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;
318 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
320 int cpu = (int)(long)hcpu;
323 case CPU_UP_CANCELED:
324 case CPU_UP_CANCELED_FROZEN:
325 case CPU_DOWN_PREPARE:
326 case CPU_DOWN_PREPARE_FROZEN:
328 case CPU_DEAD_FROZEN:
329 hrtick_clear(cpu_rq(cpu));
336 static __init void init_hrtick(void)
338 hotcpu_notifier(hotplug_hrtick, 0);
342 * Called to set the hrtick timer state.
344 * called with rq->lock held and irqs disabled
346 void hrtick_start(struct rq *rq, u64 delay)
349 * Don't schedule slices shorter than 10000ns, that just
350 * doesn't make sense. Rely on vruntime for fairness.
352 delay = max_t(u64, delay, 10000LL);
353 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
354 HRTIMER_MODE_REL_PINNED);
357 static inline void init_hrtick(void)
360 #endif /* CONFIG_SMP */
362 static void init_rq_hrtick(struct rq *rq)
365 rq->hrtick_csd_pending = 0;
367 rq->hrtick_csd.flags = 0;
368 rq->hrtick_csd.func = __hrtick_start;
369 rq->hrtick_csd.info = rq;
372 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
373 rq->hrtick_timer.function = hrtick;
375 #else /* CONFIG_SCHED_HRTICK */
376 static inline void hrtick_clear(struct rq *rq)
380 static inline void init_rq_hrtick(struct rq *rq)
384 static inline void init_hrtick(void)
387 #endif /* CONFIG_SCHED_HRTICK */
390 * cmpxchg based fetch_or, macro so it works for different integer types
392 #define fetch_or(ptr, mask) \
394 typeof(ptr) _ptr = (ptr); \
395 typeof(mask) _mask = (mask); \
396 typeof(*_ptr) _old, _val = *_ptr; \
399 _old = cmpxchg(_ptr, _val, _val | _mask); \
407 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
409 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
410 * this avoids any races wrt polling state changes and thereby avoids
413 static bool set_nr_and_not_polling(struct task_struct *p)
415 struct thread_info *ti = task_thread_info(p);
416 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
420 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
422 * If this returns true, then the idle task promises to call
423 * sched_ttwu_pending() and reschedule soon.
425 static bool set_nr_if_polling(struct task_struct *p)
427 struct thread_info *ti = task_thread_info(p);
428 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
431 if (!(val & _TIF_POLLING_NRFLAG))
433 if (val & _TIF_NEED_RESCHED)
435 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
444 static bool set_nr_and_not_polling(struct task_struct *p)
446 set_tsk_need_resched(p);
451 static bool set_nr_if_polling(struct task_struct *p)
458 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
460 struct wake_q_node *node = &task->wake_q;
463 * Atomically grab the task, if ->wake_q is !nil already it means
464 * its already queued (either by us or someone else) and will get the
465 * wakeup due to that.
467 * This cmpxchg() implies a full barrier, which pairs with the write
468 * barrier implied by the wakeup in wake_up_q().
470 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
473 get_task_struct(task);
476 * The head is context local, there can be no concurrency.
479 head->lastp = &node->next;
482 void wake_up_q(struct wake_q_head *head)
484 struct wake_q_node *node = head->first;
486 while (node != WAKE_Q_TAIL) {
487 struct task_struct *task;
489 task = container_of(node, struct task_struct, wake_q);
491 /* task can safely be re-inserted now */
493 task->wake_q.next = NULL;
496 * wake_up_process() implies a wmb() to pair with the queueing
497 * in wake_q_add() so as not to miss wakeups.
499 wake_up_process(task);
500 put_task_struct(task);
505 * resched_curr - mark rq's current task 'to be rescheduled now'.
507 * On UP this means the setting of the need_resched flag, on SMP it
508 * might also involve a cross-CPU call to trigger the scheduler on
511 void resched_curr(struct rq *rq)
513 struct task_struct *curr = rq->curr;
516 lockdep_assert_held(&rq->lock);
518 if (test_tsk_need_resched(curr))
523 if (cpu == smp_processor_id()) {
524 set_tsk_need_resched(curr);
525 set_preempt_need_resched();
529 if (set_nr_and_not_polling(curr))
530 smp_send_reschedule(cpu);
532 trace_sched_wake_idle_without_ipi(cpu);
535 void resched_cpu(int cpu)
537 struct rq *rq = cpu_rq(cpu);
540 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543 raw_spin_unlock_irqrestore(&rq->lock, flags);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int i, cpu = smp_processor_id();
559 struct sched_domain *sd;
561 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
565 for_each_domain(cpu, sd) {
566 for_each_cpu(i, sched_domain_span(sd)) {
567 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
574 if (!is_housekeeping_cpu(cpu))
575 cpu = housekeeping_any_cpu();
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 static void wake_up_idle_cpu(int cpu)
592 struct rq *rq = cpu_rq(cpu);
594 if (cpu == smp_processor_id())
597 if (set_nr_and_not_polling(rq->idle))
598 smp_send_reschedule(cpu);
600 trace_sched_wake_idle_without_ipi(cpu);
603 static bool wake_up_full_nohz_cpu(int cpu)
606 * We just need the target to call irq_exit() and re-evaluate
607 * the next tick. The nohz full kick at least implies that.
608 * If needed we can still optimize that later with an
611 if (tick_nohz_full_cpu(cpu)) {
612 if (cpu != smp_processor_id() ||
613 tick_nohz_tick_stopped())
614 tick_nohz_full_kick_cpu(cpu);
621 void wake_up_nohz_cpu(int cpu)
623 if (!wake_up_full_nohz_cpu(cpu))
624 wake_up_idle_cpu(cpu);
627 static inline bool got_nohz_idle_kick(void)
629 int cpu = smp_processor_id();
631 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
634 if (idle_cpu(cpu) && !need_resched())
638 * We can't run Idle Load Balance on this CPU for this time so we
639 * cancel it and clear NOHZ_BALANCE_KICK
641 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
645 #else /* CONFIG_NO_HZ_COMMON */
647 static inline bool got_nohz_idle_kick(void)
652 #endif /* CONFIG_NO_HZ_COMMON */
654 #ifdef CONFIG_NO_HZ_FULL
655 bool sched_can_stop_tick(struct rq *rq)
659 /* Deadline tasks, even if single, need the tick */
660 if (rq->dl.dl_nr_running)
664 * If there are more than one RR tasks, we need the tick to effect the
665 * actual RR behaviour.
667 if (rq->rt.rr_nr_running) {
668 if (rq->rt.rr_nr_running == 1)
675 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
676 * forced preemption between FIFO tasks.
678 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
683 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
684 * if there's more than one we need the tick for involuntary
687 if (rq->nr_running > 1)
692 #endif /* CONFIG_NO_HZ_FULL */
694 void sched_avg_update(struct rq *rq)
696 s64 period = sched_avg_period();
698 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
700 * Inline assembly required to prevent the compiler
701 * optimising this loop into a divmod call.
702 * See __iter_div_u64_rem() for another example of this.
704 asm("" : "+rm" (rq->age_stamp));
705 rq->age_stamp += period;
710 #endif /* CONFIG_SMP */
712 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
713 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
715 * Iterate task_group tree rooted at *from, calling @down when first entering a
716 * node and @up when leaving it for the final time.
718 * Caller must hold rcu_lock or sufficient equivalent.
720 int walk_tg_tree_from(struct task_group *from,
721 tg_visitor down, tg_visitor up, void *data)
723 struct task_group *parent, *child;
729 ret = (*down)(parent, data);
732 list_for_each_entry_rcu(child, &parent->children, siblings) {
739 ret = (*up)(parent, data);
740 if (ret || parent == from)
744 parent = parent->parent;
751 int tg_nop(struct task_group *tg, void *data)
757 static void set_load_weight(struct task_struct *p)
759 int prio = p->static_prio - MAX_RT_PRIO;
760 struct load_weight *load = &p->se.load;
763 * SCHED_IDLE tasks get minimal weight:
765 if (idle_policy(p->policy)) {
766 load->weight = scale_load(WEIGHT_IDLEPRIO);
767 load->inv_weight = WMULT_IDLEPRIO;
771 load->weight = scale_load(sched_prio_to_weight[prio]);
772 load->inv_weight = sched_prio_to_wmult[prio];
775 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
778 if (!(flags & ENQUEUE_RESTORE))
779 sched_info_queued(rq, p);
780 p->sched_class->enqueue_task(rq, p, flags);
783 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
786 if (!(flags & DEQUEUE_SAVE))
787 sched_info_dequeued(rq, p);
788 p->sched_class->dequeue_task(rq, p, flags);
791 void activate_task(struct rq *rq, struct task_struct *p, int flags)
793 if (task_contributes_to_load(p))
794 rq->nr_uninterruptible--;
796 enqueue_task(rq, p, flags);
799 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
801 if (task_contributes_to_load(p))
802 rq->nr_uninterruptible++;
804 dequeue_task(rq, p, flags);
807 static void update_rq_clock_task(struct rq *rq, s64 delta)
810 * In theory, the compile should just see 0 here, and optimize out the call
811 * to sched_rt_avg_update. But I don't trust it...
813 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
814 s64 steal = 0, irq_delta = 0;
816 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
817 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
820 * Since irq_time is only updated on {soft,}irq_exit, we might run into
821 * this case when a previous update_rq_clock() happened inside a
824 * When this happens, we stop ->clock_task and only update the
825 * prev_irq_time stamp to account for the part that fit, so that a next
826 * update will consume the rest. This ensures ->clock_task is
829 * It does however cause some slight miss-attribution of {soft,}irq
830 * time, a more accurate solution would be to update the irq_time using
831 * the current rq->clock timestamp, except that would require using
834 if (irq_delta > delta)
837 rq->prev_irq_time += irq_delta;
840 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
841 if (static_key_false((¶virt_steal_rq_enabled))) {
842 steal = paravirt_steal_clock(cpu_of(rq));
843 steal -= rq->prev_steal_time_rq;
845 if (unlikely(steal > delta))
848 rq->prev_steal_time_rq += steal;
853 rq->clock_task += delta;
855 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
856 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
857 sched_rt_avg_update(rq, irq_delta + steal);
861 void sched_set_stop_task(int cpu, struct task_struct *stop)
863 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
864 struct task_struct *old_stop = cpu_rq(cpu)->stop;
868 * Make it appear like a SCHED_FIFO task, its something
869 * userspace knows about and won't get confused about.
871 * Also, it will make PI more or less work without too
872 * much confusion -- but then, stop work should not
873 * rely on PI working anyway.
875 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
877 stop->sched_class = &stop_sched_class;
880 cpu_rq(cpu)->stop = stop;
884 * Reset it back to a normal scheduling class so that
885 * it can die in pieces.
887 old_stop->sched_class = &rt_sched_class;
892 * __normal_prio - return the priority that is based on the static prio
894 static inline int __normal_prio(struct task_struct *p)
896 return p->static_prio;
900 * Calculate the expected normal priority: i.e. priority
901 * without taking RT-inheritance into account. Might be
902 * boosted by interactivity modifiers. Changes upon fork,
903 * setprio syscalls, and whenever the interactivity
904 * estimator recalculates.
906 static inline int normal_prio(struct task_struct *p)
910 if (task_has_dl_policy(p))
911 prio = MAX_DL_PRIO-1;
912 else if (task_has_rt_policy(p))
913 prio = MAX_RT_PRIO-1 - p->rt_priority;
915 prio = __normal_prio(p);
920 * Calculate the current priority, i.e. the priority
921 * taken into account by the scheduler. This value might
922 * be boosted by RT tasks, or might be boosted by
923 * interactivity modifiers. Will be RT if the task got
924 * RT-boosted. If not then it returns p->normal_prio.
926 static int effective_prio(struct task_struct *p)
928 p->normal_prio = normal_prio(p);
930 * If we are RT tasks or we were boosted to RT priority,
931 * keep the priority unchanged. Otherwise, update priority
932 * to the normal priority:
934 if (!rt_prio(p->prio))
935 return p->normal_prio;
940 * task_curr - is this task currently executing on a CPU?
941 * @p: the task in question.
943 * Return: 1 if the task is currently executing. 0 otherwise.
945 inline int task_curr(const struct task_struct *p)
947 return cpu_curr(task_cpu(p)) == p;
951 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
952 * use the balance_callback list if you want balancing.
954 * this means any call to check_class_changed() must be followed by a call to
955 * balance_callback().
957 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
958 const struct sched_class *prev_class,
961 if (prev_class != p->sched_class) {
962 if (prev_class->switched_from)
963 prev_class->switched_from(rq, p);
965 p->sched_class->switched_to(rq, p);
966 } else if (oldprio != p->prio || dl_task(p))
967 p->sched_class->prio_changed(rq, p, oldprio);
970 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
972 const struct sched_class *class;
974 if (p->sched_class == rq->curr->sched_class) {
975 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
977 for_each_class(class) {
978 if (class == rq->curr->sched_class)
980 if (class == p->sched_class) {
988 * A queue event has occurred, and we're going to schedule. In
989 * this case, we can save a useless back to back clock update.
991 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
992 rq_clock_skip_update(rq, true);
997 * This is how migration works:
999 * 1) we invoke migration_cpu_stop() on the target CPU using
1001 * 2) stopper starts to run (implicitly forcing the migrated thread
1003 * 3) it checks whether the migrated task is still in the wrong runqueue.
1004 * 4) if it's in the wrong runqueue then the migration thread removes
1005 * it and puts it into the right queue.
1006 * 5) stopper completes and stop_one_cpu() returns and the migration
1011 * move_queued_task - move a queued task to new rq.
1013 * Returns (locked) new rq. Old rq's lock is released.
1015 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1017 lockdep_assert_held(&rq->lock);
1019 p->on_rq = TASK_ON_RQ_MIGRATING;
1020 dequeue_task(rq, p, 0);
1021 set_task_cpu(p, new_cpu);
1022 raw_spin_unlock(&rq->lock);
1024 rq = cpu_rq(new_cpu);
1026 raw_spin_lock(&rq->lock);
1027 BUG_ON(task_cpu(p) != new_cpu);
1028 enqueue_task(rq, p, 0);
1029 p->on_rq = TASK_ON_RQ_QUEUED;
1030 check_preempt_curr(rq, p, 0);
1035 struct migration_arg {
1036 struct task_struct *task;
1041 * Move (not current) task off this cpu, onto dest cpu. We're doing
1042 * this because either it can't run here any more (set_cpus_allowed()
1043 * away from this CPU, or CPU going down), or because we're
1044 * attempting to rebalance this task on exec (sched_exec).
1046 * So we race with normal scheduler movements, but that's OK, as long
1047 * as the task is no longer on this CPU.
1049 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1051 if (unlikely(!cpu_active(dest_cpu)))
1054 /* Affinity changed (again). */
1055 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1058 rq = move_queued_task(rq, p, dest_cpu);
1064 * migration_cpu_stop - this will be executed by a highprio stopper thread
1065 * and performs thread migration by bumping thread off CPU then
1066 * 'pushing' onto another runqueue.
1068 static int migration_cpu_stop(void *data)
1070 struct migration_arg *arg = data;
1071 struct task_struct *p = arg->task;
1072 struct rq *rq = this_rq();
1075 * The original target cpu might have gone down and we might
1076 * be on another cpu but it doesn't matter.
1078 local_irq_disable();
1080 * We need to explicitly wake pending tasks before running
1081 * __migrate_task() such that we will not miss enforcing cpus_allowed
1082 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1084 sched_ttwu_pending();
1086 raw_spin_lock(&p->pi_lock);
1087 raw_spin_lock(&rq->lock);
1089 * If task_rq(p) != rq, it cannot be migrated here, because we're
1090 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1091 * we're holding p->pi_lock.
1093 if (task_rq(p) == rq && task_on_rq_queued(p))
1094 rq = __migrate_task(rq, p, arg->dest_cpu);
1095 raw_spin_unlock(&rq->lock);
1096 raw_spin_unlock(&p->pi_lock);
1103 * sched_class::set_cpus_allowed must do the below, but is not required to
1104 * actually call this function.
1106 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1108 cpumask_copy(&p->cpus_allowed, new_mask);
1109 p->nr_cpus_allowed = cpumask_weight(new_mask);
1112 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1114 struct rq *rq = task_rq(p);
1115 bool queued, running;
1117 lockdep_assert_held(&p->pi_lock);
1119 queued = task_on_rq_queued(p);
1120 running = task_current(rq, p);
1124 * Because __kthread_bind() calls this on blocked tasks without
1127 lockdep_assert_held(&rq->lock);
1128 dequeue_task(rq, p, DEQUEUE_SAVE);
1131 put_prev_task(rq, p);
1133 p->sched_class->set_cpus_allowed(p, new_mask);
1136 p->sched_class->set_curr_task(rq);
1138 enqueue_task(rq, p, ENQUEUE_RESTORE);
1142 * Change a given task's CPU affinity. Migrate the thread to a
1143 * proper CPU and schedule it away if the CPU it's executing on
1144 * is removed from the allowed bitmask.
1146 * NOTE: the caller must have a valid reference to the task, the
1147 * task must not exit() & deallocate itself prematurely. The
1148 * call is not atomic; no spinlocks may be held.
1150 static int __set_cpus_allowed_ptr(struct task_struct *p,
1151 const struct cpumask *new_mask, bool check)
1153 unsigned int dest_cpu;
1158 rq = task_rq_lock(p, &rf);
1161 * Must re-check here, to close a race against __kthread_bind(),
1162 * sched_setaffinity() is not guaranteed to observe the flag.
1164 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1169 if (cpumask_equal(&p->cpus_allowed, new_mask))
1172 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1177 do_set_cpus_allowed(p, new_mask);
1179 /* Can the task run on the task's current CPU? If so, we're done */
1180 if (cpumask_test_cpu(task_cpu(p), new_mask))
1183 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1184 if (task_running(rq, p) || p->state == TASK_WAKING) {
1185 struct migration_arg arg = { p, dest_cpu };
1186 /* Need help from migration thread: drop lock and wait. */
1187 task_rq_unlock(rq, p, &rf);
1188 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1189 tlb_migrate_finish(p->mm);
1191 } else if (task_on_rq_queued(p)) {
1193 * OK, since we're going to drop the lock immediately
1194 * afterwards anyway.
1196 lockdep_unpin_lock(&rq->lock, rf.cookie);
1197 rq = move_queued_task(rq, p, dest_cpu);
1198 lockdep_repin_lock(&rq->lock, rf.cookie);
1201 task_rq_unlock(rq, p, &rf);
1206 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1208 return __set_cpus_allowed_ptr(p, new_mask, false);
1210 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1212 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1214 #ifdef CONFIG_SCHED_DEBUG
1216 * We should never call set_task_cpu() on a blocked task,
1217 * ttwu() will sort out the placement.
1219 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1223 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1224 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1225 * time relying on p->on_rq.
