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 <asm/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>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
131 const_debug unsigned int sysctl_sched_nr_migrate = 32;
134 * period over which we average the RT time consumption, measured
139 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
142 * period over which we measure -rt task cpu usage in us.
145 unsigned int sysctl_sched_rt_period = 1000000;
147 __read_mostly int scheduler_running;
150 * part of the period that we allow rt tasks to run in us.
153 int sysctl_sched_rt_runtime = 950000;
155 /* cpus with isolated domains */
156 cpumask_var_t cpu_isolated_map;
159 * this_rq_lock - lock this runqueue and disable interrupts.
161 static struct rq *this_rq_lock(void)
168 raw_spin_lock(&rq->lock);
173 #ifdef CONFIG_SCHED_HRTICK
175 * Use HR-timers to deliver accurate preemption points.
178 static void hrtick_clear(struct rq *rq)
180 if (hrtimer_active(&rq->hrtick_timer))
181 hrtimer_cancel(&rq->hrtick_timer);
185 * High-resolution timer tick.
186 * Runs from hardirq context with interrupts disabled.
188 static enum hrtimer_restart hrtick(struct hrtimer *timer)
190 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
192 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
194 raw_spin_lock(&rq->lock);
196 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
197 raw_spin_unlock(&rq->lock);
199 return HRTIMER_NORESTART;
204 static void __hrtick_restart(struct rq *rq)
206 struct hrtimer *timer = &rq->hrtick_timer;
208 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
212 * called from hardirq (IPI) context
214 static void __hrtick_start(void *arg)
218 raw_spin_lock(&rq->lock);
219 __hrtick_restart(rq);
220 rq->hrtick_csd_pending = 0;
221 raw_spin_unlock(&rq->lock);
225 * Called to set the hrtick timer state.
227 * called with rq->lock held and irqs disabled
229 void hrtick_start(struct rq *rq, u64 delay)
231 struct hrtimer *timer = &rq->hrtick_timer;
236 * Don't schedule slices shorter than 10000ns, that just
237 * doesn't make sense and can cause timer DoS.
239 delta = max_t(s64, delay, 10000LL);
240 time = ktime_add_ns(timer->base->get_time(), delta);
242 hrtimer_set_expires(timer, time);
244 if (rq == this_rq()) {
245 __hrtick_restart(rq);
246 } else if (!rq->hrtick_csd_pending) {
247 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
248 rq->hrtick_csd_pending = 1;
253 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
255 int cpu = (int)(long)hcpu;
258 case CPU_UP_CANCELED:
259 case CPU_UP_CANCELED_FROZEN:
260 case CPU_DOWN_PREPARE:
261 case CPU_DOWN_PREPARE_FROZEN:
263 case CPU_DEAD_FROZEN:
264 hrtick_clear(cpu_rq(cpu));
271 static __init void init_hrtick(void)
273 hotcpu_notifier(hotplug_hrtick, 0);
277 * Called to set the hrtick timer state.
279 * called with rq->lock held and irqs disabled
281 void hrtick_start(struct rq *rq, u64 delay)
284 * Don't schedule slices shorter than 10000ns, that just
285 * doesn't make sense. Rely on vruntime for fairness.
287 delay = max_t(u64, delay, 10000LL);
288 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
289 HRTIMER_MODE_REL_PINNED);
292 static inline void init_hrtick(void)
295 #endif /* CONFIG_SMP */
297 static void init_rq_hrtick(struct rq *rq)
300 rq->hrtick_csd_pending = 0;
302 rq->hrtick_csd.flags = 0;
303 rq->hrtick_csd.func = __hrtick_start;
304 rq->hrtick_csd.info = rq;
307 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
308 rq->hrtick_timer.function = hrtick;
310 #else /* CONFIG_SCHED_HRTICK */
311 static inline void hrtick_clear(struct rq *rq)
315 static inline void init_rq_hrtick(struct rq *rq)
319 static inline void init_hrtick(void)
322 #endif /* CONFIG_SCHED_HRTICK */
325 * cmpxchg based fetch_or, macro so it works for different integer types
327 #define fetch_or(ptr, mask) \
329 typeof(ptr) _ptr = (ptr); \
330 typeof(mask) _mask = (mask); \
331 typeof(*_ptr) _old, _val = *_ptr; \
334 _old = cmpxchg(_ptr, _val, _val | _mask); \
342 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
344 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
345 * this avoids any races wrt polling state changes and thereby avoids
348 static bool set_nr_and_not_polling(struct task_struct *p)
350 struct thread_info *ti = task_thread_info(p);
351 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
355 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
357 * If this returns true, then the idle task promises to call
358 * sched_ttwu_pending() and reschedule soon.
360 static bool set_nr_if_polling(struct task_struct *p)
362 struct thread_info *ti = task_thread_info(p);
363 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
366 if (!(val & _TIF_POLLING_NRFLAG))
368 if (val & _TIF_NEED_RESCHED)
370 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
379 static bool set_nr_and_not_polling(struct task_struct *p)
381 set_tsk_need_resched(p);
386 static bool set_nr_if_polling(struct task_struct *p)
393 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
395 struct wake_q_node *node = &task->wake_q;
398 * Atomically grab the task, if ->wake_q is !nil already it means
399 * its already queued (either by us or someone else) and will get the
400 * wakeup due to that.
402 * This cmpxchg() implies a full barrier, which pairs with the write
403 * barrier implied by the wakeup in wake_up_list().
405 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
408 get_task_struct(task);
411 * The head is context local, there can be no concurrency.
414 head->lastp = &node->next;
417 void wake_up_q(struct wake_q_head *head)
419 struct wake_q_node *node = head->first;
421 while (node != WAKE_Q_TAIL) {
422 struct task_struct *task;
424 task = container_of(node, struct task_struct, wake_q);
426 /* task can safely be re-inserted now */
428 task->wake_q.next = NULL;
431 * wake_up_process() implies a wmb() to pair with the queueing
432 * in wake_q_add() so as not to miss wakeups.
434 wake_up_process(task);
435 put_task_struct(task);
440 * resched_curr - mark rq's current task 'to be rescheduled now'.
442 * On UP this means the setting of the need_resched flag, on SMP it
443 * might also involve a cross-CPU call to trigger the scheduler on
446 void resched_curr(struct rq *rq)
448 struct task_struct *curr = rq->curr;
451 lockdep_assert_held(&rq->lock);
453 if (test_tsk_need_resched(curr))
458 if (cpu == smp_processor_id()) {
459 set_tsk_need_resched(curr);
460 set_preempt_need_resched();
464 if (set_nr_and_not_polling(curr))
465 smp_send_reschedule(cpu);
467 trace_sched_wake_idle_without_ipi(cpu);
470 void resched_cpu(int cpu)
472 struct rq *rq = cpu_rq(cpu);
475 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
478 raw_spin_unlock_irqrestore(&rq->lock, flags);
482 #ifdef CONFIG_NO_HZ_COMMON
484 * In the semi idle case, use the nearest busy cpu for migrating timers
485 * from an idle cpu. This is good for power-savings.
487 * We don't do similar optimization for completely idle system, as
488 * selecting an idle cpu will add more delays to the timers than intended
489 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
491 int get_nohz_timer_target(void)
493 int i, cpu = smp_processor_id();
494 struct sched_domain *sd;
496 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
500 for_each_domain(cpu, sd) {
501 for_each_cpu(i, sched_domain_span(sd)) {
502 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
509 if (!is_housekeeping_cpu(cpu))
510 cpu = housekeeping_any_cpu();
516 * When add_timer_on() enqueues a timer into the timer wheel of an
517 * idle CPU then this timer might expire before the next timer event
518 * which is scheduled to wake up that CPU. In case of a completely
519 * idle system the next event might even be infinite time into the
520 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
521 * leaves the inner idle loop so the newly added timer is taken into
522 * account when the CPU goes back to idle and evaluates the timer
523 * wheel for the next timer event.
525 static void wake_up_idle_cpu(int cpu)
527 struct rq *rq = cpu_rq(cpu);
529 if (cpu == smp_processor_id())
532 if (set_nr_and_not_polling(rq->idle))
533 smp_send_reschedule(cpu);
535 trace_sched_wake_idle_without_ipi(cpu);
538 static bool wake_up_full_nohz_cpu(int cpu)
541 * We just need the target to call irq_exit() and re-evaluate
542 * the next tick. The nohz full kick at least implies that.
543 * If needed we can still optimize that later with an
546 if (tick_nohz_full_cpu(cpu)) {
547 if (cpu != smp_processor_id() ||
548 tick_nohz_tick_stopped())
549 tick_nohz_full_kick_cpu(cpu);
556 void wake_up_nohz_cpu(int cpu)
558 if (!wake_up_full_nohz_cpu(cpu))
559 wake_up_idle_cpu(cpu);
562 static inline bool got_nohz_idle_kick(void)
564 int cpu = smp_processor_id();
566 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
569 if (idle_cpu(cpu) && !need_resched())
573 * We can't run Idle Load Balance on this CPU for this time so we
574 * cancel it and clear NOHZ_BALANCE_KICK
576 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
580 #else /* CONFIG_NO_HZ_COMMON */
582 static inline bool got_nohz_idle_kick(void)
587 #endif /* CONFIG_NO_HZ_COMMON */
589 #ifdef CONFIG_NO_HZ_FULL
590 bool sched_can_stop_tick(struct rq *rq)
594 /* Deadline tasks, even if single, need the tick */
595 if (rq->dl.dl_nr_running)
599 * If there are more than one RR tasks, we need the tick to effect the
600 * actual RR behaviour.
602 if (rq->rt.rr_nr_running) {
603 if (rq->rt.rr_nr_running == 1)
610 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
611 * forced preemption between FIFO tasks.
613 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
618 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
619 * if there's more than one we need the tick for involuntary
622 if (rq->nr_running > 1)
627 #endif /* CONFIG_NO_HZ_FULL */
629 void sched_avg_update(struct rq *rq)
631 s64 period = sched_avg_period();
633 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
635 * Inline assembly required to prevent the compiler
636 * optimising this loop into a divmod call.
637 * See __iter_div_u64_rem() for another example of this.
639 asm("" : "+rm" (rq->age_stamp));
640 rq->age_stamp += period;
645 #endif /* CONFIG_SMP */
647 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
653 * Caller must hold rcu_lock or sufficient equivalent.
655 int walk_tg_tree_from(struct task_group *from,
656 tg_visitor down, tg_visitor up, void *data)
658 struct task_group *parent, *child;
664 ret = (*down)(parent, data);
667 list_for_each_entry_rcu(child, &parent->children, siblings) {
674 ret = (*up)(parent, data);
675 if (ret || parent == from)
679 parent = parent->parent;
686 int tg_nop(struct task_group *tg, void *data)
692 static void set_load_weight(struct task_struct *p)
694 int prio = p->static_prio - MAX_RT_PRIO;
695 struct load_weight *load = &p->se.load;
698 * SCHED_IDLE tasks get minimal weight:
700 if (idle_policy(p->policy)) {
701 load->weight = scale_load(WEIGHT_IDLEPRIO);
702 load->inv_weight = WMULT_IDLEPRIO;
706 load->weight = scale_load(sched_prio_to_weight[prio]);
707 load->inv_weight = sched_prio_to_wmult[prio];
710 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
713 if (!(flags & ENQUEUE_RESTORE))
714 sched_info_queued(rq, p);
715 p->sched_class->enqueue_task(rq, p, flags);
718 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
721 if (!(flags & DEQUEUE_SAVE))
722 sched_info_dequeued(rq, p);
723 p->sched_class->dequeue_task(rq, p, flags);
726 void activate_task(struct rq *rq, struct task_struct *p, int flags)
728 if (task_contributes_to_load(p))
729 rq->nr_uninterruptible--;
731 enqueue_task(rq, p, flags);
734 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
736 if (task_contributes_to_load(p))
737 rq->nr_uninterruptible++;
739 dequeue_task(rq, p, flags);
742 static void update_rq_clock_task(struct rq *rq, s64 delta)
745 * In theory, the compile should just see 0 here, and optimize out the call
746 * to sched_rt_avg_update. But I don't trust it...
748 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
749 s64 steal = 0, irq_delta = 0;
751 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
752 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
755 * Since irq_time is only updated on {soft,}irq_exit, we might run into
756 * this case when a previous update_rq_clock() happened inside a
759 * When this happens, we stop ->clock_task and only update the
760 * prev_irq_time stamp to account for the part that fit, so that a next
761 * update will consume the rest. This ensures ->clock_task is
764 * It does however cause some slight miss-attribution of {soft,}irq
765 * time, a more accurate solution would be to update the irq_time using
766 * the current rq->clock timestamp, except that would require using
769 if (irq_delta > delta)
772 rq->prev_irq_time += irq_delta;
775 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
776 if (static_key_false((¶virt_steal_rq_enabled))) {
777 steal = paravirt_steal_clock(cpu_of(rq));
778 steal -= rq->prev_steal_time_rq;
780 if (unlikely(steal > delta))
783 rq->prev_steal_time_rq += steal;
788 rq->clock_task += delta;
790 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
791 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
792 sched_rt_avg_update(rq, irq_delta + steal);
796 void sched_set_stop_task(int cpu, struct task_struct *stop)
798 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
799 struct task_struct *old_stop = cpu_rq(cpu)->stop;
803 * Make it appear like a SCHED_FIFO task, its something
804 * userspace knows about and won't get confused about.
806 * Also, it will make PI more or less work without too
807 * much confusion -- but then, stop work should not
808 * rely on PI working anyway.
810 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
812 stop->sched_class = &stop_sched_class;
815 cpu_rq(cpu)->stop = stop;
819 * Reset it back to a normal scheduling class so that
820 * it can die in pieces.
822 old_stop->sched_class = &rt_sched_class;
827 * __normal_prio - return the priority that is based on the static prio
829 static inline int __normal_prio(struct task_struct *p)
831 return p->static_prio;
835 * Calculate the expected normal priority: i.e. priority
836 * without taking RT-inheritance into account. Might be
837 * boosted by interactivity modifiers. Changes upon fork,
838 * setprio syscalls, and whenever the interactivity
839 * estimator recalculates.
841 static inline int normal_prio(struct task_struct *p)
845 if (task_has_dl_policy(p))
846 prio = MAX_DL_PRIO-1;
847 else if (task_has_rt_policy(p))
848 prio = MAX_RT_PRIO-1 - p->rt_priority;
850 prio = __normal_prio(p);
855 * Calculate the current priority, i.e. the priority
856 * taken into account by the scheduler. This value might
857 * be boosted by RT tasks, or might be boosted by
858 * interactivity modifiers. Will be RT if the task got
859 * RT-boosted. If not then it returns p->normal_prio.
861 static int effective_prio(struct task_struct *p)
863 p->normal_prio = normal_prio(p);
865 * If we are RT tasks or we were boosted to RT priority,
866 * keep the priority unchanged. Otherwise, update priority
867 * to the normal priority:
869 if (!rt_prio(p->prio))
870 return p->normal_prio;
875 * task_curr - is this task currently executing on a CPU?
876 * @p: the task in question.
878 * Return: 1 if the task is currently executing. 0 otherwise.
880 inline int task_curr(const struct task_struct *p)
882 return cpu_curr(task_cpu(p)) == p;
886 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
887 * use the balance_callback list if you want balancing.
889 * this means any call to check_class_changed() must be followed by a call to
890 * balance_callback().
892 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
893 const struct sched_class *prev_class,
896 if (prev_class != p->sched_class) {
897 if (prev_class->switched_from)
898 prev_class->switched_from(rq, p);
900 p->sched_class->switched_to(rq, p);
901 } else if (oldprio != p->prio || dl_task(p))
902 p->sched_class->prio_changed(rq, p, oldprio);
905 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
907 const struct sched_class *class;
909 if (p->sched_class == rq->curr->sched_class) {
910 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
912 for_each_class(class) {
913 if (class == rq->curr->sched_class)
915 if (class == p->sched_class) {
923 * A queue event has occurred, and we're going to schedule. In
924 * this case, we can save a useless back to back clock update.
926 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
927 rq_clock_skip_update(rq, true);
932 * This is how migration works:
934 * 1) we invoke migration_cpu_stop() on the target CPU using
936 * 2) stopper starts to run (implicitly forcing the migrated thread
938 * 3) it checks whether the migrated task is still in the wrong runqueue.
939 * 4) if it's in the wrong runqueue then the migration thread removes
940 * it and puts it into the right queue.
941 * 5) stopper completes and stop_one_cpu() returns and the migration
946 * move_queued_task - move a queued task to new rq.
948 * Returns (locked) new rq. Old rq's lock is released.
950 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
952 lockdep_assert_held(&rq->lock);
954 p->on_rq = TASK_ON_RQ_MIGRATING;
955 dequeue_task(rq, p, 0);
956 set_task_cpu(p, new_cpu);
957 raw_spin_unlock(&rq->lock);
959 rq = cpu_rq(new_cpu);
961 raw_spin_lock(&rq->lock);
962 BUG_ON(task_cpu(p) != new_cpu);
963 enqueue_task(rq, p, 0);
964 p->on_rq = TASK_ON_RQ_QUEUED;
965 check_preempt_curr(rq, p, 0);
970 struct migration_arg {
971 struct task_struct *task;
976 * Move (not current) task off this cpu, onto dest cpu. We're doing
977 * this because either it can't run here any more (set_cpus_allowed()
978 * away from this CPU, or CPU going down), or because we're
979 * attempting to rebalance this task on exec (sched_exec).
981 * So we race with normal scheduler movements, but that's OK, as long
982 * as the task is no longer on this CPU.
984 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
986 if (unlikely(!cpu_active(dest_cpu)))
989 /* Affinity changed (again). */
990 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
993 rq = move_queued_task(rq, p, dest_cpu);
999 * migration_cpu_stop - this will be executed by a highprio stopper thread
1000 * and performs thread migration by bumping thread off CPU then
1001 * 'pushing' onto another runqueue.
1003 static int migration_cpu_stop(void *data)
1005 struct migration_arg *arg = data;
1006 struct task_struct *p = arg->task;
1007 struct rq *rq = this_rq();
1010 * The original target cpu might have gone down and we might
1011 * be on another cpu but it doesn't matter.
1013 local_irq_disable();
1015 * We need to explicitly wake pending tasks before running
1016 * __migrate_task() such that we will not miss enforcing cpus_allowed
1017 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1019 sched_ttwu_pending();
1021 raw_spin_lock(&p->pi_lock);
1022 raw_spin_lock(&rq->lock);
1024 * If task_rq(p) != rq, it cannot be migrated here, because we're
1025 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1026 * we're holding p->pi_lock.
1028 if (task_rq(p) == rq && task_on_rq_queued(p))
1029 rq = __migrate_task(rq, p, arg->dest_cpu);
1030 raw_spin_unlock(&rq->lock);
1031 raw_spin_unlock(&p->pi_lock);
1038 * sched_class::set_cpus_allowed must do the below, but is not required to
1039 * actually call this function.
1041 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1043 cpumask_copy(&p->cpus_allowed, new_mask);
1044 p->nr_cpus_allowed = cpumask_weight(new_mask);
1047 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1049 struct rq *rq = task_rq(p);
1050 bool queued, running;
1052 lockdep_assert_held(&p->pi_lock);
1054 queued = task_on_rq_queued(p);
1055 running = task_current(rq, p);
1059 * Because __kthread_bind() calls this on blocked tasks without
1062 lockdep_assert_held(&rq->lock);
1063 dequeue_task(rq, p, DEQUEUE_SAVE);
1066 put_prev_task(rq, p);
1068 p->sched_class->set_cpus_allowed(p, new_mask);
1071 p->sched_class->set_curr_task(rq);
1073 enqueue_task(rq, p, ENQUEUE_RESTORE);
1077 * Change a given task's CPU affinity. Migrate the thread to a
1078 * proper CPU and schedule it away if the CPU it's executing on
1079 * is removed from the allowed bitmask.
1081 * NOTE: the caller must have a valid reference to the task, the
1082 * task must not exit() & deallocate itself prematurely. The
1083 * call is not atomic; no spinlocks may be held.
1085 static int __set_cpus_allowed_ptr(struct task_struct *p,
1086 const struct cpumask *new_mask, bool check)
1088 unsigned long flags;
1090 unsigned int dest_cpu;
1093 rq = task_rq_lock(p, &flags);
1096 * Must re-check here, to close a race against __kthread_bind(),
1097 * sched_setaffinity() is not guaranteed to observe the flag.
1099 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1104 if (cpumask_equal(&p->cpus_allowed, new_mask))
1107 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1112 do_set_cpus_allowed(p, new_mask);
1114 /* Can the task run on the task's current CPU? If so, we're done */
1115 if (cpumask_test_cpu(task_cpu(p), new_mask))
1118 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1119 if (task_running(rq, p) || p->state == TASK_WAKING) {
1120 struct migration_arg arg = { p, dest_cpu };
1121 /* Need help from migration thread: drop lock and wait. */
1122 task_rq_unlock(rq, p, &flags);
1123 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1124 tlb_migrate_finish(p->mm);
1126 } else if (task_on_rq_queued(p)) {
1128 * OK, since we're going to drop the lock immediately
1129 * afterwards anyway.
1131 lockdep_unpin_lock(&rq->lock);
1132 rq = move_queued_task(rq, p, dest_cpu);
1133 lockdep_pin_lock(&rq->lock);
1136 task_rq_unlock(rq, p, &flags);
1141 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1143 return __set_cpus_allowed_ptr(p, new_mask, false);
1145 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1147 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1149 #ifdef CONFIG_SCHED_DEBUG
1151 * We should never call set_task_cpu() on a blocked task,
1152 * ttwu() will sort out the placement.
1154 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1158 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1159 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1160 * time relying on p->on_rq.
1162 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1163 p->sched_class == &fair_sched_class &&
1164 (p->on_rq && !task_on_rq_migrating(p)));
1166 #ifdef CONFIG_LOCKDEP
1168 * The caller should hold either p->pi_lock or rq->lock, when changing
1169 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1171 * sched_move_task() holds both and thus holding either pins the cgroup,
1174 * Furthermore, all task_rq users should acquire both locks, see
1177 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1178 lockdep_is_held(&task_rq(p)->lock)));
1182 trace_sched_migrate_task(p, new_cpu);
1184 if (task_cpu(p) != new_cpu) {
1185 if (p->sched_class->migrate_task_rq)
1186 p->sched_class->migrate_task_rq(p);
1187 p->se.nr_migrations++;
1188 perf_event_task_migrate(p);
1191 __set_task_cpu(p, new_cpu);
1194 static void __migrate_swap_task(struct task_struct *p, int cpu)
1196 if (task_on_rq_queued(p)) {
1197 struct rq *src_rq, *dst_rq;
1199 src_rq = task_rq(p);
1200 dst_rq = cpu_rq(cpu);
1202 p->on_rq = TASK_ON_RQ_MIGRATING;
1203 deactivate_task(src_rq, p, 0);
1204 set_task_cpu(p, cpu);
1205 activate_task(dst_rq, p, 0);
1206 p->on_rq = TASK_ON_RQ_QUEUED;
1207 check_preempt_curr(dst_rq, p, 0);
1210 * Task isn't running anymore; make it appear like we migrated
1211 * it before it went to sleep. This means on wakeup we make the
1212 * previous cpu our targer instead of where it really is.
