4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
78 #include <linux/mutex.h>
80 #include <asm/switch_to.h>
82 #include <asm/irq_regs.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
97 static void update_rq_clock_task(struct rq *rq, s64 delta);
99 void update_rq_clock(struct rq *rq)
103 lockdep_assert_held(&rq->lock);
105 if (rq->clock_update_flags & RQCF_ACT_SKIP)
108 #ifdef CONFIG_SCHED_DEBUG
109 rq->clock_update_flags |= RQCF_UPDATED;
111 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
115 update_rq_clock_task(rq, delta);
119 * Debugging: various feature bits
122 #define SCHED_FEAT(name, enabled) \
123 (1UL << __SCHED_FEAT_##name) * enabled |
125 const_debug unsigned int sysctl_sched_features =
126 #include "features.h"
132 * Number of tasks to iterate in a single balance run.
133 * Limited because this is done with IRQs disabled.
135 const_debug unsigned int sysctl_sched_nr_migrate = 32;
138 * period over which we average the RT time consumption, measured
143 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
146 * period over which we measure -rt task cpu usage in us.
149 unsigned int sysctl_sched_rt_period = 1000000;
151 __read_mostly int scheduler_running;
154 * part of the period that we allow rt tasks to run in us.
157 int sysctl_sched_rt_runtime = 950000;
159 /* cpus with isolated domains */
160 cpumask_var_t cpu_isolated_map;
163 * this_rq_lock - lock this runqueue and disable interrupts.
165 static struct rq *this_rq_lock(void)
172 raw_spin_lock(&rq->lock);
178 * __task_rq_lock - lock the rq @p resides on.
180 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
185 lockdep_assert_held(&p->pi_lock);
189 raw_spin_lock(&rq->lock);
190 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
194 raw_spin_unlock(&rq->lock);
196 while (unlikely(task_on_rq_migrating(p)))
202 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
204 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
205 __acquires(p->pi_lock)
211 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
213 raw_spin_lock(&rq->lock);
215 * move_queued_task() task_rq_lock()
218 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
219 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
220 * [S] ->cpu = new_cpu [L] task_rq()
224 * If we observe the old cpu in task_rq_lock, the acquire of
225 * the old rq->lock will fully serialize against the stores.
227 * If we observe the new cpu in task_rq_lock, the acquire will
228 * pair with the WMB to ensure we must then also see migrating.
230 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
234 raw_spin_unlock(&rq->lock);
235 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
237 while (unlikely(task_on_rq_migrating(p)))
242 #ifdef CONFIG_SCHED_HRTICK
244 * Use HR-timers to deliver accurate preemption points.
247 static void hrtick_clear(struct rq *rq)
249 if (hrtimer_active(&rq->hrtick_timer))
250 hrtimer_cancel(&rq->hrtick_timer);
254 * High-resolution timer tick.
255 * Runs from hardirq context with interrupts disabled.
257 static enum hrtimer_restart hrtick(struct hrtimer *timer)
259 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
261 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
263 raw_spin_lock(&rq->lock);
265 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
266 raw_spin_unlock(&rq->lock);
268 return HRTIMER_NORESTART;
273 static void __hrtick_restart(struct rq *rq)
275 struct hrtimer *timer = &rq->hrtick_timer;
277 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
281 * called from hardirq (IPI) context
283 static void __hrtick_start(void *arg)
287 raw_spin_lock(&rq->lock);
288 __hrtick_restart(rq);
289 rq->hrtick_csd_pending = 0;
290 raw_spin_unlock(&rq->lock);
294 * Called to set the hrtick timer state.
296 * called with rq->lock held and irqs disabled
298 void hrtick_start(struct rq *rq, u64 delay)
300 struct hrtimer *timer = &rq->hrtick_timer;
305 * Don't schedule slices shorter than 10000ns, that just
306 * doesn't make sense and can cause timer DoS.
308 delta = max_t(s64, delay, 10000LL);
309 time = ktime_add_ns(timer->base->get_time(), delta);
311 hrtimer_set_expires(timer, time);
313 if (rq == this_rq()) {
314 __hrtick_restart(rq);
315 } else if (!rq->hrtick_csd_pending) {
316 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
317 rq->hrtick_csd_pending = 1;
323 * Called to set the hrtick timer state.
325 * called with rq->lock held and irqs disabled
327 void hrtick_start(struct rq *rq, u64 delay)
330 * Don't schedule slices shorter than 10000ns, that just
331 * doesn't make sense. Rely on vruntime for fairness.
333 delay = max_t(u64, delay, 10000LL);
334 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
335 HRTIMER_MODE_REL_PINNED);
337 #endif /* CONFIG_SMP */
339 static void init_rq_hrtick(struct rq *rq)
342 rq->hrtick_csd_pending = 0;
344 rq->hrtick_csd.flags = 0;
345 rq->hrtick_csd.func = __hrtick_start;
346 rq->hrtick_csd.info = rq;
349 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
350 rq->hrtick_timer.function = hrtick;
352 #else /* CONFIG_SCHED_HRTICK */
353 static inline void hrtick_clear(struct rq *rq)
357 static inline void init_rq_hrtick(struct rq *rq)
360 #endif /* CONFIG_SCHED_HRTICK */
363 * cmpxchg based fetch_or, macro so it works for different integer types
365 #define fetch_or(ptr, mask) \
367 typeof(ptr) _ptr = (ptr); \
368 typeof(mask) _mask = (mask); \
369 typeof(*_ptr) _old, _val = *_ptr; \
372 _old = cmpxchg(_ptr, _val, _val | _mask); \
380 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
382 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
383 * this avoids any races wrt polling state changes and thereby avoids
386 static bool set_nr_and_not_polling(struct task_struct *p)
388 struct thread_info *ti = task_thread_info(p);
389 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
393 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
395 * If this returns true, then the idle task promises to call
396 * sched_ttwu_pending() and reschedule soon.
398 static bool set_nr_if_polling(struct task_struct *p)
400 struct thread_info *ti = task_thread_info(p);
401 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
404 if (!(val & _TIF_POLLING_NRFLAG))
406 if (val & _TIF_NEED_RESCHED)
408 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
417 static bool set_nr_and_not_polling(struct task_struct *p)
419 set_tsk_need_resched(p);
424 static bool set_nr_if_polling(struct task_struct *p)
431 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
433 struct wake_q_node *node = &task->wake_q;
436 * Atomically grab the task, if ->wake_q is !nil already it means
437 * its already queued (either by us or someone else) and will get the
438 * wakeup due to that.
440 * This cmpxchg() implies a full barrier, which pairs with the write
441 * barrier implied by the wakeup in wake_up_q().
443 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
446 get_task_struct(task);
449 * The head is context local, there can be no concurrency.
452 head->lastp = &node->next;
455 void wake_up_q(struct wake_q_head *head)
457 struct wake_q_node *node = head->first;
459 while (node != WAKE_Q_TAIL) {
460 struct task_struct *task;
462 task = container_of(node, struct task_struct, wake_q);
464 /* task can safely be re-inserted now */
466 task->wake_q.next = NULL;
469 * wake_up_process() implies a wmb() to pair with the queueing
470 * in wake_q_add() so as not to miss wakeups.
472 wake_up_process(task);
473 put_task_struct(task);
478 * resched_curr - mark rq's current task 'to be rescheduled now'.
480 * On UP this means the setting of the need_resched flag, on SMP it
481 * might also involve a cross-CPU call to trigger the scheduler on
484 void resched_curr(struct rq *rq)
486 struct task_struct *curr = rq->curr;
489 lockdep_assert_held(&rq->lock);
491 if (test_tsk_need_resched(curr))
496 if (cpu == smp_processor_id()) {
497 set_tsk_need_resched(curr);
498 set_preempt_need_resched();
502 if (set_nr_and_not_polling(curr))
503 smp_send_reschedule(cpu);
505 trace_sched_wake_idle_without_ipi(cpu);
508 void resched_cpu(int cpu)
510 struct rq *rq = cpu_rq(cpu);
513 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
516 raw_spin_unlock_irqrestore(&rq->lock, flags);
520 #ifdef CONFIG_NO_HZ_COMMON
522 * In the semi idle case, use the nearest busy cpu for migrating timers
523 * from an idle cpu. This is good for power-savings.
525 * We don't do similar optimization for completely idle system, as
526 * selecting an idle cpu will add more delays to the timers than intended
527 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
529 int get_nohz_timer_target(void)
531 int i, cpu = smp_processor_id();
532 struct sched_domain *sd;
534 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
538 for_each_domain(cpu, sd) {
539 for_each_cpu(i, sched_domain_span(sd)) {
543 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
550 if (!is_housekeeping_cpu(cpu))
551 cpu = housekeeping_any_cpu();
557 * When add_timer_on() enqueues a timer into the timer wheel of an
558 * idle CPU then this timer might expire before the next timer event
559 * which is scheduled to wake up that CPU. In case of a completely
560 * idle system the next event might even be infinite time into the
561 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
562 * leaves the inner idle loop so the newly added timer is taken into
563 * account when the CPU goes back to idle and evaluates the timer
564 * wheel for the next timer event.
566 static void wake_up_idle_cpu(int cpu)
568 struct rq *rq = cpu_rq(cpu);
570 if (cpu == smp_processor_id())
573 if (set_nr_and_not_polling(rq->idle))
574 smp_send_reschedule(cpu);
576 trace_sched_wake_idle_without_ipi(cpu);
579 static bool wake_up_full_nohz_cpu(int cpu)
582 * We just need the target to call irq_exit() and re-evaluate
583 * the next tick. The nohz full kick at least implies that.
584 * If needed we can still optimize that later with an
587 if (cpu_is_offline(cpu))
588 return true; /* Don't try to wake offline CPUs. */
589 if (tick_nohz_full_cpu(cpu)) {
590 if (cpu != smp_processor_id() ||
591 tick_nohz_tick_stopped())
592 tick_nohz_full_kick_cpu(cpu);
600 * Wake up the specified CPU. If the CPU is going offline, it is the
601 * caller's responsibility to deal with the lost wakeup, for example,
602 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
604 void wake_up_nohz_cpu(int cpu)
606 if (!wake_up_full_nohz_cpu(cpu))
607 wake_up_idle_cpu(cpu);
610 static inline bool got_nohz_idle_kick(void)
612 int cpu = smp_processor_id();
614 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
617 if (idle_cpu(cpu) && !need_resched())
621 * We can't run Idle Load Balance on this CPU for this time so we
622 * cancel it and clear NOHZ_BALANCE_KICK
624 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
628 #else /* CONFIG_NO_HZ_COMMON */
630 static inline bool got_nohz_idle_kick(void)
635 #endif /* CONFIG_NO_HZ_COMMON */
637 #ifdef CONFIG_NO_HZ_FULL
638 bool sched_can_stop_tick(struct rq *rq)
642 /* Deadline tasks, even if single, need the tick */
643 if (rq->dl.dl_nr_running)
647 * If there are more than one RR tasks, we need the tick to effect the
648 * actual RR behaviour.
650 if (rq->rt.rr_nr_running) {
651 if (rq->rt.rr_nr_running == 1)
658 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
659 * forced preemption between FIFO tasks.
661 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
666 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
667 * if there's more than one we need the tick for involuntary
670 if (rq->nr_running > 1)
675 #endif /* CONFIG_NO_HZ_FULL */
677 void sched_avg_update(struct rq *rq)
679 s64 period = sched_avg_period();
681 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
683 * Inline assembly required to prevent the compiler
684 * optimising this loop into a divmod call.
685 * See __iter_div_u64_rem() for another example of this.
687 asm("" : "+rm" (rq->age_stamp));
688 rq->age_stamp += period;
693 #endif /* CONFIG_SMP */
695 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
696 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
698 * Iterate task_group tree rooted at *from, calling @down when first entering a
699 * node and @up when leaving it for the final time.
701 * Caller must hold rcu_lock or sufficient equivalent.
703 int walk_tg_tree_from(struct task_group *from,
704 tg_visitor down, tg_visitor up, void *data)
706 struct task_group *parent, *child;
712 ret = (*down)(parent, data);
715 list_for_each_entry_rcu(child, &parent->children, siblings) {
722 ret = (*up)(parent, data);
723 if (ret || parent == from)
727 parent = parent->parent;
734 int tg_nop(struct task_group *tg, void *data)
740 static void set_load_weight(struct task_struct *p)
742 int prio = p->static_prio - MAX_RT_PRIO;
743 struct load_weight *load = &p->se.load;
746 * SCHED_IDLE tasks get minimal weight:
748 if (idle_policy(p->policy)) {
749 load->weight = scale_load(WEIGHT_IDLEPRIO);
750 load->inv_weight = WMULT_IDLEPRIO;
754 load->weight = scale_load(sched_prio_to_weight[prio]);
755 load->inv_weight = sched_prio_to_wmult[prio];
758 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
761 if (!(flags & ENQUEUE_RESTORE))
762 sched_info_queued(rq, p);
763 p->sched_class->enqueue_task(rq, p, flags);
766 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
769 if (!(flags & DEQUEUE_SAVE))
770 sched_info_dequeued(rq, p);
771 p->sched_class->dequeue_task(rq, p, flags);
774 void activate_task(struct rq *rq, struct task_struct *p, int flags)
776 if (task_contributes_to_load(p))
777 rq->nr_uninterruptible--;
779 enqueue_task(rq, p, flags);
782 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
784 if (task_contributes_to_load(p))
785 rq->nr_uninterruptible++;
787 dequeue_task(rq, p, flags);
790 static void update_rq_clock_task(struct rq *rq, s64 delta)
793 * In theory, the compile should just see 0 here, and optimize out the call
794 * to sched_rt_avg_update. But I don't trust it...
796 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
797 s64 steal = 0, irq_delta = 0;
799 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
800 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
803 * Since irq_time is only updated on {soft,}irq_exit, we might run into
804 * this case when a previous update_rq_clock() happened inside a
807 * When this happens, we stop ->clock_task and only update the
808 * prev_irq_time stamp to account for the part that fit, so that a next
809 * update will consume the rest. This ensures ->clock_task is
812 * It does however cause some slight miss-attribution of {soft,}irq
813 * time, a more accurate solution would be to update the irq_time using
814 * the current rq->clock timestamp, except that would require using
817 if (irq_delta > delta)
820 rq->prev_irq_time += irq_delta;
823 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
824 if (static_key_false((¶virt_steal_rq_enabled))) {
825 steal = paravirt_steal_clock(cpu_of(rq));
826 steal -= rq->prev_steal_time_rq;
828 if (unlikely(steal > delta))
831 rq->prev_steal_time_rq += steal;
836 rq->clock_task += delta;
838 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
839 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
840 sched_rt_avg_update(rq, irq_delta + steal);
844 void sched_set_stop_task(int cpu, struct task_struct *stop)
846 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
847 struct task_struct *old_stop = cpu_rq(cpu)->stop;
851 * Make it appear like a SCHED_FIFO task, its something
852 * userspace knows about and won't get confused about.
854 * Also, it will make PI more or less work without too
855 * much confusion -- but then, stop work should not
856 * rely on PI working anyway.
858 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
860 stop->sched_class = &stop_sched_class;
863 cpu_rq(cpu)->stop = stop;
867 * Reset it back to a normal scheduling class so that
868 * it can die in pieces.
870 old_stop->sched_class = &rt_sched_class;
875 * __normal_prio - return the priority that is based on the static prio
877 static inline int __normal_prio(struct task_struct *p)
879 return p->static_prio;
883 * Calculate the expected normal priority: i.e. priority
884 * without taking RT-inheritance into account. Might be
885 * boosted by interactivity modifiers. Changes upon fork,
886 * setprio syscalls, and whenever the interactivity
887 * estimator recalculates.
889 static inline int normal_prio(struct task_struct *p)
893 if (task_has_dl_policy(p))
894 prio = MAX_DL_PRIO-1;
895 else if (task_has_rt_policy(p))
896 prio = MAX_RT_PRIO-1 - p->rt_priority;
898 prio = __normal_prio(p);
903 * Calculate the current priority, i.e. the priority
904 * taken into account by the scheduler. This value might
905 * be boosted by RT tasks, or might be boosted by
906 * interactivity modifiers. Will be RT if the task got
907 * RT-boosted. If not then it returns p->normal_prio.
909 static int effective_prio(struct task_struct *p)
911 p->normal_prio = normal_prio(p);
913 * If we are RT tasks or we were boosted to RT priority,
914 * keep the priority unchanged. Otherwise, update priority
915 * to the normal priority:
917 if (!rt_prio(p->prio))
918 return p->normal_prio;
923 * task_curr - is this task currently executing on a CPU?
924 * @p: the task in question.
926 * Return: 1 if the task is currently executing. 0 otherwise.
928 inline int task_curr(const struct task_struct *p)
930 return cpu_curr(task_cpu(p)) == p;
934 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
935 * use the balance_callback list if you want balancing.
937 * this means any call to check_class_changed() must be followed by a call to
938 * balance_callback().
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
948 p->sched_class->switched_to(rq, p);
949 } else if (oldprio != p->prio || dl_task(p))
950 p->sched_class->prio_changed(rq, p, oldprio);
953 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
955 const struct sched_class *class;
957 if (p->sched_class == rq->curr->sched_class) {
958 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
960 for_each_class(class) {
961 if (class == rq->curr->sched_class)
963 if (class == p->sched_class) {
971 * A queue event has occurred, and we're going to schedule. In
972 * this case, we can save a useless back to back clock update.
974 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
975 rq_clock_skip_update(rq, true);
980 * This is how migration works:
982 * 1) we invoke migration_cpu_stop() on the target CPU using
984 * 2) stopper starts to run (implicitly forcing the migrated thread
986 * 3) it checks whether the migrated task is still in the wrong runqueue.
987 * 4) if it's in the wrong runqueue then the migration thread removes
988 * it and puts it into the right queue.
989 * 5) stopper completes and stop_one_cpu() returns and the migration
994 * move_queued_task - move a queued task to new rq.
996 * Returns (locked) new rq. Old rq's lock is released.
998 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1000 lockdep_assert_held(&rq->lock);
1002 p->on_rq = TASK_ON_RQ_MIGRATING;
1003 dequeue_task(rq, p, 0);
1004 set_task_cpu(p, new_cpu);
1005 raw_spin_unlock(&rq->lock);
1007 rq = cpu_rq(new_cpu);
1009 raw_spin_lock(&rq->lock);
1010 BUG_ON(task_cpu(p) != new_cpu);
1011 enqueue_task(rq, p, 0);
1012 p->on_rq = TASK_ON_RQ_QUEUED;
1013 check_preempt_curr(rq, p, 0);
1018 struct migration_arg {
1019 struct task_struct *task;
1024 * Move (not current) task off this cpu, onto dest cpu. We're doing
1025 * this because either it can't run here any more (set_cpus_allowed()
1026 * away from this CPU, or CPU going down), or because we're
1027 * attempting to rebalance this task on exec (sched_exec).
1029 * So we race with normal scheduler movements, but that's OK, as long
1030 * as the task is no longer on this CPU.
1032 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1034 if (unlikely(!cpu_active(dest_cpu)))
1037 /* Affinity changed (again). */
1038 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1041 rq = move_queued_task(rq, p, dest_cpu);
1047 * migration_cpu_stop - this will be executed by a highprio stopper thread
1048 * and performs thread migration by bumping thread off CPU then
1049 * 'pushing' onto another runqueue.
1051 static int migration_cpu_stop(void *data)
1053 struct migration_arg *arg = data;
1054 struct task_struct *p = arg->task;
1055 struct rq *rq = this_rq();
1058 * The original target cpu might have gone down and we might
1059 * be on another cpu but it doesn't matter.
1061 local_irq_disable();
1063 * We need to explicitly wake pending tasks before running
1064 * __migrate_task() such that we will not miss enforcing cpus_allowed
1065 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1067 sched_ttwu_pending();
1069 raw_spin_lock(&p->pi_lock);
1070 raw_spin_lock(&rq->lock);
1072 * If task_rq(p) != rq, it cannot be migrated here, because we're
1073 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1074 * we're holding p->pi_lock.
1076 if (task_rq(p) == rq) {
1077 if (task_on_rq_queued(p))
1078 rq = __migrate_task(rq, p, arg->dest_cpu);
1080 p->wake_cpu = arg->dest_cpu;
1082 raw_spin_unlock(&rq->lock);
1083 raw_spin_unlock(&p->pi_lock);
1090 * sched_class::set_cpus_allowed must do the below, but is not required to
1091 * actually call this function.
1093 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1095 cpumask_copy(&p->cpus_allowed, new_mask);
1096 p->nr_cpus_allowed = cpumask_weight(new_mask);
1099 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1101 struct rq *rq = task_rq(p);
1102 bool queued, running;
1104 lockdep_assert_held(&p->pi_lock);
1106 queued = task_on_rq_queued(p);
1107 running = task_current(rq, p);
1111 * Because __kthread_bind() calls this on blocked tasks without
1114 lockdep_assert_held(&rq->lock);
1115 dequeue_task(rq, p, DEQUEUE_SAVE);
1118 put_prev_task(rq, p);
1120 p->sched_class->set_cpus_allowed(p, new_mask);
1123 enqueue_task(rq, p, ENQUEUE_RESTORE);
1125 set_curr_task(rq, p);
1129 * Change a given task's CPU affinity. Migrate the thread to a
1130 * proper CPU and schedule it away if the CPU it's executing on
1131 * is removed from the allowed bitmask.
1133 * NOTE: the caller must have a valid reference to the task, the
1134 * task must not exit() & deallocate itself prematurely. The
1135 * call is not atomic; no spinlocks may be held.
1137 static int __set_cpus_allowed_ptr(struct task_struct *p,
1138 const struct cpumask *new_mask, bool check)
1140 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1141 unsigned int dest_cpu;
1146 rq = task_rq_lock(p, &rf);
1148 if (p->flags & PF_KTHREAD) {
1150 * Kernel threads are allowed on online && !active CPUs
1152 cpu_valid_mask = cpu_online_mask;
1156 * Must re-check here, to close a race against __kthread_bind(),
1157 * sched_setaffinity() is not guaranteed to observe the flag.
1159 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1164 if (cpumask_equal(&p->cpus_allowed, new_mask))
1167 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1172 do_set_cpus_allowed(p, new_mask);
1174 if (p->flags & PF_KTHREAD) {
1176 * For kernel threads that do indeed end up on online &&
1177 * !active we want to ensure they are strict per-cpu threads.
1179 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1180 !cpumask_intersects(new_mask, cpu_active_mask) &&
1181 p->nr_cpus_allowed != 1);
1184 /* Can the task run on the task's current CPU? If so, we're done */
1185 if (cpumask_test_cpu(task_cpu(p), new_mask))
1188 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1189 if (task_running(rq, p) || p->state == TASK_WAKING) {
1190 struct migration_arg arg = { p, dest_cpu };
1191 /* Need help from migration thread: drop lock and wait. */
1192 task_rq_unlock(rq, p, &rf);
1193 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1194 tlb_migrate_finish(p->mm);
1196 } else if (task_on_rq_queued(p)) {
1198 * OK, since we're going to drop the lock immediately
1199 * afterwards anyway.
1201 rq_unpin_lock(rq, &rf);
1202 rq = move_queued_task(rq, p, dest_cpu);
1203 rq_repin_lock(rq, &rf);
1206 task_rq_unlock(rq, p, &rf);
1211 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1213 return __set_cpus_allowed_ptr(p, new_mask, false);
1215 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1217 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1219 #ifdef CONFIG_SCHED_DEBUG
1221 * We should never call set_task_cpu() on a blocked task,
1222 * ttwu() will sort out the placement.
1224 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1228 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1229 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1230 * time relying on p->on_rq.
