4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
97 static void update_rq_clock_task(struct rq *rq, s64 delta);
99 void update_rq_clock(struct rq *rq)
103 lockdep_assert_held(&rq->lock);
105 if (rq->clock_skip_update & RQCF_ACT_SKIP)
108 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
112 update_rq_clock_task(rq, delta);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug unsigned int sysctl_sched_features =
123 #include "features.h"
129 * Number of tasks to iterate in a single balance run.
130 * Limited because this is done with IRQs disabled.
132 const_debug unsigned int sysctl_sched_nr_migrate = 32;
135 * period over which we average the RT time consumption, measured
140 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
143 * period over which we measure -rt task cpu usage in us.
146 unsigned int sysctl_sched_rt_period = 1000000;
148 __read_mostly int scheduler_running;
151 * part of the period that we allow rt tasks to run in us.
154 int sysctl_sched_rt_runtime = 950000;
156 /* cpus with isolated domains */
157 cpumask_var_t cpu_isolated_map;
160 * this_rq_lock - lock this runqueue and disable interrupts.
162 static struct rq *this_rq_lock(void)
169 raw_spin_lock(&rq->lock);
175 * __task_rq_lock - lock the rq @p resides on.
177 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
182 lockdep_assert_held(&p->pi_lock);
186 raw_spin_lock(&rq->lock);
187 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
188 rf->cookie = lockdep_pin_lock(&rq->lock);
191 raw_spin_unlock(&rq->lock);
193 while (unlikely(task_on_rq_migrating(p)))
199 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
201 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
202 __acquires(p->pi_lock)
208 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
210 raw_spin_lock(&rq->lock);
212 * move_queued_task() task_rq_lock()
215 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
216 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
217 * [S] ->cpu = new_cpu [L] task_rq()
221 * If we observe the old cpu in task_rq_lock, the acquire of
222 * the old rq->lock will fully serialize against the stores.
224 * If we observe the new cpu in task_rq_lock, the acquire will
225 * pair with the WMB to ensure we must then also see migrating.
227 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
228 rf->cookie = lockdep_pin_lock(&rq->lock);
231 raw_spin_unlock(&rq->lock);
232 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
234 while (unlikely(task_on_rq_migrating(p)))
239 #ifdef CONFIG_SCHED_HRTICK
241 * Use HR-timers to deliver accurate preemption points.
244 static void hrtick_clear(struct rq *rq)
246 if (hrtimer_active(&rq->hrtick_timer))
247 hrtimer_cancel(&rq->hrtick_timer);
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
254 static enum hrtimer_restart hrtick(struct hrtimer *timer)
256 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
258 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
260 raw_spin_lock(&rq->lock);
262 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
263 raw_spin_unlock(&rq->lock);
265 return HRTIMER_NORESTART;
270 static void __hrtick_restart(struct rq *rq)
272 struct hrtimer *timer = &rq->hrtick_timer;
274 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
278 * called from hardirq (IPI) context
280 static void __hrtick_start(void *arg)
284 raw_spin_lock(&rq->lock);
285 __hrtick_restart(rq);
286 rq->hrtick_csd_pending = 0;
287 raw_spin_unlock(&rq->lock);
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq *rq, u64 delay)
297 struct hrtimer *timer = &rq->hrtick_timer;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta = max_t(s64, delay, 10000LL);
306 time = ktime_add_ns(timer->base->get_time(), delta);
308 hrtimer_set_expires(timer, time);
310 if (rq == this_rq()) {
311 __hrtick_restart(rq);
312 } else if (!rq->hrtick_csd_pending) {
313 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
314 rq->hrtick_csd_pending = 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq *rq, u64 delay)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay = max_t(u64, delay, 10000LL);
331 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
332 HRTIMER_MODE_REL_PINNED);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq *rq)
339 rq->hrtick_csd_pending = 0;
341 rq->hrtick_csd.flags = 0;
342 rq->hrtick_csd.func = __hrtick_start;
343 rq->hrtick_csd.info = rq;
346 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
347 rq->hrtick_timer.function = hrtick;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq *rq)
354 static inline void init_rq_hrtick(struct rq *rq)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct *p)
385 struct thread_info *ti = task_thread_info(p);
386 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct *p)
397 struct thread_info *ti = task_thread_info(p);
398 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
401 if (!(val & _TIF_POLLING_NRFLAG))
403 if (val & _TIF_NEED_RESCHED)
405 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
414 static bool set_nr_and_not_polling(struct task_struct *p)
416 set_tsk_need_resched(p);
421 static bool set_nr_if_polling(struct task_struct *p)
428 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
430 struct wake_q_node *node = &task->wake_q;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
443 get_task_struct(task);
446 * The head is context local, there can be no concurrency.
449 head->lastp = &node->next;
452 void wake_up_q(struct wake_q_head *head)
454 struct wake_q_node *node = head->first;
456 while (node != WAKE_Q_TAIL) {
457 struct task_struct *task;
459 task = container_of(node, struct task_struct, wake_q);
461 /* task can safely be re-inserted now */
463 task->wake_q.next = NULL;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task);
470 put_task_struct(task);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq *rq)
483 struct task_struct *curr = rq->curr;
486 lockdep_assert_held(&rq->lock);
488 if (test_tsk_need_resched(curr))
493 if (cpu == smp_processor_id()) {
494 set_tsk_need_resched(curr);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr))
500 smp_send_reschedule(cpu);
502 trace_sched_wake_idle_without_ipi(cpu);
505 void resched_cpu(int cpu)
507 struct rq *rq = cpu_rq(cpu);
510 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
513 raw_spin_unlock_irqrestore(&rq->lock, flags);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i, cpu = smp_processor_id();
529 struct sched_domain *sd;
531 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
535 for_each_domain(cpu, sd) {
536 for_each_cpu(i, sched_domain_span(sd)) {
540 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
547 if (!is_housekeeping_cpu(cpu))
548 cpu = housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu)
565 struct rq *rq = cpu_rq(cpu);
567 if (cpu == smp_processor_id())
570 if (set_nr_and_not_polling(rq->idle))
571 smp_send_reschedule(cpu);
573 trace_sched_wake_idle_without_ipi(cpu);
576 static bool wake_up_full_nohz_cpu(int cpu)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (cpu_is_offline(cpu))
585 return true; /* Don't try to wake offline CPUs. */
586 if (tick_nohz_full_cpu(cpu)) {
587 if (cpu != smp_processor_id() ||
588 tick_nohz_tick_stopped())
589 tick_nohz_full_kick_cpu(cpu);
597 * Wake up the specified CPU. If the CPU is going offline, it is the
598 * caller's responsibility to deal with the lost wakeup, for example,
599 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
601 void wake_up_nohz_cpu(int cpu)
603 if (!wake_up_full_nohz_cpu(cpu))
604 wake_up_idle_cpu(cpu);
607 static inline bool got_nohz_idle_kick(void)
609 int cpu = smp_processor_id();
611 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
614 if (idle_cpu(cpu) && !need_resched())
618 * We can't run Idle Load Balance on this CPU for this time so we
619 * cancel it and clear NOHZ_BALANCE_KICK
621 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
625 #else /* CONFIG_NO_HZ_COMMON */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ_COMMON */
634 #ifdef CONFIG_NO_HZ_FULL
635 bool sched_can_stop_tick(struct rq *rq)
639 /* Deadline tasks, even if single, need the tick */
640 if (rq->dl.dl_nr_running)
644 * If there are more than one RR tasks, we need the tick to effect the
645 * actual RR behaviour.
647 if (rq->rt.rr_nr_running) {
648 if (rq->rt.rr_nr_running == 1)
655 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 * forced preemption between FIFO tasks.
658 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
663 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 * if there's more than one we need the tick for involuntary
667 if (rq->nr_running > 1)
672 #endif /* CONFIG_NO_HZ_FULL */
674 void sched_avg_update(struct rq *rq)
676 s64 period = sched_avg_period();
678 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
680 * Inline assembly required to prevent the compiler
681 * optimising this loop into a divmod call.
682 * See __iter_div_u64_rem() for another example of this.
684 asm("" : "+rm" (rq->age_stamp));
685 rq->age_stamp += period;
690 #endif /* CONFIG_SMP */
692 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
695 * Iterate task_group tree rooted at *from, calling @down when first entering a
696 * node and @up when leaving it for the final time.
698 * Caller must hold rcu_lock or sufficient equivalent.
700 int walk_tg_tree_from(struct task_group *from,
701 tg_visitor down, tg_visitor up, void *data)
703 struct task_group *parent, *child;
709 ret = (*down)(parent, data);
712 list_for_each_entry_rcu(child, &parent->children, siblings) {
719 ret = (*up)(parent, data);
720 if (ret || parent == from)
724 parent = parent->parent;
731 int tg_nop(struct task_group *tg, void *data)
737 static void set_load_weight(struct task_struct *p)
739 int prio = p->static_prio - MAX_RT_PRIO;
740 struct load_weight *load = &p->se.load;
743 * SCHED_IDLE tasks get minimal weight:
745 if (idle_policy(p->policy)) {
746 load->weight = scale_load(WEIGHT_IDLEPRIO);
747 load->inv_weight = WMULT_IDLEPRIO;
751 load->weight = scale_load(sched_prio_to_weight[prio]);
752 load->inv_weight = sched_prio_to_wmult[prio];
755 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
758 if (!(flags & ENQUEUE_RESTORE))
759 sched_info_queued(rq, p);
760 p->sched_class->enqueue_task(rq, p, flags);
763 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
766 if (!(flags & DEQUEUE_SAVE))
767 sched_info_dequeued(rq, p);
768 p->sched_class->dequeue_task(rq, p, flags);
771 void activate_task(struct rq *rq, struct task_struct *p, int flags)
773 if (task_contributes_to_load(p))
774 rq->nr_uninterruptible--;
776 enqueue_task(rq, p, flags);
779 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible++;
784 dequeue_task(rq, p, flags);
787 static void update_rq_clock_task(struct rq *rq, s64 delta)
790 * In theory, the compile should just see 0 here, and optimize out the call
791 * to sched_rt_avg_update. But I don't trust it...
793 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
794 s64 steal = 0, irq_delta = 0;
796 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
797 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
800 * Since irq_time is only updated on {soft,}irq_exit, we might run into
801 * this case when a previous update_rq_clock() happened inside a
804 * When this happens, we stop ->clock_task and only update the
805 * prev_irq_time stamp to account for the part that fit, so that a next
806 * update will consume the rest. This ensures ->clock_task is
809 * It does however cause some slight miss-attribution of {soft,}irq
810 * time, a more accurate solution would be to update the irq_time using
811 * the current rq->clock timestamp, except that would require using
814 if (irq_delta > delta)
817 rq->prev_irq_time += irq_delta;
820 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
821 if (static_key_false((¶virt_steal_rq_enabled))) {
822 steal = paravirt_steal_clock(cpu_of(rq));
823 steal -= rq->prev_steal_time_rq;
825 if (unlikely(steal > delta))
828 rq->prev_steal_time_rq += steal;
833 rq->clock_task += delta;
835 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
836 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
837 sched_rt_avg_update(rq, irq_delta + steal);
841 void sched_set_stop_task(int cpu, struct task_struct *stop)
843 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
844 struct task_struct *old_stop = cpu_rq(cpu)->stop;
848 * Make it appear like a SCHED_FIFO task, its something
849 * userspace knows about and won't get confused about.
851 * Also, it will make PI more or less work without too
852 * much confusion -- but then, stop work should not
853 * rely on PI working anyway.
855 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
857 stop->sched_class = &stop_sched_class;
860 cpu_rq(cpu)->stop = stop;
864 * Reset it back to a normal scheduling class so that
865 * it can die in pieces.
867 old_stop->sched_class = &rt_sched_class;
872 * __normal_prio - return the priority that is based on the static prio
874 static inline int __normal_prio(struct task_struct *p)
876 return p->static_prio;
880 * Calculate the expected normal priority: i.e. priority
881 * without taking RT-inheritance into account. Might be
882 * boosted by interactivity modifiers. Changes upon fork,
883 * setprio syscalls, and whenever the interactivity
884 * estimator recalculates.
886 static inline int normal_prio(struct task_struct *p)
890 if (task_has_dl_policy(p))
891 prio = MAX_DL_PRIO-1;
892 else if (task_has_rt_policy(p))
893 prio = MAX_RT_PRIO-1 - p->rt_priority;
895 prio = __normal_prio(p);
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
906 static int effective_prio(struct task_struct *p)
908 p->normal_prio = normal_prio(p);
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
914 if (!rt_prio(p->prio))
915 return p->normal_prio;
920 * task_curr - is this task currently executing on a CPU?
921 * @p: the task in question.
923 * Return: 1 if the task is currently executing. 0 otherwise.
925 inline int task_curr(const struct task_struct *p)
927 return cpu_curr(task_cpu(p)) == p;
931 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
932 * use the balance_callback list if you want balancing.
934 * this means any call to check_class_changed() must be followed by a call to
935 * balance_callback().
937 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
938 const struct sched_class *prev_class,
941 if (prev_class != p->sched_class) {
942 if (prev_class->switched_from)
943 prev_class->switched_from(rq, p);
945 p->sched_class->switched_to(rq, p);
946 } else if (oldprio != p->prio || dl_task(p))
947 p->sched_class->prio_changed(rq, p, oldprio);
950 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
952 const struct sched_class *class;
954 if (p->sched_class == rq->curr->sched_class) {
955 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
957 for_each_class(class) {
958 if (class == rq->curr->sched_class)
960 if (class == p->sched_class) {
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
972 rq_clock_skip_update(rq, true);
977 * This is how migration works:
979 * 1) we invoke migration_cpu_stop() on the target CPU using
981 * 2) stopper starts to run (implicitly forcing the migrated thread
983 * 3) it checks whether the migrated task is still in the wrong runqueue.
984 * 4) if it's in the wrong runqueue then the migration thread removes
985 * it and puts it into the right queue.
986 * 5) stopper completes and stop_one_cpu() returns and the migration
991 * move_queued_task - move a queued task to new rq.
993 * Returns (locked) new rq. Old rq's lock is released.
995 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
997 lockdep_assert_held(&rq->lock);
999 p->on_rq = TASK_ON_RQ_MIGRATING;
1000 dequeue_task(rq, p, 0);
1001 set_task_cpu(p, new_cpu);
1002 raw_spin_unlock(&rq->lock);
1004 rq = cpu_rq(new_cpu);
1006 raw_spin_lock(&rq->lock);
1007 BUG_ON(task_cpu(p) != new_cpu);
1008 enqueue_task(rq, p, 0);
1009 p->on_rq = TASK_ON_RQ_QUEUED;
1010 check_preempt_curr(rq, p, 0);
1015 struct migration_arg {
1016 struct task_struct *task;
1021 * Move (not current) task off this cpu, onto dest cpu. We're doing
1022 * this because either it can't run here any more (set_cpus_allowed()
1023 * away from this CPU, or CPU going down), or because we're
1024 * attempting to rebalance this task on exec (sched_exec).
1026 * So we race with normal scheduler movements, but that's OK, as long
1027 * as the task is no longer on this CPU.
1029 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1031 if (unlikely(!cpu_active(dest_cpu)))
1034 /* Affinity changed (again). */
1035 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1038 rq = move_queued_task(rq, p, dest_cpu);
1044 * migration_cpu_stop - this will be executed by a highprio stopper thread
1045 * and performs thread migration by bumping thread off CPU then
1046 * 'pushing' onto another runqueue.
1048 static int migration_cpu_stop(void *data)
1050 struct migration_arg *arg = data;
1051 struct task_struct *p = arg->task;
1052 struct rq *rq = this_rq();
1055 * The original target cpu might have gone down and we might
1056 * be on another cpu but it doesn't matter.
1058 local_irq_disable();
1060 * We need to explicitly wake pending tasks before running
1061 * __migrate_task() such that we will not miss enforcing cpus_allowed
1062 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1064 sched_ttwu_pending();
1066 raw_spin_lock(&p->pi_lock);
1067 raw_spin_lock(&rq->lock);
1069 * If task_rq(p) != rq, it cannot be migrated here, because we're
1070 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1071 * we're holding p->pi_lock.
1073 if (task_rq(p) == rq) {
1074 if (task_on_rq_queued(p))
1075 rq = __migrate_task(rq, p, arg->dest_cpu);
1077 p->wake_cpu = arg->dest_cpu;
1079 raw_spin_unlock(&rq->lock);
1080 raw_spin_unlock(&p->pi_lock);
1087 * sched_class::set_cpus_allowed must do the below, but is not required to
1088 * actually call this function.
1090 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1092 cpumask_copy(&p->cpus_allowed, new_mask);
1093 p->nr_cpus_allowed = cpumask_weight(new_mask);
1096 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1098 struct rq *rq = task_rq(p);
1099 bool queued, running;
1101 lockdep_assert_held(&p->pi_lock);
1103 queued = task_on_rq_queued(p);
1104 running = task_current(rq, p);
1108 * Because __kthread_bind() calls this on blocked tasks without
1111 lockdep_assert_held(&rq->lock);
1112 dequeue_task(rq, p, DEQUEUE_SAVE);
1115 put_prev_task(rq, p);
1117 p->sched_class->set_cpus_allowed(p, new_mask);
1120 enqueue_task(rq, p, ENQUEUE_RESTORE);
1122 set_curr_task(rq, p);
1126 * Change a given task's CPU affinity. Migrate the thread to a
1127 * proper CPU and schedule it away if the CPU it's executing on
1128 * is removed from the allowed bitmask.
1130 * NOTE: the caller must have a valid reference to the task, the
1131 * task must not exit() & deallocate itself prematurely. The
1132 * call is not atomic; no spinlocks may be held.
1134 static int __set_cpus_allowed_ptr(struct task_struct *p,
1135 const struct cpumask *new_mask, bool check)
1137 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1138 unsigned int dest_cpu;
1143 rq = task_rq_lock(p, &rf);
1145 if (p->flags & PF_KTHREAD) {
1147 * Kernel threads are allowed on online && !active CPUs
1149 cpu_valid_mask = cpu_online_mask;
1153 * Must re-check here, to close a race against __kthread_bind(),
1154 * sched_setaffinity() is not guaranteed to observe the flag.
1156 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1161 if (cpumask_equal(&p->cpus_allowed, new_mask))
1164 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1169 do_set_cpus_allowed(p, new_mask);
1171 if (p->flags & PF_KTHREAD) {
1173 * For kernel threads that do indeed end up on online &&
1174 * !active we want to ensure they are strict per-cpu threads.
1176 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1177 !cpumask_intersects(new_mask, cpu_active_mask) &&
1178 p->nr_cpus_allowed != 1);
1181 /* Can the task run on the task's current CPU? If so, we're done */
1182 if (cpumask_test_cpu(task_cpu(p), new_mask))
1185 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1186 if (task_running(rq, p) || p->state == TASK_WAKING) {
1187 struct migration_arg arg = { p, dest_cpu };
1188 /* Need help from migration thread: drop lock and wait. */
1189 task_rq_unlock(rq, p, &rf);
1190 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1191 tlb_migrate_finish(p->mm);
1193 } else if (task_on_rq_queued(p)) {
1195 * OK, since we're going to drop the lock immediately
1196 * afterwards anyway.
1198 lockdep_unpin_lock(&rq->lock, rf.cookie);
1199 rq = move_queued_task(rq, p, dest_cpu);
1200 lockdep_repin_lock(&rq->lock, rf.cookie);
1203 task_rq_unlock(rq, p, &rf);
1208 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1210 return __set_cpus_allowed_ptr(p, new_mask, false);
1212 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1214 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1216 #ifdef CONFIG_SCHED_DEBUG
1218 * We should never call set_task_cpu() on a blocked task,
1219 * ttwu() will sort out the placement.
1221 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1225 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1226 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1227 * time relying on p->on_rq.
1229 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1230 p->sched_class == &fair_sched_class &&
1231 (p->on_rq && !task_on_rq_migrating(p)));
1233 #ifdef CONFIG_LOCKDEP
1235 * The caller should hold either p->pi_lock or rq->lock, when changing
1236 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1238 * sched_move_task() holds both and thus holding either pins the cgroup,
1241 * Furthermore, all task_rq users should acquire both locks, see
1244 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1245 lockdep_is_held(&task_rq(p)->lock)));
1249 trace_sched_migrate_task(p, new_cpu);
1251 if (task_cpu(p) != new_cpu) {
1252 if (p->sched_class->migrate_task_rq)
1253 p->sched_class->migrate_task_rq(p);
1254 p->se.nr_migrations++;
1255 perf_event_task_migrate(p);
1258 __set_task_cpu(p, new_cpu);
1261 static void __migrate_swap_task(struct task_struct *p, int cpu)
1263 if (task_on_rq_queued(p)) {
1264 struct rq *src_rq, *dst_rq;
1266 src_rq = task_rq(p);
1267 dst_rq = cpu_rq(cpu);
1269 p->on_rq = TASK_ON_RQ_MIGRATING;
1270 deactivate_task(src_rq, p, 0);
1271 set_task_cpu(p, cpu);
1272 activate_task(dst_rq, p, 0);
1273 p->on_rq = TASK_ON_RQ_QUEUED;
1274 check_preempt_curr(dst_rq, p, 0);
1277 * Task isn't running anymore; make it appear like we migrated
1278 * it before it went to sleep. This means on wakeup we make the
1279 * previous cpu our target instead of where it really is.
1285 struct migration_swap_arg {
1286 struct task_struct *src_task, *dst_task;
1287 int src_cpu, dst_cpu;
1290 static int migrate_swap_stop(void *data)
1292 struct migration_swap_arg *arg = data;
1293 struct rq *src_rq, *dst_rq;
1296 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1299 src_rq = cpu_rq(arg->src_cpu);
1300 dst_rq = cpu_rq(arg->dst_cpu);
1302 double_raw_lock(&arg->src_task->pi_lock,
1303 &arg->dst_task->pi_lock);
1304 double_rq_lock(src_rq, dst_rq);
1306 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1309 if (task_cpu(arg->src_task) != arg->src_cpu)
1312 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1315 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1318 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1319 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1324 double_rq_unlock(src_rq, dst_rq);
1325 raw_spin_unlock(&arg->dst_task->pi_lock);
1326 raw_spin_unlock(&arg->src_task->pi_lock);
1332 * Cross migrate two tasks
1334 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1336 struct migration_swap_arg arg;
1339 arg = (struct migration_swap_arg){
1341 .src_cpu = task_cpu(cur),
1343 .dst_cpu = task_cpu(p),
1346 if (arg.src_cpu == arg.dst_cpu)
1350 * These three tests are all lockless; this is OK since all of them
1351 * will be re-checked with proper locks held further down the line.
1353 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1356 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1359 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1362 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1363 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1370 * wait_task_inactive - wait for a thread to unschedule.
1372 * If @match_state is nonzero, it's the @p->state value just checked and
1373 * not expected to change. If it changes, i.e. @p might have woken up,
1374 * then return zero. When we succeed in waiting for @p to be off its CPU,
1375 * we return a positive number (its total switch count). If a second call
1376 * a short while later returns the same number, the caller can be sure that
1377 * @p has remained unscheduled the whole time.
1379 * The caller must ensure that the task *will* unschedule sometime soon,
1380 * else this function might spin for a *long* time. This function can't
1381 * be called with interrupts off, or it may introduce deadlock with
1382 * smp_call_function() if an IPI is sent by the same process we are
1383 * waiting to become inactive.
