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 (!cpu_active(dest_cpu))
1544 /* No more Mr. Nice Guy. */
1547 if (IS_ENABLED(CONFIG_CPUSETS)) {
1548 cpuset_cpus_allowed_fallback(p);
1554 do_set_cpus_allowed(p, cpu_possible_mask);
1565 if (state != cpuset) {
1567 * Don't tell them about moving exiting tasks or
1568 * kernel threads (both mm NULL), since they never
1571 if (p->mm && printk_ratelimit()) {
1572 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1573 task_pid_nr(p), p->comm, cpu);
1581 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1584 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1586 lockdep_assert_held(&p->pi_lock);
1588 if (tsk_nr_cpus_allowed(p) > 1)
1589 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1591 cpu = cpumask_any(tsk_cpus_allowed(p));
1594 * In order not to call set_task_cpu() on a blocking task we need
1595 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1598 * Since this is common to all placement strategies, this lives here.
1600 * [ this allows ->select_task() to simply return task_cpu(p) and
1601 * not worry about this generic constraint ]
1603 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1605 cpu = select_fallback_rq(task_cpu(p), p);
1610 static void update_avg(u64 *avg, u64 sample)
1612 s64 diff = sample - *avg;
1618 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1619 const struct cpumask *new_mask, bool check)
1621 return set_cpus_allowed_ptr(p, new_mask);
1624 #endif /* CONFIG_SMP */
1627 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1629 #ifdef CONFIG_SCHEDSTATS
1630 struct rq *rq = this_rq();
1633 int this_cpu = smp_processor_id();
1635 if (cpu == this_cpu) {
1636 schedstat_inc(rq, ttwu_local);
1637 schedstat_inc(p, se.statistics.nr_wakeups_local);
1639 struct sched_domain *sd;
1641 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1643 for_each_domain(this_cpu, sd) {
1644 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1645 schedstat_inc(sd, ttwu_wake_remote);
1652 if (wake_flags & WF_MIGRATED)
1653 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1655 #endif /* CONFIG_SMP */
1657 schedstat_inc(rq, ttwu_count);
1658 schedstat_inc(p, se.statistics.nr_wakeups);
1660 if (wake_flags & WF_SYNC)
1661 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1663 #endif /* CONFIG_SCHEDSTATS */
1666 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1668 activate_task(rq, p, en_flags);
1669 p->on_rq = TASK_ON_RQ_QUEUED;
1671 /* if a worker is waking up, notify workqueue */
1672 if (p->flags & PF_WQ_WORKER)
1673 wq_worker_waking_up(p, cpu_of(rq));
1677 * Mark the task runnable and perform wakeup-preemption.
1679 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1680 struct pin_cookie cookie)
1682 check_preempt_curr(rq, p, wake_flags);
1683 p->state = TASK_RUNNING;
1684 trace_sched_wakeup(p);
1687 if (p->sched_class->task_woken) {
1689 * Our task @p is fully woken up and running; so its safe to
1690 * drop the rq->lock, hereafter rq is only used for statistics.
1692 lockdep_unpin_lock(&rq->lock, cookie);
1693 p->sched_class->task_woken(rq, p);
1694 lockdep_repin_lock(&rq->lock, cookie);
1697 if (rq->idle_stamp) {
1698 u64 delta = rq_clock(rq) - rq->idle_stamp;
1699 u64 max = 2*rq->max_idle_balance_cost;
1701 update_avg(&rq->avg_idle, delta);
1703 if (rq->avg_idle > max)
1712 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1713 struct pin_cookie cookie)
1715 int en_flags = ENQUEUE_WAKEUP;
1717 lockdep_assert_held(&rq->lock);
1720 if (p->sched_contributes_to_load)
1721 rq->nr_uninterruptible--;
1723 if (wake_flags & WF_MIGRATED)
1724 en_flags |= ENQUEUE_MIGRATED;
1727 ttwu_activate(rq, p, en_flags);
1728 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1732 * Called in case the task @p isn't fully descheduled from its runqueue,
1733 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1734 * since all we need to do is flip p->state to TASK_RUNNING, since
1735 * the task is still ->on_rq.
1737 static int ttwu_remote(struct task_struct *p, int wake_flags)
1743 rq = __task_rq_lock(p, &rf);
1744 if (task_on_rq_queued(p)) {
1745 /* check_preempt_curr() may use rq clock */
1746 update_rq_clock(rq);
1747 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1750 __task_rq_unlock(rq, &rf);
1756 void sched_ttwu_pending(void)
1758 struct rq *rq = this_rq();
1759 struct llist_node *llist = llist_del_all(&rq->wake_list);
1760 struct pin_cookie cookie;
1761 struct task_struct *p;
1762 unsigned long flags;
1767 raw_spin_lock_irqsave(&rq->lock, flags);
1768 cookie = lockdep_pin_lock(&rq->lock);
1771 p = llist_entry(llist, struct task_struct, wake_entry);
1772 llist = llist_next(llist);
1774 * See ttwu_queue(); we only call ttwu_queue_remote() when
1775 * its a x-cpu wakeup.
1777 ttwu_do_activate(rq, p, WF_MIGRATED, cookie);
1780 lockdep_unpin_lock(&rq->lock, cookie);
1781 raw_spin_unlock_irqrestore(&rq->lock, flags);
1784 void scheduler_ipi(void)
1787 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1788 * TIF_NEED_RESCHED remotely (for the first time) will also send
1791 preempt_fold_need_resched();
1793 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1797 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1798 * traditionally all their work was done from the interrupt return
1799 * path. Now that we actually do some work, we need to make sure
1802 * Some archs already do call them, luckily irq_enter/exit nest
1805 * Arguably we should visit all archs and update all handlers,
1806 * however a fair share of IPIs are still resched only so this would
1807 * somewhat pessimize the simple resched case.
1810 sched_ttwu_pending();
1813 * Check if someone kicked us for doing the nohz idle load balance.
1815 if (unlikely(got_nohz_idle_kick())) {
1816 this_rq()->idle_balance = 1;
1817 raise_softirq_irqoff(SCHED_SOFTIRQ);
1822 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1824 struct rq *rq = cpu_rq(cpu);
1826 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1827 if (!set_nr_if_polling(rq->idle))
1828 smp_send_reschedule(cpu);
1830 trace_sched_wake_idle_without_ipi(cpu);
1834 void wake_up_if_idle(int cpu)
1836 struct rq *rq = cpu_rq(cpu);
1837 unsigned long flags;
1841 if (!is_idle_task(rcu_dereference(rq->curr)))
1844 if (set_nr_if_polling(rq->idle)) {
1845 trace_sched_wake_idle_without_ipi(cpu);
1847 raw_spin_lock_irqsave(&rq->lock, flags);
1848 if (is_idle_task(rq->curr))
1849 smp_send_reschedule(cpu);
1850 /* Else cpu is not in idle, do nothing here */
1851 raw_spin_unlock_irqrestore(&rq->lock, flags);
1858 bool cpus_share_cache(int this_cpu, int that_cpu)
1860 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1862 #endif /* CONFIG_SMP */
1864 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1866 struct rq *rq = cpu_rq(cpu);
1867 struct pin_cookie cookie;
1869 #if defined(CONFIG_SMP)
1870 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1871 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1872 ttwu_queue_remote(p, cpu);
1877 raw_spin_lock(&rq->lock);
1878 cookie = lockdep_pin_lock(&rq->lock);
1879 ttwu_do_activate(rq, p, wake_flags, cookie);
1880 lockdep_unpin_lock(&rq->lock, cookie);
1881 raw_spin_unlock(&rq->lock);
1885 * Notes on Program-Order guarantees on SMP systems.
1889 * The basic program-order guarantee on SMP systems is that when a task [t]
1890 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1891 * execution on its new cpu [c1].
1893 * For migration (of runnable tasks) this is provided by the following means:
1895 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1896 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1897 * rq(c1)->lock (if not at the same time, then in that order).
1898 * C) LOCK of the rq(c1)->lock scheduling in task
1900 * Transitivity guarantees that B happens after A and C after B.
1901 * Note: we only require RCpc transitivity.
1902 * Note: the cpu doing B need not be c0 or c1
1911 * UNLOCK rq(0)->lock
1913 * LOCK rq(0)->lock // orders against CPU0
1915 * UNLOCK rq(0)->lock
1919 * UNLOCK rq(1)->lock
1921 * LOCK rq(1)->lock // orders against CPU2
1924 * UNLOCK rq(1)->lock
1927 * BLOCKING -- aka. SLEEP + WAKEUP
1929 * For blocking we (obviously) need to provide the same guarantee as for
1930 * migration. However the means are completely different as there is no lock
1931 * chain to provide order. Instead we do:
1933 * 1) smp_store_release(X->on_cpu, 0)
1934 * 2) smp_cond_acquire(!X->on_cpu)
1938 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1940 * LOCK rq(0)->lock LOCK X->pi_lock
1943 * smp_store_release(X->on_cpu, 0);
1945 * smp_cond_acquire(!X->on_cpu);
1951 * X->state = RUNNING
1952 * UNLOCK rq(2)->lock
1954 * LOCK rq(2)->lock // orders against CPU1
1957 * UNLOCK rq(2)->lock
1960 * UNLOCK rq(0)->lock
1963 * However; for wakeups there is a second guarantee we must provide, namely we
1964 * must observe the state that lead to our wakeup. That is, not only must our
1965 * task observe its own prior state, it must also observe the stores prior to
1968 * This means that any means of doing remote wakeups must order the CPU doing
1969 * the wakeup against the CPU the task is going to end up running on. This,
1970 * however, is already required for the regular Program-Order guarantee above,
1971 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1976 * try_to_wake_up - wake up a thread
1977 * @p: the thread to be awakened
1978 * @state: the mask of task states that can be woken
1979 * @wake_flags: wake modifier flags (WF_*)
1981 * Put it on the run-queue if it's not already there. The "current"
1982 * thread is always on the run-queue (except when the actual
1983 * re-schedule is in progress), and as such you're allowed to do
1984 * the simpler "current->state = TASK_RUNNING" to mark yourself
1985 * runnable without the overhead of this.
1987 * Return: %true if @p was woken up, %false if it was already running.
1988 * or @state didn't match @p's state.
1991 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1993 unsigned long flags;
1994 int cpu, success = 0;
1997 * If we are going to wake up a thread waiting for CONDITION we
1998 * need to ensure that CONDITION=1 done by the caller can not be
1999 * reordered with p->state check below. This pairs with mb() in
2000 * set_current_state() the waiting thread does.
2002 smp_mb__before_spinlock();
2003 raw_spin_lock_irqsave(&p->pi_lock, flags);
2004 if (!(p->state & state))
2007 trace_sched_waking(p);
2009 success = 1; /* we're going to change ->state */
2012 if (p->on_rq && ttwu_remote(p, wake_flags))
2017 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2018 * possible to, falsely, observe p->on_cpu == 0.
2020 * One must be running (->on_cpu == 1) in order to remove oneself
2021 * from the runqueue.
2023 * [S] ->on_cpu = 1; [L] ->on_rq
2027 * [S] ->on_rq = 0; [L] ->on_cpu
2029 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2030 * from the consecutive calls to schedule(); the first switching to our
2031 * task, the second putting it to sleep.
2036 * If the owning (remote) cpu is still in the middle of schedule() with
2037 * this task as prev, wait until its done referencing the task.
2039 * Pairs with the smp_store_release() in finish_lock_switch().
2041 * This ensures that tasks getting woken will be fully ordered against
2042 * their previous state and preserve Program Order.
2044 smp_cond_acquire(!p->on_cpu);
2046 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2047 p->state = TASK_WAKING;
2049 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2050 if (task_cpu(p) != cpu) {
2051 wake_flags |= WF_MIGRATED;
2052 set_task_cpu(p, cpu);
2054 #endif /* CONFIG_SMP */
2056 ttwu_queue(p, cpu, wake_flags);
2058 if (schedstat_enabled())
2059 ttwu_stat(p, cpu, wake_flags);
2061 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2067 * try_to_wake_up_local - try to wake up a local task with rq lock held
2068 * @p: the thread to be awakened
2070 * Put @p on the run-queue if it's not already there. The caller must
2071 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2074 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2076 struct rq *rq = task_rq(p);
2078 if (WARN_ON_ONCE(rq != this_rq()) ||
2079 WARN_ON_ONCE(p == current))
2082 lockdep_assert_held(&rq->lock);
2084 if (!raw_spin_trylock(&p->pi_lock)) {
2086 * This is OK, because current is on_cpu, which avoids it being
2087 * picked for load-balance and preemption/IRQs are still
2088 * disabled avoiding further scheduler activity on it and we've
2089 * not yet picked a replacement task.
2091 lockdep_unpin_lock(&rq->lock, cookie);
2092 raw_spin_unlock(&rq->lock);
2093 raw_spin_lock(&p->pi_lock);
2094 raw_spin_lock(&rq->lock);
2095 lockdep_repin_lock(&rq->lock, cookie);
2098 if (!(p->state & TASK_NORMAL))
2101 trace_sched_waking(p);
2103 if (!task_on_rq_queued(p))
2104 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2106 ttwu_do_wakeup(rq, p, 0, cookie);
2107 if (schedstat_enabled())
2108 ttwu_stat(p, smp_processor_id(), 0);
2110 raw_spin_unlock(&p->pi_lock);
2114 * wake_up_process - Wake up a specific process
2115 * @p: The process to be woken up.
2117 * Attempt to wake up the nominated process and move it to the set of runnable
2120 * Return: 1 if the process was woken up, 0 if it was already running.
2122 * It may be assumed that this function implies a write memory barrier before
2123 * changing the task state if and only if any tasks are woken up.
2125 int wake_up_process(struct task_struct *p)
2127 return try_to_wake_up(p, TASK_NORMAL, 0);
2129 EXPORT_SYMBOL(wake_up_process);
2131 int wake_up_state(struct task_struct *p, unsigned int state)
2133 return try_to_wake_up(p, state, 0);
2137 * This function clears the sched_dl_entity static params.
2139 void __dl_clear_params(struct task_struct *p)
2141 struct sched_dl_entity *dl_se = &p->dl;
2143 dl_se->dl_runtime = 0;
2144 dl_se->dl_deadline = 0;
2145 dl_se->dl_period = 0;
2149 dl_se->dl_throttled = 0;
2150 dl_se->dl_yielded = 0;
2154 * Perform scheduler related setup for a newly forked process p.
2155 * p is forked by current.
2157 * __sched_fork() is basic setup used by init_idle() too:
2159 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2164 p->se.exec_start = 0;
2165 p->se.sum_exec_runtime = 0;
2166 p->se.prev_sum_exec_runtime = 0;
2167 p->se.nr_migrations = 0;
2169 INIT_LIST_HEAD(&p->se.group_node);
2171 #ifdef CONFIG_FAIR_GROUP_SCHED
2172 p->se.cfs_rq = NULL;
2175 #ifdef CONFIG_SCHEDSTATS
2176 /* Even if schedstat is disabled, there should not be garbage */
2177 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2180 RB_CLEAR_NODE(&p->dl.rb_node);
2181 init_dl_task_timer(&p->dl);
2182 __dl_clear_params(p);
2184 INIT_LIST_HEAD(&p->rt.run_list);
2186 p->rt.time_slice = sched_rr_timeslice;
2190 #ifdef CONFIG_PREEMPT_NOTIFIERS
2191 INIT_HLIST_HEAD(&p->preempt_notifiers);
2194 #ifdef CONFIG_NUMA_BALANCING
2195 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2196 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2197 p->mm->numa_scan_seq = 0;
2200 if (clone_flags & CLONE_VM)
2201 p->numa_preferred_nid = current->numa_preferred_nid;
2203 p->numa_preferred_nid = -1;
2205 p->node_stamp = 0ULL;
2206 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2207 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2208 p->numa_work.next = &p->numa_work;
2209 p->numa_faults = NULL;
2210 p->last_task_numa_placement = 0;
2211 p->last_sum_exec_runtime = 0;
2213 p->numa_group = NULL;
2214 #endif /* CONFIG_NUMA_BALANCING */
2217 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2219 #ifdef CONFIG_NUMA_BALANCING
2221 void set_numabalancing_state(bool enabled)
2224 static_branch_enable(&sched_numa_balancing);
2226 static_branch_disable(&sched_numa_balancing);
2229 #ifdef CONFIG_PROC_SYSCTL
2230 int sysctl_numa_balancing(struct ctl_table *table, int write,
2231 void __user *buffer, size_t *lenp, loff_t *ppos)
2235 int state = static_branch_likely(&sched_numa_balancing);
2237 if (write && !capable(CAP_SYS_ADMIN))
2242 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2246 set_numabalancing_state(state);
2252 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2254 #ifdef CONFIG_SCHEDSTATS
2255 static void set_schedstats(bool enabled)
2258 static_branch_enable(&sched_schedstats);
2260 static_branch_disable(&sched_schedstats);
2263 void force_schedstat_enabled(void)
2265 if (!schedstat_enabled()) {
2266 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2267 static_branch_enable(&sched_schedstats);
2271 static int __init setup_schedstats(char *str)
2277 if (!strcmp(str, "enable")) {
2278 set_schedstats(true);
2280 } else if (!strcmp(str, "disable")) {
2281 set_schedstats(false);
2286 pr_warn("Unable to parse schedstats=\n");
2290 __setup("schedstats=", setup_schedstats);
2292 #ifdef CONFIG_PROC_SYSCTL
2293 int sysctl_schedstats(struct ctl_table *table, int write,
2294 void __user *buffer, size_t *lenp, loff_t *ppos)
2298 int state = static_branch_likely(&sched_schedstats);
2300 if (write && !capable(CAP_SYS_ADMIN))
2305 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2309 set_schedstats(state);
2316 * fork()/clone()-time setup:
2318 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2320 unsigned long flags;
2321 int cpu = get_cpu();
2323 __sched_fork(clone_flags, p);
2325 * We mark the process as running here. This guarantees that
2326 * nobody will actually run it, and a signal or other external
2327 * event cannot wake it up and insert it on the runqueue either.
2329 p->state = TASK_RUNNING;
2332 * Make sure we do not leak PI boosting priority to the child.
2334 p->prio = current->normal_prio;
2337 * Revert to default priority/policy on fork if requested.
2339 if (unlikely(p->sched_reset_on_fork)) {
2340 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2341 p->policy = SCHED_NORMAL;
2342 p->static_prio = NICE_TO_PRIO(0);
2344 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2345 p->static_prio = NICE_TO_PRIO(0);
2347 p->prio = p->normal_prio = __normal_prio(p);
2351 * We don't need the reset flag anymore after the fork. It has
2352 * fulfilled its duty:
2354 p->sched_reset_on_fork = 0;
2357 if (dl_prio(p->prio)) {
2360 } else if (rt_prio(p->prio)) {
2361 p->sched_class = &rt_sched_class;
2363 p->sched_class = &fair_sched_class;
2366 if (p->sched_class->task_fork)
2367 p->sched_class->task_fork(p);
2370 * The child is not yet in the pid-hash so no cgroup attach races,
2371 * and the cgroup is pinned to this child due to cgroup_fork()
2372 * is ran before sched_fork().
2374 * Silence PROVE_RCU.