1232 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1233 p->sched_class == &fair_sched_class &&
1234 (p->on_rq && !task_on_rq_migrating(p)));
1236 #ifdef CONFIG_LOCKDEP
1238 * The caller should hold either p->pi_lock or rq->lock, when changing
1239 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1241 * sched_move_task() holds both and thus holding either pins the cgroup,
1244 * Furthermore, all task_rq users should acquire both locks, see
1247 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1248 lockdep_is_held(&task_rq(p)->lock)));
1252 trace_sched_migrate_task(p, new_cpu);
1254 if (task_cpu(p) != new_cpu) {
1255 if (p->sched_class->migrate_task_rq)
1256 p->sched_class->migrate_task_rq(p);
1257 p->se.nr_migrations++;
1258 perf_event_task_migrate(p);
1261 __set_task_cpu(p, new_cpu);
1264 static void __migrate_swap_task(struct task_struct *p, int cpu)
1266 if (task_on_rq_queued(p)) {
1267 struct rq *src_rq, *dst_rq;
1269 src_rq = task_rq(p);
1270 dst_rq = cpu_rq(cpu);
1272 p->on_rq = TASK_ON_RQ_MIGRATING;
1273 deactivate_task(src_rq, p, 0);
1274 set_task_cpu(p, cpu);
1275 activate_task(dst_rq, p, 0);
1276 p->on_rq = TASK_ON_RQ_QUEUED;
1277 check_preempt_curr(dst_rq, p, 0);
1280 * Task isn't running anymore; make it appear like we migrated
1281 * it before it went to sleep. This means on wakeup we make the
1282 * previous cpu our target instead of where it really is.
1288 struct migration_swap_arg {
1289 struct task_struct *src_task, *dst_task;
1290 int src_cpu, dst_cpu;
1293 static int migrate_swap_stop(void *data)
1295 struct migration_swap_arg *arg = data;
1296 struct rq *src_rq, *dst_rq;
1299 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1302 src_rq = cpu_rq(arg->src_cpu);
1303 dst_rq = cpu_rq(arg->dst_cpu);
1305 double_raw_lock(&arg->src_task->pi_lock,
1306 &arg->dst_task->pi_lock);
1307 double_rq_lock(src_rq, dst_rq);
1309 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1312 if (task_cpu(arg->src_task) != arg->src_cpu)
1315 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1318 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1321 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1322 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1327 double_rq_unlock(src_rq, dst_rq);
1328 raw_spin_unlock(&arg->dst_task->pi_lock);
1329 raw_spin_unlock(&arg->src_task->pi_lock);
1335 * Cross migrate two tasks
1337 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1339 struct migration_swap_arg arg;
1342 arg = (struct migration_swap_arg){
1344 .src_cpu = task_cpu(cur),
1346 .dst_cpu = task_cpu(p),
1349 if (arg.src_cpu == arg.dst_cpu)
1353 * These three tests are all lockless; this is OK since all of them
1354 * will be re-checked with proper locks held further down the line.
1356 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1359 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1362 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1365 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1366 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1373 * wait_task_inactive - wait for a thread to unschedule.
1375 * If @match_state is nonzero, it's the @p->state value just checked and
1376 * not expected to change. If it changes, i.e. @p might have woken up,
1377 * then return zero. When we succeed in waiting for @p to be off its CPU,
1378 * we return a positive number (its total switch count). If a second call
1379 * a short while later returns the same number, the caller can be sure that
1380 * @p has remained unscheduled the whole time.
1382 * The caller must ensure that the task *will* unschedule sometime soon,
1383 * else this function might spin for a *long* time. This function can't
1384 * be called with interrupts off, or it may introduce deadlock with
1385 * smp_call_function() if an IPI is sent by the same process we are
1386 * waiting to become inactive.
1388 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1390 int running, queued;
1397 * We do the initial early heuristics without holding
1398 * any task-queue locks at all. We'll only try to get
1399 * the runqueue lock when things look like they will
1405 * If the task is actively running on another CPU
1406 * still, just relax and busy-wait without holding
1409 * NOTE! Since we don't hold any locks, it's not
1410 * even sure that "rq" stays as the right runqueue!
1411 * But we don't care, since "task_running()" will
1412 * return false if the runqueue has changed and p
1413 * is actually now running somewhere else!
1415 while (task_running(rq, p)) {
1416 if (match_state && unlikely(p->state != match_state))
1422 * Ok, time to look more closely! We need the rq
1423 * lock now, to be *sure*. If we're wrong, we'll
1424 * just go back and repeat.
1426 rq = task_rq_lock(p, &rf);
1427 trace_sched_wait_task(p);
1428 running = task_running(rq, p);
1429 queued = task_on_rq_queued(p);
1431 if (!match_state || p->state == match_state)
1432 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1433 task_rq_unlock(rq, p, &rf);
1436 * If it changed from the expected state, bail out now.
1438 if (unlikely(!ncsw))
1442 * Was it really running after all now that we
1443 * checked with the proper locks actually held?
1445 * Oops. Go back and try again..
1447 if (unlikely(running)) {
1453 * It's not enough that it's not actively running,
1454 * it must be off the runqueue _entirely_, and not
1457 * So if it was still runnable (but just not actively
1458 * running right now), it's preempted, and we should
1459 * yield - it could be a while.
1461 if (unlikely(queued)) {
1462 ktime_t to = NSEC_PER_SEC / HZ;
1464 set_current_state(TASK_UNINTERRUPTIBLE);
1465 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1470 * Ahh, all good. It wasn't running, and it wasn't
1471 * runnable, which means that it will never become
1472 * running in the future either. We're all done!
1481 * kick_process - kick a running thread to enter/exit the kernel
1482 * @p: the to-be-kicked thread
1484 * Cause a process which is running on another CPU to enter
1485 * kernel-mode, without any delay. (to get signals handled.)
1487 * NOTE: this function doesn't have to take the runqueue lock,
1488 * because all it wants to ensure is that the remote task enters
1489 * the kernel. If the IPI races and the task has been migrated
1490 * to another CPU then no harm is done and the purpose has been
1493 void kick_process(struct task_struct *p)
1499 if ((cpu != smp_processor_id()) && task_curr(p))
1500 smp_send_reschedule(cpu);
1503 EXPORT_SYMBOL_GPL(kick_process);
1506 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1508 * A few notes on cpu_active vs cpu_online:
1510 * - cpu_active must be a subset of cpu_online
1512 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1513 * see __set_cpus_allowed_ptr(). At this point the newly online
1514 * cpu isn't yet part of the sched domains, and balancing will not
1517 * - on cpu-down we clear cpu_active() to mask the sched domains and
1518 * avoid the load balancer to place new tasks on the to be removed
1519 * cpu. Existing tasks will remain running there and will be taken
1522 * This means that fallback selection must not select !active CPUs.
1523 * And can assume that any active CPU must be online. Conversely
1524 * select_task_rq() below may allow selection of !active CPUs in order
1525 * to satisfy the above rules.
1527 static int select_fallback_rq(int cpu, struct task_struct *p)
1529 int nid = cpu_to_node(cpu);
1530 const struct cpumask *nodemask = NULL;
1531 enum { cpuset, possible, fail } state = cpuset;
1535 * If the node that the cpu is on has been offlined, cpu_to_node()
1536 * will return -1. There is no cpu on the node, and we should
1537 * select the cpu on the other node.
1540 nodemask = cpumask_of_node(nid);
1542 /* Look for allowed, online CPU in same node. */
1543 for_each_cpu(dest_cpu, nodemask) {
1544 if (!cpu_active(dest_cpu))
1546 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1552 /* Any allowed, online CPU? */
1553 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1554 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1556 if (!cpu_online(dest_cpu))
1561 /* No more Mr. Nice Guy. */
1564 if (IS_ENABLED(CONFIG_CPUSETS)) {
1565 cpuset_cpus_allowed_fallback(p);
1571 do_set_cpus_allowed(p, cpu_possible_mask);
1582 if (state != cpuset) {
1584 * Don't tell them about moving exiting tasks or
1585 * kernel threads (both mm NULL), since they never
1588 if (p->mm && printk_ratelimit()) {
1589 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1590 task_pid_nr(p), p->comm, cpu);
1598 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1601 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1603 lockdep_assert_held(&p->pi_lock);
1605 if (tsk_nr_cpus_allowed(p) > 1)
1606 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1608 cpu = cpumask_any(tsk_cpus_allowed(p));
1611 * In order not to call set_task_cpu() on a blocking task we need
1612 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1615 * Since this is common to all placement strategies, this lives here.
1617 * [ this allows ->select_task() to simply return task_cpu(p) and
1618 * not worry about this generic constraint ]
1620 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1622 cpu = select_fallback_rq(task_cpu(p), p);
1627 static void update_avg(u64 *avg, u64 sample)
1629 s64 diff = sample - *avg;
1635 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1636 const struct cpumask *new_mask, bool check)
1638 return set_cpus_allowed_ptr(p, new_mask);
1641 #endif /* CONFIG_SMP */
1644 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1648 if (!schedstat_enabled())
1654 if (cpu == rq->cpu) {
1655 schedstat_inc(rq->ttwu_local);
1656 schedstat_inc(p->se.statistics.nr_wakeups_local);
1658 struct sched_domain *sd;
1660 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1662 for_each_domain(rq->cpu, sd) {
1663 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1664 schedstat_inc(sd->ttwu_wake_remote);
1671 if (wake_flags & WF_MIGRATED)
1672 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1673 #endif /* CONFIG_SMP */
1675 schedstat_inc(rq->ttwu_count);
1676 schedstat_inc(p->se.statistics.nr_wakeups);
1678 if (wake_flags & WF_SYNC)
1679 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1682 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1684 activate_task(rq, p, en_flags);
1685 p->on_rq = TASK_ON_RQ_QUEUED;
1687 /* if a worker is waking up, notify workqueue */
1688 if (p->flags & PF_WQ_WORKER)
1689 wq_worker_waking_up(p, cpu_of(rq));
1693 * Mark the task runnable and perform wakeup-preemption.
1695 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1696 struct rq_flags *rf)
1698 check_preempt_curr(rq, p, wake_flags);
1699 p->state = TASK_RUNNING;
1700 trace_sched_wakeup(p);
1703 if (p->sched_class->task_woken) {
1705 * Our task @p is fully woken up and running; so its safe to
1706 * drop the rq->lock, hereafter rq is only used for statistics.
1708 rq_unpin_lock(rq, rf);
1709 p->sched_class->task_woken(rq, p);
1710 rq_repin_lock(rq, rf);
1713 if (rq->idle_stamp) {
1714 u64 delta = rq_clock(rq) - rq->idle_stamp;
1715 u64 max = 2*rq->max_idle_balance_cost;
1717 update_avg(&rq->avg_idle, delta);
1719 if (rq->avg_idle > max)
1728 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1729 struct rq_flags *rf)
1731 int en_flags = ENQUEUE_WAKEUP;
1733 lockdep_assert_held(&rq->lock);
1736 if (p->sched_contributes_to_load)
1737 rq->nr_uninterruptible--;
1739 if (wake_flags & WF_MIGRATED)
1740 en_flags |= ENQUEUE_MIGRATED;
1743 ttwu_activate(rq, p, en_flags);
1744 ttwu_do_wakeup(rq, p, wake_flags, rf);
1748 * Called in case the task @p isn't fully descheduled from its runqueue,
1749 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1750 * since all we need to do is flip p->state to TASK_RUNNING, since
1751 * the task is still ->on_rq.
1753 static int ttwu_remote(struct task_struct *p, int wake_flags)
1759 rq = __task_rq_lock(p, &rf);
1760 if (task_on_rq_queued(p)) {
1761 /* check_preempt_curr() may use rq clock */
1762 update_rq_clock(rq);
1763 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1766 __task_rq_unlock(rq, &rf);
1772 void sched_ttwu_pending(void)
1774 struct rq *rq = this_rq();
1775 struct llist_node *llist = llist_del_all(&rq->wake_list);
1776 struct task_struct *p;
1777 unsigned long flags;
1783 raw_spin_lock_irqsave(&rq->lock, flags);
1784 rq_pin_lock(rq, &rf);
1789 p = llist_entry(llist, struct task_struct, wake_entry);
1790 llist = llist_next(llist);
1792 if (p->sched_remote_wakeup)
1793 wake_flags = WF_MIGRATED;
1795 ttwu_do_activate(rq, p, wake_flags, &rf);
1798 rq_unpin_lock(rq, &rf);
1799 raw_spin_unlock_irqrestore(&rq->lock, flags);
1802 void scheduler_ipi(void)
1805 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1806 * TIF_NEED_RESCHED remotely (for the first time) will also send
1809 preempt_fold_need_resched();
1811 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1815 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1816 * traditionally all their work was done from the interrupt return
1817 * path. Now that we actually do some work, we need to make sure
1820 * Some archs already do call them, luckily irq_enter/exit nest
1823 * Arguably we should visit all archs and update all handlers,
1824 * however a fair share of IPIs are still resched only so this would
1825 * somewhat pessimize the simple resched case.
1828 sched_ttwu_pending();
1831 * Check if someone kicked us for doing the nohz idle load balance.
1833 if (unlikely(got_nohz_idle_kick())) {
1834 this_rq()->idle_balance = 1;
1835 raise_softirq_irqoff(SCHED_SOFTIRQ);
1840 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1842 struct rq *rq = cpu_rq(cpu);
1844 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1846 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1847 if (!set_nr_if_polling(rq->idle))
1848 smp_send_reschedule(cpu);
1850 trace_sched_wake_idle_without_ipi(cpu);
1854 void wake_up_if_idle(int cpu)
1856 struct rq *rq = cpu_rq(cpu);
1857 unsigned long flags;
1861 if (!is_idle_task(rcu_dereference(rq->curr)))
1864 if (set_nr_if_polling(rq->idle)) {
1865 trace_sched_wake_idle_without_ipi(cpu);
1867 raw_spin_lock_irqsave(&rq->lock, flags);
1868 if (is_idle_task(rq->curr))
1869 smp_send_reschedule(cpu);
1870 /* Else cpu is not in idle, do nothing here */
1871 raw_spin_unlock_irqrestore(&rq->lock, flags);
1878 bool cpus_share_cache(int this_cpu, int that_cpu)
1880 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1882 #endif /* CONFIG_SMP */
1884 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1886 struct rq *rq = cpu_rq(cpu);
1889 #if defined(CONFIG_SMP)
1890 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1891 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1892 ttwu_queue_remote(p, cpu, wake_flags);
1897 raw_spin_lock(&rq->lock);
1898 rq_pin_lock(rq, &rf);
1899 ttwu_do_activate(rq, p, wake_flags, &rf);
1900 rq_unpin_lock(rq, &rf);
1901 raw_spin_unlock(&rq->lock);
1905 * Notes on Program-Order guarantees on SMP systems.
1909 * The basic program-order guarantee on SMP systems is that when a task [t]
1910 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1911 * execution on its new cpu [c1].
1913 * For migration (of runnable tasks) this is provided by the following means:
1915 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1916 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1917 * rq(c1)->lock (if not at the same time, then in that order).
1918 * C) LOCK of the rq(c1)->lock scheduling in task
1920 * Transitivity guarantees that B happens after A and C after B.
1921 * Note: we only require RCpc transitivity.
1922 * Note: the cpu doing B need not be c0 or c1
1931 * UNLOCK rq(0)->lock
1933 * LOCK rq(0)->lock // orders against CPU0
1935 * UNLOCK rq(0)->lock
1939 * UNLOCK rq(1)->lock
1941 * LOCK rq(1)->lock // orders against CPU2
1944 * UNLOCK rq(1)->lock
1947 * BLOCKING -- aka. SLEEP + WAKEUP
1949 * For blocking we (obviously) need to provide the same guarantee as for
1950 * migration. However the means are completely different as there is no lock
1951 * chain to provide order. Instead we do:
1953 * 1) smp_store_release(X->on_cpu, 0)
1954 * 2) smp_cond_load_acquire(!X->on_cpu)
1958 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1960 * LOCK rq(0)->lock LOCK X->pi_lock
1963 * smp_store_release(X->on_cpu, 0);
1965 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1971 * X->state = RUNNING
1972 * UNLOCK rq(2)->lock
1974 * LOCK rq(2)->lock // orders against CPU1
1977 * UNLOCK rq(2)->lock
1980 * UNLOCK rq(0)->lock
1983 * However; for wakeups there is a second guarantee we must provide, namely we
1984 * must observe the state that lead to our wakeup. That is, not only must our
1985 * task observe its own prior state, it must also observe the stores prior to
1988 * This means that any means of doing remote wakeups must order the CPU doing
1989 * the wakeup against the CPU the task is going to end up running on. This,
1990 * however, is already required for the regular Program-Order guarantee above,
1991 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1996 * try_to_wake_up - wake up a thread
1997 * @p: the thread to be awakened
1998 * @state: the mask of task states that can be woken
1999 * @wake_flags: wake modifier flags (WF_*)
2001 * If (@state & @p->state) @p->state = TASK_RUNNING.
2003 * If the task was not queued/runnable, also place it back on a runqueue.
2005 * Atomic against schedule() which would dequeue a task, also see
2006 * set_current_state().
2008 * Return: %true if @p->state changes (an actual wakeup was done),
2012 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2014 unsigned long flags;
2015 int cpu, success = 0;
2018 * If we are going to wake up a thread waiting for CONDITION we
2019 * need to ensure that CONDITION=1 done by the caller can not be
2020 * reordered with p->state check below. This pairs with mb() in
2021 * set_current_state() the waiting thread does.
2023 smp_mb__before_spinlock();
2024 raw_spin_lock_irqsave(&p->pi_lock, flags);
2025 if (!(p->state & state))
2028 trace_sched_waking(p);
2030 success = 1; /* we're going to change ->state */
2034 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2035 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2036 * in smp_cond_load_acquire() below.
2038 * sched_ttwu_pending() try_to_wake_up()
2039 * [S] p->on_rq = 1; [L] P->state
2040 * UNLOCK rq->lock -----.
2044 * LOCK rq->lock -----'
2048 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2050 * Pairs with the UNLOCK+LOCK on rq->lock from the
2051 * last wakeup of our task and the schedule that got our task
2055 if (p->on_rq && ttwu_remote(p, wake_flags))
2060 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2061 * possible to, falsely, observe p->on_cpu == 0.
2063 * One must be running (->on_cpu == 1) in order to remove oneself
2064 * from the runqueue.
2066 * [S] ->on_cpu = 1; [L] ->on_rq
2070 * [S] ->on_rq = 0; [L] ->on_cpu
2072 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2073 * from the consecutive calls to schedule(); the first switching to our
2074 * task, the second putting it to sleep.
2079 * If the owning (remote) cpu is still in the middle of schedule() with
2080 * this task as prev, wait until its done referencing the task.
2082 * Pairs with the smp_store_release() in finish_lock_switch().
2084 * This ensures that tasks getting woken will be fully ordered against
2085 * their previous state and preserve Program Order.
2087 smp_cond_load_acquire(&p->on_cpu, !VAL);
2089 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2090 p->state = TASK_WAKING;
2093 delayacct_blkio_end();
2094 atomic_dec(&task_rq(p)->nr_iowait);
2097 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2098 if (task_cpu(p) != cpu) {
2099 wake_flags |= WF_MIGRATED;
2100 set_task_cpu(p, cpu);
2103 #else /* CONFIG_SMP */
2106 delayacct_blkio_end();
2107 atomic_dec(&task_rq(p)->nr_iowait);
2110 #endif /* CONFIG_SMP */
2112 ttwu_queue(p, cpu, wake_flags);
2114 ttwu_stat(p, cpu, wake_flags);
2116 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2122 * try_to_wake_up_local - try to wake up a local task with rq lock held
2123 * @p: the thread to be awakened
2124 * @cookie: context's cookie for pinning
2126 * Put @p on the run-queue if it's not already there. The caller must
2127 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2130 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2132 struct rq *rq = task_rq(p);
2134 if (WARN_ON_ONCE(rq != this_rq()) ||
2135 WARN_ON_ONCE(p == current))
2138 lockdep_assert_held(&rq->lock);
2140 if (!raw_spin_trylock(&p->pi_lock)) {
2142 * This is OK, because current is on_cpu, which avoids it being
2143 * picked for load-balance and preemption/IRQs are still
2144 * disabled avoiding further scheduler activity on it and we've
2145 * not yet picked a replacement task.
2147 rq_unpin_lock(rq, rf);
2148 raw_spin_unlock(&rq->lock);
2149 raw_spin_lock(&p->pi_lock);
2150 raw_spin_lock(&rq->lock);
2151 rq_repin_lock(rq, rf);
2154 if (!(p->state & TASK_NORMAL))
2157 trace_sched_waking(p);
2159 if (!task_on_rq_queued(p)) {
2161 delayacct_blkio_end();
2162 atomic_dec(&rq->nr_iowait);
2164 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2167 ttwu_do_wakeup(rq, p, 0, rf);
2168 ttwu_stat(p, smp_processor_id(), 0);
2170 raw_spin_unlock(&p->pi_lock);
2174 * wake_up_process - Wake up a specific process
2175 * @p: The process to be woken up.
2177 * Attempt to wake up the nominated process and move it to the set of runnable
2180 * Return: 1 if the process was woken up, 0 if it was already running.
2182 * It may be assumed that this function implies a write memory barrier before
2183 * changing the task state if and only if any tasks are woken up.
2185 int wake_up_process(struct task_struct *p)
2187 return try_to_wake_up(p, TASK_NORMAL, 0);
2189 EXPORT_SYMBOL(wake_up_process);
2191 int wake_up_state(struct task_struct *p, unsigned int state)
2193 return try_to_wake_up(p, state, 0);
2197 * This function clears the sched_dl_entity static params.
2199 void __dl_clear_params(struct task_struct *p)
2201 struct sched_dl_entity *dl_se = &p->dl;
2203 dl_se->dl_runtime = 0;
2204 dl_se->dl_deadline = 0;
2205 dl_se->dl_period = 0;
2209 dl_se->dl_throttled = 0;
2210 dl_se->dl_yielded = 0;
2214 * Perform scheduler related setup for a newly forked process p.
2215 * p is forked by current.
2217 * __sched_fork() is basic setup used by init_idle() too:
2219 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2224 p->se.exec_start = 0;
2225 p->se.sum_exec_runtime = 0;
2226 p->se.prev_sum_exec_runtime = 0;
2227 p->se.nr_migrations = 0;
2229 INIT_LIST_HEAD(&p->se.group_node);
2231 #ifdef CONFIG_FAIR_GROUP_SCHED
2232 p->se.cfs_rq = NULL;
2235 #ifdef CONFIG_SCHEDSTATS
2236 /* Even if schedstat is disabled, there should not be garbage */
2237 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2240 RB_CLEAR_NODE(&p->dl.rb_node);
2241 init_dl_task_timer(&p->dl);
2242 __dl_clear_params(p);
2244 INIT_LIST_HEAD(&p->rt.run_list);
2246 p->rt.time_slice = sched_rr_timeslice;
2250 #ifdef CONFIG_PREEMPT_NOTIFIERS
2251 INIT_HLIST_HEAD(&p->preempt_notifiers);
2254 #ifdef CONFIG_NUMA_BALANCING
2255 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2256 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2257 p->mm->numa_scan_seq = 0;
2260 if (clone_flags & CLONE_VM)
2261 p->numa_preferred_nid = current->numa_preferred_nid;
2263 p->numa_preferred_nid = -1;
2265 p->node_stamp = 0ULL;
2266 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2267 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2268 p->numa_work.next = &p->numa_work;
2269 p->numa_faults = NULL;
2270 p->last_task_numa_placement = 0;
2271 p->last_sum_exec_runtime = 0;
2273 p->numa_group = NULL;
2274 #endif /* CONFIG_NUMA_BALANCING */
2277 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2279 #ifdef CONFIG_NUMA_BALANCING
2281 void set_numabalancing_state(bool enabled)
2284 static_branch_enable(&sched_numa_balancing);
2286 static_branch_disable(&sched_numa_balancing);
2289 #ifdef CONFIG_PROC_SYSCTL
2290 int sysctl_numa_balancing(struct ctl_table *table, int write,
2291 void __user *buffer, size_t *lenp, loff_t *ppos)
2295 int state = static_branch_likely(&sched_numa_balancing);
2297 if (write && !capable(CAP_SYS_ADMIN))
2302 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2306 set_numabalancing_state(state);
2312 #ifdef CONFIG_SCHEDSTATS
2314 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2315 static bool __initdata __sched_schedstats = false;
2317 static void set_schedstats(bool enabled)
2320 static_branch_enable(&sched_schedstats);
2322 static_branch_disable(&sched_schedstats);
2325 void force_schedstat_enabled(void)
2327 if (!schedstat_enabled()) {
2328 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2329 static_branch_enable(&sched_schedstats);
2333 static int __init setup_schedstats(char *str)
2340 * This code is called before jump labels have been set up, so we can't
2341 * change the static branch directly just yet. Instead set a temporary
2342 * variable so init_schedstats() can do it later.
