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>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
128 * Number of tasks to iterate in a single balance run.
129 * Limited because this is done with IRQs disabled.
131 const_debug unsigned int sysctl_sched_nr_migrate = 32;
134 * period over which we average the RT time consumption, measured
139 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
142 * period over which we measure -rt task cpu usage in us.
145 unsigned int sysctl_sched_rt_period = 1000000;
147 __read_mostly int scheduler_running;
150 * part of the period that we allow rt tasks to run in us.
153 int sysctl_sched_rt_runtime = 950000;
155 /* cpus with isolated domains */
156 cpumask_var_t cpu_isolated_map;
159 * this_rq_lock - lock this runqueue and disable interrupts.
161 static struct rq *this_rq_lock(void)
168 raw_spin_lock(&rq->lock);
174 * __task_rq_lock - lock the rq @p resides on.
176 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
181 lockdep_assert_held(&p->pi_lock);
185 raw_spin_lock(&rq->lock);
186 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
187 rf->cookie = lockdep_pin_lock(&rq->lock);
190 raw_spin_unlock(&rq->lock);
192 while (unlikely(task_on_rq_migrating(p)))
198 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
200 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
201 __acquires(p->pi_lock)
207 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
209 raw_spin_lock(&rq->lock);
211 * move_queued_task() task_rq_lock()
214 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
215 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
216 * [S] ->cpu = new_cpu [L] task_rq()
220 * If we observe the old cpu in task_rq_lock, the acquire of
221 * the old rq->lock will fully serialize against the stores.
223 * If we observe the new cpu in task_rq_lock, the acquire will
224 * pair with the WMB to ensure we must then also see migrating.
226 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
227 rf->cookie = lockdep_pin_lock(&rq->lock);
230 raw_spin_unlock(&rq->lock);
231 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
233 while (unlikely(task_on_rq_migrating(p)))
238 #ifdef CONFIG_SCHED_HRTICK
240 * Use HR-timers to deliver accurate preemption points.
243 static void hrtick_clear(struct rq *rq)
245 if (hrtimer_active(&rq->hrtick_timer))
246 hrtimer_cancel(&rq->hrtick_timer);
250 * High-resolution timer tick.
251 * Runs from hardirq context with interrupts disabled.
253 static enum hrtimer_restart hrtick(struct hrtimer *timer)
255 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
257 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
259 raw_spin_lock(&rq->lock);
261 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
262 raw_spin_unlock(&rq->lock);
264 return HRTIMER_NORESTART;
269 static void __hrtick_restart(struct rq *rq)
271 struct hrtimer *timer = &rq->hrtick_timer;
273 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
277 * called from hardirq (IPI) context
279 static void __hrtick_start(void *arg)
283 raw_spin_lock(&rq->lock);
284 __hrtick_restart(rq);
285 rq->hrtick_csd_pending = 0;
286 raw_spin_unlock(&rq->lock);
290 * Called to set the hrtick timer state.
292 * called with rq->lock held and irqs disabled
294 void hrtick_start(struct rq *rq, u64 delay)
296 struct hrtimer *timer = &rq->hrtick_timer;
301 * Don't schedule slices shorter than 10000ns, that just
302 * doesn't make sense and can cause timer DoS.
304 delta = max_t(s64, delay, 10000LL);
305 time = ktime_add_ns(timer->base->get_time(), delta);
307 hrtimer_set_expires(timer, time);
309 if (rq == this_rq()) {
310 __hrtick_restart(rq);
311 } else if (!rq->hrtick_csd_pending) {
312 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
313 rq->hrtick_csd_pending = 1;
319 * Called to set the hrtick timer state.
321 * called with rq->lock held and irqs disabled
323 void hrtick_start(struct rq *rq, u64 delay)
326 * Don't schedule slices shorter than 10000ns, that just
327 * doesn't make sense. Rely on vruntime for fairness.
329 delay = max_t(u64, delay, 10000LL);
330 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
331 HRTIMER_MODE_REL_PINNED);
333 #endif /* CONFIG_SMP */
335 static void init_rq_hrtick(struct rq *rq)
338 rq->hrtick_csd_pending = 0;
340 rq->hrtick_csd.flags = 0;
341 rq->hrtick_csd.func = __hrtick_start;
342 rq->hrtick_csd.info = rq;
345 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
346 rq->hrtick_timer.function = hrtick;
348 #else /* CONFIG_SCHED_HRTICK */
349 static inline void hrtick_clear(struct rq *rq)
353 static inline void init_rq_hrtick(struct rq *rq)
356 #endif /* CONFIG_SCHED_HRTICK */
359 * cmpxchg based fetch_or, macro so it works for different integer types
361 #define fetch_or(ptr, mask) \
363 typeof(ptr) _ptr = (ptr); \
364 typeof(mask) _mask = (mask); \
365 typeof(*_ptr) _old, _val = *_ptr; \
368 _old = cmpxchg(_ptr, _val, _val | _mask); \
376 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
378 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
379 * this avoids any races wrt polling state changes and thereby avoids
382 static bool set_nr_and_not_polling(struct task_struct *p)
384 struct thread_info *ti = task_thread_info(p);
385 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
389 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
391 * If this returns true, then the idle task promises to call
392 * sched_ttwu_pending() and reschedule soon.
394 static bool set_nr_if_polling(struct task_struct *p)
396 struct thread_info *ti = task_thread_info(p);
397 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
400 if (!(val & _TIF_POLLING_NRFLAG))
402 if (val & _TIF_NEED_RESCHED)
404 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
413 static bool set_nr_and_not_polling(struct task_struct *p)
415 set_tsk_need_resched(p);
420 static bool set_nr_if_polling(struct task_struct *p)
427 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
429 struct wake_q_node *node = &task->wake_q;
432 * Atomically grab the task, if ->wake_q is !nil already it means
433 * its already queued (either by us or someone else) and will get the
434 * wakeup due to that.
436 * This cmpxchg() implies a full barrier, which pairs with the write
437 * barrier implied by the wakeup in wake_up_q().
439 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
442 get_task_struct(task);
445 * The head is context local, there can be no concurrency.
448 head->lastp = &node->next;
451 void wake_up_q(struct wake_q_head *head)
453 struct wake_q_node *node = head->first;
455 while (node != WAKE_Q_TAIL) {
456 struct task_struct *task;
458 task = container_of(node, struct task_struct, wake_q);
460 /* task can safely be re-inserted now */
462 task->wake_q.next = NULL;
465 * wake_up_process() implies a wmb() to pair with the queueing
466 * in wake_q_add() so as not to miss wakeups.
468 wake_up_process(task);
469 put_task_struct(task);
474 * resched_curr - mark rq's current task 'to be rescheduled now'.
476 * On UP this means the setting of the need_resched flag, on SMP it
477 * might also involve a cross-CPU call to trigger the scheduler on
480 void resched_curr(struct rq *rq)
482 struct task_struct *curr = rq->curr;
485 lockdep_assert_held(&rq->lock);
487 if (test_tsk_need_resched(curr))
492 if (cpu == smp_processor_id()) {
493 set_tsk_need_resched(curr);
494 set_preempt_need_resched();
498 if (set_nr_and_not_polling(curr))
499 smp_send_reschedule(cpu);
501 trace_sched_wake_idle_without_ipi(cpu);
504 void resched_cpu(int cpu)
506 struct rq *rq = cpu_rq(cpu);
509 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
512 raw_spin_unlock_irqrestore(&rq->lock, flags);
516 #ifdef CONFIG_NO_HZ_COMMON
518 * In the semi idle case, use the nearest busy cpu for migrating timers
519 * from an idle cpu. This is good for power-savings.
521 * We don't do similar optimization for completely idle system, as
522 * selecting an idle cpu will add more delays to the timers than intended
523 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
525 int get_nohz_timer_target(void)
527 int i, cpu = smp_processor_id();
528 struct sched_domain *sd;
530 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
534 for_each_domain(cpu, sd) {
535 for_each_cpu(i, sched_domain_span(sd)) {
539 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
546 if (!is_housekeeping_cpu(cpu))
547 cpu = housekeeping_any_cpu();
553 * When add_timer_on() enqueues a timer into the timer wheel of an
554 * idle CPU then this timer might expire before the next timer event
555 * which is scheduled to wake up that CPU. In case of a completely
556 * idle system the next event might even be infinite time into the
557 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
558 * leaves the inner idle loop so the newly added timer is taken into
559 * account when the CPU goes back to idle and evaluates the timer
560 * wheel for the next timer event.
562 static void wake_up_idle_cpu(int cpu)
564 struct rq *rq = cpu_rq(cpu);
566 if (cpu == smp_processor_id())
569 if (set_nr_and_not_polling(rq->idle))
570 smp_send_reschedule(cpu);
572 trace_sched_wake_idle_without_ipi(cpu);
575 static bool wake_up_full_nohz_cpu(int cpu)
578 * We just need the target to call irq_exit() and re-evaluate
579 * the next tick. The nohz full kick at least implies that.
580 * If needed we can still optimize that later with an
583 if (tick_nohz_full_cpu(cpu)) {
584 if (cpu != smp_processor_id() ||
585 tick_nohz_tick_stopped())
586 tick_nohz_full_kick_cpu(cpu);
593 void wake_up_nohz_cpu(int cpu)
595 if (!wake_up_full_nohz_cpu(cpu))
596 wake_up_idle_cpu(cpu);
599 static inline bool got_nohz_idle_kick(void)
601 int cpu = smp_processor_id();
603 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
606 if (idle_cpu(cpu) && !need_resched())
610 * We can't run Idle Load Balance on this CPU for this time so we
611 * cancel it and clear NOHZ_BALANCE_KICK
613 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
617 #else /* CONFIG_NO_HZ_COMMON */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ_COMMON */
626 #ifdef CONFIG_NO_HZ_FULL
627 bool sched_can_stop_tick(struct rq *rq)
631 /* Deadline tasks, even if single, need the tick */
632 if (rq->dl.dl_nr_running)
636 * If there are more than one RR tasks, we need the tick to effect the
637 * actual RR behaviour.
639 if (rq->rt.rr_nr_running) {
640 if (rq->rt.rr_nr_running == 1)
647 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
648 * forced preemption between FIFO tasks.
650 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
655 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
656 * if there's more than one we need the tick for involuntary
659 if (rq->nr_running > 1)
664 #endif /* CONFIG_NO_HZ_FULL */
666 void sched_avg_update(struct rq *rq)
668 s64 period = sched_avg_period();
670 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
672 * Inline assembly required to prevent the compiler
673 * optimising this loop into a divmod call.
674 * See __iter_div_u64_rem() for another example of this.
676 asm("" : "+rm" (rq->age_stamp));
677 rq->age_stamp += period;
682 #endif /* CONFIG_SMP */
684 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
685 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
687 * Iterate task_group tree rooted at *from, calling @down when first entering a
688 * node and @up when leaving it for the final time.
690 * Caller must hold rcu_lock or sufficient equivalent.
692 int walk_tg_tree_from(struct task_group *from,
693 tg_visitor down, tg_visitor up, void *data)
695 struct task_group *parent, *child;
701 ret = (*down)(parent, data);
704 list_for_each_entry_rcu(child, &parent->children, siblings) {
711 ret = (*up)(parent, data);
712 if (ret || parent == from)
716 parent = parent->parent;
723 int tg_nop(struct task_group *tg, void *data)
729 static void set_load_weight(struct task_struct *p)
731 int prio = p->static_prio - MAX_RT_PRIO;
732 struct load_weight *load = &p->se.load;
735 * SCHED_IDLE tasks get minimal weight:
737 if (idle_policy(p->policy)) {
738 load->weight = scale_load(WEIGHT_IDLEPRIO);
739 load->inv_weight = WMULT_IDLEPRIO;
743 load->weight = scale_load(sched_prio_to_weight[prio]);
744 load->inv_weight = sched_prio_to_wmult[prio];
747 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
750 if (!(flags & ENQUEUE_RESTORE))
751 sched_info_queued(rq, p);
752 p->sched_class->enqueue_task(rq, p, flags);
755 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
758 if (!(flags & DEQUEUE_SAVE))
759 sched_info_dequeued(rq, p);
760 p->sched_class->dequeue_task(rq, p, flags);
763 void activate_task(struct rq *rq, struct task_struct *p, int flags)
765 if (task_contributes_to_load(p))
766 rq->nr_uninterruptible--;
768 enqueue_task(rq, p, flags);
771 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
773 if (task_contributes_to_load(p))
774 rq->nr_uninterruptible++;
776 dequeue_task(rq, p, flags);
779 static void update_rq_clock_task(struct rq *rq, s64 delta)
782 * In theory, the compile should just see 0 here, and optimize out the call
783 * to sched_rt_avg_update. But I don't trust it...
785 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
786 s64 steal = 0, irq_delta = 0;
788 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
789 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
792 * Since irq_time is only updated on {soft,}irq_exit, we might run into
793 * this case when a previous update_rq_clock() happened inside a
796 * When this happens, we stop ->clock_task and only update the
797 * prev_irq_time stamp to account for the part that fit, so that a next
798 * update will consume the rest. This ensures ->clock_task is
801 * It does however cause some slight miss-attribution of {soft,}irq
802 * time, a more accurate solution would be to update the irq_time using
803 * the current rq->clock timestamp, except that would require using
806 if (irq_delta > delta)
809 rq->prev_irq_time += irq_delta;
812 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
813 if (static_key_false((¶virt_steal_rq_enabled))) {
814 steal = paravirt_steal_clock(cpu_of(rq));
815 steal -= rq->prev_steal_time_rq;
817 if (unlikely(steal > delta))
820 rq->prev_steal_time_rq += steal;
825 rq->clock_task += delta;
827 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
828 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
829 sched_rt_avg_update(rq, irq_delta + steal);
833 void sched_set_stop_task(int cpu, struct task_struct *stop)
835 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
836 struct task_struct *old_stop = cpu_rq(cpu)->stop;
840 * Make it appear like a SCHED_FIFO task, its something
841 * userspace knows about and won't get confused about.
843 * Also, it will make PI more or less work without too
844 * much confusion -- but then, stop work should not
845 * rely on PI working anyway.
847 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
849 stop->sched_class = &stop_sched_class;
852 cpu_rq(cpu)->stop = stop;
856 * Reset it back to a normal scheduling class so that
857 * it can die in pieces.
859 old_stop->sched_class = &rt_sched_class;
864 * __normal_prio - return the priority that is based on the static prio
866 static inline int __normal_prio(struct task_struct *p)
868 return p->static_prio;
872 * Calculate the expected normal priority: i.e. priority
873 * without taking RT-inheritance into account. Might be
874 * boosted by interactivity modifiers. Changes upon fork,
875 * setprio syscalls, and whenever the interactivity
876 * estimator recalculates.
878 static inline int normal_prio(struct task_struct *p)
882 if (task_has_dl_policy(p))
883 prio = MAX_DL_PRIO-1;
884 else if (task_has_rt_policy(p))
885 prio = MAX_RT_PRIO-1 - p->rt_priority;
887 prio = __normal_prio(p);
892 * Calculate the current priority, i.e. the priority
893 * taken into account by the scheduler. This value might
894 * be boosted by RT tasks, or might be boosted by
895 * interactivity modifiers. Will be RT if the task got
896 * RT-boosted. If not then it returns p->normal_prio.
898 static int effective_prio(struct task_struct *p)
900 p->normal_prio = normal_prio(p);
902 * If we are RT tasks or we were boosted to RT priority,
903 * keep the priority unchanged. Otherwise, update priority
904 * to the normal priority:
906 if (!rt_prio(p->prio))
907 return p->normal_prio;
912 * task_curr - is this task currently executing on a CPU?
913 * @p: the task in question.
915 * Return: 1 if the task is currently executing. 0 otherwise.
917 inline int task_curr(const struct task_struct *p)
919 return cpu_curr(task_cpu(p)) == p;
923 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
924 * use the balance_callback list if you want balancing.
926 * this means any call to check_class_changed() must be followed by a call to
927 * balance_callback().
929 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
930 const struct sched_class *prev_class,
933 if (prev_class != p->sched_class) {
934 if (prev_class->switched_from)
935 prev_class->switched_from(rq, p);
937 p->sched_class->switched_to(rq, p);
938 } else if (oldprio != p->prio || dl_task(p))
939 p->sched_class->prio_changed(rq, p, oldprio);
942 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
944 const struct sched_class *class;
946 if (p->sched_class == rq->curr->sched_class) {
947 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
949 for_each_class(class) {
950 if (class == rq->curr->sched_class)
952 if (class == p->sched_class) {
960 * A queue event has occurred, and we're going to schedule. In
961 * this case, we can save a useless back to back clock update.
963 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
964 rq_clock_skip_update(rq, true);
969 * This is how migration works:
971 * 1) we invoke migration_cpu_stop() on the target CPU using
973 * 2) stopper starts to run (implicitly forcing the migrated thread
975 * 3) it checks whether the migrated task is still in the wrong runqueue.
976 * 4) if it's in the wrong runqueue then the migration thread removes
977 * it and puts it into the right queue.
978 * 5) stopper completes and stop_one_cpu() returns and the migration
983 * move_queued_task - move a queued task to new rq.
985 * Returns (locked) new rq. Old rq's lock is released.
987 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
989 lockdep_assert_held(&rq->lock);
991 p->on_rq = TASK_ON_RQ_MIGRATING;
992 dequeue_task(rq, p, 0);
993 set_task_cpu(p, new_cpu);
994 raw_spin_unlock(&rq->lock);
996 rq = cpu_rq(new_cpu);
998 raw_spin_lock(&rq->lock);
999 BUG_ON(task_cpu(p) != new_cpu);
1000 enqueue_task(rq, p, 0);
1001 p->on_rq = TASK_ON_RQ_QUEUED;
1002 check_preempt_curr(rq, p, 0);
1007 struct migration_arg {
1008 struct task_struct *task;
1013 * Move (not current) task off this cpu, onto dest cpu. We're doing
1014 * this because either it can't run here any more (set_cpus_allowed()
1015 * away from this CPU, or CPU going down), or because we're
1016 * attempting to rebalance this task on exec (sched_exec).
1018 * So we race with normal scheduler movements, but that's OK, as long
1019 * as the task is no longer on this CPU.
1021 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1023 if (unlikely(!cpu_active(dest_cpu)))
1026 /* Affinity changed (again). */
1027 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1030 rq = move_queued_task(rq, p, dest_cpu);
1036 * migration_cpu_stop - this will be executed by a highprio stopper thread
1037 * and performs thread migration by bumping thread off CPU then
1038 * 'pushing' onto another runqueue.
1040 static int migration_cpu_stop(void *data)
1042 struct migration_arg *arg = data;
1043 struct task_struct *p = arg->task;
1044 struct rq *rq = this_rq();
1047 * The original target cpu might have gone down and we might
1048 * be on another cpu but it doesn't matter.
1050 local_irq_disable();
1052 * We need to explicitly wake pending tasks before running
1053 * __migrate_task() such that we will not miss enforcing cpus_allowed
1054 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1056 sched_ttwu_pending();
1058 raw_spin_lock(&p->pi_lock);
1059 raw_spin_lock(&rq->lock);
1061 * If task_rq(p) != rq, it cannot be migrated here, because we're
1062 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1063 * we're holding p->pi_lock.
1065 if (task_rq(p) == rq && task_on_rq_queued(p))
1066 rq = __migrate_task(rq, p, arg->dest_cpu);
1067 raw_spin_unlock(&rq->lock);
1068 raw_spin_unlock(&p->pi_lock);
1075 * sched_class::set_cpus_allowed must do the below, but is not required to
1076 * actually call this function.
1078 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1080 cpumask_copy(&p->cpus_allowed, new_mask);
1081 p->nr_cpus_allowed = cpumask_weight(new_mask);
1084 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1086 struct rq *rq = task_rq(p);
1087 bool queued, running;
1089 lockdep_assert_held(&p->pi_lock);
1091 queued = task_on_rq_queued(p);
1092 running = task_current(rq, p);
1096 * Because __kthread_bind() calls this on blocked tasks without
1099 lockdep_assert_held(&rq->lock);
1100 dequeue_task(rq, p, DEQUEUE_SAVE);
1103 put_prev_task(rq, p);
1105 p->sched_class->set_cpus_allowed(p, new_mask);
1108 p->sched_class->set_curr_task(rq);
1110 enqueue_task(rq, p, ENQUEUE_RESTORE);
1114 * Change a given task's CPU affinity. Migrate the thread to a
1115 * proper CPU and schedule it away if the CPU it's executing on
1116 * is removed from the allowed bitmask.
1118 * NOTE: the caller must have a valid reference to the task, the
1119 * task must not exit() & deallocate itself prematurely. The
1120 * call is not atomic; no spinlocks may be held.
1122 static int __set_cpus_allowed_ptr(struct task_struct *p,
1123 const struct cpumask *new_mask, bool check)
1125 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1126 unsigned int dest_cpu;
1131 rq = task_rq_lock(p, &rf);
1133 if (p->flags & PF_KTHREAD) {
1135 * Kernel threads are allowed on online && !active CPUs
1137 cpu_valid_mask = cpu_online_mask;
1141 * Must re-check here, to close a race against __kthread_bind(),
1142 * sched_setaffinity() is not guaranteed to observe the flag.
1144 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1149 if (cpumask_equal(&p->cpus_allowed, new_mask))
1152 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1157 do_set_cpus_allowed(p, new_mask);
1159 if (p->flags & PF_KTHREAD) {
1161 * For kernel threads that do indeed end up on online &&
1162 * !active we want to ensure they are strict per-cpu threads.
1164 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1165 !cpumask_intersects(new_mask, cpu_active_mask) &&
1166 p->nr_cpus_allowed != 1);
1169 /* Can the task run on the task's current CPU? If so, we're done */
1170 if (cpumask_test_cpu(task_cpu(p), new_mask))
1173 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1174 if (task_running(rq, p) || p->state == TASK_WAKING) {
1175 struct migration_arg arg = { p, dest_cpu };
1176 /* Need help from migration thread: drop lock and wait. */
1177 task_rq_unlock(rq, p, &rf);
1178 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1179 tlb_migrate_finish(p->mm);
1181 } else if (task_on_rq_queued(p)) {
1183 * OK, since we're going to drop the lock immediately
1184 * afterwards anyway.
1186 lockdep_unpin_lock(&rq->lock, rf.cookie);
1187 rq = move_queued_task(rq, p, dest_cpu);
1188 lockdep_repin_lock(&rq->lock, rf.cookie);
1191 task_rq_unlock(rq, p, &rf);
1196 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1198 return __set_cpus_allowed_ptr(p, new_mask, false);
1200 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1202 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1204 #ifdef CONFIG_SCHED_DEBUG
1206 * We should never call set_task_cpu() on a blocked task,
1207 * ttwu() will sort out the placement.
1209 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1213 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1214 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1215 * time relying on p->on_rq.