1227 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1228 p->sched_class == &fair_sched_class &&
1229 (p->on_rq && !task_on_rq_migrating(p)));
1231 #ifdef CONFIG_LOCKDEP
1233 * The caller should hold either p->pi_lock or rq->lock, when changing
1234 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1236 * sched_move_task() holds both and thus holding either pins the cgroup,
1239 * Furthermore, all task_rq users should acquire both locks, see
1242 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1243 lockdep_is_held(&task_rq(p)->lock)));
1247 trace_sched_migrate_task(p, new_cpu);
1249 if (task_cpu(p) != new_cpu) {
1250 if (p->sched_class->migrate_task_rq)
1251 p->sched_class->migrate_task_rq(p);
1252 p->se.nr_migrations++;
1253 perf_event_task_migrate(p);
1256 __set_task_cpu(p, new_cpu);
1259 static void __migrate_swap_task(struct task_struct *p, int cpu)
1261 if (task_on_rq_queued(p)) {
1262 struct rq *src_rq, *dst_rq;
1264 src_rq = task_rq(p);
1265 dst_rq = cpu_rq(cpu);
1267 p->on_rq = TASK_ON_RQ_MIGRATING;
1268 deactivate_task(src_rq, p, 0);
1269 set_task_cpu(p, cpu);
1270 activate_task(dst_rq, p, 0);
1271 p->on_rq = TASK_ON_RQ_QUEUED;
1272 check_preempt_curr(dst_rq, p, 0);
1275 * Task isn't running anymore; make it appear like we migrated
1276 * it before it went to sleep. This means on wakeup we make the
1277 * previous cpu our targer instead of where it really is.
1283 struct migration_swap_arg {
1284 struct task_struct *src_task, *dst_task;
1285 int src_cpu, dst_cpu;
1288 static int migrate_swap_stop(void *data)
1290 struct migration_swap_arg *arg = data;
1291 struct rq *src_rq, *dst_rq;
1294 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1297 src_rq = cpu_rq(arg->src_cpu);
1298 dst_rq = cpu_rq(arg->dst_cpu);
1300 double_raw_lock(&arg->src_task->pi_lock,
1301 &arg->dst_task->pi_lock);
1302 double_rq_lock(src_rq, dst_rq);
1304 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1307 if (task_cpu(arg->src_task) != arg->src_cpu)
1310 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1313 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1316 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1317 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1322 double_rq_unlock(src_rq, dst_rq);
1323 raw_spin_unlock(&arg->dst_task->pi_lock);
1324 raw_spin_unlock(&arg->src_task->pi_lock);
1330 * Cross migrate two tasks
1332 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1334 struct migration_swap_arg arg;
1337 arg = (struct migration_swap_arg){
1339 .src_cpu = task_cpu(cur),
1341 .dst_cpu = task_cpu(p),
1344 if (arg.src_cpu == arg.dst_cpu)
1348 * These three tests are all lockless; this is OK since all of them
1349 * will be re-checked with proper locks held further down the line.
1351 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1354 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1357 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1360 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1361 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1368 * wait_task_inactive - wait for a thread to unschedule.
1370 * If @match_state is nonzero, it's the @p->state value just checked and
1371 * not expected to change. If it changes, i.e. @p might have woken up,
1372 * then return zero. When we succeed in waiting for @p to be off its CPU,
1373 * we return a positive number (its total switch count). If a second call
1374 * a short while later returns the same number, the caller can be sure that
1375 * @p has remained unscheduled the whole time.
1377 * The caller must ensure that the task *will* unschedule sometime soon,
1378 * else this function might spin for a *long* time. This function can't
1379 * be called with interrupts off, or it may introduce deadlock with
1380 * smp_call_function() if an IPI is sent by the same process we are
1381 * waiting to become inactive.
1383 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1385 int running, queued;
1392 * We do the initial early heuristics without holding
1393 * any task-queue locks at all. We'll only try to get
1394 * the runqueue lock when things look like they will
1400 * If the task is actively running on another CPU
1401 * still, just relax and busy-wait without holding
1404 * NOTE! Since we don't hold any locks, it's not
1405 * even sure that "rq" stays as the right runqueue!
1406 * But we don't care, since "task_running()" will
1407 * return false if the runqueue has changed and p
1408 * is actually now running somewhere else!
1410 while (task_running(rq, p)) {
1411 if (match_state && unlikely(p->state != match_state))
1417 * Ok, time to look more closely! We need the rq
1418 * lock now, to be *sure*. If we're wrong, we'll
1419 * just go back and repeat.
1421 rq = task_rq_lock(p, &rf);
1422 trace_sched_wait_task(p);
1423 running = task_running(rq, p);
1424 queued = task_on_rq_queued(p);
1426 if (!match_state || p->state == match_state)
1427 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1428 task_rq_unlock(rq, p, &rf);
1431 * If it changed from the expected state, bail out now.
1433 if (unlikely(!ncsw))
1437 * Was it really running after all now that we
1438 * checked with the proper locks actually held?
1440 * Oops. Go back and try again..
1442 if (unlikely(running)) {
1448 * It's not enough that it's not actively running,
1449 * it must be off the runqueue _entirely_, and not
1452 * So if it was still runnable (but just not actively
1453 * running right now), it's preempted, and we should
1454 * yield - it could be a while.
1456 if (unlikely(queued)) {
1457 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1459 set_current_state(TASK_UNINTERRUPTIBLE);
1460 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1465 * Ahh, all good. It wasn't running, and it wasn't
1466 * runnable, which means that it will never become
1467 * running in the future either. We're all done!
1476 * kick_process - kick a running thread to enter/exit the kernel
1477 * @p: the to-be-kicked thread
1479 * Cause a process which is running on another CPU to enter
1480 * kernel-mode, without any delay. (to get signals handled.)
1482 * NOTE: this function doesn't have to take the runqueue lock,
1483 * because all it wants to ensure is that the remote task enters
1484 * the kernel. If the IPI races and the task has been migrated
1485 * to another CPU then no harm is done and the purpose has been
1488 void kick_process(struct task_struct *p)
1494 if ((cpu != smp_processor_id()) && task_curr(p))
1495 smp_send_reschedule(cpu);
1498 EXPORT_SYMBOL_GPL(kick_process);
1501 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1503 static int select_fallback_rq(int cpu, struct task_struct *p)
1505 int nid = cpu_to_node(cpu);
1506 const struct cpumask *nodemask = NULL;
1507 enum { cpuset, possible, fail } state = cpuset;
1511 * If the node that the cpu is on has been offlined, cpu_to_node()
1512 * will return -1. There is no cpu on the node, and we should
1513 * select the cpu on the other node.
1516 nodemask = cpumask_of_node(nid);
1518 /* Look for allowed, online CPU in same node. */
1519 for_each_cpu(dest_cpu, nodemask) {
1520 if (!cpu_online(dest_cpu))
1522 if (!cpu_active(dest_cpu))
1524 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1530 /* Any allowed, online CPU? */
1531 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1532 if (!cpu_online(dest_cpu))
1534 if (!cpu_active(dest_cpu))
1539 /* No more Mr. Nice Guy. */
1542 if (IS_ENABLED(CONFIG_CPUSETS)) {
1543 cpuset_cpus_allowed_fallback(p);
1549 do_set_cpus_allowed(p, cpu_possible_mask);
1560 if (state != cpuset) {
1562 * Don't tell them about moving exiting tasks or
1563 * kernel threads (both mm NULL), since they never
1566 if (p->mm && printk_ratelimit()) {
1567 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1568 task_pid_nr(p), p->comm, cpu);
1576 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1579 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1581 lockdep_assert_held(&p->pi_lock);
1583 if (p->nr_cpus_allowed > 1)
1584 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1587 * In order not to call set_task_cpu() on a blocking task we need
1588 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1591 * Since this is common to all placement strategies, this lives here.
1593 * [ this allows ->select_task() to simply return task_cpu(p) and
1594 * not worry about this generic constraint ]
1596 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1598 cpu = select_fallback_rq(task_cpu(p), p);
1603 static void update_avg(u64 *avg, u64 sample)
1605 s64 diff = sample - *avg;
1611 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1612 const struct cpumask *new_mask, bool check)
1614 return set_cpus_allowed_ptr(p, new_mask);
1617 #endif /* CONFIG_SMP */
1620 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1622 #ifdef CONFIG_SCHEDSTATS
1623 struct rq *rq = this_rq();
1626 int this_cpu = smp_processor_id();
1628 if (cpu == this_cpu) {
1629 schedstat_inc(rq, ttwu_local);
1630 schedstat_inc(p, se.statistics.nr_wakeups_local);
1632 struct sched_domain *sd;
1634 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1636 for_each_domain(this_cpu, sd) {
1637 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1638 schedstat_inc(sd, ttwu_wake_remote);
1645 if (wake_flags & WF_MIGRATED)
1646 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1648 #endif /* CONFIG_SMP */
1650 schedstat_inc(rq, ttwu_count);
1651 schedstat_inc(p, se.statistics.nr_wakeups);
1653 if (wake_flags & WF_SYNC)
1654 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1656 #endif /* CONFIG_SCHEDSTATS */
1659 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1661 activate_task(rq, p, en_flags);
1662 p->on_rq = TASK_ON_RQ_QUEUED;
1664 /* if a worker is waking up, notify workqueue */
1665 if (p->flags & PF_WQ_WORKER)
1666 wq_worker_waking_up(p, cpu_of(rq));
1670 * Mark the task runnable and perform wakeup-preemption.
1672 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1673 struct pin_cookie cookie)
1675 check_preempt_curr(rq, p, wake_flags);
1676 p->state = TASK_RUNNING;
1677 trace_sched_wakeup(p);
1680 if (p->sched_class->task_woken) {
1682 * Our task @p is fully woken up and running; so its safe to
1683 * drop the rq->lock, hereafter rq is only used for statistics.
1685 lockdep_unpin_lock(&rq->lock, cookie);
1686 p->sched_class->task_woken(rq, p);
1687 lockdep_repin_lock(&rq->lock, cookie);
1690 if (rq->idle_stamp) {
1691 u64 delta = rq_clock(rq) - rq->idle_stamp;
1692 u64 max = 2*rq->max_idle_balance_cost;
1694 update_avg(&rq->avg_idle, delta);
1696 if (rq->avg_idle > max)
1705 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1706 struct pin_cookie cookie)
1708 lockdep_assert_held(&rq->lock);
1711 if (p->sched_contributes_to_load)
1712 rq->nr_uninterruptible--;
1715 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1716 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1720 * Called in case the task @p isn't fully descheduled from its runqueue,
1721 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1722 * since all we need to do is flip p->state to TASK_RUNNING, since
1723 * the task is still ->on_rq.
1725 static int ttwu_remote(struct task_struct *p, int wake_flags)
1731 rq = __task_rq_lock(p, &rf);
1732 if (task_on_rq_queued(p)) {
1733 /* check_preempt_curr() may use rq clock */
1734 update_rq_clock(rq);
1735 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1738 __task_rq_unlock(rq, &rf);
1744 void sched_ttwu_pending(void)
1746 struct rq *rq = this_rq();
1747 struct llist_node *llist = llist_del_all(&rq->wake_list);
1748 struct pin_cookie cookie;
1749 struct task_struct *p;
1750 unsigned long flags;
1755 raw_spin_lock_irqsave(&rq->lock, flags);
1756 cookie = lockdep_pin_lock(&rq->lock);
1759 p = llist_entry(llist, struct task_struct, wake_entry);
1760 llist = llist_next(llist);
1761 ttwu_do_activate(rq, p, 0, cookie);
1764 lockdep_unpin_lock(&rq->lock, cookie);
1765 raw_spin_unlock_irqrestore(&rq->lock, flags);
1768 void scheduler_ipi(void)
1771 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1772 * TIF_NEED_RESCHED remotely (for the first time) will also send
1775 preempt_fold_need_resched();
1777 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1781 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1782 * traditionally all their work was done from the interrupt return
1783 * path. Now that we actually do some work, we need to make sure
1786 * Some archs already do call them, luckily irq_enter/exit nest
1789 * Arguably we should visit all archs and update all handlers,
1790 * however a fair share of IPIs are still resched only so this would
1791 * somewhat pessimize the simple resched case.
1794 sched_ttwu_pending();
1797 * Check if someone kicked us for doing the nohz idle load balance.
1799 if (unlikely(got_nohz_idle_kick())) {
1800 this_rq()->idle_balance = 1;
1801 raise_softirq_irqoff(SCHED_SOFTIRQ);
1806 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1808 struct rq *rq = cpu_rq(cpu);
1810 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1811 if (!set_nr_if_polling(rq->idle))
1812 smp_send_reschedule(cpu);
1814 trace_sched_wake_idle_without_ipi(cpu);
1818 void wake_up_if_idle(int cpu)
1820 struct rq *rq = cpu_rq(cpu);
1821 unsigned long flags;
1825 if (!is_idle_task(rcu_dereference(rq->curr)))
1828 if (set_nr_if_polling(rq->idle)) {
1829 trace_sched_wake_idle_without_ipi(cpu);
1831 raw_spin_lock_irqsave(&rq->lock, flags);
1832 if (is_idle_task(rq->curr))
1833 smp_send_reschedule(cpu);
1834 /* Else cpu is not in idle, do nothing here */
1835 raw_spin_unlock_irqrestore(&rq->lock, flags);
1842 bool cpus_share_cache(int this_cpu, int that_cpu)
1844 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1846 #endif /* CONFIG_SMP */
1848 static void ttwu_queue(struct task_struct *p, int cpu)
1850 struct rq *rq = cpu_rq(cpu);
1851 struct pin_cookie cookie;
1853 #if defined(CONFIG_SMP)
1854 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1855 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1856 ttwu_queue_remote(p, cpu);
1861 raw_spin_lock(&rq->lock);
1862 cookie = lockdep_pin_lock(&rq->lock);
1863 ttwu_do_activate(rq, p, 0, cookie);
1864 lockdep_unpin_lock(&rq->lock, cookie);
1865 raw_spin_unlock(&rq->lock);
1869 * Notes on Program-Order guarantees on SMP systems.
1873 * The basic program-order guarantee on SMP systems is that when a task [t]
1874 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1875 * execution on its new cpu [c1].
1877 * For migration (of runnable tasks) this is provided by the following means:
1879 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1880 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1881 * rq(c1)->lock (if not at the same time, then in that order).
1882 * C) LOCK of the rq(c1)->lock scheduling in task
1884 * Transitivity guarantees that B happens after A and C after B.
1885 * Note: we only require RCpc transitivity.
1886 * Note: the cpu doing B need not be c0 or c1
1895 * UNLOCK rq(0)->lock
1897 * LOCK rq(0)->lock // orders against CPU0
1899 * UNLOCK rq(0)->lock
1903 * UNLOCK rq(1)->lock
1905 * LOCK rq(1)->lock // orders against CPU2
1908 * UNLOCK rq(1)->lock
1911 * BLOCKING -- aka. SLEEP + WAKEUP
1913 * For blocking we (obviously) need to provide the same guarantee as for
1914 * migration. However the means are completely different as there is no lock
1915 * chain to provide order. Instead we do:
1917 * 1) smp_store_release(X->on_cpu, 0)
1918 * 2) smp_cond_acquire(!X->on_cpu)
1922 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1924 * LOCK rq(0)->lock LOCK X->pi_lock
1927 * smp_store_release(X->on_cpu, 0);
1929 * smp_cond_acquire(!X->on_cpu);
1935 * X->state = RUNNING
1936 * UNLOCK rq(2)->lock
1938 * LOCK rq(2)->lock // orders against CPU1
1941 * UNLOCK rq(2)->lock
1944 * UNLOCK rq(0)->lock
1947 * However; for wakeups there is a second guarantee we must provide, namely we
1948 * must observe the state that lead to our wakeup. That is, not only must our
1949 * task observe its own prior state, it must also observe the stores prior to
1952 * This means that any means of doing remote wakeups must order the CPU doing
1953 * the wakeup against the CPU the task is going to end up running on. This,
1954 * however, is already required for the regular Program-Order guarantee above,
1955 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1960 * try_to_wake_up - wake up a thread
1961 * @p: the thread to be awakened
1962 * @state: the mask of task states that can be woken
1963 * @wake_flags: wake modifier flags (WF_*)
1965 * Put it on the run-queue if it's not already there. The "current"
1966 * thread is always on the run-queue (except when the actual
1967 * re-schedule is in progress), and as such you're allowed to do
1968 * the simpler "current->state = TASK_RUNNING" to mark yourself
1969 * runnable without the overhead of this.
1971 * Return: %true if @p was woken up, %false if it was already running.
1972 * or @state didn't match @p's state.
1975 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1977 unsigned long flags;
1978 int cpu, success = 0;
1981 * If we are going to wake up a thread waiting for CONDITION we
1982 * need to ensure that CONDITION=1 done by the caller can not be
1983 * reordered with p->state check below. This pairs with mb() in
1984 * set_current_state() the waiting thread does.
1986 smp_mb__before_spinlock();
1987 raw_spin_lock_irqsave(&p->pi_lock, flags);
1988 if (!(p->state & state))
1991 trace_sched_waking(p);
1993 success = 1; /* we're going to change ->state */
1996 if (p->on_rq && ttwu_remote(p, wake_flags))
2001 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2002 * possible to, falsely, observe p->on_cpu == 0.
2004 * One must be running (->on_cpu == 1) in order to remove oneself
2005 * from the runqueue.
2007 * [S] ->on_cpu = 1; [L] ->on_rq
2011 * [S] ->on_rq = 0; [L] ->on_cpu
2013 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2014 * from the consecutive calls to schedule(); the first switching to our
2015 * task, the second putting it to sleep.
2020 * If the owning (remote) cpu is still in the middle of schedule() with
2021 * this task as prev, wait until its done referencing the task.
2023 * Pairs with the smp_store_release() in finish_lock_switch().
2025 * This ensures that tasks getting woken will be fully ordered against
2026 * their previous state and preserve Program Order.
2028 smp_cond_acquire(!p->on_cpu);
2030 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2031 p->state = TASK_WAKING;
2033 if (p->sched_class->task_waking)
2034 p->sched_class->task_waking(p);
2036 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2037 if (task_cpu(p) != cpu) {
2038 wake_flags |= WF_MIGRATED;
2039 set_task_cpu(p, cpu);
2041 #endif /* CONFIG_SMP */
2045 if (schedstat_enabled())
2046 ttwu_stat(p, cpu, wake_flags);
2048 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2054 * try_to_wake_up_local - try to wake up a local task with rq lock held
2055 * @p: the thread to be awakened
2057 * Put @p on the run-queue if it's not already there. The caller must
2058 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2061 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2063 struct rq *rq = task_rq(p);
2065 if (WARN_ON_ONCE(rq != this_rq()) ||
2066 WARN_ON_ONCE(p == current))
2069 lockdep_assert_held(&rq->lock);
2071 if (!raw_spin_trylock(&p->pi_lock)) {
2073 * This is OK, because current is on_cpu, which avoids it being
2074 * picked for load-balance and preemption/IRQs are still
2075 * disabled avoiding further scheduler activity on it and we've
2076 * not yet picked a replacement task.