2376 raw_spin_lock_irqsave(&p->pi_lock, flags);
2377 set_task_cpu(p, cpu);
2378 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2380 #ifdef CONFIG_SCHED_INFO
2381 if (likely(sched_info_on()))
2382 memset(&p->sched_info, 0, sizeof(p->sched_info));
2384 #if defined(CONFIG_SMP)
2387 init_task_preempt_count(p);
2389 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2390 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2397 unsigned long to_ratio(u64 period, u64 runtime)
2399 if (runtime == RUNTIME_INF)
2403 * Doing this here saves a lot of checks in all
2404 * the calling paths, and returning zero seems
2405 * safe for them anyway.
2410 return div64_u64(runtime << 20, period);
2414 inline struct dl_bw *dl_bw_of(int i)
2416 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2417 "sched RCU must be held");
2418 return &cpu_rq(i)->rd->dl_bw;
2421 static inline int dl_bw_cpus(int i)
2423 struct root_domain *rd = cpu_rq(i)->rd;
2426 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2427 "sched RCU must be held");
2428 for_each_cpu_and(i, rd->span, cpu_active_mask)
2434 inline struct dl_bw *dl_bw_of(int i)
2436 return &cpu_rq(i)->dl.dl_bw;
2439 static inline int dl_bw_cpus(int i)
2446 * We must be sure that accepting a new task (or allowing changing the
2447 * parameters of an existing one) is consistent with the bandwidth
2448 * constraints. If yes, this function also accordingly updates the currently
2449 * allocated bandwidth to reflect the new situation.
2451 * This function is called while holding p's rq->lock.
2453 * XXX we should delay bw change until the task's 0-lag point, see
2456 static int dl_overflow(struct task_struct *p, int policy,
2457 const struct sched_attr *attr)
2460 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2461 u64 period = attr->sched_period ?: attr->sched_deadline;
2462 u64 runtime = attr->sched_runtime;
2463 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2466 /* !deadline task may carry old deadline bandwidth */
2467 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2471 * Either if a task, enters, leave, or stays -deadline but changes
2472 * its parameters, we may need to update accordingly the total
2473 * allocated bandwidth of the container.
2475 raw_spin_lock(&dl_b->lock);
2476 cpus = dl_bw_cpus(task_cpu(p));
2477 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2478 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2479 __dl_add(dl_b, new_bw);
2481 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2482 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2483 __dl_clear(dl_b, p->dl.dl_bw);
2484 __dl_add(dl_b, new_bw);
2486 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2487 __dl_clear(dl_b, p->dl.dl_bw);
2490 raw_spin_unlock(&dl_b->lock);
2495 extern void init_dl_bw(struct dl_bw *dl_b);
2498 * wake_up_new_task - wake up a newly created task for the first time.
2500 * This function will do some initial scheduler statistics housekeeping
2501 * that must be done for every newly created context, then puts the task
2502 * on the runqueue and wakes it.
2504 void wake_up_new_task(struct task_struct *p)
2509 /* Initialize new task's runnable average */
2510 init_entity_runnable_average(&p->se);
2511 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2514 * Fork balancing, do it here and not earlier because:
2515 * - cpus_allowed can change in the fork path
2516 * - any previously selected cpu might disappear through hotplug
2518 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2520 /* Post initialize new task's util average when its cfs_rq is set */
2521 post_init_entity_util_avg(&p->se);
2523 rq = __task_rq_lock(p, &rf);
2524 activate_task(rq, p, 0);
2525 p->on_rq = TASK_ON_RQ_QUEUED;
2526 trace_sched_wakeup_new(p);
2527 check_preempt_curr(rq, p, WF_FORK);
2529 if (p->sched_class->task_woken) {
2531 * Nothing relies on rq->lock after this, so its fine to
2534 lockdep_unpin_lock(&rq->lock, rf.cookie);
2535 p->sched_class->task_woken(rq, p);
2536 lockdep_repin_lock(&rq->lock, rf.cookie);
2539 task_rq_unlock(rq, p, &rf);
2542 #ifdef CONFIG_PREEMPT_NOTIFIERS
2544 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2546 void preempt_notifier_inc(void)
2548 static_key_slow_inc(&preempt_notifier_key);
2550 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2552 void preempt_notifier_dec(void)
2554 static_key_slow_dec(&preempt_notifier_key);
2556 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2559 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2560 * @notifier: notifier struct to register
2562 void preempt_notifier_register(struct preempt_notifier *notifier)
2564 if (!static_key_false(&preempt_notifier_key))
2565 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2567 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2569 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2572 * preempt_notifier_unregister - no longer interested in preemption notifications
2573 * @notifier: notifier struct to unregister
2575 * This is *not* safe to call from within a preemption notifier.
2577 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2579 hlist_del(¬ifier->link);
2581 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2583 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2585 struct preempt_notifier *notifier;
2587 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2588 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2591 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2593 if (static_key_false(&preempt_notifier_key))
2594 __fire_sched_in_preempt_notifiers(curr);
2598 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2599 struct task_struct *next)
2601 struct preempt_notifier *notifier;
2603 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2604 notifier->ops->sched_out(notifier, next);
2607 static __always_inline void
2608 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2609 struct task_struct *next)
2611 if (static_key_false(&preempt_notifier_key))
2612 __fire_sched_out_preempt_notifiers(curr, next);
2615 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2617 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2622 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2623 struct task_struct *next)
2627 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2630 * prepare_task_switch - prepare to switch tasks
2631 * @rq: the runqueue preparing to switch
2632 * @prev: the current task that is being switched out
2633 * @next: the task we are going to switch to.
2635 * This is called with the rq lock held and interrupts off. It must
2636 * be paired with a subsequent finish_task_switch after the context
2639 * prepare_task_switch sets up locking and calls architecture specific
2643 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2644 struct task_struct *next)
2646 sched_info_switch(rq, prev, next);
2647 perf_event_task_sched_out(prev, next);
2648 fire_sched_out_preempt_notifiers(prev, next);
2649 prepare_lock_switch(rq, next);
2650 prepare_arch_switch(next);
2654 * finish_task_switch - clean up after a task-switch
2655 * @prev: the thread we just switched away from.
2657 * finish_task_switch must be called after the context switch, paired
2658 * with a prepare_task_switch call before the context switch.
2659 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2660 * and do any other architecture-specific cleanup actions.
2662 * Note that we may have delayed dropping an mm in context_switch(). If
2663 * so, we finish that here outside of the runqueue lock. (Doing it
2664 * with the lock held can cause deadlocks; see schedule() for
2667 * The context switch have flipped the stack from under us and restored the
2668 * local variables which were saved when this task called schedule() in the
2669 * past. prev == current is still correct but we need to recalculate this_rq
2670 * because prev may have moved to another CPU.
2672 static struct rq *finish_task_switch(struct task_struct *prev)
2673 __releases(rq->lock)
2675 struct rq *rq = this_rq();
2676 struct mm_struct *mm = rq->prev_mm;
2680 * The previous task will have left us with a preempt_count of 2
2681 * because it left us after:
2684 * preempt_disable(); // 1
2686 * raw_spin_lock_irq(&rq->lock) // 2
2688 * Also, see FORK_PREEMPT_COUNT.
2690 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2691 "corrupted preempt_count: %s/%d/0x%x\n",
2692 current->comm, current->pid, preempt_count()))
2693 preempt_count_set(FORK_PREEMPT_COUNT);
2698 * A task struct has one reference for the use as "current".
2699 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2700 * schedule one last time. The schedule call will never return, and
2701 * the scheduled task must drop that reference.
2703 * We must observe prev->state before clearing prev->on_cpu (in
2704 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2705 * running on another CPU and we could rave with its RUNNING -> DEAD
2706 * transition, resulting in a double drop.
2708 prev_state = prev->state;
2709 vtime_task_switch(prev);
2710 perf_event_task_sched_in(prev, current);
2711 finish_lock_switch(rq, prev);
2712 finish_arch_post_lock_switch();
2714 fire_sched_in_preempt_notifiers(current);
2717 if (unlikely(prev_state == TASK_DEAD)) {
2718 if (prev->sched_class->task_dead)
2719 prev->sched_class->task_dead(prev);
2722 * Remove function-return probe instances associated with this
2723 * task and put them back on the free list.
2725 kprobe_flush_task(prev);
2726 put_task_struct(prev);
2729 tick_nohz_task_switch();
2735 /* rq->lock is NOT held, but preemption is disabled */
2736 static void __balance_callback(struct rq *rq)
2738 struct callback_head *head, *next;
2739 void (*func)(struct rq *rq);
2740 unsigned long flags;
2742 raw_spin_lock_irqsave(&rq->lock, flags);
2743 head = rq->balance_callback;
2744 rq->balance_callback = NULL;
2746 func = (void (*)(struct rq *))head->func;
2753 raw_spin_unlock_irqrestore(&rq->lock, flags);
2756 static inline void balance_callback(struct rq *rq)
2758 if (unlikely(rq->balance_callback))
2759 __balance_callback(rq);
2764 static inline void balance_callback(struct rq *rq)
2771 * schedule_tail - first thing a freshly forked thread must call.
2772 * @prev: the thread we just switched away from.
2774 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2775 __releases(rq->lock)
2780 * New tasks start with FORK_PREEMPT_COUNT, see there and
2781 * finish_task_switch() for details.
2783 * finish_task_switch() will drop rq->lock() and lower preempt_count
2784 * and the preempt_enable() will end up enabling preemption (on
2785 * PREEMPT_COUNT kernels).
2788 rq = finish_task_switch(prev);
2789 balance_callback(rq);
2792 if (current->set_child_tid)
2793 put_user(task_pid_vnr(current), current->set_child_tid);
2797 * context_switch - switch to the new MM and the new thread's register state.
2799 static __always_inline struct rq *
2800 context_switch(struct rq *rq, struct task_struct *prev,
2801 struct task_struct *next, struct pin_cookie cookie)
2803 struct mm_struct *mm, *oldmm;
2805 prepare_task_switch(rq, prev, next);
2808 oldmm = prev->active_mm;
2810 * For paravirt, this is coupled with an exit in switch_to to
2811 * combine the page table reload and the switch backend into
2814 arch_start_context_switch(prev);
2817 next->active_mm = oldmm;
2818 atomic_inc(&oldmm->mm_count);
2819 enter_lazy_tlb(oldmm, next);
2821 switch_mm_irqs_off(oldmm, mm, next);
2824 prev->active_mm = NULL;
2825 rq->prev_mm = oldmm;
2828 * Since the runqueue lock will be released by the next
2829 * task (which is an invalid locking op but in the case
2830 * of the scheduler it's an obvious special-case), so we
2831 * do an early lockdep release here:
2833 lockdep_unpin_lock(&rq->lock, cookie);
2834 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2836 /* Here we just switch the register state and the stack. */
2837 switch_to(prev, next, prev);
2840 return finish_task_switch(prev);
2844 * nr_running and nr_context_switches:
2846 * externally visible scheduler statistics: current number of runnable
2847 * threads, total number of context switches performed since bootup.
2849 unsigned long nr_running(void)
2851 unsigned long i, sum = 0;
2853 for_each_online_cpu(i)
2854 sum += cpu_rq(i)->nr_running;
2860 * Check if only the current task is running on the cpu.
2862 * Caution: this function does not check that the caller has disabled
2863 * preemption, thus the result might have a time-of-check-to-time-of-use
2864 * race. The caller is responsible to use it correctly, for example:
2866 * - from a non-preemptable section (of course)
2868 * - from a thread that is bound to a single CPU
2870 * - in a loop with very short iterations (e.g. a polling loop)
2872 bool single_task_running(void)
2874 return raw_rq()->nr_running == 1;
2876 EXPORT_SYMBOL(single_task_running);
2878 unsigned long long nr_context_switches(void)
2881 unsigned long long sum = 0;
2883 for_each_possible_cpu(i)
2884 sum += cpu_rq(i)->nr_switches;
2889 unsigned long nr_iowait(void)
2891 unsigned long i, sum = 0;
2893 for_each_possible_cpu(i)
2894 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2899 unsigned long nr_iowait_cpu(int cpu)
2901 struct rq *this = cpu_rq(cpu);
2902 return atomic_read(&this->nr_iowait);
2905 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2907 struct rq *rq = this_rq();
2908 *nr_waiters = atomic_read(&rq->nr_iowait);
2909 *load = rq->load.weight;
2915 * sched_exec - execve() is a valuable balancing opportunity, because at
2916 * this point the task has the smallest effective memory and cache footprint.
2918 void sched_exec(void)
2920 struct task_struct *p = current;
2921 unsigned long flags;
2924 raw_spin_lock_irqsave(&p->pi_lock, flags);
2925 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2926 if (dest_cpu == smp_processor_id())
2929 if (likely(cpu_active(dest_cpu))) {
2930 struct migration_arg arg = { p, dest_cpu };
2932 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2933 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2937 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2942 DEFINE_PER_CPU(struct kernel_stat, kstat);
2943 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2945 EXPORT_PER_CPU_SYMBOL(kstat);
2946 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2949 * Return accounted runtime for the task.
2950 * In case the task is currently running, return the runtime plus current's
2951 * pending runtime that have not been accounted yet.
2953 unsigned long long task_sched_runtime(struct task_struct *p)
2959 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2961 * 64-bit doesn't need locks to atomically read a 64bit value.
2962 * So we have a optimization chance when the task's delta_exec is 0.
2963 * Reading ->on_cpu is racy, but this is ok.
2965 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2966 * If we race with it entering cpu, unaccounted time is 0. This is
2967 * indistinguishable from the read occurring a few cycles earlier.
2968 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2969 * been accounted, so we're correct here as well.
2971 if (!p->on_cpu || !task_on_rq_queued(p))
2972 return p->se.sum_exec_runtime;
2975 rq = task_rq_lock(p, &rf);
2977 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2978 * project cycles that may never be accounted to this
2979 * thread, breaking clock_gettime().
2981 if (task_current(rq, p) && task_on_rq_queued(p)) {
2982 update_rq_clock(rq);
2983 p->sched_class->update_curr(rq);
2985 ns = p->se.sum_exec_runtime;
2986 task_rq_unlock(rq, p, &rf);
2992 * This function gets called by the timer code, with HZ frequency.
2993 * We call it with interrupts disabled.
2995 void scheduler_tick(void)
2997 int cpu = smp_processor_id();
2998 struct rq *rq = cpu_rq(cpu);
2999 struct task_struct *curr = rq->curr;
3003 raw_spin_lock(&rq->lock);
3004 update_rq_clock(rq);
3005 curr->sched_class->task_tick(rq, curr, 0);
3006 cpu_load_update_active(rq);
3007 calc_global_load_tick(rq);
3008 raw_spin_unlock(&rq->lock);
3010 perf_event_task_tick();
3013 rq->idle_balance = idle_cpu(cpu);
3014 trigger_load_balance(rq);
3016 rq_last_tick_reset(rq);
3019 #ifdef CONFIG_NO_HZ_FULL
3021 * scheduler_tick_max_deferment
3023 * Keep at least one tick per second when a single
3024 * active task is running because the scheduler doesn't
3025 * yet completely support full dynticks environment.
3027 * This makes sure that uptime, CFS vruntime, load
3028 * balancing, etc... continue to move forward, even
3029 * with a very low granularity.
3031 * Return: Maximum deferment in nanoseconds.
3033 u64 scheduler_tick_max_deferment(void)
3035 struct rq *rq = this_rq();
3036 unsigned long next, now = READ_ONCE(jiffies);
3038 next = rq->last_sched_tick + HZ;
3040 if (time_before_eq(next, now))
3043 return jiffies_to_nsecs(next - now);
3047 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3048 defined(CONFIG_PREEMPT_TRACER))
3050 * If the value passed in is equal to the current preempt count
3051 * then we just disabled preemption. Start timing the latency.
3053 static inline void preempt_latency_start(int val)
3055 if (preempt_count() == val) {
3056 unsigned long ip = get_lock_parent_ip();
3057 #ifdef CONFIG_DEBUG_PREEMPT
3058 current->preempt_disable_ip = ip;
3060 trace_preempt_off(CALLER_ADDR0, ip);
3064 void preempt_count_add(int val)
3066 #ifdef CONFIG_DEBUG_PREEMPT
3070 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3073 __preempt_count_add(val);
3074 #ifdef CONFIG_DEBUG_PREEMPT
3076 * Spinlock count overflowing soon?
3078 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3081 preempt_latency_start(val);
3083 EXPORT_SYMBOL(preempt_count_add);
3084 NOKPROBE_SYMBOL(preempt_count_add);
3087 * If the value passed in equals to the current preempt count
3088 * then we just enabled preemption. Stop timing the latency.
3090 static inline void preempt_latency_stop(int val)
3092 if (preempt_count() == val)
3093 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3096 void preempt_count_sub(int val)
3098 #ifdef CONFIG_DEBUG_PREEMPT
3102 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3105 * Is the spinlock portion underflowing?
3107 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3108 !(preempt_count() & PREEMPT_MASK)))
3112 preempt_latency_stop(val);
3113 __preempt_count_sub(val);
3115 EXPORT_SYMBOL(preempt_count_sub);
3116 NOKPROBE_SYMBOL(preempt_count_sub);
3119 static inline void preempt_latency_start(int val) { }
3120 static inline void preempt_latency_stop(int val) { }
3124 * Print scheduling while atomic bug:
3126 static noinline void __schedule_bug(struct task_struct *prev)
3128 if (oops_in_progress)
3131 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3132 prev->comm, prev->pid, preempt_count());
3134 debug_show_held_locks(prev);
3136 if (irqs_disabled())
3137 print_irqtrace_events(prev);
3138 #ifdef CONFIG_DEBUG_PREEMPT
3139 if (in_atomic_preempt_off()) {
3140 pr_err("Preemption disabled at:");
3141 print_ip_sym(current->preempt_disable_ip);
3146 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3150 * Various schedule()-time debugging checks and statistics:
3152 static inline void schedule_debug(struct task_struct *prev)
3154 #ifdef CONFIG_SCHED_STACK_END_CHECK
3155 BUG_ON(task_stack_end_corrupted(prev));
3158 if (unlikely(in_atomic_preempt_off())) {
3159 __schedule_bug(prev);
3160 preempt_count_set(PREEMPT_DISABLED);
3164 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3166 schedstat_inc(this_rq(), sched_count);
3170 * Pick up the highest-prio task:
3172 static inline struct task_struct *
3173 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3175 const struct sched_class *class = &fair_sched_class;
3176 struct task_struct *p;
3179 * Optimization: we know that if all tasks are in
3180 * the fair class we can call that function directly:
3182 if (likely(prev->sched_class == class &&
3183 rq->nr_running == rq->cfs.h_nr_running)) {
3184 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3185 if (unlikely(p == RETRY_TASK))
3188 /* assumes fair_sched_class->next == idle_sched_class */
3190 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3196 for_each_class(class) {
3197 p = class->pick_next_task(rq, prev, cookie);
3199 if (unlikely(p == RETRY_TASK))
3205 BUG(); /* the idle class will always have a runnable task */
3209 * __schedule() is the main scheduler function.
3211 * The main means of driving the scheduler and thus entering this function are:
3213 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3215 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3216 * paths. For example, see arch/x86/entry_64.S.
3218 * To drive preemption between tasks, the scheduler sets the flag in timer
3219 * interrupt handler scheduler_tick().
3221 * 3. Wakeups don't really cause entry into schedule(). They add a
3222 * task to the run-queue and that's it.
