2500 cpus = dl_bw_cpus(task_cpu(p));
2501 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2502 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2503 __dl_add(dl_b, new_bw);
2505 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2506 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2507 __dl_clear(dl_b, p->dl.dl_bw);
2508 __dl_add(dl_b, new_bw);
2510 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2511 __dl_clear(dl_b, p->dl.dl_bw);
2514 raw_spin_unlock(&dl_b->lock);
2519 extern void init_dl_bw(struct dl_bw *dl_b);
2522 * wake_up_new_task - wake up a newly created task for the first time.
2524 * This function will do some initial scheduler statistics housekeeping
2525 * that must be done for every newly created context, then puts the task
2526 * on the runqueue and wakes it.
2528 void wake_up_new_task(struct task_struct *p)
2533 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2534 p->state = TASK_RUNNING;
2537 * Fork balancing, do it here and not earlier because:
2538 * - cpus_allowed can change in the fork path
2539 * - any previously selected CPU might disappear through hotplug
2541 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2542 * as we're not fully set-up yet.
2544 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2546 rq = __task_rq_lock(p, &rf);
2547 update_rq_clock(rq);
2548 post_init_entity_util_avg(&p->se);
2550 activate_task(rq, p, 0);
2551 p->on_rq = TASK_ON_RQ_QUEUED;
2552 trace_sched_wakeup_new(p);
2553 check_preempt_curr(rq, p, WF_FORK);
2555 if (p->sched_class->task_woken) {
2557 * Nothing relies on rq->lock after this, so its fine to
2560 rq_unpin_lock(rq, &rf);
2561 p->sched_class->task_woken(rq, p);
2562 rq_repin_lock(rq, &rf);
2565 task_rq_unlock(rq, p, &rf);
2568 #ifdef CONFIG_PREEMPT_NOTIFIERS
2570 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2572 void preempt_notifier_inc(void)
2574 static_key_slow_inc(&preempt_notifier_key);
2576 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2578 void preempt_notifier_dec(void)
2580 static_key_slow_dec(&preempt_notifier_key);
2582 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2585 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2586 * @notifier: notifier struct to register
2588 void preempt_notifier_register(struct preempt_notifier *notifier)
2590 if (!static_key_false(&preempt_notifier_key))
2591 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2593 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2595 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2598 * preempt_notifier_unregister - no longer interested in preemption notifications
2599 * @notifier: notifier struct to unregister
2601 * This is *not* safe to call from within a preemption notifier.
2603 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2605 hlist_del(¬ifier->link);
2607 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2609 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2611 struct preempt_notifier *notifier;
2613 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2614 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2617 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2619 if (static_key_false(&preempt_notifier_key))
2620 __fire_sched_in_preempt_notifiers(curr);
2624 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2625 struct task_struct *next)
2627 struct preempt_notifier *notifier;
2629 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2630 notifier->ops->sched_out(notifier, next);
2633 static __always_inline void
2634 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2635 struct task_struct *next)
2637 if (static_key_false(&preempt_notifier_key))
2638 __fire_sched_out_preempt_notifiers(curr, next);
2641 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2643 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2648 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2649 struct task_struct *next)
2653 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2656 * prepare_task_switch - prepare to switch tasks
2657 * @rq: the runqueue preparing to switch
2658 * @prev: the current task that is being switched out
2659 * @next: the task we are going to switch to.
2661 * This is called with the rq lock held and interrupts off. It must
2662 * be paired with a subsequent finish_task_switch after the context
2665 * prepare_task_switch sets up locking and calls architecture specific
2669 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2670 struct task_struct *next)
2672 sched_info_switch(rq, prev, next);
2673 perf_event_task_sched_out(prev, next);
2674 fire_sched_out_preempt_notifiers(prev, next);
2675 prepare_lock_switch(rq, next);
2676 prepare_arch_switch(next);
2680 * finish_task_switch - clean up after a task-switch
2681 * @prev: the thread we just switched away from.
2683 * finish_task_switch must be called after the context switch, paired
2684 * with a prepare_task_switch call before the context switch.
2685 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2686 * and do any other architecture-specific cleanup actions.
2688 * Note that we may have delayed dropping an mm in context_switch(). If
2689 * so, we finish that here outside of the runqueue lock. (Doing it
2690 * with the lock held can cause deadlocks; see schedule() for
2693 * The context switch have flipped the stack from under us and restored the
2694 * local variables which were saved when this task called schedule() in the
2695 * past. prev == current is still correct but we need to recalculate this_rq
2696 * because prev may have moved to another CPU.
2698 static struct rq *finish_task_switch(struct task_struct *prev)
2699 __releases(rq->lock)
2701 struct rq *rq = this_rq();
2702 struct mm_struct *mm = rq->prev_mm;
2706 * The previous task will have left us with a preempt_count of 2
2707 * because it left us after:
2710 * preempt_disable(); // 1
2712 * raw_spin_lock_irq(&rq->lock) // 2
2714 * Also, see FORK_PREEMPT_COUNT.
2716 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2717 "corrupted preempt_count: %s/%d/0x%x\n",
2718 current->comm, current->pid, preempt_count()))
2719 preempt_count_set(FORK_PREEMPT_COUNT);
2724 * A task struct has one reference for the use as "current".
2725 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2726 * schedule one last time. The schedule call will never return, and
2727 * the scheduled task must drop that reference.
2729 * We must observe prev->state before clearing prev->on_cpu (in
2730 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2731 * running on another CPU and we could rave with its RUNNING -> DEAD
2732 * transition, resulting in a double drop.
2734 prev_state = prev->state;
2735 vtime_task_switch(prev);
2736 perf_event_task_sched_in(prev, current);
2737 finish_lock_switch(rq, prev);
2738 finish_arch_post_lock_switch();
2740 fire_sched_in_preempt_notifiers(current);
2743 if (unlikely(prev_state == TASK_DEAD)) {
2744 if (prev->sched_class->task_dead)
2745 prev->sched_class->task_dead(prev);
2748 * Remove function-return probe instances associated with this
2749 * task and put them back on the free list.
2751 kprobe_flush_task(prev);
2753 /* Task is done with its stack. */
2754 put_task_stack(prev);
2756 put_task_struct(prev);
2759 tick_nohz_task_switch();
2765 /* rq->lock is NOT held, but preemption is disabled */
2766 static void __balance_callback(struct rq *rq)
2768 struct callback_head *head, *next;
2769 void (*func)(struct rq *rq);
2772 rq_lock_irqsave(rq, &rf);
2773 head = rq->balance_callback;
2774 rq->balance_callback = NULL;
2776 func = (void (*)(struct rq *))head->func;
2783 rq_unlock_irqrestore(rq, &rf);
2786 static inline void balance_callback(struct rq *rq)
2788 if (unlikely(rq->balance_callback))
2789 __balance_callback(rq);
2794 static inline void balance_callback(struct rq *rq)
2801 * schedule_tail - first thing a freshly forked thread must call.
2802 * @prev: the thread we just switched away from.
2804 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2805 __releases(rq->lock)
2810 * New tasks start with FORK_PREEMPT_COUNT, see there and
2811 * finish_task_switch() for details.
2813 * finish_task_switch() will drop rq->lock() and lower preempt_count
2814 * and the preempt_enable() will end up enabling preemption (on
2815 * PREEMPT_COUNT kernels).
2818 rq = finish_task_switch(prev);
2819 balance_callback(rq);
2822 if (current->set_child_tid)
2823 put_user(task_pid_vnr(current), current->set_child_tid);
2827 * context_switch - switch to the new MM and the new thread's register state.
2829 static __always_inline struct rq *
2830 context_switch(struct rq *rq, struct task_struct *prev,
2831 struct task_struct *next, struct rq_flags *rf)
2833 struct mm_struct *mm, *oldmm;
2835 prepare_task_switch(rq, prev, next);
2838 oldmm = prev->active_mm;
2840 * For paravirt, this is coupled with an exit in switch_to to
2841 * combine the page table reload and the switch backend into
2844 arch_start_context_switch(prev);
2847 next->active_mm = oldmm;
2849 enter_lazy_tlb(oldmm, next);
2851 switch_mm_irqs_off(oldmm, mm, next);
2854 prev->active_mm = NULL;
2855 rq->prev_mm = oldmm;
2858 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2861 * Since the runqueue lock will be released by the next
2862 * task (which is an invalid locking op but in the case
2863 * of the scheduler it's an obvious special-case), so we
2864 * do an early lockdep release here:
2866 rq_unpin_lock(rq, rf);
2867 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2869 /* Here we just switch the register state and the stack. */
2870 switch_to(prev, next, prev);
2873 return finish_task_switch(prev);
2877 * nr_running and nr_context_switches:
2879 * externally visible scheduler statistics: current number of runnable
2880 * threads, total number of context switches performed since bootup.
2882 unsigned long nr_running(void)
2884 unsigned long i, sum = 0;
2886 for_each_online_cpu(i)
2887 sum += cpu_rq(i)->nr_running;
2893 * Check if only the current task is running on the CPU.
2895 * Caution: this function does not check that the caller has disabled
2896 * preemption, thus the result might have a time-of-check-to-time-of-use
2897 * race. The caller is responsible to use it correctly, for example:
2899 * - from a non-preemptable section (of course)
2901 * - from a thread that is bound to a single CPU
2903 * - in a loop with very short iterations (e.g. a polling loop)
2905 bool single_task_running(void)
2907 return raw_rq()->nr_running == 1;
2909 EXPORT_SYMBOL(single_task_running);
2911 unsigned long long nr_context_switches(void)
2914 unsigned long long sum = 0;
2916 for_each_possible_cpu(i)
2917 sum += cpu_rq(i)->nr_switches;
2923 * IO-wait accounting, and how its mostly bollocks (on SMP).
2925 * The idea behind IO-wait account is to account the idle time that we could
2926 * have spend running if it were not for IO. That is, if we were to improve the
2927 * storage performance, we'd have a proportional reduction in IO-wait time.
2929 * This all works nicely on UP, where, when a task blocks on IO, we account
2930 * idle time as IO-wait, because if the storage were faster, it could've been
2931 * running and we'd not be idle.
2933 * This has been extended to SMP, by doing the same for each CPU. This however
2936 * Imagine for instance the case where two tasks block on one CPU, only the one
2937 * CPU will have IO-wait accounted, while the other has regular idle. Even
2938 * though, if the storage were faster, both could've ran at the same time,
2939 * utilising both CPUs.
2941 * This means, that when looking globally, the current IO-wait accounting on
2942 * SMP is a lower bound, by reason of under accounting.
2944 * Worse, since the numbers are provided per CPU, they are sometimes
2945 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2946 * associated with any one particular CPU, it can wake to another CPU than it
2947 * blocked on. This means the per CPU IO-wait number is meaningless.
2949 * Task CPU affinities can make all that even more 'interesting'.
2952 unsigned long nr_iowait(void)
2954 unsigned long i, sum = 0;
2956 for_each_possible_cpu(i)
2957 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2963 * Consumers of these two interfaces, like for example the cpufreq menu
2964 * governor are using nonsensical data. Boosting frequency for a CPU that has
2965 * IO-wait which might not even end up running the task when it does become
2969 unsigned long nr_iowait_cpu(int cpu)
2971 struct rq *this = cpu_rq(cpu);
2972 return atomic_read(&this->nr_iowait);
2975 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2977 struct rq *rq = this_rq();
2978 *nr_waiters = atomic_read(&rq->nr_iowait);
2979 *load = rq->load.weight;
2985 * sched_exec - execve() is a valuable balancing opportunity, because at
2986 * this point the task has the smallest effective memory and cache footprint.
2988 void sched_exec(void)
2990 struct task_struct *p = current;
2991 unsigned long flags;
2994 raw_spin_lock_irqsave(&p->pi_lock, flags);
2995 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2996 if (dest_cpu == smp_processor_id())
2999 if (likely(cpu_active(dest_cpu))) {
3000 struct migration_arg arg = { p, dest_cpu };
3002 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3003 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3007 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3012 DEFINE_PER_CPU(struct kernel_stat, kstat);
3013 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3015 EXPORT_PER_CPU_SYMBOL(kstat);
3016 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3019 * The function fair_sched_class.update_curr accesses the struct curr
3020 * and its field curr->exec_start; when called from task_sched_runtime(),
3021 * we observe a high rate of cache misses in practice.
3022 * Prefetching this data results in improved performance.
3024 static inline void prefetch_curr_exec_start(struct task_struct *p)
3026 #ifdef CONFIG_FAIR_GROUP_SCHED
3027 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3029 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3032 prefetch(&curr->exec_start);
3036 * Return accounted runtime for the task.
3037 * In case the task is currently running, return the runtime plus current's
3038 * pending runtime that have not been accounted yet.
3040 unsigned long long task_sched_runtime(struct task_struct *p)
3046 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3048 * 64-bit doesn't need locks to atomically read a 64bit value.
3049 * So we have a optimization chance when the task's delta_exec is 0.
3050 * Reading ->on_cpu is racy, but this is ok.
3052 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3053 * If we race with it entering CPU, unaccounted time is 0. This is
3054 * indistinguishable from the read occurring a few cycles earlier.
3055 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3056 * been accounted, so we're correct here as well.