1218 struct migration_swap_arg {
1219 struct task_struct *src_task, *dst_task;
1220 int src_cpu, dst_cpu;
1223 static int migrate_swap_stop(void *data)
1225 struct migration_swap_arg *arg = data;
1226 struct rq *src_rq, *dst_rq;
1229 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1232 src_rq = cpu_rq(arg->src_cpu);
1233 dst_rq = cpu_rq(arg->dst_cpu);
1235 double_raw_lock(&arg->src_task->pi_lock,
1236 &arg->dst_task->pi_lock);
1237 double_rq_lock(src_rq, dst_rq);
1239 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1242 if (task_cpu(arg->src_task) != arg->src_cpu)
1245 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1248 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1251 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1252 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1257 double_rq_unlock(src_rq, dst_rq);
1258 raw_spin_unlock(&arg->dst_task->pi_lock);
1259 raw_spin_unlock(&arg->src_task->pi_lock);
1265 * Cross migrate two tasks
1267 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1269 struct migration_swap_arg arg;
1272 arg = (struct migration_swap_arg){
1274 .src_cpu = task_cpu(cur),
1276 .dst_cpu = task_cpu(p),
1279 if (arg.src_cpu == arg.dst_cpu)
1283 * These three tests are all lockless; this is OK since all of them
1284 * will be re-checked with proper locks held further down the line.
1286 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1289 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1292 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1295 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1296 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1303 * wait_task_inactive - wait for a thread to unschedule.
1305 * If @match_state is nonzero, it's the @p->state value just checked and
1306 * not expected to change. If it changes, i.e. @p might have woken up,
1307 * then return zero. When we succeed in waiting for @p to be off its CPU,
1308 * we return a positive number (its total switch count). If a second call
1309 * a short while later returns the same number, the caller can be sure that
1310 * @p has remained unscheduled the whole time.
1312 * The caller must ensure that the task *will* unschedule sometime soon,
1313 * else this function might spin for a *long* time. This function can't
1314 * be called with interrupts off, or it may introduce deadlock with
1315 * smp_call_function() if an IPI is sent by the same process we are
1316 * waiting to become inactive.
1318 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1320 unsigned long flags;
1321 int running, queued;
1327 * We do the initial early heuristics without holding
1328 * any task-queue locks at all. We'll only try to get
1329 * the runqueue lock when things look like they will
1335 * If the task is actively running on another CPU
1336 * still, just relax and busy-wait without holding
1339 * NOTE! Since we don't hold any locks, it's not
1340 * even sure that "rq" stays as the right runqueue!
1341 * But we don't care, since "task_running()" will
1342 * return false if the runqueue has changed and p
1343 * is actually now running somewhere else!
1345 while (task_running(rq, p)) {
1346 if (match_state && unlikely(p->state != match_state))
1352 * Ok, time to look more closely! We need the rq
1353 * lock now, to be *sure*. If we're wrong, we'll
1354 * just go back and repeat.
1356 rq = task_rq_lock(p, &flags);
1357 trace_sched_wait_task(p);
1358 running = task_running(rq, p);
1359 queued = task_on_rq_queued(p);
1361 if (!match_state || p->state == match_state)
1362 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1363 task_rq_unlock(rq, p, &flags);
1366 * If it changed from the expected state, bail out now.
1368 if (unlikely(!ncsw))
1372 * Was it really running after all now that we
1373 * checked with the proper locks actually held?
1375 * Oops. Go back and try again..
1377 if (unlikely(running)) {
1383 * It's not enough that it's not actively running,
1384 * it must be off the runqueue _entirely_, and not
1387 * So if it was still runnable (but just not actively
1388 * running right now), it's preempted, and we should
1389 * yield - it could be a while.
1391 if (unlikely(queued)) {
1392 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1394 set_current_state(TASK_UNINTERRUPTIBLE);
1395 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1400 * Ahh, all good. It wasn't running, and it wasn't
1401 * runnable, which means that it will never become
1402 * running in the future either. We're all done!
1411 * kick_process - kick a running thread to enter/exit the kernel
1412 * @p: the to-be-kicked thread
1414 * Cause a process which is running on another CPU to enter
1415 * kernel-mode, without any delay. (to get signals handled.)
1417 * NOTE: this function doesn't have to take the runqueue lock,
1418 * because all it wants to ensure is that the remote task enters
1419 * the kernel. If the IPI races and the task has been migrated
1420 * to another CPU then no harm is done and the purpose has been
1423 void kick_process(struct task_struct *p)
1429 if ((cpu != smp_processor_id()) && task_curr(p))
1430 smp_send_reschedule(cpu);
1433 EXPORT_SYMBOL_GPL(kick_process);
1436 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1438 static int select_fallback_rq(int cpu, struct task_struct *p)
1440 int nid = cpu_to_node(cpu);
1441 const struct cpumask *nodemask = NULL;
1442 enum { cpuset, possible, fail } state = cpuset;
1446 * If the node that the cpu is on has been offlined, cpu_to_node()
1447 * will return -1. There is no cpu on the node, and we should
1448 * select the cpu on the other node.
1451 nodemask = cpumask_of_node(nid);
1453 /* Look for allowed, online CPU in same node. */
1454 for_each_cpu(dest_cpu, nodemask) {
1455 if (!cpu_online(dest_cpu))
1457 if (!cpu_active(dest_cpu))
1459 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1465 /* Any allowed, online CPU? */
1466 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1467 if (!cpu_online(dest_cpu))
1469 if (!cpu_active(dest_cpu))
1474 /* No more Mr. Nice Guy. */
1477 if (IS_ENABLED(CONFIG_CPUSETS)) {
1478 cpuset_cpus_allowed_fallback(p);
1484 do_set_cpus_allowed(p, cpu_possible_mask);
1495 if (state != cpuset) {
1497 * Don't tell them about moving exiting tasks or
1498 * kernel threads (both mm NULL), since they never
1501 if (p->mm && printk_ratelimit()) {
1502 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1503 task_pid_nr(p), p->comm, cpu);
1511 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1514 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1516 lockdep_assert_held(&p->pi_lock);
1518 if (p->nr_cpus_allowed > 1)
1519 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1522 * In order not to call set_task_cpu() on a blocking task we need
1523 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1526 * Since this is common to all placement strategies, this lives here.
1528 * [ this allows ->select_task() to simply return task_cpu(p) and
1529 * not worry about this generic constraint ]
1531 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1533 cpu = select_fallback_rq(task_cpu(p), p);
1538 static void update_avg(u64 *avg, u64 sample)
1540 s64 diff = sample - *avg;
1546 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1547 const struct cpumask *new_mask, bool check)
1549 return set_cpus_allowed_ptr(p, new_mask);
1552 #endif /* CONFIG_SMP */
1555 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1557 #ifdef CONFIG_SCHEDSTATS
1558 struct rq *rq = this_rq();
1561 int this_cpu = smp_processor_id();
1563 if (cpu == this_cpu) {
1564 schedstat_inc(rq, ttwu_local);
1565 schedstat_inc(p, se.statistics.nr_wakeups_local);
1567 struct sched_domain *sd;
1569 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1571 for_each_domain(this_cpu, sd) {
1572 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1573 schedstat_inc(sd, ttwu_wake_remote);
1580 if (wake_flags & WF_MIGRATED)
1581 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1583 #endif /* CONFIG_SMP */
1585 schedstat_inc(rq, ttwu_count);
1586 schedstat_inc(p, se.statistics.nr_wakeups);
1588 if (wake_flags & WF_SYNC)
1589 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1591 #endif /* CONFIG_SCHEDSTATS */
1594 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1596 activate_task(rq, p, en_flags);
1597 p->on_rq = TASK_ON_RQ_QUEUED;
1599 /* if a worker is waking up, notify workqueue */
1600 if (p->flags & PF_WQ_WORKER)
1601 wq_worker_waking_up(p, cpu_of(rq));
1605 * Mark the task runnable and perform wakeup-preemption.
1608 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1610 check_preempt_curr(rq, p, wake_flags);
1611 p->state = TASK_RUNNING;
1612 trace_sched_wakeup(p);
1615 if (p->sched_class->task_woken) {
1617 * Our task @p is fully woken up and running; so its safe to
1618 * drop the rq->lock, hereafter rq is only used for statistics.
1620 lockdep_unpin_lock(&rq->lock);
1621 p->sched_class->task_woken(rq, p);
1622 lockdep_pin_lock(&rq->lock);
1625 if (rq->idle_stamp) {
1626 u64 delta = rq_clock(rq) - rq->idle_stamp;
1627 u64 max = 2*rq->max_idle_balance_cost;
1629 update_avg(&rq->avg_idle, delta);
1631 if (rq->avg_idle > max)
1640 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1642 lockdep_assert_held(&rq->lock);
1645 if (p->sched_contributes_to_load)
1646 rq->nr_uninterruptible--;
1649 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1650 ttwu_do_wakeup(rq, p, wake_flags);
1654 * Called in case the task @p isn't fully descheduled from its runqueue,
1655 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1656 * since all we need to do is flip p->state to TASK_RUNNING, since
1657 * the task is still ->on_rq.
1659 static int ttwu_remote(struct task_struct *p, int wake_flags)
1664 rq = __task_rq_lock(p);
1665 if (task_on_rq_queued(p)) {
1666 /* check_preempt_curr() may use rq clock */
1667 update_rq_clock(rq);
1668 ttwu_do_wakeup(rq, p, wake_flags);
1671 __task_rq_unlock(rq);
1677 void sched_ttwu_pending(void)
1679 struct rq *rq = this_rq();
1680 struct llist_node *llist = llist_del_all(&rq->wake_list);
1681 struct task_struct *p;
1682 unsigned long flags;
1687 raw_spin_lock_irqsave(&rq->lock, flags);
1688 lockdep_pin_lock(&rq->lock);
1691 p = llist_entry(llist, struct task_struct, wake_entry);
1692 llist = llist_next(llist);
1693 ttwu_do_activate(rq, p, 0);
1696 lockdep_unpin_lock(&rq->lock);
1697 raw_spin_unlock_irqrestore(&rq->lock, flags);
1700 void scheduler_ipi(void)
1703 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1704 * TIF_NEED_RESCHED remotely (for the first time) will also send
1707 preempt_fold_need_resched();
1709 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1713 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1714 * traditionally all their work was done from the interrupt return
1715 * path. Now that we actually do some work, we need to make sure
1718 * Some archs already do call them, luckily irq_enter/exit nest
1721 * Arguably we should visit all archs and update all handlers,
1722 * however a fair share of IPIs are still resched only so this would
1723 * somewhat pessimize the simple resched case.
1726 sched_ttwu_pending();
1729 * Check if someone kicked us for doing the nohz idle load balance.
1731 if (unlikely(got_nohz_idle_kick())) {
1732 this_rq()->idle_balance = 1;
1733 raise_softirq_irqoff(SCHED_SOFTIRQ);
1738 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1740 struct rq *rq = cpu_rq(cpu);
1742 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1743 if (!set_nr_if_polling(rq->idle))
1744 smp_send_reschedule(cpu);
1746 trace_sched_wake_idle_without_ipi(cpu);
1750 void wake_up_if_idle(int cpu)
1752 struct rq *rq = cpu_rq(cpu);
1753 unsigned long flags;
1757 if (!is_idle_task(rcu_dereference(rq->curr)))
1760 if (set_nr_if_polling(rq->idle)) {
1761 trace_sched_wake_idle_without_ipi(cpu);
1763 raw_spin_lock_irqsave(&rq->lock, flags);
1764 if (is_idle_task(rq->curr))
1765 smp_send_reschedule(cpu);
1766 /* Else cpu is not in idle, do nothing here */
1767 raw_spin_unlock_irqrestore(&rq->lock, flags);
1774 bool cpus_share_cache(int this_cpu, int that_cpu)
1776 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1778 #endif /* CONFIG_SMP */
1780 static void ttwu_queue(struct task_struct *p, int cpu)
1782 struct rq *rq = cpu_rq(cpu);
1784 #if defined(CONFIG_SMP)
1785 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1786 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1787 ttwu_queue_remote(p, cpu);
1792 raw_spin_lock(&rq->lock);
1793 lockdep_pin_lock(&rq->lock);
1794 ttwu_do_activate(rq, p, 0);
1795 lockdep_unpin_lock(&rq->lock);
1796 raw_spin_unlock(&rq->lock);
1800 * Notes on Program-Order guarantees on SMP systems.
1804 * The basic program-order guarantee on SMP systems is that when a task [t]
1805 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1806 * execution on its new cpu [c1].
1808 * For migration (of runnable tasks) this is provided by the following means:
1810 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1811 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1812 * rq(c1)->lock (if not at the same time, then in that order).
1813 * C) LOCK of the rq(c1)->lock scheduling in task
1815 * Transitivity guarantees that B happens after A and C after B.
1816 * Note: we only require RCpc transitivity.
1817 * Note: the cpu doing B need not be c0 or c1
1826 * UNLOCK rq(0)->lock
1828 * LOCK rq(0)->lock // orders against CPU0
1830 * UNLOCK rq(0)->lock
1834 * UNLOCK rq(1)->lock
1836 * LOCK rq(1)->lock // orders against CPU2
1839 * UNLOCK rq(1)->lock
1842 * BLOCKING -- aka. SLEEP + WAKEUP
1844 * For blocking we (obviously) need to provide the same guarantee as for
1845 * migration. However the means are completely different as there is no lock
1846 * chain to provide order. Instead we do:
1848 * 1) smp_store_release(X->on_cpu, 0)
1849 * 2) smp_cond_acquire(!X->on_cpu)
1853 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1855 * LOCK rq(0)->lock LOCK X->pi_lock
1858 * smp_store_release(X->on_cpu, 0);
1860 * smp_cond_acquire(!X->on_cpu);
1866 * X->state = RUNNING
1867 * UNLOCK rq(2)->lock
1869 * LOCK rq(2)->lock // orders against CPU1
1872 * UNLOCK rq(2)->lock
1875 * UNLOCK rq(0)->lock
1878 * However; for wakeups there is a second guarantee we must provide, namely we
1879 * must observe the state that lead to our wakeup. That is, not only must our
1880 * task observe its own prior state, it must also observe the stores prior to
1883 * This means that any means of doing remote wakeups must order the CPU doing
1884 * the wakeup against the CPU the task is going to end up running on. This,
1885 * however, is already required for the regular Program-Order guarantee above,
1886 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1891 * try_to_wake_up - wake up a thread
1892 * @p: the thread to be awakened
1893 * @state: the mask of task states that can be woken
1894 * @wake_flags: wake modifier flags (WF_*)
1896 * Put it on the run-queue if it's not already there. The "current"
1897 * thread is always on the run-queue (except when the actual
1898 * re-schedule is in progress), and as such you're allowed to do
1899 * the simpler "current->state = TASK_RUNNING" to mark yourself
1900 * runnable without the overhead of this.
1902 * Return: %true if @p was woken up, %false if it was already running.
1903 * or @state didn't match @p's state.
1906 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1908 unsigned long flags;
1909 int cpu, success = 0;
1912 * If we are going to wake up a thread waiting for CONDITION we
1913 * need to ensure that CONDITION=1 done by the caller can not be
1914 * reordered with p->state check below. This pairs with mb() in
1915 * set_current_state() the waiting thread does.
1917 smp_mb__before_spinlock();
1918 raw_spin_lock_irqsave(&p->pi_lock, flags);
1919 if (!(p->state & state))
1922 trace_sched_waking(p);
1924 success = 1; /* we're going to change ->state */
1927 if (p->on_rq && ttwu_remote(p, wake_flags))
1932 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1933 * possible to, falsely, observe p->on_cpu == 0.
1935 * One must be running (->on_cpu == 1) in order to remove oneself
1936 * from the runqueue.
1938 * [S] ->on_cpu = 1; [L] ->on_rq
1942 * [S] ->on_rq = 0; [L] ->on_cpu
1944 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1945 * from the consecutive calls to schedule(); the first switching to our
1946 * task, the second putting it to sleep.
1951 * If the owning (remote) cpu is still in the middle of schedule() with
1952 * this task as prev, wait until its done referencing the task.
1954 * Pairs with the smp_store_release() in finish_lock_switch().
1956 * This ensures that tasks getting woken will be fully ordered against
1957 * their previous state and preserve Program Order.
1959 smp_cond_acquire(!p->on_cpu);
1961 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1962 p->state = TASK_WAKING;
1964 if (p->sched_class->task_waking)
1965 p->sched_class->task_waking(p);
1967 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1968 if (task_cpu(p) != cpu) {
1969 wake_flags |= WF_MIGRATED;
1970 set_task_cpu(p, cpu);
1972 #endif /* CONFIG_SMP */
1976 if (schedstat_enabled())
1977 ttwu_stat(p, cpu, wake_flags);
1979 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1985 * try_to_wake_up_local - try to wake up a local task with rq lock held
1986 * @p: the thread to be awakened
1988 * Put @p on the run-queue if it's not already there. The caller must
1989 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1992 static void try_to_wake_up_local(struct task_struct *p)
1994 struct rq *rq = task_rq(p);
1996 if (WARN_ON_ONCE(rq != this_rq()) ||
1997 WARN_ON_ONCE(p == current))
2000 lockdep_assert_held(&rq->lock);
2002 if (!raw_spin_trylock(&p->pi_lock)) {
2004 * This is OK, because current is on_cpu, which avoids it being
2005 * picked for load-balance and preemption/IRQs are still
2006 * disabled avoiding further scheduler activity on it and we've
2007 * not yet picked a replacement task.
2009 lockdep_unpin_lock(&rq->lock);
2010 raw_spin_unlock(&rq->lock);
2011 raw_spin_lock(&p->pi_lock);
2012 raw_spin_lock(&rq->lock);
2013 lockdep_pin_lock(&rq->lock);
2016 if (!(p->state & TASK_NORMAL))
2019 trace_sched_waking(p);
2021 if (!task_on_rq_queued(p))
2022 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2024 ttwu_do_wakeup(rq, p, 0);
2025 if (schedstat_enabled())
2026 ttwu_stat(p, smp_processor_id(), 0);
2028 raw_spin_unlock(&p->pi_lock);
2032 * wake_up_process - Wake up a specific process
2033 * @p: The process to be woken up.
2035 * Attempt to wake up the nominated process and move it to the set of runnable
2038 * Return: 1 if the process was woken up, 0 if it was already running.
2040 * It may be assumed that this function implies a write memory barrier before
2041 * changing the task state if and only if any tasks are woken up.
2043 int wake_up_process(struct task_struct *p)
2045 return try_to_wake_up(p, TASK_NORMAL, 0);
2047 EXPORT_SYMBOL(wake_up_process);
2049 int wake_up_state(struct task_struct *p, unsigned int state)
2051 return try_to_wake_up(p, state, 0);
2055 * This function clears the sched_dl_entity static params.
2057 void __dl_clear_params(struct task_struct *p)
2059 struct sched_dl_entity *dl_se = &p->dl;
2061 dl_se->dl_runtime = 0;
2062 dl_se->dl_deadline = 0;
2063 dl_se->dl_period = 0;
2067 dl_se->dl_throttled = 0;
2068 dl_se->dl_yielded = 0;
2072 * Perform scheduler related setup for a newly forked process p.
2073 * p is forked by current.
2075 * __sched_fork() is basic setup used by init_idle() too:
2077 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2082 p->se.exec_start = 0;
2083 p->se.sum_exec_runtime = 0;
2084 p->se.prev_sum_exec_runtime = 0;
2085 p->se.nr_migrations = 0;
2087 INIT_LIST_HEAD(&p->se.group_node);
2089 #ifdef CONFIG_FAIR_GROUP_SCHED
2090 p->se.cfs_rq = NULL;
2093 #ifdef CONFIG_SCHEDSTATS
2094 /* Even if schedstat is disabled, there should not be garbage */
2095 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2098 RB_CLEAR_NODE(&p->dl.rb_node);
2099 init_dl_task_timer(&p->dl);
2100 __dl_clear_params(p);
2102 INIT_LIST_HEAD(&p->rt.run_list);
2104 p->rt.time_slice = sched_rr_timeslice;
2108 #ifdef CONFIG_PREEMPT_NOTIFIERS
2109 INIT_HLIST_HEAD(&p->preempt_notifiers);
2112 #ifdef CONFIG_NUMA_BALANCING
2113 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2114 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2115 p->mm->numa_scan_seq = 0;
2118 if (clone_flags & CLONE_VM)
2119 p->numa_preferred_nid = current->numa_preferred_nid;
2121 p->numa_preferred_nid = -1;
2123 p->node_stamp = 0ULL;
2124 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2125 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2126 p->numa_work.next = &p->numa_work;
2127 p->numa_faults = NULL;
2128 p->last_task_numa_placement = 0;
2129 p->last_sum_exec_runtime = 0;
2131 p->numa_group = NULL;
2132 #endif /* CONFIG_NUMA_BALANCING */
2135 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2137 #ifdef CONFIG_NUMA_BALANCING
2139 void set_numabalancing_state(bool enabled)
2142 static_branch_enable(&sched_numa_balancing);
2144 static_branch_disable(&sched_numa_balancing);
2147 #ifdef CONFIG_PROC_SYSCTL
2148 int sysctl_numa_balancing(struct ctl_table *table, int write,
2149 void __user *buffer, size_t *lenp, loff_t *ppos)
2153 int state = static_branch_likely(&sched_numa_balancing);
2155 if (write && !capable(CAP_SYS_ADMIN))
2160 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2164 set_numabalancing_state(state);
2170 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2172 #ifdef CONFIG_SCHEDSTATS
2173 static void set_schedstats(bool enabled)
2176 static_branch_enable(&sched_schedstats);
2178 static_branch_disable(&sched_schedstats);
2181 void force_schedstat_enabled(void)
2183 if (!schedstat_enabled()) {
2184 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2185 static_branch_enable(&sched_schedstats);
2189 static int __init setup_schedstats(char *str)
2195 if (!strcmp(str, "enable")) {
2196 set_schedstats(true);
2198 } else if (!strcmp(str, "disable")) {
2199 set_schedstats(false);
2204 pr_warn("Unable to parse schedstats=\n");
2208 __setup("schedstats=", setup_schedstats);
2210 #ifdef CONFIG_PROC_SYSCTL
2211 int sysctl_schedstats(struct ctl_table *table, int write,
2212 void __user *buffer, size_t *lenp, loff_t *ppos)
2216 int state = static_branch_likely(&sched_schedstats);
2218 if (write && !capable(CAP_SYS_ADMIN))
2223 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2227 set_schedstats(state);
2234 * fork()/clone()-time setup:
2236 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2238 unsigned long flags;
2239 int cpu = get_cpu();
2241 __sched_fork(clone_flags, p);
2243 * We mark the process as running here. This guarantees that
2244 * nobody will actually run it, and a signal or other external
2245 * event cannot wake it up and insert it on the runqueue either.