2078 lockdep_unpin_lock(&rq->lock, cookie);
2079 raw_spin_unlock(&rq->lock);
2080 raw_spin_lock(&p->pi_lock);
2081 raw_spin_lock(&rq->lock);
2082 lockdep_repin_lock(&rq->lock, cookie);
2085 if (!(p->state & TASK_NORMAL))
2088 trace_sched_waking(p);
2090 if (!task_on_rq_queued(p))
2091 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2093 ttwu_do_wakeup(rq, p, 0, cookie);
2094 if (schedstat_enabled())
2095 ttwu_stat(p, smp_processor_id(), 0);
2097 raw_spin_unlock(&p->pi_lock);
2101 * wake_up_process - Wake up a specific process
2102 * @p: The process to be woken up.
2104 * Attempt to wake up the nominated process and move it to the set of runnable
2107 * Return: 1 if the process was woken up, 0 if it was already running.
2109 * It may be assumed that this function implies a write memory barrier before
2110 * changing the task state if and only if any tasks are woken up.
2112 int wake_up_process(struct task_struct *p)
2114 return try_to_wake_up(p, TASK_NORMAL, 0);
2116 EXPORT_SYMBOL(wake_up_process);
2118 int wake_up_state(struct task_struct *p, unsigned int state)
2120 return try_to_wake_up(p, state, 0);
2124 * This function clears the sched_dl_entity static params.
2126 void __dl_clear_params(struct task_struct *p)
2128 struct sched_dl_entity *dl_se = &p->dl;
2130 dl_se->dl_runtime = 0;
2131 dl_se->dl_deadline = 0;
2132 dl_se->dl_period = 0;
2136 dl_se->dl_throttled = 0;
2137 dl_se->dl_yielded = 0;
2141 * Perform scheduler related setup for a newly forked process p.
2142 * p is forked by current.
2144 * __sched_fork() is basic setup used by init_idle() too:
2146 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2151 p->se.exec_start = 0;
2152 p->se.sum_exec_runtime = 0;
2153 p->se.prev_sum_exec_runtime = 0;
2154 p->se.nr_migrations = 0;
2156 INIT_LIST_HEAD(&p->se.group_node);
2158 #ifdef CONFIG_FAIR_GROUP_SCHED
2159 p->se.cfs_rq = NULL;
2162 #ifdef CONFIG_SCHEDSTATS
2163 /* Even if schedstat is disabled, there should not be garbage */
2164 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2167 RB_CLEAR_NODE(&p->dl.rb_node);
2168 init_dl_task_timer(&p->dl);
2169 __dl_clear_params(p);
2171 INIT_LIST_HEAD(&p->rt.run_list);
2173 p->rt.time_slice = sched_rr_timeslice;
2177 #ifdef CONFIG_PREEMPT_NOTIFIERS
2178 INIT_HLIST_HEAD(&p->preempt_notifiers);
2181 #ifdef CONFIG_NUMA_BALANCING
2182 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2183 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2184 p->mm->numa_scan_seq = 0;
2187 if (clone_flags & CLONE_VM)
2188 p->numa_preferred_nid = current->numa_preferred_nid;
2190 p->numa_preferred_nid = -1;
2192 p->node_stamp = 0ULL;
2193 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2194 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2195 p->numa_work.next = &p->numa_work;
2196 p->numa_faults = NULL;
2197 p->last_task_numa_placement = 0;
2198 p->last_sum_exec_runtime = 0;
2200 p->numa_group = NULL;
2201 #endif /* CONFIG_NUMA_BALANCING */
2204 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2206 #ifdef CONFIG_NUMA_BALANCING
2208 void set_numabalancing_state(bool enabled)
2211 static_branch_enable(&sched_numa_balancing);
2213 static_branch_disable(&sched_numa_balancing);
2216 #ifdef CONFIG_PROC_SYSCTL
2217 int sysctl_numa_balancing(struct ctl_table *table, int write,
2218 void __user *buffer, size_t *lenp, loff_t *ppos)
2222 int state = static_branch_likely(&sched_numa_balancing);
2224 if (write && !capable(CAP_SYS_ADMIN))
2229 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2233 set_numabalancing_state(state);
2239 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2241 #ifdef CONFIG_SCHEDSTATS
2242 static void set_schedstats(bool enabled)
2245 static_branch_enable(&sched_schedstats);
2247 static_branch_disable(&sched_schedstats);
2250 void force_schedstat_enabled(void)
2252 if (!schedstat_enabled()) {
2253 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2254 static_branch_enable(&sched_schedstats);
2258 static int __init setup_schedstats(char *str)
2264 if (!strcmp(str, "enable")) {
2265 set_schedstats(true);
2267 } else if (!strcmp(str, "disable")) {
2268 set_schedstats(false);
2273 pr_warn("Unable to parse schedstats=\n");
2277 __setup("schedstats=", setup_schedstats);
2279 #ifdef CONFIG_PROC_SYSCTL
2280 int sysctl_schedstats(struct ctl_table *table, int write,
2281 void __user *buffer, size_t *lenp, loff_t *ppos)
2285 int state = static_branch_likely(&sched_schedstats);
2287 if (write && !capable(CAP_SYS_ADMIN))
2292 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2296 set_schedstats(state);
2303 * fork()/clone()-time setup:
2305 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2307 unsigned long flags;
2308 int cpu = get_cpu();
2310 __sched_fork(clone_flags, p);
2312 * We mark the process as running here. This guarantees that
2313 * nobody will actually run it, and a signal or other external
2314 * event cannot wake it up and insert it on the runqueue either.
2316 p->state = TASK_RUNNING;
2319 * Make sure we do not leak PI boosting priority to the child.
2321 p->prio = current->normal_prio;
2324 * Revert to default priority/policy on fork if requested.
2326 if (unlikely(p->sched_reset_on_fork)) {
2327 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2328 p->policy = SCHED_NORMAL;
2329 p->static_prio = NICE_TO_PRIO(0);
2331 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2332 p->static_prio = NICE_TO_PRIO(0);
2334 p->prio = p->normal_prio = __normal_prio(p);
2338 * We don't need the reset flag anymore after the fork. It has
2339 * fulfilled its duty:
2341 p->sched_reset_on_fork = 0;
2344 if (dl_prio(p->prio)) {
2347 } else if (rt_prio(p->prio)) {
2348 p->sched_class = &rt_sched_class;
2350 p->sched_class = &fair_sched_class;
2353 if (p->sched_class->task_fork)
2354 p->sched_class->task_fork(p);
2357 * The child is not yet in the pid-hash so no cgroup attach races,
2358 * and the cgroup is pinned to this child due to cgroup_fork()
2359 * is ran before sched_fork().
2361 * Silence PROVE_RCU.
2363 raw_spin_lock_irqsave(&p->pi_lock, flags);
2364 set_task_cpu(p, cpu);
2365 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2367 #ifdef CONFIG_SCHED_INFO
2368 if (likely(sched_info_on()))
2369 memset(&p->sched_info, 0, sizeof(p->sched_info));
2371 #if defined(CONFIG_SMP)
2374 init_task_preempt_count(p);
2376 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2377 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2384 unsigned long to_ratio(u64 period, u64 runtime)
2386 if (runtime == RUNTIME_INF)
2390 * Doing this here saves a lot of checks in all
2391 * the calling paths, and returning zero seems
2392 * safe for them anyway.
2397 return div64_u64(runtime << 20, period);
2401 inline struct dl_bw *dl_bw_of(int i)
2403 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2404 "sched RCU must be held");
2405 return &cpu_rq(i)->rd->dl_bw;
2408 static inline int dl_bw_cpus(int i)
2410 struct root_domain *rd = cpu_rq(i)->rd;
2413 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2414 "sched RCU must be held");
2415 for_each_cpu_and(i, rd->span, cpu_active_mask)
2421 inline struct dl_bw *dl_bw_of(int i)
2423 return &cpu_rq(i)->dl.dl_bw;
2426 static inline int dl_bw_cpus(int i)
2433 * We must be sure that accepting a new task (or allowing changing the
2434 * parameters of an existing one) is consistent with the bandwidth
2435 * constraints. If yes, this function also accordingly updates the currently
2436 * allocated bandwidth to reflect the new situation.
2438 * This function is called while holding p's rq->lock.
2440 * XXX we should delay bw change until the task's 0-lag point, see
2443 static int dl_overflow(struct task_struct *p, int policy,
2444 const struct sched_attr *attr)
2447 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2448 u64 period = attr->sched_period ?: attr->sched_deadline;
2449 u64 runtime = attr->sched_runtime;
2450 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2453 /* !deadline task may carry old deadline bandwidth */
2454 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2458 * Either if a task, enters, leave, or stays -deadline but changes
2459 * its parameters, we may need to update accordingly the total
2460 * allocated bandwidth of the container.
2462 raw_spin_lock(&dl_b->lock);
2463 cpus = dl_bw_cpus(task_cpu(p));
2464 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2465 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2466 __dl_add(dl_b, new_bw);
2468 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2469 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2470 __dl_clear(dl_b, p->dl.dl_bw);
2471 __dl_add(dl_b, new_bw);
2473 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2474 __dl_clear(dl_b, p->dl.dl_bw);
2477 raw_spin_unlock(&dl_b->lock);
2482 extern void init_dl_bw(struct dl_bw *dl_b);
2485 * wake_up_new_task - wake up a newly created task for the first time.
2487 * This function will do some initial scheduler statistics housekeeping
2488 * that must be done for every newly created context, then puts the task
2489 * on the runqueue and wakes it.
2491 void wake_up_new_task(struct task_struct *p)
2496 /* Initialize new task's runnable average */
2497 init_entity_runnable_average(&p->se);
2498 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2501 * Fork balancing, do it here and not earlier because:
2502 * - cpus_allowed can change in the fork path
2503 * - any previously selected cpu might disappear through hotplug
2505 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2507 /* Post initialize new task's util average when its cfs_rq is set */
2508 post_init_entity_util_avg(&p->se);
2510 rq = __task_rq_lock(p, &rf);
2511 activate_task(rq, p, 0);
2512 p->on_rq = TASK_ON_RQ_QUEUED;
2513 trace_sched_wakeup_new(p);
2514 check_preempt_curr(rq, p, WF_FORK);
2516 if (p->sched_class->task_woken) {
2518 * Nothing relies on rq->lock after this, so its fine to
2521 lockdep_unpin_lock(&rq->lock, rf.cookie);
2522 p->sched_class->task_woken(rq, p);
2523 lockdep_repin_lock(&rq->lock, rf.cookie);
2526 task_rq_unlock(rq, p, &rf);
2529 #ifdef CONFIG_PREEMPT_NOTIFIERS
2531 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2533 void preempt_notifier_inc(void)
2535 static_key_slow_inc(&preempt_notifier_key);
2537 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2539 void preempt_notifier_dec(void)
2541 static_key_slow_dec(&preempt_notifier_key);
2543 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2546 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2547 * @notifier: notifier struct to register
2549 void preempt_notifier_register(struct preempt_notifier *notifier)
2551 if (!static_key_false(&preempt_notifier_key))
2552 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2554 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2556 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2559 * preempt_notifier_unregister - no longer interested in preemption notifications
2560 * @notifier: notifier struct to unregister
2562 * This is *not* safe to call from within a preemption notifier.
2564 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2566 hlist_del(¬ifier->link);
2568 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2570 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2572 struct preempt_notifier *notifier;
2574 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2575 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2578 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2580 if (static_key_false(&preempt_notifier_key))
2581 __fire_sched_in_preempt_notifiers(curr);
2585 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2586 struct task_struct *next)
2588 struct preempt_notifier *notifier;
2590 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2591 notifier->ops->sched_out(notifier, next);
2594 static __always_inline void
2595 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2596 struct task_struct *next)
2598 if (static_key_false(&preempt_notifier_key))
2599 __fire_sched_out_preempt_notifiers(curr, next);
2602 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2604 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2609 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2610 struct task_struct *next)
2614 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2617 * prepare_task_switch - prepare to switch tasks
2618 * @rq: the runqueue preparing to switch
2619 * @prev: the current task that is being switched out
2620 * @next: the task we are going to switch to.
2622 * This is called with the rq lock held and interrupts off. It must
2623 * be paired with a subsequent finish_task_switch after the context
2626 * prepare_task_switch sets up locking and calls architecture specific
2630 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2631 struct task_struct *next)
2633 sched_info_switch(rq, prev, next);
2634 perf_event_task_sched_out(prev, next);
2635 fire_sched_out_preempt_notifiers(prev, next);
2636 prepare_lock_switch(rq, next);
2637 prepare_arch_switch(next);
2641 * finish_task_switch - clean up after a task-switch
2642 * @prev: the thread we just switched away from.
2644 * finish_task_switch must be called after the context switch, paired
2645 * with a prepare_task_switch call before the context switch.
2646 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2647 * and do any other architecture-specific cleanup actions.
2649 * Note that we may have delayed dropping an mm in context_switch(). If
2650 * so, we finish that here outside of the runqueue lock. (Doing it
2651 * with the lock held can cause deadlocks; see schedule() for
2654 * The context switch have flipped the stack from under us and restored the
2655 * local variables which were saved when this task called schedule() in the
2656 * past. prev == current is still correct but we need to recalculate this_rq
2657 * because prev may have moved to another CPU.
2659 static struct rq *finish_task_switch(struct task_struct *prev)
2660 __releases(rq->lock)
2662 struct rq *rq = this_rq();
2663 struct mm_struct *mm = rq->prev_mm;
2667 * The previous task will have left us with a preempt_count of 2
2668 * because it left us after:
2671 * preempt_disable(); // 1
2673 * raw_spin_lock_irq(&rq->lock) // 2
2675 * Also, see FORK_PREEMPT_COUNT.
2677 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2678 "corrupted preempt_count: %s/%d/0x%x\n",
2679 current->comm, current->pid, preempt_count()))
2680 preempt_count_set(FORK_PREEMPT_COUNT);
2685 * A task struct has one reference for the use as "current".
2686 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2687 * schedule one last time. The schedule call will never return, and
2688 * the scheduled task must drop that reference.
2690 * We must observe prev->state before clearing prev->on_cpu (in
2691 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2692 * running on another CPU and we could rave with its RUNNING -> DEAD
2693 * transition, resulting in a double drop.
2695 prev_state = prev->state;
2696 vtime_task_switch(prev);
2697 perf_event_task_sched_in(prev, current);
2698 finish_lock_switch(rq, prev);
2699 finish_arch_post_lock_switch();
2701 fire_sched_in_preempt_notifiers(current);
2704 if (unlikely(prev_state == TASK_DEAD)) {
2705 if (prev->sched_class->task_dead)
2706 prev->sched_class->task_dead(prev);
2709 * Remove function-return probe instances associated with this
2710 * task and put them back on the free list.
2712 kprobe_flush_task(prev);
2713 put_task_struct(prev);
2716 tick_nohz_task_switch();
2722 /* rq->lock is NOT held, but preemption is disabled */
2723 static void __balance_callback(struct rq *rq)
2725 struct callback_head *head, *next;
2726 void (*func)(struct rq *rq);
2727 unsigned long flags;
2729 raw_spin_lock_irqsave(&rq->lock, flags);
2730 head = rq->balance_callback;
2731 rq->balance_callback = NULL;
2733 func = (void (*)(struct rq *))head->func;
2740 raw_spin_unlock_irqrestore(&rq->lock, flags);
2743 static inline void balance_callback(struct rq *rq)
2745 if (unlikely(rq->balance_callback))
2746 __balance_callback(rq);
2751 static inline void balance_callback(struct rq *rq)
2758 * schedule_tail - first thing a freshly forked thread must call.
2759 * @prev: the thread we just switched away from.
2761 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2762 __releases(rq->lock)
2767 * New tasks start with FORK_PREEMPT_COUNT, see there and
2768 * finish_task_switch() for details.
2770 * finish_task_switch() will drop rq->lock() and lower preempt_count
2771 * and the preempt_enable() will end up enabling preemption (on
2772 * PREEMPT_COUNT kernels).
2775 rq = finish_task_switch(prev);
2776 balance_callback(rq);
2779 if (current->set_child_tid)
2780 put_user(task_pid_vnr(current), current->set_child_tid);
2784 * context_switch - switch to the new MM and the new thread's register state.
2786 static __always_inline struct rq *
2787 context_switch(struct rq *rq, struct task_struct *prev,
2788 struct task_struct *next, struct pin_cookie cookie)
2790 struct mm_struct *mm, *oldmm;
2792 prepare_task_switch(rq, prev, next);
2795 oldmm = prev->active_mm;
2797 * For paravirt, this is coupled with an exit in switch_to to
2798 * combine the page table reload and the switch backend into
2801 arch_start_context_switch(prev);
2804 next->active_mm = oldmm;
2805 atomic_inc(&oldmm->mm_count);
2806 enter_lazy_tlb(oldmm, next);
2808 switch_mm_irqs_off(oldmm, mm, next);
2811 prev->active_mm = NULL;
2812 rq->prev_mm = oldmm;
2815 * Since the runqueue lock will be released by the next
2816 * task (which is an invalid locking op but in the case
2817 * of the scheduler it's an obvious special-case), so we
2818 * do an early lockdep release here:
2820 lockdep_unpin_lock(&rq->lock, cookie);
2821 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2823 /* Here we just switch the register state and the stack. */
2824 switch_to(prev, next, prev);
2827 return finish_task_switch(prev);
2831 * nr_running and nr_context_switches:
2833 * externally visible scheduler statistics: current number of runnable
2834 * threads, total number of context switches performed since bootup.
2836 unsigned long nr_running(void)
2838 unsigned long i, sum = 0;
2840 for_each_online_cpu(i)
2841 sum += cpu_rq(i)->nr_running;
2847 * Check if only the current task is running on the cpu.
2849 * Caution: this function does not check that the caller has disabled
2850 * preemption, thus the result might have a time-of-check-to-time-of-use
2851 * race. The caller is responsible to use it correctly, for example:
2853 * - from a non-preemptable section (of course)
2855 * - from a thread that is bound to a single CPU
2857 * - in a loop with very short iterations (e.g. a polling loop)
2859 bool single_task_running(void)
2861 return raw_rq()->nr_running == 1;
2863 EXPORT_SYMBOL(single_task_running);
2865 unsigned long long nr_context_switches(void)
2868 unsigned long long sum = 0;
2870 for_each_possible_cpu(i)
2871 sum += cpu_rq(i)->nr_switches;
2876 unsigned long nr_iowait(void)
2878 unsigned long i, sum = 0;
2880 for_each_possible_cpu(i)
2881 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2886 unsigned long nr_iowait_cpu(int cpu)
2888 struct rq *this = cpu_rq(cpu);
2889 return atomic_read(&this->nr_iowait);
2892 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2894 struct rq *rq = this_rq();
2895 *nr_waiters = atomic_read(&rq->nr_iowait);
2896 *load = rq->load.weight;
2902 * sched_exec - execve() is a valuable balancing opportunity, because at
2903 * this point the task has the smallest effective memory and cache footprint.
