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 raw_spin_lock_irqsave(&rq->lock, flags);
511 if (cpu_online(cpu) || cpu == smp_processor_id())
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);
1144 update_rq_clock(rq);
1146 if (p->flags & PF_KTHREAD) {
1148 * Kernel threads are allowed on online && !active CPUs
1150 cpu_valid_mask = cpu_online_mask;
1154 * Must re-check here, to close a race against __kthread_bind(),
1155 * sched_setaffinity() is not guaranteed to observe the flag.
1157 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1162 if (cpumask_equal(&p->cpus_allowed, new_mask))
1165 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1166 if (dest_cpu >= nr_cpu_ids) {
1171 do_set_cpus_allowed(p, new_mask);
1173 if (p->flags & PF_KTHREAD) {
1175 * For kernel threads that do indeed end up on online &&
1176 * !active we want to ensure they are strict per-cpu threads.
1178 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1179 !cpumask_intersects(new_mask, cpu_active_mask) &&
1180 p->nr_cpus_allowed != 1);
1183 /* Can the task run on the task's current CPU? If so, we're done */
1184 if (cpumask_test_cpu(task_cpu(p), new_mask))
1187 if (task_running(rq, p) || p->state == TASK_WAKING) {
1188 struct migration_arg arg = { p, dest_cpu };
1189 /* Need help from migration thread: drop lock and wait. */
1190 task_rq_unlock(rq, p, &rf);
1191 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1192 tlb_migrate_finish(p->mm);
1194 } else if (task_on_rq_queued(p)) {
1196 * OK, since we're going to drop the lock immediately
1197 * afterwards anyway.
1199 lockdep_unpin_lock(&rq->lock, rf.cookie);
1200 rq = move_queued_task(rq, p, dest_cpu);
1201 lockdep_repin_lock(&rq->lock, rf.cookie);
1204 task_rq_unlock(rq, p, &rf);
1209 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1211 return __set_cpus_allowed_ptr(p, new_mask, false);
1213 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1215 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1217 #ifdef CONFIG_SCHED_DEBUG
1219 * We should never call set_task_cpu() on a blocked task,
1220 * ttwu() will sort out the placement.
1222 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1226 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1227 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1228 * time relying on p->on_rq.
1230 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1231 p->sched_class == &fair_sched_class &&
1232 (p->on_rq && !task_on_rq_migrating(p)));
1234 #ifdef CONFIG_LOCKDEP
1236 * The caller should hold either p->pi_lock or rq->lock, when changing
1237 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1239 * sched_move_task() holds both and thus holding either pins the cgroup,
1242 * Furthermore, all task_rq users should acquire both locks, see
1245 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1246 lockdep_is_held(&task_rq(p)->lock)));
1250 trace_sched_migrate_task(p, new_cpu);
1252 if (task_cpu(p) != new_cpu) {
1253 if (p->sched_class->migrate_task_rq)
1254 p->sched_class->migrate_task_rq(p);
1255 p->se.nr_migrations++;
1256 perf_event_task_migrate(p);
1259 __set_task_cpu(p, new_cpu);
1262 static void __migrate_swap_task(struct task_struct *p, int cpu)
1264 if (task_on_rq_queued(p)) {
1265 struct rq *src_rq, *dst_rq;
1267 src_rq = task_rq(p);
1268 dst_rq = cpu_rq(cpu);
1270 p->on_rq = TASK_ON_RQ_MIGRATING;
1271 deactivate_task(src_rq, p, 0);
1272 set_task_cpu(p, cpu);
1273 activate_task(dst_rq, p, 0);
1274 p->on_rq = TASK_ON_RQ_QUEUED;
1275 check_preempt_curr(dst_rq, p, 0);
1278 * Task isn't running anymore; make it appear like we migrated
1279 * it before it went to sleep. This means on wakeup we make the
1280 * previous cpu our target instead of where it really is.
1286 struct migration_swap_arg {
1287 struct task_struct *src_task, *dst_task;
1288 int src_cpu, dst_cpu;
1291 static int migrate_swap_stop(void *data)
1293 struct migration_swap_arg *arg = data;
1294 struct rq *src_rq, *dst_rq;
1297 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1300 src_rq = cpu_rq(arg->src_cpu);
1301 dst_rq = cpu_rq(arg->dst_cpu);
1303 double_raw_lock(&arg->src_task->pi_lock,
1304 &arg->dst_task->pi_lock);
1305 double_rq_lock(src_rq, dst_rq);
1307 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1310 if (task_cpu(arg->src_task) != arg->src_cpu)
1313 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1316 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1319 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1320 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1325 double_rq_unlock(src_rq, dst_rq);
1326 raw_spin_unlock(&arg->dst_task->pi_lock);
1327 raw_spin_unlock(&arg->src_task->pi_lock);
1333 * Cross migrate two tasks
1335 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1337 struct migration_swap_arg arg;
1340 arg = (struct migration_swap_arg){
1342 .src_cpu = task_cpu(cur),
1344 .dst_cpu = task_cpu(p),
1347 if (arg.src_cpu == arg.dst_cpu)
1351 * These three tests are all lockless; this is OK since all of them
1352 * will be re-checked with proper locks held further down the line.
1354 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1357 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1360 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1363 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1364 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1371 * wait_task_inactive - wait for a thread to unschedule.
1373 * If @match_state is nonzero, it's the @p->state value just checked and
1374 * not expected to change. If it changes, i.e. @p might have woken up,
1375 * then return zero. When we succeed in waiting for @p to be off its CPU,
1376 * we return a positive number (its total switch count). If a second call
1377 * a short while later returns the same number, the caller can be sure that
1378 * @p has remained unscheduled the whole time.
1380 * The caller must ensure that the task *will* unschedule sometime soon,
1381 * else this function might spin for a *long* time. This function can't
1382 * be called with interrupts off, or it may introduce deadlock with
1383 * smp_call_function() if an IPI is sent by the same process we are
1384 * waiting to become inactive.
1386 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1388 int running, queued;
1395 * We do the initial early heuristics without holding
1396 * any task-queue locks at all. We'll only try to get
1397 * the runqueue lock when things look like they will
1403 * If the task is actively running on another CPU
1404 * still, just relax and busy-wait without holding
1407 * NOTE! Since we don't hold any locks, it's not
1408 * even sure that "rq" stays as the right runqueue!
1409 * But we don't care, since "task_running()" will
1410 * return false if the runqueue has changed and p
1411 * is actually now running somewhere else!
1413 while (task_running(rq, p)) {
1414 if (match_state && unlikely(p->state != match_state))
1420 * Ok, time to look more closely! We need the rq
1421 * lock now, to be *sure*. If we're wrong, we'll
1422 * just go back and repeat.
1424 rq = task_rq_lock(p, &rf);
1425 trace_sched_wait_task(p);
1426 running = task_running(rq, p);
1427 queued = task_on_rq_queued(p);
1429 if (!match_state || p->state == match_state)
1430 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1431 task_rq_unlock(rq, p, &rf);
1434 * If it changed from the expected state, bail out now.
1436 if (unlikely(!ncsw))
1440 * Was it really running after all now that we
1441 * checked with the proper locks actually held?
1443 * Oops. Go back and try again..
1445 if (unlikely(running)) {
1451 * It's not enough that it's not actively running,
1452 * it must be off the runqueue _entirely_, and not
1455 * So if it was still runnable (but just not actively
1456 * running right now), it's preempted, and we should
1457 * yield - it could be a while.
1459 if (unlikely(queued)) {
1460 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1462 set_current_state(TASK_UNINTERRUPTIBLE);
1463 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1468 * Ahh, all good. It wasn't running, and it wasn't
1469 * runnable, which means that it will never become
1470 * running in the future either. We're all done!
1479 * kick_process - kick a running thread to enter/exit the kernel
1480 * @p: the to-be-kicked thread
1482 * Cause a process which is running on another CPU to enter
1483 * kernel-mode, without any delay. (to get signals handled.)
1485 * NOTE: this function doesn't have to take the runqueue lock,
1486 * because all it wants to ensure is that the remote task enters
1487 * the kernel. If the IPI races and the task has been migrated
1488 * to another CPU then no harm is done and the purpose has been
1491 void kick_process(struct task_struct *p)
1497 if ((cpu != smp_processor_id()) && task_curr(p))
1498 smp_send_reschedule(cpu);
1501 EXPORT_SYMBOL_GPL(kick_process);
1504 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1506 * A few notes on cpu_active vs cpu_online:
1508 * - cpu_active must be a subset of cpu_online
1510 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1511 * see __set_cpus_allowed_ptr(). At this point the newly online
1512 * cpu isn't yet part of the sched domains, and balancing will not
1515 * - on cpu-down we clear cpu_active() to mask the sched domains and
1516 * avoid the load balancer to place new tasks on the to be removed
1517 * cpu. Existing tasks will remain running there and will be taken
1520 * This means that fallback selection must not select !active CPUs.
1521 * And can assume that any active CPU must be online. Conversely
1522 * select_task_rq() below may allow selection of !active CPUs in order
1523 * to satisfy the above rules.
1525 static int select_fallback_rq(int cpu, struct task_struct *p)
1527 int nid = cpu_to_node(cpu);
1528 const struct cpumask *nodemask = NULL;
1529 enum { cpuset, possible, fail } state = cpuset;
1533 * If the node that the cpu is on has been offlined, cpu_to_node()
1534 * will return -1. There is no cpu on the node, and we should
1535 * select the cpu on the other node.
1538 nodemask = cpumask_of_node(nid);
1540 /* Look for allowed, online CPU in same node. */
1541 for_each_cpu(dest_cpu, nodemask) {
1542 if (!cpu_active(dest_cpu))
1544 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1550 /* Any allowed, online CPU? */
1551 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1552 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1554 if (!cpu_online(dest_cpu))
1559 /* No more Mr. Nice Guy. */
1562 if (IS_ENABLED(CONFIG_CPUSETS)) {
1563 cpuset_cpus_allowed_fallback(p);
1569 do_set_cpus_allowed(p, cpu_possible_mask);
1580 if (state != cpuset) {
1582 * Don't tell them about moving exiting tasks or
1583 * kernel threads (both mm NULL), since they never
1586 if (p->mm && printk_ratelimit()) {
1587 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1588 task_pid_nr(p), p->comm, cpu);
1596 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1599 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1601 lockdep_assert_held(&p->pi_lock);
1603 if (tsk_nr_cpus_allowed(p) > 1)
1604 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1606 cpu = cpumask_any(tsk_cpus_allowed(p));
1609 * In order not to call set_task_cpu() on a blocking task we need
1610 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1613 * Since this is common to all placement strategies, this lives here.
1615 * [ this allows ->select_task() to simply return task_cpu(p) and
1616 * not worry about this generic constraint ]
1618 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1620 cpu = select_fallback_rq(task_cpu(p), p);
1625 static void update_avg(u64 *avg, u64 sample)
1627 s64 diff = sample - *avg;
1633 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1634 const struct cpumask *new_mask, bool check)
1636 return set_cpus_allowed_ptr(p, new_mask);
1639 #endif /* CONFIG_SMP */
1642 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1646 if (!schedstat_enabled())
1652 if (cpu == rq->cpu) {
1653 schedstat_inc(rq->ttwu_local);
1654 schedstat_inc(p->se.statistics.nr_wakeups_local);
1656 struct sched_domain *sd;
1658 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1660 for_each_domain(rq->cpu, sd) {
1661 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1662 schedstat_inc(sd->ttwu_wake_remote);
1669 if (wake_flags & WF_MIGRATED)
1670 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1671 #endif /* CONFIG_SMP */
1673 schedstat_inc(rq->ttwu_count);
1674 schedstat_inc(p->se.statistics.nr_wakeups);
1676 if (wake_flags & WF_SYNC)
1677 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1680 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1682 activate_task(rq, p, en_flags);
1683 p->on_rq = TASK_ON_RQ_QUEUED;
1685 /* if a worker is waking up, notify workqueue */
1686 if (p->flags & PF_WQ_WORKER)
1687 wq_worker_waking_up(p, cpu_of(rq));
1691 * Mark the task runnable and perform wakeup-preemption.
1693 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1694 struct pin_cookie cookie)
1696 check_preempt_curr(rq, p, wake_flags);
1697 p->state = TASK_RUNNING;
1698 trace_sched_wakeup(p);
1701 if (p->sched_class->task_woken) {
1703 * Our task @p is fully woken up and running; so its safe to
1704 * drop the rq->lock, hereafter rq is only used for statistics.
1706 lockdep_unpin_lock(&rq->lock, cookie);
1707 p->sched_class->task_woken(rq, p);
1708 lockdep_repin_lock(&rq->lock, cookie);
1711 if (rq->idle_stamp) {
1712 u64 delta = rq_clock(rq) - rq->idle_stamp;
1713 u64 max = 2*rq->max_idle_balance_cost;
1715 update_avg(&rq->avg_idle, delta);
1717 if (rq->avg_idle > max)
1726 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1727 struct pin_cookie cookie)
1729 int en_flags = ENQUEUE_WAKEUP;
1731 lockdep_assert_held(&rq->lock);
1734 if (p->sched_contributes_to_load)
1735 rq->nr_uninterruptible--;
1737 if (wake_flags & WF_MIGRATED)
1738 en_flags |= ENQUEUE_MIGRATED;
1741 ttwu_activate(rq, p, en_flags);
1742 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1746 * Called in case the task @p isn't fully descheduled from its runqueue,
1747 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1748 * since all we need to do is flip p->state to TASK_RUNNING, since
1749 * the task is still ->on_rq.
1751 static int ttwu_remote(struct task_struct *p, int wake_flags)
1757 rq = __task_rq_lock(p, &rf);
1758 if (task_on_rq_queued(p)) {
1759 /* check_preempt_curr() may use rq clock */
1760 update_rq_clock(rq);
1761 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1764 __task_rq_unlock(rq, &rf);
1770 void sched_ttwu_pending(void)
1772 struct rq *rq = this_rq();
1773 struct llist_node *llist = llist_del_all(&rq->wake_list);
1774 struct pin_cookie cookie;
1775 struct task_struct *p;
1776 unsigned long flags;
1781 raw_spin_lock_irqsave(&rq->lock, flags);
1782 cookie = lockdep_pin_lock(&rq->lock);
1787 p = llist_entry(llist, struct task_struct, wake_entry);
1788 llist = llist_next(llist);
1790 if (p->sched_remote_wakeup)
1791 wake_flags = WF_MIGRATED;
1793 ttwu_do_activate(rq, p, wake_flags, cookie);
1796 lockdep_unpin_lock(&rq->lock, cookie);
1797 raw_spin_unlock_irqrestore(&rq->lock, flags);
1800 void scheduler_ipi(void)
1803 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1804 * TIF_NEED_RESCHED remotely (for the first time) will also send
1807 preempt_fold_need_resched();
1809 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1813 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1814 * traditionally all their work was done from the interrupt return
1815 * path. Now that we actually do some work, we need to make sure
1818 * Some archs already do call them, luckily irq_enter/exit nest
1821 * Arguably we should visit all archs and update all handlers,
1822 * however a fair share of IPIs are still resched only so this would
1823 * somewhat pessimize the simple resched case.
1826 sched_ttwu_pending();
1829 * Check if someone kicked us for doing the nohz idle load balance.
1831 if (unlikely(got_nohz_idle_kick())) {
1832 this_rq()->idle_balance = 1;
1833 raise_softirq_irqoff(SCHED_SOFTIRQ);
1838 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1840 struct rq *rq = cpu_rq(cpu);
1842 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1844 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1845 if (!set_nr_if_polling(rq->idle))
1846 smp_send_reschedule(cpu);
1848 trace_sched_wake_idle_without_ipi(cpu);
1852 void wake_up_if_idle(int cpu)
1854 struct rq *rq = cpu_rq(cpu);
1855 unsigned long flags;
1859 if (!is_idle_task(rcu_dereference(rq->curr)))
1862 if (set_nr_if_polling(rq->idle)) {
1863 trace_sched_wake_idle_without_ipi(cpu);
1865 raw_spin_lock_irqsave(&rq->lock, flags);
1866 if (is_idle_task(rq->curr))
1867 smp_send_reschedule(cpu);
1868 /* Else cpu is not in idle, do nothing here */
1869 raw_spin_unlock_irqrestore(&rq->lock, flags);
1876 bool cpus_share_cache(int this_cpu, int that_cpu)
1878 if (this_cpu == that_cpu)
1881 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1883 #endif /* CONFIG_SMP */
1885 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1887 struct rq *rq = cpu_rq(cpu);
1888 struct pin_cookie cookie;
1890 #if defined(CONFIG_SMP)
1891 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1892 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1893 ttwu_queue_remote(p, cpu, wake_flags);
1898 raw_spin_lock(&rq->lock);
1899 cookie = lockdep_pin_lock(&rq->lock);
1900 ttwu_do_activate(rq, p, wake_flags, cookie);
1901 lockdep_unpin_lock(&rq->lock, cookie);
1902 raw_spin_unlock(&rq->lock);
1906 * Notes on Program-Order guarantees on SMP systems.
1910 * The basic program-order guarantee on SMP systems is that when a task [t]
1911 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1912 * execution on its new cpu [c1].
1914 * For migration (of runnable tasks) this is provided by the following means:
1916 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1917 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1918 * rq(c1)->lock (if not at the same time, then in that order).
1919 * C) LOCK of the rq(c1)->lock scheduling in task
1921 * Transitivity guarantees that B happens after A and C after B.
1922 * Note: we only require RCpc transitivity.
1923 * Note: the cpu doing B need not be c0 or c1
1932 * UNLOCK rq(0)->lock
1934 * LOCK rq(0)->lock // orders against CPU0
1936 * UNLOCK rq(0)->lock
1940 * UNLOCK rq(1)->lock
1942 * LOCK rq(1)->lock // orders against CPU2
1945 * UNLOCK rq(1)->lock
1948 * BLOCKING -- aka. SLEEP + WAKEUP
1950 * For blocking we (obviously) need to provide the same guarantee as for
1951 * migration. However the means are completely different as there is no lock
1952 * chain to provide order. Instead we do:
1954 * 1) smp_store_release(X->on_cpu, 0)
1955 * 2) smp_cond_load_acquire(!X->on_cpu)
1959 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1961 * LOCK rq(0)->lock LOCK X->pi_lock
1964 * smp_store_release(X->on_cpu, 0);
1966 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1972 * X->state = RUNNING
1973 * UNLOCK rq(2)->lock
1975 * LOCK rq(2)->lock // orders against CPU1
1978 * UNLOCK rq(2)->lock
1981 * UNLOCK rq(0)->lock
1984 * However; for wakeups there is a second guarantee we must provide, namely we
1985 * must observe the state that lead to our wakeup. That is, not only must our
1986 * task observe its own prior state, it must also observe the stores prior to
1989 * This means that any means of doing remote wakeups must order the CPU doing
1990 * the wakeup against the CPU the task is going to end up running on. This,
1991 * however, is already required for the regular Program-Order guarantee above,
1992 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1997 * try_to_wake_up - wake up a thread
1998 * @p: the thread to be awakened
1999 * @state: the mask of task states that can be woken
2000 * @wake_flags: wake modifier flags (WF_*)
2002 * Put it on the run-queue if it's not already there. The "current"
2003 * thread is always on the run-queue (except when the actual
2004 * re-schedule is in progress), and as such you're allowed to do
2005 * the simpler "current->state = TASK_RUNNING" to mark yourself
2006 * runnable without the overhead of this.
2008 * Return: %true if @p was woken up, %false if it was already running.
2009 * or @state didn't match @p's state.
2012 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2014 unsigned long flags;
2015 int cpu, success = 0;
2018 * If we are going to wake up a thread waiting for CONDITION we
2019 * need to ensure that CONDITION=1 done by the caller can not be
2020 * reordered with p->state check below. This pairs with mb() in
2021 * set_current_state() the waiting thread does.
2023 smp_mb__before_spinlock();
2024 raw_spin_lock_irqsave(&p->pi_lock, flags);
2025 if (!(p->state & state))
2028 trace_sched_waking(p);
2030 success = 1; /* we're going to change ->state */
2034 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2035 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2036 * in smp_cond_load_acquire() below.
2038 * sched_ttwu_pending() try_to_wake_up()
2039 * [S] p->on_rq = 1; [L] P->state
2040 * UNLOCK rq->lock -----.
2044 * LOCK rq->lock -----'
2048 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2050 * Pairs with the UNLOCK+LOCK on rq->lock from the
2051 * last wakeup of our task and the schedule that got our task
2055 if (p->on_rq && ttwu_remote(p, wake_flags))
2060 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2061 * possible to, falsely, observe p->on_cpu == 0.
2063 * One must be running (->on_cpu == 1) in order to remove oneself
2064 * from the runqueue.
2066 * [S] ->on_cpu = 1; [L] ->on_rq
2070 * [S] ->on_rq = 0; [L] ->on_cpu
2072 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2073 * from the consecutive calls to schedule(); the first switching to our
2074 * task, the second putting it to sleep.
2079 * If the owning (remote) cpu is still in the middle of schedule() with
2080 * this task as prev, wait until its done referencing the task.
2082 * Pairs with the smp_store_release() in finish_lock_switch().
2084 * This ensures that tasks getting woken will be fully ordered against
2085 * their previous state and preserve Program Order.
