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 * FIFO realtime policy runs the highest priority task (after DEADLINE).
600 * Other runnable tasks are of a lower priority. The scheduler tick
603 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
608 * Round-robin realtime tasks time slice with other tasks at the same
611 if (rq->rt.rr_nr_running) {
612 if (rq->rt.rr_nr_running == 1)
618 /* Normal multitasking need periodic preemption checks */
619 if (rq->cfs.nr_running > 1)
624 #endif /* CONFIG_NO_HZ_FULL */
626 void sched_avg_update(struct rq *rq)
628 s64 period = sched_avg_period();
630 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
642 #endif /* CONFIG_SMP */
644 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
645 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
647 * Iterate task_group tree rooted at *from, calling @down when first entering a
648 * node and @up when leaving it for the final time.
650 * Caller must hold rcu_lock or sufficient equivalent.
652 int walk_tg_tree_from(struct task_group *from,
653 tg_visitor down, tg_visitor up, void *data)
655 struct task_group *parent, *child;
661 ret = (*down)(parent, data);
664 list_for_each_entry_rcu(child, &parent->children, siblings) {
671 ret = (*up)(parent, data);
672 if (ret || parent == from)
676 parent = parent->parent;
683 int tg_nop(struct task_group *tg, void *data)
689 static void set_load_weight(struct task_struct *p)
691 int prio = p->static_prio - MAX_RT_PRIO;
692 struct load_weight *load = &p->se.load;
695 * SCHED_IDLE tasks get minimal weight:
697 if (idle_policy(p->policy)) {
698 load->weight = scale_load(WEIGHT_IDLEPRIO);
699 load->inv_weight = WMULT_IDLEPRIO;
703 load->weight = scale_load(sched_prio_to_weight[prio]);
704 load->inv_weight = sched_prio_to_wmult[prio];
707 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
710 if (!(flags & ENQUEUE_RESTORE))
711 sched_info_queued(rq, p);
712 p->sched_class->enqueue_task(rq, p, flags);
715 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
718 if (!(flags & DEQUEUE_SAVE))
719 sched_info_dequeued(rq, p);
720 p->sched_class->dequeue_task(rq, p, flags);
723 void activate_task(struct rq *rq, struct task_struct *p, int flags)
725 if (task_contributes_to_load(p))
726 rq->nr_uninterruptible--;
728 enqueue_task(rq, p, flags);
731 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
733 if (task_contributes_to_load(p))
734 rq->nr_uninterruptible++;
736 dequeue_task(rq, p, flags);
739 static void update_rq_clock_task(struct rq *rq, s64 delta)
742 * In theory, the compile should just see 0 here, and optimize out the call
743 * to sched_rt_avg_update. But I don't trust it...
745 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
746 s64 steal = 0, irq_delta = 0;
748 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
749 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
752 * Since irq_time is only updated on {soft,}irq_exit, we might run into
753 * this case when a previous update_rq_clock() happened inside a
756 * When this happens, we stop ->clock_task and only update the
757 * prev_irq_time stamp to account for the part that fit, so that a next
758 * update will consume the rest. This ensures ->clock_task is
761 * It does however cause some slight miss-attribution of {soft,}irq
762 * time, a more accurate solution would be to update the irq_time using
763 * the current rq->clock timestamp, except that would require using
766 if (irq_delta > delta)
769 rq->prev_irq_time += irq_delta;
772 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
773 if (static_key_false((¶virt_steal_rq_enabled))) {
774 steal = paravirt_steal_clock(cpu_of(rq));
775 steal -= rq->prev_steal_time_rq;
777 if (unlikely(steal > delta))
780 rq->prev_steal_time_rq += steal;
785 rq->clock_task += delta;
787 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
788 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
789 sched_rt_avg_update(rq, irq_delta + steal);
793 void sched_set_stop_task(int cpu, struct task_struct *stop)
795 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
796 struct task_struct *old_stop = cpu_rq(cpu)->stop;
800 * Make it appear like a SCHED_FIFO task, its something
801 * userspace knows about and won't get confused about.
803 * Also, it will make PI more or less work without too
804 * much confusion -- but then, stop work should not
805 * rely on PI working anyway.
807 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
809 stop->sched_class = &stop_sched_class;
812 cpu_rq(cpu)->stop = stop;
816 * Reset it back to a normal scheduling class so that
817 * it can die in pieces.
819 old_stop->sched_class = &rt_sched_class;
824 * __normal_prio - return the priority that is based on the static prio
826 static inline int __normal_prio(struct task_struct *p)
828 return p->static_prio;
832 * Calculate the expected normal priority: i.e. priority
833 * without taking RT-inheritance into account. Might be
834 * boosted by interactivity modifiers. Changes upon fork,
835 * setprio syscalls, and whenever the interactivity
836 * estimator recalculates.
838 static inline int normal_prio(struct task_struct *p)
842 if (task_has_dl_policy(p))
843 prio = MAX_DL_PRIO-1;
844 else if (task_has_rt_policy(p))
845 prio = MAX_RT_PRIO-1 - p->rt_priority;
847 prio = __normal_prio(p);
852 * Calculate the current priority, i.e. the priority
853 * taken into account by the scheduler. This value might
854 * be boosted by RT tasks, or might be boosted by
855 * interactivity modifiers. Will be RT if the task got
856 * RT-boosted. If not then it returns p->normal_prio.
858 static int effective_prio(struct task_struct *p)
860 p->normal_prio = normal_prio(p);
862 * If we are RT tasks or we were boosted to RT priority,
863 * keep the priority unchanged. Otherwise, update priority
864 * to the normal priority:
866 if (!rt_prio(p->prio))
867 return p->normal_prio;
872 * task_curr - is this task currently executing on a CPU?
873 * @p: the task in question.
875 * Return: 1 if the task is currently executing. 0 otherwise.
877 inline int task_curr(const struct task_struct *p)
879 return cpu_curr(task_cpu(p)) == p;
883 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
884 * use the balance_callback list if you want balancing.
886 * this means any call to check_class_changed() must be followed by a call to
887 * balance_callback().
889 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
890 const struct sched_class *prev_class,
893 if (prev_class != p->sched_class) {
894 if (prev_class->switched_from)
895 prev_class->switched_from(rq, p);
897 p->sched_class->switched_to(rq, p);
898 } else if (oldprio != p->prio || dl_task(p))
899 p->sched_class->prio_changed(rq, p, oldprio);
902 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
904 const struct sched_class *class;
906 if (p->sched_class == rq->curr->sched_class) {
907 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
909 for_each_class(class) {
910 if (class == rq->curr->sched_class)
912 if (class == p->sched_class) {
920 * A queue event has occurred, and we're going to schedule. In
921 * this case, we can save a useless back to back clock update.
923 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
924 rq_clock_skip_update(rq, true);
929 * This is how migration works:
931 * 1) we invoke migration_cpu_stop() on the target CPU using
933 * 2) stopper starts to run (implicitly forcing the migrated thread
935 * 3) it checks whether the migrated task is still in the wrong runqueue.
936 * 4) if it's in the wrong runqueue then the migration thread removes
937 * it and puts it into the right queue.
938 * 5) stopper completes and stop_one_cpu() returns and the migration
943 * move_queued_task - move a queued task to new rq.
945 * Returns (locked) new rq. Old rq's lock is released.
947 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
949 lockdep_assert_held(&rq->lock);
951 p->on_rq = TASK_ON_RQ_MIGRATING;
952 dequeue_task(rq, p, 0);
953 set_task_cpu(p, new_cpu);
954 raw_spin_unlock(&rq->lock);
956 rq = cpu_rq(new_cpu);
958 raw_spin_lock(&rq->lock);
959 BUG_ON(task_cpu(p) != new_cpu);
960 enqueue_task(rq, p, 0);
961 p->on_rq = TASK_ON_RQ_QUEUED;
962 check_preempt_curr(rq, p, 0);
967 struct migration_arg {
968 struct task_struct *task;
973 * Move (not current) task off this cpu, onto dest cpu. We're doing
974 * this because either it can't run here any more (set_cpus_allowed()
975 * away from this CPU, or CPU going down), or because we're
976 * attempting to rebalance this task on exec (sched_exec).
978 * So we race with normal scheduler movements, but that's OK, as long
979 * as the task is no longer on this CPU.
981 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
983 if (unlikely(!cpu_active(dest_cpu)))
986 /* Affinity changed (again). */
987 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
990 rq = move_queued_task(rq, p, dest_cpu);
996 * migration_cpu_stop - this will be executed by a highprio stopper thread
997 * and performs thread migration by bumping thread off CPU then
998 * 'pushing' onto another runqueue.
1000 static int migration_cpu_stop(void *data)
1002 struct migration_arg *arg = data;
1003 struct task_struct *p = arg->task;
1004 struct rq *rq = this_rq();
1007 * The original target cpu might have gone down and we might
1008 * be on another cpu but it doesn't matter.
1010 local_irq_disable();
1012 * We need to explicitly wake pending tasks before running
1013 * __migrate_task() such that we will not miss enforcing cpus_allowed
1014 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1016 sched_ttwu_pending();
1018 raw_spin_lock(&p->pi_lock);
1019 raw_spin_lock(&rq->lock);
1021 * If task_rq(p) != rq, it cannot be migrated here, because we're
1022 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1023 * we're holding p->pi_lock.
1025 if (task_rq(p) == rq && task_on_rq_queued(p))
1026 rq = __migrate_task(rq, p, arg->dest_cpu);
1027 raw_spin_unlock(&rq->lock);
1028 raw_spin_unlock(&p->pi_lock);
1035 * sched_class::set_cpus_allowed must do the below, but is not required to
1036 * actually call this function.
1038 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1040 cpumask_copy(&p->cpus_allowed, new_mask);
1041 p->nr_cpus_allowed = cpumask_weight(new_mask);
1044 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1046 struct rq *rq = task_rq(p);
1047 bool queued, running;
1049 lockdep_assert_held(&p->pi_lock);
1051 queued = task_on_rq_queued(p);
1052 running = task_current(rq, p);
1056 * Because __kthread_bind() calls this on blocked tasks without
1059 lockdep_assert_held(&rq->lock);
1060 dequeue_task(rq, p, DEQUEUE_SAVE);
1063 put_prev_task(rq, p);
1065 p->sched_class->set_cpus_allowed(p, new_mask);
1068 p->sched_class->set_curr_task(rq);
1070 enqueue_task(rq, p, ENQUEUE_RESTORE);
1074 * Change a given task's CPU affinity. Migrate the thread to a
1075 * proper CPU and schedule it away if the CPU it's executing on
1076 * is removed from the allowed bitmask.
1078 * NOTE: the caller must have a valid reference to the task, the
1079 * task must not exit() & deallocate itself prematurely. The
1080 * call is not atomic; no spinlocks may be held.
1082 static int __set_cpus_allowed_ptr(struct task_struct *p,
1083 const struct cpumask *new_mask, bool check)
1085 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1086 unsigned int dest_cpu;
1087 unsigned long flags;
1091 rq = task_rq_lock(p, &flags);
1093 if (p->flags & PF_KTHREAD) {
1095 * Kernel threads are allowed on online && !active CPUs
1097 cpu_valid_mask = cpu_online_mask;
1101 * Must re-check here, to close a race against __kthread_bind(),
1102 * sched_setaffinity() is not guaranteed to observe the flag.
1104 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1109 if (cpumask_equal(&p->cpus_allowed, new_mask))
1112 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1117 do_set_cpus_allowed(p, new_mask);
1119 if (p->flags & PF_KTHREAD) {
1121 * For kernel threads that do indeed end up on online &&
1122 * !active we want to ensure they are strict per-cpu threads.
1124 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1125 !cpumask_intersects(new_mask, cpu_active_mask) &&
1126 p->nr_cpus_allowed != 1);
1129 /* Can the task run on the task's current CPU? If so, we're done */
1130 if (cpumask_test_cpu(task_cpu(p), new_mask))
1133 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1134 if (task_running(rq, p) || p->state == TASK_WAKING) {
1135 struct migration_arg arg = { p, dest_cpu };
1136 /* Need help from migration thread: drop lock and wait. */
1137 task_rq_unlock(rq, p, &flags);
1138 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1139 tlb_migrate_finish(p->mm);
1141 } else if (task_on_rq_queued(p)) {
1143 * OK, since we're going to drop the lock immediately
1144 * afterwards anyway.
1146 lockdep_unpin_lock(&rq->lock);
1147 rq = move_queued_task(rq, p, dest_cpu);
1148 lockdep_pin_lock(&rq->lock);
1151 task_rq_unlock(rq, p, &flags);
1156 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1158 return __set_cpus_allowed_ptr(p, new_mask, false);
1160 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1162 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1164 #ifdef CONFIG_SCHED_DEBUG
1166 * We should never call set_task_cpu() on a blocked task,
1167 * ttwu() will sort out the placement.
1169 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1173 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1174 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1175 * time relying on p->on_rq.
1177 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1178 p->sched_class == &fair_sched_class &&
1179 (p->on_rq && !task_on_rq_migrating(p)));
1181 #ifdef CONFIG_LOCKDEP
1183 * The caller should hold either p->pi_lock or rq->lock, when changing
1184 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1186 * sched_move_task() holds both and thus holding either pins the cgroup,
1189 * Furthermore, all task_rq users should acquire both locks, see
1192 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1193 lockdep_is_held(&task_rq(p)->lock)));
1197 trace_sched_migrate_task(p, new_cpu);
1199 if (task_cpu(p) != new_cpu) {
1200 if (p->sched_class->migrate_task_rq)
1201 p->sched_class->migrate_task_rq(p);
1202 p->se.nr_migrations++;
1203 perf_event_task_migrate(p);
1206 __set_task_cpu(p, new_cpu);
1209 static void __migrate_swap_task(struct task_struct *p, int cpu)
1211 if (task_on_rq_queued(p)) {
1212 struct rq *src_rq, *dst_rq;
1214 src_rq = task_rq(p);
1215 dst_rq = cpu_rq(cpu);
1217 p->on_rq = TASK_ON_RQ_MIGRATING;
1218 deactivate_task(src_rq, p, 0);
1219 set_task_cpu(p, cpu);
1220 activate_task(dst_rq, p, 0);
1221 p->on_rq = TASK_ON_RQ_QUEUED;
1222 check_preempt_curr(dst_rq, p, 0);
1225 * Task isn't running anymore; make it appear like we migrated
1226 * it before it went to sleep. This means on wakeup we make the
1227 * previous cpu our targer instead of where it really is.
1233 struct migration_swap_arg {
1234 struct task_struct *src_task, *dst_task;
1235 int src_cpu, dst_cpu;
1238 static int migrate_swap_stop(void *data)
1240 struct migration_swap_arg *arg = data;
1241 struct rq *src_rq, *dst_rq;
1244 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1247 src_rq = cpu_rq(arg->src_cpu);
1248 dst_rq = cpu_rq(arg->dst_cpu);
1250 double_raw_lock(&arg->src_task->pi_lock,
1251 &arg->dst_task->pi_lock);
1252 double_rq_lock(src_rq, dst_rq);
1254 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1257 if (task_cpu(arg->src_task) != arg->src_cpu)
1260 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1263 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1266 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1267 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1272 double_rq_unlock(src_rq, dst_rq);
1273 raw_spin_unlock(&arg->dst_task->pi_lock);
1274 raw_spin_unlock(&arg->src_task->pi_lock);
1280 * Cross migrate two tasks
1282 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1284 struct migration_swap_arg arg;
1287 arg = (struct migration_swap_arg){
1289 .src_cpu = task_cpu(cur),
1291 .dst_cpu = task_cpu(p),
1294 if (arg.src_cpu == arg.dst_cpu)
1298 * These three tests are all lockless; this is OK since all of them
1299 * will be re-checked with proper locks held further down the line.
1301 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1304 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1307 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1310 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1311 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1318 * wait_task_inactive - wait for a thread to unschedule.
1320 * If @match_state is nonzero, it's the @p->state value just checked and
1321 * not expected to change. If it changes, i.e. @p might have woken up,
1322 * then return zero. When we succeed in waiting for @p to be off its CPU,
1323 * we return a positive number (its total switch count). If a second call
1324 * a short while later returns the same number, the caller can be sure that
1325 * @p has remained unscheduled the whole time.
1327 * The caller must ensure that the task *will* unschedule sometime soon,
1328 * else this function might spin for a *long* time. This function can't
1329 * be called with interrupts off, or it may introduce deadlock with
1330 * smp_call_function() if an IPI is sent by the same process we are
1331 * waiting to become inactive.
1333 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1335 unsigned long flags;
1336 int running, queued;
1342 * We do the initial early heuristics without holding
1343 * any task-queue locks at all. We'll only try to get
1344 * the runqueue lock when things look like they will
1350 * If the task is actively running on another CPU
1351 * still, just relax and busy-wait without holding
1354 * NOTE! Since we don't hold any locks, it's not
1355 * even sure that "rq" stays as the right runqueue!
1356 * But we don't care, since "task_running()" will
1357 * return false if the runqueue has changed and p
1358 * is actually now running somewhere else!
1360 while (task_running(rq, p)) {
1361 if (match_state && unlikely(p->state != match_state))
1367 * Ok, time to look more closely! We need the rq
1368 * lock now, to be *sure*. If we're wrong, we'll
1369 * just go back and repeat.
1371 rq = task_rq_lock(p, &flags);
1372 trace_sched_wait_task(p);
1373 running = task_running(rq, p);
1374 queued = task_on_rq_queued(p);
1376 if (!match_state || p->state == match_state)
1377 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1378 task_rq_unlock(rq, p, &flags);
1381 * If it changed from the expected state, bail out now.
1383 if (unlikely(!ncsw))
1387 * Was it really running after all now that we
1388 * checked with the proper locks actually held?
1390 * Oops. Go back and try again..
1392 if (unlikely(running)) {
1398 * It's not enough that it's not actively running,
1399 * it must be off the runqueue _entirely_, and not
1402 * So if it was still runnable (but just not actively
1403 * running right now), it's preempted, and we should
1404 * yield - it could be a while.
1406 if (unlikely(queued)) {
1407 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1409 set_current_state(TASK_UNINTERRUPTIBLE);
1410 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1415 * Ahh, all good. It wasn't running, and it wasn't
1416 * runnable, which means that it will never become
1417 * running in the future either. We're all done!
1426 * kick_process - kick a running thread to enter/exit the kernel
1427 * @p: the to-be-kicked thread
1429 * Cause a process which is running on another CPU to enter
1430 * kernel-mode, without any delay. (to get signals handled.)
1432 * NOTE: this function doesn't have to take the runqueue lock,
1433 * because all it wants to ensure is that the remote task enters
1434 * the kernel. If the IPI races and the task has been migrated
1435 * to another CPU then no harm is done and the purpose has been
1438 void kick_process(struct task_struct *p)
1444 if ((cpu != smp_processor_id()) && task_curr(p))
1445 smp_send_reschedule(cpu);
1448 EXPORT_SYMBOL_GPL(kick_process);
1451 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1453 * A few notes on cpu_active vs cpu_online:
1455 * - cpu_active must be a subset of cpu_online
1457 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1458 * see __set_cpus_allowed_ptr(). At this point the newly online
1459 * cpu isn't yet part of the sched domains, and balancing will not
1462 * - on cpu-down we clear cpu_active() to mask the sched domains and
1463 * avoid the load balancer to place new tasks on the to be removed
1464 * cpu. Existing tasks will remain running there and will be taken
1467 * This means that fallback selection must not select !active CPUs.
1468 * And can assume that any active CPU must be online. Conversely
1469 * select_task_rq() below may allow selection of !active CPUs in order
1470 * to satisfy the above rules.
1472 static int select_fallback_rq(int cpu, struct task_struct *p)
1474 int nid = cpu_to_node(cpu);
1475 const struct cpumask *nodemask = NULL;
1476 enum { cpuset, possible, fail } state = cpuset;
1480 * If the node that the cpu is on has been offlined, cpu_to_node()
1481 * will return -1. There is no cpu on the node, and we should
1482 * select the cpu on the other node.
1485 nodemask = cpumask_of_node(nid);
1487 /* Look for allowed, online CPU in same node. */
1488 for_each_cpu(dest_cpu, nodemask) {
1489 if (!cpu_active(dest_cpu))
1491 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1497 /* Any allowed, online CPU? */
1498 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1499 if (!cpu_active(dest_cpu))
1504 /* No more Mr. Nice Guy. */
1507 if (IS_ENABLED(CONFIG_CPUSETS)) {
1508 cpuset_cpus_allowed_fallback(p);
1514 do_set_cpus_allowed(p, cpu_possible_mask);
1525 if (state != cpuset) {
1527 * Don't tell them about moving exiting tasks or
1528 * kernel threads (both mm NULL), since they never
1531 if (p->mm && printk_ratelimit()) {
1532 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1533 task_pid_nr(p), p->comm, cpu);
1541 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1544 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1546 lockdep_assert_held(&p->pi_lock);
1548 if (p->nr_cpus_allowed > 1)
1549 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1551 cpu = cpumask_any(tsk_cpus_allowed(p));
1554 * In order not to call set_task_cpu() on a blocking task we need
1555 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1558 * Since this is common to all placement strategies, this lives here.