1217 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1218 p->sched_class == &fair_sched_class &&
1219 (p->on_rq && !task_on_rq_migrating(p)));
1221 #ifdef CONFIG_LOCKDEP
1223 * The caller should hold either p->pi_lock or rq->lock, when changing
1224 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1226 * sched_move_task() holds both and thus holding either pins the cgroup,
1229 * Furthermore, all task_rq users should acquire both locks, see
1232 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1233 lockdep_is_held(&task_rq(p)->lock)));
1237 trace_sched_migrate_task(p, new_cpu);
1239 if (task_cpu(p) != new_cpu) {
1240 if (p->sched_class->migrate_task_rq)
1241 p->sched_class->migrate_task_rq(p);
1242 p->se.nr_migrations++;
1243 perf_event_task_migrate(p);
1246 __set_task_cpu(p, new_cpu);
1249 static void __migrate_swap_task(struct task_struct *p, int cpu)
1251 if (task_on_rq_queued(p)) {
1252 struct rq *src_rq, *dst_rq;
1254 src_rq = task_rq(p);
1255 dst_rq = cpu_rq(cpu);
1257 p->on_rq = TASK_ON_RQ_MIGRATING;
1258 deactivate_task(src_rq, p, 0);
1259 set_task_cpu(p, cpu);
1260 activate_task(dst_rq, p, 0);
1261 p->on_rq = TASK_ON_RQ_QUEUED;
1262 check_preempt_curr(dst_rq, p, 0);
1265 * Task isn't running anymore; make it appear like we migrated
1266 * it before it went to sleep. This means on wakeup we make the
1267 * previous cpu our targer instead of where it really is.
1273 struct migration_swap_arg {
1274 struct task_struct *src_task, *dst_task;
1275 int src_cpu, dst_cpu;
1278 static int migrate_swap_stop(void *data)
1280 struct migration_swap_arg *arg = data;
1281 struct rq *src_rq, *dst_rq;
1284 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1287 src_rq = cpu_rq(arg->src_cpu);
1288 dst_rq = cpu_rq(arg->dst_cpu);
1290 double_raw_lock(&arg->src_task->pi_lock,
1291 &arg->dst_task->pi_lock);
1292 double_rq_lock(src_rq, dst_rq);
1294 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1297 if (task_cpu(arg->src_task) != arg->src_cpu)
1300 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1303 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1306 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1307 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1312 double_rq_unlock(src_rq, dst_rq);
1313 raw_spin_unlock(&arg->dst_task->pi_lock);
1314 raw_spin_unlock(&arg->src_task->pi_lock);
1320 * Cross migrate two tasks
1322 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1324 struct migration_swap_arg arg;
1327 arg = (struct migration_swap_arg){
1329 .src_cpu = task_cpu(cur),
1331 .dst_cpu = task_cpu(p),
1334 if (arg.src_cpu == arg.dst_cpu)
1338 * These three tests are all lockless; this is OK since all of them
1339 * will be re-checked with proper locks held further down the line.
1341 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1344 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1347 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1350 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1351 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1358 * wait_task_inactive - wait for a thread to unschedule.
1360 * If @match_state is nonzero, it's the @p->state value just checked and
1361 * not expected to change. If it changes, i.e. @p might have woken up,
1362 * then return zero. When we succeed in waiting for @p to be off its CPU,
1363 * we return a positive number (its total switch count). If a second call
1364 * a short while later returns the same number, the caller can be sure that
1365 * @p has remained unscheduled the whole time.
1367 * The caller must ensure that the task *will* unschedule sometime soon,
1368 * else this function might spin for a *long* time. This function can't
1369 * be called with interrupts off, or it may introduce deadlock with
1370 * smp_call_function() if an IPI is sent by the same process we are
1371 * waiting to become inactive.
1373 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1375 int running, queued;
1382 * We do the initial early heuristics without holding
1383 * any task-queue locks at all. We'll only try to get
1384 * the runqueue lock when things look like they will
1390 * If the task is actively running on another CPU
1391 * still, just relax and busy-wait without holding
1394 * NOTE! Since we don't hold any locks, it's not
1395 * even sure that "rq" stays as the right runqueue!
1396 * But we don't care, since "task_running()" will
1397 * return false if the runqueue has changed and p
1398 * is actually now running somewhere else!
1400 while (task_running(rq, p)) {
1401 if (match_state && unlikely(p->state != match_state))
1407 * Ok, time to look more closely! We need the rq
1408 * lock now, to be *sure*. If we're wrong, we'll
1409 * just go back and repeat.
1411 rq = task_rq_lock(p, &rf);
1412 trace_sched_wait_task(p);
1413 running = task_running(rq, p);
1414 queued = task_on_rq_queued(p);
1416 if (!match_state || p->state == match_state)
1417 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1418 task_rq_unlock(rq, p, &rf);
1421 * If it changed from the expected state, bail out now.
1423 if (unlikely(!ncsw))
1427 * Was it really running after all now that we
1428 * checked with the proper locks actually held?
1430 * Oops. Go back and try again..
1432 if (unlikely(running)) {
1438 * It's not enough that it's not actively running,
1439 * it must be off the runqueue _entirely_, and not
1442 * So if it was still runnable (but just not actively
1443 * running right now), it's preempted, and we should
1444 * yield - it could be a while.
1446 if (unlikely(queued)) {
1447 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1449 set_current_state(TASK_UNINTERRUPTIBLE);
1450 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1455 * Ahh, all good. It wasn't running, and it wasn't
1456 * runnable, which means that it will never become
1457 * running in the future either. We're all done!
1466 * kick_process - kick a running thread to enter/exit the kernel
1467 * @p: the to-be-kicked thread
1469 * Cause a process which is running on another CPU to enter
1470 * kernel-mode, without any delay. (to get signals handled.)
1472 * NOTE: this function doesn't have to take the runqueue lock,
1473 * because all it wants to ensure is that the remote task enters
1474 * the kernel. If the IPI races and the task has been migrated
1475 * to another CPU then no harm is done and the purpose has been
1478 void kick_process(struct task_struct *p)
1484 if ((cpu != smp_processor_id()) && task_curr(p))
1485 smp_send_reschedule(cpu);
1488 EXPORT_SYMBOL_GPL(kick_process);
1491 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1493 * A few notes on cpu_active vs cpu_online:
1495 * - cpu_active must be a subset of cpu_online
1497 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1498 * see __set_cpus_allowed_ptr(). At this point the newly online
1499 * cpu isn't yet part of the sched domains, and balancing will not
1502 * - on cpu-down we clear cpu_active() to mask the sched domains and
1503 * avoid the load balancer to place new tasks on the to be removed
1504 * cpu. Existing tasks will remain running there and will be taken
1507 * This means that fallback selection must not select !active CPUs.
1508 * And can assume that any active CPU must be online. Conversely
1509 * select_task_rq() below may allow selection of !active CPUs in order
1510 * to satisfy the above rules.
1512 static int select_fallback_rq(int cpu, struct task_struct *p)
1514 int nid = cpu_to_node(cpu);
1515 const struct cpumask *nodemask = NULL;
1516 enum { cpuset, possible, fail } state = cpuset;
1520 * If the node that the cpu is on has been offlined, cpu_to_node()
1521 * will return -1. There is no cpu on the node, and we should
1522 * select the cpu on the other node.
1525 nodemask = cpumask_of_node(nid);
1527 /* Look for allowed, online CPU in same node. */
1528 for_each_cpu(dest_cpu, nodemask) {
1529 if (!cpu_active(dest_cpu))
1531 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1537 /* Any allowed, online CPU? */
1538 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1539 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1541 if (!cpu_online(dest_cpu))
1546 /* No more Mr. Nice Guy. */
1549 if (IS_ENABLED(CONFIG_CPUSETS)) {
1550 cpuset_cpus_allowed_fallback(p);
1556 do_set_cpus_allowed(p, cpu_possible_mask);
1567 if (state != cpuset) {
1569 * Don't tell them about moving exiting tasks or
1570 * kernel threads (both mm NULL), since they never
1573 if (p->mm && printk_ratelimit()) {
1574 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1575 task_pid_nr(p), p->comm, cpu);
1583 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1586 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1588 lockdep_assert_held(&p->pi_lock);
1590 if (tsk_nr_cpus_allowed(p) > 1)
1591 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1593 cpu = cpumask_any(tsk_cpus_allowed(p));
1596 * In order not to call set_task_cpu() on a blocking task we need
1597 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1600 * Since this is common to all placement strategies, this lives here.
1602 * [ this allows ->select_task() to simply return task_cpu(p) and
1603 * not worry about this generic constraint ]
1605 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1607 cpu = select_fallback_rq(task_cpu(p), p);
1612 static void update_avg(u64 *avg, u64 sample)
1614 s64 diff = sample - *avg;
1620 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1621 const struct cpumask *new_mask, bool check)
1623 return set_cpus_allowed_ptr(p, new_mask);
1626 #endif /* CONFIG_SMP */
1629 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1631 #ifdef CONFIG_SCHEDSTATS
1632 struct rq *rq = this_rq();
1635 int this_cpu = smp_processor_id();
1637 if (cpu == this_cpu) {
1638 schedstat_inc(rq, ttwu_local);
1639 schedstat_inc(p, se.statistics.nr_wakeups_local);
1641 struct sched_domain *sd;
1643 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1645 for_each_domain(this_cpu, sd) {
1646 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1647 schedstat_inc(sd, ttwu_wake_remote);
1654 if (wake_flags & WF_MIGRATED)
1655 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1657 #endif /* CONFIG_SMP */
1659 schedstat_inc(rq, ttwu_count);
1660 schedstat_inc(p, se.statistics.nr_wakeups);
1662 if (wake_flags & WF_SYNC)
1663 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1665 #endif /* CONFIG_SCHEDSTATS */
1668 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1670 activate_task(rq, p, en_flags);
1671 p->on_rq = TASK_ON_RQ_QUEUED;
1673 /* if a worker is waking up, notify workqueue */
1674 if (p->flags & PF_WQ_WORKER)
1675 wq_worker_waking_up(p, cpu_of(rq));
1679 * Mark the task runnable and perform wakeup-preemption.
1681 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1682 struct pin_cookie cookie)
1684 check_preempt_curr(rq, p, wake_flags);
1685 p->state = TASK_RUNNING;
1686 trace_sched_wakeup(p);
1689 if (p->sched_class->task_woken) {
1691 * Our task @p is fully woken up and running; so its safe to
1692 * drop the rq->lock, hereafter rq is only used for statistics.
1694 lockdep_unpin_lock(&rq->lock, cookie);
1695 p->sched_class->task_woken(rq, p);
1696 lockdep_repin_lock(&rq->lock, cookie);
1699 if (rq->idle_stamp) {
1700 u64 delta = rq_clock(rq) - rq->idle_stamp;
1701 u64 max = 2*rq->max_idle_balance_cost;
1703 update_avg(&rq->avg_idle, delta);
1705 if (rq->avg_idle > max)
1714 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1715 struct pin_cookie cookie)
1717 int en_flags = ENQUEUE_WAKEUP;
1719 lockdep_assert_held(&rq->lock);
1722 if (p->sched_contributes_to_load)
1723 rq->nr_uninterruptible--;
1725 if (wake_flags & WF_MIGRATED)
1726 en_flags |= ENQUEUE_MIGRATED;
1729 ttwu_activate(rq, p, en_flags);
1730 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1734 * Called in case the task @p isn't fully descheduled from its runqueue,
1735 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1736 * since all we need to do is flip p->state to TASK_RUNNING, since
1737 * the task is still ->on_rq.
1739 static int ttwu_remote(struct task_struct *p, int wake_flags)
1745 rq = __task_rq_lock(p, &rf);
1746 if (task_on_rq_queued(p)) {
1747 /* check_preempt_curr() may use rq clock */
1748 update_rq_clock(rq);
1749 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1752 __task_rq_unlock(rq, &rf);
1758 void sched_ttwu_pending(void)
1760 struct rq *rq = this_rq();
1761 struct llist_node *llist = llist_del_all(&rq->wake_list);
1762 struct pin_cookie cookie;
1763 struct task_struct *p;
1764 unsigned long flags;
1769 raw_spin_lock_irqsave(&rq->lock, flags);
1770 cookie = lockdep_pin_lock(&rq->lock);
1775 p = llist_entry(llist, struct task_struct, wake_entry);
1776 llist = llist_next(llist);
1778 if (p->sched_remote_wakeup)
1779 wake_flags = WF_MIGRATED;
1781 ttwu_do_activate(rq, p, wake_flags, cookie);
1784 lockdep_unpin_lock(&rq->lock, cookie);
1785 raw_spin_unlock_irqrestore(&rq->lock, flags);
1788 void scheduler_ipi(void)
1791 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1792 * TIF_NEED_RESCHED remotely (for the first time) will also send
1795 preempt_fold_need_resched();
1797 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1801 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1802 * traditionally all their work was done from the interrupt return
1803 * path. Now that we actually do some work, we need to make sure
1806 * Some archs already do call them, luckily irq_enter/exit nest
1809 * Arguably we should visit all archs and update all handlers,
1810 * however a fair share of IPIs are still resched only so this would
1811 * somewhat pessimize the simple resched case.
1814 sched_ttwu_pending();
1817 * Check if someone kicked us for doing the nohz idle load balance.
1819 if (unlikely(got_nohz_idle_kick())) {
1820 this_rq()->idle_balance = 1;
1821 raise_softirq_irqoff(SCHED_SOFTIRQ);
1826 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1828 struct rq *rq = cpu_rq(cpu);
1830 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1832 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1833 if (!set_nr_if_polling(rq->idle))
1834 smp_send_reschedule(cpu);
1836 trace_sched_wake_idle_without_ipi(cpu);
1840 void wake_up_if_idle(int cpu)
1842 struct rq *rq = cpu_rq(cpu);
1843 unsigned long flags;
1847 if (!is_idle_task(rcu_dereference(rq->curr)))
1850 if (set_nr_if_polling(rq->idle)) {
1851 trace_sched_wake_idle_without_ipi(cpu);
1853 raw_spin_lock_irqsave(&rq->lock, flags);
1854 if (is_idle_task(rq->curr))
1855 smp_send_reschedule(cpu);
1856 /* Else cpu is not in idle, do nothing here */
1857 raw_spin_unlock_irqrestore(&rq->lock, flags);
1864 bool cpus_share_cache(int this_cpu, int that_cpu)
1866 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1868 #endif /* CONFIG_SMP */
1870 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1872 struct rq *rq = cpu_rq(cpu);
1873 struct pin_cookie cookie;
1875 #if defined(CONFIG_SMP)
1876 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1877 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1878 ttwu_queue_remote(p, cpu, wake_flags);
1883 raw_spin_lock(&rq->lock);
1884 cookie = lockdep_pin_lock(&rq->lock);
1885 ttwu_do_activate(rq, p, wake_flags, cookie);
1886 lockdep_unpin_lock(&rq->lock, cookie);
1887 raw_spin_unlock(&rq->lock);
1891 * Notes on Program-Order guarantees on SMP systems.
1895 * The basic program-order guarantee on SMP systems is that when a task [t]
1896 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1897 * execution on its new cpu [c1].
1899 * For migration (of runnable tasks) this is provided by the following means:
1901 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1902 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1903 * rq(c1)->lock (if not at the same time, then in that order).
1904 * C) LOCK of the rq(c1)->lock scheduling in task
1906 * Transitivity guarantees that B happens after A and C after B.
1907 * Note: we only require RCpc transitivity.
1908 * Note: the cpu doing B need not be c0 or c1
1917 * UNLOCK rq(0)->lock
1919 * LOCK rq(0)->lock // orders against CPU0
1921 * UNLOCK rq(0)->lock
1925 * UNLOCK rq(1)->lock
1927 * LOCK rq(1)->lock // orders against CPU2
1930 * UNLOCK rq(1)->lock
1933 * BLOCKING -- aka. SLEEP + WAKEUP
1935 * For blocking we (obviously) need to provide the same guarantee as for
1936 * migration. However the means are completely different as there is no lock
1937 * chain to provide order. Instead we do:
1939 * 1) smp_store_release(X->on_cpu, 0)
1940 * 2) smp_cond_acquire(!X->on_cpu)
1944 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1946 * LOCK rq(0)->lock LOCK X->pi_lock
1949 * smp_store_release(X->on_cpu, 0);
1951 * smp_cond_acquire(!X->on_cpu);
1957 * X->state = RUNNING
1958 * UNLOCK rq(2)->lock
1960 * LOCK rq(2)->lock // orders against CPU1
1963 * UNLOCK rq(2)->lock
1966 * UNLOCK rq(0)->lock
1969 * However; for wakeups there is a second guarantee we must provide, namely we
1970 * must observe the state that lead to our wakeup. That is, not only must our
1971 * task observe its own prior state, it must also observe the stores prior to
1974 * This means that any means of doing remote wakeups must order the CPU doing
1975 * the wakeup against the CPU the task is going to end up running on. This,
1976 * however, is already required for the regular Program-Order guarantee above,
1977 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1982 * try_to_wake_up - wake up a thread
1983 * @p: the thread to be awakened
1984 * @state: the mask of task states that can be woken
1985 * @wake_flags: wake modifier flags (WF_*)
1987 * Put it on the run-queue if it's not already there. The "current"
1988 * thread is always on the run-queue (except when the actual
1989 * re-schedule is in progress), and as such you're allowed to do
1990 * the simpler "current->state = TASK_RUNNING" to mark yourself
1991 * runnable without the overhead of this.
1993 * Return: %true if @p was woken up, %false if it was already running.
1994 * or @state didn't match @p's state.
1997 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1999 unsigned long flags;
2000 int cpu, success = 0;
2003 * If we are going to wake up a thread waiting for CONDITION we
2004 * need to ensure that CONDITION=1 done by the caller can not be
2005 * reordered with p->state check below. This pairs with mb() in
2006 * set_current_state() the waiting thread does.
2008 smp_mb__before_spinlock();
2009 raw_spin_lock_irqsave(&p->pi_lock, flags);
2010 if (!(p->state & state))
2013 trace_sched_waking(p);
2015 success = 1; /* we're going to change ->state */
2018 if (p->on_rq && ttwu_remote(p, wake_flags))
2023 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2024 * possible to, falsely, observe p->on_cpu == 0.
2026 * One must be running (->on_cpu == 1) in order to remove oneself
2027 * from the runqueue.
2029 * [S] ->on_cpu = 1; [L] ->on_rq
2033 * [S] ->on_rq = 0; [L] ->on_cpu
2035 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2036 * from the consecutive calls to schedule(); the first switching to our
2037 * task, the second putting it to sleep.
2042 * If the owning (remote) cpu is still in the middle of schedule() with
2043 * this task as prev, wait until its done referencing the task.
2045 * Pairs with the smp_store_release() in finish_lock_switch().
2047 * This ensures that tasks getting woken will be fully ordered against
2048 * their previous state and preserve Program Order.
2050 smp_cond_acquire(!p->on_cpu);
2052 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2053 p->state = TASK_WAKING;
2055 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2056 if (task_cpu(p) != cpu) {
2057 wake_flags |= WF_MIGRATED;
2058 set_task_cpu(p, cpu);
2060 #endif /* CONFIG_SMP */
2062 ttwu_queue(p, cpu, wake_flags);
2064 if (schedstat_enabled())
2065 ttwu_stat(p, cpu, wake_flags);
2067 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2073 * try_to_wake_up_local - try to wake up a local task with rq lock held
2074 * @p: the thread to be awakened
2076 * Put @p on the run-queue if it's not already there. The caller must
2077 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2080 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2082 struct rq *rq = task_rq(p);
2084 if (WARN_ON_ONCE(rq != this_rq()) ||
2085 WARN_ON_ONCE(p == current))
2088 lockdep_assert_held(&rq->lock);
2090 if (!raw_spin_trylock(&p->pi_lock)) {
2092 * This is OK, because current is on_cpu, which avoids it being
2093 * picked for load-balance and preemption/IRQs are still
2094 * disabled avoiding further scheduler activity on it and we've
2095 * not yet picked a replacement task.
2097 lockdep_unpin_lock(&rq->lock, cookie);
2098 raw_spin_unlock(&rq->lock);
2099 raw_spin_lock(&p->pi_lock);
2100 raw_spin_lock(&rq->lock);
2101 lockdep_repin_lock(&rq->lock, cookie);
2104 if (!(p->state & TASK_NORMAL))
2107 trace_sched_waking(p);
2109 if (!task_on_rq_queued(p))
2110 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2112 ttwu_do_wakeup(rq, p, 0, cookie);
2113 if (schedstat_enabled())
2114 ttwu_stat(p, smp_processor_id(), 0);
2116 raw_spin_unlock(&p->pi_lock);
2120 * wake_up_process - Wake up a specific process
2121 * @p: The process to be woken up.
2123 * Attempt to wake up the nominated process and move it to the set of runnable
2126 * Return: 1 if the process was woken up, 0 if it was already running.
2128 * It may be assumed that this function implies a write memory barrier before
2129 * changing the task state if and only if any tasks are woken up.
2131 int wake_up_process(struct task_struct *p)
2133 return try_to_wake_up(p, TASK_NORMAL, 0);
2135 EXPORT_SYMBOL(wake_up_process);
2137 int wake_up_state(struct task_struct *p, unsigned int state)
2139 return try_to_wake_up(p, state, 0);
2143 * This function clears the sched_dl_entity static params.
2145 void __dl_clear_params(struct task_struct *p)
2147 struct sched_dl_entity *dl_se = &p->dl;
2149 dl_se->dl_runtime = 0;
2150 dl_se->dl_deadline = 0;
2151 dl_se->dl_period = 0;
2155 dl_se->dl_throttled = 0;
2156 dl_se->dl_yielded = 0;
2160 * Perform scheduler related setup for a newly forked process p.
2161 * p is forked by current.
2163 * __sched_fork() is basic setup used by init_idle() too:
2165 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2170 p->se.exec_start = 0;
2171 p->se.sum_exec_runtime = 0;
2172 p->se.prev_sum_exec_runtime = 0;
2173 p->se.nr_migrations = 0;
2175 INIT_LIST_HEAD(&p->se.group_node);
2177 #ifdef CONFIG_FAIR_GROUP_SCHED
2178 p->se.cfs_rq = NULL;
2181 #ifdef CONFIG_SCHEDSTATS
2182 /* Even if schedstat is disabled, there should not be garbage */
2183 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2186 RB_CLEAR_NODE(&p->dl.rb_node);
2187 init_dl_task_timer(&p->dl);
2188 __dl_clear_params(p);
2190 INIT_LIST_HEAD(&p->rt.run_list);
2192 p->rt.time_slice = sched_rr_timeslice;
2196 #ifdef CONFIG_PREEMPT_NOTIFIERS
2197 INIT_HLIST_HEAD(&p->preempt_notifiers);
2200 #ifdef CONFIG_NUMA_BALANCING
2201 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2202 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2203 p->mm->numa_scan_seq = 0;
2206 if (clone_flags & CLONE_VM)
2207 p->numa_preferred_nid = current->numa_preferred_nid;
2209 p->numa_preferred_nid = -1;
2211 p->node_stamp = 0ULL;
2212 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2213 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2214 p->numa_work.next = &p->numa_work;
2215 p->numa_faults = NULL;
2216 p->last_task_numa_placement = 0;
2217 p->last_sum_exec_runtime = 0;
2219 p->numa_group = NULL;
2220 #endif /* CONFIG_NUMA_BALANCING */
2223 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2225 #ifdef CONFIG_NUMA_BALANCING
2227 void set_numabalancing_state(bool enabled)
2230 static_branch_enable(&sched_numa_balancing);
2232 static_branch_disable(&sched_numa_balancing);
2235 #ifdef CONFIG_PROC_SYSCTL
2236 int sysctl_numa_balancing(struct ctl_table *table, int write,
2237 void __user *buffer, size_t *lenp, loff_t *ppos)
2241 int state = static_branch_likely(&sched_numa_balancing);
2243 if (write && !capable(CAP_SYS_ADMIN))
2248 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2252 set_numabalancing_state(state);
2258 #ifdef CONFIG_SCHEDSTATS
2260 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2261 static bool __initdata __sched_schedstats = false;
2263 static void set_schedstats(bool enabled)
2266 static_branch_enable(&sched_schedstats);
2268 static_branch_disable(&sched_schedstats);
2271 void force_schedstat_enabled(void)
2273 if (!schedstat_enabled()) {
2274 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2275 static_branch_enable(&sched_schedstats);
2279 static int __init setup_schedstats(char *str)
2286 * This code is called before jump labels have been set up, so we can't
2287 * change the static branch directly just yet. Instead set a temporary
2288 * variable so init_schedstats() can do it later.