2087 smp_cond_load_acquire(&p->on_cpu, !VAL);
2089 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2090 p->state = TASK_WAKING;
2092 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2093 if (task_cpu(p) != cpu) {
2094 wake_flags |= WF_MIGRATED;
2095 set_task_cpu(p, cpu);
2097 #endif /* CONFIG_SMP */
2099 ttwu_queue(p, cpu, wake_flags);
2101 ttwu_stat(p, cpu, wake_flags);
2103 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2109 * try_to_wake_up_local - try to wake up a local task with rq lock held
2110 * @p: the thread to be awakened
2111 * @cookie: context's cookie for pinning
2113 * Put @p on the run-queue if it's not already there. The caller must
2114 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2117 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2119 struct rq *rq = task_rq(p);
2121 if (WARN_ON_ONCE(rq != this_rq()) ||
2122 WARN_ON_ONCE(p == current))
2125 lockdep_assert_held(&rq->lock);
2127 if (!raw_spin_trylock(&p->pi_lock)) {
2129 * This is OK, because current is on_cpu, which avoids it being
2130 * picked for load-balance and preemption/IRQs are still
2131 * disabled avoiding further scheduler activity on it and we've
2132 * not yet picked a replacement task.
2134 lockdep_unpin_lock(&rq->lock, cookie);
2135 raw_spin_unlock(&rq->lock);
2136 raw_spin_lock(&p->pi_lock);
2137 raw_spin_lock(&rq->lock);
2138 lockdep_repin_lock(&rq->lock, cookie);
2141 if (!(p->state & TASK_NORMAL))
2144 trace_sched_waking(p);
2146 if (!task_on_rq_queued(p))
2147 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2149 ttwu_do_wakeup(rq, p, 0, cookie);
2150 ttwu_stat(p, smp_processor_id(), 0);
2152 raw_spin_unlock(&p->pi_lock);
2156 * wake_up_process - Wake up a specific process
2157 * @p: The process to be woken up.
2159 * Attempt to wake up the nominated process and move it to the set of runnable
2162 * Return: 1 if the process was woken up, 0 if it was already running.
2164 * It may be assumed that this function implies a write memory barrier before
2165 * changing the task state if and only if any tasks are woken up.
2167 int wake_up_process(struct task_struct *p)
2169 return try_to_wake_up(p, TASK_NORMAL, 0);
2171 EXPORT_SYMBOL(wake_up_process);
2173 int wake_up_state(struct task_struct *p, unsigned int state)
2175 return try_to_wake_up(p, state, 0);
2179 * This function clears the sched_dl_entity static params.
2181 void __dl_clear_params(struct task_struct *p)
2183 struct sched_dl_entity *dl_se = &p->dl;
2185 dl_se->dl_runtime = 0;
2186 dl_se->dl_deadline = 0;
2187 dl_se->dl_period = 0;
2190 dl_se->dl_density = 0;
2192 dl_se->dl_throttled = 0;
2193 dl_se->dl_yielded = 0;
2197 * Perform scheduler related setup for a newly forked process p.
2198 * p is forked by current.
2200 * __sched_fork() is basic setup used by init_idle() too:
2202 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2207 p->se.exec_start = 0;
2208 p->se.sum_exec_runtime = 0;
2209 p->se.prev_sum_exec_runtime = 0;
2210 p->se.nr_migrations = 0;
2212 INIT_LIST_HEAD(&p->se.group_node);
2214 #ifdef CONFIG_FAIR_GROUP_SCHED
2215 p->se.cfs_rq = NULL;
2218 #ifdef CONFIG_SCHEDSTATS
2219 /* Even if schedstat is disabled, there should not be garbage */
2220 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2223 RB_CLEAR_NODE(&p->dl.rb_node);
2224 init_dl_task_timer(&p->dl);
2225 __dl_clear_params(p);
2227 INIT_LIST_HEAD(&p->rt.run_list);
2229 p->rt.time_slice = sched_rr_timeslice;
2233 #ifdef CONFIG_PREEMPT_NOTIFIERS
2234 INIT_HLIST_HEAD(&p->preempt_notifiers);
2237 #ifdef CONFIG_NUMA_BALANCING
2238 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2239 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2240 p->mm->numa_scan_seq = 0;
2243 if (clone_flags & CLONE_VM)
2244 p->numa_preferred_nid = current->numa_preferred_nid;
2246 p->numa_preferred_nid = -1;
2248 p->node_stamp = 0ULL;
2249 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2250 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2251 p->numa_work.next = &p->numa_work;
2252 p->numa_faults = NULL;
2253 p->last_task_numa_placement = 0;
2254 p->last_sum_exec_runtime = 0;
2256 p->numa_group = NULL;
2257 #endif /* CONFIG_NUMA_BALANCING */
2260 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2262 #ifdef CONFIG_NUMA_BALANCING
2264 void set_numabalancing_state(bool enabled)
2267 static_branch_enable(&sched_numa_balancing);
2269 static_branch_disable(&sched_numa_balancing);
2272 #ifdef CONFIG_PROC_SYSCTL
2273 int sysctl_numa_balancing(struct ctl_table *table, int write,
2274 void __user *buffer, size_t *lenp, loff_t *ppos)
2278 int state = static_branch_likely(&sched_numa_balancing);
2280 if (write && !capable(CAP_SYS_ADMIN))
2285 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2289 set_numabalancing_state(state);
2295 #ifdef CONFIG_SCHEDSTATS
2297 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2298 static bool __initdata __sched_schedstats = false;
2300 static void set_schedstats(bool enabled)
2303 static_branch_enable(&sched_schedstats);
2305 static_branch_disable(&sched_schedstats);
2308 void force_schedstat_enabled(void)
2310 if (!schedstat_enabled()) {
2311 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2312 static_branch_enable(&sched_schedstats);
2316 static int __init setup_schedstats(char *str)
2323 * This code is called before jump labels have been set up, so we can't
2324 * change the static branch directly just yet. Instead set a temporary
2325 * variable so init_schedstats() can do it later.
2327 if (!strcmp(str, "enable")) {
2328 __sched_schedstats = true;
2330 } else if (!strcmp(str, "disable")) {
2331 __sched_schedstats = false;
2336 pr_warn("Unable to parse schedstats=\n");
2340 __setup("schedstats=", setup_schedstats);
2342 static void __init init_schedstats(void)
2344 set_schedstats(__sched_schedstats);
2347 #ifdef CONFIG_PROC_SYSCTL
2348 int sysctl_schedstats(struct ctl_table *table, int write,
2349 void __user *buffer, size_t *lenp, loff_t *ppos)
2353 int state = static_branch_likely(&sched_schedstats);
2355 if (write && !capable(CAP_SYS_ADMIN))
2360 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2364 set_schedstats(state);
2367 #endif /* CONFIG_PROC_SYSCTL */
2368 #else /* !CONFIG_SCHEDSTATS */
2369 static inline void init_schedstats(void) {}
2370 #endif /* CONFIG_SCHEDSTATS */
2373 * fork()/clone()-time setup:
2375 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2377 unsigned long flags;
2378 int cpu = get_cpu();
2380 __sched_fork(clone_flags, p);
2382 * We mark the process as NEW here. This guarantees that
2383 * nobody will actually run it, and a signal or other external
2384 * event cannot wake it up and insert it on the runqueue either.
2386 p->state = TASK_NEW;
2389 * Make sure we do not leak PI boosting priority to the child.
2391 p->prio = current->normal_prio;
2394 * Revert to default priority/policy on fork if requested.
2396 if (unlikely(p->sched_reset_on_fork)) {
2397 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2398 p->policy = SCHED_NORMAL;
2399 p->static_prio = NICE_TO_PRIO(0);
2401 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2402 p->static_prio = NICE_TO_PRIO(0);
2404 p->prio = p->normal_prio = __normal_prio(p);
2408 * We don't need the reset flag anymore after the fork. It has
2409 * fulfilled its duty:
2411 p->sched_reset_on_fork = 0;
2414 if (dl_prio(p->prio)) {
2417 } else if (rt_prio(p->prio)) {
2418 p->sched_class = &rt_sched_class;
2420 p->sched_class = &fair_sched_class;
2423 init_entity_runnable_average(&p->se);
2426 * The child is not yet in the pid-hash so no cgroup attach races,
2427 * and the cgroup is pinned to this child due to cgroup_fork()
2428 * is ran before sched_fork().
2430 * Silence PROVE_RCU.
2432 raw_spin_lock_irqsave(&p->pi_lock, flags);
2434 * We're setting the cpu for the first time, we don't migrate,
2435 * so use __set_task_cpu().
2437 __set_task_cpu(p, cpu);
2438 if (p->sched_class->task_fork)
2439 p->sched_class->task_fork(p);
2440 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2442 #ifdef CONFIG_SCHED_INFO
2443 if (likely(sched_info_on()))
2444 memset(&p->sched_info, 0, sizeof(p->sched_info));
2446 #if defined(CONFIG_SMP)
2449 init_task_preempt_count(p);
2451 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2452 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2459 unsigned long to_ratio(u64 period, u64 runtime)
2461 if (runtime == RUNTIME_INF)
2465 * Doing this here saves a lot of checks in all
2466 * the calling paths, and returning zero seems
2467 * safe for them anyway.
2472 return div64_u64(runtime << 20, period);
2476 inline struct dl_bw *dl_bw_of(int i)
2478 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2479 "sched RCU must be held");
2480 return &cpu_rq(i)->rd->dl_bw;
2483 static inline int dl_bw_cpus(int i)
2485 struct root_domain *rd = cpu_rq(i)->rd;
2488 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2489 "sched RCU must be held");
2490 for_each_cpu_and(i, rd->span, cpu_active_mask)
2496 inline struct dl_bw *dl_bw_of(int i)
2498 return &cpu_rq(i)->dl.dl_bw;
2501 static inline int dl_bw_cpus(int i)
2508 * We must be sure that accepting a new task (or allowing changing the
2509 * parameters of an existing one) is consistent with the bandwidth
2510 * constraints. If yes, this function also accordingly updates the currently
2511 * allocated bandwidth to reflect the new situation.
2513 * This function is called while holding p's rq->lock.
2515 * XXX we should delay bw change until the task's 0-lag point, see
2518 static int dl_overflow(struct task_struct *p, int policy,
2519 const struct sched_attr *attr)
2522 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2523 u64 period = attr->sched_period ?: attr->sched_deadline;
2524 u64 runtime = attr->sched_runtime;
2525 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2528 /* !deadline task may carry old deadline bandwidth */
2529 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2533 * Either if a task, enters, leave, or stays -deadline but changes
2534 * its parameters, we may need to update accordingly the total
2535 * allocated bandwidth of the container.
2537 raw_spin_lock(&dl_b->lock);
2538 cpus = dl_bw_cpus(task_cpu(p));
2539 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2540 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2541 __dl_add(dl_b, new_bw);
2543 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2544 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2545 __dl_clear(dl_b, p->dl.dl_bw);
2546 __dl_add(dl_b, new_bw);
2548 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2549 __dl_clear(dl_b, p->dl.dl_bw);
2552 raw_spin_unlock(&dl_b->lock);
2557 extern void init_dl_bw(struct dl_bw *dl_b);
2560 * wake_up_new_task - wake up a newly created task for the first time.
2562 * This function will do some initial scheduler statistics housekeeping
2563 * that must be done for every newly created context, then puts the task
2564 * on the runqueue and wakes it.
2566 void wake_up_new_task(struct task_struct *p)
2571 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2572 p->state = TASK_RUNNING;
2575 * Fork balancing, do it here and not earlier because:
2576 * - cpus_allowed can change in the fork path
2577 * - any previously selected cpu might disappear through hotplug
2579 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2580 * as we're not fully set-up yet.
2582 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2584 rq = __task_rq_lock(p, &rf);
2585 post_init_entity_util_avg(&p->se);
2587 activate_task(rq, p, 0);
2588 p->on_rq = TASK_ON_RQ_QUEUED;
2589 trace_sched_wakeup_new(p);
2590 check_preempt_curr(rq, p, WF_FORK);
2592 if (p->sched_class->task_woken) {
2594 * Nothing relies on rq->lock after this, so its fine to
2597 lockdep_unpin_lock(&rq->lock, rf.cookie);
2598 p->sched_class->task_woken(rq, p);
2599 lockdep_repin_lock(&rq->lock, rf.cookie);
2602 task_rq_unlock(rq, p, &rf);
2605 #ifdef CONFIG_PREEMPT_NOTIFIERS
2607 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2609 void preempt_notifier_inc(void)
2611 static_key_slow_inc(&preempt_notifier_key);
2613 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2615 void preempt_notifier_dec(void)
2617 static_key_slow_dec(&preempt_notifier_key);
2619 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2622 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2623 * @notifier: notifier struct to register
2625 void preempt_notifier_register(struct preempt_notifier *notifier)
2627 if (!static_key_false(&preempt_notifier_key))
2628 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2630 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2632 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2635 * preempt_notifier_unregister - no longer interested in preemption notifications
2636 * @notifier: notifier struct to unregister
2638 * This is *not* safe to call from within a preemption notifier.
2640 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2642 hlist_del(¬ifier->link);
2644 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2646 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2648 struct preempt_notifier *notifier;
2650 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2651 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2654 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2656 if (static_key_false(&preempt_notifier_key))
2657 __fire_sched_in_preempt_notifiers(curr);
2661 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2662 struct task_struct *next)
2664 struct preempt_notifier *notifier;
2666 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2667 notifier->ops->sched_out(notifier, next);
2670 static __always_inline void
2671 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2672 struct task_struct *next)
2674 if (static_key_false(&preempt_notifier_key))
2675 __fire_sched_out_preempt_notifiers(curr, next);
2678 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2680 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2685 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2686 struct task_struct *next)
2690 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2693 * prepare_task_switch - prepare to switch tasks
2694 * @rq: the runqueue preparing to switch
2695 * @prev: the current task that is being switched out
2696 * @next: the task we are going to switch to.
2698 * This is called with the rq lock held and interrupts off. It must
2699 * be paired with a subsequent finish_task_switch after the context
2702 * prepare_task_switch sets up locking and calls architecture specific
2706 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2707 struct task_struct *next)
2709 sched_info_switch(rq, prev, next);
2710 perf_event_task_sched_out(prev, next);
2711 fire_sched_out_preempt_notifiers(prev, next);
2712 prepare_lock_switch(rq, next);
2713 prepare_arch_switch(next);
2717 * finish_task_switch - clean up after a task-switch
2718 * @prev: the thread we just switched away from.
2720 * finish_task_switch must be called after the context switch, paired
2721 * with a prepare_task_switch call before the context switch.
2722 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2723 * and do any other architecture-specific cleanup actions.
2725 * Note that we may have delayed dropping an mm in context_switch(). If
2726 * so, we finish that here outside of the runqueue lock. (Doing it
2727 * with the lock held can cause deadlocks; see schedule() for
2730 * The context switch have flipped the stack from under us and restored the
2731 * local variables which were saved when this task called schedule() in the
2732 * past. prev == current is still correct but we need to recalculate this_rq
2733 * because prev may have moved to another CPU.
2735 static struct rq *finish_task_switch(struct task_struct *prev)
2736 __releases(rq->lock)
2738 struct rq *rq = this_rq();
2739 struct mm_struct *mm = rq->prev_mm;
2743 * The previous task will have left us with a preempt_count of 2
2744 * because it left us after:
2747 * preempt_disable(); // 1
2749 * raw_spin_lock_irq(&rq->lock) // 2
2751 * Also, see FORK_PREEMPT_COUNT.
2753 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2754 "corrupted preempt_count: %s/%d/0x%x\n",
2755 current->comm, current->pid, preempt_count()))
2756 preempt_count_set(FORK_PREEMPT_COUNT);
2761 * A task struct has one reference for the use as "current".
2762 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2763 * schedule one last time. The schedule call will never return, and
2764 * the scheduled task must drop that reference.
2766 * We must observe prev->state before clearing prev->on_cpu (in
2767 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2768 * running on another CPU and we could rave with its RUNNING -> DEAD
2769 * transition, resulting in a double drop.
2771 prev_state = prev->state;
2772 vtime_task_switch(prev);
2773 perf_event_task_sched_in(prev, current);
2774 finish_lock_switch(rq, prev);
2775 finish_arch_post_lock_switch();
2777 fire_sched_in_preempt_notifiers(current);
2780 if (unlikely(prev_state == TASK_DEAD)) {
2781 if (prev->sched_class->task_dead)
2782 prev->sched_class->task_dead(prev);
2785 * Remove function-return probe instances associated with this
2786 * task and put them back on the free list.
2788 kprobe_flush_task(prev);
2790 /* Task is done with its stack. */
2791 put_task_stack(prev);
2793 put_task_struct(prev);
2796 tick_nohz_task_switch();
2802 /* rq->lock is NOT held, but preemption is disabled */
2803 static void __balance_callback(struct rq *rq)
2805 struct callback_head *head, *next;
2806 void (*func)(struct rq *rq);
2807 unsigned long flags;
2809 raw_spin_lock_irqsave(&rq->lock, flags);
2810 head = rq->balance_callback;
2811 rq->balance_callback = NULL;
2813 func = (void (*)(struct rq *))head->func;
2820 raw_spin_unlock_irqrestore(&rq->lock, flags);
2823 static inline void balance_callback(struct rq *rq)
2825 if (unlikely(rq->balance_callback))
2826 __balance_callback(rq);
2831 static inline void balance_callback(struct rq *rq)
2838 * schedule_tail - first thing a freshly forked thread must call.
2839 * @prev: the thread we just switched away from.
2841 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2842 __releases(rq->lock)
2847 * New tasks start with FORK_PREEMPT_COUNT, see there and
2848 * finish_task_switch() for details.
2850 * finish_task_switch() will drop rq->lock() and lower preempt_count
2851 * and the preempt_enable() will end up enabling preemption (on
2852 * PREEMPT_COUNT kernels).
2855 rq = finish_task_switch(prev);
2856 balance_callback(rq);
2859 if (current->set_child_tid)
2860 put_user(task_pid_vnr(current), current->set_child_tid);
2864 * context_switch - switch to the new MM and the new thread's register state.
2866 static __always_inline struct rq *
2867 context_switch(struct rq *rq, struct task_struct *prev,
2868 struct task_struct *next, struct pin_cookie cookie)
2870 struct mm_struct *mm, *oldmm;
2872 prepare_task_switch(rq, prev, next);
2875 oldmm = prev->active_mm;
2877 * For paravirt, this is coupled with an exit in switch_to to
2878 * combine the page table reload and the switch backend into
2881 arch_start_context_switch(prev);
2884 next->active_mm = oldmm;
2885 atomic_inc(&oldmm->mm_count);
2886 enter_lazy_tlb(oldmm, next);
2888 switch_mm_irqs_off(oldmm, mm, next);
2891 prev->active_mm = NULL;
2892 rq->prev_mm = oldmm;
2895 * Since the runqueue lock will be released by the next
2896 * task (which is an invalid locking op but in the case
2897 * of the scheduler it's an obvious special-case), so we
2898 * do an early lockdep release here:
2900 lockdep_unpin_lock(&rq->lock, cookie);
2901 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2903 /* Here we just switch the register state and the stack. */
2904 switch_to(prev, next, prev);
2907 return finish_task_switch(prev);
2911 * nr_running and nr_context_switches:
2913 * externally visible scheduler statistics: current number of runnable
2914 * threads, total number of context switches performed since bootup.
2916 unsigned long nr_running(void)
2918 unsigned long i, sum = 0;
2920 for_each_online_cpu(i)
2921 sum += cpu_rq(i)->nr_running;
2927 * Check if only the current task is running on the cpu.
2929 * Caution: this function does not check that the caller has disabled
2930 * preemption, thus the result might have a time-of-check-to-time-of-use
2931 * race. The caller is responsible to use it correctly, for example:
2933 * - from a non-preemptable section (of course)
2935 * - from a thread that is bound to a single CPU
2937 * - in a loop with very short iterations (e.g. a polling loop)
2939 bool single_task_running(void)
2941 return raw_rq()->nr_running == 1;
2943 EXPORT_SYMBOL(single_task_running);
2945 unsigned long long nr_context_switches(void)
2948 unsigned long long sum = 0;
2950 for_each_possible_cpu(i)
2951 sum += cpu_rq(i)->nr_switches;
2956 unsigned long nr_iowait(void)
2958 unsigned long i, sum = 0;
2960 for_each_possible_cpu(i)
2961 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2966 unsigned long nr_iowait_cpu(int cpu)
2968 struct rq *this = cpu_rq(cpu);
2969 return atomic_read(&this->nr_iowait);
2972 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2974 struct rq *rq = this_rq();
2975 *nr_waiters = atomic_read(&rq->nr_iowait);
2976 *load = rq->load.weight;
2982 * sched_exec - execve() is a valuable balancing opportunity, because at
2983 * this point the task has the smallest effective memory and cache footprint.
2985 void sched_exec(void)
2987 struct task_struct *p = current;
2988 unsigned long flags;
2991 raw_spin_lock_irqsave(&p->pi_lock, flags);
2992 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2993 if (dest_cpu == smp_processor_id())
2996 if (likely(cpu_active(dest_cpu))) {
2997 struct migration_arg arg = { p, dest_cpu };
2999 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3000 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3004 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3009 DEFINE_PER_CPU(struct kernel_stat, kstat);
3010 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3012 EXPORT_PER_CPU_SYMBOL(kstat);
3013 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3016 * The function fair_sched_class.update_curr accesses the struct curr
3017 * and its field curr->exec_start; when called from task_sched_runtime(),
3018 * we observe a high rate of cache misses in practice.