3058 if (!p->on_cpu || !task_on_rq_queued(p))
3059 return p->se.sum_exec_runtime;
3062 rq = task_rq_lock(p, &rf);
3064 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3065 * project cycles that may never be accounted to this
3066 * thread, breaking clock_gettime().
3068 if (task_current(rq, p) && task_on_rq_queued(p)) {
3069 prefetch_curr_exec_start(p);
3070 update_rq_clock(rq);
3071 p->sched_class->update_curr(rq);
3073 ns = p->se.sum_exec_runtime;
3074 task_rq_unlock(rq, p, &rf);
3080 * This function gets called by the timer code, with HZ frequency.
3081 * We call it with interrupts disabled.
3083 void scheduler_tick(void)
3085 int cpu = smp_processor_id();
3086 struct rq *rq = cpu_rq(cpu);
3087 struct task_struct *curr = rq->curr;
3094 update_rq_clock(rq);
3095 curr->sched_class->task_tick(rq, curr, 0);
3096 cpu_load_update_active(rq);
3097 calc_global_load_tick(rq);
3101 perf_event_task_tick();
3104 rq->idle_balance = idle_cpu(cpu);
3105 trigger_load_balance(rq);
3107 rq_last_tick_reset(rq);
3110 #ifdef CONFIG_NO_HZ_FULL
3112 * scheduler_tick_max_deferment
3114 * Keep at least one tick per second when a single
3115 * active task is running because the scheduler doesn't
3116 * yet completely support full dynticks environment.
3118 * This makes sure that uptime, CFS vruntime, load
3119 * balancing, etc... continue to move forward, even
3120 * with a very low granularity.
3122 * Return: Maximum deferment in nanoseconds.
3124 u64 scheduler_tick_max_deferment(void)
3126 struct rq *rq = this_rq();
3127 unsigned long next, now = READ_ONCE(jiffies);
3129 next = rq->last_sched_tick + HZ;
3131 if (time_before_eq(next, now))
3134 return jiffies_to_nsecs(next - now);
3138 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3139 defined(CONFIG_PREEMPT_TRACER))
3141 * If the value passed in is equal to the current preempt count
3142 * then we just disabled preemption. Start timing the latency.
3144 static inline void preempt_latency_start(int val)
3146 if (preempt_count() == val) {
3147 unsigned long ip = get_lock_parent_ip();
3148 #ifdef CONFIG_DEBUG_PREEMPT
3149 current->preempt_disable_ip = ip;
3151 trace_preempt_off(CALLER_ADDR0, ip);
3155 void preempt_count_add(int val)
3157 #ifdef CONFIG_DEBUG_PREEMPT
3161 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3164 __preempt_count_add(val);
3165 #ifdef CONFIG_DEBUG_PREEMPT
3167 * Spinlock count overflowing soon?
3169 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3172 preempt_latency_start(val);
3174 EXPORT_SYMBOL(preempt_count_add);
3175 NOKPROBE_SYMBOL(preempt_count_add);
3178 * If the value passed in equals to the current preempt count
3179 * then we just enabled preemption. Stop timing the latency.
3181 static inline void preempt_latency_stop(int val)
3183 if (preempt_count() == val)
3184 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3187 void preempt_count_sub(int val)
3189 #ifdef CONFIG_DEBUG_PREEMPT
3193 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3196 * Is the spinlock portion underflowing?
3198 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3199 !(preempt_count() & PREEMPT_MASK)))
3203 preempt_latency_stop(val);
3204 __preempt_count_sub(val);
3206 EXPORT_SYMBOL(preempt_count_sub);
3207 NOKPROBE_SYMBOL(preempt_count_sub);
3210 static inline void preempt_latency_start(int val) { }
3211 static inline void preempt_latency_stop(int val) { }
3214 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3216 #ifdef CONFIG_DEBUG_PREEMPT
3217 return p->preempt_disable_ip;
3224 * Print scheduling while atomic bug:
3226 static noinline void __schedule_bug(struct task_struct *prev)
3228 /* Save this before calling printk(), since that will clobber it */
3229 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3231 if (oops_in_progress)
3234 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3235 prev->comm, prev->pid, preempt_count());
3237 debug_show_held_locks(prev);
3239 if (irqs_disabled())
3240 print_irqtrace_events(prev);
3241 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3242 && in_atomic_preempt_off()) {
3243 pr_err("Preemption disabled at:");
3244 print_ip_sym(preempt_disable_ip);
3248 panic("scheduling while atomic\n");
3251 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3255 * Various schedule()-time debugging checks and statistics:
3257 static inline void schedule_debug(struct task_struct *prev)
3259 #ifdef CONFIG_SCHED_STACK_END_CHECK
3260 if (task_stack_end_corrupted(prev))
3261 panic("corrupted stack end detected inside scheduler\n");
3264 if (unlikely(in_atomic_preempt_off())) {
3265 __schedule_bug(prev);
3266 preempt_count_set(PREEMPT_DISABLED);
3270 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3272 schedstat_inc(this_rq()->sched_count);
3276 * Pick up the highest-prio task:
3278 static inline struct task_struct *
3279 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3281 const struct sched_class *class;
3282 struct task_struct *p;
3285 * Optimization: we know that if all tasks are in the fair class we can
3286 * call that function directly, but only if the @prev task wasn't of a
3287 * higher scheduling class, because otherwise those loose the
3288 * opportunity to pull in more work from other CPUs.
3290 if (likely((prev->sched_class == &idle_sched_class ||
3291 prev->sched_class == &fair_sched_class) &&
3292 rq->nr_running == rq->cfs.h_nr_running)) {
3294 p = fair_sched_class.pick_next_task(rq, prev, rf);
3295 if (unlikely(p == RETRY_TASK))
3298 /* Assumes fair_sched_class->next == idle_sched_class */
3300 p = idle_sched_class.pick_next_task(rq, prev, rf);
3306 for_each_class(class) {
3307 p = class->pick_next_task(rq, prev, rf);
3309 if (unlikely(p == RETRY_TASK))
3315 /* The idle class should always have a runnable task: */
3320 * __schedule() is the main scheduler function.
3322 * The main means of driving the scheduler and thus entering this function are:
3324 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3326 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3327 * paths. For example, see arch/x86/entry_64.S.
3329 * To drive preemption between tasks, the scheduler sets the flag in timer
3330 * interrupt handler scheduler_tick().
3332 * 3. Wakeups don't really cause entry into schedule(). They add a
3333 * task to the run-queue and that's it.
3335 * Now, if the new task added to the run-queue preempts the current
3336 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3337 * called on the nearest possible occasion:
3339 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3341 * - in syscall or exception context, at the next outmost
3342 * preempt_enable(). (this might be as soon as the wake_up()'s
3345 * - in IRQ context, return from interrupt-handler to
3346 * preemptible context
3348 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3351 * - cond_resched() call
3352 * - explicit schedule() call
3353 * - return from syscall or exception to user-space
3354 * - return from interrupt-handler to user-space
3356 * WARNING: must be called with preemption disabled!
3358 static void __sched notrace __schedule(bool preempt)
3360 struct task_struct *prev, *next;
3361 unsigned long *switch_count;
3366 cpu = smp_processor_id();
3370 schedule_debug(prev);
3372 if (sched_feat(HRTICK))
3375 local_irq_disable();
3376 rcu_note_context_switch();
3379 * Make sure that signal_pending_state()->signal_pending() below
3380 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3381 * done by the caller to avoid the race with signal_wake_up().
3383 smp_mb__before_spinlock();
3386 /* Promote REQ to ACT */
3387 rq->clock_update_flags <<= 1;
3389 switch_count = &prev->nivcsw;
3390 if (!preempt && prev->state) {
3391 if (unlikely(signal_pending_state(prev->state, prev))) {
3392 prev->state = TASK_RUNNING;
3394 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3397 if (prev->in_iowait) {
3398 atomic_inc(&rq->nr_iowait);
3399 delayacct_blkio_start();
3403 * If a worker went to sleep, notify and ask workqueue
3404 * whether it wants to wake up a task to maintain
3407 if (prev->flags & PF_WQ_WORKER) {
3408 struct task_struct *to_wakeup;
3410 to_wakeup = wq_worker_sleeping(prev);
3412 try_to_wake_up_local(to_wakeup, &rf);
3415 switch_count = &prev->nvcsw;
3418 if (task_on_rq_queued(prev))
3419 update_rq_clock(rq);
3421 next = pick_next_task(rq, prev, &rf);
3422 clear_tsk_need_resched(prev);
3423 clear_preempt_need_resched();
3425 if (likely(prev != next)) {
3430 trace_sched_switch(preempt, prev, next);
3432 /* Also unlocks the rq: */
3433 rq = context_switch(rq, prev, next, &rf);
3435 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3436 rq_unlock_irq(rq, &rf);
3439 balance_callback(rq);
3442 void __noreturn do_task_dead(void)
3445 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3446 * when the following two conditions become true.
3447 * - There is race condition of mmap_sem (It is acquired by
3449 * - SMI occurs before setting TASK_RUNINNG.
3450 * (or hypervisor of virtual machine switches to other guest)
3451 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3453 * To avoid it, we have to wait for releasing tsk->pi_lock which
3454 * is held by try_to_wake_up()
3457 raw_spin_unlock_wait(¤t->pi_lock);
3459 /* Causes final put_task_struct in finish_task_switch(): */
3460 __set_current_state(TASK_DEAD);
3462 /* Tell freezer to ignore us: */
3463 current->flags |= PF_NOFREEZE;
3468 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3473 static inline void sched_submit_work(struct task_struct *tsk)
3475 if (!tsk->state || tsk_is_pi_blocked(tsk))
3478 * If we are going to sleep and we have plugged IO queued,
3479 * make sure to submit it to avoid deadlocks.
3481 if (blk_needs_flush_plug(tsk))
3482 blk_schedule_flush_plug(tsk);
3485 asmlinkage __visible void __sched schedule(void)
3487 struct task_struct *tsk = current;
3489 sched_submit_work(tsk);
3493 sched_preempt_enable_no_resched();
3494 } while (need_resched());
3496 EXPORT_SYMBOL(schedule);
3498 #ifdef CONFIG_CONTEXT_TRACKING
3499 asmlinkage __visible void __sched schedule_user(void)
3502 * If we come here after a random call to set_need_resched(),
3503 * or we have been woken up remotely but the IPI has not yet arrived,
3504 * we haven't yet exited the RCU idle mode. Do it here manually until
3505 * we find a better solution.
3507 * NB: There are buggy callers of this function. Ideally we
3508 * should warn if prev_state != CONTEXT_USER, but that will trigger
3509 * too frequently to make sense yet.
3511 enum ctx_state prev_state = exception_enter();
3513 exception_exit(prev_state);
3518 * schedule_preempt_disabled - called with preemption disabled
3520 * Returns with preemption disabled. Note: preempt_count must be 1
3522 void __sched schedule_preempt_disabled(void)
3524 sched_preempt_enable_no_resched();
3529 static void __sched notrace preempt_schedule_common(void)
3533 * Because the function tracer can trace preempt_count_sub()
3534 * and it also uses preempt_enable/disable_notrace(), if
3535 * NEED_RESCHED is set, the preempt_enable_notrace() called
3536 * by the function tracer will call this function again and
3537 * cause infinite recursion.
3539 * Preemption must be disabled here before the function
3540 * tracer can trace. Break up preempt_disable() into two
3541 * calls. One to disable preemption without fear of being
3542 * traced. The other to still record the preemption latency,
3543 * which can also be traced by the function tracer.
3545 preempt_disable_notrace();
3546 preempt_latency_start(1);
3548 preempt_latency_stop(1);
3549 preempt_enable_no_resched_notrace();
3552 * Check again in case we missed a preemption opportunity
3553 * between schedule and now.
3555 } while (need_resched());
3558 #ifdef CONFIG_PREEMPT
3560 * this is the entry point to schedule() from in-kernel preemption
3561 * off of preempt_enable. Kernel preemptions off return from interrupt
3562 * occur there and call schedule directly.
3564 asmlinkage __visible void __sched notrace preempt_schedule(void)
3567 * If there is a non-zero preempt_count or interrupts are disabled,
3568 * we do not want to preempt the current task. Just return..
3570 if (likely(!preemptible()))
3573 preempt_schedule_common();
3575 NOKPROBE_SYMBOL(preempt_schedule);
3576 EXPORT_SYMBOL(preempt_schedule);
3579 * preempt_schedule_notrace - preempt_schedule called by tracing
3581 * The tracing infrastructure uses preempt_enable_notrace to prevent
3582 * recursion and tracing preempt enabling caused by the tracing
3583 * infrastructure itself. But as tracing can happen in areas coming
3584 * from userspace or just about to enter userspace, a preempt enable
3585 * can occur before user_exit() is called. This will cause the scheduler
3586 * to be called when the system is still in usermode.
3588 * To prevent this, the preempt_enable_notrace will use this function
3589 * instead of preempt_schedule() to exit user context if needed before
3590 * calling the scheduler.
3592 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3594 enum ctx_state prev_ctx;
3596 if (likely(!preemptible()))
3601 * Because the function tracer can trace preempt_count_sub()
3602 * and it also uses preempt_enable/disable_notrace(), if
3603 * NEED_RESCHED is set, the preempt_enable_notrace() called
3604 * by the function tracer will call this function again and
3605 * cause infinite recursion.