2905 void sched_exec(void)
2907 struct task_struct *p = current;
2908 unsigned long flags;
2911 raw_spin_lock_irqsave(&p->pi_lock, flags);
2912 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2913 if (dest_cpu == smp_processor_id())
2916 if (likely(cpu_active(dest_cpu))) {
2917 struct migration_arg arg = { p, dest_cpu };
2919 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2920 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2924 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2929 DEFINE_PER_CPU(struct kernel_stat, kstat);
2930 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2932 EXPORT_PER_CPU_SYMBOL(kstat);
2933 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2936 * Return accounted runtime for the task.
2937 * In case the task is currently running, return the runtime plus current's
2938 * pending runtime that have not been accounted yet.
2940 unsigned long long task_sched_runtime(struct task_struct *p)
2946 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2948 * 64-bit doesn't need locks to atomically read a 64bit value.
2949 * So we have a optimization chance when the task's delta_exec is 0.
2950 * Reading ->on_cpu is racy, but this is ok.
2952 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2953 * If we race with it entering cpu, unaccounted time is 0. This is
2954 * indistinguishable from the read occurring a few cycles earlier.
2955 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2956 * been accounted, so we're correct here as well.
2958 if (!p->on_cpu || !task_on_rq_queued(p))
2959 return p->se.sum_exec_runtime;
2962 rq = task_rq_lock(p, &rf);
2964 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2965 * project cycles that may never be accounted to this
2966 * thread, breaking clock_gettime().
2968 if (task_current(rq, p) && task_on_rq_queued(p)) {
2969 update_rq_clock(rq);
2970 p->sched_class->update_curr(rq);
2972 ns = p->se.sum_exec_runtime;
2973 task_rq_unlock(rq, p, &rf);
2979 * This function gets called by the timer code, with HZ frequency.
2980 * We call it with interrupts disabled.
2982 void scheduler_tick(void)
2984 int cpu = smp_processor_id();
2985 struct rq *rq = cpu_rq(cpu);
2986 struct task_struct *curr = rq->curr;
2990 raw_spin_lock(&rq->lock);
2991 update_rq_clock(rq);
2992 curr->sched_class->task_tick(rq, curr, 0);
2993 cpu_load_update_active(rq);
2994 calc_global_load_tick(rq);
2995 raw_spin_unlock(&rq->lock);
2997 perf_event_task_tick();
3000 rq->idle_balance = idle_cpu(cpu);
3001 trigger_load_balance(rq);
3003 rq_last_tick_reset(rq);
3006 #ifdef CONFIG_NO_HZ_FULL
3008 * scheduler_tick_max_deferment
3010 * Keep at least one tick per second when a single
3011 * active task is running because the scheduler doesn't
3012 * yet completely support full dynticks environment.
3014 * This makes sure that uptime, CFS vruntime, load
3015 * balancing, etc... continue to move forward, even
3016 * with a very low granularity.
3018 * Return: Maximum deferment in nanoseconds.
3020 u64 scheduler_tick_max_deferment(void)
3022 struct rq *rq = this_rq();
3023 unsigned long next, now = READ_ONCE(jiffies);
3025 next = rq->last_sched_tick + HZ;
3027 if (time_before_eq(next, now))
3030 return jiffies_to_nsecs(next - now);
3034 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3035 defined(CONFIG_PREEMPT_TRACER))
3037 * If the value passed in is equal to the current preempt count
3038 * then we just disabled preemption. Start timing the latency.
3040 static inline void preempt_latency_start(int val)
3042 if (preempt_count() == val) {
3043 unsigned long ip = get_lock_parent_ip();
3044 #ifdef CONFIG_DEBUG_PREEMPT
3045 current->preempt_disable_ip = ip;
3047 trace_preempt_off(CALLER_ADDR0, ip);
3051 void preempt_count_add(int val)
3053 #ifdef CONFIG_DEBUG_PREEMPT
3057 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3060 __preempt_count_add(val);
3061 #ifdef CONFIG_DEBUG_PREEMPT
3063 * Spinlock count overflowing soon?
3065 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3068 preempt_latency_start(val);
3070 EXPORT_SYMBOL(preempt_count_add);
3071 NOKPROBE_SYMBOL(preempt_count_add);
3074 * If the value passed in equals to the current preempt count
3075 * then we just enabled preemption. Stop timing the latency.
3077 static inline void preempt_latency_stop(int val)
3079 if (preempt_count() == val)
3080 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3083 void preempt_count_sub(int val)
3085 #ifdef CONFIG_DEBUG_PREEMPT
3089 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3092 * Is the spinlock portion underflowing?
3094 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3095 !(preempt_count() & PREEMPT_MASK)))
3099 preempt_latency_stop(val);
3100 __preempt_count_sub(val);
3102 EXPORT_SYMBOL(preempt_count_sub);
3103 NOKPROBE_SYMBOL(preempt_count_sub);
3106 static inline void preempt_latency_start(int val) { }
3107 static inline void preempt_latency_stop(int val) { }
3111 * Print scheduling while atomic bug:
3113 static noinline void __schedule_bug(struct task_struct *prev)
3115 if (oops_in_progress)
3118 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3119 prev->comm, prev->pid, preempt_count());
3121 debug_show_held_locks(prev);
3123 if (irqs_disabled())
3124 print_irqtrace_events(prev);
3125 #ifdef CONFIG_DEBUG_PREEMPT
3126 if (in_atomic_preempt_off()) {
3127 pr_err("Preemption disabled at:");
3128 print_ip_sym(current->preempt_disable_ip);
3133 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3137 * Various schedule()-time debugging checks and statistics:
3139 static inline void schedule_debug(struct task_struct *prev)
3141 #ifdef CONFIG_SCHED_STACK_END_CHECK
3142 BUG_ON(task_stack_end_corrupted(prev));
3145 if (unlikely(in_atomic_preempt_off())) {
3146 __schedule_bug(prev);
3147 preempt_count_set(PREEMPT_DISABLED);
3151 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3153 schedstat_inc(this_rq(), sched_count);
3157 * Pick up the highest-prio task:
3159 static inline struct task_struct *
3160 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3162 const struct sched_class *class = &fair_sched_class;
3163 struct task_struct *p;
3166 * Optimization: we know that if all tasks are in
3167 * the fair class we can call that function directly:
3169 if (likely(prev->sched_class == class &&
3170 rq->nr_running == rq->cfs.h_nr_running)) {
3171 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3172 if (unlikely(p == RETRY_TASK))
3175 /* assumes fair_sched_class->next == idle_sched_class */
3177 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3183 for_each_class(class) {
3184 p = class->pick_next_task(rq, prev, cookie);
3186 if (unlikely(p == RETRY_TASK))
3192 BUG(); /* the idle class will always have a runnable task */
3196 * __schedule() is the main scheduler function.
3198 * The main means of driving the scheduler and thus entering this function are:
3200 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3202 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3203 * paths. For example, see arch/x86/entry_64.S.
3205 * To drive preemption between tasks, the scheduler sets the flag in timer
3206 * interrupt handler scheduler_tick().
3208 * 3. Wakeups don't really cause entry into schedule(). They add a
3209 * task to the run-queue and that's it.
3211 * Now, if the new task added to the run-queue preempts the current
3212 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3213 * called on the nearest possible occasion:
3215 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3217 * - in syscall or exception context, at the next outmost
3218 * preempt_enable(). (this might be as soon as the wake_up()'s
3221 * - in IRQ context, return from interrupt-handler to
3222 * preemptible context
3224 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3227 * - cond_resched() call
3228 * - explicit schedule() call
3229 * - return from syscall or exception to user-space
3230 * - return from interrupt-handler to user-space
3232 * WARNING: must be called with preemption disabled!
3234 static void __sched notrace __schedule(bool preempt)
3236 struct task_struct *prev, *next;
3237 unsigned long *switch_count;
3238 struct pin_cookie cookie;
3242 cpu = smp_processor_id();
3247 * do_exit() calls schedule() with preemption disabled as an exception;
3248 * however we must fix that up, otherwise the next task will see an
3249 * inconsistent (higher) preempt count.
3251 * It also avoids the below schedule_debug() test from complaining
3254 if (unlikely(prev->state == TASK_DEAD))
3255 preempt_enable_no_resched_notrace();
3257 schedule_debug(prev);
3259 if (sched_feat(HRTICK))
3262 local_irq_disable();
3263 rcu_note_context_switch();
3266 * Make sure that signal_pending_state()->signal_pending() below
3267 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3268 * done by the caller to avoid the race with signal_wake_up().
3270 smp_mb__before_spinlock();
3271 raw_spin_lock(&rq->lock);
3272 cookie = lockdep_pin_lock(&rq->lock);
3274 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3276 switch_count = &prev->nivcsw;
3277 if (!preempt && prev->state) {
3278 if (unlikely(signal_pending_state(prev->state, prev))) {
3279 prev->state = TASK_RUNNING;
3281 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3285 * If a worker went to sleep, notify and ask workqueue
3286 * whether it wants to wake up a task to maintain
3289 if (prev->flags & PF_WQ_WORKER) {
3290 struct task_struct *to_wakeup;
3292 to_wakeup = wq_worker_sleeping(prev);
3294 try_to_wake_up_local(to_wakeup, cookie);
3297 switch_count = &prev->nvcsw;
3300 if (task_on_rq_queued(prev))
3301 update_rq_clock(rq);
3303 next = pick_next_task(rq, prev, cookie);
3304 clear_tsk_need_resched(prev);
3305 clear_preempt_need_resched();
3306 rq->clock_skip_update = 0;
3308 if (likely(prev != next)) {
3313 trace_sched_switch(preempt, prev, next);
3314 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3316 lockdep_unpin_lock(&rq->lock, cookie);
3317 raw_spin_unlock_irq(&rq->lock);
3320 balance_callback(rq);
3322 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3324 static inline void sched_submit_work(struct task_struct *tsk)
3326 if (!tsk->state || tsk_is_pi_blocked(tsk))
3329 * If we are going to sleep and we have plugged IO queued,
3330 * make sure to submit it to avoid deadlocks.
3332 if (blk_needs_flush_plug(tsk))
3333 blk_schedule_flush_plug(tsk);
3336 asmlinkage __visible void __sched schedule(void)
3338 struct task_struct *tsk = current;
3340 sched_submit_work(tsk);
3344 sched_preempt_enable_no_resched();
3345 } while (need_resched());
3347 EXPORT_SYMBOL(schedule);
3349 #ifdef CONFIG_CONTEXT_TRACKING
3350 asmlinkage __visible void __sched schedule_user(void)
3353 * If we come here after a random call to set_need_resched(),
3354 * or we have been woken up remotely but the IPI has not yet arrived,
3355 * we haven't yet exited the RCU idle mode. Do it here manually until
3356 * we find a better solution.
3358 * NB: There are buggy callers of this function. Ideally we
3359 * should warn if prev_state != CONTEXT_USER, but that will trigger
3360 * too frequently to make sense yet.
3362 enum ctx_state prev_state = exception_enter();
3364 exception_exit(prev_state);
3369 * schedule_preempt_disabled - called with preemption disabled
3371 * Returns with preemption disabled. Note: preempt_count must be 1
3373 void __sched schedule_preempt_disabled(void)
3375 sched_preempt_enable_no_resched();
3380 static void __sched notrace preempt_schedule_common(void)
3384 * Because the function tracer can trace preempt_count_sub()
3385 * and it also uses preempt_enable/disable_notrace(), if
3386 * NEED_RESCHED is set, the preempt_enable_notrace() called
3387 * by the function tracer will call this function again and
3388 * cause infinite recursion.
3390 * Preemption must be disabled here before the function
3391 * tracer can trace. Break up preempt_disable() into two
3392 * calls. One to disable preemption without fear of being
3393 * traced. The other to still record the preemption latency,
3394 * which can also be traced by the function tracer.
3396 preempt_disable_notrace();
3397 preempt_latency_start(1);
3399 preempt_latency_stop(1);
3400 preempt_enable_no_resched_notrace();
3403 * Check again in case we missed a preemption opportunity
3404 * between schedule and now.
3406 } while (need_resched());
3409 #ifdef CONFIG_PREEMPT
3411 * this is the entry point to schedule() from in-kernel preemption
3412 * off of preempt_enable. Kernel preemptions off return from interrupt
3413 * occur there and call schedule directly.
3415 asmlinkage __visible void __sched notrace preempt_schedule(void)
3418 * If there is a non-zero preempt_count or interrupts are disabled,
3419 * we do not want to preempt the current task. Just return..
3421 if (likely(!preemptible()))
3424 preempt_schedule_common();
3426 NOKPROBE_SYMBOL(preempt_schedule);
3427 EXPORT_SYMBOL(preempt_schedule);
3430 * preempt_schedule_notrace - preempt_schedule called by tracing
3432 * The tracing infrastructure uses preempt_enable_notrace to prevent
3433 * recursion and tracing preempt enabling caused by the tracing
3434 * infrastructure itself. But as tracing can happen in areas coming
3435 * from userspace or just about to enter userspace, a preempt enable
3436 * can occur before user_exit() is called. This will cause the scheduler
3437 * to be called when the system is still in usermode.
3439 * To prevent this, the preempt_enable_notrace will use this function
3440 * instead of preempt_schedule() to exit user context if needed before
3441 * calling the scheduler.
3443 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3445 enum ctx_state prev_ctx;
3447 if (likely(!preemptible()))
3452 * Because the function tracer can trace preempt_count_sub()
3453 * and it also uses preempt_enable/disable_notrace(), if
3454 * NEED_RESCHED is set, the preempt_enable_notrace() called
3455 * by the function tracer will call this function again and
3456 * cause infinite recursion.
3458 * Preemption must be disabled here before the function
3459 * tracer can trace. Break up preempt_disable() into two
3460 * calls. One to disable preemption without fear of being
3461 * traced. The other to still record the preemption latency,
3462 * which can also be traced by the function tracer.
3464 preempt_disable_notrace();
3465 preempt_latency_start(1);
3467 * Needs preempt disabled in case user_exit() is traced
3468 * and the tracer calls preempt_enable_notrace() causing
3469 * an infinite recursion.
3471 prev_ctx = exception_enter();
3473 exception_exit(prev_ctx);
3475 preempt_latency_stop(1);
3476 preempt_enable_no_resched_notrace();
3477 } while (need_resched());
3479 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3481 #endif /* CONFIG_PREEMPT */
3484 * this is the entry point to schedule() from kernel preemption
3485 * off of irq context.
3486 * Note, that this is called and return with irqs disabled. This will
3487 * protect us against recursive calling from irq.
3489 asmlinkage __visible void __sched preempt_schedule_irq(void)
3491 enum ctx_state prev_state;
3493 /* Catch callers which need to be fixed */
3494 BUG_ON(preempt_count() || !irqs_disabled());
3496 prev_state = exception_enter();
3502 local_irq_disable();
3503 sched_preempt_enable_no_resched();
3504 } while (need_resched());
3506 exception_exit(prev_state);
3509 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3512 return try_to_wake_up(curr->private, mode, wake_flags);
3514 EXPORT_SYMBOL(default_wake_function);
3516 #ifdef CONFIG_RT_MUTEXES
3519 * rt_mutex_setprio - set the current priority of a task
3521 * @prio: prio value (kernel-internal form)
3523 * This function changes the 'effective' priority of a task. It does
3524 * not touch ->normal_prio like __setscheduler().
3526 * Used by the rt_mutex code to implement priority inheritance
3527 * logic. Call site only calls if the priority of the task changed.
3529 void rt_mutex_setprio(struct task_struct *p, int prio)
3531 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3532 const struct sched_class *prev_class;
3536 BUG_ON(prio > MAX_PRIO);
3538 rq = __task_rq_lock(p, &rf);
3541 * Idle task boosting is a nono in general. There is one
3542 * exception, when PREEMPT_RT and NOHZ is active:
3544 * The idle task calls get_next_timer_interrupt() and holds
3545 * the timer wheel base->lock on the CPU and another CPU wants
3546 * to access the timer (probably to cancel it). We can safely
3547 * ignore the boosting request, as the idle CPU runs this code
3548 * with interrupts disabled and will complete the lock
3549 * protected section without being interrupted. So there is no
3550 * real need to boost.
3552 if (unlikely(p == rq->idle)) {
3553 WARN_ON(p != rq->curr);
3554 WARN_ON(p->pi_blocked_on);
3558 trace_sched_pi_setprio(p, prio);
3561 if (oldprio == prio)
3562 queue_flag &= ~DEQUEUE_MOVE;
3564 prev_class = p->sched_class;
3565 queued = task_on_rq_queued(p);
3566 running = task_current(rq, p);
3568 dequeue_task(rq, p, queue_flag);
3570 put_prev_task(rq, p);
3573 * Boosting condition are:
3574 * 1. -rt task is running and holds mutex A
3575 * --> -dl task blocks on mutex A
3577 * 2. -dl task is running and holds mutex A
3578 * --> -dl task blocks on mutex A and could preempt the
3581 if (dl_prio(prio)) {
3582 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3583 if (!dl_prio(p->normal_prio) ||
3584 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3585 p->dl.dl_boosted = 1;
3586 queue_flag |= ENQUEUE_REPLENISH;
3588 p->dl.dl_boosted = 0;
3589 p->sched_class = &dl_sched_class;
3590 } else if (rt_prio(prio)) {
3591 if (dl_prio(oldprio))
3592 p->dl.dl_boosted = 0;
3594 queue_flag |= ENQUEUE_HEAD;
3595 p->sched_class = &rt_sched_class;
3597 if (dl_prio(oldprio))
3598 p->dl.dl_boosted = 0;
3599 if (rt_prio(oldprio))
3601 p->sched_class = &fair_sched_class;
3607 p->sched_class->set_curr_task(rq);
3609 enqueue_task(rq, p, queue_flag);
3611 check_class_changed(rq, p, prev_class, oldprio);
3613 preempt_disable(); /* avoid rq from going away on us */
3614 __task_rq_unlock(rq, &rf);
3616 balance_callback(rq);
3621 void set_user_nice(struct task_struct *p, long nice)
3623 int old_prio, delta, queued;
3627 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3630 * We have to be careful, if called from sys_setpriority(),
3631 * the task might be in the middle of scheduling on another CPU.
3633 rq = task_rq_lock(p, &rf);
3635 * The RT priorities are set via sched_setscheduler(), but we still
3636 * allow the 'normal' nice value to be set - but as expected
3637 * it wont have any effect on scheduling until the task is
3638 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3640 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3641 p->static_prio = NICE_TO_PRIO(nice);
3644 queued = task_on_rq_queued(p);
3646 dequeue_task(rq, p, DEQUEUE_SAVE);
3648 p->static_prio = NICE_TO_PRIO(nice);
3651 p->prio = effective_prio(p);
3652 delta = p->prio - old_prio;
3655 enqueue_task(rq, p, ENQUEUE_RESTORE);
3657 * If the task increased its priority or is running and
3658 * lowered its priority, then reschedule its CPU:
3660 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3664 task_rq_unlock(rq, p, &rf);
3666 EXPORT_SYMBOL(set_user_nice);
3669 * can_nice - check if a task can reduce its nice value
3673 int can_nice(const struct task_struct *p, const int nice)
3675 /* convert nice value [19,-20] to rlimit style value [1,40] */
3676 int nice_rlim = nice_to_rlimit(nice);
3678 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3679 capable(CAP_SYS_NICE));
3682 #ifdef __ARCH_WANT_SYS_NICE
3685 * sys_nice - change the priority of the current process.