2344 if (!strcmp(str, "enable")) {
2345 __sched_schedstats = true;
2347 } else if (!strcmp(str, "disable")) {
2348 __sched_schedstats = false;
2353 pr_warn("Unable to parse schedstats=\n");
2357 __setup("schedstats=", setup_schedstats);
2359 static void __init init_schedstats(void)
2361 set_schedstats(__sched_schedstats);
2364 #ifdef CONFIG_PROC_SYSCTL
2365 int sysctl_schedstats(struct ctl_table *table, int write,
2366 void __user *buffer, size_t *lenp, loff_t *ppos)
2370 int state = static_branch_likely(&sched_schedstats);
2372 if (write && !capable(CAP_SYS_ADMIN))
2377 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2381 set_schedstats(state);
2384 #endif /* CONFIG_PROC_SYSCTL */
2385 #else /* !CONFIG_SCHEDSTATS */
2386 static inline void init_schedstats(void) {}
2387 #endif /* CONFIG_SCHEDSTATS */
2390 * fork()/clone()-time setup:
2392 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2394 unsigned long flags;
2395 int cpu = get_cpu();
2397 __sched_fork(clone_flags, p);
2399 * We mark the process as NEW here. This guarantees that
2400 * nobody will actually run it, and a signal or other external
2401 * event cannot wake it up and insert it on the runqueue either.
2403 p->state = TASK_NEW;
2406 * Make sure we do not leak PI boosting priority to the child.
2408 p->prio = current->normal_prio;
2411 * Revert to default priority/policy on fork if requested.
2413 if (unlikely(p->sched_reset_on_fork)) {
2414 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2415 p->policy = SCHED_NORMAL;
2416 p->static_prio = NICE_TO_PRIO(0);
2418 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2419 p->static_prio = NICE_TO_PRIO(0);
2421 p->prio = p->normal_prio = __normal_prio(p);
2425 * We don't need the reset flag anymore after the fork. It has
2426 * fulfilled its duty:
2428 p->sched_reset_on_fork = 0;
2431 if (dl_prio(p->prio)) {
2434 } else if (rt_prio(p->prio)) {
2435 p->sched_class = &rt_sched_class;
2437 p->sched_class = &fair_sched_class;
2440 init_entity_runnable_average(&p->se);
2443 * The child is not yet in the pid-hash so no cgroup attach races,
2444 * and the cgroup is pinned to this child due to cgroup_fork()
2445 * is ran before sched_fork().
2447 * Silence PROVE_RCU.
2449 raw_spin_lock_irqsave(&p->pi_lock, flags);
2451 * We're setting the cpu for the first time, we don't migrate,
2452 * so use __set_task_cpu().
2454 __set_task_cpu(p, cpu);
2455 if (p->sched_class->task_fork)
2456 p->sched_class->task_fork(p);
2457 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2459 #ifdef CONFIG_SCHED_INFO
2460 if (likely(sched_info_on()))
2461 memset(&p->sched_info, 0, sizeof(p->sched_info));
2463 #if defined(CONFIG_SMP)
2466 init_task_preempt_count(p);
2468 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2469 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2476 unsigned long to_ratio(u64 period, u64 runtime)
2478 if (runtime == RUNTIME_INF)
2482 * Doing this here saves a lot of checks in all
2483 * the calling paths, and returning zero seems
2484 * safe for them anyway.
2489 return div64_u64(runtime << 20, period);
2493 inline struct dl_bw *dl_bw_of(int i)
2495 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2496 "sched RCU must be held");
2497 return &cpu_rq(i)->rd->dl_bw;
2500 static inline int dl_bw_cpus(int i)
2502 struct root_domain *rd = cpu_rq(i)->rd;
2505 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2506 "sched RCU must be held");
2507 for_each_cpu_and(i, rd->span, cpu_active_mask)
2513 inline struct dl_bw *dl_bw_of(int i)
2515 return &cpu_rq(i)->dl.dl_bw;
2518 static inline int dl_bw_cpus(int i)
2525 * We must be sure that accepting a new task (or allowing changing the
2526 * parameters of an existing one) is consistent with the bandwidth
2527 * constraints. If yes, this function also accordingly updates the currently
2528 * allocated bandwidth to reflect the new situation.
2530 * This function is called while holding p's rq->lock.
2532 * XXX we should delay bw change until the task's 0-lag point, see
2535 static int dl_overflow(struct task_struct *p, int policy,
2536 const struct sched_attr *attr)
2539 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2540 u64 period = attr->sched_period ?: attr->sched_deadline;
2541 u64 runtime = attr->sched_runtime;
2542 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2545 /* !deadline task may carry old deadline bandwidth */
2546 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2550 * Either if a task, enters, leave, or stays -deadline but changes
2551 * its parameters, we may need to update accordingly the total
2552 * allocated bandwidth of the container.
2554 raw_spin_lock(&dl_b->lock);
2555 cpus = dl_bw_cpus(task_cpu(p));
2556 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2557 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2558 __dl_add(dl_b, new_bw);
2560 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2561 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2562 __dl_clear(dl_b, p->dl.dl_bw);
2563 __dl_add(dl_b, new_bw);
2565 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2566 __dl_clear(dl_b, p->dl.dl_bw);
2569 raw_spin_unlock(&dl_b->lock);
2574 extern void init_dl_bw(struct dl_bw *dl_b);
2577 * wake_up_new_task - wake up a newly created task for the first time.
2579 * This function will do some initial scheduler statistics housekeeping
2580 * that must be done for every newly created context, then puts the task
2581 * on the runqueue and wakes it.
2583 void wake_up_new_task(struct task_struct *p)
2588 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2589 p->state = TASK_RUNNING;
2592 * Fork balancing, do it here and not earlier because:
2593 * - cpus_allowed can change in the fork path
2594 * - any previously selected cpu might disappear through hotplug
2596 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2597 * as we're not fully set-up yet.
2599 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2601 rq = __task_rq_lock(p, &rf);
2602 update_rq_clock(rq);
2603 post_init_entity_util_avg(&p->se);
2605 activate_task(rq, p, 0);
2606 p->on_rq = TASK_ON_RQ_QUEUED;
2607 trace_sched_wakeup_new(p);
2608 check_preempt_curr(rq, p, WF_FORK);
2610 if (p->sched_class->task_woken) {
2612 * Nothing relies on rq->lock after this, so its fine to
2615 rq_unpin_lock(rq, &rf);
2616 p->sched_class->task_woken(rq, p);
2617 rq_repin_lock(rq, &rf);
2620 task_rq_unlock(rq, p, &rf);
2623 #ifdef CONFIG_PREEMPT_NOTIFIERS
2625 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2627 void preempt_notifier_inc(void)
2629 static_key_slow_inc(&preempt_notifier_key);
2631 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2633 void preempt_notifier_dec(void)
2635 static_key_slow_dec(&preempt_notifier_key);
2637 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2640 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2641 * @notifier: notifier struct to register
2643 void preempt_notifier_register(struct preempt_notifier *notifier)
2645 if (!static_key_false(&preempt_notifier_key))
2646 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2648 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2650 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2653 * preempt_notifier_unregister - no longer interested in preemption notifications
2654 * @notifier: notifier struct to unregister
2656 * This is *not* safe to call from within a preemption notifier.
2658 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2660 hlist_del(¬ifier->link);
2662 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2664 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2666 struct preempt_notifier *notifier;
2668 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2669 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2672 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2674 if (static_key_false(&preempt_notifier_key))
2675 __fire_sched_in_preempt_notifiers(curr);
2679 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2680 struct task_struct *next)
2682 struct preempt_notifier *notifier;
2684 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2685 notifier->ops->sched_out(notifier, next);
2688 static __always_inline void
2689 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2690 struct task_struct *next)
2692 if (static_key_false(&preempt_notifier_key))
2693 __fire_sched_out_preempt_notifiers(curr, next);
2696 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2698 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2703 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2704 struct task_struct *next)
2708 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2711 * prepare_task_switch - prepare to switch tasks
2712 * @rq: the runqueue preparing to switch
2713 * @prev: the current task that is being switched out
2714 * @next: the task we are going to switch to.
2716 * This is called with the rq lock held and interrupts off. It must
2717 * be paired with a subsequent finish_task_switch after the context
2720 * prepare_task_switch sets up locking and calls architecture specific
2724 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2725 struct task_struct *next)
2727 sched_info_switch(rq, prev, next);
2728 perf_event_task_sched_out(prev, next);
2729 fire_sched_out_preempt_notifiers(prev, next);
2730 prepare_lock_switch(rq, next);
2731 prepare_arch_switch(next);
2735 * finish_task_switch - clean up after a task-switch
2736 * @prev: the thread we just switched away from.
2738 * finish_task_switch must be called after the context switch, paired
2739 * with a prepare_task_switch call before the context switch.
2740 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2741 * and do any other architecture-specific cleanup actions.
2743 * Note that we may have delayed dropping an mm in context_switch(). If
2744 * so, we finish that here outside of the runqueue lock. (Doing it
2745 * with the lock held can cause deadlocks; see schedule() for
2748 * The context switch have flipped the stack from under us and restored the
2749 * local variables which were saved when this task called schedule() in the
2750 * past. prev == current is still correct but we need to recalculate this_rq
2751 * because prev may have moved to another CPU.
2753 static struct rq *finish_task_switch(struct task_struct *prev)
2754 __releases(rq->lock)
2756 struct rq *rq = this_rq();
2757 struct mm_struct *mm = rq->prev_mm;
2761 * The previous task will have left us with a preempt_count of 2
2762 * because it left us after:
2765 * preempt_disable(); // 1
2767 * raw_spin_lock_irq(&rq->lock) // 2
2769 * Also, see FORK_PREEMPT_COUNT.
2771 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2772 "corrupted preempt_count: %s/%d/0x%x\n",
2773 current->comm, current->pid, preempt_count()))
2774 preempt_count_set(FORK_PREEMPT_COUNT);
2779 * A task struct has one reference for the use as "current".
2780 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2781 * schedule one last time. The schedule call will never return, and
2782 * the scheduled task must drop that reference.
2784 * We must observe prev->state before clearing prev->on_cpu (in
2785 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2786 * running on another CPU and we could rave with its RUNNING -> DEAD
2787 * transition, resulting in a double drop.
2789 prev_state = prev->state;
2790 vtime_task_switch(prev);
2791 perf_event_task_sched_in(prev, current);
2792 finish_lock_switch(rq, prev);
2793 finish_arch_post_lock_switch();
2795 fire_sched_in_preempt_notifiers(current);
2798 if (unlikely(prev_state == TASK_DEAD)) {
2799 if (prev->sched_class->task_dead)
2800 prev->sched_class->task_dead(prev);
2803 * Remove function-return probe instances associated with this
2804 * task and put them back on the free list.
2806 kprobe_flush_task(prev);
2808 /* Task is done with its stack. */
2809 put_task_stack(prev);
2811 put_task_struct(prev);
2814 tick_nohz_task_switch();
2820 /* rq->lock is NOT held, but preemption is disabled */
2821 static void __balance_callback(struct rq *rq)
2823 struct callback_head *head, *next;
2824 void (*func)(struct rq *rq);
2825 unsigned long flags;
2827 raw_spin_lock_irqsave(&rq->lock, flags);
2828 head = rq->balance_callback;
2829 rq->balance_callback = NULL;
2831 func = (void (*)(struct rq *))head->func;
2838 raw_spin_unlock_irqrestore(&rq->lock, flags);
2841 static inline void balance_callback(struct rq *rq)
2843 if (unlikely(rq->balance_callback))
2844 __balance_callback(rq);
2849 static inline void balance_callback(struct rq *rq)
2856 * schedule_tail - first thing a freshly forked thread must call.
2857 * @prev: the thread we just switched away from.
2859 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2860 __releases(rq->lock)
2865 * New tasks start with FORK_PREEMPT_COUNT, see there and
2866 * finish_task_switch() for details.
2868 * finish_task_switch() will drop rq->lock() and lower preempt_count
2869 * and the preempt_enable() will end up enabling preemption (on
2870 * PREEMPT_COUNT kernels).
2873 rq = finish_task_switch(prev);
2874 balance_callback(rq);
2877 if (current->set_child_tid)
2878 put_user(task_pid_vnr(current), current->set_child_tid);
2882 * context_switch - switch to the new MM and the new thread's register state.
2884 static __always_inline struct rq *
2885 context_switch(struct rq *rq, struct task_struct *prev,
2886 struct task_struct *next, struct rq_flags *rf)
2888 struct mm_struct *mm, *oldmm;
2890 prepare_task_switch(rq, prev, next);
2893 oldmm = prev->active_mm;
2895 * For paravirt, this is coupled with an exit in switch_to to
2896 * combine the page table reload and the switch backend into
2899 arch_start_context_switch(prev);
2902 next->active_mm = oldmm;
2903 atomic_inc(&oldmm->mm_count);
2904 enter_lazy_tlb(oldmm, next);
2906 switch_mm_irqs_off(oldmm, mm, next);
2909 prev->active_mm = NULL;
2910 rq->prev_mm = oldmm;
2913 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2916 * Since the runqueue lock will be released by the next
2917 * task (which is an invalid locking op but in the case
2918 * of the scheduler it's an obvious special-case), so we
2919 * do an early lockdep release here:
2921 rq_unpin_lock(rq, rf);
2922 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2924 /* Here we just switch the register state and the stack. */
2925 switch_to(prev, next, prev);
2928 return finish_task_switch(prev);
2932 * nr_running and nr_context_switches:
2934 * externally visible scheduler statistics: current number of runnable
2935 * threads, total number of context switches performed since bootup.
2937 unsigned long nr_running(void)
2939 unsigned long i, sum = 0;
2941 for_each_online_cpu(i)
2942 sum += cpu_rq(i)->nr_running;
2948 * Check if only the current task is running on the cpu.
2950 * Caution: this function does not check that the caller has disabled
2951 * preemption, thus the result might have a time-of-check-to-time-of-use
2952 * race. The caller is responsible to use it correctly, for example:
2954 * - from a non-preemptable section (of course)
2956 * - from a thread that is bound to a single CPU
2958 * - in a loop with very short iterations (e.g. a polling loop)
2960 bool single_task_running(void)
2962 return raw_rq()->nr_running == 1;
2964 EXPORT_SYMBOL(single_task_running);
2966 unsigned long long nr_context_switches(void)
2969 unsigned long long sum = 0;
2971 for_each_possible_cpu(i)
2972 sum += cpu_rq(i)->nr_switches;
2978 * IO-wait accounting, and how its mostly bollocks (on SMP).
2980 * The idea behind IO-wait account is to account the idle time that we could
2981 * have spend running if it were not for IO. That is, if we were to improve the
2982 * storage performance, we'd have a proportional reduction in IO-wait time.
2984 * This all works nicely on UP, where, when a task blocks on IO, we account
2985 * idle time as IO-wait, because if the storage were faster, it could've been
2986 * running and we'd not be idle.
2988 * This has been extended to SMP, by doing the same for each CPU. This however
2991 * Imagine for instance the case where two tasks block on one CPU, only the one
2992 * CPU will have IO-wait accounted, while the other has regular idle. Even
2993 * though, if the storage were faster, both could've ran at the same time,
2994 * utilising both CPUs.
2996 * This means, that when looking globally, the current IO-wait accounting on
2997 * SMP is a lower bound, by reason of under accounting.
2999 * Worse, since the numbers are provided per CPU, they are sometimes
3000 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
3001 * associated with any one particular CPU, it can wake to another CPU than it
3002 * blocked on. This means the per CPU IO-wait number is meaningless.
3004 * Task CPU affinities can make all that even more 'interesting'.
3007 unsigned long nr_iowait(void)
3009 unsigned long i, sum = 0;
3011 for_each_possible_cpu(i)
3012 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3018 * Consumers of these two interfaces, like for example the cpufreq menu
3019 * governor are using nonsensical data. Boosting frequency for a CPU that has
3020 * IO-wait which might not even end up running the task when it does become
3024 unsigned long nr_iowait_cpu(int cpu)
3026 struct rq *this = cpu_rq(cpu);
3027 return atomic_read(&this->nr_iowait);
3030 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
3032 struct rq *rq = this_rq();
3033 *nr_waiters = atomic_read(&rq->nr_iowait);
3034 *load = rq->load.weight;
3040 * sched_exec - execve() is a valuable balancing opportunity, because at
3041 * this point the task has the smallest effective memory and cache footprint.
3043 void sched_exec(void)
3045 struct task_struct *p = current;
3046 unsigned long flags;
3049 raw_spin_lock_irqsave(&p->pi_lock, flags);
3050 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3051 if (dest_cpu == smp_processor_id())
3054 if (likely(cpu_active(dest_cpu))) {
3055 struct migration_arg arg = { p, dest_cpu };
3057 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3058 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3062 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3067 DEFINE_PER_CPU(struct kernel_stat, kstat);
3068 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3070 EXPORT_PER_CPU_SYMBOL(kstat);
3071 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3074 * The function fair_sched_class.update_curr accesses the struct curr
3075 * and its field curr->exec_start; when called from task_sched_runtime(),
3076 * we observe a high rate of cache misses in practice.
3077 * Prefetching this data results in improved performance.
3079 static inline void prefetch_curr_exec_start(struct task_struct *p)
3081 #ifdef CONFIG_FAIR_GROUP_SCHED
3082 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3084 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3087 prefetch(&curr->exec_start);
3091 * Return accounted runtime for the task.
3092 * In case the task is currently running, return the runtime plus current's
3093 * pending runtime that have not been accounted yet.
3095 unsigned long long task_sched_runtime(struct task_struct *p)
3101 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3103 * 64-bit doesn't need locks to atomically read a 64bit value.
3104 * So we have a optimization chance when the task's delta_exec is 0.
3105 * Reading ->on_cpu is racy, but this is ok.
3107 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3108 * If we race with it entering cpu, unaccounted time is 0. This is
3109 * indistinguishable from the read occurring a few cycles earlier.
3110 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3111 * been accounted, so we're correct here as well.
3113 if (!p->on_cpu || !task_on_rq_queued(p))
3114 return p->se.sum_exec_runtime;
3117 rq = task_rq_lock(p, &rf);
3119 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3120 * project cycles that may never be accounted to this
3121 * thread, breaking clock_gettime().
3123 if (task_current(rq, p) && task_on_rq_queued(p)) {
3124 prefetch_curr_exec_start(p);
3125 update_rq_clock(rq);
3126 p->sched_class->update_curr(rq);
3128 ns = p->se.sum_exec_runtime;
3129 task_rq_unlock(rq, p, &rf);
3135 * This function gets called by the timer code, with HZ frequency.
3136 * We call it with interrupts disabled.
3138 void scheduler_tick(void)
3140 int cpu = smp_processor_id();
3141 struct rq *rq = cpu_rq(cpu);
3142 struct task_struct *curr = rq->curr;
3146 raw_spin_lock(&rq->lock);
3147 update_rq_clock(rq);
3148 curr->sched_class->task_tick(rq, curr, 0);
3149 cpu_load_update_active(rq);
3150 calc_global_load_tick(rq);
3151 raw_spin_unlock(&rq->lock);
3153 perf_event_task_tick();
3156 rq->idle_balance = idle_cpu(cpu);
3157 trigger_load_balance(rq);
3159 rq_last_tick_reset(rq);
3162 #ifdef CONFIG_NO_HZ_FULL
3164 * scheduler_tick_max_deferment
3166 * Keep at least one tick per second when a single
3167 * active task is running because the scheduler doesn't
3168 * yet completely support full dynticks environment.
3170 * This makes sure that uptime, CFS vruntime, load
3171 * balancing, etc... continue to move forward, even
3172 * with a very low granularity.
3174 * Return: Maximum deferment in nanoseconds.
3176 u64 scheduler_tick_max_deferment(void)
3178 struct rq *rq = this_rq();
3179 unsigned long next, now = READ_ONCE(jiffies);
3181 next = rq->last_sched_tick + HZ;
3183 if (time_before_eq(next, now))
3186 return jiffies_to_nsecs(next - now);
3190 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3191 defined(CONFIG_PREEMPT_TRACER))
3193 * If the value passed in is equal to the current preempt count
3194 * then we just disabled preemption. Start timing the latency.
3196 static inline void preempt_latency_start(int val)
3198 if (preempt_count() == val) {
3199 unsigned long ip = get_lock_parent_ip();
3200 #ifdef CONFIG_DEBUG_PREEMPT
3201 current->preempt_disable_ip = ip;
3203 trace_preempt_off(CALLER_ADDR0, ip);
3207 void preempt_count_add(int val)
3209 #ifdef CONFIG_DEBUG_PREEMPT
3213 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3216 __preempt_count_add(val);
3217 #ifdef CONFIG_DEBUG_PREEMPT
3219 * Spinlock count overflowing soon?
3221 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3224 preempt_latency_start(val);
3226 EXPORT_SYMBOL(preempt_count_add);
3227 NOKPROBE_SYMBOL(preempt_count_add);
3230 * If the value passed in equals to the current preempt count
3231 * then we just enabled preemption. Stop timing the latency.
3233 static inline void preempt_latency_stop(int val)
3235 if (preempt_count() == val)
3236 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3239 void preempt_count_sub(int val)
3241 #ifdef CONFIG_DEBUG_PREEMPT
3245 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3248 * Is the spinlock portion underflowing?
3250 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3251 !(preempt_count() & PREEMPT_MASK)))
3255 preempt_latency_stop(val);
3256 __preempt_count_sub(val);
3258 EXPORT_SYMBOL(preempt_count_sub);
3259 NOKPROBE_SYMBOL(preempt_count_sub);
3262 static inline void preempt_latency_start(int val) { }
3263 static inline void preempt_latency_stop(int val) { }
3267 * Print scheduling while atomic bug:
3269 static noinline void __schedule_bug(struct task_struct *prev)
3271 /* Save this before calling printk(), since that will clobber it */
3272 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3274 if (oops_in_progress)
3277 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3278 prev->comm, prev->pid, preempt_count());
3280 debug_show_held_locks(prev);
3282 if (irqs_disabled())
3283 print_irqtrace_events(prev);
3284 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3285 && in_atomic_preempt_off()) {
3286 pr_err("Preemption disabled at:");
3287 print_ip_sym(preempt_disable_ip);
3291 panic("scheduling while atomic\n");
3294 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3298 * Various schedule()-time debugging checks and statistics:
3300 static inline void schedule_debug(struct task_struct *prev)
3302 #ifdef CONFIG_SCHED_STACK_END_CHECK
3303 if (task_stack_end_corrupted(prev))
3304 panic("corrupted stack end detected inside scheduler\n");
3307 if (unlikely(in_atomic_preempt_off())) {
3308 __schedule_bug(prev);
3309 preempt_count_set(PREEMPT_DISABLED);
3313 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3315 schedstat_inc(this_rq()->sched_count);
3319 * Pick up the highest-prio task:
3321 static inline struct task_struct *
3322 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3324 const struct sched_class *class;
3325 struct task_struct *p;
3328 * Optimization: we know that if all tasks are in
3329 * the fair class we can call that function directly:
3331 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3332 p = fair_sched_class.pick_next_task(rq, prev, rf);
3333 if (unlikely(p == RETRY_TASK))
3336 /* assumes fair_sched_class->next == idle_sched_class */
3338 p = idle_sched_class.pick_next_task(rq, prev, rf);
3344 for_each_class(class) {
3345 p = class->pick_next_task(rq, prev, rf);
3347 if (unlikely(p == RETRY_TASK))
3353 BUG(); /* the idle class will always have a runnable task */
3357 * __schedule() is the main scheduler function.
3359 * The main means of driving the scheduler and thus entering this function are:
3361 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3363 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3364 * paths. For example, see arch/x86/entry_64.S.
3366 * To drive preemption between tasks, the scheduler sets the flag in timer
3367 * interrupt handler scheduler_tick().
3369 * 3. Wakeups don't really cause entry into schedule(). They add a
3370 * task to the run-queue and that's it.