1385 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1387 int running, queued;
1394 * We do the initial early heuristics without holding
1395 * any task-queue locks at all. We'll only try to get
1396 * the runqueue lock when things look like they will
1402 * If the task is actively running on another CPU
1403 * still, just relax and busy-wait without holding
1406 * NOTE! Since we don't hold any locks, it's not
1407 * even sure that "rq" stays as the right runqueue!
1408 * But we don't care, since "task_running()" will
1409 * return false if the runqueue has changed and p
1410 * is actually now running somewhere else!
1412 while (task_running(rq, p)) {
1413 if (match_state && unlikely(p->state != match_state))
1419 * Ok, time to look more closely! We need the rq
1420 * lock now, to be *sure*. If we're wrong, we'll
1421 * just go back and repeat.
1423 rq = task_rq_lock(p, &rf);
1424 trace_sched_wait_task(p);
1425 running = task_running(rq, p);
1426 queued = task_on_rq_queued(p);
1428 if (!match_state || p->state == match_state)
1429 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1430 task_rq_unlock(rq, p, &rf);
1433 * If it changed from the expected state, bail out now.
1435 if (unlikely(!ncsw))
1439 * Was it really running after all now that we
1440 * checked with the proper locks actually held?
1442 * Oops. Go back and try again..
1444 if (unlikely(running)) {
1450 * It's not enough that it's not actively running,
1451 * it must be off the runqueue _entirely_, and not
1454 * So if it was still runnable (but just not actively
1455 * running right now), it's preempted, and we should
1456 * yield - it could be a while.
1458 if (unlikely(queued)) {
1459 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1461 set_current_state(TASK_UNINTERRUPTIBLE);
1462 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1467 * Ahh, all good. It wasn't running, and it wasn't
1468 * runnable, which means that it will never become
1469 * running in the future either. We're all done!
1478 * kick_process - kick a running thread to enter/exit the kernel
1479 * @p: the to-be-kicked thread
1481 * Cause a process which is running on another CPU to enter
1482 * kernel-mode, without any delay. (to get signals handled.)
1484 * NOTE: this function doesn't have to take the runqueue lock,
1485 * because all it wants to ensure is that the remote task enters
1486 * the kernel. If the IPI races and the task has been migrated
1487 * to another CPU then no harm is done and the purpose has been
1490 void kick_process(struct task_struct *p)
1496 if ((cpu != smp_processor_id()) && task_curr(p))
1497 smp_send_reschedule(cpu);
1500 EXPORT_SYMBOL_GPL(kick_process);
1503 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1505 * A few notes on cpu_active vs cpu_online:
1507 * - cpu_active must be a subset of cpu_online
1509 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1510 * see __set_cpus_allowed_ptr(). At this point the newly online
1511 * cpu isn't yet part of the sched domains, and balancing will not
1514 * - on cpu-down we clear cpu_active() to mask the sched domains and
1515 * avoid the load balancer to place new tasks on the to be removed
1516 * cpu. Existing tasks will remain running there and will be taken
1519 * This means that fallback selection must not select !active CPUs.
1520 * And can assume that any active CPU must be online. Conversely
1521 * select_task_rq() below may allow selection of !active CPUs in order
1522 * to satisfy the above rules.
1524 static int select_fallback_rq(int cpu, struct task_struct *p)
1526 int nid = cpu_to_node(cpu);
1527 const struct cpumask *nodemask = NULL;
1528 enum { cpuset, possible, fail } state = cpuset;
1532 * If the node that the cpu is on has been offlined, cpu_to_node()
1533 * will return -1. There is no cpu on the node, and we should
1534 * select the cpu on the other node.
1537 nodemask = cpumask_of_node(nid);
1539 /* Look for allowed, online CPU in same node. */
1540 for_each_cpu(dest_cpu, nodemask) {
1541 if (!cpu_active(dest_cpu))
1543 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1549 /* Any allowed, online CPU? */
1550 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1551 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1553 if (!cpu_online(dest_cpu))
1558 /* No more Mr. Nice Guy. */
1561 if (IS_ENABLED(CONFIG_CPUSETS)) {
1562 cpuset_cpus_allowed_fallback(p);
1568 do_set_cpus_allowed(p, cpu_possible_mask);
1579 if (state != cpuset) {
1581 * Don't tell them about moving exiting tasks or
1582 * kernel threads (both mm NULL), since they never
1585 if (p->mm && printk_ratelimit()) {
1586 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1587 task_pid_nr(p), p->comm, cpu);
1595 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1598 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1600 lockdep_assert_held(&p->pi_lock);
1602 if (tsk_nr_cpus_allowed(p) > 1)
1603 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1605 cpu = cpumask_any(tsk_cpus_allowed(p));
1608 * In order not to call set_task_cpu() on a blocking task we need
1609 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1612 * Since this is common to all placement strategies, this lives here.
1614 * [ this allows ->select_task() to simply return task_cpu(p) and
1615 * not worry about this generic constraint ]
1617 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1619 cpu = select_fallback_rq(task_cpu(p), p);
1624 static void update_avg(u64 *avg, u64 sample)
1626 s64 diff = sample - *avg;
1632 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1633 const struct cpumask *new_mask, bool check)
1635 return set_cpus_allowed_ptr(p, new_mask);
1638 #endif /* CONFIG_SMP */
1641 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1645 if (!schedstat_enabled())
1651 if (cpu == rq->cpu) {
1652 schedstat_inc(rq->ttwu_local);
1653 schedstat_inc(p->se.statistics.nr_wakeups_local);
1655 struct sched_domain *sd;
1657 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1659 for_each_domain(rq->cpu, sd) {
1660 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1661 schedstat_inc(sd->ttwu_wake_remote);
1668 if (wake_flags & WF_MIGRATED)
1669 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1670 #endif /* CONFIG_SMP */
1672 schedstat_inc(rq->ttwu_count);
1673 schedstat_inc(p->se.statistics.nr_wakeups);
1675 if (wake_flags & WF_SYNC)
1676 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1679 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1681 activate_task(rq, p, en_flags);
1682 p->on_rq = TASK_ON_RQ_QUEUED;
1684 /* if a worker is waking up, notify workqueue */
1685 if (p->flags & PF_WQ_WORKER)
1686 wq_worker_waking_up(p, cpu_of(rq));
1690 * Mark the task runnable and perform wakeup-preemption.
1692 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1693 struct pin_cookie cookie)
1695 check_preempt_curr(rq, p, wake_flags);
1696 p->state = TASK_RUNNING;
1697 trace_sched_wakeup(p);
1700 if (p->sched_class->task_woken) {
1702 * Our task @p is fully woken up and running; so its safe to
1703 * drop the rq->lock, hereafter rq is only used for statistics.
1705 lockdep_unpin_lock(&rq->lock, cookie);
1706 p->sched_class->task_woken(rq, p);
1707 lockdep_repin_lock(&rq->lock, cookie);
1710 if (rq->idle_stamp) {
1711 u64 delta = rq_clock(rq) - rq->idle_stamp;
1712 u64 max = 2*rq->max_idle_balance_cost;
1714 update_avg(&rq->avg_idle, delta);
1716 if (rq->avg_idle > max)
1725 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1726 struct pin_cookie cookie)
1728 int en_flags = ENQUEUE_WAKEUP;
1730 lockdep_assert_held(&rq->lock);
1733 if (p->sched_contributes_to_load)
1734 rq->nr_uninterruptible--;
1736 if (wake_flags & WF_MIGRATED)
1737 en_flags |= ENQUEUE_MIGRATED;
1740 ttwu_activate(rq, p, en_flags);
1741 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1745 * Called in case the task @p isn't fully descheduled from its runqueue,
1746 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1747 * since all we need to do is flip p->state to TASK_RUNNING, since
1748 * the task is still ->on_rq.
1750 static int ttwu_remote(struct task_struct *p, int wake_flags)
1756 rq = __task_rq_lock(p, &rf);
1757 if (task_on_rq_queued(p)) {
1758 /* check_preempt_curr() may use rq clock */
1759 update_rq_clock(rq);
1760 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1763 __task_rq_unlock(rq, &rf);
1769 void sched_ttwu_pending(void)
1771 struct rq *rq = this_rq();
1772 struct llist_node *llist = llist_del_all(&rq->wake_list);
1773 struct pin_cookie cookie;
1774 struct task_struct *p;
1775 unsigned long flags;
1780 raw_spin_lock_irqsave(&rq->lock, flags);
1781 cookie = lockdep_pin_lock(&rq->lock);
1786 p = llist_entry(llist, struct task_struct, wake_entry);
1787 llist = llist_next(llist);
1789 if (p->sched_remote_wakeup)
1790 wake_flags = WF_MIGRATED;
1792 ttwu_do_activate(rq, p, wake_flags, cookie);
1795 lockdep_unpin_lock(&rq->lock, cookie);
1796 raw_spin_unlock_irqrestore(&rq->lock, flags);
1799 void scheduler_ipi(void)
1802 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1803 * TIF_NEED_RESCHED remotely (for the first time) will also send
1806 preempt_fold_need_resched();
1808 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1812 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1813 * traditionally all their work was done from the interrupt return
1814 * path. Now that we actually do some work, we need to make sure
1817 * Some archs already do call them, luckily irq_enter/exit nest
1820 * Arguably we should visit all archs and update all handlers,
1821 * however a fair share of IPIs are still resched only so this would
1822 * somewhat pessimize the simple resched case.
1825 sched_ttwu_pending();
1828 * Check if someone kicked us for doing the nohz idle load balance.
1830 if (unlikely(got_nohz_idle_kick())) {
1831 this_rq()->idle_balance = 1;
1832 raise_softirq_irqoff(SCHED_SOFTIRQ);
1837 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1839 struct rq *rq = cpu_rq(cpu);
1841 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1843 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1844 if (!set_nr_if_polling(rq->idle))
1845 smp_send_reschedule(cpu);
1847 trace_sched_wake_idle_without_ipi(cpu);
1851 void wake_up_if_idle(int cpu)
1853 struct rq *rq = cpu_rq(cpu);
1854 unsigned long flags;
1858 if (!is_idle_task(rcu_dereference(rq->curr)))
1861 if (set_nr_if_polling(rq->idle)) {
1862 trace_sched_wake_idle_without_ipi(cpu);
1864 raw_spin_lock_irqsave(&rq->lock, flags);
1865 if (is_idle_task(rq->curr))
1866 smp_send_reschedule(cpu);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq->lock, flags);
1875 bool cpus_share_cache(int this_cpu, int that_cpu)
1877 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1879 #endif /* CONFIG_SMP */
1881 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1883 struct rq *rq = cpu_rq(cpu);
1884 struct pin_cookie cookie;
1886 #if defined(CONFIG_SMP)
1887 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1888 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1889 ttwu_queue_remote(p, cpu, wake_flags);
1894 raw_spin_lock(&rq->lock);
1895 cookie = lockdep_pin_lock(&rq->lock);
1896 ttwu_do_activate(rq, p, wake_flags, cookie);
1897 lockdep_unpin_lock(&rq->lock, cookie);
1898 raw_spin_unlock(&rq->lock);
1902 * Notes on Program-Order guarantees on SMP systems.
1906 * The basic program-order guarantee on SMP systems is that when a task [t]
1907 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1908 * execution on its new cpu [c1].
1910 * For migration (of runnable tasks) this is provided by the following means:
1912 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1913 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1914 * rq(c1)->lock (if not at the same time, then in that order).
1915 * C) LOCK of the rq(c1)->lock scheduling in task
1917 * Transitivity guarantees that B happens after A and C after B.
1918 * Note: we only require RCpc transitivity.
1919 * Note: the cpu doing B need not be c0 or c1
1928 * UNLOCK rq(0)->lock
1930 * LOCK rq(0)->lock // orders against CPU0
1932 * UNLOCK rq(0)->lock
1936 * UNLOCK rq(1)->lock
1938 * LOCK rq(1)->lock // orders against CPU2
1941 * UNLOCK rq(1)->lock
1944 * BLOCKING -- aka. SLEEP + WAKEUP
1946 * For blocking we (obviously) need to provide the same guarantee as for
1947 * migration. However the means are completely different as there is no lock
1948 * chain to provide order. Instead we do:
1950 * 1) smp_store_release(X->on_cpu, 0)
1951 * 2) smp_cond_load_acquire(!X->on_cpu)
1955 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1957 * LOCK rq(0)->lock LOCK X->pi_lock
1960 * smp_store_release(X->on_cpu, 0);
1962 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1968 * X->state = RUNNING
1969 * UNLOCK rq(2)->lock
1971 * LOCK rq(2)->lock // orders against CPU1
1974 * UNLOCK rq(2)->lock
1977 * UNLOCK rq(0)->lock
1980 * However; for wakeups there is a second guarantee we must provide, namely we
1981 * must observe the state that lead to our wakeup. That is, not only must our
1982 * task observe its own prior state, it must also observe the stores prior to
1985 * This means that any means of doing remote wakeups must order the CPU doing
1986 * the wakeup against the CPU the task is going to end up running on. This,
1987 * however, is already required for the regular Program-Order guarantee above,
1988 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1993 * try_to_wake_up - wake up a thread
1994 * @p: the thread to be awakened
1995 * @state: the mask of task states that can be woken
1996 * @wake_flags: wake modifier flags (WF_*)
1998 * Put it on the run-queue if it's not already there. The "current"
1999 * thread is always on the run-queue (except when the actual
2000 * re-schedule is in progress), and as such you're allowed to do
2001 * the simpler "current->state = TASK_RUNNING" to mark yourself
2002 * runnable without the overhead of this.
2004 * Return: %true if @p was woken up, %false if it was already running.
2005 * or @state didn't match @p's state.
2008 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2010 unsigned long flags;
2011 int cpu, success = 0;
2014 * If we are going to wake up a thread waiting for CONDITION we
2015 * need to ensure that CONDITION=1 done by the caller can not be
2016 * reordered with p->state check below. This pairs with mb() in
2017 * set_current_state() the waiting thread does.
2019 smp_mb__before_spinlock();
2020 raw_spin_lock_irqsave(&p->pi_lock, flags);
2021 if (!(p->state & state))
2024 trace_sched_waking(p);
2026 success = 1; /* we're going to change ->state */
2030 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2031 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2032 * in smp_cond_load_acquire() below.
2034 * sched_ttwu_pending() try_to_wake_up()
2035 * [S] p->on_rq = 1; [L] P->state
2036 * UNLOCK rq->lock -----.
2040 * LOCK rq->lock -----'
2044 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2046 * Pairs with the UNLOCK+LOCK on rq->lock from the
2047 * last wakeup of our task and the schedule that got our task
2051 if (p->on_rq && ttwu_remote(p, wake_flags))
2056 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2057 * possible to, falsely, observe p->on_cpu == 0.
2059 * One must be running (->on_cpu == 1) in order to remove oneself
2060 * from the runqueue.
2062 * [S] ->on_cpu = 1; [L] ->on_rq
2066 * [S] ->on_rq = 0; [L] ->on_cpu
2068 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2069 * from the consecutive calls to schedule(); the first switching to our
2070 * task, the second putting it to sleep.
2075 * If the owning (remote) cpu is still in the middle of schedule() with
2076 * this task as prev, wait until its done referencing the task.
2078 * Pairs with the smp_store_release() in finish_lock_switch().
2080 * This ensures that tasks getting woken will be fully ordered against
2081 * their previous state and preserve Program Order.
2083 smp_cond_load_acquire(&p->on_cpu, !VAL);
2085 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2086 p->state = TASK_WAKING;
2088 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2089 if (task_cpu(p) != cpu) {
2090 wake_flags |= WF_MIGRATED;
2091 set_task_cpu(p, cpu);
2093 #endif /* CONFIG_SMP */
2095 ttwu_queue(p, cpu, wake_flags);
2097 ttwu_stat(p, cpu, wake_flags);
2099 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2105 * try_to_wake_up_local - try to wake up a local task with rq lock held
2106 * @p: the thread to be awakened
2107 * @cookie: context's cookie for pinning
2109 * Put @p on the run-queue if it's not already there. The caller must
2110 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2113 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2115 struct rq *rq = task_rq(p);
2117 if (WARN_ON_ONCE(rq != this_rq()) ||
2118 WARN_ON_ONCE(p == current))
2121 lockdep_assert_held(&rq->lock);
2123 if (!raw_spin_trylock(&p->pi_lock)) {
2125 * This is OK, because current is on_cpu, which avoids it being
2126 * picked for load-balance and preemption/IRQs are still
2127 * disabled avoiding further scheduler activity on it and we've
2128 * not yet picked a replacement task.
2130 lockdep_unpin_lock(&rq->lock, cookie);
2131 raw_spin_unlock(&rq->lock);
2132 raw_spin_lock(&p->pi_lock);
2133 raw_spin_lock(&rq->lock);
2134 lockdep_repin_lock(&rq->lock, cookie);
2137 if (!(p->state & TASK_NORMAL))
2140 trace_sched_waking(p);
2142 if (!task_on_rq_queued(p))
2143 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2145 ttwu_do_wakeup(rq, p, 0, cookie);
2146 ttwu_stat(p, smp_processor_id(), 0);
2148 raw_spin_unlock(&p->pi_lock);
2152 * wake_up_process - Wake up a specific process
2153 * @p: The process to be woken up.
2155 * Attempt to wake up the nominated process and move it to the set of runnable
2158 * Return: 1 if the process was woken up, 0 if it was already running.
2160 * It may be assumed that this function implies a write memory barrier before
2161 * changing the task state if and only if any tasks are woken up.
2163 int wake_up_process(struct task_struct *p)
2165 return try_to_wake_up(p, TASK_NORMAL, 0);
2167 EXPORT_SYMBOL(wake_up_process);
2169 int wake_up_state(struct task_struct *p, unsigned int state)
2171 return try_to_wake_up(p, state, 0);
2175 * This function clears the sched_dl_entity static params.
2177 void __dl_clear_params(struct task_struct *p)
2179 struct sched_dl_entity *dl_se = &p->dl;
2181 dl_se->dl_runtime = 0;
2182 dl_se->dl_deadline = 0;
2183 dl_se->dl_period = 0;
2187 dl_se->dl_throttled = 0;
2188 dl_se->dl_yielded = 0;
2192 * Perform scheduler related setup for a newly forked process p.
2193 * p is forked by current.
2195 * __sched_fork() is basic setup used by init_idle() too:
2197 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2202 p->se.exec_start = 0;
2203 p->se.sum_exec_runtime = 0;
2204 p->se.prev_sum_exec_runtime = 0;
2205 p->se.nr_migrations = 0;
2207 INIT_LIST_HEAD(&p->se.group_node);
2209 #ifdef CONFIG_FAIR_GROUP_SCHED
2210 p->se.cfs_rq = NULL;
2213 #ifdef CONFIG_SCHEDSTATS
2214 /* Even if schedstat is disabled, there should not be garbage */
2215 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2218 RB_CLEAR_NODE(&p->dl.rb_node);
2219 init_dl_task_timer(&p->dl);
2220 __dl_clear_params(p);
2222 INIT_LIST_HEAD(&p->rt.run_list);
2224 p->rt.time_slice = sched_rr_timeslice;
2228 #ifdef CONFIG_PREEMPT_NOTIFIERS
2229 INIT_HLIST_HEAD(&p->preempt_notifiers);
2232 #ifdef CONFIG_NUMA_BALANCING
2233 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2234 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2235 p->mm->numa_scan_seq = 0;
2238 if (clone_flags & CLONE_VM)
2239 p->numa_preferred_nid = current->numa_preferred_nid;
2241 p->numa_preferred_nid = -1;
2243 p->node_stamp = 0ULL;
2244 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2245 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2246 p->numa_work.next = &p->numa_work;
2247 p->numa_faults = NULL;
2248 p->last_task_numa_placement = 0;
2249 p->last_sum_exec_runtime = 0;
2251 p->numa_group = NULL;
2252 #endif /* CONFIG_NUMA_BALANCING */
2255 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2257 #ifdef CONFIG_NUMA_BALANCING
2259 void set_numabalancing_state(bool enabled)
2262 static_branch_enable(&sched_numa_balancing);
2264 static_branch_disable(&sched_numa_balancing);
2267 #ifdef CONFIG_PROC_SYSCTL
2268 int sysctl_numa_balancing(struct ctl_table *table, int write,
2269 void __user *buffer, size_t *lenp, loff_t *ppos)
2273 int state = static_branch_likely(&sched_numa_balancing);
2275 if (write && !capable(CAP_SYS_ADMIN))
2280 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2284 set_numabalancing_state(state);
2290 #ifdef CONFIG_SCHEDSTATS
2292 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2293 static bool __initdata __sched_schedstats = false;
2295 static void set_schedstats(bool enabled)
2298 static_branch_enable(&sched_schedstats);
2300 static_branch_disable(&sched_schedstats);
2303 void force_schedstat_enabled(void)
2305 if (!schedstat_enabled()) {
2306 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2307 static_branch_enable(&sched_schedstats);
2311 static int __init setup_schedstats(char *str)
2318 * This code is called before jump labels have been set up, so we can't
2319 * change the static branch directly just yet. Instead set a temporary
2320 * variable so init_schedstats() can do it later.
2322 if (!strcmp(str, "enable")) {
2323 __sched_schedstats = true;
2325 } else if (!strcmp(str, "disable")) {
2326 __sched_schedstats = false;
2331 pr_warn("Unable to parse schedstats=\n");
2335 __setup("schedstats=", setup_schedstats);
2337 static void __init init_schedstats(void)
2339 set_schedstats(__sched_schedstats);
2342 #ifdef CONFIG_PROC_SYSCTL
2343 int sysctl_schedstats(struct ctl_table *table, int write,
2344 void __user *buffer, size_t *lenp, loff_t *ppos)
2348 int state = static_branch_likely(&sched_schedstats);
2350 if (write && !capable(CAP_SYS_ADMIN))
2355 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2359 set_schedstats(state);
2362 #endif /* CONFIG_PROC_SYSCTL */
2363 #else /* !CONFIG_SCHEDSTATS */
2364 static inline void init_schedstats(void) {}
2365 #endif /* CONFIG_SCHEDSTATS */
2368 * fork()/clone()-time setup:
2370 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2372 unsigned long flags;
2373 int cpu = get_cpu();
2375 __sched_fork(clone_flags, p);
2377 * We mark the process as NEW here. This guarantees that
2378 * nobody will actually run it, and a signal or other external
2379 * event cannot wake it up and insert it on the runqueue either.
2381 p->state = TASK_NEW;
2384 * Make sure we do not leak PI boosting priority to the child.
2386 p->prio = current->normal_prio;
2389 * Revert to default priority/policy on fork if requested.
2391 if (unlikely(p->sched_reset_on_fork)) {
2392 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2393 p->policy = SCHED_NORMAL;
2394 p->static_prio = NICE_TO_PRIO(0);
2396 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2397 p->static_prio = NICE_TO_PRIO(0);
2399 p->prio = p->normal_prio = __normal_prio(p);
2403 * We don't need the reset flag anymore after the fork. It has
2404 * fulfilled its duty:
2406 p->sched_reset_on_fork = 0;
2409 if (dl_prio(p->prio)) {
2412 } else if (rt_prio(p->prio)) {
2413 p->sched_class = &rt_sched_class;
2415 p->sched_class = &fair_sched_class;
2418 init_entity_runnable_average(&p->se);
2421 * The child is not yet in the pid-hash so no cgroup attach races,
2422 * and the cgroup is pinned to this child due to cgroup_fork()
2423 * is ran before sched_fork().