3607 * Preemption must be disabled here before the function
3608 * tracer can trace. Break up preempt_disable() into two
3609 * calls. One to disable preemption without fear of being
3610 * traced. The other to still record the preemption latency,
3611 * which can also be traced by the function tracer.
3613 preempt_disable_notrace();
3614 preempt_latency_start(1);
3616 * Needs preempt disabled in case user_exit() is traced
3617 * and the tracer calls preempt_enable_notrace() causing
3618 * an infinite recursion.
3620 prev_ctx = exception_enter();
3622 exception_exit(prev_ctx);
3624 preempt_latency_stop(1);
3625 preempt_enable_no_resched_notrace();
3626 } while (need_resched());
3628 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3630 #endif /* CONFIG_PREEMPT */
3633 * this is the entry point to schedule() from kernel preemption
3634 * off of irq context.
3635 * Note, that this is called and return with irqs disabled. This will
3636 * protect us against recursive calling from irq.
3638 asmlinkage __visible void __sched preempt_schedule_irq(void)
3640 enum ctx_state prev_state;
3642 /* Catch callers which need to be fixed */
3643 BUG_ON(preempt_count() || !irqs_disabled());
3645 prev_state = exception_enter();
3651 local_irq_disable();
3652 sched_preempt_enable_no_resched();
3653 } while (need_resched());
3655 exception_exit(prev_state);
3658 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3661 return try_to_wake_up(curr->private, mode, wake_flags);
3663 EXPORT_SYMBOL(default_wake_function);
3665 #ifdef CONFIG_RT_MUTEXES
3668 * rt_mutex_setprio - set the current priority of a task
3670 * @prio: prio value (kernel-internal form)
3672 * This function changes the 'effective' priority of a task. It does
3673 * not touch ->normal_prio like __setscheduler().
3675 * Used by the rt_mutex code to implement priority inheritance
3676 * logic. Call site only calls if the priority of the task changed.
3678 void rt_mutex_setprio(struct task_struct *p, int prio)
3680 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3681 const struct sched_class *prev_class;
3685 BUG_ON(prio > MAX_PRIO);
3687 rq = __task_rq_lock(p, &rf);
3688 update_rq_clock(rq);
3691 * Idle task boosting is a nono in general. There is one
3692 * exception, when PREEMPT_RT and NOHZ is active:
3694 * The idle task calls get_next_timer_interrupt() and holds
3695 * the timer wheel base->lock on the CPU and another CPU wants
3696 * to access the timer (probably to cancel it). We can safely
3697 * ignore the boosting request, as the idle CPU runs this code
3698 * with interrupts disabled and will complete the lock
3699 * protected section without being interrupted. So there is no
3700 * real need to boost.
3702 if (unlikely(p == rq->idle)) {
3703 WARN_ON(p != rq->curr);
3704 WARN_ON(p->pi_blocked_on);
3708 trace_sched_pi_setprio(p, prio);
3711 if (oldprio == prio)
3712 queue_flag &= ~DEQUEUE_MOVE;
3714 prev_class = p->sched_class;
3715 queued = task_on_rq_queued(p);
3716 running = task_current(rq, p);
3718 dequeue_task(rq, p, queue_flag);
3720 put_prev_task(rq, p);
3723 * Boosting condition are:
3724 * 1. -rt task is running and holds mutex A
3725 * --> -dl task blocks on mutex A
3727 * 2. -dl task is running and holds mutex A
3728 * --> -dl task blocks on mutex A and could preempt the
3731 if (dl_prio(prio)) {
3732 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3733 if (!dl_prio(p->normal_prio) ||
3734 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3735 p->dl.dl_boosted = 1;
3736 queue_flag |= ENQUEUE_REPLENISH;
3738 p->dl.dl_boosted = 0;
3739 p->sched_class = &dl_sched_class;
3740 } else if (rt_prio(prio)) {
3741 if (dl_prio(oldprio))
3742 p->dl.dl_boosted = 0;
3744 queue_flag |= ENQUEUE_HEAD;
3745 p->sched_class = &rt_sched_class;
3747 if (dl_prio(oldprio))
3748 p->dl.dl_boosted = 0;
3749 if (rt_prio(oldprio))
3751 p->sched_class = &fair_sched_class;
3757 enqueue_task(rq, p, queue_flag);
3759 set_curr_task(rq, p);
3761 check_class_changed(rq, p, prev_class, oldprio);
3763 /* Avoid rq from going away on us: */
3765 __task_rq_unlock(rq, &rf);
3767 balance_callback(rq);
3772 void set_user_nice(struct task_struct *p, long nice)
3774 bool queued, running;
3775 int old_prio, delta;
3779 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3782 * We have to be careful, if called from sys_setpriority(),
3783 * the task might be in the middle of scheduling on another CPU.
3785 rq = task_rq_lock(p, &rf);
3786 update_rq_clock(rq);
3789 * The RT priorities are set via sched_setscheduler(), but we still
3790 * allow the 'normal' nice value to be set - but as expected
3791 * it wont have any effect on scheduling until the task is
3792 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3794 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3795 p->static_prio = NICE_TO_PRIO(nice);
3798 queued = task_on_rq_queued(p);
3799 running = task_current(rq, p);
3801 dequeue_task(rq, p, DEQUEUE_SAVE);
3803 put_prev_task(rq, p);
3805 p->static_prio = NICE_TO_PRIO(nice);
3808 p->prio = effective_prio(p);
3809 delta = p->prio - old_prio;
3812 enqueue_task(rq, p, ENQUEUE_RESTORE);
3814 * If the task increased its priority or is running and
3815 * lowered its priority, then reschedule its CPU:
3817 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3821 set_curr_task(rq, p);
3823 task_rq_unlock(rq, p, &rf);
3825 EXPORT_SYMBOL(set_user_nice);
3828 * can_nice - check if a task can reduce its nice value
3832 int can_nice(const struct task_struct *p, const int nice)
3834 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3835 int nice_rlim = nice_to_rlimit(nice);
3837 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3838 capable(CAP_SYS_NICE));
3841 #ifdef __ARCH_WANT_SYS_NICE
3844 * sys_nice - change the priority of the current process.
3845 * @increment: priority increment
3847 * sys_setpriority is a more generic, but much slower function that
3848 * does similar things.
3850 SYSCALL_DEFINE1(nice, int, increment)
3855 * Setpriority might change our priority at the same moment.
3856 * We don't have to worry. Conceptually one call occurs first
3857 * and we have a single winner.
3859 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3860 nice = task_nice(current) + increment;
3862 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3863 if (increment < 0 && !can_nice(current, nice))
3866 retval = security_task_setnice(current, nice);
3870 set_user_nice(current, nice);
3877 * task_prio - return the priority value of a given task.
3878 * @p: the task in question.
3880 * Return: The priority value as seen by users in /proc.
3881 * RT tasks are offset by -200. Normal tasks are centered
3882 * around 0, value goes from -16 to +15.
3884 int task_prio(const struct task_struct *p)
3886 return p->prio - MAX_RT_PRIO;
3890 * idle_cpu - is a given CPU idle currently?
3891 * @cpu: the processor in question.
3893 * Return: 1 if the CPU is currently idle. 0 otherwise.
3895 int idle_cpu(int cpu)
3897 struct rq *rq = cpu_rq(cpu);
3899 if (rq->curr != rq->idle)
3906 if (!llist_empty(&rq->wake_list))
3914 * idle_task - return the idle task for a given CPU.
3915 * @cpu: the processor in question.
3917 * Return: The idle task for the CPU @cpu.
3919 struct task_struct *idle_task(int cpu)
3921 return cpu_rq(cpu)->idle;
3925 * find_process_by_pid - find a process with a matching PID value.
3926 * @pid: the pid in question.
3928 * The task of @pid, if found. %NULL otherwise.
3930 static struct task_struct *find_process_by_pid(pid_t pid)
3932 return pid ? find_task_by_vpid(pid) : current;
3936 * This function initializes the sched_dl_entity of a newly becoming
3937 * SCHED_DEADLINE task.
3939 * Only the static values are considered here, the actual runtime and the
3940 * absolute deadline will be properly calculated when the task is enqueued
3941 * for the first time with its new policy.
3944 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3946 struct sched_dl_entity *dl_se = &p->dl;
3948 dl_se->dl_runtime = attr->sched_runtime;
3949 dl_se->dl_deadline = attr->sched_deadline;
3950 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3951 dl_se->flags = attr->sched_flags;
3952 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3955 * Changing the parameters of a task is 'tricky' and we're not doing
3956 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3958 * What we SHOULD do is delay the bandwidth release until the 0-lag
3959 * point. This would include retaining the task_struct until that time
3960 * and change dl_overflow() to not immediately decrement the current
3963 * Instead we retain the current runtime/deadline and let the new
3964 * parameters take effect after the current reservation period lapses.
3965 * This is safe (albeit pessimistic) because the 0-lag point is always
3966 * before the current scheduling deadline.
3968 * We can still have temporary overloads because we do not delay the
3969 * change in bandwidth until that time; so admission control is
3970 * not on the safe side. It does however guarantee tasks will never
3971 * consume more than promised.
3976 * sched_setparam() passes in -1 for its policy, to let the functions
3977 * it calls know not to change it.
3979 #define SETPARAM_POLICY -1
3981 static void __setscheduler_params(struct task_struct *p,
3982 const struct sched_attr *attr)
3984 int policy = attr->sched_policy;
3986 if (policy == SETPARAM_POLICY)
3991 if (dl_policy(policy))
3992 __setparam_dl(p, attr);
3993 else if (fair_policy(policy))
3994 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3997 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3998 * !rt_policy. Always setting this ensures that things like
3999 * getparam()/getattr() don't report silly values for !rt tasks.
4001 p->rt_priority = attr->sched_priority;
4002 p->normal_prio = normal_prio(p);
4006 /* Actually do priority change: must hold pi & rq lock. */
4007 static void __setscheduler(struct rq *rq, struct task_struct *p,
4008 const struct sched_attr *attr, bool keep_boost)
4010 __setscheduler_params(p, attr);
4013 * Keep a potential priority boosting if called from
4014 * sched_setscheduler().
4017 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
4019 p->prio = normal_prio(p);
4021 if (dl_prio(p->prio))
4022 p->sched_class = &dl_sched_class;
4023 else if (rt_prio(p->prio))
4024 p->sched_class = &rt_sched_class;
4026 p->sched_class = &fair_sched_class;
4030 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4032 struct sched_dl_entity *dl_se = &p->dl;
4034 attr->sched_priority = p->rt_priority;
4035 attr->sched_runtime = dl_se->dl_runtime;
4036 attr->sched_deadline = dl_se->dl_deadline;
4037 attr->sched_period = dl_se->dl_period;
4038 attr->sched_flags = dl_se->flags;
4042 * This function validates the new parameters of a -deadline task.
4043 * We ask for the deadline not being zero, and greater or equal
4044 * than the runtime, as well as the period of being zero or
4045 * greater than deadline. Furthermore, we have to be sure that
4046 * user parameters are above the internal resolution of 1us (we
4047 * check sched_runtime only since it is always the smaller one) and
4048 * below 2^63 ns (we have to check both sched_deadline and
4049 * sched_period, as the latter can be zero).
4052 __checkparam_dl(const struct sched_attr *attr)
4055 if (attr->sched_deadline == 0)
4059 * Since we truncate DL_SCALE bits, make sure we're at least
4062 if (attr->sched_runtime < (1ULL << DL_SCALE))
4066 * Since we use the MSB for wrap-around and sign issues, make
4067 * sure it's not set (mind that period can be equal to zero).
4069 if (attr->sched_deadline & (1ULL << 63) ||
4070 attr->sched_period & (1ULL << 63))
4073 /* runtime <= deadline <= period (if period != 0) */
4074 if ((attr->sched_period != 0 &&
4075 attr->sched_period < attr->sched_deadline) ||
4076 attr->sched_deadline < attr->sched_runtime)
4083 * Check the target process has a UID that matches the current process's:
4085 static bool check_same_owner(struct task_struct *p)
4087 const struct cred *cred = current_cred(), *pcred;
4091 pcred = __task_cred(p);
4092 match = (uid_eq(cred->euid, pcred->euid) ||
4093 uid_eq(cred->euid, pcred->uid));
4098 static bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
4100 struct sched_dl_entity *dl_se = &p->dl;
4102 if (dl_se->dl_runtime != attr->sched_runtime ||
4103 dl_se->dl_deadline != attr->sched_deadline ||
4104 dl_se->dl_period != attr->sched_period ||
4105 dl_se->flags != attr->sched_flags)
4111 static int __sched_setscheduler(struct task_struct *p,
4112 const struct sched_attr *attr,
4115 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4116 MAX_RT_PRIO - 1 - attr->sched_priority;
4117 int retval, oldprio, oldpolicy = -1, queued, running;
4118 int new_effective_prio, policy = attr->sched_policy;
4119 const struct sched_class *prev_class;
4122 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4125 /* May grab non-irq protected spin_locks: */
4126 BUG_ON(in_interrupt());
4128 /* Double check policy once rq lock held: */
4130 reset_on_fork = p->sched_reset_on_fork;
4131 policy = oldpolicy = p->policy;
4133 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4135 if (!valid_policy(policy))
4139 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4143 * Valid priorities for SCHED_FIFO and SCHED_RR are
4144 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4145 * SCHED_BATCH and SCHED_IDLE is 0.