3019 * Prefetching this data results in improved performance.
3021 static inline void prefetch_curr_exec_start(struct task_struct *p)
3023 #ifdef CONFIG_FAIR_GROUP_SCHED
3024 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3026 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3029 prefetch(&curr->exec_start);
3033 * Return accounted runtime for the task.
3034 * In case the task is currently running, return the runtime plus current's
3035 * pending runtime that have not been accounted yet.
3037 unsigned long long task_sched_runtime(struct task_struct *p)
3043 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3045 * 64-bit doesn't need locks to atomically read a 64bit value.
3046 * So we have a optimization chance when the task's delta_exec is 0.
3047 * Reading ->on_cpu is racy, but this is ok.
3049 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3050 * If we race with it entering cpu, unaccounted time is 0. This is
3051 * indistinguishable from the read occurring a few cycles earlier.
3052 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3053 * been accounted, so we're correct here as well.
3055 if (!p->on_cpu || !task_on_rq_queued(p))
3056 return p->se.sum_exec_runtime;
3059 rq = task_rq_lock(p, &rf);
3061 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3062 * project cycles that may never be accounted to this
3063 * thread, breaking clock_gettime().
3065 if (task_current(rq, p) && task_on_rq_queued(p)) {
3066 prefetch_curr_exec_start(p);
3067 update_rq_clock(rq);
3068 p->sched_class->update_curr(rq);
3070 ns = p->se.sum_exec_runtime;
3071 task_rq_unlock(rq, p, &rf);
3077 * This function gets called by the timer code, with HZ frequency.
3078 * We call it with interrupts disabled.
3080 void scheduler_tick(void)
3082 int cpu = smp_processor_id();
3083 struct rq *rq = cpu_rq(cpu);
3084 struct task_struct *curr = rq->curr;
3088 raw_spin_lock(&rq->lock);
3089 update_rq_clock(rq);
3090 curr->sched_class->task_tick(rq, curr, 0);
3091 cpu_load_update_active(rq);
3092 calc_global_load_tick(rq);
3093 raw_spin_unlock(&rq->lock);
3095 perf_event_task_tick();
3098 rq->idle_balance = idle_cpu(cpu);
3099 trigger_load_balance(rq);
3101 rq_last_tick_reset(rq);
3104 #ifdef CONFIG_NO_HZ_FULL
3106 * scheduler_tick_max_deferment
3108 * Keep at least one tick per second when a single
3109 * active task is running because the scheduler doesn't
3110 * yet completely support full dynticks environment.
3112 * This makes sure that uptime, CFS vruntime, load
3113 * balancing, etc... continue to move forward, even
3114 * with a very low granularity.
3116 * Return: Maximum deferment in nanoseconds.
3118 u64 scheduler_tick_max_deferment(void)
3120 struct rq *rq = this_rq();
3121 unsigned long next, now = READ_ONCE(jiffies);
3123 next = rq->last_sched_tick + HZ;
3125 if (time_before_eq(next, now))
3128 return jiffies_to_nsecs(next - now);
3132 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3133 defined(CONFIG_PREEMPT_TRACER))
3135 * If the value passed in is equal to the current preempt count
3136 * then we just disabled preemption. Start timing the latency.
3138 static inline void preempt_latency_start(int val)
3140 if (preempt_count() == val) {
3141 unsigned long ip = get_lock_parent_ip();
3142 #ifdef CONFIG_DEBUG_PREEMPT
3143 current->preempt_disable_ip = ip;
3145 trace_preempt_off(CALLER_ADDR0, ip);
3149 void preempt_count_add(int val)
3151 #ifdef CONFIG_DEBUG_PREEMPT
3155 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3158 __preempt_count_add(val);
3159 #ifdef CONFIG_DEBUG_PREEMPT
3161 * Spinlock count overflowing soon?
3163 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3166 preempt_latency_start(val);
3168 EXPORT_SYMBOL(preempt_count_add);
3169 NOKPROBE_SYMBOL(preempt_count_add);
3172 * If the value passed in equals to the current preempt count
3173 * then we just enabled preemption. Stop timing the latency.
3175 static inline void preempt_latency_stop(int val)
3177 if (preempt_count() == val)
3178 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3181 void preempt_count_sub(int val)
3183 #ifdef CONFIG_DEBUG_PREEMPT
3187 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3190 * Is the spinlock portion underflowing?
3192 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3193 !(preempt_count() & PREEMPT_MASK)))
3197 preempt_latency_stop(val);
3198 __preempt_count_sub(val);
3200 EXPORT_SYMBOL(preempt_count_sub);
3201 NOKPROBE_SYMBOL(preempt_count_sub);
3204 static inline void preempt_latency_start(int val) { }
3205 static inline void preempt_latency_stop(int val) { }
3209 * Print scheduling while atomic bug:
3211 static noinline void __schedule_bug(struct task_struct *prev)
3213 /* Save this before calling printk(), since that will clobber it */
3214 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3216 if (oops_in_progress)
3219 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3220 prev->comm, prev->pid, preempt_count());
3222 debug_show_held_locks(prev);
3224 if (irqs_disabled())
3225 print_irqtrace_events(prev);
3226 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3227 && in_atomic_preempt_off()) {
3228 pr_err("Preemption disabled at:");
3229 print_ip_sym(preempt_disable_ip);
3233 panic("scheduling while atomic\n");
3236 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3240 * Various schedule()-time debugging checks and statistics:
3242 static inline void schedule_debug(struct task_struct *prev)
3244 #ifdef CONFIG_SCHED_STACK_END_CHECK
3245 if (task_stack_end_corrupted(prev))
3246 panic("corrupted stack end detected inside scheduler\n");
3249 if (unlikely(in_atomic_preempt_off())) {
3250 __schedule_bug(prev);
3251 preempt_count_set(PREEMPT_DISABLED);
3255 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3257 schedstat_inc(this_rq()->sched_count);
3261 * Pick up the highest-prio task:
3263 static inline struct task_struct *
3264 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3266 const struct sched_class *class = &fair_sched_class;
3267 struct task_struct *p;
3270 * Optimization: we know that if all tasks are in
3271 * the fair class we can call that function directly:
3273 if (likely(prev->sched_class == class &&
3274 rq->nr_running == rq->cfs.h_nr_running)) {
3275 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3276 if (unlikely(p == RETRY_TASK))
3279 /* assumes fair_sched_class->next == idle_sched_class */
3281 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3287 for_each_class(class) {
3288 p = class->pick_next_task(rq, prev, cookie);
3290 if (unlikely(p == RETRY_TASK))
3296 BUG(); /* the idle class will always have a runnable task */
3300 * __schedule() is the main scheduler function.
3302 * The main means of driving the scheduler and thus entering this function are:
3304 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3306 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3307 * paths. For example, see arch/x86/entry_64.S.
3309 * To drive preemption between tasks, the scheduler sets the flag in timer
3310 * interrupt handler scheduler_tick().
3312 * 3. Wakeups don't really cause entry into schedule(). They add a
3313 * task to the run-queue and that's it.
3315 * Now, if the new task added to the run-queue preempts the current
3316 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3317 * called on the nearest possible occasion:
3319 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3321 * - in syscall or exception context, at the next outmost
3322 * preempt_enable(). (this might be as soon as the wake_up()'s
3325 * - in IRQ context, return from interrupt-handler to
3326 * preemptible context
3328 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3331 * - cond_resched() call
3332 * - explicit schedule() call
3333 * - return from syscall or exception to user-space
3334 * - return from interrupt-handler to user-space
3336 * WARNING: must be called with preemption disabled!
3338 static void __sched notrace __schedule(bool preempt)
3340 struct task_struct *prev, *next;
3341 unsigned long *switch_count;
3342 struct pin_cookie cookie;
3346 cpu = smp_processor_id();
3350 schedule_debug(prev);
3352 if (sched_feat(HRTICK))
3355 local_irq_disable();
3356 rcu_note_context_switch();
3359 * Make sure that signal_pending_state()->signal_pending() below
3360 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3361 * done by the caller to avoid the race with signal_wake_up().
3363 smp_mb__before_spinlock();
3364 raw_spin_lock(&rq->lock);
3365 cookie = lockdep_pin_lock(&rq->lock);
3367 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3369 switch_count = &prev->nivcsw;
3370 if (!preempt && prev->state) {
3371 if (unlikely(signal_pending_state(prev->state, prev))) {
3372 prev->state = TASK_RUNNING;
3374 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3378 * If a worker went to sleep, notify and ask workqueue
3379 * whether it wants to wake up a task to maintain
3382 if (prev->flags & PF_WQ_WORKER) {
3383 struct task_struct *to_wakeup;
3385 to_wakeup = wq_worker_sleeping(prev);
3387 try_to_wake_up_local(to_wakeup, cookie);
3390 switch_count = &prev->nvcsw;
3393 if (task_on_rq_queued(prev))
3394 update_rq_clock(rq);
3396 next = pick_next_task(rq, prev, cookie);
3397 clear_tsk_need_resched(prev);
3398 clear_preempt_need_resched();
3399 rq->clock_skip_update = 0;
3401 if (likely(prev != next)) {
3406 trace_sched_switch(preempt, prev, next);
3407 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3409 lockdep_unpin_lock(&rq->lock, cookie);
3410 raw_spin_unlock_irq(&rq->lock);
3413 balance_callback(rq);
3416 void __noreturn do_task_dead(void)
3419 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3420 * when the following two conditions become true.
3421 * - There is race condition of mmap_sem (It is acquired by
3423 * - SMI occurs before setting TASK_RUNINNG.
3424 * (or hypervisor of virtual machine switches to other guest)
3425 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3427 * To avoid it, we have to wait for releasing tsk->pi_lock which
3428 * is held by try_to_wake_up()
3431 raw_spin_unlock_wait(¤t->pi_lock);
3433 /* causes final put_task_struct in finish_task_switch(). */
3434 __set_current_state(TASK_DEAD);
3435 current->flags |= PF_NOFREEZE; /* tell freezer to ignore us */
3438 /* Avoid "noreturn function does return". */
3440 cpu_relax(); /* For when BUG is null */
3443 static inline void sched_submit_work(struct task_struct *tsk)
3445 if (!tsk->state || tsk_is_pi_blocked(tsk))
3448 * If we are going to sleep and we have plugged IO queued,
3449 * make sure to submit it to avoid deadlocks.
3451 if (blk_needs_flush_plug(tsk))
3452 blk_schedule_flush_plug(tsk);
3455 asmlinkage __visible void __sched schedule(void)
3457 struct task_struct *tsk = current;
3459 sched_submit_work(tsk);
3463 sched_preempt_enable_no_resched();
3464 } while (need_resched());
3466 EXPORT_SYMBOL(schedule);
3468 #ifdef CONFIG_CONTEXT_TRACKING
3469 asmlinkage __visible void __sched schedule_user(void)
3472 * If we come here after a random call to set_need_resched(),
3473 * or we have been woken up remotely but the IPI has not yet arrived,
3474 * we haven't yet exited the RCU idle mode. Do it here manually until
3475 * we find a better solution.
3477 * NB: There are buggy callers of this function. Ideally we
3478 * should warn if prev_state != CONTEXT_USER, but that will trigger
3479 * too frequently to make sense yet.
3481 enum ctx_state prev_state = exception_enter();
3483 exception_exit(prev_state);
3488 * schedule_preempt_disabled - called with preemption disabled
3490 * Returns with preemption disabled. Note: preempt_count must be 1
3492 void __sched schedule_preempt_disabled(void)
3494 sched_preempt_enable_no_resched();
3499 static void __sched notrace preempt_schedule_common(void)
3503 * Because the function tracer can trace preempt_count_sub()
3504 * and it also uses preempt_enable/disable_notrace(), if
3505 * NEED_RESCHED is set, the preempt_enable_notrace() called
3506 * by the function tracer will call this function again and
3507 * cause infinite recursion.
3509 * Preemption must be disabled here before the function
3510 * tracer can trace. Break up preempt_disable() into two
3511 * calls. One to disable preemption without fear of being
3512 * traced. The other to still record the preemption latency,
3513 * which can also be traced by the function tracer.
3515 preempt_disable_notrace();
3516 preempt_latency_start(1);
3518 preempt_latency_stop(1);
3519 preempt_enable_no_resched_notrace();
3522 * Check again in case we missed a preemption opportunity
3523 * between schedule and now.
3525 } while (need_resched());
3528 #ifdef CONFIG_PREEMPT
3530 * this is the entry point to schedule() from in-kernel preemption
3531 * off of preempt_enable. Kernel preemptions off return from interrupt
3532 * occur there and call schedule directly.
3534 asmlinkage __visible void __sched notrace preempt_schedule(void)
3537 * If there is a non-zero preempt_count or interrupts are disabled,
3538 * we do not want to preempt the current task. Just return..
3540 if (likely(!preemptible()))
3543 preempt_schedule_common();
3545 NOKPROBE_SYMBOL(preempt_schedule);
3546 EXPORT_SYMBOL(preempt_schedule);
3549 * preempt_schedule_notrace - preempt_schedule called by tracing
3551 * The tracing infrastructure uses preempt_enable_notrace to prevent
3552 * recursion and tracing preempt enabling caused by the tracing
3553 * infrastructure itself. But as tracing can happen in areas coming
3554 * from userspace or just about to enter userspace, a preempt enable
3555 * can occur before user_exit() is called. This will cause the scheduler
3556 * to be called when the system is still in usermode.
3558 * To prevent this, the preempt_enable_notrace will use this function
3559 * instead of preempt_schedule() to exit user context if needed before
3560 * calling the scheduler.
3562 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3564 enum ctx_state prev_ctx;
3566 if (likely(!preemptible()))
3571 * Because the function tracer can trace preempt_count_sub()
3572 * and it also uses preempt_enable/disable_notrace(), if
3573 * NEED_RESCHED is set, the preempt_enable_notrace() called
3574 * by the function tracer will call this function again and
3575 * cause infinite recursion.
3577 * Preemption must be disabled here before the function
3578 * tracer can trace. Break up preempt_disable() into two
3579 * calls. One to disable preemption without fear of being
3580 * traced. The other to still record the preemption latency,
3581 * which can also be traced by the function tracer.
3583 preempt_disable_notrace();
3584 preempt_latency_start(1);
3586 * Needs preempt disabled in case user_exit() is traced
3587 * and the tracer calls preempt_enable_notrace() causing
3588 * an infinite recursion.
3590 prev_ctx = exception_enter();
3592 exception_exit(prev_ctx);
3594 preempt_latency_stop(1);
3595 preempt_enable_no_resched_notrace();
3596 } while (need_resched());
3598 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3600 #endif /* CONFIG_PREEMPT */
3603 * this is the entry point to schedule() from kernel preemption
3604 * off of irq context.
3605 * Note, that this is called and return with irqs disabled. This will
3606 * protect us against recursive calling from irq.
3608 asmlinkage __visible void __sched preempt_schedule_irq(void)
3610 enum ctx_state prev_state;
3612 /* Catch callers which need to be fixed */
3613 BUG_ON(preempt_count() || !irqs_disabled());
3615 prev_state = exception_enter();
3621 local_irq_disable();
3622 sched_preempt_enable_no_resched();
3623 } while (need_resched());
3625 exception_exit(prev_state);
3628 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3631 return try_to_wake_up(curr->private, mode, wake_flags);
3633 EXPORT_SYMBOL(default_wake_function);
3635 #ifdef CONFIG_RT_MUTEXES
3638 * rt_mutex_setprio - set the current priority of a task
3640 * @prio: prio value (kernel-internal form)
3642 * This function changes the 'effective' priority of a task. It does
3643 * not touch ->normal_prio like __setscheduler().
3645 * Used by the rt_mutex code to implement priority inheritance
3646 * logic. Call site only calls if the priority of the task changed.
3648 void rt_mutex_setprio(struct task_struct *p, int prio)
3650 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3651 const struct sched_class *prev_class;
3655 BUG_ON(prio > MAX_PRIO);
3657 rq = __task_rq_lock(p, &rf);
3660 * Idle task boosting is a nono in general. There is one
3661 * exception, when PREEMPT_RT and NOHZ is active:
3663 * The idle task calls get_next_timer_interrupt() and holds
3664 * the timer wheel base->lock on the CPU and another CPU wants
3665 * to access the timer (probably to cancel it). We can safely
3666 * ignore the boosting request, as the idle CPU runs this code
3667 * with interrupts disabled and will complete the lock
3668 * protected section without being interrupted. So there is no
3669 * real need to boost.
3671 if (unlikely(p == rq->idle)) {
3672 WARN_ON(p != rq->curr);
3673 WARN_ON(p->pi_blocked_on);
3677 trace_sched_pi_setprio(p, prio);
3680 if (oldprio == prio)
3681 queue_flag &= ~DEQUEUE_MOVE;
3683 prev_class = p->sched_class;
3684 queued = task_on_rq_queued(p);
3685 running = task_current(rq, p);
3687 dequeue_task(rq, p, queue_flag);
3689 put_prev_task(rq, p);
3692 * Boosting condition are:
3693 * 1. -rt task is running and holds mutex A
3694 * --> -dl task blocks on mutex A
3696 * 2. -dl task is running and holds mutex A
3697 * --> -dl task blocks on mutex A and could preempt the
3700 if (dl_prio(prio)) {
3701 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3702 if (!dl_prio(p->normal_prio) ||
3703 (pi_task && dl_prio(pi_task->prio) &&
3704 dl_entity_preempt(&pi_task->dl, &p->dl))) {
3705 p->dl.dl_boosted = 1;
3706 queue_flag |= ENQUEUE_REPLENISH;
3708 p->dl.dl_boosted = 0;
3709 p->sched_class = &dl_sched_class;
3710 } else if (rt_prio(prio)) {
3711 if (dl_prio(oldprio))
3712 p->dl.dl_boosted = 0;
3714 queue_flag |= ENQUEUE_HEAD;
3715 p->sched_class = &rt_sched_class;
3717 if (dl_prio(oldprio))
3718 p->dl.dl_boosted = 0;
3719 if (rt_prio(oldprio))
3721 p->sched_class = &fair_sched_class;
3727 enqueue_task(rq, p, queue_flag);
3729 set_curr_task(rq, p);
3731 check_class_changed(rq, p, prev_class, oldprio);
3733 preempt_disable(); /* avoid rq from going away on us */
3734 __task_rq_unlock(rq, &rf);
3736 balance_callback(rq);
3741 void set_user_nice(struct task_struct *p, long nice)
3743 bool queued, running;
3744 int old_prio, delta;
3748 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3751 * We have to be careful, if called from sys_setpriority(),
3752 * the task might be in the middle of scheduling on another CPU.
3754 rq = task_rq_lock(p, &rf);
3756 * The RT priorities are set via sched_setscheduler(), but we still
3757 * allow the 'normal' nice value to be set - but as expected
3758 * it wont have any effect on scheduling until the task is
3759 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3761 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3762 p->static_prio = NICE_TO_PRIO(nice);
3765 queued = task_on_rq_queued(p);
3766 running = task_current(rq, p);
3768 dequeue_task(rq, p, DEQUEUE_SAVE);
3770 put_prev_task(rq, p);
3772 p->static_prio = NICE_TO_PRIO(nice);
3775 p->prio = effective_prio(p);
3776 delta = p->prio - old_prio;
3779 enqueue_task(rq, p, ENQUEUE_RESTORE);
3781 * If the task increased its priority or is running and
3782 * lowered its priority, then reschedule its CPU:
3784 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3788 set_curr_task(rq, p);
3790 task_rq_unlock(rq, p, &rf);
3792 EXPORT_SYMBOL(set_user_nice);
3795 * can_nice - check if a task can reduce its nice value
3799 int can_nice(const struct task_struct *p, const int nice)
3801 /* convert nice value [19,-20] to rlimit style value [1,40] */
3802 int nice_rlim = nice_to_rlimit(nice);
3804 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3805 capable(CAP_SYS_NICE));
3808 #ifdef __ARCH_WANT_SYS_NICE
3811 * sys_nice - change the priority of the current process.
3812 * @increment: priority increment
3814 * sys_setpriority is a more generic, but much slower function that
3815 * does similar things.
3817 SYSCALL_DEFINE1(nice, int, increment)
3822 * Setpriority might change our priority at the same moment.
3823 * We don't have to worry. Conceptually one call occurs first
3824 * and we have a single winner.
3826 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3827 nice = task_nice(current) + increment;
3829 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3830 if (increment < 0 && !can_nice(current, nice))
3833 retval = security_task_setnice(current, nice);
3837 set_user_nice(current, nice);
3844 * task_prio - return the priority value of a given task.
3845 * @p: the task in question.
3847 * Return: The priority value as seen by users in /proc.
3848 * RT tasks are offset by -200. Normal tasks are centered
3849 * around 0, value goes from -16 to +15.
3851 int task_prio(const struct task_struct *p)
3853 return p->prio - MAX_RT_PRIO;
3857 * idle_cpu - is a given cpu idle currently?
3858 * @cpu: the processor in question.
3860 * Return: 1 if the CPU is currently idle. 0 otherwise.
3862 int idle_cpu(int cpu)
3864 struct rq *rq = cpu_rq(cpu);
3866 if (rq->curr != rq->idle)
3873 if (!llist_empty(&rq->wake_list))
3881 * idle_task - return the idle task for a given cpu.
3882 * @cpu: the processor in question.