2425 * Silence PROVE_RCU.
2427 raw_spin_lock_irqsave(&p->pi_lock, flags);
2429 * We're setting the cpu for the first time, we don't migrate,
2430 * so use __set_task_cpu().
2432 __set_task_cpu(p, cpu);
2433 if (p->sched_class->task_fork)
2434 p->sched_class->task_fork(p);
2435 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2437 #ifdef CONFIG_SCHED_INFO
2438 if (likely(sched_info_on()))
2439 memset(&p->sched_info, 0, sizeof(p->sched_info));
2441 #if defined(CONFIG_SMP)
2444 init_task_preempt_count(p);
2446 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2447 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2454 unsigned long to_ratio(u64 period, u64 runtime)
2456 if (runtime == RUNTIME_INF)
2460 * Doing this here saves a lot of checks in all
2461 * the calling paths, and returning zero seems
2462 * safe for them anyway.
2467 return div64_u64(runtime << 20, period);
2471 inline struct dl_bw *dl_bw_of(int i)
2473 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2474 "sched RCU must be held");
2475 return &cpu_rq(i)->rd->dl_bw;
2478 static inline int dl_bw_cpus(int i)
2480 struct root_domain *rd = cpu_rq(i)->rd;
2483 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2484 "sched RCU must be held");
2485 for_each_cpu_and(i, rd->span, cpu_active_mask)
2491 inline struct dl_bw *dl_bw_of(int i)
2493 return &cpu_rq(i)->dl.dl_bw;
2496 static inline int dl_bw_cpus(int i)
2503 * We must be sure that accepting a new task (or allowing changing the
2504 * parameters of an existing one) is consistent with the bandwidth
2505 * constraints. If yes, this function also accordingly updates the currently
2506 * allocated bandwidth to reflect the new situation.
2508 * This function is called while holding p's rq->lock.
2510 * XXX we should delay bw change until the task's 0-lag point, see
2513 static int dl_overflow(struct task_struct *p, int policy,
2514 const struct sched_attr *attr)
2517 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2518 u64 period = attr->sched_period ?: attr->sched_deadline;
2519 u64 runtime = attr->sched_runtime;
2520 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2523 /* !deadline task may carry old deadline bandwidth */
2524 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2528 * Either if a task, enters, leave, or stays -deadline but changes
2529 * its parameters, we may need to update accordingly the total
2530 * allocated bandwidth of the container.
2532 raw_spin_lock(&dl_b->lock);
2533 cpus = dl_bw_cpus(task_cpu(p));
2534 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2535 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2536 __dl_add(dl_b, new_bw);
2538 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2539 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2540 __dl_clear(dl_b, p->dl.dl_bw);
2541 __dl_add(dl_b, new_bw);
2543 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2544 __dl_clear(dl_b, p->dl.dl_bw);
2547 raw_spin_unlock(&dl_b->lock);
2552 extern void init_dl_bw(struct dl_bw *dl_b);
2555 * wake_up_new_task - wake up a newly created task for the first time.
2557 * This function will do some initial scheduler statistics housekeeping
2558 * that must be done for every newly created context, then puts the task
2559 * on the runqueue and wakes it.
2561 void wake_up_new_task(struct task_struct *p)
2566 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2567 p->state = TASK_RUNNING;
2570 * Fork balancing, do it here and not earlier because:
2571 * - cpus_allowed can change in the fork path
2572 * - any previously selected cpu might disappear through hotplug
2574 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2575 * as we're not fully set-up yet.
2577 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2579 rq = __task_rq_lock(p, &rf);
2580 post_init_entity_util_avg(&p->se);
2582 activate_task(rq, p, 0);
2583 p->on_rq = TASK_ON_RQ_QUEUED;
2584 trace_sched_wakeup_new(p);
2585 check_preempt_curr(rq, p, WF_FORK);
2587 if (p->sched_class->task_woken) {
2589 * Nothing relies on rq->lock after this, so its fine to
2592 lockdep_unpin_lock(&rq->lock, rf.cookie);
2593 p->sched_class->task_woken(rq, p);
2594 lockdep_repin_lock(&rq->lock, rf.cookie);
2597 task_rq_unlock(rq, p, &rf);
2600 #ifdef CONFIG_PREEMPT_NOTIFIERS
2602 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2604 void preempt_notifier_inc(void)
2606 static_key_slow_inc(&preempt_notifier_key);
2608 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2610 void preempt_notifier_dec(void)
2612 static_key_slow_dec(&preempt_notifier_key);
2614 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2617 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2618 * @notifier: notifier struct to register
2620 void preempt_notifier_register(struct preempt_notifier *notifier)
2622 if (!static_key_false(&preempt_notifier_key))
2623 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2625 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2627 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2630 * preempt_notifier_unregister - no longer interested in preemption notifications
2631 * @notifier: notifier struct to unregister
2633 * This is *not* safe to call from within a preemption notifier.
2635 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2637 hlist_del(¬ifier->link);
2639 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2641 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2643 struct preempt_notifier *notifier;
2645 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2646 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2649 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2651 if (static_key_false(&preempt_notifier_key))
2652 __fire_sched_in_preempt_notifiers(curr);
2656 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2657 struct task_struct *next)
2659 struct preempt_notifier *notifier;
2661 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2662 notifier->ops->sched_out(notifier, next);
2665 static __always_inline void
2666 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2667 struct task_struct *next)
2669 if (static_key_false(&preempt_notifier_key))
2670 __fire_sched_out_preempt_notifiers(curr, next);
2673 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2675 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2680 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2681 struct task_struct *next)
2685 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2688 * prepare_task_switch - prepare to switch tasks
2689 * @rq: the runqueue preparing to switch
2690 * @prev: the current task that is being switched out
2691 * @next: the task we are going to switch to.
2693 * This is called with the rq lock held and interrupts off. It must
2694 * be paired with a subsequent finish_task_switch after the context
2697 * prepare_task_switch sets up locking and calls architecture specific
2701 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2702 struct task_struct *next)
2704 sched_info_switch(rq, prev, next);
2705 perf_event_task_sched_out(prev, next);
2706 fire_sched_out_preempt_notifiers(prev, next);
2707 prepare_lock_switch(rq, next);
2708 prepare_arch_switch(next);
2712 * finish_task_switch - clean up after a task-switch
2713 * @prev: the thread we just switched away from.
2715 * finish_task_switch must be called after the context switch, paired
2716 * with a prepare_task_switch call before the context switch.
2717 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2718 * and do any other architecture-specific cleanup actions.
2720 * Note that we may have delayed dropping an mm in context_switch(). If
2721 * so, we finish that here outside of the runqueue lock. (Doing it
2722 * with the lock held can cause deadlocks; see schedule() for
2725 * The context switch have flipped the stack from under us and restored the
2726 * local variables which were saved when this task called schedule() in the
2727 * past. prev == current is still correct but we need to recalculate this_rq
2728 * because prev may have moved to another CPU.
2730 static struct rq *finish_task_switch(struct task_struct *prev)
2731 __releases(rq->lock)
2733 struct rq *rq = this_rq();
2734 struct mm_struct *mm = rq->prev_mm;
2738 * The previous task will have left us with a preempt_count of 2
2739 * because it left us after:
2742 * preempt_disable(); // 1
2744 * raw_spin_lock_irq(&rq->lock) // 2
2746 * Also, see FORK_PREEMPT_COUNT.
2748 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2749 "corrupted preempt_count: %s/%d/0x%x\n",
2750 current->comm, current->pid, preempt_count()))
2751 preempt_count_set(FORK_PREEMPT_COUNT);
2756 * A task struct has one reference for the use as "current".
2757 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2758 * schedule one last time. The schedule call will never return, and
2759 * the scheduled task must drop that reference.
2761 * We must observe prev->state before clearing prev->on_cpu (in
2762 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2763 * running on another CPU and we could rave with its RUNNING -> DEAD
2764 * transition, resulting in a double drop.
2766 prev_state = prev->state;
2767 vtime_task_switch(prev);
2768 perf_event_task_sched_in(prev, current);
2769 finish_lock_switch(rq, prev);
2770 finish_arch_post_lock_switch();
2772 fire_sched_in_preempt_notifiers(current);
2775 if (unlikely(prev_state == TASK_DEAD)) {
2776 if (prev->sched_class->task_dead)
2777 prev->sched_class->task_dead(prev);
2780 * Remove function-return probe instances associated with this
2781 * task and put them back on the free list.
2783 kprobe_flush_task(prev);
2785 /* Task is done with its stack. */
2786 put_task_stack(prev);
2788 put_task_struct(prev);
2791 tick_nohz_task_switch();
2797 /* rq->lock is NOT held, but preemption is disabled */
2798 static void __balance_callback(struct rq *rq)
2800 struct callback_head *head, *next;
2801 void (*func)(struct rq *rq);
2802 unsigned long flags;
2804 raw_spin_lock_irqsave(&rq->lock, flags);
2805 head = rq->balance_callback;
2806 rq->balance_callback = NULL;
2808 func = (void (*)(struct rq *))head->func;
2815 raw_spin_unlock_irqrestore(&rq->lock, flags);
2818 static inline void balance_callback(struct rq *rq)
2820 if (unlikely(rq->balance_callback))
2821 __balance_callback(rq);
2826 static inline void balance_callback(struct rq *rq)
2833 * schedule_tail - first thing a freshly forked thread must call.
2834 * @prev: the thread we just switched away from.
2836 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2837 __releases(rq->lock)
2842 * New tasks start with FORK_PREEMPT_COUNT, see there and
2843 * finish_task_switch() for details.
2845 * finish_task_switch() will drop rq->lock() and lower preempt_count
2846 * and the preempt_enable() will end up enabling preemption (on
2847 * PREEMPT_COUNT kernels).
2850 rq = finish_task_switch(prev);
2851 balance_callback(rq);
2854 if (current->set_child_tid)
2855 put_user(task_pid_vnr(current), current->set_child_tid);
2859 * context_switch - switch to the new MM and the new thread's register state.
2861 static __always_inline struct rq *
2862 context_switch(struct rq *rq, struct task_struct *prev,
2863 struct task_struct *next, struct pin_cookie cookie)
2865 struct mm_struct *mm, *oldmm;
2867 prepare_task_switch(rq, prev, next);
2870 oldmm = prev->active_mm;
2872 * For paravirt, this is coupled with an exit in switch_to to
2873 * combine the page table reload and the switch backend into
2876 arch_start_context_switch(prev);
2879 next->active_mm = oldmm;
2880 atomic_inc(&oldmm->mm_count);
2881 enter_lazy_tlb(oldmm, next);
2883 switch_mm_irqs_off(oldmm, mm, next);
2886 prev->active_mm = NULL;
2887 rq->prev_mm = oldmm;
2890 * Since the runqueue lock will be released by the next
2891 * task (which is an invalid locking op but in the case
2892 * of the scheduler it's an obvious special-case), so we
2893 * do an early lockdep release here:
2895 lockdep_unpin_lock(&rq->lock, cookie);
2896 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2898 /* Here we just switch the register state and the stack. */
2899 switch_to(prev, next, prev);
2902 return finish_task_switch(prev);
2906 * nr_running and nr_context_switches:
2908 * externally visible scheduler statistics: current number of runnable
2909 * threads, total number of context switches performed since bootup.
2911 unsigned long nr_running(void)
2913 unsigned long i, sum = 0;
2915 for_each_online_cpu(i)
2916 sum += cpu_rq(i)->nr_running;
2922 * Check if only the current task is running on the cpu.
2924 * Caution: this function does not check that the caller has disabled
2925 * preemption, thus the result might have a time-of-check-to-time-of-use
2926 * race. The caller is responsible to use it correctly, for example:
2928 * - from a non-preemptable section (of course)
2930 * - from a thread that is bound to a single CPU
2932 * - in a loop with very short iterations (e.g. a polling loop)
2934 bool single_task_running(void)
2936 return raw_rq()->nr_running == 1;
2938 EXPORT_SYMBOL(single_task_running);
2940 unsigned long long nr_context_switches(void)
2943 unsigned long long sum = 0;
2945 for_each_possible_cpu(i)
2946 sum += cpu_rq(i)->nr_switches;
2951 unsigned long nr_iowait(void)
2953 unsigned long i, sum = 0;
2955 for_each_possible_cpu(i)
2956 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2961 unsigned long nr_iowait_cpu(int cpu)
2963 struct rq *this = cpu_rq(cpu);
2964 return atomic_read(&this->nr_iowait);
2967 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2969 struct rq *rq = this_rq();
2970 *nr_waiters = atomic_read(&rq->nr_iowait);
2971 *load = rq->load.weight;
2977 * sched_exec - execve() is a valuable balancing opportunity, because at
2978 * this point the task has the smallest effective memory and cache footprint.
2980 void sched_exec(void)
2982 struct task_struct *p = current;
2983 unsigned long flags;
2986 raw_spin_lock_irqsave(&p->pi_lock, flags);
2987 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2988 if (dest_cpu == smp_processor_id())
2991 if (likely(cpu_active(dest_cpu))) {
2992 struct migration_arg arg = { p, dest_cpu };
2994 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2995 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2999 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3004 DEFINE_PER_CPU(struct kernel_stat, kstat);
3005 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3007 EXPORT_PER_CPU_SYMBOL(kstat);
3008 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3011 * The function fair_sched_class.update_curr accesses the struct curr
3012 * and its field curr->exec_start; when called from task_sched_runtime(),
3013 * we observe a high rate of cache misses in practice.
3014 * Prefetching this data results in improved performance.
3016 static inline void prefetch_curr_exec_start(struct task_struct *p)
3018 #ifdef CONFIG_FAIR_GROUP_SCHED
3019 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3021 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3024 prefetch(&curr->exec_start);
3028 * Return accounted runtime for the task.
3029 * In case the task is currently running, return the runtime plus current's
3030 * pending runtime that have not been accounted yet.
3032 unsigned long long task_sched_runtime(struct task_struct *p)
3038 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3040 * 64-bit doesn't need locks to atomically read a 64bit value.
3041 * So we have a optimization chance when the task's delta_exec is 0.
3042 * Reading ->on_cpu is racy, but this is ok.
3044 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3045 * If we race with it entering cpu, unaccounted time is 0. This is
3046 * indistinguishable from the read occurring a few cycles earlier.
3047 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3048 * been accounted, so we're correct here as well.
3050 if (!p->on_cpu || !task_on_rq_queued(p))
3051 return p->se.sum_exec_runtime;
3054 rq = task_rq_lock(p, &rf);
3056 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3057 * project cycles that may never be accounted to this
3058 * thread, breaking clock_gettime().
3060 if (task_current(rq, p) && task_on_rq_queued(p)) {
3061 prefetch_curr_exec_start(p);
3062 update_rq_clock(rq);
3063 p->sched_class->update_curr(rq);
3065 ns = p->se.sum_exec_runtime;
3066 task_rq_unlock(rq, p, &rf);
3072 * This function gets called by the timer code, with HZ frequency.
3073 * We call it with interrupts disabled.
3075 void scheduler_tick(void)
3077 int cpu = smp_processor_id();
3078 struct rq *rq = cpu_rq(cpu);
3079 struct task_struct *curr = rq->curr;
3083 raw_spin_lock(&rq->lock);
3084 update_rq_clock(rq);
3085 curr->sched_class->task_tick(rq, curr, 0);
3086 cpu_load_update_active(rq);
3087 calc_global_load_tick(rq);
3088 raw_spin_unlock(&rq->lock);
3090 perf_event_task_tick();
3093 rq->idle_balance = idle_cpu(cpu);
3094 trigger_load_balance(rq);
3096 rq_last_tick_reset(rq);
3099 #ifdef CONFIG_NO_HZ_FULL
3101 * scheduler_tick_max_deferment
3103 * Keep at least one tick per second when a single
3104 * active task is running because the scheduler doesn't
3105 * yet completely support full dynticks environment.
3107 * This makes sure that uptime, CFS vruntime, load
3108 * balancing, etc... continue to move forward, even
3109 * with a very low granularity.
3111 * Return: Maximum deferment in nanoseconds.
3113 u64 scheduler_tick_max_deferment(void)
3115 struct rq *rq = this_rq();
3116 unsigned long next, now = READ_ONCE(jiffies);
3118 next = rq->last_sched_tick + HZ;
3120 if (time_before_eq(next, now))
3123 return jiffies_to_nsecs(next - now);
3127 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3128 defined(CONFIG_PREEMPT_TRACER))
3130 * If the value passed in is equal to the current preempt count
3131 * then we just disabled preemption. Start timing the latency.
3133 static inline void preempt_latency_start(int val)
3135 if (preempt_count() == val) {
3136 unsigned long ip = get_lock_parent_ip();
3137 #ifdef CONFIG_DEBUG_PREEMPT
3138 current->preempt_disable_ip = ip;
3140 trace_preempt_off(CALLER_ADDR0, ip);
3144 void preempt_count_add(int val)
3146 #ifdef CONFIG_DEBUG_PREEMPT
3150 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3153 __preempt_count_add(val);
3154 #ifdef CONFIG_DEBUG_PREEMPT
3156 * Spinlock count overflowing soon?
3158 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3161 preempt_latency_start(val);
3163 EXPORT_SYMBOL(preempt_count_add);
3164 NOKPROBE_SYMBOL(preempt_count_add);
3167 * If the value passed in equals to the current preempt count
3168 * then we just enabled preemption. Stop timing the latency.
3170 static inline void preempt_latency_stop(int val)
3172 if (preempt_count() == val)
3173 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3176 void preempt_count_sub(int val)
3178 #ifdef CONFIG_DEBUG_PREEMPT
3182 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3185 * Is the spinlock portion underflowing?
3187 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3188 !(preempt_count() & PREEMPT_MASK)))
3192 preempt_latency_stop(val);
3193 __preempt_count_sub(val);
3195 EXPORT_SYMBOL(preempt_count_sub);
3196 NOKPROBE_SYMBOL(preempt_count_sub);
3199 static inline void preempt_latency_start(int val) { }
3200 static inline void preempt_latency_stop(int val) { }
3204 * Print scheduling while atomic bug:
3206 static noinline void __schedule_bug(struct task_struct *prev)
3208 /* Save this before calling printk(), since that will clobber it */
3209 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3211 if (oops_in_progress)
3214 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3215 prev->comm, prev->pid, preempt_count());
3217 debug_show_held_locks(prev);
3219 if (irqs_disabled())
3220 print_irqtrace_events(prev);
3221 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3222 && in_atomic_preempt_off()) {
3223 pr_err("Preemption disabled at:");
3224 print_ip_sym(preempt_disable_ip);
3228 panic("scheduling while atomic\n");
3231 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3235 * Various schedule()-time debugging checks and statistics:
3237 static inline void schedule_debug(struct task_struct *prev)
3239 #ifdef CONFIG_SCHED_STACK_END_CHECK
3240 if (task_stack_end_corrupted(prev))
3241 panic("corrupted stack end detected inside scheduler\n");
3244 if (unlikely(in_atomic_preempt_off())) {
3245 __schedule_bug(prev);
3246 preempt_count_set(PREEMPT_DISABLED);
3250 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3252 schedstat_inc(this_rq()->sched_count);
3256 * Pick up the highest-prio task:
3258 static inline struct task_struct *
3259 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3261 const struct sched_class *class = &fair_sched_class;
3262 struct task_struct *p;
3265 * Optimization: we know that if all tasks are in
3266 * the fair class we can call that function directly:
3268 if (likely(prev->sched_class == class &&
3269 rq->nr_running == rq->cfs.h_nr_running)) {
3270 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3271 if (unlikely(p == RETRY_TASK))
3274 /* assumes fair_sched_class->next == idle_sched_class */
3276 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3282 for_each_class(class) {
3283 p = class->pick_next_task(rq, prev, cookie);
3285 if (unlikely(p == RETRY_TASK))
3291 BUG(); /* the idle class will always have a runnable task */
3295 * __schedule() is the main scheduler function.
3297 * The main means of driving the scheduler and thus entering this function are:
3299 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3301 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3302 * paths. For example, see arch/x86/entry_64.S.
3304 * To drive preemption between tasks, the scheduler sets the flag in timer
3305 * interrupt handler scheduler_tick().
3307 * 3. Wakeups don't really cause entry into schedule(). They add a
3308 * task to the run-queue and that's it.
3310 * Now, if the new task added to the run-queue preempts the current
3311 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3312 * called on the nearest possible occasion:
3314 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3316 * - in syscall or exception context, at the next outmost
3317 * preempt_enable(). (this might be as soon as the wake_up()'s
3320 * - in IRQ context, return from interrupt-handler to
3321 * preemptible context
3323 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3326 * - cond_resched() call
3327 * - explicit schedule() call
3328 * - return from syscall or exception to user-space
3329 * - return from interrupt-handler to user-space
3331 * WARNING: must be called with preemption disabled!
3333 static void __sched notrace __schedule(bool preempt)
3335 struct task_struct *prev, *next;
3336 unsigned long *switch_count;
3337 struct pin_cookie cookie;
3341 cpu = smp_processor_id();
3345 schedule_debug(prev);
3347 if (sched_feat(HRTICK))
3350 local_irq_disable();
3351 rcu_note_context_switch();
3354 * Make sure that signal_pending_state()->signal_pending() below
3355 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3356 * done by the caller to avoid the race with signal_wake_up().
3358 smp_mb__before_spinlock();
3359 raw_spin_lock(&rq->lock);
3360 cookie = lockdep_pin_lock(&rq->lock);
3362 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3364 switch_count = &prev->nivcsw;
3365 if (!preempt && prev->state) {
3366 if (unlikely(signal_pending_state(prev->state, prev))) {
3367 prev->state = TASK_RUNNING;
3369 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3373 * If a worker went to sleep, notify and ask workqueue
3374 * whether it wants to wake up a task to maintain
3377 if (prev->flags & PF_WQ_WORKER) {
3378 struct task_struct *to_wakeup;
3380 to_wakeup = wq_worker_sleeping(prev);
3382 try_to_wake_up_local(to_wakeup, cookie);
3385 switch_count = &prev->nvcsw;
3388 if (task_on_rq_queued(prev))
3389 update_rq_clock(rq);
3391 next = pick_next_task(rq, prev, cookie);
3392 clear_tsk_need_resched(prev);
3393 clear_preempt_need_resched();
3394 rq->clock_skip_update = 0;
3396 if (likely(prev != next)) {
3401 trace_sched_switch(preempt, prev, next);
3402 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3404 lockdep_unpin_lock(&rq->lock, cookie);
3405 raw_spin_unlock_irq(&rq->lock);
3408 balance_callback(rq);
3411 void __noreturn do_task_dead(void)
3414 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3415 * when the following two conditions become true.
3416 * - There is race condition of mmap_sem (It is acquired by
3418 * - SMI occurs before setting TASK_RUNINNG.
3419 * (or hypervisor of virtual machine switches to other guest)
3420 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3422 * To avoid it, we have to wait for releasing tsk->pi_lock which
3423 * is held by try_to_wake_up()
3426 raw_spin_unlock_wait(¤t->pi_lock);
3428 /* causes final put_task_struct in finish_task_switch(). */
3429 __set_current_state(TASK_DEAD);
3430 current->flags |= PF_NOFREEZE; /* tell freezer to ignore us */
3433 /* Avoid "noreturn function does return". */
3435 cpu_relax(); /* For when BUG is null */
3438 static inline void sched_submit_work(struct task_struct *tsk)
3440 if (!tsk->state || tsk_is_pi_blocked(tsk))
3443 * If we are going to sleep and we have plugged IO queued,
3444 * make sure to submit it to avoid deadlocks.