1560 * [ this allows ->select_task() to simply return task_cpu(p) and
1561 * not worry about this generic constraint ]
1563 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1565 cpu = select_fallback_rq(task_cpu(p), p);
1570 static void update_avg(u64 *avg, u64 sample)
1572 s64 diff = sample - *avg;
1578 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1579 const struct cpumask *new_mask, bool check)
1581 return set_cpus_allowed_ptr(p, new_mask);
1584 #endif /* CONFIG_SMP */
1587 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1589 #ifdef CONFIG_SCHEDSTATS
1590 struct rq *rq = this_rq();
1593 int this_cpu = smp_processor_id();
1595 if (cpu == this_cpu) {
1596 schedstat_inc(rq, ttwu_local);
1597 schedstat_inc(p, se.statistics.nr_wakeups_local);
1599 struct sched_domain *sd;
1601 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1603 for_each_domain(this_cpu, sd) {
1604 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1605 schedstat_inc(sd, ttwu_wake_remote);
1612 if (wake_flags & WF_MIGRATED)
1613 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1615 #endif /* CONFIG_SMP */
1617 schedstat_inc(rq, ttwu_count);
1618 schedstat_inc(p, se.statistics.nr_wakeups);
1620 if (wake_flags & WF_SYNC)
1621 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1623 #endif /* CONFIG_SCHEDSTATS */
1626 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1628 activate_task(rq, p, en_flags);
1629 p->on_rq = TASK_ON_RQ_QUEUED;
1631 /* if a worker is waking up, notify workqueue */
1632 if (p->flags & PF_WQ_WORKER)
1633 wq_worker_waking_up(p, cpu_of(rq));
1637 * Mark the task runnable and perform wakeup-preemption.
1640 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1642 check_preempt_curr(rq, p, wake_flags);
1643 p->state = TASK_RUNNING;
1644 trace_sched_wakeup(p);
1647 if (p->sched_class->task_woken) {
1649 * Our task @p is fully woken up and running; so its safe to
1650 * drop the rq->lock, hereafter rq is only used for statistics.
1652 lockdep_unpin_lock(&rq->lock);
1653 p->sched_class->task_woken(rq, p);
1654 lockdep_pin_lock(&rq->lock);
1657 if (rq->idle_stamp) {
1658 u64 delta = rq_clock(rq) - rq->idle_stamp;
1659 u64 max = 2*rq->max_idle_balance_cost;
1661 update_avg(&rq->avg_idle, delta);
1663 if (rq->avg_idle > max)
1672 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1674 lockdep_assert_held(&rq->lock);
1677 if (p->sched_contributes_to_load)
1678 rq->nr_uninterruptible--;
1681 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1682 ttwu_do_wakeup(rq, p, wake_flags);
1686 * Called in case the task @p isn't fully descheduled from its runqueue,
1687 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1688 * since all we need to do is flip p->state to TASK_RUNNING, since
1689 * the task is still ->on_rq.
1691 static int ttwu_remote(struct task_struct *p, int wake_flags)
1696 rq = __task_rq_lock(p);
1697 if (task_on_rq_queued(p)) {
1698 /* check_preempt_curr() may use rq clock */
1699 update_rq_clock(rq);
1700 ttwu_do_wakeup(rq, p, wake_flags);
1703 __task_rq_unlock(rq);
1709 void sched_ttwu_pending(void)
1711 struct rq *rq = this_rq();
1712 struct llist_node *llist = llist_del_all(&rq->wake_list);
1713 struct task_struct *p;
1714 unsigned long flags;
1719 raw_spin_lock_irqsave(&rq->lock, flags);
1720 lockdep_pin_lock(&rq->lock);
1723 p = llist_entry(llist, struct task_struct, wake_entry);
1724 llist = llist_next(llist);
1725 ttwu_do_activate(rq, p, 0);
1728 lockdep_unpin_lock(&rq->lock);
1729 raw_spin_unlock_irqrestore(&rq->lock, flags);
1732 void scheduler_ipi(void)
1735 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1736 * TIF_NEED_RESCHED remotely (for the first time) will also send
1739 preempt_fold_need_resched();
1741 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1745 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1746 * traditionally all their work was done from the interrupt return
1747 * path. Now that we actually do some work, we need to make sure
1750 * Some archs already do call them, luckily irq_enter/exit nest
1753 * Arguably we should visit all archs and update all handlers,
1754 * however a fair share of IPIs are still resched only so this would
1755 * somewhat pessimize the simple resched case.
1758 sched_ttwu_pending();
1761 * Check if someone kicked us for doing the nohz idle load balance.
1763 if (unlikely(got_nohz_idle_kick())) {
1764 this_rq()->idle_balance = 1;
1765 raise_softirq_irqoff(SCHED_SOFTIRQ);
1770 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1772 struct rq *rq = cpu_rq(cpu);
1774 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1775 if (!set_nr_if_polling(rq->idle))
1776 smp_send_reschedule(cpu);
1778 trace_sched_wake_idle_without_ipi(cpu);
1782 void wake_up_if_idle(int cpu)
1784 struct rq *rq = cpu_rq(cpu);
1785 unsigned long flags;
1789 if (!is_idle_task(rcu_dereference(rq->curr)))
1792 if (set_nr_if_polling(rq->idle)) {
1793 trace_sched_wake_idle_without_ipi(cpu);
1795 raw_spin_lock_irqsave(&rq->lock, flags);
1796 if (is_idle_task(rq->curr))
1797 smp_send_reschedule(cpu);
1798 /* Else cpu is not in idle, do nothing here */
1799 raw_spin_unlock_irqrestore(&rq->lock, flags);
1806 bool cpus_share_cache(int this_cpu, int that_cpu)
1808 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1810 #endif /* CONFIG_SMP */
1812 static void ttwu_queue(struct task_struct *p, int cpu)
1814 struct rq *rq = cpu_rq(cpu);
1816 #if defined(CONFIG_SMP)
1817 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1818 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1819 ttwu_queue_remote(p, cpu);
1824 raw_spin_lock(&rq->lock);
1825 lockdep_pin_lock(&rq->lock);
1826 ttwu_do_activate(rq, p, 0);
1827 lockdep_unpin_lock(&rq->lock);
1828 raw_spin_unlock(&rq->lock);
1832 * Notes on Program-Order guarantees on SMP systems.
1836 * The basic program-order guarantee on SMP systems is that when a task [t]
1837 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1838 * execution on its new cpu [c1].
1840 * For migration (of runnable tasks) this is provided by the following means:
1842 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1843 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1844 * rq(c1)->lock (if not at the same time, then in that order).
1845 * C) LOCK of the rq(c1)->lock scheduling in task
1847 * Transitivity guarantees that B happens after A and C after B.
1848 * Note: we only require RCpc transitivity.
1849 * Note: the cpu doing B need not be c0 or c1
1858 * UNLOCK rq(0)->lock
1860 * LOCK rq(0)->lock // orders against CPU0
1862 * UNLOCK rq(0)->lock
1866 * UNLOCK rq(1)->lock
1868 * LOCK rq(1)->lock // orders against CPU2
1871 * UNLOCK rq(1)->lock
1874 * BLOCKING -- aka. SLEEP + WAKEUP
1876 * For blocking we (obviously) need to provide the same guarantee as for
1877 * migration. However the means are completely different as there is no lock
1878 * chain to provide order. Instead we do:
1880 * 1) smp_store_release(X->on_cpu, 0)
1881 * 2) smp_cond_acquire(!X->on_cpu)
1885 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1887 * LOCK rq(0)->lock LOCK X->pi_lock
1890 * smp_store_release(X->on_cpu, 0);
1892 * smp_cond_acquire(!X->on_cpu);
1898 * X->state = RUNNING
1899 * UNLOCK rq(2)->lock
1901 * LOCK rq(2)->lock // orders against CPU1
1904 * UNLOCK rq(2)->lock
1907 * UNLOCK rq(0)->lock
1910 * However; for wakeups there is a second guarantee we must provide, namely we
1911 * must observe the state that lead to our wakeup. That is, not only must our
1912 * task observe its own prior state, it must also observe the stores prior to
1915 * This means that any means of doing remote wakeups must order the CPU doing
1916 * the wakeup against the CPU the task is going to end up running on. This,
1917 * however, is already required for the regular Program-Order guarantee above,
1918 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1923 * try_to_wake_up - wake up a thread
1924 * @p: the thread to be awakened
1925 * @state: the mask of task states that can be woken
1926 * @wake_flags: wake modifier flags (WF_*)
1928 * Put it on the run-queue if it's not already there. The "current"
1929 * thread is always on the run-queue (except when the actual
1930 * re-schedule is in progress), and as such you're allowed to do
1931 * the simpler "current->state = TASK_RUNNING" to mark yourself
1932 * runnable without the overhead of this.
1934 * Return: %true if @p was woken up, %false if it was already running.
1935 * or @state didn't match @p's state.
1938 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1940 unsigned long flags;
1941 int cpu, success = 0;
1944 * If we are going to wake up a thread waiting for CONDITION we
1945 * need to ensure that CONDITION=1 done by the caller can not be
1946 * reordered with p->state check below. This pairs with mb() in
1947 * set_current_state() the waiting thread does.
1949 smp_mb__before_spinlock();
1950 raw_spin_lock_irqsave(&p->pi_lock, flags);
1951 if (!(p->state & state))
1954 trace_sched_waking(p);
1956 success = 1; /* we're going to change ->state */
1959 if (p->on_rq && ttwu_remote(p, wake_flags))
1964 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1965 * possible to, falsely, observe p->on_cpu == 0.
1967 * One must be running (->on_cpu == 1) in order to remove oneself
1968 * from the runqueue.
1970 * [S] ->on_cpu = 1; [L] ->on_rq
1974 * [S] ->on_rq = 0; [L] ->on_cpu
1976 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1977 * from the consecutive calls to schedule(); the first switching to our
1978 * task, the second putting it to sleep.
1983 * If the owning (remote) cpu is still in the middle of schedule() with
1984 * this task as prev, wait until its done referencing the task.
1986 * Pairs with the smp_store_release() in finish_lock_switch().
1988 * This ensures that tasks getting woken will be fully ordered against
1989 * their previous state and preserve Program Order.
1991 smp_cond_acquire(!p->on_cpu);
1993 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1994 p->state = TASK_WAKING;
1996 if (p->sched_class->task_waking)
1997 p->sched_class->task_waking(p);
1999 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2000 if (task_cpu(p) != cpu) {
2001 wake_flags |= WF_MIGRATED;
2002 set_task_cpu(p, cpu);
2004 #endif /* CONFIG_SMP */
2008 if (schedstat_enabled())
2009 ttwu_stat(p, cpu, wake_flags);
2011 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2017 * try_to_wake_up_local - try to wake up a local task with rq lock held
2018 * @p: the thread to be awakened
2020 * Put @p on the run-queue if it's not already there. The caller must
2021 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2024 static void try_to_wake_up_local(struct task_struct *p)
2026 struct rq *rq = task_rq(p);
2028 if (WARN_ON_ONCE(rq != this_rq()) ||
2029 WARN_ON_ONCE(p == current))
2032 lockdep_assert_held(&rq->lock);
2034 if (!raw_spin_trylock(&p->pi_lock)) {
2036 * This is OK, because current is on_cpu, which avoids it being
2037 * picked for load-balance and preemption/IRQs are still
2038 * disabled avoiding further scheduler activity on it and we've
2039 * not yet picked a replacement task.
2041 lockdep_unpin_lock(&rq->lock);
2042 raw_spin_unlock(&rq->lock);
2043 raw_spin_lock(&p->pi_lock);
2044 raw_spin_lock(&rq->lock);
2045 lockdep_pin_lock(&rq->lock);
2048 if (!(p->state & TASK_NORMAL))
2051 trace_sched_waking(p);
2053 if (!task_on_rq_queued(p))
2054 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2056 ttwu_do_wakeup(rq, p, 0);
2057 if (schedstat_enabled())
2058 ttwu_stat(p, smp_processor_id(), 0);
2060 raw_spin_unlock(&p->pi_lock);
2064 * wake_up_process - Wake up a specific process
2065 * @p: The process to be woken up.
2067 * Attempt to wake up the nominated process and move it to the set of runnable
2070 * Return: 1 if the process was woken up, 0 if it was already running.
2072 * It may be assumed that this function implies a write memory barrier before
2073 * changing the task state if and only if any tasks are woken up.
2075 int wake_up_process(struct task_struct *p)
2077 return try_to_wake_up(p, TASK_NORMAL, 0);
2079 EXPORT_SYMBOL(wake_up_process);
2081 int wake_up_state(struct task_struct *p, unsigned int state)
2083 return try_to_wake_up(p, state, 0);
2087 * This function clears the sched_dl_entity static params.
2089 void __dl_clear_params(struct task_struct *p)
2091 struct sched_dl_entity *dl_se = &p->dl;
2093 dl_se->dl_runtime = 0;
2094 dl_se->dl_deadline = 0;
2095 dl_se->dl_period = 0;
2099 dl_se->dl_throttled = 0;
2100 dl_se->dl_yielded = 0;
2104 * Perform scheduler related setup for a newly forked process p.
2105 * p is forked by current.
2107 * __sched_fork() is basic setup used by init_idle() too:
2109 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2114 p->se.exec_start = 0;
2115 p->se.sum_exec_runtime = 0;
2116 p->se.prev_sum_exec_runtime = 0;
2117 p->se.nr_migrations = 0;
2119 INIT_LIST_HEAD(&p->se.group_node);
2121 #ifdef CONFIG_FAIR_GROUP_SCHED
2122 p->se.cfs_rq = NULL;
2125 #ifdef CONFIG_SCHEDSTATS
2126 /* Even if schedstat is disabled, there should not be garbage */
2127 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2130 RB_CLEAR_NODE(&p->dl.rb_node);
2131 init_dl_task_timer(&p->dl);
2132 __dl_clear_params(p);
2134 INIT_LIST_HEAD(&p->rt.run_list);
2136 p->rt.time_slice = sched_rr_timeslice;
2140 #ifdef CONFIG_PREEMPT_NOTIFIERS
2141 INIT_HLIST_HEAD(&p->preempt_notifiers);
2144 #ifdef CONFIG_NUMA_BALANCING
2145 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2146 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2147 p->mm->numa_scan_seq = 0;
2150 if (clone_flags & CLONE_VM)
2151 p->numa_preferred_nid = current->numa_preferred_nid;
2153 p->numa_preferred_nid = -1;
2155 p->node_stamp = 0ULL;
2156 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2157 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2158 p->numa_work.next = &p->numa_work;
2159 p->numa_faults = NULL;
2160 p->last_task_numa_placement = 0;
2161 p->last_sum_exec_runtime = 0;
2163 p->numa_group = NULL;
2164 #endif /* CONFIG_NUMA_BALANCING */
2167 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2169 #ifdef CONFIG_NUMA_BALANCING
2171 void set_numabalancing_state(bool enabled)
2174 static_branch_enable(&sched_numa_balancing);
2176 static_branch_disable(&sched_numa_balancing);
2179 #ifdef CONFIG_PROC_SYSCTL
2180 int sysctl_numa_balancing(struct ctl_table *table, int write,
2181 void __user *buffer, size_t *lenp, loff_t *ppos)
2185 int state = static_branch_likely(&sched_numa_balancing);
2187 if (write && !capable(CAP_SYS_ADMIN))
2192 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2196 set_numabalancing_state(state);
2202 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2204 #ifdef CONFIG_SCHEDSTATS
2205 static void set_schedstats(bool enabled)
2208 static_branch_enable(&sched_schedstats);
2210 static_branch_disable(&sched_schedstats);
2213 void force_schedstat_enabled(void)
2215 if (!schedstat_enabled()) {
2216 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2217 static_branch_enable(&sched_schedstats);
2221 static int __init setup_schedstats(char *str)
2227 if (!strcmp(str, "enable")) {
2228 set_schedstats(true);
2230 } else if (!strcmp(str, "disable")) {
2231 set_schedstats(false);
2236 pr_warn("Unable to parse schedstats=\n");
2240 __setup("schedstats=", setup_schedstats);
2242 #ifdef CONFIG_PROC_SYSCTL
2243 int sysctl_schedstats(struct ctl_table *table, int write,
2244 void __user *buffer, size_t *lenp, loff_t *ppos)
2248 int state = static_branch_likely(&sched_schedstats);
2250 if (write && !capable(CAP_SYS_ADMIN))
2255 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2259 set_schedstats(state);
2266 * fork()/clone()-time setup:
2268 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2270 unsigned long flags;
2271 int cpu = get_cpu();
2273 __sched_fork(clone_flags, p);
2275 * We mark the process as running here. This guarantees that
2276 * nobody will actually run it, and a signal or other external
2277 * event cannot wake it up and insert it on the runqueue either.
2279 p->state = TASK_RUNNING;
2282 * Make sure we do not leak PI boosting priority to the child.
2284 p->prio = current->normal_prio;
2287 * Revert to default priority/policy on fork if requested.
2289 if (unlikely(p->sched_reset_on_fork)) {
2290 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2291 p->policy = SCHED_NORMAL;
2292 p->static_prio = NICE_TO_PRIO(0);
2294 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2295 p->static_prio = NICE_TO_PRIO(0);
2297 p->prio = p->normal_prio = __normal_prio(p);
2301 * We don't need the reset flag anymore after the fork. It has
2302 * fulfilled its duty:
2304 p->sched_reset_on_fork = 0;
2307 if (dl_prio(p->prio)) {
2310 } else if (rt_prio(p->prio)) {
2311 p->sched_class = &rt_sched_class;
2313 p->sched_class = &fair_sched_class;
2316 if (p->sched_class->task_fork)
2317 p->sched_class->task_fork(p);
2320 * The child is not yet in the pid-hash so no cgroup attach races,
2321 * and the cgroup is pinned to this child due to cgroup_fork()
2322 * is ran before sched_fork().
2324 * Silence PROVE_RCU.
2326 raw_spin_lock_irqsave(&p->pi_lock, flags);
2327 set_task_cpu(p, cpu);
2328 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2330 #ifdef CONFIG_SCHED_INFO
2331 if (likely(sched_info_on()))
2332 memset(&p->sched_info, 0, sizeof(p->sched_info));
2334 #if defined(CONFIG_SMP)
2337 init_task_preempt_count(p);
2339 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2340 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2347 unsigned long to_ratio(u64 period, u64 runtime)
2349 if (runtime == RUNTIME_INF)
2353 * Doing this here saves a lot of checks in all
2354 * the calling paths, and returning zero seems
2355 * safe for them anyway.
2360 return div64_u64(runtime << 20, period);
2364 inline struct dl_bw *dl_bw_of(int i)
2366 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2367 "sched RCU must be held");
2368 return &cpu_rq(i)->rd->dl_bw;
2371 static inline int dl_bw_cpus(int i)
2373 struct root_domain *rd = cpu_rq(i)->rd;
2376 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2377 "sched RCU must be held");
2378 for_each_cpu_and(i, rd->span, cpu_active_mask)
2384 inline struct dl_bw *dl_bw_of(int i)
2386 return &cpu_rq(i)->dl.dl_bw;
2389 static inline int dl_bw_cpus(int i)
2396 * We must be sure that accepting a new task (or allowing changing the
2397 * parameters of an existing one) is consistent with the bandwidth
2398 * constraints. If yes, this function also accordingly updates the currently
2399 * allocated bandwidth to reflect the new situation.
2401 * This function is called while holding p's rq->lock.
2403 * XXX we should delay bw change until the task's 0-lag point, see
2406 static int dl_overflow(struct task_struct *p, int policy,
2407 const struct sched_attr *attr)
2410 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2411 u64 period = attr->sched_period ?: attr->sched_deadline;
2412 u64 runtime = attr->sched_runtime;
2413 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2416 if (new_bw == p->dl.dl_bw)
2420 * Either if a task, enters, leave, or stays -deadline but changes
2421 * its parameters, we may need to update accordingly the total
2422 * allocated bandwidth of the container.
2424 raw_spin_lock(&dl_b->lock);
2425 cpus = dl_bw_cpus(task_cpu(p));
2426 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2427 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2428 __dl_add(dl_b, new_bw);
2430 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2431 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2432 __dl_clear(dl_b, p->dl.dl_bw);
2433 __dl_add(dl_b, new_bw);
2435 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2436 __dl_clear(dl_b, p->dl.dl_bw);
2439 raw_spin_unlock(&dl_b->lock);
2444 extern void init_dl_bw(struct dl_bw *dl_b);
2447 * wake_up_new_task - wake up a newly created task for the first time.
2449 * This function will do some initial scheduler statistics housekeeping
2450 * that must be done for every newly created context, then puts the task
2451 * on the runqueue and wakes it.
2453 void wake_up_new_task(struct task_struct *p)
2455 unsigned long flags;
2458 raw_spin_lock_irqsave(&p->pi_lock, flags);
2459 /* Initialize new task's runnable average */
2460 init_entity_runnable_average(&p->se);
2463 * Fork balancing, do it here and not earlier because:
2464 * - cpus_allowed can change in the fork path
2465 * - any previously selected cpu might disappear through hotplug
2467 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2470 rq = __task_rq_lock(p);
2471 activate_task(rq, p, 0);
2472 p->on_rq = TASK_ON_RQ_QUEUED;
2473 trace_sched_wakeup_new(p);
2474 check_preempt_curr(rq, p, WF_FORK);
2476 if (p->sched_class->task_woken) {
2478 * Nothing relies on rq->lock after this, so its fine to
2481 lockdep_unpin_lock(&rq->lock);
2482 p->sched_class->task_woken(rq, p);
2483 lockdep_pin_lock(&rq->lock);
2486 task_rq_unlock(rq, p, &flags);
2489 #ifdef CONFIG_PREEMPT_NOTIFIERS
2491 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2493 void preempt_notifier_inc(void)
2495 static_key_slow_inc(&preempt_notifier_key);
2497 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2499 void preempt_notifier_dec(void)
2501 static_key_slow_dec(&preempt_notifier_key);
2503 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2506 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2507 * @notifier: notifier struct to register
2509 void preempt_notifier_register(struct preempt_notifier *notifier)
2511 if (!static_key_false(&preempt_notifier_key))
2512 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2514 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2516 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2519 * preempt_notifier_unregister - no longer interested in preemption notifications
2520 * @notifier: notifier struct to unregister
2522 * This is *not* safe to call from within a preemption notifier.