3372 * Now, if the new task added to the run-queue preempts the current
3373 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3374 * called on the nearest possible occasion:
3376 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3378 * - in syscall or exception context, at the next outmost
3379 * preempt_enable(). (this might be as soon as the wake_up()'s
3382 * - in IRQ context, return from interrupt-handler to
3383 * preemptible context
3385 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3388 * - cond_resched() call
3389 * - explicit schedule() call
3390 * - return from syscall or exception to user-space
3391 * - return from interrupt-handler to user-space
3393 * WARNING: must be called with preemption disabled!
3395 static void __sched notrace __schedule(bool preempt)
3397 struct task_struct *prev, *next;
3398 unsigned long *switch_count;
3403 cpu = smp_processor_id();
3407 schedule_debug(prev);
3409 if (sched_feat(HRTICK))
3412 local_irq_disable();
3413 rcu_note_context_switch();
3416 * Make sure that signal_pending_state()->signal_pending() below
3417 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3418 * done by the caller to avoid the race with signal_wake_up().
3420 smp_mb__before_spinlock();
3421 raw_spin_lock(&rq->lock);
3422 rq_pin_lock(rq, &rf);
3424 rq->clock_update_flags <<= 1; /* promote REQ to ACT */
3426 switch_count = &prev->nivcsw;
3427 if (!preempt && prev->state) {
3428 if (unlikely(signal_pending_state(prev->state, prev))) {
3429 prev->state = TASK_RUNNING;
3431 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3434 if (prev->in_iowait) {
3435 atomic_inc(&rq->nr_iowait);
3436 delayacct_blkio_start();
3440 * If a worker went to sleep, notify and ask workqueue
3441 * whether it wants to wake up a task to maintain
3444 if (prev->flags & PF_WQ_WORKER) {
3445 struct task_struct *to_wakeup;
3447 to_wakeup = wq_worker_sleeping(prev);
3449 try_to_wake_up_local(to_wakeup, &rf);
3452 switch_count = &prev->nvcsw;
3455 if (task_on_rq_queued(prev))
3456 update_rq_clock(rq);
3458 next = pick_next_task(rq, prev, &rf);
3459 clear_tsk_need_resched(prev);
3460 clear_preempt_need_resched();
3462 if (likely(prev != next)) {
3467 trace_sched_switch(preempt, prev, next);
3468 rq = context_switch(rq, prev, next, &rf); /* unlocks the rq */
3470 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3471 rq_unpin_lock(rq, &rf);
3472 raw_spin_unlock_irq(&rq->lock);
3475 balance_callback(rq);
3478 void __noreturn do_task_dead(void)
3481 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3482 * when the following two conditions become true.
3483 * - There is race condition of mmap_sem (It is acquired by
3485 * - SMI occurs before setting TASK_RUNINNG.
3486 * (or hypervisor of virtual machine switches to other guest)
3487 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3489 * To avoid it, we have to wait for releasing tsk->pi_lock which
3490 * is held by try_to_wake_up()
3493 raw_spin_unlock_wait(¤t->pi_lock);
3495 /* causes final put_task_struct in finish_task_switch(). */
3496 __set_current_state(TASK_DEAD);
3497 current->flags |= PF_NOFREEZE; /* tell freezer to ignore us */
3500 /* Avoid "noreturn function does return". */
3502 cpu_relax(); /* For when BUG is null */
3505 static inline void sched_submit_work(struct task_struct *tsk)
3507 if (!tsk->state || tsk_is_pi_blocked(tsk))
3510 * If we are going to sleep and we have plugged IO queued,
3511 * make sure to submit it to avoid deadlocks.
3513 if (blk_needs_flush_plug(tsk))
3514 blk_schedule_flush_plug(tsk);
3517 asmlinkage __visible void __sched schedule(void)
3519 struct task_struct *tsk = current;
3521 sched_submit_work(tsk);
3525 sched_preempt_enable_no_resched();
3526 } while (need_resched());
3528 EXPORT_SYMBOL(schedule);
3530 #ifdef CONFIG_CONTEXT_TRACKING
3531 asmlinkage __visible void __sched schedule_user(void)
3534 * If we come here after a random call to set_need_resched(),
3535 * or we have been woken up remotely but the IPI has not yet arrived,
3536 * we haven't yet exited the RCU idle mode. Do it here manually until
3537 * we find a better solution.
3539 * NB: There are buggy callers of this function. Ideally we
3540 * should warn if prev_state != CONTEXT_USER, but that will trigger
3541 * too frequently to make sense yet.
3543 enum ctx_state prev_state = exception_enter();
3545 exception_exit(prev_state);
3550 * schedule_preempt_disabled - called with preemption disabled
3552 * Returns with preemption disabled. Note: preempt_count must be 1
3554 void __sched schedule_preempt_disabled(void)
3556 sched_preempt_enable_no_resched();
3561 static void __sched notrace preempt_schedule_common(void)
3565 * Because the function tracer can trace preempt_count_sub()
3566 * and it also uses preempt_enable/disable_notrace(), if
3567 * NEED_RESCHED is set, the preempt_enable_notrace() called
3568 * by the function tracer will call this function again and
3569 * cause infinite recursion.
3571 * Preemption must be disabled here before the function
3572 * tracer can trace. Break up preempt_disable() into two
3573 * calls. One to disable preemption without fear of being
3574 * traced. The other to still record the preemption latency,
3575 * which can also be traced by the function tracer.
3577 preempt_disable_notrace();
3578 preempt_latency_start(1);
3580 preempt_latency_stop(1);
3581 preempt_enable_no_resched_notrace();
3584 * Check again in case we missed a preemption opportunity
3585 * between schedule and now.
3587 } while (need_resched());
3590 #ifdef CONFIG_PREEMPT
3592 * this is the entry point to schedule() from in-kernel preemption
3593 * off of preempt_enable. Kernel preemptions off return from interrupt
3594 * occur there and call schedule directly.
3596 asmlinkage __visible void __sched notrace preempt_schedule(void)
3599 * If there is a non-zero preempt_count or interrupts are disabled,
3600 * we do not want to preempt the current task. Just return..
3602 if (likely(!preemptible()))
3605 preempt_schedule_common();
3607 NOKPROBE_SYMBOL(preempt_schedule);
3608 EXPORT_SYMBOL(preempt_schedule);
3611 * preempt_schedule_notrace - preempt_schedule called by tracing
3613 * The tracing infrastructure uses preempt_enable_notrace to prevent
3614 * recursion and tracing preempt enabling caused by the tracing
3615 * infrastructure itself. But as tracing can happen in areas coming
3616 * from userspace or just about to enter userspace, a preempt enable
3617 * can occur before user_exit() is called. This will cause the scheduler
3618 * to be called when the system is still in usermode.
3620 * To prevent this, the preempt_enable_notrace will use this function
3621 * instead of preempt_schedule() to exit user context if needed before
3622 * calling the scheduler.
3624 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3626 enum ctx_state prev_ctx;
3628 if (likely(!preemptible()))
3633 * Because the function tracer can trace preempt_count_sub()
3634 * and it also uses preempt_enable/disable_notrace(), if
3635 * NEED_RESCHED is set, the preempt_enable_notrace() called
3636 * by the function tracer will call this function again and
3637 * cause infinite recursion.
3639 * Preemption must be disabled here before the function
3640 * tracer can trace. Break up preempt_disable() into two
3641 * calls. One to disable preemption without fear of being
3642 * traced. The other to still record the preemption latency,
3643 * which can also be traced by the function tracer.
3645 preempt_disable_notrace();
3646 preempt_latency_start(1);
3648 * Needs preempt disabled in case user_exit() is traced
3649 * and the tracer calls preempt_enable_notrace() causing
3650 * an infinite recursion.
3652 prev_ctx = exception_enter();
3654 exception_exit(prev_ctx);
3656 preempt_latency_stop(1);
3657 preempt_enable_no_resched_notrace();
3658 } while (need_resched());
3660 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3662 #endif /* CONFIG_PREEMPT */
3665 * this is the entry point to schedule() from kernel preemption
3666 * off of irq context.
3667 * Note, that this is called and return with irqs disabled. This will
3668 * protect us against recursive calling from irq.
3670 asmlinkage __visible void __sched preempt_schedule_irq(void)
3672 enum ctx_state prev_state;
3674 /* Catch callers which need to be fixed */
3675 BUG_ON(preempt_count() || !irqs_disabled());
3677 prev_state = exception_enter();
3683 local_irq_disable();
3684 sched_preempt_enable_no_resched();
3685 } while (need_resched());
3687 exception_exit(prev_state);
3690 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3693 return try_to_wake_up(curr->private, mode, wake_flags);
3695 EXPORT_SYMBOL(default_wake_function);
3697 #ifdef CONFIG_RT_MUTEXES
3700 * rt_mutex_setprio - set the current priority of a task
3702 * @prio: prio value (kernel-internal form)
3704 * This function changes the 'effective' priority of a task. It does
3705 * not touch ->normal_prio like __setscheduler().
3707 * Used by the rt_mutex code to implement priority inheritance
3708 * logic. Call site only calls if the priority of the task changed.
3710 void rt_mutex_setprio(struct task_struct *p, int prio)
3712 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3713 const struct sched_class *prev_class;
3717 BUG_ON(prio > MAX_PRIO);
3719 rq = __task_rq_lock(p, &rf);
3720 update_rq_clock(rq);
3723 * Idle task boosting is a nono in general. There is one
3724 * exception, when PREEMPT_RT and NOHZ is active:
3726 * The idle task calls get_next_timer_interrupt() and holds
3727 * the timer wheel base->lock on the CPU and another CPU wants
3728 * to access the timer (probably to cancel it). We can safely
3729 * ignore the boosting request, as the idle CPU runs this code
3730 * with interrupts disabled and will complete the lock
3731 * protected section without being interrupted. So there is no
3732 * real need to boost.
3734 if (unlikely(p == rq->idle)) {
3735 WARN_ON(p != rq->curr);
3736 WARN_ON(p->pi_blocked_on);
3740 trace_sched_pi_setprio(p, prio);
3743 if (oldprio == prio)
3744 queue_flag &= ~DEQUEUE_MOVE;
3746 prev_class = p->sched_class;
3747 queued = task_on_rq_queued(p);
3748 running = task_current(rq, p);
3750 dequeue_task(rq, p, queue_flag);
3752 put_prev_task(rq, p);
3755 * Boosting condition are:
3756 * 1. -rt task is running and holds mutex A
3757 * --> -dl task blocks on mutex A
3759 * 2. -dl task is running and holds mutex A
3760 * --> -dl task blocks on mutex A and could preempt the
3763 if (dl_prio(prio)) {
3764 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3765 if (!dl_prio(p->normal_prio) ||
3766 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3767 p->dl.dl_boosted = 1;
3768 queue_flag |= ENQUEUE_REPLENISH;
3770 p->dl.dl_boosted = 0;
3771 p->sched_class = &dl_sched_class;
3772 } else if (rt_prio(prio)) {
3773 if (dl_prio(oldprio))
3774 p->dl.dl_boosted = 0;
3776 queue_flag |= ENQUEUE_HEAD;
3777 p->sched_class = &rt_sched_class;
3779 if (dl_prio(oldprio))
3780 p->dl.dl_boosted = 0;
3781 if (rt_prio(oldprio))
3783 p->sched_class = &fair_sched_class;
3789 enqueue_task(rq, p, queue_flag);
3791 set_curr_task(rq, p);
3793 check_class_changed(rq, p, prev_class, oldprio);
3795 preempt_disable(); /* avoid rq from going away on us */
3796 __task_rq_unlock(rq, &rf);
3798 balance_callback(rq);
3803 void set_user_nice(struct task_struct *p, long nice)
3805 bool queued, running;
3806 int old_prio, delta;
3810 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3813 * We have to be careful, if called from sys_setpriority(),
3814 * the task might be in the middle of scheduling on another CPU.
3816 rq = task_rq_lock(p, &rf);
3817 update_rq_clock(rq);
3820 * The RT priorities are set via sched_setscheduler(), but we still
3821 * allow the 'normal' nice value to be set - but as expected
3822 * it wont have any effect on scheduling until the task is
3823 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3825 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3826 p->static_prio = NICE_TO_PRIO(nice);
3829 queued = task_on_rq_queued(p);
3830 running = task_current(rq, p);
3832 dequeue_task(rq, p, DEQUEUE_SAVE);
3834 put_prev_task(rq, p);
3836 p->static_prio = NICE_TO_PRIO(nice);
3839 p->prio = effective_prio(p);
3840 delta = p->prio - old_prio;
3843 enqueue_task(rq, p, ENQUEUE_RESTORE);
3845 * If the task increased its priority or is running and
3846 * lowered its priority, then reschedule its CPU:
3848 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3852 set_curr_task(rq, p);
3854 task_rq_unlock(rq, p, &rf);
3856 EXPORT_SYMBOL(set_user_nice);
3859 * can_nice - check if a task can reduce its nice value
3863 int can_nice(const struct task_struct *p, const int nice)
3865 /* convert nice value [19,-20] to rlimit style value [1,40] */
3866 int nice_rlim = nice_to_rlimit(nice);
3868 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3869 capable(CAP_SYS_NICE));
3872 #ifdef __ARCH_WANT_SYS_NICE
3875 * sys_nice - change the priority of the current process.
3876 * @increment: priority increment
3878 * sys_setpriority is a more generic, but much slower function that
3879 * does similar things.
3881 SYSCALL_DEFINE1(nice, int, increment)
3886 * Setpriority might change our priority at the same moment.
3887 * We don't have to worry. Conceptually one call occurs first
3888 * and we have a single winner.
3890 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3891 nice = task_nice(current) + increment;
3893 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3894 if (increment < 0 && !can_nice(current, nice))
3897 retval = security_task_setnice(current, nice);
3901 set_user_nice(current, nice);
3908 * task_prio - return the priority value of a given task.
3909 * @p: the task in question.
3911 * Return: The priority value as seen by users in /proc.
3912 * RT tasks are offset by -200. Normal tasks are centered
3913 * around 0, value goes from -16 to +15.
3915 int task_prio(const struct task_struct *p)
3917 return p->prio - MAX_RT_PRIO;
3921 * idle_cpu - is a given cpu idle currently?
3922 * @cpu: the processor in question.
3924 * Return: 1 if the CPU is currently idle. 0 otherwise.
3926 int idle_cpu(int cpu)
3928 struct rq *rq = cpu_rq(cpu);
3930 if (rq->curr != rq->idle)
3937 if (!llist_empty(&rq->wake_list))
3945 * idle_task - return the idle task for a given cpu.
3946 * @cpu: the processor in question.
3948 * Return: The idle task for the cpu @cpu.
3950 struct task_struct *idle_task(int cpu)
3952 return cpu_rq(cpu)->idle;
3956 * find_process_by_pid - find a process with a matching PID value.
3957 * @pid: the pid in question.
3959 * The task of @pid, if found. %NULL otherwise.
3961 static struct task_struct *find_process_by_pid(pid_t pid)
3963 return pid ? find_task_by_vpid(pid) : current;
3967 * This function initializes the sched_dl_entity of a newly becoming
3968 * SCHED_DEADLINE task.
3970 * Only the static values are considered here, the actual runtime and the
3971 * absolute deadline will be properly calculated when the task is enqueued
3972 * for the first time with its new policy.
3975 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3977 struct sched_dl_entity *dl_se = &p->dl;
3979 dl_se->dl_runtime = attr->sched_runtime;
3980 dl_se->dl_deadline = attr->sched_deadline;
3981 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3982 dl_se->flags = attr->sched_flags;
3983 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3986 * Changing the parameters of a task is 'tricky' and we're not doing
3987 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3989 * What we SHOULD do is delay the bandwidth release until the 0-lag
3990 * point. This would include retaining the task_struct until that time
3991 * and change dl_overflow() to not immediately decrement the current
3994 * Instead we retain the current runtime/deadline and let the new
3995 * parameters take effect after the current reservation period lapses.
3996 * This is safe (albeit pessimistic) because the 0-lag point is always
3997 * before the current scheduling deadline.
3999 * We can still have temporary overloads because we do not delay the
4000 * change in bandwidth until that time; so admission control is
4001 * not on the safe side. It does however guarantee tasks will never
4002 * consume more than promised.
4007 * sched_setparam() passes in -1 for its policy, to let the functions
4008 * it calls know not to change it.
4010 #define SETPARAM_POLICY -1
4012 static void __setscheduler_params(struct task_struct *p,
4013 const struct sched_attr *attr)
4015 int policy = attr->sched_policy;
4017 if (policy == SETPARAM_POLICY)
4022 if (dl_policy(policy))
4023 __setparam_dl(p, attr);
4024 else if (fair_policy(policy))
4025 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4028 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4029 * !rt_policy. Always setting this ensures that things like
4030 * getparam()/getattr() don't report silly values for !rt tasks.
4032 p->rt_priority = attr->sched_priority;
4033 p->normal_prio = normal_prio(p);
4037 /* Actually do priority change: must hold pi & rq lock. */
4038 static void __setscheduler(struct rq *rq, struct task_struct *p,
4039 const struct sched_attr *attr, bool keep_boost)
4041 __setscheduler_params(p, attr);
4044 * Keep a potential priority boosting if called from
4045 * sched_setscheduler().
4048 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
4050 p->prio = normal_prio(p);
4052 if (dl_prio(p->prio))
4053 p->sched_class = &dl_sched_class;
4054 else if (rt_prio(p->prio))
4055 p->sched_class = &rt_sched_class;
4057 p->sched_class = &fair_sched_class;
4061 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4063 struct sched_dl_entity *dl_se = &p->dl;
4065 attr->sched_priority = p->rt_priority;
4066 attr->sched_runtime = dl_se->dl_runtime;
4067 attr->sched_deadline = dl_se->dl_deadline;
4068 attr->sched_period = dl_se->dl_period;
4069 attr->sched_flags = dl_se->flags;
4073 * This function validates the new parameters of a -deadline task.
4074 * We ask for the deadline not being zero, and greater or equal
4075 * than the runtime, as well as the period of being zero or
4076 * greater than deadline. Furthermore, we have to be sure that
4077 * user parameters are above the internal resolution of 1us (we
4078 * check sched_runtime only since it is always the smaller one) and
4079 * below 2^63 ns (we have to check both sched_deadline and
4080 * sched_period, as the latter can be zero).
4083 __checkparam_dl(const struct sched_attr *attr)
4086 if (attr->sched_deadline == 0)
4090 * Since we truncate DL_SCALE bits, make sure we're at least
4093 if (attr->sched_runtime < (1ULL << DL_SCALE))
4097 * Since we use the MSB for wrap-around and sign issues, make
4098 * sure it's not set (mind that period can be equal to zero).
4100 if (attr->sched_deadline & (1ULL << 63) ||
4101 attr->sched_period & (1ULL << 63))
4104 /* runtime <= deadline <= period (if period != 0) */
4105 if ((attr->sched_period != 0 &&
4106 attr->sched_period < attr->sched_deadline) ||
4107 attr->sched_deadline < attr->sched_runtime)
4114 * check the target process has a UID that matches the current process's
4116 static bool check_same_owner(struct task_struct *p)
4118 const struct cred *cred = current_cred(), *pcred;
4122 pcred = __task_cred(p);
4123 match = (uid_eq(cred->euid, pcred->euid) ||
4124 uid_eq(cred->euid, pcred->uid));
4129 static bool dl_param_changed(struct task_struct *p,
4130 const struct sched_attr *attr)
4132 struct sched_dl_entity *dl_se = &p->dl;
4134 if (dl_se->dl_runtime != attr->sched_runtime ||
4135 dl_se->dl_deadline != attr->sched_deadline ||
4136 dl_se->dl_period != attr->sched_period ||
4137 dl_se->flags != attr->sched_flags)
4143 static int __sched_setscheduler(struct task_struct *p,
4144 const struct sched_attr *attr,
4147 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4148 MAX_RT_PRIO - 1 - attr->sched_priority;
4149 int retval, oldprio, oldpolicy = -1, queued, running;
4150 int new_effective_prio, policy = attr->sched_policy;
4151 const struct sched_class *prev_class;
4154 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4157 /* may grab non-irq protected spin_locks */
4158 BUG_ON(in_interrupt());
4160 /* double check policy once rq lock held */
4162 reset_on_fork = p->sched_reset_on_fork;
4163 policy = oldpolicy = p->policy;
4165 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4167 if (!valid_policy(policy))
4171 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4175 * Valid priorities for SCHED_FIFO and SCHED_RR are
4176 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4177 * SCHED_BATCH and SCHED_IDLE is 0.
4179 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4180 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4182 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4183 (rt_policy(policy) != (attr->sched_priority != 0)))
4187 * Allow unprivileged RT tasks to decrease priority:
4189 if (user && !capable(CAP_SYS_NICE)) {
4190 if (fair_policy(policy)) {
4191 if (attr->sched_nice < task_nice(p) &&
4192 !can_nice(p, attr->sched_nice))
4196 if (rt_policy(policy)) {
4197 unsigned long rlim_rtprio =
4198 task_rlimit(p, RLIMIT_RTPRIO);
4200 /* can't set/change the rt policy */
4201 if (policy != p->policy && !rlim_rtprio)
4204 /* can't increase priority */
4205 if (attr->sched_priority > p->rt_priority &&
4206 attr->sched_priority > rlim_rtprio)
4211 * Can't set/change SCHED_DEADLINE policy at all for now
4212 * (safest behavior); in the future we would like to allow
4213 * unprivileged DL tasks to increase their relative deadline
4214 * or reduce their runtime (both ways reducing utilization)
4216 if (dl_policy(policy))
4220 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4221 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4223 if (idle_policy(p->policy) && !idle_policy(policy)) {
4224 if (!can_nice(p, task_nice(p)))
4228 /* can't change other user's priorities */
4229 if (!check_same_owner(p))
4232 /* Normal users shall not reset the sched_reset_on_fork flag */
4233 if (p->sched_reset_on_fork && !reset_on_fork)
4238 retval = security_task_setscheduler(p);
4244 * make sure no PI-waiters arrive (or leave) while we are
4245 * changing the priority of the task:
4247 * To be able to change p->policy safely, the appropriate
4248 * runqueue lock must be held.
4250 rq = task_rq_lock(p, &rf);
4251 update_rq_clock(rq);
4254 * Changing the policy of the stop threads its a very bad idea
4256 if (p == rq->stop) {
4257 task_rq_unlock(rq, p, &rf);
4262 * If not changing anything there's no need to proceed further,
4263 * but store a possible modification of reset_on_fork.
4265 if (unlikely(policy == p->policy)) {
4266 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4268 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4270 if (dl_policy(policy) && dl_param_changed(p, attr))
4273 p->sched_reset_on_fork = reset_on_fork;
4274 task_rq_unlock(rq, p, &rf);
4280 #ifdef CONFIG_RT_GROUP_SCHED
4282 * Do not allow realtime tasks into groups that have no runtime
4285 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4286 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4287 !task_group_is_autogroup(task_group(p))) {
4288 task_rq_unlock(rq, p, &rf);
4293 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4294 cpumask_t *span = rq->rd->span;
4297 * Don't allow tasks with an affinity mask smaller than
4298 * the entire root_domain to become SCHED_DEADLINE. We
4299 * will also fail if there's no bandwidth available.
4301 if (!cpumask_subset(span, &p->cpus_allowed) ||
4302 rq->rd->dl_bw.bw == 0) {
4303 task_rq_unlock(rq, p, &rf);
4310 /* recheck policy now with rq lock held */
4311 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4312 policy = oldpolicy = -1;
4313 task_rq_unlock(rq, p, &rf);
4318 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4319 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4322 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4323 task_rq_unlock(rq, p, &rf);
4327 p->sched_reset_on_fork = reset_on_fork;
4332 * Take priority boosted tasks into account. If the new
4333 * effective priority is unchanged, we just store the new
4334 * normal parameters and do not touch the scheduler class and
4335 * the runqueue. This will be done when the task deboost
4338 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4339 if (new_effective_prio == oldprio)
4340 queue_flags &= ~DEQUEUE_MOVE;
4343 queued = task_on_rq_queued(p);
4344 running = task_current(rq, p);
4346 dequeue_task(rq, p, queue_flags);
4348 put_prev_task(rq, p);
4350 prev_class = p->sched_class;
4351 __setscheduler(rq, p, attr, pi);
4355 * We enqueue to tail when the priority of a task is
4356 * increased (user space view).
