4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
78 #include <linux/mutex.h>
80 #include <asm/switch_to.h>
82 #include <asm/irq_regs.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
97 static void update_rq_clock_task(struct rq *rq, s64 delta);
99 void update_rq_clock(struct rq *rq)
103 lockdep_assert_held(&rq->lock);
105 if (rq->clock_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))) {
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))) {
231 raw_spin_unlock(&rq->lock);
232 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
234 while (unlikely(task_on_rq_migrating(p)))
239 #ifdef CONFIG_SCHED_HRTICK
241 * Use HR-timers to deliver accurate preemption points.
244 static void hrtick_clear(struct rq *rq)
246 if (hrtimer_active(&rq->hrtick_timer))
247 hrtimer_cancel(&rq->hrtick_timer);
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
254 static enum hrtimer_restart hrtick(struct hrtimer *timer)
256 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
258 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
260 raw_spin_lock(&rq->lock);
262 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
263 raw_spin_unlock(&rq->lock);
265 return HRTIMER_NORESTART;
270 static void __hrtick_restart(struct rq *rq)
272 struct hrtimer *timer = &rq->hrtick_timer;
274 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
278 * called from hardirq (IPI) context
280 static void __hrtick_start(void *arg)
284 raw_spin_lock(&rq->lock);
285 __hrtick_restart(rq);
286 rq->hrtick_csd_pending = 0;
287 raw_spin_unlock(&rq->lock);
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq *rq, u64 delay)
297 struct hrtimer *timer = &rq->hrtick_timer;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta = max_t(s64, delay, 10000LL);
306 time = ktime_add_ns(timer->base->get_time(), delta);
308 hrtimer_set_expires(timer, time);
310 if (rq == this_rq()) {
311 __hrtick_restart(rq);
312 } else if (!rq->hrtick_csd_pending) {
313 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
314 rq->hrtick_csd_pending = 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq *rq, u64 delay)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay = max_t(u64, delay, 10000LL);
331 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
332 HRTIMER_MODE_REL_PINNED);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq *rq)
339 rq->hrtick_csd_pending = 0;
341 rq->hrtick_csd.flags = 0;
342 rq->hrtick_csd.func = __hrtick_start;
343 rq->hrtick_csd.info = rq;
346 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
347 rq->hrtick_timer.function = hrtick;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq *rq)
354 static inline void init_rq_hrtick(struct rq *rq)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct *p)
385 struct thread_info *ti = task_thread_info(p);
386 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct *p)
397 struct thread_info *ti = task_thread_info(p);
398 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
401 if (!(val & _TIF_POLLING_NRFLAG))
403 if (val & _TIF_NEED_RESCHED)
405 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
414 static bool set_nr_and_not_polling(struct task_struct *p)
416 set_tsk_need_resched(p);
421 static bool set_nr_if_polling(struct task_struct *p)
428 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
430 struct wake_q_node *node = &task->wake_q;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
443 get_task_struct(task);
446 * The head is context local, there can be no concurrency.
449 head->lastp = &node->next;
452 void wake_up_q(struct wake_q_head *head)
454 struct wake_q_node *node = head->first;
456 while (node != WAKE_Q_TAIL) {
457 struct task_struct *task;
459 task = container_of(node, struct task_struct, wake_q);
461 /* task can safely be re-inserted now */
463 task->wake_q.next = NULL;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task);
470 put_task_struct(task);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq *rq)
483 struct task_struct *curr = rq->curr;
486 lockdep_assert_held(&rq->lock);
488 if (test_tsk_need_resched(curr))
493 if (cpu == smp_processor_id()) {
494 set_tsk_need_resched(curr);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr))
500 smp_send_reschedule(cpu);
502 trace_sched_wake_idle_without_ipi(cpu);
505 void resched_cpu(int cpu)
507 struct rq *rq = cpu_rq(cpu);
510 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
513 raw_spin_unlock_irqrestore(&rq->lock, flags);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i, cpu = smp_processor_id();
529 struct sched_domain *sd;
531 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
535 for_each_domain(cpu, sd) {
536 for_each_cpu(i, sched_domain_span(sd)) {
540 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
547 if (!is_housekeeping_cpu(cpu))
548 cpu = housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu)
565 struct rq *rq = cpu_rq(cpu);
567 if (cpu == smp_processor_id())
570 if (set_nr_and_not_polling(rq->idle))
571 smp_send_reschedule(cpu);
573 trace_sched_wake_idle_without_ipi(cpu);
576 static bool wake_up_full_nohz_cpu(int cpu)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (cpu_is_offline(cpu))
585 return true; /* Don't try to wake offline CPUs. */
586 if (tick_nohz_full_cpu(cpu)) {
587 if (cpu != smp_processor_id() ||
588 tick_nohz_tick_stopped())
589 tick_nohz_full_kick_cpu(cpu);
597 * Wake up the specified CPU. If the CPU is going offline, it is the
598 * caller's responsibility to deal with the lost wakeup, for example,
599 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
601 void wake_up_nohz_cpu(int cpu)
603 if (!wake_up_full_nohz_cpu(cpu))
604 wake_up_idle_cpu(cpu);
607 static inline bool got_nohz_idle_kick(void)
609 int cpu = smp_processor_id();
611 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
614 if (idle_cpu(cpu) && !need_resched())
618 * We can't run Idle Load Balance on this CPU for this time so we
619 * cancel it and clear NOHZ_BALANCE_KICK
621 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
625 #else /* CONFIG_NO_HZ_COMMON */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ_COMMON */
634 #ifdef CONFIG_NO_HZ_FULL
635 bool sched_can_stop_tick(struct rq *rq)
639 /* Deadline tasks, even if single, need the tick */
640 if (rq->dl.dl_nr_running)
644 * If there are more than one RR tasks, we need the tick to effect the
645 * actual RR behaviour.
647 if (rq->rt.rr_nr_running) {
648 if (rq->rt.rr_nr_running == 1)
655 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 * forced preemption between FIFO tasks.
658 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
663 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 * if there's more than one we need the tick for involuntary
667 if (rq->nr_running > 1)
672 #endif /* CONFIG_NO_HZ_FULL */
674 void sched_avg_update(struct rq *rq)
676 s64 period = sched_avg_period();
678 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
680 * Inline assembly required to prevent the compiler
681 * optimising this loop into a divmod call.
682 * See __iter_div_u64_rem() for another example of this.
684 asm("" : "+rm" (rq->age_stamp));
685 rq->age_stamp += period;
690 #endif /* CONFIG_SMP */
692 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
695 * Iterate task_group tree rooted at *from, calling @down when first entering a
696 * node and @up when leaving it for the final time.
698 * Caller must hold rcu_lock or sufficient equivalent.
700 int walk_tg_tree_from(struct task_group *from,
701 tg_visitor down, tg_visitor up, void *data)
703 struct task_group *parent, *child;
709 ret = (*down)(parent, data);
712 list_for_each_entry_rcu(child, &parent->children, siblings) {
719 ret = (*up)(parent, data);
720 if (ret || parent == from)
724 parent = parent->parent;
731 int tg_nop(struct task_group *tg, void *data)
737 static void set_load_weight(struct task_struct *p)
739 int prio = p->static_prio - MAX_RT_PRIO;
740 struct load_weight *load = &p->se.load;
743 * SCHED_IDLE tasks get minimal weight:
745 if (idle_policy(p->policy)) {
746 load->weight = scale_load(WEIGHT_IDLEPRIO);
747 load->inv_weight = WMULT_IDLEPRIO;
751 load->weight = scale_load(sched_prio_to_weight[prio]);
752 load->inv_weight = sched_prio_to_wmult[prio];
755 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
758 if (!(flags & ENQUEUE_RESTORE))
759 sched_info_queued(rq, p);
760 p->sched_class->enqueue_task(rq, p, flags);
763 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
766 if (!(flags & DEQUEUE_SAVE))
767 sched_info_dequeued(rq, p);
768 p->sched_class->dequeue_task(rq, p, flags);
771 void activate_task(struct rq *rq, struct task_struct *p, int flags)
773 if (task_contributes_to_load(p))
774 rq->nr_uninterruptible--;
776 enqueue_task(rq, p, flags);
779 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible++;
784 dequeue_task(rq, p, flags);
787 static void update_rq_clock_task(struct rq *rq, s64 delta)
790 * In theory, the compile should just see 0 here, and optimize out the call
791 * to sched_rt_avg_update. But I don't trust it...
793 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
794 s64 steal = 0, irq_delta = 0;
796 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
797 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
800 * Since irq_time is only updated on {soft,}irq_exit, we might run into
801 * this case when a previous update_rq_clock() happened inside a
804 * When this happens, we stop ->clock_task and only update the
805 * prev_irq_time stamp to account for the part that fit, so that a next
806 * update will consume the rest. This ensures ->clock_task is
809 * It does however cause some slight miss-attribution of {soft,}irq
810 * time, a more accurate solution would be to update the irq_time using
811 * the current rq->clock timestamp, except that would require using
814 if (irq_delta > delta)
817 rq->prev_irq_time += irq_delta;
820 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
821 if (static_key_false((¶virt_steal_rq_enabled))) {
822 steal = paravirt_steal_clock(cpu_of(rq));
823 steal -= rq->prev_steal_time_rq;
825 if (unlikely(steal > delta))
828 rq->prev_steal_time_rq += steal;
833 rq->clock_task += delta;
835 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
836 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
837 sched_rt_avg_update(rq, irq_delta + steal);
841 void sched_set_stop_task(int cpu, struct task_struct *stop)
843 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
844 struct task_struct *old_stop = cpu_rq(cpu)->stop;
848 * Make it appear like a SCHED_FIFO task, its something
849 * userspace knows about and won't get confused about.
851 * Also, it will make PI more or less work without too
852 * much confusion -- but then, stop work should not
853 * rely on PI working anyway.
855 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
857 stop->sched_class = &stop_sched_class;
860 cpu_rq(cpu)->stop = stop;
864 * Reset it back to a normal scheduling class so that
865 * it can die in pieces.
867 old_stop->sched_class = &rt_sched_class;
872 * __normal_prio - return the priority that is based on the static prio
874 static inline int __normal_prio(struct task_struct *p)
876 return p->static_prio;
880 * Calculate the expected normal priority: i.e. priority
881 * without taking RT-inheritance into account. Might be
882 * boosted by interactivity modifiers. Changes upon fork,
883 * setprio syscalls, and whenever the interactivity
884 * estimator recalculates.
886 static inline int normal_prio(struct task_struct *p)
890 if (task_has_dl_policy(p))
891 prio = MAX_DL_PRIO-1;
892 else if (task_has_rt_policy(p))
893 prio = MAX_RT_PRIO-1 - p->rt_priority;
895 prio = __normal_prio(p);
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
906 static int effective_prio(struct task_struct *p)
908 p->normal_prio = normal_prio(p);
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
914 if (!rt_prio(p->prio))
915 return p->normal_prio;
920 * task_curr - is this task currently executing on a CPU?
921 * @p: the task in question.
923 * Return: 1 if the task is currently executing. 0 otherwise.
925 inline int task_curr(const struct task_struct *p)
927 return cpu_curr(task_cpu(p)) == p;
931 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
932 * use the balance_callback list if you want balancing.
934 * this means any call to check_class_changed() must be followed by a call to
935 * balance_callback().
937 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
938 const struct sched_class *prev_class,
941 if (prev_class != p->sched_class) {
942 if (prev_class->switched_from)
943 prev_class->switched_from(rq, p);
945 p->sched_class->switched_to(rq, p);
946 } else if (oldprio != p->prio || dl_task(p))
947 p->sched_class->prio_changed(rq, p, oldprio);
950 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
952 const struct sched_class *class;
954 if (p->sched_class == rq->curr->sched_class) {
955 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
957 for_each_class(class) {
958 if (class == rq->curr->sched_class)
960 if (class == p->sched_class) {
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
972 rq_clock_skip_update(rq, true);
977 * This is how migration works:
979 * 1) we invoke migration_cpu_stop() on the target CPU using
981 * 2) stopper starts to run (implicitly forcing the migrated thread
983 * 3) it checks whether the migrated task is still in the wrong runqueue.
984 * 4) if it's in the wrong runqueue then the migration thread removes
985 * it and puts it into the right queue.
986 * 5) stopper completes and stop_one_cpu() returns and the migration
991 * move_queued_task - move a queued task to new rq.
993 * Returns (locked) new rq. Old rq's lock is released.
995 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
997 lockdep_assert_held(&rq->lock);
999 p->on_rq = TASK_ON_RQ_MIGRATING;
1000 dequeue_task(rq, p, 0);
1001 set_task_cpu(p, new_cpu);
1002 raw_spin_unlock(&rq->lock);
1004 rq = cpu_rq(new_cpu);
1006 raw_spin_lock(&rq->lock);
1007 BUG_ON(task_cpu(p) != new_cpu);
1008 enqueue_task(rq, p, 0);
1009 p->on_rq = TASK_ON_RQ_QUEUED;
1010 check_preempt_curr(rq, p, 0);
1015 struct migration_arg {
1016 struct task_struct *task;
1021 * Move (not current) task off this cpu, onto dest cpu. We're doing
1022 * this because either it can't run here any more (set_cpus_allowed()
1023 * away from this CPU, or CPU going down), or because we're
1024 * attempting to rebalance this task on exec (sched_exec).
1026 * So we race with normal scheduler movements, but that's OK, as long
1027 * as the task is no longer on this CPU.
1029 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1031 if (unlikely(!cpu_active(dest_cpu)))
1034 /* Affinity changed (again). */
1035 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1038 rq = move_queued_task(rq, p, dest_cpu);
1044 * migration_cpu_stop - this will be executed by a highprio stopper thread
1045 * and performs thread migration by bumping thread off CPU then
1046 * 'pushing' onto another runqueue.
1048 static int migration_cpu_stop(void *data)
1050 struct migration_arg *arg = data;
1051 struct task_struct *p = arg->task;
1052 struct rq *rq = this_rq();
1055 * The original target cpu might have gone down and we might
1056 * be on another cpu but it doesn't matter.
1058 local_irq_disable();
1060 * We need to explicitly wake pending tasks before running
1061 * __migrate_task() such that we will not miss enforcing cpus_allowed
1062 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1064 sched_ttwu_pending();
1066 raw_spin_lock(&p->pi_lock);
1067 raw_spin_lock(&rq->lock);
1069 * If task_rq(p) != rq, it cannot be migrated here, because we're
1070 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1071 * we're holding p->pi_lock.
1073 if (task_rq(p) == rq) {
1074 if (task_on_rq_queued(p))
1075 rq = __migrate_task(rq, p, arg->dest_cpu);
1077 p->wake_cpu = arg->dest_cpu;
1079 raw_spin_unlock(&rq->lock);
1080 raw_spin_unlock(&p->pi_lock);
1087 * sched_class::set_cpus_allowed must do the below, but is not required to
1088 * actually call this function.
1090 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1092 cpumask_copy(&p->cpus_allowed, new_mask);
1093 p->nr_cpus_allowed = cpumask_weight(new_mask);
1096 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1098 struct rq *rq = task_rq(p);
1099 bool queued, running;
1101 lockdep_assert_held(&p->pi_lock);
1103 queued = task_on_rq_queued(p);
1104 running = task_current(rq, p);
1108 * Because __kthread_bind() calls this on blocked tasks without
1111 lockdep_assert_held(&rq->lock);
1112 dequeue_task(rq, p, DEQUEUE_SAVE);
1115 put_prev_task(rq, p);
1117 p->sched_class->set_cpus_allowed(p, new_mask);
1120 enqueue_task(rq, p, ENQUEUE_RESTORE);
1122 set_curr_task(rq, p);
1126 * Change a given task's CPU affinity. Migrate the thread to a
1127 * proper CPU and schedule it away if the CPU it's executing on
1128 * is removed from the allowed bitmask.
1130 * NOTE: the caller must have a valid reference to the task, the
1131 * task must not exit() & deallocate itself prematurely. The
1132 * call is not atomic; no spinlocks may be held.
1134 static int __set_cpus_allowed_ptr(struct task_struct *p,
1135 const struct cpumask *new_mask, bool check)
1137 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1138 unsigned int dest_cpu;
1143 rq = task_rq_lock(p, &rf);
1145 if (p->flags & PF_KTHREAD) {
1147 * Kernel threads are allowed on online && !active CPUs
1149 cpu_valid_mask = cpu_online_mask;
1153 * Must re-check here, to close a race against __kthread_bind(),
1154 * sched_setaffinity() is not guaranteed to observe the flag.
1156 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1161 if (cpumask_equal(&p->cpus_allowed, new_mask))
1164 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1169 do_set_cpus_allowed(p, new_mask);
1171 if (p->flags & PF_KTHREAD) {
1173 * For kernel threads that do indeed end up on online &&
1174 * !active we want to ensure they are strict per-cpu threads.
1176 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1177 !cpumask_intersects(new_mask, cpu_active_mask) &&
1178 p->nr_cpus_allowed != 1);
1181 /* Can the task run on the task's current CPU? If so, we're done */
1182 if (cpumask_test_cpu(task_cpu(p), new_mask))
1185 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1186 if (task_running(rq, p) || p->state == TASK_WAKING) {
1187 struct migration_arg arg = { p, dest_cpu };
1188 /* Need help from migration thread: drop lock and wait. */
1189 task_rq_unlock(rq, p, &rf);
1190 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1191 tlb_migrate_finish(p->mm);
1193 } else if (task_on_rq_queued(p)) {
1195 * OK, since we're going to drop the lock immediately
1196 * afterwards anyway.
1198 rq_unpin_lock(rq, &rf);
1199 rq = move_queued_task(rq, p, dest_cpu);
1200 rq_repin_lock(rq, &rf);
1203 task_rq_unlock(rq, p, &rf);
1208 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1210 return __set_cpus_allowed_ptr(p, new_mask, false);
1212 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1214 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1216 #ifdef CONFIG_SCHED_DEBUG
1218 * We should never call set_task_cpu() on a blocked task,
1219 * ttwu() will sort out the placement.
1221 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1225 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1226 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1227 * time relying on p->on_rq.
1229 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1230 p->sched_class == &fair_sched_class &&
1231 (p->on_rq && !task_on_rq_migrating(p)));
1233 #ifdef CONFIG_LOCKDEP
1235 * The caller should hold either p->pi_lock or rq->lock, when changing
1236 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1238 * sched_move_task() holds both and thus holding either pins the cgroup,
1241 * Furthermore, all task_rq users should acquire both locks, see
1244 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1245 lockdep_is_held(&task_rq(p)->lock)));
1249 trace_sched_migrate_task(p, new_cpu);
1251 if (task_cpu(p) != new_cpu) {
1252 if (p->sched_class->migrate_task_rq)
1253 p->sched_class->migrate_task_rq(p);
1254 p->se.nr_migrations++;
1255 perf_event_task_migrate(p);
1258 __set_task_cpu(p, new_cpu);
1261 static void __migrate_swap_task(struct task_struct *p, int cpu)
1263 if (task_on_rq_queued(p)) {
1264 struct rq *src_rq, *dst_rq;
1266 src_rq = task_rq(p);
1267 dst_rq = cpu_rq(cpu);
1269 p->on_rq = TASK_ON_RQ_MIGRATING;
1270 deactivate_task(src_rq, p, 0);
1271 set_task_cpu(p, cpu);
1272 activate_task(dst_rq, p, 0);
1273 p->on_rq = TASK_ON_RQ_QUEUED;
1274 check_preempt_curr(dst_rq, p, 0);
1277 * Task isn't running anymore; make it appear like we migrated
1278 * it before it went to sleep. This means on wakeup we make the
1279 * previous cpu our target instead of where it really is.
1285 struct migration_swap_arg {
1286 struct task_struct *src_task, *dst_task;
1287 int src_cpu, dst_cpu;
1290 static int migrate_swap_stop(void *data)
1292 struct migration_swap_arg *arg = data;
1293 struct rq *src_rq, *dst_rq;
1296 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1299 src_rq = cpu_rq(arg->src_cpu);
1300 dst_rq = cpu_rq(arg->dst_cpu);
1302 double_raw_lock(&arg->src_task->pi_lock,
1303 &arg->dst_task->pi_lock);
1304 double_rq_lock(src_rq, dst_rq);
1306 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1309 if (task_cpu(arg->src_task) != arg->src_cpu)
1312 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1315 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1318 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1319 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1324 double_rq_unlock(src_rq, dst_rq);
1325 raw_spin_unlock(&arg->dst_task->pi_lock);
1326 raw_spin_unlock(&arg->src_task->pi_lock);
1332 * Cross migrate two tasks
1334 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1336 struct migration_swap_arg arg;
1339 arg = (struct migration_swap_arg){
1341 .src_cpu = task_cpu(cur),
1343 .dst_cpu = task_cpu(p),
1346 if (arg.src_cpu == arg.dst_cpu)
1350 * These three tests are all lockless; this is OK since all of them
1351 * will be re-checked with proper locks held further down the line.
1353 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1356 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1359 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1362 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1363 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1370 * wait_task_inactive - wait for a thread to unschedule.
1372 * If @match_state is nonzero, it's the @p->state value just checked and
1373 * not expected to change. If it changes, i.e. @p might have woken up,
1374 * then return zero. When we succeed in waiting for @p to be off its CPU,
1375 * we return a positive number (its total switch count). If a second call
1376 * a short while later returns the same number, the caller can be sure that
1377 * @p has remained unscheduled the whole time.
1379 * The caller must ensure that the task *will* unschedule sometime soon,
1380 * else this function might spin for a *long* time. This function can't
1381 * be called with interrupts off, or it may introduce deadlock with
1382 * smp_call_function() if an IPI is sent by the same process we are
1383 * waiting to become inactive.
1385 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1387 int running, queued;
1394 * We do the initial early heuristics without holding
1395 * any task-queue locks at all. We'll only try to get
1396 * the runqueue lock when things look like they will
1402 * If the task is actively running on another CPU
1403 * still, just relax and busy-wait without holding
1406 * NOTE! Since we don't hold any locks, it's not
1407 * even sure that "rq" stays as the right runqueue!
1408 * But we don't care, since "task_running()" will
1409 * return false if the runqueue has changed and p
1410 * is actually now running somewhere else!
1412 while (task_running(rq, p)) {
1413 if (match_state && unlikely(p->state != match_state))
1419 * Ok, time to look more closely! We need the rq
1420 * lock now, to be *sure*. If we're wrong, we'll
1421 * just go back and repeat.
1423 rq = task_rq_lock(p, &rf);
1424 trace_sched_wait_task(p);
1425 running = task_running(rq, p);
1426 queued = task_on_rq_queued(p);
1428 if (!match_state || p->state == match_state)
1429 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1430 task_rq_unlock(rq, p, &rf);
1433 * If it changed from the expected state, bail out now.
1435 if (unlikely(!ncsw))
1439 * Was it really running after all now that we
1440 * checked with the proper locks actually held?
1442 * Oops. Go back and try again..
1444 if (unlikely(running)) {
1450 * It's not enough that it's not actively running,
1451 * it must be off the runqueue _entirely_, and not
1454 * So if it was still runnable (but just not actively
1455 * running right now), it's preempted, and we should
1456 * yield - it could be a while.