4147 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4148 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4150 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4151 (rt_policy(policy) != (attr->sched_priority != 0)))
4155 * Allow unprivileged RT tasks to decrease priority:
4157 if (user && !capable(CAP_SYS_NICE)) {
4158 if (fair_policy(policy)) {
4159 if (attr->sched_nice < task_nice(p) &&
4160 !can_nice(p, attr->sched_nice))
4164 if (rt_policy(policy)) {
4165 unsigned long rlim_rtprio =
4166 task_rlimit(p, RLIMIT_RTPRIO);
4168 /* Can't set/change the rt policy: */
4169 if (policy != p->policy && !rlim_rtprio)
4172 /* Can't increase priority: */
4173 if (attr->sched_priority > p->rt_priority &&
4174 attr->sched_priority > rlim_rtprio)
4179 * Can't set/change SCHED_DEADLINE policy at all for now
4180 * (safest behavior); in the future we would like to allow
4181 * unprivileged DL tasks to increase their relative deadline
4182 * or reduce their runtime (both ways reducing utilization)
4184 if (dl_policy(policy))
4188 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4189 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4191 if (idle_policy(p->policy) && !idle_policy(policy)) {
4192 if (!can_nice(p, task_nice(p)))
4196 /* Can't change other user's priorities: */
4197 if (!check_same_owner(p))
4200 /* Normal users shall not reset the sched_reset_on_fork flag: */
4201 if (p->sched_reset_on_fork && !reset_on_fork)
4206 retval = security_task_setscheduler(p);
4212 * Make sure no PI-waiters arrive (or leave) while we are
4213 * changing the priority of the task:
4215 * To be able to change p->policy safely, the appropriate
4216 * runqueue lock must be held.
4218 rq = task_rq_lock(p, &rf);
4219 update_rq_clock(rq);
4222 * Changing the policy of the stop threads its a very bad idea:
4224 if (p == rq->stop) {
4225 task_rq_unlock(rq, p, &rf);
4230 * If not changing anything there's no need to proceed further,
4231 * but store a possible modification of reset_on_fork.
4233 if (unlikely(policy == p->policy)) {
4234 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4236 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4238 if (dl_policy(policy) && dl_param_changed(p, attr))
4241 p->sched_reset_on_fork = reset_on_fork;
4242 task_rq_unlock(rq, p, &rf);
4248 #ifdef CONFIG_RT_GROUP_SCHED
4250 * Do not allow realtime tasks into groups that have no runtime
4253 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4254 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4255 !task_group_is_autogroup(task_group(p))) {
4256 task_rq_unlock(rq, p, &rf);
4261 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4262 cpumask_t *span = rq->rd->span;
4265 * Don't allow tasks with an affinity mask smaller than
4266 * the entire root_domain to become SCHED_DEADLINE. We
4267 * will also fail if there's no bandwidth available.
4269 if (!cpumask_subset(span, &p->cpus_allowed) ||
4270 rq->rd->dl_bw.bw == 0) {
4271 task_rq_unlock(rq, p, &rf);
4278 /* Re-check policy now with rq lock held: */
4279 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4280 policy = oldpolicy = -1;
4281 task_rq_unlock(rq, p, &rf);
4286 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4287 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4290 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4291 task_rq_unlock(rq, p, &rf);
4295 p->sched_reset_on_fork = reset_on_fork;
4300 * Take priority boosted tasks into account. If the new
4301 * effective priority is unchanged, we just store the new
4302 * normal parameters and do not touch the scheduler class and
4303 * the runqueue. This will be done when the task deboost
4306 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4307 if (new_effective_prio == oldprio)
4308 queue_flags &= ~DEQUEUE_MOVE;
4311 queued = task_on_rq_queued(p);
4312 running = task_current(rq, p);
4314 dequeue_task(rq, p, queue_flags);
4316 put_prev_task(rq, p);
4318 prev_class = p->sched_class;
4319 __setscheduler(rq, p, attr, pi);
4323 * We enqueue to tail when the priority of a task is
4324 * increased (user space view).
4326 if (oldprio < p->prio)
4327 queue_flags |= ENQUEUE_HEAD;
4329 enqueue_task(rq, p, queue_flags);
4332 set_curr_task(rq, p);
4334 check_class_changed(rq, p, prev_class, oldprio);
4336 /* Avoid rq from going away on us: */
4338 task_rq_unlock(rq, p, &rf);
4341 rt_mutex_adjust_pi(p);
4343 /* Run balance callbacks after we've adjusted the PI chain: */
4344 balance_callback(rq);
4350 static int _sched_setscheduler(struct task_struct *p, int policy,
4351 const struct sched_param *param, bool check)
4353 struct sched_attr attr = {
4354 .sched_policy = policy,
4355 .sched_priority = param->sched_priority,
4356 .sched_nice = PRIO_TO_NICE(p->static_prio),
4359 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4360 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4361 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4362 policy &= ~SCHED_RESET_ON_FORK;
4363 attr.sched_policy = policy;
4366 return __sched_setscheduler(p, &attr, check, true);
4369 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4370 * @p: the task in question.
4371 * @policy: new policy.
4372 * @param: structure containing the new RT priority.
4374 * Return: 0 on success. An error code otherwise.
4376 * NOTE that the task may be already dead.
4378 int sched_setscheduler(struct task_struct *p, int policy,
4379 const struct sched_param *param)
4381 return _sched_setscheduler(p, policy, param, true);
4383 EXPORT_SYMBOL_GPL(sched_setscheduler);
4385 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4387 return __sched_setscheduler(p, attr, true, true);
4389 EXPORT_SYMBOL_GPL(sched_setattr);
4392 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4393 * @p: the task in question.
4394 * @policy: new policy.
4395 * @param: structure containing the new RT priority.
4397 * Just like sched_setscheduler, only don't bother checking if the
4398 * current context has permission. For example, this is needed in
4399 * stop_machine(): we create temporary high priority worker threads,
4400 * but our caller might not have that capability.
4402 * Return: 0 on success. An error code otherwise.
4404 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4405 const struct sched_param *param)
4407 return _sched_setscheduler(p, policy, param, false);
4409 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4412 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4414 struct sched_param lparam;
4415 struct task_struct *p;
4418 if (!param || pid < 0)
4420 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4425 p = find_process_by_pid(pid);
4427 retval = sched_setscheduler(p, policy, &lparam);
4434 * Mimics kernel/events/core.c perf_copy_attr().
4436 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4441 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4444 /* Zero the full structure, so that a short copy will be nice: */
4445 memset(attr, 0, sizeof(*attr));
4447 ret = get_user(size, &uattr->size);
4451 /* Bail out on silly large: */
4452 if (size > PAGE_SIZE)
4455 /* ABI compatibility quirk: */
4457 size = SCHED_ATTR_SIZE_VER0;
4459 if (size < SCHED_ATTR_SIZE_VER0)
4463 * If we're handed a bigger struct than we know of,
4464 * ensure all the unknown bits are 0 - i.e. new
4465 * user-space does not rely on any kernel feature
4466 * extensions we dont know about yet.
4468 if (size > sizeof(*attr)) {
4469 unsigned char __user *addr;
4470 unsigned char __user *end;
4473 addr = (void __user *)uattr + sizeof(*attr);
4474 end = (void __user *)uattr + size;
4476 for (; addr < end; addr++) {
4477 ret = get_user(val, addr);
4483 size = sizeof(*attr);
4486 ret = copy_from_user(attr, uattr, size);
4491 * XXX: Do we want to be lenient like existing syscalls; or do we want
4492 * to be strict and return an error on out-of-bounds values?
4494 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4499 put_user(sizeof(*attr), &uattr->size);
4504 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4505 * @pid: the pid in question.
4506 * @policy: new policy.
4507 * @param: structure containing the new RT priority.
4509 * Return: 0 on success. An error code otherwise.
4511 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4516 return do_sched_setscheduler(pid, policy, param);
4520 * sys_sched_setparam - set/change the RT priority of a thread
4521 * @pid: the pid in question.
4522 * @param: structure containing the new RT priority.
4524 * Return: 0 on success. An error code otherwise.
4526 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4528 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4532 * sys_sched_setattr - same as above, but with extended sched_attr
4533 * @pid: the pid in question.
4534 * @uattr: structure containing the extended parameters.
4535 * @flags: for future extension.
4537 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4538 unsigned int, flags)
4540 struct sched_attr attr;
4541 struct task_struct *p;
4544 if (!uattr || pid < 0 || flags)
4547 retval = sched_copy_attr(uattr, &attr);
4551 if ((int)attr.sched_policy < 0)
4556 p = find_process_by_pid(pid);
4558 retval = sched_setattr(p, &attr);
4565 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4566 * @pid: the pid in question.
4568 * Return: On success, the policy of the thread. Otherwise, a negative error
4571 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4573 struct task_struct *p;
4581 p = find_process_by_pid(pid);
4583 retval = security_task_getscheduler(p);
4586 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4593 * sys_sched_getparam - get the RT priority of a thread
4594 * @pid: the pid in question.
4595 * @param: structure containing the RT priority.
4597 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4600 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4602 struct sched_param lp = { .sched_priority = 0 };
4603 struct task_struct *p;
4606 if (!param || pid < 0)
4610 p = find_process_by_pid(pid);
4615 retval = security_task_getscheduler(p);
4619 if (task_has_rt_policy(p))
4620 lp.sched_priority = p->rt_priority;
4624 * This one might sleep, we cannot do it with a spinlock held ...
4626 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4635 static int sched_read_attr(struct sched_attr __user *uattr,
4636 struct sched_attr *attr,
4641 if (!access_ok(VERIFY_WRITE, uattr, usize))
4645 * If we're handed a smaller struct than we know of,
4646 * ensure all the unknown bits are 0 - i.e. old
4647 * user-space does not get uncomplete information.
4649 if (usize < sizeof(*attr)) {
4650 unsigned char *addr;
4653 addr = (void *)attr + usize;
4654 end = (void *)attr + sizeof(*attr);
4656 for (; addr < end; addr++) {
4664 ret = copy_to_user(uattr, attr, attr->size);
4672 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4673 * @pid: the pid in question.
4674 * @uattr: structure containing the extended parameters.
4675 * @size: sizeof(attr) for fwd/bwd comp.
4676 * @flags: for future extension.
4678 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4679 unsigned int, size, unsigned int, flags)
4681 struct sched_attr attr = {
4682 .size = sizeof(struct sched_attr),
4684 struct task_struct *p;
4687 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4688 size < SCHED_ATTR_SIZE_VER0 || flags)
4692 p = find_process_by_pid(pid);
4697 retval = security_task_getscheduler(p);
4701 attr.sched_policy = p->policy;
4702 if (p->sched_reset_on_fork)
4703 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4704 if (task_has_dl_policy(p))
4705 __getparam_dl(p, &attr);
4706 else if (task_has_rt_policy(p))
4707 attr.sched_priority = p->rt_priority;
4709 attr.sched_nice = task_nice(p);
4713 retval = sched_read_attr(uattr, &attr, size);
4721 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4723 cpumask_var_t cpus_allowed, new_mask;
4724 struct task_struct *p;
4729 p = find_process_by_pid(pid);
4735 /* Prevent p going away */
4739 if (p->flags & PF_NO_SETAFFINITY) {
4743 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4747 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4749 goto out_free_cpus_allowed;
4752 if (!check_same_owner(p)) {
4754 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4756 goto out_free_new_mask;
4761 retval = security_task_setscheduler(p);
4763 goto out_free_new_mask;
4766 cpuset_cpus_allowed(p, cpus_allowed);
4767 cpumask_and(new_mask, in_mask, cpus_allowed);
4770 * Since bandwidth control happens on root_domain basis,
4771 * if admission test is enabled, we only admit -deadline
4772 * tasks allowed to run on all the CPUs in the task's
4776 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4778 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4781 goto out_free_new_mask;
4787 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4790 cpuset_cpus_allowed(p, cpus_allowed);
4791 if (!cpumask_subset(new_mask, cpus_allowed)) {
4793 * We must have raced with a concurrent cpuset
4794 * update. Just reset the cpus_allowed to the
4795 * cpuset's cpus_allowed
4797 cpumask_copy(new_mask, cpus_allowed);
4802 free_cpumask_var(new_mask);
4803 out_free_cpus_allowed:
4804 free_cpumask_var(cpus_allowed);
4810 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4811 struct cpumask *new_mask)
4813 if (len < cpumask_size())
4814 cpumask_clear(new_mask);
4815 else if (len > cpumask_size())
4816 len = cpumask_size();
4818 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4822 * sys_sched_setaffinity - set the CPU affinity of a process
4823 * @pid: pid of the process
4824 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4825 * @user_mask_ptr: user-space pointer to the new CPU mask
4827 * Return: 0 on success. An error code otherwise.
4829 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4830 unsigned long __user *, user_mask_ptr)
4832 cpumask_var_t new_mask;
4835 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4838 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4840 retval = sched_setaffinity(pid, new_mask);
4841 free_cpumask_var(new_mask);
4845 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4847 struct task_struct *p;
4848 unsigned long flags;
4854 p = find_process_by_pid(pid);
4858 retval = security_task_getscheduler(p);
4862 raw_spin_lock_irqsave(&p->pi_lock, flags);
4863 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4864 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4873 * sys_sched_getaffinity - get the CPU affinity of a process
4874 * @pid: pid of the process
4875 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4876 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4878 * Return: size of CPU mask copied to user_mask_ptr on success. An
4879 * error code otherwise.