2524 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2526 hlist_del(¬ifier->link);
2528 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2530 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2532 struct preempt_notifier *notifier;
2534 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2535 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2538 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2540 if (static_key_false(&preempt_notifier_key))
2541 __fire_sched_in_preempt_notifiers(curr);
2545 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2546 struct task_struct *next)
2548 struct preempt_notifier *notifier;
2550 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2551 notifier->ops->sched_out(notifier, next);
2554 static __always_inline void
2555 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2556 struct task_struct *next)
2558 if (static_key_false(&preempt_notifier_key))
2559 __fire_sched_out_preempt_notifiers(curr, next);
2562 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2564 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2569 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2570 struct task_struct *next)
2574 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2577 * prepare_task_switch - prepare to switch tasks
2578 * @rq: the runqueue preparing to switch
2579 * @prev: the current task that is being switched out
2580 * @next: the task we are going to switch to.
2582 * This is called with the rq lock held and interrupts off. It must
2583 * be paired with a subsequent finish_task_switch after the context
2586 * prepare_task_switch sets up locking and calls architecture specific
2590 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2591 struct task_struct *next)
2593 sched_info_switch(rq, prev, next);
2594 perf_event_task_sched_out(prev, next);
2595 fire_sched_out_preempt_notifiers(prev, next);
2596 prepare_lock_switch(rq, next);
2597 prepare_arch_switch(next);
2601 * finish_task_switch - clean up after a task-switch
2602 * @prev: the thread we just switched away from.
2604 * finish_task_switch must be called after the context switch, paired
2605 * with a prepare_task_switch call before the context switch.
2606 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2607 * and do any other architecture-specific cleanup actions.
2609 * Note that we may have delayed dropping an mm in context_switch(). If
2610 * so, we finish that here outside of the runqueue lock. (Doing it
2611 * with the lock held can cause deadlocks; see schedule() for
2614 * The context switch have flipped the stack from under us and restored the
2615 * local variables which were saved when this task called schedule() in the
2616 * past. prev == current is still correct but we need to recalculate this_rq
2617 * because prev may have moved to another CPU.
2619 static struct rq *finish_task_switch(struct task_struct *prev)
2620 __releases(rq->lock)
2622 struct rq *rq = this_rq();
2623 struct mm_struct *mm = rq->prev_mm;
2627 * The previous task will have left us with a preempt_count of 2
2628 * because it left us after:
2631 * preempt_disable(); // 1
2633 * raw_spin_lock_irq(&rq->lock) // 2
2635 * Also, see FORK_PREEMPT_COUNT.
2637 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2638 "corrupted preempt_count: %s/%d/0x%x\n",
2639 current->comm, current->pid, preempt_count()))
2640 preempt_count_set(FORK_PREEMPT_COUNT);
2645 * A task struct has one reference for the use as "current".
2646 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2647 * schedule one last time. The schedule call will never return, and
2648 * the scheduled task must drop that reference.
2650 * We must observe prev->state before clearing prev->on_cpu (in
2651 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2652 * running on another CPU and we could rave with its RUNNING -> DEAD
2653 * transition, resulting in a double drop.
2655 prev_state = prev->state;
2656 vtime_task_switch(prev);
2657 perf_event_task_sched_in(prev, current);
2658 finish_lock_switch(rq, prev);
2659 finish_arch_post_lock_switch();
2661 fire_sched_in_preempt_notifiers(current);
2664 if (unlikely(prev_state == TASK_DEAD)) {
2665 if (prev->sched_class->task_dead)
2666 prev->sched_class->task_dead(prev);
2669 * Remove function-return probe instances associated with this
2670 * task and put them back on the free list.
2672 kprobe_flush_task(prev);
2673 put_task_struct(prev);
2676 tick_nohz_task_switch();
2682 /* rq->lock is NOT held, but preemption is disabled */
2683 static void __balance_callback(struct rq *rq)
2685 struct callback_head *head, *next;
2686 void (*func)(struct rq *rq);
2687 unsigned long flags;
2689 raw_spin_lock_irqsave(&rq->lock, flags);
2690 head = rq->balance_callback;
2691 rq->balance_callback = NULL;
2693 func = (void (*)(struct rq *))head->func;
2700 raw_spin_unlock_irqrestore(&rq->lock, flags);
2703 static inline void balance_callback(struct rq *rq)
2705 if (unlikely(rq->balance_callback))
2706 __balance_callback(rq);
2711 static inline void balance_callback(struct rq *rq)
2718 * schedule_tail - first thing a freshly forked thread must call.
2719 * @prev: the thread we just switched away from.
2721 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2722 __releases(rq->lock)
2727 * New tasks start with FORK_PREEMPT_COUNT, see there and
2728 * finish_task_switch() for details.
2730 * finish_task_switch() will drop rq->lock() and lower preempt_count
2731 * and the preempt_enable() will end up enabling preemption (on
2732 * PREEMPT_COUNT kernels).
2735 rq = finish_task_switch(prev);
2736 balance_callback(rq);
2739 if (current->set_child_tid)
2740 put_user(task_pid_vnr(current), current->set_child_tid);
2744 * context_switch - switch to the new MM and the new thread's register state.
2746 static __always_inline struct rq *
2747 context_switch(struct rq *rq, struct task_struct *prev,
2748 struct task_struct *next)
2750 struct mm_struct *mm, *oldmm;
2752 prepare_task_switch(rq, prev, next);
2755 oldmm = prev->active_mm;
2757 * For paravirt, this is coupled with an exit in switch_to to
2758 * combine the page table reload and the switch backend into
2761 arch_start_context_switch(prev);
2764 next->active_mm = oldmm;
2765 atomic_inc(&oldmm->mm_count);
2766 enter_lazy_tlb(oldmm, next);
2768 switch_mm(oldmm, mm, next);
2771 prev->active_mm = NULL;
2772 rq->prev_mm = oldmm;
2775 * Since the runqueue lock will be released by the next
2776 * task (which is an invalid locking op but in the case
2777 * of the scheduler it's an obvious special-case), so we
2778 * do an early lockdep release here:
2780 lockdep_unpin_lock(&rq->lock);
2781 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2783 /* Here we just switch the register state and the stack. */
2784 switch_to(prev, next, prev);
2787 return finish_task_switch(prev);
2791 * nr_running and nr_context_switches:
2793 * externally visible scheduler statistics: current number of runnable
2794 * threads, total number of context switches performed since bootup.
2796 unsigned long nr_running(void)
2798 unsigned long i, sum = 0;
2800 for_each_online_cpu(i)
2801 sum += cpu_rq(i)->nr_running;
2807 * Check if only the current task is running on the cpu.
2809 * Caution: this function does not check that the caller has disabled
2810 * preemption, thus the result might have a time-of-check-to-time-of-use
2811 * race. The caller is responsible to use it correctly, for example:
2813 * - from a non-preemptable section (of course)
2815 * - from a thread that is bound to a single CPU
2817 * - in a loop with very short iterations (e.g. a polling loop)
2819 bool single_task_running(void)
2821 return raw_rq()->nr_running == 1;
2823 EXPORT_SYMBOL(single_task_running);
2825 unsigned long long nr_context_switches(void)
2828 unsigned long long sum = 0;
2830 for_each_possible_cpu(i)
2831 sum += cpu_rq(i)->nr_switches;
2836 unsigned long nr_iowait(void)
2838 unsigned long i, sum = 0;
2840 for_each_possible_cpu(i)
2841 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2846 unsigned long nr_iowait_cpu(int cpu)
2848 struct rq *this = cpu_rq(cpu);
2849 return atomic_read(&this->nr_iowait);
2852 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2854 struct rq *rq = this_rq();
2855 *nr_waiters = atomic_read(&rq->nr_iowait);
2856 *load = rq->load.weight;
2862 * sched_exec - execve() is a valuable balancing opportunity, because at
2863 * this point the task has the smallest effective memory and cache footprint.
2865 void sched_exec(void)
2867 struct task_struct *p = current;
2868 unsigned long flags;
2871 raw_spin_lock_irqsave(&p->pi_lock, flags);
2872 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2873 if (dest_cpu == smp_processor_id())
2876 if (likely(cpu_active(dest_cpu))) {
2877 struct migration_arg arg = { p, dest_cpu };
2879 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2880 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2884 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2889 DEFINE_PER_CPU(struct kernel_stat, kstat);
2890 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2892 EXPORT_PER_CPU_SYMBOL(kstat);
2893 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2896 * Return accounted runtime for the task.
2897 * In case the task is currently running, return the runtime plus current's
2898 * pending runtime that have not been accounted yet.
2900 unsigned long long task_sched_runtime(struct task_struct *p)
2902 unsigned long flags;
2906 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2908 * 64-bit doesn't need locks to atomically read a 64bit value.
2909 * So we have a optimization chance when the task's delta_exec is 0.
2910 * Reading ->on_cpu is racy, but this is ok.
2912 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2913 * If we race with it entering cpu, unaccounted time is 0. This is
2914 * indistinguishable from the read occurring a few cycles earlier.
2915 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2916 * been accounted, so we're correct here as well.
2918 if (!p->on_cpu || !task_on_rq_queued(p))
2919 return p->se.sum_exec_runtime;
2922 rq = task_rq_lock(p, &flags);
2924 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2925 * project cycles that may never be accounted to this
2926 * thread, breaking clock_gettime().
2928 if (task_current(rq, p) && task_on_rq_queued(p)) {
2929 update_rq_clock(rq);
2930 p->sched_class->update_curr(rq);
2932 ns = p->se.sum_exec_runtime;
2933 task_rq_unlock(rq, p, &flags);
2939 * This function gets called by the timer code, with HZ frequency.
2940 * We call it with interrupts disabled.
2942 void scheduler_tick(void)
2944 int cpu = smp_processor_id();
2945 struct rq *rq = cpu_rq(cpu);
2946 struct task_struct *curr = rq->curr;
2950 raw_spin_lock(&rq->lock);
2951 update_rq_clock(rq);
2952 curr->sched_class->task_tick(rq, curr, 0);
2953 update_cpu_load_active(rq);
2954 calc_global_load_tick(rq);
2955 raw_spin_unlock(&rq->lock);
2957 perf_event_task_tick();
2960 rq->idle_balance = idle_cpu(cpu);
2961 trigger_load_balance(rq);
2963 rq_last_tick_reset(rq);
2966 #ifdef CONFIG_NO_HZ_FULL
2968 * scheduler_tick_max_deferment
2970 * Keep at least one tick per second when a single
2971 * active task is running because the scheduler doesn't
2972 * yet completely support full dynticks environment.
2974 * This makes sure that uptime, CFS vruntime, load
2975 * balancing, etc... continue to move forward, even
2976 * with a very low granularity.
2978 * Return: Maximum deferment in nanoseconds.
2980 u64 scheduler_tick_max_deferment(void)
2982 struct rq *rq = this_rq();
2983 unsigned long next, now = READ_ONCE(jiffies);
2985 next = rq->last_sched_tick + HZ;
2987 if (time_before_eq(next, now))
2990 return jiffies_to_nsecs(next - now);
2994 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2995 defined(CONFIG_PREEMPT_TRACER))
2997 void preempt_count_add(int val)
2999 #ifdef CONFIG_DEBUG_PREEMPT
3003 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3006 __preempt_count_add(val);
3007 #ifdef CONFIG_DEBUG_PREEMPT
3009 * Spinlock count overflowing soon?
3011 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3014 if (preempt_count() == val) {
3015 unsigned long ip = get_lock_parent_ip();
3016 #ifdef CONFIG_DEBUG_PREEMPT
3017 current->preempt_disable_ip = ip;
3019 trace_preempt_off(CALLER_ADDR0, ip);
3022 EXPORT_SYMBOL(preempt_count_add);
3023 NOKPROBE_SYMBOL(preempt_count_add);
3025 void preempt_count_sub(int val)
3027 #ifdef CONFIG_DEBUG_PREEMPT
3031 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3034 * Is the spinlock portion underflowing?
3036 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3037 !(preempt_count() & PREEMPT_MASK)))
3041 if (preempt_count() == val)
3042 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3043 __preempt_count_sub(val);
3045 EXPORT_SYMBOL(preempt_count_sub);
3046 NOKPROBE_SYMBOL(preempt_count_sub);
3051 * Print scheduling while atomic bug:
3053 static noinline void __schedule_bug(struct task_struct *prev)
3055 if (oops_in_progress)
3058 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3059 prev->comm, prev->pid, preempt_count());
3061 debug_show_held_locks(prev);
3063 if (irqs_disabled())
3064 print_irqtrace_events(prev);
3065 #ifdef CONFIG_DEBUG_PREEMPT
3066 if (in_atomic_preempt_off()) {
3067 pr_err("Preemption disabled at:");
3068 print_ip_sym(current->preempt_disable_ip);
3073 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3077 * Various schedule()-time debugging checks and statistics:
3079 static inline void schedule_debug(struct task_struct *prev)
3081 #ifdef CONFIG_SCHED_STACK_END_CHECK
3082 BUG_ON(task_stack_end_corrupted(prev));
3085 if (unlikely(in_atomic_preempt_off())) {
3086 __schedule_bug(prev);
3087 preempt_count_set(PREEMPT_DISABLED);
3091 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3093 schedstat_inc(this_rq(), sched_count);
3097 * Pick up the highest-prio task:
3099 static inline struct task_struct *
3100 pick_next_task(struct rq *rq, struct task_struct *prev)
3102 const struct sched_class *class = &fair_sched_class;
3103 struct task_struct *p;
3106 * Optimization: we know that if all tasks are in
3107 * the fair class we can call that function directly:
3109 if (likely(prev->sched_class == class &&
3110 rq->nr_running == rq->cfs.h_nr_running)) {
3111 p = fair_sched_class.pick_next_task(rq, prev);
3112 if (unlikely(p == RETRY_TASK))
3115 /* assumes fair_sched_class->next == idle_sched_class */
3117 p = idle_sched_class.pick_next_task(rq, prev);
3123 for_each_class(class) {
3124 p = class->pick_next_task(rq, prev);
3126 if (unlikely(p == RETRY_TASK))
3132 BUG(); /* the idle class will always have a runnable task */
3136 * __schedule() is the main scheduler function.
3138 * The main means of driving the scheduler and thus entering this function are:
3140 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3142 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3143 * paths. For example, see arch/x86/entry_64.S.
3145 * To drive preemption between tasks, the scheduler sets the flag in timer
3146 * interrupt handler scheduler_tick().
3148 * 3. Wakeups don't really cause entry into schedule(). They add a
3149 * task to the run-queue and that's it.
3151 * Now, if the new task added to the run-queue preempts the current
3152 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3153 * called on the nearest possible occasion:
3155 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3157 * - in syscall or exception context, at the next outmost
3158 * preempt_enable(). (this might be as soon as the wake_up()'s
3161 * - in IRQ context, return from interrupt-handler to
3162 * preemptible context
3164 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3167 * - cond_resched() call
3168 * - explicit schedule() call
3169 * - return from syscall or exception to user-space
3170 * - return from interrupt-handler to user-space
3172 * WARNING: must be called with preemption disabled!
3174 static void __sched notrace __schedule(bool preempt)
3176 struct task_struct *prev, *next;
3177 unsigned long *switch_count;
3181 cpu = smp_processor_id();
3186 * do_exit() calls schedule() with preemption disabled as an exception;
3187 * however we must fix that up, otherwise the next task will see an
3188 * inconsistent (higher) preempt count.
3190 * It also avoids the below schedule_debug() test from complaining
3193 if (unlikely(prev->state == TASK_DEAD))
3194 preempt_enable_no_resched_notrace();
3196 schedule_debug(prev);
3198 if (sched_feat(HRTICK))
3201 local_irq_disable();
3202 rcu_note_context_switch();
3205 * Make sure that signal_pending_state()->signal_pending() below
3206 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3207 * done by the caller to avoid the race with signal_wake_up().
3209 smp_mb__before_spinlock();
3210 raw_spin_lock(&rq->lock);
3211 lockdep_pin_lock(&rq->lock);
3213 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3215 switch_count = &prev->nivcsw;
3216 if (!preempt && prev->state) {
3217 if (unlikely(signal_pending_state(prev->state, prev))) {
3218 prev->state = TASK_RUNNING;
3220 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3224 * If a worker went to sleep, notify and ask workqueue
3225 * whether it wants to wake up a task to maintain
3228 if (prev->flags & PF_WQ_WORKER) {
3229 struct task_struct *to_wakeup;
3231 to_wakeup = wq_worker_sleeping(prev);
3233 try_to_wake_up_local(to_wakeup);
3236 switch_count = &prev->nvcsw;
3239 if (task_on_rq_queued(prev))
3240 update_rq_clock(rq);
3242 next = pick_next_task(rq, prev);
3243 clear_tsk_need_resched(prev);
3244 clear_preempt_need_resched();
3245 rq->clock_skip_update = 0;
3247 if (likely(prev != next)) {
3252 trace_sched_switch(preempt, prev, next);
3253 rq = context_switch(rq, prev, next); /* unlocks the rq */
3255 lockdep_unpin_lock(&rq->lock);
3256 raw_spin_unlock_irq(&rq->lock);
3259 balance_callback(rq);
3261 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3263 static inline void sched_submit_work(struct task_struct *tsk)
3265 if (!tsk->state || tsk_is_pi_blocked(tsk))
3268 * If we are going to sleep and we have plugged IO queued,
3269 * make sure to submit it to avoid deadlocks.
3271 if (blk_needs_flush_plug(tsk))
3272 blk_schedule_flush_plug(tsk);
3275 asmlinkage __visible void __sched schedule(void)
3277 struct task_struct *tsk = current;
3279 sched_submit_work(tsk);
3283 sched_preempt_enable_no_resched();
3284 } while (need_resched());
3286 EXPORT_SYMBOL(schedule);
3288 #ifdef CONFIG_CONTEXT_TRACKING
3289 asmlinkage __visible void __sched schedule_user(void)
3292 * If we come here after a random call to set_need_resched(),
3293 * or we have been woken up remotely but the IPI has not yet arrived,
3294 * we haven't yet exited the RCU idle mode. Do it here manually until
3295 * we find a better solution.
3297 * NB: There are buggy callers of this function. Ideally we
3298 * should warn if prev_state != CONTEXT_USER, but that will trigger
3299 * too frequently to make sense yet.
3301 enum ctx_state prev_state = exception_enter();
3303 exception_exit(prev_state);
3308 * schedule_preempt_disabled - called with preemption disabled
3310 * Returns with preemption disabled. Note: preempt_count must be 1
3312 void __sched schedule_preempt_disabled(void)
3314 sched_preempt_enable_no_resched();
3319 static void __sched notrace preempt_schedule_common(void)
3322 preempt_disable_notrace();
3324 preempt_enable_no_resched_notrace();
3327 * Check again in case we missed a preemption opportunity
3328 * between schedule and now.
3330 } while (need_resched());
3333 #ifdef CONFIG_PREEMPT
3335 * this is the entry point to schedule() from in-kernel preemption
3336 * off of preempt_enable. Kernel preemptions off return from interrupt
3337 * occur there and call schedule directly.
3339 asmlinkage __visible void __sched notrace preempt_schedule(void)
3342 * If there is a non-zero preempt_count or interrupts are disabled,
3343 * we do not want to preempt the current task. Just return..
3345 if (likely(!preemptible()))
3348 preempt_schedule_common();
3350 NOKPROBE_SYMBOL(preempt_schedule);
3351 EXPORT_SYMBOL(preempt_schedule);
3354 * preempt_schedule_notrace - preempt_schedule called by tracing
3356 * The tracing infrastructure uses preempt_enable_notrace to prevent
3357 * recursion and tracing preempt enabling caused by the tracing
3358 * infrastructure itself. But as tracing can happen in areas coming
3359 * from userspace or just about to enter userspace, a preempt enable
3360 * can occur before user_exit() is called. This will cause the scheduler
3361 * to be called when the system is still in usermode.
3363 * To prevent this, the preempt_enable_notrace will use this function
3364 * instead of preempt_schedule() to exit user context if needed before
3365 * calling the scheduler.
3367 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3369 enum ctx_state prev_ctx;
3371 if (likely(!preemptible()))
3375 preempt_disable_notrace();
3377 * Needs preempt disabled in case user_exit() is traced
3378 * and the tracer calls preempt_enable_notrace() causing
3379 * an infinite recursion.
3381 prev_ctx = exception_enter();
3383 exception_exit(prev_ctx);
3385 preempt_enable_no_resched_notrace();
3386 } while (need_resched());
3388 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3390 #endif /* CONFIG_PREEMPT */
3393 * this is the entry point to schedule() from kernel preemption
3394 * off of irq context.
3395 * Note, that this is called and return with irqs disabled. This will
3396 * protect us against recursive calling from irq.
3398 asmlinkage __visible void __sched preempt_schedule_irq(void)
3400 enum ctx_state prev_state;
3402 /* Catch callers which need to be fixed */
3403 BUG_ON(preempt_count() || !irqs_disabled());
3405 prev_state = exception_enter();
3411 local_irq_disable();
3412 sched_preempt_enable_no_resched();
3413 } while (need_resched());
3415 exception_exit(prev_state);
3418 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3421 return try_to_wake_up(curr->private, mode, wake_flags);
3423 EXPORT_SYMBOL(default_wake_function);
3425 #ifdef CONFIG_RT_MUTEXES
3428 * rt_mutex_setprio - set the current priority of a task
3430 * @prio: prio value (kernel-internal form)
3432 * This function changes the 'effective' priority of a task. It does
3433 * not touch ->normal_prio like __setscheduler().