1458 if (unlikely(queued)) {
1459 ktime_t to = NSEC_PER_SEC / HZ;
1461 set_current_state(TASK_UNINTERRUPTIBLE);
1462 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1467 * Ahh, all good. It wasn't running, and it wasn't
1468 * runnable, which means that it will never become
1469 * running in the future either. We're all done!
1478 * kick_process - kick a running thread to enter/exit the kernel
1479 * @p: the to-be-kicked thread
1481 * Cause a process which is running on another CPU to enter
1482 * kernel-mode, without any delay. (to get signals handled.)
1484 * NOTE: this function doesn't have to take the runqueue lock,
1485 * because all it wants to ensure is that the remote task enters
1486 * the kernel. If the IPI races and the task has been migrated
1487 * to another CPU then no harm is done and the purpose has been
1490 void kick_process(struct task_struct *p)
1496 if ((cpu != smp_processor_id()) && task_curr(p))
1497 smp_send_reschedule(cpu);
1500 EXPORT_SYMBOL_GPL(kick_process);
1503 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1505 * A few notes on cpu_active vs cpu_online:
1507 * - cpu_active must be a subset of cpu_online
1509 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1510 * see __set_cpus_allowed_ptr(). At this point the newly online
1511 * cpu isn't yet part of the sched domains, and balancing will not
1514 * - on cpu-down we clear cpu_active() to mask the sched domains and
1515 * avoid the load balancer to place new tasks on the to be removed
1516 * cpu. Existing tasks will remain running there and will be taken
1519 * This means that fallback selection must not select !active CPUs.
1520 * And can assume that any active CPU must be online. Conversely
1521 * select_task_rq() below may allow selection of !active CPUs in order
1522 * to satisfy the above rules.
1524 static int select_fallback_rq(int cpu, struct task_struct *p)
1526 int nid = cpu_to_node(cpu);
1527 const struct cpumask *nodemask = NULL;
1528 enum { cpuset, possible, fail } state = cpuset;
1532 * If the node that the cpu is on has been offlined, cpu_to_node()
1533 * will return -1. There is no cpu on the node, and we should
1534 * select the cpu on the other node.
1537 nodemask = cpumask_of_node(nid);
1539 /* Look for allowed, online CPU in same node. */
1540 for_each_cpu(dest_cpu, nodemask) {
1541 if (!cpu_active(dest_cpu))
1543 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1549 /* Any allowed, online CPU? */
1550 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1551 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1553 if (!cpu_online(dest_cpu))
1558 /* No more Mr. Nice Guy. */
1561 if (IS_ENABLED(CONFIG_CPUSETS)) {
1562 cpuset_cpus_allowed_fallback(p);
1568 do_set_cpus_allowed(p, cpu_possible_mask);
1579 if (state != cpuset) {
1581 * Don't tell them about moving exiting tasks or
1582 * kernel threads (both mm NULL), since they never
1585 if (p->mm && printk_ratelimit()) {
1586 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1587 task_pid_nr(p), p->comm, cpu);
1595 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1598 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1600 lockdep_assert_held(&p->pi_lock);
1602 if (tsk_nr_cpus_allowed(p) > 1)
1603 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1605 cpu = cpumask_any(tsk_cpus_allowed(p));
1608 * In order not to call set_task_cpu() on a blocking task we need
1609 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1612 * Since this is common to all placement strategies, this lives here.
1614 * [ this allows ->select_task() to simply return task_cpu(p) and
1615 * not worry about this generic constraint ]
1617 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1619 cpu = select_fallback_rq(task_cpu(p), p);
1624 static void update_avg(u64 *avg, u64 sample)
1626 s64 diff = sample - *avg;
1632 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1633 const struct cpumask *new_mask, bool check)
1635 return set_cpus_allowed_ptr(p, new_mask);
1638 #endif /* CONFIG_SMP */
1641 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1645 if (!schedstat_enabled())
1651 if (cpu == rq->cpu) {
1652 schedstat_inc(rq->ttwu_local);
1653 schedstat_inc(p->se.statistics.nr_wakeups_local);
1655 struct sched_domain *sd;
1657 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1659 for_each_domain(rq->cpu, sd) {
1660 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1661 schedstat_inc(sd->ttwu_wake_remote);
1668 if (wake_flags & WF_MIGRATED)
1669 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1670 #endif /* CONFIG_SMP */
1672 schedstat_inc(rq->ttwu_count);
1673 schedstat_inc(p->se.statistics.nr_wakeups);
1675 if (wake_flags & WF_SYNC)
1676 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1679 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1681 activate_task(rq, p, en_flags);
1682 p->on_rq = TASK_ON_RQ_QUEUED;
1684 /* if a worker is waking up, notify workqueue */
1685 if (p->flags & PF_WQ_WORKER)
1686 wq_worker_waking_up(p, cpu_of(rq));
1690 * Mark the task runnable and perform wakeup-preemption.
1692 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1693 struct rq_flags *rf)
1695 check_preempt_curr(rq, p, wake_flags);
1696 p->state = TASK_RUNNING;
1697 trace_sched_wakeup(p);
1700 if (p->sched_class->task_woken) {
1702 * Our task @p is fully woken up and running; so its safe to
1703 * drop the rq->lock, hereafter rq is only used for statistics.
1705 rq_unpin_lock(rq, rf);
1706 p->sched_class->task_woken(rq, p);
1707 rq_repin_lock(rq, rf);
1710 if (rq->idle_stamp) {
1711 u64 delta = rq_clock(rq) - rq->idle_stamp;
1712 u64 max = 2*rq->max_idle_balance_cost;
1714 update_avg(&rq->avg_idle, delta);
1716 if (rq->avg_idle > max)
1725 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1726 struct rq_flags *rf)
1728 int en_flags = ENQUEUE_WAKEUP;
1730 lockdep_assert_held(&rq->lock);
1733 if (p->sched_contributes_to_load)
1734 rq->nr_uninterruptible--;
1736 if (wake_flags & WF_MIGRATED)
1737 en_flags |= ENQUEUE_MIGRATED;
1740 ttwu_activate(rq, p, en_flags);
1741 ttwu_do_wakeup(rq, p, wake_flags, rf);
1745 * Called in case the task @p isn't fully descheduled from its runqueue,
1746 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1747 * since all we need to do is flip p->state to TASK_RUNNING, since
1748 * the task is still ->on_rq.
1750 static int ttwu_remote(struct task_struct *p, int wake_flags)
1756 rq = __task_rq_lock(p, &rf);
1757 if (task_on_rq_queued(p)) {
1758 /* check_preempt_curr() may use rq clock */
1759 update_rq_clock(rq);
1760 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1763 __task_rq_unlock(rq, &rf);
1769 void sched_ttwu_pending(void)
1771 struct rq *rq = this_rq();
1772 struct llist_node *llist = llist_del_all(&rq->wake_list);
1773 struct task_struct *p;
1774 unsigned long flags;
1780 raw_spin_lock_irqsave(&rq->lock, flags);
1781 rq_pin_lock(rq, &rf);
1786 p = llist_entry(llist, struct task_struct, wake_entry);
1787 llist = llist_next(llist);
1789 if (p->sched_remote_wakeup)
1790 wake_flags = WF_MIGRATED;
1792 ttwu_do_activate(rq, p, wake_flags, &rf);
1795 rq_unpin_lock(rq, &rf);
1796 raw_spin_unlock_irqrestore(&rq->lock, flags);
1799 void scheduler_ipi(void)
1802 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1803 * TIF_NEED_RESCHED remotely (for the first time) will also send
1806 preempt_fold_need_resched();
1808 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1812 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1813 * traditionally all their work was done from the interrupt return
1814 * path. Now that we actually do some work, we need to make sure
1817 * Some archs already do call them, luckily irq_enter/exit nest
1820 * Arguably we should visit all archs and update all handlers,
1821 * however a fair share of IPIs are still resched only so this would
1822 * somewhat pessimize the simple resched case.
1825 sched_ttwu_pending();
1828 * Check if someone kicked us for doing the nohz idle load balance.
1830 if (unlikely(got_nohz_idle_kick())) {
1831 this_rq()->idle_balance = 1;
1832 raise_softirq_irqoff(SCHED_SOFTIRQ);
1837 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1839 struct rq *rq = cpu_rq(cpu);
1841 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1843 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1844 if (!set_nr_if_polling(rq->idle))
1845 smp_send_reschedule(cpu);
1847 trace_sched_wake_idle_without_ipi(cpu);
1851 void wake_up_if_idle(int cpu)
1853 struct rq *rq = cpu_rq(cpu);
1854 unsigned long flags;
1858 if (!is_idle_task(rcu_dereference(rq->curr)))
1861 if (set_nr_if_polling(rq->idle)) {
1862 trace_sched_wake_idle_without_ipi(cpu);
1864 raw_spin_lock_irqsave(&rq->lock, flags);
1865 if (is_idle_task(rq->curr))
1866 smp_send_reschedule(cpu);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq->lock, flags);
1875 bool cpus_share_cache(int this_cpu, int that_cpu)
1877 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1879 #endif /* CONFIG_SMP */
1881 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1883 struct rq *rq = cpu_rq(cpu);
1886 #if defined(CONFIG_SMP)
1887 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1888 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1889 ttwu_queue_remote(p, cpu, wake_flags);
1894 raw_spin_lock(&rq->lock);
1895 rq_pin_lock(rq, &rf);
1896 ttwu_do_activate(rq, p, wake_flags, &rf);
1897 rq_unpin_lock(rq, &rf);
1898 raw_spin_unlock(&rq->lock);
1902 * Notes on Program-Order guarantees on SMP systems.
1906 * The basic program-order guarantee on SMP systems is that when a task [t]
1907 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1908 * execution on its new cpu [c1].
1910 * For migration (of runnable tasks) this is provided by the following means:
1912 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1913 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1914 * rq(c1)->lock (if not at the same time, then in that order).
1915 * C) LOCK of the rq(c1)->lock scheduling in task
1917 * Transitivity guarantees that B happens after A and C after B.
1918 * Note: we only require RCpc transitivity.
1919 * Note: the cpu doing B need not be c0 or c1
1928 * UNLOCK rq(0)->lock
1930 * LOCK rq(0)->lock // orders against CPU0
1932 * UNLOCK rq(0)->lock
1936 * UNLOCK rq(1)->lock
1938 * LOCK rq(1)->lock // orders against CPU2
1941 * UNLOCK rq(1)->lock
1944 * BLOCKING -- aka. SLEEP + WAKEUP
1946 * For blocking we (obviously) need to provide the same guarantee as for
1947 * migration. However the means are completely different as there is no lock
1948 * chain to provide order. Instead we do:
1950 * 1) smp_store_release(X->on_cpu, 0)
1951 * 2) smp_cond_load_acquire(!X->on_cpu)
1955 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1957 * LOCK rq(0)->lock LOCK X->pi_lock
1960 * smp_store_release(X->on_cpu, 0);
1962 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1968 * X->state = RUNNING
1969 * UNLOCK rq(2)->lock
1971 * LOCK rq(2)->lock // orders against CPU1
1974 * UNLOCK rq(2)->lock
1977 * UNLOCK rq(0)->lock
1980 * However; for wakeups there is a second guarantee we must provide, namely we
1981 * must observe the state that lead to our wakeup. That is, not only must our
1982 * task observe its own prior state, it must also observe the stores prior to
1985 * This means that any means of doing remote wakeups must order the CPU doing
1986 * the wakeup against the CPU the task is going to end up running on. This,
1987 * however, is already required for the regular Program-Order guarantee above,
1988 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1993 * try_to_wake_up - wake up a thread
1994 * @p: the thread to be awakened
1995 * @state: the mask of task states that can be woken
1996 * @wake_flags: wake modifier flags (WF_*)
1998 * If (@state & @p->state) @p->state = TASK_RUNNING.
2000 * If the task was not queued/runnable, also place it back on a runqueue.
2002 * Atomic against schedule() which would dequeue a task, also see
2003 * set_current_state().
2005 * Return: %true if @p->state changes (an actual wakeup was done),
2009 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2011 unsigned long flags;
2012 int cpu, success = 0;
2015 * If we are going to wake up a thread waiting for CONDITION we
2016 * need to ensure that CONDITION=1 done by the caller can not be
2017 * reordered with p->state check below. This pairs with mb() in
2018 * set_current_state() the waiting thread does.
2020 smp_mb__before_spinlock();
2021 raw_spin_lock_irqsave(&p->pi_lock, flags);
2022 if (!(p->state & state))
2025 trace_sched_waking(p);
2027 success = 1; /* we're going to change ->state */
2031 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2032 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2033 * in smp_cond_load_acquire() below.
2035 * sched_ttwu_pending() try_to_wake_up()
2036 * [S] p->on_rq = 1; [L] P->state
2037 * UNLOCK rq->lock -----.
2041 * LOCK rq->lock -----'
2045 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2047 * Pairs with the UNLOCK+LOCK on rq->lock from the
2048 * last wakeup of our task and the schedule that got our task
2052 if (p->on_rq && ttwu_remote(p, wake_flags))
2057 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2058 * possible to, falsely, observe p->on_cpu == 0.
2060 * One must be running (->on_cpu == 1) in order to remove oneself
2061 * from the runqueue.
2063 * [S] ->on_cpu = 1; [L] ->on_rq
2067 * [S] ->on_rq = 0; [L] ->on_cpu
2069 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2070 * from the consecutive calls to schedule(); the first switching to our
2071 * task, the second putting it to sleep.
2076 * If the owning (remote) cpu is still in the middle of schedule() with
2077 * this task as prev, wait until its done referencing the task.
2079 * Pairs with the smp_store_release() in finish_lock_switch().
2081 * This ensures that tasks getting woken will be fully ordered against
2082 * their previous state and preserve Program Order.
2084 smp_cond_load_acquire(&p->on_cpu, !VAL);
2086 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2087 p->state = TASK_WAKING;
2089 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2090 if (task_cpu(p) != cpu) {
2091 wake_flags |= WF_MIGRATED;
2092 set_task_cpu(p, cpu);
2094 #endif /* CONFIG_SMP */
2096 ttwu_queue(p, cpu, wake_flags);
2098 ttwu_stat(p, cpu, wake_flags);
2100 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2106 * try_to_wake_up_local - try to wake up a local task with rq lock held
2107 * @p: the thread to be awakened
2108 * @cookie: context's cookie for pinning
2110 * Put @p on the run-queue if it's not already there. The caller must
2111 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2114 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2116 struct rq *rq = task_rq(p);
2118 if (WARN_ON_ONCE(rq != this_rq()) ||
2119 WARN_ON_ONCE(p == current))
2122 lockdep_assert_held(&rq->lock);
2124 if (!raw_spin_trylock(&p->pi_lock)) {
2126 * This is OK, because current is on_cpu, which avoids it being
2127 * picked for load-balance and preemption/IRQs are still
2128 * disabled avoiding further scheduler activity on it and we've
2129 * not yet picked a replacement task.
2131 rq_unpin_lock(rq, rf);
2132 raw_spin_unlock(&rq->lock);
2133 raw_spin_lock(&p->pi_lock);
2134 raw_spin_lock(&rq->lock);
2135 rq_repin_lock(rq, rf);
2138 if (!(p->state & TASK_NORMAL))
2141 trace_sched_waking(p);
2143 if (!task_on_rq_queued(p))
2144 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2146 ttwu_do_wakeup(rq, p, 0, rf);
2147 ttwu_stat(p, smp_processor_id(), 0);
2149 raw_spin_unlock(&p->pi_lock);
2153 * wake_up_process - Wake up a specific process
2154 * @p: The process to be woken up.
2156 * Attempt to wake up the nominated process and move it to the set of runnable
2159 * Return: 1 if the process was woken up, 0 if it was already running.
2161 * It may be assumed that this function implies a write memory barrier before
2162 * changing the task state if and only if any tasks are woken up.
2164 int wake_up_process(struct task_struct *p)
2166 return try_to_wake_up(p, TASK_NORMAL, 0);
2168 EXPORT_SYMBOL(wake_up_process);
2170 int wake_up_state(struct task_struct *p, unsigned int state)
2172 return try_to_wake_up(p, state, 0);
2176 * This function clears the sched_dl_entity static params.
2178 void __dl_clear_params(struct task_struct *p)
2180 struct sched_dl_entity *dl_se = &p->dl;
2182 dl_se->dl_runtime = 0;
2183 dl_se->dl_deadline = 0;
2184 dl_se->dl_period = 0;
2188 dl_se->dl_throttled = 0;
2189 dl_se->dl_yielded = 0;
2193 * Perform scheduler related setup for a newly forked process p.
2194 * p is forked by current.
2196 * __sched_fork() is basic setup used by init_idle() too:
2198 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2203 p->se.exec_start = 0;
2204 p->se.sum_exec_runtime = 0;
2205 p->se.prev_sum_exec_runtime = 0;
2206 p->se.nr_migrations = 0;
2208 INIT_LIST_HEAD(&p->se.group_node);
2210 #ifdef CONFIG_FAIR_GROUP_SCHED
2211 p->se.cfs_rq = NULL;
2214 #ifdef CONFIG_SCHEDSTATS
2215 /* Even if schedstat is disabled, there should not be garbage */
2216 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2219 RB_CLEAR_NODE(&p->dl.rb_node);
2220 init_dl_task_timer(&p->dl);
2221 __dl_clear_params(p);
2223 INIT_LIST_HEAD(&p->rt.run_list);
2225 p->rt.time_slice = sched_rr_timeslice;
2229 #ifdef CONFIG_PREEMPT_NOTIFIERS
2230 INIT_HLIST_HEAD(&p->preempt_notifiers);
2233 #ifdef CONFIG_NUMA_BALANCING
2234 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2235 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2236 p->mm->numa_scan_seq = 0;
2239 if (clone_flags & CLONE_VM)
2240 p->numa_preferred_nid = current->numa_preferred_nid;
2242 p->numa_preferred_nid = -1;
2244 p->node_stamp = 0ULL;
2245 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2246 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2247 p->numa_work.next = &p->numa_work;
2248 p->numa_faults = NULL;
2249 p->last_task_numa_placement = 0;
2250 p->last_sum_exec_runtime = 0;
2252 p->numa_group = NULL;
2253 #endif /* CONFIG_NUMA_BALANCING */
2256 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2258 #ifdef CONFIG_NUMA_BALANCING
2260 void set_numabalancing_state(bool enabled)
2263 static_branch_enable(&sched_numa_balancing);
2265 static_branch_disable(&sched_numa_balancing);
2268 #ifdef CONFIG_PROC_SYSCTL
2269 int sysctl_numa_balancing(struct ctl_table *table, int write,
2270 void __user *buffer, size_t *lenp, loff_t *ppos)
2274 int state = static_branch_likely(&sched_numa_balancing);
2276 if (write && !capable(CAP_SYS_ADMIN))
2281 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2285 set_numabalancing_state(state);
2291 #ifdef CONFIG_SCHEDSTATS
2293 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2294 static bool __initdata __sched_schedstats = false;
2296 static void set_schedstats(bool enabled)
2299 static_branch_enable(&sched_schedstats);
2301 static_branch_disable(&sched_schedstats);
2304 void force_schedstat_enabled(void)
2306 if (!schedstat_enabled()) {
2307 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2308 static_branch_enable(&sched_schedstats);
2312 static int __init setup_schedstats(char *str)
2319 * This code is called before jump labels have been set up, so we can't
2320 * change the static branch directly just yet. Instead set a temporary
2321 * variable so init_schedstats() can do it later.
2323 if (!strcmp(str, "enable")) {
2324 __sched_schedstats = true;
2326 } else if (!strcmp(str, "disable")) {
2327 __sched_schedstats = false;
2332 pr_warn("Unable to parse schedstats=\n");
2336 __setup("schedstats=", setup_schedstats);
2338 static void __init init_schedstats(void)
2340 set_schedstats(__sched_schedstats);
2343 #ifdef CONFIG_PROC_SYSCTL
2344 int sysctl_schedstats(struct ctl_table *table, int write,
2345 void __user *buffer, size_t *lenp, loff_t *ppos)
2349 int state = static_branch_likely(&sched_schedstats);
2351 if (write && !capable(CAP_SYS_ADMIN))
2356 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2360 set_schedstats(state);
2363 #endif /* CONFIG_PROC_SYSCTL */
2364 #else /* !CONFIG_SCHEDSTATS */
2365 static inline void init_schedstats(void) {}
2366 #endif /* CONFIG_SCHEDSTATS */
2369 * fork()/clone()-time setup:
2371 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2373 unsigned long flags;
2374 int cpu = get_cpu();
2376 __sched_fork(clone_flags, p);
2378 * We mark the process as NEW here. This guarantees that
2379 * nobody will actually run it, and a signal or other external
2380 * event cannot wake it up and insert it on the runqueue either.
2382 p->state = TASK_NEW;
2385 * Make sure we do not leak PI boosting priority to the child.
2387 p->prio = current->normal_prio;
2390 * Revert to default priority/policy on fork if requested.
2392 if (unlikely(p->sched_reset_on_fork)) {
2393 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2394 p->policy = SCHED_NORMAL;
2395 p->static_prio = NICE_TO_PRIO(0);
2397 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2398 p->static_prio = NICE_TO_PRIO(0);
2400 p->prio = p->normal_prio = __normal_prio(p);
2404 * We don't need the reset flag anymore after the fork. It has
2405 * fulfilled its duty:
2407 p->sched_reset_on_fork = 0;
2410 if (dl_prio(p->prio)) {
2413 } else if (rt_prio(p->prio)) {
2414 p->sched_class = &rt_sched_class;
2416 p->sched_class = &fair_sched_class;
2419 init_entity_runnable_average(&p->se);
2422 * The child is not yet in the pid-hash so no cgroup attach races,
2423 * and the cgroup is pinned to this child due to cgroup_fork()
2424 * is ran before sched_fork().
2426 * Silence PROVE_RCU.
2428 raw_spin_lock_irqsave(&p->pi_lock, flags);
2430 * We're setting the cpu for the first time, we don't migrate,
2431 * so use __set_task_cpu().
2433 __set_task_cpu(p, cpu);
2434 if (p->sched_class->task_fork)
2435 p->sched_class->task_fork(p);
2436 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2438 #ifdef CONFIG_SCHED_INFO
2439 if (likely(sched_info_on()))
2440 memset(&p->sched_info, 0, sizeof(p->sched_info));
2442 #if defined(CONFIG_SMP)
2445 init_task_preempt_count(p);
2447 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2448 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2455 unsigned long to_ratio(u64 period, u64 runtime)
2457 if (runtime == RUNTIME_INF)
2461 * Doing this here saves a lot of checks in all
2462 * the calling paths, and returning zero seems
2463 * safe for them anyway.
2468 return div64_u64(runtime << 20, period);
2472 inline struct dl_bw *dl_bw_of(int i)
2474 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2475 "sched RCU must be held");
2476 return &cpu_rq(i)->rd->dl_bw;
2479 static inline int dl_bw_cpus(int i)
2481 struct root_domain *rd = cpu_rq(i)->rd;
2484 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2485 "sched RCU must be held");
2486 for_each_cpu_and(i, rd->span, cpu_active_mask)
2492 inline struct dl_bw *dl_bw_of(int i)
2494 return &cpu_rq(i)->dl.dl_bw;
2497 static inline int dl_bw_cpus(int i)
2504 * We must be sure that accepting a new task (or allowing changing the
2505 * parameters of an existing one) is consistent with the bandwidth
2506 * constraints. If yes, this function also accordingly updates the currently
2507 * allocated bandwidth to reflect the new situation.