3446 if (blk_needs_flush_plug(tsk))
3447 blk_schedule_flush_plug(tsk);
3450 asmlinkage __visible void __sched schedule(void)
3452 struct task_struct *tsk = current;
3454 sched_submit_work(tsk);
3458 sched_preempt_enable_no_resched();
3459 } while (need_resched());
3461 EXPORT_SYMBOL(schedule);
3463 #ifdef CONFIG_CONTEXT_TRACKING
3464 asmlinkage __visible void __sched schedule_user(void)
3467 * If we come here after a random call to set_need_resched(),
3468 * or we have been woken up remotely but the IPI has not yet arrived,
3469 * we haven't yet exited the RCU idle mode. Do it here manually until
3470 * we find a better solution.
3472 * NB: There are buggy callers of this function. Ideally we
3473 * should warn if prev_state != CONTEXT_USER, but that will trigger
3474 * too frequently to make sense yet.
3476 enum ctx_state prev_state = exception_enter();
3478 exception_exit(prev_state);
3483 * schedule_preempt_disabled - called with preemption disabled
3485 * Returns with preemption disabled. Note: preempt_count must be 1
3487 void __sched schedule_preempt_disabled(void)
3489 sched_preempt_enable_no_resched();
3494 static void __sched notrace preempt_schedule_common(void)
3498 * Because the function tracer can trace preempt_count_sub()
3499 * and it also uses preempt_enable/disable_notrace(), if
3500 * NEED_RESCHED is set, the preempt_enable_notrace() called
3501 * by the function tracer will call this function again and
3502 * cause infinite recursion.
3504 * Preemption must be disabled here before the function
3505 * tracer can trace. Break up preempt_disable() into two
3506 * calls. One to disable preemption without fear of being
3507 * traced. The other to still record the preemption latency,
3508 * which can also be traced by the function tracer.
3510 preempt_disable_notrace();
3511 preempt_latency_start(1);
3513 preempt_latency_stop(1);
3514 preempt_enable_no_resched_notrace();
3517 * Check again in case we missed a preemption opportunity
3518 * between schedule and now.
3520 } while (need_resched());
3523 #ifdef CONFIG_PREEMPT
3525 * this is the entry point to schedule() from in-kernel preemption
3526 * off of preempt_enable. Kernel preemptions off return from interrupt
3527 * occur there and call schedule directly.
3529 asmlinkage __visible void __sched notrace preempt_schedule(void)
3532 * If there is a non-zero preempt_count or interrupts are disabled,
3533 * we do not want to preempt the current task. Just return..
3535 if (likely(!preemptible()))
3538 preempt_schedule_common();
3540 NOKPROBE_SYMBOL(preempt_schedule);
3541 EXPORT_SYMBOL(preempt_schedule);
3544 * preempt_schedule_notrace - preempt_schedule called by tracing
3546 * The tracing infrastructure uses preempt_enable_notrace to prevent
3547 * recursion and tracing preempt enabling caused by the tracing
3548 * infrastructure itself. But as tracing can happen in areas coming
3549 * from userspace or just about to enter userspace, a preempt enable
3550 * can occur before user_exit() is called. This will cause the scheduler
3551 * to be called when the system is still in usermode.
3553 * To prevent this, the preempt_enable_notrace will use this function
3554 * instead of preempt_schedule() to exit user context if needed before
3555 * calling the scheduler.
3557 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3559 enum ctx_state prev_ctx;
3561 if (likely(!preemptible()))
3566 * Because the function tracer can trace preempt_count_sub()
3567 * and it also uses preempt_enable/disable_notrace(), if
3568 * NEED_RESCHED is set, the preempt_enable_notrace() called
3569 * by the function tracer will call this function again and
3570 * cause infinite recursion.
3572 * Preemption must be disabled here before the function
3573 * tracer can trace. Break up preempt_disable() into two
3574 * calls. One to disable preemption without fear of being
3575 * traced. The other to still record the preemption latency,
3576 * which can also be traced by the function tracer.
3578 preempt_disable_notrace();
3579 preempt_latency_start(1);
3581 * Needs preempt disabled in case user_exit() is traced
3582 * and the tracer calls preempt_enable_notrace() causing
3583 * an infinite recursion.
3585 prev_ctx = exception_enter();
3587 exception_exit(prev_ctx);
3589 preempt_latency_stop(1);
3590 preempt_enable_no_resched_notrace();
3591 } while (need_resched());
3593 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3595 #endif /* CONFIG_PREEMPT */
3598 * this is the entry point to schedule() from kernel preemption
3599 * off of irq context.
3600 * Note, that this is called and return with irqs disabled. This will
3601 * protect us against recursive calling from irq.
3603 asmlinkage __visible void __sched preempt_schedule_irq(void)
3605 enum ctx_state prev_state;
3607 /* Catch callers which need to be fixed */
3608 BUG_ON(preempt_count() || !irqs_disabled());
3610 prev_state = exception_enter();
3616 local_irq_disable();
3617 sched_preempt_enable_no_resched();
3618 } while (need_resched());
3620 exception_exit(prev_state);
3623 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3626 return try_to_wake_up(curr->private, mode, wake_flags);
3628 EXPORT_SYMBOL(default_wake_function);
3630 #ifdef CONFIG_RT_MUTEXES
3633 * rt_mutex_setprio - set the current priority of a task
3635 * @prio: prio value (kernel-internal form)
3637 * This function changes the 'effective' priority of a task. It does
3638 * not touch ->normal_prio like __setscheduler().
3640 * Used by the rt_mutex code to implement priority inheritance
3641 * logic. Call site only calls if the priority of the task changed.
3643 void rt_mutex_setprio(struct task_struct *p, int prio)
3645 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3646 const struct sched_class *prev_class;
3650 BUG_ON(prio > MAX_PRIO);
3652 rq = __task_rq_lock(p, &rf);
3655 * Idle task boosting is a nono in general. There is one
3656 * exception, when PREEMPT_RT and NOHZ is active:
3658 * The idle task calls get_next_timer_interrupt() and holds
3659 * the timer wheel base->lock on the CPU and another CPU wants
3660 * to access the timer (probably to cancel it). We can safely
3661 * ignore the boosting request, as the idle CPU runs this code
3662 * with interrupts disabled and will complete the lock
3663 * protected section without being interrupted. So there is no
3664 * real need to boost.
3666 if (unlikely(p == rq->idle)) {
3667 WARN_ON(p != rq->curr);
3668 WARN_ON(p->pi_blocked_on);
3672 trace_sched_pi_setprio(p, prio);
3675 if (oldprio == prio)
3676 queue_flag &= ~DEQUEUE_MOVE;
3678 prev_class = p->sched_class;
3679 queued = task_on_rq_queued(p);
3680 running = task_current(rq, p);
3682 dequeue_task(rq, p, queue_flag);
3684 put_prev_task(rq, p);
3687 * Boosting condition are:
3688 * 1. -rt task is running and holds mutex A
3689 * --> -dl task blocks on mutex A
3691 * 2. -dl task is running and holds mutex A
3692 * --> -dl task blocks on mutex A and could preempt the
3695 if (dl_prio(prio)) {
3696 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3697 if (!dl_prio(p->normal_prio) ||
3698 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3699 p->dl.dl_boosted = 1;
3700 queue_flag |= ENQUEUE_REPLENISH;
3702 p->dl.dl_boosted = 0;
3703 p->sched_class = &dl_sched_class;
3704 } else if (rt_prio(prio)) {
3705 if (dl_prio(oldprio))
3706 p->dl.dl_boosted = 0;
3708 queue_flag |= ENQUEUE_HEAD;
3709 p->sched_class = &rt_sched_class;
3711 if (dl_prio(oldprio))
3712 p->dl.dl_boosted = 0;
3713 if (rt_prio(oldprio))
3715 p->sched_class = &fair_sched_class;
3721 enqueue_task(rq, p, queue_flag);
3723 set_curr_task(rq, p);
3725 check_class_changed(rq, p, prev_class, oldprio);
3727 preempt_disable(); /* avoid rq from going away on us */
3728 __task_rq_unlock(rq, &rf);
3730 balance_callback(rq);
3735 void set_user_nice(struct task_struct *p, long nice)
3737 bool queued, running;
3738 int old_prio, delta;
3742 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3745 * We have to be careful, if called from sys_setpriority(),
3746 * the task might be in the middle of scheduling on another CPU.
3748 rq = task_rq_lock(p, &rf);
3750 * The RT priorities are set via sched_setscheduler(), but we still
3751 * allow the 'normal' nice value to be set - but as expected
3752 * it wont have any effect on scheduling until the task is
3753 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3755 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3756 p->static_prio = NICE_TO_PRIO(nice);
3759 queued = task_on_rq_queued(p);
3760 running = task_current(rq, p);
3762 dequeue_task(rq, p, DEQUEUE_SAVE);
3764 put_prev_task(rq, p);
3766 p->static_prio = NICE_TO_PRIO(nice);
3769 p->prio = effective_prio(p);
3770 delta = p->prio - old_prio;
3773 enqueue_task(rq, p, ENQUEUE_RESTORE);
3775 * If the task increased its priority or is running and
3776 * lowered its priority, then reschedule its CPU:
3778 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3782 set_curr_task(rq, p);
3784 task_rq_unlock(rq, p, &rf);
3786 EXPORT_SYMBOL(set_user_nice);
3789 * can_nice - check if a task can reduce its nice value
3793 int can_nice(const struct task_struct *p, const int nice)
3795 /* convert nice value [19,-20] to rlimit style value [1,40] */
3796 int nice_rlim = nice_to_rlimit(nice);
3798 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3799 capable(CAP_SYS_NICE));
3802 #ifdef __ARCH_WANT_SYS_NICE
3805 * sys_nice - change the priority of the current process.
3806 * @increment: priority increment
3808 * sys_setpriority is a more generic, but much slower function that
3809 * does similar things.
3811 SYSCALL_DEFINE1(nice, int, increment)
3816 * Setpriority might change our priority at the same moment.
3817 * We don't have to worry. Conceptually one call occurs first
3818 * and we have a single winner.
3820 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3821 nice = task_nice(current) + increment;
3823 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3824 if (increment < 0 && !can_nice(current, nice))
3827 retval = security_task_setnice(current, nice);
3831 set_user_nice(current, nice);
3838 * task_prio - return the priority value of a given task.
3839 * @p: the task in question.
3841 * Return: The priority value as seen by users in /proc.
3842 * RT tasks are offset by -200. Normal tasks are centered
3843 * around 0, value goes from -16 to +15.
3845 int task_prio(const struct task_struct *p)
3847 return p->prio - MAX_RT_PRIO;
3851 * idle_cpu - is a given cpu idle currently?
3852 * @cpu: the processor in question.
3854 * Return: 1 if the CPU is currently idle. 0 otherwise.
3856 int idle_cpu(int cpu)
3858 struct rq *rq = cpu_rq(cpu);
3860 if (rq->curr != rq->idle)
3867 if (!llist_empty(&rq->wake_list))
3875 * idle_task - return the idle task for a given cpu.
3876 * @cpu: the processor in question.
3878 * Return: The idle task for the cpu @cpu.
3880 struct task_struct *idle_task(int cpu)
3882 return cpu_rq(cpu)->idle;
3886 * find_process_by_pid - find a process with a matching PID value.
3887 * @pid: the pid in question.
3889 * The task of @pid, if found. %NULL otherwise.
3891 static struct task_struct *find_process_by_pid(pid_t pid)
3893 return pid ? find_task_by_vpid(pid) : current;
3897 * This function initializes the sched_dl_entity of a newly becoming
3898 * SCHED_DEADLINE task.
3900 * Only the static values are considered here, the actual runtime and the
3901 * absolute deadline will be properly calculated when the task is enqueued
3902 * for the first time with its new policy.
3905 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3907 struct sched_dl_entity *dl_se = &p->dl;
3909 dl_se->dl_runtime = attr->sched_runtime;
3910 dl_se->dl_deadline = attr->sched_deadline;
3911 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3912 dl_se->flags = attr->sched_flags;
3913 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3916 * Changing the parameters of a task is 'tricky' and we're not doing
3917 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3919 * What we SHOULD do is delay the bandwidth release until the 0-lag
3920 * point. This would include retaining the task_struct until that time
3921 * and change dl_overflow() to not immediately decrement the current
3924 * Instead we retain the current runtime/deadline and let the new
3925 * parameters take effect after the current reservation period lapses.
3926 * This is safe (albeit pessimistic) because the 0-lag point is always
3927 * before the current scheduling deadline.
3929 * We can still have temporary overloads because we do not delay the
3930 * change in bandwidth until that time; so admission control is
3931 * not on the safe side. It does however guarantee tasks will never
3932 * consume more than promised.
3937 * sched_setparam() passes in -1 for its policy, to let the functions
3938 * it calls know not to change it.
3940 #define SETPARAM_POLICY -1
3942 static void __setscheduler_params(struct task_struct *p,
3943 const struct sched_attr *attr)
3945 int policy = attr->sched_policy;
3947 if (policy == SETPARAM_POLICY)
3952 if (dl_policy(policy))
3953 __setparam_dl(p, attr);
3954 else if (fair_policy(policy))
3955 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3958 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3959 * !rt_policy. Always setting this ensures that things like
3960 * getparam()/getattr() don't report silly values for !rt tasks.
3962 p->rt_priority = attr->sched_priority;
3963 p->normal_prio = normal_prio(p);
3967 /* Actually do priority change: must hold pi & rq lock. */
3968 static void __setscheduler(struct rq *rq, struct task_struct *p,
3969 const struct sched_attr *attr, bool keep_boost)
3971 __setscheduler_params(p, attr);
3974 * Keep a potential priority boosting if called from
3975 * sched_setscheduler().
3978 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3980 p->prio = normal_prio(p);
3982 if (dl_prio(p->prio))
3983 p->sched_class = &dl_sched_class;
3984 else if (rt_prio(p->prio))
3985 p->sched_class = &rt_sched_class;
3987 p->sched_class = &fair_sched_class;
3991 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3993 struct sched_dl_entity *dl_se = &p->dl;
3995 attr->sched_priority = p->rt_priority;
3996 attr->sched_runtime = dl_se->dl_runtime;
3997 attr->sched_deadline = dl_se->dl_deadline;
3998 attr->sched_period = dl_se->dl_period;
3999 attr->sched_flags = dl_se->flags;
4003 * This function validates the new parameters of a -deadline task.
4004 * We ask for the deadline not being zero, and greater or equal
4005 * than the runtime, as well as the period of being zero or
4006 * greater than deadline. Furthermore, we have to be sure that
4007 * user parameters are above the internal resolution of 1us (we
4008 * check sched_runtime only since it is always the smaller one) and
4009 * below 2^63 ns (we have to check both sched_deadline and
4010 * sched_period, as the latter can be zero).
4013 __checkparam_dl(const struct sched_attr *attr)
4016 if (attr->sched_deadline == 0)
4020 * Since we truncate DL_SCALE bits, make sure we're at least
4023 if (attr->sched_runtime < (1ULL << DL_SCALE))
4027 * Since we use the MSB for wrap-around and sign issues, make
4028 * sure it's not set (mind that period can be equal to zero).
4030 if (attr->sched_deadline & (1ULL << 63) ||
4031 attr->sched_period & (1ULL << 63))
4034 /* runtime <= deadline <= period (if period != 0) */
4035 if ((attr->sched_period != 0 &&
4036 attr->sched_period < attr->sched_deadline) ||
4037 attr->sched_deadline < attr->sched_runtime)
4044 * check the target process has a UID that matches the current process's
4046 static bool check_same_owner(struct task_struct *p)
4048 const struct cred *cred = current_cred(), *pcred;
4052 pcred = __task_cred(p);
4053 match = (uid_eq(cred->euid, pcred->euid) ||
4054 uid_eq(cred->euid, pcred->uid));
4059 static bool dl_param_changed(struct task_struct *p,
4060 const struct sched_attr *attr)
4062 struct sched_dl_entity *dl_se = &p->dl;
4064 if (dl_se->dl_runtime != attr->sched_runtime ||
4065 dl_se->dl_deadline != attr->sched_deadline ||
4066 dl_se->dl_period != attr->sched_period ||
4067 dl_se->flags != attr->sched_flags)
4073 static int __sched_setscheduler(struct task_struct *p,
4074 const struct sched_attr *attr,
4077 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4078 MAX_RT_PRIO - 1 - attr->sched_priority;
4079 int retval, oldprio, oldpolicy = -1, queued, running;
4080 int new_effective_prio, policy = attr->sched_policy;
4081 const struct sched_class *prev_class;
4084 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4087 /* may grab non-irq protected spin_locks */
4088 BUG_ON(in_interrupt());
4090 /* double check policy once rq lock held */
4092 reset_on_fork = p->sched_reset_on_fork;
4093 policy = oldpolicy = p->policy;
4095 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4097 if (!valid_policy(policy))
4101 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4105 * Valid priorities for SCHED_FIFO and SCHED_RR are
4106 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4107 * SCHED_BATCH and SCHED_IDLE is 0.
4109 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4110 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4112 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4113 (rt_policy(policy) != (attr->sched_priority != 0)))
4117 * Allow unprivileged RT tasks to decrease priority:
4119 if (user && !capable(CAP_SYS_NICE)) {
4120 if (fair_policy(policy)) {
4121 if (attr->sched_nice < task_nice(p) &&
4122 !can_nice(p, attr->sched_nice))
4126 if (rt_policy(policy)) {
4127 unsigned long rlim_rtprio =
4128 task_rlimit(p, RLIMIT_RTPRIO);
4130 /* can't set/change the rt policy */
4131 if (policy != p->policy && !rlim_rtprio)
4134 /* can't increase priority */
4135 if (attr->sched_priority > p->rt_priority &&
4136 attr->sched_priority > rlim_rtprio)
4141 * Can't set/change SCHED_DEADLINE policy at all for now
4142 * (safest behavior); in the future we would like to allow
4143 * unprivileged DL tasks to increase their relative deadline
4144 * or reduce their runtime (both ways reducing utilization)
4146 if (dl_policy(policy))
4150 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4151 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4153 if (idle_policy(p->policy) && !idle_policy(policy)) {
4154 if (!can_nice(p, task_nice(p)))
4158 /* can't change other user's priorities */
4159 if (!check_same_owner(p))
4162 /* Normal users shall not reset the sched_reset_on_fork flag */
4163 if (p->sched_reset_on_fork && !reset_on_fork)
4168 retval = security_task_setscheduler(p);
4174 * make sure no PI-waiters arrive (or leave) while we are
4175 * changing the priority of the task:
4177 * To be able to change p->policy safely, the appropriate
4178 * runqueue lock must be held.
4180 rq = task_rq_lock(p, &rf);
4183 * Changing the policy of the stop threads its a very bad idea
4185 if (p == rq->stop) {
4186 task_rq_unlock(rq, p, &rf);
4191 * If not changing anything there's no need to proceed further,
4192 * but store a possible modification of reset_on_fork.
4194 if (unlikely(policy == p->policy)) {
4195 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4197 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4199 if (dl_policy(policy) && dl_param_changed(p, attr))
4202 p->sched_reset_on_fork = reset_on_fork;
4203 task_rq_unlock(rq, p, &rf);
4209 #ifdef CONFIG_RT_GROUP_SCHED
4211 * Do not allow realtime tasks into groups that have no runtime
4214 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4215 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4216 !task_group_is_autogroup(task_group(p))) {
4217 task_rq_unlock(rq, p, &rf);
4222 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4223 cpumask_t *span = rq->rd->span;
4226 * Don't allow tasks with an affinity mask smaller than
4227 * the entire root_domain to become SCHED_DEADLINE. We
4228 * will also fail if there's no bandwidth available.
4230 if (!cpumask_subset(span, &p->cpus_allowed) ||
4231 rq->rd->dl_bw.bw == 0) {
4232 task_rq_unlock(rq, p, &rf);
4239 /* recheck policy now with rq lock held */
4240 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4241 policy = oldpolicy = -1;
4242 task_rq_unlock(rq, p, &rf);
4247 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4248 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4251 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4252 task_rq_unlock(rq, p, &rf);
4256 p->sched_reset_on_fork = reset_on_fork;
4261 * Take priority boosted tasks into account. If the new
4262 * effective priority is unchanged, we just store the new
4263 * normal parameters and do not touch the scheduler class and
4264 * the runqueue. This will be done when the task deboost
4267 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4268 if (new_effective_prio == oldprio)
4269 queue_flags &= ~DEQUEUE_MOVE;
4272 queued = task_on_rq_queued(p);
4273 running = task_current(rq, p);
4275 dequeue_task(rq, p, queue_flags);
4277 put_prev_task(rq, p);
4279 prev_class = p->sched_class;
4280 __setscheduler(rq, p, attr, pi);
4284 * We enqueue to tail when the priority of a task is
4285 * increased (user space view).
4287 if (oldprio < p->prio)
4288 queue_flags |= ENQUEUE_HEAD;
4290 enqueue_task(rq, p, queue_flags);
4293 set_curr_task(rq, p);
4295 check_class_changed(rq, p, prev_class, oldprio);
4296 preempt_disable(); /* avoid rq from going away on us */
4297 task_rq_unlock(rq, p, &rf);
4300 rt_mutex_adjust_pi(p);
4303 * Run balance callbacks after we've adjusted the PI chain.
4305 balance_callback(rq);
4311 static int _sched_setscheduler(struct task_struct *p, int policy,
4312 const struct sched_param *param, bool check)
4314 struct sched_attr attr = {
4315 .sched_policy = policy,
4316 .sched_priority = param->sched_priority,
4317 .sched_nice = PRIO_TO_NICE(p->static_prio),
4320 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4321 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4322 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4323 policy &= ~SCHED_RESET_ON_FORK;
4324 attr.sched_policy = policy;
4327 return __sched_setscheduler(p, &attr, check, true);
4330 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4331 * @p: the task in question.
4332 * @policy: new policy.
4333 * @param: structure containing the new RT priority.
4335 * Return: 0 on success. An error code otherwise.
4337 * NOTE that the task may be already dead.
4339 int sched_setscheduler(struct task_struct *p, int policy,
4340 const struct sched_param *param)
4342 return _sched_setscheduler(p, policy, param, true);
4344 EXPORT_SYMBOL_GPL(sched_setscheduler);
4346 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4348 return __sched_setscheduler(p, attr, true, true);
4350 EXPORT_SYMBOL_GPL(sched_setattr);
4353 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4354 * @p: the task in question.
4355 * @policy: new policy.
4356 * @param: structure containing the new RT priority.
4358 * Just like sched_setscheduler, only don't bother checking if the
4359 * current context has permission. For example, this is needed in
4360 * stop_machine(): we create temporary high priority worker threads,
4361 * but our caller might not have that capability.
4363 * Return: 0 on success. An error code otherwise.
4365 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4366 const struct sched_param *param)
4368 return _sched_setscheduler(p, policy, param, false);
4370 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4373 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4375 struct sched_param lparam;
4376 struct task_struct *p;
4379 if (!param || pid < 0)
4381 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4386 p = find_process_by_pid(pid);
4388 retval = sched_setscheduler(p, policy, &lparam);
4395 * Mimics kernel/events/core.c perf_copy_attr().