3435 * Used by the rt_mutex code to implement priority inheritance
3436 * logic. Call site only calls if the priority of the task changed.
3438 void rt_mutex_setprio(struct task_struct *p, int prio)
3440 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3442 const struct sched_class *prev_class;
3444 BUG_ON(prio > MAX_PRIO);
3446 rq = __task_rq_lock(p);
3449 * Idle task boosting is a nono in general. There is one
3450 * exception, when PREEMPT_RT and NOHZ is active:
3452 * The idle task calls get_next_timer_interrupt() and holds
3453 * the timer wheel base->lock on the CPU and another CPU wants
3454 * to access the timer (probably to cancel it). We can safely
3455 * ignore the boosting request, as the idle CPU runs this code
3456 * with interrupts disabled and will complete the lock
3457 * protected section without being interrupted. So there is no
3458 * real need to boost.
3460 if (unlikely(p == rq->idle)) {
3461 WARN_ON(p != rq->curr);
3462 WARN_ON(p->pi_blocked_on);
3466 trace_sched_pi_setprio(p, prio);
3469 if (oldprio == prio)
3470 queue_flag &= ~DEQUEUE_MOVE;
3472 prev_class = p->sched_class;
3473 queued = task_on_rq_queued(p);
3474 running = task_current(rq, p);
3476 dequeue_task(rq, p, queue_flag);
3478 put_prev_task(rq, p);
3481 * Boosting condition are:
3482 * 1. -rt task is running and holds mutex A
3483 * --> -dl task blocks on mutex A
3485 * 2. -dl task is running and holds mutex A
3486 * --> -dl task blocks on mutex A and could preempt the
3489 if (dl_prio(prio)) {
3490 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3491 if (!dl_prio(p->normal_prio) ||
3492 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3493 p->dl.dl_boosted = 1;
3494 queue_flag |= ENQUEUE_REPLENISH;
3496 p->dl.dl_boosted = 0;
3497 p->sched_class = &dl_sched_class;
3498 } else if (rt_prio(prio)) {
3499 if (dl_prio(oldprio))
3500 p->dl.dl_boosted = 0;
3502 queue_flag |= ENQUEUE_HEAD;
3503 p->sched_class = &rt_sched_class;
3505 if (dl_prio(oldprio))
3506 p->dl.dl_boosted = 0;
3507 if (rt_prio(oldprio))
3509 p->sched_class = &fair_sched_class;
3515 p->sched_class->set_curr_task(rq);
3517 enqueue_task(rq, p, queue_flag);
3519 check_class_changed(rq, p, prev_class, oldprio);
3521 preempt_disable(); /* avoid rq from going away on us */
3522 __task_rq_unlock(rq);
3524 balance_callback(rq);
3529 void set_user_nice(struct task_struct *p, long nice)
3531 int old_prio, delta, queued;
3532 unsigned long flags;
3535 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3538 * We have to be careful, if called from sys_setpriority(),
3539 * the task might be in the middle of scheduling on another CPU.
3541 rq = task_rq_lock(p, &flags);
3543 * The RT priorities are set via sched_setscheduler(), but we still
3544 * allow the 'normal' nice value to be set - but as expected
3545 * it wont have any effect on scheduling until the task is
3546 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3548 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3549 p->static_prio = NICE_TO_PRIO(nice);
3552 queued = task_on_rq_queued(p);
3554 dequeue_task(rq, p, DEQUEUE_SAVE);
3556 p->static_prio = NICE_TO_PRIO(nice);
3559 p->prio = effective_prio(p);
3560 delta = p->prio - old_prio;
3563 enqueue_task(rq, p, ENQUEUE_RESTORE);
3565 * If the task increased its priority or is running and
3566 * lowered its priority, then reschedule its CPU:
3568 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3572 task_rq_unlock(rq, p, &flags);
3574 EXPORT_SYMBOL(set_user_nice);
3577 * can_nice - check if a task can reduce its nice value
3581 int can_nice(const struct task_struct *p, const int nice)
3583 /* convert nice value [19,-20] to rlimit style value [1,40] */
3584 int nice_rlim = nice_to_rlimit(nice);
3586 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3587 capable(CAP_SYS_NICE));
3590 #ifdef __ARCH_WANT_SYS_NICE
3593 * sys_nice - change the priority of the current process.
3594 * @increment: priority increment
3596 * sys_setpriority is a more generic, but much slower function that
3597 * does similar things.
3599 SYSCALL_DEFINE1(nice, int, increment)
3604 * Setpriority might change our priority at the same moment.
3605 * We don't have to worry. Conceptually one call occurs first
3606 * and we have a single winner.
3608 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3609 nice = task_nice(current) + increment;
3611 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3612 if (increment < 0 && !can_nice(current, nice))
3615 retval = security_task_setnice(current, nice);
3619 set_user_nice(current, nice);
3626 * task_prio - return the priority value of a given task.
3627 * @p: the task in question.
3629 * Return: The priority value as seen by users in /proc.
3630 * RT tasks are offset by -200. Normal tasks are centered
3631 * around 0, value goes from -16 to +15.
3633 int task_prio(const struct task_struct *p)
3635 return p->prio - MAX_RT_PRIO;
3639 * idle_cpu - is a given cpu idle currently?
3640 * @cpu: the processor in question.
3642 * Return: 1 if the CPU is currently idle. 0 otherwise.
3644 int idle_cpu(int cpu)
3646 struct rq *rq = cpu_rq(cpu);
3648 if (rq->curr != rq->idle)
3655 if (!llist_empty(&rq->wake_list))
3663 * idle_task - return the idle task for a given cpu.
3664 * @cpu: the processor in question.
3666 * Return: The idle task for the cpu @cpu.
3668 struct task_struct *idle_task(int cpu)
3670 return cpu_rq(cpu)->idle;
3674 * find_process_by_pid - find a process with a matching PID value.
3675 * @pid: the pid in question.
3677 * The task of @pid, if found. %NULL otherwise.
3679 static struct task_struct *find_process_by_pid(pid_t pid)
3681 return pid ? find_task_by_vpid(pid) : current;
3685 * This function initializes the sched_dl_entity of a newly becoming
3686 * SCHED_DEADLINE task.
3688 * Only the static values are considered here, the actual runtime and the
3689 * absolute deadline will be properly calculated when the task is enqueued
3690 * for the first time with its new policy.
3693 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3695 struct sched_dl_entity *dl_se = &p->dl;
3697 dl_se->dl_runtime = attr->sched_runtime;
3698 dl_se->dl_deadline = attr->sched_deadline;
3699 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3700 dl_se->flags = attr->sched_flags;
3701 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3704 * Changing the parameters of a task is 'tricky' and we're not doing
3705 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3707 * What we SHOULD do is delay the bandwidth release until the 0-lag
3708 * point. This would include retaining the task_struct until that time
3709 * and change dl_overflow() to not immediately decrement the current
3712 * Instead we retain the current runtime/deadline and let the new
3713 * parameters take effect after the current reservation period lapses.
3714 * This is safe (albeit pessimistic) because the 0-lag point is always
3715 * before the current scheduling deadline.
3717 * We can still have temporary overloads because we do not delay the
3718 * change in bandwidth until that time; so admission control is
3719 * not on the safe side. It does however guarantee tasks will never
3720 * consume more than promised.
3725 * sched_setparam() passes in -1 for its policy, to let the functions
3726 * it calls know not to change it.
3728 #define SETPARAM_POLICY -1
3730 static void __setscheduler_params(struct task_struct *p,
3731 const struct sched_attr *attr)
3733 int policy = attr->sched_policy;
3735 if (policy == SETPARAM_POLICY)
3740 if (dl_policy(policy))
3741 __setparam_dl(p, attr);
3742 else if (fair_policy(policy))
3743 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3746 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3747 * !rt_policy. Always setting this ensures that things like
3748 * getparam()/getattr() don't report silly values for !rt tasks.
3750 p->rt_priority = attr->sched_priority;
3751 p->normal_prio = normal_prio(p);
3755 /* Actually do priority change: must hold pi & rq lock. */
3756 static void __setscheduler(struct rq *rq, struct task_struct *p,
3757 const struct sched_attr *attr, bool keep_boost)
3759 __setscheduler_params(p, attr);
3762 * Keep a potential priority boosting if called from
3763 * sched_setscheduler().
3766 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3768 p->prio = normal_prio(p);
3770 if (dl_prio(p->prio))
3771 p->sched_class = &dl_sched_class;
3772 else if (rt_prio(p->prio))
3773 p->sched_class = &rt_sched_class;
3775 p->sched_class = &fair_sched_class;
3779 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3781 struct sched_dl_entity *dl_se = &p->dl;
3783 attr->sched_priority = p->rt_priority;
3784 attr->sched_runtime = dl_se->dl_runtime;
3785 attr->sched_deadline = dl_se->dl_deadline;
3786 attr->sched_period = dl_se->dl_period;
3787 attr->sched_flags = dl_se->flags;
3791 * This function validates the new parameters of a -deadline task.
3792 * We ask for the deadline not being zero, and greater or equal
3793 * than the runtime, as well as the period of being zero or
3794 * greater than deadline. Furthermore, we have to be sure that
3795 * user parameters are above the internal resolution of 1us (we
3796 * check sched_runtime only since it is always the smaller one) and
3797 * below 2^63 ns (we have to check both sched_deadline and
3798 * sched_period, as the latter can be zero).
3801 __checkparam_dl(const struct sched_attr *attr)
3804 if (attr->sched_deadline == 0)
3808 * Since we truncate DL_SCALE bits, make sure we're at least
3811 if (attr->sched_runtime < (1ULL << DL_SCALE))
3815 * Since we use the MSB for wrap-around and sign issues, make
3816 * sure it's not set (mind that period can be equal to zero).
3818 if (attr->sched_deadline & (1ULL << 63) ||
3819 attr->sched_period & (1ULL << 63))
3822 /* runtime <= deadline <= period (if period != 0) */
3823 if ((attr->sched_period != 0 &&
3824 attr->sched_period < attr->sched_deadline) ||
3825 attr->sched_deadline < attr->sched_runtime)
3832 * check the target process has a UID that matches the current process's
3834 static bool check_same_owner(struct task_struct *p)
3836 const struct cred *cred = current_cred(), *pcred;
3840 pcred = __task_cred(p);
3841 match = (uid_eq(cred->euid, pcred->euid) ||
3842 uid_eq(cred->euid, pcred->uid));
3847 static bool dl_param_changed(struct task_struct *p,
3848 const struct sched_attr *attr)
3850 struct sched_dl_entity *dl_se = &p->dl;
3852 if (dl_se->dl_runtime != attr->sched_runtime ||
3853 dl_se->dl_deadline != attr->sched_deadline ||
3854 dl_se->dl_period != attr->sched_period ||
3855 dl_se->flags != attr->sched_flags)
3861 static int __sched_setscheduler(struct task_struct *p,
3862 const struct sched_attr *attr,
3865 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3866 MAX_RT_PRIO - 1 - attr->sched_priority;
3867 int retval, oldprio, oldpolicy = -1, queued, running;
3868 int new_effective_prio, policy = attr->sched_policy;
3869 unsigned long flags;
3870 const struct sched_class *prev_class;
3873 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3875 /* may grab non-irq protected spin_locks */
3876 BUG_ON(in_interrupt());
3878 /* double check policy once rq lock held */
3880 reset_on_fork = p->sched_reset_on_fork;
3881 policy = oldpolicy = p->policy;
3883 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3885 if (!valid_policy(policy))
3889 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3893 * Valid priorities for SCHED_FIFO and SCHED_RR are
3894 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3895 * SCHED_BATCH and SCHED_IDLE is 0.
3897 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3898 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3900 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3901 (rt_policy(policy) != (attr->sched_priority != 0)))
3905 * Allow unprivileged RT tasks to decrease priority:
3907 if (user && !capable(CAP_SYS_NICE)) {
3908 if (fair_policy(policy)) {
3909 if (attr->sched_nice < task_nice(p) &&
3910 !can_nice(p, attr->sched_nice))
3914 if (rt_policy(policy)) {
3915 unsigned long rlim_rtprio =
3916 task_rlimit(p, RLIMIT_RTPRIO);
3918 /* can't set/change the rt policy */
3919 if (policy != p->policy && !rlim_rtprio)
3922 /* can't increase priority */
3923 if (attr->sched_priority > p->rt_priority &&
3924 attr->sched_priority > rlim_rtprio)
3929 * Can't set/change SCHED_DEADLINE policy at all for now
3930 * (safest behavior); in the future we would like to allow
3931 * unprivileged DL tasks to increase their relative deadline
3932 * or reduce their runtime (both ways reducing utilization)
3934 if (dl_policy(policy))
3938 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3939 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3941 if (idle_policy(p->policy) && !idle_policy(policy)) {
3942 if (!can_nice(p, task_nice(p)))
3946 /* can't change other user's priorities */
3947 if (!check_same_owner(p))
3950 /* Normal users shall not reset the sched_reset_on_fork flag */
3951 if (p->sched_reset_on_fork && !reset_on_fork)
3956 retval = security_task_setscheduler(p);
3962 * make sure no PI-waiters arrive (or leave) while we are
3963 * changing the priority of the task:
3965 * To be able to change p->policy safely, the appropriate
3966 * runqueue lock must be held.
3968 rq = task_rq_lock(p, &flags);
3971 * Changing the policy of the stop threads its a very bad idea
3973 if (p == rq->stop) {
3974 task_rq_unlock(rq, p, &flags);
3979 * If not changing anything there's no need to proceed further,
3980 * but store a possible modification of reset_on_fork.
3982 if (unlikely(policy == p->policy)) {
3983 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3985 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3987 if (dl_policy(policy) && dl_param_changed(p, attr))
3990 p->sched_reset_on_fork = reset_on_fork;
3991 task_rq_unlock(rq, p, &flags);
3997 #ifdef CONFIG_RT_GROUP_SCHED
3999 * Do not allow realtime tasks into groups that have no runtime
4002 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4003 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4004 !task_group_is_autogroup(task_group(p))) {
4005 task_rq_unlock(rq, p, &flags);
4010 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4011 cpumask_t *span = rq->rd->span;
4014 * Don't allow tasks with an affinity mask smaller than
4015 * the entire root_domain to become SCHED_DEADLINE. We
4016 * will also fail if there's no bandwidth available.
4018 if (!cpumask_subset(span, &p->cpus_allowed) ||
4019 rq->rd->dl_bw.bw == 0) {
4020 task_rq_unlock(rq, p, &flags);
4027 /* recheck policy now with rq lock held */
4028 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4029 policy = oldpolicy = -1;
4030 task_rq_unlock(rq, p, &flags);
4035 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4036 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4039 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4040 task_rq_unlock(rq, p, &flags);
4044 p->sched_reset_on_fork = reset_on_fork;
4049 * Take priority boosted tasks into account. If the new
4050 * effective priority is unchanged, we just store the new
4051 * normal parameters and do not touch the scheduler class and
4052 * the runqueue. This will be done when the task deboost
4055 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4056 if (new_effective_prio == oldprio)
4057 queue_flags &= ~DEQUEUE_MOVE;
4060 queued = task_on_rq_queued(p);
4061 running = task_current(rq, p);
4063 dequeue_task(rq, p, queue_flags);
4065 put_prev_task(rq, p);
4067 prev_class = p->sched_class;
4068 __setscheduler(rq, p, attr, pi);
4071 p->sched_class->set_curr_task(rq);
4074 * We enqueue to tail when the priority of a task is
4075 * increased (user space view).
4077 if (oldprio < p->prio)
4078 queue_flags |= ENQUEUE_HEAD;
4080 enqueue_task(rq, p, queue_flags);
4083 check_class_changed(rq, p, prev_class, oldprio);
4084 preempt_disable(); /* avoid rq from going away on us */
4085 task_rq_unlock(rq, p, &flags);
4088 rt_mutex_adjust_pi(p);
4091 * Run balance callbacks after we've adjusted the PI chain.
4093 balance_callback(rq);
4099 static int _sched_setscheduler(struct task_struct *p, int policy,
4100 const struct sched_param *param, bool check)
4102 struct sched_attr attr = {
4103 .sched_policy = policy,
4104 .sched_priority = param->sched_priority,
4105 .sched_nice = PRIO_TO_NICE(p->static_prio),
4108 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4109 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4110 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4111 policy &= ~SCHED_RESET_ON_FORK;
4112 attr.sched_policy = policy;
4115 return __sched_setscheduler(p, &attr, check, true);
4118 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4119 * @p: the task in question.
4120 * @policy: new policy.
4121 * @param: structure containing the new RT priority.
4123 * Return: 0 on success. An error code otherwise.
4125 * NOTE that the task may be already dead.
4127 int sched_setscheduler(struct task_struct *p, int policy,
4128 const struct sched_param *param)
4130 return _sched_setscheduler(p, policy, param, true);
4132 EXPORT_SYMBOL_GPL(sched_setscheduler);
4134 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4136 return __sched_setscheduler(p, attr, true, true);
4138 EXPORT_SYMBOL_GPL(sched_setattr);
4141 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4142 * @p: the task in question.
4143 * @policy: new policy.
4144 * @param: structure containing the new RT priority.
4146 * Just like sched_setscheduler, only don't bother checking if the
4147 * current context has permission. For example, this is needed in
4148 * stop_machine(): we create temporary high priority worker threads,
4149 * but our caller might not have that capability.
4151 * Return: 0 on success. An error code otherwise.
4153 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4154 const struct sched_param *param)
4156 return _sched_setscheduler(p, policy, param, false);
4158 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4161 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4163 struct sched_param lparam;
4164 struct task_struct *p;
4167 if (!param || pid < 0)
4169 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4174 p = find_process_by_pid(pid);
4176 retval = sched_setscheduler(p, policy, &lparam);
4183 * Mimics kernel/events/core.c perf_copy_attr().
4185 static int sched_copy_attr(struct sched_attr __user *uattr,
4186 struct sched_attr *attr)
4191 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4195 * zero the full structure, so that a short copy will be nice.
4197 memset(attr, 0, sizeof(*attr));
4199 ret = get_user(size, &uattr->size);
4203 if (size > PAGE_SIZE) /* silly large */
4206 if (!size) /* abi compat */
4207 size = SCHED_ATTR_SIZE_VER0;
4209 if (size < SCHED_ATTR_SIZE_VER0)
4213 * If we're handed a bigger struct than we know of,
4214 * ensure all the unknown bits are 0 - i.e. new
4215 * user-space does not rely on any kernel feature
4216 * extensions we dont know about yet.
4218 if (size > sizeof(*attr)) {
4219 unsigned char __user *addr;
4220 unsigned char __user *end;
4223 addr = (void __user *)uattr + sizeof(*attr);
4224 end = (void __user *)uattr + size;
4226 for (; addr < end; addr++) {
4227 ret = get_user(val, addr);
4233 size = sizeof(*attr);
4236 ret = copy_from_user(attr, uattr, size);
4241 * XXX: do we want to be lenient like existing syscalls; or do we want
4242 * to be strict and return an error on out-of-bounds values?
4244 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4249 put_user(sizeof(*attr), &uattr->size);
4254 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4255 * @pid: the pid in question.
4256 * @policy: new policy.
4257 * @param: structure containing the new RT priority.
4259 * Return: 0 on success. An error code otherwise.
4261 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4262 struct sched_param __user *, param)
4264 /* negative values for policy are not valid */
4268 return do_sched_setscheduler(pid, policy, param);
4272 * sys_sched_setparam - set/change the RT priority of a thread
4273 * @pid: the pid in question.
4274 * @param: structure containing the new RT priority.
4276 * Return: 0 on success. An error code otherwise.
4278 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4280 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4284 * sys_sched_setattr - same as above, but with extended sched_attr
4285 * @pid: the pid in question.
4286 * @uattr: structure containing the extended parameters.
4287 * @flags: for future extension.
4289 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4290 unsigned int, flags)
4292 struct sched_attr attr;
4293 struct task_struct *p;
4296 if (!uattr || pid < 0 || flags)
4299 retval = sched_copy_attr(uattr, &attr);
4303 if ((int)attr.sched_policy < 0)
4308 p = find_process_by_pid(pid);
4310 retval = sched_setattr(p, &attr);
4317 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4318 * @pid: the pid in question.
4320 * Return: On success, the policy of the thread. Otherwise, a negative error
4323 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4325 struct task_struct *p;
4333 p = find_process_by_pid(pid);
4335 retval = security_task_getscheduler(p);
4338 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4345 * sys_sched_getparam - get the RT priority of a thread
4346 * @pid: the pid in question.
4347 * @param: structure containing the RT priority.
4349 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4352 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4354 struct sched_param lp = { .sched_priority = 0 };
4355 struct task_struct *p;
4358 if (!param || pid < 0)
4362 p = find_process_by_pid(pid);
4367 retval = security_task_getscheduler(p);
4371 if (task_has_rt_policy(p))
4372 lp.sched_priority = p->rt_priority;
4376 * This one might sleep, we cannot do it with a spinlock held ...
4378 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4387 static int sched_read_attr(struct sched_attr __user *uattr,
4388 struct sched_attr *attr,
4393 if (!access_ok(VERIFY_WRITE, uattr, usize))
4397 * If we're handed a smaller struct than we know of,
4398 * ensure all the unknown bits are 0 - i.e. old
4399 * user-space does not get uncomplete information.