4881 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4882 unsigned long __user *, user_mask_ptr)
4887 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4889 if (len & (sizeof(unsigned long)-1))
4892 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4895 ret = sched_getaffinity(pid, mask);
4897 size_t retlen = min_t(size_t, len, cpumask_size());
4899 if (copy_to_user(user_mask_ptr, mask, retlen))
4904 free_cpumask_var(mask);
4910 * sys_sched_yield - yield the current processor to other threads.
4912 * This function yields the current CPU to other tasks. If there are no
4913 * other threads running on this CPU then this function will return.
4917 SYSCALL_DEFINE0(sched_yield)
4922 local_irq_disable();
4926 schedstat_inc(rq->yld_count);
4927 current->sched_class->yield_task(rq);
4930 * Since we are going to call schedule() anyway, there's
4931 * no need to preempt or enable interrupts:
4935 sched_preempt_enable_no_resched();
4942 #ifndef CONFIG_PREEMPT
4943 int __sched _cond_resched(void)
4945 if (should_resched(0)) {
4946 preempt_schedule_common();
4951 EXPORT_SYMBOL(_cond_resched);
4955 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4956 * call schedule, and on return reacquire the lock.
4958 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4959 * operations here to prevent schedule() from being called twice (once via
4960 * spin_unlock(), once by hand).
4962 int __cond_resched_lock(spinlock_t *lock)
4964 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4967 lockdep_assert_held(lock);
4969 if (spin_needbreak(lock) || resched) {
4972 preempt_schedule_common();
4980 EXPORT_SYMBOL(__cond_resched_lock);
4982 int __sched __cond_resched_softirq(void)
4984 BUG_ON(!in_softirq());
4986 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4988 preempt_schedule_common();
4994 EXPORT_SYMBOL(__cond_resched_softirq);
4997 * yield - yield the current processor to other threads.
4999 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5001 * The scheduler is at all times free to pick the calling task as the most
5002 * eligible task to run, if removing the yield() call from your code breaks
5003 * it, its already broken.
5005 * Typical broken usage is:
5010 * where one assumes that yield() will let 'the other' process run that will
5011 * make event true. If the current task is a SCHED_FIFO task that will never
5012 * happen. Never use yield() as a progress guarantee!!
5014 * If you want to use yield() to wait for something, use wait_event().
5015 * If you want to use yield() to be 'nice' for others, use cond_resched().
5016 * If you still want to use yield(), do not!
5018 void __sched yield(void)
5020 set_current_state(TASK_RUNNING);
5023 EXPORT_SYMBOL(yield);
5026 * yield_to - yield the current processor to another thread in
5027 * your thread group, or accelerate that thread toward the
5028 * processor it's on.
5030 * @preempt: whether task preemption is allowed or not
5032 * It's the caller's job to ensure that the target task struct
5033 * can't go away on us before we can do any checks.
5036 * true (>0) if we indeed boosted the target task.
5037 * false (0) if we failed to boost the target.
5038 * -ESRCH if there's no task to yield to.
5040 int __sched yield_to(struct task_struct *p, bool preempt)
5042 struct task_struct *curr = current;
5043 struct rq *rq, *p_rq;
5044 unsigned long flags;
5047 local_irq_save(flags);
5053 * If we're the only runnable task on the rq and target rq also
5054 * has only one task, there's absolutely no point in yielding.
5056 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5061 double_rq_lock(rq, p_rq);
5062 if (task_rq(p) != p_rq) {
5063 double_rq_unlock(rq, p_rq);
5067 if (!curr->sched_class->yield_to_task)
5070 if (curr->sched_class != p->sched_class)
5073 if (task_running(p_rq, p) || p->state)
5076 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5078 schedstat_inc(rq->yld_count);
5080 * Make p's CPU reschedule; pick_next_entity takes care of
5083 if (preempt && rq != p_rq)
5088 double_rq_unlock(rq, p_rq);
5090 local_irq_restore(flags);
5097 EXPORT_SYMBOL_GPL(yield_to);
5099 int io_schedule_prepare(void)
5101 int old_iowait = current->in_iowait;
5103 current->in_iowait = 1;
5104 blk_schedule_flush_plug(current);
5109 void io_schedule_finish(int token)
5111 current->in_iowait = token;
5115 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5116 * that process accounting knows that this is a task in IO wait state.
5118 long __sched io_schedule_timeout(long timeout)
5123 token = io_schedule_prepare();
5124 ret = schedule_timeout(timeout);
5125 io_schedule_finish(token);
5129 EXPORT_SYMBOL(io_schedule_timeout);
5131 void io_schedule(void)
5135 token = io_schedule_prepare();
5137 io_schedule_finish(token);
5139 EXPORT_SYMBOL(io_schedule);
5142 * sys_sched_get_priority_max - return maximum RT priority.
5143 * @policy: scheduling class.
5145 * Return: On success, this syscall returns the maximum
5146 * rt_priority that can be used by a given scheduling class.
5147 * On failure, a negative error code is returned.
5149 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5156 ret = MAX_USER_RT_PRIO-1;
5158 case SCHED_DEADLINE:
5169 * sys_sched_get_priority_min - return minimum RT priority.
5170 * @policy: scheduling class.
5172 * Return: On success, this syscall returns the minimum
5173 * rt_priority that can be used by a given scheduling class.
5174 * On failure, a negative error code is returned.
5176 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5185 case SCHED_DEADLINE:
5195 * sys_sched_rr_get_interval - return the default timeslice of a process.
5196 * @pid: pid of the process.
5197 * @interval: userspace pointer to the timeslice value.
5199 * this syscall writes the default timeslice value of a given process
5200 * into the user-space timespec buffer. A value of '0' means infinity.
5202 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5205 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5206 struct timespec __user *, interval)
5208 struct task_struct *p;
5209 unsigned int time_slice;
5220 p = find_process_by_pid(pid);
5224 retval = security_task_getscheduler(p);
5228 rq = task_rq_lock(p, &rf);
5230 if (p->sched_class->get_rr_interval)
5231 time_slice = p->sched_class->get_rr_interval(rq, p);
5232 task_rq_unlock(rq, p, &rf);
5235 jiffies_to_timespec(time_slice, &t);
5236 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5244 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5246 void sched_show_task(struct task_struct *p)
5248 unsigned long free = 0;
5250 unsigned long state = p->state;
5252 /* Make sure the string lines up properly with the number of task states: */
5253 BUILD_BUG_ON(sizeof(TASK_STATE_TO_CHAR_STR)-1 != ilog2(TASK_STATE_MAX)+1);
5255 if (!try_get_task_stack(p))
5258 state = __ffs(state) + 1;
5259 printk(KERN_INFO "%-15.15s %c", p->comm,
5260 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5261 if (state == TASK_RUNNING)
5262 printk(KERN_CONT " running task ");
5263 #ifdef CONFIG_DEBUG_STACK_USAGE
5264 free = stack_not_used(p);
5269 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5271 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5272 task_pid_nr(p), ppid,
5273 (unsigned long)task_thread_info(p)->flags);
5275 print_worker_info(KERN_INFO, p);
5276 show_stack(p, NULL);
5280 void show_state_filter(unsigned long state_filter)
5282 struct task_struct *g, *p;
5284 #if BITS_PER_LONG == 32
5286 " task PC stack pid father\n");
5289 " task PC stack pid father\n");
5292 for_each_process_thread(g, p) {
5294 * reset the NMI-timeout, listing all files on a slow
5295 * console might take a lot of time:
5296 * Also, reset softlockup watchdogs on all CPUs, because
5297 * another CPU might be blocked waiting for us to process
5300 touch_nmi_watchdog();
5301 touch_all_softlockup_watchdogs();
5302 if (!state_filter || (p->state & state_filter))
5306 #ifdef CONFIG_SCHED_DEBUG
5308 sysrq_sched_debug_show();
5312 * Only show locks if all tasks are dumped:
5315 debug_show_all_locks();
5318 void init_idle_bootup_task(struct task_struct *idle)
5320 idle->sched_class = &idle_sched_class;
5324 * init_idle - set up an idle thread for a given CPU
5325 * @idle: task in question
5326 * @cpu: CPU the idle task belongs to
5328 * NOTE: this function does not set the idle thread's NEED_RESCHED
5329 * flag, to make booting more robust.
5331 void init_idle(struct task_struct *idle, int cpu)
5333 struct rq *rq = cpu_rq(cpu);
5334 unsigned long flags;
5336 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5337 raw_spin_lock(&rq->lock);
5339 __sched_fork(0, idle);
5340 idle->state = TASK_RUNNING;
5341 idle->se.exec_start = sched_clock();
5342 idle->flags |= PF_IDLE;
5344 kasan_unpoison_task_stack(idle);
5348 * Its possible that init_idle() gets called multiple times on a task,
5349 * in that case do_set_cpus_allowed() will not do the right thing.
5351 * And since this is boot we can forgo the serialization.
5353 set_cpus_allowed_common(idle, cpumask_of(cpu));
5356 * We're having a chicken and egg problem, even though we are
5357 * holding rq->lock, the CPU isn't yet set to this CPU so the
5358 * lockdep check in task_group() will fail.
5360 * Similar case to sched_fork(). / Alternatively we could
5361 * use task_rq_lock() here and obtain the other rq->lock.
5366 __set_task_cpu(idle, cpu);
5369 rq->curr = rq->idle = idle;
5370 idle->on_rq = TASK_ON_RQ_QUEUED;
5374 raw_spin_unlock(&rq->lock);
5375 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5377 /* Set the preempt count _outside_ the spinlocks! */
5378 init_idle_preempt_count(idle, cpu);
5381 * The idle tasks have their own, simple scheduling class:
5383 idle->sched_class = &idle_sched_class;
5384 ftrace_graph_init_idle_task(idle, cpu);
5385 vtime_init_idle(idle, cpu);
5387 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5391 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5392 const struct cpumask *trial)
5394 int ret = 1, trial_cpus;
5395 struct dl_bw *cur_dl_b;
5396 unsigned long flags;
5398 if (!cpumask_weight(cur))
5401 rcu_read_lock_sched();
5402 cur_dl_b = dl_bw_of(cpumask_any(cur));
5403 trial_cpus = cpumask_weight(trial);
5405 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5406 if (cur_dl_b->bw != -1 &&
5407 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5409 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5410 rcu_read_unlock_sched();
5415 int task_can_attach(struct task_struct *p,
5416 const struct cpumask *cs_cpus_allowed)
5421 * Kthreads which disallow setaffinity shouldn't be moved
5422 * to a new cpuset; we don't want to change their CPU
5423 * affinity and isolating such threads by their set of
5424 * allowed nodes is unnecessary. Thus, cpusets are not
5425 * applicable for such threads. This prevents checking for
5426 * success of set_cpus_allowed_ptr() on all attached tasks
5427 * before cpus_allowed may be changed.
5429 if (p->flags & PF_NO_SETAFFINITY) {
5435 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5437 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5442 unsigned long flags;
5444 rcu_read_lock_sched();
5445 dl_b = dl_bw_of(dest_cpu);
5446 raw_spin_lock_irqsave(&dl_b->lock, flags);
5447 cpus = dl_bw_cpus(dest_cpu);
5448 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5453 * We reserve space for this task in the destination
5454 * root_domain, as we can't fail after this point.
5455 * We will free resources in the source root_domain
5456 * later on (see set_cpus_allowed_dl()).
5458 __dl_add(dl_b, p->dl.dl_bw);
5460 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5461 rcu_read_unlock_sched();
5471 bool sched_smp_initialized __read_mostly;
5473 #ifdef CONFIG_NUMA_BALANCING
5474 /* Migrate current task p to target_cpu */
5475 int migrate_task_to(struct task_struct *p, int target_cpu)
5477 struct migration_arg arg = { p, target_cpu };
5478 int curr_cpu = task_cpu(p);
5480 if (curr_cpu == target_cpu)
5483 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5486 /* TODO: This is not properly updating schedstats */
5488 trace_sched_move_numa(p, curr_cpu, target_cpu);
5489 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5493 * Requeue a task on a given node and accurately track the number of NUMA
5494 * tasks on the runqueues
5496 void sched_setnuma(struct task_struct *p, int nid)
5498 bool queued, running;
5502 rq = task_rq_lock(p, &rf);
5503 queued = task_on_rq_queued(p);
5504 running = task_current(rq, p);
5507 dequeue_task(rq, p, DEQUEUE_SAVE);
5509 put_prev_task(rq, p);
5511 p->numa_preferred_nid = nid;
5514 enqueue_task(rq, p, ENQUEUE_RESTORE);
5516 set_curr_task(rq, p);
5517 task_rq_unlock(rq, p, &rf);
5519 #endif /* CONFIG_NUMA_BALANCING */
5521 #ifdef CONFIG_HOTPLUG_CPU
5523 * Ensure that the idle task is using init_mm right before its CPU goes
5526 void idle_task_exit(void)
5528 struct mm_struct *mm = current->active_mm;
5530 BUG_ON(cpu_online(smp_processor_id()));
5532 if (mm != &init_mm) {
5533 switch_mm_irqs_off(mm, &init_mm, current);
5534 finish_arch_post_lock_switch();
5540 * Since this CPU is going 'away' for a while, fold any nr_active delta
5541 * we might have. Assumes we're called after migrate_tasks() so that the
5542 * nr_active count is stable. We need to take the teardown thread which
5543 * is calling this into account, so we hand in adjust = 1 to the load
5546 * Also see the comment "Global load-average calculations".