2290 if (!strcmp(str, "enable")) {
2291 __sched_schedstats = true;
2293 } else if (!strcmp(str, "disable")) {
2294 __sched_schedstats = false;
2299 pr_warn("Unable to parse schedstats=\n");
2303 __setup("schedstats=", setup_schedstats);
2305 static void __init init_schedstats(void)
2307 set_schedstats(__sched_schedstats);
2310 #ifdef CONFIG_PROC_SYSCTL
2311 int sysctl_schedstats(struct ctl_table *table, int write,
2312 void __user *buffer, size_t *lenp, loff_t *ppos)
2316 int state = static_branch_likely(&sched_schedstats);
2318 if (write && !capable(CAP_SYS_ADMIN))
2323 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2327 set_schedstats(state);
2330 #endif /* CONFIG_PROC_SYSCTL */
2331 #else /* !CONFIG_SCHEDSTATS */
2332 static inline void init_schedstats(void) {}
2333 #endif /* CONFIG_SCHEDSTATS */
2336 * fork()/clone()-time setup:
2338 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2340 unsigned long flags;
2341 int cpu = get_cpu();
2343 __sched_fork(clone_flags, p);
2345 * We mark the process as running here. This guarantees that
2346 * nobody will actually run it, and a signal or other external
2347 * event cannot wake it up and insert it on the runqueue either.
2349 p->state = TASK_RUNNING;
2352 * Make sure we do not leak PI boosting priority to the child.
2354 p->prio = current->normal_prio;
2357 * Revert to default priority/policy on fork if requested.
2359 if (unlikely(p->sched_reset_on_fork)) {
2360 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2361 p->policy = SCHED_NORMAL;
2362 p->static_prio = NICE_TO_PRIO(0);
2364 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2365 p->static_prio = NICE_TO_PRIO(0);
2367 p->prio = p->normal_prio = __normal_prio(p);
2371 * We don't need the reset flag anymore after the fork. It has
2372 * fulfilled its duty:
2374 p->sched_reset_on_fork = 0;
2377 if (dl_prio(p->prio)) {
2380 } else if (rt_prio(p->prio)) {
2381 p->sched_class = &rt_sched_class;
2383 p->sched_class = &fair_sched_class;
2387 * The child is not yet in the pid-hash so no cgroup attach races,
2388 * and the cgroup is pinned to this child due to cgroup_fork()
2389 * is ran before sched_fork().
2391 * Silence PROVE_RCU.
2393 raw_spin_lock_irqsave(&p->pi_lock, flags);
2395 * We're setting the cpu for the first time, we don't migrate,
2396 * so use __set_task_cpu().
2398 __set_task_cpu(p, cpu);
2399 if (p->sched_class->task_fork)
2400 p->sched_class->task_fork(p);
2401 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2403 #ifdef CONFIG_SCHED_INFO
2404 if (likely(sched_info_on()))
2405 memset(&p->sched_info, 0, sizeof(p->sched_info));
2407 #if defined(CONFIG_SMP)
2410 init_task_preempt_count(p);
2412 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2413 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2420 unsigned long to_ratio(u64 period, u64 runtime)
2422 if (runtime == RUNTIME_INF)
2426 * Doing this here saves a lot of checks in all
2427 * the calling paths, and returning zero seems
2428 * safe for them anyway.
2433 return div64_u64(runtime << 20, period);
2437 inline struct dl_bw *dl_bw_of(int i)
2439 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2440 "sched RCU must be held");
2441 return &cpu_rq(i)->rd->dl_bw;
2444 static inline int dl_bw_cpus(int i)
2446 struct root_domain *rd = cpu_rq(i)->rd;
2449 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2450 "sched RCU must be held");
2451 for_each_cpu_and(i, rd->span, cpu_active_mask)
2457 inline struct dl_bw *dl_bw_of(int i)
2459 return &cpu_rq(i)->dl.dl_bw;
2462 static inline int dl_bw_cpus(int i)
2469 * We must be sure that accepting a new task (or allowing changing the
2470 * parameters of an existing one) is consistent with the bandwidth
2471 * constraints. If yes, this function also accordingly updates the currently
2472 * allocated bandwidth to reflect the new situation.
2474 * This function is called while holding p's rq->lock.
2476 * XXX we should delay bw change until the task's 0-lag point, see
2479 static int dl_overflow(struct task_struct *p, int policy,
2480 const struct sched_attr *attr)
2483 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2484 u64 period = attr->sched_period ?: attr->sched_deadline;
2485 u64 runtime = attr->sched_runtime;
2486 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2489 /* !deadline task may carry old deadline bandwidth */
2490 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2494 * Either if a task, enters, leave, or stays -deadline but changes
2495 * its parameters, we may need to update accordingly the total
2496 * allocated bandwidth of the container.
2498 raw_spin_lock(&dl_b->lock);
2499 cpus = dl_bw_cpus(task_cpu(p));
2500 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2501 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2502 __dl_add(dl_b, new_bw);
2504 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2505 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2506 __dl_clear(dl_b, p->dl.dl_bw);
2507 __dl_add(dl_b, new_bw);
2509 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2510 __dl_clear(dl_b, p->dl.dl_bw);
2513 raw_spin_unlock(&dl_b->lock);
2518 extern void init_dl_bw(struct dl_bw *dl_b);
2521 * wake_up_new_task - wake up a newly created task for the first time.
2523 * This function will do some initial scheduler statistics housekeeping
2524 * that must be done for every newly created context, then puts the task
2525 * on the runqueue and wakes it.
2527 void wake_up_new_task(struct task_struct *p)
2532 /* Initialize new task's runnable average */
2533 init_entity_runnable_average(&p->se);
2534 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2537 * Fork balancing, do it here and not earlier because:
2538 * - cpus_allowed can change in the fork path
2539 * - any previously selected cpu might disappear through hotplug
2541 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2542 * as we're not fully set-up yet.
2544 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2546 rq = __task_rq_lock(p, &rf);
2547 post_init_entity_util_avg(&p->se);
2549 activate_task(rq, p, 0);
2550 p->on_rq = TASK_ON_RQ_QUEUED;
2551 trace_sched_wakeup_new(p);
2552 check_preempt_curr(rq, p, WF_FORK);
2554 if (p->sched_class->task_woken) {
2556 * Nothing relies on rq->lock after this, so its fine to
2559 lockdep_unpin_lock(&rq->lock, rf.cookie);
2560 p->sched_class->task_woken(rq, p);
2561 lockdep_repin_lock(&rq->lock, rf.cookie);
2564 task_rq_unlock(rq, p, &rf);
2567 #ifdef CONFIG_PREEMPT_NOTIFIERS
2569 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2571 void preempt_notifier_inc(void)
2573 static_key_slow_inc(&preempt_notifier_key);
2575 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2577 void preempt_notifier_dec(void)
2579 static_key_slow_dec(&preempt_notifier_key);
2581 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2584 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2585 * @notifier: notifier struct to register
2587 void preempt_notifier_register(struct preempt_notifier *notifier)
2589 if (!static_key_false(&preempt_notifier_key))
2590 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2592 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2594 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2597 * preempt_notifier_unregister - no longer interested in preemption notifications
2598 * @notifier: notifier struct to unregister
2600 * This is *not* safe to call from within a preemption notifier.
2602 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2604 hlist_del(¬ifier->link);
2606 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2608 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2610 struct preempt_notifier *notifier;
2612 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2613 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2616 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2618 if (static_key_false(&preempt_notifier_key))
2619 __fire_sched_in_preempt_notifiers(curr);
2623 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2624 struct task_struct *next)
2626 struct preempt_notifier *notifier;
2628 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2629 notifier->ops->sched_out(notifier, next);
2632 static __always_inline void
2633 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2634 struct task_struct *next)
2636 if (static_key_false(&preempt_notifier_key))
2637 __fire_sched_out_preempt_notifiers(curr, next);
2640 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2642 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2647 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2648 struct task_struct *next)
2652 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2655 * prepare_task_switch - prepare to switch tasks
2656 * @rq: the runqueue preparing to switch
2657 * @prev: the current task that is being switched out
2658 * @next: the task we are going to switch to.
2660 * This is called with the rq lock held and interrupts off. It must
2661 * be paired with a subsequent finish_task_switch after the context
2664 * prepare_task_switch sets up locking and calls architecture specific
2668 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2669 struct task_struct *next)
2671 sched_info_switch(rq, prev, next);
2672 perf_event_task_sched_out(prev, next);
2673 fire_sched_out_preempt_notifiers(prev, next);
2674 prepare_lock_switch(rq, next);
2675 prepare_arch_switch(next);
2679 * finish_task_switch - clean up after a task-switch
2680 * @prev: the thread we just switched away from.
2682 * finish_task_switch must be called after the context switch, paired
2683 * with a prepare_task_switch call before the context switch.
2684 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2685 * and do any other architecture-specific cleanup actions.
2687 * Note that we may have delayed dropping an mm in context_switch(). If
2688 * so, we finish that here outside of the runqueue lock. (Doing it
2689 * with the lock held can cause deadlocks; see schedule() for
2692 * The context switch have flipped the stack from under us and restored the
2693 * local variables which were saved when this task called schedule() in the
2694 * past. prev == current is still correct but we need to recalculate this_rq
2695 * because prev may have moved to another CPU.
2697 static struct rq *finish_task_switch(struct task_struct *prev)
2698 __releases(rq->lock)
2700 struct rq *rq = this_rq();
2701 struct mm_struct *mm = rq->prev_mm;
2705 * The previous task will have left us with a preempt_count of 2
2706 * because it left us after:
2709 * preempt_disable(); // 1
2711 * raw_spin_lock_irq(&rq->lock) // 2
2713 * Also, see FORK_PREEMPT_COUNT.
2715 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2716 "corrupted preempt_count: %s/%d/0x%x\n",
2717 current->comm, current->pid, preempt_count()))
2718 preempt_count_set(FORK_PREEMPT_COUNT);
2723 * A task struct has one reference for the use as "current".
2724 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2725 * schedule one last time. The schedule call will never return, and
2726 * the scheduled task must drop that reference.
2728 * We must observe prev->state before clearing prev->on_cpu (in
2729 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2730 * running on another CPU and we could rave with its RUNNING -> DEAD
2731 * transition, resulting in a double drop.
2733 prev_state = prev->state;
2734 vtime_task_switch(prev);
2735 perf_event_task_sched_in(prev, current);
2736 finish_lock_switch(rq, prev);
2737 finish_arch_post_lock_switch();
2739 fire_sched_in_preempt_notifiers(current);
2742 if (unlikely(prev_state == TASK_DEAD)) {
2743 if (prev->sched_class->task_dead)
2744 prev->sched_class->task_dead(prev);
2747 * Remove function-return probe instances associated with this
2748 * task and put them back on the free list.
2750 kprobe_flush_task(prev);
2751 put_task_struct(prev);
2754 tick_nohz_task_switch();
2760 /* rq->lock is NOT held, but preemption is disabled */
2761 static void __balance_callback(struct rq *rq)
2763 struct callback_head *head, *next;
2764 void (*func)(struct rq *rq);
2765 unsigned long flags;
2767 raw_spin_lock_irqsave(&rq->lock, flags);
2768 head = rq->balance_callback;
2769 rq->balance_callback = NULL;
2771 func = (void (*)(struct rq *))head->func;
2778 raw_spin_unlock_irqrestore(&rq->lock, flags);
2781 static inline void balance_callback(struct rq *rq)
2783 if (unlikely(rq->balance_callback))
2784 __balance_callback(rq);
2789 static inline void balance_callback(struct rq *rq)
2796 * schedule_tail - first thing a freshly forked thread must call.
2797 * @prev: the thread we just switched away from.
2799 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2800 __releases(rq->lock)
2805 * New tasks start with FORK_PREEMPT_COUNT, see there and
2806 * finish_task_switch() for details.
2808 * finish_task_switch() will drop rq->lock() and lower preempt_count
2809 * and the preempt_enable() will end up enabling preemption (on
2810 * PREEMPT_COUNT kernels).
2813 rq = finish_task_switch(prev);
2814 balance_callback(rq);
2817 if (current->set_child_tid)
2818 put_user(task_pid_vnr(current), current->set_child_tid);
2822 * context_switch - switch to the new MM and the new thread's register state.
2824 static __always_inline struct rq *
2825 context_switch(struct rq *rq, struct task_struct *prev,
2826 struct task_struct *next, struct pin_cookie cookie)
2828 struct mm_struct *mm, *oldmm;
2830 prepare_task_switch(rq, prev, next);
2833 oldmm = prev->active_mm;
2835 * For paravirt, this is coupled with an exit in switch_to to
2836 * combine the page table reload and the switch backend into
2839 arch_start_context_switch(prev);
2842 next->active_mm = oldmm;
2843 atomic_inc(&oldmm->mm_count);
2844 enter_lazy_tlb(oldmm, next);
2846 switch_mm_irqs_off(oldmm, mm, next);
2849 prev->active_mm = NULL;
2850 rq->prev_mm = oldmm;
2853 * Since the runqueue lock will be released by the next
2854 * task (which is an invalid locking op but in the case
2855 * of the scheduler it's an obvious special-case), so we
2856 * do an early lockdep release here:
2858 lockdep_unpin_lock(&rq->lock, cookie);
2859 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2861 /* Here we just switch the register state and the stack. */
2862 switch_to(prev, next, prev);
2865 return finish_task_switch(prev);
2869 * nr_running and nr_context_switches:
2871 * externally visible scheduler statistics: current number of runnable
2872 * threads, total number of context switches performed since bootup.
2874 unsigned long nr_running(void)
2876 unsigned long i, sum = 0;
2878 for_each_online_cpu(i)
2879 sum += cpu_rq(i)->nr_running;
2885 * Check if only the current task is running on the cpu.
2887 * Caution: this function does not check that the caller has disabled
2888 * preemption, thus the result might have a time-of-check-to-time-of-use
2889 * race. The caller is responsible to use it correctly, for example:
2891 * - from a non-preemptable section (of course)
2893 * - from a thread that is bound to a single CPU
2895 * - in a loop with very short iterations (e.g. a polling loop)
2897 bool single_task_running(void)
2899 return raw_rq()->nr_running == 1;
2901 EXPORT_SYMBOL(single_task_running);
2903 unsigned long long nr_context_switches(void)
2906 unsigned long long sum = 0;
2908 for_each_possible_cpu(i)
2909 sum += cpu_rq(i)->nr_switches;
2914 unsigned long nr_iowait(void)
2916 unsigned long i, sum = 0;
2918 for_each_possible_cpu(i)
2919 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2924 unsigned long nr_iowait_cpu(int cpu)
2926 struct rq *this = cpu_rq(cpu);
2927 return atomic_read(&this->nr_iowait);
2930 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2932 struct rq *rq = this_rq();
2933 *nr_waiters = atomic_read(&rq->nr_iowait);
2934 *load = rq->load.weight;
2940 * sched_exec - execve() is a valuable balancing opportunity, because at
2941 * this point the task has the smallest effective memory and cache footprint.
2943 void sched_exec(void)
2945 struct task_struct *p = current;
2946 unsigned long flags;
2949 raw_spin_lock_irqsave(&p->pi_lock, flags);
2950 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2951 if (dest_cpu == smp_processor_id())
2954 if (likely(cpu_active(dest_cpu))) {
2955 struct migration_arg arg = { p, dest_cpu };
2957 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2958 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2962 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2967 DEFINE_PER_CPU(struct kernel_stat, kstat);
2968 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2970 EXPORT_PER_CPU_SYMBOL(kstat);
2971 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2974 * Return accounted runtime for the task.
2975 * In case the task is currently running, return the runtime plus current's
2976 * pending runtime that have not been accounted yet.
2978 unsigned long long task_sched_runtime(struct task_struct *p)
2984 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2986 * 64-bit doesn't need locks to atomically read a 64bit value.
2987 * So we have a optimization chance when the task's delta_exec is 0.
2988 * Reading ->on_cpu is racy, but this is ok.
2990 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2991 * If we race with it entering cpu, unaccounted time is 0. This is
2992 * indistinguishable from the read occurring a few cycles earlier.
2993 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2994 * been accounted, so we're correct here as well.
2996 if (!p->on_cpu || !task_on_rq_queued(p))
2997 return p->se.sum_exec_runtime;
3000 rq = task_rq_lock(p, &rf);
3002 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3003 * project cycles that may never be accounted to this
3004 * thread, breaking clock_gettime().
3006 if (task_current(rq, p) && task_on_rq_queued(p)) {
3007 update_rq_clock(rq);
3008 p->sched_class->update_curr(rq);
3010 ns = p->se.sum_exec_runtime;
3011 task_rq_unlock(rq, p, &rf);
3017 * This function gets called by the timer code, with HZ frequency.
3018 * We call it with interrupts disabled.
3020 void scheduler_tick(void)
3022 int cpu = smp_processor_id();
3023 struct rq *rq = cpu_rq(cpu);
3024 struct task_struct *curr = rq->curr;
3028 raw_spin_lock(&rq->lock);
3029 update_rq_clock(rq);
3030 curr->sched_class->task_tick(rq, curr, 0);
3031 cpu_load_update_active(rq);
3032 calc_global_load_tick(rq);
3033 raw_spin_unlock(&rq->lock);
3035 perf_event_task_tick();
3038 rq->idle_balance = idle_cpu(cpu);
3039 trigger_load_balance(rq);
3041 rq_last_tick_reset(rq);
3044 #ifdef CONFIG_NO_HZ_FULL
3046 * scheduler_tick_max_deferment
3048 * Keep at least one tick per second when a single
3049 * active task is running because the scheduler doesn't
3050 * yet completely support full dynticks environment.
3052 * This makes sure that uptime, CFS vruntime, load
3053 * balancing, etc... continue to move forward, even
3054 * with a very low granularity.
3056 * Return: Maximum deferment in nanoseconds.
3058 u64 scheduler_tick_max_deferment(void)
3060 struct rq *rq = this_rq();
3061 unsigned long next, now = READ_ONCE(jiffies);
3063 next = rq->last_sched_tick + HZ;
3065 if (time_before_eq(next, now))
3068 return jiffies_to_nsecs(next - now);
3072 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3073 defined(CONFIG_PREEMPT_TRACER))
3075 * If the value passed in is equal to the current preempt count
3076 * then we just disabled preemption. Start timing the latency.
3078 static inline void preempt_latency_start(int val)
3080 if (preempt_count() == val) {
3081 unsigned long ip = get_lock_parent_ip();
3082 #ifdef CONFIG_DEBUG_PREEMPT
3083 current->preempt_disable_ip = ip;
3085 trace_preempt_off(CALLER_ADDR0, ip);
3089 void preempt_count_add(int val)
3091 #ifdef CONFIG_DEBUG_PREEMPT
3095 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3098 __preempt_count_add(val);
3099 #ifdef CONFIG_DEBUG_PREEMPT
3101 * Spinlock count overflowing soon?
3103 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3106 preempt_latency_start(val);
3108 EXPORT_SYMBOL(preempt_count_add);
3109 NOKPROBE_SYMBOL(preempt_count_add);
3112 * If the value passed in equals to the current preempt count
3113 * then we just enabled preemption. Stop timing the latency.
3115 static inline void preempt_latency_stop(int val)
3117 if (preempt_count() == val)
3118 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3121 void preempt_count_sub(int val)
3123 #ifdef CONFIG_DEBUG_PREEMPT
3127 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3130 * Is the spinlock portion underflowing?
3132 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3133 !(preempt_count() & PREEMPT_MASK)))
3137 preempt_latency_stop(val);
3138 __preempt_count_sub(val);
3140 EXPORT_SYMBOL(preempt_count_sub);
3141 NOKPROBE_SYMBOL(preempt_count_sub);
3144 static inline void preempt_latency_start(int val) { }
3145 static inline void preempt_latency_stop(int val) { }
3149 * Print scheduling while atomic bug:
3151 static noinline void __schedule_bug(struct task_struct *prev)
3153 if (oops_in_progress)
3156 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3157 prev->comm, prev->pid, preempt_count());
3159 debug_show_held_locks(prev);
3161 if (irqs_disabled())
3162 print_irqtrace_events(prev);
3163 #ifdef CONFIG_DEBUG_PREEMPT
3164 if (in_atomic_preempt_off()) {
3165 pr_err("Preemption disabled at:");
3166 print_ip_sym(current->preempt_disable_ip);
3171 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3175 * Various schedule()-time debugging checks and statistics:
3177 static inline void schedule_debug(struct task_struct *prev)
3179 #ifdef CONFIG_SCHED_STACK_END_CHECK
3180 if (task_stack_end_corrupted(prev))
3181 panic("corrupted stack end detected inside scheduler\n");
3184 if (unlikely(in_atomic_preempt_off())) {
3185 __schedule_bug(prev);
3186 preempt_count_set(PREEMPT_DISABLED);
3190 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3192 schedstat_inc(this_rq(), sched_count);
3196 * Pick up the highest-prio task:
3198 static inline struct task_struct *
3199 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3201 const struct sched_class *class = &fair_sched_class;
3202 struct task_struct *p;
3205 * Optimization: we know that if all tasks are in
3206 * the fair class we can call that function directly:
3208 if (likely(prev->sched_class == class &&
3209 rq->nr_running == rq->cfs.h_nr_running)) {
3210 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3211 if (unlikely(p == RETRY_TASK))
3214 /* assumes fair_sched_class->next == idle_sched_class */
3216 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3222 for_each_class(class) {
3223 p = class->pick_next_task(rq, prev, cookie);
3225 if (unlikely(p == RETRY_TASK))
3231 BUG(); /* the idle class will always have a runnable task */
3235 * __schedule() is the main scheduler function.
3237 * The main means of driving the scheduler and thus entering this function are:
3239 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3241 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3242 * paths. For example, see arch/x86/entry_64.S.
3244 * To drive preemption between tasks, the scheduler sets the flag in timer
3245 * interrupt handler scheduler_tick().
3247 * 3. Wakeups don't really cause entry into schedule(). They add a
3248 * task to the run-queue and that's it.