4358 if (oldprio < p->prio)
4359 queue_flags |= ENQUEUE_HEAD;
4361 enqueue_task(rq, p, queue_flags);
4364 set_curr_task(rq, p);
4366 check_class_changed(rq, p, prev_class, oldprio);
4367 preempt_disable(); /* avoid rq from going away on us */
4368 task_rq_unlock(rq, p, &rf);
4371 rt_mutex_adjust_pi(p);
4374 * Run balance callbacks after we've adjusted the PI chain.
4376 balance_callback(rq);
4382 static int _sched_setscheduler(struct task_struct *p, int policy,
4383 const struct sched_param *param, bool check)
4385 struct sched_attr attr = {
4386 .sched_policy = policy,
4387 .sched_priority = param->sched_priority,
4388 .sched_nice = PRIO_TO_NICE(p->static_prio),
4391 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4392 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4393 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4394 policy &= ~SCHED_RESET_ON_FORK;
4395 attr.sched_policy = policy;
4398 return __sched_setscheduler(p, &attr, check, true);
4401 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4402 * @p: the task in question.
4403 * @policy: new policy.
4404 * @param: structure containing the new RT priority.
4406 * Return: 0 on success. An error code otherwise.
4408 * NOTE that the task may be already dead.
4410 int sched_setscheduler(struct task_struct *p, int policy,
4411 const struct sched_param *param)
4413 return _sched_setscheduler(p, policy, param, true);
4415 EXPORT_SYMBOL_GPL(sched_setscheduler);
4417 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4419 return __sched_setscheduler(p, attr, true, true);
4421 EXPORT_SYMBOL_GPL(sched_setattr);
4424 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4425 * @p: the task in question.
4426 * @policy: new policy.
4427 * @param: structure containing the new RT priority.
4429 * Just like sched_setscheduler, only don't bother checking if the
4430 * current context has permission. For example, this is needed in
4431 * stop_machine(): we create temporary high priority worker threads,
4432 * but our caller might not have that capability.
4434 * Return: 0 on success. An error code otherwise.
4436 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4437 const struct sched_param *param)
4439 return _sched_setscheduler(p, policy, param, false);
4441 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4444 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4446 struct sched_param lparam;
4447 struct task_struct *p;
4450 if (!param || pid < 0)
4452 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4457 p = find_process_by_pid(pid);
4459 retval = sched_setscheduler(p, policy, &lparam);
4466 * Mimics kernel/events/core.c perf_copy_attr().
4468 static int sched_copy_attr(struct sched_attr __user *uattr,
4469 struct sched_attr *attr)
4474 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4478 * zero the full structure, so that a short copy will be nice.
4480 memset(attr, 0, sizeof(*attr));
4482 ret = get_user(size, &uattr->size);
4486 if (size > PAGE_SIZE) /* silly large */
4489 if (!size) /* abi compat */
4490 size = SCHED_ATTR_SIZE_VER0;
4492 if (size < SCHED_ATTR_SIZE_VER0)
4496 * If we're handed a bigger struct than we know of,
4497 * ensure all the unknown bits are 0 - i.e. new
4498 * user-space does not rely on any kernel feature
4499 * extensions we dont know about yet.
4501 if (size > sizeof(*attr)) {
4502 unsigned char __user *addr;
4503 unsigned char __user *end;
4506 addr = (void __user *)uattr + sizeof(*attr);
4507 end = (void __user *)uattr + size;
4509 for (; addr < end; addr++) {
4510 ret = get_user(val, addr);
4516 size = sizeof(*attr);
4519 ret = copy_from_user(attr, uattr, size);
4524 * XXX: do we want to be lenient like existing syscalls; or do we want
4525 * to be strict and return an error on out-of-bounds values?
4527 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4532 put_user(sizeof(*attr), &uattr->size);
4537 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4538 * @pid: the pid in question.
4539 * @policy: new policy.
4540 * @param: structure containing the new RT priority.
4542 * Return: 0 on success. An error code otherwise.
4544 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4545 struct sched_param __user *, param)
4547 /* negative values for policy are not valid */
4551 return do_sched_setscheduler(pid, policy, param);
4555 * sys_sched_setparam - set/change the RT priority of a thread
4556 * @pid: the pid in question.
4557 * @param: structure containing the new RT priority.
4559 * Return: 0 on success. An error code otherwise.
4561 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4563 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4567 * sys_sched_setattr - same as above, but with extended sched_attr
4568 * @pid: the pid in question.
4569 * @uattr: structure containing the extended parameters.
4570 * @flags: for future extension.
4572 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4573 unsigned int, flags)
4575 struct sched_attr attr;
4576 struct task_struct *p;
4579 if (!uattr || pid < 0 || flags)
4582 retval = sched_copy_attr(uattr, &attr);
4586 if ((int)attr.sched_policy < 0)
4591 p = find_process_by_pid(pid);
4593 retval = sched_setattr(p, &attr);
4600 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4601 * @pid: the pid in question.
4603 * Return: On success, the policy of the thread. Otherwise, a negative error
4606 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4608 struct task_struct *p;
4616 p = find_process_by_pid(pid);
4618 retval = security_task_getscheduler(p);
4621 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4628 * sys_sched_getparam - get the RT priority of a thread
4629 * @pid: the pid in question.
4630 * @param: structure containing the RT priority.
4632 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4635 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4637 struct sched_param lp = { .sched_priority = 0 };
4638 struct task_struct *p;
4641 if (!param || pid < 0)
4645 p = find_process_by_pid(pid);
4650 retval = security_task_getscheduler(p);
4654 if (task_has_rt_policy(p))
4655 lp.sched_priority = p->rt_priority;
4659 * This one might sleep, we cannot do it with a spinlock held ...
4661 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4670 static int sched_read_attr(struct sched_attr __user *uattr,
4671 struct sched_attr *attr,
4676 if (!access_ok(VERIFY_WRITE, uattr, usize))
4680 * If we're handed a smaller struct than we know of,
4681 * ensure all the unknown bits are 0 - i.e. old
4682 * user-space does not get uncomplete information.
4684 if (usize < sizeof(*attr)) {
4685 unsigned char *addr;
4688 addr = (void *)attr + usize;
4689 end = (void *)attr + sizeof(*attr);
4691 for (; addr < end; addr++) {
4699 ret = copy_to_user(uattr, attr, attr->size);
4707 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4708 * @pid: the pid in question.
4709 * @uattr: structure containing the extended parameters.
4710 * @size: sizeof(attr) for fwd/bwd comp.
4711 * @flags: for future extension.
4713 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4714 unsigned int, size, unsigned int, flags)
4716 struct sched_attr attr = {
4717 .size = sizeof(struct sched_attr),
4719 struct task_struct *p;
4722 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4723 size < SCHED_ATTR_SIZE_VER0 || flags)
4727 p = find_process_by_pid(pid);
4732 retval = security_task_getscheduler(p);
4736 attr.sched_policy = p->policy;
4737 if (p->sched_reset_on_fork)
4738 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4739 if (task_has_dl_policy(p))
4740 __getparam_dl(p, &attr);
4741 else if (task_has_rt_policy(p))
4742 attr.sched_priority = p->rt_priority;
4744 attr.sched_nice = task_nice(p);
4748 retval = sched_read_attr(uattr, &attr, size);
4756 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4758 cpumask_var_t cpus_allowed, new_mask;
4759 struct task_struct *p;
4764 p = find_process_by_pid(pid);
4770 /* Prevent p going away */
4774 if (p->flags & PF_NO_SETAFFINITY) {
4778 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4782 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4784 goto out_free_cpus_allowed;
4787 if (!check_same_owner(p)) {
4789 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4791 goto out_free_new_mask;
4796 retval = security_task_setscheduler(p);
4798 goto out_free_new_mask;
4801 cpuset_cpus_allowed(p, cpus_allowed);
4802 cpumask_and(new_mask, in_mask, cpus_allowed);
4805 * Since bandwidth control happens on root_domain basis,
4806 * if admission test is enabled, we only admit -deadline
4807 * tasks allowed to run on all the CPUs in the task's
4811 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4813 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4816 goto out_free_new_mask;
4822 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4825 cpuset_cpus_allowed(p, cpus_allowed);
4826 if (!cpumask_subset(new_mask, cpus_allowed)) {
4828 * We must have raced with a concurrent cpuset
4829 * update. Just reset the cpus_allowed to the
4830 * cpuset's cpus_allowed
4832 cpumask_copy(new_mask, cpus_allowed);
4837 free_cpumask_var(new_mask);
4838 out_free_cpus_allowed:
4839 free_cpumask_var(cpus_allowed);
4845 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4846 struct cpumask *new_mask)
4848 if (len < cpumask_size())
4849 cpumask_clear(new_mask);
4850 else if (len > cpumask_size())
4851 len = cpumask_size();
4853 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4857 * sys_sched_setaffinity - set the cpu affinity of a process
4858 * @pid: pid of the process
4859 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4860 * @user_mask_ptr: user-space pointer to the new cpu mask
4862 * Return: 0 on success. An error code otherwise.
4864 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4865 unsigned long __user *, user_mask_ptr)
4867 cpumask_var_t new_mask;
4870 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4873 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4875 retval = sched_setaffinity(pid, new_mask);
4876 free_cpumask_var(new_mask);
4880 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4882 struct task_struct *p;
4883 unsigned long flags;
4889 p = find_process_by_pid(pid);
4893 retval = security_task_getscheduler(p);
4897 raw_spin_lock_irqsave(&p->pi_lock, flags);
4898 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4899 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4908 * sys_sched_getaffinity - get the cpu affinity of a process
4909 * @pid: pid of the process
4910 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4911 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4913 * Return: size of CPU mask copied to user_mask_ptr on success. An
4914 * error code otherwise.
4916 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4917 unsigned long __user *, user_mask_ptr)
4922 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4924 if (len & (sizeof(unsigned long)-1))
4927 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4930 ret = sched_getaffinity(pid, mask);
4932 size_t retlen = min_t(size_t, len, cpumask_size());
4934 if (copy_to_user(user_mask_ptr, mask, retlen))
4939 free_cpumask_var(mask);
4945 * sys_sched_yield - yield the current processor to other threads.
4947 * This function yields the current CPU to other tasks. If there are no
4948 * other threads running on this CPU then this function will return.
4952 SYSCALL_DEFINE0(sched_yield)
4954 struct rq *rq = this_rq_lock();
4956 schedstat_inc(rq->yld_count);
4957 current->sched_class->yield_task(rq);
4960 * Since we are going to call schedule() anyway, there's
4961 * no need to preempt or enable interrupts:
4963 __release(rq->lock);
4964 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4965 do_raw_spin_unlock(&rq->lock);
4966 sched_preempt_enable_no_resched();
4973 #ifndef CONFIG_PREEMPT
4974 int __sched _cond_resched(void)
4976 if (should_resched(0)) {
4977 preempt_schedule_common();
4982 EXPORT_SYMBOL(_cond_resched);
4986 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4987 * call schedule, and on return reacquire the lock.
4989 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4990 * operations here to prevent schedule() from being called twice (once via
4991 * spin_unlock(), once by hand).
4993 int __cond_resched_lock(spinlock_t *lock)
4995 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4998 lockdep_assert_held(lock);
5000 if (spin_needbreak(lock) || resched) {
5003 preempt_schedule_common();
5011 EXPORT_SYMBOL(__cond_resched_lock);
5013 int __sched __cond_resched_softirq(void)
5015 BUG_ON(!in_softirq());
5017 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
5019 preempt_schedule_common();
5025 EXPORT_SYMBOL(__cond_resched_softirq);
5028 * yield - yield the current processor to other threads.
5030 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5032 * The scheduler is at all times free to pick the calling task as the most
5033 * eligible task to run, if removing the yield() call from your code breaks
5034 * it, its already broken.
5036 * Typical broken usage is:
5041 * where one assumes that yield() will let 'the other' process run that will
5042 * make event true. If the current task is a SCHED_FIFO task that will never
5043 * happen. Never use yield() as a progress guarantee!!
5045 * If you want to use yield() to wait for something, use wait_event().
5046 * If you want to use yield() to be 'nice' for others, use cond_resched().
5047 * If you still want to use yield(), do not!
5049 void __sched yield(void)
5051 set_current_state(TASK_RUNNING);
5054 EXPORT_SYMBOL(yield);
5057 * yield_to - yield the current processor to another thread in
5058 * your thread group, or accelerate that thread toward the
5059 * processor it's on.
5061 * @preempt: whether task preemption is allowed or not
5063 * It's the caller's job to ensure that the target task struct
5064 * can't go away on us before we can do any checks.
5067 * true (>0) if we indeed boosted the target task.
5068 * false (0) if we failed to boost the target.
5069 * -ESRCH if there's no task to yield to.
5071 int __sched yield_to(struct task_struct *p, bool preempt)
5073 struct task_struct *curr = current;
5074 struct rq *rq, *p_rq;
5075 unsigned long flags;
5078 local_irq_save(flags);
5084 * If we're the only runnable task on the rq and target rq also
5085 * has only one task, there's absolutely no point in yielding.
5087 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5092 double_rq_lock(rq, p_rq);
5093 if (task_rq(p) != p_rq) {
5094 double_rq_unlock(rq, p_rq);
5098 if (!curr->sched_class->yield_to_task)
5101 if (curr->sched_class != p->sched_class)
5104 if (task_running(p_rq, p) || p->state)
5107 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5109 schedstat_inc(rq->yld_count);
5111 * Make p's CPU reschedule; pick_next_entity takes care of
5114 if (preempt && rq != p_rq)
5119 double_rq_unlock(rq, p_rq);
5121 local_irq_restore(flags);
5128 EXPORT_SYMBOL_GPL(yield_to);
5130 int io_schedule_prepare(void)
5132 int old_iowait = current->in_iowait;
5134 current->in_iowait = 1;
5135 blk_schedule_flush_plug(current);
5140 void io_schedule_finish(int token)
5142 current->in_iowait = token;
5146 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5147 * that process accounting knows that this is a task in IO wait state.
5149 long __sched io_schedule_timeout(long timeout)
5154 token = io_schedule_prepare();
5155 ret = schedule_timeout(timeout);
5156 io_schedule_finish(token);
5160 EXPORT_SYMBOL(io_schedule_timeout);
5162 void io_schedule(void)
5166 token = io_schedule_prepare();
5168 io_schedule_finish(token);
5170 EXPORT_SYMBOL(io_schedule);
5173 * sys_sched_get_priority_max - return maximum RT priority.
5174 * @policy: scheduling class.
5176 * Return: On success, this syscall returns the maximum
5177 * rt_priority that can be used by a given scheduling class.
5178 * On failure, a negative error code is returned.
5180 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5187 ret = MAX_USER_RT_PRIO-1;
5189 case SCHED_DEADLINE:
5200 * sys_sched_get_priority_min - return minimum RT priority.
5201 * @policy: scheduling class.
5203 * Return: On success, this syscall returns the minimum
5204 * rt_priority that can be used by a given scheduling class.
5205 * On failure, a negative error code is returned.
5207 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5216 case SCHED_DEADLINE:
5226 * sys_sched_rr_get_interval - return the default timeslice of a process.
5227 * @pid: pid of the process.
5228 * @interval: userspace pointer to the timeslice value.
5230 * this syscall writes the default timeslice value of a given process
5231 * into the user-space timespec buffer. A value of '0' means infinity.
5233 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5236 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5237 struct timespec __user *, interval)
5239 struct task_struct *p;
5240 unsigned int time_slice;
5251 p = find_process_by_pid(pid);
5255 retval = security_task_getscheduler(p);
5259 rq = task_rq_lock(p, &rf);
5261 if (p->sched_class->get_rr_interval)
5262 time_slice = p->sched_class->get_rr_interval(rq, p);
5263 task_rq_unlock(rq, p, &rf);
5266 jiffies_to_timespec(time_slice, &t);
5267 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5275 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5277 void sched_show_task(struct task_struct *p)
5279 unsigned long free = 0;
5281 unsigned long state = p->state;
5283 if (!try_get_task_stack(p))
5286 state = __ffs(state) + 1;
5287 printk(KERN_INFO "%-15.15s %c", p->comm,
5288 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5289 if (state == TASK_RUNNING)
5290 printk(KERN_CONT " running task ");
5291 #ifdef CONFIG_DEBUG_STACK_USAGE
5292 free = stack_not_used(p);
5297 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5299 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5300 task_pid_nr(p), ppid,
5301 (unsigned long)task_thread_info(p)->flags);
5303 print_worker_info(KERN_INFO, p);
5304 show_stack(p, NULL);
5308 void show_state_filter(unsigned long state_filter)
5310 struct task_struct *g, *p;
5312 #if BITS_PER_LONG == 32
5314 " task PC stack pid father\n");
5317 " task PC stack pid father\n");
5320 for_each_process_thread(g, p) {
5322 * reset the NMI-timeout, listing all files on a slow
5323 * console might take a lot of time:
5324 * Also, reset softlockup watchdogs on all CPUs, because
5325 * another CPU might be blocked waiting for us to process
5328 touch_nmi_watchdog();
5329 touch_all_softlockup_watchdogs();
5330 if (!state_filter || (p->state & state_filter))
5334 #ifdef CONFIG_SCHED_DEBUG
5336 sysrq_sched_debug_show();
5340 * Only show locks if all tasks are dumped:
5343 debug_show_all_locks();
5346 void init_idle_bootup_task(struct task_struct *idle)
5348 idle->sched_class = &idle_sched_class;
5352 * init_idle - set up an idle thread for a given CPU
5353 * @idle: task in question
5354 * @cpu: cpu the idle task belongs to
5356 * NOTE: this function does not set the idle thread's NEED_RESCHED
5357 * flag, to make booting more robust.
5359 void init_idle(struct task_struct *idle, int cpu)
5361 struct rq *rq = cpu_rq(cpu);
5362 unsigned long flags;
5364 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5365 raw_spin_lock(&rq->lock);
5367 __sched_fork(0, idle);
5368 idle->state = TASK_RUNNING;
5369 idle->se.exec_start = sched_clock();
5370 idle->flags |= PF_IDLE;
5372 kasan_unpoison_task_stack(idle);
5376 * Its possible that init_idle() gets called multiple times on a task,
5377 * in that case do_set_cpus_allowed() will not do the right thing.
5379 * And since this is boot we can forgo the serialization.
5381 set_cpus_allowed_common(idle, cpumask_of(cpu));
5384 * We're having a chicken and egg problem, even though we are
5385 * holding rq->lock, the cpu isn't yet set to this cpu so the
5386 * lockdep check in task_group() will fail.
5388 * Similar case to sched_fork(). / Alternatively we could
5389 * use task_rq_lock() here and obtain the other rq->lock.
5394 __set_task_cpu(idle, cpu);
5397 rq->curr = rq->idle = idle;
5398 idle->on_rq = TASK_ON_RQ_QUEUED;
5402 raw_spin_unlock(&rq->lock);
5403 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5405 /* Set the preempt count _outside_ the spinlocks! */
5406 init_idle_preempt_count(idle, cpu);
5409 * The idle tasks have their own, simple scheduling class:
5411 idle->sched_class = &idle_sched_class;
5412 ftrace_graph_init_idle_task(idle, cpu);
5413 vtime_init_idle(idle, cpu);
5415 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5419 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5420 const struct cpumask *trial)
5422 int ret = 1, trial_cpus;
5423 struct dl_bw *cur_dl_b;
5424 unsigned long flags;
5426 if (!cpumask_weight(cur))
5429 rcu_read_lock_sched();
5430 cur_dl_b = dl_bw_of(cpumask_any(cur));
5431 trial_cpus = cpumask_weight(trial);
5433 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5434 if (cur_dl_b->bw != -1 &&
5435 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5437 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5438 rcu_read_unlock_sched();
5443 int task_can_attach(struct task_struct *p,
5444 const struct cpumask *cs_cpus_allowed)
5449 * Kthreads which disallow setaffinity shouldn't be moved
5450 * to a new cpuset; we don't want to change their cpu
5451 * affinity and isolating such threads by their set of
5452 * allowed nodes is unnecessary. Thus, cpusets are not
5453 * applicable for such threads. This prevents checking for
5454 * success of set_cpus_allowed_ptr() on all attached tasks
5455 * before cpus_allowed may be changed.
5457 if (p->flags & PF_NO_SETAFFINITY) {
5463 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5465 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5470 unsigned long flags;
5472 rcu_read_lock_sched();
5473 dl_b = dl_bw_of(dest_cpu);
5474 raw_spin_lock_irqsave(&dl_b->lock, flags);
5475 cpus = dl_bw_cpus(dest_cpu);
5476 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5481 * We reserve space for this task in the destination
5482 * root_domain, as we can't fail after this point.
5483 * We will free resources in the source root_domain
5484 * later on (see set_cpus_allowed_dl()).
5486 __dl_add(dl_b, p->dl.dl_bw);
5488 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5489 rcu_read_unlock_sched();
5499 static bool sched_smp_initialized __read_mostly;
5501 #ifdef CONFIG_NUMA_BALANCING
5502 /* Migrate current task p to target_cpu */
5503 int migrate_task_to(struct task_struct *p, int target_cpu)
5505 struct migration_arg arg = { p, target_cpu };
5506 int curr_cpu = task_cpu(p);
5508 if (curr_cpu == target_cpu)
5511 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5514 /* TODO: This is not properly updating schedstats */
5516 trace_sched_move_numa(p, curr_cpu, target_cpu);
5517 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5521 * Requeue a task on a given node and accurately track the number of NUMA
5522 * tasks on the runqueues
5524 void sched_setnuma(struct task_struct *p, int nid)
5526 bool queued, running;
5530 rq = task_rq_lock(p, &rf);
5531 queued = task_on_rq_queued(p);
5532 running = task_current(rq, p);
5535 dequeue_task(rq, p, DEQUEUE_SAVE);
5537 put_prev_task(rq, p);
5539 p->numa_preferred_nid = nid;
5542 enqueue_task(rq, p, ENQUEUE_RESTORE);
5544 set_curr_task(rq, p);
5545 task_rq_unlock(rq, p, &rf);
5547 #endif /* CONFIG_NUMA_BALANCING */
5549 #ifdef CONFIG_HOTPLUG_CPU
5551 * Ensures that the idle task is using init_mm right before its cpu goes
5554 void idle_task_exit(void)
5556 struct mm_struct *mm = current->active_mm;
5558 BUG_ON(cpu_online(smp_processor_id()));
5560 if (mm != &init_mm) {
5561 switch_mm_irqs_off(mm, &init_mm, current);
5562 finish_arch_post_lock_switch();
5568 * Since this CPU is going 'away' for a while, fold any nr_active delta
5569 * we might have. Assumes we're called after migrate_tasks() so that the
5570 * nr_active count is stable. We need to take the teardown thread which
5571 * is calling this into account, so we hand in adjust = 1 to the load
5574 * Also see the comment "Global load-average calculations".
5576 static void calc_load_migrate(struct rq *rq)
5578 long delta = calc_load_fold_active(rq, 1);
5580 atomic_long_add(delta, &calc_load_tasks);
5583 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5587 static const struct sched_class fake_sched_class = {
5588 .put_prev_task = put_prev_task_fake,
5591 static struct task_struct fake_task = {
5593 * Avoid pull_{rt,dl}_task()
5595 .prio = MAX_PRIO + 1,
5596 .sched_class = &fake_sched_class,
5600 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5601 * try_to_wake_up()->select_task_rq().
5603 * Called with rq->lock held even though we'er in stop_machine() and
5604 * there's no concurrency possible, we hold the required locks anyway
5605 * because of lock validation efforts.
5607 static void migrate_tasks(struct rq *dead_rq)
5609 struct rq *rq = dead_rq;
5610 struct task_struct *next, *stop = rq->stop;
5611 struct rq_flags rf, old_rf;
5615 * Fudge the rq selection such that the below task selection loop
5616 * doesn't get stuck on the currently eligible stop task.