2509 * This function is called while holding p's rq->lock.
2511 * XXX we should delay bw change until the task's 0-lag point, see
2514 static int dl_overflow(struct task_struct *p, int policy,
2515 const struct sched_attr *attr)
2518 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2519 u64 period = attr->sched_period ?: attr->sched_deadline;
2520 u64 runtime = attr->sched_runtime;
2521 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2524 /* !deadline task may carry old deadline bandwidth */
2525 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2529 * Either if a task, enters, leave, or stays -deadline but changes
2530 * its parameters, we may need to update accordingly the total
2531 * allocated bandwidth of the container.
2533 raw_spin_lock(&dl_b->lock);
2534 cpus = dl_bw_cpus(task_cpu(p));
2535 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2536 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2537 __dl_add(dl_b, new_bw);
2539 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2540 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2541 __dl_clear(dl_b, p->dl.dl_bw);
2542 __dl_add(dl_b, new_bw);
2544 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2545 __dl_clear(dl_b, p->dl.dl_bw);
2548 raw_spin_unlock(&dl_b->lock);
2553 extern void init_dl_bw(struct dl_bw *dl_b);
2556 * wake_up_new_task - wake up a newly created task for the first time.
2558 * This function will do some initial scheduler statistics housekeeping
2559 * that must be done for every newly created context, then puts the task
2560 * on the runqueue and wakes it.
2562 void wake_up_new_task(struct task_struct *p)
2567 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2568 p->state = TASK_RUNNING;
2571 * Fork balancing, do it here and not earlier because:
2572 * - cpus_allowed can change in the fork path
2573 * - any previously selected cpu might disappear through hotplug
2575 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2576 * as we're not fully set-up yet.
2578 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2580 rq = __task_rq_lock(p, &rf);
2581 post_init_entity_util_avg(&p->se);
2583 activate_task(rq, p, 0);
2584 p->on_rq = TASK_ON_RQ_QUEUED;
2585 trace_sched_wakeup_new(p);
2586 check_preempt_curr(rq, p, WF_FORK);
2588 if (p->sched_class->task_woken) {
2590 * Nothing relies on rq->lock after this, so its fine to
2593 rq_unpin_lock(rq, &rf);
2594 p->sched_class->task_woken(rq, p);
2595 rq_repin_lock(rq, &rf);
2598 task_rq_unlock(rq, p, &rf);
2601 #ifdef CONFIG_PREEMPT_NOTIFIERS
2603 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2605 void preempt_notifier_inc(void)
2607 static_key_slow_inc(&preempt_notifier_key);
2609 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2611 void preempt_notifier_dec(void)
2613 static_key_slow_dec(&preempt_notifier_key);
2615 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2618 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2619 * @notifier: notifier struct to register
2621 void preempt_notifier_register(struct preempt_notifier *notifier)
2623 if (!static_key_false(&preempt_notifier_key))
2624 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2626 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2628 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2631 * preempt_notifier_unregister - no longer interested in preemption notifications
2632 * @notifier: notifier struct to unregister
2634 * This is *not* safe to call from within a preemption notifier.
2636 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2638 hlist_del(¬ifier->link);
2640 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2642 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2644 struct preempt_notifier *notifier;
2646 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2647 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2650 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2652 if (static_key_false(&preempt_notifier_key))
2653 __fire_sched_in_preempt_notifiers(curr);
2657 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2658 struct task_struct *next)
2660 struct preempt_notifier *notifier;
2662 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2663 notifier->ops->sched_out(notifier, next);
2666 static __always_inline void
2667 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2668 struct task_struct *next)
2670 if (static_key_false(&preempt_notifier_key))
2671 __fire_sched_out_preempt_notifiers(curr, next);
2674 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2676 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2681 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2682 struct task_struct *next)
2686 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2689 * prepare_task_switch - prepare to switch tasks
2690 * @rq: the runqueue preparing to switch
2691 * @prev: the current task that is being switched out
2692 * @next: the task we are going to switch to.
2694 * This is called with the rq lock held and interrupts off. It must
2695 * be paired with a subsequent finish_task_switch after the context
2698 * prepare_task_switch sets up locking and calls architecture specific
2702 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2703 struct task_struct *next)
2705 sched_info_switch(rq, prev, next);
2706 perf_event_task_sched_out(prev, next);
2707 fire_sched_out_preempt_notifiers(prev, next);
2708 prepare_lock_switch(rq, next);
2709 prepare_arch_switch(next);
2713 * finish_task_switch - clean up after a task-switch
2714 * @prev: the thread we just switched away from.
2716 * finish_task_switch must be called after the context switch, paired
2717 * with a prepare_task_switch call before the context switch.
2718 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2719 * and do any other architecture-specific cleanup actions.
2721 * Note that we may have delayed dropping an mm in context_switch(). If
2722 * so, we finish that here outside of the runqueue lock. (Doing it
2723 * with the lock held can cause deadlocks; see schedule() for
2726 * The context switch have flipped the stack from under us and restored the
2727 * local variables which were saved when this task called schedule() in the
2728 * past. prev == current is still correct but we need to recalculate this_rq
2729 * because prev may have moved to another CPU.
2731 static struct rq *finish_task_switch(struct task_struct *prev)
2732 __releases(rq->lock)
2734 struct rq *rq = this_rq();
2735 struct mm_struct *mm = rq->prev_mm;
2739 * The previous task will have left us with a preempt_count of 2
2740 * because it left us after:
2743 * preempt_disable(); // 1
2745 * raw_spin_lock_irq(&rq->lock) // 2
2747 * Also, see FORK_PREEMPT_COUNT.
2749 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2750 "corrupted preempt_count: %s/%d/0x%x\n",
2751 current->comm, current->pid, preempt_count()))
2752 preempt_count_set(FORK_PREEMPT_COUNT);
2757 * A task struct has one reference for the use as "current".
2758 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2759 * schedule one last time. The schedule call will never return, and
2760 * the scheduled task must drop that reference.
2762 * We must observe prev->state before clearing prev->on_cpu (in
2763 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2764 * running on another CPU and we could rave with its RUNNING -> DEAD
2765 * transition, resulting in a double drop.
2767 prev_state = prev->state;
2768 vtime_task_switch(prev);
2769 perf_event_task_sched_in(prev, current);
2770 finish_lock_switch(rq, prev);
2771 finish_arch_post_lock_switch();
2773 fire_sched_in_preempt_notifiers(current);
2776 if (unlikely(prev_state == TASK_DEAD)) {
2777 if (prev->sched_class->task_dead)
2778 prev->sched_class->task_dead(prev);
2781 * Remove function-return probe instances associated with this
2782 * task and put them back on the free list.
2784 kprobe_flush_task(prev);
2786 /* Task is done with its stack. */
2787 put_task_stack(prev);
2789 put_task_struct(prev);
2792 tick_nohz_task_switch();
2798 /* rq->lock is NOT held, but preemption is disabled */
2799 static void __balance_callback(struct rq *rq)
2801 struct callback_head *head, *next;
2802 void (*func)(struct rq *rq);
2803 unsigned long flags;
2805 raw_spin_lock_irqsave(&rq->lock, flags);
2806 head = rq->balance_callback;
2807 rq->balance_callback = NULL;
2809 func = (void (*)(struct rq *))head->func;
2816 raw_spin_unlock_irqrestore(&rq->lock, flags);
2819 static inline void balance_callback(struct rq *rq)
2821 if (unlikely(rq->balance_callback))
2822 __balance_callback(rq);
2827 static inline void balance_callback(struct rq *rq)
2834 * schedule_tail - first thing a freshly forked thread must call.
2835 * @prev: the thread we just switched away from.
2837 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2838 __releases(rq->lock)
2843 * New tasks start with FORK_PREEMPT_COUNT, see there and
2844 * finish_task_switch() for details.
2846 * finish_task_switch() will drop rq->lock() and lower preempt_count
2847 * and the preempt_enable() will end up enabling preemption (on
2848 * PREEMPT_COUNT kernels).
2851 rq = finish_task_switch(prev);
2852 balance_callback(rq);
2855 if (current->set_child_tid)
2856 put_user(task_pid_vnr(current), current->set_child_tid);
2860 * context_switch - switch to the new MM and the new thread's register state.
2862 static __always_inline struct rq *
2863 context_switch(struct rq *rq, struct task_struct *prev,
2864 struct task_struct *next, struct rq_flags *rf)
2866 struct mm_struct *mm, *oldmm;
2868 prepare_task_switch(rq, prev, next);
2871 oldmm = prev->active_mm;
2873 * For paravirt, this is coupled with an exit in switch_to to
2874 * combine the page table reload and the switch backend into
2877 arch_start_context_switch(prev);
2880 next->active_mm = oldmm;
2881 atomic_inc(&oldmm->mm_count);
2882 enter_lazy_tlb(oldmm, next);
2884 switch_mm_irqs_off(oldmm, mm, next);
2887 prev->active_mm = NULL;
2888 rq->prev_mm = oldmm;
2891 rq->clock_skip_update = 0;
2894 * Since the runqueue lock will be released by the next
2895 * task (which is an invalid locking op but in the case
2896 * of the scheduler it's an obvious special-case), so we
2897 * do an early lockdep release here:
2899 rq_unpin_lock(rq, rf);
2900 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2902 /* Here we just switch the register state and the stack. */
2903 switch_to(prev, next, prev);
2906 return finish_task_switch(prev);
2910 * nr_running and nr_context_switches:
2912 * externally visible scheduler statistics: current number of runnable
2913 * threads, total number of context switches performed since bootup.
2915 unsigned long nr_running(void)
2917 unsigned long i, sum = 0;
2919 for_each_online_cpu(i)
2920 sum += cpu_rq(i)->nr_running;
2926 * Check if only the current task is running on the cpu.
2928 * Caution: this function does not check that the caller has disabled
2929 * preemption, thus the result might have a time-of-check-to-time-of-use
2930 * race. The caller is responsible to use it correctly, for example:
2932 * - from a non-preemptable section (of course)
2934 * - from a thread that is bound to a single CPU
2936 * - in a loop with very short iterations (e.g. a polling loop)
2938 bool single_task_running(void)
2940 return raw_rq()->nr_running == 1;
2942 EXPORT_SYMBOL(single_task_running);
2944 unsigned long long nr_context_switches(void)
2947 unsigned long long sum = 0;
2949 for_each_possible_cpu(i)
2950 sum += cpu_rq(i)->nr_switches;
2955 unsigned long nr_iowait(void)
2957 unsigned long i, sum = 0;
2959 for_each_possible_cpu(i)
2960 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2965 unsigned long nr_iowait_cpu(int cpu)
2967 struct rq *this = cpu_rq(cpu);
2968 return atomic_read(&this->nr_iowait);
2971 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2973 struct rq *rq = this_rq();
2974 *nr_waiters = atomic_read(&rq->nr_iowait);
2975 *load = rq->load.weight;
2981 * sched_exec - execve() is a valuable balancing opportunity, because at
2982 * this point the task has the smallest effective memory and cache footprint.
2984 void sched_exec(void)
2986 struct task_struct *p = current;
2987 unsigned long flags;
2990 raw_spin_lock_irqsave(&p->pi_lock, flags);
2991 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2992 if (dest_cpu == smp_processor_id())
2995 if (likely(cpu_active(dest_cpu))) {
2996 struct migration_arg arg = { p, dest_cpu };
2998 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2999 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3003 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3008 DEFINE_PER_CPU(struct kernel_stat, kstat);
3009 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3011 EXPORT_PER_CPU_SYMBOL(kstat);
3012 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3015 * The function fair_sched_class.update_curr accesses the struct curr
3016 * and its field curr->exec_start; when called from task_sched_runtime(),
3017 * we observe a high rate of cache misses in practice.
3018 * Prefetching this data results in improved performance.
3020 static inline void prefetch_curr_exec_start(struct task_struct *p)
3022 #ifdef CONFIG_FAIR_GROUP_SCHED
3023 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3025 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3028 prefetch(&curr->exec_start);
3032 * Return accounted runtime for the task.
3033 * In case the task is currently running, return the runtime plus current's
3034 * pending runtime that have not been accounted yet.
3036 unsigned long long task_sched_runtime(struct task_struct *p)
3042 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3044 * 64-bit doesn't need locks to atomically read a 64bit value.
3045 * So we have a optimization chance when the task's delta_exec is 0.
3046 * Reading ->on_cpu is racy, but this is ok.
3048 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3049 * If we race with it entering cpu, unaccounted time is 0. This is
3050 * indistinguishable from the read occurring a few cycles earlier.
3051 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3052 * been accounted, so we're correct here as well.
3054 if (!p->on_cpu || !task_on_rq_queued(p))
3055 return p->se.sum_exec_runtime;
3058 rq = task_rq_lock(p, &rf);
3060 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3061 * project cycles that may never be accounted to this
3062 * thread, breaking clock_gettime().
3064 if (task_current(rq, p) && task_on_rq_queued(p)) {
3065 prefetch_curr_exec_start(p);
3066 update_rq_clock(rq);
3067 p->sched_class->update_curr(rq);
3069 ns = p->se.sum_exec_runtime;
3070 task_rq_unlock(rq, p, &rf);
3076 * This function gets called by the timer code, with HZ frequency.
3077 * We call it with interrupts disabled.
3079 void scheduler_tick(void)
3081 int cpu = smp_processor_id();
3082 struct rq *rq = cpu_rq(cpu);
3083 struct task_struct *curr = rq->curr;
3087 raw_spin_lock(&rq->lock);
3088 update_rq_clock(rq);
3089 curr->sched_class->task_tick(rq, curr, 0);
3090 cpu_load_update_active(rq);
3091 calc_global_load_tick(rq);
3092 raw_spin_unlock(&rq->lock);
3094 perf_event_task_tick();
3097 rq->idle_balance = idle_cpu(cpu);
3098 trigger_load_balance(rq);
3100 rq_last_tick_reset(rq);
3103 #ifdef CONFIG_NO_HZ_FULL
3105 * scheduler_tick_max_deferment
3107 * Keep at least one tick per second when a single
3108 * active task is running because the scheduler doesn't
3109 * yet completely support full dynticks environment.
3111 * This makes sure that uptime, CFS vruntime, load
3112 * balancing, etc... continue to move forward, even
3113 * with a very low granularity.
3115 * Return: Maximum deferment in nanoseconds.
3117 u64 scheduler_tick_max_deferment(void)
3119 struct rq *rq = this_rq();
3120 unsigned long next, now = READ_ONCE(jiffies);
3122 next = rq->last_sched_tick + HZ;
3124 if (time_before_eq(next, now))
3127 return jiffies_to_nsecs(next - now);
3131 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3132 defined(CONFIG_PREEMPT_TRACER))
3134 * If the value passed in is equal to the current preempt count
3135 * then we just disabled preemption. Start timing the latency.
3137 static inline void preempt_latency_start(int val)
3139 if (preempt_count() == val) {
3140 unsigned long ip = get_lock_parent_ip();
3141 #ifdef CONFIG_DEBUG_PREEMPT
3142 current->preempt_disable_ip = ip;
3144 trace_preempt_off(CALLER_ADDR0, ip);
3148 void preempt_count_add(int val)
3150 #ifdef CONFIG_DEBUG_PREEMPT
3154 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3157 __preempt_count_add(val);
3158 #ifdef CONFIG_DEBUG_PREEMPT
3160 * Spinlock count overflowing soon?
3162 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3165 preempt_latency_start(val);
3167 EXPORT_SYMBOL(preempt_count_add);
3168 NOKPROBE_SYMBOL(preempt_count_add);
3171 * If the value passed in equals to the current preempt count
3172 * then we just enabled preemption. Stop timing the latency.
3174 static inline void preempt_latency_stop(int val)
3176 if (preempt_count() == val)
3177 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3180 void preempt_count_sub(int val)
3182 #ifdef CONFIG_DEBUG_PREEMPT
3186 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3189 * Is the spinlock portion underflowing?
3191 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3192 !(preempt_count() & PREEMPT_MASK)))
3196 preempt_latency_stop(val);
3197 __preempt_count_sub(val);
3199 EXPORT_SYMBOL(preempt_count_sub);
3200 NOKPROBE_SYMBOL(preempt_count_sub);
3203 static inline void preempt_latency_start(int val) { }
3204 static inline void preempt_latency_stop(int val) { }
3208 * Print scheduling while atomic bug:
3210 static noinline void __schedule_bug(struct task_struct *prev)
3212 /* Save this before calling printk(), since that will clobber it */
3213 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3215 if (oops_in_progress)
3218 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3219 prev->comm, prev->pid, preempt_count());
3221 debug_show_held_locks(prev);
3223 if (irqs_disabled())
3224 print_irqtrace_events(prev);
3225 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3226 && in_atomic_preempt_off()) {
3227 pr_err("Preemption disabled at:");
3228 print_ip_sym(preempt_disable_ip);
3232 panic("scheduling while atomic\n");
3235 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3239 * Various schedule()-time debugging checks and statistics:
3241 static inline void schedule_debug(struct task_struct *prev)
3243 #ifdef CONFIG_SCHED_STACK_END_CHECK
3244 if (task_stack_end_corrupted(prev))
3245 panic("corrupted stack end detected inside scheduler\n");
3248 if (unlikely(in_atomic_preempt_off())) {
3249 __schedule_bug(prev);
3250 preempt_count_set(PREEMPT_DISABLED);
3254 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3256 schedstat_inc(this_rq()->sched_count);
3260 * Pick up the highest-prio task:
3262 static inline struct task_struct *
3263 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3265 const struct sched_class *class = &fair_sched_class;
3266 struct task_struct *p;
3269 * Optimization: we know that if all tasks are in
3270 * the fair class we can call that function directly:
3272 if (likely(prev->sched_class == class &&
3273 rq->nr_running == rq->cfs.h_nr_running)) {
3274 p = fair_sched_class.pick_next_task(rq, prev, rf);
3275 if (unlikely(p == RETRY_TASK))
3278 /* assumes fair_sched_class->next == idle_sched_class */
3280 p = idle_sched_class.pick_next_task(rq, prev, rf);
3286 for_each_class(class) {
3287 p = class->pick_next_task(rq, prev, rf);
3289 if (unlikely(p == RETRY_TASK))
3295 BUG(); /* the idle class will always have a runnable task */
3299 * __schedule() is the main scheduler function.
3301 * The main means of driving the scheduler and thus entering this function are:
3303 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3305 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3306 * paths. For example, see arch/x86/entry_64.S.
3308 * To drive preemption between tasks, the scheduler sets the flag in timer
3309 * interrupt handler scheduler_tick().
3311 * 3. Wakeups don't really cause entry into schedule(). They add a
3312 * task to the run-queue and that's it.
3314 * Now, if the new task added to the run-queue preempts the current
3315 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3316 * called on the nearest possible occasion:
3318 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3320 * - in syscall or exception context, at the next outmost
3321 * preempt_enable(). (this might be as soon as the wake_up()'s
3324 * - in IRQ context, return from interrupt-handler to
3325 * preemptible context
3327 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3330 * - cond_resched() call
3331 * - explicit schedule() call
3332 * - return from syscall or exception to user-space
3333 * - return from interrupt-handler to user-space
3335 * WARNING: must be called with preemption disabled!
3337 static void __sched notrace __schedule(bool preempt)
3339 struct task_struct *prev, *next;
3340 unsigned long *switch_count;
3345 cpu = smp_processor_id();
3349 schedule_debug(prev);
3351 if (sched_feat(HRTICK))
3354 local_irq_disable();
3355 rcu_note_context_switch();
3358 * Make sure that signal_pending_state()->signal_pending() below
3359 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3360 * done by the caller to avoid the race with signal_wake_up().
3362 smp_mb__before_spinlock();
3363 raw_spin_lock(&rq->lock);
3364 rq_pin_lock(rq, &rf);
3366 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3368 switch_count = &prev->nivcsw;
3369 if (!preempt && prev->state) {
3370 if (unlikely(signal_pending_state(prev->state, prev))) {
3371 prev->state = TASK_RUNNING;
3373 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3377 * If a worker went to sleep, notify and ask workqueue
3378 * whether it wants to wake up a task to maintain
3381 if (prev->flags & PF_WQ_WORKER) {
3382 struct task_struct *to_wakeup;
3384 to_wakeup = wq_worker_sleeping(prev);
3386 try_to_wake_up_local(to_wakeup, &rf);
3389 switch_count = &prev->nvcsw;
3392 if (task_on_rq_queued(prev))
3393 update_rq_clock(rq);
3395 next = pick_next_task(rq, prev, &rf);
3396 clear_tsk_need_resched(prev);
3397 clear_preempt_need_resched();
3399 if (likely(prev != next)) {
3404 trace_sched_switch(preempt, prev, next);
3405 rq = context_switch(rq, prev, next, &rf); /* unlocks the rq */
3407 rq->clock_skip_update = 0;
3408 rq_unpin_lock(rq, &rf);
3409 raw_spin_unlock_irq(&rq->lock);
3412 balance_callback(rq);
3415 void __noreturn do_task_dead(void)
3418 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3419 * when the following two conditions become true.
3420 * - There is race condition of mmap_sem (It is acquired by
3422 * - SMI occurs before setting TASK_RUNINNG.
3423 * (or hypervisor of virtual machine switches to other guest)
3424 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3426 * To avoid it, we have to wait for releasing tsk->pi_lock which
3427 * is held by try_to_wake_up()
3430 raw_spin_unlock_wait(¤t->pi_lock);
3432 /* causes final put_task_struct in finish_task_switch(). */
3433 __set_current_state(TASK_DEAD);
3434 current->flags |= PF_NOFREEZE; /* tell freezer to ignore us */
3437 /* Avoid "noreturn function does return". */
3439 cpu_relax(); /* For when BUG is null */
3442 static inline void sched_submit_work(struct task_struct *tsk)
3444 if (!tsk->state || tsk_is_pi_blocked(tsk))
3447 * If we are going to sleep and we have plugged IO queued,
3448 * make sure to submit it to avoid deadlocks.
3450 if (blk_needs_flush_plug(tsk))
3451 blk_schedule_flush_plug(tsk);
3454 asmlinkage __visible void __sched schedule(void)
3456 struct task_struct *tsk = current;
3458 sched_submit_work(tsk);
3462 sched_preempt_enable_no_resched();
3463 } while (need_resched());
3465 EXPORT_SYMBOL(schedule);
3467 #ifdef CONFIG_CONTEXT_TRACKING
3468 asmlinkage __visible void __sched schedule_user(void)
3471 * If we come here after a random call to set_need_resched(),
3472 * or we have been woken up remotely but the IPI has not yet arrived,
3473 * we haven't yet exited the RCU idle mode. Do it here manually until
3474 * we find a better solution.
3476 * NB: There are buggy callers of this function. Ideally we
3477 * should warn if prev_state != CONTEXT_USER, but that will trigger
3478 * too frequently to make sense yet.