4401 if (usize < sizeof(*attr)) {
4402 unsigned char *addr;
4405 addr = (void *)attr + usize;
4406 end = (void *)attr + sizeof(*attr);
4408 for (; addr < end; addr++) {
4416 ret = copy_to_user(uattr, attr, attr->size);
4424 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4425 * @pid: the pid in question.
4426 * @uattr: structure containing the extended parameters.
4427 * @size: sizeof(attr) for fwd/bwd comp.
4428 * @flags: for future extension.
4430 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4431 unsigned int, size, unsigned int, flags)
4433 struct sched_attr attr = {
4434 .size = sizeof(struct sched_attr),
4436 struct task_struct *p;
4439 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4440 size < SCHED_ATTR_SIZE_VER0 || flags)
4444 p = find_process_by_pid(pid);
4449 retval = security_task_getscheduler(p);
4453 attr.sched_policy = p->policy;
4454 if (p->sched_reset_on_fork)
4455 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4456 if (task_has_dl_policy(p))
4457 __getparam_dl(p, &attr);
4458 else if (task_has_rt_policy(p))
4459 attr.sched_priority = p->rt_priority;
4461 attr.sched_nice = task_nice(p);
4465 retval = sched_read_attr(uattr, &attr, size);
4473 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4475 cpumask_var_t cpus_allowed, new_mask;
4476 struct task_struct *p;
4481 p = find_process_by_pid(pid);
4487 /* Prevent p going away */
4491 if (p->flags & PF_NO_SETAFFINITY) {
4495 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4499 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4501 goto out_free_cpus_allowed;
4504 if (!check_same_owner(p)) {
4506 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4508 goto out_free_new_mask;
4513 retval = security_task_setscheduler(p);
4515 goto out_free_new_mask;
4518 cpuset_cpus_allowed(p, cpus_allowed);
4519 cpumask_and(new_mask, in_mask, cpus_allowed);
4522 * Since bandwidth control happens on root_domain basis,
4523 * if admission test is enabled, we only admit -deadline
4524 * tasks allowed to run on all the CPUs in the task's
4528 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4530 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4533 goto out_free_new_mask;
4539 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4542 cpuset_cpus_allowed(p, cpus_allowed);
4543 if (!cpumask_subset(new_mask, cpus_allowed)) {
4545 * We must have raced with a concurrent cpuset
4546 * update. Just reset the cpus_allowed to the
4547 * cpuset's cpus_allowed
4549 cpumask_copy(new_mask, cpus_allowed);
4554 free_cpumask_var(new_mask);
4555 out_free_cpus_allowed:
4556 free_cpumask_var(cpus_allowed);
4562 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4563 struct cpumask *new_mask)
4565 if (len < cpumask_size())
4566 cpumask_clear(new_mask);
4567 else if (len > cpumask_size())
4568 len = cpumask_size();
4570 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4574 * sys_sched_setaffinity - set the cpu affinity of a process
4575 * @pid: pid of the process
4576 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4577 * @user_mask_ptr: user-space pointer to the new cpu mask
4579 * Return: 0 on success. An error code otherwise.
4581 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4582 unsigned long __user *, user_mask_ptr)
4584 cpumask_var_t new_mask;
4587 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4590 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4592 retval = sched_setaffinity(pid, new_mask);
4593 free_cpumask_var(new_mask);
4597 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4599 struct task_struct *p;
4600 unsigned long flags;
4606 p = find_process_by_pid(pid);
4610 retval = security_task_getscheduler(p);
4614 raw_spin_lock_irqsave(&p->pi_lock, flags);
4615 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4616 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4625 * sys_sched_getaffinity - get the cpu affinity of a process
4626 * @pid: pid of the process
4627 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4628 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4630 * Return: 0 on success. An error code otherwise.
4632 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4633 unsigned long __user *, user_mask_ptr)
4638 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4640 if (len & (sizeof(unsigned long)-1))
4643 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4646 ret = sched_getaffinity(pid, mask);
4648 size_t retlen = min_t(size_t, len, cpumask_size());
4650 if (copy_to_user(user_mask_ptr, mask, retlen))
4655 free_cpumask_var(mask);
4661 * sys_sched_yield - yield the current processor to other threads.
4663 * This function yields the current CPU to other tasks. If there are no
4664 * other threads running on this CPU then this function will return.
4668 SYSCALL_DEFINE0(sched_yield)
4670 struct rq *rq = this_rq_lock();
4672 schedstat_inc(rq, yld_count);
4673 current->sched_class->yield_task(rq);
4676 * Since we are going to call schedule() anyway, there's
4677 * no need to preempt or enable interrupts:
4679 __release(rq->lock);
4680 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4681 do_raw_spin_unlock(&rq->lock);
4682 sched_preempt_enable_no_resched();
4689 int __sched _cond_resched(void)
4691 if (should_resched(0)) {
4692 preempt_schedule_common();
4697 EXPORT_SYMBOL(_cond_resched);
4700 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4701 * call schedule, and on return reacquire the lock.
4703 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4704 * operations here to prevent schedule() from being called twice (once via
4705 * spin_unlock(), once by hand).
4707 int __cond_resched_lock(spinlock_t *lock)
4709 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4712 lockdep_assert_held(lock);
4714 if (spin_needbreak(lock) || resched) {
4717 preempt_schedule_common();
4725 EXPORT_SYMBOL(__cond_resched_lock);
4727 int __sched __cond_resched_softirq(void)
4729 BUG_ON(!in_softirq());
4731 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4733 preempt_schedule_common();
4739 EXPORT_SYMBOL(__cond_resched_softirq);
4742 * yield - yield the current processor to other threads.
4744 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4746 * The scheduler is at all times free to pick the calling task as the most
4747 * eligible task to run, if removing the yield() call from your code breaks
4748 * it, its already broken.
4750 * Typical broken usage is:
4755 * where one assumes that yield() will let 'the other' process run that will
4756 * make event true. If the current task is a SCHED_FIFO task that will never
4757 * happen. Never use yield() as a progress guarantee!!
4759 * If you want to use yield() to wait for something, use wait_event().
4760 * If you want to use yield() to be 'nice' for others, use cond_resched().
4761 * If you still want to use yield(), do not!
4763 void __sched yield(void)
4765 set_current_state(TASK_RUNNING);
4768 EXPORT_SYMBOL(yield);
4771 * yield_to - yield the current processor to another thread in
4772 * your thread group, or accelerate that thread toward the
4773 * processor it's on.
4775 * @preempt: whether task preemption is allowed or not
4777 * It's the caller's job to ensure that the target task struct
4778 * can't go away on us before we can do any checks.
4781 * true (>0) if we indeed boosted the target task.
4782 * false (0) if we failed to boost the target.
4783 * -ESRCH if there's no task to yield to.
4785 int __sched yield_to(struct task_struct *p, bool preempt)
4787 struct task_struct *curr = current;
4788 struct rq *rq, *p_rq;
4789 unsigned long flags;
4792 local_irq_save(flags);
4798 * If we're the only runnable task on the rq and target rq also
4799 * has only one task, there's absolutely no point in yielding.
4801 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4806 double_rq_lock(rq, p_rq);
4807 if (task_rq(p) != p_rq) {
4808 double_rq_unlock(rq, p_rq);
4812 if (!curr->sched_class->yield_to_task)
4815 if (curr->sched_class != p->sched_class)
4818 if (task_running(p_rq, p) || p->state)
4821 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4823 schedstat_inc(rq, yld_count);
4825 * Make p's CPU reschedule; pick_next_entity takes care of
4828 if (preempt && rq != p_rq)
4833 double_rq_unlock(rq, p_rq);
4835 local_irq_restore(flags);
4842 EXPORT_SYMBOL_GPL(yield_to);
4845 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4846 * that process accounting knows that this is a task in IO wait state.
4848 long __sched io_schedule_timeout(long timeout)
4850 int old_iowait = current->in_iowait;
4854 current->in_iowait = 1;
4855 blk_schedule_flush_plug(current);
4857 delayacct_blkio_start();
4859 atomic_inc(&rq->nr_iowait);
4860 ret = schedule_timeout(timeout);
4861 current->in_iowait = old_iowait;
4862 atomic_dec(&rq->nr_iowait);
4863 delayacct_blkio_end();
4867 EXPORT_SYMBOL(io_schedule_timeout);
4870 * sys_sched_get_priority_max - return maximum RT priority.
4871 * @policy: scheduling class.
4873 * Return: On success, this syscall returns the maximum
4874 * rt_priority that can be used by a given scheduling class.
4875 * On failure, a negative error code is returned.
4877 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4884 ret = MAX_USER_RT_PRIO-1;
4886 case SCHED_DEADLINE:
4897 * sys_sched_get_priority_min - return minimum RT priority.
4898 * @policy: scheduling class.
4900 * Return: On success, this syscall returns the minimum
4901 * rt_priority that can be used by a given scheduling class.
4902 * On failure, a negative error code is returned.
4904 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4913 case SCHED_DEADLINE:
4923 * sys_sched_rr_get_interval - return the default timeslice of a process.
4924 * @pid: pid of the process.
4925 * @interval: userspace pointer to the timeslice value.
4927 * this syscall writes the default timeslice value of a given process
4928 * into the user-space timespec buffer. A value of '0' means infinity.
4930 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4933 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4934 struct timespec __user *, interval)
4936 struct task_struct *p;
4937 unsigned int time_slice;
4938 unsigned long flags;
4948 p = find_process_by_pid(pid);
4952 retval = security_task_getscheduler(p);
4956 rq = task_rq_lock(p, &flags);
4958 if (p->sched_class->get_rr_interval)
4959 time_slice = p->sched_class->get_rr_interval(rq, p);
4960 task_rq_unlock(rq, p, &flags);
4963 jiffies_to_timespec(time_slice, &t);
4964 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4972 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4974 void sched_show_task(struct task_struct *p)
4976 unsigned long free = 0;
4978 unsigned long state = p->state;
4981 state = __ffs(state) + 1;
4982 printk(KERN_INFO "%-15.15s %c", p->comm,
4983 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4984 #if BITS_PER_LONG == 32
4985 if (state == TASK_RUNNING)
4986 printk(KERN_CONT " running ");
4988 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4990 if (state == TASK_RUNNING)
4991 printk(KERN_CONT " running task ");
4993 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4995 #ifdef CONFIG_DEBUG_STACK_USAGE
4996 free = stack_not_used(p);
5001 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5003 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5004 task_pid_nr(p), ppid,
5005 (unsigned long)task_thread_info(p)->flags);
5007 print_worker_info(KERN_INFO, p);
5008 show_stack(p, NULL);
5011 void show_state_filter(unsigned long state_filter)
5013 struct task_struct *g, *p;
5015 #if BITS_PER_LONG == 32
5017 " task PC stack pid father\n");
5020 " task PC stack pid father\n");
5023 for_each_process_thread(g, p) {
5025 * reset the NMI-timeout, listing all files on a slow
5026 * console might take a lot of time:
5028 touch_nmi_watchdog();
5029 if (!state_filter || (p->state & state_filter))
5033 touch_all_softlockup_watchdogs();
5035 #ifdef CONFIG_SCHED_DEBUG
5036 sysrq_sched_debug_show();
5040 * Only show locks if all tasks are dumped:
5043 debug_show_all_locks();
5046 void init_idle_bootup_task(struct task_struct *idle)
5048 idle->sched_class = &idle_sched_class;
5052 * init_idle - set up an idle thread for a given CPU
5053 * @idle: task in question
5054 * @cpu: cpu the idle task belongs to
5056 * NOTE: this function does not set the idle thread's NEED_RESCHED
5057 * flag, to make booting more robust.
5059 void init_idle(struct task_struct *idle, int cpu)
5061 struct rq *rq = cpu_rq(cpu);
5062 unsigned long flags;
5064 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5065 raw_spin_lock(&rq->lock);
5067 __sched_fork(0, idle);
5068 idle->state = TASK_RUNNING;
5069 idle->se.exec_start = sched_clock();
5071 kasan_unpoison_task_stack(idle);
5075 * Its possible that init_idle() gets called multiple times on a task,
5076 * in that case do_set_cpus_allowed() will not do the right thing.
5078 * And since this is boot we can forgo the serialization.
5080 set_cpus_allowed_common(idle, cpumask_of(cpu));
5083 * We're having a chicken and egg problem, even though we are
5084 * holding rq->lock, the cpu isn't yet set to this cpu so the
5085 * lockdep check in task_group() will fail.
5087 * Similar case to sched_fork(). / Alternatively we could
5088 * use task_rq_lock() here and obtain the other rq->lock.
5093 __set_task_cpu(idle, cpu);
5096 rq->curr = rq->idle = idle;
5097 idle->on_rq = TASK_ON_RQ_QUEUED;
5101 raw_spin_unlock(&rq->lock);
5102 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5104 /* Set the preempt count _outside_ the spinlocks! */
5105 init_idle_preempt_count(idle, cpu);
5108 * The idle tasks have their own, simple scheduling class:
5110 idle->sched_class = &idle_sched_class;
5111 ftrace_graph_init_idle_task(idle, cpu);
5112 vtime_init_idle(idle, cpu);
5114 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5118 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5119 const struct cpumask *trial)
5121 int ret = 1, trial_cpus;
5122 struct dl_bw *cur_dl_b;
5123 unsigned long flags;
5125 if (!cpumask_weight(cur))
5128 rcu_read_lock_sched();
5129 cur_dl_b = dl_bw_of(cpumask_any(cur));
5130 trial_cpus = cpumask_weight(trial);
5132 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5133 if (cur_dl_b->bw != -1 &&
5134 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5136 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5137 rcu_read_unlock_sched();
5142 int task_can_attach(struct task_struct *p,
5143 const struct cpumask *cs_cpus_allowed)
5148 * Kthreads which disallow setaffinity shouldn't be moved
5149 * to a new cpuset; we don't want to change their cpu
5150 * affinity and isolating such threads by their set of
5151 * allowed nodes is unnecessary. Thus, cpusets are not
5152 * applicable for such threads. This prevents checking for
5153 * success of set_cpus_allowed_ptr() on all attached tasks
5154 * before cpus_allowed may be changed.
5156 if (p->flags & PF_NO_SETAFFINITY) {
5162 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5164 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5169 unsigned long flags;
5171 rcu_read_lock_sched();
5172 dl_b = dl_bw_of(dest_cpu);
5173 raw_spin_lock_irqsave(&dl_b->lock, flags);
5174 cpus = dl_bw_cpus(dest_cpu);
5175 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5180 * We reserve space for this task in the destination
5181 * root_domain, as we can't fail after this point.
5182 * We will free resources in the source root_domain
5183 * later on (see set_cpus_allowed_dl()).
5185 __dl_add(dl_b, p->dl.dl_bw);
5187 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5188 rcu_read_unlock_sched();
5198 static bool sched_smp_initialized __read_mostly;
5200 #ifdef CONFIG_NUMA_BALANCING
5201 /* Migrate current task p to target_cpu */
5202 int migrate_task_to(struct task_struct *p, int target_cpu)
5204 struct migration_arg arg = { p, target_cpu };
5205 int curr_cpu = task_cpu(p);
5207 if (curr_cpu == target_cpu)
5210 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5213 /* TODO: This is not properly updating schedstats */
5215 trace_sched_move_numa(p, curr_cpu, target_cpu);
5216 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5220 * Requeue a task on a given node and accurately track the number of NUMA
5221 * tasks on the runqueues
5223 void sched_setnuma(struct task_struct *p, int nid)
5226 unsigned long flags;
5227 bool queued, running;
5229 rq = task_rq_lock(p, &flags);
5230 queued = task_on_rq_queued(p);
5231 running = task_current(rq, p);
5234 dequeue_task(rq, p, DEQUEUE_SAVE);
5236 put_prev_task(rq, p);
5238 p->numa_preferred_nid = nid;
5241 p->sched_class->set_curr_task(rq);
5243 enqueue_task(rq, p, ENQUEUE_RESTORE);
5244 task_rq_unlock(rq, p, &flags);
5246 #endif /* CONFIG_NUMA_BALANCING */
5248 #ifdef CONFIG_HOTPLUG_CPU
5250 * Ensures that the idle task is using init_mm right before its cpu goes
5253 void idle_task_exit(void)
5255 struct mm_struct *mm = current->active_mm;
5257 BUG_ON(cpu_online(smp_processor_id()));
5259 if (mm != &init_mm) {
5260 switch_mm(mm, &init_mm, current);
5261 finish_arch_post_lock_switch();
5267 * Since this CPU is going 'away' for a while, fold any nr_active delta
5268 * we might have. Assumes we're called after migrate_tasks() so that the
5269 * nr_active count is stable.
5271 * Also see the comment "Global load-average calculations".
5273 static void calc_load_migrate(struct rq *rq)
5275 long delta = calc_load_fold_active(rq);
5277 atomic_long_add(delta, &calc_load_tasks);
5280 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5284 static const struct sched_class fake_sched_class = {
5285 .put_prev_task = put_prev_task_fake,
5288 static struct task_struct fake_task = {
5290 * Avoid pull_{rt,dl}_task()
5292 .prio = MAX_PRIO + 1,
5293 .sched_class = &fake_sched_class,
5297 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5298 * try_to_wake_up()->select_task_rq().
5300 * Called with rq->lock held even though we'er in stop_machine() and
5301 * there's no concurrency possible, we hold the required locks anyway
5302 * because of lock validation efforts.
5304 static void migrate_tasks(struct rq *dead_rq)
5306 struct rq *rq = dead_rq;
5307 struct task_struct *next, *stop = rq->stop;
5311 * Fudge the rq selection such that the below task selection loop
5312 * doesn't get stuck on the currently eligible stop task.
5314 * We're currently inside stop_machine() and the rq is either stuck
5315 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5316 * either way we should never end up calling schedule() until we're
5322 * put_prev_task() and pick_next_task() sched
5323 * class method both need to have an up-to-date
5324 * value of rq->clock[_task]
5326 update_rq_clock(rq);
5330 * There's this thread running, bail when that's the only
5333 if (rq->nr_running == 1)
5337 * pick_next_task assumes pinned rq->lock.
5339 lockdep_pin_lock(&rq->lock);
5340 next = pick_next_task(rq, &fake_task);
5342 next->sched_class->put_prev_task(rq, next);
5345 * Rules for changing task_struct::cpus_allowed are holding
5346 * both pi_lock and rq->lock, such that holding either
5347 * stabilizes the mask.
5349 * Drop rq->lock is not quite as disastrous as it usually is
5350 * because !cpu_active at this point, which means load-balance
5351 * will not interfere. Also, stop-machine.
5353 lockdep_unpin_lock(&rq->lock);
5354 raw_spin_unlock(&rq->lock);
5355 raw_spin_lock(&next->pi_lock);
5356 raw_spin_lock(&rq->lock);
5359 * Since we're inside stop-machine, _nothing_ should have
5360 * changed the task, WARN if weird stuff happened, because in
5361 * that case the above rq->lock drop is a fail too.