3686 * @increment: priority increment
3688 * sys_setpriority is a more generic, but much slower function that
3689 * does similar things.
3691 SYSCALL_DEFINE1(nice, int, increment)
3696 * Setpriority might change our priority at the same moment.
3697 * We don't have to worry. Conceptually one call occurs first
3698 * and we have a single winner.
3700 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3701 nice = task_nice(current) + increment;
3703 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3704 if (increment < 0 && !can_nice(current, nice))
3707 retval = security_task_setnice(current, nice);
3711 set_user_nice(current, nice);
3718 * task_prio - return the priority value of a given task.
3719 * @p: the task in question.
3721 * Return: The priority value as seen by users in /proc.
3722 * RT tasks are offset by -200. Normal tasks are centered
3723 * around 0, value goes from -16 to +15.
3725 int task_prio(const struct task_struct *p)
3727 return p->prio - MAX_RT_PRIO;
3731 * idle_cpu - is a given cpu idle currently?
3732 * @cpu: the processor in question.
3734 * Return: 1 if the CPU is currently idle. 0 otherwise.
3736 int idle_cpu(int cpu)
3738 struct rq *rq = cpu_rq(cpu);
3740 if (rq->curr != rq->idle)
3747 if (!llist_empty(&rq->wake_list))
3755 * idle_task - return the idle task for a given cpu.
3756 * @cpu: the processor in question.
3758 * Return: The idle task for the cpu @cpu.
3760 struct task_struct *idle_task(int cpu)
3762 return cpu_rq(cpu)->idle;
3766 * find_process_by_pid - find a process with a matching PID value.
3767 * @pid: the pid in question.
3769 * The task of @pid, if found. %NULL otherwise.
3771 static struct task_struct *find_process_by_pid(pid_t pid)
3773 return pid ? find_task_by_vpid(pid) : current;
3777 * This function initializes the sched_dl_entity of a newly becoming
3778 * SCHED_DEADLINE task.
3780 * Only the static values are considered here, the actual runtime and the
3781 * absolute deadline will be properly calculated when the task is enqueued
3782 * for the first time with its new policy.
3785 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3787 struct sched_dl_entity *dl_se = &p->dl;
3789 dl_se->dl_runtime = attr->sched_runtime;
3790 dl_se->dl_deadline = attr->sched_deadline;
3791 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3792 dl_se->flags = attr->sched_flags;
3793 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3796 * Changing the parameters of a task is 'tricky' and we're not doing
3797 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3799 * What we SHOULD do is delay the bandwidth release until the 0-lag
3800 * point. This would include retaining the task_struct until that time
3801 * and change dl_overflow() to not immediately decrement the current
3804 * Instead we retain the current runtime/deadline and let the new
3805 * parameters take effect after the current reservation period lapses.
3806 * This is safe (albeit pessimistic) because the 0-lag point is always
3807 * before the current scheduling deadline.
3809 * We can still have temporary overloads because we do not delay the
3810 * change in bandwidth until that time; so admission control is
3811 * not on the safe side. It does however guarantee tasks will never
3812 * consume more than promised.
3817 * sched_setparam() passes in -1 for its policy, to let the functions
3818 * it calls know not to change it.
3820 #define SETPARAM_POLICY -1
3822 static void __setscheduler_params(struct task_struct *p,
3823 const struct sched_attr *attr)
3825 int policy = attr->sched_policy;
3827 if (policy == SETPARAM_POLICY)
3832 if (dl_policy(policy))
3833 __setparam_dl(p, attr);
3834 else if (fair_policy(policy))
3835 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3838 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3839 * !rt_policy. Always setting this ensures that things like
3840 * getparam()/getattr() don't report silly values for !rt tasks.
3842 p->rt_priority = attr->sched_priority;
3843 p->normal_prio = normal_prio(p);
3847 /* Actually do priority change: must hold pi & rq lock. */
3848 static void __setscheduler(struct rq *rq, struct task_struct *p,
3849 const struct sched_attr *attr, bool keep_boost)
3851 __setscheduler_params(p, attr);
3854 * Keep a potential priority boosting if called from
3855 * sched_setscheduler().
3858 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3860 p->prio = normal_prio(p);
3862 if (dl_prio(p->prio))
3863 p->sched_class = &dl_sched_class;
3864 else if (rt_prio(p->prio))
3865 p->sched_class = &rt_sched_class;
3867 p->sched_class = &fair_sched_class;
3871 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3873 struct sched_dl_entity *dl_se = &p->dl;
3875 attr->sched_priority = p->rt_priority;
3876 attr->sched_runtime = dl_se->dl_runtime;
3877 attr->sched_deadline = dl_se->dl_deadline;
3878 attr->sched_period = dl_se->dl_period;
3879 attr->sched_flags = dl_se->flags;
3883 * This function validates the new parameters of a -deadline task.
3884 * We ask for the deadline not being zero, and greater or equal
3885 * than the runtime, as well as the period of being zero or
3886 * greater than deadline. Furthermore, we have to be sure that
3887 * user parameters are above the internal resolution of 1us (we
3888 * check sched_runtime only since it is always the smaller one) and
3889 * below 2^63 ns (we have to check both sched_deadline and
3890 * sched_period, as the latter can be zero).
3893 __checkparam_dl(const struct sched_attr *attr)
3896 if (attr->sched_deadline == 0)
3900 * Since we truncate DL_SCALE bits, make sure we're at least
3903 if (attr->sched_runtime < (1ULL << DL_SCALE))
3907 * Since we use the MSB for wrap-around and sign issues, make
3908 * sure it's not set (mind that period can be equal to zero).
3910 if (attr->sched_deadline & (1ULL << 63) ||
3911 attr->sched_period & (1ULL << 63))
3914 /* runtime <= deadline <= period (if period != 0) */
3915 if ((attr->sched_period != 0 &&
3916 attr->sched_period < attr->sched_deadline) ||
3917 attr->sched_deadline < attr->sched_runtime)
3924 * check the target process has a UID that matches the current process's
3926 static bool check_same_owner(struct task_struct *p)
3928 const struct cred *cred = current_cred(), *pcred;
3932 pcred = __task_cred(p);
3933 match = (uid_eq(cred->euid, pcred->euid) ||
3934 uid_eq(cred->euid, pcred->uid));
3939 static bool dl_param_changed(struct task_struct *p,
3940 const struct sched_attr *attr)
3942 struct sched_dl_entity *dl_se = &p->dl;
3944 if (dl_se->dl_runtime != attr->sched_runtime ||
3945 dl_se->dl_deadline != attr->sched_deadline ||
3946 dl_se->dl_period != attr->sched_period ||
3947 dl_se->flags != attr->sched_flags)
3953 static int __sched_setscheduler(struct task_struct *p,
3954 const struct sched_attr *attr,
3957 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3958 MAX_RT_PRIO - 1 - attr->sched_priority;
3959 int retval, oldprio, oldpolicy = -1, queued, running;
3960 int new_effective_prio, policy = attr->sched_policy;
3961 const struct sched_class *prev_class;
3964 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3967 /* may grab non-irq protected spin_locks */
3968 BUG_ON(in_interrupt());
3970 /* double check policy once rq lock held */
3972 reset_on_fork = p->sched_reset_on_fork;
3973 policy = oldpolicy = p->policy;
3975 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3977 if (!valid_policy(policy))
3981 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3985 * Valid priorities for SCHED_FIFO and SCHED_RR are
3986 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3987 * SCHED_BATCH and SCHED_IDLE is 0.
3989 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3990 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3992 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3993 (rt_policy(policy) != (attr->sched_priority != 0)))
3997 * Allow unprivileged RT tasks to decrease priority:
3999 if (user && !capable(CAP_SYS_NICE)) {
4000 if (fair_policy(policy)) {
4001 if (attr->sched_nice < task_nice(p) &&
4002 !can_nice(p, attr->sched_nice))
4006 if (rt_policy(policy)) {
4007 unsigned long rlim_rtprio =
4008 task_rlimit(p, RLIMIT_RTPRIO);
4010 /* can't set/change the rt policy */
4011 if (policy != p->policy && !rlim_rtprio)
4014 /* can't increase priority */
4015 if (attr->sched_priority > p->rt_priority &&
4016 attr->sched_priority > rlim_rtprio)
4021 * Can't set/change SCHED_DEADLINE policy at all for now
4022 * (safest behavior); in the future we would like to allow
4023 * unprivileged DL tasks to increase their relative deadline
4024 * or reduce their runtime (both ways reducing utilization)
4026 if (dl_policy(policy))
4030 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4031 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4033 if (idle_policy(p->policy) && !idle_policy(policy)) {
4034 if (!can_nice(p, task_nice(p)))
4038 /* can't change other user's priorities */
4039 if (!check_same_owner(p))
4042 /* Normal users shall not reset the sched_reset_on_fork flag */
4043 if (p->sched_reset_on_fork && !reset_on_fork)
4048 retval = security_task_setscheduler(p);
4054 * make sure no PI-waiters arrive (or leave) while we are
4055 * changing the priority of the task:
4057 * To be able to change p->policy safely, the appropriate
4058 * runqueue lock must be held.
4060 rq = task_rq_lock(p, &rf);
4063 * Changing the policy of the stop threads its a very bad idea
4065 if (p == rq->stop) {
4066 task_rq_unlock(rq, p, &rf);
4071 * If not changing anything there's no need to proceed further,
4072 * but store a possible modification of reset_on_fork.
4074 if (unlikely(policy == p->policy)) {
4075 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4077 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4079 if (dl_policy(policy) && dl_param_changed(p, attr))
4082 p->sched_reset_on_fork = reset_on_fork;
4083 task_rq_unlock(rq, p, &rf);
4089 #ifdef CONFIG_RT_GROUP_SCHED
4091 * Do not allow realtime tasks into groups that have no runtime
4094 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4095 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4096 !task_group_is_autogroup(task_group(p))) {
4097 task_rq_unlock(rq, p, &rf);
4102 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4103 cpumask_t *span = rq->rd->span;
4106 * Don't allow tasks with an affinity mask smaller than
4107 * the entire root_domain to become SCHED_DEADLINE. We
4108 * will also fail if there's no bandwidth available.
4110 if (!cpumask_subset(span, &p->cpus_allowed) ||
4111 rq->rd->dl_bw.bw == 0) {
4112 task_rq_unlock(rq, p, &rf);
4119 /* recheck policy now with rq lock held */
4120 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4121 policy = oldpolicy = -1;
4122 task_rq_unlock(rq, p, &rf);
4127 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4128 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4131 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4132 task_rq_unlock(rq, p, &rf);
4136 p->sched_reset_on_fork = reset_on_fork;
4141 * Take priority boosted tasks into account. If the new
4142 * effective priority is unchanged, we just store the new
4143 * normal parameters and do not touch the scheduler class and
4144 * the runqueue. This will be done when the task deboost
4147 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4148 if (new_effective_prio == oldprio)
4149 queue_flags &= ~DEQUEUE_MOVE;
4152 queued = task_on_rq_queued(p);
4153 running = task_current(rq, p);
4155 dequeue_task(rq, p, queue_flags);
4157 put_prev_task(rq, p);
4159 prev_class = p->sched_class;
4160 __setscheduler(rq, p, attr, pi);
4163 p->sched_class->set_curr_task(rq);
4166 * We enqueue to tail when the priority of a task is
4167 * increased (user space view).
4169 if (oldprio < p->prio)
4170 queue_flags |= ENQUEUE_HEAD;
4172 enqueue_task(rq, p, queue_flags);
4175 check_class_changed(rq, p, prev_class, oldprio);
4176 preempt_disable(); /* avoid rq from going away on us */
4177 task_rq_unlock(rq, p, &rf);
4180 rt_mutex_adjust_pi(p);
4183 * Run balance callbacks after we've adjusted the PI chain.
4185 balance_callback(rq);
4191 static int _sched_setscheduler(struct task_struct *p, int policy,
4192 const struct sched_param *param, bool check)
4194 struct sched_attr attr = {
4195 .sched_policy = policy,
4196 .sched_priority = param->sched_priority,
4197 .sched_nice = PRIO_TO_NICE(p->static_prio),
4200 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4201 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4202 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4203 policy &= ~SCHED_RESET_ON_FORK;
4204 attr.sched_policy = policy;
4207 return __sched_setscheduler(p, &attr, check, true);
4210 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4211 * @p: the task in question.
4212 * @policy: new policy.
4213 * @param: structure containing the new RT priority.
4215 * Return: 0 on success. An error code otherwise.
4217 * NOTE that the task may be already dead.
4219 int sched_setscheduler(struct task_struct *p, int policy,
4220 const struct sched_param *param)
4222 return _sched_setscheduler(p, policy, param, true);
4224 EXPORT_SYMBOL_GPL(sched_setscheduler);
4226 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4228 return __sched_setscheduler(p, attr, true, true);
4230 EXPORT_SYMBOL_GPL(sched_setattr);
4233 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4234 * @p: the task in question.
4235 * @policy: new policy.
4236 * @param: structure containing the new RT priority.
4238 * Just like sched_setscheduler, only don't bother checking if the
4239 * current context has permission. For example, this is needed in
4240 * stop_machine(): we create temporary high priority worker threads,
4241 * but our caller might not have that capability.
4243 * Return: 0 on success. An error code otherwise.
4245 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4246 const struct sched_param *param)
4248 return _sched_setscheduler(p, policy, param, false);
4250 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4253 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4255 struct sched_param lparam;
4256 struct task_struct *p;
4259 if (!param || pid < 0)
4261 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4266 p = find_process_by_pid(pid);
4268 retval = sched_setscheduler(p, policy, &lparam);
4275 * Mimics kernel/events/core.c perf_copy_attr().
4277 static int sched_copy_attr(struct sched_attr __user *uattr,
4278 struct sched_attr *attr)
4283 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4287 * zero the full structure, so that a short copy will be nice.
4289 memset(attr, 0, sizeof(*attr));
4291 ret = get_user(size, &uattr->size);
4295 if (size > PAGE_SIZE) /* silly large */
4298 if (!size) /* abi compat */
4299 size = SCHED_ATTR_SIZE_VER0;
4301 if (size < SCHED_ATTR_SIZE_VER0)
4305 * If we're handed a bigger struct than we know of,
4306 * ensure all the unknown bits are 0 - i.e. new
4307 * user-space does not rely on any kernel feature
4308 * extensions we dont know about yet.
4310 if (size > sizeof(*attr)) {
4311 unsigned char __user *addr;
4312 unsigned char __user *end;
4315 addr = (void __user *)uattr + sizeof(*attr);
4316 end = (void __user *)uattr + size;
4318 for (; addr < end; addr++) {
4319 ret = get_user(val, addr);
4325 size = sizeof(*attr);
4328 ret = copy_from_user(attr, uattr, size);
4333 * XXX: do we want to be lenient like existing syscalls; or do we want
4334 * to be strict and return an error on out-of-bounds values?
4336 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4341 put_user(sizeof(*attr), &uattr->size);
4346 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4347 * @pid: the pid in question.
4348 * @policy: new policy.
4349 * @param: structure containing the new RT priority.
4351 * Return: 0 on success. An error code otherwise.
4353 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4354 struct sched_param __user *, param)
4356 /* negative values for policy are not valid */
4360 return do_sched_setscheduler(pid, policy, param);
4364 * sys_sched_setparam - set/change the RT priority of a thread
4365 * @pid: the pid in question.
4366 * @param: structure containing the new RT priority.
4368 * Return: 0 on success. An error code otherwise.
4370 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4372 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4376 * sys_sched_setattr - same as above, but with extended sched_attr
4377 * @pid: the pid in question.
4378 * @uattr: structure containing the extended parameters.
4379 * @flags: for future extension.
4381 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4382 unsigned int, flags)
4384 struct sched_attr attr;
4385 struct task_struct *p;
4388 if (!uattr || pid < 0 || flags)
4391 retval = sched_copy_attr(uattr, &attr);
4395 if ((int)attr.sched_policy < 0)
4400 p = find_process_by_pid(pid);
4402 retval = sched_setattr(p, &attr);
4409 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4410 * @pid: the pid in question.
4412 * Return: On success, the policy of the thread. Otherwise, a negative error
4415 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4417 struct task_struct *p;
4425 p = find_process_by_pid(pid);
4427 retval = security_task_getscheduler(p);
4430 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4437 * sys_sched_getparam - get the RT priority of a thread
4438 * @pid: the pid in question.
4439 * @param: structure containing the RT priority.
4441 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4444 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4446 struct sched_param lp = { .sched_priority = 0 };
4447 struct task_struct *p;
4450 if (!param || pid < 0)
4454 p = find_process_by_pid(pid);
4459 retval = security_task_getscheduler(p);
4463 if (task_has_rt_policy(p))
4464 lp.sched_priority = p->rt_priority;
4468 * This one might sleep, we cannot do it with a spinlock held ...
4470 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4479 static int sched_read_attr(struct sched_attr __user *uattr,
4480 struct sched_attr *attr,
4485 if (!access_ok(VERIFY_WRITE, uattr, usize))
4489 * If we're handed a smaller struct than we know of,
4490 * ensure all the unknown bits are 0 - i.e. old
4491 * user-space does not get uncomplete information.
4493 if (usize < sizeof(*attr)) {
4494 unsigned char *addr;
4497 addr = (void *)attr + usize;
4498 end = (void *)attr + sizeof(*attr);
4500 for (; addr < end; addr++) {
4508 ret = copy_to_user(uattr, attr, attr->size);
4516 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4517 * @pid: the pid in question.
4518 * @uattr: structure containing the extended parameters.
4519 * @size: sizeof(attr) for fwd/bwd comp.
4520 * @flags: for future extension.