3480 enum ctx_state prev_state = exception_enter();
3482 exception_exit(prev_state);
3487 * schedule_preempt_disabled - called with preemption disabled
3489 * Returns with preemption disabled. Note: preempt_count must be 1
3491 void __sched schedule_preempt_disabled(void)
3493 sched_preempt_enable_no_resched();
3498 static void __sched notrace preempt_schedule_common(void)
3502 * Because the function tracer can trace preempt_count_sub()
3503 * and it also uses preempt_enable/disable_notrace(), if
3504 * NEED_RESCHED is set, the preempt_enable_notrace() called
3505 * by the function tracer will call this function again and
3506 * cause infinite recursion.
3508 * Preemption must be disabled here before the function
3509 * tracer can trace. Break up preempt_disable() into two
3510 * calls. One to disable preemption without fear of being
3511 * traced. The other to still record the preemption latency,
3512 * which can also be traced by the function tracer.
3514 preempt_disable_notrace();
3515 preempt_latency_start(1);
3517 preempt_latency_stop(1);
3518 preempt_enable_no_resched_notrace();
3521 * Check again in case we missed a preemption opportunity
3522 * between schedule and now.
3524 } while (need_resched());
3527 #ifdef CONFIG_PREEMPT
3529 * this is the entry point to schedule() from in-kernel preemption
3530 * off of preempt_enable. Kernel preemptions off return from interrupt
3531 * occur there and call schedule directly.
3533 asmlinkage __visible void __sched notrace preempt_schedule(void)
3536 * If there is a non-zero preempt_count or interrupts are disabled,
3537 * we do not want to preempt the current task. Just return..
3539 if (likely(!preemptible()))
3542 preempt_schedule_common();
3544 NOKPROBE_SYMBOL(preempt_schedule);
3545 EXPORT_SYMBOL(preempt_schedule);
3548 * preempt_schedule_notrace - preempt_schedule called by tracing
3550 * The tracing infrastructure uses preempt_enable_notrace to prevent
3551 * recursion and tracing preempt enabling caused by the tracing
3552 * infrastructure itself. But as tracing can happen in areas coming
3553 * from userspace or just about to enter userspace, a preempt enable
3554 * can occur before user_exit() is called. This will cause the scheduler
3555 * to be called when the system is still in usermode.
3557 * To prevent this, the preempt_enable_notrace will use this function
3558 * instead of preempt_schedule() to exit user context if needed before
3559 * calling the scheduler.
3561 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3563 enum ctx_state prev_ctx;
3565 if (likely(!preemptible()))
3570 * Because the function tracer can trace preempt_count_sub()
3571 * and it also uses preempt_enable/disable_notrace(), if
3572 * NEED_RESCHED is set, the preempt_enable_notrace() called
3573 * by the function tracer will call this function again and
3574 * cause infinite recursion.
3576 * Preemption must be disabled here before the function
3577 * tracer can trace. Break up preempt_disable() into two
3578 * calls. One to disable preemption without fear of being
3579 * traced. The other to still record the preemption latency,
3580 * which can also be traced by the function tracer.
3582 preempt_disable_notrace();
3583 preempt_latency_start(1);
3585 * Needs preempt disabled in case user_exit() is traced
3586 * and the tracer calls preempt_enable_notrace() causing
3587 * an infinite recursion.
3589 prev_ctx = exception_enter();
3591 exception_exit(prev_ctx);
3593 preempt_latency_stop(1);
3594 preempt_enable_no_resched_notrace();
3595 } while (need_resched());
3597 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3599 #endif /* CONFIG_PREEMPT */
3602 * this is the entry point to schedule() from kernel preemption
3603 * off of irq context.
3604 * Note, that this is called and return with irqs disabled. This will
3605 * protect us against recursive calling from irq.
3607 asmlinkage __visible void __sched preempt_schedule_irq(void)
3609 enum ctx_state prev_state;
3611 /* Catch callers which need to be fixed */
3612 BUG_ON(preempt_count() || !irqs_disabled());
3614 prev_state = exception_enter();
3620 local_irq_disable();
3621 sched_preempt_enable_no_resched();
3622 } while (need_resched());
3624 exception_exit(prev_state);
3627 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3630 return try_to_wake_up(curr->private, mode, wake_flags);
3632 EXPORT_SYMBOL(default_wake_function);
3634 #ifdef CONFIG_RT_MUTEXES
3637 * rt_mutex_setprio - set the current priority of a task
3639 * @prio: prio value (kernel-internal form)
3641 * This function changes the 'effective' priority of a task. It does
3642 * not touch ->normal_prio like __setscheduler().
3644 * Used by the rt_mutex code to implement priority inheritance
3645 * logic. Call site only calls if the priority of the task changed.
3647 void rt_mutex_setprio(struct task_struct *p, int prio)
3649 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3650 const struct sched_class *prev_class;
3654 BUG_ON(prio > MAX_PRIO);
3656 rq = __task_rq_lock(p, &rf);
3659 * Idle task boosting is a nono in general. There is one
3660 * exception, when PREEMPT_RT and NOHZ is active:
3662 * The idle task calls get_next_timer_interrupt() and holds
3663 * the timer wheel base->lock on the CPU and another CPU wants
3664 * to access the timer (probably to cancel it). We can safely
3665 * ignore the boosting request, as the idle CPU runs this code
3666 * with interrupts disabled and will complete the lock
3667 * protected section without being interrupted. So there is no
3668 * real need to boost.
3670 if (unlikely(p == rq->idle)) {
3671 WARN_ON(p != rq->curr);
3672 WARN_ON(p->pi_blocked_on);
3676 trace_sched_pi_setprio(p, prio);
3679 if (oldprio == prio)
3680 queue_flag &= ~DEQUEUE_MOVE;
3682 prev_class = p->sched_class;
3683 queued = task_on_rq_queued(p);
3684 running = task_current(rq, p);
3686 dequeue_task(rq, p, queue_flag);
3688 put_prev_task(rq, p);
3691 * Boosting condition are:
3692 * 1. -rt task is running and holds mutex A
3693 * --> -dl task blocks on mutex A
3695 * 2. -dl task is running and holds mutex A
3696 * --> -dl task blocks on mutex A and could preempt the
3699 if (dl_prio(prio)) {
3700 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3701 if (!dl_prio(p->normal_prio) ||
3702 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3703 p->dl.dl_boosted = 1;
3704 queue_flag |= ENQUEUE_REPLENISH;
3706 p->dl.dl_boosted = 0;
3707 p->sched_class = &dl_sched_class;
3708 } else if (rt_prio(prio)) {
3709 if (dl_prio(oldprio))
3710 p->dl.dl_boosted = 0;
3712 queue_flag |= ENQUEUE_HEAD;
3713 p->sched_class = &rt_sched_class;
3715 if (dl_prio(oldprio))
3716 p->dl.dl_boosted = 0;
3717 if (rt_prio(oldprio))
3719 p->sched_class = &fair_sched_class;
3725 enqueue_task(rq, p, queue_flag);
3727 set_curr_task(rq, p);
3729 check_class_changed(rq, p, prev_class, oldprio);
3731 preempt_disable(); /* avoid rq from going away on us */
3732 __task_rq_unlock(rq, &rf);
3734 balance_callback(rq);
3739 void set_user_nice(struct task_struct *p, long nice)
3741 bool queued, running;
3742 int old_prio, delta;
3746 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3749 * We have to be careful, if called from sys_setpriority(),
3750 * the task might be in the middle of scheduling on another CPU.
3752 rq = task_rq_lock(p, &rf);
3754 * The RT priorities are set via sched_setscheduler(), but we still
3755 * allow the 'normal' nice value to be set - but as expected
3756 * it wont have any effect on scheduling until the task is
3757 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3759 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3760 p->static_prio = NICE_TO_PRIO(nice);
3763 queued = task_on_rq_queued(p);
3764 running = task_current(rq, p);
3766 dequeue_task(rq, p, DEQUEUE_SAVE);
3768 put_prev_task(rq, p);
3770 p->static_prio = NICE_TO_PRIO(nice);
3773 p->prio = effective_prio(p);
3774 delta = p->prio - old_prio;
3777 enqueue_task(rq, p, ENQUEUE_RESTORE);
3779 * If the task increased its priority or is running and
3780 * lowered its priority, then reschedule its CPU:
3782 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3786 set_curr_task(rq, p);
3788 task_rq_unlock(rq, p, &rf);
3790 EXPORT_SYMBOL(set_user_nice);
3793 * can_nice - check if a task can reduce its nice value
3797 int can_nice(const struct task_struct *p, const int nice)
3799 /* convert nice value [19,-20] to rlimit style value [1,40] */
3800 int nice_rlim = nice_to_rlimit(nice);
3802 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3803 capable(CAP_SYS_NICE));
3806 #ifdef __ARCH_WANT_SYS_NICE
3809 * sys_nice - change the priority of the current process.
3810 * @increment: priority increment
3812 * sys_setpriority is a more generic, but much slower function that
3813 * does similar things.
3815 SYSCALL_DEFINE1(nice, int, increment)
3820 * Setpriority might change our priority at the same moment.
3821 * We don't have to worry. Conceptually one call occurs first
3822 * and we have a single winner.
3824 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3825 nice = task_nice(current) + increment;
3827 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3828 if (increment < 0 && !can_nice(current, nice))
3831 retval = security_task_setnice(current, nice);
3835 set_user_nice(current, nice);
3842 * task_prio - return the priority value of a given task.
3843 * @p: the task in question.
3845 * Return: The priority value as seen by users in /proc.
3846 * RT tasks are offset by -200. Normal tasks are centered
3847 * around 0, value goes from -16 to +15.
3849 int task_prio(const struct task_struct *p)
3851 return p->prio - MAX_RT_PRIO;
3855 * idle_cpu - is a given cpu idle currently?
3856 * @cpu: the processor in question.
3858 * Return: 1 if the CPU is currently idle. 0 otherwise.
3860 int idle_cpu(int cpu)
3862 struct rq *rq = cpu_rq(cpu);
3864 if (rq->curr != rq->idle)
3871 if (!llist_empty(&rq->wake_list))
3879 * idle_task - return the idle task for a given cpu.
3880 * @cpu: the processor in question.
3882 * Return: The idle task for the cpu @cpu.
3884 struct task_struct *idle_task(int cpu)
3886 return cpu_rq(cpu)->idle;
3890 * find_process_by_pid - find a process with a matching PID value.
3891 * @pid: the pid in question.
3893 * The task of @pid, if found. %NULL otherwise.
3895 static struct task_struct *find_process_by_pid(pid_t pid)
3897 return pid ? find_task_by_vpid(pid) : current;
3901 * This function initializes the sched_dl_entity of a newly becoming
3902 * SCHED_DEADLINE task.
3904 * Only the static values are considered here, the actual runtime and the
3905 * absolute deadline will be properly calculated when the task is enqueued
3906 * for the first time with its new policy.
3909 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3911 struct sched_dl_entity *dl_se = &p->dl;
3913 dl_se->dl_runtime = attr->sched_runtime;
3914 dl_se->dl_deadline = attr->sched_deadline;
3915 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3916 dl_se->flags = attr->sched_flags;
3917 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3920 * Changing the parameters of a task is 'tricky' and we're not doing
3921 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3923 * What we SHOULD do is delay the bandwidth release until the 0-lag
3924 * point. This would include retaining the task_struct until that time
3925 * and change dl_overflow() to not immediately decrement the current
3928 * Instead we retain the current runtime/deadline and let the new
3929 * parameters take effect after the current reservation period lapses.
3930 * This is safe (albeit pessimistic) because the 0-lag point is always
3931 * before the current scheduling deadline.
3933 * We can still have temporary overloads because we do not delay the
3934 * change in bandwidth until that time; so admission control is
3935 * not on the safe side. It does however guarantee tasks will never
3936 * consume more than promised.
3941 * sched_setparam() passes in -1 for its policy, to let the functions
3942 * it calls know not to change it.
3944 #define SETPARAM_POLICY -1
3946 static void __setscheduler_params(struct task_struct *p,
3947 const struct sched_attr *attr)
3949 int policy = attr->sched_policy;
3951 if (policy == SETPARAM_POLICY)
3956 if (dl_policy(policy))
3957 __setparam_dl(p, attr);
3958 else if (fair_policy(policy))
3959 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3962 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3963 * !rt_policy. Always setting this ensures that things like
3964 * getparam()/getattr() don't report silly values for !rt tasks.
3966 p->rt_priority = attr->sched_priority;
3967 p->normal_prio = normal_prio(p);
3971 /* Actually do priority change: must hold pi & rq lock. */
3972 static void __setscheduler(struct rq *rq, struct task_struct *p,
3973 const struct sched_attr *attr, bool keep_boost)
3975 __setscheduler_params(p, attr);
3978 * Keep a potential priority boosting if called from
3979 * sched_setscheduler().
3982 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3984 p->prio = normal_prio(p);
3986 if (dl_prio(p->prio))
3987 p->sched_class = &dl_sched_class;
3988 else if (rt_prio(p->prio))
3989 p->sched_class = &rt_sched_class;
3991 p->sched_class = &fair_sched_class;
3995 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3997 struct sched_dl_entity *dl_se = &p->dl;
3999 attr->sched_priority = p->rt_priority;
4000 attr->sched_runtime = dl_se->dl_runtime;
4001 attr->sched_deadline = dl_se->dl_deadline;
4002 attr->sched_period = dl_se->dl_period;
4003 attr->sched_flags = dl_se->flags;
4007 * This function validates the new parameters of a -deadline task.
4008 * We ask for the deadline not being zero, and greater or equal
4009 * than the runtime, as well as the period of being zero or
4010 * greater than deadline. Furthermore, we have to be sure that
4011 * user parameters are above the internal resolution of 1us (we
4012 * check sched_runtime only since it is always the smaller one) and
4013 * below 2^63 ns (we have to check both sched_deadline and
4014 * sched_period, as the latter can be zero).
4017 __checkparam_dl(const struct sched_attr *attr)
4020 if (attr->sched_deadline == 0)
4024 * Since we truncate DL_SCALE bits, make sure we're at least
4027 if (attr->sched_runtime < (1ULL << DL_SCALE))
4031 * Since we use the MSB for wrap-around and sign issues, make
4032 * sure it's not set (mind that period can be equal to zero).
4034 if (attr->sched_deadline & (1ULL << 63) ||
4035 attr->sched_period & (1ULL << 63))
4038 /* runtime <= deadline <= period (if period != 0) */
4039 if ((attr->sched_period != 0 &&
4040 attr->sched_period < attr->sched_deadline) ||
4041 attr->sched_deadline < attr->sched_runtime)
4048 * check the target process has a UID that matches the current process's
4050 static bool check_same_owner(struct task_struct *p)
4052 const struct cred *cred = current_cred(), *pcred;
4056 pcred = __task_cred(p);
4057 match = (uid_eq(cred->euid, pcred->euid) ||
4058 uid_eq(cred->euid, pcred->uid));
4063 static bool dl_param_changed(struct task_struct *p,
4064 const struct sched_attr *attr)
4066 struct sched_dl_entity *dl_se = &p->dl;
4068 if (dl_se->dl_runtime != attr->sched_runtime ||
4069 dl_se->dl_deadline != attr->sched_deadline ||
4070 dl_se->dl_period != attr->sched_period ||
4071 dl_se->flags != attr->sched_flags)
4077 static int __sched_setscheduler(struct task_struct *p,
4078 const struct sched_attr *attr,
4081 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4082 MAX_RT_PRIO - 1 - attr->sched_priority;
4083 int retval, oldprio, oldpolicy = -1, queued, running;
4084 int new_effective_prio, policy = attr->sched_policy;
4085 const struct sched_class *prev_class;
4088 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4091 /* may grab non-irq protected spin_locks */
4092 BUG_ON(in_interrupt());
4094 /* double check policy once rq lock held */
4096 reset_on_fork = p->sched_reset_on_fork;
4097 policy = oldpolicy = p->policy;
4099 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4101 if (!valid_policy(policy))
4105 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4109 * Valid priorities for SCHED_FIFO and SCHED_RR are
4110 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4111 * SCHED_BATCH and SCHED_IDLE is 0.
4113 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4114 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4116 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4117 (rt_policy(policy) != (attr->sched_priority != 0)))
4121 * Allow unprivileged RT tasks to decrease priority:
4123 if (user && !capable(CAP_SYS_NICE)) {
4124 if (fair_policy(policy)) {
4125 if (attr->sched_nice < task_nice(p) &&
4126 !can_nice(p, attr->sched_nice))
4130 if (rt_policy(policy)) {
4131 unsigned long rlim_rtprio =
4132 task_rlimit(p, RLIMIT_RTPRIO);
4134 /* can't set/change the rt policy */
4135 if (policy != p->policy && !rlim_rtprio)
4138 /* can't increase priority */
4139 if (attr->sched_priority > p->rt_priority &&
4140 attr->sched_priority > rlim_rtprio)
4145 * Can't set/change SCHED_DEADLINE policy at all for now
4146 * (safest behavior); in the future we would like to allow
4147 * unprivileged DL tasks to increase their relative deadline
4148 * or reduce their runtime (both ways reducing utilization)
4150 if (dl_policy(policy))
4154 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4155 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4157 if (idle_policy(p->policy) && !idle_policy(policy)) {
4158 if (!can_nice(p, task_nice(p)))
4162 /* can't change other user's priorities */
4163 if (!check_same_owner(p))
4166 /* Normal users shall not reset the sched_reset_on_fork flag */
4167 if (p->sched_reset_on_fork && !reset_on_fork)
4172 retval = security_task_setscheduler(p);
4178 * make sure no PI-waiters arrive (or leave) while we are
4179 * changing the priority of the task:
4181 * To be able to change p->policy safely, the appropriate
4182 * runqueue lock must be held.
4184 rq = task_rq_lock(p, &rf);
4187 * Changing the policy of the stop threads its a very bad idea
4189 if (p == rq->stop) {
4190 task_rq_unlock(rq, p, &rf);
4195 * If not changing anything there's no need to proceed further,
4196 * but store a possible modification of reset_on_fork.
4198 if (unlikely(policy == p->policy)) {
4199 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4201 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4203 if (dl_policy(policy) && dl_param_changed(p, attr))
4206 p->sched_reset_on_fork = reset_on_fork;
4207 task_rq_unlock(rq, p, &rf);
4213 #ifdef CONFIG_RT_GROUP_SCHED
4215 * Do not allow realtime tasks into groups that have no runtime
4218 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4219 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4220 !task_group_is_autogroup(task_group(p))) {
4221 task_rq_unlock(rq, p, &rf);
4226 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4227 cpumask_t *span = rq->rd->span;
4230 * Don't allow tasks with an affinity mask smaller than
4231 * the entire root_domain to become SCHED_DEADLINE. We
4232 * will also fail if there's no bandwidth available.
4234 if (!cpumask_subset(span, &p->cpus_allowed) ||
4235 rq->rd->dl_bw.bw == 0) {
4236 task_rq_unlock(rq, p, &rf);
4243 /* recheck policy now with rq lock held */
4244 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4245 policy = oldpolicy = -1;
4246 task_rq_unlock(rq, p, &rf);
4251 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4252 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4255 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4256 task_rq_unlock(rq, p, &rf);
4260 p->sched_reset_on_fork = reset_on_fork;
4265 * Take priority boosted tasks into account. If the new
4266 * effective priority is unchanged, we just store the new
4267 * normal parameters and do not touch the scheduler class and
4268 * the runqueue. This will be done when the task deboost
4271 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4272 if (new_effective_prio == oldprio)
4273 queue_flags &= ~DEQUEUE_MOVE;
4276 queued = task_on_rq_queued(p);
4277 running = task_current(rq, p);
4279 dequeue_task(rq, p, queue_flags);
4281 put_prev_task(rq, p);
4283 prev_class = p->sched_class;
4284 __setscheduler(rq, p, attr, pi);
4288 * We enqueue to tail when the priority of a task is
4289 * increased (user space view).
4291 if (oldprio < p->prio)
4292 queue_flags |= ENQUEUE_HEAD;
4294 enqueue_task(rq, p, queue_flags);
4297 set_curr_task(rq, p);
4299 check_class_changed(rq, p, prev_class, oldprio);
4300 preempt_disable(); /* avoid rq from going away on us */
4301 task_rq_unlock(rq, p, &rf);
4304 rt_mutex_adjust_pi(p);
4307 * Run balance callbacks after we've adjusted the PI chain.
4309 balance_callback(rq);
4315 static int _sched_setscheduler(struct task_struct *p, int policy,
4316 const struct sched_param *param, bool check)
4318 struct sched_attr attr = {
4319 .sched_policy = policy,
4320 .sched_priority = param->sched_priority,
4321 .sched_nice = PRIO_TO_NICE(p->static_prio),
4324 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4325 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4326 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4327 policy &= ~SCHED_RESET_ON_FORK;
4328 attr.sched_policy = policy;
4331 return __sched_setscheduler(p, &attr, check, true);
4334 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4335 * @p: the task in question.
4336 * @policy: new policy.
4337 * @param: structure containing the new RT priority.
4339 * Return: 0 on success. An error code otherwise.
4341 * NOTE that the task may be already dead.
4343 int sched_setscheduler(struct task_struct *p, int policy,
4344 const struct sched_param *param)
4346 return _sched_setscheduler(p, policy, param, true);
4348 EXPORT_SYMBOL_GPL(sched_setscheduler);
4350 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4352 return __sched_setscheduler(p, attr, true, true);
4354 EXPORT_SYMBOL_GPL(sched_setattr);
4357 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4358 * @p: the task in question.
4359 * @policy: new policy.
4360 * @param: structure containing the new RT priority.
4362 * Just like sched_setscheduler, only don't bother checking if the
4363 * current context has permission. For example, this is needed in
4364 * stop_machine(): we create temporary high priority worker threads,
4365 * but our caller might not have that capability.
4367 * Return: 0 on success. An error code otherwise.
4369 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4370 const struct sched_param *param)
4372 return _sched_setscheduler(p, policy, param, false);
4374 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4377 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4379 struct sched_param lparam;
4380 struct task_struct *p;
4383 if (!param || pid < 0)
4385 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4390 p = find_process_by_pid(pid);
4392 retval = sched_setscheduler(p, policy, &lparam);
4399 * Mimics kernel/events/core.c perf_copy_attr().
4401 static int sched_copy_attr(struct sched_attr __user *uattr,
4402 struct sched_attr *attr)
4407 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4411 * zero the full structure, so that a short copy will be nice.
4413 memset(attr, 0, sizeof(*attr));
4415 ret = get_user(size, &uattr->size);
4419 if (size > PAGE_SIZE) /* silly large */
4422 if (!size) /* abi compat */
4423 size = SCHED_ATTR_SIZE_VER0;
4425 if (size < SCHED_ATTR_SIZE_VER0)
4429 * If we're handed a bigger struct than we know of,
4430 * ensure all the unknown bits are 0 - i.e. new
4431 * user-space does not rely on any kernel feature
4432 * extensions we dont know about yet.