2247 p->state = TASK_RUNNING;
2250 * Make sure we do not leak PI boosting priority to the child.
2252 p->prio = current->normal_prio;
2255 * Revert to default priority/policy on fork if requested.
2257 if (unlikely(p->sched_reset_on_fork)) {
2258 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2259 p->policy = SCHED_NORMAL;
2260 p->static_prio = NICE_TO_PRIO(0);
2262 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2263 p->static_prio = NICE_TO_PRIO(0);
2265 p->prio = p->normal_prio = __normal_prio(p);
2269 * We don't need the reset flag anymore after the fork. It has
2270 * fulfilled its duty:
2272 p->sched_reset_on_fork = 0;
2275 if (dl_prio(p->prio)) {
2278 } else if (rt_prio(p->prio)) {
2279 p->sched_class = &rt_sched_class;
2281 p->sched_class = &fair_sched_class;
2284 if (p->sched_class->task_fork)
2285 p->sched_class->task_fork(p);
2288 * The child is not yet in the pid-hash so no cgroup attach races,
2289 * and the cgroup is pinned to this child due to cgroup_fork()
2290 * is ran before sched_fork().
2292 * Silence PROVE_RCU.
2294 raw_spin_lock_irqsave(&p->pi_lock, flags);
2295 set_task_cpu(p, cpu);
2296 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2298 #ifdef CONFIG_SCHED_INFO
2299 if (likely(sched_info_on()))
2300 memset(&p->sched_info, 0, sizeof(p->sched_info));
2302 #if defined(CONFIG_SMP)
2305 init_task_preempt_count(p);
2307 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2308 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2315 unsigned long to_ratio(u64 period, u64 runtime)
2317 if (runtime == RUNTIME_INF)
2321 * Doing this here saves a lot of checks in all
2322 * the calling paths, and returning zero seems
2323 * safe for them anyway.
2328 return div64_u64(runtime << 20, period);
2332 inline struct dl_bw *dl_bw_of(int i)
2334 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2335 "sched RCU must be held");
2336 return &cpu_rq(i)->rd->dl_bw;
2339 static inline int dl_bw_cpus(int i)
2341 struct root_domain *rd = cpu_rq(i)->rd;
2344 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2345 "sched RCU must be held");
2346 for_each_cpu_and(i, rd->span, cpu_active_mask)
2352 inline struct dl_bw *dl_bw_of(int i)
2354 return &cpu_rq(i)->dl.dl_bw;
2357 static inline int dl_bw_cpus(int i)
2364 * We must be sure that accepting a new task (or allowing changing the
2365 * parameters of an existing one) is consistent with the bandwidth
2366 * constraints. If yes, this function also accordingly updates the currently
2367 * allocated bandwidth to reflect the new situation.
2369 * This function is called while holding p's rq->lock.
2371 * XXX we should delay bw change until the task's 0-lag point, see
2374 static int dl_overflow(struct task_struct *p, int policy,
2375 const struct sched_attr *attr)
2378 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2379 u64 period = attr->sched_period ?: attr->sched_deadline;
2380 u64 runtime = attr->sched_runtime;
2381 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2384 if (new_bw == p->dl.dl_bw)
2388 * Either if a task, enters, leave, or stays -deadline but changes
2389 * its parameters, we may need to update accordingly the total
2390 * allocated bandwidth of the container.
2392 raw_spin_lock(&dl_b->lock);
2393 cpus = dl_bw_cpus(task_cpu(p));
2394 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2395 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2396 __dl_add(dl_b, new_bw);
2398 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2399 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2400 __dl_clear(dl_b, p->dl.dl_bw);
2401 __dl_add(dl_b, new_bw);
2403 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2404 __dl_clear(dl_b, p->dl.dl_bw);
2407 raw_spin_unlock(&dl_b->lock);
2412 extern void init_dl_bw(struct dl_bw *dl_b);
2415 * wake_up_new_task - wake up a newly created task for the first time.
2417 * This function will do some initial scheduler statistics housekeeping
2418 * that must be done for every newly created context, then puts the task
2419 * on the runqueue and wakes it.
2421 void wake_up_new_task(struct task_struct *p)
2423 unsigned long flags;
2426 raw_spin_lock_irqsave(&p->pi_lock, flags);
2427 /* Initialize new task's runnable average */
2428 init_entity_runnable_average(&p->se);
2431 * Fork balancing, do it here and not earlier because:
2432 * - cpus_allowed can change in the fork path
2433 * - any previously selected cpu might disappear through hotplug
2435 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2438 rq = __task_rq_lock(p);
2439 activate_task(rq, p, 0);
2440 p->on_rq = TASK_ON_RQ_QUEUED;
2441 trace_sched_wakeup_new(p);
2442 check_preempt_curr(rq, p, WF_FORK);
2444 if (p->sched_class->task_woken) {
2446 * Nothing relies on rq->lock after this, so its fine to
2449 lockdep_unpin_lock(&rq->lock);
2450 p->sched_class->task_woken(rq, p);
2451 lockdep_pin_lock(&rq->lock);
2454 task_rq_unlock(rq, p, &flags);
2457 #ifdef CONFIG_PREEMPT_NOTIFIERS
2459 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2461 void preempt_notifier_inc(void)
2463 static_key_slow_inc(&preempt_notifier_key);
2465 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2467 void preempt_notifier_dec(void)
2469 static_key_slow_dec(&preempt_notifier_key);
2471 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2474 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2475 * @notifier: notifier struct to register
2477 void preempt_notifier_register(struct preempt_notifier *notifier)
2479 if (!static_key_false(&preempt_notifier_key))
2480 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2482 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2484 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2487 * preempt_notifier_unregister - no longer interested in preemption notifications
2488 * @notifier: notifier struct to unregister
2490 * This is *not* safe to call from within a preemption notifier.
2492 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2494 hlist_del(¬ifier->link);
2496 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2498 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2500 struct preempt_notifier *notifier;
2502 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2503 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2506 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2508 if (static_key_false(&preempt_notifier_key))
2509 __fire_sched_in_preempt_notifiers(curr);
2513 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2514 struct task_struct *next)
2516 struct preempt_notifier *notifier;
2518 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2519 notifier->ops->sched_out(notifier, next);
2522 static __always_inline void
2523 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2524 struct task_struct *next)
2526 if (static_key_false(&preempt_notifier_key))
2527 __fire_sched_out_preempt_notifiers(curr, next);
2530 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2532 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2537 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2538 struct task_struct *next)
2542 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2545 * prepare_task_switch - prepare to switch tasks
2546 * @rq: the runqueue preparing to switch
2547 * @prev: the current task that is being switched out
2548 * @next: the task we are going to switch to.
2550 * This is called with the rq lock held and interrupts off. It must
2551 * be paired with a subsequent finish_task_switch after the context
2554 * prepare_task_switch sets up locking and calls architecture specific
2558 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2559 struct task_struct *next)
2561 sched_info_switch(rq, prev, next);
2562 perf_event_task_sched_out(prev, next);
2563 fire_sched_out_preempt_notifiers(prev, next);
2564 prepare_lock_switch(rq, next);
2565 prepare_arch_switch(next);
2569 * finish_task_switch - clean up after a task-switch
2570 * @prev: the thread we just switched away from.
2572 * finish_task_switch must be called after the context switch, paired
2573 * with a prepare_task_switch call before the context switch.
2574 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2575 * and do any other architecture-specific cleanup actions.
2577 * Note that we may have delayed dropping an mm in context_switch(). If
2578 * so, we finish that here outside of the runqueue lock. (Doing it
2579 * with the lock held can cause deadlocks; see schedule() for
2582 * The context switch have flipped the stack from under us and restored the
2583 * local variables which were saved when this task called schedule() in the
2584 * past. prev == current is still correct but we need to recalculate this_rq
2585 * because prev may have moved to another CPU.
2587 static struct rq *finish_task_switch(struct task_struct *prev)
2588 __releases(rq->lock)
2590 struct rq *rq = this_rq();
2591 struct mm_struct *mm = rq->prev_mm;
2595 * The previous task will have left us with a preempt_count of 2
2596 * because it left us after:
2599 * preempt_disable(); // 1
2601 * raw_spin_lock_irq(&rq->lock) // 2
2603 * Also, see FORK_PREEMPT_COUNT.
2605 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2606 "corrupted preempt_count: %s/%d/0x%x\n",
2607 current->comm, current->pid, preempt_count()))
2608 preempt_count_set(FORK_PREEMPT_COUNT);
2613 * A task struct has one reference for the use as "current".
2614 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2615 * schedule one last time. The schedule call will never return, and
2616 * the scheduled task must drop that reference.
2618 * We must observe prev->state before clearing prev->on_cpu (in
2619 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2620 * running on another CPU and we could rave with its RUNNING -> DEAD
2621 * transition, resulting in a double drop.
2623 prev_state = prev->state;
2624 vtime_task_switch(prev);
2625 perf_event_task_sched_in(prev, current);
2626 finish_lock_switch(rq, prev);
2627 finish_arch_post_lock_switch();
2629 fire_sched_in_preempt_notifiers(current);
2632 if (unlikely(prev_state == TASK_DEAD)) {
2633 if (prev->sched_class->task_dead)
2634 prev->sched_class->task_dead(prev);
2637 * Remove function-return probe instances associated with this
2638 * task and put them back on the free list.
2640 kprobe_flush_task(prev);
2641 put_task_struct(prev);
2644 tick_nohz_task_switch();
2650 /* rq->lock is NOT held, but preemption is disabled */
2651 static void __balance_callback(struct rq *rq)
2653 struct callback_head *head, *next;
2654 void (*func)(struct rq *rq);
2655 unsigned long flags;
2657 raw_spin_lock_irqsave(&rq->lock, flags);
2658 head = rq->balance_callback;
2659 rq->balance_callback = NULL;
2661 func = (void (*)(struct rq *))head->func;
2668 raw_spin_unlock_irqrestore(&rq->lock, flags);
2671 static inline void balance_callback(struct rq *rq)
2673 if (unlikely(rq->balance_callback))
2674 __balance_callback(rq);
2679 static inline void balance_callback(struct rq *rq)
2686 * schedule_tail - first thing a freshly forked thread must call.
2687 * @prev: the thread we just switched away from.
2689 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2690 __releases(rq->lock)
2695 * New tasks start with FORK_PREEMPT_COUNT, see there and
2696 * finish_task_switch() for details.
2698 * finish_task_switch() will drop rq->lock() and lower preempt_count
2699 * and the preempt_enable() will end up enabling preemption (on
2700 * PREEMPT_COUNT kernels).
2703 rq = finish_task_switch(prev);
2704 balance_callback(rq);
2707 if (current->set_child_tid)
2708 put_user(task_pid_vnr(current), current->set_child_tid);
2712 * context_switch - switch to the new MM and the new thread's register state.
2714 static __always_inline struct rq *
2715 context_switch(struct rq *rq, struct task_struct *prev,
2716 struct task_struct *next)
2718 struct mm_struct *mm, *oldmm;
2720 prepare_task_switch(rq, prev, next);
2723 oldmm = prev->active_mm;
2725 * For paravirt, this is coupled with an exit in switch_to to
2726 * combine the page table reload and the switch backend into
2729 arch_start_context_switch(prev);
2732 next->active_mm = oldmm;
2733 atomic_inc(&oldmm->mm_count);
2734 enter_lazy_tlb(oldmm, next);
2736 switch_mm(oldmm, mm, next);
2739 prev->active_mm = NULL;
2740 rq->prev_mm = oldmm;
2743 * Since the runqueue lock will be released by the next
2744 * task (which is an invalid locking op but in the case
2745 * of the scheduler it's an obvious special-case), so we
2746 * do an early lockdep release here:
2748 lockdep_unpin_lock(&rq->lock);
2749 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2751 /* Here we just switch the register state and the stack. */
2752 switch_to(prev, next, prev);
2755 return finish_task_switch(prev);
2759 * nr_running and nr_context_switches:
2761 * externally visible scheduler statistics: current number of runnable
2762 * threads, total number of context switches performed since bootup.
2764 unsigned long nr_running(void)
2766 unsigned long i, sum = 0;
2768 for_each_online_cpu(i)
2769 sum += cpu_rq(i)->nr_running;
2775 * Check if only the current task is running on the cpu.
2777 * Caution: this function does not check that the caller has disabled
2778 * preemption, thus the result might have a time-of-check-to-time-of-use
2779 * race. The caller is responsible to use it correctly, for example:
2781 * - from a non-preemptable section (of course)
2783 * - from a thread that is bound to a single CPU
2785 * - in a loop with very short iterations (e.g. a polling loop)
2787 bool single_task_running(void)
2789 return raw_rq()->nr_running == 1;
2791 EXPORT_SYMBOL(single_task_running);
2793 unsigned long long nr_context_switches(void)
2796 unsigned long long sum = 0;
2798 for_each_possible_cpu(i)
2799 sum += cpu_rq(i)->nr_switches;
2804 unsigned long nr_iowait(void)
2806 unsigned long i, sum = 0;
2808 for_each_possible_cpu(i)
2809 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2814 unsigned long nr_iowait_cpu(int cpu)
2816 struct rq *this = cpu_rq(cpu);
2817 return atomic_read(&this->nr_iowait);
2820 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2822 struct rq *rq = this_rq();
2823 *nr_waiters = atomic_read(&rq->nr_iowait);
2824 *load = rq->load.weight;
2830 * sched_exec - execve() is a valuable balancing opportunity, because at
2831 * this point the task has the smallest effective memory and cache footprint.
2833 void sched_exec(void)
2835 struct task_struct *p = current;
2836 unsigned long flags;
2839 raw_spin_lock_irqsave(&p->pi_lock, flags);
2840 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2841 if (dest_cpu == smp_processor_id())
2844 if (likely(cpu_active(dest_cpu))) {
2845 struct migration_arg arg = { p, dest_cpu };
2847 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2848 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2852 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2857 DEFINE_PER_CPU(struct kernel_stat, kstat);
2858 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2860 EXPORT_PER_CPU_SYMBOL(kstat);
2861 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2864 * Return accounted runtime for the task.
2865 * In case the task is currently running, return the runtime plus current's
2866 * pending runtime that have not been accounted yet.
2868 unsigned long long task_sched_runtime(struct task_struct *p)
2870 unsigned long flags;
2874 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2876 * 64-bit doesn't need locks to atomically read a 64bit value.
2877 * So we have a optimization chance when the task's delta_exec is 0.
2878 * Reading ->on_cpu is racy, but this is ok.
2880 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2881 * If we race with it entering cpu, unaccounted time is 0. This is
2882 * indistinguishable from the read occurring a few cycles earlier.
2883 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2884 * been accounted, so we're correct here as well.
2886 if (!p->on_cpu || !task_on_rq_queued(p))
2887 return p->se.sum_exec_runtime;
2890 rq = task_rq_lock(p, &flags);
2892 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2893 * project cycles that may never be accounted to this
2894 * thread, breaking clock_gettime().
2896 if (task_current(rq, p) && task_on_rq_queued(p)) {
2897 update_rq_clock(rq);
2898 p->sched_class->update_curr(rq);
2900 ns = p->se.sum_exec_runtime;
2901 task_rq_unlock(rq, p, &flags);
2907 * This function gets called by the timer code, with HZ frequency.
2908 * We call it with interrupts disabled.
2910 void scheduler_tick(void)
2912 int cpu = smp_processor_id();
2913 struct rq *rq = cpu_rq(cpu);
2914 struct task_struct *curr = rq->curr;
2918 raw_spin_lock(&rq->lock);
2919 update_rq_clock(rq);
2920 curr->sched_class->task_tick(rq, curr, 0);
2921 update_cpu_load_active(rq);
2922 calc_global_load_tick(rq);
2923 raw_spin_unlock(&rq->lock);
2925 perf_event_task_tick();
2928 rq->idle_balance = idle_cpu(cpu);
2929 trigger_load_balance(rq);
2931 rq_last_tick_reset(rq);
2934 #ifdef CONFIG_NO_HZ_FULL
2936 * scheduler_tick_max_deferment
2938 * Keep at least one tick per second when a single
2939 * active task is running because the scheduler doesn't
2940 * yet completely support full dynticks environment.
2942 * This makes sure that uptime, CFS vruntime, load
2943 * balancing, etc... continue to move forward, even
2944 * with a very low granularity.
2946 * Return: Maximum deferment in nanoseconds.
2948 u64 scheduler_tick_max_deferment(void)
2950 struct rq *rq = this_rq();
2951 unsigned long next, now = READ_ONCE(jiffies);
2953 next = rq->last_sched_tick + HZ;
2955 if (time_before_eq(next, now))
2958 return jiffies_to_nsecs(next - now);
2962 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2963 defined(CONFIG_PREEMPT_TRACER))
2965 void preempt_count_add(int val)
2967 #ifdef CONFIG_DEBUG_PREEMPT
2971 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2974 __preempt_count_add(val);
2975 #ifdef CONFIG_DEBUG_PREEMPT
2977 * Spinlock count overflowing soon?
2979 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2982 if (preempt_count() == val) {
2983 unsigned long ip = get_lock_parent_ip();
2984 #ifdef CONFIG_DEBUG_PREEMPT
2985 current->preempt_disable_ip = ip;
2987 trace_preempt_off(CALLER_ADDR0, ip);
2990 EXPORT_SYMBOL(preempt_count_add);
2991 NOKPROBE_SYMBOL(preempt_count_add);
2993 void preempt_count_sub(int val)
2995 #ifdef CONFIG_DEBUG_PREEMPT
2999 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3002 * Is the spinlock portion underflowing?
3004 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3005 !(preempt_count() & PREEMPT_MASK)))
3009 if (preempt_count() == val)
3010 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3011 __preempt_count_sub(val);
3013 EXPORT_SYMBOL(preempt_count_sub);
3014 NOKPROBE_SYMBOL(preempt_count_sub);
3019 * Print scheduling while atomic bug:
3021 static noinline void __schedule_bug(struct task_struct *prev)
3023 if (oops_in_progress)
3026 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3027 prev->comm, prev->pid, preempt_count());
3029 debug_show_held_locks(prev);
3031 if (irqs_disabled())
3032 print_irqtrace_events(prev);
3033 #ifdef CONFIG_DEBUG_PREEMPT
3034 if (in_atomic_preempt_off()) {
3035 pr_err("Preemption disabled at:");
3036 print_ip_sym(current->preempt_disable_ip);
3041 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3045 * Various schedule()-time debugging checks and statistics:
3047 static inline void schedule_debug(struct task_struct *prev)
3049 #ifdef CONFIG_SCHED_STACK_END_CHECK
3050 if (task_stack_end_corrupted(prev))
3051 panic("corrupted stack end detected inside scheduler\n");
3054 if (unlikely(in_atomic_preempt_off())) {
3055 __schedule_bug(prev);
3056 preempt_count_set(PREEMPT_DISABLED);
3060 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3062 schedstat_inc(this_rq(), sched_count);
3066 * Pick up the highest-prio task:
3068 static inline struct task_struct *
3069 pick_next_task(struct rq *rq, struct task_struct *prev)
3071 const struct sched_class *class = &fair_sched_class;
3072 struct task_struct *p;
3075 * Optimization: we know that if all tasks are in
3076 * the fair class we can call that function directly:
3078 if (likely(prev->sched_class == class &&
3079 rq->nr_running == rq->cfs.h_nr_running)) {
3080 p = fair_sched_class.pick_next_task(rq, prev);
3081 if (unlikely(p == RETRY_TASK))
3084 /* assumes fair_sched_class->next == idle_sched_class */
3086 p = idle_sched_class.pick_next_task(rq, prev);
3092 for_each_class(class) {
3093 p = class->pick_next_task(rq, prev);
3095 if (unlikely(p == RETRY_TASK))
3101 BUG(); /* the idle class will always have a runnable task */
3105 * __schedule() is the main scheduler function.
3107 * The main means of driving the scheduler and thus entering this function are:
3109 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3111 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3112 * paths. For example, see arch/x86/entry_64.S.
3114 * To drive preemption between tasks, the scheduler sets the flag in timer
3115 * interrupt handler scheduler_tick().
3117 * 3. Wakeups don't really cause entry into schedule(). They add a
3118 * task to the run-queue and that's it.
3120 * Now, if the new task added to the run-queue preempts the current
3121 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3122 * called on the nearest possible occasion:
3124 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3126 * - in syscall or exception context, at the next outmost
3127 * preempt_enable(). (this might be as soon as the wake_up()'s
3130 * - in IRQ context, return from interrupt-handler to
3131 * preemptible context
3133 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3136 * - cond_resched() call
3137 * - explicit schedule() call
3138 * - return from syscall or exception to user-space
3139 * - return from interrupt-handler to user-space
3141 * WARNING: must be called with preemption disabled!
3143 static void __sched notrace __schedule(bool preempt)
3145 struct task_struct *prev, *next;
3146 unsigned long *switch_count;
3150 cpu = smp_processor_id();
3155 * do_exit() calls schedule() with preemption disabled as an exception;
3156 * however we must fix that up, otherwise the next task will see an
3157 * inconsistent (higher) preempt count.
3159 * It also avoids the below schedule_debug() test from complaining
3162 if (unlikely(prev->state == TASK_DEAD))
3163 preempt_enable_no_resched_notrace();
3165 schedule_debug(prev);
3167 if (sched_feat(HRTICK))
3170 local_irq_disable();
3171 rcu_note_context_switch();
3174 * Make sure that signal_pending_state()->signal_pending() below
3175 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3176 * done by the caller to avoid the race with signal_wake_up().