5548 static void calc_load_migrate(struct rq *rq)
5550 long delta = calc_load_fold_active(rq, 1);
5552 atomic_long_add(delta, &calc_load_tasks);
5555 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5559 static const struct sched_class fake_sched_class = {
5560 .put_prev_task = put_prev_task_fake,
5563 static struct task_struct fake_task = {
5565 * Avoid pull_{rt,dl}_task()
5567 .prio = MAX_PRIO + 1,
5568 .sched_class = &fake_sched_class,
5572 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5573 * try_to_wake_up()->select_task_rq().
5575 * Called with rq->lock held even though we'er in stop_machine() and
5576 * there's no concurrency possible, we hold the required locks anyway
5577 * because of lock validation efforts.
5579 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5581 struct rq *rq = dead_rq;
5582 struct task_struct *next, *stop = rq->stop;
5583 struct rq_flags orf = *rf;
5587 * Fudge the rq selection such that the below task selection loop
5588 * doesn't get stuck on the currently eligible stop task.
5590 * We're currently inside stop_machine() and the rq is either stuck
5591 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5592 * either way we should never end up calling schedule() until we're
5598 * put_prev_task() and pick_next_task() sched
5599 * class method both need to have an up-to-date
5600 * value of rq->clock[_task]
5602 update_rq_clock(rq);
5606 * There's this thread running, bail when that's the only
5609 if (rq->nr_running == 1)
5613 * pick_next_task() assumes pinned rq->lock:
5615 next = pick_next_task(rq, &fake_task, rf);
5617 next->sched_class->put_prev_task(rq, next);
5620 * Rules for changing task_struct::cpus_allowed are holding
5621 * both pi_lock and rq->lock, such that holding either
5622 * stabilizes the mask.
5624 * Drop rq->lock is not quite as disastrous as it usually is
5625 * because !cpu_active at this point, which means load-balance
5626 * will not interfere. Also, stop-machine.
5629 raw_spin_lock(&next->pi_lock);
5633 * Since we're inside stop-machine, _nothing_ should have
5634 * changed the task, WARN if weird stuff happened, because in
5635 * that case the above rq->lock drop is a fail too.
5637 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5638 raw_spin_unlock(&next->pi_lock);
5642 /* Find suitable destination for @next, with force if needed. */
5643 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5644 rq = __migrate_task(rq, rf, next, dest_cpu);
5645 if (rq != dead_rq) {
5651 raw_spin_unlock(&next->pi_lock);
5656 #endif /* CONFIG_HOTPLUG_CPU */
5658 void set_rq_online(struct rq *rq)
5661 const struct sched_class *class;
5663 cpumask_set_cpu(rq->cpu, rq->rd->online);
5666 for_each_class(class) {
5667 if (class->rq_online)
5668 class->rq_online(rq);
5673 void set_rq_offline(struct rq *rq)
5676 const struct sched_class *class;
5678 for_each_class(class) {
5679 if (class->rq_offline)
5680 class->rq_offline(rq);
5683 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5688 static void set_cpu_rq_start_time(unsigned int cpu)
5690 struct rq *rq = cpu_rq(cpu);
5692 rq->age_stamp = sched_clock_cpu(cpu);
5696 * used to mark begin/end of suspend/resume:
5698 static int num_cpus_frozen;
5701 * Update cpusets according to cpu_active mask. If cpusets are
5702 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5703 * around partition_sched_domains().
5705 * If we come here as part of a suspend/resume, don't touch cpusets because we
5706 * want to restore it back to its original state upon resume anyway.
5708 static void cpuset_cpu_active(void)
5710 if (cpuhp_tasks_frozen) {
5712 * num_cpus_frozen tracks how many CPUs are involved in suspend
5713 * resume sequence. As long as this is not the last online
5714 * operation in the resume sequence, just build a single sched
5715 * domain, ignoring cpusets.
5718 if (likely(num_cpus_frozen)) {
5719 partition_sched_domains(1, NULL, NULL);
5723 * This is the last CPU online operation. So fall through and
5724 * restore the original sched domains by considering the
5725 * cpuset configurations.
5728 cpuset_update_active_cpus(true);
5731 static int cpuset_cpu_inactive(unsigned int cpu)
5733 unsigned long flags;
5738 if (!cpuhp_tasks_frozen) {
5739 rcu_read_lock_sched();
5740 dl_b = dl_bw_of(cpu);
5742 raw_spin_lock_irqsave(&dl_b->lock, flags);
5743 cpus = dl_bw_cpus(cpu);
5744 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5745 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5747 rcu_read_unlock_sched();
5751 cpuset_update_active_cpus(false);
5754 partition_sched_domains(1, NULL, NULL);
5759 int sched_cpu_activate(unsigned int cpu)
5761 struct rq *rq = cpu_rq(cpu);
5764 set_cpu_active(cpu, true);
5766 if (sched_smp_initialized) {
5767 sched_domains_numa_masks_set(cpu);
5768 cpuset_cpu_active();
5772 * Put the rq online, if not already. This happens:
5774 * 1) In the early boot process, because we build the real domains
5775 * after all CPUs have been brought up.
5777 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5780 rq_lock_irqsave(rq, &rf);
5782 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5785 rq_unlock_irqrestore(rq, &rf);
5787 update_max_interval();
5792 int sched_cpu_deactivate(unsigned int cpu)
5796 set_cpu_active(cpu, false);
5798 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5799 * users of this state to go away such that all new such users will
5802 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5803 * not imply sync_sched(), so wait for both.
5805 * Do sync before park smpboot threads to take care the rcu boost case.
5807 if (IS_ENABLED(CONFIG_PREEMPT))
5808 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5812 if (!sched_smp_initialized)
5815 ret = cpuset_cpu_inactive(cpu);
5817 set_cpu_active(cpu, true);
5820 sched_domains_numa_masks_clear(cpu);
5824 static void sched_rq_cpu_starting(unsigned int cpu)
5826 struct rq *rq = cpu_rq(cpu);
5828 rq->calc_load_update = calc_load_update;
5829 update_max_interval();
5832 int sched_cpu_starting(unsigned int cpu)
5834 set_cpu_rq_start_time(cpu);
5835 sched_rq_cpu_starting(cpu);
5839 #ifdef CONFIG_HOTPLUG_CPU
5840 int sched_cpu_dying(unsigned int cpu)
5842 struct rq *rq = cpu_rq(cpu);
5845 /* Handle pending wakeups and then migrate everything off */
5846 sched_ttwu_pending();
5848 rq_lock_irqsave(rq, &rf);
5850 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5853 migrate_tasks(rq, &rf);
5854 BUG_ON(rq->nr_running != 1);
5855 rq_unlock_irqrestore(rq, &rf);
5857 calc_load_migrate(rq);
5858 update_max_interval();
5859 nohz_balance_exit_idle(cpu);
5865 #ifdef CONFIG_SCHED_SMT
5866 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5868 static void sched_init_smt(void)
5871 * We've enumerated all CPUs and will assume that if any CPU
5872 * has SMT siblings, CPU0 will too.
5874 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5875 static_branch_enable(&sched_smt_present);
5878 static inline void sched_init_smt(void) { }
5881 void __init sched_init_smp(void)
5883 cpumask_var_t non_isolated_cpus;
5885 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
5886 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
5891 * There's no userspace yet to cause hotplug operations; hence all the
5892 * CPU masks are stable and all blatant races in the below code cannot
5895 mutex_lock(&sched_domains_mutex);
5896 init_sched_domains(cpu_active_mask);
5897 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
5898 if (cpumask_empty(non_isolated_cpus))
5899 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
5900 mutex_unlock(&sched_domains_mutex);
5902 /* Move init over to a non-isolated CPU */
5903 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
5905 sched_init_granularity();
5906 free_cpumask_var(non_isolated_cpus);
5908 init_sched_rt_class();
5909 init_sched_dl_class();
5912 sched_clock_init_late();
5914 sched_smp_initialized = true;
5917 static int __init migration_init(void)
5919 sched_rq_cpu_starting(smp_processor_id());
5922 early_initcall(migration_init);
5925 void __init sched_init_smp(void)
5927 sched_init_granularity();
5928 sched_clock_init_late();
5930 #endif /* CONFIG_SMP */
5932 int in_sched_functions(unsigned long addr)
5934 return in_lock_functions(addr) ||
5935 (addr >= (unsigned long)__sched_text_start
5936 && addr < (unsigned long)__sched_text_end);
5939 #ifdef CONFIG_CGROUP_SCHED
5941 * Default task group.
5942 * Every task in system belongs to this group at bootup.
5944 struct task_group root_task_group;
5945 LIST_HEAD(task_groups);
5947 /* Cacheline aligned slab cache for task_group */
5948 static struct kmem_cache *task_group_cache __read_mostly;
5951 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5952 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5954 #define WAIT_TABLE_BITS 8
5955 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
5956 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
5958 wait_queue_head_t *bit_waitqueue(void *word, int bit)
5960 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
5961 unsigned long val = (unsigned long)word << shift | bit;
5963 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
5965 EXPORT_SYMBOL(bit_waitqueue);
5967 void __init sched_init(void)
5970 unsigned long alloc_size = 0, ptr;
5974 for (i = 0; i < WAIT_TABLE_SIZE; i++)
5975 init_waitqueue_head(bit_wait_table + i);
5977 #ifdef CONFIG_FAIR_GROUP_SCHED
5978 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5980 #ifdef CONFIG_RT_GROUP_SCHED
5981 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5984 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5986 #ifdef CONFIG_FAIR_GROUP_SCHED
5987 root_task_group.se = (struct sched_entity **)ptr;
5988 ptr += nr_cpu_ids * sizeof(void **);
5990 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5991 ptr += nr_cpu_ids * sizeof(void **);
5993 #endif /* CONFIG_FAIR_GROUP_SCHED */
5994 #ifdef CONFIG_RT_GROUP_SCHED
5995 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5996 ptr += nr_cpu_ids * sizeof(void **);
5998 root_task_group.rt_rq = (struct rt_rq **)ptr;
5999 ptr += nr_cpu_ids * sizeof(void **);
6001 #endif /* CONFIG_RT_GROUP_SCHED */
6003 #ifdef CONFIG_CPUMASK_OFFSTACK
6004 for_each_possible_cpu(i) {
6005 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6006 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6007 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6008 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6010 #endif /* CONFIG_CPUMASK_OFFSTACK */
6012 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6013 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6016 init_defrootdomain();
6019 #ifdef CONFIG_RT_GROUP_SCHED
6020 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6021 global_rt_period(), global_rt_runtime());
6022 #endif /* CONFIG_RT_GROUP_SCHED */
6024 #ifdef CONFIG_CGROUP_SCHED
6025 task_group_cache = KMEM_CACHE(task_group, 0);
6027 list_add(&root_task_group.list, &task_groups);
6028 INIT_LIST_HEAD(&root_task_group.children);
6029 INIT_LIST_HEAD(&root_task_group.siblings);
6030 autogroup_init(&init_task);
6031 #endif /* CONFIG_CGROUP_SCHED */
6033 for_each_possible_cpu(i) {
6037 raw_spin_lock_init(&rq->lock);
6039 rq->calc_load_active = 0;
6040 rq->calc_load_update = jiffies + LOAD_FREQ;
6041 init_cfs_rq(&rq->cfs);
6042 init_rt_rq(&rq->rt);
6043 init_dl_rq(&rq->dl);
6044 #ifdef CONFIG_FAIR_GROUP_SCHED
6045 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6046 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6047 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6049 * How much CPU bandwidth does root_task_group get?
6051 * In case of task-groups formed thr' the cgroup filesystem, it
6052 * gets 100% of the CPU resources in the system. This overall
6053 * system CPU resource is divided among the tasks of
6054 * root_task_group and its child task-groups in a fair manner,
6055 * based on each entity's (task or task-group's) weight
6056 * (se->load.weight).
6058 * In other words, if root_task_group has 10 tasks of weight
6059 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6060 * then A0's share of the CPU resource is:
6062 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6064 * We achieve this by letting root_task_group's tasks sit
6065 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6067 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6068 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6069 #endif /* CONFIG_FAIR_GROUP_SCHED */
6071 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6072 #ifdef CONFIG_RT_GROUP_SCHED
6073 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6076 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6077 rq->cpu_load[j] = 0;
6082 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6083 rq->balance_callback = NULL;
6084 rq->active_balance = 0;
6085 rq->next_balance = jiffies;
6090 rq->avg_idle = 2*sysctl_sched_migration_cost;
6091 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6093 INIT_LIST_HEAD(&rq->cfs_tasks);
6095 rq_attach_root(rq, &def_root_domain);
6096 #ifdef CONFIG_NO_HZ_COMMON
6097 rq->last_load_update_tick = jiffies;
6100 #ifdef CONFIG_NO_HZ_FULL
6101 rq->last_sched_tick = 0;
6103 #endif /* CONFIG_SMP */
6105 atomic_set(&rq->nr_iowait, 0);
6108 set_load_weight(&init_task);
6111 * The boot idle thread does lazy MMU switching as well:
6114 enter_lazy_tlb(&init_mm, current);
6117 * Make us the idle thread. Technically, schedule() should not be
6118 * called from this thread, however somewhere below it might be,
6119 * but because we are the idle thread, we just pick up running again
6120 * when this runqueue becomes "idle".