4397 static int sched_copy_attr(struct sched_attr __user *uattr,
4398 struct sched_attr *attr)
4403 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4407 * zero the full structure, so that a short copy will be nice.
4409 memset(attr, 0, sizeof(*attr));
4411 ret = get_user(size, &uattr->size);
4415 if (size > PAGE_SIZE) /* silly large */
4418 if (!size) /* abi compat */
4419 size = SCHED_ATTR_SIZE_VER0;
4421 if (size < SCHED_ATTR_SIZE_VER0)
4425 * If we're handed a bigger struct than we know of,
4426 * ensure all the unknown bits are 0 - i.e. new
4427 * user-space does not rely on any kernel feature
4428 * extensions we dont know about yet.
4430 if (size > sizeof(*attr)) {
4431 unsigned char __user *addr;
4432 unsigned char __user *end;
4435 addr = (void __user *)uattr + sizeof(*attr);
4436 end = (void __user *)uattr + size;
4438 for (; addr < end; addr++) {
4439 ret = get_user(val, addr);
4445 size = sizeof(*attr);
4448 ret = copy_from_user(attr, uattr, size);
4453 * XXX: do we want to be lenient like existing syscalls; or do we want
4454 * to be strict and return an error on out-of-bounds values?
4456 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4461 put_user(sizeof(*attr), &uattr->size);
4466 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4467 * @pid: the pid in question.
4468 * @policy: new policy.
4469 * @param: structure containing the new RT priority.
4471 * Return: 0 on success. An error code otherwise.
4473 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4474 struct sched_param __user *, param)
4476 /* negative values for policy are not valid */
4480 return do_sched_setscheduler(pid, policy, param);
4484 * sys_sched_setparam - set/change the RT priority of a thread
4485 * @pid: the pid in question.
4486 * @param: structure containing the new RT priority.
4488 * Return: 0 on success. An error code otherwise.
4490 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4492 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4496 * sys_sched_setattr - same as above, but with extended sched_attr
4497 * @pid: the pid in question.
4498 * @uattr: structure containing the extended parameters.
4499 * @flags: for future extension.
4501 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4502 unsigned int, flags)
4504 struct sched_attr attr;
4505 struct task_struct *p;
4508 if (!uattr || pid < 0 || flags)
4511 retval = sched_copy_attr(uattr, &attr);
4515 if ((int)attr.sched_policy < 0)
4520 p = find_process_by_pid(pid);
4522 retval = sched_setattr(p, &attr);
4529 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4530 * @pid: the pid in question.
4532 * Return: On success, the policy of the thread. Otherwise, a negative error
4535 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4537 struct task_struct *p;
4545 p = find_process_by_pid(pid);
4547 retval = security_task_getscheduler(p);
4550 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4557 * sys_sched_getparam - get the RT priority of a thread
4558 * @pid: the pid in question.
4559 * @param: structure containing the RT priority.
4561 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4564 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4566 struct sched_param lp = { .sched_priority = 0 };
4567 struct task_struct *p;
4570 if (!param || pid < 0)
4574 p = find_process_by_pid(pid);
4579 retval = security_task_getscheduler(p);
4583 if (task_has_rt_policy(p))
4584 lp.sched_priority = p->rt_priority;
4588 * This one might sleep, we cannot do it with a spinlock held ...
4590 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4599 static int sched_read_attr(struct sched_attr __user *uattr,
4600 struct sched_attr *attr,
4605 if (!access_ok(VERIFY_WRITE, uattr, usize))
4609 * If we're handed a smaller struct than we know of,
4610 * ensure all the unknown bits are 0 - i.e. old
4611 * user-space does not get uncomplete information.
4613 if (usize < sizeof(*attr)) {
4614 unsigned char *addr;
4617 addr = (void *)attr + usize;
4618 end = (void *)attr + sizeof(*attr);
4620 for (; addr < end; addr++) {
4628 ret = copy_to_user(uattr, attr, attr->size);
4636 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4637 * @pid: the pid in question.
4638 * @uattr: structure containing the extended parameters.
4639 * @size: sizeof(attr) for fwd/bwd comp.
4640 * @flags: for future extension.
4642 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4643 unsigned int, size, unsigned int, flags)
4645 struct sched_attr attr = {
4646 .size = sizeof(struct sched_attr),
4648 struct task_struct *p;
4651 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4652 size < SCHED_ATTR_SIZE_VER0 || flags)
4656 p = find_process_by_pid(pid);
4661 retval = security_task_getscheduler(p);
4665 attr.sched_policy = p->policy;
4666 if (p->sched_reset_on_fork)
4667 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4668 if (task_has_dl_policy(p))
4669 __getparam_dl(p, &attr);
4670 else if (task_has_rt_policy(p))
4671 attr.sched_priority = p->rt_priority;
4673 attr.sched_nice = task_nice(p);
4677 retval = sched_read_attr(uattr, &attr, size);
4685 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4687 cpumask_var_t cpus_allowed, new_mask;
4688 struct task_struct *p;
4693 p = find_process_by_pid(pid);
4699 /* Prevent p going away */
4703 if (p->flags & PF_NO_SETAFFINITY) {
4707 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4711 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4713 goto out_free_cpus_allowed;
4716 if (!check_same_owner(p)) {
4718 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4720 goto out_free_new_mask;
4725 retval = security_task_setscheduler(p);
4727 goto out_free_new_mask;
4730 cpuset_cpus_allowed(p, cpus_allowed);
4731 cpumask_and(new_mask, in_mask, cpus_allowed);
4734 * Since bandwidth control happens on root_domain basis,
4735 * if admission test is enabled, we only admit -deadline
4736 * tasks allowed to run on all the CPUs in the task's
4740 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4742 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4745 goto out_free_new_mask;
4751 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4754 cpuset_cpus_allowed(p, cpus_allowed);
4755 if (!cpumask_subset(new_mask, cpus_allowed)) {
4757 * We must have raced with a concurrent cpuset
4758 * update. Just reset the cpus_allowed to the
4759 * cpuset's cpus_allowed
4761 cpumask_copy(new_mask, cpus_allowed);
4766 free_cpumask_var(new_mask);
4767 out_free_cpus_allowed:
4768 free_cpumask_var(cpus_allowed);
4774 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4775 struct cpumask *new_mask)
4777 if (len < cpumask_size())
4778 cpumask_clear(new_mask);
4779 else if (len > cpumask_size())
4780 len = cpumask_size();
4782 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4786 * sys_sched_setaffinity - set the cpu affinity of a process
4787 * @pid: pid of the process
4788 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4789 * @user_mask_ptr: user-space pointer to the new cpu mask
4791 * Return: 0 on success. An error code otherwise.
4793 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4794 unsigned long __user *, user_mask_ptr)
4796 cpumask_var_t new_mask;
4799 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4802 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4804 retval = sched_setaffinity(pid, new_mask);
4805 free_cpumask_var(new_mask);
4809 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4811 struct task_struct *p;
4812 unsigned long flags;
4818 p = find_process_by_pid(pid);
4822 retval = security_task_getscheduler(p);
4826 raw_spin_lock_irqsave(&p->pi_lock, flags);
4827 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4828 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4837 * sys_sched_getaffinity - get the cpu affinity of a process
4838 * @pid: pid of the process
4839 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4840 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4842 * Return: size of CPU mask copied to user_mask_ptr on success. An
4843 * error code otherwise.
4845 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4846 unsigned long __user *, user_mask_ptr)
4851 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4853 if (len & (sizeof(unsigned long)-1))
4856 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4859 ret = sched_getaffinity(pid, mask);
4861 size_t retlen = min_t(size_t, len, cpumask_size());
4863 if (copy_to_user(user_mask_ptr, mask, retlen))
4868 free_cpumask_var(mask);
4874 * sys_sched_yield - yield the current processor to other threads.
4876 * This function yields the current CPU to other tasks. If there are no
4877 * other threads running on this CPU then this function will return.
4881 SYSCALL_DEFINE0(sched_yield)
4883 struct rq *rq = this_rq_lock();
4885 schedstat_inc(rq->yld_count);
4886 current->sched_class->yield_task(rq);
4889 * Since we are going to call schedule() anyway, there's
4890 * no need to preempt or enable interrupts:
4892 __release(rq->lock);
4893 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4894 do_raw_spin_unlock(&rq->lock);
4895 sched_preempt_enable_no_resched();
4902 #ifndef CONFIG_PREEMPT
4903 int __sched _cond_resched(void)
4905 if (should_resched(0)) {
4906 preempt_schedule_common();
4911 EXPORT_SYMBOL(_cond_resched);
4915 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4916 * call schedule, and on return reacquire the lock.
4918 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4919 * operations here to prevent schedule() from being called twice (once via
4920 * spin_unlock(), once by hand).
4922 int __cond_resched_lock(spinlock_t *lock)
4924 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4927 lockdep_assert_held(lock);
4929 if (spin_needbreak(lock) || resched) {
4932 preempt_schedule_common();
4940 EXPORT_SYMBOL(__cond_resched_lock);
4942 int __sched __cond_resched_softirq(void)
4944 BUG_ON(!in_softirq());
4946 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4948 preempt_schedule_common();
4954 EXPORT_SYMBOL(__cond_resched_softirq);
4957 * yield - yield the current processor to other threads.
4959 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4961 * The scheduler is at all times free to pick the calling task as the most
4962 * eligible task to run, if removing the yield() call from your code breaks
4963 * it, its already broken.
4965 * Typical broken usage is:
4970 * where one assumes that yield() will let 'the other' process run that will
4971 * make event true. If the current task is a SCHED_FIFO task that will never
4972 * happen. Never use yield() as a progress guarantee!!
4974 * If you want to use yield() to wait for something, use wait_event().
4975 * If you want to use yield() to be 'nice' for others, use cond_resched().
4976 * If you still want to use yield(), do not!
4978 void __sched yield(void)
4980 set_current_state(TASK_RUNNING);
4983 EXPORT_SYMBOL(yield);
4986 * yield_to - yield the current processor to another thread in
4987 * your thread group, or accelerate that thread toward the
4988 * processor it's on.
4990 * @preempt: whether task preemption is allowed or not
4992 * It's the caller's job to ensure that the target task struct
4993 * can't go away on us before we can do any checks.
4996 * true (>0) if we indeed boosted the target task.
4997 * false (0) if we failed to boost the target.
4998 * -ESRCH if there's no task to yield to.
5000 int __sched yield_to(struct task_struct *p, bool preempt)
5002 struct task_struct *curr = current;
5003 struct rq *rq, *p_rq;
5004 unsigned long flags;
5007 local_irq_save(flags);
5013 * If we're the only runnable task on the rq and target rq also
5014 * has only one task, there's absolutely no point in yielding.
5016 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5021 double_rq_lock(rq, p_rq);
5022 if (task_rq(p) != p_rq) {
5023 double_rq_unlock(rq, p_rq);
5027 if (!curr->sched_class->yield_to_task)
5030 if (curr->sched_class != p->sched_class)
5033 if (task_running(p_rq, p) || p->state)
5036 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5038 schedstat_inc(rq->yld_count);
5040 * Make p's CPU reschedule; pick_next_entity takes care of
5043 if (preempt && rq != p_rq)
5048 double_rq_unlock(rq, p_rq);
5050 local_irq_restore(flags);
5057 EXPORT_SYMBOL_GPL(yield_to);
5060 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5061 * that process accounting knows that this is a task in IO wait state.
5063 long __sched io_schedule_timeout(long timeout)
5065 int old_iowait = current->in_iowait;
5069 current->in_iowait = 1;
5070 blk_schedule_flush_plug(current);
5072 delayacct_blkio_start();
5074 atomic_inc(&rq->nr_iowait);
5075 ret = schedule_timeout(timeout);
5076 current->in_iowait = old_iowait;
5077 atomic_dec(&rq->nr_iowait);
5078 delayacct_blkio_end();
5082 EXPORT_SYMBOL(io_schedule_timeout);
5085 * sys_sched_get_priority_max - return maximum RT priority.
5086 * @policy: scheduling class.
5088 * Return: On success, this syscall returns the maximum
5089 * rt_priority that can be used by a given scheduling class.
5090 * On failure, a negative error code is returned.
5092 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5099 ret = MAX_USER_RT_PRIO-1;
5101 case SCHED_DEADLINE:
5112 * sys_sched_get_priority_min - return minimum RT priority.
5113 * @policy: scheduling class.
5115 * Return: On success, this syscall returns the minimum
5116 * rt_priority that can be used by a given scheduling class.
5117 * On failure, a negative error code is returned.
5119 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5128 case SCHED_DEADLINE:
5138 * sys_sched_rr_get_interval - return the default timeslice of a process.
5139 * @pid: pid of the process.
5140 * @interval: userspace pointer to the timeslice value.
5142 * this syscall writes the default timeslice value of a given process
5143 * into the user-space timespec buffer. A value of '0' means infinity.
5145 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5148 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5149 struct timespec __user *, interval)
5151 struct task_struct *p;
5152 unsigned int time_slice;
5163 p = find_process_by_pid(pid);
5167 retval = security_task_getscheduler(p);
5171 rq = task_rq_lock(p, &rf);
5173 if (p->sched_class->get_rr_interval)
5174 time_slice = p->sched_class->get_rr_interval(rq, p);
5175 task_rq_unlock(rq, p, &rf);
5178 jiffies_to_timespec(time_slice, &t);
5179 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5187 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5189 void sched_show_task(struct task_struct *p)
5191 unsigned long free = 0;
5193 unsigned long state = p->state;
5195 if (!try_get_task_stack(p))
5198 state = __ffs(state) + 1;
5199 printk(KERN_INFO "%-15.15s %c", p->comm,
5200 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5201 if (state == TASK_RUNNING)
5202 printk(KERN_CONT " running task ");
5203 #ifdef CONFIG_DEBUG_STACK_USAGE
5204 free = stack_not_used(p);
5209 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5211 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5212 task_pid_nr(p), ppid,
5213 (unsigned long)task_thread_info(p)->flags);
5215 print_worker_info(KERN_INFO, p);
5216 show_stack(p, NULL);
5220 void show_state_filter(unsigned long state_filter)
5222 struct task_struct *g, *p;
5224 #if BITS_PER_LONG == 32
5226 " task PC stack pid father\n");
5229 " task PC stack pid father\n");
5232 for_each_process_thread(g, p) {
5234 * reset the NMI-timeout, listing all files on a slow
5235 * console might take a lot of time:
5236 * Also, reset softlockup watchdogs on all CPUs, because
5237 * another CPU might be blocked waiting for us to process
5240 touch_nmi_watchdog();
5241 touch_all_softlockup_watchdogs();
5242 if (!state_filter || (p->state & state_filter))
5246 #ifdef CONFIG_SCHED_DEBUG
5248 sysrq_sched_debug_show();
5252 * Only show locks if all tasks are dumped:
5255 debug_show_all_locks();
5258 void init_idle_bootup_task(struct task_struct *idle)
5260 idle->sched_class = &idle_sched_class;
5264 * init_idle - set up an idle thread for a given CPU
5265 * @idle: task in question
5266 * @cpu: cpu the idle task belongs to
5268 * NOTE: this function does not set the idle thread's NEED_RESCHED
5269 * flag, to make booting more robust.
5271 void init_idle(struct task_struct *idle, int cpu)
5273 struct rq *rq = cpu_rq(cpu);
5274 unsigned long flags;
5276 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5277 raw_spin_lock(&rq->lock);
5279 __sched_fork(0, idle);
5280 idle->state = TASK_RUNNING;
5281 idle->se.exec_start = sched_clock();
5283 kasan_unpoison_task_stack(idle);
5287 * Its possible that init_idle() gets called multiple times on a task,
5288 * in that case do_set_cpus_allowed() will not do the right thing.
5290 * And since this is boot we can forgo the serialization.
5292 set_cpus_allowed_common(idle, cpumask_of(cpu));
5295 * We're having a chicken and egg problem, even though we are
5296 * holding rq->lock, the cpu isn't yet set to this cpu so the
5297 * lockdep check in task_group() will fail.
5299 * Similar case to sched_fork(). / Alternatively we could
5300 * use task_rq_lock() here and obtain the other rq->lock.
5305 __set_task_cpu(idle, cpu);
5308 rq->curr = rq->idle = idle;
5309 idle->on_rq = TASK_ON_RQ_QUEUED;
5313 raw_spin_unlock(&rq->lock);
5314 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5316 /* Set the preempt count _outside_ the spinlocks! */
5317 init_idle_preempt_count(idle, cpu);
5320 * The idle tasks have their own, simple scheduling class:
5322 idle->sched_class = &idle_sched_class;
5323 ftrace_graph_init_idle_task(idle, cpu);
5324 vtime_init_idle(idle, cpu);
5326 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5330 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5331 const struct cpumask *trial)
5333 int ret = 1, trial_cpus;
5334 struct dl_bw *cur_dl_b;
5335 unsigned long flags;
5337 if (!cpumask_weight(cur))
5340 rcu_read_lock_sched();
5341 cur_dl_b = dl_bw_of(cpumask_any(cur));
5342 trial_cpus = cpumask_weight(trial);
5344 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5345 if (cur_dl_b->bw != -1 &&
5346 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5348 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5349 rcu_read_unlock_sched();
5354 int task_can_attach(struct task_struct *p,
5355 const struct cpumask *cs_cpus_allowed)
5360 * Kthreads which disallow setaffinity shouldn't be moved
5361 * to a new cpuset; we don't want to change their cpu
5362 * affinity and isolating such threads by their set of
5363 * allowed nodes is unnecessary. Thus, cpusets are not
5364 * applicable for such threads. This prevents checking for
5365 * success of set_cpus_allowed_ptr() on all attached tasks
5366 * before cpus_allowed may be changed.
5368 if (p->flags & PF_NO_SETAFFINITY) {
5374 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5376 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5381 unsigned long flags;
5383 rcu_read_lock_sched();
5384 dl_b = dl_bw_of(dest_cpu);
5385 raw_spin_lock_irqsave(&dl_b->lock, flags);
5386 cpus = dl_bw_cpus(dest_cpu);
5387 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5392 * We reserve space for this task in the destination
5393 * root_domain, as we can't fail after this point.
5394 * We will free resources in the source root_domain
5395 * later on (see set_cpus_allowed_dl()).
5397 __dl_add(dl_b, p->dl.dl_bw);
5399 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5400 rcu_read_unlock_sched();
5410 static bool sched_smp_initialized __read_mostly;
5412 #ifdef CONFIG_NUMA_BALANCING
5413 /* Migrate current task p to target_cpu */
5414 int migrate_task_to(struct task_struct *p, int target_cpu)
5416 struct migration_arg arg = { p, target_cpu };
5417 int curr_cpu = task_cpu(p);
5419 if (curr_cpu == target_cpu)
5422 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5425 /* TODO: This is not properly updating schedstats */
5427 trace_sched_move_numa(p, curr_cpu, target_cpu);
5428 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5432 * Requeue a task on a given node and accurately track the number of NUMA
5433 * tasks on the runqueues
5435 void sched_setnuma(struct task_struct *p, int nid)
5437 bool queued, running;
5441 rq = task_rq_lock(p, &rf);
5442 queued = task_on_rq_queued(p);
5443 running = task_current(rq, p);
5446 dequeue_task(rq, p, DEQUEUE_SAVE);
5448 put_prev_task(rq, p);
5450 p->numa_preferred_nid = nid;
5453 enqueue_task(rq, p, ENQUEUE_RESTORE);
5455 set_curr_task(rq, p);
5456 task_rq_unlock(rq, p, &rf);
5458 #endif /* CONFIG_NUMA_BALANCING */
5460 #ifdef CONFIG_HOTPLUG_CPU
5462 * Ensures that the idle task is using init_mm right before its cpu goes
5465 void idle_task_exit(void)
5467 struct mm_struct *mm = current->active_mm;
5469 BUG_ON(cpu_online(smp_processor_id()));
5471 if (mm != &init_mm) {
5472 switch_mm_irqs_off(mm, &init_mm, current);
5473 finish_arch_post_lock_switch();
5479 * Since this CPU is going 'away' for a while, fold any nr_active delta
5480 * we might have. Assumes we're called after migrate_tasks() so that the
5481 * nr_active count is stable. We need to take the teardown thread which
5482 * is calling this into account, so we hand in adjust = 1 to the load
5485 * Also see the comment "Global load-average calculations".
5487 static void calc_load_migrate(struct rq *rq)
5489 long delta = calc_load_fold_active(rq, 1);
5491 atomic_long_add(delta, &calc_load_tasks);
5494 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5498 static const struct sched_class fake_sched_class = {
5499 .put_prev_task = put_prev_task_fake,
5502 static struct task_struct fake_task = {
5504 * Avoid pull_{rt,dl}_task()
5506 .prio = MAX_PRIO + 1,
5507 .sched_class = &fake_sched_class,
5511 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5512 * try_to_wake_up()->select_task_rq().
5514 * Called with rq->lock held even though we'er in stop_machine() and
5515 * there's no concurrency possible, we hold the required locks anyway
5516 * because of lock validation efforts.
5518 static void migrate_tasks(struct rq *dead_rq)
5520 struct rq *rq = dead_rq;
5521 struct task_struct *next, *stop = rq->stop;
5522 struct pin_cookie cookie;
5526 * Fudge the rq selection such that the below task selection loop
5527 * doesn't get stuck on the currently eligible stop task.
5529 * We're currently inside stop_machine() and the rq is either stuck
5530 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5531 * either way we should never end up calling schedule() until we're
5537 * put_prev_task() and pick_next_task() sched
5538 * class method both need to have an up-to-date
5539 * value of rq->clock[_task]
5541 update_rq_clock(rq);
5545 * There's this thread running, bail when that's the only
5548 if (rq->nr_running == 1)
5552 * pick_next_task assumes pinned rq->lock.
5554 cookie = lockdep_pin_lock(&rq->lock);
5555 next = pick_next_task(rq, &fake_task, cookie);
5557 next->sched_class->put_prev_task(rq, next);
5560 * Rules for changing task_struct::cpus_allowed are holding
5561 * both pi_lock and rq->lock, such that holding either
5562 * stabilizes the mask.
5564 * Drop rq->lock is not quite as disastrous as it usually is
5565 * because !cpu_active at this point, which means load-balance
5566 * will not interfere. Also, stop-machine.
5568 lockdep_unpin_lock(&rq->lock, cookie);
5569 raw_spin_unlock(&rq->lock);
5570 raw_spin_lock(&next->pi_lock);
5571 raw_spin_lock(&rq->lock);
5574 * Since we're inside stop-machine, _nothing_ should have
5575 * changed the task, WARN if weird stuff happened, because in
5576 * that case the above rq->lock drop is a fail too.