3884 * Return: The idle task for the cpu @cpu.
3886 struct task_struct *idle_task(int cpu)
3888 return cpu_rq(cpu)->idle;
3892 * find_process_by_pid - find a process with a matching PID value.
3893 * @pid: the pid in question.
3895 * The task of @pid, if found. %NULL otherwise.
3897 static struct task_struct *find_process_by_pid(pid_t pid)
3899 return pid ? find_task_by_vpid(pid) : current;
3903 * This function initializes the sched_dl_entity of a newly becoming
3904 * SCHED_DEADLINE task.
3906 * Only the static values are considered here, the actual runtime and the
3907 * absolute deadline will be properly calculated when the task is enqueued
3908 * for the first time with its new policy.
3911 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3913 struct sched_dl_entity *dl_se = &p->dl;
3915 dl_se->dl_runtime = attr->sched_runtime;
3916 dl_se->dl_deadline = attr->sched_deadline;
3917 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3918 dl_se->flags = attr->sched_flags;
3919 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3920 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3923 * Changing the parameters of a task is 'tricky' and we're not doing
3924 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3926 * What we SHOULD do is delay the bandwidth release until the 0-lag
3927 * point. This would include retaining the task_struct until that time
3928 * and change dl_overflow() to not immediately decrement the current
3931 * Instead we retain the current runtime/deadline and let the new
3932 * parameters take effect after the current reservation period lapses.
3933 * This is safe (albeit pessimistic) because the 0-lag point is always
3934 * before the current scheduling deadline.
3936 * We can still have temporary overloads because we do not delay the
3937 * change in bandwidth until that time; so admission control is
3938 * not on the safe side. It does however guarantee tasks will never
3939 * consume more than promised.
3944 * sched_setparam() passes in -1 for its policy, to let the functions
3945 * it calls know not to change it.
3947 #define SETPARAM_POLICY -1
3949 static void __setscheduler_params(struct task_struct *p,
3950 const struct sched_attr *attr)
3952 int policy = attr->sched_policy;
3954 if (policy == SETPARAM_POLICY)
3959 if (dl_policy(policy))
3960 __setparam_dl(p, attr);
3961 else if (fair_policy(policy))
3962 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3965 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3966 * !rt_policy. Always setting this ensures that things like
3967 * getparam()/getattr() don't report silly values for !rt tasks.
3969 p->rt_priority = attr->sched_priority;
3970 p->normal_prio = normal_prio(p);
3974 /* Actually do priority change: must hold pi & rq lock. */
3975 static void __setscheduler(struct rq *rq, struct task_struct *p,
3976 const struct sched_attr *attr, bool keep_boost)
3978 __setscheduler_params(p, attr);
3981 * Keep a potential priority boosting if called from
3982 * sched_setscheduler().
3985 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3987 p->prio = normal_prio(p);
3989 if (dl_prio(p->prio))
3990 p->sched_class = &dl_sched_class;
3991 else if (rt_prio(p->prio))
3992 p->sched_class = &rt_sched_class;
3994 p->sched_class = &fair_sched_class;
3998 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4000 struct sched_dl_entity *dl_se = &p->dl;
4002 attr->sched_priority = p->rt_priority;
4003 attr->sched_runtime = dl_se->dl_runtime;
4004 attr->sched_deadline = dl_se->dl_deadline;
4005 attr->sched_period = dl_se->dl_period;
4006 attr->sched_flags = dl_se->flags;
4010 * This function validates the new parameters of a -deadline task.
4011 * We ask for the deadline not being zero, and greater or equal
4012 * than the runtime, as well as the period of being zero or
4013 * greater than deadline. Furthermore, we have to be sure that
4014 * user parameters are above the internal resolution of 1us (we
4015 * check sched_runtime only since it is always the smaller one) and
4016 * below 2^63 ns (we have to check both sched_deadline and
4017 * sched_period, as the latter can be zero).
4020 __checkparam_dl(const struct sched_attr *attr)
4023 if (attr->sched_deadline == 0)
4027 * Since we truncate DL_SCALE bits, make sure we're at least
4030 if (attr->sched_runtime < (1ULL << DL_SCALE))
4034 * Since we use the MSB for wrap-around and sign issues, make
4035 * sure it's not set (mind that period can be equal to zero).
4037 if (attr->sched_deadline & (1ULL << 63) ||
4038 attr->sched_period & (1ULL << 63))
4041 /* runtime <= deadline <= period (if period != 0) */
4042 if ((attr->sched_period != 0 &&
4043 attr->sched_period < attr->sched_deadline) ||
4044 attr->sched_deadline < attr->sched_runtime)
4051 * check the target process has a UID that matches the current process's
4053 static bool check_same_owner(struct task_struct *p)
4055 const struct cred *cred = current_cred(), *pcred;
4059 pcred = __task_cred(p);
4060 match = (uid_eq(cred->euid, pcred->euid) ||
4061 uid_eq(cred->euid, pcred->uid));
4066 static bool dl_param_changed(struct task_struct *p,
4067 const struct sched_attr *attr)
4069 struct sched_dl_entity *dl_se = &p->dl;
4071 if (dl_se->dl_runtime != attr->sched_runtime ||
4072 dl_se->dl_deadline != attr->sched_deadline ||
4073 dl_se->dl_period != attr->sched_period ||
4074 dl_se->flags != attr->sched_flags)
4080 static int __sched_setscheduler(struct task_struct *p,
4081 const struct sched_attr *attr,
4084 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4085 MAX_RT_PRIO - 1 - attr->sched_priority;
4086 int retval, oldprio, oldpolicy = -1, queued, running;
4087 int new_effective_prio, policy = attr->sched_policy;
4088 const struct sched_class *prev_class;
4091 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4094 /* The pi code expects interrupts enabled */
4095 BUG_ON(pi && in_interrupt());
4097 /* double check policy once rq lock held */
4099 reset_on_fork = p->sched_reset_on_fork;
4100 policy = oldpolicy = p->policy;
4102 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4104 if (!valid_policy(policy))
4108 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4112 * Valid priorities for SCHED_FIFO and SCHED_RR are
4113 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4114 * SCHED_BATCH and SCHED_IDLE is 0.
4116 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4117 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4119 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4120 (rt_policy(policy) != (attr->sched_priority != 0)))
4124 * Allow unprivileged RT tasks to decrease priority:
4126 if (user && !capable(CAP_SYS_NICE)) {
4127 if (fair_policy(policy)) {
4128 if (attr->sched_nice < task_nice(p) &&
4129 !can_nice(p, attr->sched_nice))
4133 if (rt_policy(policy)) {
4134 unsigned long rlim_rtprio =
4135 task_rlimit(p, RLIMIT_RTPRIO);
4137 /* can't set/change the rt policy */
4138 if (policy != p->policy && !rlim_rtprio)
4141 /* can't increase priority */
4142 if (attr->sched_priority > p->rt_priority &&
4143 attr->sched_priority > rlim_rtprio)
4148 * Can't set/change SCHED_DEADLINE policy at all for now
4149 * (safest behavior); in the future we would like to allow
4150 * unprivileged DL tasks to increase their relative deadline
4151 * or reduce their runtime (both ways reducing utilization)
4153 if (dl_policy(policy))
4157 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4158 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4160 if (idle_policy(p->policy) && !idle_policy(policy)) {
4161 if (!can_nice(p, task_nice(p)))
4165 /* can't change other user's priorities */
4166 if (!check_same_owner(p))
4169 /* Normal users shall not reset the sched_reset_on_fork flag */
4170 if (p->sched_reset_on_fork && !reset_on_fork)
4175 retval = security_task_setscheduler(p);
4181 * make sure no PI-waiters arrive (or leave) while we are
4182 * changing the priority of the task:
4184 * To be able to change p->policy safely, the appropriate
4185 * runqueue lock must be held.
4187 rq = task_rq_lock(p, &rf);
4190 * Changing the policy of the stop threads its a very bad idea
4192 if (p == rq->stop) {
4193 task_rq_unlock(rq, p, &rf);
4198 * If not changing anything there's no need to proceed further,
4199 * but store a possible modification of reset_on_fork.
4201 if (unlikely(policy == p->policy)) {
4202 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4204 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4206 if (dl_policy(policy) && dl_param_changed(p, attr))
4209 p->sched_reset_on_fork = reset_on_fork;
4210 task_rq_unlock(rq, p, &rf);
4216 #ifdef CONFIG_RT_GROUP_SCHED
4218 * Do not allow realtime tasks into groups that have no runtime
4221 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4222 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4223 !task_group_is_autogroup(task_group(p))) {
4224 task_rq_unlock(rq, p, &rf);
4229 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4230 cpumask_t *span = rq->rd->span;
4233 * Don't allow tasks with an affinity mask smaller than
4234 * the entire root_domain to become SCHED_DEADLINE. We
4235 * will also fail if there's no bandwidth available.
4237 if (!cpumask_subset(span, &p->cpus_allowed) ||
4238 rq->rd->dl_bw.bw == 0) {
4239 task_rq_unlock(rq, p, &rf);
4246 /* recheck policy now with rq lock held */
4247 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4248 policy = oldpolicy = -1;
4249 task_rq_unlock(rq, p, &rf);
4254 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4255 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4258 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4259 task_rq_unlock(rq, p, &rf);
4263 p->sched_reset_on_fork = reset_on_fork;
4268 * Take priority boosted tasks into account. If the new
4269 * effective priority is unchanged, we just store the new
4270 * normal parameters and do not touch the scheduler class and
4271 * the runqueue. This will be done when the task deboost
4274 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4275 if (new_effective_prio == oldprio)
4276 queue_flags &= ~DEQUEUE_MOVE;
4279 queued = task_on_rq_queued(p);
4280 running = task_current(rq, p);
4282 dequeue_task(rq, p, queue_flags);
4284 put_prev_task(rq, p);
4286 prev_class = p->sched_class;
4287 __setscheduler(rq, p, attr, pi);
4291 * We enqueue to tail when the priority of a task is
4292 * increased (user space view).
4294 if (oldprio < p->prio)
4295 queue_flags |= ENQUEUE_HEAD;
4297 enqueue_task(rq, p, queue_flags);
4300 set_curr_task(rq, p);
4302 check_class_changed(rq, p, prev_class, oldprio);
4303 preempt_disable(); /* avoid rq from going away on us */
4304 task_rq_unlock(rq, p, &rf);
4307 rt_mutex_adjust_pi(p);
4310 * Run balance callbacks after we've adjusted the PI chain.
4312 balance_callback(rq);
4318 static int _sched_setscheduler(struct task_struct *p, int policy,
4319 const struct sched_param *param, bool check)
4321 struct sched_attr attr = {
4322 .sched_policy = policy,
4323 .sched_priority = param->sched_priority,
4324 .sched_nice = PRIO_TO_NICE(p->static_prio),
4327 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4328 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4329 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4330 policy &= ~SCHED_RESET_ON_FORK;
4331 attr.sched_policy = policy;
4334 return __sched_setscheduler(p, &attr, check, true);
4337 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4338 * @p: the task in question.
4339 * @policy: new policy.
4340 * @param: structure containing the new RT priority.
4342 * Return: 0 on success. An error code otherwise.
4344 * NOTE that the task may be already dead.
4346 int sched_setscheduler(struct task_struct *p, int policy,
4347 const struct sched_param *param)
4349 return _sched_setscheduler(p, policy, param, true);
4351 EXPORT_SYMBOL_GPL(sched_setscheduler);
4353 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4355 return __sched_setscheduler(p, attr, true, true);
4357 EXPORT_SYMBOL_GPL(sched_setattr);
4360 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4361 * @p: the task in question.
4362 * @policy: new policy.
4363 * @param: structure containing the new RT priority.
4365 * Just like sched_setscheduler, only don't bother checking if the
4366 * current context has permission. For example, this is needed in
4367 * stop_machine(): we create temporary high priority worker threads,
4368 * but our caller might not have that capability.
4370 * Return: 0 on success. An error code otherwise.
4372 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4373 const struct sched_param *param)
4375 return _sched_setscheduler(p, policy, param, false);
4377 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4380 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4382 struct sched_param lparam;
4383 struct task_struct *p;
4386 if (!param || pid < 0)
4388 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4393 p = find_process_by_pid(pid);
4395 retval = sched_setscheduler(p, policy, &lparam);
4402 * Mimics kernel/events/core.c perf_copy_attr().
4404 static int sched_copy_attr(struct sched_attr __user *uattr,
4405 struct sched_attr *attr)
4410 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4414 * zero the full structure, so that a short copy will be nice.
4416 memset(attr, 0, sizeof(*attr));
4418 ret = get_user(size, &uattr->size);
4422 if (size > PAGE_SIZE) /* silly large */
4425 if (!size) /* abi compat */
4426 size = SCHED_ATTR_SIZE_VER0;
4428 if (size < SCHED_ATTR_SIZE_VER0)
4432 * If we're handed a bigger struct than we know of,
4433 * ensure all the unknown bits are 0 - i.e. new
4434 * user-space does not rely on any kernel feature
4435 * extensions we dont know about yet.
4437 if (size > sizeof(*attr)) {
4438 unsigned char __user *addr;
4439 unsigned char __user *end;
4442 addr = (void __user *)uattr + sizeof(*attr);
4443 end = (void __user *)uattr + size;
4445 for (; addr < end; addr++) {
4446 ret = get_user(val, addr);
4452 size = sizeof(*attr);
4455 ret = copy_from_user(attr, uattr, size);
4460 * XXX: do we want to be lenient like existing syscalls; or do we want
4461 * to be strict and return an error on out-of-bounds values?
4463 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4468 put_user(sizeof(*attr), &uattr->size);
4473 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4474 * @pid: the pid in question.
4475 * @policy: new policy.
4476 * @param: structure containing the new RT priority.
4478 * Return: 0 on success. An error code otherwise.
4480 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4481 struct sched_param __user *, param)
4483 /* negative values for policy are not valid */
4487 return do_sched_setscheduler(pid, policy, param);
4491 * sys_sched_setparam - set/change the RT priority of a thread
4492 * @pid: the pid in question.
4493 * @param: structure containing the new RT priority.
4495 * Return: 0 on success. An error code otherwise.
4497 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4499 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4503 * sys_sched_setattr - same as above, but with extended sched_attr
4504 * @pid: the pid in question.
4505 * @uattr: structure containing the extended parameters.
4506 * @flags: for future extension.
4508 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4509 unsigned int, flags)
4511 struct sched_attr attr;
4512 struct task_struct *p;
4515 if (!uattr || pid < 0 || flags)
4518 retval = sched_copy_attr(uattr, &attr);
4522 if ((int)attr.sched_policy < 0)
4527 p = find_process_by_pid(pid);
4529 retval = sched_setattr(p, &attr);
4536 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4537 * @pid: the pid in question.
4539 * Return: On success, the policy of the thread. Otherwise, a negative error
4542 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4544 struct task_struct *p;
4552 p = find_process_by_pid(pid);
4554 retval = security_task_getscheduler(p);
4557 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4564 * sys_sched_getparam - get the RT priority of a thread
4565 * @pid: the pid in question.
4566 * @param: structure containing the RT priority.
4568 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4571 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4573 struct sched_param lp = { .sched_priority = 0 };
4574 struct task_struct *p;
4577 if (!param || pid < 0)
4581 p = find_process_by_pid(pid);
4586 retval = security_task_getscheduler(p);
4590 if (task_has_rt_policy(p))
4591 lp.sched_priority = p->rt_priority;
4595 * This one might sleep, we cannot do it with a spinlock held ...
4597 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4606 static int sched_read_attr(struct sched_attr __user *uattr,
4607 struct sched_attr *attr,
4612 if (!access_ok(VERIFY_WRITE, uattr, usize))
4616 * If we're handed a smaller struct than we know of,
4617 * ensure all the unknown bits are 0 - i.e. old
4618 * user-space does not get uncomplete information.
4620 if (usize < sizeof(*attr)) {
4621 unsigned char *addr;
4624 addr = (void *)attr + usize;
4625 end = (void *)attr + sizeof(*attr);
4627 for (; addr < end; addr++) {
4635 ret = copy_to_user(uattr, attr, attr->size);
4643 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4644 * @pid: the pid in question.
4645 * @uattr: structure containing the extended parameters.
4646 * @size: sizeof(attr) for fwd/bwd comp.
4647 * @flags: for future extension.
4649 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4650 unsigned int, size, unsigned int, flags)
4652 struct sched_attr attr = {
4653 .size = sizeof(struct sched_attr),
4655 struct task_struct *p;
4658 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4659 size < SCHED_ATTR_SIZE_VER0 || flags)
4663 p = find_process_by_pid(pid);
4668 retval = security_task_getscheduler(p);
4672 attr.sched_policy = p->policy;
4673 if (p->sched_reset_on_fork)
4674 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4675 if (task_has_dl_policy(p))
4676 __getparam_dl(p, &attr);
4677 else if (task_has_rt_policy(p))
4678 attr.sched_priority = p->rt_priority;
4680 attr.sched_nice = task_nice(p);
4684 retval = sched_read_attr(uattr, &attr, size);
4692 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4694 cpumask_var_t cpus_allowed, new_mask;
4695 struct task_struct *p;
4700 p = find_process_by_pid(pid);
4706 /* Prevent p going away */
4710 if (p->flags & PF_NO_SETAFFINITY) {
4714 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4718 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4720 goto out_free_cpus_allowed;
4723 if (!check_same_owner(p)) {
4725 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4727 goto out_free_new_mask;
4732 retval = security_task_setscheduler(p);
4734 goto out_free_new_mask;
4737 cpuset_cpus_allowed(p, cpus_allowed);
4738 cpumask_and(new_mask, in_mask, cpus_allowed);
4741 * Since bandwidth control happens on root_domain basis,
4742 * if admission test is enabled, we only admit -deadline
4743 * tasks allowed to run on all the CPUs in the task's
4747 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4749 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4752 goto out_free_new_mask;
4758 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4761 cpuset_cpus_allowed(p, cpus_allowed);
4762 if (!cpumask_subset(new_mask, cpus_allowed)) {
4764 * We must have raced with a concurrent cpuset
4765 * update. Just reset the cpus_allowed to the
4766 * cpuset's cpus_allowed
4768 cpumask_copy(new_mask, cpus_allowed);
4773 free_cpumask_var(new_mask);
4774 out_free_cpus_allowed:
4775 free_cpumask_var(cpus_allowed);
4781 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4782 struct cpumask *new_mask)
4784 if (len < cpumask_size())
4785 cpumask_clear(new_mask);
4786 else if (len > cpumask_size())
4787 len = cpumask_size();
4789 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4793 * sys_sched_setaffinity - set the cpu affinity of a process
4794 * @pid: pid of the process
4795 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4796 * @user_mask_ptr: user-space pointer to the new cpu mask
4798 * Return: 0 on success. An error code otherwise.
4800 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4801 unsigned long __user *, user_mask_ptr)
4803 cpumask_var_t new_mask;
4806 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4809 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4811 retval = sched_setaffinity(pid, new_mask);
4812 free_cpumask_var(new_mask);
4816 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4818 struct task_struct *p;
4819 unsigned long flags;
4825 p = find_process_by_pid(pid);
4829 retval = security_task_getscheduler(p);
4833 raw_spin_lock_irqsave(&p->pi_lock, flags);
4834 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4835 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4844 * sys_sched_getaffinity - get the cpu affinity of a process
4845 * @pid: pid of the process
4846 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4847 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4849 * Return: size of CPU mask copied to user_mask_ptr on success. An
4850 * error code otherwise.
4852 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4853 unsigned long __user *, user_mask_ptr)
4858 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4860 if (len & (sizeof(unsigned long)-1))
4863 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4866 ret = sched_getaffinity(pid, mask);
4868 size_t retlen = min_t(size_t, len, cpumask_size());
4870 if (copy_to_user(user_mask_ptr, mask, retlen))
4875 free_cpumask_var(mask);
4881 * sys_sched_yield - yield the current processor to other threads.
4883 * This function yields the current CPU to other tasks. If there are no
4884 * other threads running on this CPU then this function will return.
4888 SYSCALL_DEFINE0(sched_yield)
4890 struct rq *rq = this_rq_lock();
4892 schedstat_inc(rq->yld_count);
4893 current->sched_class->yield_task(rq);
4896 * Since we are going to call schedule() anyway, there's
4897 * no need to preempt or enable interrupts:
4899 __release(rq->lock);
4900 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4901 do_raw_spin_unlock(&rq->lock);
4902 sched_preempt_enable_no_resched();
4909 #ifndef CONFIG_PREEMPT
4910 int __sched _cond_resched(void)
4912 if (should_resched(0)) {
4913 preempt_schedule_common();
4918 EXPORT_SYMBOL(_cond_resched);
4922 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4923 * call schedule, and on return reacquire the lock.
4925 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4926 * operations here to prevent schedule() from being called twice (once via
4927 * spin_unlock(), once by hand).
4929 int __cond_resched_lock(spinlock_t *lock)
4931 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4934 lockdep_assert_held(lock);
4936 if (spin_needbreak(lock) || resched) {
4939 preempt_schedule_common();
4947 EXPORT_SYMBOL(__cond_resched_lock);
4949 int __sched __cond_resched_softirq(void)
4951 BUG_ON(!in_softirq());
4953 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4955 preempt_schedule_common();
4961 EXPORT_SYMBOL(__cond_resched_softirq);
4964 * yield - yield the current processor to other threads.