5618 * We're currently inside stop_machine() and the rq is either stuck
5619 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5620 * either way we should never end up calling schedule() until we're
5626 * put_prev_task() and pick_next_task() sched
5627 * class method both need to have an up-to-date
5628 * value of rq->clock[_task]
5630 update_rq_clock(rq);
5634 * There's this thread running, bail when that's the only
5637 if (rq->nr_running == 1)
5641 * pick_next_task assumes pinned rq->lock.
5643 rq_pin_lock(rq, &rf);
5644 next = pick_next_task(rq, &fake_task, &rf);
5646 next->sched_class->put_prev_task(rq, next);
5649 * Rules for changing task_struct::cpus_allowed are holding
5650 * both pi_lock and rq->lock, such that holding either
5651 * stabilizes the mask.
5653 * Drop rq->lock is not quite as disastrous as it usually is
5654 * because !cpu_active at this point, which means load-balance
5655 * will not interfere. Also, stop-machine.
5657 rq_unpin_lock(rq, &rf);
5658 raw_spin_unlock(&rq->lock);
5659 raw_spin_lock(&next->pi_lock);
5660 raw_spin_lock(&rq->lock);
5663 * Since we're inside stop-machine, _nothing_ should have
5664 * changed the task, WARN if weird stuff happened, because in
5665 * that case the above rq->lock drop is a fail too.
5667 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5668 raw_spin_unlock(&next->pi_lock);
5673 * __migrate_task() may return with a different
5674 * rq->lock held and a new cookie in 'rf', but we need
5675 * to preserve rf::clock_update_flags for 'dead_rq'.
5679 /* Find suitable destination for @next, with force if needed. */
5680 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5682 rq = __migrate_task(rq, next, dest_cpu);
5683 if (rq != dead_rq) {
5684 raw_spin_unlock(&rq->lock);
5686 raw_spin_lock(&rq->lock);
5689 raw_spin_unlock(&next->pi_lock);
5694 #endif /* CONFIG_HOTPLUG_CPU */
5696 static void set_rq_online(struct rq *rq)
5699 const struct sched_class *class;
5701 cpumask_set_cpu(rq->cpu, rq->rd->online);
5704 for_each_class(class) {
5705 if (class->rq_online)
5706 class->rq_online(rq);
5711 static void set_rq_offline(struct rq *rq)
5714 const struct sched_class *class;
5716 for_each_class(class) {
5717 if (class->rq_offline)
5718 class->rq_offline(rq);
5721 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5726 static void set_cpu_rq_start_time(unsigned int cpu)
5728 struct rq *rq = cpu_rq(cpu);
5730 rq->age_stamp = sched_clock_cpu(cpu);
5733 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5735 #ifdef CONFIG_SCHED_DEBUG
5737 static __read_mostly int sched_debug_enabled;
5739 static int __init sched_debug_setup(char *str)
5741 sched_debug_enabled = 1;
5745 early_param("sched_debug", sched_debug_setup);
5747 static inline bool sched_debug(void)
5749 return sched_debug_enabled;
5752 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5753 struct cpumask *groupmask)
5755 struct sched_group *group = sd->groups;
5757 cpumask_clear(groupmask);
5759 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5761 if (!(sd->flags & SD_LOAD_BALANCE)) {
5762 printk("does not load-balance\n");
5764 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5769 printk(KERN_CONT "span %*pbl level %s\n",
5770 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5772 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5773 printk(KERN_ERR "ERROR: domain->span does not contain "
5776 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5777 printk(KERN_ERR "ERROR: domain->groups does not contain"
5781 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5785 printk(KERN_ERR "ERROR: group is NULL\n");
5789 if (!cpumask_weight(sched_group_cpus(group))) {
5790 printk(KERN_CONT "\n");
5791 printk(KERN_ERR "ERROR: empty group\n");
5795 if (!(sd->flags & SD_OVERLAP) &&
5796 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5797 printk(KERN_CONT "\n");
5798 printk(KERN_ERR "ERROR: repeated CPUs\n");
5802 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5804 printk(KERN_CONT " %*pbl",
5805 cpumask_pr_args(sched_group_cpus(group)));
5806 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5807 printk(KERN_CONT " (cpu_capacity = %lu)",
5808 group->sgc->capacity);
5811 group = group->next;
5812 } while (group != sd->groups);
5813 printk(KERN_CONT "\n");
5815 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5816 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5819 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5820 printk(KERN_ERR "ERROR: parent span is not a superset "
5821 "of domain->span\n");
5825 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5829 if (!sched_debug_enabled)
5833 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5837 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5840 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5848 #else /* !CONFIG_SCHED_DEBUG */
5850 # define sched_debug_enabled 0
5851 # define sched_domain_debug(sd, cpu) do { } while (0)
5852 static inline bool sched_debug(void)
5856 #endif /* CONFIG_SCHED_DEBUG */
5858 static int sd_degenerate(struct sched_domain *sd)
5860 if (cpumask_weight(sched_domain_span(sd)) == 1)
5863 /* Following flags need at least 2 groups */
5864 if (sd->flags & (SD_LOAD_BALANCE |
5865 SD_BALANCE_NEWIDLE |
5868 SD_SHARE_CPUCAPACITY |
5869 SD_ASYM_CPUCAPACITY |
5870 SD_SHARE_PKG_RESOURCES |
5871 SD_SHARE_POWERDOMAIN)) {
5872 if (sd->groups != sd->groups->next)
5876 /* Following flags don't use groups */
5877 if (sd->flags & (SD_WAKE_AFFINE))
5884 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5886 unsigned long cflags = sd->flags, pflags = parent->flags;
5888 if (sd_degenerate(parent))
5891 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5894 /* Flags needing groups don't count if only 1 group in parent */
5895 if (parent->groups == parent->groups->next) {
5896 pflags &= ~(SD_LOAD_BALANCE |
5897 SD_BALANCE_NEWIDLE |
5900 SD_ASYM_CPUCAPACITY |
5901 SD_SHARE_CPUCAPACITY |
5902 SD_SHARE_PKG_RESOURCES |
5904 SD_SHARE_POWERDOMAIN);
5905 if (nr_node_ids == 1)
5906 pflags &= ~SD_SERIALIZE;
5908 if (~cflags & pflags)
5914 static void free_rootdomain(struct rcu_head *rcu)
5916 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5918 cpupri_cleanup(&rd->cpupri);
5919 cpudl_cleanup(&rd->cpudl);
5920 free_cpumask_var(rd->dlo_mask);
5921 free_cpumask_var(rd->rto_mask);
5922 free_cpumask_var(rd->online);
5923 free_cpumask_var(rd->span);
5927 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5929 struct root_domain *old_rd = NULL;
5930 unsigned long flags;
5932 raw_spin_lock_irqsave(&rq->lock, flags);
5937 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5940 cpumask_clear_cpu(rq->cpu, old_rd->span);
5943 * If we dont want to free the old_rd yet then
5944 * set old_rd to NULL to skip the freeing later
5947 if (!atomic_dec_and_test(&old_rd->refcount))
5951 atomic_inc(&rd->refcount);
5954 cpumask_set_cpu(rq->cpu, rd->span);
5955 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5958 raw_spin_unlock_irqrestore(&rq->lock, flags);
5961 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5964 static int init_rootdomain(struct root_domain *rd)
5966 memset(rd, 0, sizeof(*rd));
5968 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5970 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5972 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5974 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5977 init_dl_bw(&rd->dl_bw);
5978 if (cpudl_init(&rd->cpudl) != 0)
5981 if (cpupri_init(&rd->cpupri) != 0)
5986 cpudl_cleanup(&rd->cpudl);
5988 free_cpumask_var(rd->rto_mask);
5990 free_cpumask_var(rd->dlo_mask);
5992 free_cpumask_var(rd->online);
5994 free_cpumask_var(rd->span);
6000 * By default the system creates a single root-domain with all cpus as
6001 * members (mimicking the global state we have today).
6003 struct root_domain def_root_domain;
6005 static void init_defrootdomain(void)
6007 init_rootdomain(&def_root_domain);
6009 atomic_set(&def_root_domain.refcount, 1);
6012 static struct root_domain *alloc_rootdomain(void)
6014 struct root_domain *rd;
6016 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6020 if (init_rootdomain(rd) != 0) {
6028 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6030 struct sched_group *tmp, *first;
6039 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6044 } while (sg != first);
6047 static void destroy_sched_domain(struct sched_domain *sd)
6050 * If its an overlapping domain it has private groups, iterate and
6053 if (sd->flags & SD_OVERLAP) {
6054 free_sched_groups(sd->groups, 1);
6055 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6056 kfree(sd->groups->sgc);
6059 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
6064 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
6066 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6069 struct sched_domain *parent = sd->parent;
6070 destroy_sched_domain(sd);
6075 static void destroy_sched_domains(struct sched_domain *sd)
6078 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
6082 * Keep a special pointer to the highest sched_domain that has
6083 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6084 * allows us to avoid some pointer chasing select_idle_sibling().
6086 * Also keep a unique ID per domain (we use the first cpu number in
6087 * the cpumask of the domain), this allows us to quickly tell if
6088 * two cpus are in the same cache domain, see cpus_share_cache().
6090 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6091 DEFINE_PER_CPU(int, sd_llc_size);
6092 DEFINE_PER_CPU(int, sd_llc_id);
6093 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
6094 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6095 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6097 static void update_top_cache_domain(int cpu)
6099 struct sched_domain_shared *sds = NULL;
6100 struct sched_domain *sd;
6104 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6106 id = cpumask_first(sched_domain_span(sd));
6107 size = cpumask_weight(sched_domain_span(sd));
6111 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6112 per_cpu(sd_llc_size, cpu) = size;
6113 per_cpu(sd_llc_id, cpu) = id;
6114 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
6116 sd = lowest_flag_domain(cpu, SD_NUMA);
6117 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6119 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6120 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6124 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6125 * hold the hotplug lock.
6128 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6130 struct rq *rq = cpu_rq(cpu);
6131 struct sched_domain *tmp;
6133 /* Remove the sched domains which do not contribute to scheduling. */
6134 for (tmp = sd; tmp; ) {
6135 struct sched_domain *parent = tmp->parent;
6139 if (sd_parent_degenerate(tmp, parent)) {
6140 tmp->parent = parent->parent;
6142 parent->parent->child = tmp;
6144 * Transfer SD_PREFER_SIBLING down in case of a
6145 * degenerate parent; the spans match for this
6146 * so the property transfers.
6148 if (parent->flags & SD_PREFER_SIBLING)
6149 tmp->flags |= SD_PREFER_SIBLING;
6150 destroy_sched_domain(parent);
6155 if (sd && sd_degenerate(sd)) {
6158 destroy_sched_domain(tmp);
6163 sched_domain_debug(sd, cpu);
6165 rq_attach_root(rq, rd);
6167 rcu_assign_pointer(rq->sd, sd);
6168 destroy_sched_domains(tmp);
6170 update_top_cache_domain(cpu);
6173 /* Setup the mask of cpus configured for isolated domains */
6174 static int __init isolated_cpu_setup(char *str)
6178 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6179 ret = cpulist_parse(str, cpu_isolated_map);
6181 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6186 __setup("isolcpus=", isolated_cpu_setup);
6189 struct sched_domain ** __percpu sd;
6190 struct root_domain *rd;
6201 * Build an iteration mask that can exclude certain CPUs from the upwards
6204 * Asymmetric node setups can result in situations where the domain tree is of
6205 * unequal depth, make sure to skip domains that already cover the entire
6208 * In that case build_sched_domains() will have terminated the iteration early
6209 * and our sibling sd spans will be empty. Domains should always include the
6210 * cpu they're built on, so check that.
6213 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6215 const struct cpumask *span = sched_domain_span(sd);
6216 struct sd_data *sdd = sd->private;
6217 struct sched_domain *sibling;
6220 for_each_cpu(i, span) {
6221 sibling = *per_cpu_ptr(sdd->sd, i);
6222 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6225 cpumask_set_cpu(i, sched_group_mask(sg));
6230 * Return the canonical balance cpu for this group, this is the first cpu
6231 * of this group that's also in the iteration mask.
6233 int group_balance_cpu(struct sched_group *sg)
6235 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6239 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6241 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6242 const struct cpumask *span = sched_domain_span(sd);
6243 struct cpumask *covered = sched_domains_tmpmask;
6244 struct sd_data *sdd = sd->private;
6245 struct sched_domain *sibling;
6248 cpumask_clear(covered);
6250 for_each_cpu(i, span) {
6251 struct cpumask *sg_span;
6253 if (cpumask_test_cpu(i, covered))
6256 sibling = *per_cpu_ptr(sdd->sd, i);
6258 /* See the comment near build_group_mask(). */
6259 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6262 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6263 GFP_KERNEL, cpu_to_node(cpu));
6268 sg_span = sched_group_cpus(sg);
6270 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6272 cpumask_set_cpu(i, sg_span);
6274 cpumask_or(covered, covered, sg_span);
6276 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6277 if (atomic_inc_return(&sg->sgc->ref) == 1)
6278 build_group_mask(sd, sg);
6281 * Initialize sgc->capacity such that even if we mess up the
6282 * domains and no possible iteration will get us here, we won't
6285 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6286 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6289 * Make sure the first group of this domain contains the
6290 * canonical balance cpu. Otherwise the sched_domain iteration
6291 * breaks. See update_sg_lb_stats().
6293 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6294 group_balance_cpu(sg) == cpu)
6304 sd->groups = groups;
6309 free_sched_groups(first, 0);
6314 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6316 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6317 struct sched_domain *child = sd->child;
6320 cpu = cpumask_first(sched_domain_span(child));
6323 *sg = *per_cpu_ptr(sdd->sg, cpu);
6324 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6325 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6332 * build_sched_groups will build a circular linked list of the groups
6333 * covered by the given span, and will set each group's ->cpumask correctly,
6334 * and ->cpu_capacity to 0.
6336 * Assumes the sched_domain tree is fully constructed
6339 build_sched_groups(struct sched_domain *sd, int cpu)
6341 struct sched_group *first = NULL, *last = NULL;
6342 struct sd_data *sdd = sd->private;
6343 const struct cpumask *span = sched_domain_span(sd);
6344 struct cpumask *covered;
6347 get_group(cpu, sdd, &sd->groups);
6348 atomic_inc(&sd->groups->ref);
6350 if (cpu != cpumask_first(span))
6353 lockdep_assert_held(&sched_domains_mutex);
6354 covered = sched_domains_tmpmask;
6356 cpumask_clear(covered);
6358 for_each_cpu(i, span) {
6359 struct sched_group *sg;
6362 if (cpumask_test_cpu(i, covered))
6365 group = get_group(i, sdd, &sg);
6366 cpumask_setall(sched_group_mask(sg));
6368 for_each_cpu(j, span) {
6369 if (get_group(j, sdd, NULL) != group)
6372 cpumask_set_cpu(j, covered);
6373 cpumask_set_cpu(j, sched_group_cpus(sg));
6388 * Initialize sched groups cpu_capacity.
6390 * cpu_capacity indicates the capacity of sched group, which is used while
6391 * distributing the load between different sched groups in a sched domain.
6392 * Typically cpu_capacity for all the groups in a sched domain will be same
6393 * unless there are asymmetries in the topology. If there are asymmetries,
6394 * group having more cpu_capacity will pickup more load compared to the
6395 * group having less cpu_capacity.
6397 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6399 struct sched_group *sg = sd->groups;
6404 int cpu, max_cpu = -1;
6406 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6408 if (!(sd->flags & SD_ASYM_PACKING))
6411 for_each_cpu(cpu, sched_group_cpus(sg)) {
6414 else if (sched_asym_prefer(cpu, max_cpu))
6417 sg->asym_prefer_cpu = max_cpu;
6421 } while (sg != sd->groups);
6423 if (cpu != group_balance_cpu(sg))
6426 update_group_capacity(sd, cpu);
6430 * Initializers for schedule domains
6431 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6434 static int default_relax_domain_level = -1;
6435 int sched_domain_level_max;
6437 static int __init setup_relax_domain_level(char *str)
6439 if (kstrtoint(str, 0, &default_relax_domain_level))
6440 pr_warn("Unable to set relax_domain_level\n");
6444 __setup("relax_domain_level=", setup_relax_domain_level);
6446 static void set_domain_attribute(struct sched_domain *sd,
6447 struct sched_domain_attr *attr)
6451 if (!attr || attr->relax_domain_level < 0) {
6452 if (default_relax_domain_level < 0)
6455 request = default_relax_domain_level;
6457 request = attr->relax_domain_level;
6458 if (request < sd->level) {
6459 /* turn off idle balance on this domain */
6460 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6462 /* turn on idle balance on this domain */
6463 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6467 static void __sdt_free(const struct cpumask *cpu_map);
6468 static int __sdt_alloc(const struct cpumask *cpu_map);
6470 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6471 const struct cpumask *cpu_map)
6475 if (!atomic_read(&d->rd->refcount))
6476 free_rootdomain(&d->rd->rcu); /* fall through */
6478 free_percpu(d->sd); /* fall through */
6480 __sdt_free(cpu_map); /* fall through */
6486 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6487 const struct cpumask *cpu_map)
6489 memset(d, 0, sizeof(*d));
6491 if (__sdt_alloc(cpu_map))
6492 return sa_sd_storage;
6493 d->sd = alloc_percpu(struct sched_domain *);
6495 return sa_sd_storage;
6496 d->rd = alloc_rootdomain();
6499 return sa_rootdomain;
6503 * NULL the sd_data elements we've used to build the sched_domain and
6504 * sched_group structure so that the subsequent __free_domain_allocs()
6505 * will not free the data we're using.
6507 static void claim_allocations(int cpu, struct sched_domain *sd)
6509 struct sd_data *sdd = sd->private;
6511 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6512 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6514 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
6515 *per_cpu_ptr(sdd->sds, cpu) = NULL;
6517 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6518 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6520 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6521 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6525 static int sched_domains_numa_levels;
6526 enum numa_topology_type sched_numa_topology_type;
6527 static int *sched_domains_numa_distance;
6528 int sched_max_numa_distance;
6529 static struct cpumask ***sched_domains_numa_masks;
6530 static int sched_domains_curr_level;
6534 * SD_flags allowed in topology descriptions.
6536 * These flags are purely descriptive of the topology and do not prescribe
6537 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6540 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6541 * SD_SHARE_PKG_RESOURCES - describes shared caches
6542 * SD_NUMA - describes NUMA topologies
6543 * SD_SHARE_POWERDOMAIN - describes shared power domain
6544 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6546 * Odd one out, which beside describing the topology has a quirk also
6547 * prescribes the desired behaviour that goes along with it:
6549 * SD_ASYM_PACKING - describes SMT quirks
6551 #define TOPOLOGY_SD_FLAGS \
6552 (SD_SHARE_CPUCAPACITY | \
6553 SD_SHARE_PKG_RESOURCES | \
6556 SD_ASYM_CPUCAPACITY | \
6557 SD_SHARE_POWERDOMAIN)
6559 static struct sched_domain *
6560 sd_init(struct sched_domain_topology_level *tl,
6561 const struct cpumask *cpu_map,
6562 struct sched_domain *child, int cpu)
6564 struct sd_data *sdd = &tl->data;
6565 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6566 int sd_id, sd_weight, sd_flags = 0;
6570 * Ugly hack to pass state to sd_numa_mask()...
6572 sched_domains_curr_level = tl->numa_level;
6575 sd_weight = cpumask_weight(tl->mask(cpu));
6578 sd_flags = (*tl->sd_flags)();
6579 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6580 "wrong sd_flags in topology description\n"))
6581 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6583 *sd = (struct sched_domain){
6584 .min_interval = sd_weight,
6585 .max_interval = 2*sd_weight,
6587 .imbalance_pct = 125,
6589 .cache_nice_tries = 0,
6596 .flags = 1*SD_LOAD_BALANCE
6597 | 1*SD_BALANCE_NEWIDLE
6602 | 0*SD_SHARE_CPUCAPACITY
6603 | 0*SD_SHARE_PKG_RESOURCES
6605 | 0*SD_PREFER_SIBLING
6610 .last_balance = jiffies,
6611 .balance_interval = sd_weight,
6613 .max_newidle_lb_cost = 0,
6614 .next_decay_max_lb_cost = jiffies,
6616 #ifdef CONFIG_SCHED_DEBUG
6621 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6622 sd_id = cpumask_first(sched_domain_span(sd));
6625 * Convert topological properties into behaviour.
6628 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6629 struct sched_domain *t = sd;
6631 for_each_lower_domain(t)
6632 t->flags |= SD_BALANCE_WAKE;
6635 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6636 sd->flags |= SD_PREFER_SIBLING;
6637 sd->imbalance_pct = 110;
6638 sd->smt_gain = 1178; /* ~15% */
6640 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6641 sd->imbalance_pct = 117;
6642 sd->cache_nice_tries = 1;
6646 } else if (sd->flags & SD_NUMA) {
6647 sd->cache_nice_tries = 2;
6651 sd->flags |= SD_SERIALIZE;
6652 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6653 sd->flags &= ~(SD_BALANCE_EXEC |
6660 sd->flags |= SD_PREFER_SIBLING;
6661 sd->cache_nice_tries = 1;
6667 * For all levels sharing cache; connect a sched_domain_shared
6670 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6671 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
6672 atomic_inc(&sd->shared->ref);
6673 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
6682 * Topology list, bottom-up.
6684 static struct sched_domain_topology_level default_topology[] = {
6685 #ifdef CONFIG_SCHED_SMT
6686 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6688 #ifdef CONFIG_SCHED_MC
6689 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6691 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6695 static struct sched_domain_topology_level *sched_domain_topology =
6698 #define for_each_sd_topology(tl) \
6699 for (tl = sched_domain_topology; tl->mask; tl++)
6701 void set_sched_topology(struct sched_domain_topology_level *tl)
6703 if (WARN_ON_ONCE(sched_smp_initialized))
6706 sched_domain_topology = tl;
6711 static const struct cpumask *sd_numa_mask(int cpu)
6713 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6716 static void sched_numa_warn(const char *str)
6718 static int done = false;
6726 printk(KERN_WARNING "ERROR: %s\n\n", str);
6728 for (i = 0; i < nr_node_ids; i++) {
6729 printk(KERN_WARNING " ");
6730 for (j = 0; j < nr_node_ids; j++)
6731 printk(KERN_CONT "%02d ", node_distance(i,j));
6732 printk(KERN_CONT "\n");
6734 printk(KERN_WARNING "\n");
6737 bool find_numa_distance(int distance)
6741 if (distance == node_distance(0, 0))
6744 for (i = 0; i < sched_domains_numa_levels; i++) {
6745 if (sched_domains_numa_distance[i] == distance)
6753 * A system can have three types of NUMA topology:
6754 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6755 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6756 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6758 * The difference between a glueless mesh topology and a backplane
6759 * topology lies in whether communication between not directly
6760 * connected nodes goes through intermediary nodes (where programs
6761 * could run), or through backplane controllers. This affects
6762 * placement of programs.
6764 * The type of topology can be discerned with the following tests:
6765 * - If the maximum distance between any nodes is 1 hop, the system
6766 * is directly connected.
6767 * - If for two nodes A and B, located N > 1 hops away from each other,
6768 * there is an intermediary node C, which is < N hops away from both
6769 * nodes A and B, the system is a glueless mesh.
6771 static void init_numa_topology_type(void)
6775 n = sched_max_numa_distance;
6777 if (sched_domains_numa_levels <= 1) {
6778 sched_numa_topology_type = NUMA_DIRECT;
6782 for_each_online_node(a) {
6783 for_each_online_node(b) {
6784 /* Find two nodes furthest removed from each other. */
6785 if (node_distance(a, b) < n)
6788 /* Is there an intermediary node between a and b? */
6789 for_each_online_node(c) {
6790 if (node_distance(a, c) < n &&
6791 node_distance(b, c) < n) {
6792 sched_numa_topology_type =
6798 sched_numa_topology_type = NUMA_BACKPLANE;
6804 static void sched_init_numa(void)
6806 int next_distance, curr_distance = node_distance(0, 0);
6807 struct sched_domain_topology_level *tl;
6811 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6812 if (!sched_domains_numa_distance)
6816 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6817 * unique distances in the node_distance() table.
6819 * Assumes node_distance(0,j) includes all distances in
6820 * node_distance(i,j) in order to avoid cubic time.