5578 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5579 raw_spin_unlock(&next->pi_lock);
5583 /* Find suitable destination for @next, with force if needed. */
5584 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5586 rq = __migrate_task(rq, next, dest_cpu);
5587 if (rq != dead_rq) {
5588 raw_spin_unlock(&rq->lock);
5590 raw_spin_lock(&rq->lock);
5592 raw_spin_unlock(&next->pi_lock);
5597 #endif /* CONFIG_HOTPLUG_CPU */
5599 static void set_rq_online(struct rq *rq)
5602 const struct sched_class *class;
5604 cpumask_set_cpu(rq->cpu, rq->rd->online);
5607 for_each_class(class) {
5608 if (class->rq_online)
5609 class->rq_online(rq);
5614 static void set_rq_offline(struct rq *rq)
5617 const struct sched_class *class;
5619 for_each_class(class) {
5620 if (class->rq_offline)
5621 class->rq_offline(rq);
5624 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5629 static void set_cpu_rq_start_time(unsigned int cpu)
5631 struct rq *rq = cpu_rq(cpu);
5633 rq->age_stamp = sched_clock_cpu(cpu);
5636 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5638 #ifdef CONFIG_SCHED_DEBUG
5640 static __read_mostly int sched_debug_enabled;
5642 static int __init sched_debug_setup(char *str)
5644 sched_debug_enabled = 1;
5648 early_param("sched_debug", sched_debug_setup);
5650 static inline bool sched_debug(void)
5652 return sched_debug_enabled;
5655 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5656 struct cpumask *groupmask)
5658 struct sched_group *group = sd->groups;
5660 cpumask_clear(groupmask);
5662 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5664 if (!(sd->flags & SD_LOAD_BALANCE)) {
5665 printk("does not load-balance\n");
5667 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5672 printk(KERN_CONT "span %*pbl level %s\n",
5673 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5675 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5676 printk(KERN_ERR "ERROR: domain->span does not contain "
5679 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5680 printk(KERN_ERR "ERROR: domain->groups does not contain"
5684 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5688 printk(KERN_ERR "ERROR: group is NULL\n");
5692 if (!cpumask_weight(sched_group_cpus(group))) {
5693 printk(KERN_CONT "\n");
5694 printk(KERN_ERR "ERROR: empty group\n");
5698 if (!(sd->flags & SD_OVERLAP) &&
5699 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5700 printk(KERN_CONT "\n");
5701 printk(KERN_ERR "ERROR: repeated CPUs\n");
5705 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5707 printk(KERN_CONT " %*pbl",
5708 cpumask_pr_args(sched_group_cpus(group)));
5709 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5710 printk(KERN_CONT " (cpu_capacity = %d)",
5711 group->sgc->capacity);
5714 group = group->next;
5715 } while (group != sd->groups);
5716 printk(KERN_CONT "\n");
5718 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5719 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5722 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5723 printk(KERN_ERR "ERROR: parent span is not a superset "
5724 "of domain->span\n");
5728 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5732 if (!sched_debug_enabled)
5736 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5740 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5743 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5751 #else /* !CONFIG_SCHED_DEBUG */
5753 # define sched_debug_enabled 0
5754 # define sched_domain_debug(sd, cpu) do { } while (0)
5755 static inline bool sched_debug(void)
5759 #endif /* CONFIG_SCHED_DEBUG */
5761 static int sd_degenerate(struct sched_domain *sd)
5763 if (cpumask_weight(sched_domain_span(sd)) == 1)
5766 /* Following flags need at least 2 groups */
5767 if (sd->flags & (SD_LOAD_BALANCE |
5768 SD_BALANCE_NEWIDLE |
5771 SD_SHARE_CPUCAPACITY |
5772 SD_ASYM_CPUCAPACITY |
5773 SD_SHARE_PKG_RESOURCES |
5774 SD_SHARE_POWERDOMAIN)) {
5775 if (sd->groups != sd->groups->next)
5779 /* Following flags don't use groups */
5780 if (sd->flags & (SD_WAKE_AFFINE))
5787 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5789 unsigned long cflags = sd->flags, pflags = parent->flags;
5791 if (sd_degenerate(parent))
5794 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5797 /* Flags needing groups don't count if only 1 group in parent */
5798 if (parent->groups == parent->groups->next) {
5799 pflags &= ~(SD_LOAD_BALANCE |
5800 SD_BALANCE_NEWIDLE |
5803 SD_ASYM_CPUCAPACITY |
5804 SD_SHARE_CPUCAPACITY |
5805 SD_SHARE_PKG_RESOURCES |
5807 SD_SHARE_POWERDOMAIN);
5808 if (nr_node_ids == 1)
5809 pflags &= ~SD_SERIALIZE;
5811 if (~cflags & pflags)
5817 static void free_rootdomain(struct rcu_head *rcu)
5819 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5821 cpupri_cleanup(&rd->cpupri);
5822 cpudl_cleanup(&rd->cpudl);
5823 free_cpumask_var(rd->dlo_mask);
5824 free_cpumask_var(rd->rto_mask);
5825 free_cpumask_var(rd->online);
5826 free_cpumask_var(rd->span);
5830 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5832 struct root_domain *old_rd = NULL;
5833 unsigned long flags;
5835 raw_spin_lock_irqsave(&rq->lock, flags);
5840 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5843 cpumask_clear_cpu(rq->cpu, old_rd->span);
5846 * If we dont want to free the old_rd yet then
5847 * set old_rd to NULL to skip the freeing later
5850 if (!atomic_dec_and_test(&old_rd->refcount))
5854 atomic_inc(&rd->refcount);
5857 cpumask_set_cpu(rq->cpu, rd->span);
5858 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5861 raw_spin_unlock_irqrestore(&rq->lock, flags);
5864 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5867 static int init_rootdomain(struct root_domain *rd)
5869 memset(rd, 0, sizeof(*rd));
5871 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5873 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5875 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5877 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5880 init_dl_bw(&rd->dl_bw);
5881 if (cpudl_init(&rd->cpudl) != 0)
5884 if (cpupri_init(&rd->cpupri) != 0)
5889 free_cpumask_var(rd->rto_mask);
5891 free_cpumask_var(rd->dlo_mask);
5893 free_cpumask_var(rd->online);
5895 free_cpumask_var(rd->span);
5901 * By default the system creates a single root-domain with all cpus as
5902 * members (mimicking the global state we have today).
5904 struct root_domain def_root_domain;
5906 static void init_defrootdomain(void)
5908 init_rootdomain(&def_root_domain);
5910 atomic_set(&def_root_domain.refcount, 1);
5913 static struct root_domain *alloc_rootdomain(void)
5915 struct root_domain *rd;
5917 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5921 if (init_rootdomain(rd) != 0) {
5929 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5931 struct sched_group *tmp, *first;
5940 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5945 } while (sg != first);
5948 static void destroy_sched_domain(struct sched_domain *sd)
5951 * If its an overlapping domain it has private groups, iterate and
5954 if (sd->flags & SD_OVERLAP) {
5955 free_sched_groups(sd->groups, 1);
5956 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5957 kfree(sd->groups->sgc);
5960 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
5965 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
5967 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5970 struct sched_domain *parent = sd->parent;
5971 destroy_sched_domain(sd);
5976 static void destroy_sched_domains(struct sched_domain *sd)
5979 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
5983 * Keep a special pointer to the highest sched_domain that has
5984 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5985 * allows us to avoid some pointer chasing select_idle_sibling().
5987 * Also keep a unique ID per domain (we use the first cpu number in
5988 * the cpumask of the domain), this allows us to quickly tell if
5989 * two cpus are in the same cache domain, see cpus_share_cache().
5991 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5992 DEFINE_PER_CPU(int, sd_llc_size);
5993 DEFINE_PER_CPU(int, sd_llc_id);
5994 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
5995 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5996 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5998 static void update_top_cache_domain(int cpu)
6000 struct sched_domain_shared *sds = NULL;
6001 struct sched_domain *sd;
6005 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6007 id = cpumask_first(sched_domain_span(sd));
6008 size = cpumask_weight(sched_domain_span(sd));
6012 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6013 per_cpu(sd_llc_size, cpu) = size;
6014 per_cpu(sd_llc_id, cpu) = id;
6015 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
6017 sd = lowest_flag_domain(cpu, SD_NUMA);
6018 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6020 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6021 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6025 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6026 * hold the hotplug lock.
6029 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6031 struct rq *rq = cpu_rq(cpu);
6032 struct sched_domain *tmp;
6034 /* Remove the sched domains which do not contribute to scheduling. */
6035 for (tmp = sd; tmp; ) {
6036 struct sched_domain *parent = tmp->parent;
6040 if (sd_parent_degenerate(tmp, parent)) {
6041 tmp->parent = parent->parent;
6043 parent->parent->child = tmp;
6045 * Transfer SD_PREFER_SIBLING down in case of a
6046 * degenerate parent; the spans match for this
6047 * so the property transfers.
6049 if (parent->flags & SD_PREFER_SIBLING)
6050 tmp->flags |= SD_PREFER_SIBLING;
6051 destroy_sched_domain(parent);
6056 if (sd && sd_degenerate(sd)) {
6059 destroy_sched_domain(tmp);
6064 sched_domain_debug(sd, cpu);
6066 rq_attach_root(rq, rd);
6068 rcu_assign_pointer(rq->sd, sd);
6069 destroy_sched_domains(tmp);
6071 update_top_cache_domain(cpu);
6074 /* Setup the mask of cpus configured for isolated domains */
6075 static int __init isolated_cpu_setup(char *str)
6079 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6080 ret = cpulist_parse(str, cpu_isolated_map);
6082 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6087 __setup("isolcpus=", isolated_cpu_setup);
6090 struct sched_domain ** __percpu sd;
6091 struct root_domain *rd;
6102 * Build an iteration mask that can exclude certain CPUs from the upwards
6105 * Asymmetric node setups can result in situations where the domain tree is of
6106 * unequal depth, make sure to skip domains that already cover the entire
6109 * In that case build_sched_domains() will have terminated the iteration early
6110 * and our sibling sd spans will be empty. Domains should always include the
6111 * cpu they're built on, so check that.
6114 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6116 const struct cpumask *span = sched_domain_span(sd);
6117 struct sd_data *sdd = sd->private;
6118 struct sched_domain *sibling;
6121 for_each_cpu(i, span) {
6122 sibling = *per_cpu_ptr(sdd->sd, i);
6123 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6126 cpumask_set_cpu(i, sched_group_mask(sg));
6131 * Return the canonical balance cpu for this group, this is the first cpu
6132 * of this group that's also in the iteration mask.
6134 int group_balance_cpu(struct sched_group *sg)
6136 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6140 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6142 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6143 const struct cpumask *span = sched_domain_span(sd);
6144 struct cpumask *covered = sched_domains_tmpmask;
6145 struct sd_data *sdd = sd->private;
6146 struct sched_domain *sibling;
6149 cpumask_clear(covered);
6151 for_each_cpu(i, span) {
6152 struct cpumask *sg_span;
6154 if (cpumask_test_cpu(i, covered))
6157 sibling = *per_cpu_ptr(sdd->sd, i);
6159 /* See the comment near build_group_mask(). */
6160 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6163 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6164 GFP_KERNEL, cpu_to_node(cpu));
6169 sg_span = sched_group_cpus(sg);
6171 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6173 cpumask_set_cpu(i, sg_span);
6175 cpumask_or(covered, covered, sg_span);
6177 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6178 if (atomic_inc_return(&sg->sgc->ref) == 1)
6179 build_group_mask(sd, sg);
6182 * Initialize sgc->capacity such that even if we mess up the
6183 * domains and no possible iteration will get us here, we won't
6186 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6189 * Make sure the first group of this domain contains the
6190 * canonical balance cpu. Otherwise the sched_domain iteration
6191 * breaks. See update_sg_lb_stats().
6193 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6194 group_balance_cpu(sg) == cpu)
6204 sd->groups = groups;
6209 free_sched_groups(first, 0);
6214 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6216 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6217 struct sched_domain *child = sd->child;
6220 cpu = cpumask_first(sched_domain_span(child));
6223 *sg = *per_cpu_ptr(sdd->sg, cpu);
6224 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6225 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6232 * build_sched_groups will build a circular linked list of the groups
6233 * covered by the given span, and will set each group's ->cpumask correctly,
6234 * and ->cpu_capacity to 0.
6236 * Assumes the sched_domain tree is fully constructed
6239 build_sched_groups(struct sched_domain *sd, int cpu)
6241 struct sched_group *first = NULL, *last = NULL;
6242 struct sd_data *sdd = sd->private;
6243 const struct cpumask *span = sched_domain_span(sd);
6244 struct cpumask *covered;
6247 get_group(cpu, sdd, &sd->groups);
6248 atomic_inc(&sd->groups->ref);
6250 if (cpu != cpumask_first(span))
6253 lockdep_assert_held(&sched_domains_mutex);
6254 covered = sched_domains_tmpmask;
6256 cpumask_clear(covered);
6258 for_each_cpu(i, span) {
6259 struct sched_group *sg;
6262 if (cpumask_test_cpu(i, covered))
6265 group = get_group(i, sdd, &sg);
6266 cpumask_setall(sched_group_mask(sg));
6268 for_each_cpu(j, span) {
6269 if (get_group(j, sdd, NULL) != group)
6272 cpumask_set_cpu(j, covered);
6273 cpumask_set_cpu(j, sched_group_cpus(sg));
6288 * Initialize sched groups cpu_capacity.
6290 * cpu_capacity indicates the capacity of sched group, which is used while
6291 * distributing the load between different sched groups in a sched domain.
6292 * Typically cpu_capacity for all the groups in a sched domain will be same
6293 * unless there are asymmetries in the topology. If there are asymmetries,
6294 * group having more cpu_capacity will pickup more load compared to the
6295 * group having less cpu_capacity.
6297 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6299 struct sched_group *sg = sd->groups;
6304 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6306 } while (sg != sd->groups);
6308 if (cpu != group_balance_cpu(sg))
6311 update_group_capacity(sd, cpu);
6315 * Initializers for schedule domains
6316 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6319 static int default_relax_domain_level = -1;
6320 int sched_domain_level_max;
6322 static int __init setup_relax_domain_level(char *str)
6324 if (kstrtoint(str, 0, &default_relax_domain_level))
6325 pr_warn("Unable to set relax_domain_level\n");
6329 __setup("relax_domain_level=", setup_relax_domain_level);
6331 static void set_domain_attribute(struct sched_domain *sd,
6332 struct sched_domain_attr *attr)
6336 if (!attr || attr->relax_domain_level < 0) {
6337 if (default_relax_domain_level < 0)
6340 request = default_relax_domain_level;
6342 request = attr->relax_domain_level;
6343 if (request < sd->level) {
6344 /* turn off idle balance on this domain */
6345 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6347 /* turn on idle balance on this domain */
6348 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6352 static void __sdt_free(const struct cpumask *cpu_map);
6353 static int __sdt_alloc(const struct cpumask *cpu_map);
6355 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6356 const struct cpumask *cpu_map)
6360 if (!atomic_read(&d->rd->refcount))
6361 free_rootdomain(&d->rd->rcu); /* fall through */
6363 free_percpu(d->sd); /* fall through */
6365 __sdt_free(cpu_map); /* fall through */
6371 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6372 const struct cpumask *cpu_map)
6374 memset(d, 0, sizeof(*d));
6376 if (__sdt_alloc(cpu_map))
6377 return sa_sd_storage;
6378 d->sd = alloc_percpu(struct sched_domain *);
6380 return sa_sd_storage;
6381 d->rd = alloc_rootdomain();
6384 return sa_rootdomain;
6388 * NULL the sd_data elements we've used to build the sched_domain and
6389 * sched_group structure so that the subsequent __free_domain_allocs()
6390 * will not free the data we're using.
6392 static void claim_allocations(int cpu, struct sched_domain *sd)
6394 struct sd_data *sdd = sd->private;
6396 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6397 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6399 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
6400 *per_cpu_ptr(sdd->sds, cpu) = NULL;
6402 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6403 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6405 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6406 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6410 static int sched_domains_numa_levels;
6411 enum numa_topology_type sched_numa_topology_type;
6412 static int *sched_domains_numa_distance;
6413 int sched_max_numa_distance;
6414 static struct cpumask ***sched_domains_numa_masks;
6415 static int sched_domains_curr_level;
6419 * SD_flags allowed in topology descriptions.
6421 * These flags are purely descriptive of the topology and do not prescribe
6422 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6425 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6426 * SD_SHARE_PKG_RESOURCES - describes shared caches
6427 * SD_NUMA - describes NUMA topologies
6428 * SD_SHARE_POWERDOMAIN - describes shared power domain
6429 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6431 * Odd one out, which beside describing the topology has a quirk also
6432 * prescribes the desired behaviour that goes along with it:
6434 * SD_ASYM_PACKING - describes SMT quirks
6436 #define TOPOLOGY_SD_FLAGS \
6437 (SD_SHARE_CPUCAPACITY | \
6438 SD_SHARE_PKG_RESOURCES | \
6441 SD_ASYM_CPUCAPACITY | \
6442 SD_SHARE_POWERDOMAIN)
6444 static struct sched_domain *
6445 sd_init(struct sched_domain_topology_level *tl,
6446 const struct cpumask *cpu_map,
6447 struct sched_domain *child, int cpu)
6449 struct sd_data *sdd = &tl->data;
6450 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6451 int sd_id, sd_weight, sd_flags = 0;
6455 * Ugly hack to pass state to sd_numa_mask()...
6457 sched_domains_curr_level = tl->numa_level;
6460 sd_weight = cpumask_weight(tl->mask(cpu));
6463 sd_flags = (*tl->sd_flags)();
6464 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6465 "wrong sd_flags in topology description\n"))
6466 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6468 *sd = (struct sched_domain){
6469 .min_interval = sd_weight,
6470 .max_interval = 2*sd_weight,
6472 .imbalance_pct = 125,
6474 .cache_nice_tries = 0,
6481 .flags = 1*SD_LOAD_BALANCE
6482 | 1*SD_BALANCE_NEWIDLE
6487 | 0*SD_SHARE_CPUCAPACITY
6488 | 0*SD_SHARE_PKG_RESOURCES
6490 | 0*SD_PREFER_SIBLING
6495 .last_balance = jiffies,
6496 .balance_interval = sd_weight,
6498 .max_newidle_lb_cost = 0,
6499 .next_decay_max_lb_cost = jiffies,
6501 #ifdef CONFIG_SCHED_DEBUG
6506 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6507 sd_id = cpumask_first(sched_domain_span(sd));
6510 * Convert topological properties into behaviour.
6513 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6514 struct sched_domain *t = sd;
6516 for_each_lower_domain(t)
6517 t->flags |= SD_BALANCE_WAKE;
6520 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6521 sd->flags |= SD_PREFER_SIBLING;
6522 sd->imbalance_pct = 110;
6523 sd->smt_gain = 1178; /* ~15% */
6525 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6526 sd->imbalance_pct = 117;
6527 sd->cache_nice_tries = 1;
6531 } else if (sd->flags & SD_NUMA) {
6532 sd->cache_nice_tries = 2;
6536 sd->flags |= SD_SERIALIZE;
6537 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6538 sd->flags &= ~(SD_BALANCE_EXEC |
6545 sd->flags |= SD_PREFER_SIBLING;
6546 sd->cache_nice_tries = 1;
6552 * For all levels sharing cache; connect a sched_domain_shared
6555 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6556 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
6557 atomic_inc(&sd->shared->ref);
6558 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
6567 * Topology list, bottom-up.
6569 static struct sched_domain_topology_level default_topology[] = {
6570 #ifdef CONFIG_SCHED_SMT
6571 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6573 #ifdef CONFIG_SCHED_MC
6574 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6576 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6580 static struct sched_domain_topology_level *sched_domain_topology =
6583 #define for_each_sd_topology(tl) \
6584 for (tl = sched_domain_topology; tl->mask; tl++)
6586 void set_sched_topology(struct sched_domain_topology_level *tl)
6588 if (WARN_ON_ONCE(sched_smp_initialized))
6591 sched_domain_topology = tl;
6596 static const struct cpumask *sd_numa_mask(int cpu)
6598 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6601 static void sched_numa_warn(const char *str)
6603 static int done = false;
6611 printk(KERN_WARNING "ERROR: %s\n\n", str);
6613 for (i = 0; i < nr_node_ids; i++) {
6614 printk(KERN_WARNING " ");
6615 for (j = 0; j < nr_node_ids; j++)
6616 printk(KERN_CONT "%02d ", node_distance(i,j));
6617 printk(KERN_CONT "\n");
6619 printk(KERN_WARNING "\n");
6622 bool find_numa_distance(int distance)
6626 if (distance == node_distance(0, 0))
6629 for (i = 0; i < sched_domains_numa_levels; i++) {
6630 if (sched_domains_numa_distance[i] == distance)
6638 * A system can have three types of NUMA topology:
6639 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6640 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6641 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6643 * The difference between a glueless mesh topology and a backplane
6644 * topology lies in whether communication between not directly
6645 * connected nodes goes through intermediary nodes (where programs
6646 * could run), or through backplane controllers. This affects
6647 * placement of programs.
6649 * The type of topology can be discerned with the following tests:
6650 * - If the maximum distance between any nodes is 1 hop, the system
6651 * is directly connected.
6652 * - If for two nodes A and B, located N > 1 hops away from each other,
6653 * there is an intermediary node C, which is < N hops away from both
6654 * nodes A and B, the system is a glueless mesh.
6656 static void init_numa_topology_type(void)
6660 n = sched_max_numa_distance;
6662 if (sched_domains_numa_levels <= 1) {
6663 sched_numa_topology_type = NUMA_DIRECT;
6667 for_each_online_node(a) {
6668 for_each_online_node(b) {
6669 /* Find two nodes furthest removed from each other. */
6670 if (node_distance(a, b) < n)
6673 /* Is there an intermediary node between a and b? */
6674 for_each_online_node(c) {
6675 if (node_distance(a, c) < n &&
6676 node_distance(b, c) < n) {
6677 sched_numa_topology_type =
6683 sched_numa_topology_type = NUMA_BACKPLANE;
6689 static void sched_init_numa(void)
6691 int next_distance, curr_distance = node_distance(0, 0);
6692 struct sched_domain_topology_level *tl;
6696 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6697 if (!sched_domains_numa_distance)
6701 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6702 * unique distances in the node_distance() table.
6704 * Assumes node_distance(0,j) includes all distances in
6705 * node_distance(i,j) in order to avoid cubic time.
6707 next_distance = curr_distance;
6708 for (i = 0; i < nr_node_ids; i++) {
6709 for (j = 0; j < nr_node_ids; j++) {
6710 for (k = 0; k < nr_node_ids; k++) {
6711 int distance = node_distance(i, k);
6713 if (distance > curr_distance &&
6714 (distance < next_distance ||
6715 next_distance == curr_distance))
6716 next_distance = distance;
6719 * While not a strong assumption it would be nice to know
6720 * about cases where if node A is connected to B, B is not
6721 * equally connected to A.
6723 if (sched_debug() && node_distance(k, i) != distance)
6724 sched_numa_warn("Node-distance not symmetric");
6726 if (sched_debug() && i && !find_numa_distance(distance))
6727 sched_numa_warn("Node-0 not representative");
6729 if (next_distance != curr_distance) {
6730 sched_domains_numa_distance[level++] = next_distance;
6731 sched_domains_numa_levels = level;
6732 curr_distance = next_distance;
6737 * In case of sched_debug() we verify the above assumption.
6747 * 'level' contains the number of unique distances, excluding the
6748 * identity distance node_distance(i,i).
6750 * The sched_domains_numa_distance[] array includes the actual distance
6755 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6756 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6757 * the array will contain less then 'level' members. This could be
6758 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6759 * in other functions.
6761 * We reset it to 'level' at the end of this function.
6763 sched_domains_numa_levels = 0;
6765 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6766 if (!sched_domains_numa_masks)
6770 * Now for each level, construct a mask per node which contains all
6771 * cpus of nodes that are that many hops away from us.