3224 * Now, if the new task added to the run-queue preempts the current
3225 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3226 * called on the nearest possible occasion:
3228 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3230 * - in syscall or exception context, at the next outmost
3231 * preempt_enable(). (this might be as soon as the wake_up()'s
3234 * - in IRQ context, return from interrupt-handler to
3235 * preemptible context
3237 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3240 * - cond_resched() call
3241 * - explicit schedule() call
3242 * - return from syscall or exception to user-space
3243 * - return from interrupt-handler to user-space
3245 * WARNING: must be called with preemption disabled!
3247 static void __sched notrace __schedule(bool preempt)
3249 struct task_struct *prev, *next;
3250 unsigned long *switch_count;
3251 struct pin_cookie cookie;
3255 cpu = smp_processor_id();
3260 * do_exit() calls schedule() with preemption disabled as an exception;
3261 * however we must fix that up, otherwise the next task will see an
3262 * inconsistent (higher) preempt count.
3264 * It also avoids the below schedule_debug() test from complaining
3267 if (unlikely(prev->state == TASK_DEAD))
3268 preempt_enable_no_resched_notrace();
3270 schedule_debug(prev);
3272 if (sched_feat(HRTICK))
3275 local_irq_disable();
3276 rcu_note_context_switch();
3279 * Make sure that signal_pending_state()->signal_pending() below
3280 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3281 * done by the caller to avoid the race with signal_wake_up().
3283 smp_mb__before_spinlock();
3284 raw_spin_lock(&rq->lock);
3285 cookie = lockdep_pin_lock(&rq->lock);
3287 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3289 switch_count = &prev->nivcsw;
3290 if (!preempt && prev->state) {
3291 if (unlikely(signal_pending_state(prev->state, prev))) {
3292 prev->state = TASK_RUNNING;
3294 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3298 * If a worker went to sleep, notify and ask workqueue
3299 * whether it wants to wake up a task to maintain
3302 if (prev->flags & PF_WQ_WORKER) {
3303 struct task_struct *to_wakeup;
3305 to_wakeup = wq_worker_sleeping(prev);
3307 try_to_wake_up_local(to_wakeup, cookie);
3310 switch_count = &prev->nvcsw;
3313 if (task_on_rq_queued(prev))
3314 update_rq_clock(rq);
3316 next = pick_next_task(rq, prev, cookie);
3317 clear_tsk_need_resched(prev);
3318 clear_preempt_need_resched();
3319 rq->clock_skip_update = 0;
3321 if (likely(prev != next)) {
3326 trace_sched_switch(preempt, prev, next);
3327 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3329 lockdep_unpin_lock(&rq->lock, cookie);
3330 raw_spin_unlock_irq(&rq->lock);
3333 balance_callback(rq);
3335 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3337 static inline void sched_submit_work(struct task_struct *tsk)
3339 if (!tsk->state || tsk_is_pi_blocked(tsk))
3342 * If we are going to sleep and we have plugged IO queued,
3343 * make sure to submit it to avoid deadlocks.
3345 if (blk_needs_flush_plug(tsk))
3346 blk_schedule_flush_plug(tsk);
3349 asmlinkage __visible void __sched schedule(void)
3351 struct task_struct *tsk = current;
3353 sched_submit_work(tsk);
3357 sched_preempt_enable_no_resched();
3358 } while (need_resched());
3360 EXPORT_SYMBOL(schedule);
3362 #ifdef CONFIG_CONTEXT_TRACKING
3363 asmlinkage __visible void __sched schedule_user(void)
3366 * If we come here after a random call to set_need_resched(),
3367 * or we have been woken up remotely but the IPI has not yet arrived,
3368 * we haven't yet exited the RCU idle mode. Do it here manually until
3369 * we find a better solution.
3371 * NB: There are buggy callers of this function. Ideally we
3372 * should warn if prev_state != CONTEXT_USER, but that will trigger
3373 * too frequently to make sense yet.
3375 enum ctx_state prev_state = exception_enter();
3377 exception_exit(prev_state);
3382 * schedule_preempt_disabled - called with preemption disabled
3384 * Returns with preemption disabled. Note: preempt_count must be 1
3386 void __sched schedule_preempt_disabled(void)
3388 sched_preempt_enable_no_resched();
3393 static void __sched notrace preempt_schedule_common(void)
3397 * Because the function tracer can trace preempt_count_sub()
3398 * and it also uses preempt_enable/disable_notrace(), if
3399 * NEED_RESCHED is set, the preempt_enable_notrace() called
3400 * by the function tracer will call this function again and
3401 * cause infinite recursion.
3403 * Preemption must be disabled here before the function
3404 * tracer can trace. Break up preempt_disable() into two
3405 * calls. One to disable preemption without fear of being
3406 * traced. The other to still record the preemption latency,
3407 * which can also be traced by the function tracer.
3409 preempt_disable_notrace();
3410 preempt_latency_start(1);
3412 preempt_latency_stop(1);
3413 preempt_enable_no_resched_notrace();
3416 * Check again in case we missed a preemption opportunity
3417 * between schedule and now.
3419 } while (need_resched());
3422 #ifdef CONFIG_PREEMPT
3424 * this is the entry point to schedule() from in-kernel preemption
3425 * off of preempt_enable. Kernel preemptions off return from interrupt
3426 * occur there and call schedule directly.
3428 asmlinkage __visible void __sched notrace preempt_schedule(void)
3431 * If there is a non-zero preempt_count or interrupts are disabled,
3432 * we do not want to preempt the current task. Just return..
3434 if (likely(!preemptible()))
3437 preempt_schedule_common();
3439 NOKPROBE_SYMBOL(preempt_schedule);
3440 EXPORT_SYMBOL(preempt_schedule);
3443 * preempt_schedule_notrace - preempt_schedule called by tracing
3445 * The tracing infrastructure uses preempt_enable_notrace to prevent
3446 * recursion and tracing preempt enabling caused by the tracing
3447 * infrastructure itself. But as tracing can happen in areas coming
3448 * from userspace or just about to enter userspace, a preempt enable
3449 * can occur before user_exit() is called. This will cause the scheduler
3450 * to be called when the system is still in usermode.
3452 * To prevent this, the preempt_enable_notrace will use this function
3453 * instead of preempt_schedule() to exit user context if needed before
3454 * calling the scheduler.
3456 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3458 enum ctx_state prev_ctx;
3460 if (likely(!preemptible()))
3465 * Because the function tracer can trace preempt_count_sub()
3466 * and it also uses preempt_enable/disable_notrace(), if
3467 * NEED_RESCHED is set, the preempt_enable_notrace() called
3468 * by the function tracer will call this function again and
3469 * cause infinite recursion.
3471 * Preemption must be disabled here before the function
3472 * tracer can trace. Break up preempt_disable() into two
3473 * calls. One to disable preemption without fear of being
3474 * traced. The other to still record the preemption latency,
3475 * which can also be traced by the function tracer.
3477 preempt_disable_notrace();
3478 preempt_latency_start(1);
3480 * Needs preempt disabled in case user_exit() is traced
3481 * and the tracer calls preempt_enable_notrace() causing
3482 * an infinite recursion.
3484 prev_ctx = exception_enter();
3486 exception_exit(prev_ctx);
3488 preempt_latency_stop(1);
3489 preempt_enable_no_resched_notrace();
3490 } while (need_resched());
3492 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3494 #endif /* CONFIG_PREEMPT */
3497 * this is the entry point to schedule() from kernel preemption
3498 * off of irq context.
3499 * Note, that this is called and return with irqs disabled. This will
3500 * protect us against recursive calling from irq.
3502 asmlinkage __visible void __sched preempt_schedule_irq(void)
3504 enum ctx_state prev_state;
3506 /* Catch callers which need to be fixed */
3507 BUG_ON(preempt_count() || !irqs_disabled());
3509 prev_state = exception_enter();
3515 local_irq_disable();
3516 sched_preempt_enable_no_resched();
3517 } while (need_resched());
3519 exception_exit(prev_state);
3522 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3525 return try_to_wake_up(curr->private, mode, wake_flags);
3527 EXPORT_SYMBOL(default_wake_function);
3529 #ifdef CONFIG_RT_MUTEXES
3532 * rt_mutex_setprio - set the current priority of a task
3534 * @prio: prio value (kernel-internal form)
3536 * This function changes the 'effective' priority of a task. It does
3537 * not touch ->normal_prio like __setscheduler().
3539 * Used by the rt_mutex code to implement priority inheritance
3540 * logic. Call site only calls if the priority of the task changed.
3542 void rt_mutex_setprio(struct task_struct *p, int prio)
3544 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3545 const struct sched_class *prev_class;
3549 BUG_ON(prio > MAX_PRIO);
3551 rq = __task_rq_lock(p, &rf);
3554 * Idle task boosting is a nono in general. There is one
3555 * exception, when PREEMPT_RT and NOHZ is active:
3557 * The idle task calls get_next_timer_interrupt() and holds
3558 * the timer wheel base->lock on the CPU and another CPU wants
3559 * to access the timer (probably to cancel it). We can safely
3560 * ignore the boosting request, as the idle CPU runs this code
3561 * with interrupts disabled and will complete the lock
3562 * protected section without being interrupted. So there is no
3563 * real need to boost.
3565 if (unlikely(p == rq->idle)) {
3566 WARN_ON(p != rq->curr);
3567 WARN_ON(p->pi_blocked_on);
3571 trace_sched_pi_setprio(p, prio);
3574 if (oldprio == prio)
3575 queue_flag &= ~DEQUEUE_MOVE;
3577 prev_class = p->sched_class;
3578 queued = task_on_rq_queued(p);
3579 running = task_current(rq, p);
3581 dequeue_task(rq, p, queue_flag);
3583 put_prev_task(rq, p);
3586 * Boosting condition are:
3587 * 1. -rt task is running and holds mutex A
3588 * --> -dl task blocks on mutex A
3590 * 2. -dl task is running and holds mutex A
3591 * --> -dl task blocks on mutex A and could preempt the
3594 if (dl_prio(prio)) {
3595 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3596 if (!dl_prio(p->normal_prio) ||
3597 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3598 p->dl.dl_boosted = 1;
3599 queue_flag |= ENQUEUE_REPLENISH;
3601 p->dl.dl_boosted = 0;
3602 p->sched_class = &dl_sched_class;
3603 } else if (rt_prio(prio)) {
3604 if (dl_prio(oldprio))
3605 p->dl.dl_boosted = 0;
3607 queue_flag |= ENQUEUE_HEAD;
3608 p->sched_class = &rt_sched_class;
3610 if (dl_prio(oldprio))
3611 p->dl.dl_boosted = 0;
3612 if (rt_prio(oldprio))
3614 p->sched_class = &fair_sched_class;
3620 p->sched_class->set_curr_task(rq);
3622 enqueue_task(rq, p, queue_flag);
3624 check_class_changed(rq, p, prev_class, oldprio);
3626 preempt_disable(); /* avoid rq from going away on us */
3627 __task_rq_unlock(rq, &rf);
3629 balance_callback(rq);
3634 void set_user_nice(struct task_struct *p, long nice)
3636 int old_prio, delta, queued;
3640 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3643 * We have to be careful, if called from sys_setpriority(),
3644 * the task might be in the middle of scheduling on another CPU.
3646 rq = task_rq_lock(p, &rf);
3648 * The RT priorities are set via sched_setscheduler(), but we still
3649 * allow the 'normal' nice value to be set - but as expected
3650 * it wont have any effect on scheduling until the task is
3651 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3653 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3654 p->static_prio = NICE_TO_PRIO(nice);
3657 queued = task_on_rq_queued(p);
3659 dequeue_task(rq, p, DEQUEUE_SAVE);
3661 p->static_prio = NICE_TO_PRIO(nice);
3664 p->prio = effective_prio(p);
3665 delta = p->prio - old_prio;
3668 enqueue_task(rq, p, ENQUEUE_RESTORE);
3670 * If the task increased its priority or is running and
3671 * lowered its priority, then reschedule its CPU:
3673 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3677 task_rq_unlock(rq, p, &rf);
3679 EXPORT_SYMBOL(set_user_nice);
3682 * can_nice - check if a task can reduce its nice value
3686 int can_nice(const struct task_struct *p, const int nice)
3688 /* convert nice value [19,-20] to rlimit style value [1,40] */
3689 int nice_rlim = nice_to_rlimit(nice);
3691 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3692 capable(CAP_SYS_NICE));
3695 #ifdef __ARCH_WANT_SYS_NICE
3698 * sys_nice - change the priority of the current process.
3699 * @increment: priority increment
3701 * sys_setpriority is a more generic, but much slower function that
3702 * does similar things.
3704 SYSCALL_DEFINE1(nice, int, increment)
3709 * Setpriority might change our priority at the same moment.
3710 * We don't have to worry. Conceptually one call occurs first
3711 * and we have a single winner.
3713 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3714 nice = task_nice(current) + increment;
3716 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3717 if (increment < 0 && !can_nice(current, nice))
3720 retval = security_task_setnice(current, nice);
3724 set_user_nice(current, nice);
3731 * task_prio - return the priority value of a given task.
3732 * @p: the task in question.
3734 * Return: The priority value as seen by users in /proc.
3735 * RT tasks are offset by -200. Normal tasks are centered
3736 * around 0, value goes from -16 to +15.
3738 int task_prio(const struct task_struct *p)
3740 return p->prio - MAX_RT_PRIO;
3744 * idle_cpu - is a given cpu idle currently?
3745 * @cpu: the processor in question.
3747 * Return: 1 if the CPU is currently idle. 0 otherwise.
3749 int idle_cpu(int cpu)
3751 struct rq *rq = cpu_rq(cpu);
3753 if (rq->curr != rq->idle)
3760 if (!llist_empty(&rq->wake_list))
3768 * idle_task - return the idle task for a given cpu.
3769 * @cpu: the processor in question.
3771 * Return: The idle task for the cpu @cpu.
3773 struct task_struct *idle_task(int cpu)
3775 return cpu_rq(cpu)->idle;
3779 * find_process_by_pid - find a process with a matching PID value.
3780 * @pid: the pid in question.
3782 * The task of @pid, if found. %NULL otherwise.
3784 static struct task_struct *find_process_by_pid(pid_t pid)
3786 return pid ? find_task_by_vpid(pid) : current;
3790 * This function initializes the sched_dl_entity of a newly becoming
3791 * SCHED_DEADLINE task.
3793 * Only the static values are considered here, the actual runtime and the
3794 * absolute deadline will be properly calculated when the task is enqueued
3795 * for the first time with its new policy.
3798 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3800 struct sched_dl_entity *dl_se = &p->dl;
3802 dl_se->dl_runtime = attr->sched_runtime;
3803 dl_se->dl_deadline = attr->sched_deadline;
3804 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3805 dl_se->flags = attr->sched_flags;
3806 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3809 * Changing the parameters of a task is 'tricky' and we're not doing
3810 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3812 * What we SHOULD do is delay the bandwidth release until the 0-lag
3813 * point. This would include retaining the task_struct until that time
3814 * and change dl_overflow() to not immediately decrement the current
3817 * Instead we retain the current runtime/deadline and let the new
3818 * parameters take effect after the current reservation period lapses.
3819 * This is safe (albeit pessimistic) because the 0-lag point is always
3820 * before the current scheduling deadline.
3822 * We can still have temporary overloads because we do not delay the
3823 * change in bandwidth until that time; so admission control is
3824 * not on the safe side. It does however guarantee tasks will never
3825 * consume more than promised.
3830 * sched_setparam() passes in -1 for its policy, to let the functions
3831 * it calls know not to change it.
3833 #define SETPARAM_POLICY -1
3835 static void __setscheduler_params(struct task_struct *p,
3836 const struct sched_attr *attr)
3838 int policy = attr->sched_policy;
3840 if (policy == SETPARAM_POLICY)
3845 if (dl_policy(policy))
3846 __setparam_dl(p, attr);
3847 else if (fair_policy(policy))
3848 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3851 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3852 * !rt_policy. Always setting this ensures that things like
3853 * getparam()/getattr() don't report silly values for !rt tasks.
3855 p->rt_priority = attr->sched_priority;
3856 p->normal_prio = normal_prio(p);
3860 /* Actually do priority change: must hold pi & rq lock. */
3861 static void __setscheduler(struct rq *rq, struct task_struct *p,
3862 const struct sched_attr *attr, bool keep_boost)
3864 __setscheduler_params(p, attr);
3867 * Keep a potential priority boosting if called from
3868 * sched_setscheduler().
3871 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3873 p->prio = normal_prio(p);
3875 if (dl_prio(p->prio))
3876 p->sched_class = &dl_sched_class;
3877 else if (rt_prio(p->prio))
3878 p->sched_class = &rt_sched_class;
3880 p->sched_class = &fair_sched_class;
3884 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3886 struct sched_dl_entity *dl_se = &p->dl;
3888 attr->sched_priority = p->rt_priority;
3889 attr->sched_runtime = dl_se->dl_runtime;
3890 attr->sched_deadline = dl_se->dl_deadline;
3891 attr->sched_period = dl_se->dl_period;
3892 attr->sched_flags = dl_se->flags;
3896 * This function validates the new parameters of a -deadline task.
3897 * We ask for the deadline not being zero, and greater or equal
3898 * than the runtime, as well as the period of being zero or
3899 * greater than deadline. Furthermore, we have to be sure that
3900 * user parameters are above the internal resolution of 1us (we
3901 * check sched_runtime only since it is always the smaller one) and
3902 * below 2^63 ns (we have to check both sched_deadline and
3903 * sched_period, as the latter can be zero).
3906 __checkparam_dl(const struct sched_attr *attr)
3909 if (attr->sched_deadline == 0)
3913 * Since we truncate DL_SCALE bits, make sure we're at least
3916 if (attr->sched_runtime < (1ULL << DL_SCALE))
3920 * Since we use the MSB for wrap-around and sign issues, make
3921 * sure it's not set (mind that period can be equal to zero).
3923 if (attr->sched_deadline & (1ULL << 63) ||
3924 attr->sched_period & (1ULL << 63))
3927 /* runtime <= deadline <= period (if period != 0) */
3928 if ((attr->sched_period != 0 &&
3929 attr->sched_period < attr->sched_deadline) ||
3930 attr->sched_deadline < attr->sched_runtime)
3937 * check the target process has a UID that matches the current process's
3939 static bool check_same_owner(struct task_struct *p)
3941 const struct cred *cred = current_cred(), *pcred;
3945 pcred = __task_cred(p);
3946 match = (uid_eq(cred->euid, pcred->euid) ||
3947 uid_eq(cred->euid, pcred->uid));
3952 static bool dl_param_changed(struct task_struct *p,
3953 const struct sched_attr *attr)
3955 struct sched_dl_entity *dl_se = &p->dl;
3957 if (dl_se->dl_runtime != attr->sched_runtime ||
3958 dl_se->dl_deadline != attr->sched_deadline ||
3959 dl_se->dl_period != attr->sched_period ||
3960 dl_se->flags != attr->sched_flags)
3966 static int __sched_setscheduler(struct task_struct *p,
3967 const struct sched_attr *attr,
3970 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3971 MAX_RT_PRIO - 1 - attr->sched_priority;
3972 int retval, oldprio, oldpolicy = -1, queued, running;
3973 int new_effective_prio, policy = attr->sched_policy;
3974 const struct sched_class *prev_class;
3977 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3980 /* may grab non-irq protected spin_locks */
3981 BUG_ON(in_interrupt());
3983 /* double check policy once rq lock held */
3985 reset_on_fork = p->sched_reset_on_fork;
3986 policy = oldpolicy = p->policy;
3988 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3990 if (!valid_policy(policy))
3994 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3998 * Valid priorities for SCHED_FIFO and SCHED_RR are
3999 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4000 * SCHED_BATCH and SCHED_IDLE is 0.