3250 * Now, if the new task added to the run-queue preempts the current
3251 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3252 * called on the nearest possible occasion:
3254 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3256 * - in syscall or exception context, at the next outmost
3257 * preempt_enable(). (this might be as soon as the wake_up()'s
3260 * - in IRQ context, return from interrupt-handler to
3261 * preemptible context
3263 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3266 * - cond_resched() call
3267 * - explicit schedule() call
3268 * - return from syscall or exception to user-space
3269 * - return from interrupt-handler to user-space
3271 * WARNING: must be called with preemption disabled!
3273 static void __sched notrace __schedule(bool preempt)
3275 struct task_struct *prev, *next;
3276 unsigned long *switch_count;
3277 struct pin_cookie cookie;
3281 cpu = smp_processor_id();
3286 * do_exit() calls schedule() with preemption disabled as an exception;
3287 * however we must fix that up, otherwise the next task will see an
3288 * inconsistent (higher) preempt count.
3290 * It also avoids the below schedule_debug() test from complaining
3293 if (unlikely(prev->state == TASK_DEAD))
3294 preempt_enable_no_resched_notrace();
3296 schedule_debug(prev);
3298 if (sched_feat(HRTICK))
3301 local_irq_disable();
3302 rcu_note_context_switch();
3305 * Make sure that signal_pending_state()->signal_pending() below
3306 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3307 * done by the caller to avoid the race with signal_wake_up().
3309 smp_mb__before_spinlock();
3310 raw_spin_lock(&rq->lock);
3311 cookie = lockdep_pin_lock(&rq->lock);
3313 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3315 switch_count = &prev->nivcsw;
3316 if (!preempt && prev->state) {
3317 if (unlikely(signal_pending_state(prev->state, prev))) {
3318 prev->state = TASK_RUNNING;
3320 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3324 * If a worker went to sleep, notify and ask workqueue
3325 * whether it wants to wake up a task to maintain
3328 if (prev->flags & PF_WQ_WORKER) {
3329 struct task_struct *to_wakeup;
3331 to_wakeup = wq_worker_sleeping(prev);
3333 try_to_wake_up_local(to_wakeup, cookie);
3336 switch_count = &prev->nvcsw;
3339 if (task_on_rq_queued(prev))
3340 update_rq_clock(rq);
3342 next = pick_next_task(rq, prev, cookie);
3343 clear_tsk_need_resched(prev);
3344 clear_preempt_need_resched();
3345 rq->clock_skip_update = 0;
3347 if (likely(prev != next)) {
3352 trace_sched_switch(preempt, prev, next);
3353 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3355 lockdep_unpin_lock(&rq->lock, cookie);
3356 raw_spin_unlock_irq(&rq->lock);
3359 balance_callback(rq);
3361 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3363 static inline void sched_submit_work(struct task_struct *tsk)
3365 if (!tsk->state || tsk_is_pi_blocked(tsk))
3368 * If we are going to sleep and we have plugged IO queued,
3369 * make sure to submit it to avoid deadlocks.
3371 if (blk_needs_flush_plug(tsk))
3372 blk_schedule_flush_plug(tsk);
3375 asmlinkage __visible void __sched schedule(void)
3377 struct task_struct *tsk = current;
3379 sched_submit_work(tsk);
3383 sched_preempt_enable_no_resched();
3384 } while (need_resched());
3386 EXPORT_SYMBOL(schedule);
3388 #ifdef CONFIG_CONTEXT_TRACKING
3389 asmlinkage __visible void __sched schedule_user(void)
3392 * If we come here after a random call to set_need_resched(),
3393 * or we have been woken up remotely but the IPI has not yet arrived,
3394 * we haven't yet exited the RCU idle mode. Do it here manually until
3395 * we find a better solution.
3397 * NB: There are buggy callers of this function. Ideally we
3398 * should warn if prev_state != CONTEXT_USER, but that will trigger
3399 * too frequently to make sense yet.
3401 enum ctx_state prev_state = exception_enter();
3403 exception_exit(prev_state);
3408 * schedule_preempt_disabled - called with preemption disabled
3410 * Returns with preemption disabled. Note: preempt_count must be 1
3412 void __sched schedule_preempt_disabled(void)
3414 sched_preempt_enable_no_resched();
3419 static void __sched notrace preempt_schedule_common(void)
3423 * Because the function tracer can trace preempt_count_sub()
3424 * and it also uses preempt_enable/disable_notrace(), if
3425 * NEED_RESCHED is set, the preempt_enable_notrace() called
3426 * by the function tracer will call this function again and
3427 * cause infinite recursion.
3429 * Preemption must be disabled here before the function
3430 * tracer can trace. Break up preempt_disable() into two
3431 * calls. One to disable preemption without fear of being
3432 * traced. The other to still record the preemption latency,
3433 * which can also be traced by the function tracer.
3435 preempt_disable_notrace();
3436 preempt_latency_start(1);
3438 preempt_latency_stop(1);
3439 preempt_enable_no_resched_notrace();
3442 * Check again in case we missed a preemption opportunity
3443 * between schedule and now.
3445 } while (need_resched());
3448 #ifdef CONFIG_PREEMPT
3450 * this is the entry point to schedule() from in-kernel preemption
3451 * off of preempt_enable. Kernel preemptions off return from interrupt
3452 * occur there and call schedule directly.
3454 asmlinkage __visible void __sched notrace preempt_schedule(void)
3457 * If there is a non-zero preempt_count or interrupts are disabled,
3458 * we do not want to preempt the current task. Just return..
3460 if (likely(!preemptible()))
3463 preempt_schedule_common();
3465 NOKPROBE_SYMBOL(preempt_schedule);
3466 EXPORT_SYMBOL(preempt_schedule);
3469 * preempt_schedule_notrace - preempt_schedule called by tracing
3471 * The tracing infrastructure uses preempt_enable_notrace to prevent
3472 * recursion and tracing preempt enabling caused by the tracing
3473 * infrastructure itself. But as tracing can happen in areas coming
3474 * from userspace or just about to enter userspace, a preempt enable
3475 * can occur before user_exit() is called. This will cause the scheduler
3476 * to be called when the system is still in usermode.
3478 * To prevent this, the preempt_enable_notrace will use this function
3479 * instead of preempt_schedule() to exit user context if needed before
3480 * calling the scheduler.
3482 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3484 enum ctx_state prev_ctx;
3486 if (likely(!preemptible()))
3491 * Because the function tracer can trace preempt_count_sub()
3492 * and it also uses preempt_enable/disable_notrace(), if
3493 * NEED_RESCHED is set, the preempt_enable_notrace() called
3494 * by the function tracer will call this function again and
3495 * cause infinite recursion.
3497 * Preemption must be disabled here before the function
3498 * tracer can trace. Break up preempt_disable() into two
3499 * calls. One to disable preemption without fear of being
3500 * traced. The other to still record the preemption latency,
3501 * which can also be traced by the function tracer.
3503 preempt_disable_notrace();
3504 preempt_latency_start(1);
3506 * Needs preempt disabled in case user_exit() is traced
3507 * and the tracer calls preempt_enable_notrace() causing
3508 * an infinite recursion.
3510 prev_ctx = exception_enter();
3512 exception_exit(prev_ctx);
3514 preempt_latency_stop(1);
3515 preempt_enable_no_resched_notrace();
3516 } while (need_resched());
3518 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3520 #endif /* CONFIG_PREEMPT */
3523 * this is the entry point to schedule() from kernel preemption
3524 * off of irq context.
3525 * Note, that this is called and return with irqs disabled. This will
3526 * protect us against recursive calling from irq.
3528 asmlinkage __visible void __sched preempt_schedule_irq(void)
3530 enum ctx_state prev_state;
3532 /* Catch callers which need to be fixed */
3533 BUG_ON(preempt_count() || !irqs_disabled());
3535 prev_state = exception_enter();
3541 local_irq_disable();
3542 sched_preempt_enable_no_resched();
3543 } while (need_resched());
3545 exception_exit(prev_state);
3548 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3551 return try_to_wake_up(curr->private, mode, wake_flags);
3553 EXPORT_SYMBOL(default_wake_function);
3555 #ifdef CONFIG_RT_MUTEXES
3558 * rt_mutex_setprio - set the current priority of a task
3560 * @prio: prio value (kernel-internal form)
3562 * This function changes the 'effective' priority of a task. It does
3563 * not touch ->normal_prio like __setscheduler().
3565 * Used by the rt_mutex code to implement priority inheritance
3566 * logic. Call site only calls if the priority of the task changed.
3568 void rt_mutex_setprio(struct task_struct *p, int prio)
3570 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3571 const struct sched_class *prev_class;
3575 BUG_ON(prio > MAX_PRIO);
3577 rq = __task_rq_lock(p, &rf);
3580 * Idle task boosting is a nono in general. There is one
3581 * exception, when PREEMPT_RT and NOHZ is active:
3583 * The idle task calls get_next_timer_interrupt() and holds
3584 * the timer wheel base->lock on the CPU and another CPU wants
3585 * to access the timer (probably to cancel it). We can safely
3586 * ignore the boosting request, as the idle CPU runs this code
3587 * with interrupts disabled and will complete the lock
3588 * protected section without being interrupted. So there is no
3589 * real need to boost.
3591 if (unlikely(p == rq->idle)) {
3592 WARN_ON(p != rq->curr);
3593 WARN_ON(p->pi_blocked_on);
3597 trace_sched_pi_setprio(p, prio);
3600 if (oldprio == prio)
3601 queue_flag &= ~DEQUEUE_MOVE;
3603 prev_class = p->sched_class;
3604 queued = task_on_rq_queued(p);
3605 running = task_current(rq, p);
3607 dequeue_task(rq, p, queue_flag);
3609 put_prev_task(rq, p);
3612 * Boosting condition are:
3613 * 1. -rt task is running and holds mutex A
3614 * --> -dl task blocks on mutex A
3616 * 2. -dl task is running and holds mutex A
3617 * --> -dl task blocks on mutex A and could preempt the
3620 if (dl_prio(prio)) {
3621 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3622 if (!dl_prio(p->normal_prio) ||
3623 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3624 p->dl.dl_boosted = 1;
3625 queue_flag |= ENQUEUE_REPLENISH;
3627 p->dl.dl_boosted = 0;
3628 p->sched_class = &dl_sched_class;
3629 } else if (rt_prio(prio)) {
3630 if (dl_prio(oldprio))
3631 p->dl.dl_boosted = 0;
3633 queue_flag |= ENQUEUE_HEAD;
3634 p->sched_class = &rt_sched_class;
3636 if (dl_prio(oldprio))
3637 p->dl.dl_boosted = 0;
3638 if (rt_prio(oldprio))
3640 p->sched_class = &fair_sched_class;
3646 p->sched_class->set_curr_task(rq);
3648 enqueue_task(rq, p, queue_flag);
3650 check_class_changed(rq, p, prev_class, oldprio);
3652 preempt_disable(); /* avoid rq from going away on us */
3653 __task_rq_unlock(rq, &rf);
3655 balance_callback(rq);
3660 void set_user_nice(struct task_struct *p, long nice)
3662 int old_prio, delta, queued;
3666 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3669 * We have to be careful, if called from sys_setpriority(),
3670 * the task might be in the middle of scheduling on another CPU.
3672 rq = task_rq_lock(p, &rf);
3674 * The RT priorities are set via sched_setscheduler(), but we still
3675 * allow the 'normal' nice value to be set - but as expected
3676 * it wont have any effect on scheduling until the task is
3677 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3679 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3680 p->static_prio = NICE_TO_PRIO(nice);
3683 queued = task_on_rq_queued(p);
3685 dequeue_task(rq, p, DEQUEUE_SAVE);
3687 p->static_prio = NICE_TO_PRIO(nice);
3690 p->prio = effective_prio(p);
3691 delta = p->prio - old_prio;
3694 enqueue_task(rq, p, ENQUEUE_RESTORE);
3696 * If the task increased its priority or is running and
3697 * lowered its priority, then reschedule its CPU:
3699 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3703 task_rq_unlock(rq, p, &rf);
3705 EXPORT_SYMBOL(set_user_nice);
3708 * can_nice - check if a task can reduce its nice value
3712 int can_nice(const struct task_struct *p, const int nice)
3714 /* convert nice value [19,-20] to rlimit style value [1,40] */
3715 int nice_rlim = nice_to_rlimit(nice);
3717 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3718 capable(CAP_SYS_NICE));
3721 #ifdef __ARCH_WANT_SYS_NICE
3724 * sys_nice - change the priority of the current process.
3725 * @increment: priority increment
3727 * sys_setpriority is a more generic, but much slower function that
3728 * does similar things.
3730 SYSCALL_DEFINE1(nice, int, increment)
3735 * Setpriority might change our priority at the same moment.
3736 * We don't have to worry. Conceptually one call occurs first
3737 * and we have a single winner.
3739 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3740 nice = task_nice(current) + increment;
3742 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3743 if (increment < 0 && !can_nice(current, nice))
3746 retval = security_task_setnice(current, nice);
3750 set_user_nice(current, nice);
3757 * task_prio - return the priority value of a given task.
3758 * @p: the task in question.
3760 * Return: The priority value as seen by users in /proc.
3761 * RT tasks are offset by -200. Normal tasks are centered
3762 * around 0, value goes from -16 to +15.
3764 int task_prio(const struct task_struct *p)
3766 return p->prio - MAX_RT_PRIO;
3770 * idle_cpu - is a given cpu idle currently?
3771 * @cpu: the processor in question.
3773 * Return: 1 if the CPU is currently idle. 0 otherwise.
3775 int idle_cpu(int cpu)
3777 struct rq *rq = cpu_rq(cpu);
3779 if (rq->curr != rq->idle)
3786 if (!llist_empty(&rq->wake_list))
3794 * idle_task - return the idle task for a given cpu.
3795 * @cpu: the processor in question.
3797 * Return: The idle task for the cpu @cpu.
3799 struct task_struct *idle_task(int cpu)
3801 return cpu_rq(cpu)->idle;
3805 * find_process_by_pid - find a process with a matching PID value.
3806 * @pid: the pid in question.
3808 * The task of @pid, if found. %NULL otherwise.
3810 static struct task_struct *find_process_by_pid(pid_t pid)
3812 return pid ? find_task_by_vpid(pid) : current;
3816 * This function initializes the sched_dl_entity of a newly becoming
3817 * SCHED_DEADLINE task.
3819 * Only the static values are considered here, the actual runtime and the
3820 * absolute deadline will be properly calculated when the task is enqueued
3821 * for the first time with its new policy.
3824 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3826 struct sched_dl_entity *dl_se = &p->dl;
3828 dl_se->dl_runtime = attr->sched_runtime;
3829 dl_se->dl_deadline = attr->sched_deadline;
3830 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3831 dl_se->flags = attr->sched_flags;
3832 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3835 * Changing the parameters of a task is 'tricky' and we're not doing
3836 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3838 * What we SHOULD do is delay the bandwidth release until the 0-lag
3839 * point. This would include retaining the task_struct until that time
3840 * and change dl_overflow() to not immediately decrement the current
3843 * Instead we retain the current runtime/deadline and let the new
3844 * parameters take effect after the current reservation period lapses.
3845 * This is safe (albeit pessimistic) because the 0-lag point is always
3846 * before the current scheduling deadline.
3848 * We can still have temporary overloads because we do not delay the
3849 * change in bandwidth until that time; so admission control is
3850 * not on the safe side. It does however guarantee tasks will never
3851 * consume more than promised.
3856 * sched_setparam() passes in -1 for its policy, to let the functions
3857 * it calls know not to change it.
3859 #define SETPARAM_POLICY -1
3861 static void __setscheduler_params(struct task_struct *p,
3862 const struct sched_attr *attr)
3864 int policy = attr->sched_policy;
3866 if (policy == SETPARAM_POLICY)
3871 if (dl_policy(policy))
3872 __setparam_dl(p, attr);
3873 else if (fair_policy(policy))
3874 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3877 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3878 * !rt_policy. Always setting this ensures that things like
3879 * getparam()/getattr() don't report silly values for !rt tasks.
3881 p->rt_priority = attr->sched_priority;
3882 p->normal_prio = normal_prio(p);
3886 /* Actually do priority change: must hold pi & rq lock. */
3887 static void __setscheduler(struct rq *rq, struct task_struct *p,
3888 const struct sched_attr *attr, bool keep_boost)
3890 __setscheduler_params(p, attr);
3893 * Keep a potential priority boosting if called from
3894 * sched_setscheduler().
3897 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3899 p->prio = normal_prio(p);
3901 if (dl_prio(p->prio))
3902 p->sched_class = &dl_sched_class;
3903 else if (rt_prio(p->prio))
3904 p->sched_class = &rt_sched_class;
3906 p->sched_class = &fair_sched_class;
3910 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3912 struct sched_dl_entity *dl_se = &p->dl;
3914 attr->sched_priority = p->rt_priority;
3915 attr->sched_runtime = dl_se->dl_runtime;
3916 attr->sched_deadline = dl_se->dl_deadline;
3917 attr->sched_period = dl_se->dl_period;
3918 attr->sched_flags = dl_se->flags;
3922 * This function validates the new parameters of a -deadline task.
3923 * We ask for the deadline not being zero, and greater or equal
3924 * than the runtime, as well as the period of being zero or
3925 * greater than deadline. Furthermore, we have to be sure that
3926 * user parameters are above the internal resolution of 1us (we
3927 * check sched_runtime only since it is always the smaller one) and
3928 * below 2^63 ns (we have to check both sched_deadline and
3929 * sched_period, as the latter can be zero).
3932 __checkparam_dl(const struct sched_attr *attr)
3935 if (attr->sched_deadline == 0)
3939 * Since we truncate DL_SCALE bits, make sure we're at least
3942 if (attr->sched_runtime < (1ULL << DL_SCALE))
3946 * Since we use the MSB for wrap-around and sign issues, make
3947 * sure it's not set (mind that period can be equal to zero).
3949 if (attr->sched_deadline & (1ULL << 63) ||
3950 attr->sched_period & (1ULL << 63))
3953 /* runtime <= deadline <= period (if period != 0) */
3954 if ((attr->sched_period != 0 &&
3955 attr->sched_period < attr->sched_deadline) ||
3956 attr->sched_deadline < attr->sched_runtime)
3963 * check the target process has a UID that matches the current process's
3965 static bool check_same_owner(struct task_struct *p)
3967 const struct cred *cred = current_cred(), *pcred;
3971 pcred = __task_cred(p);
3972 match = (uid_eq(cred->euid, pcred->euid) ||
3973 uid_eq(cred->euid, pcred->uid));
3978 static bool dl_param_changed(struct task_struct *p,
3979 const struct sched_attr *attr)
3981 struct sched_dl_entity *dl_se = &p->dl;
3983 if (dl_se->dl_runtime != attr->sched_runtime ||
3984 dl_se->dl_deadline != attr->sched_deadline ||
3985 dl_se->dl_period != attr->sched_period ||
3986 dl_se->flags != attr->sched_flags)
3992 static int __sched_setscheduler(struct task_struct *p,
3993 const struct sched_attr *attr,
3996 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3997 MAX_RT_PRIO - 1 - attr->sched_priority;
3998 int retval, oldprio, oldpolicy = -1, queued, running;
3999 int new_effective_prio, policy = attr->sched_policy;
4000 const struct sched_class *prev_class;
4003 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4006 /* may grab non-irq protected spin_locks */
4007 BUG_ON(in_interrupt());
4009 /* double check policy once rq lock held */
4011 reset_on_fork = p->sched_reset_on_fork;
4012 policy = oldpolicy = p->policy;
4014 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4016 if (!valid_policy(policy))
4020 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4024 * Valid priorities for SCHED_FIFO and SCHED_RR are
4025 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4026 * SCHED_BATCH and SCHED_IDLE is 0.
4028 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4029 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4031 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4032 (rt_policy(policy) != (attr->sched_priority != 0)))
4036 * Allow unprivileged RT tasks to decrease priority:
4038 if (user && !capable(CAP_SYS_NICE)) {
4039 if (fair_policy(policy)) {
4040 if (attr->sched_nice < task_nice(p) &&
4041 !can_nice(p, attr->sched_nice))
4045 if (rt_policy(policy)) {
4046 unsigned long rlim_rtprio =
4047 task_rlimit(p, RLIMIT_RTPRIO);
4049 /* can't set/change the rt policy */
4050 if (policy != p->policy && !rlim_rtprio)
4053 /* can't increase priority */
4054 if (attr->sched_priority > p->rt_priority &&
4055 attr->sched_priority > rlim_rtprio)
4060 * Can't set/change SCHED_DEADLINE policy at all for now
4061 * (safest behavior); in the future we would like to allow
4062 * unprivileged DL tasks to increase their relative deadline
4063 * or reduce their runtime (both ways reducing utilization)
4065 if (dl_policy(policy))
4069 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4070 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4072 if (idle_policy(p->policy) && !idle_policy(policy)) {
4073 if (!can_nice(p, task_nice(p)))
4077 /* can't change other user's priorities */
4078 if (!check_same_owner(p))
4081 /* Normal users shall not reset the sched_reset_on_fork flag */
4082 if (p->sched_reset_on_fork && !reset_on_fork)
4087 retval = security_task_setscheduler(p);
4093 * make sure no PI-waiters arrive (or leave) while we are
4094 * changing the priority of the task:
4096 * To be able to change p->policy safely, the appropriate
4097 * runqueue lock must be held.
4099 rq = task_rq_lock(p, &rf);
4102 * Changing the policy of the stop threads its a very bad idea
4104 if (p == rq->stop) {
4105 task_rq_unlock(rq, p, &rf);
4110 * If not changing anything there's no need to proceed further,
4111 * but store a possible modification of reset_on_fork.
4113 if (unlikely(policy == p->policy)) {
4114 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4116 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4118 if (dl_policy(policy) && dl_param_changed(p, attr))
4121 p->sched_reset_on_fork = reset_on_fork;
4122 task_rq_unlock(rq, p, &rf);
4128 #ifdef CONFIG_RT_GROUP_SCHED
4130 * Do not allow realtime tasks into groups that have no runtime
4133 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4134 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4135 !task_group_is_autogroup(task_group(p))) {
4136 task_rq_unlock(rq, p, &rf);
4141 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4142 cpumask_t *span = rq->rd->span;
4145 * Don't allow tasks with an affinity mask smaller than
4146 * the entire root_domain to become SCHED_DEADLINE. We
4147 * will also fail if there's no bandwidth available.
4149 if (!cpumask_subset(span, &p->cpus_allowed) ||
4150 rq->rd->dl_bw.bw == 0) {
4151 task_rq_unlock(rq, p, &rf);
4158 /* recheck policy now with rq lock held */
4159 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4160 policy = oldpolicy = -1;
4161 task_rq_unlock(rq, p, &rf);
4166 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4167 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4170 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4171 task_rq_unlock(rq, p, &rf);
4175 p->sched_reset_on_fork = reset_on_fork;
4180 * Take priority boosted tasks into account. If the new
4181 * effective priority is unchanged, we just store the new
4182 * normal parameters and do not touch the scheduler class and
4183 * the runqueue. This will be done when the task deboost
4186 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4187 if (new_effective_prio == oldprio)
4188 queue_flags &= ~DEQUEUE_MOVE;
4191 queued = task_on_rq_queued(p);
4192 running = task_current(rq, p);
4194 dequeue_task(rq, p, queue_flags);
4196 put_prev_task(rq, p);
4198 prev_class = p->sched_class;
4199 __setscheduler(rq, p, attr, pi);
4202 p->sched_class->set_curr_task(rq);
4205 * We enqueue to tail when the priority of a task is
4206 * increased (user space view).