6773 for (i = 0; i < level; i++) {
6774 sched_domains_numa_masks[i] =
6775 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6776 if (!sched_domains_numa_masks[i])
6779 for (j = 0; j < nr_node_ids; j++) {
6780 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6784 sched_domains_numa_masks[i][j] = mask;
6787 if (node_distance(j, k) > sched_domains_numa_distance[i])
6790 cpumask_or(mask, mask, cpumask_of_node(k));
6795 /* Compute default topology size */
6796 for (i = 0; sched_domain_topology[i].mask; i++);
6798 tl = kzalloc((i + level + 1) *
6799 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6804 * Copy the default topology bits..
6806 for (i = 0; sched_domain_topology[i].mask; i++)
6807 tl[i] = sched_domain_topology[i];
6810 * .. and append 'j' levels of NUMA goodness.
6812 for (j = 0; j < level; i++, j++) {
6813 tl[i] = (struct sched_domain_topology_level){
6814 .mask = sd_numa_mask,
6815 .sd_flags = cpu_numa_flags,
6816 .flags = SDTL_OVERLAP,
6822 sched_domain_topology = tl;
6824 sched_domains_numa_levels = level;
6825 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6827 init_numa_topology_type();
6830 static void sched_domains_numa_masks_set(unsigned int cpu)
6832 int node = cpu_to_node(cpu);
6835 for (i = 0; i < sched_domains_numa_levels; i++) {
6836 for (j = 0; j < nr_node_ids; j++) {
6837 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6838 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6843 static void sched_domains_numa_masks_clear(unsigned int cpu)
6847 for (i = 0; i < sched_domains_numa_levels; i++) {
6848 for (j = 0; j < nr_node_ids; j++)
6849 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6854 static inline void sched_init_numa(void) { }
6855 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6856 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6857 #endif /* CONFIG_NUMA */
6859 static int __sdt_alloc(const struct cpumask *cpu_map)
6861 struct sched_domain_topology_level *tl;
6864 for_each_sd_topology(tl) {
6865 struct sd_data *sdd = &tl->data;
6867 sdd->sd = alloc_percpu(struct sched_domain *);
6871 sdd->sds = alloc_percpu(struct sched_domain_shared *);
6875 sdd->sg = alloc_percpu(struct sched_group *);
6879 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6883 for_each_cpu(j, cpu_map) {
6884 struct sched_domain *sd;
6885 struct sched_domain_shared *sds;
6886 struct sched_group *sg;
6887 struct sched_group_capacity *sgc;
6889 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6890 GFP_KERNEL, cpu_to_node(j));
6894 *per_cpu_ptr(sdd->sd, j) = sd;
6896 sds = kzalloc_node(sizeof(struct sched_domain_shared),
6897 GFP_KERNEL, cpu_to_node(j));
6901 *per_cpu_ptr(sdd->sds, j) = sds;
6903 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6904 GFP_KERNEL, cpu_to_node(j));
6910 *per_cpu_ptr(sdd->sg, j) = sg;
6912 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6913 GFP_KERNEL, cpu_to_node(j));
6917 *per_cpu_ptr(sdd->sgc, j) = sgc;
6924 static void __sdt_free(const struct cpumask *cpu_map)
6926 struct sched_domain_topology_level *tl;
6929 for_each_sd_topology(tl) {
6930 struct sd_data *sdd = &tl->data;
6932 for_each_cpu(j, cpu_map) {
6933 struct sched_domain *sd;
6936 sd = *per_cpu_ptr(sdd->sd, j);
6937 if (sd && (sd->flags & SD_OVERLAP))
6938 free_sched_groups(sd->groups, 0);
6939 kfree(*per_cpu_ptr(sdd->sd, j));
6943 kfree(*per_cpu_ptr(sdd->sds, j));
6945 kfree(*per_cpu_ptr(sdd->sg, j));
6947 kfree(*per_cpu_ptr(sdd->sgc, j));
6949 free_percpu(sdd->sd);
6951 free_percpu(sdd->sds);
6953 free_percpu(sdd->sg);
6955 free_percpu(sdd->sgc);
6960 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6961 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6962 struct sched_domain *child, int cpu)
6964 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
6967 sd->level = child->level + 1;
6968 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6971 if (!cpumask_subset(sched_domain_span(child),
6972 sched_domain_span(sd))) {
6973 pr_err("BUG: arch topology borken\n");
6974 #ifdef CONFIG_SCHED_DEBUG
6975 pr_err(" the %s domain not a subset of the %s domain\n",
6976 child->name, sd->name);
6978 /* Fixup, ensure @sd has at least @child cpus. */
6979 cpumask_or(sched_domain_span(sd),
6980 sched_domain_span(sd),
6981 sched_domain_span(child));
6985 set_domain_attribute(sd, attr);
6991 * Build sched domains for a given set of cpus and attach the sched domains
6992 * to the individual cpus
6994 static int build_sched_domains(const struct cpumask *cpu_map,
6995 struct sched_domain_attr *attr)
6997 enum s_alloc alloc_state;
6998 struct sched_domain *sd;
7000 struct rq *rq = NULL;
7001 int i, ret = -ENOMEM;
7003 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7004 if (alloc_state != sa_rootdomain)
7007 /* Set up domains for cpus specified by the cpu_map. */
7008 for_each_cpu(i, cpu_map) {
7009 struct sched_domain_topology_level *tl;
7012 for_each_sd_topology(tl) {
7013 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7014 if (tl == sched_domain_topology)
7015 *per_cpu_ptr(d.sd, i) = sd;
7016 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7017 sd->flags |= SD_OVERLAP;
7018 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7023 /* Build the groups for the domains */
7024 for_each_cpu(i, cpu_map) {
7025 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7026 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7027 if (sd->flags & SD_OVERLAP) {
7028 if (build_overlap_sched_groups(sd, i))
7031 if (build_sched_groups(sd, i))
7037 /* Calculate CPU capacity for physical packages and nodes */
7038 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7039 if (!cpumask_test_cpu(i, cpu_map))
7042 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7043 claim_allocations(i, sd);
7044 init_sched_groups_capacity(i, sd);
7048 /* Attach the domains */
7050 for_each_cpu(i, cpu_map) {
7052 sd = *per_cpu_ptr(d.sd, i);
7054 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7055 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
7056 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
7058 cpu_attach_domain(sd, d.rd, i);
7062 if (rq && sched_debug_enabled) {
7063 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7064 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
7069 __free_domain_allocs(&d, alloc_state, cpu_map);
7073 static cpumask_var_t *doms_cur; /* current sched domains */
7074 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7075 static struct sched_domain_attr *dattr_cur;
7076 /* attribues of custom domains in 'doms_cur' */
7079 * Special case: If a kmalloc of a doms_cur partition (array of
7080 * cpumask) fails, then fallback to a single sched domain,
7081 * as determined by the single cpumask fallback_doms.
7083 static cpumask_var_t fallback_doms;
7086 * arch_update_cpu_topology lets virtualized architectures update the
7087 * cpu core maps. It is supposed to return 1 if the topology changed
7088 * or 0 if it stayed the same.
7090 int __weak arch_update_cpu_topology(void)
7095 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7098 cpumask_var_t *doms;
7100 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7103 for (i = 0; i < ndoms; i++) {
7104 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7105 free_sched_domains(doms, i);
7112 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7115 for (i = 0; i < ndoms; i++)
7116 free_cpumask_var(doms[i]);
7121 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7122 * For now this just excludes isolated cpus, but could be used to
7123 * exclude other special cases in the future.
7125 static int init_sched_domains(const struct cpumask *cpu_map)
7129 arch_update_cpu_topology();
7131 doms_cur = alloc_sched_domains(ndoms_cur);
7133 doms_cur = &fallback_doms;
7134 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7135 err = build_sched_domains(doms_cur[0], NULL);
7136 register_sched_domain_sysctl();
7142 * Detach sched domains from a group of cpus specified in cpu_map
7143 * These cpus will now be attached to the NULL domain
7145 static void detach_destroy_domains(const struct cpumask *cpu_map)
7150 for_each_cpu(i, cpu_map)
7151 cpu_attach_domain(NULL, &def_root_domain, i);
7155 /* handle null as "default" */
7156 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7157 struct sched_domain_attr *new, int idx_new)
7159 struct sched_domain_attr tmp;
7166 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7167 new ? (new + idx_new) : &tmp,
7168 sizeof(struct sched_domain_attr));
7172 * Partition sched domains as specified by the 'ndoms_new'
7173 * cpumasks in the array doms_new[] of cpumasks. This compares
7174 * doms_new[] to the current sched domain partitioning, doms_cur[].
7175 * It destroys each deleted domain and builds each new domain.
7177 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7178 * The masks don't intersect (don't overlap.) We should setup one
7179 * sched domain for each mask. CPUs not in any of the cpumasks will
7180 * not be load balanced. If the same cpumask appears both in the
7181 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7184 * The passed in 'doms_new' should be allocated using
7185 * alloc_sched_domains. This routine takes ownership of it and will
7186 * free_sched_domains it when done with it. If the caller failed the
7187 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7188 * and partition_sched_domains() will fallback to the single partition
7189 * 'fallback_doms', it also forces the domains to be rebuilt.
7191 * If doms_new == NULL it will be replaced with cpu_online_mask.
7192 * ndoms_new == 0 is a special case for destroying existing domains,
7193 * and it will not create the default domain.
7195 * Call with hotplug lock held
7197 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7198 struct sched_domain_attr *dattr_new)
7203 mutex_lock(&sched_domains_mutex);
7205 /* always unregister in case we don't destroy any domains */
7206 unregister_sched_domain_sysctl();
7208 /* Let architecture update cpu core mappings. */
7209 new_topology = arch_update_cpu_topology();
7211 n = doms_new ? ndoms_new : 0;
7213 /* Destroy deleted domains */
7214 for (i = 0; i < ndoms_cur; i++) {
7215 for (j = 0; j < n && !new_topology; j++) {
7216 if (cpumask_equal(doms_cur[i], doms_new[j])
7217 && dattrs_equal(dattr_cur, i, dattr_new, j))
7220 /* no match - a current sched domain not in new doms_new[] */
7221 detach_destroy_domains(doms_cur[i]);
7227 if (doms_new == NULL) {
7229 doms_new = &fallback_doms;
7230 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7231 WARN_ON_ONCE(dattr_new);
7234 /* Build new domains */
7235 for (i = 0; i < ndoms_new; i++) {
7236 for (j = 0; j < n && !new_topology; j++) {
7237 if (cpumask_equal(doms_new[i], doms_cur[j])
7238 && dattrs_equal(dattr_new, i, dattr_cur, j))
7241 /* no match - add a new doms_new */
7242 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7247 /* Remember the new sched domains */
7248 if (doms_cur != &fallback_doms)
7249 free_sched_domains(doms_cur, ndoms_cur);
7250 kfree(dattr_cur); /* kfree(NULL) is safe */
7251 doms_cur = doms_new;
7252 dattr_cur = dattr_new;
7253 ndoms_cur = ndoms_new;
7255 register_sched_domain_sysctl();
7257 mutex_unlock(&sched_domains_mutex);
7260 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7263 * Update cpusets according to cpu_active mask. If cpusets are
7264 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7265 * around partition_sched_domains().
7267 * If we come here as part of a suspend/resume, don't touch cpusets because we
7268 * want to restore it back to its original state upon resume anyway.
7270 static void cpuset_cpu_active(void)
7272 if (cpuhp_tasks_frozen) {
7274 * num_cpus_frozen tracks how many CPUs are involved in suspend
7275 * resume sequence. As long as this is not the last online
7276 * operation in the resume sequence, just build a single sched
7277 * domain, ignoring cpusets.
7280 if (likely(num_cpus_frozen)) {
7281 partition_sched_domains(1, NULL, NULL);
7285 * This is the last CPU online operation. So fall through and
7286 * restore the original sched domains by considering the
7287 * cpuset configurations.
7290 cpuset_update_active_cpus(true);
7293 static int cpuset_cpu_inactive(unsigned int cpu)
7295 unsigned long flags;
7300 if (!cpuhp_tasks_frozen) {
7301 rcu_read_lock_sched();
7302 dl_b = dl_bw_of(cpu);
7304 raw_spin_lock_irqsave(&dl_b->lock, flags);
7305 cpus = dl_bw_cpus(cpu);
7306 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7307 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7309 rcu_read_unlock_sched();
7313 cpuset_update_active_cpus(false);
7316 partition_sched_domains(1, NULL, NULL);
7321 int sched_cpu_activate(unsigned int cpu)
7323 struct rq *rq = cpu_rq(cpu);
7324 unsigned long flags;
7326 set_cpu_active(cpu, true);
7328 if (sched_smp_initialized) {
7329 sched_domains_numa_masks_set(cpu);
7330 cpuset_cpu_active();
7334 * Put the rq online, if not already. This happens:
7336 * 1) In the early boot process, because we build the real domains
7337 * after all cpus have been brought up.
7339 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7342 raw_spin_lock_irqsave(&rq->lock, flags);
7344 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7347 raw_spin_unlock_irqrestore(&rq->lock, flags);
7349 update_max_interval();
7354 int sched_cpu_deactivate(unsigned int cpu)
7358 set_cpu_active(cpu, false);
7360 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7361 * users of this state to go away such that all new such users will
7364 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7365 * not imply sync_sched(), so wait for both.
7367 * Do sync before park smpboot threads to take care the rcu boost case.
7369 if (IS_ENABLED(CONFIG_PREEMPT))
7370 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7374 if (!sched_smp_initialized)
7377 ret = cpuset_cpu_inactive(cpu);
7379 set_cpu_active(cpu, true);
7382 sched_domains_numa_masks_clear(cpu);
7386 static void sched_rq_cpu_starting(unsigned int cpu)
7388 struct rq *rq = cpu_rq(cpu);
7390 rq->calc_load_update = calc_load_update;
7391 update_max_interval();
7394 int sched_cpu_starting(unsigned int cpu)
7396 set_cpu_rq_start_time(cpu);
7397 sched_rq_cpu_starting(cpu);
7401 #ifdef CONFIG_HOTPLUG_CPU
7402 int sched_cpu_dying(unsigned int cpu)
7404 struct rq *rq = cpu_rq(cpu);
7405 unsigned long flags;
7407 /* Handle pending wakeups and then migrate everything off */
7408 sched_ttwu_pending();
7409 raw_spin_lock_irqsave(&rq->lock, flags);
7411 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7415 BUG_ON(rq->nr_running != 1);
7416 raw_spin_unlock_irqrestore(&rq->lock, flags);
7417 calc_load_migrate(rq);
7418 update_max_interval();
7419 nohz_balance_exit_idle(cpu);
7425 #ifdef CONFIG_SCHED_SMT
7426 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7428 static void sched_init_smt(void)
7431 * We've enumerated all CPUs and will assume that if any CPU
7432 * has SMT siblings, CPU0 will too.
7434 if (cpumask_weight(cpu_smt_mask(0)) > 1)
7435 static_branch_enable(&sched_smt_present);
7438 static inline void sched_init_smt(void) { }
7441 void __init sched_init_smp(void)
7443 cpumask_var_t non_isolated_cpus;
7445 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7446 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7451 * There's no userspace yet to cause hotplug operations; hence all the
7452 * cpu masks are stable and all blatant races in the below code cannot
7455 mutex_lock(&sched_domains_mutex);
7456 init_sched_domains(cpu_active_mask);
7457 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7458 if (cpumask_empty(non_isolated_cpus))
7459 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7460 mutex_unlock(&sched_domains_mutex);
7462 /* Move init over to a non-isolated CPU */
7463 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7465 sched_init_granularity();
7466 free_cpumask_var(non_isolated_cpus);
7468 init_sched_rt_class();
7469 init_sched_dl_class();
7473 sched_smp_initialized = true;
7476 static int __init migration_init(void)
7478 sched_rq_cpu_starting(smp_processor_id());
7481 early_initcall(migration_init);
7484 void __init sched_init_smp(void)
7486 sched_init_granularity();
7488 #endif /* CONFIG_SMP */
7490 int in_sched_functions(unsigned long addr)
7492 return in_lock_functions(addr) ||
7493 (addr >= (unsigned long)__sched_text_start
7494 && addr < (unsigned long)__sched_text_end);
7497 #ifdef CONFIG_CGROUP_SCHED
7499 * Default task group.
7500 * Every task in system belongs to this group at bootup.
7502 struct task_group root_task_group;
7503 LIST_HEAD(task_groups);
7505 /* Cacheline aligned slab cache for task_group */
7506 static struct kmem_cache *task_group_cache __read_mostly;
7509 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7510 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7512 #define WAIT_TABLE_BITS 8
7513 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7514 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
7516 wait_queue_head_t *bit_waitqueue(void *word, int bit)
7518 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
7519 unsigned long val = (unsigned long)word << shift | bit;
7521 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
7523 EXPORT_SYMBOL(bit_waitqueue);
7525 void __init sched_init(void)
7528 unsigned long alloc_size = 0, ptr;
7530 for (i = 0; i < WAIT_TABLE_SIZE; i++)
7531 init_waitqueue_head(bit_wait_table + i);
7533 #ifdef CONFIG_FAIR_GROUP_SCHED
7534 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7536 #ifdef CONFIG_RT_GROUP_SCHED
7537 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7540 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7542 #ifdef CONFIG_FAIR_GROUP_SCHED
7543 root_task_group.se = (struct sched_entity **)ptr;
7544 ptr += nr_cpu_ids * sizeof(void **);
7546 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7547 ptr += nr_cpu_ids * sizeof(void **);
7549 #endif /* CONFIG_FAIR_GROUP_SCHED */
7550 #ifdef CONFIG_RT_GROUP_SCHED
7551 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7552 ptr += nr_cpu_ids * sizeof(void **);
7554 root_task_group.rt_rq = (struct rt_rq **)ptr;
7555 ptr += nr_cpu_ids * sizeof(void **);
7557 #endif /* CONFIG_RT_GROUP_SCHED */
7559 #ifdef CONFIG_CPUMASK_OFFSTACK
7560 for_each_possible_cpu(i) {
7561 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7562 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7563 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7564 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7566 #endif /* CONFIG_CPUMASK_OFFSTACK */
7568 init_rt_bandwidth(&def_rt_bandwidth,
7569 global_rt_period(), global_rt_runtime());
7570 init_dl_bandwidth(&def_dl_bandwidth,
7571 global_rt_period(), global_rt_runtime());
7574 init_defrootdomain();
7577 #ifdef CONFIG_RT_GROUP_SCHED
7578 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7579 global_rt_period(), global_rt_runtime());
7580 #endif /* CONFIG_RT_GROUP_SCHED */
7582 #ifdef CONFIG_CGROUP_SCHED
7583 task_group_cache = KMEM_CACHE(task_group, 0);
7585 list_add(&root_task_group.list, &task_groups);
7586 INIT_LIST_HEAD(&root_task_group.children);
7587 INIT_LIST_HEAD(&root_task_group.siblings);
7588 autogroup_init(&init_task);
7589 #endif /* CONFIG_CGROUP_SCHED */
7591 for_each_possible_cpu(i) {
7595 raw_spin_lock_init(&rq->lock);
7597 rq->calc_load_active = 0;
7598 rq->calc_load_update = jiffies + LOAD_FREQ;
7599 init_cfs_rq(&rq->cfs);
7600 init_rt_rq(&rq->rt);
7601 init_dl_rq(&rq->dl);
7602 #ifdef CONFIG_FAIR_GROUP_SCHED
7603 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7604 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7606 * How much cpu bandwidth does root_task_group get?
7608 * In case of task-groups formed thr' the cgroup filesystem, it
7609 * gets 100% of the cpu resources in the system. This overall
7610 * system cpu resource is divided among the tasks of
7611 * root_task_group and its child task-groups in a fair manner,
7612 * based on each entity's (task or task-group's) weight
7613 * (se->load.weight).
7615 * In other words, if root_task_group has 10 tasks of weight
7616 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7617 * then A0's share of the cpu resource is:
7619 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7621 * We achieve this by letting root_task_group's tasks sit
7622 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7624 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7625 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7626 #endif /* CONFIG_FAIR_GROUP_SCHED */
7628 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7629 #ifdef CONFIG_RT_GROUP_SCHED
7630 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7633 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7634 rq->cpu_load[j] = 0;
7639 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7640 rq->balance_callback = NULL;
7641 rq->active_balance = 0;
7642 rq->next_balance = jiffies;
7647 rq->avg_idle = 2*sysctl_sched_migration_cost;
7648 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7650 INIT_LIST_HEAD(&rq->cfs_tasks);
7652 rq_attach_root(rq, &def_root_domain);
7653 #ifdef CONFIG_NO_HZ_COMMON
7654 rq->last_load_update_tick = jiffies;
7657 #ifdef CONFIG_NO_HZ_FULL
7658 rq->last_sched_tick = 0;
7660 #endif /* CONFIG_SMP */
7662 atomic_set(&rq->nr_iowait, 0);
7665 set_load_weight(&init_task);
7668 * The boot idle thread does lazy MMU switching as well:
7670 atomic_inc(&init_mm.mm_count);
7671 enter_lazy_tlb(&init_mm, current);
7674 * Make us the idle thread. Technically, schedule() should not be
7675 * called from this thread, however somewhere below it might be,
7676 * but because we are the idle thread, we just pick up running again
7677 * when this runqueue becomes "idle".
7679 init_idle(current, smp_processor_id());
7681 calc_load_update = jiffies + LOAD_FREQ;
7684 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7685 /* May be allocated at isolcpus cmdline parse time */
7686 if (cpu_isolated_map == NULL)
7687 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7688 idle_thread_set_boot_cpu();
7689 set_cpu_rq_start_time(smp_processor_id());
7691 init_sched_fair_class();
7695 scheduler_running = 1;
7698 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7699 static inline int preempt_count_equals(int preempt_offset)
7701 int nested = preempt_count() + rcu_preempt_depth();
7703 return (nested == preempt_offset);
7706 void __might_sleep(const char *file, int line, int preempt_offset)
7709 * Blocking primitives will set (and therefore destroy) current->state,
7710 * since we will exit with TASK_RUNNING make sure we enter with it,
7711 * otherwise we will destroy state.