5363 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5364 raw_spin_unlock(&next->pi_lock);
5368 /* Find suitable destination for @next, with force if needed. */
5369 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5371 rq = __migrate_task(rq, next, dest_cpu);
5372 if (rq != dead_rq) {
5373 raw_spin_unlock(&rq->lock);
5375 raw_spin_lock(&rq->lock);
5377 raw_spin_unlock(&next->pi_lock);
5382 #endif /* CONFIG_HOTPLUG_CPU */
5384 static void set_rq_online(struct rq *rq)
5387 const struct sched_class *class;
5389 cpumask_set_cpu(rq->cpu, rq->rd->online);
5392 for_each_class(class) {
5393 if (class->rq_online)
5394 class->rq_online(rq);
5399 static void set_rq_offline(struct rq *rq)
5402 const struct sched_class *class;
5404 for_each_class(class) {
5405 if (class->rq_offline)
5406 class->rq_offline(rq);
5409 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5414 static void set_cpu_rq_start_time(unsigned int cpu)
5416 struct rq *rq = cpu_rq(cpu);
5418 rq->age_stamp = sched_clock_cpu(cpu);
5421 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5423 #ifdef CONFIG_SCHED_DEBUG
5425 static __read_mostly int sched_debug_enabled;
5427 static int __init sched_debug_setup(char *str)
5429 sched_debug_enabled = 1;
5433 early_param("sched_debug", sched_debug_setup);
5435 static inline bool sched_debug(void)
5437 return sched_debug_enabled;
5440 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5441 struct cpumask *groupmask)
5443 struct sched_group *group = sd->groups;
5445 cpumask_clear(groupmask);
5447 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5449 if (!(sd->flags & SD_LOAD_BALANCE)) {
5450 printk("does not load-balance\n");
5452 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5457 printk(KERN_CONT "span %*pbl level %s\n",
5458 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5460 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5461 printk(KERN_ERR "ERROR: domain->span does not contain "
5464 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5465 printk(KERN_ERR "ERROR: domain->groups does not contain"
5469 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5473 printk(KERN_ERR "ERROR: group is NULL\n");
5477 if (!cpumask_weight(sched_group_cpus(group))) {
5478 printk(KERN_CONT "\n");
5479 printk(KERN_ERR "ERROR: empty group\n");
5483 if (!(sd->flags & SD_OVERLAP) &&
5484 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5485 printk(KERN_CONT "\n");
5486 printk(KERN_ERR "ERROR: repeated CPUs\n");
5490 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5492 printk(KERN_CONT " %*pbl",
5493 cpumask_pr_args(sched_group_cpus(group)));
5494 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5495 printk(KERN_CONT " (cpu_capacity = %d)",
5496 group->sgc->capacity);
5499 group = group->next;
5500 } while (group != sd->groups);
5501 printk(KERN_CONT "\n");
5503 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5504 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5507 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5508 printk(KERN_ERR "ERROR: parent span is not a superset "
5509 "of domain->span\n");
5513 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5517 if (!sched_debug_enabled)
5521 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5525 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5528 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5536 #else /* !CONFIG_SCHED_DEBUG */
5537 # define sched_domain_debug(sd, cpu) do { } while (0)
5538 static inline bool sched_debug(void)
5542 #endif /* CONFIG_SCHED_DEBUG */
5544 static int sd_degenerate(struct sched_domain *sd)
5546 if (cpumask_weight(sched_domain_span(sd)) == 1)
5549 /* Following flags need at least 2 groups */
5550 if (sd->flags & (SD_LOAD_BALANCE |
5551 SD_BALANCE_NEWIDLE |
5554 SD_SHARE_CPUCAPACITY |
5555 SD_SHARE_PKG_RESOURCES |
5556 SD_SHARE_POWERDOMAIN)) {
5557 if (sd->groups != sd->groups->next)
5561 /* Following flags don't use groups */
5562 if (sd->flags & (SD_WAKE_AFFINE))
5569 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5571 unsigned long cflags = sd->flags, pflags = parent->flags;
5573 if (sd_degenerate(parent))
5576 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5579 /* Flags needing groups don't count if only 1 group in parent */
5580 if (parent->groups == parent->groups->next) {
5581 pflags &= ~(SD_LOAD_BALANCE |
5582 SD_BALANCE_NEWIDLE |
5585 SD_SHARE_CPUCAPACITY |
5586 SD_SHARE_PKG_RESOURCES |
5588 SD_SHARE_POWERDOMAIN);
5589 if (nr_node_ids == 1)
5590 pflags &= ~SD_SERIALIZE;
5592 if (~cflags & pflags)
5598 static void free_rootdomain(struct rcu_head *rcu)
5600 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5602 cpupri_cleanup(&rd->cpupri);
5603 cpudl_cleanup(&rd->cpudl);
5604 free_cpumask_var(rd->dlo_mask);
5605 free_cpumask_var(rd->rto_mask);
5606 free_cpumask_var(rd->online);
5607 free_cpumask_var(rd->span);
5611 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5613 struct root_domain *old_rd = NULL;
5614 unsigned long flags;
5616 raw_spin_lock_irqsave(&rq->lock, flags);
5621 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5624 cpumask_clear_cpu(rq->cpu, old_rd->span);
5627 * If we dont want to free the old_rd yet then
5628 * set old_rd to NULL to skip the freeing later
5631 if (!atomic_dec_and_test(&old_rd->refcount))
5635 atomic_inc(&rd->refcount);
5638 cpumask_set_cpu(rq->cpu, rd->span);
5639 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5642 raw_spin_unlock_irqrestore(&rq->lock, flags);
5645 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5648 static int init_rootdomain(struct root_domain *rd)
5650 memset(rd, 0, sizeof(*rd));
5652 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5654 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5656 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5658 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5661 init_dl_bw(&rd->dl_bw);
5662 if (cpudl_init(&rd->cpudl) != 0)
5665 if (cpupri_init(&rd->cpupri) != 0)
5670 free_cpumask_var(rd->rto_mask);
5672 free_cpumask_var(rd->dlo_mask);
5674 free_cpumask_var(rd->online);
5676 free_cpumask_var(rd->span);
5682 * By default the system creates a single root-domain with all cpus as
5683 * members (mimicking the global state we have today).
5685 struct root_domain def_root_domain;
5687 static void init_defrootdomain(void)
5689 init_rootdomain(&def_root_domain);
5691 atomic_set(&def_root_domain.refcount, 1);
5694 static struct root_domain *alloc_rootdomain(void)
5696 struct root_domain *rd;
5698 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5702 if (init_rootdomain(rd) != 0) {
5710 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5712 struct sched_group *tmp, *first;
5721 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5726 } while (sg != first);
5729 static void free_sched_domain(struct rcu_head *rcu)
5731 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5734 * If its an overlapping domain it has private groups, iterate and
5737 if (sd->flags & SD_OVERLAP) {
5738 free_sched_groups(sd->groups, 1);
5739 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5740 kfree(sd->groups->sgc);
5746 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5748 call_rcu(&sd->rcu, free_sched_domain);
5751 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5753 for (; sd; sd = sd->parent)
5754 destroy_sched_domain(sd, cpu);
5758 * Keep a special pointer to the highest sched_domain that has
5759 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5760 * allows us to avoid some pointer chasing select_idle_sibling().
5762 * Also keep a unique ID per domain (we use the first cpu number in
5763 * the cpumask of the domain), this allows us to quickly tell if
5764 * two cpus are in the same cache domain, see cpus_share_cache().
5766 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5767 DEFINE_PER_CPU(int, sd_llc_size);
5768 DEFINE_PER_CPU(int, sd_llc_id);
5769 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5770 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5771 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5773 static void update_top_cache_domain(int cpu)
5775 struct sched_domain *sd;
5776 struct sched_domain *busy_sd = NULL;
5780 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5782 id = cpumask_first(sched_domain_span(sd));
5783 size = cpumask_weight(sched_domain_span(sd));
5784 busy_sd = sd->parent; /* sd_busy */
5786 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5788 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5789 per_cpu(sd_llc_size, cpu) = size;
5790 per_cpu(sd_llc_id, cpu) = id;
5792 sd = lowest_flag_domain(cpu, SD_NUMA);
5793 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5795 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5796 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5800 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5801 * hold the hotplug lock.
5804 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5806 struct rq *rq = cpu_rq(cpu);
5807 struct sched_domain *tmp;
5809 /* Remove the sched domains which do not contribute to scheduling. */
5810 for (tmp = sd; tmp; ) {
5811 struct sched_domain *parent = tmp->parent;
5815 if (sd_parent_degenerate(tmp, parent)) {
5816 tmp->parent = parent->parent;
5818 parent->parent->child = tmp;
5820 * Transfer SD_PREFER_SIBLING down in case of a
5821 * degenerate parent; the spans match for this
5822 * so the property transfers.
5824 if (parent->flags & SD_PREFER_SIBLING)
5825 tmp->flags |= SD_PREFER_SIBLING;
5826 destroy_sched_domain(parent, cpu);
5831 if (sd && sd_degenerate(sd)) {
5834 destroy_sched_domain(tmp, cpu);
5839 sched_domain_debug(sd, cpu);
5841 rq_attach_root(rq, rd);
5843 rcu_assign_pointer(rq->sd, sd);
5844 destroy_sched_domains(tmp, cpu);
5846 update_top_cache_domain(cpu);
5849 /* Setup the mask of cpus configured for isolated domains */
5850 static int __init isolated_cpu_setup(char *str)
5854 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5855 ret = cpulist_parse(str, cpu_isolated_map);
5857 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5862 __setup("isolcpus=", isolated_cpu_setup);
5865 struct sched_domain ** __percpu sd;
5866 struct root_domain *rd;
5877 * Build an iteration mask that can exclude certain CPUs from the upwards
5880 * Asymmetric node setups can result in situations where the domain tree is of
5881 * unequal depth, make sure to skip domains that already cover the entire
5884 * In that case build_sched_domains() will have terminated the iteration early
5885 * and our sibling sd spans will be empty. Domains should always include the
5886 * cpu they're built on, so check that.
5889 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5891 const struct cpumask *span = sched_domain_span(sd);
5892 struct sd_data *sdd = sd->private;
5893 struct sched_domain *sibling;
5896 for_each_cpu(i, span) {
5897 sibling = *per_cpu_ptr(sdd->sd, i);
5898 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5901 cpumask_set_cpu(i, sched_group_mask(sg));
5906 * Return the canonical balance cpu for this group, this is the first cpu
5907 * of this group that's also in the iteration mask.
5909 int group_balance_cpu(struct sched_group *sg)
5911 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5915 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5917 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5918 const struct cpumask *span = sched_domain_span(sd);
5919 struct cpumask *covered = sched_domains_tmpmask;
5920 struct sd_data *sdd = sd->private;
5921 struct sched_domain *sibling;
5924 cpumask_clear(covered);
5926 for_each_cpu(i, span) {
5927 struct cpumask *sg_span;
5929 if (cpumask_test_cpu(i, covered))
5932 sibling = *per_cpu_ptr(sdd->sd, i);
5934 /* See the comment near build_group_mask(). */
5935 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5938 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5939 GFP_KERNEL, cpu_to_node(cpu));
5944 sg_span = sched_group_cpus(sg);
5946 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5948 cpumask_set_cpu(i, sg_span);
5950 cpumask_or(covered, covered, sg_span);
5952 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5953 if (atomic_inc_return(&sg->sgc->ref) == 1)
5954 build_group_mask(sd, sg);
5957 * Initialize sgc->capacity such that even if we mess up the
5958 * domains and no possible iteration will get us here, we won't
5961 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5964 * Make sure the first group of this domain contains the
5965 * canonical balance cpu. Otherwise the sched_domain iteration
5966 * breaks. See update_sg_lb_stats().
5968 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5969 group_balance_cpu(sg) == cpu)
5979 sd->groups = groups;
5984 free_sched_groups(first, 0);
5989 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5991 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5992 struct sched_domain *child = sd->child;
5995 cpu = cpumask_first(sched_domain_span(child));
5998 *sg = *per_cpu_ptr(sdd->sg, cpu);
5999 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6000 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6007 * build_sched_groups will build a circular linked list of the groups
6008 * covered by the given span, and will set each group's ->cpumask correctly,
6009 * and ->cpu_capacity to 0.
6011 * Assumes the sched_domain tree is fully constructed
6014 build_sched_groups(struct sched_domain *sd, int cpu)
6016 struct sched_group *first = NULL, *last = NULL;
6017 struct sd_data *sdd = sd->private;
6018 const struct cpumask *span = sched_domain_span(sd);
6019 struct cpumask *covered;
6022 get_group(cpu, sdd, &sd->groups);
6023 atomic_inc(&sd->groups->ref);
6025 if (cpu != cpumask_first(span))
6028 lockdep_assert_held(&sched_domains_mutex);
6029 covered = sched_domains_tmpmask;
6031 cpumask_clear(covered);
6033 for_each_cpu(i, span) {
6034 struct sched_group *sg;
6037 if (cpumask_test_cpu(i, covered))
6040 group = get_group(i, sdd, &sg);
6041 cpumask_setall(sched_group_mask(sg));
6043 for_each_cpu(j, span) {
6044 if (get_group(j, sdd, NULL) != group)
6047 cpumask_set_cpu(j, covered);
6048 cpumask_set_cpu(j, sched_group_cpus(sg));
6063 * Initialize sched groups cpu_capacity.
6065 * cpu_capacity indicates the capacity of sched group, which is used while
6066 * distributing the load between different sched groups in a sched domain.
6067 * Typically cpu_capacity for all the groups in a sched domain will be same
6068 * unless there are asymmetries in the topology. If there are asymmetries,
6069 * group having more cpu_capacity will pickup more load compared to the
6070 * group having less cpu_capacity.
6072 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6074 struct sched_group *sg = sd->groups;
6079 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6081 } while (sg != sd->groups);
6083 if (cpu != group_balance_cpu(sg))
6086 update_group_capacity(sd, cpu);
6087 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6091 * Initializers for schedule domains
6092 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6095 static int default_relax_domain_level = -1;
6096 int sched_domain_level_max;
6098 static int __init setup_relax_domain_level(char *str)
6100 if (kstrtoint(str, 0, &default_relax_domain_level))
6101 pr_warn("Unable to set relax_domain_level\n");
6105 __setup("relax_domain_level=", setup_relax_domain_level);
6107 static void set_domain_attribute(struct sched_domain *sd,
6108 struct sched_domain_attr *attr)
6112 if (!attr || attr->relax_domain_level < 0) {
6113 if (default_relax_domain_level < 0)
6116 request = default_relax_domain_level;
6118 request = attr->relax_domain_level;
6119 if (request < sd->level) {
6120 /* turn off idle balance on this domain */
6121 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6123 /* turn on idle balance on this domain */
6124 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6128 static void __sdt_free(const struct cpumask *cpu_map);
6129 static int __sdt_alloc(const struct cpumask *cpu_map);
6131 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6132 const struct cpumask *cpu_map)
6136 if (!atomic_read(&d->rd->refcount))
6137 free_rootdomain(&d->rd->rcu); /* fall through */
6139 free_percpu(d->sd); /* fall through */
6141 __sdt_free(cpu_map); /* fall through */
6147 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6148 const struct cpumask *cpu_map)
6150 memset(d, 0, sizeof(*d));
6152 if (__sdt_alloc(cpu_map))
6153 return sa_sd_storage;
6154 d->sd = alloc_percpu(struct sched_domain *);
6156 return sa_sd_storage;
6157 d->rd = alloc_rootdomain();
6160 return sa_rootdomain;
6164 * NULL the sd_data elements we've used to build the sched_domain and
6165 * sched_group structure so that the subsequent __free_domain_allocs()
6166 * will not free the data we're using.
6168 static void claim_allocations(int cpu, struct sched_domain *sd)
6170 struct sd_data *sdd = sd->private;
6172 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6173 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6175 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6176 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6178 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6179 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6183 static int sched_domains_numa_levels;
6184 enum numa_topology_type sched_numa_topology_type;
6185 static int *sched_domains_numa_distance;
6186 int sched_max_numa_distance;
6187 static struct cpumask ***sched_domains_numa_masks;
6188 static int sched_domains_curr_level;
6192 * SD_flags allowed in topology descriptions.
6194 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6195 * SD_SHARE_PKG_RESOURCES - describes shared caches
6196 * SD_NUMA - describes NUMA topologies
6197 * SD_SHARE_POWERDOMAIN - describes shared power domain
6200 * SD_ASYM_PACKING - describes SMT quirks
6202 #define TOPOLOGY_SD_FLAGS \
6203 (SD_SHARE_CPUCAPACITY | \
6204 SD_SHARE_PKG_RESOURCES | \
6207 SD_SHARE_POWERDOMAIN)
6209 static struct sched_domain *
6210 sd_init(struct sched_domain_topology_level *tl, int cpu)
6212 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6213 int sd_weight, sd_flags = 0;
6217 * Ugly hack to pass state to sd_numa_mask()...
6219 sched_domains_curr_level = tl->numa_level;
6222 sd_weight = cpumask_weight(tl->mask(cpu));
6225 sd_flags = (*tl->sd_flags)();
6226 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6227 "wrong sd_flags in topology description\n"))
6228 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6230 *sd = (struct sched_domain){
6231 .min_interval = sd_weight,
6232 .max_interval = 2*sd_weight,
6234 .imbalance_pct = 125,
6236 .cache_nice_tries = 0,
6243 .flags = 1*SD_LOAD_BALANCE
6244 | 1*SD_BALANCE_NEWIDLE
6249 | 0*SD_SHARE_CPUCAPACITY
6250 | 0*SD_SHARE_PKG_RESOURCES
6252 | 0*SD_PREFER_SIBLING
6257 .last_balance = jiffies,
6258 .balance_interval = sd_weight,
6260 .max_newidle_lb_cost = 0,
6261 .next_decay_max_lb_cost = jiffies,
6262 #ifdef CONFIG_SCHED_DEBUG
6268 * Convert topological properties into behaviour.
6271 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6272 sd->flags |= SD_PREFER_SIBLING;
6273 sd->imbalance_pct = 110;
6274 sd->smt_gain = 1178; /* ~15% */
6276 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6277 sd->imbalance_pct = 117;
6278 sd->cache_nice_tries = 1;
6282 } else if (sd->flags & SD_NUMA) {
6283 sd->cache_nice_tries = 2;
6287 sd->flags |= SD_SERIALIZE;
6288 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6289 sd->flags &= ~(SD_BALANCE_EXEC |
6296 sd->flags |= SD_PREFER_SIBLING;
6297 sd->cache_nice_tries = 1;
6302 sd->private = &tl->data;
6308 * Topology list, bottom-up.
6310 static struct sched_domain_topology_level default_topology[] = {
6311 #ifdef CONFIG_SCHED_SMT
6312 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6314 #ifdef CONFIG_SCHED_MC
6315 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6317 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6321 static struct sched_domain_topology_level *sched_domain_topology =
6324 #define for_each_sd_topology(tl) \
6325 for (tl = sched_domain_topology; tl->mask; tl++)
6327 void set_sched_topology(struct sched_domain_topology_level *tl)
6329 sched_domain_topology = tl;
6334 static const struct cpumask *sd_numa_mask(int cpu)
6336 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6339 static void sched_numa_warn(const char *str)
6341 static int done = false;
6349 printk(KERN_WARNING "ERROR: %s\n\n", str);
6351 for (i = 0; i < nr_node_ids; i++) {
6352 printk(KERN_WARNING " ");
6353 for (j = 0; j < nr_node_ids; j++)
6354 printk(KERN_CONT "%02d ", node_distance(i,j));
6355 printk(KERN_CONT "\n");
6357 printk(KERN_WARNING "\n");
6360 bool find_numa_distance(int distance)
6364 if (distance == node_distance(0, 0))
6367 for (i = 0; i < sched_domains_numa_levels; i++) {
6368 if (sched_domains_numa_distance[i] == distance)
6376 * A system can have three types of NUMA topology:
6377 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6378 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6379 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6381 * The difference between a glueless mesh topology and a backplane
6382 * topology lies in whether communication between not directly
6383 * connected nodes goes through intermediary nodes (where programs
6384 * could run), or through backplane controllers. This affects
6385 * placement of programs.
6387 * The type of topology can be discerned with the following tests:
6388 * - If the maximum distance between any nodes is 1 hop, the system
6389 * is directly connected.
6390 * - If for two nodes A and B, located N > 1 hops away from each other,
6391 * there is an intermediary node C, which is < N hops away from both
6392 * nodes A and B, the system is a glueless mesh.
6394 static void init_numa_topology_type(void)
6398 n = sched_max_numa_distance;
6400 if (sched_domains_numa_levels <= 1) {
6401 sched_numa_topology_type = NUMA_DIRECT;
6405 for_each_online_node(a) {
6406 for_each_online_node(b) {
6407 /* Find two nodes furthest removed from each other. */
6408 if (node_distance(a, b) < n)
6411 /* Is there an intermediary node between a and b? */
6412 for_each_online_node(c) {
6413 if (node_distance(a, c) < n &&
6414 node_distance(b, c) < n) {
6415 sched_numa_topology_type =
6421 sched_numa_topology_type = NUMA_BACKPLANE;
6427 static void sched_init_numa(void)
6429 int next_distance, curr_distance = node_distance(0, 0);
6430 struct sched_domain_topology_level *tl;
6434 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6435 if (!sched_domains_numa_distance)
6439 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6440 * unique distances in the node_distance() table.
6442 * Assumes node_distance(0,j) includes all distances in
6443 * node_distance(i,j) in order to avoid cubic time.
6445 next_distance = curr_distance;
6446 for (i = 0; i < nr_node_ids; i++) {
6447 for (j = 0; j < nr_node_ids; j++) {
6448 for (k = 0; k < nr_node_ids; k++) {
6449 int distance = node_distance(i, k);
6451 if (distance > curr_distance &&
6452 (distance < next_distance ||
6453 next_distance == curr_distance))
6454 next_distance = distance;
6457 * While not a strong assumption it would be nice to know
6458 * about cases where if node A is connected to B, B is not
6459 * equally connected to A.
6461 if (sched_debug() && node_distance(k, i) != distance)
6462 sched_numa_warn("Node-distance not symmetric");
6464 if (sched_debug() && i && !find_numa_distance(distance))
6465 sched_numa_warn("Node-0 not representative");
6467 if (next_distance != curr_distance) {
6468 sched_domains_numa_distance[level++] = next_distance;
6469 sched_domains_numa_levels = level;
6470 curr_distance = next_distance;
6475 * In case of sched_debug() we verify the above assumption.
6485 * 'level' contains the number of unique distances, excluding the
6486 * identity distance node_distance(i,i).
6488 * The sched_domains_numa_distance[] array includes the actual distance
6493 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6494 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6495 * the array will contain less then 'level' members. This could be
6496 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6497 * in other functions.
6499 * We reset it to 'level' at the end of this function.
6501 sched_domains_numa_levels = 0;
6503 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6504 if (!sched_domains_numa_masks)
6508 * Now for each level, construct a mask per node which contains all
6509 * cpus of nodes that are that many hops away from us.
6511 for (i = 0; i < level; i++) {
6512 sched_domains_numa_masks[i] =
6513 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6514 if (!sched_domains_numa_masks[i])
6517 for (j = 0; j < nr_node_ids; j++) {
6518 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6522 sched_domains_numa_masks[i][j] = mask;
6525 if (node_distance(j, k) > sched_domains_numa_distance[i])
6528 cpumask_or(mask, mask, cpumask_of_node(k));
6533 /* Compute default topology size */
6534 for (i = 0; sched_domain_topology[i].mask; i++);
6536 tl = kzalloc((i + level + 1) *
6537 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6542 * Copy the default topology bits..
6544 for (i = 0; sched_domain_topology[i].mask; i++)
6545 tl[i] = sched_domain_topology[i];
6548 * .. and append 'j' levels of NUMA goodness.