4002 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4003 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4005 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4006 (rt_policy(policy) != (attr->sched_priority != 0)))
4010 * Allow unprivileged RT tasks to decrease priority:
4012 if (user && !capable(CAP_SYS_NICE)) {
4013 if (fair_policy(policy)) {
4014 if (attr->sched_nice < task_nice(p) &&
4015 !can_nice(p, attr->sched_nice))
4019 if (rt_policy(policy)) {
4020 unsigned long rlim_rtprio =
4021 task_rlimit(p, RLIMIT_RTPRIO);
4023 /* can't set/change the rt policy */
4024 if (policy != p->policy && !rlim_rtprio)
4027 /* can't increase priority */
4028 if (attr->sched_priority > p->rt_priority &&
4029 attr->sched_priority > rlim_rtprio)
4034 * Can't set/change SCHED_DEADLINE policy at all for now
4035 * (safest behavior); in the future we would like to allow
4036 * unprivileged DL tasks to increase their relative deadline
4037 * or reduce their runtime (both ways reducing utilization)
4039 if (dl_policy(policy))
4043 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4044 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4046 if (idle_policy(p->policy) && !idle_policy(policy)) {
4047 if (!can_nice(p, task_nice(p)))
4051 /* can't change other user's priorities */
4052 if (!check_same_owner(p))
4055 /* Normal users shall not reset the sched_reset_on_fork flag */
4056 if (p->sched_reset_on_fork && !reset_on_fork)
4061 retval = security_task_setscheduler(p);
4067 * make sure no PI-waiters arrive (or leave) while we are
4068 * changing the priority of the task:
4070 * To be able to change p->policy safely, the appropriate
4071 * runqueue lock must be held.
4073 rq = task_rq_lock(p, &rf);
4076 * Changing the policy of the stop threads its a very bad idea
4078 if (p == rq->stop) {
4079 task_rq_unlock(rq, p, &rf);
4084 * If not changing anything there's no need to proceed further,
4085 * but store a possible modification of reset_on_fork.
4087 if (unlikely(policy == p->policy)) {
4088 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4090 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4092 if (dl_policy(policy) && dl_param_changed(p, attr))
4095 p->sched_reset_on_fork = reset_on_fork;
4096 task_rq_unlock(rq, p, &rf);
4102 #ifdef CONFIG_RT_GROUP_SCHED
4104 * Do not allow realtime tasks into groups that have no runtime
4107 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4108 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4109 !task_group_is_autogroup(task_group(p))) {
4110 task_rq_unlock(rq, p, &rf);
4115 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4116 cpumask_t *span = rq->rd->span;
4119 * Don't allow tasks with an affinity mask smaller than
4120 * the entire root_domain to become SCHED_DEADLINE. We
4121 * will also fail if there's no bandwidth available.
4123 if (!cpumask_subset(span, &p->cpus_allowed) ||
4124 rq->rd->dl_bw.bw == 0) {
4125 task_rq_unlock(rq, p, &rf);
4132 /* recheck policy now with rq lock held */
4133 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4134 policy = oldpolicy = -1;
4135 task_rq_unlock(rq, p, &rf);
4140 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4141 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4144 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4145 task_rq_unlock(rq, p, &rf);
4149 p->sched_reset_on_fork = reset_on_fork;
4154 * Take priority boosted tasks into account. If the new
4155 * effective priority is unchanged, we just store the new
4156 * normal parameters and do not touch the scheduler class and
4157 * the runqueue. This will be done when the task deboost
4160 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4161 if (new_effective_prio == oldprio)
4162 queue_flags &= ~DEQUEUE_MOVE;
4165 queued = task_on_rq_queued(p);
4166 running = task_current(rq, p);
4168 dequeue_task(rq, p, queue_flags);
4170 put_prev_task(rq, p);
4172 prev_class = p->sched_class;
4173 __setscheduler(rq, p, attr, pi);
4176 p->sched_class->set_curr_task(rq);
4179 * We enqueue to tail when the priority of a task is
4180 * increased (user space view).
4182 if (oldprio < p->prio)
4183 queue_flags |= ENQUEUE_HEAD;
4185 enqueue_task(rq, p, queue_flags);
4188 check_class_changed(rq, p, prev_class, oldprio);
4189 preempt_disable(); /* avoid rq from going away on us */
4190 task_rq_unlock(rq, p, &rf);
4193 rt_mutex_adjust_pi(p);
4196 * Run balance callbacks after we've adjusted the PI chain.
4198 balance_callback(rq);
4204 static int _sched_setscheduler(struct task_struct *p, int policy,
4205 const struct sched_param *param, bool check)
4207 struct sched_attr attr = {
4208 .sched_policy = policy,
4209 .sched_priority = param->sched_priority,
4210 .sched_nice = PRIO_TO_NICE(p->static_prio),
4213 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4214 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4215 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4216 policy &= ~SCHED_RESET_ON_FORK;
4217 attr.sched_policy = policy;
4220 return __sched_setscheduler(p, &attr, check, true);
4223 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4224 * @p: the task in question.
4225 * @policy: new policy.
4226 * @param: structure containing the new RT priority.
4228 * Return: 0 on success. An error code otherwise.
4230 * NOTE that the task may be already dead.
4232 int sched_setscheduler(struct task_struct *p, int policy,
4233 const struct sched_param *param)
4235 return _sched_setscheduler(p, policy, param, true);
4237 EXPORT_SYMBOL_GPL(sched_setscheduler);
4239 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4241 return __sched_setscheduler(p, attr, true, true);
4243 EXPORT_SYMBOL_GPL(sched_setattr);
4246 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4247 * @p: the task in question.
4248 * @policy: new policy.
4249 * @param: structure containing the new RT priority.
4251 * Just like sched_setscheduler, only don't bother checking if the
4252 * current context has permission. For example, this is needed in
4253 * stop_machine(): we create temporary high priority worker threads,
4254 * but our caller might not have that capability.
4256 * Return: 0 on success. An error code otherwise.
4258 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4259 const struct sched_param *param)
4261 return _sched_setscheduler(p, policy, param, false);
4263 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4266 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4268 struct sched_param lparam;
4269 struct task_struct *p;
4272 if (!param || pid < 0)
4274 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4279 p = find_process_by_pid(pid);
4281 retval = sched_setscheduler(p, policy, &lparam);
4288 * Mimics kernel/events/core.c perf_copy_attr().
4290 static int sched_copy_attr(struct sched_attr __user *uattr,
4291 struct sched_attr *attr)
4296 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4300 * zero the full structure, so that a short copy will be nice.
4302 memset(attr, 0, sizeof(*attr));
4304 ret = get_user(size, &uattr->size);
4308 if (size > PAGE_SIZE) /* silly large */
4311 if (!size) /* abi compat */
4312 size = SCHED_ATTR_SIZE_VER0;
4314 if (size < SCHED_ATTR_SIZE_VER0)
4318 * If we're handed a bigger struct than we know of,
4319 * ensure all the unknown bits are 0 - i.e. new
4320 * user-space does not rely on any kernel feature
4321 * extensions we dont know about yet.
4323 if (size > sizeof(*attr)) {
4324 unsigned char __user *addr;
4325 unsigned char __user *end;
4328 addr = (void __user *)uattr + sizeof(*attr);
4329 end = (void __user *)uattr + size;
4331 for (; addr < end; addr++) {
4332 ret = get_user(val, addr);
4338 size = sizeof(*attr);
4341 ret = copy_from_user(attr, uattr, size);
4346 * XXX: do we want to be lenient like existing syscalls; or do we want
4347 * to be strict and return an error on out-of-bounds values?
4349 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4354 put_user(sizeof(*attr), &uattr->size);
4359 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4360 * @pid: the pid in question.
4361 * @policy: new policy.
4362 * @param: structure containing the new RT priority.
4364 * Return: 0 on success. An error code otherwise.
4366 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4367 struct sched_param __user *, param)
4369 /* negative values for policy are not valid */
4373 return do_sched_setscheduler(pid, policy, param);
4377 * sys_sched_setparam - set/change the RT priority of a thread
4378 * @pid: the pid in question.
4379 * @param: structure containing the new RT priority.
4381 * Return: 0 on success. An error code otherwise.
4383 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4385 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4389 * sys_sched_setattr - same as above, but with extended sched_attr
4390 * @pid: the pid in question.
4391 * @uattr: structure containing the extended parameters.
4392 * @flags: for future extension.
4394 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4395 unsigned int, flags)
4397 struct sched_attr attr;
4398 struct task_struct *p;
4401 if (!uattr || pid < 0 || flags)
4404 retval = sched_copy_attr(uattr, &attr);
4408 if ((int)attr.sched_policy < 0)
4413 p = find_process_by_pid(pid);
4415 retval = sched_setattr(p, &attr);
4422 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4423 * @pid: the pid in question.
4425 * Return: On success, the policy of the thread. Otherwise, a negative error
4428 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4430 struct task_struct *p;
4438 p = find_process_by_pid(pid);
4440 retval = security_task_getscheduler(p);
4443 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4450 * sys_sched_getparam - get the RT priority of a thread
4451 * @pid: the pid in question.
4452 * @param: structure containing the RT priority.
4454 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4457 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4459 struct sched_param lp = { .sched_priority = 0 };
4460 struct task_struct *p;
4463 if (!param || pid < 0)
4467 p = find_process_by_pid(pid);
4472 retval = security_task_getscheduler(p);
4476 if (task_has_rt_policy(p))
4477 lp.sched_priority = p->rt_priority;
4481 * This one might sleep, we cannot do it with a spinlock held ...
4483 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4492 static int sched_read_attr(struct sched_attr __user *uattr,
4493 struct sched_attr *attr,
4498 if (!access_ok(VERIFY_WRITE, uattr, usize))
4502 * If we're handed a smaller struct than we know of,
4503 * ensure all the unknown bits are 0 - i.e. old
4504 * user-space does not get uncomplete information.
4506 if (usize < sizeof(*attr)) {
4507 unsigned char *addr;
4510 addr = (void *)attr + usize;
4511 end = (void *)attr + sizeof(*attr);
4513 for (; addr < end; addr++) {
4521 ret = copy_to_user(uattr, attr, attr->size);
4529 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4530 * @pid: the pid in question.
4531 * @uattr: structure containing the extended parameters.
4532 * @size: sizeof(attr) for fwd/bwd comp.
4533 * @flags: for future extension.
4535 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4536 unsigned int, size, unsigned int, flags)
4538 struct sched_attr attr = {
4539 .size = sizeof(struct sched_attr),
4541 struct task_struct *p;
4544 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4545 size < SCHED_ATTR_SIZE_VER0 || flags)
4549 p = find_process_by_pid(pid);
4554 retval = security_task_getscheduler(p);
4558 attr.sched_policy = p->policy;
4559 if (p->sched_reset_on_fork)
4560 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4561 if (task_has_dl_policy(p))
4562 __getparam_dl(p, &attr);
4563 else if (task_has_rt_policy(p))
4564 attr.sched_priority = p->rt_priority;
4566 attr.sched_nice = task_nice(p);
4570 retval = sched_read_attr(uattr, &attr, size);
4578 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4580 cpumask_var_t cpus_allowed, new_mask;
4581 struct task_struct *p;
4586 p = find_process_by_pid(pid);
4592 /* Prevent p going away */
4596 if (p->flags & PF_NO_SETAFFINITY) {
4600 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4604 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4606 goto out_free_cpus_allowed;
4609 if (!check_same_owner(p)) {
4611 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4613 goto out_free_new_mask;
4618 retval = security_task_setscheduler(p);
4620 goto out_free_new_mask;
4623 cpuset_cpus_allowed(p, cpus_allowed);
4624 cpumask_and(new_mask, in_mask, cpus_allowed);
4627 * Since bandwidth control happens on root_domain basis,
4628 * if admission test is enabled, we only admit -deadline
4629 * tasks allowed to run on all the CPUs in the task's
4633 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4635 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4638 goto out_free_new_mask;
4644 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4647 cpuset_cpus_allowed(p, cpus_allowed);
4648 if (!cpumask_subset(new_mask, cpus_allowed)) {
4650 * We must have raced with a concurrent cpuset
4651 * update. Just reset the cpus_allowed to the
4652 * cpuset's cpus_allowed
4654 cpumask_copy(new_mask, cpus_allowed);
4659 free_cpumask_var(new_mask);
4660 out_free_cpus_allowed:
4661 free_cpumask_var(cpus_allowed);
4667 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4668 struct cpumask *new_mask)
4670 if (len < cpumask_size())
4671 cpumask_clear(new_mask);
4672 else if (len > cpumask_size())
4673 len = cpumask_size();
4675 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4679 * sys_sched_setaffinity - set the cpu affinity of a process
4680 * @pid: pid of the process
4681 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4682 * @user_mask_ptr: user-space pointer to the new cpu mask
4684 * Return: 0 on success. An error code otherwise.
4686 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4687 unsigned long __user *, user_mask_ptr)
4689 cpumask_var_t new_mask;
4692 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4695 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4697 retval = sched_setaffinity(pid, new_mask);
4698 free_cpumask_var(new_mask);
4702 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4704 struct task_struct *p;
4705 unsigned long flags;
4711 p = find_process_by_pid(pid);
4715 retval = security_task_getscheduler(p);
4719 raw_spin_lock_irqsave(&p->pi_lock, flags);
4720 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4721 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4730 * sys_sched_getaffinity - get the cpu affinity of a process
4731 * @pid: pid of the process
4732 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4733 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4735 * Return: 0 on success. An error code otherwise.
4737 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4738 unsigned long __user *, user_mask_ptr)
4743 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4745 if (len & (sizeof(unsigned long)-1))
4748 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4751 ret = sched_getaffinity(pid, mask);
4753 size_t retlen = min_t(size_t, len, cpumask_size());
4755 if (copy_to_user(user_mask_ptr, mask, retlen))
4760 free_cpumask_var(mask);
4766 * sys_sched_yield - yield the current processor to other threads.
4768 * This function yields the current CPU to other tasks. If there are no
4769 * other threads running on this CPU then this function will return.
4773 SYSCALL_DEFINE0(sched_yield)
4775 struct rq *rq = this_rq_lock();
4777 schedstat_inc(rq, yld_count);
4778 current->sched_class->yield_task(rq);
4781 * Since we are going to call schedule() anyway, there's
4782 * no need to preempt or enable interrupts:
4784 __release(rq->lock);
4785 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4786 do_raw_spin_unlock(&rq->lock);
4787 sched_preempt_enable_no_resched();
4794 int __sched _cond_resched(void)
4796 if (should_resched(0)) {
4797 preempt_schedule_common();
4802 EXPORT_SYMBOL(_cond_resched);
4805 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4806 * call schedule, and on return reacquire the lock.
4808 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4809 * operations here to prevent schedule() from being called twice (once via
4810 * spin_unlock(), once by hand).
4812 int __cond_resched_lock(spinlock_t *lock)
4814 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4817 lockdep_assert_held(lock);
4819 if (spin_needbreak(lock) || resched) {
4822 preempt_schedule_common();
4830 EXPORT_SYMBOL(__cond_resched_lock);
4832 int __sched __cond_resched_softirq(void)
4834 BUG_ON(!in_softirq());
4836 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4838 preempt_schedule_common();
4844 EXPORT_SYMBOL(__cond_resched_softirq);
4847 * yield - yield the current processor to other threads.
4849 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4851 * The scheduler is at all times free to pick the calling task as the most
4852 * eligible task to run, if removing the yield() call from your code breaks
4853 * it, its already broken.
4855 * Typical broken usage is:
4860 * where one assumes that yield() will let 'the other' process run that will
4861 * make event true. If the current task is a SCHED_FIFO task that will never
4862 * happen. Never use yield() as a progress guarantee!!
4864 * If you want to use yield() to wait for something, use wait_event().
4865 * If you want to use yield() to be 'nice' for others, use cond_resched().
4866 * If you still want to use yield(), do not!
4868 void __sched yield(void)
4870 set_current_state(TASK_RUNNING);
4873 EXPORT_SYMBOL(yield);
4876 * yield_to - yield the current processor to another thread in
4877 * your thread group, or accelerate that thread toward the
4878 * processor it's on.
4880 * @preempt: whether task preemption is allowed or not
4882 * It's the caller's job to ensure that the target task struct
4883 * can't go away on us before we can do any checks.
4886 * true (>0) if we indeed boosted the target task.
4887 * false (0) if we failed to boost the target.
4888 * -ESRCH if there's no task to yield to.
4890 int __sched yield_to(struct task_struct *p, bool preempt)
4892 struct task_struct *curr = current;
4893 struct rq *rq, *p_rq;
4894 unsigned long flags;
4897 local_irq_save(flags);
4903 * If we're the only runnable task on the rq and target rq also
4904 * has only one task, there's absolutely no point in yielding.
4906 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4911 double_rq_lock(rq, p_rq);
4912 if (task_rq(p) != p_rq) {
4913 double_rq_unlock(rq, p_rq);
4917 if (!curr->sched_class->yield_to_task)
4920 if (curr->sched_class != p->sched_class)
4923 if (task_running(p_rq, p) || p->state)
4926 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4928 schedstat_inc(rq, yld_count);
4930 * Make p's CPU reschedule; pick_next_entity takes care of
4933 if (preempt && rq != p_rq)
4938 double_rq_unlock(rq, p_rq);
4940 local_irq_restore(flags);
4947 EXPORT_SYMBOL_GPL(yield_to);
4950 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4951 * that process accounting knows that this is a task in IO wait state.
4953 long __sched io_schedule_timeout(long timeout)
4955 int old_iowait = current->in_iowait;
4959 current->in_iowait = 1;
4960 blk_schedule_flush_plug(current);
4962 delayacct_blkio_start();
4964 atomic_inc(&rq->nr_iowait);
4965 ret = schedule_timeout(timeout);
4966 current->in_iowait = old_iowait;
4967 atomic_dec(&rq->nr_iowait);
4968 delayacct_blkio_end();
4972 EXPORT_SYMBOL(io_schedule_timeout);
4975 * sys_sched_get_priority_max - return maximum RT priority.
4976 * @policy: scheduling class.
4978 * Return: On success, this syscall returns the maximum
4979 * rt_priority that can be used by a given scheduling class.
4980 * On failure, a negative error code is returned.
4982 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4989 ret = MAX_USER_RT_PRIO-1;
4991 case SCHED_DEADLINE:
5002 * sys_sched_get_priority_min - return minimum RT priority.
5003 * @policy: scheduling class.
5005 * Return: On success, this syscall returns the minimum
5006 * rt_priority that can be used by a given scheduling class.
5007 * On failure, a negative error code is returned.
5009 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5018 case SCHED_DEADLINE:
5028 * sys_sched_rr_get_interval - return the default timeslice of a process.
5029 * @pid: pid of the process.
5030 * @interval: userspace pointer to the timeslice value.
5032 * this syscall writes the default timeslice value of a given process
5033 * into the user-space timespec buffer. A value of '0' means infinity.