3178 smp_mb__before_spinlock();
3179 raw_spin_lock(&rq->lock);
3180 lockdep_pin_lock(&rq->lock);
3182 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3184 switch_count = &prev->nivcsw;
3185 if (!preempt && prev->state) {
3186 if (unlikely(signal_pending_state(prev->state, prev))) {
3187 prev->state = TASK_RUNNING;
3189 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3193 * If a worker went to sleep, notify and ask workqueue
3194 * whether it wants to wake up a task to maintain
3197 if (prev->flags & PF_WQ_WORKER) {
3198 struct task_struct *to_wakeup;
3200 to_wakeup = wq_worker_sleeping(prev);
3202 try_to_wake_up_local(to_wakeup);
3205 switch_count = &prev->nvcsw;
3208 if (task_on_rq_queued(prev))
3209 update_rq_clock(rq);
3211 next = pick_next_task(rq, prev);
3212 clear_tsk_need_resched(prev);
3213 clear_preempt_need_resched();
3214 rq->clock_skip_update = 0;
3216 if (likely(prev != next)) {
3221 trace_sched_switch(preempt, prev, next);
3222 rq = context_switch(rq, prev, next); /* unlocks the rq */
3224 lockdep_unpin_lock(&rq->lock);
3225 raw_spin_unlock_irq(&rq->lock);
3228 balance_callback(rq);
3230 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3232 static inline void sched_submit_work(struct task_struct *tsk)
3234 if (!tsk->state || tsk_is_pi_blocked(tsk))
3237 * If we are going to sleep and we have plugged IO queued,
3238 * make sure to submit it to avoid deadlocks.
3240 if (blk_needs_flush_plug(tsk))
3241 blk_schedule_flush_plug(tsk);
3244 asmlinkage __visible void __sched schedule(void)
3246 struct task_struct *tsk = current;
3248 sched_submit_work(tsk);
3252 sched_preempt_enable_no_resched();
3253 } while (need_resched());
3255 EXPORT_SYMBOL(schedule);
3257 #ifdef CONFIG_CONTEXT_TRACKING
3258 asmlinkage __visible void __sched schedule_user(void)
3261 * If we come here after a random call to set_need_resched(),
3262 * or we have been woken up remotely but the IPI has not yet arrived,
3263 * we haven't yet exited the RCU idle mode. Do it here manually until
3264 * we find a better solution.
3266 * NB: There are buggy callers of this function. Ideally we
3267 * should warn if prev_state != CONTEXT_USER, but that will trigger
3268 * too frequently to make sense yet.
3270 enum ctx_state prev_state = exception_enter();
3272 exception_exit(prev_state);
3277 * schedule_preempt_disabled - called with preemption disabled
3279 * Returns with preemption disabled. Note: preempt_count must be 1
3281 void __sched schedule_preempt_disabled(void)
3283 sched_preempt_enable_no_resched();
3288 static void __sched notrace preempt_schedule_common(void)
3291 preempt_disable_notrace();
3293 preempt_enable_no_resched_notrace();
3296 * Check again in case we missed a preemption opportunity
3297 * between schedule and now.
3299 } while (need_resched());
3302 #ifdef CONFIG_PREEMPT
3304 * this is the entry point to schedule() from in-kernel preemption
3305 * off of preempt_enable. Kernel preemptions off return from interrupt
3306 * occur there and call schedule directly.
3308 asmlinkage __visible void __sched notrace preempt_schedule(void)
3311 * If there is a non-zero preempt_count or interrupts are disabled,
3312 * we do not want to preempt the current task. Just return..
3314 if (likely(!preemptible()))
3317 preempt_schedule_common();
3319 NOKPROBE_SYMBOL(preempt_schedule);
3320 EXPORT_SYMBOL(preempt_schedule);
3323 * preempt_schedule_notrace - preempt_schedule called by tracing
3325 * The tracing infrastructure uses preempt_enable_notrace to prevent
3326 * recursion and tracing preempt enabling caused by the tracing
3327 * infrastructure itself. But as tracing can happen in areas coming
3328 * from userspace or just about to enter userspace, a preempt enable
3329 * can occur before user_exit() is called. This will cause the scheduler
3330 * to be called when the system is still in usermode.
3332 * To prevent this, the preempt_enable_notrace will use this function
3333 * instead of preempt_schedule() to exit user context if needed before
3334 * calling the scheduler.
3336 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3338 enum ctx_state prev_ctx;
3340 if (likely(!preemptible()))
3344 preempt_disable_notrace();
3346 * Needs preempt disabled in case user_exit() is traced
3347 * and the tracer calls preempt_enable_notrace() causing
3348 * an infinite recursion.
3350 prev_ctx = exception_enter();
3352 exception_exit(prev_ctx);
3354 preempt_enable_no_resched_notrace();
3355 } while (need_resched());
3357 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3359 #endif /* CONFIG_PREEMPT */
3362 * this is the entry point to schedule() from kernel preemption
3363 * off of irq context.
3364 * Note, that this is called and return with irqs disabled. This will
3365 * protect us against recursive calling from irq.
3367 asmlinkage __visible void __sched preempt_schedule_irq(void)
3369 enum ctx_state prev_state;
3371 /* Catch callers which need to be fixed */
3372 BUG_ON(preempt_count() || !irqs_disabled());
3374 prev_state = exception_enter();
3380 local_irq_disable();
3381 sched_preempt_enable_no_resched();
3382 } while (need_resched());
3384 exception_exit(prev_state);
3387 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3390 return try_to_wake_up(curr->private, mode, wake_flags);
3392 EXPORT_SYMBOL(default_wake_function);
3394 #ifdef CONFIG_RT_MUTEXES
3397 * rt_mutex_setprio - set the current priority of a task
3399 * @prio: prio value (kernel-internal form)
3401 * This function changes the 'effective' priority of a task. It does
3402 * not touch ->normal_prio like __setscheduler().
3404 * Used by the rt_mutex code to implement priority inheritance
3405 * logic. Call site only calls if the priority of the task changed.
3407 void rt_mutex_setprio(struct task_struct *p, int prio)
3409 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3411 const struct sched_class *prev_class;
3413 BUG_ON(prio > MAX_PRIO);
3415 rq = __task_rq_lock(p);
3418 * Idle task boosting is a nono in general. There is one
3419 * exception, when PREEMPT_RT and NOHZ is active:
3421 * The idle task calls get_next_timer_interrupt() and holds
3422 * the timer wheel base->lock on the CPU and another CPU wants
3423 * to access the timer (probably to cancel it). We can safely
3424 * ignore the boosting request, as the idle CPU runs this code
3425 * with interrupts disabled and will complete the lock
3426 * protected section without being interrupted. So there is no
3427 * real need to boost.
3429 if (unlikely(p == rq->idle)) {
3430 WARN_ON(p != rq->curr);
3431 WARN_ON(p->pi_blocked_on);
3435 trace_sched_pi_setprio(p, prio);
3438 if (oldprio == prio)
3439 queue_flag &= ~DEQUEUE_MOVE;
3441 prev_class = p->sched_class;
3442 queued = task_on_rq_queued(p);
3443 running = task_current(rq, p);
3445 dequeue_task(rq, p, queue_flag);
3447 put_prev_task(rq, p);
3450 * Boosting condition are:
3451 * 1. -rt task is running and holds mutex A
3452 * --> -dl task blocks on mutex A
3454 * 2. -dl task is running and holds mutex A
3455 * --> -dl task blocks on mutex A and could preempt the
3458 if (dl_prio(prio)) {
3459 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3460 if (!dl_prio(p->normal_prio) ||
3461 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3462 p->dl.dl_boosted = 1;
3463 queue_flag |= ENQUEUE_REPLENISH;
3465 p->dl.dl_boosted = 0;
3466 p->sched_class = &dl_sched_class;
3467 } else if (rt_prio(prio)) {
3468 if (dl_prio(oldprio))
3469 p->dl.dl_boosted = 0;
3471 queue_flag |= ENQUEUE_HEAD;
3472 p->sched_class = &rt_sched_class;
3474 if (dl_prio(oldprio))
3475 p->dl.dl_boosted = 0;
3476 if (rt_prio(oldprio))
3478 p->sched_class = &fair_sched_class;
3484 p->sched_class->set_curr_task(rq);
3486 enqueue_task(rq, p, queue_flag);
3488 check_class_changed(rq, p, prev_class, oldprio);
3490 preempt_disable(); /* avoid rq from going away on us */
3491 __task_rq_unlock(rq);
3493 balance_callback(rq);
3498 void set_user_nice(struct task_struct *p, long nice)
3500 int old_prio, delta, queued;
3501 unsigned long flags;
3504 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3507 * We have to be careful, if called from sys_setpriority(),
3508 * the task might be in the middle of scheduling on another CPU.
3510 rq = task_rq_lock(p, &flags);
3512 * The RT priorities are set via sched_setscheduler(), but we still
3513 * allow the 'normal' nice value to be set - but as expected
3514 * it wont have any effect on scheduling until the task is
3515 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3517 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3518 p->static_prio = NICE_TO_PRIO(nice);
3521 queued = task_on_rq_queued(p);
3523 dequeue_task(rq, p, DEQUEUE_SAVE);
3525 p->static_prio = NICE_TO_PRIO(nice);
3528 p->prio = effective_prio(p);
3529 delta = p->prio - old_prio;
3532 enqueue_task(rq, p, ENQUEUE_RESTORE);
3534 * If the task increased its priority or is running and
3535 * lowered its priority, then reschedule its CPU:
3537 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3541 task_rq_unlock(rq, p, &flags);
3543 EXPORT_SYMBOL(set_user_nice);
3546 * can_nice - check if a task can reduce its nice value
3550 int can_nice(const struct task_struct *p, const int nice)
3552 /* convert nice value [19,-20] to rlimit style value [1,40] */
3553 int nice_rlim = nice_to_rlimit(nice);
3555 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3556 capable(CAP_SYS_NICE));
3559 #ifdef __ARCH_WANT_SYS_NICE
3562 * sys_nice - change the priority of the current process.
3563 * @increment: priority increment
3565 * sys_setpriority is a more generic, but much slower function that
3566 * does similar things.
3568 SYSCALL_DEFINE1(nice, int, increment)
3573 * Setpriority might change our priority at the same moment.
3574 * We don't have to worry. Conceptually one call occurs first
3575 * and we have a single winner.
3577 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3578 nice = task_nice(current) + increment;
3580 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3581 if (increment < 0 && !can_nice(current, nice))
3584 retval = security_task_setnice(current, nice);
3588 set_user_nice(current, nice);
3595 * task_prio - return the priority value of a given task.
3596 * @p: the task in question.
3598 * Return: The priority value as seen by users in /proc.
3599 * RT tasks are offset by -200. Normal tasks are centered
3600 * around 0, value goes from -16 to +15.
3602 int task_prio(const struct task_struct *p)
3604 return p->prio - MAX_RT_PRIO;
3608 * idle_cpu - is a given cpu idle currently?
3609 * @cpu: the processor in question.
3611 * Return: 1 if the CPU is currently idle. 0 otherwise.
3613 int idle_cpu(int cpu)
3615 struct rq *rq = cpu_rq(cpu);
3617 if (rq->curr != rq->idle)
3624 if (!llist_empty(&rq->wake_list))
3632 * idle_task - return the idle task for a given cpu.
3633 * @cpu: the processor in question.
3635 * Return: The idle task for the cpu @cpu.
3637 struct task_struct *idle_task(int cpu)
3639 return cpu_rq(cpu)->idle;
3643 * find_process_by_pid - find a process with a matching PID value.
3644 * @pid: the pid in question.
3646 * The task of @pid, if found. %NULL otherwise.
3648 static struct task_struct *find_process_by_pid(pid_t pid)
3650 return pid ? find_task_by_vpid(pid) : current;
3654 * This function initializes the sched_dl_entity of a newly becoming
3655 * SCHED_DEADLINE task.
3657 * Only the static values are considered here, the actual runtime and the
3658 * absolute deadline will be properly calculated when the task is enqueued
3659 * for the first time with its new policy.
3662 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3664 struct sched_dl_entity *dl_se = &p->dl;
3666 dl_se->dl_runtime = attr->sched_runtime;
3667 dl_se->dl_deadline = attr->sched_deadline;
3668 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3669 dl_se->flags = attr->sched_flags;
3670 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3673 * Changing the parameters of a task is 'tricky' and we're not doing
3674 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3676 * What we SHOULD do is delay the bandwidth release until the 0-lag
3677 * point. This would include retaining the task_struct until that time
3678 * and change dl_overflow() to not immediately decrement the current
3681 * Instead we retain the current runtime/deadline and let the new
3682 * parameters take effect after the current reservation period lapses.
3683 * This is safe (albeit pessimistic) because the 0-lag point is always
3684 * before the current scheduling deadline.
3686 * We can still have temporary overloads because we do not delay the
3687 * change in bandwidth until that time; so admission control is
3688 * not on the safe side. It does however guarantee tasks will never
3689 * consume more than promised.
3694 * sched_setparam() passes in -1 for its policy, to let the functions
3695 * it calls know not to change it.
3697 #define SETPARAM_POLICY -1
3699 static void __setscheduler_params(struct task_struct *p,
3700 const struct sched_attr *attr)
3702 int policy = attr->sched_policy;
3704 if (policy == SETPARAM_POLICY)
3709 if (dl_policy(policy))
3710 __setparam_dl(p, attr);
3711 else if (fair_policy(policy))
3712 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3715 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3716 * !rt_policy. Always setting this ensures that things like
3717 * getparam()/getattr() don't report silly values for !rt tasks.
3719 p->rt_priority = attr->sched_priority;
3720 p->normal_prio = normal_prio(p);
3724 /* Actually do priority change: must hold pi & rq lock. */
3725 static void __setscheduler(struct rq *rq, struct task_struct *p,
3726 const struct sched_attr *attr, bool keep_boost)
3728 __setscheduler_params(p, attr);
3731 * Keep a potential priority boosting if called from
3732 * sched_setscheduler().
3735 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3737 p->prio = normal_prio(p);
3739 if (dl_prio(p->prio))
3740 p->sched_class = &dl_sched_class;
3741 else if (rt_prio(p->prio))
3742 p->sched_class = &rt_sched_class;
3744 p->sched_class = &fair_sched_class;
3748 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3750 struct sched_dl_entity *dl_se = &p->dl;
3752 attr->sched_priority = p->rt_priority;
3753 attr->sched_runtime = dl_se->dl_runtime;
3754 attr->sched_deadline = dl_se->dl_deadline;
3755 attr->sched_period = dl_se->dl_period;
3756 attr->sched_flags = dl_se->flags;
3760 * This function validates the new parameters of a -deadline task.
3761 * We ask for the deadline not being zero, and greater or equal
3762 * than the runtime, as well as the period of being zero or
3763 * greater than deadline. Furthermore, we have to be sure that
3764 * user parameters are above the internal resolution of 1us (we
3765 * check sched_runtime only since it is always the smaller one) and
3766 * below 2^63 ns (we have to check both sched_deadline and
3767 * sched_period, as the latter can be zero).
3770 __checkparam_dl(const struct sched_attr *attr)
3773 if (attr->sched_deadline == 0)
3777 * Since we truncate DL_SCALE bits, make sure we're at least
3780 if (attr->sched_runtime < (1ULL << DL_SCALE))
3784 * Since we use the MSB for wrap-around and sign issues, make
3785 * sure it's not set (mind that period can be equal to zero).
3787 if (attr->sched_deadline & (1ULL << 63) ||
3788 attr->sched_period & (1ULL << 63))
3791 /* runtime <= deadline <= period (if period != 0) */
3792 if ((attr->sched_period != 0 &&
3793 attr->sched_period < attr->sched_deadline) ||
3794 attr->sched_deadline < attr->sched_runtime)
3801 * check the target process has a UID that matches the current process's
3803 static bool check_same_owner(struct task_struct *p)
3805 const struct cred *cred = current_cred(), *pcred;
3809 pcred = __task_cred(p);
3810 match = (uid_eq(cred->euid, pcred->euid) ||
3811 uid_eq(cred->euid, pcred->uid));
3816 static bool dl_param_changed(struct task_struct *p,
3817 const struct sched_attr *attr)
3819 struct sched_dl_entity *dl_se = &p->dl;
3821 if (dl_se->dl_runtime != attr->sched_runtime ||
3822 dl_se->dl_deadline != attr->sched_deadline ||
3823 dl_se->dl_period != attr->sched_period ||
3824 dl_se->flags != attr->sched_flags)
3830 static int __sched_setscheduler(struct task_struct *p,
3831 const struct sched_attr *attr,
3834 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3835 MAX_RT_PRIO - 1 - attr->sched_priority;
3836 int retval, oldprio, oldpolicy = -1, queued, running;
3837 int new_effective_prio, policy = attr->sched_policy;
3838 unsigned long flags;
3839 const struct sched_class *prev_class;
3842 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3844 /* may grab non-irq protected spin_locks */
3845 BUG_ON(in_interrupt());
3847 /* double check policy once rq lock held */
3849 reset_on_fork = p->sched_reset_on_fork;
3850 policy = oldpolicy = p->policy;
3852 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3854 if (!valid_policy(policy))
3858 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3862 * Valid priorities for SCHED_FIFO and SCHED_RR are
3863 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3864 * SCHED_BATCH and SCHED_IDLE is 0.
3866 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3867 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3869 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3870 (rt_policy(policy) != (attr->sched_priority != 0)))
3874 * Allow unprivileged RT tasks to decrease priority:
3876 if (user && !capable(CAP_SYS_NICE)) {
3877 if (fair_policy(policy)) {
3878 if (attr->sched_nice < task_nice(p) &&
3879 !can_nice(p, attr->sched_nice))
3883 if (rt_policy(policy)) {
3884 unsigned long rlim_rtprio =
3885 task_rlimit(p, RLIMIT_RTPRIO);
3887 /* can't set/change the rt policy */
3888 if (policy != p->policy && !rlim_rtprio)
3891 /* can't increase priority */
3892 if (attr->sched_priority > p->rt_priority &&
3893 attr->sched_priority > rlim_rtprio)
3898 * Can't set/change SCHED_DEADLINE policy at all for now
3899 * (safest behavior); in the future we would like to allow
3900 * unprivileged DL tasks to increase their relative deadline
3901 * or reduce their runtime (both ways reducing utilization)
3903 if (dl_policy(policy))
3907 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3908 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3910 if (idle_policy(p->policy) && !idle_policy(policy)) {
3911 if (!can_nice(p, task_nice(p)))
3915 /* can't change other user's priorities */
3916 if (!check_same_owner(p))
3919 /* Normal users shall not reset the sched_reset_on_fork flag */
3920 if (p->sched_reset_on_fork && !reset_on_fork)
3925 retval = security_task_setscheduler(p);
3931 * make sure no PI-waiters arrive (or leave) while we are
3932 * changing the priority of the task:
3934 * To be able to change p->policy safely, the appropriate
3935 * runqueue lock must be held.
3937 rq = task_rq_lock(p, &flags);
3940 * Changing the policy of the stop threads its a very bad idea
3942 if (p == rq->stop) {
3943 task_rq_unlock(rq, p, &flags);
3948 * If not changing anything there's no need to proceed further,
3949 * but store a possible modification of reset_on_fork.
3951 if (unlikely(policy == p->policy)) {
3952 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3954 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3956 if (dl_policy(policy) && dl_param_changed(p, attr))
3959 p->sched_reset_on_fork = reset_on_fork;
3960 task_rq_unlock(rq, p, &flags);
3966 #ifdef CONFIG_RT_GROUP_SCHED
3968 * Do not allow realtime tasks into groups that have no runtime
3971 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3972 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3973 !task_group_is_autogroup(task_group(p))) {
3974 task_rq_unlock(rq, p, &flags);
3979 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3980 cpumask_t *span = rq->rd->span;
3983 * Don't allow tasks with an affinity mask smaller than
3984 * the entire root_domain to become SCHED_DEADLINE. We
3985 * will also fail if there's no bandwidth available.
3987 if (!cpumask_subset(span, &p->cpus_allowed) ||
3988 rq->rd->dl_bw.bw == 0) {
3989 task_rq_unlock(rq, p, &flags);
3996 /* recheck policy now with rq lock held */
3997 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3998 policy = oldpolicy = -1;
3999 task_rq_unlock(rq, p, &flags);
4004 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4005 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4008 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4009 task_rq_unlock(rq, p, &flags);
4013 p->sched_reset_on_fork = reset_on_fork;
4018 * Take priority boosted tasks into account. If the new
4019 * effective priority is unchanged, we just store the new
4020 * normal parameters and do not touch the scheduler class and
4021 * the runqueue. This will be done when the task deboost
4024 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4025 if (new_effective_prio == oldprio)
4026 queue_flags &= ~DEQUEUE_MOVE;
4029 queued = task_on_rq_queued(p);
4030 running = task_current(rq, p);
4032 dequeue_task(rq, p, queue_flags);
4034 put_prev_task(rq, p);
4036 prev_class = p->sched_class;
4037 __setscheduler(rq, p, attr, pi);
4040 p->sched_class->set_curr_task(rq);
4043 * We enqueue to tail when the priority of a task is
4044 * increased (user space view).
4046 if (oldprio < p->prio)
4047 queue_flags |= ENQUEUE_HEAD;
4049 enqueue_task(rq, p, queue_flags);
4052 check_class_changed(rq, p, prev_class, oldprio);
4053 preempt_disable(); /* avoid rq from going away on us */
4054 task_rq_unlock(rq, p, &flags);
4057 rt_mutex_adjust_pi(p);
4060 * Run balance callbacks after we've adjusted the PI chain.
4062 balance_callback(rq);
4068 static int _sched_setscheduler(struct task_struct *p, int policy,
4069 const struct sched_param *param, bool check)
4071 struct sched_attr attr = {
4072 .sched_policy = policy,
4073 .sched_priority = param->sched_priority,
4074 .sched_nice = PRIO_TO_NICE(p->static_prio),
4077 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4078 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4079 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4080 policy &= ~SCHED_RESET_ON_FORK;
4081 attr.sched_policy = policy;
4084 return __sched_setscheduler(p, &attr, check, true);
4087 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4088 * @p: the task in question.
4089 * @policy: new policy.
4090 * @param: structure containing the new RT priority.
4092 * Return: 0 on success. An error code otherwise.
4094 * NOTE that the task may be already dead.
4096 int sched_setscheduler(struct task_struct *p, int policy,
4097 const struct sched_param *param)
4099 return _sched_setscheduler(p, policy, param, true);
4101 EXPORT_SYMBOL_GPL(sched_setscheduler);
4103 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4105 return __sched_setscheduler(p, attr, true, true);
4107 EXPORT_SYMBOL_GPL(sched_setattr);
4110 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4111 * @p: the task in question.
4112 * @policy: new policy.
4113 * @param: structure containing the new RT priority.
4115 * Just like sched_setscheduler, only don't bother checking if the
4116 * current context has permission. For example, this is needed in
4117 * stop_machine(): we create temporary high priority worker threads,
4118 * but our caller might not have that capability.
4120 * Return: 0 on success. An error code otherwise.