4434 if (size > sizeof(*attr)) {
4435 unsigned char __user *addr;
4436 unsigned char __user *end;
4439 addr = (void __user *)uattr + sizeof(*attr);
4440 end = (void __user *)uattr + size;
4442 for (; addr < end; addr++) {
4443 ret = get_user(val, addr);
4449 size = sizeof(*attr);
4452 ret = copy_from_user(attr, uattr, size);
4457 * XXX: do we want to be lenient like existing syscalls; or do we want
4458 * to be strict and return an error on out-of-bounds values?
4460 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4465 put_user(sizeof(*attr), &uattr->size);
4470 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4471 * @pid: the pid in question.
4472 * @policy: new policy.
4473 * @param: structure containing the new RT priority.
4475 * Return: 0 on success. An error code otherwise.
4477 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4478 struct sched_param __user *, param)
4480 /* negative values for policy are not valid */
4484 return do_sched_setscheduler(pid, policy, param);
4488 * sys_sched_setparam - set/change the RT priority of a thread
4489 * @pid: the pid in question.
4490 * @param: structure containing the new RT priority.
4492 * Return: 0 on success. An error code otherwise.
4494 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4496 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4500 * sys_sched_setattr - same as above, but with extended sched_attr
4501 * @pid: the pid in question.
4502 * @uattr: structure containing the extended parameters.
4503 * @flags: for future extension.
4505 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4506 unsigned int, flags)
4508 struct sched_attr attr;
4509 struct task_struct *p;
4512 if (!uattr || pid < 0 || flags)
4515 retval = sched_copy_attr(uattr, &attr);
4519 if ((int)attr.sched_policy < 0)
4524 p = find_process_by_pid(pid);
4526 retval = sched_setattr(p, &attr);
4533 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4534 * @pid: the pid in question.
4536 * Return: On success, the policy of the thread. Otherwise, a negative error
4539 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4541 struct task_struct *p;
4549 p = find_process_by_pid(pid);
4551 retval = security_task_getscheduler(p);
4554 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4561 * sys_sched_getparam - get the RT priority of a thread
4562 * @pid: the pid in question.
4563 * @param: structure containing the RT priority.
4565 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4568 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4570 struct sched_param lp = { .sched_priority = 0 };
4571 struct task_struct *p;
4574 if (!param || pid < 0)
4578 p = find_process_by_pid(pid);
4583 retval = security_task_getscheduler(p);
4587 if (task_has_rt_policy(p))
4588 lp.sched_priority = p->rt_priority;
4592 * This one might sleep, we cannot do it with a spinlock held ...
4594 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4603 static int sched_read_attr(struct sched_attr __user *uattr,
4604 struct sched_attr *attr,
4609 if (!access_ok(VERIFY_WRITE, uattr, usize))
4613 * If we're handed a smaller struct than we know of,
4614 * ensure all the unknown bits are 0 - i.e. old
4615 * user-space does not get uncomplete information.
4617 if (usize < sizeof(*attr)) {
4618 unsigned char *addr;
4621 addr = (void *)attr + usize;
4622 end = (void *)attr + sizeof(*attr);
4624 for (; addr < end; addr++) {
4632 ret = copy_to_user(uattr, attr, attr->size);
4640 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4641 * @pid: the pid in question.
4642 * @uattr: structure containing the extended parameters.
4643 * @size: sizeof(attr) for fwd/bwd comp.
4644 * @flags: for future extension.
4646 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4647 unsigned int, size, unsigned int, flags)
4649 struct sched_attr attr = {
4650 .size = sizeof(struct sched_attr),
4652 struct task_struct *p;
4655 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4656 size < SCHED_ATTR_SIZE_VER0 || flags)
4660 p = find_process_by_pid(pid);
4665 retval = security_task_getscheduler(p);
4669 attr.sched_policy = p->policy;
4670 if (p->sched_reset_on_fork)
4671 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4672 if (task_has_dl_policy(p))
4673 __getparam_dl(p, &attr);
4674 else if (task_has_rt_policy(p))
4675 attr.sched_priority = p->rt_priority;
4677 attr.sched_nice = task_nice(p);
4681 retval = sched_read_attr(uattr, &attr, size);
4689 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4691 cpumask_var_t cpus_allowed, new_mask;
4692 struct task_struct *p;
4697 p = find_process_by_pid(pid);
4703 /* Prevent p going away */
4707 if (p->flags & PF_NO_SETAFFINITY) {
4711 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4715 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4717 goto out_free_cpus_allowed;
4720 if (!check_same_owner(p)) {
4722 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4724 goto out_free_new_mask;
4729 retval = security_task_setscheduler(p);
4731 goto out_free_new_mask;
4734 cpuset_cpus_allowed(p, cpus_allowed);
4735 cpumask_and(new_mask, in_mask, cpus_allowed);
4738 * Since bandwidth control happens on root_domain basis,
4739 * if admission test is enabled, we only admit -deadline
4740 * tasks allowed to run on all the CPUs in the task's
4744 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4746 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4749 goto out_free_new_mask;
4755 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4758 cpuset_cpus_allowed(p, cpus_allowed);
4759 if (!cpumask_subset(new_mask, cpus_allowed)) {
4761 * We must have raced with a concurrent cpuset
4762 * update. Just reset the cpus_allowed to the
4763 * cpuset's cpus_allowed
4765 cpumask_copy(new_mask, cpus_allowed);
4770 free_cpumask_var(new_mask);
4771 out_free_cpus_allowed:
4772 free_cpumask_var(cpus_allowed);
4778 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4779 struct cpumask *new_mask)
4781 if (len < cpumask_size())
4782 cpumask_clear(new_mask);
4783 else if (len > cpumask_size())
4784 len = cpumask_size();
4786 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4790 * sys_sched_setaffinity - set the cpu affinity of a process
4791 * @pid: pid of the process
4792 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4793 * @user_mask_ptr: user-space pointer to the new cpu mask
4795 * Return: 0 on success. An error code otherwise.
4797 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4798 unsigned long __user *, user_mask_ptr)
4800 cpumask_var_t new_mask;
4803 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4806 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4808 retval = sched_setaffinity(pid, new_mask);
4809 free_cpumask_var(new_mask);
4813 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4815 struct task_struct *p;
4816 unsigned long flags;
4822 p = find_process_by_pid(pid);
4826 retval = security_task_getscheduler(p);
4830 raw_spin_lock_irqsave(&p->pi_lock, flags);
4831 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4832 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4841 * sys_sched_getaffinity - get the cpu affinity of a process
4842 * @pid: pid of the process
4843 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4844 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4846 * Return: size of CPU mask copied to user_mask_ptr on success. An
4847 * error code otherwise.
4849 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4850 unsigned long __user *, user_mask_ptr)
4855 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4857 if (len & (sizeof(unsigned long)-1))
4860 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4863 ret = sched_getaffinity(pid, mask);
4865 size_t retlen = min_t(size_t, len, cpumask_size());
4867 if (copy_to_user(user_mask_ptr, mask, retlen))
4872 free_cpumask_var(mask);
4878 * sys_sched_yield - yield the current processor to other threads.
4880 * This function yields the current CPU to other tasks. If there are no
4881 * other threads running on this CPU then this function will return.
4885 SYSCALL_DEFINE0(sched_yield)
4887 struct rq *rq = this_rq_lock();
4889 schedstat_inc(rq->yld_count);
4890 current->sched_class->yield_task(rq);
4893 * Since we are going to call schedule() anyway, there's
4894 * no need to preempt or enable interrupts:
4896 __release(rq->lock);
4897 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4898 do_raw_spin_unlock(&rq->lock);
4899 sched_preempt_enable_no_resched();
4906 #ifndef CONFIG_PREEMPT
4907 int __sched _cond_resched(void)
4909 if (should_resched(0)) {
4910 preempt_schedule_common();
4915 EXPORT_SYMBOL(_cond_resched);
4919 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4920 * call schedule, and on return reacquire the lock.
4922 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4923 * operations here to prevent schedule() from being called twice (once via
4924 * spin_unlock(), once by hand).
4926 int __cond_resched_lock(spinlock_t *lock)
4928 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4931 lockdep_assert_held(lock);
4933 if (spin_needbreak(lock) || resched) {
4936 preempt_schedule_common();
4944 EXPORT_SYMBOL(__cond_resched_lock);
4946 int __sched __cond_resched_softirq(void)
4948 BUG_ON(!in_softirq());
4950 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4952 preempt_schedule_common();
4958 EXPORT_SYMBOL(__cond_resched_softirq);
4961 * yield - yield the current processor to other threads.
4963 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4965 * The scheduler is at all times free to pick the calling task as the most
4966 * eligible task to run, if removing the yield() call from your code breaks
4967 * it, its already broken.
4969 * Typical broken usage is:
4974 * where one assumes that yield() will let 'the other' process run that will
4975 * make event true. If the current task is a SCHED_FIFO task that will never
4976 * happen. Never use yield() as a progress guarantee!!
4978 * If you want to use yield() to wait for something, use wait_event().
4979 * If you want to use yield() to be 'nice' for others, use cond_resched().
4980 * If you still want to use yield(), do not!
4982 void __sched yield(void)
4984 set_current_state(TASK_RUNNING);
4987 EXPORT_SYMBOL(yield);
4990 * yield_to - yield the current processor to another thread in
4991 * your thread group, or accelerate that thread toward the
4992 * processor it's on.
4994 * @preempt: whether task preemption is allowed or not
4996 * It's the caller's job to ensure that the target task struct
4997 * can't go away on us before we can do any checks.
5000 * true (>0) if we indeed boosted the target task.
5001 * false (0) if we failed to boost the target.
5002 * -ESRCH if there's no task to yield to.
5004 int __sched yield_to(struct task_struct *p, bool preempt)
5006 struct task_struct *curr = current;
5007 struct rq *rq, *p_rq;
5008 unsigned long flags;
5011 local_irq_save(flags);
5017 * If we're the only runnable task on the rq and target rq also
5018 * has only one task, there's absolutely no point in yielding.
5020 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5025 double_rq_lock(rq, p_rq);
5026 if (task_rq(p) != p_rq) {
5027 double_rq_unlock(rq, p_rq);
5031 if (!curr->sched_class->yield_to_task)
5034 if (curr->sched_class != p->sched_class)
5037 if (task_running(p_rq, p) || p->state)
5040 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5042 schedstat_inc(rq->yld_count);
5044 * Make p's CPU reschedule; pick_next_entity takes care of
5047 if (preempt && rq != p_rq)
5052 double_rq_unlock(rq, p_rq);
5054 local_irq_restore(flags);
5061 EXPORT_SYMBOL_GPL(yield_to);
5064 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5065 * that process accounting knows that this is a task in IO wait state.
5067 long __sched io_schedule_timeout(long timeout)
5069 int old_iowait = current->in_iowait;
5073 current->in_iowait = 1;
5074 blk_schedule_flush_plug(current);
5076 delayacct_blkio_start();
5078 atomic_inc(&rq->nr_iowait);
5079 ret = schedule_timeout(timeout);
5080 current->in_iowait = old_iowait;
5081 atomic_dec(&rq->nr_iowait);
5082 delayacct_blkio_end();
5086 EXPORT_SYMBOL(io_schedule_timeout);
5089 * sys_sched_get_priority_max - return maximum RT priority.
5090 * @policy: scheduling class.
5092 * Return: On success, this syscall returns the maximum
5093 * rt_priority that can be used by a given scheduling class.
5094 * On failure, a negative error code is returned.
5096 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5103 ret = MAX_USER_RT_PRIO-1;
5105 case SCHED_DEADLINE:
5116 * sys_sched_get_priority_min - return minimum RT priority.
5117 * @policy: scheduling class.
5119 * Return: On success, this syscall returns the minimum
5120 * rt_priority that can be used by a given scheduling class.
5121 * On failure, a negative error code is returned.
5123 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5132 case SCHED_DEADLINE:
5142 * sys_sched_rr_get_interval - return the default timeslice of a process.
5143 * @pid: pid of the process.
5144 * @interval: userspace pointer to the timeslice value.
5146 * this syscall writes the default timeslice value of a given process
5147 * into the user-space timespec buffer. A value of '0' means infinity.
5149 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5152 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5153 struct timespec __user *, interval)
5155 struct task_struct *p;
5156 unsigned int time_slice;
5167 p = find_process_by_pid(pid);
5171 retval = security_task_getscheduler(p);
5175 rq = task_rq_lock(p, &rf);
5177 if (p->sched_class->get_rr_interval)
5178 time_slice = p->sched_class->get_rr_interval(rq, p);
5179 task_rq_unlock(rq, p, &rf);
5182 jiffies_to_timespec(time_slice, &t);
5183 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5191 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5193 void sched_show_task(struct task_struct *p)
5195 unsigned long free = 0;
5197 unsigned long state = p->state;
5199 if (!try_get_task_stack(p))
5202 state = __ffs(state) + 1;
5203 printk(KERN_INFO "%-15.15s %c", p->comm,
5204 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5205 if (state == TASK_RUNNING)
5206 printk(KERN_CONT " running task ");
5207 #ifdef CONFIG_DEBUG_STACK_USAGE
5208 free = stack_not_used(p);
5213 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5215 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5216 task_pid_nr(p), ppid,
5217 (unsigned long)task_thread_info(p)->flags);
5219 print_worker_info(KERN_INFO, p);
5220 show_stack(p, NULL);
5224 void show_state_filter(unsigned long state_filter)
5226 struct task_struct *g, *p;
5228 #if BITS_PER_LONG == 32
5230 " task PC stack pid father\n");
5233 " task PC stack pid father\n");
5236 for_each_process_thread(g, p) {
5238 * reset the NMI-timeout, listing all files on a slow
5239 * console might take a lot of time:
5240 * Also, reset softlockup watchdogs on all CPUs, because
5241 * another CPU might be blocked waiting for us to process
5244 touch_nmi_watchdog();
5245 touch_all_softlockup_watchdogs();
5246 if (!state_filter || (p->state & state_filter))
5250 #ifdef CONFIG_SCHED_DEBUG
5252 sysrq_sched_debug_show();
5256 * Only show locks if all tasks are dumped:
5259 debug_show_all_locks();
5262 void init_idle_bootup_task(struct task_struct *idle)
5264 idle->sched_class = &idle_sched_class;
5268 * init_idle - set up an idle thread for a given CPU
5269 * @idle: task in question
5270 * @cpu: cpu the idle task belongs to
5272 * NOTE: this function does not set the idle thread's NEED_RESCHED
5273 * flag, to make booting more robust.
5275 void init_idle(struct task_struct *idle, int cpu)
5277 struct rq *rq = cpu_rq(cpu);
5278 unsigned long flags;
5280 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5281 raw_spin_lock(&rq->lock);
5283 __sched_fork(0, idle);
5284 idle->state = TASK_RUNNING;
5285 idle->se.exec_start = sched_clock();
5286 idle->flags |= PF_IDLE;
5288 kasan_unpoison_task_stack(idle);
5292 * Its possible that init_idle() gets called multiple times on a task,
5293 * in that case do_set_cpus_allowed() will not do the right thing.
5295 * And since this is boot we can forgo the serialization.
5297 set_cpus_allowed_common(idle, cpumask_of(cpu));
5300 * We're having a chicken and egg problem, even though we are
5301 * holding rq->lock, the cpu isn't yet set to this cpu so the
5302 * lockdep check in task_group() will fail.
5304 * Similar case to sched_fork(). / Alternatively we could
5305 * use task_rq_lock() here and obtain the other rq->lock.
5310 __set_task_cpu(idle, cpu);
5313 rq->curr = rq->idle = idle;
5314 idle->on_rq = TASK_ON_RQ_QUEUED;
5318 raw_spin_unlock(&rq->lock);
5319 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5321 /* Set the preempt count _outside_ the spinlocks! */
5322 init_idle_preempt_count(idle, cpu);
5325 * The idle tasks have their own, simple scheduling class:
5327 idle->sched_class = &idle_sched_class;
5328 ftrace_graph_init_idle_task(idle, cpu);
5329 vtime_init_idle(idle, cpu);
5331 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5335 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5336 const struct cpumask *trial)
5338 int ret = 1, trial_cpus;
5339 struct dl_bw *cur_dl_b;
5340 unsigned long flags;
5342 if (!cpumask_weight(cur))
5345 rcu_read_lock_sched();
5346 cur_dl_b = dl_bw_of(cpumask_any(cur));
5347 trial_cpus = cpumask_weight(trial);
5349 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5350 if (cur_dl_b->bw != -1 &&
5351 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5353 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5354 rcu_read_unlock_sched();
5359 int task_can_attach(struct task_struct *p,
5360 const struct cpumask *cs_cpus_allowed)
5365 * Kthreads which disallow setaffinity shouldn't be moved
5366 * to a new cpuset; we don't want to change their cpu
5367 * affinity and isolating such threads by their set of
5368 * allowed nodes is unnecessary. Thus, cpusets are not
5369 * applicable for such threads. This prevents checking for
5370 * success of set_cpus_allowed_ptr() on all attached tasks
5371 * before cpus_allowed may be changed.
5373 if (p->flags & PF_NO_SETAFFINITY) {
5379 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5381 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5386 unsigned long flags;
5388 rcu_read_lock_sched();
5389 dl_b = dl_bw_of(dest_cpu);
5390 raw_spin_lock_irqsave(&dl_b->lock, flags);
5391 cpus = dl_bw_cpus(dest_cpu);
5392 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5397 * We reserve space for this task in the destination
5398 * root_domain, as we can't fail after this point.
5399 * We will free resources in the source root_domain
5400 * later on (see set_cpus_allowed_dl()).
5402 __dl_add(dl_b, p->dl.dl_bw);
5404 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5405 rcu_read_unlock_sched();
5415 static bool sched_smp_initialized __read_mostly;
5417 #ifdef CONFIG_NUMA_BALANCING
5418 /* Migrate current task p to target_cpu */
5419 int migrate_task_to(struct task_struct *p, int target_cpu)
5421 struct migration_arg arg = { p, target_cpu };
5422 int curr_cpu = task_cpu(p);
5424 if (curr_cpu == target_cpu)
5427 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5430 /* TODO: This is not properly updating schedstats */
5432 trace_sched_move_numa(p, curr_cpu, target_cpu);
5433 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5437 * Requeue a task on a given node and accurately track the number of NUMA
5438 * tasks on the runqueues
5440 void sched_setnuma(struct task_struct *p, int nid)
5442 bool queued, running;
5446 rq = task_rq_lock(p, &rf);
5447 queued = task_on_rq_queued(p);
5448 running = task_current(rq, p);
5451 dequeue_task(rq, p, DEQUEUE_SAVE);
5453 put_prev_task(rq, p);
5455 p->numa_preferred_nid = nid;
5458 enqueue_task(rq, p, ENQUEUE_RESTORE);
5460 set_curr_task(rq, p);
5461 task_rq_unlock(rq, p, &rf);
5463 #endif /* CONFIG_NUMA_BALANCING */
5465 #ifdef CONFIG_HOTPLUG_CPU
5467 * Ensures that the idle task is using init_mm right before its cpu goes
5470 void idle_task_exit(void)
5472 struct mm_struct *mm = current->active_mm;
5474 BUG_ON(cpu_online(smp_processor_id()));
5476 if (mm != &init_mm) {
5477 switch_mm_irqs_off(mm, &init_mm, current);
5478 finish_arch_post_lock_switch();
5484 * Since this CPU is going 'away' for a while, fold any nr_active delta
5485 * we might have. Assumes we're called after migrate_tasks() so that the
5486 * nr_active count is stable. We need to take the teardown thread which
5487 * is calling this into account, so we hand in adjust = 1 to the load
5490 * Also see the comment "Global load-average calculations".
5492 static void calc_load_migrate(struct rq *rq)
5494 long delta = calc_load_fold_active(rq, 1);
5496 atomic_long_add(delta, &calc_load_tasks);
5499 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5503 static const struct sched_class fake_sched_class = {
5504 .put_prev_task = put_prev_task_fake,
5507 static struct task_struct fake_task = {
5509 * Avoid pull_{rt,dl}_task()
5511 .prio = MAX_PRIO + 1,
5512 .sched_class = &fake_sched_class,
5516 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5517 * try_to_wake_up()->select_task_rq().
5519 * Called with rq->lock held even though we'er in stop_machine() and
5520 * there's no concurrency possible, we hold the required locks anyway
5521 * because of lock validation efforts.
5523 static void migrate_tasks(struct rq *dead_rq)
5525 struct rq *rq = dead_rq;
5526 struct task_struct *next, *stop = rq->stop;
5531 * Fudge the rq selection such that the below task selection loop
5532 * doesn't get stuck on the currently eligible stop task.
5534 * We're currently inside stop_machine() and the rq is either stuck
5535 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5536 * either way we should never end up calling schedule() until we're
5542 * put_prev_task() and pick_next_task() sched
5543 * class method both need to have an up-to-date
5544 * value of rq->clock[_task]
5546 update_rq_clock(rq);
5550 * There's this thread running, bail when that's the only
5553 if (rq->nr_running == 1)
5557 * pick_next_task assumes pinned rq->lock.
5559 rq_pin_lock(rq, &rf);
5560 next = pick_next_task(rq, &fake_task, &rf);
5562 next->sched_class->put_prev_task(rq, next);
5565 * Rules for changing task_struct::cpus_allowed are holding
5566 * both pi_lock and rq->lock, such that holding either
5567 * stabilizes the mask.
5569 * Drop rq->lock is not quite as disastrous as it usually is
5570 * because !cpu_active at this point, which means load-balance
5571 * will not interfere. Also, stop-machine.
5573 rq_unpin_lock(rq, &rf);
5574 raw_spin_unlock(&rq->lock);
5575 raw_spin_lock(&next->pi_lock);
5576 raw_spin_lock(&rq->lock);
5579 * Since we're inside stop-machine, _nothing_ should have
5580 * changed the task, WARN if weird stuff happened, because in
5581 * that case the above rq->lock drop is a fail too.