4208 if (oldprio < p->prio)
4209 queue_flags |= ENQUEUE_HEAD;
4211 enqueue_task(rq, p, queue_flags);
4214 check_class_changed(rq, p, prev_class, oldprio);
4215 preempt_disable(); /* avoid rq from going away on us */
4216 task_rq_unlock(rq, p, &rf);
4219 rt_mutex_adjust_pi(p);
4222 * Run balance callbacks after we've adjusted the PI chain.
4224 balance_callback(rq);
4230 static int _sched_setscheduler(struct task_struct *p, int policy,
4231 const struct sched_param *param, bool check)
4233 struct sched_attr attr = {
4234 .sched_policy = policy,
4235 .sched_priority = param->sched_priority,
4236 .sched_nice = PRIO_TO_NICE(p->static_prio),
4239 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4240 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4241 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4242 policy &= ~SCHED_RESET_ON_FORK;
4243 attr.sched_policy = policy;
4246 return __sched_setscheduler(p, &attr, check, true);
4249 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4250 * @p: the task in question.
4251 * @policy: new policy.
4252 * @param: structure containing the new RT priority.
4254 * Return: 0 on success. An error code otherwise.
4256 * NOTE that the task may be already dead.
4258 int sched_setscheduler(struct task_struct *p, int policy,
4259 const struct sched_param *param)
4261 return _sched_setscheduler(p, policy, param, true);
4263 EXPORT_SYMBOL_GPL(sched_setscheduler);
4265 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4267 return __sched_setscheduler(p, attr, true, true);
4269 EXPORT_SYMBOL_GPL(sched_setattr);
4272 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4273 * @p: the task in question.
4274 * @policy: new policy.
4275 * @param: structure containing the new RT priority.
4277 * Just like sched_setscheduler, only don't bother checking if the
4278 * current context has permission. For example, this is needed in
4279 * stop_machine(): we create temporary high priority worker threads,
4280 * but our caller might not have that capability.
4282 * Return: 0 on success. An error code otherwise.
4284 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4285 const struct sched_param *param)
4287 return _sched_setscheduler(p, policy, param, false);
4289 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4292 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4294 struct sched_param lparam;
4295 struct task_struct *p;
4298 if (!param || pid < 0)
4300 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4305 p = find_process_by_pid(pid);
4307 retval = sched_setscheduler(p, policy, &lparam);
4314 * Mimics kernel/events/core.c perf_copy_attr().
4316 static int sched_copy_attr(struct sched_attr __user *uattr,
4317 struct sched_attr *attr)
4322 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4326 * zero the full structure, so that a short copy will be nice.
4328 memset(attr, 0, sizeof(*attr));
4330 ret = get_user(size, &uattr->size);
4334 if (size > PAGE_SIZE) /* silly large */
4337 if (!size) /* abi compat */
4338 size = SCHED_ATTR_SIZE_VER0;
4340 if (size < SCHED_ATTR_SIZE_VER0)
4344 * If we're handed a bigger struct than we know of,
4345 * ensure all the unknown bits are 0 - i.e. new
4346 * user-space does not rely on any kernel feature
4347 * extensions we dont know about yet.
4349 if (size > sizeof(*attr)) {
4350 unsigned char __user *addr;
4351 unsigned char __user *end;
4354 addr = (void __user *)uattr + sizeof(*attr);
4355 end = (void __user *)uattr + size;
4357 for (; addr < end; addr++) {
4358 ret = get_user(val, addr);
4364 size = sizeof(*attr);
4367 ret = copy_from_user(attr, uattr, size);
4372 * XXX: do we want to be lenient like existing syscalls; or do we want
4373 * to be strict and return an error on out-of-bounds values?
4375 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4380 put_user(sizeof(*attr), &uattr->size);
4385 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4386 * @pid: the pid in question.
4387 * @policy: new policy.
4388 * @param: structure containing the new RT priority.
4390 * Return: 0 on success. An error code otherwise.
4392 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4393 struct sched_param __user *, param)
4395 /* negative values for policy are not valid */
4399 return do_sched_setscheduler(pid, policy, param);
4403 * sys_sched_setparam - set/change the RT priority of a thread
4404 * @pid: the pid in question.
4405 * @param: structure containing the new RT priority.
4407 * Return: 0 on success. An error code otherwise.
4409 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4411 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4415 * sys_sched_setattr - same as above, but with extended sched_attr
4416 * @pid: the pid in question.
4417 * @uattr: structure containing the extended parameters.
4418 * @flags: for future extension.
4420 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4421 unsigned int, flags)
4423 struct sched_attr attr;
4424 struct task_struct *p;
4427 if (!uattr || pid < 0 || flags)
4430 retval = sched_copy_attr(uattr, &attr);
4434 if ((int)attr.sched_policy < 0)
4439 p = find_process_by_pid(pid);
4441 retval = sched_setattr(p, &attr);
4448 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4449 * @pid: the pid in question.
4451 * Return: On success, the policy of the thread. Otherwise, a negative error
4454 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4456 struct task_struct *p;
4464 p = find_process_by_pid(pid);
4466 retval = security_task_getscheduler(p);
4469 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4476 * sys_sched_getparam - get the RT priority of a thread
4477 * @pid: the pid in question.
4478 * @param: structure containing the RT priority.
4480 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4483 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4485 struct sched_param lp = { .sched_priority = 0 };
4486 struct task_struct *p;
4489 if (!param || pid < 0)
4493 p = find_process_by_pid(pid);
4498 retval = security_task_getscheduler(p);
4502 if (task_has_rt_policy(p))
4503 lp.sched_priority = p->rt_priority;
4507 * This one might sleep, we cannot do it with a spinlock held ...
4509 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4518 static int sched_read_attr(struct sched_attr __user *uattr,
4519 struct sched_attr *attr,
4524 if (!access_ok(VERIFY_WRITE, uattr, usize))
4528 * If we're handed a smaller struct than we know of,
4529 * ensure all the unknown bits are 0 - i.e. old
4530 * user-space does not get uncomplete information.
4532 if (usize < sizeof(*attr)) {
4533 unsigned char *addr;
4536 addr = (void *)attr + usize;
4537 end = (void *)attr + sizeof(*attr);
4539 for (; addr < end; addr++) {
4547 ret = copy_to_user(uattr, attr, attr->size);
4555 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4556 * @pid: the pid in question.
4557 * @uattr: structure containing the extended parameters.
4558 * @size: sizeof(attr) for fwd/bwd comp.
4559 * @flags: for future extension.
4561 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4562 unsigned int, size, unsigned int, flags)
4564 struct sched_attr attr = {
4565 .size = sizeof(struct sched_attr),
4567 struct task_struct *p;
4570 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4571 size < SCHED_ATTR_SIZE_VER0 || flags)
4575 p = find_process_by_pid(pid);
4580 retval = security_task_getscheduler(p);
4584 attr.sched_policy = p->policy;
4585 if (p->sched_reset_on_fork)
4586 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4587 if (task_has_dl_policy(p))
4588 __getparam_dl(p, &attr);
4589 else if (task_has_rt_policy(p))
4590 attr.sched_priority = p->rt_priority;
4592 attr.sched_nice = task_nice(p);
4596 retval = sched_read_attr(uattr, &attr, size);
4604 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4606 cpumask_var_t cpus_allowed, new_mask;
4607 struct task_struct *p;
4612 p = find_process_by_pid(pid);
4618 /* Prevent p going away */
4622 if (p->flags & PF_NO_SETAFFINITY) {
4626 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4630 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4632 goto out_free_cpus_allowed;
4635 if (!check_same_owner(p)) {
4637 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4639 goto out_free_new_mask;
4644 retval = security_task_setscheduler(p);
4646 goto out_free_new_mask;
4649 cpuset_cpus_allowed(p, cpus_allowed);
4650 cpumask_and(new_mask, in_mask, cpus_allowed);
4653 * Since bandwidth control happens on root_domain basis,
4654 * if admission test is enabled, we only admit -deadline
4655 * tasks allowed to run on all the CPUs in the task's
4659 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4661 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4664 goto out_free_new_mask;
4670 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4673 cpuset_cpus_allowed(p, cpus_allowed);
4674 if (!cpumask_subset(new_mask, cpus_allowed)) {
4676 * We must have raced with a concurrent cpuset
4677 * update. Just reset the cpus_allowed to the
4678 * cpuset's cpus_allowed
4680 cpumask_copy(new_mask, cpus_allowed);
4685 free_cpumask_var(new_mask);
4686 out_free_cpus_allowed:
4687 free_cpumask_var(cpus_allowed);
4693 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4694 struct cpumask *new_mask)
4696 if (len < cpumask_size())
4697 cpumask_clear(new_mask);
4698 else if (len > cpumask_size())
4699 len = cpumask_size();
4701 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4705 * sys_sched_setaffinity - set the cpu affinity of a process
4706 * @pid: pid of the process
4707 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4708 * @user_mask_ptr: user-space pointer to the new cpu mask
4710 * Return: 0 on success. An error code otherwise.
4712 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4713 unsigned long __user *, user_mask_ptr)
4715 cpumask_var_t new_mask;
4718 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4721 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4723 retval = sched_setaffinity(pid, new_mask);
4724 free_cpumask_var(new_mask);
4728 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4730 struct task_struct *p;
4731 unsigned long flags;
4737 p = find_process_by_pid(pid);
4741 retval = security_task_getscheduler(p);
4745 raw_spin_lock_irqsave(&p->pi_lock, flags);
4746 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4747 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4756 * sys_sched_getaffinity - get the cpu affinity of a process
4757 * @pid: pid of the process
4758 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4759 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4761 * Return: 0 on success. An error code otherwise.
4763 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4764 unsigned long __user *, user_mask_ptr)
4769 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4771 if (len & (sizeof(unsigned long)-1))
4774 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4777 ret = sched_getaffinity(pid, mask);
4779 size_t retlen = min_t(size_t, len, cpumask_size());
4781 if (copy_to_user(user_mask_ptr, mask, retlen))
4786 free_cpumask_var(mask);
4792 * sys_sched_yield - yield the current processor to other threads.
4794 * This function yields the current CPU to other tasks. If there are no
4795 * other threads running on this CPU then this function will return.
4799 SYSCALL_DEFINE0(sched_yield)
4801 struct rq *rq = this_rq_lock();
4803 schedstat_inc(rq, yld_count);
4804 current->sched_class->yield_task(rq);
4807 * Since we are going to call schedule() anyway, there's
4808 * no need to preempt or enable interrupts:
4810 __release(rq->lock);
4811 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4812 do_raw_spin_unlock(&rq->lock);
4813 sched_preempt_enable_no_resched();
4820 int __sched _cond_resched(void)
4822 if (should_resched(0)) {
4823 preempt_schedule_common();
4828 EXPORT_SYMBOL(_cond_resched);
4831 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4832 * call schedule, and on return reacquire the lock.
4834 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4835 * operations here to prevent schedule() from being called twice (once via
4836 * spin_unlock(), once by hand).
4838 int __cond_resched_lock(spinlock_t *lock)
4840 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4843 lockdep_assert_held(lock);
4845 if (spin_needbreak(lock) || resched) {
4848 preempt_schedule_common();
4856 EXPORT_SYMBOL(__cond_resched_lock);
4858 int __sched __cond_resched_softirq(void)
4860 BUG_ON(!in_softirq());
4862 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4864 preempt_schedule_common();
4870 EXPORT_SYMBOL(__cond_resched_softirq);
4873 * yield - yield the current processor to other threads.
4875 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4877 * The scheduler is at all times free to pick the calling task as the most
4878 * eligible task to run, if removing the yield() call from your code breaks
4879 * it, its already broken.
4881 * Typical broken usage is:
4886 * where one assumes that yield() will let 'the other' process run that will
4887 * make event true. If the current task is a SCHED_FIFO task that will never
4888 * happen. Never use yield() as a progress guarantee!!
4890 * If you want to use yield() to wait for something, use wait_event().
4891 * If you want to use yield() to be 'nice' for others, use cond_resched().
4892 * If you still want to use yield(), do not!
4894 void __sched yield(void)
4896 set_current_state(TASK_RUNNING);
4899 EXPORT_SYMBOL(yield);
4902 * yield_to - yield the current processor to another thread in
4903 * your thread group, or accelerate that thread toward the
4904 * processor it's on.
4906 * @preempt: whether task preemption is allowed or not
4908 * It's the caller's job to ensure that the target task struct
4909 * can't go away on us before we can do any checks.
4912 * true (>0) if we indeed boosted the target task.
4913 * false (0) if we failed to boost the target.
4914 * -ESRCH if there's no task to yield to.
4916 int __sched yield_to(struct task_struct *p, bool preempt)
4918 struct task_struct *curr = current;
4919 struct rq *rq, *p_rq;
4920 unsigned long flags;
4923 local_irq_save(flags);
4929 * If we're the only runnable task on the rq and target rq also
4930 * has only one task, there's absolutely no point in yielding.
4932 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4937 double_rq_lock(rq, p_rq);
4938 if (task_rq(p) != p_rq) {
4939 double_rq_unlock(rq, p_rq);
4943 if (!curr->sched_class->yield_to_task)
4946 if (curr->sched_class != p->sched_class)
4949 if (task_running(p_rq, p) || p->state)
4952 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4954 schedstat_inc(rq, yld_count);
4956 * Make p's CPU reschedule; pick_next_entity takes care of
4959 if (preempt && rq != p_rq)
4964 double_rq_unlock(rq, p_rq);
4966 local_irq_restore(flags);
4973 EXPORT_SYMBOL_GPL(yield_to);
4976 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4977 * that process accounting knows that this is a task in IO wait state.
4979 long __sched io_schedule_timeout(long timeout)
4981 int old_iowait = current->in_iowait;
4985 current->in_iowait = 1;
4986 blk_schedule_flush_plug(current);
4988 delayacct_blkio_start();
4990 atomic_inc(&rq->nr_iowait);
4991 ret = schedule_timeout(timeout);
4992 current->in_iowait = old_iowait;
4993 atomic_dec(&rq->nr_iowait);
4994 delayacct_blkio_end();
4998 EXPORT_SYMBOL(io_schedule_timeout);
5001 * sys_sched_get_priority_max - return maximum RT priority.
5002 * @policy: scheduling class.
5004 * Return: On success, this syscall returns the maximum
5005 * rt_priority that can be used by a given scheduling class.
5006 * On failure, a negative error code is returned.
5008 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5015 ret = MAX_USER_RT_PRIO-1;
5017 case SCHED_DEADLINE:
5028 * sys_sched_get_priority_min - return minimum RT priority.
5029 * @policy: scheduling class.
5031 * Return: On success, this syscall returns the minimum
5032 * rt_priority that can be used by a given scheduling class.
5033 * On failure, a negative error code is returned.
5035 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5044 case SCHED_DEADLINE:
5054 * sys_sched_rr_get_interval - return the default timeslice of a process.
5055 * @pid: pid of the process.
5056 * @interval: userspace pointer to the timeslice value.
5058 * this syscall writes the default timeslice value of a given process
5059 * into the user-space timespec buffer. A value of '0' means infinity.
5061 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5064 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5065 struct timespec __user *, interval)
5067 struct task_struct *p;
5068 unsigned int time_slice;
5079 p = find_process_by_pid(pid);
5083 retval = security_task_getscheduler(p);
5087 rq = task_rq_lock(p, &rf);
5089 if (p->sched_class->get_rr_interval)
5090 time_slice = p->sched_class->get_rr_interval(rq, p);
5091 task_rq_unlock(rq, p, &rf);
5094 jiffies_to_timespec(time_slice, &t);
5095 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5103 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5105 void sched_show_task(struct task_struct *p)
5107 unsigned long free = 0;
5109 unsigned long state = p->state;
5112 state = __ffs(state) + 1;
5113 printk(KERN_INFO "%-15.15s %c", p->comm,
5114 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5115 #if BITS_PER_LONG == 32
5116 if (state == TASK_RUNNING)
5117 printk(KERN_CONT " running ");
5119 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5121 if (state == TASK_RUNNING)
5122 printk(KERN_CONT " running task ");
5124 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5126 #ifdef CONFIG_DEBUG_STACK_USAGE
5127 free = stack_not_used(p);
5132 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5134 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5135 task_pid_nr(p), ppid,
5136 (unsigned long)task_thread_info(p)->flags);
5138 print_worker_info(KERN_INFO, p);
5139 show_stack(p, NULL);
5142 void show_state_filter(unsigned long state_filter)
5144 struct task_struct *g, *p;
5146 #if BITS_PER_LONG == 32
5148 " task PC stack pid father\n");
5151 " task PC stack pid father\n");
5154 for_each_process_thread(g, p) {
5156 * reset the NMI-timeout, listing all files on a slow
5157 * console might take a lot of time:
5158 * Also, reset softlockup watchdogs on all CPUs, because
5159 * another CPU might be blocked waiting for us to process
5162 touch_nmi_watchdog();
5163 touch_all_softlockup_watchdogs();
5164 if (!state_filter || (p->state & state_filter))
5168 #ifdef CONFIG_SCHED_DEBUG
5170 sysrq_sched_debug_show();
5174 * Only show locks if all tasks are dumped:
5177 debug_show_all_locks();
5180 void init_idle_bootup_task(struct task_struct *idle)
5182 idle->sched_class = &idle_sched_class;
5186 * init_idle - set up an idle thread for a given CPU
5187 * @idle: task in question
5188 * @cpu: cpu the idle task belongs to
5190 * NOTE: this function does not set the idle thread's NEED_RESCHED
5191 * flag, to make booting more robust.
5193 void init_idle(struct task_struct *idle, int cpu)
5195 struct rq *rq = cpu_rq(cpu);
5196 unsigned long flags;
5198 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5199 raw_spin_lock(&rq->lock);
5201 __sched_fork(0, idle);
5202 idle->state = TASK_RUNNING;
5203 idle->se.exec_start = sched_clock();
5205 kasan_unpoison_task_stack(idle);
5209 * Its possible that init_idle() gets called multiple times on a task,
5210 * in that case do_set_cpus_allowed() will not do the right thing.
5212 * And since this is boot we can forgo the serialization.
5214 set_cpus_allowed_common(idle, cpumask_of(cpu));
5217 * We're having a chicken and egg problem, even though we are
5218 * holding rq->lock, the cpu isn't yet set to this cpu so the
5219 * lockdep check in task_group() will fail.
5221 * Similar case to sched_fork(). / Alternatively we could
5222 * use task_rq_lock() here and obtain the other rq->lock.
5227 __set_task_cpu(idle, cpu);
5230 rq->curr = rq->idle = idle;
5231 idle->on_rq = TASK_ON_RQ_QUEUED;
5235 raw_spin_unlock(&rq->lock);
5236 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5238 /* Set the preempt count _outside_ the spinlocks! */
5239 init_idle_preempt_count(idle, cpu);
5242 * The idle tasks have their own, simple scheduling class:
5244 idle->sched_class = &idle_sched_class;
5245 ftrace_graph_init_idle_task(idle, cpu);
5246 vtime_init_idle(idle, cpu);
5248 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5252 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5253 const struct cpumask *trial)
5255 int ret = 1, trial_cpus;
5256 struct dl_bw *cur_dl_b;
5257 unsigned long flags;
5259 if (!cpumask_weight(cur))
5262 rcu_read_lock_sched();
5263 cur_dl_b = dl_bw_of(cpumask_any(cur));
5264 trial_cpus = cpumask_weight(trial);
5266 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5267 if (cur_dl_b->bw != -1 &&
5268 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5270 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5271 rcu_read_unlock_sched();
5276 int task_can_attach(struct task_struct *p,
5277 const struct cpumask *cs_cpus_allowed)
5282 * Kthreads which disallow setaffinity shouldn't be moved
5283 * to a new cpuset; we don't want to change their cpu
5284 * affinity and isolating such threads by their set of
5285 * allowed nodes is unnecessary. Thus, cpusets are not
5286 * applicable for such threads. This prevents checking for
5287 * success of set_cpus_allowed_ptr() on all attached tasks
5288 * before cpus_allowed may be changed.
5290 if (p->flags & PF_NO_SETAFFINITY) {
5296 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5298 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5303 unsigned long flags;
5305 rcu_read_lock_sched();
5306 dl_b = dl_bw_of(dest_cpu);
5307 raw_spin_lock_irqsave(&dl_b->lock, flags);
5308 cpus = dl_bw_cpus(dest_cpu);
5309 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5314 * We reserve space for this task in the destination
5315 * root_domain, as we can't fail after this point.
5316 * We will free resources in the source root_domain
5317 * later on (see set_cpus_allowed_dl()).
5319 __dl_add(dl_b, p->dl.dl_bw);
5321 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5322 rcu_read_unlock_sched();
5332 static bool sched_smp_initialized __read_mostly;
5334 #ifdef CONFIG_NUMA_BALANCING
5335 /* Migrate current task p to target_cpu */
5336 int migrate_task_to(struct task_struct *p, int target_cpu)
5338 struct migration_arg arg = { p, target_cpu };
5339 int curr_cpu = task_cpu(p);
5341 if (curr_cpu == target_cpu)
5344 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5347 /* TODO: This is not properly updating schedstats */
5349 trace_sched_move_numa(p, curr_cpu, target_cpu);
5350 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5354 * Requeue a task on a given node and accurately track the number of NUMA
5355 * tasks on the runqueues
5357 void sched_setnuma(struct task_struct *p, int nid)
5359 bool queued, running;
5363 rq = task_rq_lock(p, &rf);
5364 queued = task_on_rq_queued(p);
5365 running = task_current(rq, p);
5368 dequeue_task(rq, p, DEQUEUE_SAVE);
5370 put_prev_task(rq, p);
5372 p->numa_preferred_nid = nid;
5375 p->sched_class->set_curr_task(rq);
5377 enqueue_task(rq, p, ENQUEUE_RESTORE);
5378 task_rq_unlock(rq, p, &rf);
5380 #endif /* CONFIG_NUMA_BALANCING */
5382 #ifdef CONFIG_HOTPLUG_CPU
5384 * Ensures that the idle task is using init_mm right before its cpu goes
5387 void idle_task_exit(void)
5389 struct mm_struct *mm = current->active_mm;
5391 BUG_ON(cpu_online(smp_processor_id()));
5393 if (mm != &init_mm) {
5394 switch_mm_irqs_off(mm, &init_mm, current);
5395 finish_arch_post_lock_switch();
5401 * Since this CPU is going 'away' for a while, fold any nr_active delta
5402 * we might have. Assumes we're called after migrate_tasks() so that the
5403 * nr_active count is stable.
5405 * Also see the comment "Global load-average calculations".
5407 static void calc_load_migrate(struct rq *rq)
5409 long delta = calc_load_fold_active(rq);
5411 atomic_long_add(delta, &calc_load_tasks);
5414 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5418 static const struct sched_class fake_sched_class = {
5419 .put_prev_task = put_prev_task_fake,
5422 static struct task_struct fake_task = {
5424 * Avoid pull_{rt,dl}_task()
5426 .prio = MAX_PRIO + 1,
5427 .sched_class = &fake_sched_class,
5431 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5432 * try_to_wake_up()->select_task_rq().