4122 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4123 const struct sched_param *param)
4125 return _sched_setscheduler(p, policy, param, false);
4127 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4130 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4132 struct sched_param lparam;
4133 struct task_struct *p;
4136 if (!param || pid < 0)
4138 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4143 p = find_process_by_pid(pid);
4145 retval = sched_setscheduler(p, policy, &lparam);
4152 * Mimics kernel/events/core.c perf_copy_attr().
4154 static int sched_copy_attr(struct sched_attr __user *uattr,
4155 struct sched_attr *attr)
4160 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4164 * zero the full structure, so that a short copy will be nice.
4166 memset(attr, 0, sizeof(*attr));
4168 ret = get_user(size, &uattr->size);
4172 if (size > PAGE_SIZE) /* silly large */
4175 if (!size) /* abi compat */
4176 size = SCHED_ATTR_SIZE_VER0;
4178 if (size < SCHED_ATTR_SIZE_VER0)
4182 * If we're handed a bigger struct than we know of,
4183 * ensure all the unknown bits are 0 - i.e. new
4184 * user-space does not rely on any kernel feature
4185 * extensions we dont know about yet.
4187 if (size > sizeof(*attr)) {
4188 unsigned char __user *addr;
4189 unsigned char __user *end;
4192 addr = (void __user *)uattr + sizeof(*attr);
4193 end = (void __user *)uattr + size;
4195 for (; addr < end; addr++) {
4196 ret = get_user(val, addr);
4202 size = sizeof(*attr);
4205 ret = copy_from_user(attr, uattr, size);
4210 * XXX: do we want to be lenient like existing syscalls; or do we want
4211 * to be strict and return an error on out-of-bounds values?
4213 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4218 put_user(sizeof(*attr), &uattr->size);
4223 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4224 * @pid: the pid in question.
4225 * @policy: new policy.
4226 * @param: structure containing the new RT priority.
4228 * Return: 0 on success. An error code otherwise.
4230 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4231 struct sched_param __user *, param)
4233 /* negative values for policy are not valid */
4237 return do_sched_setscheduler(pid, policy, param);
4241 * sys_sched_setparam - set/change the RT priority of a thread
4242 * @pid: the pid in question.
4243 * @param: structure containing the new RT priority.
4245 * Return: 0 on success. An error code otherwise.
4247 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4249 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4253 * sys_sched_setattr - same as above, but with extended sched_attr
4254 * @pid: the pid in question.
4255 * @uattr: structure containing the extended parameters.
4256 * @flags: for future extension.
4258 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4259 unsigned int, flags)
4261 struct sched_attr attr;
4262 struct task_struct *p;
4265 if (!uattr || pid < 0 || flags)
4268 retval = sched_copy_attr(uattr, &attr);
4272 if ((int)attr.sched_policy < 0)
4277 p = find_process_by_pid(pid);
4279 retval = sched_setattr(p, &attr);
4286 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4287 * @pid: the pid in question.
4289 * Return: On success, the policy of the thread. Otherwise, a negative error
4292 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4294 struct task_struct *p;
4302 p = find_process_by_pid(pid);
4304 retval = security_task_getscheduler(p);
4307 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4314 * sys_sched_getparam - get the RT priority of a thread
4315 * @pid: the pid in question.
4316 * @param: structure containing the RT priority.
4318 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4321 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4323 struct sched_param lp = { .sched_priority = 0 };
4324 struct task_struct *p;
4327 if (!param || pid < 0)
4331 p = find_process_by_pid(pid);
4336 retval = security_task_getscheduler(p);
4340 if (task_has_rt_policy(p))
4341 lp.sched_priority = p->rt_priority;
4345 * This one might sleep, we cannot do it with a spinlock held ...
4347 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4356 static int sched_read_attr(struct sched_attr __user *uattr,
4357 struct sched_attr *attr,
4362 if (!access_ok(VERIFY_WRITE, uattr, usize))
4366 * If we're handed a smaller struct than we know of,
4367 * ensure all the unknown bits are 0 - i.e. old
4368 * user-space does not get uncomplete information.
4370 if (usize < sizeof(*attr)) {
4371 unsigned char *addr;
4374 addr = (void *)attr + usize;
4375 end = (void *)attr + sizeof(*attr);
4377 for (; addr < end; addr++) {
4385 ret = copy_to_user(uattr, attr, attr->size);
4393 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4394 * @pid: the pid in question.
4395 * @uattr: structure containing the extended parameters.
4396 * @size: sizeof(attr) for fwd/bwd comp.
4397 * @flags: for future extension.
4399 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4400 unsigned int, size, unsigned int, flags)
4402 struct sched_attr attr = {
4403 .size = sizeof(struct sched_attr),
4405 struct task_struct *p;
4408 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4409 size < SCHED_ATTR_SIZE_VER0 || flags)
4413 p = find_process_by_pid(pid);
4418 retval = security_task_getscheduler(p);
4422 attr.sched_policy = p->policy;
4423 if (p->sched_reset_on_fork)
4424 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4425 if (task_has_dl_policy(p))
4426 __getparam_dl(p, &attr);
4427 else if (task_has_rt_policy(p))
4428 attr.sched_priority = p->rt_priority;
4430 attr.sched_nice = task_nice(p);
4434 retval = sched_read_attr(uattr, &attr, size);
4442 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4444 cpumask_var_t cpus_allowed, new_mask;
4445 struct task_struct *p;
4450 p = find_process_by_pid(pid);
4456 /* Prevent p going away */
4460 if (p->flags & PF_NO_SETAFFINITY) {
4464 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4468 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4470 goto out_free_cpus_allowed;
4473 if (!check_same_owner(p)) {
4475 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4477 goto out_free_new_mask;
4482 retval = security_task_setscheduler(p);
4484 goto out_free_new_mask;
4487 cpuset_cpus_allowed(p, cpus_allowed);
4488 cpumask_and(new_mask, in_mask, cpus_allowed);
4491 * Since bandwidth control happens on root_domain basis,
4492 * if admission test is enabled, we only admit -deadline
4493 * tasks allowed to run on all the CPUs in the task's
4497 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4499 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4502 goto out_free_new_mask;
4508 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4511 cpuset_cpus_allowed(p, cpus_allowed);
4512 if (!cpumask_subset(new_mask, cpus_allowed)) {
4514 * We must have raced with a concurrent cpuset
4515 * update. Just reset the cpus_allowed to the
4516 * cpuset's cpus_allowed
4518 cpumask_copy(new_mask, cpus_allowed);
4523 free_cpumask_var(new_mask);
4524 out_free_cpus_allowed:
4525 free_cpumask_var(cpus_allowed);
4531 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4532 struct cpumask *new_mask)
4534 if (len < cpumask_size())
4535 cpumask_clear(new_mask);
4536 else if (len > cpumask_size())
4537 len = cpumask_size();
4539 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4543 * sys_sched_setaffinity - set the cpu affinity of a process
4544 * @pid: pid of the process
4545 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4546 * @user_mask_ptr: user-space pointer to the new cpu mask
4548 * Return: 0 on success. An error code otherwise.
4550 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4551 unsigned long __user *, user_mask_ptr)
4553 cpumask_var_t new_mask;
4556 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4559 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4561 retval = sched_setaffinity(pid, new_mask);
4562 free_cpumask_var(new_mask);
4566 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4568 struct task_struct *p;
4569 unsigned long flags;
4575 p = find_process_by_pid(pid);
4579 retval = security_task_getscheduler(p);
4583 raw_spin_lock_irqsave(&p->pi_lock, flags);
4584 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4585 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4594 * sys_sched_getaffinity - get the cpu affinity of a process
4595 * @pid: pid of the process
4596 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4597 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4599 * Return: 0 on success. An error code otherwise.
4601 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4602 unsigned long __user *, user_mask_ptr)
4607 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4609 if (len & (sizeof(unsigned long)-1))
4612 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4615 ret = sched_getaffinity(pid, mask);
4617 size_t retlen = min_t(size_t, len, cpumask_size());
4619 if (copy_to_user(user_mask_ptr, mask, retlen))
4624 free_cpumask_var(mask);
4630 * sys_sched_yield - yield the current processor to other threads.
4632 * This function yields the current CPU to other tasks. If there are no
4633 * other threads running on this CPU then this function will return.
4637 SYSCALL_DEFINE0(sched_yield)
4639 struct rq *rq = this_rq_lock();
4641 schedstat_inc(rq, yld_count);
4642 current->sched_class->yield_task(rq);
4645 * Since we are going to call schedule() anyway, there's
4646 * no need to preempt or enable interrupts:
4648 __release(rq->lock);
4649 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4650 do_raw_spin_unlock(&rq->lock);
4651 sched_preempt_enable_no_resched();
4658 int __sched _cond_resched(void)
4660 if (should_resched(0)) {
4661 preempt_schedule_common();
4666 EXPORT_SYMBOL(_cond_resched);
4669 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4670 * call schedule, and on return reacquire the lock.
4672 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4673 * operations here to prevent schedule() from being called twice (once via
4674 * spin_unlock(), once by hand).
4676 int __cond_resched_lock(spinlock_t *lock)
4678 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4681 lockdep_assert_held(lock);
4683 if (spin_needbreak(lock) || resched) {
4686 preempt_schedule_common();
4694 EXPORT_SYMBOL(__cond_resched_lock);
4696 int __sched __cond_resched_softirq(void)
4698 BUG_ON(!in_softirq());
4700 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4702 preempt_schedule_common();
4708 EXPORT_SYMBOL(__cond_resched_softirq);
4711 * yield - yield the current processor to other threads.
4713 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4715 * The scheduler is at all times free to pick the calling task as the most
4716 * eligible task to run, if removing the yield() call from your code breaks
4717 * it, its already broken.
4719 * Typical broken usage is:
4724 * where one assumes that yield() will let 'the other' process run that will
4725 * make event true. If the current task is a SCHED_FIFO task that will never
4726 * happen. Never use yield() as a progress guarantee!!
4728 * If you want to use yield() to wait for something, use wait_event().
4729 * If you want to use yield() to be 'nice' for others, use cond_resched().
4730 * If you still want to use yield(), do not!
4732 void __sched yield(void)
4734 set_current_state(TASK_RUNNING);
4737 EXPORT_SYMBOL(yield);
4740 * yield_to - yield the current processor to another thread in
4741 * your thread group, or accelerate that thread toward the
4742 * processor it's on.
4744 * @preempt: whether task preemption is allowed or not
4746 * It's the caller's job to ensure that the target task struct
4747 * can't go away on us before we can do any checks.
4750 * true (>0) if we indeed boosted the target task.
4751 * false (0) if we failed to boost the target.
4752 * -ESRCH if there's no task to yield to.
4754 int __sched yield_to(struct task_struct *p, bool preempt)
4756 struct task_struct *curr = current;
4757 struct rq *rq, *p_rq;
4758 unsigned long flags;
4761 local_irq_save(flags);
4767 * If we're the only runnable task on the rq and target rq also
4768 * has only one task, there's absolutely no point in yielding.
4770 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4775 double_rq_lock(rq, p_rq);
4776 if (task_rq(p) != p_rq) {
4777 double_rq_unlock(rq, p_rq);
4781 if (!curr->sched_class->yield_to_task)
4784 if (curr->sched_class != p->sched_class)
4787 if (task_running(p_rq, p) || p->state)
4790 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4792 schedstat_inc(rq, yld_count);
4794 * Make p's CPU reschedule; pick_next_entity takes care of
4797 if (preempt && rq != p_rq)
4802 double_rq_unlock(rq, p_rq);
4804 local_irq_restore(flags);
4811 EXPORT_SYMBOL_GPL(yield_to);
4814 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4815 * that process accounting knows that this is a task in IO wait state.
4817 long __sched io_schedule_timeout(long timeout)
4819 int old_iowait = current->in_iowait;
4823 current->in_iowait = 1;
4824 blk_schedule_flush_plug(current);
4826 delayacct_blkio_start();
4828 atomic_inc(&rq->nr_iowait);
4829 ret = schedule_timeout(timeout);
4830 current->in_iowait = old_iowait;
4831 atomic_dec(&rq->nr_iowait);
4832 delayacct_blkio_end();
4836 EXPORT_SYMBOL(io_schedule_timeout);
4839 * sys_sched_get_priority_max - return maximum RT priority.
4840 * @policy: scheduling class.
4842 * Return: On success, this syscall returns the maximum
4843 * rt_priority that can be used by a given scheduling class.
4844 * On failure, a negative error code is returned.
4846 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4853 ret = MAX_USER_RT_PRIO-1;
4855 case SCHED_DEADLINE:
4866 * sys_sched_get_priority_min - return minimum RT priority.
4867 * @policy: scheduling class.
4869 * Return: On success, this syscall returns the minimum
4870 * rt_priority that can be used by a given scheduling class.
4871 * On failure, a negative error code is returned.
4873 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4882 case SCHED_DEADLINE:
4892 * sys_sched_rr_get_interval - return the default timeslice of a process.
4893 * @pid: pid of the process.
4894 * @interval: userspace pointer to the timeslice value.
4896 * this syscall writes the default timeslice value of a given process
4897 * into the user-space timespec buffer. A value of '0' means infinity.
4899 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4902 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4903 struct timespec __user *, interval)
4905 struct task_struct *p;
4906 unsigned int time_slice;
4907 unsigned long flags;
4917 p = find_process_by_pid(pid);
4921 retval = security_task_getscheduler(p);
4925 rq = task_rq_lock(p, &flags);
4927 if (p->sched_class->get_rr_interval)
4928 time_slice = p->sched_class->get_rr_interval(rq, p);
4929 task_rq_unlock(rq, p, &flags);
4932 jiffies_to_timespec(time_slice, &t);
4933 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4941 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4943 void sched_show_task(struct task_struct *p)
4945 unsigned long free = 0;
4947 unsigned long state = p->state;
4950 state = __ffs(state) + 1;
4951 printk(KERN_INFO "%-15.15s %c", p->comm,
4952 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4953 #if BITS_PER_LONG == 32
4954 if (state == TASK_RUNNING)
4955 printk(KERN_CONT " running ");
4957 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4959 if (state == TASK_RUNNING)
4960 printk(KERN_CONT " running task ");
4962 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4964 #ifdef CONFIG_DEBUG_STACK_USAGE
4965 free = stack_not_used(p);
4970 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4972 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4973 task_pid_nr(p), ppid,
4974 (unsigned long)task_thread_info(p)->flags);
4976 print_worker_info(KERN_INFO, p);
4977 show_stack(p, NULL);
4980 void show_state_filter(unsigned long state_filter)
4982 struct task_struct *g, *p;
4984 #if BITS_PER_LONG == 32
4986 " task PC stack pid father\n");
4989 " task PC stack pid father\n");
4992 for_each_process_thread(g, p) {
4994 * reset the NMI-timeout, listing all files on a slow
4995 * console might take a lot of time:
4997 touch_nmi_watchdog();
4998 if (!state_filter || (p->state & state_filter))
5002 touch_all_softlockup_watchdogs();
5004 #ifdef CONFIG_SCHED_DEBUG
5005 sysrq_sched_debug_show();
5009 * Only show locks if all tasks are dumped:
5012 debug_show_all_locks();
5015 void init_idle_bootup_task(struct task_struct *idle)
5017 idle->sched_class = &idle_sched_class;
5021 * init_idle - set up an idle thread for a given CPU
5022 * @idle: task in question
5023 * @cpu: cpu the idle task belongs to
5025 * NOTE: this function does not set the idle thread's NEED_RESCHED
5026 * flag, to make booting more robust.
5028 void init_idle(struct task_struct *idle, int cpu)
5030 struct rq *rq = cpu_rq(cpu);
5031 unsigned long flags;
5033 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5034 raw_spin_lock(&rq->lock);
5036 __sched_fork(0, idle);
5037 idle->state = TASK_RUNNING;
5038 idle->se.exec_start = sched_clock();
5040 kasan_unpoison_task_stack(idle);
5044 * Its possible that init_idle() gets called multiple times on a task,
5045 * in that case do_set_cpus_allowed() will not do the right thing.
5047 * And since this is boot we can forgo the serialization.
5049 set_cpus_allowed_common(idle, cpumask_of(cpu));
5052 * We're having a chicken and egg problem, even though we are
5053 * holding rq->lock, the cpu isn't yet set to this cpu so the
5054 * lockdep check in task_group() will fail.
5056 * Similar case to sched_fork(). / Alternatively we could
5057 * use task_rq_lock() here and obtain the other rq->lock.
5062 __set_task_cpu(idle, cpu);
5065 rq->curr = rq->idle = idle;
5066 idle->on_rq = TASK_ON_RQ_QUEUED;
5070 raw_spin_unlock(&rq->lock);
5071 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5073 /* Set the preempt count _outside_ the spinlocks! */
5074 init_idle_preempt_count(idle, cpu);
5077 * The idle tasks have their own, simple scheduling class:
5079 idle->sched_class = &idle_sched_class;
5080 ftrace_graph_init_idle_task(idle, cpu);
5081 vtime_init_idle(idle, cpu);
5083 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5087 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5088 const struct cpumask *trial)
5090 int ret = 1, trial_cpus;
5091 struct dl_bw *cur_dl_b;
5092 unsigned long flags;
5094 if (!cpumask_weight(cur))
5097 rcu_read_lock_sched();
5098 cur_dl_b = dl_bw_of(cpumask_any(cur));
5099 trial_cpus = cpumask_weight(trial);
5101 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5102 if (cur_dl_b->bw != -1 &&
5103 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5105 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5106 rcu_read_unlock_sched();
5111 int task_can_attach(struct task_struct *p,
5112 const struct cpumask *cs_cpus_allowed)
5117 * Kthreads which disallow setaffinity shouldn't be moved
5118 * to a new cpuset; we don't want to change their cpu
5119 * affinity and isolating such threads by their set of
5120 * allowed nodes is unnecessary. Thus, cpusets are not
5121 * applicable for such threads. This prevents checking for
5122 * success of set_cpus_allowed_ptr() on all attached tasks
5123 * before cpus_allowed may be changed.
5125 if (p->flags & PF_NO_SETAFFINITY) {
5131 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5133 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5138 unsigned long flags;
5140 rcu_read_lock_sched();
5141 dl_b = dl_bw_of(dest_cpu);
5142 raw_spin_lock_irqsave(&dl_b->lock, flags);
5143 cpus = dl_bw_cpus(dest_cpu);
5144 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5149 * We reserve space for this task in the destination
5150 * root_domain, as we can't fail after this point.
5151 * We will free resources in the source root_domain
5152 * later on (see set_cpus_allowed_dl()).
5154 __dl_add(dl_b, p->dl.dl_bw);
5156 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5157 rcu_read_unlock_sched();
5167 #ifdef CONFIG_NUMA_BALANCING
5168 /* Migrate current task p to target_cpu */
5169 int migrate_task_to(struct task_struct *p, int target_cpu)
5171 struct migration_arg arg = { p, target_cpu };
5172 int curr_cpu = task_cpu(p);
5174 if (curr_cpu == target_cpu)
5177 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5180 /* TODO: This is not properly updating schedstats */
5182 trace_sched_move_numa(p, curr_cpu, target_cpu);
5183 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5187 * Requeue a task on a given node and accurately track the number of NUMA
5188 * tasks on the runqueues
5190 void sched_setnuma(struct task_struct *p, int nid)
5193 unsigned long flags;
5194 bool queued, running;
5196 rq = task_rq_lock(p, &flags);
5197 queued = task_on_rq_queued(p);
5198 running = task_current(rq, p);
5201 dequeue_task(rq, p, DEQUEUE_SAVE);
5203 put_prev_task(rq, p);
5205 p->numa_preferred_nid = nid;
5208 p->sched_class->set_curr_task(rq);
5210 enqueue_task(rq, p, ENQUEUE_RESTORE);
5211 task_rq_unlock(rq, p, &flags);
5213 #endif /* CONFIG_NUMA_BALANCING */
5215 #ifdef CONFIG_HOTPLUG_CPU
5217 * Ensures that the idle task is using init_mm right before its cpu goes
5220 void idle_task_exit(void)
5222 struct mm_struct *mm = current->active_mm;
5224 BUG_ON(cpu_online(smp_processor_id()));
5226 if (mm != &init_mm) {
5227 switch_mm(mm, &init_mm, current);
5228 finish_arch_post_lock_switch();
5234 * Since this CPU is going 'away' for a while, fold any nr_active delta
5235 * we might have. Assumes we're called after migrate_tasks() so that the
5236 * nr_active count is stable.
5238 * Also see the comment "Global load-average calculations".
5240 static void calc_load_migrate(struct rq *rq)
5242 long delta = calc_load_fold_active(rq);
5244 atomic_long_add(delta, &calc_load_tasks);
5247 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5251 static const struct sched_class fake_sched_class = {
5252 .put_prev_task = put_prev_task_fake,
5255 static struct task_struct fake_task = {
5257 * Avoid pull_{rt,dl}_task()
5259 .prio = MAX_PRIO + 1,
5260 .sched_class = &fake_sched_class,
5264 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5265 * try_to_wake_up()->select_task_rq().
5267 * Called with rq->lock held even though we'er in stop_machine() and
5268 * there's no concurrency possible, we hold the required locks anyway
5269 * because of lock validation efforts.
5271 static void migrate_tasks(struct rq *dead_rq)
5273 struct rq *rq = dead_rq;
5274 struct task_struct *next, *stop = rq->stop;
5278 * Fudge the rq selection such that the below task selection loop
5279 * doesn't get stuck on the currently eligible stop task.
5281 * We're currently inside stop_machine() and the rq is either stuck
5282 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5283 * either way we should never end up calling schedule() until we're
5289 * put_prev_task() and pick_next_task() sched
5290 * class method both need to have an up-to-date
5291 * value of rq->clock[_task]
5293 update_rq_clock(rq);
5297 * There's this thread running, bail when that's the only
5300 if (rq->nr_running == 1)
5304 * pick_next_task assumes pinned rq->lock.
5306 lockdep_pin_lock(&rq->lock);
5307 next = pick_next_task(rq, &fake_task);
5309 next->sched_class->put_prev_task(rq, next);
5312 * Rules for changing task_struct::cpus_allowed are holding
5313 * both pi_lock and rq->lock, such that holding either
5314 * stabilizes the mask.
5316 * Drop rq->lock is not quite as disastrous as it usually is
5317 * because !cpu_active at this point, which means load-balance
5318 * will not interfere. Also, stop-machine.
5320 lockdep_unpin_lock(&rq->lock);
5321 raw_spin_unlock(&rq->lock);
5322 raw_spin_lock(&next->pi_lock);
5323 raw_spin_lock(&rq->lock);
5326 * Since we're inside stop-machine, _nothing_ should have
5327 * changed the task, WARN if weird stuff happened, because in
5328 * that case the above rq->lock drop is a fail too.