5035 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5038 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5039 struct timespec __user *, interval)
5041 struct task_struct *p;
5042 unsigned int time_slice;
5053 p = find_process_by_pid(pid);
5057 retval = security_task_getscheduler(p);
5061 rq = task_rq_lock(p, &rf);
5063 if (p->sched_class->get_rr_interval)
5064 time_slice = p->sched_class->get_rr_interval(rq, p);
5065 task_rq_unlock(rq, p, &rf);
5068 jiffies_to_timespec(time_slice, &t);
5069 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5077 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5079 void sched_show_task(struct task_struct *p)
5081 unsigned long free = 0;
5083 unsigned long state = p->state;
5086 state = __ffs(state) + 1;
5087 printk(KERN_INFO "%-15.15s %c", p->comm,
5088 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5089 #if BITS_PER_LONG == 32
5090 if (state == TASK_RUNNING)
5091 printk(KERN_CONT " running ");
5093 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5095 if (state == TASK_RUNNING)
5096 printk(KERN_CONT " running task ");
5098 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5100 #ifdef CONFIG_DEBUG_STACK_USAGE
5101 free = stack_not_used(p);
5106 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5108 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5109 task_pid_nr(p), ppid,
5110 (unsigned long)task_thread_info(p)->flags);
5112 print_worker_info(KERN_INFO, p);
5113 show_stack(p, NULL);
5116 void show_state_filter(unsigned long state_filter)
5118 struct task_struct *g, *p;
5120 #if BITS_PER_LONG == 32
5122 " task PC stack pid father\n");
5125 " task PC stack pid father\n");
5128 for_each_process_thread(g, p) {
5130 * reset the NMI-timeout, listing all files on a slow
5131 * console might take a lot of time:
5133 touch_nmi_watchdog();
5134 if (!state_filter || (p->state & state_filter))
5138 touch_all_softlockup_watchdogs();
5140 #ifdef CONFIG_SCHED_DEBUG
5142 sysrq_sched_debug_show();
5146 * Only show locks if all tasks are dumped:
5149 debug_show_all_locks();
5152 void init_idle_bootup_task(struct task_struct *idle)
5154 idle->sched_class = &idle_sched_class;
5158 * init_idle - set up an idle thread for a given CPU
5159 * @idle: task in question
5160 * @cpu: cpu the idle task belongs to
5162 * NOTE: this function does not set the idle thread's NEED_RESCHED
5163 * flag, to make booting more robust.
5165 void init_idle(struct task_struct *idle, int cpu)
5167 struct rq *rq = cpu_rq(cpu);
5168 unsigned long flags;
5170 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5171 raw_spin_lock(&rq->lock);
5173 __sched_fork(0, idle);
5174 idle->state = TASK_RUNNING;
5175 idle->se.exec_start = sched_clock();
5177 kasan_unpoison_task_stack(idle);
5181 * Its possible that init_idle() gets called multiple times on a task,
5182 * in that case do_set_cpus_allowed() will not do the right thing.
5184 * And since this is boot we can forgo the serialization.
5186 set_cpus_allowed_common(idle, cpumask_of(cpu));
5189 * We're having a chicken and egg problem, even though we are
5190 * holding rq->lock, the cpu isn't yet set to this cpu so the
5191 * lockdep check in task_group() will fail.
5193 * Similar case to sched_fork(). / Alternatively we could
5194 * use task_rq_lock() here and obtain the other rq->lock.
5199 __set_task_cpu(idle, cpu);
5202 rq->curr = rq->idle = idle;
5203 idle->on_rq = TASK_ON_RQ_QUEUED;
5207 raw_spin_unlock(&rq->lock);
5208 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5210 /* Set the preempt count _outside_ the spinlocks! */
5211 init_idle_preempt_count(idle, cpu);
5214 * The idle tasks have their own, simple scheduling class:
5216 idle->sched_class = &idle_sched_class;
5217 ftrace_graph_init_idle_task(idle, cpu);
5218 vtime_init_idle(idle, cpu);
5220 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5224 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5225 const struct cpumask *trial)
5227 int ret = 1, trial_cpus;
5228 struct dl_bw *cur_dl_b;
5229 unsigned long flags;
5231 if (!cpumask_weight(cur))
5234 rcu_read_lock_sched();
5235 cur_dl_b = dl_bw_of(cpumask_any(cur));
5236 trial_cpus = cpumask_weight(trial);
5238 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5239 if (cur_dl_b->bw != -1 &&
5240 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5242 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5243 rcu_read_unlock_sched();
5248 int task_can_attach(struct task_struct *p,
5249 const struct cpumask *cs_cpus_allowed)
5254 * Kthreads which disallow setaffinity shouldn't be moved
5255 * to a new cpuset; we don't want to change their cpu
5256 * affinity and isolating such threads by their set of
5257 * allowed nodes is unnecessary. Thus, cpusets are not
5258 * applicable for such threads. This prevents checking for
5259 * success of set_cpus_allowed_ptr() on all attached tasks
5260 * before cpus_allowed may be changed.
5262 if (p->flags & PF_NO_SETAFFINITY) {
5268 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5270 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5275 unsigned long flags;
5277 rcu_read_lock_sched();
5278 dl_b = dl_bw_of(dest_cpu);
5279 raw_spin_lock_irqsave(&dl_b->lock, flags);
5280 cpus = dl_bw_cpus(dest_cpu);
5281 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5286 * We reserve space for this task in the destination
5287 * root_domain, as we can't fail after this point.
5288 * We will free resources in the source root_domain
5289 * later on (see set_cpus_allowed_dl()).
5291 __dl_add(dl_b, p->dl.dl_bw);
5293 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5294 rcu_read_unlock_sched();
5304 static bool sched_smp_initialized __read_mostly;
5306 #ifdef CONFIG_NUMA_BALANCING
5307 /* Migrate current task p to target_cpu */
5308 int migrate_task_to(struct task_struct *p, int target_cpu)
5310 struct migration_arg arg = { p, target_cpu };
5311 int curr_cpu = task_cpu(p);
5313 if (curr_cpu == target_cpu)
5316 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5319 /* TODO: This is not properly updating schedstats */
5321 trace_sched_move_numa(p, curr_cpu, target_cpu);
5322 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5326 * Requeue a task on a given node and accurately track the number of NUMA
5327 * tasks on the runqueues
5329 void sched_setnuma(struct task_struct *p, int nid)
5331 bool queued, running;
5335 rq = task_rq_lock(p, &rf);
5336 queued = task_on_rq_queued(p);
5337 running = task_current(rq, p);
5340 dequeue_task(rq, p, DEQUEUE_SAVE);
5342 put_prev_task(rq, p);
5344 p->numa_preferred_nid = nid;
5347 p->sched_class->set_curr_task(rq);
5349 enqueue_task(rq, p, ENQUEUE_RESTORE);
5350 task_rq_unlock(rq, p, &rf);
5352 #endif /* CONFIG_NUMA_BALANCING */
5354 #ifdef CONFIG_HOTPLUG_CPU
5356 * Ensures that the idle task is using init_mm right before its cpu goes
5359 void idle_task_exit(void)
5361 struct mm_struct *mm = current->active_mm;
5363 BUG_ON(cpu_online(smp_processor_id()));
5365 if (mm != &init_mm) {
5366 switch_mm_irqs_off(mm, &init_mm, current);
5367 finish_arch_post_lock_switch();
5373 * Since this CPU is going 'away' for a while, fold any nr_active delta
5374 * we might have. Assumes we're called after migrate_tasks() so that the
5375 * nr_active count is stable.
5377 * Also see the comment "Global load-average calculations".
5379 static void calc_load_migrate(struct rq *rq)
5381 long delta = calc_load_fold_active(rq);
5383 atomic_long_add(delta, &calc_load_tasks);
5386 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5390 static const struct sched_class fake_sched_class = {
5391 .put_prev_task = put_prev_task_fake,
5394 static struct task_struct fake_task = {
5396 * Avoid pull_{rt,dl}_task()
5398 .prio = MAX_PRIO + 1,
5399 .sched_class = &fake_sched_class,
5403 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5404 * try_to_wake_up()->select_task_rq().
5406 * Called with rq->lock held even though we'er in stop_machine() and
5407 * there's no concurrency possible, we hold the required locks anyway
5408 * because of lock validation efforts.
5410 static void migrate_tasks(struct rq *dead_rq)
5412 struct rq *rq = dead_rq;
5413 struct task_struct *next, *stop = rq->stop;
5414 struct pin_cookie cookie;
5418 * Fudge the rq selection such that the below task selection loop
5419 * doesn't get stuck on the currently eligible stop task.
5421 * We're currently inside stop_machine() and the rq is either stuck
5422 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5423 * either way we should never end up calling schedule() until we're
5429 * put_prev_task() and pick_next_task() sched
5430 * class method both need to have an up-to-date
5431 * value of rq->clock[_task]
5433 update_rq_clock(rq);
5437 * There's this thread running, bail when that's the only
5440 if (rq->nr_running == 1)
5444 * pick_next_task assumes pinned rq->lock.
5446 cookie = lockdep_pin_lock(&rq->lock);
5447 next = pick_next_task(rq, &fake_task, cookie);
5449 next->sched_class->put_prev_task(rq, next);
5452 * Rules for changing task_struct::cpus_allowed are holding
5453 * both pi_lock and rq->lock, such that holding either
5454 * stabilizes the mask.
5456 * Drop rq->lock is not quite as disastrous as it usually is
5457 * because !cpu_active at this point, which means load-balance
5458 * will not interfere. Also, stop-machine.
5460 lockdep_unpin_lock(&rq->lock, cookie);
5461 raw_spin_unlock(&rq->lock);
5462 raw_spin_lock(&next->pi_lock);
5463 raw_spin_lock(&rq->lock);
5466 * Since we're inside stop-machine, _nothing_ should have
5467 * changed the task, WARN if weird stuff happened, because in
5468 * that case the above rq->lock drop is a fail too.
5470 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5471 raw_spin_unlock(&next->pi_lock);
5475 /* Find suitable destination for @next, with force if needed. */
5476 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5478 rq = __migrate_task(rq, next, dest_cpu);
5479 if (rq != dead_rq) {
5480 raw_spin_unlock(&rq->lock);
5482 raw_spin_lock(&rq->lock);
5484 raw_spin_unlock(&next->pi_lock);
5489 #endif /* CONFIG_HOTPLUG_CPU */
5491 static void set_rq_online(struct rq *rq)
5494 const struct sched_class *class;
5496 cpumask_set_cpu(rq->cpu, rq->rd->online);
5499 for_each_class(class) {
5500 if (class->rq_online)
5501 class->rq_online(rq);
5506 static void set_rq_offline(struct rq *rq)
5509 const struct sched_class *class;
5511 for_each_class(class) {
5512 if (class->rq_offline)
5513 class->rq_offline(rq);
5516 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5521 static void set_cpu_rq_start_time(unsigned int cpu)
5523 struct rq *rq = cpu_rq(cpu);
5525 rq->age_stamp = sched_clock_cpu(cpu);
5528 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5530 #ifdef CONFIG_SCHED_DEBUG
5532 static __read_mostly int sched_debug_enabled;
5534 static int __init sched_debug_setup(char *str)
5536 sched_debug_enabled = 1;
5540 early_param("sched_debug", sched_debug_setup);
5542 static inline bool sched_debug(void)
5544 return sched_debug_enabled;
5547 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5548 struct cpumask *groupmask)
5550 struct sched_group *group = sd->groups;
5552 cpumask_clear(groupmask);
5554 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5556 if (!(sd->flags & SD_LOAD_BALANCE)) {
5557 printk("does not load-balance\n");
5559 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5564 printk(KERN_CONT "span %*pbl level %s\n",
5565 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5567 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5568 printk(KERN_ERR "ERROR: domain->span does not contain "
5571 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5572 printk(KERN_ERR "ERROR: domain->groups does not contain"
5576 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5580 printk(KERN_ERR "ERROR: group is NULL\n");
5584 if (!cpumask_weight(sched_group_cpus(group))) {
5585 printk(KERN_CONT "\n");
5586 printk(KERN_ERR "ERROR: empty group\n");
5590 if (!(sd->flags & SD_OVERLAP) &&
5591 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5592 printk(KERN_CONT "\n");
5593 printk(KERN_ERR "ERROR: repeated CPUs\n");
5597 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5599 printk(KERN_CONT " %*pbl",
5600 cpumask_pr_args(sched_group_cpus(group)));
5601 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5602 printk(KERN_CONT " (cpu_capacity = %d)",
5603 group->sgc->capacity);
5606 group = group->next;
5607 } while (group != sd->groups);
5608 printk(KERN_CONT "\n");
5610 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5611 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5614 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5615 printk(KERN_ERR "ERROR: parent span is not a superset "
5616 "of domain->span\n");
5620 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5624 if (!sched_debug_enabled)
5628 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5632 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5635 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5643 #else /* !CONFIG_SCHED_DEBUG */
5644 # define sched_domain_debug(sd, cpu) do { } while (0)
5645 static inline bool sched_debug(void)
5649 #endif /* CONFIG_SCHED_DEBUG */
5651 static int sd_degenerate(struct sched_domain *sd)
5653 if (cpumask_weight(sched_domain_span(sd)) == 1)
5656 /* Following flags need at least 2 groups */
5657 if (sd->flags & (SD_LOAD_BALANCE |
5658 SD_BALANCE_NEWIDLE |
5661 SD_SHARE_CPUCAPACITY |
5662 SD_SHARE_PKG_RESOURCES |
5663 SD_SHARE_POWERDOMAIN)) {
5664 if (sd->groups != sd->groups->next)
5668 /* Following flags don't use groups */
5669 if (sd->flags & (SD_WAKE_AFFINE))
5676 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5678 unsigned long cflags = sd->flags, pflags = parent->flags;
5680 if (sd_degenerate(parent))
5683 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5686 /* Flags needing groups don't count if only 1 group in parent */
5687 if (parent->groups == parent->groups->next) {
5688 pflags &= ~(SD_LOAD_BALANCE |
5689 SD_BALANCE_NEWIDLE |
5692 SD_SHARE_CPUCAPACITY |
5693 SD_SHARE_PKG_RESOURCES |
5695 SD_SHARE_POWERDOMAIN);
5696 if (nr_node_ids == 1)
5697 pflags &= ~SD_SERIALIZE;
5699 if (~cflags & pflags)
5705 static void free_rootdomain(struct rcu_head *rcu)
5707 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5709 cpupri_cleanup(&rd->cpupri);
5710 cpudl_cleanup(&rd->cpudl);
5711 free_cpumask_var(rd->dlo_mask);
5712 free_cpumask_var(rd->rto_mask);
5713 free_cpumask_var(rd->online);
5714 free_cpumask_var(rd->span);
5718 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5720 struct root_domain *old_rd = NULL;
5721 unsigned long flags;
5723 raw_spin_lock_irqsave(&rq->lock, flags);
5728 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5731 cpumask_clear_cpu(rq->cpu, old_rd->span);
5734 * If we dont want to free the old_rd yet then
5735 * set old_rd to NULL to skip the freeing later
5738 if (!atomic_dec_and_test(&old_rd->refcount))
5742 atomic_inc(&rd->refcount);
5745 cpumask_set_cpu(rq->cpu, rd->span);
5746 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5749 raw_spin_unlock_irqrestore(&rq->lock, flags);
5752 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5755 static int init_rootdomain(struct root_domain *rd)
5757 memset(rd, 0, sizeof(*rd));
5759 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5761 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5763 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5765 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5768 init_dl_bw(&rd->dl_bw);
5769 if (cpudl_init(&rd->cpudl) != 0)
5772 if (cpupri_init(&rd->cpupri) != 0)
5777 free_cpumask_var(rd->rto_mask);
5779 free_cpumask_var(rd->dlo_mask);
5781 free_cpumask_var(rd->online);
5783 free_cpumask_var(rd->span);
5789 * By default the system creates a single root-domain with all cpus as
5790 * members (mimicking the global state we have today).
5792 struct root_domain def_root_domain;
5794 static void init_defrootdomain(void)
5796 init_rootdomain(&def_root_domain);
5798 atomic_set(&def_root_domain.refcount, 1);
5801 static struct root_domain *alloc_rootdomain(void)
5803 struct root_domain *rd;
5805 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5809 if (init_rootdomain(rd) != 0) {
5817 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5819 struct sched_group *tmp, *first;
5828 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5833 } while (sg != first);
5836 static void free_sched_domain(struct rcu_head *rcu)
5838 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5841 * If its an overlapping domain it has private groups, iterate and
5844 if (sd->flags & SD_OVERLAP) {
5845 free_sched_groups(sd->groups, 1);
5846 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5847 kfree(sd->groups->sgc);
5853 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5855 call_rcu(&sd->rcu, free_sched_domain);
5858 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5860 for (; sd; sd = sd->parent)
5861 destroy_sched_domain(sd, cpu);
5865 * Keep a special pointer to the highest sched_domain that has
5866 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5867 * allows us to avoid some pointer chasing select_idle_sibling().
5869 * Also keep a unique ID per domain (we use the first cpu number in
5870 * the cpumask of the domain), this allows us to quickly tell if
5871 * two cpus are in the same cache domain, see cpus_share_cache().
5873 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5874 DEFINE_PER_CPU(int, sd_llc_size);
5875 DEFINE_PER_CPU(int, sd_llc_id);
5876 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5877 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5878 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5880 static void update_top_cache_domain(int cpu)
5882 struct sched_domain *sd;
5883 struct sched_domain *busy_sd = NULL;
5887 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5889 id = cpumask_first(sched_domain_span(sd));
5890 size = cpumask_weight(sched_domain_span(sd));
5891 busy_sd = sd->parent; /* sd_busy */
5893 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5895 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5896 per_cpu(sd_llc_size, cpu) = size;
5897 per_cpu(sd_llc_id, cpu) = id;
5899 sd = lowest_flag_domain(cpu, SD_NUMA);
5900 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5902 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5903 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5907 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5908 * hold the hotplug lock.
5911 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5913 struct rq *rq = cpu_rq(cpu);
5914 struct sched_domain *tmp;
5916 /* Remove the sched domains which do not contribute to scheduling. */
5917 for (tmp = sd; tmp; ) {
5918 struct sched_domain *parent = tmp->parent;
5922 if (sd_parent_degenerate(tmp, parent)) {
5923 tmp->parent = parent->parent;
5925 parent->parent->child = tmp;
5927 * Transfer SD_PREFER_SIBLING down in case of a
5928 * degenerate parent; the spans match for this
5929 * so the property transfers.
5931 if (parent->flags & SD_PREFER_SIBLING)
5932 tmp->flags |= SD_PREFER_SIBLING;
5933 destroy_sched_domain(parent, cpu);
5938 if (sd && sd_degenerate(sd)) {
5941 destroy_sched_domain(tmp, cpu);
5946 sched_domain_debug(sd, cpu);
5948 rq_attach_root(rq, rd);
5950 rcu_assign_pointer(rq->sd, sd);
5951 destroy_sched_domains(tmp, cpu);
5953 update_top_cache_domain(cpu);
5956 /* Setup the mask of cpus configured for isolated domains */
5957 static int __init isolated_cpu_setup(char *str)
5961 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5962 ret = cpulist_parse(str, cpu_isolated_map);
5964 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5969 __setup("isolcpus=", isolated_cpu_setup);
5972 struct sched_domain ** __percpu sd;
5973 struct root_domain *rd;
5984 * Build an iteration mask that can exclude certain CPUs from the upwards
5987 * Asymmetric node setups can result in situations where the domain tree is of
5988 * unequal depth, make sure to skip domains that already cover the entire
5991 * In that case build_sched_domains() will have terminated the iteration early
5992 * and our sibling sd spans will be empty. Domains should always include the
5993 * cpu they're built on, so check that.