5434 * Called with rq->lock held even though we'er in stop_machine() and
5435 * there's no concurrency possible, we hold the required locks anyway
5436 * because of lock validation efforts.
5438 static void migrate_tasks(struct rq *dead_rq)
5440 struct rq *rq = dead_rq;
5441 struct task_struct *next, *stop = rq->stop;
5442 struct pin_cookie cookie;
5446 * Fudge the rq selection such that the below task selection loop
5447 * doesn't get stuck on the currently eligible stop task.
5449 * We're currently inside stop_machine() and the rq is either stuck
5450 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5451 * either way we should never end up calling schedule() until we're
5457 * put_prev_task() and pick_next_task() sched
5458 * class method both need to have an up-to-date
5459 * value of rq->clock[_task]
5461 update_rq_clock(rq);
5465 * There's this thread running, bail when that's the only
5468 if (rq->nr_running == 1)
5472 * pick_next_task assumes pinned rq->lock.
5474 cookie = lockdep_pin_lock(&rq->lock);
5475 next = pick_next_task(rq, &fake_task, cookie);
5477 next->sched_class->put_prev_task(rq, next);
5480 * Rules for changing task_struct::cpus_allowed are holding
5481 * both pi_lock and rq->lock, such that holding either
5482 * stabilizes the mask.
5484 * Drop rq->lock is not quite as disastrous as it usually is
5485 * because !cpu_active at this point, which means load-balance
5486 * will not interfere. Also, stop-machine.
5488 lockdep_unpin_lock(&rq->lock, cookie);
5489 raw_spin_unlock(&rq->lock);
5490 raw_spin_lock(&next->pi_lock);
5491 raw_spin_lock(&rq->lock);
5494 * Since we're inside stop-machine, _nothing_ should have
5495 * changed the task, WARN if weird stuff happened, because in
5496 * that case the above rq->lock drop is a fail too.
5498 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5499 raw_spin_unlock(&next->pi_lock);
5503 /* Find suitable destination for @next, with force if needed. */
5504 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5506 rq = __migrate_task(rq, next, dest_cpu);
5507 if (rq != dead_rq) {
5508 raw_spin_unlock(&rq->lock);
5510 raw_spin_lock(&rq->lock);
5512 raw_spin_unlock(&next->pi_lock);
5517 #endif /* CONFIG_HOTPLUG_CPU */
5519 static void set_rq_online(struct rq *rq)
5522 const struct sched_class *class;
5524 cpumask_set_cpu(rq->cpu, rq->rd->online);
5527 for_each_class(class) {
5528 if (class->rq_online)
5529 class->rq_online(rq);
5534 static void set_rq_offline(struct rq *rq)
5537 const struct sched_class *class;
5539 for_each_class(class) {
5540 if (class->rq_offline)
5541 class->rq_offline(rq);
5544 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5549 static void set_cpu_rq_start_time(unsigned int cpu)
5551 struct rq *rq = cpu_rq(cpu);
5553 rq->age_stamp = sched_clock_cpu(cpu);
5556 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5558 #ifdef CONFIG_SCHED_DEBUG
5560 static __read_mostly int sched_debug_enabled;
5562 static int __init sched_debug_setup(char *str)
5564 sched_debug_enabled = 1;
5568 early_param("sched_debug", sched_debug_setup);
5570 static inline bool sched_debug(void)
5572 return sched_debug_enabled;
5575 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5576 struct cpumask *groupmask)
5578 struct sched_group *group = sd->groups;
5580 cpumask_clear(groupmask);
5582 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5584 if (!(sd->flags & SD_LOAD_BALANCE)) {
5585 printk("does not load-balance\n");
5587 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5592 printk(KERN_CONT "span %*pbl level %s\n",
5593 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5595 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5596 printk(KERN_ERR "ERROR: domain->span does not contain "
5599 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5600 printk(KERN_ERR "ERROR: domain->groups does not contain"
5604 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5608 printk(KERN_ERR "ERROR: group is NULL\n");
5612 if (!cpumask_weight(sched_group_cpus(group))) {
5613 printk(KERN_CONT "\n");
5614 printk(KERN_ERR "ERROR: empty group\n");
5618 if (!(sd->flags & SD_OVERLAP) &&
5619 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5620 printk(KERN_CONT "\n");
5621 printk(KERN_ERR "ERROR: repeated CPUs\n");
5625 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5627 printk(KERN_CONT " %*pbl",
5628 cpumask_pr_args(sched_group_cpus(group)));
5629 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5630 printk(KERN_CONT " (cpu_capacity = %d)",
5631 group->sgc->capacity);
5634 group = group->next;
5635 } while (group != sd->groups);
5636 printk(KERN_CONT "\n");
5638 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5639 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5642 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5643 printk(KERN_ERR "ERROR: parent span is not a superset "
5644 "of domain->span\n");
5648 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5652 if (!sched_debug_enabled)
5656 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5660 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5663 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5671 #else /* !CONFIG_SCHED_DEBUG */
5672 # define sched_domain_debug(sd, cpu) do { } while (0)
5673 static inline bool sched_debug(void)
5677 #endif /* CONFIG_SCHED_DEBUG */
5679 static int sd_degenerate(struct sched_domain *sd)
5681 if (cpumask_weight(sched_domain_span(sd)) == 1)
5684 /* Following flags need at least 2 groups */
5685 if (sd->flags & (SD_LOAD_BALANCE |
5686 SD_BALANCE_NEWIDLE |
5689 SD_SHARE_CPUCAPACITY |
5690 SD_SHARE_PKG_RESOURCES |
5691 SD_SHARE_POWERDOMAIN)) {
5692 if (sd->groups != sd->groups->next)
5696 /* Following flags don't use groups */
5697 if (sd->flags & (SD_WAKE_AFFINE))
5704 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5706 unsigned long cflags = sd->flags, pflags = parent->flags;
5708 if (sd_degenerate(parent))
5711 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5714 /* Flags needing groups don't count if only 1 group in parent */
5715 if (parent->groups == parent->groups->next) {
5716 pflags &= ~(SD_LOAD_BALANCE |
5717 SD_BALANCE_NEWIDLE |
5720 SD_SHARE_CPUCAPACITY |
5721 SD_SHARE_PKG_RESOURCES |
5723 SD_SHARE_POWERDOMAIN);
5724 if (nr_node_ids == 1)
5725 pflags &= ~SD_SERIALIZE;
5727 if (~cflags & pflags)
5733 static void free_rootdomain(struct rcu_head *rcu)
5735 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5737 cpupri_cleanup(&rd->cpupri);
5738 cpudl_cleanup(&rd->cpudl);
5739 free_cpumask_var(rd->dlo_mask);
5740 free_cpumask_var(rd->rto_mask);
5741 free_cpumask_var(rd->online);
5742 free_cpumask_var(rd->span);
5746 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5748 struct root_domain *old_rd = NULL;
5749 unsigned long flags;
5751 raw_spin_lock_irqsave(&rq->lock, flags);
5756 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5759 cpumask_clear_cpu(rq->cpu, old_rd->span);
5762 * If we dont want to free the old_rd yet then
5763 * set old_rd to NULL to skip the freeing later
5766 if (!atomic_dec_and_test(&old_rd->refcount))
5770 atomic_inc(&rd->refcount);
5773 cpumask_set_cpu(rq->cpu, rd->span);
5774 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5777 raw_spin_unlock_irqrestore(&rq->lock, flags);
5780 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5783 static int init_rootdomain(struct root_domain *rd)
5785 memset(rd, 0, sizeof(*rd));
5787 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5789 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5791 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5793 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5796 init_dl_bw(&rd->dl_bw);
5797 if (cpudl_init(&rd->cpudl) != 0)
5800 if (cpupri_init(&rd->cpupri) != 0)
5805 free_cpumask_var(rd->rto_mask);
5807 free_cpumask_var(rd->dlo_mask);
5809 free_cpumask_var(rd->online);
5811 free_cpumask_var(rd->span);
5817 * By default the system creates a single root-domain with all cpus as
5818 * members (mimicking the global state we have today).
5820 struct root_domain def_root_domain;
5822 static void init_defrootdomain(void)
5824 init_rootdomain(&def_root_domain);
5826 atomic_set(&def_root_domain.refcount, 1);
5829 static struct root_domain *alloc_rootdomain(void)
5831 struct root_domain *rd;
5833 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5837 if (init_rootdomain(rd) != 0) {
5845 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5847 struct sched_group *tmp, *first;
5856 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5861 } while (sg != first);
5864 static void free_sched_domain(struct rcu_head *rcu)
5866 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5869 * If its an overlapping domain it has private groups, iterate and
5872 if (sd->flags & SD_OVERLAP) {
5873 free_sched_groups(sd->groups, 1);
5874 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5875 kfree(sd->groups->sgc);
5881 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5883 call_rcu(&sd->rcu, free_sched_domain);
5886 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5888 for (; sd; sd = sd->parent)
5889 destroy_sched_domain(sd, cpu);
5893 * Keep a special pointer to the highest sched_domain that has
5894 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5895 * allows us to avoid some pointer chasing select_idle_sibling().
5897 * Also keep a unique ID per domain (we use the first cpu number in
5898 * the cpumask of the domain), this allows us to quickly tell if
5899 * two cpus are in the same cache domain, see cpus_share_cache().
5901 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5902 DEFINE_PER_CPU(int, sd_llc_size);
5903 DEFINE_PER_CPU(int, sd_llc_id);
5904 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5905 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5906 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5908 static void update_top_cache_domain(int cpu)
5910 struct sched_domain *sd;
5911 struct sched_domain *busy_sd = NULL;
5915 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5917 id = cpumask_first(sched_domain_span(sd));
5918 size = cpumask_weight(sched_domain_span(sd));
5919 busy_sd = sd->parent; /* sd_busy */
5921 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5923 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5924 per_cpu(sd_llc_size, cpu) = size;
5925 per_cpu(sd_llc_id, cpu) = id;
5927 sd = lowest_flag_domain(cpu, SD_NUMA);
5928 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5930 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5931 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5935 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5936 * hold the hotplug lock.
5939 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5941 struct rq *rq = cpu_rq(cpu);
5942 struct sched_domain *tmp;
5944 /* Remove the sched domains which do not contribute to scheduling. */
5945 for (tmp = sd; tmp; ) {
5946 struct sched_domain *parent = tmp->parent;
5950 if (sd_parent_degenerate(tmp, parent)) {
5951 tmp->parent = parent->parent;
5953 parent->parent->child = tmp;
5955 * Transfer SD_PREFER_SIBLING down in case of a
5956 * degenerate parent; the spans match for this
5957 * so the property transfers.
5959 if (parent->flags & SD_PREFER_SIBLING)
5960 tmp->flags |= SD_PREFER_SIBLING;
5961 destroy_sched_domain(parent, cpu);
5966 if (sd && sd_degenerate(sd)) {
5969 destroy_sched_domain(tmp, cpu);
5974 sched_domain_debug(sd, cpu);
5976 rq_attach_root(rq, rd);
5978 rcu_assign_pointer(rq->sd, sd);
5979 destroy_sched_domains(tmp, cpu);
5981 update_top_cache_domain(cpu);
5984 /* Setup the mask of cpus configured for isolated domains */
5985 static int __init isolated_cpu_setup(char *str)
5989 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5990 ret = cpulist_parse(str, cpu_isolated_map);
5992 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5997 __setup("isolcpus=", isolated_cpu_setup);
6000 struct sched_domain ** __percpu sd;
6001 struct root_domain *rd;
6012 * Build an iteration mask that can exclude certain CPUs from the upwards
6015 * Asymmetric node setups can result in situations where the domain tree is of
6016 * unequal depth, make sure to skip domains that already cover the entire
6019 * In that case build_sched_domains() will have terminated the iteration early
6020 * and our sibling sd spans will be empty. Domains should always include the
6021 * cpu they're built on, so check that.
6024 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6026 const struct cpumask *span = sched_domain_span(sd);
6027 struct sd_data *sdd = sd->private;
6028 struct sched_domain *sibling;
6031 for_each_cpu(i, span) {
6032 sibling = *per_cpu_ptr(sdd->sd, i);
6033 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6036 cpumask_set_cpu(i, sched_group_mask(sg));
6041 * Return the canonical balance cpu for this group, this is the first cpu
6042 * of this group that's also in the iteration mask.
6044 int group_balance_cpu(struct sched_group *sg)
6046 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6050 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6052 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6053 const struct cpumask *span = sched_domain_span(sd);
6054 struct cpumask *covered = sched_domains_tmpmask;
6055 struct sd_data *sdd = sd->private;
6056 struct sched_domain *sibling;
6059 cpumask_clear(covered);
6061 for_each_cpu(i, span) {
6062 struct cpumask *sg_span;
6064 if (cpumask_test_cpu(i, covered))
6067 sibling = *per_cpu_ptr(sdd->sd, i);
6069 /* See the comment near build_group_mask(). */
6070 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6073 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6074 GFP_KERNEL, cpu_to_node(cpu));
6079 sg_span = sched_group_cpus(sg);
6081 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6083 cpumask_set_cpu(i, sg_span);
6085 cpumask_or(covered, covered, sg_span);
6087 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6088 if (atomic_inc_return(&sg->sgc->ref) == 1)
6089 build_group_mask(sd, sg);
6092 * Initialize sgc->capacity such that even if we mess up the
6093 * domains and no possible iteration will get us here, we won't
6096 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6099 * Make sure the first group of this domain contains the
6100 * canonical balance cpu. Otherwise the sched_domain iteration
6101 * breaks. See update_sg_lb_stats().
6103 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6104 group_balance_cpu(sg) == cpu)
6114 sd->groups = groups;
6119 free_sched_groups(first, 0);
6124 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6126 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6127 struct sched_domain *child = sd->child;
6130 cpu = cpumask_first(sched_domain_span(child));
6133 *sg = *per_cpu_ptr(sdd->sg, cpu);
6134 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6135 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6142 * build_sched_groups will build a circular linked list of the groups
6143 * covered by the given span, and will set each group's ->cpumask correctly,
6144 * and ->cpu_capacity to 0.
6146 * Assumes the sched_domain tree is fully constructed
6149 build_sched_groups(struct sched_domain *sd, int cpu)
6151 struct sched_group *first = NULL, *last = NULL;
6152 struct sd_data *sdd = sd->private;
6153 const struct cpumask *span = sched_domain_span(sd);
6154 struct cpumask *covered;
6157 get_group(cpu, sdd, &sd->groups);
6158 atomic_inc(&sd->groups->ref);
6160 if (cpu != cpumask_first(span))
6163 lockdep_assert_held(&sched_domains_mutex);
6164 covered = sched_domains_tmpmask;
6166 cpumask_clear(covered);
6168 for_each_cpu(i, span) {
6169 struct sched_group *sg;
6172 if (cpumask_test_cpu(i, covered))
6175 group = get_group(i, sdd, &sg);
6176 cpumask_setall(sched_group_mask(sg));
6178 for_each_cpu(j, span) {
6179 if (get_group(j, sdd, NULL) != group)
6182 cpumask_set_cpu(j, covered);
6183 cpumask_set_cpu(j, sched_group_cpus(sg));
6198 * Initialize sched groups cpu_capacity.
6200 * cpu_capacity indicates the capacity of sched group, which is used while
6201 * distributing the load between different sched groups in a sched domain.
6202 * Typically cpu_capacity for all the groups in a sched domain will be same
6203 * unless there are asymmetries in the topology. If there are asymmetries,
6204 * group having more cpu_capacity will pickup more load compared to the
6205 * group having less cpu_capacity.
6207 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6209 struct sched_group *sg = sd->groups;
6214 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6216 } while (sg != sd->groups);
6218 if (cpu != group_balance_cpu(sg))
6221 update_group_capacity(sd, cpu);
6222 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6226 * Initializers for schedule domains
6227 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6230 static int default_relax_domain_level = -1;
6231 int sched_domain_level_max;
6233 static int __init setup_relax_domain_level(char *str)
6235 if (kstrtoint(str, 0, &default_relax_domain_level))
6236 pr_warn("Unable to set relax_domain_level\n");
6240 __setup("relax_domain_level=", setup_relax_domain_level);
6242 static void set_domain_attribute(struct sched_domain *sd,
6243 struct sched_domain_attr *attr)
6247 if (!attr || attr->relax_domain_level < 0) {
6248 if (default_relax_domain_level < 0)
6251 request = default_relax_domain_level;
6253 request = attr->relax_domain_level;
6254 if (request < sd->level) {
6255 /* turn off idle balance on this domain */
6256 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6258 /* turn on idle balance on this domain */
6259 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6263 static void __sdt_free(const struct cpumask *cpu_map);
6264 static int __sdt_alloc(const struct cpumask *cpu_map);
6266 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6267 const struct cpumask *cpu_map)
6271 if (!atomic_read(&d->rd->refcount))
6272 free_rootdomain(&d->rd->rcu); /* fall through */
6274 free_percpu(d->sd); /* fall through */
6276 __sdt_free(cpu_map); /* fall through */
6282 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6283 const struct cpumask *cpu_map)
6285 memset(d, 0, sizeof(*d));
6287 if (__sdt_alloc(cpu_map))
6288 return sa_sd_storage;
6289 d->sd = alloc_percpu(struct sched_domain *);
6291 return sa_sd_storage;
6292 d->rd = alloc_rootdomain();
6295 return sa_rootdomain;
6299 * NULL the sd_data elements we've used to build the sched_domain and
6300 * sched_group structure so that the subsequent __free_domain_allocs()
6301 * will not free the data we're using.
6303 static void claim_allocations(int cpu, struct sched_domain *sd)
6305 struct sd_data *sdd = sd->private;
6307 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6308 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6310 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6311 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6313 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6314 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6318 static int sched_domains_numa_levels;
6319 enum numa_topology_type sched_numa_topology_type;
6320 static int *sched_domains_numa_distance;
6321 int sched_max_numa_distance;
6322 static struct cpumask ***sched_domains_numa_masks;
6323 static int sched_domains_curr_level;
6327 * SD_flags allowed in topology descriptions.
6329 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6330 * SD_SHARE_PKG_RESOURCES - describes shared caches
6331 * SD_NUMA - describes NUMA topologies
6332 * SD_SHARE_POWERDOMAIN - describes shared power domain
6335 * SD_ASYM_PACKING - describes SMT quirks
6337 #define TOPOLOGY_SD_FLAGS \
6338 (SD_SHARE_CPUCAPACITY | \
6339 SD_SHARE_PKG_RESOURCES | \
6342 SD_SHARE_POWERDOMAIN)
6344 static struct sched_domain *
6345 sd_init(struct sched_domain_topology_level *tl, int cpu)
6347 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6348 int sd_weight, sd_flags = 0;
6352 * Ugly hack to pass state to sd_numa_mask()...
6354 sched_domains_curr_level = tl->numa_level;
6357 sd_weight = cpumask_weight(tl->mask(cpu));
6360 sd_flags = (*tl->sd_flags)();
6361 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6362 "wrong sd_flags in topology description\n"))
6363 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6365 *sd = (struct sched_domain){
6366 .min_interval = sd_weight,
6367 .max_interval = 2*sd_weight,
6369 .imbalance_pct = 125,
6371 .cache_nice_tries = 0,
6378 .flags = 1*SD_LOAD_BALANCE
6379 | 1*SD_BALANCE_NEWIDLE
6384 | 0*SD_SHARE_CPUCAPACITY
6385 | 0*SD_SHARE_PKG_RESOURCES
6387 | 0*SD_PREFER_SIBLING
6392 .last_balance = jiffies,
6393 .balance_interval = sd_weight,
6395 .max_newidle_lb_cost = 0,
6396 .next_decay_max_lb_cost = jiffies,
6397 #ifdef CONFIG_SCHED_DEBUG
6403 * Convert topological properties into behaviour.
6406 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6407 sd->flags |= SD_PREFER_SIBLING;
6408 sd->imbalance_pct = 110;
6409 sd->smt_gain = 1178; /* ~15% */
6411 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6412 sd->imbalance_pct = 117;
6413 sd->cache_nice_tries = 1;
6417 } else if (sd->flags & SD_NUMA) {
6418 sd->cache_nice_tries = 2;
6422 sd->flags |= SD_SERIALIZE;
6423 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6424 sd->flags &= ~(SD_BALANCE_EXEC |
6431 sd->flags |= SD_PREFER_SIBLING;
6432 sd->cache_nice_tries = 1;
6437 sd->private = &tl->data;
6443 * Topology list, bottom-up.
6445 static struct sched_domain_topology_level default_topology[] = {
6446 #ifdef CONFIG_SCHED_SMT
6447 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6449 #ifdef CONFIG_SCHED_MC
6450 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6452 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6456 static struct sched_domain_topology_level *sched_domain_topology =
6459 #define for_each_sd_topology(tl) \
6460 for (tl = sched_domain_topology; tl->mask; tl++)
6462 void set_sched_topology(struct sched_domain_topology_level *tl)
6464 sched_domain_topology = tl;
6469 static const struct cpumask *sd_numa_mask(int cpu)
6471 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6474 static void sched_numa_warn(const char *str)
6476 static int done = false;
6484 printk(KERN_WARNING "ERROR: %s\n\n", str);
6486 for (i = 0; i < nr_node_ids; i++) {
6487 printk(KERN_WARNING " ");
6488 for (j = 0; j < nr_node_ids; j++)
6489 printk(KERN_CONT "%02d ", node_distance(i,j));
6490 printk(KERN_CONT "\n");
6492 printk(KERN_WARNING "\n");
6495 bool find_numa_distance(int distance)
6499 if (distance == node_distance(0, 0))
6502 for (i = 0; i < sched_domains_numa_levels; i++) {
6503 if (sched_domains_numa_distance[i] == distance)
6511 * A system can have three types of NUMA topology:
6512 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6513 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6514 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6516 * The difference between a glueless mesh topology and a backplane
6517 * topology lies in whether communication between not directly
6518 * connected nodes goes through intermediary nodes (where programs
6519 * could run), or through backplane controllers. This affects
6520 * placement of programs.
6522 * The type of topology can be discerned with the following tests:
6523 * - If the maximum distance between any nodes is 1 hop, the system
6524 * is directly connected.
6525 * - If for two nodes A and B, located N > 1 hops away from each other,
6526 * there is an intermediary node C, which is < N hops away from both
6527 * nodes A and B, the system is a glueless mesh.
6529 static void init_numa_topology_type(void)
6533 n = sched_max_numa_distance;
6535 if (sched_domains_numa_levels <= 1) {
6536 sched_numa_topology_type = NUMA_DIRECT;
6540 for_each_online_node(a) {
6541 for_each_online_node(b) {
6542 /* Find two nodes furthest removed from each other. */
6543 if (node_distance(a, b) < n)
6546 /* Is there an intermediary node between a and b? */
6547 for_each_online_node(c) {
6548 if (node_distance(a, c) < n &&
6549 node_distance(b, c) < n) {
6550 sched_numa_topology_type =
6556 sched_numa_topology_type = NUMA_BACKPLANE;
6562 static void sched_init_numa(void)
6564 int next_distance, curr_distance = node_distance(0, 0);
6565 struct sched_domain_topology_level *tl;
6569 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6570 if (!sched_domains_numa_distance)
6574 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6575 * unique distances in the node_distance() table.