7713 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7714 "do not call blocking ops when !TASK_RUNNING; "
7715 "state=%lx set at [<%p>] %pS\n",
7717 (void *)current->task_state_change,
7718 (void *)current->task_state_change);
7720 ___might_sleep(file, line, preempt_offset);
7722 EXPORT_SYMBOL(__might_sleep);
7724 void ___might_sleep(const char *file, int line, int preempt_offset)
7726 static unsigned long prev_jiffy; /* ratelimiting */
7727 unsigned long preempt_disable_ip;
7729 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7730 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7731 !is_idle_task(current)) ||
7732 system_state != SYSTEM_RUNNING || oops_in_progress)
7734 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7736 prev_jiffy = jiffies;
7738 /* Save this before calling printk(), since that will clobber it */
7739 preempt_disable_ip = get_preempt_disable_ip(current);
7742 "BUG: sleeping function called from invalid context at %s:%d\n",
7745 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7746 in_atomic(), irqs_disabled(),
7747 current->pid, current->comm);
7749 if (task_stack_end_corrupted(current))
7750 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7752 debug_show_held_locks(current);
7753 if (irqs_disabled())
7754 print_irqtrace_events(current);
7755 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7756 && !preempt_count_equals(preempt_offset)) {
7757 pr_err("Preemption disabled at:");
7758 print_ip_sym(preempt_disable_ip);
7762 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7764 EXPORT_SYMBOL(___might_sleep);
7767 #ifdef CONFIG_MAGIC_SYSRQ
7768 void normalize_rt_tasks(void)
7770 struct task_struct *g, *p;
7771 struct sched_attr attr = {
7772 .sched_policy = SCHED_NORMAL,
7775 read_lock(&tasklist_lock);
7776 for_each_process_thread(g, p) {
7778 * Only normalize user tasks:
7780 if (p->flags & PF_KTHREAD)
7783 p->se.exec_start = 0;
7784 schedstat_set(p->se.statistics.wait_start, 0);
7785 schedstat_set(p->se.statistics.sleep_start, 0);
7786 schedstat_set(p->se.statistics.block_start, 0);
7788 if (!dl_task(p) && !rt_task(p)) {
7790 * Renice negative nice level userspace
7793 if (task_nice(p) < 0)
7794 set_user_nice(p, 0);
7798 __sched_setscheduler(p, &attr, false, false);
7800 read_unlock(&tasklist_lock);
7803 #endif /* CONFIG_MAGIC_SYSRQ */
7805 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7807 * These functions are only useful for the IA64 MCA handling, or kdb.
7809 * They can only be called when the whole system has been
7810 * stopped - every CPU needs to be quiescent, and no scheduling
7811 * activity can take place. Using them for anything else would
7812 * be a serious bug, and as a result, they aren't even visible
7813 * under any other configuration.
7817 * curr_task - return the current task for a given cpu.
7818 * @cpu: the processor in question.
7820 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7822 * Return: The current task for @cpu.
7824 struct task_struct *curr_task(int cpu)
7826 return cpu_curr(cpu);
7829 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7833 * set_curr_task - set the current task for a given cpu.
7834 * @cpu: the processor in question.
7835 * @p: the task pointer to set.
7837 * Description: This function must only be used when non-maskable interrupts
7838 * are serviced on a separate stack. It allows the architecture to switch the
7839 * notion of the current task on a cpu in a non-blocking manner. This function
7840 * must be called with all CPU's synchronized, and interrupts disabled, the
7841 * and caller must save the original value of the current task (see
7842 * curr_task() above) and restore that value before reenabling interrupts and
7843 * re-starting the system.
7845 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7847 void ia64_set_curr_task(int cpu, struct task_struct *p)
7854 #ifdef CONFIG_CGROUP_SCHED
7855 /* task_group_lock serializes the addition/removal of task groups */
7856 static DEFINE_SPINLOCK(task_group_lock);
7858 static void sched_free_group(struct task_group *tg)
7860 free_fair_sched_group(tg);
7861 free_rt_sched_group(tg);
7863 kmem_cache_free(task_group_cache, tg);
7866 /* allocate runqueue etc for a new task group */
7867 struct task_group *sched_create_group(struct task_group *parent)
7869 struct task_group *tg;
7871 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7873 return ERR_PTR(-ENOMEM);
7875 if (!alloc_fair_sched_group(tg, parent))
7878 if (!alloc_rt_sched_group(tg, parent))
7884 sched_free_group(tg);
7885 return ERR_PTR(-ENOMEM);
7888 void sched_online_group(struct task_group *tg, struct task_group *parent)
7890 unsigned long flags;
7892 spin_lock_irqsave(&task_group_lock, flags);
7893 list_add_rcu(&tg->list, &task_groups);
7895 WARN_ON(!parent); /* root should already exist */
7897 tg->parent = parent;
7898 INIT_LIST_HEAD(&tg->children);
7899 list_add_rcu(&tg->siblings, &parent->children);
7900 spin_unlock_irqrestore(&task_group_lock, flags);
7902 online_fair_sched_group(tg);
7905 /* rcu callback to free various structures associated with a task group */
7906 static void sched_free_group_rcu(struct rcu_head *rhp)
7908 /* now it should be safe to free those cfs_rqs */
7909 sched_free_group(container_of(rhp, struct task_group, rcu));
7912 void sched_destroy_group(struct task_group *tg)
7914 /* wait for possible concurrent references to cfs_rqs complete */
7915 call_rcu(&tg->rcu, sched_free_group_rcu);
7918 void sched_offline_group(struct task_group *tg)
7920 unsigned long flags;
7922 /* end participation in shares distribution */
7923 unregister_fair_sched_group(tg);
7925 spin_lock_irqsave(&task_group_lock, flags);
7926 list_del_rcu(&tg->list);
7927 list_del_rcu(&tg->siblings);
7928 spin_unlock_irqrestore(&task_group_lock, flags);
7931 static void sched_change_group(struct task_struct *tsk, int type)
7933 struct task_group *tg;
7936 * All callers are synchronized by task_rq_lock(); we do not use RCU
7937 * which is pointless here. Thus, we pass "true" to task_css_check()
7938 * to prevent lockdep warnings.
7940 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7941 struct task_group, css);
7942 tg = autogroup_task_group(tsk, tg);
7943 tsk->sched_task_group = tg;
7945 #ifdef CONFIG_FAIR_GROUP_SCHED
7946 if (tsk->sched_class->task_change_group)
7947 tsk->sched_class->task_change_group(tsk, type);
7950 set_task_rq(tsk, task_cpu(tsk));
7954 * Change task's runqueue when it moves between groups.
7956 * The caller of this function should have put the task in its new group by
7957 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7960 void sched_move_task(struct task_struct *tsk)
7962 int queued, running;
7966 rq = task_rq_lock(tsk, &rf);
7968 running = task_current(rq, tsk);
7969 queued = task_on_rq_queued(tsk);
7972 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7973 if (unlikely(running))
7974 put_prev_task(rq, tsk);
7976 sched_change_group(tsk, TASK_MOVE_GROUP);
7979 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7980 if (unlikely(running))
7981 set_curr_task(rq, tsk);
7983 task_rq_unlock(rq, tsk, &rf);
7985 #endif /* CONFIG_CGROUP_SCHED */
7987 #ifdef CONFIG_RT_GROUP_SCHED
7989 * Ensure that the real time constraints are schedulable.
7991 static DEFINE_MUTEX(rt_constraints_mutex);
7993 /* Must be called with tasklist_lock held */
7994 static inline int tg_has_rt_tasks(struct task_group *tg)
7996 struct task_struct *g, *p;
7999 * Autogroups do not have RT tasks; see autogroup_create().
8001 if (task_group_is_autogroup(tg))
8004 for_each_process_thread(g, p) {
8005 if (rt_task(p) && task_group(p) == tg)
8012 struct rt_schedulable_data {
8013 struct task_group *tg;
8018 static int tg_rt_schedulable(struct task_group *tg, void *data)
8020 struct rt_schedulable_data *d = data;
8021 struct task_group *child;
8022 unsigned long total, sum = 0;
8023 u64 period, runtime;
8025 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8026 runtime = tg->rt_bandwidth.rt_runtime;
8029 period = d->rt_period;
8030 runtime = d->rt_runtime;
8034 * Cannot have more runtime than the period.
8036 if (runtime > period && runtime != RUNTIME_INF)
8040 * Ensure we don't starve existing RT tasks.
8042 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8045 total = to_ratio(period, runtime);
8048 * Nobody can have more than the global setting allows.
8050 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8054 * The sum of our children's runtime should not exceed our own.
8056 list_for_each_entry_rcu(child, &tg->children, siblings) {
8057 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8058 runtime = child->rt_bandwidth.rt_runtime;
8060 if (child == d->tg) {
8061 period = d->rt_period;
8062 runtime = d->rt_runtime;
8065 sum += to_ratio(period, runtime);
8074 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8078 struct rt_schedulable_data data = {
8080 .rt_period = period,
8081 .rt_runtime = runtime,
8085 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8091 static int tg_set_rt_bandwidth(struct task_group *tg,
8092 u64 rt_period, u64 rt_runtime)
8097 * Disallowing the root group RT runtime is BAD, it would disallow the
8098 * kernel creating (and or operating) RT threads.
8100 if (tg == &root_task_group && rt_runtime == 0)
8103 /* No period doesn't make any sense. */
8107 mutex_lock(&rt_constraints_mutex);
8108 read_lock(&tasklist_lock);
8109 err = __rt_schedulable(tg, rt_period, rt_runtime);
8113 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8114 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8115 tg->rt_bandwidth.rt_runtime = rt_runtime;
8117 for_each_possible_cpu(i) {
8118 struct rt_rq *rt_rq = tg->rt_rq[i];
8120 raw_spin_lock(&rt_rq->rt_runtime_lock);
8121 rt_rq->rt_runtime = rt_runtime;
8122 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8124 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8126 read_unlock(&tasklist_lock);
8127 mutex_unlock(&rt_constraints_mutex);
8132 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8134 u64 rt_runtime, rt_period;
8136 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8137 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8138 if (rt_runtime_us < 0)
8139 rt_runtime = RUNTIME_INF;
8141 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8144 static long sched_group_rt_runtime(struct task_group *tg)
8148 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8151 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8152 do_div(rt_runtime_us, NSEC_PER_USEC);
8153 return rt_runtime_us;
8156 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8158 u64 rt_runtime, rt_period;
8160 rt_period = rt_period_us * NSEC_PER_USEC;
8161 rt_runtime = tg->rt_bandwidth.rt_runtime;
8163 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8166 static long sched_group_rt_period(struct task_group *tg)
8170 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8171 do_div(rt_period_us, NSEC_PER_USEC);
8172 return rt_period_us;
8174 #endif /* CONFIG_RT_GROUP_SCHED */
8176 #ifdef CONFIG_RT_GROUP_SCHED
8177 static int sched_rt_global_constraints(void)
8181 mutex_lock(&rt_constraints_mutex);
8182 read_lock(&tasklist_lock);
8183 ret = __rt_schedulable(NULL, 0, 0);
8184 read_unlock(&tasklist_lock);
8185 mutex_unlock(&rt_constraints_mutex);
8190 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8192 /* Don't accept realtime tasks when there is no way for them to run */
8193 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8199 #else /* !CONFIG_RT_GROUP_SCHED */
8200 static int sched_rt_global_constraints(void)
8202 unsigned long flags;
8205 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8206 for_each_possible_cpu(i) {
8207 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8209 raw_spin_lock(&rt_rq->rt_runtime_lock);
8210 rt_rq->rt_runtime = global_rt_runtime();
8211 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8213 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8217 #endif /* CONFIG_RT_GROUP_SCHED */
8219 static int sched_dl_global_validate(void)
8221 u64 runtime = global_rt_runtime();
8222 u64 period = global_rt_period();
8223 u64 new_bw = to_ratio(period, runtime);
8226 unsigned long flags;
8229 * Here we want to check the bandwidth not being set to some
8230 * value smaller than the currently allocated bandwidth in
8231 * any of the root_domains.
8233 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8234 * cycling on root_domains... Discussion on different/better
8235 * solutions is welcome!
8237 for_each_possible_cpu(cpu) {
8238 rcu_read_lock_sched();
8239 dl_b = dl_bw_of(cpu);
8241 raw_spin_lock_irqsave(&dl_b->lock, flags);
8242 if (new_bw < dl_b->total_bw)
8244 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8246 rcu_read_unlock_sched();
8255 static void sched_dl_do_global(void)
8260 unsigned long flags;
8262 def_dl_bandwidth.dl_period = global_rt_period();
8263 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8265 if (global_rt_runtime() != RUNTIME_INF)
8266 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8269 * FIXME: As above...
8271 for_each_possible_cpu(cpu) {
8272 rcu_read_lock_sched();
8273 dl_b = dl_bw_of(cpu);
8275 raw_spin_lock_irqsave(&dl_b->lock, flags);
8277 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8279 rcu_read_unlock_sched();
8283 static int sched_rt_global_validate(void)
8285 if (sysctl_sched_rt_period <= 0)
8288 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8289 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8295 static void sched_rt_do_global(void)
8297 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8298 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8301 int sched_rt_handler(struct ctl_table *table, int write,
8302 void __user *buffer, size_t *lenp,
8305 int old_period, old_runtime;
8306 static DEFINE_MUTEX(mutex);
8310 old_period = sysctl_sched_rt_period;
8311 old_runtime = sysctl_sched_rt_runtime;
8313 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8315 if (!ret && write) {
8316 ret = sched_rt_global_validate();
8320 ret = sched_dl_global_validate();
8324 ret = sched_rt_global_constraints();
8328 sched_rt_do_global();
8329 sched_dl_do_global();
8333 sysctl_sched_rt_period = old_period;
8334 sysctl_sched_rt_runtime = old_runtime;
8336 mutex_unlock(&mutex);
8341 int sched_rr_handler(struct ctl_table *table, int write,
8342 void __user *buffer, size_t *lenp,
8346 static DEFINE_MUTEX(mutex);
8349 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8350 /* make sure that internally we keep jiffies */
8351 /* also, writing zero resets timeslice to default */
8352 if (!ret && write) {
8353 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8354 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8356 mutex_unlock(&mutex);
8360 #ifdef CONFIG_CGROUP_SCHED
8362 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8364 return css ? container_of(css, struct task_group, css) : NULL;
8367 static struct cgroup_subsys_state *
8368 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8370 struct task_group *parent = css_tg(parent_css);
8371 struct task_group *tg;
8374 /* This is early initialization for the top cgroup */
8375 return &root_task_group.css;
8378 tg = sched_create_group(parent);
8380 return ERR_PTR(-ENOMEM);
8382 sched_online_group(tg, parent);
8387 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8389 struct task_group *tg = css_tg(css);
8391 sched_offline_group(tg);
8394 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8396 struct task_group *tg = css_tg(css);
8399 * Relies on the RCU grace period between css_released() and this.
8401 sched_free_group(tg);
8405 * This is called before wake_up_new_task(), therefore we really only
8406 * have to set its group bits, all the other stuff does not apply.
8408 static void cpu_cgroup_fork(struct task_struct *task)
8413 rq = task_rq_lock(task, &rf);
8415 sched_change_group(task, TASK_SET_GROUP);
8417 task_rq_unlock(rq, task, &rf);
8420 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8422 struct task_struct *task;
8423 struct cgroup_subsys_state *css;
8426 cgroup_taskset_for_each(task, css, tset) {
8427 #ifdef CONFIG_RT_GROUP_SCHED
8428 if (!sched_rt_can_attach(css_tg(css), task))
8431 /* We don't support RT-tasks being in separate groups */
8432 if (task->sched_class != &fair_sched_class)
8436 * Serialize against wake_up_new_task() such that if its
8437 * running, we're sure to observe its full state.
8439 raw_spin_lock_irq(&task->pi_lock);
8441 * Avoid calling sched_move_task() before wake_up_new_task()
8442 * has happened. This would lead to problems with PELT, due to
8443 * move wanting to detach+attach while we're not attached yet.
8445 if (task->state == TASK_NEW)
8447 raw_spin_unlock_irq(&task->pi_lock);
8455 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8457 struct task_struct *task;
8458 struct cgroup_subsys_state *css;
8460 cgroup_taskset_for_each(task, css, tset)
8461 sched_move_task(task);
8464 #ifdef CONFIG_FAIR_GROUP_SCHED
8465 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8466 struct cftype *cftype, u64 shareval)
8468 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8471 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8474 struct task_group *tg = css_tg(css);
8476 return (u64) scale_load_down(tg->shares);
8479 #ifdef CONFIG_CFS_BANDWIDTH
8480 static DEFINE_MUTEX(cfs_constraints_mutex);
8482 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8483 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8485 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8487 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8489 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8490 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8492 if (tg == &root_task_group)
8496 * Ensure we have at some amount of bandwidth every period. This is
8497 * to prevent reaching a state of large arrears when throttled via
8498 * entity_tick() resulting in prolonged exit starvation.
8500 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8504 * Likewise, bound things on the otherside by preventing insane quota
8505 * periods. This also allows us to normalize in computing quota
8508 if (period > max_cfs_quota_period)
8512 * Prevent race between setting of cfs_rq->runtime_enabled and
8513 * unthrottle_offline_cfs_rqs().
8516 mutex_lock(&cfs_constraints_mutex);
8517 ret = __cfs_schedulable(tg, period, quota);
8521 runtime_enabled = quota != RUNTIME_INF;
8522 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8524 * If we need to toggle cfs_bandwidth_used, off->on must occur
8525 * before making related changes, and on->off must occur afterwards
8527 if (runtime_enabled && !runtime_was_enabled)
8528 cfs_bandwidth_usage_inc();
8529 raw_spin_lock_irq(&cfs_b->lock);
8530 cfs_b->period = ns_to_ktime(period);
8531 cfs_b->quota = quota;
8533 __refill_cfs_bandwidth_runtime(cfs_b);
8534 /* restart the period timer (if active) to handle new period expiry */
8535 if (runtime_enabled)
8536 start_cfs_bandwidth(cfs_b);
8537 raw_spin_unlock_irq(&cfs_b->lock);
8539 for_each_online_cpu(i) {
8540 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8541 struct rq *rq = cfs_rq->rq;
8543 raw_spin_lock_irq(&rq->lock);
8544 cfs_rq->runtime_enabled = runtime_enabled;
8545 cfs_rq->runtime_remaining = 0;
8547 if (cfs_rq->throttled)
8548 unthrottle_cfs_rq(cfs_rq);
8549 raw_spin_unlock_irq(&rq->lock);
8551 if (runtime_was_enabled && !runtime_enabled)
8552 cfs_bandwidth_usage_dec();
8554 mutex_unlock(&cfs_constraints_mutex);
8560 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8564 period = ktime_to_ns(tg->cfs_bandwidth.period);
8565 if (cfs_quota_us < 0)
8566 quota = RUNTIME_INF;
8568 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8570 return tg_set_cfs_bandwidth(tg, period, quota);
8573 long tg_get_cfs_quota(struct task_group *tg)
8577 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8580 quota_us = tg->cfs_bandwidth.quota;
8581 do_div(quota_us, NSEC_PER_USEC);
8586 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8590 period = (u64)cfs_period_us * NSEC_PER_USEC;
8591 quota = tg->cfs_bandwidth.quota;
8593 return tg_set_cfs_bandwidth(tg, period, quota);
8596 long tg_get_cfs_period(struct task_group *tg)
8600 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8601 do_div(cfs_period_us, NSEC_PER_USEC);
8603 return cfs_period_us;
8606 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8609 return tg_get_cfs_quota(css_tg(css));
8612 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8613 struct cftype *cftype, s64 cfs_quota_us)
8615 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8618 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8621 return tg_get_cfs_period(css_tg(css));
8624 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8625 struct cftype *cftype, u64 cfs_period_us)
8627 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8630 struct cfs_schedulable_data {
8631 struct task_group *tg;
8636 * normalize group quota/period to be quota/max_period
8637 * note: units are usecs
8639 static u64 normalize_cfs_quota(struct task_group *tg,
8640 struct cfs_schedulable_data *d)
8648 period = tg_get_cfs_period(tg);
8649 quota = tg_get_cfs_quota(tg);
8652 /* note: these should typically be equivalent */
8653 if (quota == RUNTIME_INF || quota == -1)
8656 return to_ratio(period, quota);
8659 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8661 struct cfs_schedulable_data *d = data;
8662 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8663 s64 quota = 0, parent_quota = -1;
8666 quota = RUNTIME_INF;
8668 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8670 quota = normalize_cfs_quota(tg, d);
8671 parent_quota = parent_b->hierarchical_quota;
8674 * ensure max(child_quota) <= parent_quota, inherit when no
8677 if (quota == RUNTIME_INF)
8678 quota = parent_quota;
8679 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8682 cfs_b->hierarchical_quota = quota;
8687 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8690 struct cfs_schedulable_data data = {
8696 if (quota != RUNTIME_INF) {
8697 do_div(data.period, NSEC_PER_USEC);
8698 do_div(data.quota, NSEC_PER_USEC);
8702 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8708 static int cpu_stats_show(struct seq_file *sf, void *v)
8710 struct task_group *tg = css_tg(seq_css(sf));
8711 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8713 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8714 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8715 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8719 #endif /* CONFIG_CFS_BANDWIDTH */
8720 #endif /* CONFIG_FAIR_GROUP_SCHED */
8722 #ifdef CONFIG_RT_GROUP_SCHED
8723 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8724 struct cftype *cft, s64 val)
8726 return sched_group_set_rt_runtime(css_tg(css), val);
8729 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8732 return sched_group_rt_runtime(css_tg(css));
8735 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8736 struct cftype *cftype, u64 rt_period_us)
8738 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8741 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8744 return sched_group_rt_period(css_tg(css));
8746 #endif /* CONFIG_RT_GROUP_SCHED */
8748 static struct cftype cpu_files[] = {
8749 #ifdef CONFIG_FAIR_GROUP_SCHED
8752 .read_u64 = cpu_shares_read_u64,
8753 .write_u64 = cpu_shares_write_u64,
8756 #ifdef CONFIG_CFS_BANDWIDTH
8758 .name = "cfs_quota_us",
8759 .read_s64 = cpu_cfs_quota_read_s64,
8760 .write_s64 = cpu_cfs_quota_write_s64,
8763 .name = "cfs_period_us",
8764 .read_u64 = cpu_cfs_period_read_u64,
8765 .write_u64 = cpu_cfs_period_write_u64,
8769 .seq_show = cpu_stats_show,
8772 #ifdef CONFIG_RT_GROUP_SCHED
8774 .name = "rt_runtime_us",
8775 .read_s64 = cpu_rt_runtime_read,
8776 .write_s64 = cpu_rt_runtime_write,
8779 .name = "rt_period_us",
8780 .read_u64 = cpu_rt_period_read_uint,
8781 .write_u64 = cpu_rt_period_write_uint,
8787 struct cgroup_subsys cpu_cgrp_subsys = {
8788 .css_alloc = cpu_cgroup_css_alloc,
8789 .css_released = cpu_cgroup_css_released,
8790 .css_free = cpu_cgroup_css_free,
8791 .fork = cpu_cgroup_fork,
8792 .can_attach = cpu_cgroup_can_attach,
8793 .attach = cpu_cgroup_attach,
8794 .legacy_cftypes = cpu_files,
8798 #endif /* CONFIG_CGROUP_SCHED */
8800 void dump_cpu_task(int cpu)
8802 pr_info("Task dump for CPU %d:\n", cpu);
8803 sched_show_task(cpu_curr(cpu));
8807 * Nice levels are multiplicative, with a gentle 10% change for every
8808 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8809 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8810 * that remained on nice 0.
8812 * The "10% effect" is relative and cumulative: from _any_ nice level,
8813 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8814 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8815 * If a task goes up by ~10% and another task goes down by ~10% then
8816 * the relative distance between them is ~25%.)
8818 const int sched_prio_to_weight[40] = {
8819 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8820 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8821 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8822 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8823 /* 0 */ 1024, 820, 655, 526, 423,
8824 /* 5 */ 335, 272, 215, 172, 137,
8825 /* 10 */ 110, 87, 70, 56, 45,
8826 /* 15 */ 36, 29, 23, 18, 15,
8830 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8832 * In cases where the weight does not change often, we can use the
8833 * precalculated inverse to speed up arithmetics by turning divisions
8834 * into multiplications:
8836 const u32 sched_prio_to_wmult[40] = {
8837 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8838 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8839 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8840 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8841 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8842 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8843 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8844 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,