5996 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5998 const struct cpumask *span = sched_domain_span(sd);
5999 struct sd_data *sdd = sd->private;
6000 struct sched_domain *sibling;
6003 for_each_cpu(i, span) {
6004 sibling = *per_cpu_ptr(sdd->sd, i);
6005 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6008 cpumask_set_cpu(i, sched_group_mask(sg));
6013 * Return the canonical balance cpu for this group, this is the first cpu
6014 * of this group that's also in the iteration mask.
6016 int group_balance_cpu(struct sched_group *sg)
6018 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6022 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6024 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6025 const struct cpumask *span = sched_domain_span(sd);
6026 struct cpumask *covered = sched_domains_tmpmask;
6027 struct sd_data *sdd = sd->private;
6028 struct sched_domain *sibling;
6031 cpumask_clear(covered);
6033 for_each_cpu(i, span) {
6034 struct cpumask *sg_span;
6036 if (cpumask_test_cpu(i, covered))
6039 sibling = *per_cpu_ptr(sdd->sd, i);
6041 /* See the comment near build_group_mask(). */
6042 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6045 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6046 GFP_KERNEL, cpu_to_node(cpu));
6051 sg_span = sched_group_cpus(sg);
6053 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6055 cpumask_set_cpu(i, sg_span);
6057 cpumask_or(covered, covered, sg_span);
6059 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6060 if (atomic_inc_return(&sg->sgc->ref) == 1)
6061 build_group_mask(sd, sg);
6064 * Initialize sgc->capacity such that even if we mess up the
6065 * domains and no possible iteration will get us here, we won't
6068 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6071 * Make sure the first group of this domain contains the
6072 * canonical balance cpu. Otherwise the sched_domain iteration
6073 * breaks. See update_sg_lb_stats().
6075 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6076 group_balance_cpu(sg) == cpu)
6086 sd->groups = groups;
6091 free_sched_groups(first, 0);
6096 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6098 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6099 struct sched_domain *child = sd->child;
6102 cpu = cpumask_first(sched_domain_span(child));
6105 *sg = *per_cpu_ptr(sdd->sg, cpu);
6106 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6107 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6114 * build_sched_groups will build a circular linked list of the groups
6115 * covered by the given span, and will set each group's ->cpumask correctly,
6116 * and ->cpu_capacity to 0.
6118 * Assumes the sched_domain tree is fully constructed
6121 build_sched_groups(struct sched_domain *sd, int cpu)
6123 struct sched_group *first = NULL, *last = NULL;
6124 struct sd_data *sdd = sd->private;
6125 const struct cpumask *span = sched_domain_span(sd);
6126 struct cpumask *covered;
6129 get_group(cpu, sdd, &sd->groups);
6130 atomic_inc(&sd->groups->ref);
6132 if (cpu != cpumask_first(span))
6135 lockdep_assert_held(&sched_domains_mutex);
6136 covered = sched_domains_tmpmask;
6138 cpumask_clear(covered);
6140 for_each_cpu(i, span) {
6141 struct sched_group *sg;
6144 if (cpumask_test_cpu(i, covered))
6147 group = get_group(i, sdd, &sg);
6148 cpumask_setall(sched_group_mask(sg));
6150 for_each_cpu(j, span) {
6151 if (get_group(j, sdd, NULL) != group)
6154 cpumask_set_cpu(j, covered);
6155 cpumask_set_cpu(j, sched_group_cpus(sg));
6170 * Initialize sched groups cpu_capacity.
6172 * cpu_capacity indicates the capacity of sched group, which is used while
6173 * distributing the load between different sched groups in a sched domain.
6174 * Typically cpu_capacity for all the groups in a sched domain will be same
6175 * unless there are asymmetries in the topology. If there are asymmetries,
6176 * group having more cpu_capacity will pickup more load compared to the
6177 * group having less cpu_capacity.
6179 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6181 struct sched_group *sg = sd->groups;
6186 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6188 } while (sg != sd->groups);
6190 if (cpu != group_balance_cpu(sg))
6193 update_group_capacity(sd, cpu);
6194 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6198 * Initializers for schedule domains
6199 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6202 static int default_relax_domain_level = -1;
6203 int sched_domain_level_max;
6205 static int __init setup_relax_domain_level(char *str)
6207 if (kstrtoint(str, 0, &default_relax_domain_level))
6208 pr_warn("Unable to set relax_domain_level\n");
6212 __setup("relax_domain_level=", setup_relax_domain_level);
6214 static void set_domain_attribute(struct sched_domain *sd,
6215 struct sched_domain_attr *attr)
6219 if (!attr || attr->relax_domain_level < 0) {
6220 if (default_relax_domain_level < 0)
6223 request = default_relax_domain_level;
6225 request = attr->relax_domain_level;
6226 if (request < sd->level) {
6227 /* turn off idle balance on this domain */
6228 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6230 /* turn on idle balance on this domain */
6231 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6235 static void __sdt_free(const struct cpumask *cpu_map);
6236 static int __sdt_alloc(const struct cpumask *cpu_map);
6238 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6239 const struct cpumask *cpu_map)
6243 if (!atomic_read(&d->rd->refcount))
6244 free_rootdomain(&d->rd->rcu); /* fall through */
6246 free_percpu(d->sd); /* fall through */
6248 __sdt_free(cpu_map); /* fall through */
6254 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6255 const struct cpumask *cpu_map)
6257 memset(d, 0, sizeof(*d));
6259 if (__sdt_alloc(cpu_map))
6260 return sa_sd_storage;
6261 d->sd = alloc_percpu(struct sched_domain *);
6263 return sa_sd_storage;
6264 d->rd = alloc_rootdomain();
6267 return sa_rootdomain;
6271 * NULL the sd_data elements we've used to build the sched_domain and
6272 * sched_group structure so that the subsequent __free_domain_allocs()
6273 * will not free the data we're using.
6275 static void claim_allocations(int cpu, struct sched_domain *sd)
6277 struct sd_data *sdd = sd->private;
6279 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6280 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6282 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6283 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6285 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6286 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6290 static int sched_domains_numa_levels;
6291 enum numa_topology_type sched_numa_topology_type;
6292 static int *sched_domains_numa_distance;
6293 int sched_max_numa_distance;
6294 static struct cpumask ***sched_domains_numa_masks;
6295 static int sched_domains_curr_level;
6299 * SD_flags allowed in topology descriptions.
6301 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6302 * SD_SHARE_PKG_RESOURCES - describes shared caches
6303 * SD_NUMA - describes NUMA topologies
6304 * SD_SHARE_POWERDOMAIN - describes shared power domain
6307 * SD_ASYM_PACKING - describes SMT quirks
6309 #define TOPOLOGY_SD_FLAGS \
6310 (SD_SHARE_CPUCAPACITY | \
6311 SD_SHARE_PKG_RESOURCES | \
6314 SD_SHARE_POWERDOMAIN)
6316 static struct sched_domain *
6317 sd_init(struct sched_domain_topology_level *tl, int cpu)
6319 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6320 int sd_weight, sd_flags = 0;
6324 * Ugly hack to pass state to sd_numa_mask()...
6326 sched_domains_curr_level = tl->numa_level;
6329 sd_weight = cpumask_weight(tl->mask(cpu));
6332 sd_flags = (*tl->sd_flags)();
6333 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6334 "wrong sd_flags in topology description\n"))
6335 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6337 *sd = (struct sched_domain){
6338 .min_interval = sd_weight,
6339 .max_interval = 2*sd_weight,
6341 .imbalance_pct = 125,
6343 .cache_nice_tries = 0,
6350 .flags = 1*SD_LOAD_BALANCE
6351 | 1*SD_BALANCE_NEWIDLE
6356 | 0*SD_SHARE_CPUCAPACITY
6357 | 0*SD_SHARE_PKG_RESOURCES
6359 | 0*SD_PREFER_SIBLING
6364 .last_balance = jiffies,
6365 .balance_interval = sd_weight,
6367 .max_newidle_lb_cost = 0,
6368 .next_decay_max_lb_cost = jiffies,
6369 #ifdef CONFIG_SCHED_DEBUG
6375 * Convert topological properties into behaviour.
6378 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6379 sd->flags |= SD_PREFER_SIBLING;
6380 sd->imbalance_pct = 110;
6381 sd->smt_gain = 1178; /* ~15% */
6383 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6384 sd->imbalance_pct = 117;
6385 sd->cache_nice_tries = 1;
6389 } else if (sd->flags & SD_NUMA) {
6390 sd->cache_nice_tries = 2;
6394 sd->flags |= SD_SERIALIZE;
6395 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6396 sd->flags &= ~(SD_BALANCE_EXEC |
6403 sd->flags |= SD_PREFER_SIBLING;
6404 sd->cache_nice_tries = 1;
6409 sd->private = &tl->data;
6415 * Topology list, bottom-up.
6417 static struct sched_domain_topology_level default_topology[] = {
6418 #ifdef CONFIG_SCHED_SMT
6419 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6421 #ifdef CONFIG_SCHED_MC
6422 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6424 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6428 static struct sched_domain_topology_level *sched_domain_topology =
6431 #define for_each_sd_topology(tl) \
6432 for (tl = sched_domain_topology; tl->mask; tl++)
6434 void set_sched_topology(struct sched_domain_topology_level *tl)
6436 sched_domain_topology = tl;
6441 static const struct cpumask *sd_numa_mask(int cpu)
6443 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6446 static void sched_numa_warn(const char *str)
6448 static int done = false;
6456 printk(KERN_WARNING "ERROR: %s\n\n", str);
6458 for (i = 0; i < nr_node_ids; i++) {
6459 printk(KERN_WARNING " ");
6460 for (j = 0; j < nr_node_ids; j++)
6461 printk(KERN_CONT "%02d ", node_distance(i,j));
6462 printk(KERN_CONT "\n");
6464 printk(KERN_WARNING "\n");
6467 bool find_numa_distance(int distance)
6471 if (distance == node_distance(0, 0))
6474 for (i = 0; i < sched_domains_numa_levels; i++) {
6475 if (sched_domains_numa_distance[i] == distance)
6483 * A system can have three types of NUMA topology:
6484 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6485 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6486 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6488 * The difference between a glueless mesh topology and a backplane
6489 * topology lies in whether communication between not directly
6490 * connected nodes goes through intermediary nodes (where programs
6491 * could run), or through backplane controllers. This affects
6492 * placement of programs.
6494 * The type of topology can be discerned with the following tests:
6495 * - If the maximum distance between any nodes is 1 hop, the system
6496 * is directly connected.
6497 * - If for two nodes A and B, located N > 1 hops away from each other,
6498 * there is an intermediary node C, which is < N hops away from both
6499 * nodes A and B, the system is a glueless mesh.
6501 static void init_numa_topology_type(void)
6505 n = sched_max_numa_distance;
6507 if (sched_domains_numa_levels <= 1) {
6508 sched_numa_topology_type = NUMA_DIRECT;
6512 for_each_online_node(a) {
6513 for_each_online_node(b) {
6514 /* Find two nodes furthest removed from each other. */
6515 if (node_distance(a, b) < n)
6518 /* Is there an intermediary node between a and b? */
6519 for_each_online_node(c) {
6520 if (node_distance(a, c) < n &&
6521 node_distance(b, c) < n) {
6522 sched_numa_topology_type =
6528 sched_numa_topology_type = NUMA_BACKPLANE;
6534 static void sched_init_numa(void)
6536 int next_distance, curr_distance = node_distance(0, 0);
6537 struct sched_domain_topology_level *tl;
6541 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6542 if (!sched_domains_numa_distance)
6546 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6547 * unique distances in the node_distance() table.
6549 * Assumes node_distance(0,j) includes all distances in
6550 * node_distance(i,j) in order to avoid cubic time.
6552 next_distance = curr_distance;
6553 for (i = 0; i < nr_node_ids; i++) {
6554 for (j = 0; j < nr_node_ids; j++) {
6555 for (k = 0; k < nr_node_ids; k++) {
6556 int distance = node_distance(i, k);
6558 if (distance > curr_distance &&
6559 (distance < next_distance ||
6560 next_distance == curr_distance))
6561 next_distance = distance;
6564 * While not a strong assumption it would be nice to know
6565 * about cases where if node A is connected to B, B is not
6566 * equally connected to A.
6568 if (sched_debug() && node_distance(k, i) != distance)
6569 sched_numa_warn("Node-distance not symmetric");
6571 if (sched_debug() && i && !find_numa_distance(distance))
6572 sched_numa_warn("Node-0 not representative");
6574 if (next_distance != curr_distance) {
6575 sched_domains_numa_distance[level++] = next_distance;
6576 sched_domains_numa_levels = level;
6577 curr_distance = next_distance;
6582 * In case of sched_debug() we verify the above assumption.
6592 * 'level' contains the number of unique distances, excluding the
6593 * identity distance node_distance(i,i).
6595 * The sched_domains_numa_distance[] array includes the actual distance
6600 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6601 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6602 * the array will contain less then 'level' members. This could be
6603 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6604 * in other functions.
6606 * We reset it to 'level' at the end of this function.
6608 sched_domains_numa_levels = 0;
6610 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6611 if (!sched_domains_numa_masks)
6615 * Now for each level, construct a mask per node which contains all
6616 * cpus of nodes that are that many hops away from us.
6618 for (i = 0; i < level; i++) {
6619 sched_domains_numa_masks[i] =
6620 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6621 if (!sched_domains_numa_masks[i])
6624 for (j = 0; j < nr_node_ids; j++) {
6625 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6629 sched_domains_numa_masks[i][j] = mask;
6632 if (node_distance(j, k) > sched_domains_numa_distance[i])
6635 cpumask_or(mask, mask, cpumask_of_node(k));
6640 /* Compute default topology size */
6641 for (i = 0; sched_domain_topology[i].mask; i++);
6643 tl = kzalloc((i + level + 1) *
6644 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6649 * Copy the default topology bits..
6651 for (i = 0; sched_domain_topology[i].mask; i++)
6652 tl[i] = sched_domain_topology[i];
6655 * .. and append 'j' levels of NUMA goodness.
6657 for (j = 0; j < level; i++, j++) {
6658 tl[i] = (struct sched_domain_topology_level){
6659 .mask = sd_numa_mask,
6660 .sd_flags = cpu_numa_flags,
6661 .flags = SDTL_OVERLAP,
6667 sched_domain_topology = tl;
6669 sched_domains_numa_levels = level;
6670 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6672 init_numa_topology_type();
6675 static void sched_domains_numa_masks_set(unsigned int cpu)
6677 int node = cpu_to_node(cpu);
6680 for (i = 0; i < sched_domains_numa_levels; i++) {
6681 for (j = 0; j < nr_node_ids; j++) {
6682 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6683 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6688 static void sched_domains_numa_masks_clear(unsigned int cpu)
6692 for (i = 0; i < sched_domains_numa_levels; i++) {
6693 for (j = 0; j < nr_node_ids; j++)
6694 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6699 static inline void sched_init_numa(void) { }
6700 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6701 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6702 #endif /* CONFIG_NUMA */
6704 static int __sdt_alloc(const struct cpumask *cpu_map)
6706 struct sched_domain_topology_level *tl;
6709 for_each_sd_topology(tl) {
6710 struct sd_data *sdd = &tl->data;
6712 sdd->sd = alloc_percpu(struct sched_domain *);
6716 sdd->sg = alloc_percpu(struct sched_group *);
6720 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6724 for_each_cpu(j, cpu_map) {
6725 struct sched_domain *sd;
6726 struct sched_group *sg;
6727 struct sched_group_capacity *sgc;
6729 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6730 GFP_KERNEL, cpu_to_node(j));
6734 *per_cpu_ptr(sdd->sd, j) = sd;
6736 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6737 GFP_KERNEL, cpu_to_node(j));
6743 *per_cpu_ptr(sdd->sg, j) = sg;
6745 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6746 GFP_KERNEL, cpu_to_node(j));
6750 *per_cpu_ptr(sdd->sgc, j) = sgc;
6757 static void __sdt_free(const struct cpumask *cpu_map)
6759 struct sched_domain_topology_level *tl;
6762 for_each_sd_topology(tl) {
6763 struct sd_data *sdd = &tl->data;
6765 for_each_cpu(j, cpu_map) {
6766 struct sched_domain *sd;
6769 sd = *per_cpu_ptr(sdd->sd, j);
6770 if (sd && (sd->flags & SD_OVERLAP))
6771 free_sched_groups(sd->groups, 0);
6772 kfree(*per_cpu_ptr(sdd->sd, j));
6776 kfree(*per_cpu_ptr(sdd->sg, j));
6778 kfree(*per_cpu_ptr(sdd->sgc, j));
6780 free_percpu(sdd->sd);
6782 free_percpu(sdd->sg);
6784 free_percpu(sdd->sgc);
6789 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6790 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6791 struct sched_domain *child, int cpu)
6793 struct sched_domain *sd = sd_init(tl, cpu);
6797 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6799 sd->level = child->level + 1;
6800 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6804 if (!cpumask_subset(sched_domain_span(child),
6805 sched_domain_span(sd))) {
6806 pr_err("BUG: arch topology borken\n");
6807 #ifdef CONFIG_SCHED_DEBUG
6808 pr_err(" the %s domain not a subset of the %s domain\n",
6809 child->name, sd->name);
6811 /* Fixup, ensure @sd has at least @child cpus. */
6812 cpumask_or(sched_domain_span(sd),
6813 sched_domain_span(sd),
6814 sched_domain_span(child));
6818 set_domain_attribute(sd, attr);
6824 * Build sched domains for a given set of cpus and attach the sched domains
6825 * to the individual cpus
6827 static int build_sched_domains(const struct cpumask *cpu_map,
6828 struct sched_domain_attr *attr)
6830 enum s_alloc alloc_state;
6831 struct sched_domain *sd;
6833 int i, ret = -ENOMEM;
6835 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6836 if (alloc_state != sa_rootdomain)
6839 /* Set up domains for cpus specified by the cpu_map. */
6840 for_each_cpu(i, cpu_map) {
6841 struct sched_domain_topology_level *tl;
6844 for_each_sd_topology(tl) {
6845 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6846 if (tl == sched_domain_topology)
6847 *per_cpu_ptr(d.sd, i) = sd;
6848 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6849 sd->flags |= SD_OVERLAP;
6850 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6855 /* Build the groups for the domains */
6856 for_each_cpu(i, cpu_map) {
6857 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6858 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6859 if (sd->flags & SD_OVERLAP) {
6860 if (build_overlap_sched_groups(sd, i))
6863 if (build_sched_groups(sd, i))
6869 /* Calculate CPU capacity for physical packages and nodes */
6870 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6871 if (!cpumask_test_cpu(i, cpu_map))
6874 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6875 claim_allocations(i, sd);
6876 init_sched_groups_capacity(i, sd);
6880 /* Attach the domains */
6882 for_each_cpu(i, cpu_map) {
6883 sd = *per_cpu_ptr(d.sd, i);
6884 cpu_attach_domain(sd, d.rd, i);
6890 __free_domain_allocs(&d, alloc_state, cpu_map);
6894 static cpumask_var_t *doms_cur; /* current sched domains */
6895 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6896 static struct sched_domain_attr *dattr_cur;
6897 /* attribues of custom domains in 'doms_cur' */
6900 * Special case: If a kmalloc of a doms_cur partition (array of
6901 * cpumask) fails, then fallback to a single sched domain,
6902 * as determined by the single cpumask fallback_doms.