6577 * Assumes node_distance(0,j) includes all distances in
6578 * node_distance(i,j) in order to avoid cubic time.
6580 next_distance = curr_distance;
6581 for (i = 0; i < nr_node_ids; i++) {
6582 for (j = 0; j < nr_node_ids; j++) {
6583 for (k = 0; k < nr_node_ids; k++) {
6584 int distance = node_distance(i, k);
6586 if (distance > curr_distance &&
6587 (distance < next_distance ||
6588 next_distance == curr_distance))
6589 next_distance = distance;
6592 * While not a strong assumption it would be nice to know
6593 * about cases where if node A is connected to B, B is not
6594 * equally connected to A.
6596 if (sched_debug() && node_distance(k, i) != distance)
6597 sched_numa_warn("Node-distance not symmetric");
6599 if (sched_debug() && i && !find_numa_distance(distance))
6600 sched_numa_warn("Node-0 not representative");
6602 if (next_distance != curr_distance) {
6603 sched_domains_numa_distance[level++] = next_distance;
6604 sched_domains_numa_levels = level;
6605 curr_distance = next_distance;
6610 * In case of sched_debug() we verify the above assumption.
6620 * 'level' contains the number of unique distances, excluding the
6621 * identity distance node_distance(i,i).
6623 * The sched_domains_numa_distance[] array includes the actual distance
6628 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6629 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6630 * the array will contain less then 'level' members. This could be
6631 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6632 * in other functions.
6634 * We reset it to 'level' at the end of this function.
6636 sched_domains_numa_levels = 0;
6638 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6639 if (!sched_domains_numa_masks)
6643 * Now for each level, construct a mask per node which contains all
6644 * cpus of nodes that are that many hops away from us.
6646 for (i = 0; i < level; i++) {
6647 sched_domains_numa_masks[i] =
6648 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6649 if (!sched_domains_numa_masks[i])
6652 for (j = 0; j < nr_node_ids; j++) {
6653 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6657 sched_domains_numa_masks[i][j] = mask;
6660 if (node_distance(j, k) > sched_domains_numa_distance[i])
6663 cpumask_or(mask, mask, cpumask_of_node(k));
6668 /* Compute default topology size */
6669 for (i = 0; sched_domain_topology[i].mask; i++);
6671 tl = kzalloc((i + level + 1) *
6672 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6677 * Copy the default topology bits..
6679 for (i = 0; sched_domain_topology[i].mask; i++)
6680 tl[i] = sched_domain_topology[i];
6683 * .. and append 'j' levels of NUMA goodness.
6685 for (j = 0; j < level; i++, j++) {
6686 tl[i] = (struct sched_domain_topology_level){
6687 .mask = sd_numa_mask,
6688 .sd_flags = cpu_numa_flags,
6689 .flags = SDTL_OVERLAP,
6695 sched_domain_topology = tl;
6697 sched_domains_numa_levels = level;
6698 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6700 init_numa_topology_type();
6703 static void sched_domains_numa_masks_set(unsigned int cpu)
6705 int node = cpu_to_node(cpu);
6708 for (i = 0; i < sched_domains_numa_levels; i++) {
6709 for (j = 0; j < nr_node_ids; j++) {
6710 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6711 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6716 static void sched_domains_numa_masks_clear(unsigned int cpu)
6720 for (i = 0; i < sched_domains_numa_levels; i++) {
6721 for (j = 0; j < nr_node_ids; j++)
6722 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6727 static inline void sched_init_numa(void) { }
6728 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6729 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6730 #endif /* CONFIG_NUMA */
6732 static int __sdt_alloc(const struct cpumask *cpu_map)
6734 struct sched_domain_topology_level *tl;
6737 for_each_sd_topology(tl) {
6738 struct sd_data *sdd = &tl->data;
6740 sdd->sd = alloc_percpu(struct sched_domain *);
6744 sdd->sg = alloc_percpu(struct sched_group *);
6748 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6752 for_each_cpu(j, cpu_map) {
6753 struct sched_domain *sd;
6754 struct sched_group *sg;
6755 struct sched_group_capacity *sgc;
6757 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6758 GFP_KERNEL, cpu_to_node(j));
6762 *per_cpu_ptr(sdd->sd, j) = sd;
6764 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6765 GFP_KERNEL, cpu_to_node(j));
6771 *per_cpu_ptr(sdd->sg, j) = sg;
6773 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6774 GFP_KERNEL, cpu_to_node(j));
6778 *per_cpu_ptr(sdd->sgc, j) = sgc;
6785 static void __sdt_free(const struct cpumask *cpu_map)
6787 struct sched_domain_topology_level *tl;
6790 for_each_sd_topology(tl) {
6791 struct sd_data *sdd = &tl->data;
6793 for_each_cpu(j, cpu_map) {
6794 struct sched_domain *sd;
6797 sd = *per_cpu_ptr(sdd->sd, j);
6798 if (sd && (sd->flags & SD_OVERLAP))
6799 free_sched_groups(sd->groups, 0);
6800 kfree(*per_cpu_ptr(sdd->sd, j));
6804 kfree(*per_cpu_ptr(sdd->sg, j));
6806 kfree(*per_cpu_ptr(sdd->sgc, j));
6808 free_percpu(sdd->sd);
6810 free_percpu(sdd->sg);
6812 free_percpu(sdd->sgc);
6817 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6818 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6819 struct sched_domain *child, int cpu)
6821 struct sched_domain *sd = sd_init(tl, cpu);
6825 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6827 sd->level = child->level + 1;
6828 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6832 if (!cpumask_subset(sched_domain_span(child),
6833 sched_domain_span(sd))) {
6834 pr_err("BUG: arch topology borken\n");
6835 #ifdef CONFIG_SCHED_DEBUG
6836 pr_err(" the %s domain not a subset of the %s domain\n",
6837 child->name, sd->name);
6839 /* Fixup, ensure @sd has at least @child cpus. */
6840 cpumask_or(sched_domain_span(sd),
6841 sched_domain_span(sd),
6842 sched_domain_span(child));
6846 set_domain_attribute(sd, attr);
6852 * Build sched domains for a given set of cpus and attach the sched domains
6853 * to the individual cpus
6855 static int build_sched_domains(const struct cpumask *cpu_map,
6856 struct sched_domain_attr *attr)
6858 enum s_alloc alloc_state;
6859 struct sched_domain *sd;
6861 int i, ret = -ENOMEM;
6863 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6864 if (alloc_state != sa_rootdomain)
6867 /* Set up domains for cpus specified by the cpu_map. */
6868 for_each_cpu(i, cpu_map) {
6869 struct sched_domain_topology_level *tl;
6872 for_each_sd_topology(tl) {
6873 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6874 if (tl == sched_domain_topology)
6875 *per_cpu_ptr(d.sd, i) = sd;
6876 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6877 sd->flags |= SD_OVERLAP;
6878 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6883 /* Build the groups for the domains */
6884 for_each_cpu(i, cpu_map) {
6885 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6886 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6887 if (sd->flags & SD_OVERLAP) {
6888 if (build_overlap_sched_groups(sd, i))
6891 if (build_sched_groups(sd, i))
6897 /* Calculate CPU capacity for physical packages and nodes */
6898 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6899 if (!cpumask_test_cpu(i, cpu_map))
6902 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6903 claim_allocations(i, sd);
6904 init_sched_groups_capacity(i, sd);
6908 /* Attach the domains */
6910 for_each_cpu(i, cpu_map) {
6911 sd = *per_cpu_ptr(d.sd, i);
6912 cpu_attach_domain(sd, d.rd, i);
6918 __free_domain_allocs(&d, alloc_state, cpu_map);
6922 static cpumask_var_t *doms_cur; /* current sched domains */
6923 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6924 static struct sched_domain_attr *dattr_cur;
6925 /* attribues of custom domains in 'doms_cur' */
6928 * Special case: If a kmalloc of a doms_cur partition (array of
6929 * cpumask) fails, then fallback to a single sched domain,
6930 * as determined by the single cpumask fallback_doms.
6932 static cpumask_var_t fallback_doms;
6935 * arch_update_cpu_topology lets virtualized architectures update the
6936 * cpu core maps. It is supposed to return 1 if the topology changed
6937 * or 0 if it stayed the same.
6939 int __weak arch_update_cpu_topology(void)
6944 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6947 cpumask_var_t *doms;
6949 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6952 for (i = 0; i < ndoms; i++) {
6953 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6954 free_sched_domains(doms, i);
6961 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6964 for (i = 0; i < ndoms; i++)
6965 free_cpumask_var(doms[i]);
6970 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6971 * For now this just excludes isolated cpus, but could be used to
6972 * exclude other special cases in the future.
6974 static int init_sched_domains(const struct cpumask *cpu_map)
6978 arch_update_cpu_topology();
6980 doms_cur = alloc_sched_domains(ndoms_cur);
6982 doms_cur = &fallback_doms;
6983 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6984 err = build_sched_domains(doms_cur[0], NULL);
6985 register_sched_domain_sysctl();
6991 * Detach sched domains from a group of cpus specified in cpu_map
6992 * These cpus will now be attached to the NULL domain
6994 static void detach_destroy_domains(const struct cpumask *cpu_map)
6999 for_each_cpu(i, cpu_map)
7000 cpu_attach_domain(NULL, &def_root_domain, i);
7004 /* handle null as "default" */
7005 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7006 struct sched_domain_attr *new, int idx_new)
7008 struct sched_domain_attr tmp;
7015 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7016 new ? (new + idx_new) : &tmp,
7017 sizeof(struct sched_domain_attr));
7021 * Partition sched domains as specified by the 'ndoms_new'
7022 * cpumasks in the array doms_new[] of cpumasks. This compares
7023 * doms_new[] to the current sched domain partitioning, doms_cur[].
7024 * It destroys each deleted domain and builds each new domain.
7026 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7027 * The masks don't intersect (don't overlap.) We should setup one
7028 * sched domain for each mask. CPUs not in any of the cpumasks will
7029 * not be load balanced. If the same cpumask appears both in the
7030 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7033 * The passed in 'doms_new' should be allocated using
7034 * alloc_sched_domains. This routine takes ownership of it and will
7035 * free_sched_domains it when done with it. If the caller failed the
7036 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7037 * and partition_sched_domains() will fallback to the single partition
7038 * 'fallback_doms', it also forces the domains to be rebuilt.
7040 * If doms_new == NULL it will be replaced with cpu_online_mask.
7041 * ndoms_new == 0 is a special case for destroying existing domains,
7042 * and it will not create the default domain.
7044 * Call with hotplug lock held
7046 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7047 struct sched_domain_attr *dattr_new)
7052 mutex_lock(&sched_domains_mutex);
7054 /* always unregister in case we don't destroy any domains */
7055 unregister_sched_domain_sysctl();
7057 /* Let architecture update cpu core mappings. */
7058 new_topology = arch_update_cpu_topology();
7060 n = doms_new ? ndoms_new : 0;
7062 /* Destroy deleted domains */
7063 for (i = 0; i < ndoms_cur; i++) {
7064 for (j = 0; j < n && !new_topology; j++) {
7065 if (cpumask_equal(doms_cur[i], doms_new[j])
7066 && dattrs_equal(dattr_cur, i, dattr_new, j))
7069 /* no match - a current sched domain not in new doms_new[] */
7070 detach_destroy_domains(doms_cur[i]);
7076 if (doms_new == NULL) {
7078 doms_new = &fallback_doms;
7079 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7080 WARN_ON_ONCE(dattr_new);
7083 /* Build new domains */
7084 for (i = 0; i < ndoms_new; i++) {
7085 for (j = 0; j < n && !new_topology; j++) {
7086 if (cpumask_equal(doms_new[i], doms_cur[j])
7087 && dattrs_equal(dattr_new, i, dattr_cur, j))
7090 /* no match - add a new doms_new */
7091 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7096 /* Remember the new sched domains */
7097 if (doms_cur != &fallback_doms)
7098 free_sched_domains(doms_cur, ndoms_cur);
7099 kfree(dattr_cur); /* kfree(NULL) is safe */
7100 doms_cur = doms_new;
7101 dattr_cur = dattr_new;
7102 ndoms_cur = ndoms_new;
7104 register_sched_domain_sysctl();
7106 mutex_unlock(&sched_domains_mutex);
7109 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7112 * Update cpusets according to cpu_active mask. If cpusets are
7113 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7114 * around partition_sched_domains().
7116 * If we come here as part of a suspend/resume, don't touch cpusets because we
7117 * want to restore it back to its original state upon resume anyway.
7119 static void cpuset_cpu_active(void)
7121 if (cpuhp_tasks_frozen) {
7123 * num_cpus_frozen tracks how many CPUs are involved in suspend
7124 * resume sequence. As long as this is not the last online
7125 * operation in the resume sequence, just build a single sched
7126 * domain, ignoring cpusets.
7129 if (likely(num_cpus_frozen)) {
7130 partition_sched_domains(1, NULL, NULL);
7134 * This is the last CPU online operation. So fall through and
7135 * restore the original sched domains by considering the
7136 * cpuset configurations.
7139 cpuset_update_active_cpus(true);
7142 static int cpuset_cpu_inactive(unsigned int cpu)
7144 unsigned long flags;
7149 if (!cpuhp_tasks_frozen) {
7150 rcu_read_lock_sched();
7151 dl_b = dl_bw_of(cpu);
7153 raw_spin_lock_irqsave(&dl_b->lock, flags);
7154 cpus = dl_bw_cpus(cpu);
7155 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7156 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7158 rcu_read_unlock_sched();
7162 cpuset_update_active_cpus(false);
7165 partition_sched_domains(1, NULL, NULL);
7170 int sched_cpu_activate(unsigned int cpu)
7172 struct rq *rq = cpu_rq(cpu);
7173 unsigned long flags;
7175 set_cpu_active(cpu, true);
7177 if (sched_smp_initialized) {
7178 sched_domains_numa_masks_set(cpu);
7179 cpuset_cpu_active();
7183 * Put the rq online, if not already. This happens:
7185 * 1) In the early boot process, because we build the real domains
7186 * after all cpus have been brought up.
7188 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7191 raw_spin_lock_irqsave(&rq->lock, flags);
7193 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7196 raw_spin_unlock_irqrestore(&rq->lock, flags);
7198 update_max_interval();
7203 int sched_cpu_deactivate(unsigned int cpu)
7207 set_cpu_active(cpu, false);
7209 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7210 * users of this state to go away such that all new such users will
7213 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7214 * not imply sync_sched(), so wait for both.
7216 * Do sync before park smpboot threads to take care the rcu boost case.
7218 if (IS_ENABLED(CONFIG_PREEMPT))
7219 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7223 if (!sched_smp_initialized)
7226 ret = cpuset_cpu_inactive(cpu);
7228 set_cpu_active(cpu, true);
7231 sched_domains_numa_masks_clear(cpu);
7235 static void sched_rq_cpu_starting(unsigned int cpu)
7237 struct rq *rq = cpu_rq(cpu);
7239 rq->calc_load_update = calc_load_update;
7240 update_max_interval();
7243 int sched_cpu_starting(unsigned int cpu)
7245 set_cpu_rq_start_time(cpu);
7246 sched_rq_cpu_starting(cpu);
7250 #ifdef CONFIG_HOTPLUG_CPU
7251 int sched_cpu_dying(unsigned int cpu)
7253 struct rq *rq = cpu_rq(cpu);
7254 unsigned long flags;
7256 /* Handle pending wakeups and then migrate everything off */
7257 sched_ttwu_pending();
7258 raw_spin_lock_irqsave(&rq->lock, flags);
7260 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7264 BUG_ON(rq->nr_running != 1);
7265 raw_spin_unlock_irqrestore(&rq->lock, flags);
7266 calc_load_migrate(rq);
7267 update_max_interval();
7268 nohz_balance_exit_idle(cpu);
7274 void __init sched_init_smp(void)
7276 cpumask_var_t non_isolated_cpus;
7278 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7279 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7284 * There's no userspace yet to cause hotplug operations; hence all the
7285 * cpu masks are stable and all blatant races in the below code cannot
7288 mutex_lock(&sched_domains_mutex);
7289 init_sched_domains(cpu_active_mask);
7290 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7291 if (cpumask_empty(non_isolated_cpus))
7292 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7293 mutex_unlock(&sched_domains_mutex);
7295 /* Move init over to a non-isolated CPU */
7296 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7298 sched_init_granularity();
7299 free_cpumask_var(non_isolated_cpus);
7301 init_sched_rt_class();
7302 init_sched_dl_class();
7303 sched_smp_initialized = true;
7306 static int __init migration_init(void)
7308 sched_rq_cpu_starting(smp_processor_id());
7311 early_initcall(migration_init);
7314 void __init sched_init_smp(void)
7316 sched_init_granularity();
7318 #endif /* CONFIG_SMP */
7320 int in_sched_functions(unsigned long addr)
7322 return in_lock_functions(addr) ||
7323 (addr >= (unsigned long)__sched_text_start
7324 && addr < (unsigned long)__sched_text_end);
7327 #ifdef CONFIG_CGROUP_SCHED
7329 * Default task group.
7330 * Every task in system belongs to this group at bootup.
7332 struct task_group root_task_group;
7333 LIST_HEAD(task_groups);
7335 /* Cacheline aligned slab cache for task_group */
7336 static struct kmem_cache *task_group_cache __read_mostly;
7339 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7341 void __init sched_init(void)
7344 unsigned long alloc_size = 0, ptr;
7346 #ifdef CONFIG_FAIR_GROUP_SCHED
7347 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7349 #ifdef CONFIG_RT_GROUP_SCHED
7350 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7353 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7355 #ifdef CONFIG_FAIR_GROUP_SCHED
7356 root_task_group.se = (struct sched_entity **)ptr;
7357 ptr += nr_cpu_ids * sizeof(void **);
7359 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7360 ptr += nr_cpu_ids * sizeof(void **);
7362 #endif /* CONFIG_FAIR_GROUP_SCHED */
7363 #ifdef CONFIG_RT_GROUP_SCHED
7364 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7365 ptr += nr_cpu_ids * sizeof(void **);
7367 root_task_group.rt_rq = (struct rt_rq **)ptr;
7368 ptr += nr_cpu_ids * sizeof(void **);
7370 #endif /* CONFIG_RT_GROUP_SCHED */
7372 #ifdef CONFIG_CPUMASK_OFFSTACK
7373 for_each_possible_cpu(i) {
7374 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7375 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7377 #endif /* CONFIG_CPUMASK_OFFSTACK */
7379 init_rt_bandwidth(&def_rt_bandwidth,
7380 global_rt_period(), global_rt_runtime());
7381 init_dl_bandwidth(&def_dl_bandwidth,
7382 global_rt_period(), global_rt_runtime());
7385 init_defrootdomain();
7388 #ifdef CONFIG_RT_GROUP_SCHED
7389 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7390 global_rt_period(), global_rt_runtime());
7391 #endif /* CONFIG_RT_GROUP_SCHED */
7393 #ifdef CONFIG_CGROUP_SCHED
7394 task_group_cache = KMEM_CACHE(task_group, 0);
7396 list_add(&root_task_group.list, &task_groups);
7397 INIT_LIST_HEAD(&root_task_group.children);
7398 INIT_LIST_HEAD(&root_task_group.siblings);
7399 autogroup_init(&init_task);
7400 #endif /* CONFIG_CGROUP_SCHED */
7402 for_each_possible_cpu(i) {
7406 raw_spin_lock_init(&rq->lock);
7408 rq->calc_load_active = 0;
7409 rq->calc_load_update = jiffies + LOAD_FREQ;
7410 init_cfs_rq(&rq->cfs);
7411 init_rt_rq(&rq->rt);
7412 init_dl_rq(&rq->dl);
7413 #ifdef CONFIG_FAIR_GROUP_SCHED
7414 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7415 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7417 * How much cpu bandwidth does root_task_group get?
7419 * In case of task-groups formed thr' the cgroup filesystem, it
7420 * gets 100% of the cpu resources in the system. This overall
7421 * system cpu resource is divided among the tasks of
7422 * root_task_group and its child task-groups in a fair manner,
7423 * based on each entity's (task or task-group's) weight
7424 * (se->load.weight).
7426 * In other words, if root_task_group has 10 tasks of weight
7427 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7428 * then A0's share of the cpu resource is:
7430 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7432 * We achieve this by letting root_task_group's tasks sit
7433 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7435 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7436 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7437 #endif /* CONFIG_FAIR_GROUP_SCHED */
7439 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7440 #ifdef CONFIG_RT_GROUP_SCHED
7441 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7444 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7445 rq->cpu_load[j] = 0;
7450 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7451 rq->balance_callback = NULL;
7452 rq->active_balance = 0;
7453 rq->next_balance = jiffies;
7458 rq->avg_idle = 2*sysctl_sched_migration_cost;
7459 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7461 INIT_LIST_HEAD(&rq->cfs_tasks);
7463 rq_attach_root(rq, &def_root_domain);
7464 #ifdef CONFIG_NO_HZ_COMMON
7465 rq->last_load_update_tick = jiffies;
7468 #ifdef CONFIG_NO_HZ_FULL
7469 rq->last_sched_tick = 0;
7471 #endif /* CONFIG_SMP */
7473 atomic_set(&rq->nr_iowait, 0);
7476 set_load_weight(&init_task);
7478 #ifdef CONFIG_PREEMPT_NOTIFIERS
7479 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7483 * The boot idle thread does lazy MMU switching as well:
7485 atomic_inc(&init_mm.mm_count);
7486 enter_lazy_tlb(&init_mm, current);
7489 * During early bootup we pretend to be a normal task:
7491 current->sched_class = &fair_sched_class;
7494 * Make us the idle thread. Technically, schedule() should not be
7495 * called from this thread, however somewhere below it might be,
7496 * but because we are the idle thread, we just pick up running again
7497 * when this runqueue becomes "idle".
7499 init_idle(current, smp_processor_id());
7501 calc_load_update = jiffies + LOAD_FREQ;
7504 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7505 /* May be allocated at isolcpus cmdline parse time */
7506 if (cpu_isolated_map == NULL)
7507 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7508 idle_thread_set_boot_cpu();
7509 set_cpu_rq_start_time(smp_processor_id());
7511 init_sched_fair_class();
7515 scheduler_running = 1;
7518 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7519 static inline int preempt_count_equals(int preempt_offset)
7521 int nested = preempt_count() + rcu_preempt_depth();
7523 return (nested == preempt_offset);
7526 void __might_sleep(const char *file, int line, int preempt_offset)
7529 * Blocking primitives will set (and therefore destroy) current->state,
7530 * since we will exit with TASK_RUNNING make sure we enter with it,
7531 * otherwise we will destroy state.