5330 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5331 raw_spin_unlock(&next->pi_lock);
5335 /* Find suitable destination for @next, with force if needed. */
5336 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5338 rq = __migrate_task(rq, next, dest_cpu);
5339 if (rq != dead_rq) {
5340 raw_spin_unlock(&rq->lock);
5342 raw_spin_lock(&rq->lock);
5344 raw_spin_unlock(&next->pi_lock);
5349 #endif /* CONFIG_HOTPLUG_CPU */
5351 static void set_rq_online(struct rq *rq)
5354 const struct sched_class *class;
5356 cpumask_set_cpu(rq->cpu, rq->rd->online);
5359 for_each_class(class) {
5360 if (class->rq_online)
5361 class->rq_online(rq);
5366 static void set_rq_offline(struct rq *rq)
5369 const struct sched_class *class;
5371 for_each_class(class) {
5372 if (class->rq_offline)
5373 class->rq_offline(rq);
5376 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5382 * migration_call - callback that gets triggered when a CPU is added.
5383 * Here we can start up the necessary migration thread for the new CPU.
5386 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5388 int cpu = (long)hcpu;
5389 unsigned long flags;
5390 struct rq *rq = cpu_rq(cpu);
5392 switch (action & ~CPU_TASKS_FROZEN) {
5394 case CPU_UP_PREPARE:
5395 rq->calc_load_update = calc_load_update;
5396 account_reset_rq(rq);
5400 /* Update our root-domain */
5401 raw_spin_lock_irqsave(&rq->lock, flags);
5403 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5407 raw_spin_unlock_irqrestore(&rq->lock, flags);
5410 #ifdef CONFIG_HOTPLUG_CPU
5412 sched_ttwu_pending();
5413 /* Update our root-domain */
5414 raw_spin_lock_irqsave(&rq->lock, flags);
5416 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5420 BUG_ON(rq->nr_running != 1); /* the migration thread */
5421 raw_spin_unlock_irqrestore(&rq->lock, flags);
5425 calc_load_migrate(rq);
5430 update_max_interval();
5436 * Register at high priority so that task migration (migrate_all_tasks)
5437 * happens before everything else. This has to be lower priority than
5438 * the notifier in the perf_event subsystem, though.
5440 static struct notifier_block migration_notifier = {
5441 .notifier_call = migration_call,
5442 .priority = CPU_PRI_MIGRATION,
5445 static void set_cpu_rq_start_time(void)
5447 int cpu = smp_processor_id();
5448 struct rq *rq = cpu_rq(cpu);
5449 rq->age_stamp = sched_clock_cpu(cpu);
5452 static int sched_cpu_active(struct notifier_block *nfb,
5453 unsigned long action, void *hcpu)
5455 int cpu = (long)hcpu;
5457 switch (action & ~CPU_TASKS_FROZEN) {
5459 set_cpu_rq_start_time();
5462 case CPU_DOWN_FAILED:
5463 set_cpu_active(cpu, true);
5471 static int sched_cpu_inactive(struct notifier_block *nfb,
5472 unsigned long action, void *hcpu)
5474 switch (action & ~CPU_TASKS_FROZEN) {
5475 case CPU_DOWN_PREPARE:
5476 set_cpu_active((long)hcpu, false);
5483 static int __init migration_init(void)
5485 void *cpu = (void *)(long)smp_processor_id();
5488 /* Initialize migration for the boot CPU */
5489 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5490 BUG_ON(err == NOTIFY_BAD);
5491 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5492 register_cpu_notifier(&migration_notifier);
5494 /* Register cpu active notifiers */
5495 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5496 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5500 early_initcall(migration_init);
5502 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5504 #ifdef CONFIG_SCHED_DEBUG
5506 static __read_mostly int sched_debug_enabled;
5508 static int __init sched_debug_setup(char *str)
5510 sched_debug_enabled = 1;
5514 early_param("sched_debug", sched_debug_setup);
5516 static inline bool sched_debug(void)
5518 return sched_debug_enabled;
5521 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5522 struct cpumask *groupmask)
5524 struct sched_group *group = sd->groups;
5526 cpumask_clear(groupmask);
5528 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5530 if (!(sd->flags & SD_LOAD_BALANCE)) {
5531 printk("does not load-balance\n");
5533 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5538 printk(KERN_CONT "span %*pbl level %s\n",
5539 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5541 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5542 printk(KERN_ERR "ERROR: domain->span does not contain "
5545 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5546 printk(KERN_ERR "ERROR: domain->groups does not contain"
5550 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5554 printk(KERN_ERR "ERROR: group is NULL\n");
5558 if (!cpumask_weight(sched_group_cpus(group))) {
5559 printk(KERN_CONT "\n");
5560 printk(KERN_ERR "ERROR: empty group\n");
5564 if (!(sd->flags & SD_OVERLAP) &&
5565 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5566 printk(KERN_CONT "\n");
5567 printk(KERN_ERR "ERROR: repeated CPUs\n");
5571 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5573 printk(KERN_CONT " %*pbl",
5574 cpumask_pr_args(sched_group_cpus(group)));
5575 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5576 printk(KERN_CONT " (cpu_capacity = %d)",
5577 group->sgc->capacity);
5580 group = group->next;
5581 } while (group != sd->groups);
5582 printk(KERN_CONT "\n");
5584 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5585 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5588 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5589 printk(KERN_ERR "ERROR: parent span is not a superset "
5590 "of domain->span\n");
5594 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5598 if (!sched_debug_enabled)
5602 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5606 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5609 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5617 #else /* !CONFIG_SCHED_DEBUG */
5618 # define sched_domain_debug(sd, cpu) do { } while (0)
5619 static inline bool sched_debug(void)
5623 #endif /* CONFIG_SCHED_DEBUG */
5625 static int sd_degenerate(struct sched_domain *sd)
5627 if (cpumask_weight(sched_domain_span(sd)) == 1)
5630 /* Following flags need at least 2 groups */
5631 if (sd->flags & (SD_LOAD_BALANCE |
5632 SD_BALANCE_NEWIDLE |
5635 SD_SHARE_CPUCAPACITY |
5636 SD_SHARE_PKG_RESOURCES |
5637 SD_SHARE_POWERDOMAIN)) {
5638 if (sd->groups != sd->groups->next)
5642 /* Following flags don't use groups */
5643 if (sd->flags & (SD_WAKE_AFFINE))
5650 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5652 unsigned long cflags = sd->flags, pflags = parent->flags;
5654 if (sd_degenerate(parent))
5657 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5660 /* Flags needing groups don't count if only 1 group in parent */
5661 if (parent->groups == parent->groups->next) {
5662 pflags &= ~(SD_LOAD_BALANCE |
5663 SD_BALANCE_NEWIDLE |
5666 SD_SHARE_CPUCAPACITY |
5667 SD_SHARE_PKG_RESOURCES |
5669 SD_SHARE_POWERDOMAIN);
5670 if (nr_node_ids == 1)
5671 pflags &= ~SD_SERIALIZE;
5673 if (~cflags & pflags)
5679 static void free_rootdomain(struct rcu_head *rcu)
5681 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5683 cpupri_cleanup(&rd->cpupri);
5684 cpudl_cleanup(&rd->cpudl);
5685 free_cpumask_var(rd->dlo_mask);
5686 free_cpumask_var(rd->rto_mask);
5687 free_cpumask_var(rd->online);
5688 free_cpumask_var(rd->span);
5692 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5694 struct root_domain *old_rd = NULL;
5695 unsigned long flags;
5697 raw_spin_lock_irqsave(&rq->lock, flags);
5702 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5705 cpumask_clear_cpu(rq->cpu, old_rd->span);
5708 * If we dont want to free the old_rd yet then
5709 * set old_rd to NULL to skip the freeing later
5712 if (!atomic_dec_and_test(&old_rd->refcount))
5716 atomic_inc(&rd->refcount);
5719 cpumask_set_cpu(rq->cpu, rd->span);
5720 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5723 raw_spin_unlock_irqrestore(&rq->lock, flags);
5726 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5729 static int init_rootdomain(struct root_domain *rd)
5731 memset(rd, 0, sizeof(*rd));
5733 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5735 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5737 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5739 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5742 init_dl_bw(&rd->dl_bw);
5743 if (cpudl_init(&rd->cpudl) != 0)
5746 if (cpupri_init(&rd->cpupri) != 0)
5751 free_cpumask_var(rd->rto_mask);
5753 free_cpumask_var(rd->dlo_mask);
5755 free_cpumask_var(rd->online);
5757 free_cpumask_var(rd->span);
5763 * By default the system creates a single root-domain with all cpus as
5764 * members (mimicking the global state we have today).
5766 struct root_domain def_root_domain;
5768 static void init_defrootdomain(void)
5770 init_rootdomain(&def_root_domain);
5772 atomic_set(&def_root_domain.refcount, 1);
5775 static struct root_domain *alloc_rootdomain(void)
5777 struct root_domain *rd;
5779 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5783 if (init_rootdomain(rd) != 0) {
5791 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5793 struct sched_group *tmp, *first;
5802 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5807 } while (sg != first);
5810 static void free_sched_domain(struct rcu_head *rcu)
5812 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5815 * If its an overlapping domain it has private groups, iterate and
5818 if (sd->flags & SD_OVERLAP) {
5819 free_sched_groups(sd->groups, 1);
5820 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5821 kfree(sd->groups->sgc);
5827 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5829 call_rcu(&sd->rcu, free_sched_domain);
5832 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5834 for (; sd; sd = sd->parent)
5835 destroy_sched_domain(sd, cpu);
5839 * Keep a special pointer to the highest sched_domain that has
5840 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5841 * allows us to avoid some pointer chasing select_idle_sibling().
5843 * Also keep a unique ID per domain (we use the first cpu number in
5844 * the cpumask of the domain), this allows us to quickly tell if
5845 * two cpus are in the same cache domain, see cpus_share_cache().
5847 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5848 DEFINE_PER_CPU(int, sd_llc_size);
5849 DEFINE_PER_CPU(int, sd_llc_id);
5850 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5851 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5852 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5854 static void update_top_cache_domain(int cpu)
5856 struct sched_domain *sd;
5857 struct sched_domain *busy_sd = NULL;
5861 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5863 id = cpumask_first(sched_domain_span(sd));
5864 size = cpumask_weight(sched_domain_span(sd));
5865 busy_sd = sd->parent; /* sd_busy */
5867 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5869 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5870 per_cpu(sd_llc_size, cpu) = size;
5871 per_cpu(sd_llc_id, cpu) = id;
5873 sd = lowest_flag_domain(cpu, SD_NUMA);
5874 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5876 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5877 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5881 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5882 * hold the hotplug lock.
5885 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5887 struct rq *rq = cpu_rq(cpu);
5888 struct sched_domain *tmp;
5890 /* Remove the sched domains which do not contribute to scheduling. */
5891 for (tmp = sd; tmp; ) {
5892 struct sched_domain *parent = tmp->parent;
5896 if (sd_parent_degenerate(tmp, parent)) {
5897 tmp->parent = parent->parent;
5899 parent->parent->child = tmp;
5901 * Transfer SD_PREFER_SIBLING down in case of a
5902 * degenerate parent; the spans match for this
5903 * so the property transfers.
5905 if (parent->flags & SD_PREFER_SIBLING)
5906 tmp->flags |= SD_PREFER_SIBLING;
5907 destroy_sched_domain(parent, cpu);
5912 if (sd && sd_degenerate(sd)) {
5915 destroy_sched_domain(tmp, cpu);
5920 sched_domain_debug(sd, cpu);
5922 rq_attach_root(rq, rd);
5924 rcu_assign_pointer(rq->sd, sd);
5925 destroy_sched_domains(tmp, cpu);
5927 update_top_cache_domain(cpu);
5930 /* Setup the mask of cpus configured for isolated domains */
5931 static int __init isolated_cpu_setup(char *str)
5935 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5936 ret = cpulist_parse(str, cpu_isolated_map);
5938 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5943 __setup("isolcpus=", isolated_cpu_setup);
5946 struct sched_domain ** __percpu sd;
5947 struct root_domain *rd;
5958 * Build an iteration mask that can exclude certain CPUs from the upwards
5961 * Asymmetric node setups can result in situations where the domain tree is of
5962 * unequal depth, make sure to skip domains that already cover the entire
5965 * In that case build_sched_domains() will have terminated the iteration early
5966 * and our sibling sd spans will be empty. Domains should always include the
5967 * cpu they're built on, so check that.
5970 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5972 const struct cpumask *span = sched_domain_span(sd);
5973 struct sd_data *sdd = sd->private;
5974 struct sched_domain *sibling;
5977 for_each_cpu(i, span) {
5978 sibling = *per_cpu_ptr(sdd->sd, i);
5979 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5982 cpumask_set_cpu(i, sched_group_mask(sg));
5987 * Return the canonical balance cpu for this group, this is the first cpu
5988 * of this group that's also in the iteration mask.
5990 int group_balance_cpu(struct sched_group *sg)
5992 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5996 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5998 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5999 const struct cpumask *span = sched_domain_span(sd);
6000 struct cpumask *covered = sched_domains_tmpmask;
6001 struct sd_data *sdd = sd->private;
6002 struct sched_domain *sibling;
6005 cpumask_clear(covered);
6007 for_each_cpu(i, span) {
6008 struct cpumask *sg_span;
6010 if (cpumask_test_cpu(i, covered))
6013 sibling = *per_cpu_ptr(sdd->sd, i);
6015 /* See the comment near build_group_mask(). */
6016 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6019 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6020 GFP_KERNEL, cpu_to_node(cpu));
6025 sg_span = sched_group_cpus(sg);
6027 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6029 cpumask_set_cpu(i, sg_span);
6031 cpumask_or(covered, covered, sg_span);
6033 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6034 if (atomic_inc_return(&sg->sgc->ref) == 1)
6035 build_group_mask(sd, sg);
6038 * Initialize sgc->capacity such that even if we mess up the
6039 * domains and no possible iteration will get us here, we won't
6042 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6045 * Make sure the first group of this domain contains the
6046 * canonical balance cpu. Otherwise the sched_domain iteration
6047 * breaks. See update_sg_lb_stats().
6049 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6050 group_balance_cpu(sg) == cpu)
6060 sd->groups = groups;
6065 free_sched_groups(first, 0);
6070 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6072 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6073 struct sched_domain *child = sd->child;
6076 cpu = cpumask_first(sched_domain_span(child));
6079 *sg = *per_cpu_ptr(sdd->sg, cpu);
6080 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6081 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6088 * build_sched_groups will build a circular linked list of the groups
6089 * covered by the given span, and will set each group's ->cpumask correctly,
6090 * and ->cpu_capacity to 0.
6092 * Assumes the sched_domain tree is fully constructed
6095 build_sched_groups(struct sched_domain *sd, int cpu)
6097 struct sched_group *first = NULL, *last = NULL;
6098 struct sd_data *sdd = sd->private;
6099 const struct cpumask *span = sched_domain_span(sd);
6100 struct cpumask *covered;
6103 get_group(cpu, sdd, &sd->groups);
6104 atomic_inc(&sd->groups->ref);
6106 if (cpu != cpumask_first(span))
6109 lockdep_assert_held(&sched_domains_mutex);
6110 covered = sched_domains_tmpmask;
6112 cpumask_clear(covered);
6114 for_each_cpu(i, span) {
6115 struct sched_group *sg;
6118 if (cpumask_test_cpu(i, covered))
6121 group = get_group(i, sdd, &sg);
6122 cpumask_setall(sched_group_mask(sg));
6124 for_each_cpu(j, span) {
6125 if (get_group(j, sdd, NULL) != group)
6128 cpumask_set_cpu(j, covered);
6129 cpumask_set_cpu(j, sched_group_cpus(sg));
6144 * Initialize sched groups cpu_capacity.
6146 * cpu_capacity indicates the capacity of sched group, which is used while
6147 * distributing the load between different sched groups in a sched domain.
6148 * Typically cpu_capacity for all the groups in a sched domain will be same
6149 * unless there are asymmetries in the topology. If there are asymmetries,
6150 * group having more cpu_capacity will pickup more load compared to the
6151 * group having less cpu_capacity.
6153 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6155 struct sched_group *sg = sd->groups;
6160 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6162 } while (sg != sd->groups);
6164 if (cpu != group_balance_cpu(sg))
6167 update_group_capacity(sd, cpu);
6168 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6172 * Initializers for schedule domains
6173 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6176 static int default_relax_domain_level = -1;
6177 int sched_domain_level_max;
6179 static int __init setup_relax_domain_level(char *str)
6181 if (kstrtoint(str, 0, &default_relax_domain_level))
6182 pr_warn("Unable to set relax_domain_level\n");
6186 __setup("relax_domain_level=", setup_relax_domain_level);
6188 static void set_domain_attribute(struct sched_domain *sd,
6189 struct sched_domain_attr *attr)
6193 if (!attr || attr->relax_domain_level < 0) {
6194 if (default_relax_domain_level < 0)
6197 request = default_relax_domain_level;
6199 request = attr->relax_domain_level;
6200 if (request < sd->level) {
6201 /* turn off idle balance on this domain */
6202 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6204 /* turn on idle balance on this domain */
6205 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6209 static void __sdt_free(const struct cpumask *cpu_map);
6210 static int __sdt_alloc(const struct cpumask *cpu_map);
6212 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6213 const struct cpumask *cpu_map)
6217 if (!atomic_read(&d->rd->refcount))
6218 free_rootdomain(&d->rd->rcu); /* fall through */
6220 free_percpu(d->sd); /* fall through */
6222 __sdt_free(cpu_map); /* fall through */
6228 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6229 const struct cpumask *cpu_map)
6231 memset(d, 0, sizeof(*d));
6233 if (__sdt_alloc(cpu_map))
6234 return sa_sd_storage;
6235 d->sd = alloc_percpu(struct sched_domain *);
6237 return sa_sd_storage;
6238 d->rd = alloc_rootdomain();
6241 return sa_rootdomain;
6245 * NULL the sd_data elements we've used to build the sched_domain and
6246 * sched_group structure so that the subsequent __free_domain_allocs()
6247 * will not free the data we're using.
6249 static void claim_allocations(int cpu, struct sched_domain *sd)
6251 struct sd_data *sdd = sd->private;
6253 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6254 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6256 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6257 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6259 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6260 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6264 static int sched_domains_numa_levels;
6265 enum numa_topology_type sched_numa_topology_type;
6266 static int *sched_domains_numa_distance;
6267 int sched_max_numa_distance;
6268 static struct cpumask ***sched_domains_numa_masks;
6269 static int sched_domains_curr_level;
6273 * SD_flags allowed in topology descriptions.
6275 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6276 * SD_SHARE_PKG_RESOURCES - describes shared caches
6277 * SD_NUMA - describes NUMA topologies
6278 * SD_SHARE_POWERDOMAIN - describes shared power domain
6281 * SD_ASYM_PACKING - describes SMT quirks
6283 #define TOPOLOGY_SD_FLAGS \
6284 (SD_SHARE_CPUCAPACITY | \
6285 SD_SHARE_PKG_RESOURCES | \
6288 SD_SHARE_POWERDOMAIN)
6290 static struct sched_domain *
6291 sd_init(struct sched_domain_topology_level *tl, int cpu)
6293 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6294 int sd_weight, sd_flags = 0;
6298 * Ugly hack to pass state to sd_numa_mask()...
6300 sched_domains_curr_level = tl->numa_level;
6303 sd_weight = cpumask_weight(tl->mask(cpu));
6306 sd_flags = (*tl->sd_flags)();
6307 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6308 "wrong sd_flags in topology description\n"))
6309 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6311 *sd = (struct sched_domain){
6312 .min_interval = sd_weight,
6313 .max_interval = 2*sd_weight,
6315 .imbalance_pct = 125,
6317 .cache_nice_tries = 0,
6324 .flags = 1*SD_LOAD_BALANCE
6325 | 1*SD_BALANCE_NEWIDLE
6330 | 0*SD_SHARE_CPUCAPACITY
6331 | 0*SD_SHARE_PKG_RESOURCES
6333 | 0*SD_PREFER_SIBLING
6338 .last_balance = jiffies,
6339 .balance_interval = sd_weight,
6341 .max_newidle_lb_cost = 0,
6342 .next_decay_max_lb_cost = jiffies,
6343 #ifdef CONFIG_SCHED_DEBUG
6349 * Convert topological properties into behaviour.
6352 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6353 sd->flags |= SD_PREFER_SIBLING;
6354 sd->imbalance_pct = 110;
6355 sd->smt_gain = 1178; /* ~15% */
6357 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6358 sd->imbalance_pct = 117;
6359 sd->cache_nice_tries = 1;
6363 } else if (sd->flags & SD_NUMA) {
6364 sd->cache_nice_tries = 2;
6368 sd->flags |= SD_SERIALIZE;
6369 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6370 sd->flags &= ~(SD_BALANCE_EXEC |
6377 sd->flags |= SD_PREFER_SIBLING;
6378 sd->cache_nice_tries = 1;
6383 sd->private = &tl->data;
6389 * Topology list, bottom-up.
6391 static struct sched_domain_topology_level default_topology[] = {
6392 #ifdef CONFIG_SCHED_SMT
6393 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6395 #ifdef CONFIG_SCHED_MC
6396 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6398 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6402 static struct sched_domain_topology_level *sched_domain_topology =
6405 #define for_each_sd_topology(tl) \
6406 for (tl = sched_domain_topology; tl->mask; tl++)
6408 void set_sched_topology(struct sched_domain_topology_level *tl)
6410 sched_domain_topology = tl;
6415 static const struct cpumask *sd_numa_mask(int cpu)
6417 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6420 static void sched_numa_warn(const char *str)
6422 static int done = false;
6430 printk(KERN_WARNING "ERROR: %s\n\n", str);
6432 for (i = 0; i < nr_node_ids; i++) {
6433 printk(KERN_WARNING " ");
6434 for (j = 0; j < nr_node_ids; j++)
6435 printk(KERN_CONT "%02d ", node_distance(i,j));
6436 printk(KERN_CONT "\n");
6438 printk(KERN_WARNING "\n");
6441 bool find_numa_distance(int distance)
6445 if (distance == node_distance(0, 0))
6448 for (i = 0; i < sched_domains_numa_levels; i++) {
6449 if (sched_domains_numa_distance[i] == distance)
6457 * A system can have three types of NUMA topology:
6458 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6459 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6460 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6462 * The difference between a glueless mesh topology and a backplane
6463 * topology lies in whether communication between not directly
6464 * connected nodes goes through intermediary nodes (where programs
6465 * could run), or through backplane controllers. This affects
6466 * placement of programs.
6468 * The type of topology can be discerned with the following tests:
6469 * - If the maximum distance between any nodes is 1 hop, the system
6470 * is directly connected.