6904 static cpumask_var_t fallback_doms;
6907 * arch_update_cpu_topology lets virtualized architectures update the
6908 * cpu core maps. It is supposed to return 1 if the topology changed
6909 * or 0 if it stayed the same.
6911 int __weak arch_update_cpu_topology(void)
6916 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6919 cpumask_var_t *doms;
6921 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6924 for (i = 0; i < ndoms; i++) {
6925 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6926 free_sched_domains(doms, i);
6933 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6936 for (i = 0; i < ndoms; i++)
6937 free_cpumask_var(doms[i]);
6942 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6943 * For now this just excludes isolated cpus, but could be used to
6944 * exclude other special cases in the future.
6946 static int init_sched_domains(const struct cpumask *cpu_map)
6950 arch_update_cpu_topology();
6952 doms_cur = alloc_sched_domains(ndoms_cur);
6954 doms_cur = &fallback_doms;
6955 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6956 err = build_sched_domains(doms_cur[0], NULL);
6957 register_sched_domain_sysctl();
6963 * Detach sched domains from a group of cpus specified in cpu_map
6964 * These cpus will now be attached to the NULL domain
6966 static void detach_destroy_domains(const struct cpumask *cpu_map)
6971 for_each_cpu(i, cpu_map)
6972 cpu_attach_domain(NULL, &def_root_domain, i);
6976 /* handle null as "default" */
6977 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6978 struct sched_domain_attr *new, int idx_new)
6980 struct sched_domain_attr tmp;
6987 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6988 new ? (new + idx_new) : &tmp,
6989 sizeof(struct sched_domain_attr));
6993 * Partition sched domains as specified by the 'ndoms_new'
6994 * cpumasks in the array doms_new[] of cpumasks. This compares
6995 * doms_new[] to the current sched domain partitioning, doms_cur[].
6996 * It destroys each deleted domain and builds each new domain.
6998 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6999 * The masks don't intersect (don't overlap.) We should setup one
7000 * sched domain for each mask. CPUs not in any of the cpumasks will
7001 * not be load balanced. If the same cpumask appears both in the
7002 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7005 * The passed in 'doms_new' should be allocated using
7006 * alloc_sched_domains. This routine takes ownership of it and will
7007 * free_sched_domains it when done with it. If the caller failed the
7008 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7009 * and partition_sched_domains() will fallback to the single partition
7010 * 'fallback_doms', it also forces the domains to be rebuilt.
7012 * If doms_new == NULL it will be replaced with cpu_online_mask.
7013 * ndoms_new == 0 is a special case for destroying existing domains,
7014 * and it will not create the default domain.
7016 * Call with hotplug lock held
7018 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7019 struct sched_domain_attr *dattr_new)
7024 mutex_lock(&sched_domains_mutex);
7026 /* always unregister in case we don't destroy any domains */
7027 unregister_sched_domain_sysctl();
7029 /* Let architecture update cpu core mappings. */
7030 new_topology = arch_update_cpu_topology();
7032 n = doms_new ? ndoms_new : 0;
7034 /* Destroy deleted domains */
7035 for (i = 0; i < ndoms_cur; i++) {
7036 for (j = 0; j < n && !new_topology; j++) {
7037 if (cpumask_equal(doms_cur[i], doms_new[j])
7038 && dattrs_equal(dattr_cur, i, dattr_new, j))
7041 /* no match - a current sched domain not in new doms_new[] */
7042 detach_destroy_domains(doms_cur[i]);
7048 if (doms_new == NULL) {
7050 doms_new = &fallback_doms;
7051 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7052 WARN_ON_ONCE(dattr_new);
7055 /* Build new domains */
7056 for (i = 0; i < ndoms_new; i++) {
7057 for (j = 0; j < n && !new_topology; j++) {
7058 if (cpumask_equal(doms_new[i], doms_cur[j])
7059 && dattrs_equal(dattr_new, i, dattr_cur, j))
7062 /* no match - add a new doms_new */
7063 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7068 /* Remember the new sched domains */
7069 if (doms_cur != &fallback_doms)
7070 free_sched_domains(doms_cur, ndoms_cur);
7071 kfree(dattr_cur); /* kfree(NULL) is safe */
7072 doms_cur = doms_new;
7073 dattr_cur = dattr_new;
7074 ndoms_cur = ndoms_new;
7076 register_sched_domain_sysctl();
7078 mutex_unlock(&sched_domains_mutex);
7081 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7084 * Update cpusets according to cpu_active mask. If cpusets are
7085 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7086 * around partition_sched_domains().
7088 * If we come here as part of a suspend/resume, don't touch cpusets because we
7089 * want to restore it back to its original state upon resume anyway.
7091 static void cpuset_cpu_active(void)
7093 if (cpuhp_tasks_frozen) {
7095 * num_cpus_frozen tracks how many CPUs are involved in suspend
7096 * resume sequence. As long as this is not the last online
7097 * operation in the resume sequence, just build a single sched
7098 * domain, ignoring cpusets.
7101 if (likely(num_cpus_frozen)) {
7102 partition_sched_domains(1, NULL, NULL);
7106 * This is the last CPU online operation. So fall through and
7107 * restore the original sched domains by considering the
7108 * cpuset configurations.
7111 cpuset_update_active_cpus(true);
7114 static int cpuset_cpu_inactive(unsigned int cpu)
7116 unsigned long flags;
7121 if (!cpuhp_tasks_frozen) {
7122 rcu_read_lock_sched();
7123 dl_b = dl_bw_of(cpu);
7125 raw_spin_lock_irqsave(&dl_b->lock, flags);
7126 cpus = dl_bw_cpus(cpu);
7127 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7128 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7130 rcu_read_unlock_sched();
7134 cpuset_update_active_cpus(false);
7137 partition_sched_domains(1, NULL, NULL);
7142 int sched_cpu_activate(unsigned int cpu)
7144 struct rq *rq = cpu_rq(cpu);
7145 unsigned long flags;
7147 set_cpu_active(cpu, true);
7149 if (sched_smp_initialized) {
7150 sched_domains_numa_masks_set(cpu);
7151 cpuset_cpu_active();
7155 * Put the rq online, if not already. This happens:
7157 * 1) In the early boot process, because we build the real domains
7158 * after all cpus have been brought up.
7160 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7163 raw_spin_lock_irqsave(&rq->lock, flags);
7165 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7168 raw_spin_unlock_irqrestore(&rq->lock, flags);
7170 update_max_interval();
7175 int sched_cpu_deactivate(unsigned int cpu)
7179 set_cpu_active(cpu, false);
7181 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7182 * users of this state to go away such that all new such users will
7185 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7186 * not imply sync_sched(), so wait for both.
7188 * Do sync before park smpboot threads to take care the rcu boost case.
7190 if (IS_ENABLED(CONFIG_PREEMPT))
7191 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7195 if (!sched_smp_initialized)
7198 ret = cpuset_cpu_inactive(cpu);
7200 set_cpu_active(cpu, true);
7203 sched_domains_numa_masks_clear(cpu);
7207 static void sched_rq_cpu_starting(unsigned int cpu)
7209 struct rq *rq = cpu_rq(cpu);
7211 rq->calc_load_update = calc_load_update;
7212 account_reset_rq(rq);
7213 update_max_interval();
7216 int sched_cpu_starting(unsigned int cpu)
7218 set_cpu_rq_start_time(cpu);
7219 sched_rq_cpu_starting(cpu);
7223 #ifdef CONFIG_HOTPLUG_CPU
7224 int sched_cpu_dying(unsigned int cpu)
7226 struct rq *rq = cpu_rq(cpu);
7227 unsigned long flags;
7229 /* Handle pending wakeups and then migrate everything off */
7230 sched_ttwu_pending();
7231 raw_spin_lock_irqsave(&rq->lock, flags);
7233 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7237 BUG_ON(rq->nr_running != 1);
7238 raw_spin_unlock_irqrestore(&rq->lock, flags);
7239 calc_load_migrate(rq);
7240 update_max_interval();
7241 nohz_balance_exit_idle(cpu);
7247 void __init sched_init_smp(void)
7249 cpumask_var_t non_isolated_cpus;
7251 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7252 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7257 * There's no userspace yet to cause hotplug operations; hence all the
7258 * cpu masks are stable and all blatant races in the below code cannot
7261 mutex_lock(&sched_domains_mutex);
7262 init_sched_domains(cpu_active_mask);
7263 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7264 if (cpumask_empty(non_isolated_cpus))
7265 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7266 mutex_unlock(&sched_domains_mutex);
7268 /* Move init over to a non-isolated CPU */
7269 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7271 sched_init_granularity();
7272 free_cpumask_var(non_isolated_cpus);
7274 init_sched_rt_class();
7275 init_sched_dl_class();
7276 sched_smp_initialized = true;
7279 static int __init migration_init(void)
7281 sched_rq_cpu_starting(smp_processor_id());
7284 early_initcall(migration_init);
7287 void __init sched_init_smp(void)
7289 sched_init_granularity();
7291 #endif /* CONFIG_SMP */
7293 int in_sched_functions(unsigned long addr)
7295 return in_lock_functions(addr) ||
7296 (addr >= (unsigned long)__sched_text_start
7297 && addr < (unsigned long)__sched_text_end);
7300 #ifdef CONFIG_CGROUP_SCHED
7302 * Default task group.
7303 * Every task in system belongs to this group at bootup.
7305 struct task_group root_task_group;
7306 LIST_HEAD(task_groups);
7308 /* Cacheline aligned slab cache for task_group */
7309 static struct kmem_cache *task_group_cache __read_mostly;
7312 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7314 void __init sched_init(void)
7317 unsigned long alloc_size = 0, ptr;
7319 #ifdef CONFIG_FAIR_GROUP_SCHED
7320 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7322 #ifdef CONFIG_RT_GROUP_SCHED
7323 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7326 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7328 #ifdef CONFIG_FAIR_GROUP_SCHED
7329 root_task_group.se = (struct sched_entity **)ptr;
7330 ptr += nr_cpu_ids * sizeof(void **);
7332 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7333 ptr += nr_cpu_ids * sizeof(void **);
7335 #endif /* CONFIG_FAIR_GROUP_SCHED */
7336 #ifdef CONFIG_RT_GROUP_SCHED
7337 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7338 ptr += nr_cpu_ids * sizeof(void **);
7340 root_task_group.rt_rq = (struct rt_rq **)ptr;
7341 ptr += nr_cpu_ids * sizeof(void **);
7343 #endif /* CONFIG_RT_GROUP_SCHED */
7345 #ifdef CONFIG_CPUMASK_OFFSTACK
7346 for_each_possible_cpu(i) {
7347 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7348 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7350 #endif /* CONFIG_CPUMASK_OFFSTACK */
7352 init_rt_bandwidth(&def_rt_bandwidth,
7353 global_rt_period(), global_rt_runtime());
7354 init_dl_bandwidth(&def_dl_bandwidth,
7355 global_rt_period(), global_rt_runtime());
7358 init_defrootdomain();
7361 #ifdef CONFIG_RT_GROUP_SCHED
7362 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7363 global_rt_period(), global_rt_runtime());
7364 #endif /* CONFIG_RT_GROUP_SCHED */
7366 #ifdef CONFIG_CGROUP_SCHED
7367 task_group_cache = KMEM_CACHE(task_group, 0);
7369 list_add(&root_task_group.list, &task_groups);
7370 INIT_LIST_HEAD(&root_task_group.children);
7371 INIT_LIST_HEAD(&root_task_group.siblings);
7372 autogroup_init(&init_task);
7373 #endif /* CONFIG_CGROUP_SCHED */
7375 for_each_possible_cpu(i) {
7379 raw_spin_lock_init(&rq->lock);
7381 rq->calc_load_active = 0;
7382 rq->calc_load_update = jiffies + LOAD_FREQ;
7383 init_cfs_rq(&rq->cfs);
7384 init_rt_rq(&rq->rt);
7385 init_dl_rq(&rq->dl);
7386 #ifdef CONFIG_FAIR_GROUP_SCHED
7387 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7388 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7390 * How much cpu bandwidth does root_task_group get?
7392 * In case of task-groups formed thr' the cgroup filesystem, it
7393 * gets 100% of the cpu resources in the system. This overall
7394 * system cpu resource is divided among the tasks of
7395 * root_task_group and its child task-groups in a fair manner,
7396 * based on each entity's (task or task-group's) weight
7397 * (se->load.weight).
7399 * In other words, if root_task_group has 10 tasks of weight
7400 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7401 * then A0's share of the cpu resource is:
7403 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7405 * We achieve this by letting root_task_group's tasks sit
7406 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7408 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7409 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7410 #endif /* CONFIG_FAIR_GROUP_SCHED */
7412 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7413 #ifdef CONFIG_RT_GROUP_SCHED
7414 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7417 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7418 rq->cpu_load[j] = 0;
7423 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7424 rq->balance_callback = NULL;
7425 rq->active_balance = 0;
7426 rq->next_balance = jiffies;
7431 rq->avg_idle = 2*sysctl_sched_migration_cost;
7432 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7434 INIT_LIST_HEAD(&rq->cfs_tasks);
7436 rq_attach_root(rq, &def_root_domain);
7437 #ifdef CONFIG_NO_HZ_COMMON
7438 rq->last_load_update_tick = jiffies;
7441 #ifdef CONFIG_NO_HZ_FULL
7442 rq->last_sched_tick = 0;
7444 #endif /* CONFIG_SMP */
7446 atomic_set(&rq->nr_iowait, 0);
7449 set_load_weight(&init_task);
7451 #ifdef CONFIG_PREEMPT_NOTIFIERS
7452 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7456 * The boot idle thread does lazy MMU switching as well:
7458 atomic_inc(&init_mm.mm_count);
7459 enter_lazy_tlb(&init_mm, current);
7462 * During early bootup we pretend to be a normal task:
7464 current->sched_class = &fair_sched_class;
7467 * Make us the idle thread. Technically, schedule() should not be
7468 * called from this thread, however somewhere below it might be,
7469 * but because we are the idle thread, we just pick up running again
7470 * when this runqueue becomes "idle".
7472 init_idle(current, smp_processor_id());
7474 calc_load_update = jiffies + LOAD_FREQ;
7477 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7478 /* May be allocated at isolcpus cmdline parse time */
7479 if (cpu_isolated_map == NULL)
7480 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7481 idle_thread_set_boot_cpu();
7482 set_cpu_rq_start_time(smp_processor_id());
7484 init_sched_fair_class();
7486 scheduler_running = 1;
7489 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7490 static inline int preempt_count_equals(int preempt_offset)
7492 int nested = preempt_count() + rcu_preempt_depth();
7494 return (nested == preempt_offset);
7497 void __might_sleep(const char *file, int line, int preempt_offset)
7500 * Blocking primitives will set (and therefore destroy) current->state,
7501 * since we will exit with TASK_RUNNING make sure we enter with it,
7502 * otherwise we will destroy state.
7504 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7505 "do not call blocking ops when !TASK_RUNNING; "
7506 "state=%lx set at [<%p>] %pS\n",
7508 (void *)current->task_state_change,
7509 (void *)current->task_state_change);
7511 ___might_sleep(file, line, preempt_offset);
7513 EXPORT_SYMBOL(__might_sleep);
7515 void ___might_sleep(const char *file, int line, int preempt_offset)
7517 static unsigned long prev_jiffy; /* ratelimiting */
7519 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7520 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7521 !is_idle_task(current)) ||
7522 system_state != SYSTEM_RUNNING || oops_in_progress)
7524 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7526 prev_jiffy = jiffies;
7529 "BUG: sleeping function called from invalid context at %s:%d\n",
7532 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7533 in_atomic(), irqs_disabled(),
7534 current->pid, current->comm);
7536 if (task_stack_end_corrupted(current))
7537 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7539 debug_show_held_locks(current);
7540 if (irqs_disabled())
7541 print_irqtrace_events(current);
7542 #ifdef CONFIG_DEBUG_PREEMPT
7543 if (!preempt_count_equals(preempt_offset)) {
7544 pr_err("Preemption disabled at:");
7545 print_ip_sym(current->preempt_disable_ip);
7551 EXPORT_SYMBOL(___might_sleep);
7554 #ifdef CONFIG_MAGIC_SYSRQ
7555 void normalize_rt_tasks(void)
7557 struct task_struct *g, *p;
7558 struct sched_attr attr = {
7559 .sched_policy = SCHED_NORMAL,
7562 read_lock(&tasklist_lock);
7563 for_each_process_thread(g, p) {
7565 * Only normalize user tasks:
7567 if (p->flags & PF_KTHREAD)
7570 p->se.exec_start = 0;
7571 #ifdef CONFIG_SCHEDSTATS
7572 p->se.statistics.wait_start = 0;
7573 p->se.statistics.sleep_start = 0;
7574 p->se.statistics.block_start = 0;
7577 if (!dl_task(p) && !rt_task(p)) {
7579 * Renice negative nice level userspace
7582 if (task_nice(p) < 0)
7583 set_user_nice(p, 0);
7587 __sched_setscheduler(p, &attr, false, false);
7589 read_unlock(&tasklist_lock);
7592 #endif /* CONFIG_MAGIC_SYSRQ */
7594 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7596 * These functions are only useful for the IA64 MCA handling, or kdb.
7598 * They can only be called when the whole system has been
7599 * stopped - every CPU needs to be quiescent, and no scheduling
7600 * activity can take place. Using them for anything else would
7601 * be a serious bug, and as a result, they aren't even visible
7602 * under any other configuration.
7606 * curr_task - return the current task for a given cpu.
7607 * @cpu: the processor in question.
7609 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7611 * Return: The current task for @cpu.
7613 struct task_struct *curr_task(int cpu)
7615 return cpu_curr(cpu);
7618 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7622 * set_curr_task - set the current task for a given cpu.
7623 * @cpu: the processor in question.
7624 * @p: the task pointer to set.
7626 * Description: This function must only be used when non-maskable interrupts
7627 * are serviced on a separate stack. It allows the architecture to switch the
7628 * notion of the current task on a cpu in a non-blocking manner. This function
7629 * must be called with all CPU's synchronized, and interrupts disabled, the
7630 * and caller must save the original value of the current task (see
7631 * curr_task() above) and restore that value before reenabling interrupts and
7632 * re-starting the system.