5583 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5584 raw_spin_unlock(&next->pi_lock);
5588 /* Find suitable destination for @next, with force if needed. */
5589 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5591 rq = __migrate_task(rq, next, dest_cpu);
5592 if (rq != dead_rq) {
5593 raw_spin_unlock(&rq->lock);
5595 raw_spin_lock(&rq->lock);
5597 raw_spin_unlock(&next->pi_lock);
5602 #endif /* CONFIG_HOTPLUG_CPU */
5604 static void set_rq_online(struct rq *rq)
5607 const struct sched_class *class;
5609 cpumask_set_cpu(rq->cpu, rq->rd->online);
5612 for_each_class(class) {
5613 if (class->rq_online)
5614 class->rq_online(rq);
5619 static void set_rq_offline(struct rq *rq)
5622 const struct sched_class *class;
5624 for_each_class(class) {
5625 if (class->rq_offline)
5626 class->rq_offline(rq);
5629 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5634 static void set_cpu_rq_start_time(unsigned int cpu)
5636 struct rq *rq = cpu_rq(cpu);
5638 rq->age_stamp = sched_clock_cpu(cpu);
5641 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5643 #ifdef CONFIG_SCHED_DEBUG
5645 static __read_mostly int sched_debug_enabled;
5647 static int __init sched_debug_setup(char *str)
5649 sched_debug_enabled = 1;
5653 early_param("sched_debug", sched_debug_setup);
5655 static inline bool sched_debug(void)
5657 return sched_debug_enabled;
5660 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5661 struct cpumask *groupmask)
5663 struct sched_group *group = sd->groups;
5665 cpumask_clear(groupmask);
5667 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5669 if (!(sd->flags & SD_LOAD_BALANCE)) {
5670 printk("does not load-balance\n");
5672 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5677 printk(KERN_CONT "span %*pbl level %s\n",
5678 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5680 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5681 printk(KERN_ERR "ERROR: domain->span does not contain "
5684 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5685 printk(KERN_ERR "ERROR: domain->groups does not contain"
5689 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5693 printk(KERN_ERR "ERROR: group is NULL\n");
5697 if (!cpumask_weight(sched_group_cpus(group))) {
5698 printk(KERN_CONT "\n");
5699 printk(KERN_ERR "ERROR: empty group\n");
5703 if (!(sd->flags & SD_OVERLAP) &&
5704 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5705 printk(KERN_CONT "\n");
5706 printk(KERN_ERR "ERROR: repeated CPUs\n");
5710 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5712 printk(KERN_CONT " %*pbl",
5713 cpumask_pr_args(sched_group_cpus(group)));
5714 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5715 printk(KERN_CONT " (cpu_capacity = %lu)",
5716 group->sgc->capacity);
5719 group = group->next;
5720 } while (group != sd->groups);
5721 printk(KERN_CONT "\n");
5723 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5724 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5727 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5728 printk(KERN_ERR "ERROR: parent span is not a superset "
5729 "of domain->span\n");
5733 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5737 if (!sched_debug_enabled)
5741 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5745 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5748 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5756 #else /* !CONFIG_SCHED_DEBUG */
5758 # define sched_debug_enabled 0
5759 # define sched_domain_debug(sd, cpu) do { } while (0)
5760 static inline bool sched_debug(void)
5764 #endif /* CONFIG_SCHED_DEBUG */
5766 static int sd_degenerate(struct sched_domain *sd)
5768 if (cpumask_weight(sched_domain_span(sd)) == 1)
5771 /* Following flags need at least 2 groups */
5772 if (sd->flags & (SD_LOAD_BALANCE |
5773 SD_BALANCE_NEWIDLE |
5776 SD_SHARE_CPUCAPACITY |
5777 SD_ASYM_CPUCAPACITY |
5778 SD_SHARE_PKG_RESOURCES |
5779 SD_SHARE_POWERDOMAIN)) {
5780 if (sd->groups != sd->groups->next)
5784 /* Following flags don't use groups */
5785 if (sd->flags & (SD_WAKE_AFFINE))
5792 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5794 unsigned long cflags = sd->flags, pflags = parent->flags;
5796 if (sd_degenerate(parent))
5799 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5802 /* Flags needing groups don't count if only 1 group in parent */
5803 if (parent->groups == parent->groups->next) {
5804 pflags &= ~(SD_LOAD_BALANCE |
5805 SD_BALANCE_NEWIDLE |
5808 SD_ASYM_CPUCAPACITY |
5809 SD_SHARE_CPUCAPACITY |
5810 SD_SHARE_PKG_RESOURCES |
5812 SD_SHARE_POWERDOMAIN);
5813 if (nr_node_ids == 1)
5814 pflags &= ~SD_SERIALIZE;
5816 if (~cflags & pflags)
5822 static void free_rootdomain(struct rcu_head *rcu)
5824 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5826 cpupri_cleanup(&rd->cpupri);
5827 cpudl_cleanup(&rd->cpudl);
5828 free_cpumask_var(rd->dlo_mask);
5829 free_cpumask_var(rd->rto_mask);
5830 free_cpumask_var(rd->online);
5831 free_cpumask_var(rd->span);
5835 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5837 struct root_domain *old_rd = NULL;
5838 unsigned long flags;
5840 raw_spin_lock_irqsave(&rq->lock, flags);
5845 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5848 cpumask_clear_cpu(rq->cpu, old_rd->span);
5851 * If we dont want to free the old_rd yet then
5852 * set old_rd to NULL to skip the freeing later
5855 if (!atomic_dec_and_test(&old_rd->refcount))
5859 atomic_inc(&rd->refcount);
5862 cpumask_set_cpu(rq->cpu, rd->span);
5863 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5866 raw_spin_unlock_irqrestore(&rq->lock, flags);
5869 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5872 static int init_rootdomain(struct root_domain *rd)
5874 memset(rd, 0, sizeof(*rd));
5876 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5878 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5880 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5882 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5885 init_dl_bw(&rd->dl_bw);
5886 if (cpudl_init(&rd->cpudl) != 0)
5889 if (cpupri_init(&rd->cpupri) != 0)
5894 free_cpumask_var(rd->rto_mask);
5896 free_cpumask_var(rd->dlo_mask);
5898 free_cpumask_var(rd->online);
5900 free_cpumask_var(rd->span);
5906 * By default the system creates a single root-domain with all cpus as
5907 * members (mimicking the global state we have today).
5909 struct root_domain def_root_domain;
5911 static void init_defrootdomain(void)
5913 init_rootdomain(&def_root_domain);
5915 atomic_set(&def_root_domain.refcount, 1);
5918 static struct root_domain *alloc_rootdomain(void)
5920 struct root_domain *rd;
5922 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5926 if (init_rootdomain(rd) != 0) {
5934 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5936 struct sched_group *tmp, *first;
5945 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5950 } while (sg != first);
5953 static void destroy_sched_domain(struct sched_domain *sd)
5956 * If its an overlapping domain it has private groups, iterate and
5959 if (sd->flags & SD_OVERLAP) {
5960 free_sched_groups(sd->groups, 1);
5961 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5962 kfree(sd->groups->sgc);
5965 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
5970 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
5972 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5975 struct sched_domain *parent = sd->parent;
5976 destroy_sched_domain(sd);
5981 static void destroy_sched_domains(struct sched_domain *sd)
5984 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
5988 * Keep a special pointer to the highest sched_domain that has
5989 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5990 * allows us to avoid some pointer chasing select_idle_sibling().
5992 * Also keep a unique ID per domain (we use the first cpu number in
5993 * the cpumask of the domain), this allows us to quickly tell if
5994 * two cpus are in the same cache domain, see cpus_share_cache().
5996 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5997 DEFINE_PER_CPU(int, sd_llc_size);
5998 DEFINE_PER_CPU(int, sd_llc_id);
5999 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
6000 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6001 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6003 static void update_top_cache_domain(int cpu)
6005 struct sched_domain_shared *sds = NULL;
6006 struct sched_domain *sd;
6010 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6012 id = cpumask_first(sched_domain_span(sd));
6013 size = cpumask_weight(sched_domain_span(sd));
6017 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6018 per_cpu(sd_llc_size, cpu) = size;
6019 per_cpu(sd_llc_id, cpu) = id;
6020 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
6022 sd = lowest_flag_domain(cpu, SD_NUMA);
6023 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6025 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6026 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6030 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6031 * hold the hotplug lock.
6034 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6036 struct rq *rq = cpu_rq(cpu);
6037 struct sched_domain *tmp;
6039 /* Remove the sched domains which do not contribute to scheduling. */
6040 for (tmp = sd; tmp; ) {
6041 struct sched_domain *parent = tmp->parent;
6045 if (sd_parent_degenerate(tmp, parent)) {
6046 tmp->parent = parent->parent;
6048 parent->parent->child = tmp;
6050 * Transfer SD_PREFER_SIBLING down in case of a
6051 * degenerate parent; the spans match for this
6052 * so the property transfers.
6054 if (parent->flags & SD_PREFER_SIBLING)
6055 tmp->flags |= SD_PREFER_SIBLING;
6056 destroy_sched_domain(parent);
6061 if (sd && sd_degenerate(sd)) {
6064 destroy_sched_domain(tmp);
6069 sched_domain_debug(sd, cpu);
6071 rq_attach_root(rq, rd);
6073 rcu_assign_pointer(rq->sd, sd);
6074 destroy_sched_domains(tmp);
6076 update_top_cache_domain(cpu);
6079 /* Setup the mask of cpus configured for isolated domains */
6080 static int __init isolated_cpu_setup(char *str)
6084 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6085 ret = cpulist_parse(str, cpu_isolated_map);
6087 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6092 __setup("isolcpus=", isolated_cpu_setup);
6095 struct sched_domain ** __percpu sd;
6096 struct root_domain *rd;
6107 * Build an iteration mask that can exclude certain CPUs from the upwards
6110 * Asymmetric node setups can result in situations where the domain tree is of
6111 * unequal depth, make sure to skip domains that already cover the entire
6114 * In that case build_sched_domains() will have terminated the iteration early
6115 * and our sibling sd spans will be empty. Domains should always include the
6116 * cpu they're built on, so check that.
6119 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6121 const struct cpumask *span = sched_domain_span(sd);
6122 struct sd_data *sdd = sd->private;
6123 struct sched_domain *sibling;
6126 for_each_cpu(i, span) {
6127 sibling = *per_cpu_ptr(sdd->sd, i);
6128 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6131 cpumask_set_cpu(i, sched_group_mask(sg));
6136 * Return the canonical balance cpu for this group, this is the first cpu
6137 * of this group that's also in the iteration mask.
6139 int group_balance_cpu(struct sched_group *sg)
6141 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6145 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6147 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6148 const struct cpumask *span = sched_domain_span(sd);
6149 struct cpumask *covered = sched_domains_tmpmask;
6150 struct sd_data *sdd = sd->private;
6151 struct sched_domain *sibling;
6154 cpumask_clear(covered);
6156 for_each_cpu(i, span) {
6157 struct cpumask *sg_span;
6159 if (cpumask_test_cpu(i, covered))
6162 sibling = *per_cpu_ptr(sdd->sd, i);
6164 /* See the comment near build_group_mask(). */
6165 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6168 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6169 GFP_KERNEL, cpu_to_node(cpu));
6174 sg_span = sched_group_cpus(sg);
6176 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6178 cpumask_set_cpu(i, sg_span);
6180 cpumask_or(covered, covered, sg_span);
6182 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6183 if (atomic_inc_return(&sg->sgc->ref) == 1)
6184 build_group_mask(sd, sg);
6187 * Initialize sgc->capacity such that even if we mess up the
6188 * domains and no possible iteration will get us here, we won't
6191 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6192 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6195 * Make sure the first group of this domain contains the
6196 * canonical balance cpu. Otherwise the sched_domain iteration
6197 * breaks. See update_sg_lb_stats().
6199 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6200 group_balance_cpu(sg) == cpu)
6210 sd->groups = groups;
6215 free_sched_groups(first, 0);
6220 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6222 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6223 struct sched_domain *child = sd->child;
6226 cpu = cpumask_first(sched_domain_span(child));
6229 *sg = *per_cpu_ptr(sdd->sg, cpu);
6230 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6231 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6238 * build_sched_groups will build a circular linked list of the groups
6239 * covered by the given span, and will set each group's ->cpumask correctly,
6240 * and ->cpu_capacity to 0.
6242 * Assumes the sched_domain tree is fully constructed
6245 build_sched_groups(struct sched_domain *sd, int cpu)
6247 struct sched_group *first = NULL, *last = NULL;
6248 struct sd_data *sdd = sd->private;
6249 const struct cpumask *span = sched_domain_span(sd);
6250 struct cpumask *covered;
6253 get_group(cpu, sdd, &sd->groups);
6254 atomic_inc(&sd->groups->ref);
6256 if (cpu != cpumask_first(span))
6259 lockdep_assert_held(&sched_domains_mutex);
6260 covered = sched_domains_tmpmask;
6262 cpumask_clear(covered);
6264 for_each_cpu(i, span) {
6265 struct sched_group *sg;
6268 if (cpumask_test_cpu(i, covered))
6271 group = get_group(i, sdd, &sg);
6272 cpumask_setall(sched_group_mask(sg));
6274 for_each_cpu(j, span) {
6275 if (get_group(j, sdd, NULL) != group)
6278 cpumask_set_cpu(j, covered);
6279 cpumask_set_cpu(j, sched_group_cpus(sg));
6294 * Initialize sched groups cpu_capacity.
6296 * cpu_capacity indicates the capacity of sched group, which is used while
6297 * distributing the load between different sched groups in a sched domain.
6298 * Typically cpu_capacity for all the groups in a sched domain will be same
6299 * unless there are asymmetries in the topology. If there are asymmetries,
6300 * group having more cpu_capacity will pickup more load compared to the
6301 * group having less cpu_capacity.
6303 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6305 struct sched_group *sg = sd->groups;
6310 int cpu, max_cpu = -1;
6312 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6314 if (!(sd->flags & SD_ASYM_PACKING))
6317 for_each_cpu(cpu, sched_group_cpus(sg)) {
6320 else if (sched_asym_prefer(cpu, max_cpu))
6323 sg->asym_prefer_cpu = max_cpu;
6327 } while (sg != sd->groups);
6329 if (cpu != group_balance_cpu(sg))
6332 update_group_capacity(sd, cpu);
6336 * Initializers for schedule domains
6337 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6340 static int default_relax_domain_level = -1;
6341 int sched_domain_level_max;
6343 static int __init setup_relax_domain_level(char *str)
6345 if (kstrtoint(str, 0, &default_relax_domain_level))
6346 pr_warn("Unable to set relax_domain_level\n");
6350 __setup("relax_domain_level=", setup_relax_domain_level);
6352 static void set_domain_attribute(struct sched_domain *sd,
6353 struct sched_domain_attr *attr)
6357 if (!attr || attr->relax_domain_level < 0) {
6358 if (default_relax_domain_level < 0)
6361 request = default_relax_domain_level;
6363 request = attr->relax_domain_level;
6364 if (request < sd->level) {
6365 /* turn off idle balance on this domain */
6366 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6368 /* turn on idle balance on this domain */
6369 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6373 static void __sdt_free(const struct cpumask *cpu_map);
6374 static int __sdt_alloc(const struct cpumask *cpu_map);
6376 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6377 const struct cpumask *cpu_map)
6381 if (!atomic_read(&d->rd->refcount))
6382 free_rootdomain(&d->rd->rcu); /* fall through */
6384 free_percpu(d->sd); /* fall through */
6386 __sdt_free(cpu_map); /* fall through */
6392 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6393 const struct cpumask *cpu_map)
6395 memset(d, 0, sizeof(*d));
6397 if (__sdt_alloc(cpu_map))
6398 return sa_sd_storage;
6399 d->sd = alloc_percpu(struct sched_domain *);
6401 return sa_sd_storage;
6402 d->rd = alloc_rootdomain();
6405 return sa_rootdomain;
6409 * NULL the sd_data elements we've used to build the sched_domain and
6410 * sched_group structure so that the subsequent __free_domain_allocs()
6411 * will not free the data we're using.
6413 static void claim_allocations(int cpu, struct sched_domain *sd)
6415 struct sd_data *sdd = sd->private;
6417 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6418 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6420 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
6421 *per_cpu_ptr(sdd->sds, cpu) = NULL;
6423 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6424 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6426 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6427 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6431 static int sched_domains_numa_levels;
6432 enum numa_topology_type sched_numa_topology_type;
6433 static int *sched_domains_numa_distance;
6434 int sched_max_numa_distance;
6435 static struct cpumask ***sched_domains_numa_masks;
6436 static int sched_domains_curr_level;
6440 * SD_flags allowed in topology descriptions.
6442 * These flags are purely descriptive of the topology and do not prescribe
6443 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6446 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6447 * SD_SHARE_PKG_RESOURCES - describes shared caches
6448 * SD_NUMA - describes NUMA topologies
6449 * SD_SHARE_POWERDOMAIN - describes shared power domain
6450 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6452 * Odd one out, which beside describing the topology has a quirk also
6453 * prescribes the desired behaviour that goes along with it:
6455 * SD_ASYM_PACKING - describes SMT quirks
6457 #define TOPOLOGY_SD_FLAGS \
6458 (SD_SHARE_CPUCAPACITY | \
6459 SD_SHARE_PKG_RESOURCES | \
6462 SD_ASYM_CPUCAPACITY | \
6463 SD_SHARE_POWERDOMAIN)
6465 static struct sched_domain *
6466 sd_init(struct sched_domain_topology_level *tl,
6467 const struct cpumask *cpu_map,
6468 struct sched_domain *child, int cpu)
6470 struct sd_data *sdd = &tl->data;
6471 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6472 int sd_id, sd_weight, sd_flags = 0;
6476 * Ugly hack to pass state to sd_numa_mask()...
6478 sched_domains_curr_level = tl->numa_level;
6481 sd_weight = cpumask_weight(tl->mask(cpu));
6484 sd_flags = (*tl->sd_flags)();
6485 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6486 "wrong sd_flags in topology description\n"))
6487 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6489 *sd = (struct sched_domain){
6490 .min_interval = sd_weight,
6491 .max_interval = 2*sd_weight,
6493 .imbalance_pct = 125,
6495 .cache_nice_tries = 0,
6502 .flags = 1*SD_LOAD_BALANCE
6503 | 1*SD_BALANCE_NEWIDLE
6508 | 0*SD_SHARE_CPUCAPACITY
6509 | 0*SD_SHARE_PKG_RESOURCES
6511 | 0*SD_PREFER_SIBLING
6516 .last_balance = jiffies,
6517 .balance_interval = sd_weight,
6519 .max_newidle_lb_cost = 0,
6520 .next_decay_max_lb_cost = jiffies,
6522 #ifdef CONFIG_SCHED_DEBUG
6527 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6528 sd_id = cpumask_first(sched_domain_span(sd));
6531 * Convert topological properties into behaviour.
6534 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6535 struct sched_domain *t = sd;
6537 for_each_lower_domain(t)
6538 t->flags |= SD_BALANCE_WAKE;
6541 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6542 sd->flags |= SD_PREFER_SIBLING;
6543 sd->imbalance_pct = 110;
6544 sd->smt_gain = 1178; /* ~15% */
6546 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6547 sd->imbalance_pct = 117;
6548 sd->cache_nice_tries = 1;
6552 } else if (sd->flags & SD_NUMA) {
6553 sd->cache_nice_tries = 2;
6557 sd->flags |= SD_SERIALIZE;
6558 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6559 sd->flags &= ~(SD_BALANCE_EXEC |
6566 sd->flags |= SD_PREFER_SIBLING;
6567 sd->cache_nice_tries = 1;
6573 * For all levels sharing cache; connect a sched_domain_shared
6576 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6577 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
6578 atomic_inc(&sd->shared->ref);
6579 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
6588 * Topology list, bottom-up.
6590 static struct sched_domain_topology_level default_topology[] = {
6591 #ifdef CONFIG_SCHED_SMT
6592 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6594 #ifdef CONFIG_SCHED_MC
6595 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6597 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6601 static struct sched_domain_topology_level *sched_domain_topology =
6604 #define for_each_sd_topology(tl) \
6605 for (tl = sched_domain_topology; tl->mask; tl++)
6607 void set_sched_topology(struct sched_domain_topology_level *tl)
6609 if (WARN_ON_ONCE(sched_smp_initialized))
6612 sched_domain_topology = tl;
6617 static const struct cpumask *sd_numa_mask(int cpu)
6619 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6622 static void sched_numa_warn(const char *str)
6624 static int done = false;
6632 printk(KERN_WARNING "ERROR: %s\n\n", str);
6634 for (i = 0; i < nr_node_ids; i++) {
6635 printk(KERN_WARNING " ");
6636 for (j = 0; j < nr_node_ids; j++)
6637 printk(KERN_CONT "%02d ", node_distance(i,j));
6638 printk(KERN_CONT "\n");
6640 printk(KERN_WARNING "\n");
6643 bool find_numa_distance(int distance)
6647 if (distance == node_distance(0, 0))
6650 for (i = 0; i < sched_domains_numa_levels; i++) {
6651 if (sched_domains_numa_distance[i] == distance)
6659 * A system can have three types of NUMA topology:
6660 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6661 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6662 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6664 * The difference between a glueless mesh topology and a backplane
6665 * topology lies in whether communication between not directly
6666 * connected nodes goes through intermediary nodes (where programs
6667 * could run), or through backplane controllers. This affects
6668 * placement of programs.
6670 * The type of topology can be discerned with the following tests:
6671 * - If the maximum distance between any nodes is 1 hop, the system
6672 * is directly connected.
6673 * - If for two nodes A and B, located N > 1 hops away from each other,
6674 * there is an intermediary node C, which is < N hops away from both
6675 * nodes A and B, the system is a glueless mesh.
6677 static void init_numa_topology_type(void)
6681 n = sched_max_numa_distance;
6683 if (sched_domains_numa_levels <= 1) {
6684 sched_numa_topology_type = NUMA_DIRECT;
6688 for_each_online_node(a) {
6689 for_each_online_node(b) {
6690 /* Find two nodes furthest removed from each other. */
6691 if (node_distance(a, b) < n)
6694 /* Is there an intermediary node between a and b? */
6695 for_each_online_node(c) {
6696 if (node_distance(a, c) < n &&
6697 node_distance(b, c) < n) {
6698 sched_numa_topology_type =
6704 sched_numa_topology_type = NUMA_BACKPLANE;
6710 static void sched_init_numa(void)
6712 int next_distance, curr_distance = node_distance(0, 0);
6713 struct sched_domain_topology_level *tl;
6717 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6718 if (!sched_domains_numa_distance)
6722 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6723 * unique distances in the node_distance() table.
6725 * Assumes node_distance(0,j) includes all distances in
6726 * node_distance(i,j) in order to avoid cubic time.
6728 next_distance = curr_distance;
6729 for (i = 0; i < nr_node_ids; i++) {
6730 for (j = 0; j < nr_node_ids; j++) {
6731 for (k = 0; k < nr_node_ids; k++) {
6732 int distance = node_distance(i, k);
6734 if (distance > curr_distance &&
6735 (distance < next_distance ||
6736 next_distance == curr_distance))
6737 next_distance = distance;
6740 * While not a strong assumption it would be nice to know
6741 * about cases where if node A is connected to B, B is not
6742 * equally connected to A.
6744 if (sched_debug() && node_distance(k, i) != distance)
6745 sched_numa_warn("Node-distance not symmetric");
6747 if (sched_debug() && i && !find_numa_distance(distance))
6748 sched_numa_warn("Node-0 not representative");
6750 if (next_distance != curr_distance) {
6751 sched_domains_numa_distance[level++] = next_distance;
6752 sched_domains_numa_levels = level;
6753 curr_distance = next_distance;
6758 * In case of sched_debug() we verify the above assumption.
6768 * 'level' contains the number of unique distances, excluding the
6769 * identity distance node_distance(i,i).
6771 * The sched_domains_numa_distance[] array includes the actual distance
6776 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6777 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6778 * the array will contain less then 'level' members. This could be
6779 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6780 * in other functions.
6782 * We reset it to 'level' at the end of this function.
6784 sched_domains_numa_levels = 0;
6786 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6787 if (!sched_domains_numa_masks)
6791 * Now for each level, construct a mask per node which contains all
6792 * cpus of nodes that are that many hops away from us.