6471 * - If for two nodes A and B, located N > 1 hops away from each other,
6472 * there is an intermediary node C, which is < N hops away from both
6473 * nodes A and B, the system is a glueless mesh.
6475 static void init_numa_topology_type(void)
6479 n = sched_max_numa_distance;
6481 if (sched_domains_numa_levels <= 1) {
6482 sched_numa_topology_type = NUMA_DIRECT;
6486 for_each_online_node(a) {
6487 for_each_online_node(b) {
6488 /* Find two nodes furthest removed from each other. */
6489 if (node_distance(a, b) < n)
6492 /* Is there an intermediary node between a and b? */
6493 for_each_online_node(c) {
6494 if (node_distance(a, c) < n &&
6495 node_distance(b, c) < n) {
6496 sched_numa_topology_type =
6502 sched_numa_topology_type = NUMA_BACKPLANE;
6508 static void sched_init_numa(void)
6510 int next_distance, curr_distance = node_distance(0, 0);
6511 struct sched_domain_topology_level *tl;
6515 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6516 if (!sched_domains_numa_distance)
6520 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6521 * unique distances in the node_distance() table.
6523 * Assumes node_distance(0,j) includes all distances in
6524 * node_distance(i,j) in order to avoid cubic time.
6526 next_distance = curr_distance;
6527 for (i = 0; i < nr_node_ids; i++) {
6528 for (j = 0; j < nr_node_ids; j++) {
6529 for (k = 0; k < nr_node_ids; k++) {
6530 int distance = node_distance(i, k);
6532 if (distance > curr_distance &&
6533 (distance < next_distance ||
6534 next_distance == curr_distance))
6535 next_distance = distance;
6538 * While not a strong assumption it would be nice to know
6539 * about cases where if node A is connected to B, B is not
6540 * equally connected to A.
6542 if (sched_debug() && node_distance(k, i) != distance)
6543 sched_numa_warn("Node-distance not symmetric");
6545 if (sched_debug() && i && !find_numa_distance(distance))
6546 sched_numa_warn("Node-0 not representative");
6548 if (next_distance != curr_distance) {
6549 sched_domains_numa_distance[level++] = next_distance;
6550 sched_domains_numa_levels = level;
6551 curr_distance = next_distance;
6556 * In case of sched_debug() we verify the above assumption.
6566 * 'level' contains the number of unique distances, excluding the
6567 * identity distance node_distance(i,i).
6569 * The sched_domains_numa_distance[] array includes the actual distance
6574 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6575 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6576 * the array will contain less then 'level' members. This could be
6577 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6578 * in other functions.
6580 * We reset it to 'level' at the end of this function.
6582 sched_domains_numa_levels = 0;
6584 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6585 if (!sched_domains_numa_masks)
6589 * Now for each level, construct a mask per node which contains all
6590 * cpus of nodes that are that many hops away from us.
6592 for (i = 0; i < level; i++) {
6593 sched_domains_numa_masks[i] =
6594 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6595 if (!sched_domains_numa_masks[i])
6598 for (j = 0; j < nr_node_ids; j++) {
6599 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6603 sched_domains_numa_masks[i][j] = mask;
6606 if (node_distance(j, k) > sched_domains_numa_distance[i])
6609 cpumask_or(mask, mask, cpumask_of_node(k));
6614 /* Compute default topology size */
6615 for (i = 0; sched_domain_topology[i].mask; i++);
6617 tl = kzalloc((i + level + 1) *
6618 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6623 * Copy the default topology bits..
6625 for (i = 0; sched_domain_topology[i].mask; i++)
6626 tl[i] = sched_domain_topology[i];
6629 * .. and append 'j' levels of NUMA goodness.
6631 for (j = 0; j < level; i++, j++) {
6632 tl[i] = (struct sched_domain_topology_level){
6633 .mask = sd_numa_mask,
6634 .sd_flags = cpu_numa_flags,
6635 .flags = SDTL_OVERLAP,
6641 sched_domain_topology = tl;
6643 sched_domains_numa_levels = level;
6644 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6646 init_numa_topology_type();
6649 static void sched_domains_numa_masks_set(int cpu)
6652 int node = cpu_to_node(cpu);
6654 for (i = 0; i < sched_domains_numa_levels; i++) {
6655 for (j = 0; j < nr_node_ids; j++) {
6656 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6657 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6662 static void sched_domains_numa_masks_clear(int cpu)
6665 for (i = 0; i < sched_domains_numa_levels; i++) {
6666 for (j = 0; j < nr_node_ids; j++)
6667 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6672 * Update sched_domains_numa_masks[level][node] array when new cpus
6675 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6676 unsigned long action,
6679 int cpu = (long)hcpu;
6681 switch (action & ~CPU_TASKS_FROZEN) {
6683 sched_domains_numa_masks_set(cpu);
6687 sched_domains_numa_masks_clear(cpu);
6697 static inline void sched_init_numa(void)
6701 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6702 unsigned long action,
6707 #endif /* CONFIG_NUMA */
6709 static int __sdt_alloc(const struct cpumask *cpu_map)
6711 struct sched_domain_topology_level *tl;
6714 for_each_sd_topology(tl) {
6715 struct sd_data *sdd = &tl->data;
6717 sdd->sd = alloc_percpu(struct sched_domain *);
6721 sdd->sg = alloc_percpu(struct sched_group *);
6725 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6729 for_each_cpu(j, cpu_map) {
6730 struct sched_domain *sd;
6731 struct sched_group *sg;
6732 struct sched_group_capacity *sgc;
6734 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6735 GFP_KERNEL, cpu_to_node(j));
6739 *per_cpu_ptr(sdd->sd, j) = sd;
6741 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6742 GFP_KERNEL, cpu_to_node(j));
6748 *per_cpu_ptr(sdd->sg, j) = sg;
6750 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6751 GFP_KERNEL, cpu_to_node(j));
6755 *per_cpu_ptr(sdd->sgc, j) = sgc;
6762 static void __sdt_free(const struct cpumask *cpu_map)
6764 struct sched_domain_topology_level *tl;
6767 for_each_sd_topology(tl) {
6768 struct sd_data *sdd = &tl->data;
6770 for_each_cpu(j, cpu_map) {
6771 struct sched_domain *sd;
6774 sd = *per_cpu_ptr(sdd->sd, j);
6775 if (sd && (sd->flags & SD_OVERLAP))
6776 free_sched_groups(sd->groups, 0);
6777 kfree(*per_cpu_ptr(sdd->sd, j));
6781 kfree(*per_cpu_ptr(sdd->sg, j));
6783 kfree(*per_cpu_ptr(sdd->sgc, j));
6785 free_percpu(sdd->sd);
6787 free_percpu(sdd->sg);
6789 free_percpu(sdd->sgc);
6794 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6795 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6796 struct sched_domain *child, int cpu)
6798 struct sched_domain *sd = sd_init(tl, cpu);
6802 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6804 sd->level = child->level + 1;
6805 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6809 if (!cpumask_subset(sched_domain_span(child),
6810 sched_domain_span(sd))) {
6811 pr_err("BUG: arch topology borken\n");
6812 #ifdef CONFIG_SCHED_DEBUG
6813 pr_err(" the %s domain not a subset of the %s domain\n",
6814 child->name, sd->name);
6816 /* Fixup, ensure @sd has at least @child cpus. */
6817 cpumask_or(sched_domain_span(sd),
6818 sched_domain_span(sd),
6819 sched_domain_span(child));
6823 set_domain_attribute(sd, attr);
6829 * Build sched domains for a given set of cpus and attach the sched domains
6830 * to the individual cpus
6832 static int build_sched_domains(const struct cpumask *cpu_map,
6833 struct sched_domain_attr *attr)
6835 enum s_alloc alloc_state;
6836 struct sched_domain *sd;
6838 int i, ret = -ENOMEM;
6840 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6841 if (alloc_state != sa_rootdomain)
6844 /* Set up domains for cpus specified by the cpu_map. */
6845 for_each_cpu(i, cpu_map) {
6846 struct sched_domain_topology_level *tl;
6849 for_each_sd_topology(tl) {
6850 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6851 if (tl == sched_domain_topology)
6852 *per_cpu_ptr(d.sd, i) = sd;
6853 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6854 sd->flags |= SD_OVERLAP;
6855 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6860 /* Build the groups for the domains */
6861 for_each_cpu(i, cpu_map) {
6862 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6863 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6864 if (sd->flags & SD_OVERLAP) {
6865 if (build_overlap_sched_groups(sd, i))
6868 if (build_sched_groups(sd, i))
6874 /* Calculate CPU capacity for physical packages and nodes */
6875 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6876 if (!cpumask_test_cpu(i, cpu_map))
6879 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6880 claim_allocations(i, sd);
6881 init_sched_groups_capacity(i, sd);
6885 /* Attach the domains */
6887 for_each_cpu(i, cpu_map) {
6888 sd = *per_cpu_ptr(d.sd, i);
6889 cpu_attach_domain(sd, d.rd, i);
6895 __free_domain_allocs(&d, alloc_state, cpu_map);
6899 static cpumask_var_t *doms_cur; /* current sched domains */
6900 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6901 static struct sched_domain_attr *dattr_cur;
6902 /* attribues of custom domains in 'doms_cur' */
6905 * Special case: If a kmalloc of a doms_cur partition (array of
6906 * cpumask) fails, then fallback to a single sched domain,
6907 * as determined by the single cpumask fallback_doms.
6909 static cpumask_var_t fallback_doms;
6912 * arch_update_cpu_topology lets virtualized architectures update the
6913 * cpu core maps. It is supposed to return 1 if the topology changed
6914 * or 0 if it stayed the same.
6916 int __weak arch_update_cpu_topology(void)
6921 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6924 cpumask_var_t *doms;
6926 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6929 for (i = 0; i < ndoms; i++) {
6930 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6931 free_sched_domains(doms, i);
6938 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6941 for (i = 0; i < ndoms; i++)
6942 free_cpumask_var(doms[i]);
6947 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6948 * For now this just excludes isolated cpus, but could be used to
6949 * exclude other special cases in the future.
6951 static int init_sched_domains(const struct cpumask *cpu_map)
6955 arch_update_cpu_topology();
6957 doms_cur = alloc_sched_domains(ndoms_cur);
6959 doms_cur = &fallback_doms;
6960 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6961 err = build_sched_domains(doms_cur[0], NULL);
6962 register_sched_domain_sysctl();
6968 * Detach sched domains from a group of cpus specified in cpu_map
6969 * These cpus will now be attached to the NULL domain
6971 static void detach_destroy_domains(const struct cpumask *cpu_map)
6976 for_each_cpu(i, cpu_map)
6977 cpu_attach_domain(NULL, &def_root_domain, i);
6981 /* handle null as "default" */
6982 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6983 struct sched_domain_attr *new, int idx_new)
6985 struct sched_domain_attr tmp;
6992 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6993 new ? (new + idx_new) : &tmp,
6994 sizeof(struct sched_domain_attr));
6998 * Partition sched domains as specified by the 'ndoms_new'
6999 * cpumasks in the array doms_new[] of cpumasks. This compares
7000 * doms_new[] to the current sched domain partitioning, doms_cur[].
7001 * It destroys each deleted domain and builds each new domain.
7003 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7004 * The masks don't intersect (don't overlap.) We should setup one
7005 * sched domain for each mask. CPUs not in any of the cpumasks will
7006 * not be load balanced. If the same cpumask appears both in the
7007 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7010 * The passed in 'doms_new' should be allocated using
7011 * alloc_sched_domains. This routine takes ownership of it and will
7012 * free_sched_domains it when done with it. If the caller failed the
7013 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7014 * and partition_sched_domains() will fallback to the single partition
7015 * 'fallback_doms', it also forces the domains to be rebuilt.
7017 * If doms_new == NULL it will be replaced with cpu_online_mask.
7018 * ndoms_new == 0 is a special case for destroying existing domains,
7019 * and it will not create the default domain.
7021 * Call with hotplug lock held
7023 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7024 struct sched_domain_attr *dattr_new)
7029 mutex_lock(&sched_domains_mutex);
7031 /* always unregister in case we don't destroy any domains */
7032 unregister_sched_domain_sysctl();
7034 /* Let architecture update cpu core mappings. */
7035 new_topology = arch_update_cpu_topology();
7037 n = doms_new ? ndoms_new : 0;
7039 /* Destroy deleted domains */
7040 for (i = 0; i < ndoms_cur; i++) {
7041 for (j = 0; j < n && !new_topology; j++) {
7042 if (cpumask_equal(doms_cur[i], doms_new[j])
7043 && dattrs_equal(dattr_cur, i, dattr_new, j))
7046 /* no match - a current sched domain not in new doms_new[] */
7047 detach_destroy_domains(doms_cur[i]);
7053 if (doms_new == NULL) {
7055 doms_new = &fallback_doms;
7056 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7057 WARN_ON_ONCE(dattr_new);
7060 /* Build new domains */
7061 for (i = 0; i < ndoms_new; i++) {
7062 for (j = 0; j < n && !new_topology; j++) {
7063 if (cpumask_equal(doms_new[i], doms_cur[j])
7064 && dattrs_equal(dattr_new, i, dattr_cur, j))
7067 /* no match - add a new doms_new */
7068 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7073 /* Remember the new sched domains */
7074 if (doms_cur != &fallback_doms)
7075 free_sched_domains(doms_cur, ndoms_cur);
7076 kfree(dattr_cur); /* kfree(NULL) is safe */
7077 doms_cur = doms_new;
7078 dattr_cur = dattr_new;
7079 ndoms_cur = ndoms_new;
7081 register_sched_domain_sysctl();
7083 mutex_unlock(&sched_domains_mutex);
7086 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7089 * Update cpusets according to cpu_active mask. If cpusets are
7090 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7091 * around partition_sched_domains().
7093 * If we come here as part of a suspend/resume, don't touch cpusets because we
7094 * want to restore it back to its original state upon resume anyway.
7096 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7100 case CPU_ONLINE_FROZEN:
7101 case CPU_DOWN_FAILED_FROZEN:
7104 * num_cpus_frozen tracks how many CPUs are involved in suspend
7105 * resume sequence. As long as this is not the last online
7106 * operation in the resume sequence, just build a single sched
7107 * domain, ignoring cpusets.
7110 if (likely(num_cpus_frozen)) {
7111 partition_sched_domains(1, NULL, NULL);
7116 * This is the last CPU online operation. So fall through and
7117 * restore the original sched domains by considering the
7118 * cpuset configurations.
7122 cpuset_update_active_cpus(true);
7130 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7133 unsigned long flags;
7134 long cpu = (long)hcpu;
7140 case CPU_DOWN_PREPARE:
7141 rcu_read_lock_sched();
7142 dl_b = dl_bw_of(cpu);
7144 raw_spin_lock_irqsave(&dl_b->lock, flags);
7145 cpus = dl_bw_cpus(cpu);
7146 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7147 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7149 rcu_read_unlock_sched();
7152 return notifier_from_errno(-EBUSY);
7153 cpuset_update_active_cpus(false);
7155 case CPU_DOWN_PREPARE_FROZEN:
7157 partition_sched_domains(1, NULL, NULL);
7165 void __init sched_init_smp(void)
7167 cpumask_var_t non_isolated_cpus;
7169 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7170 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7175 * There's no userspace yet to cause hotplug operations; hence all the
7176 * cpu masks are stable and all blatant races in the below code cannot
7179 mutex_lock(&sched_domains_mutex);
7180 init_sched_domains(cpu_active_mask);
7181 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7182 if (cpumask_empty(non_isolated_cpus))
7183 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7184 mutex_unlock(&sched_domains_mutex);
7186 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7187 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7188 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7192 /* Move init over to a non-isolated CPU */
7193 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7195 sched_init_granularity();
7196 free_cpumask_var(non_isolated_cpus);
7198 init_sched_rt_class();
7199 init_sched_dl_class();
7202 void __init sched_init_smp(void)
7204 sched_init_granularity();
7206 #endif /* CONFIG_SMP */
7208 int in_sched_functions(unsigned long addr)
7210 return in_lock_functions(addr) ||
7211 (addr >= (unsigned long)__sched_text_start
7212 && addr < (unsigned long)__sched_text_end);
7215 #ifdef CONFIG_CGROUP_SCHED
7217 * Default task group.
7218 * Every task in system belongs to this group at bootup.
7220 struct task_group root_task_group;
7221 LIST_HEAD(task_groups);
7223 /* Cacheline aligned slab cache for task_group */
7224 static struct kmem_cache *task_group_cache __read_mostly;
7227 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7229 void __init sched_init(void)
7232 unsigned long alloc_size = 0, ptr;
7234 #ifdef CONFIG_FAIR_GROUP_SCHED
7235 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7237 #ifdef CONFIG_RT_GROUP_SCHED
7238 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7241 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7243 #ifdef CONFIG_FAIR_GROUP_SCHED
7244 root_task_group.se = (struct sched_entity **)ptr;
7245 ptr += nr_cpu_ids * sizeof(void **);
7247 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7248 ptr += nr_cpu_ids * sizeof(void **);
7250 #endif /* CONFIG_FAIR_GROUP_SCHED */
7251 #ifdef CONFIG_RT_GROUP_SCHED
7252 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7253 ptr += nr_cpu_ids * sizeof(void **);
7255 root_task_group.rt_rq = (struct rt_rq **)ptr;
7256 ptr += nr_cpu_ids * sizeof(void **);
7258 #endif /* CONFIG_RT_GROUP_SCHED */
7260 #ifdef CONFIG_CPUMASK_OFFSTACK
7261 for_each_possible_cpu(i) {
7262 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7263 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7265 #endif /* CONFIG_CPUMASK_OFFSTACK */
7267 init_rt_bandwidth(&def_rt_bandwidth,
7268 global_rt_period(), global_rt_runtime());
7269 init_dl_bandwidth(&def_dl_bandwidth,
7270 global_rt_period(), global_rt_runtime());
7273 init_defrootdomain();
7276 #ifdef CONFIG_RT_GROUP_SCHED
7277 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7278 global_rt_period(), global_rt_runtime());
7279 #endif /* CONFIG_RT_GROUP_SCHED */
7281 #ifdef CONFIG_CGROUP_SCHED
7282 task_group_cache = KMEM_CACHE(task_group, 0);
7284 list_add(&root_task_group.list, &task_groups);
7285 INIT_LIST_HEAD(&root_task_group.children);
7286 INIT_LIST_HEAD(&root_task_group.siblings);
7287 autogroup_init(&init_task);
7288 #endif /* CONFIG_CGROUP_SCHED */
7290 for_each_possible_cpu(i) {
7294 raw_spin_lock_init(&rq->lock);
7296 rq->calc_load_active = 0;
7297 rq->calc_load_update = jiffies + LOAD_FREQ;
7298 init_cfs_rq(&rq->cfs);
7299 init_rt_rq(&rq->rt);
7300 init_dl_rq(&rq->dl);
7301 #ifdef CONFIG_FAIR_GROUP_SCHED
7302 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7303 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7305 * How much cpu bandwidth does root_task_group get?
7307 * In case of task-groups formed thr' the cgroup filesystem, it
7308 * gets 100% of the cpu resources in the system. This overall
7309 * system cpu resource is divided among the tasks of
7310 * root_task_group and its child task-groups in a fair manner,
7311 * based on each entity's (task or task-group's) weight
7312 * (se->load.weight).
7314 * In other words, if root_task_group has 10 tasks of weight
7315 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7316 * then A0's share of the cpu resource is:
7318 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7320 * We achieve this by letting root_task_group's tasks sit
7321 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7323 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7324 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7325 #endif /* CONFIG_FAIR_GROUP_SCHED */
7327 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7328 #ifdef CONFIG_RT_GROUP_SCHED
7329 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7332 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7333 rq->cpu_load[j] = 0;
7335 rq->last_load_update_tick = jiffies;
7340 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7341 rq->balance_callback = NULL;
7342 rq->active_balance = 0;
7343 rq->next_balance = jiffies;
7348 rq->avg_idle = 2*sysctl_sched_migration_cost;
7349 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7351 INIT_LIST_HEAD(&rq->cfs_tasks);
7353 rq_attach_root(rq, &def_root_domain);
7354 #ifdef CONFIG_NO_HZ_COMMON
7357 #ifdef CONFIG_NO_HZ_FULL
7358 rq->last_sched_tick = 0;
7362 atomic_set(&rq->nr_iowait, 0);
7365 set_load_weight(&init_task);
7367 #ifdef CONFIG_PREEMPT_NOTIFIERS
7368 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7372 * The boot idle thread does lazy MMU switching as well:
7374 atomic_inc(&init_mm.mm_count);
7375 enter_lazy_tlb(&init_mm, current);
7378 * During early bootup we pretend to be a normal task:
7380 current->sched_class = &fair_sched_class;
7383 * Make us the idle thread. Technically, schedule() should not be
7384 * called from this thread, however somewhere below it might be,
7385 * but because we are the idle thread, we just pick up running again
7386 * when this runqueue becomes "idle".
7388 init_idle(current, smp_processor_id());
7390 calc_load_update = jiffies + LOAD_FREQ;
7393 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7394 /* May be allocated at isolcpus cmdline parse time */
7395 if (cpu_isolated_map == NULL)
7396 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7397 idle_thread_set_boot_cpu();
7398 set_cpu_rq_start_time();
7400 init_sched_fair_class();
7402 scheduler_running = 1;
7405 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7406 static inline int preempt_count_equals(int preempt_offset)
7408 int nested = preempt_count() + rcu_preempt_depth();
7410 return (nested == preempt_offset);
7413 void __might_sleep(const char *file, int line, int preempt_offset)
7416 * Blocking primitives will set (and therefore destroy) current->state,
7417 * since we will exit with TASK_RUNNING make sure we enter with it,
7418 * otherwise we will destroy state.