6550 for (j = 0; j < level; i++, j++) {
6551 tl[i] = (struct sched_domain_topology_level){
6552 .mask = sd_numa_mask,
6553 .sd_flags = cpu_numa_flags,
6554 .flags = SDTL_OVERLAP,
6560 sched_domain_topology = tl;
6562 sched_domains_numa_levels = level;
6563 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6565 init_numa_topology_type();
6568 static void sched_domains_numa_masks_set(unsigned int cpu)
6570 int node = cpu_to_node(cpu);
6573 for (i = 0; i < sched_domains_numa_levels; i++) {
6574 for (j = 0; j < nr_node_ids; j++) {
6575 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6576 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6581 static void sched_domains_numa_masks_clear(unsigned int cpu)
6585 for (i = 0; i < sched_domains_numa_levels; i++) {
6586 for (j = 0; j < nr_node_ids; j++)
6587 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6592 static inline void sched_init_numa(void) { }
6593 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6594 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6595 #endif /* CONFIG_NUMA */
6597 static int __sdt_alloc(const struct cpumask *cpu_map)
6599 struct sched_domain_topology_level *tl;
6602 for_each_sd_topology(tl) {
6603 struct sd_data *sdd = &tl->data;
6605 sdd->sd = alloc_percpu(struct sched_domain *);
6609 sdd->sg = alloc_percpu(struct sched_group *);
6613 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6617 for_each_cpu(j, cpu_map) {
6618 struct sched_domain *sd;
6619 struct sched_group *sg;
6620 struct sched_group_capacity *sgc;
6622 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6623 GFP_KERNEL, cpu_to_node(j));
6627 *per_cpu_ptr(sdd->sd, j) = sd;
6629 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6630 GFP_KERNEL, cpu_to_node(j));
6636 *per_cpu_ptr(sdd->sg, j) = sg;
6638 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6639 GFP_KERNEL, cpu_to_node(j));
6643 *per_cpu_ptr(sdd->sgc, j) = sgc;
6650 static void __sdt_free(const struct cpumask *cpu_map)
6652 struct sched_domain_topology_level *tl;
6655 for_each_sd_topology(tl) {
6656 struct sd_data *sdd = &tl->data;
6658 for_each_cpu(j, cpu_map) {
6659 struct sched_domain *sd;
6662 sd = *per_cpu_ptr(sdd->sd, j);
6663 if (sd && (sd->flags & SD_OVERLAP))
6664 free_sched_groups(sd->groups, 0);
6665 kfree(*per_cpu_ptr(sdd->sd, j));
6669 kfree(*per_cpu_ptr(sdd->sg, j));
6671 kfree(*per_cpu_ptr(sdd->sgc, j));
6673 free_percpu(sdd->sd);
6675 free_percpu(sdd->sg);
6677 free_percpu(sdd->sgc);
6682 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6683 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6684 struct sched_domain *child, int cpu)
6686 struct sched_domain *sd = sd_init(tl, cpu);
6690 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6692 sd->level = child->level + 1;
6693 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6697 if (!cpumask_subset(sched_domain_span(child),
6698 sched_domain_span(sd))) {
6699 pr_err("BUG: arch topology borken\n");
6700 #ifdef CONFIG_SCHED_DEBUG
6701 pr_err(" the %s domain not a subset of the %s domain\n",
6702 child->name, sd->name);
6704 /* Fixup, ensure @sd has at least @child cpus. */
6705 cpumask_or(sched_domain_span(sd),
6706 sched_domain_span(sd),
6707 sched_domain_span(child));
6711 set_domain_attribute(sd, attr);
6717 * Build sched domains for a given set of cpus and attach the sched domains
6718 * to the individual cpus
6720 static int build_sched_domains(const struct cpumask *cpu_map,
6721 struct sched_domain_attr *attr)
6723 enum s_alloc alloc_state;
6724 struct sched_domain *sd;
6726 int i, ret = -ENOMEM;
6728 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6729 if (alloc_state != sa_rootdomain)
6732 /* Set up domains for cpus specified by the cpu_map. */
6733 for_each_cpu(i, cpu_map) {
6734 struct sched_domain_topology_level *tl;
6737 for_each_sd_topology(tl) {
6738 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6739 if (tl == sched_domain_topology)
6740 *per_cpu_ptr(d.sd, i) = sd;
6741 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6742 sd->flags |= SD_OVERLAP;
6743 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6748 /* Build the groups for the domains */
6749 for_each_cpu(i, cpu_map) {
6750 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6751 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6752 if (sd->flags & SD_OVERLAP) {
6753 if (build_overlap_sched_groups(sd, i))
6756 if (build_sched_groups(sd, i))
6762 /* Calculate CPU capacity for physical packages and nodes */
6763 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6764 if (!cpumask_test_cpu(i, cpu_map))
6767 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6768 claim_allocations(i, sd);
6769 init_sched_groups_capacity(i, sd);
6773 /* Attach the domains */
6775 for_each_cpu(i, cpu_map) {
6776 sd = *per_cpu_ptr(d.sd, i);
6777 cpu_attach_domain(sd, d.rd, i);
6783 __free_domain_allocs(&d, alloc_state, cpu_map);
6787 static cpumask_var_t *doms_cur; /* current sched domains */
6788 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6789 static struct sched_domain_attr *dattr_cur;
6790 /* attribues of custom domains in 'doms_cur' */
6793 * Special case: If a kmalloc of a doms_cur partition (array of
6794 * cpumask) fails, then fallback to a single sched domain,
6795 * as determined by the single cpumask fallback_doms.
6797 static cpumask_var_t fallback_doms;
6800 * arch_update_cpu_topology lets virtualized architectures update the
6801 * cpu core maps. It is supposed to return 1 if the topology changed
6802 * or 0 if it stayed the same.
6804 int __weak arch_update_cpu_topology(void)
6809 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6812 cpumask_var_t *doms;
6814 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6817 for (i = 0; i < ndoms; i++) {
6818 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6819 free_sched_domains(doms, i);
6826 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6829 for (i = 0; i < ndoms; i++)
6830 free_cpumask_var(doms[i]);
6835 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6836 * For now this just excludes isolated cpus, but could be used to
6837 * exclude other special cases in the future.
6839 static int init_sched_domains(const struct cpumask *cpu_map)
6843 arch_update_cpu_topology();
6845 doms_cur = alloc_sched_domains(ndoms_cur);
6847 doms_cur = &fallback_doms;
6848 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6849 err = build_sched_domains(doms_cur[0], NULL);
6850 register_sched_domain_sysctl();
6856 * Detach sched domains from a group of cpus specified in cpu_map
6857 * These cpus will now be attached to the NULL domain
6859 static void detach_destroy_domains(const struct cpumask *cpu_map)
6864 for_each_cpu(i, cpu_map)
6865 cpu_attach_domain(NULL, &def_root_domain, i);
6869 /* handle null as "default" */
6870 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6871 struct sched_domain_attr *new, int idx_new)
6873 struct sched_domain_attr tmp;
6880 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6881 new ? (new + idx_new) : &tmp,
6882 sizeof(struct sched_domain_attr));
6886 * Partition sched domains as specified by the 'ndoms_new'
6887 * cpumasks in the array doms_new[] of cpumasks. This compares
6888 * doms_new[] to the current sched domain partitioning, doms_cur[].
6889 * It destroys each deleted domain and builds each new domain.
6891 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6892 * The masks don't intersect (don't overlap.) We should setup one
6893 * sched domain for each mask. CPUs not in any of the cpumasks will
6894 * not be load balanced. If the same cpumask appears both in the
6895 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6898 * The passed in 'doms_new' should be allocated using
6899 * alloc_sched_domains. This routine takes ownership of it and will
6900 * free_sched_domains it when done with it. If the caller failed the
6901 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6902 * and partition_sched_domains() will fallback to the single partition
6903 * 'fallback_doms', it also forces the domains to be rebuilt.
6905 * If doms_new == NULL it will be replaced with cpu_online_mask.
6906 * ndoms_new == 0 is a special case for destroying existing domains,
6907 * and it will not create the default domain.
6909 * Call with hotplug lock held
6911 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6912 struct sched_domain_attr *dattr_new)
6917 mutex_lock(&sched_domains_mutex);
6919 /* always unregister in case we don't destroy any domains */
6920 unregister_sched_domain_sysctl();
6922 /* Let architecture update cpu core mappings. */
6923 new_topology = arch_update_cpu_topology();
6925 n = doms_new ? ndoms_new : 0;
6927 /* Destroy deleted domains */
6928 for (i = 0; i < ndoms_cur; i++) {
6929 for (j = 0; j < n && !new_topology; j++) {
6930 if (cpumask_equal(doms_cur[i], doms_new[j])
6931 && dattrs_equal(dattr_cur, i, dattr_new, j))
6934 /* no match - a current sched domain not in new doms_new[] */
6935 detach_destroy_domains(doms_cur[i]);
6941 if (doms_new == NULL) {
6943 doms_new = &fallback_doms;
6944 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6945 WARN_ON_ONCE(dattr_new);
6948 /* Build new domains */
6949 for (i = 0; i < ndoms_new; i++) {
6950 for (j = 0; j < n && !new_topology; j++) {
6951 if (cpumask_equal(doms_new[i], doms_cur[j])
6952 && dattrs_equal(dattr_new, i, dattr_cur, j))
6955 /* no match - add a new doms_new */
6956 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6961 /* Remember the new sched domains */
6962 if (doms_cur != &fallback_doms)
6963 free_sched_domains(doms_cur, ndoms_cur);
6964 kfree(dattr_cur); /* kfree(NULL) is safe */
6965 doms_cur = doms_new;
6966 dattr_cur = dattr_new;
6967 ndoms_cur = ndoms_new;
6969 register_sched_domain_sysctl();
6971 mutex_unlock(&sched_domains_mutex);
6974 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6977 * Update cpusets according to cpu_active mask. If cpusets are
6978 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6979 * around partition_sched_domains().
6981 * If we come here as part of a suspend/resume, don't touch cpusets because we
6982 * want to restore it back to its original state upon resume anyway.
6984 static void cpuset_cpu_active(void)
6986 if (cpuhp_tasks_frozen) {
6988 * num_cpus_frozen tracks how many CPUs are involved in suspend
6989 * resume sequence. As long as this is not the last online
6990 * operation in the resume sequence, just build a single sched
6991 * domain, ignoring cpusets.
6994 if (likely(num_cpus_frozen)) {
6995 partition_sched_domains(1, NULL, NULL);
6999 * This is the last CPU online operation. So fall through and
7000 * restore the original sched domains by considering the
7001 * cpuset configurations.
7004 cpuset_update_active_cpus(true);
7007 static int cpuset_cpu_inactive(unsigned int cpu)
7009 unsigned long flags;
7014 if (!cpuhp_tasks_frozen) {
7015 rcu_read_lock_sched();
7016 dl_b = dl_bw_of(cpu);
7018 raw_spin_lock_irqsave(&dl_b->lock, flags);
7019 cpus = dl_bw_cpus(cpu);
7020 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7021 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7023 rcu_read_unlock_sched();
7027 cpuset_update_active_cpus(false);
7030 partition_sched_domains(1, NULL, NULL);
7035 int sched_cpu_activate(unsigned int cpu)
7037 struct rq *rq = cpu_rq(cpu);
7038 unsigned long flags;
7040 set_cpu_active(cpu, true);
7042 if (sched_smp_initialized) {
7043 sched_domains_numa_masks_set(cpu);
7044 cpuset_cpu_active();
7048 * Put the rq online, if not already. This happens:
7050 * 1) In the early boot process, because we build the real domains
7051 * after all cpus have been brought up.
7053 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7056 raw_spin_lock_irqsave(&rq->lock, flags);
7058 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7061 raw_spin_unlock_irqrestore(&rq->lock, flags);
7063 update_max_interval();
7068 int sched_cpu_deactivate(unsigned int cpu)
7072 set_cpu_active(cpu, false);
7074 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7075 * users of this state to go away such that all new such users will
7078 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7079 * not imply sync_sched(), so wait for both.
7081 * Do sync before park smpboot threads to take care the rcu boost case.
7083 if (IS_ENABLED(CONFIG_PREEMPT))
7084 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7088 if (!sched_smp_initialized)
7091 ret = cpuset_cpu_inactive(cpu);
7093 set_cpu_active(cpu, true);
7096 sched_domains_numa_masks_clear(cpu);
7100 static void sched_rq_cpu_starting(unsigned int cpu)
7102 struct rq *rq = cpu_rq(cpu);
7104 rq->calc_load_update = calc_load_update;
7105 account_reset_rq(rq);
7106 update_max_interval();
7109 int sched_cpu_starting(unsigned int cpu)
7111 set_cpu_rq_start_time(cpu);
7112 sched_rq_cpu_starting(cpu);
7116 #ifdef CONFIG_HOTPLUG_CPU
7117 int sched_cpu_dying(unsigned int cpu)
7119 struct rq *rq = cpu_rq(cpu);
7120 unsigned long flags;
7122 /* Handle pending wakeups and then migrate everything off */
7123 sched_ttwu_pending();
7124 raw_spin_lock_irqsave(&rq->lock, flags);
7126 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7130 BUG_ON(rq->nr_running != 1);
7131 raw_spin_unlock_irqrestore(&rq->lock, flags);
7132 calc_load_migrate(rq);
7133 update_max_interval();
7138 void __init sched_init_smp(void)
7140 cpumask_var_t non_isolated_cpus;
7142 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7143 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7148 * There's no userspace yet to cause hotplug operations; hence all the
7149 * cpu masks are stable and all blatant races in the below code cannot
7152 mutex_lock(&sched_domains_mutex);
7153 init_sched_domains(cpu_active_mask);
7154 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7155 if (cpumask_empty(non_isolated_cpus))
7156 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7157 mutex_unlock(&sched_domains_mutex);
7161 /* Move init over to a non-isolated CPU */
7162 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7164 sched_init_granularity();
7165 free_cpumask_var(non_isolated_cpus);
7167 init_sched_rt_class();
7168 init_sched_dl_class();
7169 sched_smp_initialized = true;
7172 static int __init migration_init(void)
7174 sched_rq_cpu_starting(smp_processor_id());
7177 early_initcall(migration_init);
7180 void __init sched_init_smp(void)
7182 sched_init_granularity();
7184 #endif /* CONFIG_SMP */
7186 int in_sched_functions(unsigned long addr)
7188 return in_lock_functions(addr) ||
7189 (addr >= (unsigned long)__sched_text_start
7190 && addr < (unsigned long)__sched_text_end);
7193 #ifdef CONFIG_CGROUP_SCHED
7195 * Default task group.
7196 * Every task in system belongs to this group at bootup.
7198 struct task_group root_task_group;
7199 LIST_HEAD(task_groups);
7201 /* Cacheline aligned slab cache for task_group */
7202 static struct kmem_cache *task_group_cache __read_mostly;
7205 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7207 void __init sched_init(void)
7210 unsigned long alloc_size = 0, ptr;
7212 #ifdef CONFIG_FAIR_GROUP_SCHED
7213 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7215 #ifdef CONFIG_RT_GROUP_SCHED
7216 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7219 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7221 #ifdef CONFIG_FAIR_GROUP_SCHED
7222 root_task_group.se = (struct sched_entity **)ptr;
7223 ptr += nr_cpu_ids * sizeof(void **);
7225 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7226 ptr += nr_cpu_ids * sizeof(void **);
7228 #endif /* CONFIG_FAIR_GROUP_SCHED */
7229 #ifdef CONFIG_RT_GROUP_SCHED
7230 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7231 ptr += nr_cpu_ids * sizeof(void **);
7233 root_task_group.rt_rq = (struct rt_rq **)ptr;
7234 ptr += nr_cpu_ids * sizeof(void **);
7236 #endif /* CONFIG_RT_GROUP_SCHED */
7238 #ifdef CONFIG_CPUMASK_OFFSTACK
7239 for_each_possible_cpu(i) {
7240 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7241 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7243 #endif /* CONFIG_CPUMASK_OFFSTACK */
7245 init_rt_bandwidth(&def_rt_bandwidth,
7246 global_rt_period(), global_rt_runtime());
7247 init_dl_bandwidth(&def_dl_bandwidth,
7248 global_rt_period(), global_rt_runtime());
7251 init_defrootdomain();
7254 #ifdef CONFIG_RT_GROUP_SCHED
7255 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7256 global_rt_period(), global_rt_runtime());
7257 #endif /* CONFIG_RT_GROUP_SCHED */
7259 #ifdef CONFIG_CGROUP_SCHED
7260 task_group_cache = KMEM_CACHE(task_group, 0);
7262 list_add(&root_task_group.list, &task_groups);
7263 INIT_LIST_HEAD(&root_task_group.children);
7264 INIT_LIST_HEAD(&root_task_group.siblings);
7265 autogroup_init(&init_task);
7266 #endif /* CONFIG_CGROUP_SCHED */
7268 for_each_possible_cpu(i) {
7272 raw_spin_lock_init(&rq->lock);
7274 rq->calc_load_active = 0;
7275 rq->calc_load_update = jiffies + LOAD_FREQ;
7276 init_cfs_rq(&rq->cfs);
7277 init_rt_rq(&rq->rt);
7278 init_dl_rq(&rq->dl);
7279 #ifdef CONFIG_FAIR_GROUP_SCHED
7280 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7281 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7283 * How much cpu bandwidth does root_task_group get?
7285 * In case of task-groups formed thr' the cgroup filesystem, it
7286 * gets 100% of the cpu resources in the system. This overall
7287 * system cpu resource is divided among the tasks of
7288 * root_task_group and its child task-groups in a fair manner,
7289 * based on each entity's (task or task-group's) weight
7290 * (se->load.weight).
7292 * In other words, if root_task_group has 10 tasks of weight
7293 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7294 * then A0's share of the cpu resource is:
7296 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7298 * We achieve this by letting root_task_group's tasks sit
7299 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7301 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7302 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7303 #endif /* CONFIG_FAIR_GROUP_SCHED */
7305 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7306 #ifdef CONFIG_RT_GROUP_SCHED
7307 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7310 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7311 rq->cpu_load[j] = 0;
7313 rq->last_load_update_tick = jiffies;
7318 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7319 rq->balance_callback = NULL;
7320 rq->active_balance = 0;
7321 rq->next_balance = jiffies;
7326 rq->avg_idle = 2*sysctl_sched_migration_cost;
7327 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7329 INIT_LIST_HEAD(&rq->cfs_tasks);
7331 rq_attach_root(rq, &def_root_domain);
7332 #ifdef CONFIG_NO_HZ_COMMON
7335 #ifdef CONFIG_NO_HZ_FULL
7336 rq->last_sched_tick = 0;
7340 atomic_set(&rq->nr_iowait, 0);
7343 set_load_weight(&init_task);
7345 #ifdef CONFIG_PREEMPT_NOTIFIERS
7346 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7350 * The boot idle thread does lazy MMU switching as well:
7352 atomic_inc(&init_mm.mm_count);
7353 enter_lazy_tlb(&init_mm, current);
7356 * During early bootup we pretend to be a normal task:
7358 current->sched_class = &fair_sched_class;
7361 * Make us the idle thread. Technically, schedule() should not be
7362 * called from this thread, however somewhere below it might be,
7363 * but because we are the idle thread, we just pick up running again
7364 * when this runqueue becomes "idle".
7366 init_idle(current, smp_processor_id());
7368 calc_load_update = jiffies + LOAD_FREQ;
7371 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7372 /* May be allocated at isolcpus cmdline parse time */
7373 if (cpu_isolated_map == NULL)
7374 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7375 idle_thread_set_boot_cpu();
7376 set_cpu_rq_start_time(smp_processor_id());
7378 init_sched_fair_class();
7380 scheduler_running = 1;
7383 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7384 static inline int preempt_count_equals(int preempt_offset)
7386 int nested = preempt_count() + rcu_preempt_depth();
7388 return (nested == preempt_offset);
7391 void __might_sleep(const char *file, int line, int preempt_offset)
7394 * Blocking primitives will set (and therefore destroy) current->state,
7395 * since we will exit with TASK_RUNNING make sure we enter with it,
7396 * otherwise we will destroy state.
7398 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7399 "do not call blocking ops when !TASK_RUNNING; "
7400 "state=%lx set at [<%p>] %pS\n",
7402 (void *)current->task_state_change,
7403 (void *)current->task_state_change);
7405 ___might_sleep(file, line, preempt_offset);
7407 EXPORT_SYMBOL(__might_sleep);
7409 void ___might_sleep(const char *file, int line, int preempt_offset)
7411 static unsigned long prev_jiffy; /* ratelimiting */
7413 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7414 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7415 !is_idle_task(current)) ||
7416 system_state != SYSTEM_RUNNING || oops_in_progress)
7418 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7420 prev_jiffy = jiffies;
7423 "BUG: sleeping function called from invalid context at %s:%d\n",
7426 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7427 in_atomic(), irqs_disabled(),
7428 current->pid, current->comm);
7430 if (task_stack_end_corrupted(current))
7431 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7433 debug_show_held_locks(current);
7434 if (irqs_disabled())
7435 print_irqtrace_events(current);
7436 #ifdef CONFIG_DEBUG_PREEMPT
7437 if (!preempt_count_equals(preempt_offset)) {
7438 pr_err("Preemption disabled at:");
7439 print_ip_sym(current->preempt_disable_ip);
7445 EXPORT_SYMBOL(___might_sleep);
7448 #ifdef CONFIG_MAGIC_SYSRQ
7449 void normalize_rt_tasks(void)
7451 struct task_struct *g, *p;
7452 struct sched_attr attr = {
7453 .sched_policy = SCHED_NORMAL,
7456 read_lock(&tasklist_lock);
7457 for_each_process_thread(g, p) {
7459 * Only normalize user tasks:
7461 if (p->flags & PF_KTHREAD)
7464 p->se.exec_start = 0;
7465 #ifdef CONFIG_SCHEDSTATS
7466 p->se.statistics.wait_start = 0;
7467 p->se.statistics.sleep_start = 0;
7468 p->se.statistics.block_start = 0;
7471 if (!dl_task(p) && !rt_task(p)) {
7473 * Renice negative nice level userspace
7476 if (task_nice(p) < 0)
7477 set_user_nice(p, 0);
7481 __sched_setscheduler(p, &attr, false, false);
7483 read_unlock(&tasklist_lock);
7486 #endif /* CONFIG_MAGIC_SYSRQ */
7488 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7490 * These functions are only useful for the IA64 MCA handling, or kdb.