6794 for (i = 0; i < level; i++) {
6795 sched_domains_numa_masks[i] =
6796 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6797 if (!sched_domains_numa_masks[i])
6800 for (j = 0; j < nr_node_ids; j++) {
6801 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6805 sched_domains_numa_masks[i][j] = mask;
6808 if (node_distance(j, k) > sched_domains_numa_distance[i])
6811 cpumask_or(mask, mask, cpumask_of_node(k));
6816 /* Compute default topology size */
6817 for (i = 0; sched_domain_topology[i].mask; i++);
6819 tl = kzalloc((i + level + 1) *
6820 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6825 * Copy the default topology bits..
6827 for (i = 0; sched_domain_topology[i].mask; i++)
6828 tl[i] = sched_domain_topology[i];
6831 * .. and append 'j' levels of NUMA goodness.
6833 for (j = 0; j < level; i++, j++) {
6834 tl[i] = (struct sched_domain_topology_level){
6835 .mask = sd_numa_mask,
6836 .sd_flags = cpu_numa_flags,
6837 .flags = SDTL_OVERLAP,
6843 sched_domain_topology = tl;
6845 sched_domains_numa_levels = level;
6846 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6848 init_numa_topology_type();
6851 static void sched_domains_numa_masks_set(unsigned int cpu)
6853 int node = cpu_to_node(cpu);
6856 for (i = 0; i < sched_domains_numa_levels; i++) {
6857 for (j = 0; j < nr_node_ids; j++) {
6858 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6859 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6864 static void sched_domains_numa_masks_clear(unsigned int cpu)
6868 for (i = 0; i < sched_domains_numa_levels; i++) {
6869 for (j = 0; j < nr_node_ids; j++)
6870 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6875 static inline void sched_init_numa(void) { }
6876 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6877 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6878 #endif /* CONFIG_NUMA */
6880 static int __sdt_alloc(const struct cpumask *cpu_map)
6882 struct sched_domain_topology_level *tl;
6885 for_each_sd_topology(tl) {
6886 struct sd_data *sdd = &tl->data;
6888 sdd->sd = alloc_percpu(struct sched_domain *);
6892 sdd->sds = alloc_percpu(struct sched_domain_shared *);
6896 sdd->sg = alloc_percpu(struct sched_group *);
6900 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6904 for_each_cpu(j, cpu_map) {
6905 struct sched_domain *sd;
6906 struct sched_domain_shared *sds;
6907 struct sched_group *sg;
6908 struct sched_group_capacity *sgc;
6910 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6911 GFP_KERNEL, cpu_to_node(j));
6915 *per_cpu_ptr(sdd->sd, j) = sd;
6917 sds = kzalloc_node(sizeof(struct sched_domain_shared),
6918 GFP_KERNEL, cpu_to_node(j));
6922 *per_cpu_ptr(sdd->sds, j) = sds;
6924 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6925 GFP_KERNEL, cpu_to_node(j));
6931 *per_cpu_ptr(sdd->sg, j) = sg;
6933 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6934 GFP_KERNEL, cpu_to_node(j));
6938 *per_cpu_ptr(sdd->sgc, j) = sgc;
6945 static void __sdt_free(const struct cpumask *cpu_map)
6947 struct sched_domain_topology_level *tl;
6950 for_each_sd_topology(tl) {
6951 struct sd_data *sdd = &tl->data;
6953 for_each_cpu(j, cpu_map) {
6954 struct sched_domain *sd;
6957 sd = *per_cpu_ptr(sdd->sd, j);
6958 if (sd && (sd->flags & SD_OVERLAP))
6959 free_sched_groups(sd->groups, 0);
6960 kfree(*per_cpu_ptr(sdd->sd, j));
6964 kfree(*per_cpu_ptr(sdd->sds, j));
6966 kfree(*per_cpu_ptr(sdd->sg, j));
6968 kfree(*per_cpu_ptr(sdd->sgc, j));
6970 free_percpu(sdd->sd);
6972 free_percpu(sdd->sds);
6974 free_percpu(sdd->sg);
6976 free_percpu(sdd->sgc);
6981 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6982 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6983 struct sched_domain *child, int cpu)
6985 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
6988 sd->level = child->level + 1;
6989 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6992 if (!cpumask_subset(sched_domain_span(child),
6993 sched_domain_span(sd))) {
6994 pr_err("BUG: arch topology borken\n");
6995 #ifdef CONFIG_SCHED_DEBUG
6996 pr_err(" the %s domain not a subset of the %s domain\n",
6997 child->name, sd->name);
6999 /* Fixup, ensure @sd has at least @child cpus. */
7000 cpumask_or(sched_domain_span(sd),
7001 sched_domain_span(sd),
7002 sched_domain_span(child));
7006 set_domain_attribute(sd, attr);
7012 * Build sched domains for a given set of cpus and attach the sched domains
7013 * to the individual cpus
7015 static int build_sched_domains(const struct cpumask *cpu_map,
7016 struct sched_domain_attr *attr)
7018 enum s_alloc alloc_state;
7019 struct sched_domain *sd;
7021 struct rq *rq = NULL;
7022 int i, ret = -ENOMEM;
7024 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7025 if (alloc_state != sa_rootdomain)
7028 /* Set up domains for cpus specified by the cpu_map. */
7029 for_each_cpu(i, cpu_map) {
7030 struct sched_domain_topology_level *tl;
7033 for_each_sd_topology(tl) {
7034 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7035 if (tl == sched_domain_topology)
7036 *per_cpu_ptr(d.sd, i) = sd;
7037 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7038 sd->flags |= SD_OVERLAP;
7039 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7044 /* Build the groups for the domains */
7045 for_each_cpu(i, cpu_map) {
7046 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7047 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7048 if (sd->flags & SD_OVERLAP) {
7049 if (build_overlap_sched_groups(sd, i))
7052 if (build_sched_groups(sd, i))
7058 /* Calculate CPU capacity for physical packages and nodes */
7059 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7060 if (!cpumask_test_cpu(i, cpu_map))
7063 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7064 claim_allocations(i, sd);
7065 init_sched_groups_capacity(i, sd);
7069 /* Attach the domains */
7071 for_each_cpu(i, cpu_map) {
7073 sd = *per_cpu_ptr(d.sd, i);
7075 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7076 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
7077 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
7079 cpu_attach_domain(sd, d.rd, i);
7083 if (rq && sched_debug_enabled) {
7084 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7085 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
7090 __free_domain_allocs(&d, alloc_state, cpu_map);
7094 static cpumask_var_t *doms_cur; /* current sched domains */
7095 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7096 static struct sched_domain_attr *dattr_cur;
7097 /* attribues of custom domains in 'doms_cur' */
7100 * Special case: If a kmalloc of a doms_cur partition (array of
7101 * cpumask) fails, then fallback to a single sched domain,
7102 * as determined by the single cpumask fallback_doms.
7104 static cpumask_var_t fallback_doms;
7107 * arch_update_cpu_topology lets virtualized architectures update the
7108 * cpu core maps. It is supposed to return 1 if the topology changed
7109 * or 0 if it stayed the same.
7111 int __weak arch_update_cpu_topology(void)
7116 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7119 cpumask_var_t *doms;
7121 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7124 for (i = 0; i < ndoms; i++) {
7125 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7126 free_sched_domains(doms, i);
7133 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7136 for (i = 0; i < ndoms; i++)
7137 free_cpumask_var(doms[i]);
7142 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7143 * For now this just excludes isolated cpus, but could be used to
7144 * exclude other special cases in the future.
7146 static int init_sched_domains(const struct cpumask *cpu_map)
7150 arch_update_cpu_topology();
7152 doms_cur = alloc_sched_domains(ndoms_cur);
7154 doms_cur = &fallback_doms;
7155 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7156 err = build_sched_domains(doms_cur[0], NULL);
7157 register_sched_domain_sysctl();
7163 * Detach sched domains from a group of cpus specified in cpu_map
7164 * These cpus will now be attached to the NULL domain
7166 static void detach_destroy_domains(const struct cpumask *cpu_map)
7171 for_each_cpu(i, cpu_map)
7172 cpu_attach_domain(NULL, &def_root_domain, i);
7176 /* handle null as "default" */
7177 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7178 struct sched_domain_attr *new, int idx_new)
7180 struct sched_domain_attr tmp;
7187 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7188 new ? (new + idx_new) : &tmp,
7189 sizeof(struct sched_domain_attr));
7193 * Partition sched domains as specified by the 'ndoms_new'
7194 * cpumasks in the array doms_new[] of cpumasks. This compares
7195 * doms_new[] to the current sched domain partitioning, doms_cur[].
7196 * It destroys each deleted domain and builds each new domain.
7198 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7199 * The masks don't intersect (don't overlap.) We should setup one
7200 * sched domain for each mask. CPUs not in any of the cpumasks will
7201 * not be load balanced. If the same cpumask appears both in the
7202 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7205 * The passed in 'doms_new' should be allocated using
7206 * alloc_sched_domains. This routine takes ownership of it and will
7207 * free_sched_domains it when done with it. If the caller failed the
7208 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7209 * and partition_sched_domains() will fallback to the single partition
7210 * 'fallback_doms', it also forces the domains to be rebuilt.
7212 * If doms_new == NULL it will be replaced with cpu_online_mask.
7213 * ndoms_new == 0 is a special case for destroying existing domains,
7214 * and it will not create the default domain.
7216 * Call with hotplug lock held
7218 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7219 struct sched_domain_attr *dattr_new)
7224 mutex_lock(&sched_domains_mutex);
7226 /* always unregister in case we don't destroy any domains */
7227 unregister_sched_domain_sysctl();
7229 /* Let architecture update cpu core mappings. */
7230 new_topology = arch_update_cpu_topology();
7232 n = doms_new ? ndoms_new : 0;
7234 /* Destroy deleted domains */
7235 for (i = 0; i < ndoms_cur; i++) {
7236 for (j = 0; j < n && !new_topology; j++) {
7237 if (cpumask_equal(doms_cur[i], doms_new[j])
7238 && dattrs_equal(dattr_cur, i, dattr_new, j))
7241 /* no match - a current sched domain not in new doms_new[] */
7242 detach_destroy_domains(doms_cur[i]);
7248 if (doms_new == NULL) {
7250 doms_new = &fallback_doms;
7251 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7252 WARN_ON_ONCE(dattr_new);
7255 /* Build new domains */
7256 for (i = 0; i < ndoms_new; i++) {
7257 for (j = 0; j < n && !new_topology; j++) {
7258 if (cpumask_equal(doms_new[i], doms_cur[j])
7259 && dattrs_equal(dattr_new, i, dattr_cur, j))
7262 /* no match - add a new doms_new */
7263 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7268 /* Remember the new sched domains */
7269 if (doms_cur != &fallback_doms)
7270 free_sched_domains(doms_cur, ndoms_cur);
7271 kfree(dattr_cur); /* kfree(NULL) is safe */
7272 doms_cur = doms_new;
7273 dattr_cur = dattr_new;
7274 ndoms_cur = ndoms_new;
7276 register_sched_domain_sysctl();
7278 mutex_unlock(&sched_domains_mutex);
7281 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7284 * Update cpusets according to cpu_active mask. If cpusets are
7285 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7286 * around partition_sched_domains().
7288 * If we come here as part of a suspend/resume, don't touch cpusets because we
7289 * want to restore it back to its original state upon resume anyway.
7291 static void cpuset_cpu_active(void)
7293 if (cpuhp_tasks_frozen) {
7295 * num_cpus_frozen tracks how many CPUs are involved in suspend
7296 * resume sequence. As long as this is not the last online
7297 * operation in the resume sequence, just build a single sched
7298 * domain, ignoring cpusets.
7301 if (likely(num_cpus_frozen)) {
7302 partition_sched_domains(1, NULL, NULL);
7306 * This is the last CPU online operation. So fall through and
7307 * restore the original sched domains by considering the
7308 * cpuset configurations.
7311 cpuset_update_active_cpus(true);
7314 static int cpuset_cpu_inactive(unsigned int cpu)
7316 unsigned long flags;
7321 if (!cpuhp_tasks_frozen) {
7322 rcu_read_lock_sched();
7323 dl_b = dl_bw_of(cpu);
7325 raw_spin_lock_irqsave(&dl_b->lock, flags);
7326 cpus = dl_bw_cpus(cpu);
7327 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7328 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7330 rcu_read_unlock_sched();
7334 cpuset_update_active_cpus(false);
7337 partition_sched_domains(1, NULL, NULL);
7342 int sched_cpu_activate(unsigned int cpu)
7344 struct rq *rq = cpu_rq(cpu);
7345 unsigned long flags;
7347 set_cpu_active(cpu, true);
7349 if (sched_smp_initialized) {
7350 sched_domains_numa_masks_set(cpu);
7351 cpuset_cpu_active();
7355 * Put the rq online, if not already. This happens:
7357 * 1) In the early boot process, because we build the real domains
7358 * after all cpus have been brought up.
7360 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7363 raw_spin_lock_irqsave(&rq->lock, flags);
7365 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7368 raw_spin_unlock_irqrestore(&rq->lock, flags);
7370 update_max_interval();
7375 int sched_cpu_deactivate(unsigned int cpu)
7379 set_cpu_active(cpu, false);
7381 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7382 * users of this state to go away such that all new such users will
7385 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7386 * not imply sync_sched(), so wait for both.
7388 * Do sync before park smpboot threads to take care the rcu boost case.
7390 if (IS_ENABLED(CONFIG_PREEMPT))
7391 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7395 if (!sched_smp_initialized)
7398 ret = cpuset_cpu_inactive(cpu);
7400 set_cpu_active(cpu, true);
7403 sched_domains_numa_masks_clear(cpu);
7407 static void sched_rq_cpu_starting(unsigned int cpu)
7409 struct rq *rq = cpu_rq(cpu);
7411 rq->calc_load_update = calc_load_update;
7412 update_max_interval();
7415 int sched_cpu_starting(unsigned int cpu)
7417 set_cpu_rq_start_time(cpu);
7418 sched_rq_cpu_starting(cpu);
7422 #ifdef CONFIG_HOTPLUG_CPU
7423 int sched_cpu_dying(unsigned int cpu)
7425 struct rq *rq = cpu_rq(cpu);
7426 unsigned long flags;
7428 /* Handle pending wakeups and then migrate everything off */
7429 sched_ttwu_pending();
7430 raw_spin_lock_irqsave(&rq->lock, flags);
7432 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7436 BUG_ON(rq->nr_running != 1);
7437 raw_spin_unlock_irqrestore(&rq->lock, flags);
7438 calc_load_migrate(rq);
7439 update_max_interval();
7440 nohz_balance_exit_idle(cpu);
7446 #ifdef CONFIG_SCHED_SMT
7447 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7449 static void sched_init_smt(void)
7452 * We've enumerated all CPUs and will assume that if any CPU
7453 * has SMT siblings, CPU0 will too.
7455 if (cpumask_weight(cpu_smt_mask(0)) > 1)
7456 static_branch_enable(&sched_smt_present);
7459 static inline void sched_init_smt(void) { }
7462 void __init sched_init_smp(void)
7464 cpumask_var_t non_isolated_cpus;
7466 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7467 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7472 * There's no userspace yet to cause hotplug operations; hence all the
7473 * cpu masks are stable and all blatant races in the below code cannot
7476 mutex_lock(&sched_domains_mutex);
7477 init_sched_domains(cpu_active_mask);
7478 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7479 if (cpumask_empty(non_isolated_cpus))
7480 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7481 mutex_unlock(&sched_domains_mutex);
7483 /* Move init over to a non-isolated CPU */
7484 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7486 sched_init_granularity();
7487 free_cpumask_var(non_isolated_cpus);
7489 init_sched_rt_class();
7490 init_sched_dl_class();
7494 sched_smp_initialized = true;
7497 static int __init migration_init(void)
7499 sched_rq_cpu_starting(smp_processor_id());
7502 early_initcall(migration_init);
7505 void __init sched_init_smp(void)
7507 sched_init_granularity();
7509 #endif /* CONFIG_SMP */
7511 int in_sched_functions(unsigned long addr)
7513 return in_lock_functions(addr) ||
7514 (addr >= (unsigned long)__sched_text_start
7515 && addr < (unsigned long)__sched_text_end);
7518 #ifdef CONFIG_CGROUP_SCHED
7520 * Default task group.
7521 * Every task in system belongs to this group at bootup.
7523 struct task_group root_task_group;
7524 LIST_HEAD(task_groups);
7526 /* Cacheline aligned slab cache for task_group */
7527 static struct kmem_cache *task_group_cache __read_mostly;
7530 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7531 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7533 #define WAIT_TABLE_BITS 8
7534 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
7535 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
7537 wait_queue_head_t *bit_waitqueue(void *word, int bit)
7539 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
7540 unsigned long val = (unsigned long)word << shift | bit;
7542 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
7544 EXPORT_SYMBOL(bit_waitqueue);
7546 void __init sched_init(void)
7549 unsigned long alloc_size = 0, ptr;
7551 for (i = 0; i < WAIT_TABLE_SIZE; i++)
7552 init_waitqueue_head(bit_wait_table + i);
7554 #ifdef CONFIG_FAIR_GROUP_SCHED
7555 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7557 #ifdef CONFIG_RT_GROUP_SCHED
7558 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7561 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7563 #ifdef CONFIG_FAIR_GROUP_SCHED
7564 root_task_group.se = (struct sched_entity **)ptr;
7565 ptr += nr_cpu_ids * sizeof(void **);
7567 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7568 ptr += nr_cpu_ids * sizeof(void **);
7570 #endif /* CONFIG_FAIR_GROUP_SCHED */
7571 #ifdef CONFIG_RT_GROUP_SCHED
7572 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7573 ptr += nr_cpu_ids * sizeof(void **);
7575 root_task_group.rt_rq = (struct rt_rq **)ptr;
7576 ptr += nr_cpu_ids * sizeof(void **);
7578 #endif /* CONFIG_RT_GROUP_SCHED */
7580 #ifdef CONFIG_CPUMASK_OFFSTACK
7581 for_each_possible_cpu(i) {
7582 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7583 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7584 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7585 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7587 #endif /* CONFIG_CPUMASK_OFFSTACK */
7589 init_rt_bandwidth(&def_rt_bandwidth,
7590 global_rt_period(), global_rt_runtime());
7591 init_dl_bandwidth(&def_dl_bandwidth,
7592 global_rt_period(), global_rt_runtime());
7595 init_defrootdomain();
7598 #ifdef CONFIG_RT_GROUP_SCHED
7599 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7600 global_rt_period(), global_rt_runtime());
7601 #endif /* CONFIG_RT_GROUP_SCHED */
7603 #ifdef CONFIG_CGROUP_SCHED
7604 task_group_cache = KMEM_CACHE(task_group, 0);
7606 list_add(&root_task_group.list, &task_groups);
7607 INIT_LIST_HEAD(&root_task_group.children);
7608 INIT_LIST_HEAD(&root_task_group.siblings);
7609 autogroup_init(&init_task);
7610 #endif /* CONFIG_CGROUP_SCHED */
7612 for_each_possible_cpu(i) {
7616 raw_spin_lock_init(&rq->lock);
7618 rq->calc_load_active = 0;
7619 rq->calc_load_update = jiffies + LOAD_FREQ;
7620 init_cfs_rq(&rq->cfs);
7621 init_rt_rq(&rq->rt);
7622 init_dl_rq(&rq->dl);
7623 #ifdef CONFIG_FAIR_GROUP_SCHED
7624 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7625 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7626 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7628 * How much cpu bandwidth does root_task_group get?
7630 * In case of task-groups formed thr' the cgroup filesystem, it
7631 * gets 100% of the cpu resources in the system. This overall
7632 * system cpu resource is divided among the tasks of
7633 * root_task_group and its child task-groups in a fair manner,
7634 * based on each entity's (task or task-group's) weight
7635 * (se->load.weight).
7637 * In other words, if root_task_group has 10 tasks of weight
7638 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7639 * then A0's share of the cpu resource is:
7641 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7643 * We achieve this by letting root_task_group's tasks sit
7644 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7646 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7647 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7648 #endif /* CONFIG_FAIR_GROUP_SCHED */
7650 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7651 #ifdef CONFIG_RT_GROUP_SCHED
7652 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7655 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7656 rq->cpu_load[j] = 0;
7661 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7662 rq->balance_callback = NULL;
7663 rq->active_balance = 0;
7664 rq->next_balance = jiffies;
7669 rq->avg_idle = 2*sysctl_sched_migration_cost;
7670 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7672 INIT_LIST_HEAD(&rq->cfs_tasks);
7674 rq_attach_root(rq, &def_root_domain);
7675 #ifdef CONFIG_NO_HZ_COMMON
7676 rq->last_load_update_tick = jiffies;
7679 #ifdef CONFIG_NO_HZ_FULL
7680 rq->last_sched_tick = 0;
7682 #endif /* CONFIG_SMP */
7684 atomic_set(&rq->nr_iowait, 0);
7687 set_load_weight(&init_task);
7690 * The boot idle thread does lazy MMU switching as well:
7692 atomic_inc(&init_mm.mm_count);
7693 enter_lazy_tlb(&init_mm, current);
7696 * Make us the idle thread. Technically, schedule() should not be
7697 * called from this thread, however somewhere below it might be,
7698 * but because we are the idle thread, we just pick up running again
7699 * when this runqueue becomes "idle".
7701 init_idle(current, smp_processor_id());
7703 calc_load_update = jiffies + LOAD_FREQ;
7706 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7707 /* May be allocated at isolcpus cmdline parse time */
7708 if (cpu_isolated_map == NULL)
7709 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7710 idle_thread_set_boot_cpu();
7711 set_cpu_rq_start_time(smp_processor_id());
7713 init_sched_fair_class();
7717 scheduler_running = 1;
7720 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7721 static inline int preempt_count_equals(int preempt_offset)
7723 int nested = preempt_count() + rcu_preempt_depth();
7725 return (nested == preempt_offset);
7728 void __might_sleep(const char *file, int line, int preempt_offset)
7731 * Blocking primitives will set (and therefore destroy) current->state,
7732 * since we will exit with TASK_RUNNING make sure we enter with it,
7733 * otherwise we will destroy state.
7735 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7736 "do not call blocking ops when !TASK_RUNNING; "
7737 "state=%lx set at [<%p>] %pS\n",
7739 (void *)current->task_state_change,
7740 (void *)current->task_state_change);
7742 ___might_sleep(file, line, preempt_offset);
7744 EXPORT_SYMBOL(__might_sleep);
7746 void ___might_sleep(const char *file, int line, int preempt_offset)
7748 static unsigned long prev_jiffy; /* ratelimiting */
7749 unsigned long preempt_disable_ip;
7751 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7752 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7753 !is_idle_task(current)) ||
7754 system_state != SYSTEM_RUNNING || oops_in_progress)
7756 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7758 prev_jiffy = jiffies;
7760 /* Save this before calling printk(), since that will clobber it */
7761 preempt_disable_ip = get_preempt_disable_ip(current);
7764 "BUG: sleeping function called from invalid context at %s:%d\n",
7767 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7768 in_atomic(), irqs_disabled(),
7769 current->pid, current->comm);
7771 if (task_stack_end_corrupted(current))
7772 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7774 debug_show_held_locks(current);
7775 if (irqs_disabled())
7776 print_irqtrace_events(current);
7777 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7778 && !preempt_count_equals(preempt_offset)) {
7779 pr_err("Preemption disabled at:");
7780 print_ip_sym(preempt_disable_ip);
7784 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7786 EXPORT_SYMBOL(___might_sleep);
7789 #ifdef CONFIG_MAGIC_SYSRQ
7790 void normalize_rt_tasks(void)
7792 struct task_struct *g, *p;
7793 struct sched_attr attr = {
7794 .sched_policy = SCHED_NORMAL,
7797 read_lock(&tasklist_lock);
7798 for_each_process_thread(g, p) {
7800 * Only normalize user tasks:
7802 if (p->flags & PF_KTHREAD)
7805 p->se.exec_start = 0;
7806 schedstat_set(p->se.statistics.wait_start, 0);
7807 schedstat_set(p->se.statistics.sleep_start, 0);
7808 schedstat_set(p->se.statistics.block_start, 0);
7810 if (!dl_task(p) && !rt_task(p)) {
7812 * Renice negative nice level userspace
7815 if (task_nice(p) < 0)
7816 set_user_nice(p, 0);
7820 __sched_setscheduler(p, &attr, false, false);
7822 read_unlock(&tasklist_lock);
7825 #endif /* CONFIG_MAGIC_SYSRQ */
7827 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7829 * These functions are only useful for the IA64 MCA handling, or kdb.