7420 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7421 "do not call blocking ops when !TASK_RUNNING; "
7422 "state=%lx set at [<%p>] %pS\n",
7424 (void *)current->task_state_change,
7425 (void *)current->task_state_change);
7427 ___might_sleep(file, line, preempt_offset);
7429 EXPORT_SYMBOL(__might_sleep);
7431 void ___might_sleep(const char *file, int line, int preempt_offset)
7433 static unsigned long prev_jiffy; /* ratelimiting */
7435 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7436 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7437 !is_idle_task(current)) ||
7438 system_state != SYSTEM_RUNNING || oops_in_progress)
7440 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7442 prev_jiffy = jiffies;
7445 "BUG: sleeping function called from invalid context at %s:%d\n",
7448 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7449 in_atomic(), irqs_disabled(),
7450 current->pid, current->comm);
7452 if (task_stack_end_corrupted(current))
7453 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7455 debug_show_held_locks(current);
7456 if (irqs_disabled())
7457 print_irqtrace_events(current);
7458 #ifdef CONFIG_DEBUG_PREEMPT
7459 if (!preempt_count_equals(preempt_offset)) {
7460 pr_err("Preemption disabled at:");
7461 print_ip_sym(current->preempt_disable_ip);
7467 EXPORT_SYMBOL(___might_sleep);
7470 #ifdef CONFIG_MAGIC_SYSRQ
7471 void normalize_rt_tasks(void)
7473 struct task_struct *g, *p;
7474 struct sched_attr attr = {
7475 .sched_policy = SCHED_NORMAL,
7478 read_lock(&tasklist_lock);
7479 for_each_process_thread(g, p) {
7481 * Only normalize user tasks:
7483 if (p->flags & PF_KTHREAD)
7486 p->se.exec_start = 0;
7487 #ifdef CONFIG_SCHEDSTATS
7488 p->se.statistics.wait_start = 0;
7489 p->se.statistics.sleep_start = 0;
7490 p->se.statistics.block_start = 0;
7493 if (!dl_task(p) && !rt_task(p)) {
7495 * Renice negative nice level userspace
7498 if (task_nice(p) < 0)
7499 set_user_nice(p, 0);
7503 __sched_setscheduler(p, &attr, false, false);
7505 read_unlock(&tasklist_lock);
7508 #endif /* CONFIG_MAGIC_SYSRQ */
7510 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7512 * These functions are only useful for the IA64 MCA handling, or kdb.
7514 * They can only be called when the whole system has been
7515 * stopped - every CPU needs to be quiescent, and no scheduling
7516 * activity can take place. Using them for anything else would
7517 * be a serious bug, and as a result, they aren't even visible
7518 * under any other configuration.
7522 * curr_task - return the current task for a given cpu.
7523 * @cpu: the processor in question.
7525 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7527 * Return: The current task for @cpu.
7529 struct task_struct *curr_task(int cpu)
7531 return cpu_curr(cpu);
7534 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7538 * set_curr_task - set the current task for a given cpu.
7539 * @cpu: the processor in question.
7540 * @p: the task pointer to set.
7542 * Description: This function must only be used when non-maskable interrupts
7543 * are serviced on a separate stack. It allows the architecture to switch the
7544 * notion of the current task on a cpu in a non-blocking manner. This function
7545 * must be called with all CPU's synchronized, and interrupts disabled, the
7546 * and caller must save the original value of the current task (see
7547 * curr_task() above) and restore that value before reenabling interrupts and
7548 * re-starting the system.
7550 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7552 void set_curr_task(int cpu, struct task_struct *p)
7559 #ifdef CONFIG_CGROUP_SCHED
7560 /* task_group_lock serializes the addition/removal of task groups */
7561 static DEFINE_SPINLOCK(task_group_lock);
7563 static void sched_free_group(struct task_group *tg)
7565 free_fair_sched_group(tg);
7566 free_rt_sched_group(tg);
7568 kmem_cache_free(task_group_cache, tg);
7571 /* allocate runqueue etc for a new task group */
7572 struct task_group *sched_create_group(struct task_group *parent)
7574 struct task_group *tg;
7576 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7578 return ERR_PTR(-ENOMEM);
7580 if (!alloc_fair_sched_group(tg, parent))
7583 if (!alloc_rt_sched_group(tg, parent))
7589 sched_free_group(tg);
7590 return ERR_PTR(-ENOMEM);
7593 void sched_online_group(struct task_group *tg, struct task_group *parent)
7595 unsigned long flags;
7597 spin_lock_irqsave(&task_group_lock, flags);
7598 list_add_rcu(&tg->list, &task_groups);
7600 WARN_ON(!parent); /* root should already exist */
7602 tg->parent = parent;
7603 INIT_LIST_HEAD(&tg->children);
7604 list_add_rcu(&tg->siblings, &parent->children);
7605 spin_unlock_irqrestore(&task_group_lock, flags);
7608 /* rcu callback to free various structures associated with a task group */
7609 static void sched_free_group_rcu(struct rcu_head *rhp)
7611 /* now it should be safe to free those cfs_rqs */
7612 sched_free_group(container_of(rhp, struct task_group, rcu));
7615 void sched_destroy_group(struct task_group *tg)
7617 /* wait for possible concurrent references to cfs_rqs complete */
7618 call_rcu(&tg->rcu, sched_free_group_rcu);
7621 void sched_offline_group(struct task_group *tg)
7623 unsigned long flags;
7625 /* end participation in shares distribution */
7626 unregister_fair_sched_group(tg);
7628 spin_lock_irqsave(&task_group_lock, flags);
7629 list_del_rcu(&tg->list);
7630 list_del_rcu(&tg->siblings);
7631 spin_unlock_irqrestore(&task_group_lock, flags);
7634 /* change task's runqueue when it moves between groups.
7635 * The caller of this function should have put the task in its new group
7636 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7637 * reflect its new group.
7639 void sched_move_task(struct task_struct *tsk)
7641 struct task_group *tg;
7642 int queued, running;
7643 unsigned long flags;
7646 rq = task_rq_lock(tsk, &flags);
7648 running = task_current(rq, tsk);
7649 queued = task_on_rq_queued(tsk);
7652 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7653 if (unlikely(running))
7654 put_prev_task(rq, tsk);
7657 * All callers are synchronized by task_rq_lock(); we do not use RCU
7658 * which is pointless here. Thus, we pass "true" to task_css_check()
7659 * to prevent lockdep warnings.
7661 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7662 struct task_group, css);
7663 tg = autogroup_task_group(tsk, tg);
7664 tsk->sched_task_group = tg;
7666 #ifdef CONFIG_FAIR_GROUP_SCHED
7667 if (tsk->sched_class->task_move_group)
7668 tsk->sched_class->task_move_group(tsk);
7671 set_task_rq(tsk, task_cpu(tsk));
7673 if (unlikely(running))
7674 tsk->sched_class->set_curr_task(rq);
7676 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7678 task_rq_unlock(rq, tsk, &flags);
7680 #endif /* CONFIG_CGROUP_SCHED */
7682 #ifdef CONFIG_RT_GROUP_SCHED
7684 * Ensure that the real time constraints are schedulable.
7686 static DEFINE_MUTEX(rt_constraints_mutex);
7688 /* Must be called with tasklist_lock held */
7689 static inline int tg_has_rt_tasks(struct task_group *tg)
7691 struct task_struct *g, *p;
7694 * Autogroups do not have RT tasks; see autogroup_create().
7696 if (task_group_is_autogroup(tg))
7699 for_each_process_thread(g, p) {
7700 if (rt_task(p) && task_group(p) == tg)
7707 struct rt_schedulable_data {
7708 struct task_group *tg;
7713 static int tg_rt_schedulable(struct task_group *tg, void *data)
7715 struct rt_schedulable_data *d = data;
7716 struct task_group *child;
7717 unsigned long total, sum = 0;
7718 u64 period, runtime;
7720 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7721 runtime = tg->rt_bandwidth.rt_runtime;
7724 period = d->rt_period;
7725 runtime = d->rt_runtime;
7729 * Cannot have more runtime than the period.
7731 if (runtime > period && runtime != RUNTIME_INF)
7735 * Ensure we don't starve existing RT tasks.
7737 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7740 total = to_ratio(period, runtime);
7743 * Nobody can have more than the global setting allows.
7745 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7749 * The sum of our children's runtime should not exceed our own.
7751 list_for_each_entry_rcu(child, &tg->children, siblings) {
7752 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7753 runtime = child->rt_bandwidth.rt_runtime;
7755 if (child == d->tg) {
7756 period = d->rt_period;
7757 runtime = d->rt_runtime;
7760 sum += to_ratio(period, runtime);
7769 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7773 struct rt_schedulable_data data = {
7775 .rt_period = period,
7776 .rt_runtime = runtime,
7780 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7786 static int tg_set_rt_bandwidth(struct task_group *tg,
7787 u64 rt_period, u64 rt_runtime)
7792 * Disallowing the root group RT runtime is BAD, it would disallow the
7793 * kernel creating (and or operating) RT threads.
7795 if (tg == &root_task_group && rt_runtime == 0)
7798 /* No period doesn't make any sense. */
7802 mutex_lock(&rt_constraints_mutex);
7803 read_lock(&tasklist_lock);
7804 err = __rt_schedulable(tg, rt_period, rt_runtime);
7808 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7809 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7810 tg->rt_bandwidth.rt_runtime = rt_runtime;
7812 for_each_possible_cpu(i) {
7813 struct rt_rq *rt_rq = tg->rt_rq[i];
7815 raw_spin_lock(&rt_rq->rt_runtime_lock);
7816 rt_rq->rt_runtime = rt_runtime;
7817 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7819 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7821 read_unlock(&tasklist_lock);
7822 mutex_unlock(&rt_constraints_mutex);
7827 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7829 u64 rt_runtime, rt_period;
7831 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7832 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7833 if (rt_runtime_us < 0)
7834 rt_runtime = RUNTIME_INF;
7836 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7839 static long sched_group_rt_runtime(struct task_group *tg)
7843 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7846 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7847 do_div(rt_runtime_us, NSEC_PER_USEC);
7848 return rt_runtime_us;
7851 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7853 u64 rt_runtime, rt_period;
7855 rt_period = rt_period_us * NSEC_PER_USEC;
7856 rt_runtime = tg->rt_bandwidth.rt_runtime;
7858 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7861 static long sched_group_rt_period(struct task_group *tg)
7865 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7866 do_div(rt_period_us, NSEC_PER_USEC);
7867 return rt_period_us;
7869 #endif /* CONFIG_RT_GROUP_SCHED */
7871 #ifdef CONFIG_RT_GROUP_SCHED
7872 static int sched_rt_global_constraints(void)
7876 mutex_lock(&rt_constraints_mutex);
7877 read_lock(&tasklist_lock);
7878 ret = __rt_schedulable(NULL, 0, 0);
7879 read_unlock(&tasklist_lock);
7880 mutex_unlock(&rt_constraints_mutex);
7885 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7887 /* Don't accept realtime tasks when there is no way for them to run */
7888 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7894 #else /* !CONFIG_RT_GROUP_SCHED */
7895 static int sched_rt_global_constraints(void)
7897 unsigned long flags;
7900 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7901 for_each_possible_cpu(i) {
7902 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7904 raw_spin_lock(&rt_rq->rt_runtime_lock);
7905 rt_rq->rt_runtime = global_rt_runtime();
7906 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7908 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7912 #endif /* CONFIG_RT_GROUP_SCHED */
7914 static int sched_dl_global_validate(void)
7916 u64 runtime = global_rt_runtime();
7917 u64 period = global_rt_period();
7918 u64 new_bw = to_ratio(period, runtime);
7921 unsigned long flags;
7924 * Here we want to check the bandwidth not being set to some
7925 * value smaller than the currently allocated bandwidth in
7926 * any of the root_domains.
7928 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7929 * cycling on root_domains... Discussion on different/better
7930 * solutions is welcome!
7932 for_each_possible_cpu(cpu) {
7933 rcu_read_lock_sched();
7934 dl_b = dl_bw_of(cpu);
7936 raw_spin_lock_irqsave(&dl_b->lock, flags);
7937 if (new_bw < dl_b->total_bw)
7939 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7941 rcu_read_unlock_sched();
7950 static void sched_dl_do_global(void)
7955 unsigned long flags;
7957 def_dl_bandwidth.dl_period = global_rt_period();
7958 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7960 if (global_rt_runtime() != RUNTIME_INF)
7961 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7964 * FIXME: As above...
7966 for_each_possible_cpu(cpu) {
7967 rcu_read_lock_sched();
7968 dl_b = dl_bw_of(cpu);
7970 raw_spin_lock_irqsave(&dl_b->lock, flags);
7972 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7974 rcu_read_unlock_sched();
7978 static int sched_rt_global_validate(void)
7980 if (sysctl_sched_rt_period <= 0)
7983 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7984 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7990 static void sched_rt_do_global(void)
7992 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7993 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7996 int sched_rt_handler(struct ctl_table *table, int write,
7997 void __user *buffer, size_t *lenp,
8000 int old_period, old_runtime;
8001 static DEFINE_MUTEX(mutex);
8005 old_period = sysctl_sched_rt_period;
8006 old_runtime = sysctl_sched_rt_runtime;
8008 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8010 if (!ret && write) {
8011 ret = sched_rt_global_validate();
8015 ret = sched_dl_global_validate();
8019 ret = sched_rt_global_constraints();
8023 sched_rt_do_global();
8024 sched_dl_do_global();
8028 sysctl_sched_rt_period = old_period;
8029 sysctl_sched_rt_runtime = old_runtime;
8031 mutex_unlock(&mutex);
8036 int sched_rr_handler(struct ctl_table *table, int write,
8037 void __user *buffer, size_t *lenp,
8041 static DEFINE_MUTEX(mutex);
8044 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8045 /* make sure that internally we keep jiffies */
8046 /* also, writing zero resets timeslice to default */
8047 if (!ret && write) {
8048 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8049 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8051 mutex_unlock(&mutex);
8055 #ifdef CONFIG_CGROUP_SCHED
8057 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8059 return css ? container_of(css, struct task_group, css) : NULL;
8062 static struct cgroup_subsys_state *
8063 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8065 struct task_group *parent = css_tg(parent_css);
8066 struct task_group *tg;
8069 /* This is early initialization for the top cgroup */
8070 return &root_task_group.css;
8073 tg = sched_create_group(parent);
8075 return ERR_PTR(-ENOMEM);
8077 sched_online_group(tg, parent);
8082 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8084 struct task_group *tg = css_tg(css);
8086 sched_offline_group(tg);
8089 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8091 struct task_group *tg = css_tg(css);
8094 * Relies on the RCU grace period between css_released() and this.
8096 sched_free_group(tg);
8099 static void cpu_cgroup_fork(struct task_struct *task)
8101 sched_move_task(task);
8104 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8106 struct task_struct *task;
8107 struct cgroup_subsys_state *css;
8109 cgroup_taskset_for_each(task, css, tset) {
8110 #ifdef CONFIG_RT_GROUP_SCHED
8111 if (!sched_rt_can_attach(css_tg(css), task))
8114 /* We don't support RT-tasks being in separate groups */
8115 if (task->sched_class != &fair_sched_class)
8122 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8124 struct task_struct *task;
8125 struct cgroup_subsys_state *css;
8127 cgroup_taskset_for_each(task, css, tset)
8128 sched_move_task(task);
8131 #ifdef CONFIG_FAIR_GROUP_SCHED
8132 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8133 struct cftype *cftype, u64 shareval)
8135 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8138 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8141 struct task_group *tg = css_tg(css);
8143 return (u64) scale_load_down(tg->shares);
8146 #ifdef CONFIG_CFS_BANDWIDTH
8147 static DEFINE_MUTEX(cfs_constraints_mutex);
8149 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8150 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8152 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8154 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8156 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8157 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8159 if (tg == &root_task_group)
8163 * Ensure we have at some amount of bandwidth every period. This is
8164 * to prevent reaching a state of large arrears when throttled via
8165 * entity_tick() resulting in prolonged exit starvation.
8167 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8171 * Likewise, bound things on the otherside by preventing insane quota
8172 * periods. This also allows us to normalize in computing quota
8175 if (period > max_cfs_quota_period)
8179 * Prevent race between setting of cfs_rq->runtime_enabled and
8180 * unthrottle_offline_cfs_rqs().
8183 mutex_lock(&cfs_constraints_mutex);
8184 ret = __cfs_schedulable(tg, period, quota);
8188 runtime_enabled = quota != RUNTIME_INF;
8189 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8191 * If we need to toggle cfs_bandwidth_used, off->on must occur
8192 * before making related changes, and on->off must occur afterwards
8194 if (runtime_enabled && !runtime_was_enabled)
8195 cfs_bandwidth_usage_inc();
8196 raw_spin_lock_irq(&cfs_b->lock);
8197 cfs_b->period = ns_to_ktime(period);
8198 cfs_b->quota = quota;
8200 __refill_cfs_bandwidth_runtime(cfs_b);
8201 /* restart the period timer (if active) to handle new period expiry */
8202 if (runtime_enabled)
8203 start_cfs_bandwidth(cfs_b);
8204 raw_spin_unlock_irq(&cfs_b->lock);
8206 for_each_online_cpu(i) {
8207 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8208 struct rq *rq = cfs_rq->rq;
8210 raw_spin_lock_irq(&rq->lock);
8211 cfs_rq->runtime_enabled = runtime_enabled;
8212 cfs_rq->runtime_remaining = 0;
8214 if (cfs_rq->throttled)
8215 unthrottle_cfs_rq(cfs_rq);
8216 raw_spin_unlock_irq(&rq->lock);
8218 if (runtime_was_enabled && !runtime_enabled)
8219 cfs_bandwidth_usage_dec();
8221 mutex_unlock(&cfs_constraints_mutex);
8227 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8231 period = ktime_to_ns(tg->cfs_bandwidth.period);
8232 if (cfs_quota_us < 0)
8233 quota = RUNTIME_INF;
8235 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8237 return tg_set_cfs_bandwidth(tg, period, quota);
8240 long tg_get_cfs_quota(struct task_group *tg)
8244 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8247 quota_us = tg->cfs_bandwidth.quota;
8248 do_div(quota_us, NSEC_PER_USEC);
8253 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8257 period = (u64)cfs_period_us * NSEC_PER_USEC;
8258 quota = tg->cfs_bandwidth.quota;
8260 return tg_set_cfs_bandwidth(tg, period, quota);
8263 long tg_get_cfs_period(struct task_group *tg)
8267 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8268 do_div(cfs_period_us, NSEC_PER_USEC);
8270 return cfs_period_us;
8273 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8276 return tg_get_cfs_quota(css_tg(css));
8279 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8280 struct cftype *cftype, s64 cfs_quota_us)
8282 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8285 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8288 return tg_get_cfs_period(css_tg(css));
8291 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8292 struct cftype *cftype, u64 cfs_period_us)
8294 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8297 struct cfs_schedulable_data {
8298 struct task_group *tg;
8303 * normalize group quota/period to be quota/max_period
8304 * note: units are usecs
8306 static u64 normalize_cfs_quota(struct task_group *tg,
8307 struct cfs_schedulable_data *d)
8315 period = tg_get_cfs_period(tg);
8316 quota = tg_get_cfs_quota(tg);
8319 /* note: these should typically be equivalent */
8320 if (quota == RUNTIME_INF || quota == -1)
8323 return to_ratio(period, quota);
8326 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8328 struct cfs_schedulable_data *d = data;
8329 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8330 s64 quota = 0, parent_quota = -1;
8333 quota = RUNTIME_INF;
8335 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8337 quota = normalize_cfs_quota(tg, d);
8338 parent_quota = parent_b->hierarchical_quota;
8341 * ensure max(child_quota) <= parent_quota, inherit when no
8344 if (quota == RUNTIME_INF)
8345 quota = parent_quota;
8346 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8349 cfs_b->hierarchical_quota = quota;
8354 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8357 struct cfs_schedulable_data data = {
8363 if (quota != RUNTIME_INF) {
8364 do_div(data.period, NSEC_PER_USEC);
8365 do_div(data.quota, NSEC_PER_USEC);
8369 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8375 static int cpu_stats_show(struct seq_file *sf, void *v)
8377 struct task_group *tg = css_tg(seq_css(sf));
8378 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8380 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8381 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8382 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8386 #endif /* CONFIG_CFS_BANDWIDTH */
8387 #endif /* CONFIG_FAIR_GROUP_SCHED */
8389 #ifdef CONFIG_RT_GROUP_SCHED
8390 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8391 struct cftype *cft, s64 val)
8393 return sched_group_set_rt_runtime(css_tg(css), val);
8396 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8399 return sched_group_rt_runtime(css_tg(css));
8402 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8403 struct cftype *cftype, u64 rt_period_us)
8405 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8408 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8411 return sched_group_rt_period(css_tg(css));
8413 #endif /* CONFIG_RT_GROUP_SCHED */
8415 static struct cftype cpu_files[] = {
8416 #ifdef CONFIG_FAIR_GROUP_SCHED
8419 .read_u64 = cpu_shares_read_u64,
8420 .write_u64 = cpu_shares_write_u64,
8423 #ifdef CONFIG_CFS_BANDWIDTH
8425 .name = "cfs_quota_us",
8426 .read_s64 = cpu_cfs_quota_read_s64,
8427 .write_s64 = cpu_cfs_quota_write_s64,
8430 .name = "cfs_period_us",
8431 .read_u64 = cpu_cfs_period_read_u64,
8432 .write_u64 = cpu_cfs_period_write_u64,
8436 .seq_show = cpu_stats_show,
8439 #ifdef CONFIG_RT_GROUP_SCHED
8441 .name = "rt_runtime_us",
8442 .read_s64 = cpu_rt_runtime_read,
8443 .write_s64 = cpu_rt_runtime_write,
8446 .name = "rt_period_us",
8447 .read_u64 = cpu_rt_period_read_uint,
8448 .write_u64 = cpu_rt_period_write_uint,
8454 struct cgroup_subsys cpu_cgrp_subsys = {
8455 .css_alloc = cpu_cgroup_css_alloc,
8456 .css_released = cpu_cgroup_css_released,
8457 .css_free = cpu_cgroup_css_free,
8458 .fork = cpu_cgroup_fork,
8459 .can_attach = cpu_cgroup_can_attach,
8460 .attach = cpu_cgroup_attach,
8461 .legacy_cftypes = cpu_files,
8465 #endif /* CONFIG_CGROUP_SCHED */
8467 void dump_cpu_task(int cpu)
8469 pr_info("Task dump for CPU %d:\n", cpu);
8470 sched_show_task(cpu_curr(cpu));
8474 * Nice levels are multiplicative, with a gentle 10% change for every
8475 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8476 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8477 * that remained on nice 0.
8479 * The "10% effect" is relative and cumulative: from _any_ nice level,
8480 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8481 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8482 * If a task goes up by ~10% and another task goes down by ~10% then
8483 * the relative distance between them is ~25%.)
8485 const int sched_prio_to_weight[40] = {
8486 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8487 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8488 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8489 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8490 /* 0 */ 1024, 820, 655, 526, 423,
8491 /* 5 */ 335, 272, 215, 172, 137,
8492 /* 10 */ 110, 87, 70, 56, 45,
8493 /* 15 */ 36, 29, 23, 18, 15,
8497 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8499 * In cases where the weight does not change often, we can use the
8500 * precalculated inverse to speed up arithmetics by turning divisions
8501 * into multiplications:
8503 const u32 sched_prio_to_wmult[40] = {
8504 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8505 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8506 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8507 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8508 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8509 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8510 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8511 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,