4966 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4968 * The scheduler is at all times free to pick the calling task as the most
4969 * eligible task to run, if removing the yield() call from your code breaks
4970 * it, its already broken.
4972 * Typical broken usage is:
4977 * where one assumes that yield() will let 'the other' process run that will
4978 * make event true. If the current task is a SCHED_FIFO task that will never
4979 * happen. Never use yield() as a progress guarantee!!
4981 * If you want to use yield() to wait for something, use wait_event().
4982 * If you want to use yield() to be 'nice' for others, use cond_resched().
4983 * If you still want to use yield(), do not!
4985 void __sched yield(void)
4987 set_current_state(TASK_RUNNING);
4990 EXPORT_SYMBOL(yield);
4993 * yield_to - yield the current processor to another thread in
4994 * your thread group, or accelerate that thread toward the
4995 * processor it's on.
4997 * @preempt: whether task preemption is allowed or not
4999 * It's the caller's job to ensure that the target task struct
5000 * can't go away on us before we can do any checks.
5003 * true (>0) if we indeed boosted the target task.
5004 * false (0) if we failed to boost the target.
5005 * -ESRCH if there's no task to yield to.
5007 int __sched yield_to(struct task_struct *p, bool preempt)
5009 struct task_struct *curr = current;
5010 struct rq *rq, *p_rq;
5011 unsigned long flags;
5014 local_irq_save(flags);
5020 * If we're the only runnable task on the rq and target rq also
5021 * has only one task, there's absolutely no point in yielding.
5023 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5028 double_rq_lock(rq, p_rq);
5029 if (task_rq(p) != p_rq) {
5030 double_rq_unlock(rq, p_rq);
5034 if (!curr->sched_class->yield_to_task)
5037 if (curr->sched_class != p->sched_class)
5040 if (task_running(p_rq, p) || p->state)
5043 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5045 schedstat_inc(rq->yld_count);
5047 * Make p's CPU reschedule; pick_next_entity takes care of
5050 if (preempt && rq != p_rq)
5055 double_rq_unlock(rq, p_rq);
5057 local_irq_restore(flags);
5064 EXPORT_SYMBOL_GPL(yield_to);
5067 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5068 * that process accounting knows that this is a task in IO wait state.
5070 long __sched io_schedule_timeout(long timeout)
5072 int old_iowait = current->in_iowait;
5076 current->in_iowait = 1;
5077 blk_schedule_flush_plug(current);
5079 delayacct_blkio_start();
5081 atomic_inc(&rq->nr_iowait);
5082 ret = schedule_timeout(timeout);
5083 current->in_iowait = old_iowait;
5084 atomic_dec(&rq->nr_iowait);
5085 delayacct_blkio_end();
5089 EXPORT_SYMBOL(io_schedule_timeout);
5092 * sys_sched_get_priority_max - return maximum RT priority.
5093 * @policy: scheduling class.
5095 * Return: On success, this syscall returns the maximum
5096 * rt_priority that can be used by a given scheduling class.
5097 * On failure, a negative error code is returned.
5099 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5106 ret = MAX_USER_RT_PRIO-1;
5108 case SCHED_DEADLINE:
5119 * sys_sched_get_priority_min - return minimum RT priority.
5120 * @policy: scheduling class.
5122 * Return: On success, this syscall returns the minimum
5123 * rt_priority that can be used by a given scheduling class.
5124 * On failure, a negative error code is returned.
5126 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5135 case SCHED_DEADLINE:
5145 * sys_sched_rr_get_interval - return the default timeslice of a process.
5146 * @pid: pid of the process.
5147 * @interval: userspace pointer to the timeslice value.
5149 * this syscall writes the default timeslice value of a given process
5150 * into the user-space timespec buffer. A value of '0' means infinity.
5152 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5155 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5156 struct timespec __user *, interval)
5158 struct task_struct *p;
5159 unsigned int time_slice;
5170 p = find_process_by_pid(pid);
5174 retval = security_task_getscheduler(p);
5178 rq = task_rq_lock(p, &rf);
5180 if (p->sched_class->get_rr_interval)
5181 time_slice = p->sched_class->get_rr_interval(rq, p);
5182 task_rq_unlock(rq, p, &rf);
5185 jiffies_to_timespec(time_slice, &t);
5186 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5194 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5196 void sched_show_task(struct task_struct *p)
5198 unsigned long free = 0;
5200 unsigned long state = p->state;
5202 if (!try_get_task_stack(p))
5205 state = __ffs(state) + 1;
5206 printk(KERN_INFO "%-15.15s %c", p->comm,
5207 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5208 if (state == TASK_RUNNING)
5209 printk(KERN_CONT " running task ");
5210 #ifdef CONFIG_DEBUG_STACK_USAGE
5211 free = stack_not_used(p);
5216 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5218 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5219 task_pid_nr(p), ppid,
5220 (unsigned long)task_thread_info(p)->flags);
5222 print_worker_info(KERN_INFO, p);
5223 show_stack(p, NULL);
5227 void show_state_filter(unsigned long state_filter)
5229 struct task_struct *g, *p;
5231 #if BITS_PER_LONG == 32
5233 " task PC stack pid father\n");
5236 " task PC stack pid father\n");
5239 for_each_process_thread(g, p) {
5241 * reset the NMI-timeout, listing all files on a slow
5242 * console might take a lot of time:
5243 * Also, reset softlockup watchdogs on all CPUs, because
5244 * another CPU might be blocked waiting for us to process
5247 touch_nmi_watchdog();
5248 touch_all_softlockup_watchdogs();
5249 if (!state_filter || (p->state & state_filter))
5253 #ifdef CONFIG_SCHED_DEBUG
5255 sysrq_sched_debug_show();
5259 * Only show locks if all tasks are dumped:
5262 debug_show_all_locks();
5265 void init_idle_bootup_task(struct task_struct *idle)
5267 idle->sched_class = &idle_sched_class;
5271 * init_idle - set up an idle thread for a given CPU
5272 * @idle: task in question
5273 * @cpu: cpu the idle task belongs to
5275 * NOTE: this function does not set the idle thread's NEED_RESCHED
5276 * flag, to make booting more robust.
5278 void init_idle(struct task_struct *idle, int cpu)
5280 struct rq *rq = cpu_rq(cpu);
5281 unsigned long flags;
5283 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5284 raw_spin_lock(&rq->lock);
5286 __sched_fork(0, idle);
5287 idle->state = TASK_RUNNING;
5288 idle->se.exec_start = sched_clock();
5290 kasan_unpoison_task_stack(idle);
5294 * Its possible that init_idle() gets called multiple times on a task,
5295 * in that case do_set_cpus_allowed() will not do the right thing.
5297 * And since this is boot we can forgo the serialization.
5299 set_cpus_allowed_common(idle, cpumask_of(cpu));
5302 * We're having a chicken and egg problem, even though we are
5303 * holding rq->lock, the cpu isn't yet set to this cpu so the
5304 * lockdep check in task_group() will fail.
5306 * Similar case to sched_fork(). / Alternatively we could
5307 * use task_rq_lock() here and obtain the other rq->lock.
5312 __set_task_cpu(idle, cpu);
5315 rq->curr = rq->idle = idle;
5316 idle->on_rq = TASK_ON_RQ_QUEUED;
5320 raw_spin_unlock(&rq->lock);
5321 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5323 /* Set the preempt count _outside_ the spinlocks! */
5324 init_idle_preempt_count(idle, cpu);
5327 * The idle tasks have their own, simple scheduling class:
5329 idle->sched_class = &idle_sched_class;
5330 ftrace_graph_init_idle_task(idle, cpu);
5331 vtime_init_idle(idle, cpu);
5333 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5337 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5338 const struct cpumask *trial)
5340 int ret = 1, trial_cpus;
5341 struct dl_bw *cur_dl_b;
5342 unsigned long flags;
5344 if (!cpumask_weight(cur))
5347 rcu_read_lock_sched();
5348 cur_dl_b = dl_bw_of(cpumask_any(cur));
5349 trial_cpus = cpumask_weight(trial);
5351 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5352 if (cur_dl_b->bw != -1 &&
5353 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5355 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5356 rcu_read_unlock_sched();
5361 int task_can_attach(struct task_struct *p,
5362 const struct cpumask *cs_cpus_allowed)
5367 * Kthreads which disallow setaffinity shouldn't be moved
5368 * to a new cpuset; we don't want to change their cpu
5369 * affinity and isolating such threads by their set of
5370 * allowed nodes is unnecessary. Thus, cpusets are not
5371 * applicable for such threads. This prevents checking for
5372 * success of set_cpus_allowed_ptr() on all attached tasks
5373 * before cpus_allowed may be changed.
5375 if (p->flags & PF_NO_SETAFFINITY) {
5381 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5383 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5388 unsigned long flags;
5390 rcu_read_lock_sched();
5391 dl_b = dl_bw_of(dest_cpu);
5392 raw_spin_lock_irqsave(&dl_b->lock, flags);
5393 cpus = dl_bw_cpus(dest_cpu);
5394 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5399 * We reserve space for this task in the destination
5400 * root_domain, as we can't fail after this point.
5401 * We will free resources in the source root_domain
5402 * later on (see set_cpus_allowed_dl()).
5404 __dl_add(dl_b, p->dl.dl_bw);
5406 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5407 rcu_read_unlock_sched();
5417 static bool sched_smp_initialized __read_mostly;
5419 #ifdef CONFIG_NUMA_BALANCING
5420 /* Migrate current task p to target_cpu */
5421 int migrate_task_to(struct task_struct *p, int target_cpu)
5423 struct migration_arg arg = { p, target_cpu };
5424 int curr_cpu = task_cpu(p);
5426 if (curr_cpu == target_cpu)
5429 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5432 /* TODO: This is not properly updating schedstats */
5434 trace_sched_move_numa(p, curr_cpu, target_cpu);
5435 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5439 * Requeue a task on a given node and accurately track the number of NUMA
5440 * tasks on the runqueues
5442 void sched_setnuma(struct task_struct *p, int nid)
5444 bool queued, running;
5448 rq = task_rq_lock(p, &rf);
5449 queued = task_on_rq_queued(p);
5450 running = task_current(rq, p);
5453 dequeue_task(rq, p, DEQUEUE_SAVE);
5455 put_prev_task(rq, p);
5457 p->numa_preferred_nid = nid;
5460 enqueue_task(rq, p, ENQUEUE_RESTORE);
5462 set_curr_task(rq, p);
5463 task_rq_unlock(rq, p, &rf);
5465 #endif /* CONFIG_NUMA_BALANCING */
5467 #ifdef CONFIG_HOTPLUG_CPU
5469 * Ensures that the idle task is using init_mm right before its cpu goes
5472 void idle_task_exit(void)
5474 struct mm_struct *mm = current->active_mm;
5476 BUG_ON(cpu_online(smp_processor_id()));
5478 if (mm != &init_mm) {
5479 switch_mm(mm, &init_mm, current);
5480 finish_arch_post_lock_switch();
5486 * Since this CPU is going 'away' for a while, fold any nr_active delta
5487 * we might have. Assumes we're called after migrate_tasks() so that the
5488 * nr_active count is stable. We need to take the teardown thread which
5489 * is calling this into account, so we hand in adjust = 1 to the load
5492 * Also see the comment "Global load-average calculations".
5494 static void calc_load_migrate(struct rq *rq)
5496 long delta = calc_load_fold_active(rq, 1);
5498 atomic_long_add(delta, &calc_load_tasks);
5501 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5505 static const struct sched_class fake_sched_class = {
5506 .put_prev_task = put_prev_task_fake,
5509 static struct task_struct fake_task = {
5511 * Avoid pull_{rt,dl}_task()
5513 .prio = MAX_PRIO + 1,
5514 .sched_class = &fake_sched_class,
5518 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5519 * try_to_wake_up()->select_task_rq().
5521 * Called with rq->lock held even though we'er in stop_machine() and
5522 * there's no concurrency possible, we hold the required locks anyway
5523 * because of lock validation efforts.
5525 static void migrate_tasks(struct rq *dead_rq)
5527 struct rq *rq = dead_rq;
5528 struct task_struct *next, *stop = rq->stop;
5529 struct pin_cookie cookie;
5533 * Fudge the rq selection such that the below task selection loop
5534 * doesn't get stuck on the currently eligible stop task.
5536 * We're currently inside stop_machine() and the rq is either stuck
5537 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5538 * either way we should never end up calling schedule() until we're
5544 * put_prev_task() and pick_next_task() sched
5545 * class method both need to have an up-to-date
5546 * value of rq->clock[_task]
5548 update_rq_clock(rq);
5552 * There's this thread running, bail when that's the only
5555 if (rq->nr_running == 1)
5559 * pick_next_task assumes pinned rq->lock.
5561 cookie = lockdep_pin_lock(&rq->lock);
5562 next = pick_next_task(rq, &fake_task, cookie);
5564 next->sched_class->put_prev_task(rq, next);
5567 * Rules for changing task_struct::cpus_allowed are holding
5568 * both pi_lock and rq->lock, such that holding either
5569 * stabilizes the mask.
5571 * Drop rq->lock is not quite as disastrous as it usually is
5572 * because !cpu_active at this point, which means load-balance
5573 * will not interfere. Also, stop-machine.
5575 lockdep_unpin_lock(&rq->lock, cookie);
5576 raw_spin_unlock(&rq->lock);
5577 raw_spin_lock(&next->pi_lock);
5578 raw_spin_lock(&rq->lock);
5581 * Since we're inside stop-machine, _nothing_ should have
5582 * changed the task, WARN if weird stuff happened, because in
5583 * that case the above rq->lock drop is a fail too.
5585 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5586 raw_spin_unlock(&next->pi_lock);
5590 /* Find suitable destination for @next, with force if needed. */
5591 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5593 rq = __migrate_task(rq, next, dest_cpu);
5594 if (rq != dead_rq) {
5595 raw_spin_unlock(&rq->lock);
5597 raw_spin_lock(&rq->lock);
5599 raw_spin_unlock(&next->pi_lock);
5604 #endif /* CONFIG_HOTPLUG_CPU */
5606 static void set_rq_online(struct rq *rq)
5609 const struct sched_class *class;
5611 cpumask_set_cpu(rq->cpu, rq->rd->online);
5614 for_each_class(class) {
5615 if (class->rq_online)
5616 class->rq_online(rq);
5621 static void set_rq_offline(struct rq *rq)
5624 const struct sched_class *class;
5626 for_each_class(class) {
5627 if (class->rq_offline)
5628 class->rq_offline(rq);
5631 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5636 static void set_cpu_rq_start_time(unsigned int cpu)
5638 struct rq *rq = cpu_rq(cpu);
5640 rq->age_stamp = sched_clock_cpu(cpu);
5643 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5645 #ifdef CONFIG_SCHED_DEBUG
5647 static __read_mostly int sched_debug_enabled;
5649 static int __init sched_debug_setup(char *str)
5651 sched_debug_enabled = 1;
5655 early_param("sched_debug", sched_debug_setup);
5657 static inline bool sched_debug(void)
5659 return sched_debug_enabled;
5662 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5663 struct cpumask *groupmask)
5665 struct sched_group *group = sd->groups;
5667 cpumask_clear(groupmask);
5669 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5671 if (!(sd->flags & SD_LOAD_BALANCE)) {
5672 printk("does not load-balance\n");
5674 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5679 printk(KERN_CONT "span %*pbl level %s\n",
5680 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5682 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5683 printk(KERN_ERR "ERROR: domain->span does not contain "
5686 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5687 printk(KERN_ERR "ERROR: domain->groups does not contain"
5691 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5695 printk(KERN_ERR "ERROR: group is NULL\n");
5699 if (!cpumask_weight(sched_group_cpus(group))) {
5700 printk(KERN_CONT "\n");
5701 printk(KERN_ERR "ERROR: empty group\n");
5705 if (!(sd->flags & SD_OVERLAP) &&
5706 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5707 printk(KERN_CONT "\n");
5708 printk(KERN_ERR "ERROR: repeated CPUs\n");
5712 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5714 printk(KERN_CONT " %*pbl",
5715 cpumask_pr_args(sched_group_cpus(group)));
5716 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5717 printk(KERN_CONT " (cpu_capacity = %d)",
5718 group->sgc->capacity);
5721 group = group->next;
5722 } while (group != sd->groups);
5723 printk(KERN_CONT "\n");
5725 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5726 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5729 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5730 printk(KERN_ERR "ERROR: parent span is not a superset "
5731 "of domain->span\n");
5735 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5739 if (!sched_debug_enabled)
5743 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5747 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5750 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5758 #else /* !CONFIG_SCHED_DEBUG */
5760 # define sched_debug_enabled 0
5761 # define sched_domain_debug(sd, cpu) do { } while (0)
5762 static inline bool sched_debug(void)
5766 #endif /* CONFIG_SCHED_DEBUG */
5768 static int sd_degenerate(struct sched_domain *sd)
5770 if (cpumask_weight(sched_domain_span(sd)) == 1)
5773 /* Following flags need at least 2 groups */
5774 if (sd->flags & (SD_LOAD_BALANCE |
5775 SD_BALANCE_NEWIDLE |
5778 SD_SHARE_CPUCAPACITY |
5779 SD_ASYM_CPUCAPACITY |
5780 SD_SHARE_PKG_RESOURCES |
5781 SD_SHARE_POWERDOMAIN)) {
5782 if (sd->groups != sd->groups->next)
5786 /* Following flags don't use groups */
5787 if (sd->flags & (SD_WAKE_AFFINE))
5794 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5796 unsigned long cflags = sd->flags, pflags = parent->flags;
5798 if (sd_degenerate(parent))
5801 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5804 /* Flags needing groups don't count if only 1 group in parent */
5805 if (parent->groups == parent->groups->next) {
5806 pflags &= ~(SD_LOAD_BALANCE |
5807 SD_BALANCE_NEWIDLE |
5810 SD_ASYM_CPUCAPACITY |
5811 SD_SHARE_CPUCAPACITY |
5812 SD_SHARE_PKG_RESOURCES |
5814 SD_SHARE_POWERDOMAIN);
5815 if (nr_node_ids == 1)
5816 pflags &= ~SD_SERIALIZE;
5818 if (~cflags & pflags)
5824 static void free_rootdomain(struct rcu_head *rcu)
5826 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5828 cpupri_cleanup(&rd->cpupri);
5829 cpudl_cleanup(&rd->cpudl);
5830 free_cpumask_var(rd->dlo_mask);
5831 free_cpumask_var(rd->rto_mask);
5832 free_cpumask_var(rd->online);
5833 free_cpumask_var(rd->span);
5837 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5839 struct root_domain *old_rd = NULL;
5840 unsigned long flags;
5842 raw_spin_lock_irqsave(&rq->lock, flags);
5847 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5850 cpumask_clear_cpu(rq->cpu, old_rd->span);
5853 * If we dont want to free the old_rd yet then
5854 * set old_rd to NULL to skip the freeing later
5857 if (!atomic_dec_and_test(&old_rd->refcount))
5861 atomic_inc(&rd->refcount);
5864 cpumask_set_cpu(rq->cpu, rd->span);
5865 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5868 raw_spin_unlock_irqrestore(&rq->lock, flags);
5871 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5874 void sched_get_rd(struct root_domain *rd)
5876 atomic_inc(&rd->refcount);
5879 void sched_put_rd(struct root_domain *rd)
5881 if (!atomic_dec_and_test(&rd->refcount))
5884 call_rcu_sched(&rd->rcu, free_rootdomain);
5887 static int init_rootdomain(struct root_domain *rd)
5889 memset(rd, 0, sizeof(*rd));
5891 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5893 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5895 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5897 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5900 #ifdef HAVE_RT_PUSH_IPI
5902 raw_spin_lock_init(&rd->rto_lock);
5903 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
5906 init_dl_bw(&rd->dl_bw);
5907 if (cpudl_init(&rd->cpudl) != 0)
5910 if (cpupri_init(&rd->cpupri) != 0)
5915 free_cpumask_var(rd->rto_mask);
5917 free_cpumask_var(rd->dlo_mask);
5919 free_cpumask_var(rd->online);
5921 free_cpumask_var(rd->span);
5927 * By default the system creates a single root-domain with all cpus as
5928 * members (mimicking the global state we have today).
5930 struct root_domain def_root_domain;
5932 static void init_defrootdomain(void)
5934 init_rootdomain(&def_root_domain);
5936 atomic_set(&def_root_domain.refcount, 1);
5939 static struct root_domain *alloc_rootdomain(void)
5941 struct root_domain *rd;
5943 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5947 if (init_rootdomain(rd) != 0) {
5955 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5957 struct sched_group *tmp, *first;
5966 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5971 } while (sg != first);
5974 static void destroy_sched_domain(struct sched_domain *sd)
5977 * If its an overlapping domain it has private groups, iterate and
5980 if (sd->flags & SD_OVERLAP) {
5981 free_sched_groups(sd->groups, 1);
5982 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5983 kfree(sd->groups->sgc);
5986 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
5991 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
5993 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5996 struct sched_domain *parent = sd->parent;
5997 destroy_sched_domain(sd);
6002 static void destroy_sched_domains(struct sched_domain *sd)
6005 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
6009 * Keep a special pointer to the highest sched_domain that has
6010 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6011 * allows us to avoid some pointer chasing select_idle_sibling().
6013 * Also keep a unique ID per domain (we use the first cpu number in
6014 * the cpumask of the domain), this allows us to quickly tell if
6015 * two cpus are in the same cache domain, see cpus_share_cache().