4522 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4523 unsigned int, size, unsigned int, flags)
4525 struct sched_attr attr = {
4526 .size = sizeof(struct sched_attr),
4528 struct task_struct *p;
4531 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4532 size < SCHED_ATTR_SIZE_VER0 || flags)
4536 p = find_process_by_pid(pid);
4541 retval = security_task_getscheduler(p);
4545 attr.sched_policy = p->policy;
4546 if (p->sched_reset_on_fork)
4547 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4548 if (task_has_dl_policy(p))
4549 __getparam_dl(p, &attr);
4550 else if (task_has_rt_policy(p))
4551 attr.sched_priority = p->rt_priority;
4553 attr.sched_nice = task_nice(p);
4557 retval = sched_read_attr(uattr, &attr, size);
4565 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4567 cpumask_var_t cpus_allowed, new_mask;
4568 struct task_struct *p;
4573 p = find_process_by_pid(pid);
4579 /* Prevent p going away */
4583 if (p->flags & PF_NO_SETAFFINITY) {
4587 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4591 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4593 goto out_free_cpus_allowed;
4596 if (!check_same_owner(p)) {
4598 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4600 goto out_free_new_mask;
4605 retval = security_task_setscheduler(p);
4607 goto out_free_new_mask;
4610 cpuset_cpus_allowed(p, cpus_allowed);
4611 cpumask_and(new_mask, in_mask, cpus_allowed);
4614 * Since bandwidth control happens on root_domain basis,
4615 * if admission test is enabled, we only admit -deadline
4616 * tasks allowed to run on all the CPUs in the task's
4620 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4622 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4625 goto out_free_new_mask;
4631 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4634 cpuset_cpus_allowed(p, cpus_allowed);
4635 if (!cpumask_subset(new_mask, cpus_allowed)) {
4637 * We must have raced with a concurrent cpuset
4638 * update. Just reset the cpus_allowed to the
4639 * cpuset's cpus_allowed
4641 cpumask_copy(new_mask, cpus_allowed);
4646 free_cpumask_var(new_mask);
4647 out_free_cpus_allowed:
4648 free_cpumask_var(cpus_allowed);
4654 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4655 struct cpumask *new_mask)
4657 if (len < cpumask_size())
4658 cpumask_clear(new_mask);
4659 else if (len > cpumask_size())
4660 len = cpumask_size();
4662 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4666 * sys_sched_setaffinity - set the cpu affinity of a process
4667 * @pid: pid of the process
4668 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4669 * @user_mask_ptr: user-space pointer to the new cpu mask
4671 * Return: 0 on success. An error code otherwise.
4673 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4674 unsigned long __user *, user_mask_ptr)
4676 cpumask_var_t new_mask;
4679 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4682 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4684 retval = sched_setaffinity(pid, new_mask);
4685 free_cpumask_var(new_mask);
4689 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4691 struct task_struct *p;
4692 unsigned long flags;
4698 p = find_process_by_pid(pid);
4702 retval = security_task_getscheduler(p);
4706 raw_spin_lock_irqsave(&p->pi_lock, flags);
4707 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4708 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4717 * sys_sched_getaffinity - get the cpu affinity of a process
4718 * @pid: pid of the process
4719 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4720 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4722 * Return: 0 on success. An error code otherwise.
4724 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4725 unsigned long __user *, user_mask_ptr)
4730 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4732 if (len & (sizeof(unsigned long)-1))
4735 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4738 ret = sched_getaffinity(pid, mask);
4740 size_t retlen = min_t(size_t, len, cpumask_size());
4742 if (copy_to_user(user_mask_ptr, mask, retlen))
4747 free_cpumask_var(mask);
4753 * sys_sched_yield - yield the current processor to other threads.
4755 * This function yields the current CPU to other tasks. If there are no
4756 * other threads running on this CPU then this function will return.
4760 SYSCALL_DEFINE0(sched_yield)
4762 struct rq *rq = this_rq_lock();
4764 schedstat_inc(rq, yld_count);
4765 current->sched_class->yield_task(rq);
4768 * Since we are going to call schedule() anyway, there's
4769 * no need to preempt or enable interrupts:
4771 __release(rq->lock);
4772 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4773 do_raw_spin_unlock(&rq->lock);
4774 sched_preempt_enable_no_resched();
4781 int __sched _cond_resched(void)
4783 if (should_resched(0)) {
4784 preempt_schedule_common();
4789 EXPORT_SYMBOL(_cond_resched);
4792 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4793 * call schedule, and on return reacquire the lock.
4795 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4796 * operations here to prevent schedule() from being called twice (once via
4797 * spin_unlock(), once by hand).
4799 int __cond_resched_lock(spinlock_t *lock)
4801 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4804 lockdep_assert_held(lock);
4806 if (spin_needbreak(lock) || resched) {
4809 preempt_schedule_common();
4817 EXPORT_SYMBOL(__cond_resched_lock);
4819 int __sched __cond_resched_softirq(void)
4821 BUG_ON(!in_softirq());
4823 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4825 preempt_schedule_common();
4831 EXPORT_SYMBOL(__cond_resched_softirq);
4834 * yield - yield the current processor to other threads.
4836 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4838 * The scheduler is at all times free to pick the calling task as the most
4839 * eligible task to run, if removing the yield() call from your code breaks
4840 * it, its already broken.
4842 * Typical broken usage is:
4847 * where one assumes that yield() will let 'the other' process run that will
4848 * make event true. If the current task is a SCHED_FIFO task that will never
4849 * happen. Never use yield() as a progress guarantee!!
4851 * If you want to use yield() to wait for something, use wait_event().
4852 * If you want to use yield() to be 'nice' for others, use cond_resched().
4853 * If you still want to use yield(), do not!
4855 void __sched yield(void)
4857 set_current_state(TASK_RUNNING);
4860 EXPORT_SYMBOL(yield);
4863 * yield_to - yield the current processor to another thread in
4864 * your thread group, or accelerate that thread toward the
4865 * processor it's on.
4867 * @preempt: whether task preemption is allowed or not
4869 * It's the caller's job to ensure that the target task struct
4870 * can't go away on us before we can do any checks.
4873 * true (>0) if we indeed boosted the target task.
4874 * false (0) if we failed to boost the target.
4875 * -ESRCH if there's no task to yield to.
4877 int __sched yield_to(struct task_struct *p, bool preempt)
4879 struct task_struct *curr = current;
4880 struct rq *rq, *p_rq;
4881 unsigned long flags;
4884 local_irq_save(flags);
4890 * If we're the only runnable task on the rq and target rq also
4891 * has only one task, there's absolutely no point in yielding.
4893 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4898 double_rq_lock(rq, p_rq);
4899 if (task_rq(p) != p_rq) {
4900 double_rq_unlock(rq, p_rq);
4904 if (!curr->sched_class->yield_to_task)
4907 if (curr->sched_class != p->sched_class)
4910 if (task_running(p_rq, p) || p->state)
4913 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4915 schedstat_inc(rq, yld_count);
4917 * Make p's CPU reschedule; pick_next_entity takes care of
4920 if (preempt && rq != p_rq)
4925 double_rq_unlock(rq, p_rq);
4927 local_irq_restore(flags);
4934 EXPORT_SYMBOL_GPL(yield_to);
4937 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4938 * that process accounting knows that this is a task in IO wait state.
4940 long __sched io_schedule_timeout(long timeout)
4942 int old_iowait = current->in_iowait;
4946 current->in_iowait = 1;
4947 blk_schedule_flush_plug(current);
4949 delayacct_blkio_start();
4951 atomic_inc(&rq->nr_iowait);
4952 ret = schedule_timeout(timeout);
4953 current->in_iowait = old_iowait;
4954 atomic_dec(&rq->nr_iowait);
4955 delayacct_blkio_end();
4959 EXPORT_SYMBOL(io_schedule_timeout);
4962 * sys_sched_get_priority_max - return maximum RT priority.
4963 * @policy: scheduling class.
4965 * Return: On success, this syscall returns the maximum
4966 * rt_priority that can be used by a given scheduling class.
4967 * On failure, a negative error code is returned.
4969 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4976 ret = MAX_USER_RT_PRIO-1;
4978 case SCHED_DEADLINE:
4989 * sys_sched_get_priority_min - return minimum RT priority.
4990 * @policy: scheduling class.
4992 * Return: On success, this syscall returns the minimum
4993 * rt_priority that can be used by a given scheduling class.
4994 * On failure, a negative error code is returned.
4996 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5005 case SCHED_DEADLINE:
5015 * sys_sched_rr_get_interval - return the default timeslice of a process.
5016 * @pid: pid of the process.
5017 * @interval: userspace pointer to the timeslice value.
5019 * this syscall writes the default timeslice value of a given process
5020 * into the user-space timespec buffer. A value of '0' means infinity.
5022 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5025 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5026 struct timespec __user *, interval)
5028 struct task_struct *p;
5029 unsigned int time_slice;
5040 p = find_process_by_pid(pid);
5044 retval = security_task_getscheduler(p);
5048 rq = task_rq_lock(p, &rf);
5050 if (p->sched_class->get_rr_interval)
5051 time_slice = p->sched_class->get_rr_interval(rq, p);
5052 task_rq_unlock(rq, p, &rf);
5055 jiffies_to_timespec(time_slice, &t);
5056 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5064 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5066 void sched_show_task(struct task_struct *p)
5068 unsigned long free = 0;
5070 unsigned long state = p->state;
5073 state = __ffs(state) + 1;
5074 printk(KERN_INFO "%-15.15s %c", p->comm,
5075 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5076 #if BITS_PER_LONG == 32
5077 if (state == TASK_RUNNING)
5078 printk(KERN_CONT " running ");
5080 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5082 if (state == TASK_RUNNING)
5083 printk(KERN_CONT " running task ");
5085 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5087 #ifdef CONFIG_DEBUG_STACK_USAGE
5088 free = stack_not_used(p);
5093 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5095 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5096 task_pid_nr(p), ppid,
5097 (unsigned long)task_thread_info(p)->flags);
5099 print_worker_info(KERN_INFO, p);
5100 show_stack(p, NULL);
5103 void show_state_filter(unsigned long state_filter)
5105 struct task_struct *g, *p;
5107 #if BITS_PER_LONG == 32
5109 " task PC stack pid father\n");
5112 " task PC stack pid father\n");
5115 for_each_process_thread(g, p) {
5117 * reset the NMI-timeout, listing all files on a slow
5118 * console might take a lot of time:
5120 touch_nmi_watchdog();
5121 if (!state_filter || (p->state & state_filter))
5125 touch_all_softlockup_watchdogs();
5127 #ifdef CONFIG_SCHED_DEBUG
5129 sysrq_sched_debug_show();
5133 * Only show locks if all tasks are dumped:
5136 debug_show_all_locks();
5139 void init_idle_bootup_task(struct task_struct *idle)
5141 idle->sched_class = &idle_sched_class;
5145 * init_idle - set up an idle thread for a given CPU
5146 * @idle: task in question
5147 * @cpu: cpu the idle task belongs to
5149 * NOTE: this function does not set the idle thread's NEED_RESCHED
5150 * flag, to make booting more robust.
5152 void init_idle(struct task_struct *idle, int cpu)
5154 struct rq *rq = cpu_rq(cpu);
5155 unsigned long flags;
5157 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5158 raw_spin_lock(&rq->lock);
5160 __sched_fork(0, idle);
5161 idle->state = TASK_RUNNING;
5162 idle->se.exec_start = sched_clock();
5164 kasan_unpoison_task_stack(idle);
5168 * Its possible that init_idle() gets called multiple times on a task,
5169 * in that case do_set_cpus_allowed() will not do the right thing.
5171 * And since this is boot we can forgo the serialization.
5173 set_cpus_allowed_common(idle, cpumask_of(cpu));
5176 * We're having a chicken and egg problem, even though we are
5177 * holding rq->lock, the cpu isn't yet set to this cpu so the
5178 * lockdep check in task_group() will fail.
5180 * Similar case to sched_fork(). / Alternatively we could
5181 * use task_rq_lock() here and obtain the other rq->lock.
5186 __set_task_cpu(idle, cpu);
5189 rq->curr = rq->idle = idle;
5190 idle->on_rq = TASK_ON_RQ_QUEUED;
5194 raw_spin_unlock(&rq->lock);
5195 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5197 /* Set the preempt count _outside_ the spinlocks! */
5198 init_idle_preempt_count(idle, cpu);
5201 * The idle tasks have their own, simple scheduling class:
5203 idle->sched_class = &idle_sched_class;
5204 ftrace_graph_init_idle_task(idle, cpu);
5205 vtime_init_idle(idle, cpu);
5207 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5211 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5212 const struct cpumask *trial)
5214 int ret = 1, trial_cpus;
5215 struct dl_bw *cur_dl_b;
5216 unsigned long flags;
5218 if (!cpumask_weight(cur))
5221 rcu_read_lock_sched();
5222 cur_dl_b = dl_bw_of(cpumask_any(cur));
5223 trial_cpus = cpumask_weight(trial);
5225 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5226 if (cur_dl_b->bw != -1 &&
5227 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5229 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5230 rcu_read_unlock_sched();
5235 int task_can_attach(struct task_struct *p,
5236 const struct cpumask *cs_cpus_allowed)
5241 * Kthreads which disallow setaffinity shouldn't be moved
5242 * to a new cpuset; we don't want to change their cpu
5243 * affinity and isolating such threads by their set of
5244 * allowed nodes is unnecessary. Thus, cpusets are not
5245 * applicable for such threads. This prevents checking for
5246 * success of set_cpus_allowed_ptr() on all attached tasks
5247 * before cpus_allowed may be changed.
5249 if (p->flags & PF_NO_SETAFFINITY) {
5255 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5257 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5262 unsigned long flags;
5264 rcu_read_lock_sched();
5265 dl_b = dl_bw_of(dest_cpu);
5266 raw_spin_lock_irqsave(&dl_b->lock, flags);
5267 cpus = dl_bw_cpus(dest_cpu);
5268 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5273 * We reserve space for this task in the destination
5274 * root_domain, as we can't fail after this point.
5275 * We will free resources in the source root_domain
5276 * later on (see set_cpus_allowed_dl()).
5278 __dl_add(dl_b, p->dl.dl_bw);
5280 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5281 rcu_read_unlock_sched();
5291 #ifdef CONFIG_NUMA_BALANCING
5292 /* Migrate current task p to target_cpu */
5293 int migrate_task_to(struct task_struct *p, int target_cpu)
5295 struct migration_arg arg = { p, target_cpu };
5296 int curr_cpu = task_cpu(p);
5298 if (curr_cpu == target_cpu)
5301 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5304 /* TODO: This is not properly updating schedstats */
5306 trace_sched_move_numa(p, curr_cpu, target_cpu);
5307 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5311 * Requeue a task on a given node and accurately track the number of NUMA
5312 * tasks on the runqueues
5314 void sched_setnuma(struct task_struct *p, int nid)
5316 bool queued, running;
5320 rq = task_rq_lock(p, &rf);
5321 queued = task_on_rq_queued(p);
5322 running = task_current(rq, p);
5325 dequeue_task(rq, p, DEQUEUE_SAVE);
5327 put_prev_task(rq, p);
5329 p->numa_preferred_nid = nid;
5332 p->sched_class->set_curr_task(rq);
5334 enqueue_task(rq, p, ENQUEUE_RESTORE);
5335 task_rq_unlock(rq, p, &rf);
5337 #endif /* CONFIG_NUMA_BALANCING */
5339 #ifdef CONFIG_HOTPLUG_CPU
5341 * Ensures that the idle task is using init_mm right before its cpu goes
5344 void idle_task_exit(void)
5346 struct mm_struct *mm = current->active_mm;
5348 BUG_ON(cpu_online(smp_processor_id()));
5350 if (mm != &init_mm) {
5351 switch_mm_irqs_off(mm, &init_mm, current);
5352 finish_arch_post_lock_switch();
5358 * Since this CPU is going 'away' for a while, fold any nr_active delta
5359 * we might have. Assumes we're called after migrate_tasks() so that the
5360 * nr_active count is stable.
5362 * Also see the comment "Global load-average calculations".
5364 static void calc_load_migrate(struct rq *rq)
5366 long delta = calc_load_fold_active(rq);
5368 atomic_long_add(delta, &calc_load_tasks);
5371 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5375 static const struct sched_class fake_sched_class = {
5376 .put_prev_task = put_prev_task_fake,
5379 static struct task_struct fake_task = {
5381 * Avoid pull_{rt,dl}_task()
5383 .prio = MAX_PRIO + 1,
5384 .sched_class = &fake_sched_class,
5388 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5389 * try_to_wake_up()->select_task_rq().
5391 * Called with rq->lock held even though we'er in stop_machine() and
5392 * there's no concurrency possible, we hold the required locks anyway
5393 * because of lock validation efforts.
5395 static void migrate_tasks(struct rq *dead_rq)
5397 struct rq *rq = dead_rq;
5398 struct task_struct *next, *stop = rq->stop;
5399 struct pin_cookie cookie;
5403 * Fudge the rq selection such that the below task selection loop
5404 * doesn't get stuck on the currently eligible stop task.
5406 * We're currently inside stop_machine() and the rq is either stuck
5407 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5408 * either way we should never end up calling schedule() until we're
5414 * put_prev_task() and pick_next_task() sched
5415 * class method both need to have an up-to-date
5416 * value of rq->clock[_task]
5418 update_rq_clock(rq);
5422 * There's this thread running, bail when that's the only
5425 if (rq->nr_running == 1)
5429 * pick_next_task assumes pinned rq->lock.
5431 cookie = lockdep_pin_lock(&rq->lock);
5432 next = pick_next_task(rq, &fake_task, cookie);
5434 next->sched_class->put_prev_task(rq, next);
5437 * Rules for changing task_struct::cpus_allowed are holding
5438 * both pi_lock and rq->lock, such that holding either
5439 * stabilizes the mask.
5441 * Drop rq->lock is not quite as disastrous as it usually is
5442 * because !cpu_active at this point, which means load-balance
5443 * will not interfere. Also, stop-machine.
5445 lockdep_unpin_lock(&rq->lock, cookie);
5446 raw_spin_unlock(&rq->lock);
5447 raw_spin_lock(&next->pi_lock);
5448 raw_spin_lock(&rq->lock);
5451 * Since we're inside stop-machine, _nothing_ should have
5452 * changed the task, WARN if weird stuff happened, because in
5453 * that case the above rq->lock drop is a fail too.
5455 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5456 raw_spin_unlock(&next->pi_lock);
5460 /* Find suitable destination for @next, with force if needed. */
5461 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5463 rq = __migrate_task(rq, next, dest_cpu);
5464 if (rq != dead_rq) {
5465 raw_spin_unlock(&rq->lock);
5467 raw_spin_lock(&rq->lock);
5469 raw_spin_unlock(&next->pi_lock);
5474 #endif /* CONFIG_HOTPLUG_CPU */
5476 static void set_rq_online(struct rq *rq)
5479 const struct sched_class *class;
5481 cpumask_set_cpu(rq->cpu, rq->rd->online);
5484 for_each_class(class) {
5485 if (class->rq_online)
5486 class->rq_online(rq);
5491 static void set_rq_offline(struct rq *rq)
5494 const struct sched_class *class;
5496 for_each_class(class) {
5497 if (class->rq_offline)
5498 class->rq_offline(rq);
5501 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5507 * migration_call - callback that gets triggered when a CPU is added.
5508 * Here we can start up the necessary migration thread for the new CPU.
5511 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5513 int cpu = (long)hcpu;
5514 unsigned long flags;
5515 struct rq *rq = cpu_rq(cpu);
5517 switch (action & ~CPU_TASKS_FROZEN) {
5519 case CPU_UP_PREPARE:
5520 rq->calc_load_update = calc_load_update;
5521 account_reset_rq(rq);
5525 /* Update our root-domain */
5526 raw_spin_lock_irqsave(&rq->lock, flags);
5528 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5532 raw_spin_unlock_irqrestore(&rq->lock, flags);
5535 #ifdef CONFIG_HOTPLUG_CPU
5537 sched_ttwu_pending();
5538 /* Update our root-domain */
5539 raw_spin_lock_irqsave(&rq->lock, flags);
5541 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5545 BUG_ON(rq->nr_running != 1); /* the migration thread */
5546 raw_spin_unlock_irqrestore(&rq->lock, flags);
5550 calc_load_migrate(rq);
5555 update_max_interval();
5561 * Register at high priority so that task migration (migrate_all_tasks)
5562 * happens before everything else. This has to be lower priority than
5563 * the notifier in the perf_event subsystem, though.