7492 * They can only be called when the whole system has been
7493 * stopped - every CPU needs to be quiescent, and no scheduling
7494 * activity can take place. Using them for anything else would
7495 * be a serious bug, and as a result, they aren't even visible
7496 * under any other configuration.
7500 * curr_task - return the current task for a given cpu.
7501 * @cpu: the processor in question.
7503 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7505 * Return: The current task for @cpu.
7507 struct task_struct *curr_task(int cpu)
7509 return cpu_curr(cpu);
7512 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7516 * set_curr_task - set the current task for a given cpu.
7517 * @cpu: the processor in question.
7518 * @p: the task pointer to set.
7520 * Description: This function must only be used when non-maskable interrupts
7521 * are serviced on a separate stack. It allows the architecture to switch the
7522 * notion of the current task on a cpu in a non-blocking manner. This function
7523 * must be called with all CPU's synchronized, and interrupts disabled, the
7524 * and caller must save the original value of the current task (see
7525 * curr_task() above) and restore that value before reenabling interrupts and
7526 * re-starting the system.
7528 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7530 void set_curr_task(int cpu, struct task_struct *p)
7537 #ifdef CONFIG_CGROUP_SCHED
7538 /* task_group_lock serializes the addition/removal of task groups */
7539 static DEFINE_SPINLOCK(task_group_lock);
7541 static void sched_free_group(struct task_group *tg)
7543 free_fair_sched_group(tg);
7544 free_rt_sched_group(tg);
7546 kmem_cache_free(task_group_cache, tg);
7549 /* allocate runqueue etc for a new task group */
7550 struct task_group *sched_create_group(struct task_group *parent)
7552 struct task_group *tg;
7554 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7556 return ERR_PTR(-ENOMEM);
7558 if (!alloc_fair_sched_group(tg, parent))
7561 if (!alloc_rt_sched_group(tg, parent))
7567 sched_free_group(tg);
7568 return ERR_PTR(-ENOMEM);
7571 void sched_online_group(struct task_group *tg, struct task_group *parent)
7573 unsigned long flags;
7575 spin_lock_irqsave(&task_group_lock, flags);
7576 list_add_rcu(&tg->list, &task_groups);
7578 WARN_ON(!parent); /* root should already exist */
7580 tg->parent = parent;
7581 INIT_LIST_HEAD(&tg->children);
7582 list_add_rcu(&tg->siblings, &parent->children);
7583 spin_unlock_irqrestore(&task_group_lock, flags);
7586 /* rcu callback to free various structures associated with a task group */
7587 static void sched_free_group_rcu(struct rcu_head *rhp)
7589 /* now it should be safe to free those cfs_rqs */
7590 sched_free_group(container_of(rhp, struct task_group, rcu));
7593 void sched_destroy_group(struct task_group *tg)
7595 /* wait for possible concurrent references to cfs_rqs complete */
7596 call_rcu(&tg->rcu, sched_free_group_rcu);
7599 void sched_offline_group(struct task_group *tg)
7601 unsigned long flags;
7603 /* end participation in shares distribution */
7604 unregister_fair_sched_group(tg);
7606 spin_lock_irqsave(&task_group_lock, flags);
7607 list_del_rcu(&tg->list);
7608 list_del_rcu(&tg->siblings);
7609 spin_unlock_irqrestore(&task_group_lock, flags);
7612 /* change task's runqueue when it moves between groups.
7613 * The caller of this function should have put the task in its new group
7614 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7615 * reflect its new group.
7617 void sched_move_task(struct task_struct *tsk)
7619 struct task_group *tg;
7620 int queued, running;
7621 unsigned long flags;
7624 rq = task_rq_lock(tsk, &flags);
7626 running = task_current(rq, tsk);
7627 queued = task_on_rq_queued(tsk);
7630 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7631 if (unlikely(running))
7632 put_prev_task(rq, tsk);
7635 * All callers are synchronized by task_rq_lock(); we do not use RCU
7636 * which is pointless here. Thus, we pass "true" to task_css_check()
7637 * to prevent lockdep warnings.
7639 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7640 struct task_group, css);
7641 tg = autogroup_task_group(tsk, tg);
7642 tsk->sched_task_group = tg;
7644 #ifdef CONFIG_FAIR_GROUP_SCHED
7645 if (tsk->sched_class->task_move_group)
7646 tsk->sched_class->task_move_group(tsk);
7649 set_task_rq(tsk, task_cpu(tsk));
7651 if (unlikely(running))
7652 tsk->sched_class->set_curr_task(rq);
7654 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7656 task_rq_unlock(rq, tsk, &flags);
7658 #endif /* CONFIG_CGROUP_SCHED */
7660 #ifdef CONFIG_RT_GROUP_SCHED
7662 * Ensure that the real time constraints are schedulable.
7664 static DEFINE_MUTEX(rt_constraints_mutex);
7666 /* Must be called with tasklist_lock held */
7667 static inline int tg_has_rt_tasks(struct task_group *tg)
7669 struct task_struct *g, *p;
7672 * Autogroups do not have RT tasks; see autogroup_create().
7674 if (task_group_is_autogroup(tg))
7677 for_each_process_thread(g, p) {
7678 if (rt_task(p) && task_group(p) == tg)
7685 struct rt_schedulable_data {
7686 struct task_group *tg;
7691 static int tg_rt_schedulable(struct task_group *tg, void *data)
7693 struct rt_schedulable_data *d = data;
7694 struct task_group *child;
7695 unsigned long total, sum = 0;
7696 u64 period, runtime;
7698 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7699 runtime = tg->rt_bandwidth.rt_runtime;
7702 period = d->rt_period;
7703 runtime = d->rt_runtime;
7707 * Cannot have more runtime than the period.
7709 if (runtime > period && runtime != RUNTIME_INF)
7713 * Ensure we don't starve existing RT tasks.
7715 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7718 total = to_ratio(period, runtime);
7721 * Nobody can have more than the global setting allows.
7723 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7727 * The sum of our children's runtime should not exceed our own.
7729 list_for_each_entry_rcu(child, &tg->children, siblings) {
7730 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7731 runtime = child->rt_bandwidth.rt_runtime;
7733 if (child == d->tg) {
7734 period = d->rt_period;
7735 runtime = d->rt_runtime;
7738 sum += to_ratio(period, runtime);
7747 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7751 struct rt_schedulable_data data = {
7753 .rt_period = period,
7754 .rt_runtime = runtime,
7758 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7764 static int tg_set_rt_bandwidth(struct task_group *tg,
7765 u64 rt_period, u64 rt_runtime)
7770 * Disallowing the root group RT runtime is BAD, it would disallow the
7771 * kernel creating (and or operating) RT threads.
7773 if (tg == &root_task_group && rt_runtime == 0)
7776 /* No period doesn't make any sense. */
7780 mutex_lock(&rt_constraints_mutex);
7781 read_lock(&tasklist_lock);
7782 err = __rt_schedulable(tg, rt_period, rt_runtime);
7786 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7787 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7788 tg->rt_bandwidth.rt_runtime = rt_runtime;
7790 for_each_possible_cpu(i) {
7791 struct rt_rq *rt_rq = tg->rt_rq[i];
7793 raw_spin_lock(&rt_rq->rt_runtime_lock);
7794 rt_rq->rt_runtime = rt_runtime;
7795 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7797 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7799 read_unlock(&tasklist_lock);
7800 mutex_unlock(&rt_constraints_mutex);
7805 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7807 u64 rt_runtime, rt_period;
7809 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7810 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7811 if (rt_runtime_us < 0)
7812 rt_runtime = RUNTIME_INF;
7814 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7817 static long sched_group_rt_runtime(struct task_group *tg)
7821 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7824 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7825 do_div(rt_runtime_us, NSEC_PER_USEC);
7826 return rt_runtime_us;
7829 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7831 u64 rt_runtime, rt_period;
7833 rt_period = rt_period_us * NSEC_PER_USEC;
7834 rt_runtime = tg->rt_bandwidth.rt_runtime;
7836 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7839 static long sched_group_rt_period(struct task_group *tg)
7843 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7844 do_div(rt_period_us, NSEC_PER_USEC);
7845 return rt_period_us;
7847 #endif /* CONFIG_RT_GROUP_SCHED */
7849 #ifdef CONFIG_RT_GROUP_SCHED
7850 static int sched_rt_global_constraints(void)
7854 mutex_lock(&rt_constraints_mutex);
7855 read_lock(&tasklist_lock);
7856 ret = __rt_schedulable(NULL, 0, 0);
7857 read_unlock(&tasklist_lock);
7858 mutex_unlock(&rt_constraints_mutex);
7863 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7865 /* Don't accept realtime tasks when there is no way for them to run */
7866 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7872 #else /* !CONFIG_RT_GROUP_SCHED */
7873 static int sched_rt_global_constraints(void)
7875 unsigned long flags;
7878 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7879 for_each_possible_cpu(i) {
7880 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7882 raw_spin_lock(&rt_rq->rt_runtime_lock);
7883 rt_rq->rt_runtime = global_rt_runtime();
7884 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7886 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7890 #endif /* CONFIG_RT_GROUP_SCHED */
7892 static int sched_dl_global_validate(void)
7894 u64 runtime = global_rt_runtime();
7895 u64 period = global_rt_period();
7896 u64 new_bw = to_ratio(period, runtime);
7899 unsigned long flags;
7902 * Here we want to check the bandwidth not being set to some
7903 * value smaller than the currently allocated bandwidth in
7904 * any of the root_domains.
7906 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7907 * cycling on root_domains... Discussion on different/better
7908 * solutions is welcome!
7910 for_each_possible_cpu(cpu) {
7911 rcu_read_lock_sched();
7912 dl_b = dl_bw_of(cpu);
7914 raw_spin_lock_irqsave(&dl_b->lock, flags);
7915 if (new_bw < dl_b->total_bw)
7917 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7919 rcu_read_unlock_sched();
7928 static void sched_dl_do_global(void)
7933 unsigned long flags;
7935 def_dl_bandwidth.dl_period = global_rt_period();
7936 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7938 if (global_rt_runtime() != RUNTIME_INF)
7939 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7942 * FIXME: As above...
7944 for_each_possible_cpu(cpu) {
7945 rcu_read_lock_sched();
7946 dl_b = dl_bw_of(cpu);
7948 raw_spin_lock_irqsave(&dl_b->lock, flags);
7950 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7952 rcu_read_unlock_sched();
7956 static int sched_rt_global_validate(void)
7958 if (sysctl_sched_rt_period <= 0)
7961 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7962 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7968 static void sched_rt_do_global(void)
7970 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7971 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7974 int sched_rt_handler(struct ctl_table *table, int write,
7975 void __user *buffer, size_t *lenp,
7978 int old_period, old_runtime;
7979 static DEFINE_MUTEX(mutex);
7983 old_period = sysctl_sched_rt_period;
7984 old_runtime = sysctl_sched_rt_runtime;
7986 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7988 if (!ret && write) {
7989 ret = sched_rt_global_validate();
7993 ret = sched_dl_global_validate();
7997 ret = sched_rt_global_constraints();
8001 sched_rt_do_global();
8002 sched_dl_do_global();
8006 sysctl_sched_rt_period = old_period;
8007 sysctl_sched_rt_runtime = old_runtime;
8009 mutex_unlock(&mutex);
8014 int sched_rr_handler(struct ctl_table *table, int write,
8015 void __user *buffer, size_t *lenp,
8019 static DEFINE_MUTEX(mutex);
8022 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8023 /* make sure that internally we keep jiffies */
8024 /* also, writing zero resets timeslice to default */
8025 if (!ret && write) {
8026 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8027 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8029 mutex_unlock(&mutex);
8033 #ifdef CONFIG_CGROUP_SCHED
8035 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8037 return css ? container_of(css, struct task_group, css) : NULL;
8040 static struct cgroup_subsys_state *
8041 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8043 struct task_group *parent = css_tg(parent_css);
8044 struct task_group *tg;
8047 /* This is early initialization for the top cgroup */
8048 return &root_task_group.css;
8051 tg = sched_create_group(parent);
8053 return ERR_PTR(-ENOMEM);
8055 sched_online_group(tg, parent);
8060 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8062 struct task_group *tg = css_tg(css);
8064 sched_offline_group(tg);
8067 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8069 struct task_group *tg = css_tg(css);
8072 * Relies on the RCU grace period between css_released() and this.
8074 sched_free_group(tg);
8077 static void cpu_cgroup_fork(struct task_struct *task)
8079 sched_move_task(task);
8082 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8084 struct task_struct *task;
8085 struct cgroup_subsys_state *css;
8087 cgroup_taskset_for_each(task, css, tset) {
8088 #ifdef CONFIG_RT_GROUP_SCHED
8089 if (!sched_rt_can_attach(css_tg(css), task))
8092 /* We don't support RT-tasks being in separate groups */
8093 if (task->sched_class != &fair_sched_class)
8100 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8102 struct task_struct *task;
8103 struct cgroup_subsys_state *css;
8105 cgroup_taskset_for_each(task, css, tset)
8106 sched_move_task(task);
8109 #ifdef CONFIG_FAIR_GROUP_SCHED
8110 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8111 struct cftype *cftype, u64 shareval)
8113 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8116 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8119 struct task_group *tg = css_tg(css);
8121 return (u64) scale_load_down(tg->shares);
8124 #ifdef CONFIG_CFS_BANDWIDTH
8125 static DEFINE_MUTEX(cfs_constraints_mutex);
8127 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8128 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8130 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8132 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8134 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8135 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8137 if (tg == &root_task_group)
8141 * Ensure we have at some amount of bandwidth every period. This is
8142 * to prevent reaching a state of large arrears when throttled via
8143 * entity_tick() resulting in prolonged exit starvation.
8145 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8149 * Likewise, bound things on the otherside by preventing insane quota
8150 * periods. This also allows us to normalize in computing quota
8153 if (period > max_cfs_quota_period)
8157 * Prevent race between setting of cfs_rq->runtime_enabled and
8158 * unthrottle_offline_cfs_rqs().
8161 mutex_lock(&cfs_constraints_mutex);
8162 ret = __cfs_schedulable(tg, period, quota);
8166 runtime_enabled = quota != RUNTIME_INF;
8167 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8169 * If we need to toggle cfs_bandwidth_used, off->on must occur
8170 * before making related changes, and on->off must occur afterwards
8172 if (runtime_enabled && !runtime_was_enabled)
8173 cfs_bandwidth_usage_inc();
8174 raw_spin_lock_irq(&cfs_b->lock);
8175 cfs_b->period = ns_to_ktime(period);
8176 cfs_b->quota = quota;
8178 __refill_cfs_bandwidth_runtime(cfs_b);
8179 /* restart the period timer (if active) to handle new period expiry */
8180 if (runtime_enabled)
8181 start_cfs_bandwidth(cfs_b);
8182 raw_spin_unlock_irq(&cfs_b->lock);
8184 for_each_online_cpu(i) {
8185 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8186 struct rq *rq = cfs_rq->rq;
8188 raw_spin_lock_irq(&rq->lock);
8189 cfs_rq->runtime_enabled = runtime_enabled;
8190 cfs_rq->runtime_remaining = 0;
8192 if (cfs_rq->throttled)
8193 unthrottle_cfs_rq(cfs_rq);
8194 raw_spin_unlock_irq(&rq->lock);
8196 if (runtime_was_enabled && !runtime_enabled)
8197 cfs_bandwidth_usage_dec();
8199 mutex_unlock(&cfs_constraints_mutex);
8205 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8209 period = ktime_to_ns(tg->cfs_bandwidth.period);
8210 if (cfs_quota_us < 0)
8211 quota = RUNTIME_INF;
8213 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8215 return tg_set_cfs_bandwidth(tg, period, quota);
8218 long tg_get_cfs_quota(struct task_group *tg)
8222 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8225 quota_us = tg->cfs_bandwidth.quota;
8226 do_div(quota_us, NSEC_PER_USEC);
8231 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8235 period = (u64)cfs_period_us * NSEC_PER_USEC;
8236 quota = tg->cfs_bandwidth.quota;
8238 return tg_set_cfs_bandwidth(tg, period, quota);
8241 long tg_get_cfs_period(struct task_group *tg)
8245 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8246 do_div(cfs_period_us, NSEC_PER_USEC);
8248 return cfs_period_us;
8251 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8254 return tg_get_cfs_quota(css_tg(css));
8257 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8258 struct cftype *cftype, s64 cfs_quota_us)
8260 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8263 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8266 return tg_get_cfs_period(css_tg(css));
8269 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8270 struct cftype *cftype, u64 cfs_period_us)
8272 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8275 struct cfs_schedulable_data {
8276 struct task_group *tg;
8281 * normalize group quota/period to be quota/max_period
8282 * note: units are usecs
8284 static u64 normalize_cfs_quota(struct task_group *tg,
8285 struct cfs_schedulable_data *d)
8293 period = tg_get_cfs_period(tg);
8294 quota = tg_get_cfs_quota(tg);
8297 /* note: these should typically be equivalent */
8298 if (quota == RUNTIME_INF || quota == -1)
8301 return to_ratio(period, quota);
8304 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8306 struct cfs_schedulable_data *d = data;
8307 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8308 s64 quota = 0, parent_quota = -1;
8311 quota = RUNTIME_INF;
8313 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8315 quota = normalize_cfs_quota(tg, d);
8316 parent_quota = parent_b->hierarchical_quota;
8319 * ensure max(child_quota) <= parent_quota, inherit when no
8322 if (quota == RUNTIME_INF)
8323 quota = parent_quota;
8324 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8327 cfs_b->hierarchical_quota = quota;
8332 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8335 struct cfs_schedulable_data data = {
8341 if (quota != RUNTIME_INF) {
8342 do_div(data.period, NSEC_PER_USEC);
8343 do_div(data.quota, NSEC_PER_USEC);
8347 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8353 static int cpu_stats_show(struct seq_file *sf, void *v)
8355 struct task_group *tg = css_tg(seq_css(sf));
8356 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8358 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8359 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8360 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8364 #endif /* CONFIG_CFS_BANDWIDTH */
8365 #endif /* CONFIG_FAIR_GROUP_SCHED */
8367 #ifdef CONFIG_RT_GROUP_SCHED
8368 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8369 struct cftype *cft, s64 val)
8371 return sched_group_set_rt_runtime(css_tg(css), val);
8374 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8377 return sched_group_rt_runtime(css_tg(css));
8380 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8381 struct cftype *cftype, u64 rt_period_us)
8383 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8386 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8389 return sched_group_rt_period(css_tg(css));
8391 #endif /* CONFIG_RT_GROUP_SCHED */
8393 static struct cftype cpu_files[] = {
8394 #ifdef CONFIG_FAIR_GROUP_SCHED
8397 .read_u64 = cpu_shares_read_u64,
8398 .write_u64 = cpu_shares_write_u64,
8401 #ifdef CONFIG_CFS_BANDWIDTH
8403 .name = "cfs_quota_us",
8404 .read_s64 = cpu_cfs_quota_read_s64,
8405 .write_s64 = cpu_cfs_quota_write_s64,
8408 .name = "cfs_period_us",
8409 .read_u64 = cpu_cfs_period_read_u64,
8410 .write_u64 = cpu_cfs_period_write_u64,
8414 .seq_show = cpu_stats_show,
8417 #ifdef CONFIG_RT_GROUP_SCHED
8419 .name = "rt_runtime_us",
8420 .read_s64 = cpu_rt_runtime_read,
8421 .write_s64 = cpu_rt_runtime_write,
8424 .name = "rt_period_us",
8425 .read_u64 = cpu_rt_period_read_uint,
8426 .write_u64 = cpu_rt_period_write_uint,
8432 struct cgroup_subsys cpu_cgrp_subsys = {
8433 .css_alloc = cpu_cgroup_css_alloc,
8434 .css_released = cpu_cgroup_css_released,
8435 .css_free = cpu_cgroup_css_free,
8436 .fork = cpu_cgroup_fork,
8437 .can_attach = cpu_cgroup_can_attach,
8438 .attach = cpu_cgroup_attach,
8439 .legacy_cftypes = cpu_files,
8443 #endif /* CONFIG_CGROUP_SCHED */
8445 void dump_cpu_task(int cpu)
8447 pr_info("Task dump for CPU %d:\n", cpu);
8448 sched_show_task(cpu_curr(cpu));
8452 * Nice levels are multiplicative, with a gentle 10% change for every
8453 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8454 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8455 * that remained on nice 0.
8457 * The "10% effect" is relative and cumulative: from _any_ nice level,
8458 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8459 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8460 * If a task goes up by ~10% and another task goes down by ~10% then
8461 * the relative distance between them is ~25%.)
8463 const int sched_prio_to_weight[40] = {
8464 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8465 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8466 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8467 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8468 /* 0 */ 1024, 820, 655, 526, 423,
8469 /* 5 */ 335, 272, 215, 172, 137,
8470 /* 10 */ 110, 87, 70, 56, 45,
8471 /* 15 */ 36, 29, 23, 18, 15,
8475 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8477 * In cases where the weight does not change often, we can use the
8478 * precalculated inverse to speed up arithmetics by turning divisions
8479 * into multiplications:
8481 const u32 sched_prio_to_wmult[40] = {
8482 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8483 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8484 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8485 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8486 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8487 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8488 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8489 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,