6017 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6018 DEFINE_PER_CPU(int, sd_llc_size);
6019 DEFINE_PER_CPU(int, sd_llc_id);
6020 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
6021 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6022 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6024 static void update_top_cache_domain(int cpu)
6026 struct sched_domain_shared *sds = NULL;
6027 struct sched_domain *sd;
6031 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6033 id = cpumask_first(sched_domain_span(sd));
6034 size = cpumask_weight(sched_domain_span(sd));
6038 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6039 per_cpu(sd_llc_size, cpu) = size;
6040 per_cpu(sd_llc_id, cpu) = id;
6041 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
6043 sd = lowest_flag_domain(cpu, SD_NUMA);
6044 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6046 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6047 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6051 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6052 * hold the hotplug lock.
6055 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6057 struct rq *rq = cpu_rq(cpu);
6058 struct sched_domain *tmp;
6060 /* Remove the sched domains which do not contribute to scheduling. */
6061 for (tmp = sd; tmp; ) {
6062 struct sched_domain *parent = tmp->parent;
6066 if (sd_parent_degenerate(tmp, parent)) {
6067 tmp->parent = parent->parent;
6069 parent->parent->child = tmp;
6071 * Transfer SD_PREFER_SIBLING down in case of a
6072 * degenerate parent; the spans match for this
6073 * so the property transfers.
6075 if (parent->flags & SD_PREFER_SIBLING)
6076 tmp->flags |= SD_PREFER_SIBLING;
6077 destroy_sched_domain(parent);
6082 if (sd && sd_degenerate(sd)) {
6085 destroy_sched_domain(tmp);
6090 sched_domain_debug(sd, cpu);
6092 rq_attach_root(rq, rd);
6094 rcu_assign_pointer(rq->sd, sd);
6095 destroy_sched_domains(tmp);
6097 update_top_cache_domain(cpu);
6100 /* Setup the mask of cpus configured for isolated domains */
6101 static int __init isolated_cpu_setup(char *str)
6105 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6106 ret = cpulist_parse(str, cpu_isolated_map);
6108 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6113 __setup("isolcpus=", isolated_cpu_setup);
6116 struct sched_domain ** __percpu sd;
6117 struct root_domain *rd;
6128 * Build an iteration mask that can exclude certain CPUs from the upwards
6131 * Only CPUs that can arrive at this group should be considered to continue
6134 * Asymmetric node setups can result in situations where the domain tree is of
6135 * unequal depth, make sure to skip domains that already cover the entire
6138 * In that case build_sched_domains() will have terminated the iteration early
6139 * and our sibling sd spans will be empty. Domains should always include the
6140 * cpu they're built on, so check that.
6143 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6145 const struct cpumask *sg_span = sched_group_cpus(sg);
6146 struct sd_data *sdd = sd->private;
6147 struct sched_domain *sibling;
6150 for_each_cpu(i, sg_span) {
6151 sibling = *per_cpu_ptr(sdd->sd, i);
6154 * Can happen in the asymmetric case, where these siblings are
6155 * unused. The mask will not be empty because those CPUs that
6156 * do have the top domain _should_ span the domain.
6158 if (!sibling->child)
6161 /* If we would not end up here, we can't continue from here */
6162 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6165 cpumask_set_cpu(i, sched_group_mask(sg));
6168 /* We must not have empty masks here */
6169 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6173 * Return the canonical balance cpu for this group, this is the first cpu
6174 * of this group that's also in the iteration mask.
6176 int group_balance_cpu(struct sched_group *sg)
6178 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6182 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6184 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6185 const struct cpumask *span = sched_domain_span(sd);
6186 struct cpumask *covered = sched_domains_tmpmask;
6187 struct sd_data *sdd = sd->private;
6188 struct sched_domain *sibling;
6191 cpumask_clear(covered);
6193 for_each_cpu_wrap(i, span, cpu) {
6194 struct cpumask *sg_span;
6196 if (cpumask_test_cpu(i, covered))
6199 sibling = *per_cpu_ptr(sdd->sd, i);
6201 /* See the comment near build_group_mask(). */
6202 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6205 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6206 GFP_KERNEL, cpu_to_node(cpu));
6211 sg_span = sched_group_cpus(sg);
6213 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6215 cpumask_set_cpu(i, sg_span);
6217 cpumask_or(covered, covered, sg_span);
6219 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6220 if (atomic_inc_return(&sg->sgc->ref) == 1)
6221 build_group_mask(sd, sg);
6224 * Initialize sgc->capacity such that even if we mess up the
6225 * domains and no possible iteration will get us here, we won't
6228 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6231 * Make sure the first group of this domain contains the
6232 * canonical balance cpu. Otherwise the sched_domain iteration
6233 * breaks. See update_sg_lb_stats().
6235 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6236 group_balance_cpu(sg) == cpu)
6246 sd->groups = groups;
6251 free_sched_groups(first, 0);
6256 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6258 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6259 struct sched_domain *child = sd->child;
6262 cpu = cpumask_first(sched_domain_span(child));
6265 *sg = *per_cpu_ptr(sdd->sg, cpu);
6266 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6267 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6274 * build_sched_groups will build a circular linked list of the groups
6275 * covered by the given span, and will set each group's ->cpumask correctly,
6276 * and ->cpu_capacity to 0.
6278 * Assumes the sched_domain tree is fully constructed
6281 build_sched_groups(struct sched_domain *sd, int cpu)
6283 struct sched_group *first = NULL, *last = NULL;
6284 struct sd_data *sdd = sd->private;
6285 const struct cpumask *span = sched_domain_span(sd);
6286 struct cpumask *covered;
6289 get_group(cpu, sdd, &sd->groups);
6290 atomic_inc(&sd->groups->ref);
6292 if (cpu != cpumask_first(span))
6295 lockdep_assert_held(&sched_domains_mutex);
6296 covered = sched_domains_tmpmask;
6298 cpumask_clear(covered);
6300 for_each_cpu(i, span) {
6301 struct sched_group *sg;
6304 if (cpumask_test_cpu(i, covered))
6307 group = get_group(i, sdd, &sg);
6308 cpumask_setall(sched_group_mask(sg));
6310 for_each_cpu(j, span) {
6311 if (get_group(j, sdd, NULL) != group)
6314 cpumask_set_cpu(j, covered);
6315 cpumask_set_cpu(j, sched_group_cpus(sg));
6330 * Initialize sched groups cpu_capacity.
6332 * cpu_capacity indicates the capacity of sched group, which is used while
6333 * distributing the load between different sched groups in a sched domain.
6334 * Typically cpu_capacity for all the groups in a sched domain will be same
6335 * unless there are asymmetries in the topology. If there are asymmetries,
6336 * group having more cpu_capacity will pickup more load compared to the
6337 * group having less cpu_capacity.
6339 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6341 struct sched_group *sg = sd->groups;
6346 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6348 } while (sg != sd->groups);
6350 if (cpu != group_balance_cpu(sg))
6353 update_group_capacity(sd, cpu);
6357 * Initializers for schedule domains
6358 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6361 static int default_relax_domain_level = -1;
6362 int sched_domain_level_max;
6364 static int __init setup_relax_domain_level(char *str)
6366 if (kstrtoint(str, 0, &default_relax_domain_level))
6367 pr_warn("Unable to set relax_domain_level\n");
6371 __setup("relax_domain_level=", setup_relax_domain_level);
6373 static void set_domain_attribute(struct sched_domain *sd,
6374 struct sched_domain_attr *attr)
6378 if (!attr || attr->relax_domain_level < 0) {
6379 if (default_relax_domain_level < 0)
6382 request = default_relax_domain_level;
6384 request = attr->relax_domain_level;
6385 if (request < sd->level) {
6386 /* turn off idle balance on this domain */
6387 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6389 /* turn on idle balance on this domain */
6390 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6394 static void __sdt_free(const struct cpumask *cpu_map);
6395 static int __sdt_alloc(const struct cpumask *cpu_map);
6397 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6398 const struct cpumask *cpu_map)
6402 if (!atomic_read(&d->rd->refcount))
6403 free_rootdomain(&d->rd->rcu); /* fall through */
6405 free_percpu(d->sd); /* fall through */
6407 __sdt_free(cpu_map); /* fall through */
6413 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6414 const struct cpumask *cpu_map)
6416 memset(d, 0, sizeof(*d));
6418 if (__sdt_alloc(cpu_map))
6419 return sa_sd_storage;
6420 d->sd = alloc_percpu(struct sched_domain *);
6422 return sa_sd_storage;
6423 d->rd = alloc_rootdomain();
6426 return sa_rootdomain;
6430 * NULL the sd_data elements we've used to build the sched_domain and
6431 * sched_group structure so that the subsequent __free_domain_allocs()
6432 * will not free the data we're using.
6434 static void claim_allocations(int cpu, struct sched_domain *sd)
6436 struct sd_data *sdd = sd->private;
6438 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6439 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6441 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
6442 *per_cpu_ptr(sdd->sds, cpu) = NULL;
6444 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6445 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6447 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6448 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6452 static int sched_domains_numa_levels;
6453 enum numa_topology_type sched_numa_topology_type;
6454 static int *sched_domains_numa_distance;
6455 int sched_max_numa_distance;
6456 static struct cpumask ***sched_domains_numa_masks;
6457 static int sched_domains_curr_level;
6461 * SD_flags allowed in topology descriptions.
6463 * These flags are purely descriptive of the topology and do not prescribe
6464 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6467 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6468 * SD_SHARE_PKG_RESOURCES - describes shared caches
6469 * SD_NUMA - describes NUMA topologies
6470 * SD_SHARE_POWERDOMAIN - describes shared power domain
6471 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6473 * Odd one out, which beside describing the topology has a quirk also
6474 * prescribes the desired behaviour that goes along with it:
6476 * SD_ASYM_PACKING - describes SMT quirks
6478 #define TOPOLOGY_SD_FLAGS \
6479 (SD_SHARE_CPUCAPACITY | \
6480 SD_SHARE_PKG_RESOURCES | \
6483 SD_ASYM_CPUCAPACITY | \
6484 SD_SHARE_POWERDOMAIN)
6486 static struct sched_domain *
6487 sd_init(struct sched_domain_topology_level *tl,
6488 const struct cpumask *cpu_map,
6489 struct sched_domain *child, int cpu)
6491 struct sd_data *sdd = &tl->data;
6492 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6493 int sd_id, sd_weight, sd_flags = 0;
6497 * Ugly hack to pass state to sd_numa_mask()...
6499 sched_domains_curr_level = tl->numa_level;
6502 sd_weight = cpumask_weight(tl->mask(cpu));
6505 sd_flags = (*tl->sd_flags)();
6506 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6507 "wrong sd_flags in topology description\n"))
6508 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6510 *sd = (struct sched_domain){
6511 .min_interval = sd_weight,
6512 .max_interval = 2*sd_weight,
6514 .imbalance_pct = 125,
6516 .cache_nice_tries = 0,
6523 .flags = 1*SD_LOAD_BALANCE
6524 | 1*SD_BALANCE_NEWIDLE
6529 | 0*SD_SHARE_CPUCAPACITY
6530 | 0*SD_SHARE_PKG_RESOURCES
6532 | 0*SD_PREFER_SIBLING
6537 .last_balance = jiffies,
6538 .balance_interval = sd_weight,
6540 .max_newidle_lb_cost = 0,
6541 .next_decay_max_lb_cost = jiffies,
6543 #ifdef CONFIG_SCHED_DEBUG
6548 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6549 sd_id = cpumask_first(sched_domain_span(sd));
6552 * Convert topological properties into behaviour.
6555 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6556 struct sched_domain *t = sd;
6558 for_each_lower_domain(t)
6559 t->flags |= SD_BALANCE_WAKE;
6562 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6563 sd->flags |= SD_PREFER_SIBLING;
6564 sd->imbalance_pct = 110;
6565 sd->smt_gain = 1178; /* ~15% */
6567 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6568 sd->imbalance_pct = 117;
6569 sd->cache_nice_tries = 1;
6573 } else if (sd->flags & SD_NUMA) {
6574 sd->cache_nice_tries = 2;
6578 sd->flags |= SD_SERIALIZE;
6579 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6580 sd->flags &= ~(SD_BALANCE_EXEC |
6587 sd->flags |= SD_PREFER_SIBLING;
6588 sd->cache_nice_tries = 1;
6594 * For all levels sharing cache; connect a sched_domain_shared
6597 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6598 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
6599 atomic_inc(&sd->shared->ref);
6600 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
6609 * Topology list, bottom-up.
6611 static struct sched_domain_topology_level default_topology[] = {
6612 #ifdef CONFIG_SCHED_SMT
6613 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6615 #ifdef CONFIG_SCHED_MC
6616 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6618 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6622 static struct sched_domain_topology_level *sched_domain_topology =
6625 #define for_each_sd_topology(tl) \
6626 for (tl = sched_domain_topology; tl->mask; tl++)
6628 void set_sched_topology(struct sched_domain_topology_level *tl)
6630 if (WARN_ON_ONCE(sched_smp_initialized))
6633 sched_domain_topology = tl;
6638 static const struct cpumask *sd_numa_mask(int cpu)
6640 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6643 static void sched_numa_warn(const char *str)
6645 static int done = false;
6653 printk(KERN_WARNING "ERROR: %s\n\n", str);
6655 for (i = 0; i < nr_node_ids; i++) {
6656 printk(KERN_WARNING " ");
6657 for (j = 0; j < nr_node_ids; j++)
6658 printk(KERN_CONT "%02d ", node_distance(i,j));
6659 printk(KERN_CONT "\n");
6661 printk(KERN_WARNING "\n");
6664 bool find_numa_distance(int distance)
6668 if (distance == node_distance(0, 0))
6671 for (i = 0; i < sched_domains_numa_levels; i++) {
6672 if (sched_domains_numa_distance[i] == distance)
6680 * A system can have three types of NUMA topology:
6681 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6682 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6683 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6685 * The difference between a glueless mesh topology and a backplane
6686 * topology lies in whether communication between not directly
6687 * connected nodes goes through intermediary nodes (where programs
6688 * could run), or through backplane controllers. This affects
6689 * placement of programs.
6691 * The type of topology can be discerned with the following tests:
6692 * - If the maximum distance between any nodes is 1 hop, the system
6693 * is directly connected.
6694 * - If for two nodes A and B, located N > 1 hops away from each other,
6695 * there is an intermediary node C, which is < N hops away from both
6696 * nodes A and B, the system is a glueless mesh.
6698 static void init_numa_topology_type(void)
6702 n = sched_max_numa_distance;
6704 if (sched_domains_numa_levels <= 1) {
6705 sched_numa_topology_type = NUMA_DIRECT;
6709 for_each_online_node(a) {
6710 for_each_online_node(b) {
6711 /* Find two nodes furthest removed from each other. */
6712 if (node_distance(a, b) < n)
6715 /* Is there an intermediary node between a and b? */
6716 for_each_online_node(c) {
6717 if (node_distance(a, c) < n &&
6718 node_distance(b, c) < n) {
6719 sched_numa_topology_type =
6725 sched_numa_topology_type = NUMA_BACKPLANE;
6731 static void sched_init_numa(void)
6733 int next_distance, curr_distance = node_distance(0, 0);
6734 struct sched_domain_topology_level *tl;
6738 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6739 if (!sched_domains_numa_distance)
6743 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6744 * unique distances in the node_distance() table.
6746 * Assumes node_distance(0,j) includes all distances in
6747 * node_distance(i,j) in order to avoid cubic time.
6749 next_distance = curr_distance;
6750 for (i = 0; i < nr_node_ids; i++) {
6751 for (j = 0; j < nr_node_ids; j++) {
6752 for (k = 0; k < nr_node_ids; k++) {
6753 int distance = node_distance(i, k);
6755 if (distance > curr_distance &&
6756 (distance < next_distance ||
6757 next_distance == curr_distance))
6758 next_distance = distance;
6761 * While not a strong assumption it would be nice to know
6762 * about cases where if node A is connected to B, B is not
6763 * equally connected to A.
6765 if (sched_debug() && node_distance(k, i) != distance)
6766 sched_numa_warn("Node-distance not symmetric");
6768 if (sched_debug() && i && !find_numa_distance(distance))
6769 sched_numa_warn("Node-0 not representative");
6771 if (next_distance != curr_distance) {
6772 sched_domains_numa_distance[level++] = next_distance;
6773 sched_domains_numa_levels = level;
6774 curr_distance = next_distance;
6779 * In case of sched_debug() we verify the above assumption.
6789 * 'level' contains the number of unique distances, excluding the
6790 * identity distance node_distance(i,i).
6792 * The sched_domains_numa_distance[] array includes the actual distance
6797 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6798 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6799 * the array will contain less then 'level' members. This could be
6800 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6801 * in other functions.
6803 * We reset it to 'level' at the end of this function.
6805 sched_domains_numa_levels = 0;
6807 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6808 if (!sched_domains_numa_masks)
6812 * Now for each level, construct a mask per node which contains all
6813 * cpus of nodes that are that many hops away from us.
6815 for (i = 0; i < level; i++) {
6816 sched_domains_numa_masks[i] =
6817 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6818 if (!sched_domains_numa_masks[i])
6821 for (j = 0; j < nr_node_ids; j++) {
6822 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6826 sched_domains_numa_masks[i][j] = mask;
6829 if (node_distance(j, k) > sched_domains_numa_distance[i])
6832 cpumask_or(mask, mask, cpumask_of_node(k));
6837 /* Compute default topology size */
6838 for (i = 0; sched_domain_topology[i].mask; i++);
6840 tl = kzalloc((i + level + 1) *
6841 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6846 * Copy the default topology bits..
6848 for (i = 0; sched_domain_topology[i].mask; i++)
6849 tl[i] = sched_domain_topology[i];
6852 * .. and append 'j' levels of NUMA goodness.
6854 for (j = 0; j < level; i++, j++) {
6855 tl[i] = (struct sched_domain_topology_level){
6856 .mask = sd_numa_mask,
6857 .sd_flags = cpu_numa_flags,
6858 .flags = SDTL_OVERLAP,
6864 sched_domain_topology = tl;
6866 sched_domains_numa_levels = level;
6867 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6869 init_numa_topology_type();
6872 static void sched_domains_numa_masks_set(unsigned int cpu)
6874 int node = cpu_to_node(cpu);
6877 for (i = 0; i < sched_domains_numa_levels; i++) {
6878 for (j = 0; j < nr_node_ids; j++) {
6879 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6880 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6885 static void sched_domains_numa_masks_clear(unsigned int cpu)
6889 for (i = 0; i < sched_domains_numa_levels; i++) {
6890 for (j = 0; j < nr_node_ids; j++)
6891 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6896 static inline void sched_init_numa(void) { }
6897 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6898 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6899 #endif /* CONFIG_NUMA */
6901 static int __sdt_alloc(const struct cpumask *cpu_map)
6903 struct sched_domain_topology_level *tl;
6906 for_each_sd_topology(tl) {
6907 struct sd_data *sdd = &tl->data;
6909 sdd->sd = alloc_percpu(struct sched_domain *);
6913 sdd->sds = alloc_percpu(struct sched_domain_shared *);
6917 sdd->sg = alloc_percpu(struct sched_group *);
6921 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6925 for_each_cpu(j, cpu_map) {
6926 struct sched_domain *sd;
6927 struct sched_domain_shared *sds;
6928 struct sched_group *sg;
6929 struct sched_group_capacity *sgc;
6931 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6932 GFP_KERNEL, cpu_to_node(j));
6936 *per_cpu_ptr(sdd->sd, j) = sd;
6938 sds = kzalloc_node(sizeof(struct sched_domain_shared),
6939 GFP_KERNEL, cpu_to_node(j));
6943 *per_cpu_ptr(sdd->sds, j) = sds;
6945 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6946 GFP_KERNEL, cpu_to_node(j));
6952 *per_cpu_ptr(sdd->sg, j) = sg;
6954 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6955 GFP_KERNEL, cpu_to_node(j));
6959 *per_cpu_ptr(sdd->sgc, j) = sgc;
6966 static void __sdt_free(const struct cpumask *cpu_map)
6968 struct sched_domain_topology_level *tl;
6971 for_each_sd_topology(tl) {
6972 struct sd_data *sdd = &tl->data;
6974 for_each_cpu(j, cpu_map) {
6975 struct sched_domain *sd;
6978 sd = *per_cpu_ptr(sdd->sd, j);
6979 if (sd && (sd->flags & SD_OVERLAP))
6980 free_sched_groups(sd->groups, 0);
6981 kfree(*per_cpu_ptr(sdd->sd, j));
6985 kfree(*per_cpu_ptr(sdd->sds, j));
6987 kfree(*per_cpu_ptr(sdd->sg, j));
6989 kfree(*per_cpu_ptr(sdd->sgc, j));
6991 free_percpu(sdd->sd);
6993 free_percpu(sdd->sds);
6995 free_percpu(sdd->sg);
6997 free_percpu(sdd->sgc);
7002 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7003 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7004 struct sched_domain *child, int cpu)
7006 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
7009 sd->level = child->level + 1;
7010 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7013 if (!cpumask_subset(sched_domain_span(child),
7014 sched_domain_span(sd))) {
7015 pr_err("BUG: arch topology borken\n");
7016 #ifdef CONFIG_SCHED_DEBUG
7017 pr_err(" the %s domain not a subset of the %s domain\n",
7018 child->name, sd->name);
7020 /* Fixup, ensure @sd has at least @child cpus. */
7021 cpumask_or(sched_domain_span(sd),
7022 sched_domain_span(sd),
7023 sched_domain_span(child));
7027 set_domain_attribute(sd, attr);
7033 * Build sched domains for a given set of cpus and attach the sched domains
7034 * to the individual cpus
7036 static int build_sched_domains(const struct cpumask *cpu_map,
7037 struct sched_domain_attr *attr)
7039 enum s_alloc alloc_state;
7040 struct sched_domain *sd;
7042 struct rq *rq = NULL;
7043 int i, ret = -ENOMEM;
7045 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7046 if (alloc_state != sa_rootdomain)
7049 /* Set up domains for cpus specified by the cpu_map. */
7050 for_each_cpu(i, cpu_map) {
7051 struct sched_domain_topology_level *tl;
7054 for_each_sd_topology(tl) {
7055 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7056 if (tl == sched_domain_topology)
7057 *per_cpu_ptr(d.sd, i) = sd;
7058 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7059 sd->flags |= SD_OVERLAP;
7060 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7065 /* Build the groups for the domains */
7066 for_each_cpu(i, cpu_map) {
7067 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7068 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7069 if (sd->flags & SD_OVERLAP) {
7070 if (build_overlap_sched_groups(sd, i))
7073 if (build_sched_groups(sd, i))
7079 /* Calculate CPU capacity for physical packages and nodes */
7080 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7081 if (!cpumask_test_cpu(i, cpu_map))
7084 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7085 claim_allocations(i, sd);
7086 init_sched_groups_capacity(i, sd);
7090 /* Attach the domains */
7092 for_each_cpu(i, cpu_map) {
7094 sd = *per_cpu_ptr(d.sd, i);
7096 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7097 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
7098 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
7100 cpu_attach_domain(sd, d.rd, i);
7104 if (rq && sched_debug_enabled) {
7105 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7106 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
7111 __free_domain_allocs(&d, alloc_state, cpu_map);
7115 static cpumask_var_t *doms_cur; /* current sched domains */
7116 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7117 static struct sched_domain_attr *dattr_cur;
7118 /* attribues of custom domains in 'doms_cur' */
7121 * Special case: If a kmalloc of a doms_cur partition (array of
7122 * cpumask) fails, then fallback to a single sched domain,
7123 * as determined by the single cpumask fallback_doms.