7533 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7534 "do not call blocking ops when !TASK_RUNNING; "
7535 "state=%lx set at [<%p>] %pS\n",
7537 (void *)current->task_state_change,
7538 (void *)current->task_state_change);
7540 ___might_sleep(file, line, preempt_offset);
7542 EXPORT_SYMBOL(__might_sleep);
7544 void ___might_sleep(const char *file, int line, int preempt_offset)
7546 static unsigned long prev_jiffy; /* ratelimiting */
7548 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7549 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7550 !is_idle_task(current)) ||
7551 system_state != SYSTEM_RUNNING || oops_in_progress)
7553 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7555 prev_jiffy = jiffies;
7558 "BUG: sleeping function called from invalid context at %s:%d\n",
7561 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7562 in_atomic(), irqs_disabled(),
7563 current->pid, current->comm);
7565 if (task_stack_end_corrupted(current))
7566 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7568 debug_show_held_locks(current);
7569 if (irqs_disabled())
7570 print_irqtrace_events(current);
7571 #ifdef CONFIG_DEBUG_PREEMPT
7572 if (!preempt_count_equals(preempt_offset)) {
7573 pr_err("Preemption disabled at:");
7574 print_ip_sym(current->preempt_disable_ip);
7580 EXPORT_SYMBOL(___might_sleep);
7583 #ifdef CONFIG_MAGIC_SYSRQ
7584 void normalize_rt_tasks(void)
7586 struct task_struct *g, *p;
7587 struct sched_attr attr = {
7588 .sched_policy = SCHED_NORMAL,
7591 read_lock(&tasklist_lock);
7592 for_each_process_thread(g, p) {
7594 * Only normalize user tasks:
7596 if (p->flags & PF_KTHREAD)
7599 p->se.exec_start = 0;
7600 #ifdef CONFIG_SCHEDSTATS
7601 p->se.statistics.wait_start = 0;
7602 p->se.statistics.sleep_start = 0;
7603 p->se.statistics.block_start = 0;
7606 if (!dl_task(p) && !rt_task(p)) {
7608 * Renice negative nice level userspace
7611 if (task_nice(p) < 0)
7612 set_user_nice(p, 0);
7616 __sched_setscheduler(p, &attr, false, false);
7618 read_unlock(&tasklist_lock);
7621 #endif /* CONFIG_MAGIC_SYSRQ */
7623 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7625 * These functions are only useful for the IA64 MCA handling, or kdb.
7627 * They can only be called when the whole system has been
7628 * stopped - every CPU needs to be quiescent, and no scheduling
7629 * activity can take place. Using them for anything else would
7630 * be a serious bug, and as a result, they aren't even visible
7631 * under any other configuration.
7635 * curr_task - return the current task for a given cpu.
7636 * @cpu: the processor in question.
7638 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7640 * Return: The current task for @cpu.
7642 struct task_struct *curr_task(int cpu)
7644 return cpu_curr(cpu);
7647 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7651 * set_curr_task - set the current task for a given cpu.
7652 * @cpu: the processor in question.
7653 * @p: the task pointer to set.
7655 * Description: This function must only be used when non-maskable interrupts
7656 * are serviced on a separate stack. It allows the architecture to switch the
7657 * notion of the current task on a cpu in a non-blocking manner. This function
7658 * must be called with all CPU's synchronized, and interrupts disabled, the
7659 * and caller must save the original value of the current task (see
7660 * curr_task() above) and restore that value before reenabling interrupts and
7661 * re-starting the system.
7663 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7665 void set_curr_task(int cpu, struct task_struct *p)
7672 #ifdef CONFIG_CGROUP_SCHED
7673 /* task_group_lock serializes the addition/removal of task groups */
7674 static DEFINE_SPINLOCK(task_group_lock);
7676 static void sched_free_group(struct task_group *tg)
7678 free_fair_sched_group(tg);
7679 free_rt_sched_group(tg);
7681 kmem_cache_free(task_group_cache, tg);
7684 /* allocate runqueue etc for a new task group */
7685 struct task_group *sched_create_group(struct task_group *parent)
7687 struct task_group *tg;
7689 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7691 return ERR_PTR(-ENOMEM);
7693 if (!alloc_fair_sched_group(tg, parent))
7696 if (!alloc_rt_sched_group(tg, parent))
7702 sched_free_group(tg);
7703 return ERR_PTR(-ENOMEM);
7706 void sched_online_group(struct task_group *tg, struct task_group *parent)
7708 unsigned long flags;
7710 spin_lock_irqsave(&task_group_lock, flags);
7711 list_add_rcu(&tg->list, &task_groups);
7713 WARN_ON(!parent); /* root should already exist */
7715 tg->parent = parent;
7716 INIT_LIST_HEAD(&tg->children);
7717 list_add_rcu(&tg->siblings, &parent->children);
7718 spin_unlock_irqrestore(&task_group_lock, flags);
7721 /* rcu callback to free various structures associated with a task group */
7722 static void sched_free_group_rcu(struct rcu_head *rhp)
7724 /* now it should be safe to free those cfs_rqs */
7725 sched_free_group(container_of(rhp, struct task_group, rcu));
7728 void sched_destroy_group(struct task_group *tg)
7730 /* wait for possible concurrent references to cfs_rqs complete */
7731 call_rcu(&tg->rcu, sched_free_group_rcu);
7734 void sched_offline_group(struct task_group *tg)
7736 unsigned long flags;
7738 /* end participation in shares distribution */
7739 unregister_fair_sched_group(tg);
7741 spin_lock_irqsave(&task_group_lock, flags);
7742 list_del_rcu(&tg->list);
7743 list_del_rcu(&tg->siblings);
7744 spin_unlock_irqrestore(&task_group_lock, flags);
7747 static void sched_change_group(struct task_struct *tsk, int type)
7749 struct task_group *tg;
7752 * All callers are synchronized by task_rq_lock(); we do not use RCU
7753 * which is pointless here. Thus, we pass "true" to task_css_check()
7754 * to prevent lockdep warnings.
7756 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7757 struct task_group, css);
7758 tg = autogroup_task_group(tsk, tg);
7759 tsk->sched_task_group = tg;
7761 #ifdef CONFIG_FAIR_GROUP_SCHED
7762 if (tsk->sched_class->task_change_group)
7763 tsk->sched_class->task_change_group(tsk, type);
7766 set_task_rq(tsk, task_cpu(tsk));
7770 * Change task's runqueue when it moves between groups.
7772 * The caller of this function should have put the task in its new group by
7773 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7776 void sched_move_task(struct task_struct *tsk)
7778 int queued, running;
7782 rq = task_rq_lock(tsk, &rf);
7784 running = task_current(rq, tsk);
7785 queued = task_on_rq_queued(tsk);
7788 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7789 if (unlikely(running))
7790 put_prev_task(rq, tsk);
7792 sched_change_group(tsk, TASK_MOVE_GROUP);
7794 if (unlikely(running))
7795 tsk->sched_class->set_curr_task(rq);
7797 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7799 task_rq_unlock(rq, tsk, &rf);
7801 #endif /* CONFIG_CGROUP_SCHED */
7803 #ifdef CONFIG_RT_GROUP_SCHED
7805 * Ensure that the real time constraints are schedulable.
7807 static DEFINE_MUTEX(rt_constraints_mutex);
7809 /* Must be called with tasklist_lock held */
7810 static inline int tg_has_rt_tasks(struct task_group *tg)
7812 struct task_struct *g, *p;
7815 * Autogroups do not have RT tasks; see autogroup_create().
7817 if (task_group_is_autogroup(tg))
7820 for_each_process_thread(g, p) {
7821 if (rt_task(p) && task_group(p) == tg)
7828 struct rt_schedulable_data {
7829 struct task_group *tg;
7834 static int tg_rt_schedulable(struct task_group *tg, void *data)
7836 struct rt_schedulable_data *d = data;
7837 struct task_group *child;
7838 unsigned long total, sum = 0;
7839 u64 period, runtime;
7841 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7842 runtime = tg->rt_bandwidth.rt_runtime;
7845 period = d->rt_period;
7846 runtime = d->rt_runtime;
7850 * Cannot have more runtime than the period.
7852 if (runtime > period && runtime != RUNTIME_INF)
7856 * Ensure we don't starve existing RT tasks.
7858 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7861 total = to_ratio(period, runtime);
7864 * Nobody can have more than the global setting allows.
7866 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7870 * The sum of our children's runtime should not exceed our own.
7872 list_for_each_entry_rcu(child, &tg->children, siblings) {
7873 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7874 runtime = child->rt_bandwidth.rt_runtime;
7876 if (child == d->tg) {
7877 period = d->rt_period;
7878 runtime = d->rt_runtime;
7881 sum += to_ratio(period, runtime);
7890 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7894 struct rt_schedulable_data data = {
7896 .rt_period = period,
7897 .rt_runtime = runtime,
7901 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7907 static int tg_set_rt_bandwidth(struct task_group *tg,
7908 u64 rt_period, u64 rt_runtime)
7913 * Disallowing the root group RT runtime is BAD, it would disallow the
7914 * kernel creating (and or operating) RT threads.
7916 if (tg == &root_task_group && rt_runtime == 0)
7919 /* No period doesn't make any sense. */
7923 mutex_lock(&rt_constraints_mutex);
7924 read_lock(&tasklist_lock);
7925 err = __rt_schedulable(tg, rt_period, rt_runtime);
7929 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7930 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7931 tg->rt_bandwidth.rt_runtime = rt_runtime;
7933 for_each_possible_cpu(i) {
7934 struct rt_rq *rt_rq = tg->rt_rq[i];
7936 raw_spin_lock(&rt_rq->rt_runtime_lock);
7937 rt_rq->rt_runtime = rt_runtime;
7938 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7940 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7942 read_unlock(&tasklist_lock);
7943 mutex_unlock(&rt_constraints_mutex);
7948 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7950 u64 rt_runtime, rt_period;
7952 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7953 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7954 if (rt_runtime_us < 0)
7955 rt_runtime = RUNTIME_INF;
7957 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7960 static long sched_group_rt_runtime(struct task_group *tg)
7964 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7967 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7968 do_div(rt_runtime_us, NSEC_PER_USEC);
7969 return rt_runtime_us;
7972 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7974 u64 rt_runtime, rt_period;
7976 rt_period = rt_period_us * NSEC_PER_USEC;
7977 rt_runtime = tg->rt_bandwidth.rt_runtime;
7979 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7982 static long sched_group_rt_period(struct task_group *tg)
7986 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7987 do_div(rt_period_us, NSEC_PER_USEC);
7988 return rt_period_us;
7990 #endif /* CONFIG_RT_GROUP_SCHED */
7992 #ifdef CONFIG_RT_GROUP_SCHED
7993 static int sched_rt_global_constraints(void)
7997 mutex_lock(&rt_constraints_mutex);
7998 read_lock(&tasklist_lock);
7999 ret = __rt_schedulable(NULL, 0, 0);
8000 read_unlock(&tasklist_lock);
8001 mutex_unlock(&rt_constraints_mutex);
8006 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8008 /* Don't accept realtime tasks when there is no way for them to run */
8009 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8015 #else /* !CONFIG_RT_GROUP_SCHED */
8016 static int sched_rt_global_constraints(void)
8018 unsigned long flags;
8021 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8022 for_each_possible_cpu(i) {
8023 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8025 raw_spin_lock(&rt_rq->rt_runtime_lock);
8026 rt_rq->rt_runtime = global_rt_runtime();
8027 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8029 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8033 #endif /* CONFIG_RT_GROUP_SCHED */
8035 static int sched_dl_global_validate(void)
8037 u64 runtime = global_rt_runtime();
8038 u64 period = global_rt_period();
8039 u64 new_bw = to_ratio(period, runtime);
8042 unsigned long flags;
8045 * Here we want to check the bandwidth not being set to some
8046 * value smaller than the currently allocated bandwidth in
8047 * any of the root_domains.
8049 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8050 * cycling on root_domains... Discussion on different/better
8051 * solutions is welcome!
8053 for_each_possible_cpu(cpu) {
8054 rcu_read_lock_sched();
8055 dl_b = dl_bw_of(cpu);
8057 raw_spin_lock_irqsave(&dl_b->lock, flags);
8058 if (new_bw < dl_b->total_bw)
8060 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8062 rcu_read_unlock_sched();
8071 static void sched_dl_do_global(void)
8076 unsigned long flags;
8078 def_dl_bandwidth.dl_period = global_rt_period();
8079 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8081 if (global_rt_runtime() != RUNTIME_INF)
8082 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8085 * FIXME: As above...
8087 for_each_possible_cpu(cpu) {
8088 rcu_read_lock_sched();
8089 dl_b = dl_bw_of(cpu);
8091 raw_spin_lock_irqsave(&dl_b->lock, flags);
8093 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8095 rcu_read_unlock_sched();
8099 static int sched_rt_global_validate(void)
8101 if (sysctl_sched_rt_period <= 0)
8104 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8105 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8111 static void sched_rt_do_global(void)
8113 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8114 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8117 int sched_rt_handler(struct ctl_table *table, int write,
8118 void __user *buffer, size_t *lenp,
8121 int old_period, old_runtime;
8122 static DEFINE_MUTEX(mutex);
8126 old_period = sysctl_sched_rt_period;
8127 old_runtime = sysctl_sched_rt_runtime;
8129 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8131 if (!ret && write) {
8132 ret = sched_rt_global_validate();
8136 ret = sched_dl_global_validate();
8140 ret = sched_rt_global_constraints();
8144 sched_rt_do_global();
8145 sched_dl_do_global();
8149 sysctl_sched_rt_period = old_period;
8150 sysctl_sched_rt_runtime = old_runtime;
8152 mutex_unlock(&mutex);
8157 int sched_rr_handler(struct ctl_table *table, int write,
8158 void __user *buffer, size_t *lenp,
8162 static DEFINE_MUTEX(mutex);
8165 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8166 /* make sure that internally we keep jiffies */
8167 /* also, writing zero resets timeslice to default */
8168 if (!ret && write) {
8169 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8170 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8172 mutex_unlock(&mutex);
8176 #ifdef CONFIG_CGROUP_SCHED
8178 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8180 return css ? container_of(css, struct task_group, css) : NULL;
8183 static struct cgroup_subsys_state *
8184 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8186 struct task_group *parent = css_tg(parent_css);
8187 struct task_group *tg;
8190 /* This is early initialization for the top cgroup */
8191 return &root_task_group.css;
8194 tg = sched_create_group(parent);
8196 return ERR_PTR(-ENOMEM);
8198 sched_online_group(tg, parent);
8203 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8205 struct task_group *tg = css_tg(css);
8207 sched_offline_group(tg);
8210 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8212 struct task_group *tg = css_tg(css);
8215 * Relies on the RCU grace period between css_released() and this.
8217 sched_free_group(tg);
8221 * This is called before wake_up_new_task(), therefore we really only
8222 * have to set its group bits, all the other stuff does not apply.
8224 static void cpu_cgroup_fork(struct task_struct *task)
8229 rq = task_rq_lock(task, &rf);
8231 sched_change_group(task, TASK_SET_GROUP);
8233 task_rq_unlock(rq, task, &rf);
8236 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8238 struct task_struct *task;
8239 struct cgroup_subsys_state *css;
8241 cgroup_taskset_for_each(task, css, tset) {
8242 #ifdef CONFIG_RT_GROUP_SCHED
8243 if (!sched_rt_can_attach(css_tg(css), task))
8246 /* We don't support RT-tasks being in separate groups */
8247 if (task->sched_class != &fair_sched_class)
8254 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8256 struct task_struct *task;
8257 struct cgroup_subsys_state *css;
8259 cgroup_taskset_for_each(task, css, tset)
8260 sched_move_task(task);
8263 #ifdef CONFIG_FAIR_GROUP_SCHED
8264 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8265 struct cftype *cftype, u64 shareval)
8267 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8270 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8273 struct task_group *tg = css_tg(css);
8275 return (u64) scale_load_down(tg->shares);
8278 #ifdef CONFIG_CFS_BANDWIDTH
8279 static DEFINE_MUTEX(cfs_constraints_mutex);
8281 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8282 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8284 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8286 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8288 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8289 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8291 if (tg == &root_task_group)
8295 * Ensure we have at some amount of bandwidth every period. This is
8296 * to prevent reaching a state of large arrears when throttled via
8297 * entity_tick() resulting in prolonged exit starvation.
8299 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8303 * Likewise, bound things on the otherside by preventing insane quota
8304 * periods. This also allows us to normalize in computing quota
8307 if (period > max_cfs_quota_period)
8311 * Prevent race between setting of cfs_rq->runtime_enabled and
8312 * unthrottle_offline_cfs_rqs().
8315 mutex_lock(&cfs_constraints_mutex);
8316 ret = __cfs_schedulable(tg, period, quota);
8320 runtime_enabled = quota != RUNTIME_INF;
8321 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8323 * If we need to toggle cfs_bandwidth_used, off->on must occur
8324 * before making related changes, and on->off must occur afterwards
8326 if (runtime_enabled && !runtime_was_enabled)
8327 cfs_bandwidth_usage_inc();
8328 raw_spin_lock_irq(&cfs_b->lock);
8329 cfs_b->period = ns_to_ktime(period);
8330 cfs_b->quota = quota;
8332 __refill_cfs_bandwidth_runtime(cfs_b);
8333 /* restart the period timer (if active) to handle new period expiry */
8334 if (runtime_enabled)
8335 start_cfs_bandwidth(cfs_b);
8336 raw_spin_unlock_irq(&cfs_b->lock);
8338 for_each_online_cpu(i) {
8339 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8340 struct rq *rq = cfs_rq->rq;
8342 raw_spin_lock_irq(&rq->lock);
8343 cfs_rq->runtime_enabled = runtime_enabled;
8344 cfs_rq->runtime_remaining = 0;
8346 if (cfs_rq->throttled)
8347 unthrottle_cfs_rq(cfs_rq);
8348 raw_spin_unlock_irq(&rq->lock);
8350 if (runtime_was_enabled && !runtime_enabled)
8351 cfs_bandwidth_usage_dec();
8353 mutex_unlock(&cfs_constraints_mutex);
8359 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8363 period = ktime_to_ns(tg->cfs_bandwidth.period);
8364 if (cfs_quota_us < 0)
8365 quota = RUNTIME_INF;
8367 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8369 return tg_set_cfs_bandwidth(tg, period, quota);
8372 long tg_get_cfs_quota(struct task_group *tg)
8376 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8379 quota_us = tg->cfs_bandwidth.quota;
8380 do_div(quota_us, NSEC_PER_USEC);
8385 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8389 period = (u64)cfs_period_us * NSEC_PER_USEC;
8390 quota = tg->cfs_bandwidth.quota;
8392 return tg_set_cfs_bandwidth(tg, period, quota);
8395 long tg_get_cfs_period(struct task_group *tg)
8399 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8400 do_div(cfs_period_us, NSEC_PER_USEC);
8402 return cfs_period_us;
8405 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8408 return tg_get_cfs_quota(css_tg(css));
8411 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8412 struct cftype *cftype, s64 cfs_quota_us)
8414 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8417 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8420 return tg_get_cfs_period(css_tg(css));
8423 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8424 struct cftype *cftype, u64 cfs_period_us)
8426 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8429 struct cfs_schedulable_data {
8430 struct task_group *tg;
8435 * normalize group quota/period to be quota/max_period
8436 * note: units are usecs
8438 static u64 normalize_cfs_quota(struct task_group *tg,
8439 struct cfs_schedulable_data *d)
8447 period = tg_get_cfs_period(tg);
8448 quota = tg_get_cfs_quota(tg);
8451 /* note: these should typically be equivalent */
8452 if (quota == RUNTIME_INF || quota == -1)
8455 return to_ratio(period, quota);
8458 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8460 struct cfs_schedulable_data *d = data;
8461 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8462 s64 quota = 0, parent_quota = -1;
8465 quota = RUNTIME_INF;
8467 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8469 quota = normalize_cfs_quota(tg, d);
8470 parent_quota = parent_b->hierarchical_quota;
8473 * ensure max(child_quota) <= parent_quota, inherit when no
8476 if (quota == RUNTIME_INF)
8477 quota = parent_quota;
8478 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8481 cfs_b->hierarchical_quota = quota;
8486 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8489 struct cfs_schedulable_data data = {
8495 if (quota != RUNTIME_INF) {
8496 do_div(data.period, NSEC_PER_USEC);
8497 do_div(data.quota, NSEC_PER_USEC);
8501 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8507 static int cpu_stats_show(struct seq_file *sf, void *v)
8509 struct task_group *tg = css_tg(seq_css(sf));
8510 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8512 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8513 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8514 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8518 #endif /* CONFIG_CFS_BANDWIDTH */
8519 #endif /* CONFIG_FAIR_GROUP_SCHED */
8521 #ifdef CONFIG_RT_GROUP_SCHED
8522 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8523 struct cftype *cft, s64 val)
8525 return sched_group_set_rt_runtime(css_tg(css), val);
8528 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8531 return sched_group_rt_runtime(css_tg(css));
8534 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8535 struct cftype *cftype, u64 rt_period_us)
8537 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8540 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8543 return sched_group_rt_period(css_tg(css));
8545 #endif /* CONFIG_RT_GROUP_SCHED */
8547 static struct cftype cpu_files[] = {
8548 #ifdef CONFIG_FAIR_GROUP_SCHED
8551 .read_u64 = cpu_shares_read_u64,
8552 .write_u64 = cpu_shares_write_u64,
8555 #ifdef CONFIG_CFS_BANDWIDTH
8557 .name = "cfs_quota_us",
8558 .read_s64 = cpu_cfs_quota_read_s64,
8559 .write_s64 = cpu_cfs_quota_write_s64,
8562 .name = "cfs_period_us",
8563 .read_u64 = cpu_cfs_period_read_u64,
8564 .write_u64 = cpu_cfs_period_write_u64,
8568 .seq_show = cpu_stats_show,
8571 #ifdef CONFIG_RT_GROUP_SCHED
8573 .name = "rt_runtime_us",
8574 .read_s64 = cpu_rt_runtime_read,
8575 .write_s64 = cpu_rt_runtime_write,
8578 .name = "rt_period_us",
8579 .read_u64 = cpu_rt_period_read_uint,
8580 .write_u64 = cpu_rt_period_write_uint,
8586 struct cgroup_subsys cpu_cgrp_subsys = {
8587 .css_alloc = cpu_cgroup_css_alloc,
8588 .css_released = cpu_cgroup_css_released,
8589 .css_free = cpu_cgroup_css_free,
8590 .fork = cpu_cgroup_fork,
8591 .can_attach = cpu_cgroup_can_attach,
8592 .attach = cpu_cgroup_attach,
8593 .legacy_cftypes = cpu_files,
8597 #endif /* CONFIG_CGROUP_SCHED */
8599 void dump_cpu_task(int cpu)
8601 pr_info("Task dump for CPU %d:\n", cpu);
8602 sched_show_task(cpu_curr(cpu));
8606 * Nice levels are multiplicative, with a gentle 10% change for every
8607 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8608 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8609 * that remained on nice 0.
8611 * The "10% effect" is relative and cumulative: from _any_ nice level,
8612 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8613 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8614 * If a task goes up by ~10% and another task goes down by ~10% then
8615 * the relative distance between them is ~25%.)
8617 const int sched_prio_to_weight[40] = {
8618 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8619 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8620 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8621 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8622 /* 0 */ 1024, 820, 655, 526, 423,
8623 /* 5 */ 335, 272, 215, 172, 137,
8624 /* 10 */ 110, 87, 70, 56, 45,
8625 /* 15 */ 36, 29, 23, 18, 15,
8629 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8631 * In cases where the weight does not change often, we can use the
8632 * precalculated inverse to speed up arithmetics by turning divisions
8633 * into multiplications:
8635 const u32 sched_prio_to_wmult[40] = {
8636 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8637 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8638 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8639 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8640 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8641 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8642 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8643 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,