5565 static struct notifier_block migration_notifier = {
5566 .notifier_call = migration_call,
5567 .priority = CPU_PRI_MIGRATION,
5570 static void set_cpu_rq_start_time(void)
5572 int cpu = smp_processor_id();
5573 struct rq *rq = cpu_rq(cpu);
5574 rq->age_stamp = sched_clock_cpu(cpu);
5577 static int sched_cpu_active(struct notifier_block *nfb,
5578 unsigned long action, void *hcpu)
5580 int cpu = (long)hcpu;
5582 switch (action & ~CPU_TASKS_FROZEN) {
5584 set_cpu_rq_start_time();
5587 case CPU_DOWN_FAILED:
5588 set_cpu_active(cpu, true);
5596 static int sched_cpu_inactive(struct notifier_block *nfb,
5597 unsigned long action, void *hcpu)
5599 switch (action & ~CPU_TASKS_FROZEN) {
5600 case CPU_DOWN_PREPARE:
5601 set_cpu_active((long)hcpu, false);
5608 static int __init migration_init(void)
5610 void *cpu = (void *)(long)smp_processor_id();
5613 /* Initialize migration for the boot CPU */
5614 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5615 BUG_ON(err == NOTIFY_BAD);
5616 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5617 register_cpu_notifier(&migration_notifier);
5619 /* Register cpu active notifiers */
5620 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5621 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5625 early_initcall(migration_init);
5627 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5629 #ifdef CONFIG_SCHED_DEBUG
5631 static __read_mostly int sched_debug_enabled;
5633 static int __init sched_debug_setup(char *str)
5635 sched_debug_enabled = 1;
5639 early_param("sched_debug", sched_debug_setup);
5641 static inline bool sched_debug(void)
5643 return sched_debug_enabled;
5646 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5647 struct cpumask *groupmask)
5649 struct sched_group *group = sd->groups;
5651 cpumask_clear(groupmask);
5653 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5655 if (!(sd->flags & SD_LOAD_BALANCE)) {
5656 printk("does not load-balance\n");
5658 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5663 printk(KERN_CONT "span %*pbl level %s\n",
5664 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5666 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5667 printk(KERN_ERR "ERROR: domain->span does not contain "
5670 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5671 printk(KERN_ERR "ERROR: domain->groups does not contain"
5675 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5679 printk(KERN_ERR "ERROR: group is NULL\n");
5683 if (!cpumask_weight(sched_group_cpus(group))) {
5684 printk(KERN_CONT "\n");
5685 printk(KERN_ERR "ERROR: empty group\n");
5689 if (!(sd->flags & SD_OVERLAP) &&
5690 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5691 printk(KERN_CONT "\n");
5692 printk(KERN_ERR "ERROR: repeated CPUs\n");
5696 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5698 printk(KERN_CONT " %*pbl",
5699 cpumask_pr_args(sched_group_cpus(group)));
5700 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5701 printk(KERN_CONT " (cpu_capacity = %d)",
5702 group->sgc->capacity);
5705 group = group->next;
5706 } while (group != sd->groups);
5707 printk(KERN_CONT "\n");
5709 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5710 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5713 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5714 printk(KERN_ERR "ERROR: parent span is not a superset "
5715 "of domain->span\n");
5719 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5723 if (!sched_debug_enabled)
5727 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5731 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5734 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5742 #else /* !CONFIG_SCHED_DEBUG */
5743 # define sched_domain_debug(sd, cpu) do { } while (0)
5744 static inline bool sched_debug(void)
5748 #endif /* CONFIG_SCHED_DEBUG */
5750 static int sd_degenerate(struct sched_domain *sd)
5752 if (cpumask_weight(sched_domain_span(sd)) == 1)
5755 /* Following flags need at least 2 groups */
5756 if (sd->flags & (SD_LOAD_BALANCE |
5757 SD_BALANCE_NEWIDLE |
5760 SD_SHARE_CPUCAPACITY |
5761 SD_SHARE_PKG_RESOURCES |
5762 SD_SHARE_POWERDOMAIN)) {
5763 if (sd->groups != sd->groups->next)
5767 /* Following flags don't use groups */
5768 if (sd->flags & (SD_WAKE_AFFINE))
5775 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5777 unsigned long cflags = sd->flags, pflags = parent->flags;
5779 if (sd_degenerate(parent))
5782 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5785 /* Flags needing groups don't count if only 1 group in parent */
5786 if (parent->groups == parent->groups->next) {
5787 pflags &= ~(SD_LOAD_BALANCE |
5788 SD_BALANCE_NEWIDLE |
5791 SD_SHARE_CPUCAPACITY |
5792 SD_SHARE_PKG_RESOURCES |
5794 SD_SHARE_POWERDOMAIN);
5795 if (nr_node_ids == 1)
5796 pflags &= ~SD_SERIALIZE;
5798 if (~cflags & pflags)
5804 static void free_rootdomain(struct rcu_head *rcu)
5806 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5808 cpupri_cleanup(&rd->cpupri);
5809 cpudl_cleanup(&rd->cpudl);
5810 free_cpumask_var(rd->dlo_mask);
5811 free_cpumask_var(rd->rto_mask);
5812 free_cpumask_var(rd->online);
5813 free_cpumask_var(rd->span);
5817 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5819 struct root_domain *old_rd = NULL;
5820 unsigned long flags;
5822 raw_spin_lock_irqsave(&rq->lock, flags);
5827 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5830 cpumask_clear_cpu(rq->cpu, old_rd->span);
5833 * If we dont want to free the old_rd yet then
5834 * set old_rd to NULL to skip the freeing later
5837 if (!atomic_dec_and_test(&old_rd->refcount))
5841 atomic_inc(&rd->refcount);
5844 cpumask_set_cpu(rq->cpu, rd->span);
5845 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5848 raw_spin_unlock_irqrestore(&rq->lock, flags);
5851 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5854 static int init_rootdomain(struct root_domain *rd)
5856 memset(rd, 0, sizeof(*rd));
5858 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5860 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5862 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5864 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5867 init_dl_bw(&rd->dl_bw);
5868 if (cpudl_init(&rd->cpudl) != 0)
5871 if (cpupri_init(&rd->cpupri) != 0)
5876 free_cpumask_var(rd->rto_mask);
5878 free_cpumask_var(rd->dlo_mask);
5880 free_cpumask_var(rd->online);
5882 free_cpumask_var(rd->span);
5888 * By default the system creates a single root-domain with all cpus as
5889 * members (mimicking the global state we have today).
5891 struct root_domain def_root_domain;
5893 static void init_defrootdomain(void)
5895 init_rootdomain(&def_root_domain);
5897 atomic_set(&def_root_domain.refcount, 1);
5900 static struct root_domain *alloc_rootdomain(void)
5902 struct root_domain *rd;
5904 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5908 if (init_rootdomain(rd) != 0) {
5916 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5918 struct sched_group *tmp, *first;
5927 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5932 } while (sg != first);
5935 static void free_sched_domain(struct rcu_head *rcu)
5937 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5940 * If its an overlapping domain it has private groups, iterate and
5943 if (sd->flags & SD_OVERLAP) {
5944 free_sched_groups(sd->groups, 1);
5945 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5946 kfree(sd->groups->sgc);
5952 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5954 call_rcu(&sd->rcu, free_sched_domain);
5957 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5959 for (; sd; sd = sd->parent)
5960 destroy_sched_domain(sd, cpu);
5964 * Keep a special pointer to the highest sched_domain that has
5965 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5966 * allows us to avoid some pointer chasing select_idle_sibling().
5968 * Also keep a unique ID per domain (we use the first cpu number in
5969 * the cpumask of the domain), this allows us to quickly tell if
5970 * two cpus are in the same cache domain, see cpus_share_cache().
5972 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5973 DEFINE_PER_CPU(int, sd_llc_size);
5974 DEFINE_PER_CPU(int, sd_llc_id);
5975 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5976 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5977 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5979 static void update_top_cache_domain(int cpu)
5981 struct sched_domain *sd;
5982 struct sched_domain *busy_sd = NULL;
5986 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5988 id = cpumask_first(sched_domain_span(sd));
5989 size = cpumask_weight(sched_domain_span(sd));
5990 busy_sd = sd->parent; /* sd_busy */
5992 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5994 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5995 per_cpu(sd_llc_size, cpu) = size;
5996 per_cpu(sd_llc_id, cpu) = id;
5998 sd = lowest_flag_domain(cpu, SD_NUMA);
5999 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6001 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6002 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6006 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6007 * hold the hotplug lock.
6010 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6012 struct rq *rq = cpu_rq(cpu);
6013 struct sched_domain *tmp;
6015 /* Remove the sched domains which do not contribute to scheduling. */
6016 for (tmp = sd; tmp; ) {
6017 struct sched_domain *parent = tmp->parent;
6021 if (sd_parent_degenerate(tmp, parent)) {
6022 tmp->parent = parent->parent;
6024 parent->parent->child = tmp;
6026 * Transfer SD_PREFER_SIBLING down in case of a
6027 * degenerate parent; the spans match for this
6028 * so the property transfers.
6030 if (parent->flags & SD_PREFER_SIBLING)
6031 tmp->flags |= SD_PREFER_SIBLING;
6032 destroy_sched_domain(parent, cpu);
6037 if (sd && sd_degenerate(sd)) {
6040 destroy_sched_domain(tmp, cpu);
6045 sched_domain_debug(sd, cpu);
6047 rq_attach_root(rq, rd);
6049 rcu_assign_pointer(rq->sd, sd);
6050 destroy_sched_domains(tmp, cpu);
6052 update_top_cache_domain(cpu);
6055 /* Setup the mask of cpus configured for isolated domains */
6056 static int __init isolated_cpu_setup(char *str)
6060 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6061 ret = cpulist_parse(str, cpu_isolated_map);
6063 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6068 __setup("isolcpus=", isolated_cpu_setup);
6071 struct sched_domain ** __percpu sd;
6072 struct root_domain *rd;
6083 * Build an iteration mask that can exclude certain CPUs from the upwards
6086 * Asymmetric node setups can result in situations where the domain tree is of
6087 * unequal depth, make sure to skip domains that already cover the entire
6090 * In that case build_sched_domains() will have terminated the iteration early
6091 * and our sibling sd spans will be empty. Domains should always include the
6092 * cpu they're built on, so check that.
6095 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6097 const struct cpumask *span = sched_domain_span(sd);
6098 struct sd_data *sdd = sd->private;
6099 struct sched_domain *sibling;
6102 for_each_cpu(i, span) {
6103 sibling = *per_cpu_ptr(sdd->sd, i);
6104 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6107 cpumask_set_cpu(i, sched_group_mask(sg));
6112 * Return the canonical balance cpu for this group, this is the first cpu
6113 * of this group that's also in the iteration mask.
6115 int group_balance_cpu(struct sched_group *sg)
6117 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6121 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6123 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6124 const struct cpumask *span = sched_domain_span(sd);
6125 struct cpumask *covered = sched_domains_tmpmask;
6126 struct sd_data *sdd = sd->private;
6127 struct sched_domain *sibling;
6130 cpumask_clear(covered);
6132 for_each_cpu(i, span) {
6133 struct cpumask *sg_span;
6135 if (cpumask_test_cpu(i, covered))
6138 sibling = *per_cpu_ptr(sdd->sd, i);
6140 /* See the comment near build_group_mask(). */
6141 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6144 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6145 GFP_KERNEL, cpu_to_node(cpu));
6150 sg_span = sched_group_cpus(sg);
6152 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6154 cpumask_set_cpu(i, sg_span);
6156 cpumask_or(covered, covered, sg_span);
6158 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6159 if (atomic_inc_return(&sg->sgc->ref) == 1)
6160 build_group_mask(sd, sg);
6163 * Initialize sgc->capacity such that even if we mess up the
6164 * domains and no possible iteration will get us here, we won't
6167 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6170 * Make sure the first group of this domain contains the
6171 * canonical balance cpu. Otherwise the sched_domain iteration
6172 * breaks. See update_sg_lb_stats().
6174 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6175 group_balance_cpu(sg) == cpu)
6185 sd->groups = groups;
6190 free_sched_groups(first, 0);
6195 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6197 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6198 struct sched_domain *child = sd->child;
6201 cpu = cpumask_first(sched_domain_span(child));
6204 *sg = *per_cpu_ptr(sdd->sg, cpu);
6205 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6206 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6213 * build_sched_groups will build a circular linked list of the groups
6214 * covered by the given span, and will set each group's ->cpumask correctly,
6215 * and ->cpu_capacity to 0.
6217 * Assumes the sched_domain tree is fully constructed
6220 build_sched_groups(struct sched_domain *sd, int cpu)
6222 struct sched_group *first = NULL, *last = NULL;
6223 struct sd_data *sdd = sd->private;
6224 const struct cpumask *span = sched_domain_span(sd);
6225 struct cpumask *covered;
6228 get_group(cpu, sdd, &sd->groups);
6229 atomic_inc(&sd->groups->ref);
6231 if (cpu != cpumask_first(span))
6234 lockdep_assert_held(&sched_domains_mutex);
6235 covered = sched_domains_tmpmask;
6237 cpumask_clear(covered);
6239 for_each_cpu(i, span) {
6240 struct sched_group *sg;
6243 if (cpumask_test_cpu(i, covered))
6246 group = get_group(i, sdd, &sg);
6247 cpumask_setall(sched_group_mask(sg));
6249 for_each_cpu(j, span) {
6250 if (get_group(j, sdd, NULL) != group)
6253 cpumask_set_cpu(j, covered);
6254 cpumask_set_cpu(j, sched_group_cpus(sg));
6269 * Initialize sched groups cpu_capacity.
6271 * cpu_capacity indicates the capacity of sched group, which is used while
6272 * distributing the load between different sched groups in a sched domain.
6273 * Typically cpu_capacity for all the groups in a sched domain will be same
6274 * unless there are asymmetries in the topology. If there are asymmetries,
6275 * group having more cpu_capacity will pickup more load compared to the
6276 * group having less cpu_capacity.
6278 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6280 struct sched_group *sg = sd->groups;
6285 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6287 } while (sg != sd->groups);
6289 if (cpu != group_balance_cpu(sg))
6292 update_group_capacity(sd, cpu);
6293 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6297 * Initializers for schedule domains
6298 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6301 static int default_relax_domain_level = -1;
6302 int sched_domain_level_max;
6304 static int __init setup_relax_domain_level(char *str)
6306 if (kstrtoint(str, 0, &default_relax_domain_level))
6307 pr_warn("Unable to set relax_domain_level\n");
6311 __setup("relax_domain_level=", setup_relax_domain_level);
6313 static void set_domain_attribute(struct sched_domain *sd,
6314 struct sched_domain_attr *attr)
6318 if (!attr || attr->relax_domain_level < 0) {
6319 if (default_relax_domain_level < 0)
6322 request = default_relax_domain_level;
6324 request = attr->relax_domain_level;
6325 if (request < sd->level) {
6326 /* turn off idle balance on this domain */
6327 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6329 /* turn on idle balance on this domain */
6330 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6334 static void __sdt_free(const struct cpumask *cpu_map);
6335 static int __sdt_alloc(const struct cpumask *cpu_map);
6337 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6338 const struct cpumask *cpu_map)
6342 if (!atomic_read(&d->rd->refcount))
6343 free_rootdomain(&d->rd->rcu); /* fall through */
6345 free_percpu(d->sd); /* fall through */
6347 __sdt_free(cpu_map); /* fall through */
6353 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6354 const struct cpumask *cpu_map)
6356 memset(d, 0, sizeof(*d));
6358 if (__sdt_alloc(cpu_map))
6359 return sa_sd_storage;
6360 d->sd = alloc_percpu(struct sched_domain *);
6362 return sa_sd_storage;
6363 d->rd = alloc_rootdomain();
6366 return sa_rootdomain;
6370 * NULL the sd_data elements we've used to build the sched_domain and
6371 * sched_group structure so that the subsequent __free_domain_allocs()
6372 * will not free the data we're using.
6374 static void claim_allocations(int cpu, struct sched_domain *sd)
6376 struct sd_data *sdd = sd->private;
6378 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6379 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6381 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6382 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6384 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6385 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6389 static int sched_domains_numa_levels;
6390 enum numa_topology_type sched_numa_topology_type;
6391 static int *sched_domains_numa_distance;
6392 int sched_max_numa_distance;
6393 static struct cpumask ***sched_domains_numa_masks;
6394 static int sched_domains_curr_level;
6398 * SD_flags allowed in topology descriptions.
6400 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6401 * SD_SHARE_PKG_RESOURCES - describes shared caches
6402 * SD_NUMA - describes NUMA topologies
6403 * SD_SHARE_POWERDOMAIN - describes shared power domain
6406 * SD_ASYM_PACKING - describes SMT quirks
6408 #define TOPOLOGY_SD_FLAGS \
6409 (SD_SHARE_CPUCAPACITY | \
6410 SD_SHARE_PKG_RESOURCES | \
6413 SD_SHARE_POWERDOMAIN)
6415 static struct sched_domain *
6416 sd_init(struct sched_domain_topology_level *tl, int cpu)
6418 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6419 int sd_weight, sd_flags = 0;
6423 * Ugly hack to pass state to sd_numa_mask()...
6425 sched_domains_curr_level = tl->numa_level;
6428 sd_weight = cpumask_weight(tl->mask(cpu));
6431 sd_flags = (*tl->sd_flags)();
6432 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6433 "wrong sd_flags in topology description\n"))
6434 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6436 *sd = (struct sched_domain){
6437 .min_interval = sd_weight,
6438 .max_interval = 2*sd_weight,
6440 .imbalance_pct = 125,
6442 .cache_nice_tries = 0,
6449 .flags = 1*SD_LOAD_BALANCE
6450 | 1*SD_BALANCE_NEWIDLE
6455 | 0*SD_SHARE_CPUCAPACITY
6456 | 0*SD_SHARE_PKG_RESOURCES
6458 | 0*SD_PREFER_SIBLING
6463 .last_balance = jiffies,
6464 .balance_interval = sd_weight,
6466 .max_newidle_lb_cost = 0,
6467 .next_decay_max_lb_cost = jiffies,
6468 #ifdef CONFIG_SCHED_DEBUG
6474 * Convert topological properties into behaviour.
6477 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6478 sd->flags |= SD_PREFER_SIBLING;
6479 sd->imbalance_pct = 110;
6480 sd->smt_gain = 1178; /* ~15% */
6482 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6483 sd->imbalance_pct = 117;
6484 sd->cache_nice_tries = 1;
6488 } else if (sd->flags & SD_NUMA) {
6489 sd->cache_nice_tries = 2;
6493 sd->flags |= SD_SERIALIZE;
6494 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6495 sd->flags &= ~(SD_BALANCE_EXEC |
6502 sd->flags |= SD_PREFER_SIBLING;
6503 sd->cache_nice_tries = 1;
6508 sd->private = &tl->data;
6514 * Topology list, bottom-up.