7125 static cpumask_var_t fallback_doms;
7128 * arch_update_cpu_topology lets virtualized architectures update the
7129 * cpu core maps. It is supposed to return 1 if the topology changed
7130 * or 0 if it stayed the same.
7132 int __weak arch_update_cpu_topology(void)
7137 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7140 cpumask_var_t *doms;
7142 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7145 for (i = 0; i < ndoms; i++) {
7146 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7147 free_sched_domains(doms, i);
7154 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7157 for (i = 0; i < ndoms; i++)
7158 free_cpumask_var(doms[i]);
7163 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7164 * For now this just excludes isolated cpus, but could be used to
7165 * exclude other special cases in the future.
7167 static int init_sched_domains(const struct cpumask *cpu_map)
7171 arch_update_cpu_topology();
7173 doms_cur = alloc_sched_domains(ndoms_cur);
7175 doms_cur = &fallback_doms;
7176 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7177 err = build_sched_domains(doms_cur[0], NULL);
7178 register_sched_domain_sysctl();
7184 * Detach sched domains from a group of cpus specified in cpu_map
7185 * These cpus will now be attached to the NULL domain
7187 static void detach_destroy_domains(const struct cpumask *cpu_map)
7192 for_each_cpu(i, cpu_map)
7193 cpu_attach_domain(NULL, &def_root_domain, i);
7197 /* handle null as "default" */
7198 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7199 struct sched_domain_attr *new, int idx_new)
7201 struct sched_domain_attr tmp;
7208 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7209 new ? (new + idx_new) : &tmp,
7210 sizeof(struct sched_domain_attr));
7214 * Partition sched domains as specified by the 'ndoms_new'
7215 * cpumasks in the array doms_new[] of cpumasks. This compares
7216 * doms_new[] to the current sched domain partitioning, doms_cur[].
7217 * It destroys each deleted domain and builds each new domain.
7219 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7220 * The masks don't intersect (don't overlap.) We should setup one
7221 * sched domain for each mask. CPUs not in any of the cpumasks will
7222 * not be load balanced. If the same cpumask appears both in the
7223 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7226 * The passed in 'doms_new' should be allocated using
7227 * alloc_sched_domains. This routine takes ownership of it and will
7228 * free_sched_domains it when done with it. If the caller failed the
7229 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7230 * and partition_sched_domains() will fallback to the single partition
7231 * 'fallback_doms', it also forces the domains to be rebuilt.
7233 * If doms_new == NULL it will be replaced with cpu_online_mask.
7234 * ndoms_new == 0 is a special case for destroying existing domains,
7235 * and it will not create the default domain.
7237 * Call with hotplug lock held
7239 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7240 struct sched_domain_attr *dattr_new)
7245 mutex_lock(&sched_domains_mutex);
7247 /* always unregister in case we don't destroy any domains */
7248 unregister_sched_domain_sysctl();
7250 /* Let architecture update cpu core mappings. */
7251 new_topology = arch_update_cpu_topology();
7253 n = doms_new ? ndoms_new : 0;
7255 /* Destroy deleted domains */
7256 for (i = 0; i < ndoms_cur; i++) {
7257 for (j = 0; j < n && !new_topology; j++) {
7258 if (cpumask_equal(doms_cur[i], doms_new[j])
7259 && dattrs_equal(dattr_cur, i, dattr_new, j))
7262 /* no match - a current sched domain not in new doms_new[] */
7263 detach_destroy_domains(doms_cur[i]);
7269 if (doms_new == NULL) {
7271 doms_new = &fallback_doms;
7272 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7273 WARN_ON_ONCE(dattr_new);
7276 /* Build new domains */
7277 for (i = 0; i < ndoms_new; i++) {
7278 for (j = 0; j < n && !new_topology; j++) {
7279 if (cpumask_equal(doms_new[i], doms_cur[j])
7280 && dattrs_equal(dattr_new, i, dattr_cur, j))
7283 /* no match - add a new doms_new */
7284 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7289 /* Remember the new sched domains */
7290 if (doms_cur != &fallback_doms)
7291 free_sched_domains(doms_cur, ndoms_cur);
7292 kfree(dattr_cur); /* kfree(NULL) is safe */
7293 doms_cur = doms_new;
7294 dattr_cur = dattr_new;
7295 ndoms_cur = ndoms_new;
7297 register_sched_domain_sysctl();
7299 mutex_unlock(&sched_domains_mutex);
7302 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7305 * Update cpusets according to cpu_active mask. If cpusets are
7306 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7307 * around partition_sched_domains().
7309 * If we come here as part of a suspend/resume, don't touch cpusets because we
7310 * want to restore it back to its original state upon resume anyway.
7312 static void cpuset_cpu_active(void)
7314 if (cpuhp_tasks_frozen) {
7316 * num_cpus_frozen tracks how many CPUs are involved in suspend
7317 * resume sequence. As long as this is not the last online
7318 * operation in the resume sequence, just build a single sched
7319 * domain, ignoring cpusets.
7321 partition_sched_domains(1, NULL, NULL);
7322 if (--num_cpus_frozen)
7325 * This is the last CPU online operation. So fall through and
7326 * restore the original sched domains by considering the
7327 * cpuset configurations.
7329 cpuset_force_rebuild();
7331 cpuset_update_active_cpus(true);
7334 static int cpuset_cpu_inactive(unsigned int cpu)
7336 unsigned long flags;
7341 if (!cpuhp_tasks_frozen) {
7342 rcu_read_lock_sched();
7343 dl_b = dl_bw_of(cpu);
7345 raw_spin_lock_irqsave(&dl_b->lock, flags);
7346 cpus = dl_bw_cpus(cpu);
7347 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7348 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7350 rcu_read_unlock_sched();
7354 cpuset_update_active_cpus(false);
7357 partition_sched_domains(1, NULL, NULL);
7362 #ifdef CONFIG_SCHED_SMT
7363 atomic_t sched_smt_present = ATOMIC_INIT(0);
7366 int sched_cpu_activate(unsigned int cpu)
7368 struct rq *rq = cpu_rq(cpu);
7369 unsigned long flags;
7371 #ifdef CONFIG_SCHED_SMT
7373 * When going up, increment the number of cores with SMT present.
7375 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7376 atomic_inc(&sched_smt_present);
7378 set_cpu_active(cpu, true);
7380 if (sched_smp_initialized) {
7381 sched_domains_numa_masks_set(cpu);
7382 cpuset_cpu_active();
7386 * Put the rq online, if not already. This happens:
7388 * 1) In the early boot process, because we build the real domains
7389 * after all cpus have been brought up.
7391 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7394 raw_spin_lock_irqsave(&rq->lock, flags);
7396 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7399 raw_spin_unlock_irqrestore(&rq->lock, flags);
7401 update_max_interval();
7406 int sched_cpu_deactivate(unsigned int cpu)
7410 set_cpu_active(cpu, false);
7412 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7413 * users of this state to go away such that all new such users will
7416 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7417 * not imply sync_sched(), so wait for both.
7419 * Do sync before park smpboot threads to take care the rcu boost case.
7421 if (IS_ENABLED(CONFIG_PREEMPT))
7422 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7426 #ifdef CONFIG_SCHED_SMT
7428 * When going down, decrement the number of cores with SMT present.
7430 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
7431 atomic_dec(&sched_smt_present);
7434 if (!sched_smp_initialized)
7437 ret = cpuset_cpu_inactive(cpu);
7439 set_cpu_active(cpu, true);
7442 sched_domains_numa_masks_clear(cpu);
7446 static void sched_rq_cpu_starting(unsigned int cpu)
7448 struct rq *rq = cpu_rq(cpu);
7450 rq->calc_load_update = calc_load_update;
7451 update_max_interval();
7454 int sched_cpu_starting(unsigned int cpu)
7456 set_cpu_rq_start_time(cpu);
7457 sched_rq_cpu_starting(cpu);
7461 #ifdef CONFIG_HOTPLUG_CPU
7462 int sched_cpu_dying(unsigned int cpu)
7464 struct rq *rq = cpu_rq(cpu);
7465 unsigned long flags;
7467 /* Handle pending wakeups and then migrate everything off */
7468 sched_ttwu_pending();
7469 raw_spin_lock_irqsave(&rq->lock, flags);
7471 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7475 BUG_ON(rq->nr_running != 1);
7476 raw_spin_unlock_irqrestore(&rq->lock, flags);
7477 calc_load_migrate(rq);
7478 update_max_interval();
7479 nohz_balance_exit_idle(cpu);
7485 void __init sched_init_smp(void)
7487 cpumask_var_t non_isolated_cpus;
7489 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7490 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7495 * There's no userspace yet to cause hotplug operations; hence all the
7496 * cpu masks are stable and all blatant races in the below code cannot
7499 mutex_lock(&sched_domains_mutex);
7500 init_sched_domains(cpu_active_mask);
7501 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7502 if (cpumask_empty(non_isolated_cpus))
7503 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7504 mutex_unlock(&sched_domains_mutex);
7506 /* Move init over to a non-isolated CPU */
7507 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7509 sched_init_granularity();
7510 free_cpumask_var(non_isolated_cpus);
7512 init_sched_rt_class();
7513 init_sched_dl_class();
7514 sched_smp_initialized = true;
7517 static int __init migration_init(void)
7519 sched_rq_cpu_starting(smp_processor_id());
7522 early_initcall(migration_init);
7525 void __init sched_init_smp(void)
7527 sched_init_granularity();
7529 #endif /* CONFIG_SMP */
7531 int in_sched_functions(unsigned long addr)
7533 return in_lock_functions(addr) ||
7534 (addr >= (unsigned long)__sched_text_start
7535 && addr < (unsigned long)__sched_text_end);
7538 #ifdef CONFIG_CGROUP_SCHED
7540 * Default task group.
7541 * Every task in system belongs to this group at bootup.
7543 struct task_group root_task_group;
7544 LIST_HEAD(task_groups);
7546 /* Cacheline aligned slab cache for task_group */
7547 static struct kmem_cache *task_group_cache __read_mostly;
7550 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7551 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7553 #define WAIT_TABLE_BITS 8
7554 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7555 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
7557 wait_queue_head_t *bit_waitqueue(void *word, int bit)
7559 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
7560 unsigned long val = (unsigned long)word << shift | bit;
7562 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
7564 EXPORT_SYMBOL(bit_waitqueue);
7566 void __init sched_init(void)
7569 unsigned long alloc_size = 0, ptr;
7571 for (i = 0; i < WAIT_TABLE_SIZE; i++)
7572 init_waitqueue_head(bit_wait_table + i);
7574 #ifdef CONFIG_FAIR_GROUP_SCHED
7575 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7577 #ifdef CONFIG_RT_GROUP_SCHED
7578 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7581 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7583 #ifdef CONFIG_FAIR_GROUP_SCHED
7584 root_task_group.se = (struct sched_entity **)ptr;
7585 ptr += nr_cpu_ids * sizeof(void **);
7587 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7588 ptr += nr_cpu_ids * sizeof(void **);
7590 #endif /* CONFIG_FAIR_GROUP_SCHED */
7591 #ifdef CONFIG_RT_GROUP_SCHED
7592 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7593 ptr += nr_cpu_ids * sizeof(void **);
7595 root_task_group.rt_rq = (struct rt_rq **)ptr;
7596 ptr += nr_cpu_ids * sizeof(void **);
7598 #endif /* CONFIG_RT_GROUP_SCHED */
7600 #ifdef CONFIG_CPUMASK_OFFSTACK
7601 for_each_possible_cpu(i) {
7602 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7603 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7604 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7605 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7607 #endif /* CONFIG_CPUMASK_OFFSTACK */
7609 init_rt_bandwidth(&def_rt_bandwidth,
7610 global_rt_period(), global_rt_runtime());
7611 init_dl_bandwidth(&def_dl_bandwidth,
7612 global_rt_period(), global_rt_runtime());
7615 init_defrootdomain();
7618 #ifdef CONFIG_RT_GROUP_SCHED
7619 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7620 global_rt_period(), global_rt_runtime());
7621 #endif /* CONFIG_RT_GROUP_SCHED */
7623 #ifdef CONFIG_CGROUP_SCHED
7624 task_group_cache = KMEM_CACHE(task_group, 0);
7626 list_add(&root_task_group.list, &task_groups);
7627 INIT_LIST_HEAD(&root_task_group.children);
7628 INIT_LIST_HEAD(&root_task_group.siblings);
7629 autogroup_init(&init_task);
7630 #endif /* CONFIG_CGROUP_SCHED */
7632 for_each_possible_cpu(i) {
7636 raw_spin_lock_init(&rq->lock);
7638 rq->calc_load_active = 0;
7639 rq->calc_load_update = jiffies + LOAD_FREQ;
7640 init_cfs_rq(&rq->cfs);
7641 init_rt_rq(&rq->rt);
7642 init_dl_rq(&rq->dl);
7643 #ifdef CONFIG_FAIR_GROUP_SCHED
7644 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7645 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7647 * How much cpu bandwidth does root_task_group get?
7649 * In case of task-groups formed thr' the cgroup filesystem, it
7650 * gets 100% of the cpu resources in the system. This overall
7651 * system cpu resource is divided among the tasks of
7652 * root_task_group and its child task-groups in a fair manner,
7653 * based on each entity's (task or task-group's) weight
7654 * (se->load.weight).
7656 * In other words, if root_task_group has 10 tasks of weight
7657 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7658 * then A0's share of the cpu resource is:
7660 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7662 * We achieve this by letting root_task_group's tasks sit
7663 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7665 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7666 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7667 #endif /* CONFIG_FAIR_GROUP_SCHED */
7669 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7670 #ifdef CONFIG_RT_GROUP_SCHED
7671 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7674 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7675 rq->cpu_load[j] = 0;
7680 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7681 rq->balance_callback = NULL;
7682 rq->active_balance = 0;
7683 rq->next_balance = jiffies;
7688 rq->avg_idle = 2*sysctl_sched_migration_cost;
7689 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7691 INIT_LIST_HEAD(&rq->cfs_tasks);
7693 rq_attach_root(rq, &def_root_domain);
7694 #ifdef CONFIG_NO_HZ_COMMON
7695 rq->last_load_update_tick = jiffies;
7698 #ifdef CONFIG_NO_HZ_FULL
7699 rq->last_sched_tick = 0;
7701 #endif /* CONFIG_SMP */
7703 atomic_set(&rq->nr_iowait, 0);
7706 set_load_weight(&init_task);
7709 * The boot idle thread does lazy MMU switching as well:
7711 atomic_inc(&init_mm.mm_count);
7712 enter_lazy_tlb(&init_mm, current);
7715 * Make us the idle thread. Technically, schedule() should not be
7716 * called from this thread, however somewhere below it might be,
7717 * but because we are the idle thread, we just pick up running again
7718 * when this runqueue becomes "idle".
7720 init_idle(current, smp_processor_id());
7722 calc_load_update = jiffies + LOAD_FREQ;
7725 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7726 /* May be allocated at isolcpus cmdline parse time */
7727 if (cpu_isolated_map == NULL)
7728 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7729 idle_thread_set_boot_cpu();
7730 set_cpu_rq_start_time(smp_processor_id());
7732 init_sched_fair_class();
7736 scheduler_running = 1;
7739 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7740 static inline int preempt_count_equals(int preempt_offset)
7742 int nested = preempt_count() + rcu_preempt_depth();
7744 return (nested == preempt_offset);
7747 void __might_sleep(const char *file, int line, int preempt_offset)
7750 * Blocking primitives will set (and therefore destroy) current->state,
7751 * since we will exit with TASK_RUNNING make sure we enter with it,
7752 * otherwise we will destroy state.
7754 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7755 "do not call blocking ops when !TASK_RUNNING; "
7756 "state=%lx set at [<%p>] %pS\n",
7758 (void *)current->task_state_change,
7759 (void *)current->task_state_change);
7761 ___might_sleep(file, line, preempt_offset);
7763 EXPORT_SYMBOL(__might_sleep);
7765 void ___might_sleep(const char *file, int line, int preempt_offset)
7767 static unsigned long prev_jiffy; /* ratelimiting */
7768 unsigned long preempt_disable_ip;
7770 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7771 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7772 !is_idle_task(current)) ||
7773 system_state != SYSTEM_RUNNING || oops_in_progress)
7775 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7777 prev_jiffy = jiffies;
7779 /* Save this before calling printk(), since that will clobber it */
7780 preempt_disable_ip = get_preempt_disable_ip(current);
7783 "BUG: sleeping function called from invalid context at %s:%d\n",
7786 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7787 in_atomic(), irqs_disabled(),
7788 current->pid, current->comm);
7790 if (task_stack_end_corrupted(current))
7791 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7793 debug_show_held_locks(current);
7794 if (irqs_disabled())
7795 print_irqtrace_events(current);
7796 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7797 && !preempt_count_equals(preempt_offset)) {
7798 pr_err("Preemption disabled at:");
7799 print_ip_sym(preempt_disable_ip);
7803 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7805 EXPORT_SYMBOL(___might_sleep);
7808 #ifdef CONFIG_MAGIC_SYSRQ
7809 void normalize_rt_tasks(void)
7811 struct task_struct *g, *p;
7812 struct sched_attr attr = {
7813 .sched_policy = SCHED_NORMAL,
7816 read_lock(&tasklist_lock);
7817 for_each_process_thread(g, p) {
7819 * Only normalize user tasks:
7821 if (p->flags & PF_KTHREAD)
7824 p->se.exec_start = 0;
7825 schedstat_set(p->se.statistics.wait_start, 0);
7826 schedstat_set(p->se.statistics.sleep_start, 0);
7827 schedstat_set(p->se.statistics.block_start, 0);
7829 if (!dl_task(p) && !rt_task(p)) {
7831 * Renice negative nice level userspace
7834 if (task_nice(p) < 0)
7835 set_user_nice(p, 0);
7839 __sched_setscheduler(p, &attr, false, false);
7841 read_unlock(&tasklist_lock);
7844 #endif /* CONFIG_MAGIC_SYSRQ */
7846 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7848 * These functions are only useful for the IA64 MCA handling, or kdb.
7850 * They can only be called when the whole system has been
7851 * stopped - every CPU needs to be quiescent, and no scheduling
7852 * activity can take place. Using them for anything else would
7853 * be a serious bug, and as a result, they aren't even visible
7854 * under any other configuration.
7858 * curr_task - return the current task for a given cpu.
7859 * @cpu: the processor in question.
7861 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7863 * Return: The current task for @cpu.
7865 struct task_struct *curr_task(int cpu)
7867 return cpu_curr(cpu);
7870 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7874 * set_curr_task - set the current task for a given cpu.
7875 * @cpu: the processor in question.
7876 * @p: the task pointer to set.
7878 * Description: This function must only be used when non-maskable interrupts
7879 * are serviced on a separate stack. It allows the architecture to switch the
7880 * notion of the current task on a cpu in a non-blocking manner. This function
7881 * must be called with all CPU's synchronized, and interrupts disabled, the
7882 * and caller must save the original value of the current task (see
7883 * curr_task() above) and restore that value before reenabling interrupts and
7884 * re-starting the system.