7831 * They can only be called when the whole system has been
7832 * stopped - every CPU needs to be quiescent, and no scheduling
7833 * activity can take place. Using them for anything else would
7834 * be a serious bug, and as a result, they aren't even visible
7835 * under any other configuration.
7839 * curr_task - return the current task for a given cpu.
7840 * @cpu: the processor in question.
7842 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7844 * Return: The current task for @cpu.
7846 struct task_struct *curr_task(int cpu)
7848 return cpu_curr(cpu);
7851 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7855 * set_curr_task - set the current task for a given cpu.
7856 * @cpu: the processor in question.
7857 * @p: the task pointer to set.
7859 * Description: This function must only be used when non-maskable interrupts
7860 * are serviced on a separate stack. It allows the architecture to switch the
7861 * notion of the current task on a cpu in a non-blocking manner. This function
7862 * must be called with all CPU's synchronized, and interrupts disabled, the
7863 * and caller must save the original value of the current task (see
7864 * curr_task() above) and restore that value before reenabling interrupts and
7865 * re-starting the system.
7867 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7869 void ia64_set_curr_task(int cpu, struct task_struct *p)
7876 #ifdef CONFIG_CGROUP_SCHED
7877 /* task_group_lock serializes the addition/removal of task groups */
7878 static DEFINE_SPINLOCK(task_group_lock);
7880 static void sched_free_group(struct task_group *tg)
7882 free_fair_sched_group(tg);
7883 free_rt_sched_group(tg);
7885 kmem_cache_free(task_group_cache, tg);
7888 /* allocate runqueue etc for a new task group */
7889 struct task_group *sched_create_group(struct task_group *parent)
7891 struct task_group *tg;
7893 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7895 return ERR_PTR(-ENOMEM);
7897 if (!alloc_fair_sched_group(tg, parent))
7900 if (!alloc_rt_sched_group(tg, parent))
7906 sched_free_group(tg);
7907 return ERR_PTR(-ENOMEM);
7910 void sched_online_group(struct task_group *tg, struct task_group *parent)
7912 unsigned long flags;
7914 spin_lock_irqsave(&task_group_lock, flags);
7915 list_add_rcu(&tg->list, &task_groups);
7917 WARN_ON(!parent); /* root should already exist */
7919 tg->parent = parent;
7920 INIT_LIST_HEAD(&tg->children);
7921 list_add_rcu(&tg->siblings, &parent->children);
7922 spin_unlock_irqrestore(&task_group_lock, flags);
7924 online_fair_sched_group(tg);
7927 /* rcu callback to free various structures associated with a task group */
7928 static void sched_free_group_rcu(struct rcu_head *rhp)
7930 /* now it should be safe to free those cfs_rqs */
7931 sched_free_group(container_of(rhp, struct task_group, rcu));
7934 void sched_destroy_group(struct task_group *tg)
7936 /* wait for possible concurrent references to cfs_rqs complete */
7937 call_rcu(&tg->rcu, sched_free_group_rcu);
7940 void sched_offline_group(struct task_group *tg)
7942 unsigned long flags;
7944 /* end participation in shares distribution */
7945 unregister_fair_sched_group(tg);
7947 spin_lock_irqsave(&task_group_lock, flags);
7948 list_del_rcu(&tg->list);
7949 list_del_rcu(&tg->siblings);
7950 spin_unlock_irqrestore(&task_group_lock, flags);
7953 static void sched_change_group(struct task_struct *tsk, int type)
7955 struct task_group *tg;
7958 * All callers are synchronized by task_rq_lock(); we do not use RCU
7959 * which is pointless here. Thus, we pass "true" to task_css_check()
7960 * to prevent lockdep warnings.
7962 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7963 struct task_group, css);
7964 tg = autogroup_task_group(tsk, tg);
7965 tsk->sched_task_group = tg;
7967 #ifdef CONFIG_FAIR_GROUP_SCHED
7968 if (tsk->sched_class->task_change_group)
7969 tsk->sched_class->task_change_group(tsk, type);
7972 set_task_rq(tsk, task_cpu(tsk));
7976 * Change task's runqueue when it moves between groups.
7978 * The caller of this function should have put the task in its new group by
7979 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7982 void sched_move_task(struct task_struct *tsk)
7984 int queued, running;
7988 rq = task_rq_lock(tsk, &rf);
7990 running = task_current(rq, tsk);
7991 queued = task_on_rq_queued(tsk);
7994 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7995 if (unlikely(running))
7996 put_prev_task(rq, tsk);
7998 sched_change_group(tsk, TASK_MOVE_GROUP);
8001 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
8002 if (unlikely(running))
8003 set_curr_task(rq, tsk);
8005 task_rq_unlock(rq, tsk, &rf);
8007 #endif /* CONFIG_CGROUP_SCHED */
8009 #ifdef CONFIG_RT_GROUP_SCHED
8011 * Ensure that the real time constraints are schedulable.
8013 static DEFINE_MUTEX(rt_constraints_mutex);
8015 /* Must be called with tasklist_lock held */
8016 static inline int tg_has_rt_tasks(struct task_group *tg)
8018 struct task_struct *g, *p;
8021 * Autogroups do not have RT tasks; see autogroup_create().
8023 if (task_group_is_autogroup(tg))
8026 for_each_process_thread(g, p) {
8027 if (rt_task(p) && task_group(p) == tg)
8034 struct rt_schedulable_data {
8035 struct task_group *tg;
8040 static int tg_rt_schedulable(struct task_group *tg, void *data)
8042 struct rt_schedulable_data *d = data;
8043 struct task_group *child;
8044 unsigned long total, sum = 0;
8045 u64 period, runtime;
8047 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8048 runtime = tg->rt_bandwidth.rt_runtime;
8051 period = d->rt_period;
8052 runtime = d->rt_runtime;
8056 * Cannot have more runtime than the period.
8058 if (runtime > period && runtime != RUNTIME_INF)
8062 * Ensure we don't starve existing RT tasks.
8064 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8067 total = to_ratio(period, runtime);
8070 * Nobody can have more than the global setting allows.
8072 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8076 * The sum of our children's runtime should not exceed our own.
8078 list_for_each_entry_rcu(child, &tg->children, siblings) {
8079 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8080 runtime = child->rt_bandwidth.rt_runtime;
8082 if (child == d->tg) {
8083 period = d->rt_period;
8084 runtime = d->rt_runtime;
8087 sum += to_ratio(period, runtime);
8096 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8100 struct rt_schedulable_data data = {
8102 .rt_period = period,
8103 .rt_runtime = runtime,
8107 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8113 static int tg_set_rt_bandwidth(struct task_group *tg,
8114 u64 rt_period, u64 rt_runtime)
8119 * Disallowing the root group RT runtime is BAD, it would disallow the
8120 * kernel creating (and or operating) RT threads.
8122 if (tg == &root_task_group && rt_runtime == 0)
8125 /* No period doesn't make any sense. */
8129 mutex_lock(&rt_constraints_mutex);
8130 read_lock(&tasklist_lock);
8131 err = __rt_schedulable(tg, rt_period, rt_runtime);
8135 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8136 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8137 tg->rt_bandwidth.rt_runtime = rt_runtime;
8139 for_each_possible_cpu(i) {
8140 struct rt_rq *rt_rq = tg->rt_rq[i];
8142 raw_spin_lock(&rt_rq->rt_runtime_lock);
8143 rt_rq->rt_runtime = rt_runtime;
8144 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8146 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8148 read_unlock(&tasklist_lock);
8149 mutex_unlock(&rt_constraints_mutex);
8154 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8156 u64 rt_runtime, rt_period;
8158 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8159 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8160 if (rt_runtime_us < 0)
8161 rt_runtime = RUNTIME_INF;
8163 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8166 static long sched_group_rt_runtime(struct task_group *tg)
8170 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8173 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8174 do_div(rt_runtime_us, NSEC_PER_USEC);
8175 return rt_runtime_us;
8178 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8180 u64 rt_runtime, rt_period;
8182 rt_period = rt_period_us * NSEC_PER_USEC;
8183 rt_runtime = tg->rt_bandwidth.rt_runtime;
8185 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8188 static long sched_group_rt_period(struct task_group *tg)
8192 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8193 do_div(rt_period_us, NSEC_PER_USEC);
8194 return rt_period_us;
8196 #endif /* CONFIG_RT_GROUP_SCHED */
8198 #ifdef CONFIG_RT_GROUP_SCHED
8199 static int sched_rt_global_constraints(void)
8203 mutex_lock(&rt_constraints_mutex);
8204 read_lock(&tasklist_lock);
8205 ret = __rt_schedulable(NULL, 0, 0);
8206 read_unlock(&tasklist_lock);
8207 mutex_unlock(&rt_constraints_mutex);
8212 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8214 /* Don't accept realtime tasks when there is no way for them to run */
8215 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8221 #else /* !CONFIG_RT_GROUP_SCHED */
8222 static int sched_rt_global_constraints(void)
8224 unsigned long flags;
8227 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8228 for_each_possible_cpu(i) {
8229 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8231 raw_spin_lock(&rt_rq->rt_runtime_lock);
8232 rt_rq->rt_runtime = global_rt_runtime();
8233 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8235 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8239 #endif /* CONFIG_RT_GROUP_SCHED */
8241 static int sched_dl_global_validate(void)
8243 u64 runtime = global_rt_runtime();
8244 u64 period = global_rt_period();
8245 u64 new_bw = to_ratio(period, runtime);
8248 unsigned long flags;
8251 * Here we want to check the bandwidth not being set to some
8252 * value smaller than the currently allocated bandwidth in
8253 * any of the root_domains.
8255 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8256 * cycling on root_domains... Discussion on different/better
8257 * solutions is welcome!
8259 for_each_possible_cpu(cpu) {
8260 rcu_read_lock_sched();
8261 dl_b = dl_bw_of(cpu);
8263 raw_spin_lock_irqsave(&dl_b->lock, flags);
8264 if (new_bw < dl_b->total_bw)
8266 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8268 rcu_read_unlock_sched();
8277 static void sched_dl_do_global(void)
8282 unsigned long flags;
8284 def_dl_bandwidth.dl_period = global_rt_period();
8285 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8287 if (global_rt_runtime() != RUNTIME_INF)
8288 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8291 * FIXME: As above...
8293 for_each_possible_cpu(cpu) {
8294 rcu_read_lock_sched();
8295 dl_b = dl_bw_of(cpu);
8297 raw_spin_lock_irqsave(&dl_b->lock, flags);
8299 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8301 rcu_read_unlock_sched();
8305 static int sched_rt_global_validate(void)
8307 if (sysctl_sched_rt_period <= 0)
8310 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8311 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8317 static void sched_rt_do_global(void)
8319 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8320 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8323 int sched_rt_handler(struct ctl_table *table, int write,
8324 void __user *buffer, size_t *lenp,
8327 int old_period, old_runtime;
8328 static DEFINE_MUTEX(mutex);
8332 old_period = sysctl_sched_rt_period;
8333 old_runtime = sysctl_sched_rt_runtime;
8335 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8337 if (!ret && write) {
8338 ret = sched_rt_global_validate();
8342 ret = sched_dl_global_validate();
8346 ret = sched_rt_global_constraints();
8350 sched_rt_do_global();
8351 sched_dl_do_global();
8355 sysctl_sched_rt_period = old_period;
8356 sysctl_sched_rt_runtime = old_runtime;
8358 mutex_unlock(&mutex);
8363 int sched_rr_handler(struct ctl_table *table, int write,
8364 void __user *buffer, size_t *lenp,
8368 static DEFINE_MUTEX(mutex);
8371 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8372 /* make sure that internally we keep jiffies */
8373 /* also, writing zero resets timeslice to default */
8374 if (!ret && write) {
8375 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8376 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8378 mutex_unlock(&mutex);
8382 #ifdef CONFIG_CGROUP_SCHED
8384 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8386 return css ? container_of(css, struct task_group, css) : NULL;
8389 static struct cgroup_subsys_state *
8390 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8392 struct task_group *parent = css_tg(parent_css);
8393 struct task_group *tg;
8396 /* This is early initialization for the top cgroup */
8397 return &root_task_group.css;
8400 tg = sched_create_group(parent);
8402 return ERR_PTR(-ENOMEM);
8404 sched_online_group(tg, parent);
8409 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8411 struct task_group *tg = css_tg(css);
8413 sched_offline_group(tg);
8416 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8418 struct task_group *tg = css_tg(css);
8421 * Relies on the RCU grace period between css_released() and this.
8423 sched_free_group(tg);
8427 * This is called before wake_up_new_task(), therefore we really only
8428 * have to set its group bits, all the other stuff does not apply.
8430 static void cpu_cgroup_fork(struct task_struct *task)
8435 rq = task_rq_lock(task, &rf);
8437 sched_change_group(task, TASK_SET_GROUP);
8439 task_rq_unlock(rq, task, &rf);
8442 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8444 struct task_struct *task;
8445 struct cgroup_subsys_state *css;
8448 cgroup_taskset_for_each(task, css, tset) {
8449 #ifdef CONFIG_RT_GROUP_SCHED
8450 if (!sched_rt_can_attach(css_tg(css), task))
8453 /* We don't support RT-tasks being in separate groups */
8454 if (task->sched_class != &fair_sched_class)
8458 * Serialize against wake_up_new_task() such that if its
8459 * running, we're sure to observe its full state.
8461 raw_spin_lock_irq(&task->pi_lock);
8463 * Avoid calling sched_move_task() before wake_up_new_task()
8464 * has happened. This would lead to problems with PELT, due to
8465 * move wanting to detach+attach while we're not attached yet.
8467 if (task->state == TASK_NEW)
8469 raw_spin_unlock_irq(&task->pi_lock);
8477 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8479 struct task_struct *task;
8480 struct cgroup_subsys_state *css;
8482 cgroup_taskset_for_each(task, css, tset)
8483 sched_move_task(task);
8486 #ifdef CONFIG_FAIR_GROUP_SCHED
8487 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8488 struct cftype *cftype, u64 shareval)
8490 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8493 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8496 struct task_group *tg = css_tg(css);
8498 return (u64) scale_load_down(tg->shares);
8501 #ifdef CONFIG_CFS_BANDWIDTH
8502 static DEFINE_MUTEX(cfs_constraints_mutex);
8504 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8505 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8507 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8509 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8511 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8512 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8514 if (tg == &root_task_group)
8518 * Ensure we have at some amount of bandwidth every period. This is
8519 * to prevent reaching a state of large arrears when throttled via
8520 * entity_tick() resulting in prolonged exit starvation.
8522 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8526 * Likewise, bound things on the otherside by preventing insane quota
8527 * periods. This also allows us to normalize in computing quota
8530 if (period > max_cfs_quota_period)
8534 * Prevent race between setting of cfs_rq->runtime_enabled and
8535 * unthrottle_offline_cfs_rqs().
8538 mutex_lock(&cfs_constraints_mutex);
8539 ret = __cfs_schedulable(tg, period, quota);
8543 runtime_enabled = quota != RUNTIME_INF;
8544 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8546 * If we need to toggle cfs_bandwidth_used, off->on must occur
8547 * before making related changes, and on->off must occur afterwards
8549 if (runtime_enabled && !runtime_was_enabled)
8550 cfs_bandwidth_usage_inc();
8551 raw_spin_lock_irq(&cfs_b->lock);
8552 cfs_b->period = ns_to_ktime(period);
8553 cfs_b->quota = quota;
8555 __refill_cfs_bandwidth_runtime(cfs_b);
8556 /* restart the period timer (if active) to handle new period expiry */
8557 if (runtime_enabled)
8558 start_cfs_bandwidth(cfs_b);
8559 raw_spin_unlock_irq(&cfs_b->lock);
8561 for_each_online_cpu(i) {
8562 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8563 struct rq *rq = cfs_rq->rq;
8565 raw_spin_lock_irq(&rq->lock);
8566 cfs_rq->runtime_enabled = runtime_enabled;
8567 cfs_rq->runtime_remaining = 0;
8569 if (cfs_rq->throttled)
8570 unthrottle_cfs_rq(cfs_rq);
8571 raw_spin_unlock_irq(&rq->lock);
8573 if (runtime_was_enabled && !runtime_enabled)
8574 cfs_bandwidth_usage_dec();
8576 mutex_unlock(&cfs_constraints_mutex);
8582 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8586 period = ktime_to_ns(tg->cfs_bandwidth.period);
8587 if (cfs_quota_us < 0)
8588 quota = RUNTIME_INF;
8590 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8592 return tg_set_cfs_bandwidth(tg, period, quota);
8595 long tg_get_cfs_quota(struct task_group *tg)
8599 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8602 quota_us = tg->cfs_bandwidth.quota;
8603 do_div(quota_us, NSEC_PER_USEC);
8608 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8612 period = (u64)cfs_period_us * NSEC_PER_USEC;
8613 quota = tg->cfs_bandwidth.quota;
8615 return tg_set_cfs_bandwidth(tg, period, quota);
8618 long tg_get_cfs_period(struct task_group *tg)
8622 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8623 do_div(cfs_period_us, NSEC_PER_USEC);
8625 return cfs_period_us;
8628 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8631 return tg_get_cfs_quota(css_tg(css));
8634 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8635 struct cftype *cftype, s64 cfs_quota_us)
8637 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8640 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8643 return tg_get_cfs_period(css_tg(css));
8646 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8647 struct cftype *cftype, u64 cfs_period_us)
8649 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8652 struct cfs_schedulable_data {
8653 struct task_group *tg;
8658 * normalize group quota/period to be quota/max_period
8659 * note: units are usecs
8661 static u64 normalize_cfs_quota(struct task_group *tg,
8662 struct cfs_schedulable_data *d)
8670 period = tg_get_cfs_period(tg);
8671 quota = tg_get_cfs_quota(tg);
8674 /* note: these should typically be equivalent */
8675 if (quota == RUNTIME_INF || quota == -1)
8678 return to_ratio(period, quota);
8681 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8683 struct cfs_schedulable_data *d = data;
8684 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8685 s64 quota = 0, parent_quota = -1;
8688 quota = RUNTIME_INF;
8690 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8692 quota = normalize_cfs_quota(tg, d);
8693 parent_quota = parent_b->hierarchical_quota;
8696 * ensure max(child_quota) <= parent_quota, inherit when no
8699 if (quota == RUNTIME_INF)
8700 quota = parent_quota;
8701 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8704 cfs_b->hierarchical_quota = quota;
8709 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8712 struct cfs_schedulable_data data = {
8718 if (quota != RUNTIME_INF) {
8719 do_div(data.period, NSEC_PER_USEC);
8720 do_div(data.quota, NSEC_PER_USEC);
8724 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8730 static int cpu_stats_show(struct seq_file *sf, void *v)
8732 struct task_group *tg = css_tg(seq_css(sf));
8733 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8735 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8736 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8737 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8741 #endif /* CONFIG_CFS_BANDWIDTH */
8742 #endif /* CONFIG_FAIR_GROUP_SCHED */
8744 #ifdef CONFIG_RT_GROUP_SCHED
8745 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8746 struct cftype *cft, s64 val)
8748 return sched_group_set_rt_runtime(css_tg(css), val);
8751 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8754 return sched_group_rt_runtime(css_tg(css));
8757 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8758 struct cftype *cftype, u64 rt_period_us)
8760 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8763 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8766 return sched_group_rt_period(css_tg(css));
8768 #endif /* CONFIG_RT_GROUP_SCHED */
8770 static struct cftype cpu_files[] = {
8771 #ifdef CONFIG_FAIR_GROUP_SCHED
8774 .read_u64 = cpu_shares_read_u64,
8775 .write_u64 = cpu_shares_write_u64,
8778 #ifdef CONFIG_CFS_BANDWIDTH
8780 .name = "cfs_quota_us",
8781 .read_s64 = cpu_cfs_quota_read_s64,
8782 .write_s64 = cpu_cfs_quota_write_s64,
8785 .name = "cfs_period_us",
8786 .read_u64 = cpu_cfs_period_read_u64,
8787 .write_u64 = cpu_cfs_period_write_u64,
8791 .seq_show = cpu_stats_show,
8794 #ifdef CONFIG_RT_GROUP_SCHED
8796 .name = "rt_runtime_us",
8797 .read_s64 = cpu_rt_runtime_read,
8798 .write_s64 = cpu_rt_runtime_write,
8801 .name = "rt_period_us",
8802 .read_u64 = cpu_rt_period_read_uint,
8803 .write_u64 = cpu_rt_period_write_uint,
8809 struct cgroup_subsys cpu_cgrp_subsys = {
8810 .css_alloc = cpu_cgroup_css_alloc,
8811 .css_released = cpu_cgroup_css_released,
8812 .css_free = cpu_cgroup_css_free,
8813 .fork = cpu_cgroup_fork,
8814 .can_attach = cpu_cgroup_can_attach,
8815 .attach = cpu_cgroup_attach,
8816 .legacy_cftypes = cpu_files,
8820 #endif /* CONFIG_CGROUP_SCHED */
8822 void dump_cpu_task(int cpu)
8824 pr_info("Task dump for CPU %d:\n", cpu);
8825 sched_show_task(cpu_curr(cpu));
8829 * Nice levels are multiplicative, with a gentle 10% change for every
8830 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8831 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8832 * that remained on nice 0.
8834 * The "10% effect" is relative and cumulative: from _any_ nice level,
8835 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8836 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8837 * If a task goes up by ~10% and another task goes down by ~10% then
8838 * the relative distance between them is ~25%.)
8840 const int sched_prio_to_weight[40] = {
8841 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8842 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8843 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8844 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8845 /* 0 */ 1024, 820, 655, 526, 423,
8846 /* 5 */ 335, 272, 215, 172, 137,
8847 /* 10 */ 110, 87, 70, 56, 45,
8848 /* 15 */ 36, 29, 23, 18, 15,
8852 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8854 * In cases where the weight does not change often, we can use the
8855 * precalculated inverse to speed up arithmetics by turning divisions
8856 * into multiplications:
8858 const u32 sched_prio_to_wmult[40] = {
8859 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8860 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8861 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8862 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8863 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8864 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8865 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8866 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,