6122 init_idle(current, smp_processor_id());
6124 calc_load_update = jiffies + LOAD_FREQ;
6127 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6128 /* May be allocated at isolcpus cmdline parse time */
6129 if (cpu_isolated_map == NULL)
6130 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6131 idle_thread_set_boot_cpu();
6132 set_cpu_rq_start_time(smp_processor_id());
6134 init_sched_fair_class();
6138 scheduler_running = 1;
6141 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6142 static inline int preempt_count_equals(int preempt_offset)
6144 int nested = preempt_count() + rcu_preempt_depth();
6146 return (nested == preempt_offset);
6149 void __might_sleep(const char *file, int line, int preempt_offset)
6152 * Blocking primitives will set (and therefore destroy) current->state,
6153 * since we will exit with TASK_RUNNING make sure we enter with it,
6154 * otherwise we will destroy state.
6156 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6157 "do not call blocking ops when !TASK_RUNNING; "
6158 "state=%lx set at [<%p>] %pS\n",
6160 (void *)current->task_state_change,
6161 (void *)current->task_state_change);
6163 ___might_sleep(file, line, preempt_offset);
6165 EXPORT_SYMBOL(__might_sleep);
6167 void ___might_sleep(const char *file, int line, int preempt_offset)
6169 /* Ratelimiting timestamp: */
6170 static unsigned long prev_jiffy;
6172 unsigned long preempt_disable_ip;
6174 /* WARN_ON_ONCE() by default, no rate limit required: */
6177 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6178 !is_idle_task(current)) ||
6179 system_state != SYSTEM_RUNNING || oops_in_progress)
6181 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6183 prev_jiffy = jiffies;
6185 /* Save this before calling printk(), since that will clobber it: */
6186 preempt_disable_ip = get_preempt_disable_ip(current);
6189 "BUG: sleeping function called from invalid context at %s:%d\n",
6192 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6193 in_atomic(), irqs_disabled(),
6194 current->pid, current->comm);
6196 if (task_stack_end_corrupted(current))
6197 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6199 debug_show_held_locks(current);
6200 if (irqs_disabled())
6201 print_irqtrace_events(current);
6202 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6203 && !preempt_count_equals(preempt_offset)) {
6204 pr_err("Preemption disabled at:");
6205 print_ip_sym(preempt_disable_ip);
6209 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6211 EXPORT_SYMBOL(___might_sleep);
6214 #ifdef CONFIG_MAGIC_SYSRQ
6215 void normalize_rt_tasks(void)
6217 struct task_struct *g, *p;
6218 struct sched_attr attr = {
6219 .sched_policy = SCHED_NORMAL,
6222 read_lock(&tasklist_lock);
6223 for_each_process_thread(g, p) {
6225 * Only normalize user tasks:
6227 if (p->flags & PF_KTHREAD)
6230 p->se.exec_start = 0;
6231 schedstat_set(p->se.statistics.wait_start, 0);
6232 schedstat_set(p->se.statistics.sleep_start, 0);
6233 schedstat_set(p->se.statistics.block_start, 0);
6235 if (!dl_task(p) && !rt_task(p)) {
6237 * Renice negative nice level userspace
6240 if (task_nice(p) < 0)
6241 set_user_nice(p, 0);
6245 __sched_setscheduler(p, &attr, false, false);
6247 read_unlock(&tasklist_lock);
6250 #endif /* CONFIG_MAGIC_SYSRQ */
6252 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6254 * These functions are only useful for the IA64 MCA handling, or kdb.
6256 * They can only be called when the whole system has been
6257 * stopped - every CPU needs to be quiescent, and no scheduling
6258 * activity can take place. Using them for anything else would
6259 * be a serious bug, and as a result, they aren't even visible
6260 * under any other configuration.
6264 * curr_task - return the current task for a given CPU.
6265 * @cpu: the processor in question.
6267 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6269 * Return: The current task for @cpu.
6271 struct task_struct *curr_task(int cpu)
6273 return cpu_curr(cpu);
6276 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6280 * set_curr_task - set the current task for a given CPU.
6281 * @cpu: the processor in question.
6282 * @p: the task pointer to set.
6284 * Description: This function must only be used when non-maskable interrupts
6285 * are serviced on a separate stack. It allows the architecture to switch the
6286 * notion of the current task on a CPU in a non-blocking manner. This function
6287 * must be called with all CPU's synchronized, and interrupts disabled, the
6288 * and caller must save the original value of the current task (see
6289 * curr_task() above) and restore that value before reenabling interrupts and
6290 * re-starting the system.
6292 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6294 void ia64_set_curr_task(int cpu, struct task_struct *p)
6301 #ifdef CONFIG_CGROUP_SCHED
6302 /* task_group_lock serializes the addition/removal of task groups */
6303 static DEFINE_SPINLOCK(task_group_lock);
6305 static void sched_free_group(struct task_group *tg)
6307 free_fair_sched_group(tg);
6308 free_rt_sched_group(tg);
6310 kmem_cache_free(task_group_cache, tg);
6313 /* allocate runqueue etc for a new task group */
6314 struct task_group *sched_create_group(struct task_group *parent)
6316 struct task_group *tg;
6318 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6320 return ERR_PTR(-ENOMEM);
6322 if (!alloc_fair_sched_group(tg, parent))
6325 if (!alloc_rt_sched_group(tg, parent))
6331 sched_free_group(tg);
6332 return ERR_PTR(-ENOMEM);
6335 void sched_online_group(struct task_group *tg, struct task_group *parent)
6337 unsigned long flags;
6339 spin_lock_irqsave(&task_group_lock, flags);
6340 list_add_rcu(&tg->list, &task_groups);
6342 /* Root should already exist: */
6345 tg->parent = parent;
6346 INIT_LIST_HEAD(&tg->children);
6347 list_add_rcu(&tg->siblings, &parent->children);
6348 spin_unlock_irqrestore(&task_group_lock, flags);
6350 online_fair_sched_group(tg);
6353 /* rcu callback to free various structures associated with a task group */
6354 static void sched_free_group_rcu(struct rcu_head *rhp)
6356 /* Now it should be safe to free those cfs_rqs: */
6357 sched_free_group(container_of(rhp, struct task_group, rcu));
6360 void sched_destroy_group(struct task_group *tg)
6362 /* Wait for possible concurrent references to cfs_rqs complete: */
6363 call_rcu(&tg->rcu, sched_free_group_rcu);
6366 void sched_offline_group(struct task_group *tg)
6368 unsigned long flags;
6370 /* End participation in shares distribution: */
6371 unregister_fair_sched_group(tg);
6373 spin_lock_irqsave(&task_group_lock, flags);
6374 list_del_rcu(&tg->list);
6375 list_del_rcu(&tg->siblings);
6376 spin_unlock_irqrestore(&task_group_lock, flags);
6379 static void sched_change_group(struct task_struct *tsk, int type)
6381 struct task_group *tg;
6384 * All callers are synchronized by task_rq_lock(); we do not use RCU
6385 * which is pointless here. Thus, we pass "true" to task_css_check()
6386 * to prevent lockdep warnings.
6388 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6389 struct task_group, css);
6390 tg = autogroup_task_group(tsk, tg);
6391 tsk->sched_task_group = tg;
6393 #ifdef CONFIG_FAIR_GROUP_SCHED
6394 if (tsk->sched_class->task_change_group)
6395 tsk->sched_class->task_change_group(tsk, type);
6398 set_task_rq(tsk, task_cpu(tsk));
6402 * Change task's runqueue when it moves between groups.
6404 * The caller of this function should have put the task in its new group by
6405 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6408 void sched_move_task(struct task_struct *tsk)
6410 int queued, running;
6414 rq = task_rq_lock(tsk, &rf);
6415 update_rq_clock(rq);
6417 running = task_current(rq, tsk);
6418 queued = task_on_rq_queued(tsk);
6421 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
6423 put_prev_task(rq, tsk);
6425 sched_change_group(tsk, TASK_MOVE_GROUP);
6428 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
6430 set_curr_task(rq, tsk);
6432 task_rq_unlock(rq, tsk, &rf);
6434 #endif /* CONFIG_CGROUP_SCHED */
6436 #ifdef CONFIG_RT_GROUP_SCHED
6438 * Ensure that the real time constraints are schedulable.
6440 static DEFINE_MUTEX(rt_constraints_mutex);
6442 /* Must be called with tasklist_lock held */
6443 static inline int tg_has_rt_tasks(struct task_group *tg)
6445 struct task_struct *g, *p;
6448 * Autogroups do not have RT tasks; see autogroup_create().
6450 if (task_group_is_autogroup(tg))
6453 for_each_process_thread(g, p) {
6454 if (rt_task(p) && task_group(p) == tg)
6461 struct rt_schedulable_data {
6462 struct task_group *tg;
6467 static int tg_rt_schedulable(struct task_group *tg, void *data)
6469 struct rt_schedulable_data *d = data;
6470 struct task_group *child;
6471 unsigned long total, sum = 0;
6472 u64 period, runtime;
6474 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6475 runtime = tg->rt_bandwidth.rt_runtime;
6478 period = d->rt_period;
6479 runtime = d->rt_runtime;
6483 * Cannot have more runtime than the period.
6485 if (runtime > period && runtime != RUNTIME_INF)
6489 * Ensure we don't starve existing RT tasks.
6491 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6494 total = to_ratio(period, runtime);
6497 * Nobody can have more than the global setting allows.
6499 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6503 * The sum of our children's runtime should not exceed our own.
6505 list_for_each_entry_rcu(child, &tg->children, siblings) {
6506 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6507 runtime = child->rt_bandwidth.rt_runtime;
6509 if (child == d->tg) {
6510 period = d->rt_period;
6511 runtime = d->rt_runtime;
6514 sum += to_ratio(period, runtime);
6523 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6527 struct rt_schedulable_data data = {
6529 .rt_period = period,
6530 .rt_runtime = runtime,
6534 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6540 static int tg_set_rt_bandwidth(struct task_group *tg,
6541 u64 rt_period, u64 rt_runtime)
6546 * Disallowing the root group RT runtime is BAD, it would disallow the
6547 * kernel creating (and or operating) RT threads.
6549 if (tg == &root_task_group && rt_runtime == 0)
6552 /* No period doesn't make any sense. */
6556 mutex_lock(&rt_constraints_mutex);
6557 read_lock(&tasklist_lock);
6558 err = __rt_schedulable(tg, rt_period, rt_runtime);
6562 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6563 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6564 tg->rt_bandwidth.rt_runtime = rt_runtime;
6566 for_each_possible_cpu(i) {
6567 struct rt_rq *rt_rq = tg->rt_rq[i];
6569 raw_spin_lock(&rt_rq->rt_runtime_lock);
6570 rt_rq->rt_runtime = rt_runtime;
6571 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6573 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6575 read_unlock(&tasklist_lock);
6576 mutex_unlock(&rt_constraints_mutex);
6581 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6583 u64 rt_runtime, rt_period;
6585 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6586 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6587 if (rt_runtime_us < 0)
6588 rt_runtime = RUNTIME_INF;
6590 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6593 static long sched_group_rt_runtime(struct task_group *tg)
6597 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6600 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6601 do_div(rt_runtime_us, NSEC_PER_USEC);
6602 return rt_runtime_us;
6605 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
6607 u64 rt_runtime, rt_period;
6609 rt_period = rt_period_us * NSEC_PER_USEC;
6610 rt_runtime = tg->rt_bandwidth.rt_runtime;
6612 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6615 static long sched_group_rt_period(struct task_group *tg)
6619 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6620 do_div(rt_period_us, NSEC_PER_USEC);
6621 return rt_period_us;
6623 #endif /* CONFIG_RT_GROUP_SCHED */
6625 #ifdef CONFIG_RT_GROUP_SCHED
6626 static int sched_rt_global_constraints(void)
6630 mutex_lock(&rt_constraints_mutex);
6631 read_lock(&tasklist_lock);
6632 ret = __rt_schedulable(NULL, 0, 0);
6633 read_unlock(&tasklist_lock);
6634 mutex_unlock(&rt_constraints_mutex);
6639 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6641 /* Don't accept realtime tasks when there is no way for them to run */
6642 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6648 #else /* !CONFIG_RT_GROUP_SCHED */
6649 static int sched_rt_global_constraints(void)
6651 unsigned long flags;
6654 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6655 for_each_possible_cpu(i) {
6656 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6658 raw_spin_lock(&rt_rq->rt_runtime_lock);
6659 rt_rq->rt_runtime = global_rt_runtime();
6660 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6662 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6666 #endif /* CONFIG_RT_GROUP_SCHED */
6668 static int sched_dl_global_validate(void)
6670 u64 runtime = global_rt_runtime();
6671 u64 period = global_rt_period();
6672 u64 new_bw = to_ratio(period, runtime);
6675 unsigned long flags;
6678 * Here we want to check the bandwidth not being set to some
6679 * value smaller than the currently allocated bandwidth in
6680 * any of the root_domains.