7886 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7888 void ia64_set_curr_task(int cpu, struct task_struct *p)
7895 #ifdef CONFIG_CGROUP_SCHED
7896 /* task_group_lock serializes the addition/removal of task groups */
7897 static DEFINE_SPINLOCK(task_group_lock);
7899 static void sched_free_group(struct task_group *tg)
7901 free_fair_sched_group(tg);
7902 free_rt_sched_group(tg);
7904 kmem_cache_free(task_group_cache, tg);
7907 /* allocate runqueue etc for a new task group */
7908 struct task_group *sched_create_group(struct task_group *parent)
7910 struct task_group *tg;
7912 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7914 return ERR_PTR(-ENOMEM);
7916 if (!alloc_fair_sched_group(tg, parent))
7919 if (!alloc_rt_sched_group(tg, parent))
7925 sched_free_group(tg);
7926 return ERR_PTR(-ENOMEM);
7929 void sched_online_group(struct task_group *tg, struct task_group *parent)
7931 unsigned long flags;
7933 spin_lock_irqsave(&task_group_lock, flags);
7934 list_add_rcu(&tg->list, &task_groups);
7936 WARN_ON(!parent); /* root should already exist */
7938 tg->parent = parent;
7939 INIT_LIST_HEAD(&tg->children);
7940 list_add_rcu(&tg->siblings, &parent->children);
7941 spin_unlock_irqrestore(&task_group_lock, flags);
7943 online_fair_sched_group(tg);
7946 /* rcu callback to free various structures associated with a task group */
7947 static void sched_free_group_rcu(struct rcu_head *rhp)
7949 /* now it should be safe to free those cfs_rqs */
7950 sched_free_group(container_of(rhp, struct task_group, rcu));
7953 void sched_destroy_group(struct task_group *tg)
7955 /* wait for possible concurrent references to cfs_rqs complete */
7956 call_rcu(&tg->rcu, sched_free_group_rcu);
7959 void sched_offline_group(struct task_group *tg)
7961 unsigned long flags;
7963 /* end participation in shares distribution */
7964 unregister_fair_sched_group(tg);
7966 spin_lock_irqsave(&task_group_lock, flags);
7967 list_del_rcu(&tg->list);
7968 list_del_rcu(&tg->siblings);
7969 spin_unlock_irqrestore(&task_group_lock, flags);
7972 static void sched_change_group(struct task_struct *tsk, int type)
7974 struct task_group *tg;
7977 * All callers are synchronized by task_rq_lock(); we do not use RCU
7978 * which is pointless here. Thus, we pass "true" to task_css_check()
7979 * to prevent lockdep warnings.
7981 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7982 struct task_group, css);
7983 tg = autogroup_task_group(tsk, tg);
7984 tsk->sched_task_group = tg;
7986 #ifdef CONFIG_FAIR_GROUP_SCHED
7987 if (tsk->sched_class->task_change_group)
7988 tsk->sched_class->task_change_group(tsk, type);
7991 set_task_rq(tsk, task_cpu(tsk));
7995 * Change task's runqueue when it moves between groups.
7997 * The caller of this function should have put the task in its new group by
7998 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8001 void sched_move_task(struct task_struct *tsk)
8003 int queued, running;
8007 rq = task_rq_lock(tsk, &rf);
8008 update_rq_clock(rq);
8010 running = task_current(rq, tsk);
8011 queued = task_on_rq_queued(tsk);
8014 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
8015 if (unlikely(running))
8016 put_prev_task(rq, tsk);
8018 sched_change_group(tsk, TASK_MOVE_GROUP);
8021 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
8022 if (unlikely(running))
8023 set_curr_task(rq, tsk);
8025 task_rq_unlock(rq, tsk, &rf);
8027 #endif /* CONFIG_CGROUP_SCHED */
8029 #ifdef CONFIG_RT_GROUP_SCHED
8031 * Ensure that the real time constraints are schedulable.
8033 static DEFINE_MUTEX(rt_constraints_mutex);
8035 /* Must be called with tasklist_lock held */
8036 static inline int tg_has_rt_tasks(struct task_group *tg)
8038 struct task_struct *g, *p;
8041 * Autogroups do not have RT tasks; see autogroup_create().
8043 if (task_group_is_autogroup(tg))
8046 for_each_process_thread(g, p) {
8047 if (rt_task(p) && task_group(p) == tg)
8054 struct rt_schedulable_data {
8055 struct task_group *tg;
8060 static int tg_rt_schedulable(struct task_group *tg, void *data)
8062 struct rt_schedulable_data *d = data;
8063 struct task_group *child;
8064 unsigned long total, sum = 0;
8065 u64 period, runtime;
8067 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8068 runtime = tg->rt_bandwidth.rt_runtime;
8071 period = d->rt_period;
8072 runtime = d->rt_runtime;
8076 * Cannot have more runtime than the period.
8078 if (runtime > period && runtime != RUNTIME_INF)
8082 * Ensure we don't starve existing RT tasks.
8084 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8087 total = to_ratio(period, runtime);
8090 * Nobody can have more than the global setting allows.
8092 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8096 * The sum of our children's runtime should not exceed our own.
8098 list_for_each_entry_rcu(child, &tg->children, siblings) {
8099 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8100 runtime = child->rt_bandwidth.rt_runtime;
8102 if (child == d->tg) {
8103 period = d->rt_period;
8104 runtime = d->rt_runtime;
8107 sum += to_ratio(period, runtime);
8116 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8120 struct rt_schedulable_data data = {
8122 .rt_period = period,
8123 .rt_runtime = runtime,
8127 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8133 static int tg_set_rt_bandwidth(struct task_group *tg,
8134 u64 rt_period, u64 rt_runtime)
8139 * Disallowing the root group RT runtime is BAD, it would disallow the
8140 * kernel creating (and or operating) RT threads.
8142 if (tg == &root_task_group && rt_runtime == 0)
8145 /* No period doesn't make any sense. */
8149 mutex_lock(&rt_constraints_mutex);
8150 read_lock(&tasklist_lock);
8151 err = __rt_schedulable(tg, rt_period, rt_runtime);
8155 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8156 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8157 tg->rt_bandwidth.rt_runtime = rt_runtime;
8159 for_each_possible_cpu(i) {
8160 struct rt_rq *rt_rq = tg->rt_rq[i];
8162 raw_spin_lock(&rt_rq->rt_runtime_lock);
8163 rt_rq->rt_runtime = rt_runtime;
8164 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8166 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8168 read_unlock(&tasklist_lock);
8169 mutex_unlock(&rt_constraints_mutex);
8174 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8176 u64 rt_runtime, rt_period;
8178 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8179 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8180 if (rt_runtime_us < 0)
8181 rt_runtime = RUNTIME_INF;
8183 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8186 static long sched_group_rt_runtime(struct task_group *tg)
8190 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8193 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8194 do_div(rt_runtime_us, NSEC_PER_USEC);
8195 return rt_runtime_us;
8198 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8200 u64 rt_runtime, rt_period;
8202 rt_period = rt_period_us * NSEC_PER_USEC;
8203 rt_runtime = tg->rt_bandwidth.rt_runtime;
8205 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8208 static long sched_group_rt_period(struct task_group *tg)
8212 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8213 do_div(rt_period_us, NSEC_PER_USEC);
8214 return rt_period_us;
8216 #endif /* CONFIG_RT_GROUP_SCHED */
8218 #ifdef CONFIG_RT_GROUP_SCHED
8219 static int sched_rt_global_constraints(void)
8223 mutex_lock(&rt_constraints_mutex);
8224 read_lock(&tasklist_lock);
8225 ret = __rt_schedulable(NULL, 0, 0);
8226 read_unlock(&tasklist_lock);
8227 mutex_unlock(&rt_constraints_mutex);
8232 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8234 /* Don't accept realtime tasks when there is no way for them to run */
8235 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8241 #else /* !CONFIG_RT_GROUP_SCHED */
8242 static int sched_rt_global_constraints(void)
8244 unsigned long flags;
8247 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8248 for_each_possible_cpu(i) {
8249 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8251 raw_spin_lock(&rt_rq->rt_runtime_lock);
8252 rt_rq->rt_runtime = global_rt_runtime();
8253 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8255 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8259 #endif /* CONFIG_RT_GROUP_SCHED */
8261 static int sched_dl_global_validate(void)
8263 u64 runtime = global_rt_runtime();
8264 u64 period = global_rt_period();
8265 u64 new_bw = to_ratio(period, runtime);
8268 unsigned long flags;
8271 * Here we want to check the bandwidth not being set to some
8272 * value smaller than the currently allocated bandwidth in
8273 * any of the root_domains.
8275 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8276 * cycling on root_domains... Discussion on different/better
8277 * solutions is welcome!
8279 for_each_possible_cpu(cpu) {
8280 rcu_read_lock_sched();
8281 dl_b = dl_bw_of(cpu);
8283 raw_spin_lock_irqsave(&dl_b->lock, flags);
8284 if (new_bw < dl_b->total_bw)
8286 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8288 rcu_read_unlock_sched();
8297 static void sched_dl_do_global(void)
8302 unsigned long flags;
8304 def_dl_bandwidth.dl_period = global_rt_period();
8305 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8307 if (global_rt_runtime() != RUNTIME_INF)
8308 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8311 * FIXME: As above...
8313 for_each_possible_cpu(cpu) {
8314 rcu_read_lock_sched();
8315 dl_b = dl_bw_of(cpu);
8317 raw_spin_lock_irqsave(&dl_b->lock, flags);
8319 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8321 rcu_read_unlock_sched();
8325 static int sched_rt_global_validate(void)
8327 if (sysctl_sched_rt_period <= 0)
8330 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8331 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8337 static void sched_rt_do_global(void)
8339 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8340 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8343 int sched_rt_handler(struct ctl_table *table, int write,
8344 void __user *buffer, size_t *lenp,
8347 int old_period, old_runtime;
8348 static DEFINE_MUTEX(mutex);
8352 old_period = sysctl_sched_rt_period;
8353 old_runtime = sysctl_sched_rt_runtime;
8355 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8357 if (!ret && write) {
8358 ret = sched_rt_global_validate();
8362 ret = sched_dl_global_validate();
8366 ret = sched_rt_global_constraints();
8370 sched_rt_do_global();
8371 sched_dl_do_global();
8375 sysctl_sched_rt_period = old_period;
8376 sysctl_sched_rt_runtime = old_runtime;
8378 mutex_unlock(&mutex);
8383 int sched_rr_handler(struct ctl_table *table, int write,
8384 void __user *buffer, size_t *lenp,
8388 static DEFINE_MUTEX(mutex);
8391 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8392 /* make sure that internally we keep jiffies */
8393 /* also, writing zero resets timeslice to default */
8394 if (!ret && write) {
8395 sched_rr_timeslice =
8396 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
8397 msecs_to_jiffies(sysctl_sched_rr_timeslice);
8399 mutex_unlock(&mutex);
8403 #ifdef CONFIG_CGROUP_SCHED
8405 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8407 return css ? container_of(css, struct task_group, css) : NULL;
8410 static struct cgroup_subsys_state *
8411 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8413 struct task_group *parent = css_tg(parent_css);
8414 struct task_group *tg;
8417 /* This is early initialization for the top cgroup */
8418 return &root_task_group.css;
8421 tg = sched_create_group(parent);
8423 return ERR_PTR(-ENOMEM);
8428 /* Expose task group only after completing cgroup initialization */
8429 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8431 struct task_group *tg = css_tg(css);
8432 struct task_group *parent = css_tg(css->parent);
8435 sched_online_group(tg, parent);
8439 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8441 struct task_group *tg = css_tg(css);
8443 sched_offline_group(tg);
8446 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8448 struct task_group *tg = css_tg(css);
8451 * Relies on the RCU grace period between css_released() and this.
8453 sched_free_group(tg);
8457 * This is called before wake_up_new_task(), therefore we really only
8458 * have to set its group bits, all the other stuff does not apply.
8460 static void cpu_cgroup_fork(struct task_struct *task)
8465 rq = task_rq_lock(task, &rf);
8467 sched_change_group(task, TASK_SET_GROUP);
8469 task_rq_unlock(rq, task, &rf);
8472 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8474 struct task_struct *task;
8475 struct cgroup_subsys_state *css;
8478 cgroup_taskset_for_each(task, css, tset) {
8479 #ifdef CONFIG_RT_GROUP_SCHED
8480 if (!sched_rt_can_attach(css_tg(css), task))
8484 * Serialize against wake_up_new_task() such that if its
8485 * running, we're sure to observe its full state.
8487 raw_spin_lock_irq(&task->pi_lock);
8489 * Avoid calling sched_move_task() before wake_up_new_task()
8490 * has happened. This would lead to problems with PELT, due to
8491 * move wanting to detach+attach while we're not attached yet.
8493 if (task->state == TASK_NEW)
8495 raw_spin_unlock_irq(&task->pi_lock);
8503 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8505 struct task_struct *task;
8506 struct cgroup_subsys_state *css;
8508 cgroup_taskset_for_each(task, css, tset)
8509 sched_move_task(task);
8512 #ifdef CONFIG_FAIR_GROUP_SCHED
8513 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8514 struct cftype *cftype, u64 shareval)
8516 if (shareval > scale_load_down(ULONG_MAX))
8517 shareval = MAX_SHARES;
8518 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8521 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8524 struct task_group *tg = css_tg(css);
8526 return (u64) scale_load_down(tg->shares);
8529 #ifdef CONFIG_CFS_BANDWIDTH
8530 static DEFINE_MUTEX(cfs_constraints_mutex);
8532 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8533 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8535 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8537 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8539 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8540 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8542 if (tg == &root_task_group)
8546 * Ensure we have at some amount of bandwidth every period. This is
8547 * to prevent reaching a state of large arrears when throttled via
8548 * entity_tick() resulting in prolonged exit starvation.
8550 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8554 * Likewise, bound things on the otherside by preventing insane quota
8555 * periods. This also allows us to normalize in computing quota
8558 if (period > max_cfs_quota_period)
8562 * Prevent race between setting of cfs_rq->runtime_enabled and
8563 * unthrottle_offline_cfs_rqs().
8566 mutex_lock(&cfs_constraints_mutex);
8567 ret = __cfs_schedulable(tg, period, quota);
8571 runtime_enabled = quota != RUNTIME_INF;
8572 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8574 * If we need to toggle cfs_bandwidth_used, off->on must occur
8575 * before making related changes, and on->off must occur afterwards
8577 if (runtime_enabled && !runtime_was_enabled)
8578 cfs_bandwidth_usage_inc();
8579 raw_spin_lock_irq(&cfs_b->lock);
8580 cfs_b->period = ns_to_ktime(period);
8581 cfs_b->quota = quota;
8583 __refill_cfs_bandwidth_runtime(cfs_b);
8584 /* restart the period timer (if active) to handle new period expiry */
8585 if (runtime_enabled)
8586 start_cfs_bandwidth(cfs_b);
8587 raw_spin_unlock_irq(&cfs_b->lock);
8589 for_each_online_cpu(i) {
8590 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8591 struct rq *rq = cfs_rq->rq;
8593 raw_spin_lock_irq(&rq->lock);
8594 cfs_rq->runtime_enabled = runtime_enabled;
8595 cfs_rq->runtime_remaining = 0;
8597 if (cfs_rq->throttled)
8598 unthrottle_cfs_rq(cfs_rq);
8599 raw_spin_unlock_irq(&rq->lock);
8601 if (runtime_was_enabled && !runtime_enabled)
8602 cfs_bandwidth_usage_dec();
8604 mutex_unlock(&cfs_constraints_mutex);
8610 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8614 period = ktime_to_ns(tg->cfs_bandwidth.period);
8615 if (cfs_quota_us < 0)
8616 quota = RUNTIME_INF;
8617 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
8618 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8622 return tg_set_cfs_bandwidth(tg, period, quota);
8625 long tg_get_cfs_quota(struct task_group *tg)
8629 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8632 quota_us = tg->cfs_bandwidth.quota;
8633 do_div(quota_us, NSEC_PER_USEC);
8638 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8642 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
8645 period = (u64)cfs_period_us * NSEC_PER_USEC;
8646 quota = tg->cfs_bandwidth.quota;
8648 return tg_set_cfs_bandwidth(tg, period, quota);
8651 long tg_get_cfs_period(struct task_group *tg)
8655 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8656 do_div(cfs_period_us, NSEC_PER_USEC);
8658 return cfs_period_us;
8661 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8664 return tg_get_cfs_quota(css_tg(css));
8667 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8668 struct cftype *cftype, s64 cfs_quota_us)
8670 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8673 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8676 return tg_get_cfs_period(css_tg(css));
8679 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8680 struct cftype *cftype, u64 cfs_period_us)
8682 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8685 struct cfs_schedulable_data {
8686 struct task_group *tg;
8691 * normalize group quota/period to be quota/max_period
8692 * note: units are usecs
8694 static u64 normalize_cfs_quota(struct task_group *tg,
8695 struct cfs_schedulable_data *d)
8703 period = tg_get_cfs_period(tg);
8704 quota = tg_get_cfs_quota(tg);
8707 /* note: these should typically be equivalent */
8708 if (quota == RUNTIME_INF || quota == -1)
8711 return to_ratio(period, quota);
8714 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8716 struct cfs_schedulable_data *d = data;
8717 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8718 s64 quota = 0, parent_quota = -1;
8721 quota = RUNTIME_INF;
8723 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8725 quota = normalize_cfs_quota(tg, d);
8726 parent_quota = parent_b->hierarchical_quota;
8729 * ensure max(child_quota) <= parent_quota, inherit when no
8732 if (quota == RUNTIME_INF)
8733 quota = parent_quota;
8734 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8737 cfs_b->hierarchical_quota = quota;
8742 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8745 struct cfs_schedulable_data data = {
8751 if (quota != RUNTIME_INF) {
8752 do_div(data.period, NSEC_PER_USEC);
8753 do_div(data.quota, NSEC_PER_USEC);
8757 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8763 static int cpu_stats_show(struct seq_file *sf, void *v)
8765 struct task_group *tg = css_tg(seq_css(sf));
8766 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8768 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8769 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8770 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8774 #endif /* CONFIG_CFS_BANDWIDTH */
8775 #endif /* CONFIG_FAIR_GROUP_SCHED */
8777 #ifdef CONFIG_RT_GROUP_SCHED
8778 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8779 struct cftype *cft, s64 val)
8781 return sched_group_set_rt_runtime(css_tg(css), val);
8784 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8787 return sched_group_rt_runtime(css_tg(css));
8790 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8791 struct cftype *cftype, u64 rt_period_us)
8793 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8796 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8799 return sched_group_rt_period(css_tg(css));
8801 #endif /* CONFIG_RT_GROUP_SCHED */
8803 static struct cftype cpu_files[] = {
8804 #ifdef CONFIG_FAIR_GROUP_SCHED
8807 .read_u64 = cpu_shares_read_u64,
8808 .write_u64 = cpu_shares_write_u64,
8811 #ifdef CONFIG_CFS_BANDWIDTH
8813 .name = "cfs_quota_us",
8814 .read_s64 = cpu_cfs_quota_read_s64,
8815 .write_s64 = cpu_cfs_quota_write_s64,
8818 .name = "cfs_period_us",
8819 .read_u64 = cpu_cfs_period_read_u64,
8820 .write_u64 = cpu_cfs_period_write_u64,
8824 .seq_show = cpu_stats_show,
8827 #ifdef CONFIG_RT_GROUP_SCHED
8829 .name = "rt_runtime_us",
8830 .read_s64 = cpu_rt_runtime_read,
8831 .write_s64 = cpu_rt_runtime_write,
8834 .name = "rt_period_us",
8835 .read_u64 = cpu_rt_period_read_uint,
8836 .write_u64 = cpu_rt_period_write_uint,
8842 struct cgroup_subsys cpu_cgrp_subsys = {
8843 .css_alloc = cpu_cgroup_css_alloc,
8844 .css_online = cpu_cgroup_css_online,
8845 .css_released = cpu_cgroup_css_released,
8846 .css_free = cpu_cgroup_css_free,
8847 .fork = cpu_cgroup_fork,
8848 .can_attach = cpu_cgroup_can_attach,
8849 .attach = cpu_cgroup_attach,
8850 .legacy_cftypes = cpu_files,
8854 #endif /* CONFIG_CGROUP_SCHED */
8856 void dump_cpu_task(int cpu)
8858 pr_info("Task dump for CPU %d:\n", cpu);
8859 sched_show_task(cpu_curr(cpu));
8863 * Nice levels are multiplicative, with a gentle 10% change for every
8864 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8865 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8866 * that remained on nice 0.
8868 * The "10% effect" is relative and cumulative: from _any_ nice level,
8869 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8870 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8871 * If a task goes up by ~10% and another task goes down by ~10% then
8872 * the relative distance between them is ~25%.)
8874 const int sched_prio_to_weight[40] = {
8875 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8876 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8877 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8878 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8879 /* 0 */ 1024, 820, 655, 526, 423,
8880 /* 5 */ 335, 272, 215, 172, 137,
8881 /* 10 */ 110, 87, 70, 56, 45,
8882 /* 15 */ 36, 29, 23, 18, 15,
8886 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8888 * In cases where the weight does not change often, we can use the
8889 * precalculated inverse to speed up arithmetics by turning divisions
8890 * into multiplications:
8892 const u32 sched_prio_to_wmult[40] = {
8893 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8894 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8895 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8896 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8897 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8898 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8899 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8900 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,