7634 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7636 void set_curr_task(int cpu, struct task_struct *p)
7643 #ifdef CONFIG_CGROUP_SCHED
7644 /* task_group_lock serializes the addition/removal of task groups */
7645 static DEFINE_SPINLOCK(task_group_lock);
7647 static void sched_free_group(struct task_group *tg)
7649 free_fair_sched_group(tg);
7650 free_rt_sched_group(tg);
7652 kmem_cache_free(task_group_cache, tg);
7655 /* allocate runqueue etc for a new task group */
7656 struct task_group *sched_create_group(struct task_group *parent)
7658 struct task_group *tg;
7660 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7662 return ERR_PTR(-ENOMEM);
7664 if (!alloc_fair_sched_group(tg, parent))
7667 if (!alloc_rt_sched_group(tg, parent))
7673 sched_free_group(tg);
7674 return ERR_PTR(-ENOMEM);
7677 void sched_online_group(struct task_group *tg, struct task_group *parent)
7679 unsigned long flags;
7681 spin_lock_irqsave(&task_group_lock, flags);
7682 list_add_rcu(&tg->list, &task_groups);
7684 WARN_ON(!parent); /* root should already exist */
7686 tg->parent = parent;
7687 INIT_LIST_HEAD(&tg->children);
7688 list_add_rcu(&tg->siblings, &parent->children);
7689 spin_unlock_irqrestore(&task_group_lock, flags);
7692 /* rcu callback to free various structures associated with a task group */
7693 static void sched_free_group_rcu(struct rcu_head *rhp)
7695 /* now it should be safe to free those cfs_rqs */
7696 sched_free_group(container_of(rhp, struct task_group, rcu));
7699 void sched_destroy_group(struct task_group *tg)
7701 /* wait for possible concurrent references to cfs_rqs complete */
7702 call_rcu(&tg->rcu, sched_free_group_rcu);
7705 void sched_offline_group(struct task_group *tg)
7707 unsigned long flags;
7709 /* end participation in shares distribution */
7710 unregister_fair_sched_group(tg);
7712 spin_lock_irqsave(&task_group_lock, flags);
7713 list_del_rcu(&tg->list);
7714 list_del_rcu(&tg->siblings);
7715 spin_unlock_irqrestore(&task_group_lock, flags);
7718 /* change task's runqueue when it moves between groups.
7719 * The caller of this function should have put the task in its new group
7720 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7721 * reflect its new group.
7723 void sched_move_task(struct task_struct *tsk)
7725 struct task_group *tg;
7726 int queued, running;
7730 rq = task_rq_lock(tsk, &rf);
7732 running = task_current(rq, tsk);
7733 queued = task_on_rq_queued(tsk);
7736 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7737 if (unlikely(running))
7738 put_prev_task(rq, tsk);
7741 * All callers are synchronized by task_rq_lock(); we do not use RCU
7742 * which is pointless here. Thus, we pass "true" to task_css_check()
7743 * to prevent lockdep warnings.
7745 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7746 struct task_group, css);
7747 tg = autogroup_task_group(tsk, tg);
7748 tsk->sched_task_group = tg;
7750 #ifdef CONFIG_FAIR_GROUP_SCHED
7751 if (tsk->sched_class->task_move_group)
7752 tsk->sched_class->task_move_group(tsk);
7755 set_task_rq(tsk, task_cpu(tsk));
7757 if (unlikely(running))
7758 tsk->sched_class->set_curr_task(rq);
7760 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7762 task_rq_unlock(rq, tsk, &rf);
7764 #endif /* CONFIG_CGROUP_SCHED */
7766 #ifdef CONFIG_RT_GROUP_SCHED
7768 * Ensure that the real time constraints are schedulable.
7770 static DEFINE_MUTEX(rt_constraints_mutex);
7772 /* Must be called with tasklist_lock held */
7773 static inline int tg_has_rt_tasks(struct task_group *tg)
7775 struct task_struct *g, *p;
7778 * Autogroups do not have RT tasks; see autogroup_create().
7780 if (task_group_is_autogroup(tg))
7783 for_each_process_thread(g, p) {
7784 if (rt_task(p) && task_group(p) == tg)
7791 struct rt_schedulable_data {
7792 struct task_group *tg;
7797 static int tg_rt_schedulable(struct task_group *tg, void *data)
7799 struct rt_schedulable_data *d = data;
7800 struct task_group *child;
7801 unsigned long total, sum = 0;
7802 u64 period, runtime;
7804 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7805 runtime = tg->rt_bandwidth.rt_runtime;
7808 period = d->rt_period;
7809 runtime = d->rt_runtime;
7813 * Cannot have more runtime than the period.
7815 if (runtime > period && runtime != RUNTIME_INF)
7819 * Ensure we don't starve existing RT tasks.
7821 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7824 total = to_ratio(period, runtime);
7827 * Nobody can have more than the global setting allows.
7829 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7833 * The sum of our children's runtime should not exceed our own.
7835 list_for_each_entry_rcu(child, &tg->children, siblings) {
7836 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7837 runtime = child->rt_bandwidth.rt_runtime;
7839 if (child == d->tg) {
7840 period = d->rt_period;
7841 runtime = d->rt_runtime;
7844 sum += to_ratio(period, runtime);
7853 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7857 struct rt_schedulable_data data = {
7859 .rt_period = period,
7860 .rt_runtime = runtime,
7864 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7870 static int tg_set_rt_bandwidth(struct task_group *tg,
7871 u64 rt_period, u64 rt_runtime)
7876 * Disallowing the root group RT runtime is BAD, it would disallow the
7877 * kernel creating (and or operating) RT threads.
7879 if (tg == &root_task_group && rt_runtime == 0)
7882 /* No period doesn't make any sense. */
7886 mutex_lock(&rt_constraints_mutex);
7887 read_lock(&tasklist_lock);
7888 err = __rt_schedulable(tg, rt_period, rt_runtime);
7892 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7893 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7894 tg->rt_bandwidth.rt_runtime = rt_runtime;
7896 for_each_possible_cpu(i) {
7897 struct rt_rq *rt_rq = tg->rt_rq[i];
7899 raw_spin_lock(&rt_rq->rt_runtime_lock);
7900 rt_rq->rt_runtime = rt_runtime;
7901 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7903 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7905 read_unlock(&tasklist_lock);
7906 mutex_unlock(&rt_constraints_mutex);
7911 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7913 u64 rt_runtime, rt_period;
7915 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7916 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7917 if (rt_runtime_us < 0)
7918 rt_runtime = RUNTIME_INF;
7920 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7923 static long sched_group_rt_runtime(struct task_group *tg)
7927 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7930 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7931 do_div(rt_runtime_us, NSEC_PER_USEC);
7932 return rt_runtime_us;
7935 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7937 u64 rt_runtime, rt_period;
7939 rt_period = rt_period_us * NSEC_PER_USEC;
7940 rt_runtime = tg->rt_bandwidth.rt_runtime;
7942 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7945 static long sched_group_rt_period(struct task_group *tg)
7949 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7950 do_div(rt_period_us, NSEC_PER_USEC);
7951 return rt_period_us;
7953 #endif /* CONFIG_RT_GROUP_SCHED */
7955 #ifdef CONFIG_RT_GROUP_SCHED
7956 static int sched_rt_global_constraints(void)
7960 mutex_lock(&rt_constraints_mutex);
7961 read_lock(&tasklist_lock);
7962 ret = __rt_schedulable(NULL, 0, 0);
7963 read_unlock(&tasklist_lock);
7964 mutex_unlock(&rt_constraints_mutex);
7969 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7971 /* Don't accept realtime tasks when there is no way for them to run */
7972 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7978 #else /* !CONFIG_RT_GROUP_SCHED */
7979 static int sched_rt_global_constraints(void)
7981 unsigned long flags;
7984 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7985 for_each_possible_cpu(i) {
7986 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7988 raw_spin_lock(&rt_rq->rt_runtime_lock);
7989 rt_rq->rt_runtime = global_rt_runtime();
7990 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7992 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7996 #endif /* CONFIG_RT_GROUP_SCHED */
7998 static int sched_dl_global_validate(void)
8000 u64 runtime = global_rt_runtime();
8001 u64 period = global_rt_period();
8002 u64 new_bw = to_ratio(period, runtime);
8005 unsigned long flags;
8008 * Here we want to check the bandwidth not being set to some
8009 * value smaller than the currently allocated bandwidth in
8010 * any of the root_domains.
8012 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8013 * cycling on root_domains... Discussion on different/better
8014 * solutions is welcome!
8016 for_each_possible_cpu(cpu) {
8017 rcu_read_lock_sched();
8018 dl_b = dl_bw_of(cpu);
8020 raw_spin_lock_irqsave(&dl_b->lock, flags);
8021 if (new_bw < dl_b->total_bw)
8023 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8025 rcu_read_unlock_sched();
8034 static void sched_dl_do_global(void)
8039 unsigned long flags;
8041 def_dl_bandwidth.dl_period = global_rt_period();
8042 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8044 if (global_rt_runtime() != RUNTIME_INF)
8045 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8048 * FIXME: As above...
8050 for_each_possible_cpu(cpu) {
8051 rcu_read_lock_sched();
8052 dl_b = dl_bw_of(cpu);
8054 raw_spin_lock_irqsave(&dl_b->lock, flags);
8056 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8058 rcu_read_unlock_sched();
8062 static int sched_rt_global_validate(void)
8064 if (sysctl_sched_rt_period <= 0)
8067 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8068 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8074 static void sched_rt_do_global(void)
8076 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8077 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8080 int sched_rt_handler(struct ctl_table *table, int write,
8081 void __user *buffer, size_t *lenp,
8084 int old_period, old_runtime;
8085 static DEFINE_MUTEX(mutex);
8089 old_period = sysctl_sched_rt_period;
8090 old_runtime = sysctl_sched_rt_runtime;
8092 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8094 if (!ret && write) {
8095 ret = sched_rt_global_validate();
8099 ret = sched_dl_global_validate();
8103 ret = sched_rt_global_constraints();
8107 sched_rt_do_global();
8108 sched_dl_do_global();
8112 sysctl_sched_rt_period = old_period;
8113 sysctl_sched_rt_runtime = old_runtime;
8115 mutex_unlock(&mutex);
8120 int sched_rr_handler(struct ctl_table *table, int write,
8121 void __user *buffer, size_t *lenp,
8125 static DEFINE_MUTEX(mutex);
8128 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8129 /* make sure that internally we keep jiffies */
8130 /* also, writing zero resets timeslice to default */
8131 if (!ret && write) {
8132 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8133 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8135 mutex_unlock(&mutex);
8139 #ifdef CONFIG_CGROUP_SCHED
8141 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8143 return css ? container_of(css, struct task_group, css) : NULL;
8146 static struct cgroup_subsys_state *
8147 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8149 struct task_group *parent = css_tg(parent_css);
8150 struct task_group *tg;
8153 /* This is early initialization for the top cgroup */
8154 return &root_task_group.css;
8157 tg = sched_create_group(parent);
8159 return ERR_PTR(-ENOMEM);
8161 sched_online_group(tg, parent);
8166 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8168 struct task_group *tg = css_tg(css);
8170 sched_offline_group(tg);
8173 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8175 struct task_group *tg = css_tg(css);
8178 * Relies on the RCU grace period between css_released() and this.
8180 sched_free_group(tg);
8183 static void cpu_cgroup_fork(struct task_struct *task)
8185 sched_move_task(task);
8188 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8190 struct task_struct *task;
8191 struct cgroup_subsys_state *css;
8193 cgroup_taskset_for_each(task, css, tset) {
8194 #ifdef CONFIG_RT_GROUP_SCHED
8195 if (!sched_rt_can_attach(css_tg(css), task))
8198 /* We don't support RT-tasks being in separate groups */
8199 if (task->sched_class != &fair_sched_class)
8206 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8208 struct task_struct *task;
8209 struct cgroup_subsys_state *css;
8211 cgroup_taskset_for_each(task, css, tset)
8212 sched_move_task(task);
8215 #ifdef CONFIG_FAIR_GROUP_SCHED
8216 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8217 struct cftype *cftype, u64 shareval)
8219 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8222 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8225 struct task_group *tg = css_tg(css);
8227 return (u64) scale_load_down(tg->shares);
8230 #ifdef CONFIG_CFS_BANDWIDTH
8231 static DEFINE_MUTEX(cfs_constraints_mutex);
8233 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8234 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8236 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8238 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8240 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8241 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8243 if (tg == &root_task_group)
8247 * Ensure we have at some amount of bandwidth every period. This is
8248 * to prevent reaching a state of large arrears when throttled via
8249 * entity_tick() resulting in prolonged exit starvation.
8251 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8255 * Likewise, bound things on the otherside by preventing insane quota
8256 * periods. This also allows us to normalize in computing quota
8259 if (period > max_cfs_quota_period)
8263 * Prevent race between setting of cfs_rq->runtime_enabled and
8264 * unthrottle_offline_cfs_rqs().
8267 mutex_lock(&cfs_constraints_mutex);
8268 ret = __cfs_schedulable(tg, period, quota);
8272 runtime_enabled = quota != RUNTIME_INF;
8273 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8275 * If we need to toggle cfs_bandwidth_used, off->on must occur
8276 * before making related changes, and on->off must occur afterwards
8278 if (runtime_enabled && !runtime_was_enabled)
8279 cfs_bandwidth_usage_inc();
8280 raw_spin_lock_irq(&cfs_b->lock);
8281 cfs_b->period = ns_to_ktime(period);
8282 cfs_b->quota = quota;
8284 __refill_cfs_bandwidth_runtime(cfs_b);
8285 /* restart the period timer (if active) to handle new period expiry */
8286 if (runtime_enabled)
8287 start_cfs_bandwidth(cfs_b);
8288 raw_spin_unlock_irq(&cfs_b->lock);
8290 for_each_online_cpu(i) {
8291 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8292 struct rq *rq = cfs_rq->rq;
8294 raw_spin_lock_irq(&rq->lock);
8295 cfs_rq->runtime_enabled = runtime_enabled;
8296 cfs_rq->runtime_remaining = 0;
8298 if (cfs_rq->throttled)
8299 unthrottle_cfs_rq(cfs_rq);
8300 raw_spin_unlock_irq(&rq->lock);
8302 if (runtime_was_enabled && !runtime_enabled)
8303 cfs_bandwidth_usage_dec();
8305 mutex_unlock(&cfs_constraints_mutex);
8311 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8315 period = ktime_to_ns(tg->cfs_bandwidth.period);
8316 if (cfs_quota_us < 0)
8317 quota = RUNTIME_INF;
8319 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8321 return tg_set_cfs_bandwidth(tg, period, quota);
8324 long tg_get_cfs_quota(struct task_group *tg)
8328 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8331 quota_us = tg->cfs_bandwidth.quota;
8332 do_div(quota_us, NSEC_PER_USEC);
8337 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8341 period = (u64)cfs_period_us * NSEC_PER_USEC;
8342 quota = tg->cfs_bandwidth.quota;
8344 return tg_set_cfs_bandwidth(tg, period, quota);
8347 long tg_get_cfs_period(struct task_group *tg)
8351 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8352 do_div(cfs_period_us, NSEC_PER_USEC);
8354 return cfs_period_us;
8357 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8360 return tg_get_cfs_quota(css_tg(css));
8363 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8364 struct cftype *cftype, s64 cfs_quota_us)
8366 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8369 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8372 return tg_get_cfs_period(css_tg(css));
8375 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8376 struct cftype *cftype, u64 cfs_period_us)
8378 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8381 struct cfs_schedulable_data {
8382 struct task_group *tg;
8387 * normalize group quota/period to be quota/max_period
8388 * note: units are usecs
8390 static u64 normalize_cfs_quota(struct task_group *tg,
8391 struct cfs_schedulable_data *d)
8399 period = tg_get_cfs_period(tg);
8400 quota = tg_get_cfs_quota(tg);
8403 /* note: these should typically be equivalent */
8404 if (quota == RUNTIME_INF || quota == -1)
8407 return to_ratio(period, quota);
8410 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8412 struct cfs_schedulable_data *d = data;
8413 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8414 s64 quota = 0, parent_quota = -1;
8417 quota = RUNTIME_INF;
8419 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8421 quota = normalize_cfs_quota(tg, d);
8422 parent_quota = parent_b->hierarchical_quota;
8425 * ensure max(child_quota) <= parent_quota, inherit when no
8428 if (quota == RUNTIME_INF)
8429 quota = parent_quota;
8430 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8433 cfs_b->hierarchical_quota = quota;
8438 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8441 struct cfs_schedulable_data data = {
8447 if (quota != RUNTIME_INF) {
8448 do_div(data.period, NSEC_PER_USEC);
8449 do_div(data.quota, NSEC_PER_USEC);
8453 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8459 static int cpu_stats_show(struct seq_file *sf, void *v)
8461 struct task_group *tg = css_tg(seq_css(sf));
8462 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8464 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8465 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8466 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8470 #endif /* CONFIG_CFS_BANDWIDTH */
8471 #endif /* CONFIG_FAIR_GROUP_SCHED */
8473 #ifdef CONFIG_RT_GROUP_SCHED
8474 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8475 struct cftype *cft, s64 val)
8477 return sched_group_set_rt_runtime(css_tg(css), val);
8480 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8483 return sched_group_rt_runtime(css_tg(css));
8486 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8487 struct cftype *cftype, u64 rt_period_us)
8489 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8492 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8495 return sched_group_rt_period(css_tg(css));
8497 #endif /* CONFIG_RT_GROUP_SCHED */
8499 static struct cftype cpu_files[] = {
8500 #ifdef CONFIG_FAIR_GROUP_SCHED
8503 .read_u64 = cpu_shares_read_u64,
8504 .write_u64 = cpu_shares_write_u64,
8507 #ifdef CONFIG_CFS_BANDWIDTH
8509 .name = "cfs_quota_us",
8510 .read_s64 = cpu_cfs_quota_read_s64,
8511 .write_s64 = cpu_cfs_quota_write_s64,
8514 .name = "cfs_period_us",
8515 .read_u64 = cpu_cfs_period_read_u64,
8516 .write_u64 = cpu_cfs_period_write_u64,
8520 .seq_show = cpu_stats_show,
8523 #ifdef CONFIG_RT_GROUP_SCHED
8525 .name = "rt_runtime_us",
8526 .read_s64 = cpu_rt_runtime_read,
8527 .write_s64 = cpu_rt_runtime_write,
8530 .name = "rt_period_us",
8531 .read_u64 = cpu_rt_period_read_uint,
8532 .write_u64 = cpu_rt_period_write_uint,
8538 struct cgroup_subsys cpu_cgrp_subsys = {
8539 .css_alloc = cpu_cgroup_css_alloc,
8540 .css_released = cpu_cgroup_css_released,
8541 .css_free = cpu_cgroup_css_free,
8542 .fork = cpu_cgroup_fork,
8543 .can_attach = cpu_cgroup_can_attach,
8544 .attach = cpu_cgroup_attach,
8545 .legacy_cftypes = cpu_files,
8549 #endif /* CONFIG_CGROUP_SCHED */
8551 void dump_cpu_task(int cpu)
8553 pr_info("Task dump for CPU %d:\n", cpu);
8554 sched_show_task(cpu_curr(cpu));
8558 * Nice levels are multiplicative, with a gentle 10% change for every
8559 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8560 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8561 * that remained on nice 0.
8563 * The "10% effect" is relative and cumulative: from _any_ nice level,
8564 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8565 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8566 * If a task goes up by ~10% and another task goes down by ~10% then
8567 * the relative distance between them is ~25%.)
8569 const int sched_prio_to_weight[40] = {
8570 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8571 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8572 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8573 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8574 /* 0 */ 1024, 820, 655, 526, 423,
8575 /* 5 */ 335, 272, 215, 172, 137,
8576 /* 10 */ 110, 87, 70, 56, 45,
8577 /* 15 */ 36, 29, 23, 18, 15,
8581 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8583 * In cases where the weight does not change often, we can use the
8584 * precalculated inverse to speed up arithmetics by turning divisions
8585 * into multiplications:
8587 const u32 sched_prio_to_wmult[40] = {
8588 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8589 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8590 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8591 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8592 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8593 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8594 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8595 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,