6682 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6683 * cycling on root_domains... Discussion on different/better
6684 * solutions is welcome!
6686 for_each_possible_cpu(cpu) {
6687 rcu_read_lock_sched();
6688 dl_b = dl_bw_of(cpu);
6690 raw_spin_lock_irqsave(&dl_b->lock, flags);
6691 if (new_bw < dl_b->total_bw)
6693 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6695 rcu_read_unlock_sched();
6704 static void sched_dl_do_global(void)
6709 unsigned long flags;
6711 def_dl_bandwidth.dl_period = global_rt_period();
6712 def_dl_bandwidth.dl_runtime = global_rt_runtime();
6714 if (global_rt_runtime() != RUNTIME_INF)
6715 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
6718 * FIXME: As above...
6720 for_each_possible_cpu(cpu) {
6721 rcu_read_lock_sched();
6722 dl_b = dl_bw_of(cpu);
6724 raw_spin_lock_irqsave(&dl_b->lock, flags);
6726 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6728 rcu_read_unlock_sched();
6732 static int sched_rt_global_validate(void)
6734 if (sysctl_sched_rt_period <= 0)
6737 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
6738 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
6744 static void sched_rt_do_global(void)
6746 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6747 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
6750 int sched_rt_handler(struct ctl_table *table, int write,
6751 void __user *buffer, size_t *lenp,
6754 int old_period, old_runtime;
6755 static DEFINE_MUTEX(mutex);
6759 old_period = sysctl_sched_rt_period;
6760 old_runtime = sysctl_sched_rt_runtime;
6762 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6764 if (!ret && write) {
6765 ret = sched_rt_global_validate();
6769 ret = sched_dl_global_validate();
6773 ret = sched_rt_global_constraints();
6777 sched_rt_do_global();
6778 sched_dl_do_global();
6782 sysctl_sched_rt_period = old_period;
6783 sysctl_sched_rt_runtime = old_runtime;
6785 mutex_unlock(&mutex);
6790 int sched_rr_handler(struct ctl_table *table, int write,
6791 void __user *buffer, size_t *lenp,
6795 static DEFINE_MUTEX(mutex);
6798 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6800 * Make sure that internally we keep jiffies.
6801 * Also, writing zero resets the timeslice to default:
6803 if (!ret && write) {
6804 sched_rr_timeslice =
6805 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
6806 msecs_to_jiffies(sysctl_sched_rr_timeslice);
6808 mutex_unlock(&mutex);
6812 #ifdef CONFIG_CGROUP_SCHED
6814 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6816 return css ? container_of(css, struct task_group, css) : NULL;
6819 static struct cgroup_subsys_state *
6820 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6822 struct task_group *parent = css_tg(parent_css);
6823 struct task_group *tg;
6826 /* This is early initialization for the top cgroup */
6827 return &root_task_group.css;
6830 tg = sched_create_group(parent);
6832 return ERR_PTR(-ENOMEM);
6837 /* Expose task group only after completing cgroup initialization */
6838 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6840 struct task_group *tg = css_tg(css);
6841 struct task_group *parent = css_tg(css->parent);
6844 sched_online_group(tg, parent);
6848 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6850 struct task_group *tg = css_tg(css);
6852 sched_offline_group(tg);
6855 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6857 struct task_group *tg = css_tg(css);
6860 * Relies on the RCU grace period between css_released() and this.
6862 sched_free_group(tg);
6866 * This is called before wake_up_new_task(), therefore we really only
6867 * have to set its group bits, all the other stuff does not apply.
6869 static void cpu_cgroup_fork(struct task_struct *task)
6874 rq = task_rq_lock(task, &rf);
6876 update_rq_clock(rq);
6877 sched_change_group(task, TASK_SET_GROUP);
6879 task_rq_unlock(rq, task, &rf);
6882 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6884 struct task_struct *task;
6885 struct cgroup_subsys_state *css;
6888 cgroup_taskset_for_each(task, css, tset) {
6889 #ifdef CONFIG_RT_GROUP_SCHED
6890 if (!sched_rt_can_attach(css_tg(css), task))
6893 /* We don't support RT-tasks being in separate groups */
6894 if (task->sched_class != &fair_sched_class)
6898 * Serialize against wake_up_new_task() such that if its
6899 * running, we're sure to observe its full state.
6901 raw_spin_lock_irq(&task->pi_lock);
6903 * Avoid calling sched_move_task() before wake_up_new_task()
6904 * has happened. This would lead to problems with PELT, due to
6905 * move wanting to detach+attach while we're not attached yet.
6907 if (task->state == TASK_NEW)
6909 raw_spin_unlock_irq(&task->pi_lock);
6917 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6919 struct task_struct *task;
6920 struct cgroup_subsys_state *css;
6922 cgroup_taskset_for_each(task, css, tset)
6923 sched_move_task(task);
6926 #ifdef CONFIG_FAIR_GROUP_SCHED
6927 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6928 struct cftype *cftype, u64 shareval)
6930 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6933 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6936 struct task_group *tg = css_tg(css);
6938 return (u64) scale_load_down(tg->shares);
6941 #ifdef CONFIG_CFS_BANDWIDTH
6942 static DEFINE_MUTEX(cfs_constraints_mutex);
6944 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6945 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6947 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6949 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6951 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6952 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6954 if (tg == &root_task_group)
6958 * Ensure we have at some amount of bandwidth every period. This is
6959 * to prevent reaching a state of large arrears when throttled via
6960 * entity_tick() resulting in prolonged exit starvation.
6962 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6966 * Likewise, bound things on the otherside by preventing insane quota
6967 * periods. This also allows us to normalize in computing quota
6970 if (period > max_cfs_quota_period)
6974 * Prevent race between setting of cfs_rq->runtime_enabled and
6975 * unthrottle_offline_cfs_rqs().
6978 mutex_lock(&cfs_constraints_mutex);
6979 ret = __cfs_schedulable(tg, period, quota);
6983 runtime_enabled = quota != RUNTIME_INF;
6984 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6986 * If we need to toggle cfs_bandwidth_used, off->on must occur
6987 * before making related changes, and on->off must occur afterwards
6989 if (runtime_enabled && !runtime_was_enabled)
6990 cfs_bandwidth_usage_inc();
6991 raw_spin_lock_irq(&cfs_b->lock);
6992 cfs_b->period = ns_to_ktime(period);
6993 cfs_b->quota = quota;
6995 __refill_cfs_bandwidth_runtime(cfs_b);
6997 /* Restart the period timer (if active) to handle new period expiry: */
6998 if (runtime_enabled)
6999 start_cfs_bandwidth(cfs_b);
7001 raw_spin_unlock_irq(&cfs_b->lock);
7003 for_each_online_cpu(i) {
7004 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7005 struct rq *rq = cfs_rq->rq;
7008 rq_lock_irq(rq, &rf);
7009 cfs_rq->runtime_enabled = runtime_enabled;
7010 cfs_rq->runtime_remaining = 0;
7012 if (cfs_rq->throttled)
7013 unthrottle_cfs_rq(cfs_rq);
7014 rq_unlock_irq(rq, &rf);
7016 if (runtime_was_enabled && !runtime_enabled)
7017 cfs_bandwidth_usage_dec();
7019 mutex_unlock(&cfs_constraints_mutex);
7025 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7029 period = ktime_to_ns(tg->cfs_bandwidth.period);
7030 if (cfs_quota_us < 0)
7031 quota = RUNTIME_INF;
7033 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7035 return tg_set_cfs_bandwidth(tg, period, quota);
7038 long tg_get_cfs_quota(struct task_group *tg)
7042 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7045 quota_us = tg->cfs_bandwidth.quota;
7046 do_div(quota_us, NSEC_PER_USEC);
7051 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7055 period = (u64)cfs_period_us * NSEC_PER_USEC;
7056 quota = tg->cfs_bandwidth.quota;
7058 return tg_set_cfs_bandwidth(tg, period, quota);
7061 long tg_get_cfs_period(struct task_group *tg)
7065 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7066 do_div(cfs_period_us, NSEC_PER_USEC);
7068 return cfs_period_us;
7071 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7074 return tg_get_cfs_quota(css_tg(css));
7077 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7078 struct cftype *cftype, s64 cfs_quota_us)
7080 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7083 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7086 return tg_get_cfs_period(css_tg(css));
7089 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7090 struct cftype *cftype, u64 cfs_period_us)
7092 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7095 struct cfs_schedulable_data {
7096 struct task_group *tg;
7101 * normalize group quota/period to be quota/max_period
7102 * note: units are usecs
7104 static u64 normalize_cfs_quota(struct task_group *tg,
7105 struct cfs_schedulable_data *d)
7113 period = tg_get_cfs_period(tg);
7114 quota = tg_get_cfs_quota(tg);
7117 /* note: these should typically be equivalent */
7118 if (quota == RUNTIME_INF || quota == -1)
7121 return to_ratio(period, quota);
7124 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7126 struct cfs_schedulable_data *d = data;
7127 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7128 s64 quota = 0, parent_quota = -1;
7131 quota = RUNTIME_INF;
7133 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7135 quota = normalize_cfs_quota(tg, d);
7136 parent_quota = parent_b->hierarchical_quota;
7139 * Ensure max(child_quota) <= parent_quota, inherit when no
7142 if (quota == RUNTIME_INF)
7143 quota = parent_quota;
7144 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7147 cfs_b->hierarchical_quota = quota;
7152 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7155 struct cfs_schedulable_data data = {
7161 if (quota != RUNTIME_INF) {
7162 do_div(data.period, NSEC_PER_USEC);
7163 do_div(data.quota, NSEC_PER_USEC);
7167 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7173 static int cpu_stats_show(struct seq_file *sf, void *v)
7175 struct task_group *tg = css_tg(seq_css(sf));
7176 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7178 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7179 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7180 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7184 #endif /* CONFIG_CFS_BANDWIDTH */
7185 #endif /* CONFIG_FAIR_GROUP_SCHED */
7187 #ifdef CONFIG_RT_GROUP_SCHED
7188 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7189 struct cftype *cft, s64 val)
7191 return sched_group_set_rt_runtime(css_tg(css), val);
7194 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7197 return sched_group_rt_runtime(css_tg(css));
7200 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7201 struct cftype *cftype, u64 rt_period_us)
7203 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7206 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7209 return sched_group_rt_period(css_tg(css));
7211 #endif /* CONFIG_RT_GROUP_SCHED */
7213 static struct cftype cpu_files[] = {
7214 #ifdef CONFIG_FAIR_GROUP_SCHED
7217 .read_u64 = cpu_shares_read_u64,
7218 .write_u64 = cpu_shares_write_u64,
7221 #ifdef CONFIG_CFS_BANDWIDTH
7223 .name = "cfs_quota_us",
7224 .read_s64 = cpu_cfs_quota_read_s64,
7225 .write_s64 = cpu_cfs_quota_write_s64,
7228 .name = "cfs_period_us",
7229 .read_u64 = cpu_cfs_period_read_u64,
7230 .write_u64 = cpu_cfs_period_write_u64,
7234 .seq_show = cpu_stats_show,
7237 #ifdef CONFIG_RT_GROUP_SCHED
7239 .name = "rt_runtime_us",
7240 .read_s64 = cpu_rt_runtime_read,
7241 .write_s64 = cpu_rt_runtime_write,
7244 .name = "rt_period_us",
7245 .read_u64 = cpu_rt_period_read_uint,
7246 .write_u64 = cpu_rt_period_write_uint,
7252 struct cgroup_subsys cpu_cgrp_subsys = {
7253 .css_alloc = cpu_cgroup_css_alloc,
7254 .css_online = cpu_cgroup_css_online,
7255 .css_released = cpu_cgroup_css_released,
7256 .css_free = cpu_cgroup_css_free,
7257 .fork = cpu_cgroup_fork,
7258 .can_attach = cpu_cgroup_can_attach,
7259 .attach = cpu_cgroup_attach,
7260 .legacy_cftypes = cpu_files,
7264 #endif /* CONFIG_CGROUP_SCHED */
7266 void dump_cpu_task(int cpu)
7268 pr_info("Task dump for CPU %d:\n", cpu);
7269 sched_show_task(cpu_curr(cpu));
7273 * Nice levels are multiplicative, with a gentle 10% change for every
7274 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7275 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7276 * that remained on nice 0.
7278 * The "10% effect" is relative and cumulative: from _any_ nice level,
7279 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7280 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7281 * If a task goes up by ~10% and another task goes down by ~10% then
7282 * the relative distance between them is ~25%.)
7284 const int sched_prio_to_weight[40] = {
7285 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7286 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7287 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7288 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7289 /* 0 */ 1024, 820, 655, 526, 423,
7290 /* 5 */ 335, 272, 215, 172, 137,
7291 /* 10 */ 110, 87, 70, 56, 45,
7292 /* 15 */ 36, 29, 23, 18, 15,
7296 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7298 * In cases where the weight does not change often, we can use the
7299 * precalculated inverse to speed up arithmetics by turning divisions
7300 * into multiplications:
7302 const u32 sched_prio_to_wmult[40] = {
7303 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7304 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7305 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7306 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7307 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7308 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7309 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7310 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,