6822 next_distance = curr_distance;
6823 for (i = 0; i < nr_node_ids; i++) {
6824 for (j = 0; j < nr_node_ids; j++) {
6825 for (k = 0; k < nr_node_ids; k++) {
6826 int distance = node_distance(i, k);
6828 if (distance > curr_distance &&
6829 (distance < next_distance ||
6830 next_distance == curr_distance))
6831 next_distance = distance;
6834 * While not a strong assumption it would be nice to know
6835 * about cases where if node A is connected to B, B is not
6836 * equally connected to A.
6838 if (sched_debug() && node_distance(k, i) != distance)
6839 sched_numa_warn("Node-distance not symmetric");
6841 if (sched_debug() && i && !find_numa_distance(distance))
6842 sched_numa_warn("Node-0 not representative");
6844 if (next_distance != curr_distance) {
6845 sched_domains_numa_distance[level++] = next_distance;
6846 sched_domains_numa_levels = level;
6847 curr_distance = next_distance;
6852 * In case of sched_debug() we verify the above assumption.
6862 * 'level' contains the number of unique distances, excluding the
6863 * identity distance node_distance(i,i).
6865 * The sched_domains_numa_distance[] array includes the actual distance
6870 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6871 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6872 * the array will contain less then 'level' members. This could be
6873 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6874 * in other functions.
6876 * We reset it to 'level' at the end of this function.
6878 sched_domains_numa_levels = 0;
6880 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6881 if (!sched_domains_numa_masks)
6885 * Now for each level, construct a mask per node which contains all
6886 * cpus of nodes that are that many hops away from us.
6888 for (i = 0; i < level; i++) {
6889 sched_domains_numa_masks[i] =
6890 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6891 if (!sched_domains_numa_masks[i])
6894 for (j = 0; j < nr_node_ids; j++) {
6895 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6899 sched_domains_numa_masks[i][j] = mask;
6902 if (node_distance(j, k) > sched_domains_numa_distance[i])
6905 cpumask_or(mask, mask, cpumask_of_node(k));
6910 /* Compute default topology size */
6911 for (i = 0; sched_domain_topology[i].mask; i++);
6913 tl = kzalloc((i + level + 1) *
6914 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6919 * Copy the default topology bits..
6921 for (i = 0; sched_domain_topology[i].mask; i++)
6922 tl[i] = sched_domain_topology[i];
6925 * .. and append 'j' levels of NUMA goodness.
6927 for (j = 0; j < level; i++, j++) {
6928 tl[i] = (struct sched_domain_topology_level){
6929 .mask = sd_numa_mask,
6930 .sd_flags = cpu_numa_flags,
6931 .flags = SDTL_OVERLAP,
6937 sched_domain_topology = tl;
6939 sched_domains_numa_levels = level;
6940 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6942 init_numa_topology_type();
6945 static void sched_domains_numa_masks_set(unsigned int cpu)
6947 int node = cpu_to_node(cpu);
6950 for (i = 0; i < sched_domains_numa_levels; i++) {
6951 for (j = 0; j < nr_node_ids; j++) {
6952 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6953 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6958 static void sched_domains_numa_masks_clear(unsigned int cpu)
6962 for (i = 0; i < sched_domains_numa_levels; i++) {
6963 for (j = 0; j < nr_node_ids; j++)
6964 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6969 static inline void sched_init_numa(void) { }
6970 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6971 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6972 #endif /* CONFIG_NUMA */
6974 static int __sdt_alloc(const struct cpumask *cpu_map)
6976 struct sched_domain_topology_level *tl;
6979 for_each_sd_topology(tl) {
6980 struct sd_data *sdd = &tl->data;
6982 sdd->sd = alloc_percpu(struct sched_domain *);
6986 sdd->sds = alloc_percpu(struct sched_domain_shared *);
6990 sdd->sg = alloc_percpu(struct sched_group *);
6994 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6998 for_each_cpu(j, cpu_map) {
6999 struct sched_domain *sd;
7000 struct sched_domain_shared *sds;
7001 struct sched_group *sg;
7002 struct sched_group_capacity *sgc;
7004 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7005 GFP_KERNEL, cpu_to_node(j));
7009 *per_cpu_ptr(sdd->sd, j) = sd;
7011 sds = kzalloc_node(sizeof(struct sched_domain_shared),
7012 GFP_KERNEL, cpu_to_node(j));
7016 *per_cpu_ptr(sdd->sds, j) = sds;
7018 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7019 GFP_KERNEL, cpu_to_node(j));
7025 *per_cpu_ptr(sdd->sg, j) = sg;
7027 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7028 GFP_KERNEL, cpu_to_node(j));
7032 *per_cpu_ptr(sdd->sgc, j) = sgc;
7039 static void __sdt_free(const struct cpumask *cpu_map)
7041 struct sched_domain_topology_level *tl;
7044 for_each_sd_topology(tl) {
7045 struct sd_data *sdd = &tl->data;
7047 for_each_cpu(j, cpu_map) {
7048 struct sched_domain *sd;
7051 sd = *per_cpu_ptr(sdd->sd, j);
7052 if (sd && (sd->flags & SD_OVERLAP))
7053 free_sched_groups(sd->groups, 0);
7054 kfree(*per_cpu_ptr(sdd->sd, j));
7058 kfree(*per_cpu_ptr(sdd->sds, j));
7060 kfree(*per_cpu_ptr(sdd->sg, j));
7062 kfree(*per_cpu_ptr(sdd->sgc, j));
7064 free_percpu(sdd->sd);
7066 free_percpu(sdd->sds);
7068 free_percpu(sdd->sg);
7070 free_percpu(sdd->sgc);
7075 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7076 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7077 struct sched_domain *child, int cpu)
7079 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
7082 sd->level = child->level + 1;
7083 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7086 if (!cpumask_subset(sched_domain_span(child),
7087 sched_domain_span(sd))) {
7088 pr_err("BUG: arch topology borken\n");
7089 #ifdef CONFIG_SCHED_DEBUG
7090 pr_err(" the %s domain not a subset of the %s domain\n",
7091 child->name, sd->name);
7093 /* Fixup, ensure @sd has at least @child cpus. */
7094 cpumask_or(sched_domain_span(sd),
7095 sched_domain_span(sd),
7096 sched_domain_span(child));
7100 set_domain_attribute(sd, attr);
7106 * Build sched domains for a given set of cpus and attach the sched domains
7107 * to the individual cpus
7109 static int build_sched_domains(const struct cpumask *cpu_map,
7110 struct sched_domain_attr *attr)
7112 enum s_alloc alloc_state;
7113 struct sched_domain *sd;
7115 struct rq *rq = NULL;
7116 int i, ret = -ENOMEM;
7118 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7119 if (alloc_state != sa_rootdomain)
7122 /* Set up domains for cpus specified by the cpu_map. */
7123 for_each_cpu(i, cpu_map) {
7124 struct sched_domain_topology_level *tl;
7127 for_each_sd_topology(tl) {
7128 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7129 if (tl == sched_domain_topology)
7130 *per_cpu_ptr(d.sd, i) = sd;
7131 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7132 sd->flags |= SD_OVERLAP;
7133 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7138 /* Build the groups for the domains */
7139 for_each_cpu(i, cpu_map) {
7140 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7141 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7142 if (sd->flags & SD_OVERLAP) {
7143 if (build_overlap_sched_groups(sd, i))
7146 if (build_sched_groups(sd, i))
7152 /* Calculate CPU capacity for physical packages and nodes */
7153 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7154 if (!cpumask_test_cpu(i, cpu_map))
7157 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7158 claim_allocations(i, sd);
7159 init_sched_groups_capacity(i, sd);
7163 /* Attach the domains */
7165 for_each_cpu(i, cpu_map) {
7167 sd = *per_cpu_ptr(d.sd, i);
7169 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7170 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
7171 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
7173 cpu_attach_domain(sd, d.rd, i);
7177 if (rq && sched_debug_enabled) {
7178 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7179 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
7184 __free_domain_allocs(&d, alloc_state, cpu_map);
7188 static cpumask_var_t *doms_cur; /* current sched domains */
7189 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7190 static struct sched_domain_attr *dattr_cur;
7191 /* attribues of custom domains in 'doms_cur' */
7194 * Special case: If a kmalloc of a doms_cur partition (array of
7195 * cpumask) fails, then fallback to a single sched domain,
7196 * as determined by the single cpumask fallback_doms.
7198 static cpumask_var_t fallback_doms;
7201 * arch_update_cpu_topology lets virtualized architectures update the
7202 * cpu core maps. It is supposed to return 1 if the topology changed
7203 * or 0 if it stayed the same.
7205 int __weak arch_update_cpu_topology(void)
7210 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7213 cpumask_var_t *doms;
7215 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7218 for (i = 0; i < ndoms; i++) {
7219 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7220 free_sched_domains(doms, i);
7227 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7230 for (i = 0; i < ndoms; i++)
7231 free_cpumask_var(doms[i]);
7236 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7237 * For now this just excludes isolated cpus, but could be used to
7238 * exclude other special cases in the future.
7240 static int init_sched_domains(const struct cpumask *cpu_map)
7244 arch_update_cpu_topology();
7246 doms_cur = alloc_sched_domains(ndoms_cur);
7248 doms_cur = &fallback_doms;
7249 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7250 err = build_sched_domains(doms_cur[0], NULL);
7251 register_sched_domain_sysctl();
7257 * Detach sched domains from a group of cpus specified in cpu_map
7258 * These cpus will now be attached to the NULL domain
7260 static void detach_destroy_domains(const struct cpumask *cpu_map)
7265 for_each_cpu(i, cpu_map)
7266 cpu_attach_domain(NULL, &def_root_domain, i);
7270 /* handle null as "default" */
7271 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7272 struct sched_domain_attr *new, int idx_new)
7274 struct sched_domain_attr tmp;
7281 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7282 new ? (new + idx_new) : &tmp,
7283 sizeof(struct sched_domain_attr));
7287 * Partition sched domains as specified by the 'ndoms_new'
7288 * cpumasks in the array doms_new[] of cpumasks. This compares
7289 * doms_new[] to the current sched domain partitioning, doms_cur[].
7290 * It destroys each deleted domain and builds each new domain.
7292 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7293 * The masks don't intersect (don't overlap.) We should setup one
7294 * sched domain for each mask. CPUs not in any of the cpumasks will
7295 * not be load balanced. If the same cpumask appears both in the
7296 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7299 * The passed in 'doms_new' should be allocated using
7300 * alloc_sched_domains. This routine takes ownership of it and will
7301 * free_sched_domains it when done with it. If the caller failed the
7302 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7303 * and partition_sched_domains() will fallback to the single partition
7304 * 'fallback_doms', it also forces the domains to be rebuilt.
7306 * If doms_new == NULL it will be replaced with cpu_online_mask.
7307 * ndoms_new == 0 is a special case for destroying existing domains,
7308 * and it will not create the default domain.
7310 * Call with hotplug lock held
7312 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7313 struct sched_domain_attr *dattr_new)
7318 mutex_lock(&sched_domains_mutex);
7320 /* always unregister in case we don't destroy any domains */
7321 unregister_sched_domain_sysctl();
7323 /* Let architecture update cpu core mappings. */
7324 new_topology = arch_update_cpu_topology();
7326 n = doms_new ? ndoms_new : 0;
7328 /* Destroy deleted domains */
7329 for (i = 0; i < ndoms_cur; i++) {
7330 for (j = 0; j < n && !new_topology; j++) {
7331 if (cpumask_equal(doms_cur[i], doms_new[j])
7332 && dattrs_equal(dattr_cur, i, dattr_new, j))
7335 /* no match - a current sched domain not in new doms_new[] */
7336 detach_destroy_domains(doms_cur[i]);
7342 if (doms_new == NULL) {
7344 doms_new = &fallback_doms;
7345 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7346 WARN_ON_ONCE(dattr_new);
7349 /* Build new domains */
7350 for (i = 0; i < ndoms_new; i++) {
7351 for (j = 0; j < n && !new_topology; j++) {
7352 if (cpumask_equal(doms_new[i], doms_cur[j])
7353 && dattrs_equal(dattr_new, i, dattr_cur, j))
7356 /* no match - add a new doms_new */
7357 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7362 /* Remember the new sched domains */
7363 if (doms_cur != &fallback_doms)
7364 free_sched_domains(doms_cur, ndoms_cur);
7365 kfree(dattr_cur); /* kfree(NULL) is safe */
7366 doms_cur = doms_new;
7367 dattr_cur = dattr_new;
7368 ndoms_cur = ndoms_new;
7370 register_sched_domain_sysctl();
7372 mutex_unlock(&sched_domains_mutex);
7375 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7378 * Update cpusets according to cpu_active mask. If cpusets are
7379 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7380 * around partition_sched_domains().
7382 * If we come here as part of a suspend/resume, don't touch cpusets because we
7383 * want to restore it back to its original state upon resume anyway.
7385 static void cpuset_cpu_active(void)
7387 if (cpuhp_tasks_frozen) {
7389 * num_cpus_frozen tracks how many CPUs are involved in suspend
7390 * resume sequence. As long as this is not the last online
7391 * operation in the resume sequence, just build a single sched
7392 * domain, ignoring cpusets.
7395 if (likely(num_cpus_frozen)) {
7396 partition_sched_domains(1, NULL, NULL);
7400 * This is the last CPU online operation. So fall through and
7401 * restore the original sched domains by considering the
7402 * cpuset configurations.
7405 cpuset_update_active_cpus(true);
7408 static int cpuset_cpu_inactive(unsigned int cpu)
7410 unsigned long flags;
7415 if (!cpuhp_tasks_frozen) {
7416 rcu_read_lock_sched();
7417 dl_b = dl_bw_of(cpu);
7419 raw_spin_lock_irqsave(&dl_b->lock, flags);
7420 cpus = dl_bw_cpus(cpu);
7421 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7422 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7424 rcu_read_unlock_sched();
7428 cpuset_update_active_cpus(false);
7431 partition_sched_domains(1, NULL, NULL);
7436 int sched_cpu_activate(unsigned int cpu)
7438 struct rq *rq = cpu_rq(cpu);
7439 unsigned long flags;
7441 set_cpu_active(cpu, true);
7443 if (sched_smp_initialized) {
7444 sched_domains_numa_masks_set(cpu);
7445 cpuset_cpu_active();
7449 * Put the rq online, if not already. This happens:
7451 * 1) In the early boot process, because we build the real domains
7452 * after all cpus have been brought up.
7454 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7457 raw_spin_lock_irqsave(&rq->lock, flags);
7459 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7462 raw_spin_unlock_irqrestore(&rq->lock, flags);
7464 update_max_interval();
7469 int sched_cpu_deactivate(unsigned int cpu)
7473 set_cpu_active(cpu, false);
7475 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7476 * users of this state to go away such that all new such users will
7479 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7480 * not imply sync_sched(), so wait for both.
7482 * Do sync before park smpboot threads to take care the rcu boost case.
7484 if (IS_ENABLED(CONFIG_PREEMPT))
7485 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7489 if (!sched_smp_initialized)
7492 ret = cpuset_cpu_inactive(cpu);
7494 set_cpu_active(cpu, true);
7497 sched_domains_numa_masks_clear(cpu);
7501 static void sched_rq_cpu_starting(unsigned int cpu)
7503 struct rq *rq = cpu_rq(cpu);
7505 rq->calc_load_update = calc_load_update;
7506 update_max_interval();
7509 int sched_cpu_starting(unsigned int cpu)
7511 set_cpu_rq_start_time(cpu);
7512 sched_rq_cpu_starting(cpu);
7516 #ifdef CONFIG_HOTPLUG_CPU
7517 int sched_cpu_dying(unsigned int cpu)
7519 struct rq *rq = cpu_rq(cpu);
7520 unsigned long flags;
7522 /* Handle pending wakeups and then migrate everything off */
7523 sched_ttwu_pending();
7524 raw_spin_lock_irqsave(&rq->lock, flags);
7526 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7530 BUG_ON(rq->nr_running != 1);
7531 raw_spin_unlock_irqrestore(&rq->lock, flags);
7532 calc_load_migrate(rq);
7533 update_max_interval();
7534 nohz_balance_exit_idle(cpu);
7540 #ifdef CONFIG_SCHED_SMT
7541 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7543 static void sched_init_smt(void)
7546 * We've enumerated all CPUs and will assume that if any CPU
7547 * has SMT siblings, CPU0 will too.
7549 if (cpumask_weight(cpu_smt_mask(0)) > 1)
7550 static_branch_enable(&sched_smt_present);
7553 static inline void sched_init_smt(void) { }
7556 void __init sched_init_smp(void)
7558 cpumask_var_t non_isolated_cpus;
7560 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7561 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7566 * There's no userspace yet to cause hotplug operations; hence all the
7567 * cpu masks are stable and all blatant races in the below code cannot
7570 mutex_lock(&sched_domains_mutex);
7571 init_sched_domains(cpu_active_mask);
7572 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7573 if (cpumask_empty(non_isolated_cpus))
7574 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7575 mutex_unlock(&sched_domains_mutex);
7577 /* Move init over to a non-isolated CPU */
7578 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7580 sched_init_granularity();
7581 free_cpumask_var(non_isolated_cpus);
7583 init_sched_rt_class();
7584 init_sched_dl_class();
7587 sched_clock_init_late();
7589 sched_smp_initialized = true;
7592 static int __init migration_init(void)
7594 sched_rq_cpu_starting(smp_processor_id());
7597 early_initcall(migration_init);
7600 void __init sched_init_smp(void)
7602 sched_init_granularity();
7603 sched_clock_init_late();
7605 #endif /* CONFIG_SMP */
7607 int in_sched_functions(unsigned long addr)
7609 return in_lock_functions(addr) ||
7610 (addr >= (unsigned long)__sched_text_start
7611 && addr < (unsigned long)__sched_text_end);
7614 #ifdef CONFIG_CGROUP_SCHED
7616 * Default task group.
7617 * Every task in system belongs to this group at bootup.
7619 struct task_group root_task_group;
7620 LIST_HEAD(task_groups);
7622 /* Cacheline aligned slab cache for task_group */
7623 static struct kmem_cache *task_group_cache __read_mostly;
7626 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7627 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7629 #define WAIT_TABLE_BITS 8
7630 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7631 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
7633 wait_queue_head_t *bit_waitqueue(void *word, int bit)
7635 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
7636 unsigned long val = (unsigned long)word << shift | bit;
7638 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
7640 EXPORT_SYMBOL(bit_waitqueue);
7642 void __init sched_init(void)
7645 unsigned long alloc_size = 0, ptr;
7649 for (i = 0; i < WAIT_TABLE_SIZE; i++)
7650 init_waitqueue_head(bit_wait_table + i);
7652 #ifdef CONFIG_FAIR_GROUP_SCHED
7653 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7655 #ifdef CONFIG_RT_GROUP_SCHED
7656 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7659 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7661 #ifdef CONFIG_FAIR_GROUP_SCHED
7662 root_task_group.se = (struct sched_entity **)ptr;
7663 ptr += nr_cpu_ids * sizeof(void **);
7665 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7666 ptr += nr_cpu_ids * sizeof(void **);
7668 #endif /* CONFIG_FAIR_GROUP_SCHED */
7669 #ifdef CONFIG_RT_GROUP_SCHED
7670 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7671 ptr += nr_cpu_ids * sizeof(void **);
7673 root_task_group.rt_rq = (struct rt_rq **)ptr;
7674 ptr += nr_cpu_ids * sizeof(void **);
7676 #endif /* CONFIG_RT_GROUP_SCHED */
7678 #ifdef CONFIG_CPUMASK_OFFSTACK
7679 for_each_possible_cpu(i) {
7680 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7681 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7682 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7683 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7685 #endif /* CONFIG_CPUMASK_OFFSTACK */
7687 init_rt_bandwidth(&def_rt_bandwidth,
7688 global_rt_period(), global_rt_runtime());
7689 init_dl_bandwidth(&def_dl_bandwidth,
7690 global_rt_period(), global_rt_runtime());
7693 init_defrootdomain();
7696 #ifdef CONFIG_RT_GROUP_SCHED
7697 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7698 global_rt_period(), global_rt_runtime());
7699 #endif /* CONFIG_RT_GROUP_SCHED */
7701 #ifdef CONFIG_CGROUP_SCHED
7702 task_group_cache = KMEM_CACHE(task_group, 0);
7704 list_add(&root_task_group.list, &task_groups);
7705 INIT_LIST_HEAD(&root_task_group.children);
7706 INIT_LIST_HEAD(&root_task_group.siblings);
7707 autogroup_init(&init_task);
7708 #endif /* CONFIG_CGROUP_SCHED */
7710 for_each_possible_cpu(i) {
7714 raw_spin_lock_init(&rq->lock);
7716 rq->calc_load_active = 0;
7717 rq->calc_load_update = jiffies + LOAD_FREQ;
7718 init_cfs_rq(&rq->cfs);
7719 init_rt_rq(&rq->rt);
7720 init_dl_rq(&rq->dl);
7721 #ifdef CONFIG_FAIR_GROUP_SCHED
7722 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7723 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7724 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7726 * How much cpu bandwidth does root_task_group get?
7728 * In case of task-groups formed thr' the cgroup filesystem, it
7729 * gets 100% of the cpu resources in the system. This overall
7730 * system cpu resource is divided among the tasks of
7731 * root_task_group and its child task-groups in a fair manner,
7732 * based on each entity's (task or task-group's) weight
7733 * (se->load.weight).
7735 * In other words, if root_task_group has 10 tasks of weight
7736 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7737 * then A0's share of the cpu resource is:
7739 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7741 * We achieve this by letting root_task_group's tasks sit
7742 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7744 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7745 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7746 #endif /* CONFIG_FAIR_GROUP_SCHED */
7748 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7749 #ifdef CONFIG_RT_GROUP_SCHED
7750 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7753 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7754 rq->cpu_load[j] = 0;
7759 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7760 rq->balance_callback = NULL;
7761 rq->active_balance = 0;
7762 rq->next_balance = jiffies;
7767 rq->avg_idle = 2*sysctl_sched_migration_cost;
7768 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7770 INIT_LIST_HEAD(&rq->cfs_tasks);
7772 rq_attach_root(rq, &def_root_domain);
7773 #ifdef CONFIG_NO_HZ_COMMON
7774 rq->last_load_update_tick = jiffies;
7777 #ifdef CONFIG_NO_HZ_FULL
7778 rq->last_sched_tick = 0;
7780 #endif /* CONFIG_SMP */
7782 atomic_set(&rq->nr_iowait, 0);
7785 set_load_weight(&init_task);
7788 * The boot idle thread does lazy MMU switching as well:
7790 atomic_inc(&init_mm.mm_count);
7791 enter_lazy_tlb(&init_mm, current);
7794 * Make us the idle thread. Technically, schedule() should not be
7795 * called from this thread, however somewhere below it might be,
7796 * but because we are the idle thread, we just pick up running again
7797 * when this runqueue becomes "idle".
7799 init_idle(current, smp_processor_id());
7801 calc_load_update = jiffies + LOAD_FREQ;
7804 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7805 /* May be allocated at isolcpus cmdline parse time */
7806 if (cpu_isolated_map == NULL)
7807 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7808 idle_thread_set_boot_cpu();
7809 set_cpu_rq_start_time(smp_processor_id());
7811 init_sched_fair_class();
7815 scheduler_running = 1;
7818 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7819 static inline int preempt_count_equals(int preempt_offset)
7821 int nested = preempt_count() + rcu_preempt_depth();
7823 return (nested == preempt_offset);
7826 void __might_sleep(const char *file, int line, int preempt_offset)
7829 * Blocking primitives will set (and therefore destroy) current->state,
7830 * since we will exit with TASK_RUNNING make sure we enter with it,
7831 * otherwise we will destroy state.
7833 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7834 "do not call blocking ops when !TASK_RUNNING; "
7835 "state=%lx set at [<%p>] %pS\n",
7837 (void *)current->task_state_change,
7838 (void *)current->task_state_change);
7840 ___might_sleep(file, line, preempt_offset);
7842 EXPORT_SYMBOL(__might_sleep);
7844 void ___might_sleep(const char *file, int line, int preempt_offset)
7846 static unsigned long prev_jiffy; /* ratelimiting */
7847 unsigned long preempt_disable_ip;
7849 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7850 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7851 !is_idle_task(current)) ||
7852 system_state != SYSTEM_RUNNING || oops_in_progress)
7854 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7856 prev_jiffy = jiffies;
7858 /* Save this before calling printk(), since that will clobber it */
7859 preempt_disable_ip = get_preempt_disable_ip(current);
7862 "BUG: sleeping function called from invalid context at %s:%d\n",
7865 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7866 in_atomic(), irqs_disabled(),
7867 current->pid, current->comm);
7869 if (task_stack_end_corrupted(current))
7870 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7872 debug_show_held_locks(current);
7873 if (irqs_disabled())
7874 print_irqtrace_events(current);
7875 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7876 && !preempt_count_equals(preempt_offset)) {
7877 pr_err("Preemption disabled at:");
7878 print_ip_sym(preempt_disable_ip);
7882 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7884 EXPORT_SYMBOL(___might_sleep);
7887 #ifdef CONFIG_MAGIC_SYSRQ
7888 void normalize_rt_tasks(void)
7890 struct task_struct *g, *p;
7891 struct sched_attr attr = {
7892 .sched_policy = SCHED_NORMAL,
7895 read_lock(&tasklist_lock);
7896 for_each_process_thread(g, p) {
7898 * Only normalize user tasks:
7900 if (p->flags & PF_KTHREAD)
7903 p->se.exec_start = 0;
7904 schedstat_set(p->se.statistics.wait_start, 0);
7905 schedstat_set(p->se.statistics.sleep_start, 0);
7906 schedstat_set(p->se.statistics.block_start, 0);
7908 if (!dl_task(p) && !rt_task(p)) {
7910 * Renice negative nice level userspace
7913 if (task_nice(p) < 0)
7914 set_user_nice(p, 0);
7918 __sched_setscheduler(p, &attr, false, false);
7920 read_unlock(&tasklist_lock);
7923 #endif /* CONFIG_MAGIC_SYSRQ */
7925 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7927 * These functions are only useful for the IA64 MCA handling, or kdb.
7929 * They can only be called when the whole system has been
7930 * stopped - every CPU needs to be quiescent, and no scheduling
7931 * activity can take place. Using them for anything else would
7932 * be a serious bug, and as a result, they aren't even visible
7933 * under any other configuration.
7937 * curr_task - return the current task for a given cpu.
7938 * @cpu: the processor in question.
7940 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7942 * Return: The current task for @cpu.
7944 struct task_struct *curr_task(int cpu)
7946 return cpu_curr(cpu);
7949 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7953 * set_curr_task - set the current task for a given cpu.
7954 * @cpu: the processor in question.
7955 * @p: the task pointer to set.
7957 * Description: This function must only be used when non-maskable interrupts
7958 * are serviced on a separate stack. It allows the architecture to switch the
7959 * notion of the current task on a cpu in a non-blocking manner. This function
7960 * must be called with all CPU's synchronized, and interrupts disabled, the
7961 * and caller must save the original value of the current task (see
7962 * curr_task() above) and restore that value before reenabling interrupts and
7963 * re-starting the system.
7965 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7967 void ia64_set_curr_task(int cpu, struct task_struct *p)
7974 #ifdef CONFIG_CGROUP_SCHED
7975 /* task_group_lock serializes the addition/removal of task groups */
7976 static DEFINE_SPINLOCK(task_group_lock);
7978 static void sched_free_group(struct task_group *tg)
7980 free_fair_sched_group(tg);
7981 free_rt_sched_group(tg);
7983 kmem_cache_free(task_group_cache, tg);
7986 /* allocate runqueue etc for a new task group */
7987 struct task_group *sched_create_group(struct task_group *parent)
7989 struct task_group *tg;
7991 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7993 return ERR_PTR(-ENOMEM);
7995 if (!alloc_fair_sched_group(tg, parent))
7998 if (!alloc_rt_sched_group(tg, parent))
8004 sched_free_group(tg);
8005 return ERR_PTR(-ENOMEM);
8008 void sched_online_group(struct task_group *tg, struct task_group *parent)
8010 unsigned long flags;
8012 spin_lock_irqsave(&task_group_lock, flags);
8013 list_add_rcu(&tg->list, &task_groups);
8015 WARN_ON(!parent); /* root should already exist */
8017 tg->parent = parent;
8018 INIT_LIST_HEAD(&tg->children);
8019 list_add_rcu(&tg->siblings, &parent->children);
8020 spin_unlock_irqrestore(&task_group_lock, flags);
8022 online_fair_sched_group(tg);
8025 /* rcu callback to free various structures associated with a task group */
8026 static void sched_free_group_rcu(struct rcu_head *rhp)
8028 /* now it should be safe to free those cfs_rqs */
8029 sched_free_group(container_of(rhp, struct task_group, rcu));
8032 void sched_destroy_group(struct task_group *tg)
8034 /* wait for possible concurrent references to cfs_rqs complete */
8035 call_rcu(&tg->rcu, sched_free_group_rcu);
8038 void sched_offline_group(struct task_group *tg)
8040 unsigned long flags;
8042 /* end participation in shares distribution */
8043 unregister_fair_sched_group(tg);
8045 spin_lock_irqsave(&task_group_lock, flags);
8046 list_del_rcu(&tg->list);
8047 list_del_rcu(&tg->siblings);
8048 spin_unlock_irqrestore(&task_group_lock, flags);
8051 static void sched_change_group(struct task_struct *tsk, int type)
8053 struct task_group *tg;
8056 * All callers are synchronized by task_rq_lock(); we do not use RCU
8057 * which is pointless here. Thus, we pass "true" to task_css_check()
8058 * to prevent lockdep warnings.
8060 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8061 struct task_group, css);
8062 tg = autogroup_task_group(tsk, tg);
8063 tsk->sched_task_group = tg;
8065 #ifdef CONFIG_FAIR_GROUP_SCHED
8066 if (tsk->sched_class->task_change_group)
8067 tsk->sched_class->task_change_group(tsk, type);
8070 set_task_rq(tsk, task_cpu(tsk));
8074 * Change task's runqueue when it moves between groups.
8076 * The caller of this function should have put the task in its new group by
8077 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8080 void sched_move_task(struct task_struct *tsk)
8082 int queued, running;
8086 rq = task_rq_lock(tsk, &rf);
8087 update_rq_clock(rq);
8089 running = task_current(rq, tsk);
8090 queued = task_on_rq_queued(tsk);
8093 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
8094 if (unlikely(running))
8095 put_prev_task(rq, tsk);
8097 sched_change_group(tsk, TASK_MOVE_GROUP);
8100 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
8101 if (unlikely(running))
8102 set_curr_task(rq, tsk);
8104 task_rq_unlock(rq, tsk, &rf);
8106 #endif /* CONFIG_CGROUP_SCHED */
8108 #ifdef CONFIG_RT_GROUP_SCHED
8110 * Ensure that the real time constraints are schedulable.
8112 static DEFINE_MUTEX(rt_constraints_mutex);
8114 /* Must be called with tasklist_lock held */
8115 static inline int tg_has_rt_tasks(struct task_group *tg)
8117 struct task_struct *g, *p;
8120 * Autogroups do not have RT tasks; see autogroup_create().
8122 if (task_group_is_autogroup(tg))
8125 for_each_process_thread(g, p) {
8126 if (rt_task(p) && task_group(p) == tg)
8133 struct rt_schedulable_data {
8134 struct task_group *tg;
8139 static int tg_rt_schedulable(struct task_group *tg, void *data)
8141 struct rt_schedulable_data *d = data;
8142 struct task_group *child;
8143 unsigned long total, sum = 0;
8144 u64 period, runtime;
8146 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8147 runtime = tg->rt_bandwidth.rt_runtime;
8150 period = d->rt_period;
8151 runtime = d->rt_runtime;
8155 * Cannot have more runtime than the period.
8157 if (runtime > period && runtime != RUNTIME_INF)
8161 * Ensure we don't starve existing RT tasks.
8163 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8166 total = to_ratio(period, runtime);
8169 * Nobody can have more than the global setting allows.
8171 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8175 * The sum of our children's runtime should not exceed our own.
8177 list_for_each_entry_rcu(child, &tg->children, siblings) {
8178 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8179 runtime = child->rt_bandwidth.rt_runtime;
8181 if (child == d->tg) {
8182 period = d->rt_period;
8183 runtime = d->rt_runtime;
8186 sum += to_ratio(period, runtime);
8195 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8199 struct rt_schedulable_data data = {
8201 .rt_period = period,
8202 .rt_runtime = runtime,
8206 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8212 static int tg_set_rt_bandwidth(struct task_group *tg,
8213 u64 rt_period, u64 rt_runtime)
8218 * Disallowing the root group RT runtime is BAD, it would disallow the
8219 * kernel creating (and or operating) RT threads.
8221 if (tg == &root_task_group && rt_runtime == 0)
8224 /* No period doesn't make any sense. */
8228 mutex_lock(&rt_constraints_mutex);
8229 read_lock(&tasklist_lock);
8230 err = __rt_schedulable(tg, rt_period, rt_runtime);
8234 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8235 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8236 tg->rt_bandwidth.rt_runtime = rt_runtime;
8238 for_each_possible_cpu(i) {
8239 struct rt_rq *rt_rq = tg->rt_rq[i];
8241 raw_spin_lock(&rt_rq->rt_runtime_lock);
8242 rt_rq->rt_runtime = rt_runtime;
8243 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8245 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8247 read_unlock(&tasklist_lock);
8248 mutex_unlock(&rt_constraints_mutex);
8253 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8255 u64 rt_runtime, rt_period;
8257 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8258 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8259 if (rt_runtime_us < 0)
8260 rt_runtime = RUNTIME_INF;
8262 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8265 static long sched_group_rt_runtime(struct task_group *tg)
8269 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8272 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8273 do_div(rt_runtime_us, NSEC_PER_USEC);
8274 return rt_runtime_us;
8277 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8279 u64 rt_runtime, rt_period;
8281 rt_period = rt_period_us * NSEC_PER_USEC;
8282 rt_runtime = tg->rt_bandwidth.rt_runtime;
8284 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8287 static long sched_group_rt_period(struct task_group *tg)
8291 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8292 do_div(rt_period_us, NSEC_PER_USEC);
8293 return rt_period_us;
8295 #endif /* CONFIG_RT_GROUP_SCHED */
8297 #ifdef CONFIG_RT_GROUP_SCHED
8298 static int sched_rt_global_constraints(void)
8302 mutex_lock(&rt_constraints_mutex);
8303 read_lock(&tasklist_lock);
8304 ret = __rt_schedulable(NULL, 0, 0);
8305 read_unlock(&tasklist_lock);
8306 mutex_unlock(&rt_constraints_mutex);
8311 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8313 /* Don't accept realtime tasks when there is no way for them to run */
8314 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8320 #else /* !CONFIG_RT_GROUP_SCHED */
8321 static int sched_rt_global_constraints(void)
8323 unsigned long flags;
8326 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8327 for_each_possible_cpu(i) {
8328 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8330 raw_spin_lock(&rt_rq->rt_runtime_lock);
8331 rt_rq->rt_runtime = global_rt_runtime();
8332 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8334 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8338 #endif /* CONFIG_RT_GROUP_SCHED */
8340 static int sched_dl_global_validate(void)
8342 u64 runtime = global_rt_runtime();
8343 u64 period = global_rt_period();
8344 u64 new_bw = to_ratio(period, runtime);
8347 unsigned long flags;
8350 * Here we want to check the bandwidth not being set to some
8351 * value smaller than the currently allocated bandwidth in
8352 * any of the root_domains.
8354 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8355 * cycling on root_domains... Discussion on different/better
8356 * solutions is welcome!
8358 for_each_possible_cpu(cpu) {
8359 rcu_read_lock_sched();
8360 dl_b = dl_bw_of(cpu);
8362 raw_spin_lock_irqsave(&dl_b->lock, flags);
8363 if (new_bw < dl_b->total_bw)
8365 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8367 rcu_read_unlock_sched();
8376 static void sched_dl_do_global(void)
8381 unsigned long flags;
8383 def_dl_bandwidth.dl_period = global_rt_period();
8384 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8386 if (global_rt_runtime() != RUNTIME_INF)
8387 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8390 * FIXME: As above...
8392 for_each_possible_cpu(cpu) {
8393 rcu_read_lock_sched();
8394 dl_b = dl_bw_of(cpu);
8396 raw_spin_lock_irqsave(&dl_b->lock, flags);
8398 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8400 rcu_read_unlock_sched();
8404 static int sched_rt_global_validate(void)
8406 if (sysctl_sched_rt_period <= 0)
8409 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8410 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8416 static void sched_rt_do_global(void)
8418 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8419 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8422 int sched_rt_handler(struct ctl_table *table, int write,
8423 void __user *buffer, size_t *lenp,
8426 int old_period, old_runtime;
8427 static DEFINE_MUTEX(mutex);
8431 old_period = sysctl_sched_rt_period;
8432 old_runtime = sysctl_sched_rt_runtime;
8434 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8436 if (!ret && write) {
8437 ret = sched_rt_global_validate();
8441 ret = sched_dl_global_validate();
8445 ret = sched_rt_global_constraints();
8449 sched_rt_do_global();
8450 sched_dl_do_global();
8454 sysctl_sched_rt_period = old_period;
8455 sysctl_sched_rt_runtime = old_runtime;
8457 mutex_unlock(&mutex);
8462 int sched_rr_handler(struct ctl_table *table, int write,
8463 void __user *buffer, size_t *lenp,
8467 static DEFINE_MUTEX(mutex);
8470 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8471 /* make sure that internally we keep jiffies */
8472 /* also, writing zero resets timeslice to default */
8473 if (!ret && write) {
8474 sched_rr_timeslice =
8475 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
8476 msecs_to_jiffies(sysctl_sched_rr_timeslice);
8478 mutex_unlock(&mutex);
8482 #ifdef CONFIG_CGROUP_SCHED
8484 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8486 return css ? container_of(css, struct task_group, css) : NULL;
8489 static struct cgroup_subsys_state *
8490 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8492 struct task_group *parent = css_tg(parent_css);
8493 struct task_group *tg;
8496 /* This is early initialization for the top cgroup */
8497 return &root_task_group.css;
8500 tg = sched_create_group(parent);
8502 return ERR_PTR(-ENOMEM);
8504 sched_online_group(tg, parent);
8509 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8511 struct task_group *tg = css_tg(css);
8513 sched_offline_group(tg);
8516 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8518 struct task_group *tg = css_tg(css);
8521 * Relies on the RCU grace period between css_released() and this.
8523 sched_free_group(tg);
8527 * This is called before wake_up_new_task(), therefore we really only
8528 * have to set its group bits, all the other stuff does not apply.
8530 static void cpu_cgroup_fork(struct task_struct *task)
8535 rq = task_rq_lock(task, &rf);
8537 update_rq_clock(rq);
8538 sched_change_group(task, TASK_SET_GROUP);
8540 task_rq_unlock(rq, task, &rf);
8543 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8545 struct task_struct *task;
8546 struct cgroup_subsys_state *css;
8549 cgroup_taskset_for_each(task, css, tset) {
8550 #ifdef CONFIG_RT_GROUP_SCHED
8551 if (!sched_rt_can_attach(css_tg(css), task))
8554 /* We don't support RT-tasks being in separate groups */
8555 if (task->sched_class != &fair_sched_class)
8559 * Serialize against wake_up_new_task() such that if its
8560 * running, we're sure to observe its full state.
8562 raw_spin_lock_irq(&task->pi_lock);
8564 * Avoid calling sched_move_task() before wake_up_new_task()
8565 * has happened. This would lead to problems with PELT, due to
8566 * move wanting to detach+attach while we're not attached yet.
8568 if (task->state == TASK_NEW)
8570 raw_spin_unlock_irq(&task->pi_lock);
8578 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8580 struct task_struct *task;
8581 struct cgroup_subsys_state *css;
8583 cgroup_taskset_for_each(task, css, tset)
8584 sched_move_task(task);
8587 #ifdef CONFIG_FAIR_GROUP_SCHED
8588 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8589 struct cftype *cftype, u64 shareval)
8591 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8594 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8597 struct task_group *tg = css_tg(css);
8599 return (u64) scale_load_down(tg->shares);
8602 #ifdef CONFIG_CFS_BANDWIDTH
8603 static DEFINE_MUTEX(cfs_constraints_mutex);
8605 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8606 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8608 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8610 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8612 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8613 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8615 if (tg == &root_task_group)
8619 * Ensure we have at some amount of bandwidth every period. This is
8620 * to prevent reaching a state of large arrears when throttled via
8621 * entity_tick() resulting in prolonged exit starvation.
8623 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8627 * Likewise, bound things on the otherside by preventing insane quota
8628 * periods. This also allows us to normalize in computing quota
8631 if (period > max_cfs_quota_period)
8635 * Prevent race between setting of cfs_rq->runtime_enabled and
8636 * unthrottle_offline_cfs_rqs().
8639 mutex_lock(&cfs_constraints_mutex);
8640 ret = __cfs_schedulable(tg, period, quota);
8644 runtime_enabled = quota != RUNTIME_INF;
8645 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8647 * If we need to toggle cfs_bandwidth_used, off->on must occur
8648 * before making related changes, and on->off must occur afterwards
8650 if (runtime_enabled && !runtime_was_enabled)
8651 cfs_bandwidth_usage_inc();
8652 raw_spin_lock_irq(&cfs_b->lock);
8653 cfs_b->period = ns_to_ktime(period);
8654 cfs_b->quota = quota;
8656 __refill_cfs_bandwidth_runtime(cfs_b);
8657 /* restart the period timer (if active) to handle new period expiry */
8658 if (runtime_enabled)
8659 start_cfs_bandwidth(cfs_b);
8660 raw_spin_unlock_irq(&cfs_b->lock);
8662 for_each_online_cpu(i) {
8663 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8664 struct rq *rq = cfs_rq->rq;
8666 raw_spin_lock_irq(&rq->lock);
8667 cfs_rq->runtime_enabled = runtime_enabled;
8668 cfs_rq->runtime_remaining = 0;
8670 if (cfs_rq->throttled)
8671 unthrottle_cfs_rq(cfs_rq);
8672 raw_spin_unlock_irq(&rq->lock);
8674 if (runtime_was_enabled && !runtime_enabled)
8675 cfs_bandwidth_usage_dec();
8677 mutex_unlock(&cfs_constraints_mutex);
8683 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8687 period = ktime_to_ns(tg->cfs_bandwidth.period);
8688 if (cfs_quota_us < 0)
8689 quota = RUNTIME_INF;
8691 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8693 return tg_set_cfs_bandwidth(tg, period, quota);
8696 long tg_get_cfs_quota(struct task_group *tg)
8700 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8703 quota_us = tg->cfs_bandwidth.quota;
8704 do_div(quota_us, NSEC_PER_USEC);
8709 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8713 period = (u64)cfs_period_us * NSEC_PER_USEC;
8714 quota = tg->cfs_bandwidth.quota;
8716 return tg_set_cfs_bandwidth(tg, period, quota);
8719 long tg_get_cfs_period(struct task_group *tg)
8723 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8724 do_div(cfs_period_us, NSEC_PER_USEC);
8726 return cfs_period_us;
8729 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8732 return tg_get_cfs_quota(css_tg(css));
8735 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8736 struct cftype *cftype, s64 cfs_quota_us)
8738 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8741 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8744 return tg_get_cfs_period(css_tg(css));
8747 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8748 struct cftype *cftype, u64 cfs_period_us)
8750 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8753 struct cfs_schedulable_data {
8754 struct task_group *tg;
8759 * normalize group quota/period to be quota/max_period
8760 * note: units are usecs
8762 static u64 normalize_cfs_quota(struct task_group *tg,
8763 struct cfs_schedulable_data *d)
8771 period = tg_get_cfs_period(tg);
8772 quota = tg_get_cfs_quota(tg);
8775 /* note: these should typically be equivalent */
8776 if (quota == RUNTIME_INF || quota == -1)
8779 return to_ratio(period, quota);
8782 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8784 struct cfs_schedulable_data *d = data;
8785 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8786 s64 quota = 0, parent_quota = -1;
8789 quota = RUNTIME_INF;
8791 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8793 quota = normalize_cfs_quota(tg, d);
8794 parent_quota = parent_b->hierarchical_quota;
8797 * ensure max(child_quota) <= parent_quota, inherit when no
8800 if (quota == RUNTIME_INF)
8801 quota = parent_quota;
8802 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8805 cfs_b->hierarchical_quota = quota;
8810 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8813 struct cfs_schedulable_data data = {
8819 if (quota != RUNTIME_INF) {
8820 do_div(data.period, NSEC_PER_USEC);
8821 do_div(data.quota, NSEC_PER_USEC);
8825 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8831 static int cpu_stats_show(struct seq_file *sf, void *v)
8833 struct task_group *tg = css_tg(seq_css(sf));
8834 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8836 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8837 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8838 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8842 #endif /* CONFIG_CFS_BANDWIDTH */
8843 #endif /* CONFIG_FAIR_GROUP_SCHED */
8845 #ifdef CONFIG_RT_GROUP_SCHED
8846 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8847 struct cftype *cft, s64 val)
8849 return sched_group_set_rt_runtime(css_tg(css), val);
8852 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8855 return sched_group_rt_runtime(css_tg(css));
8858 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8859 struct cftype *cftype, u64 rt_period_us)
8861 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8864 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8867 return sched_group_rt_period(css_tg(css));
8869 #endif /* CONFIG_RT_GROUP_SCHED */
8871 static struct cftype cpu_files[] = {
8872 #ifdef CONFIG_FAIR_GROUP_SCHED
8875 .read_u64 = cpu_shares_read_u64,
8876 .write_u64 = cpu_shares_write_u64,
8879 #ifdef CONFIG_CFS_BANDWIDTH
8881 .name = "cfs_quota_us",
8882 .read_s64 = cpu_cfs_quota_read_s64,
8883 .write_s64 = cpu_cfs_quota_write_s64,
8886 .name = "cfs_period_us",
8887 .read_u64 = cpu_cfs_period_read_u64,
8888 .write_u64 = cpu_cfs_period_write_u64,
8892 .seq_show = cpu_stats_show,
8895 #ifdef CONFIG_RT_GROUP_SCHED
8897 .name = "rt_runtime_us",
8898 .read_s64 = cpu_rt_runtime_read,
8899 .write_s64 = cpu_rt_runtime_write,
8902 .name = "rt_period_us",
8903 .read_u64 = cpu_rt_period_read_uint,
8904 .write_u64 = cpu_rt_period_write_uint,
8910 struct cgroup_subsys cpu_cgrp_subsys = {
8911 .css_alloc = cpu_cgroup_css_alloc,
8912 .css_released = cpu_cgroup_css_released,
8913 .css_free = cpu_cgroup_css_free,
8914 .fork = cpu_cgroup_fork,
8915 .can_attach = cpu_cgroup_can_attach,
8916 .attach = cpu_cgroup_attach,
8917 .legacy_cftypes = cpu_files,
8921 #endif /* CONFIG_CGROUP_SCHED */
8923 void dump_cpu_task(int cpu)
8925 pr_info("Task dump for CPU %d:\n", cpu);
8926 sched_show_task(cpu_curr(cpu));
8930 * Nice levels are multiplicative, with a gentle 10% change for every
8931 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8932 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8933 * that remained on nice 0.
8935 * The "10% effect" is relative and cumulative: from _any_ nice level,
8936 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8937 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8938 * If a task goes up by ~10% and another task goes down by ~10% then
8939 * the relative distance between them is ~25%.)
8941 const int sched_prio_to_weight[40] = {
8942 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8943 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8944 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8945 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8946 /* 0 */ 1024, 820, 655, 526, 423,
8947 /* 5 */ 335, 272, 215, 172, 137,
8948 /* 10 */ 110, 87, 70, 56, 45,
8949 /* 15 */ 36, 29, 23, 18, 15,
8953 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8955 * In cases where the weight does not change often, we can use the
8956 * precalculated inverse to speed up arithmetics by turning divisions
8957 * into multiplications:
8959 const u32 sched_prio_to_wmult[40] = {
8960 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8961 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8962 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8963 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8964 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8965 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8966 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8967 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,