6516 static struct sched_domain_topology_level default_topology[] = {
6517 #ifdef CONFIG_SCHED_SMT
6518 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6520 #ifdef CONFIG_SCHED_MC
6521 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6523 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6527 static struct sched_domain_topology_level *sched_domain_topology =
6530 #define for_each_sd_topology(tl) \
6531 for (tl = sched_domain_topology; tl->mask; tl++)
6533 void set_sched_topology(struct sched_domain_topology_level *tl)
6535 sched_domain_topology = tl;
6540 static const struct cpumask *sd_numa_mask(int cpu)
6542 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6545 static void sched_numa_warn(const char *str)
6547 static int done = false;
6555 printk(KERN_WARNING "ERROR: %s\n\n", str);
6557 for (i = 0; i < nr_node_ids; i++) {
6558 printk(KERN_WARNING " ");
6559 for (j = 0; j < nr_node_ids; j++)
6560 printk(KERN_CONT "%02d ", node_distance(i,j));
6561 printk(KERN_CONT "\n");
6563 printk(KERN_WARNING "\n");
6566 bool find_numa_distance(int distance)
6570 if (distance == node_distance(0, 0))
6573 for (i = 0; i < sched_domains_numa_levels; i++) {
6574 if (sched_domains_numa_distance[i] == distance)
6582 * A system can have three types of NUMA topology:
6583 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6584 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6585 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6587 * The difference between a glueless mesh topology and a backplane
6588 * topology lies in whether communication between not directly
6589 * connected nodes goes through intermediary nodes (where programs
6590 * could run), or through backplane controllers. This affects
6591 * placement of programs.
6593 * The type of topology can be discerned with the following tests:
6594 * - If the maximum distance between any nodes is 1 hop, the system
6595 * is directly connected.
6596 * - If for two nodes A and B, located N > 1 hops away from each other,
6597 * there is an intermediary node C, which is < N hops away from both
6598 * nodes A and B, the system is a glueless mesh.
6600 static void init_numa_topology_type(void)
6604 n = sched_max_numa_distance;
6606 if (sched_domains_numa_levels <= 1) {
6607 sched_numa_topology_type = NUMA_DIRECT;
6611 for_each_online_node(a) {
6612 for_each_online_node(b) {
6613 /* Find two nodes furthest removed from each other. */
6614 if (node_distance(a, b) < n)
6617 /* Is there an intermediary node between a and b? */
6618 for_each_online_node(c) {
6619 if (node_distance(a, c) < n &&
6620 node_distance(b, c) < n) {
6621 sched_numa_topology_type =
6627 sched_numa_topology_type = NUMA_BACKPLANE;
6633 static void sched_init_numa(void)
6635 int next_distance, curr_distance = node_distance(0, 0);
6636 struct sched_domain_topology_level *tl;
6640 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6641 if (!sched_domains_numa_distance)
6645 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6646 * unique distances in the node_distance() table.
6648 * Assumes node_distance(0,j) includes all distances in
6649 * node_distance(i,j) in order to avoid cubic time.
6651 next_distance = curr_distance;
6652 for (i = 0; i < nr_node_ids; i++) {
6653 for (j = 0; j < nr_node_ids; j++) {
6654 for (k = 0; k < nr_node_ids; k++) {
6655 int distance = node_distance(i, k);
6657 if (distance > curr_distance &&
6658 (distance < next_distance ||
6659 next_distance == curr_distance))
6660 next_distance = distance;
6663 * While not a strong assumption it would be nice to know
6664 * about cases where if node A is connected to B, B is not
6665 * equally connected to A.
6667 if (sched_debug() && node_distance(k, i) != distance)
6668 sched_numa_warn("Node-distance not symmetric");
6670 if (sched_debug() && i && !find_numa_distance(distance))
6671 sched_numa_warn("Node-0 not representative");
6673 if (next_distance != curr_distance) {
6674 sched_domains_numa_distance[level++] = next_distance;
6675 sched_domains_numa_levels = level;
6676 curr_distance = next_distance;
6681 * In case of sched_debug() we verify the above assumption.
6691 * 'level' contains the number of unique distances, excluding the
6692 * identity distance node_distance(i,i).
6694 * The sched_domains_numa_distance[] array includes the actual distance
6699 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6700 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6701 * the array will contain less then 'level' members. This could be
6702 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6703 * in other functions.
6705 * We reset it to 'level' at the end of this function.
6707 sched_domains_numa_levels = 0;
6709 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6710 if (!sched_domains_numa_masks)
6714 * Now for each level, construct a mask per node which contains all
6715 * cpus of nodes that are that many hops away from us.
6717 for (i = 0; i < level; i++) {
6718 sched_domains_numa_masks[i] =
6719 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6720 if (!sched_domains_numa_masks[i])
6723 for (j = 0; j < nr_node_ids; j++) {
6724 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6728 sched_domains_numa_masks[i][j] = mask;
6731 if (node_distance(j, k) > sched_domains_numa_distance[i])
6734 cpumask_or(mask, mask, cpumask_of_node(k));
6739 /* Compute default topology size */
6740 for (i = 0; sched_domain_topology[i].mask; i++);
6742 tl = kzalloc((i + level + 1) *
6743 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6748 * Copy the default topology bits..
6750 for (i = 0; sched_domain_topology[i].mask; i++)
6751 tl[i] = sched_domain_topology[i];
6754 * .. and append 'j' levels of NUMA goodness.
6756 for (j = 0; j < level; i++, j++) {
6757 tl[i] = (struct sched_domain_topology_level){
6758 .mask = sd_numa_mask,
6759 .sd_flags = cpu_numa_flags,
6760 .flags = SDTL_OVERLAP,
6766 sched_domain_topology = tl;
6768 sched_domains_numa_levels = level;
6769 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6771 init_numa_topology_type();
6774 static void sched_domains_numa_masks_set(int cpu)
6777 int node = cpu_to_node(cpu);
6779 for (i = 0; i < sched_domains_numa_levels; i++) {
6780 for (j = 0; j < nr_node_ids; j++) {
6781 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6782 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6787 static void sched_domains_numa_masks_clear(int cpu)
6790 for (i = 0; i < sched_domains_numa_levels; i++) {
6791 for (j = 0; j < nr_node_ids; j++)
6792 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6797 * Update sched_domains_numa_masks[level][node] array when new cpus
6800 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6801 unsigned long action,
6804 int cpu = (long)hcpu;
6806 switch (action & ~CPU_TASKS_FROZEN) {
6808 sched_domains_numa_masks_set(cpu);
6812 sched_domains_numa_masks_clear(cpu);
6822 static inline void sched_init_numa(void)
6826 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6827 unsigned long action,
6832 #endif /* CONFIG_NUMA */
6834 static int __sdt_alloc(const struct cpumask *cpu_map)
6836 struct sched_domain_topology_level *tl;
6839 for_each_sd_topology(tl) {
6840 struct sd_data *sdd = &tl->data;
6842 sdd->sd = alloc_percpu(struct sched_domain *);
6846 sdd->sg = alloc_percpu(struct sched_group *);
6850 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6854 for_each_cpu(j, cpu_map) {
6855 struct sched_domain *sd;
6856 struct sched_group *sg;
6857 struct sched_group_capacity *sgc;
6859 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6860 GFP_KERNEL, cpu_to_node(j));
6864 *per_cpu_ptr(sdd->sd, j) = sd;
6866 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6867 GFP_KERNEL, cpu_to_node(j));
6873 *per_cpu_ptr(sdd->sg, j) = sg;
6875 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6876 GFP_KERNEL, cpu_to_node(j));
6880 *per_cpu_ptr(sdd->sgc, j) = sgc;
6887 static void __sdt_free(const struct cpumask *cpu_map)
6889 struct sched_domain_topology_level *tl;
6892 for_each_sd_topology(tl) {
6893 struct sd_data *sdd = &tl->data;
6895 for_each_cpu(j, cpu_map) {
6896 struct sched_domain *sd;
6899 sd = *per_cpu_ptr(sdd->sd, j);
6900 if (sd && (sd->flags & SD_OVERLAP))
6901 free_sched_groups(sd->groups, 0);
6902 kfree(*per_cpu_ptr(sdd->sd, j));
6906 kfree(*per_cpu_ptr(sdd->sg, j));
6908 kfree(*per_cpu_ptr(sdd->sgc, j));
6910 free_percpu(sdd->sd);
6912 free_percpu(sdd->sg);
6914 free_percpu(sdd->sgc);
6919 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6920 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6921 struct sched_domain *child, int cpu)
6923 struct sched_domain *sd = sd_init(tl, cpu);
6927 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6929 sd->level = child->level + 1;
6930 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6934 if (!cpumask_subset(sched_domain_span(child),
6935 sched_domain_span(sd))) {
6936 pr_err("BUG: arch topology borken\n");
6937 #ifdef CONFIG_SCHED_DEBUG
6938 pr_err(" the %s domain not a subset of the %s domain\n",
6939 child->name, sd->name);
6941 /* Fixup, ensure @sd has at least @child cpus. */
6942 cpumask_or(sched_domain_span(sd),
6943 sched_domain_span(sd),
6944 sched_domain_span(child));
6948 set_domain_attribute(sd, attr);
6954 * Build sched domains for a given set of cpus and attach the sched domains
6955 * to the individual cpus
6957 static int build_sched_domains(const struct cpumask *cpu_map,
6958 struct sched_domain_attr *attr)
6960 enum s_alloc alloc_state;
6961 struct sched_domain *sd;
6963 int i, ret = -ENOMEM;
6965 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6966 if (alloc_state != sa_rootdomain)
6969 /* Set up domains for cpus specified by the cpu_map. */
6970 for_each_cpu(i, cpu_map) {
6971 struct sched_domain_topology_level *tl;
6974 for_each_sd_topology(tl) {
6975 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6976 if (tl == sched_domain_topology)
6977 *per_cpu_ptr(d.sd, i) = sd;
6978 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6979 sd->flags |= SD_OVERLAP;
6980 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6985 /* Build the groups for the domains */
6986 for_each_cpu(i, cpu_map) {
6987 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6988 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6989 if (sd->flags & SD_OVERLAP) {
6990 if (build_overlap_sched_groups(sd, i))
6993 if (build_sched_groups(sd, i))
6999 /* Calculate CPU capacity for physical packages and nodes */
7000 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7001 if (!cpumask_test_cpu(i, cpu_map))
7004 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7005 claim_allocations(i, sd);
7006 init_sched_groups_capacity(i, sd);
7010 /* Attach the domains */
7012 for_each_cpu(i, cpu_map) {
7013 sd = *per_cpu_ptr(d.sd, i);
7014 cpu_attach_domain(sd, d.rd, i);
7020 __free_domain_allocs(&d, alloc_state, cpu_map);
7024 static cpumask_var_t *doms_cur; /* current sched domains */
7025 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7026 static struct sched_domain_attr *dattr_cur;
7027 /* attribues of custom domains in 'doms_cur' */
7030 * Special case: If a kmalloc of a doms_cur partition (array of
7031 * cpumask) fails, then fallback to a single sched domain,
7032 * as determined by the single cpumask fallback_doms.
7034 static cpumask_var_t fallback_doms;
7037 * arch_update_cpu_topology lets virtualized architectures update the
7038 * cpu core maps. It is supposed to return 1 if the topology changed
7039 * or 0 if it stayed the same.
7041 int __weak arch_update_cpu_topology(void)
7046 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7049 cpumask_var_t *doms;
7051 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7054 for (i = 0; i < ndoms; i++) {
7055 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7056 free_sched_domains(doms, i);
7063 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7066 for (i = 0; i < ndoms; i++)
7067 free_cpumask_var(doms[i]);
7072 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7073 * For now this just excludes isolated cpus, but could be used to
7074 * exclude other special cases in the future.
7076 static int init_sched_domains(const struct cpumask *cpu_map)
7080 arch_update_cpu_topology();
7082 doms_cur = alloc_sched_domains(ndoms_cur);
7084 doms_cur = &fallback_doms;
7085 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7086 err = build_sched_domains(doms_cur[0], NULL);
7087 register_sched_domain_sysctl();
7093 * Detach sched domains from a group of cpus specified in cpu_map
7094 * These cpus will now be attached to the NULL domain
7096 static void detach_destroy_domains(const struct cpumask *cpu_map)
7101 for_each_cpu(i, cpu_map)
7102 cpu_attach_domain(NULL, &def_root_domain, i);
7106 /* handle null as "default" */
7107 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7108 struct sched_domain_attr *new, int idx_new)
7110 struct sched_domain_attr tmp;
7117 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7118 new ? (new + idx_new) : &tmp,
7119 sizeof(struct sched_domain_attr));
7123 * Partition sched domains as specified by the 'ndoms_new'
7124 * cpumasks in the array doms_new[] of cpumasks. This compares
7125 * doms_new[] to the current sched domain partitioning, doms_cur[].
7126 * It destroys each deleted domain and builds each new domain.
7128 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7129 * The masks don't intersect (don't overlap.) We should setup one
7130 * sched domain for each mask. CPUs not in any of the cpumasks will
7131 * not be load balanced. If the same cpumask appears both in the
7132 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7135 * The passed in 'doms_new' should be allocated using
7136 * alloc_sched_domains. This routine takes ownership of it and will
7137 * free_sched_domains it when done with it. If the caller failed the
7138 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7139 * and partition_sched_domains() will fallback to the single partition
7140 * 'fallback_doms', it also forces the domains to be rebuilt.
7142 * If doms_new == NULL it will be replaced with cpu_online_mask.
7143 * ndoms_new == 0 is a special case for destroying existing domains,
7144 * and it will not create the default domain.
7146 * Call with hotplug lock held
7148 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7149 struct sched_domain_attr *dattr_new)
7154 mutex_lock(&sched_domains_mutex);
7156 /* always unregister in case we don't destroy any domains */
7157 unregister_sched_domain_sysctl();
7159 /* Let architecture update cpu core mappings. */
7160 new_topology = arch_update_cpu_topology();
7162 n = doms_new ? ndoms_new : 0;
7164 /* Destroy deleted domains */
7165 for (i = 0; i < ndoms_cur; i++) {
7166 for (j = 0; j < n && !new_topology; j++) {
7167 if (cpumask_equal(doms_cur[i], doms_new[j])
7168 && dattrs_equal(dattr_cur, i, dattr_new, j))
7171 /* no match - a current sched domain not in new doms_new[] */
7172 detach_destroy_domains(doms_cur[i]);
7178 if (doms_new == NULL) {
7180 doms_new = &fallback_doms;
7181 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7182 WARN_ON_ONCE(dattr_new);
7185 /* Build new domains */
7186 for (i = 0; i < ndoms_new; i++) {
7187 for (j = 0; j < n && !new_topology; j++) {
7188 if (cpumask_equal(doms_new[i], doms_cur[j])
7189 && dattrs_equal(dattr_new, i, dattr_cur, j))
7192 /* no match - add a new doms_new */
7193 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7198 /* Remember the new sched domains */
7199 if (doms_cur != &fallback_doms)
7200 free_sched_domains(doms_cur, ndoms_cur);
7201 kfree(dattr_cur); /* kfree(NULL) is safe */
7202 doms_cur = doms_new;
7203 dattr_cur = dattr_new;
7204 ndoms_cur = ndoms_new;
7206 register_sched_domain_sysctl();
7208 mutex_unlock(&sched_domains_mutex);
7211 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7214 * Update cpusets according to cpu_active mask. If cpusets are
7215 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7216 * around partition_sched_domains().
7218 * If we come here as part of a suspend/resume, don't touch cpusets because we
7219 * want to restore it back to its original state upon resume anyway.
7221 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7225 case CPU_ONLINE_FROZEN:
7226 case CPU_DOWN_FAILED_FROZEN:
7229 * num_cpus_frozen tracks how many CPUs are involved in suspend
7230 * resume sequence. As long as this is not the last online
7231 * operation in the resume sequence, just build a single sched
7232 * domain, ignoring cpusets.
7235 if (likely(num_cpus_frozen)) {
7236 partition_sched_domains(1, NULL, NULL);
7241 * This is the last CPU online operation. So fall through and
7242 * restore the original sched domains by considering the
7243 * cpuset configurations.
7247 cpuset_update_active_cpus(true);
7255 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7258 unsigned long flags;
7259 long cpu = (long)hcpu;
7265 case CPU_DOWN_PREPARE:
7266 rcu_read_lock_sched();
7267 dl_b = dl_bw_of(cpu);
7269 raw_spin_lock_irqsave(&dl_b->lock, flags);
7270 cpus = dl_bw_cpus(cpu);
7271 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7272 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7274 rcu_read_unlock_sched();
7277 return notifier_from_errno(-EBUSY);
7278 cpuset_update_active_cpus(false);
7280 case CPU_DOWN_PREPARE_FROZEN:
7282 partition_sched_domains(1, NULL, NULL);
7290 void __init sched_init_smp(void)
7292 cpumask_var_t non_isolated_cpus;
7294 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7295 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7300 * There's no userspace yet to cause hotplug operations; hence all the
7301 * cpu masks are stable and all blatant races in the below code cannot
7304 mutex_lock(&sched_domains_mutex);
7305 init_sched_domains(cpu_active_mask);
7306 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7307 if (cpumask_empty(non_isolated_cpus))
7308 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7309 mutex_unlock(&sched_domains_mutex);
7311 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7312 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7313 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7317 /* Move init over to a non-isolated CPU */
7318 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7320 sched_init_granularity();
7321 free_cpumask_var(non_isolated_cpus);
7323 init_sched_rt_class();
7324 init_sched_dl_class();
7327 void __init sched_init_smp(void)
7329 sched_init_granularity();
7331 #endif /* CONFIG_SMP */
7333 int in_sched_functions(unsigned long addr)
7335 return in_lock_functions(addr) ||
7336 (addr >= (unsigned long)__sched_text_start
7337 && addr < (unsigned long)__sched_text_end);
7340 #ifdef CONFIG_CGROUP_SCHED
7342 * Default task group.
7343 * Every task in system belongs to this group at bootup.
7345 struct task_group root_task_group;
7346 LIST_HEAD(task_groups);
7348 /* Cacheline aligned slab cache for task_group */
7349 static struct kmem_cache *task_group_cache __read_mostly;
7352 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7354 void __init sched_init(void)
7357 unsigned long alloc_size = 0, ptr;
7359 #ifdef CONFIG_FAIR_GROUP_SCHED
7360 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7362 #ifdef CONFIG_RT_GROUP_SCHED
7363 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7366 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7368 #ifdef CONFIG_FAIR_GROUP_SCHED
7369 root_task_group.se = (struct sched_entity **)ptr;
7370 ptr += nr_cpu_ids * sizeof(void **);
7372 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7373 ptr += nr_cpu_ids * sizeof(void **);
7375 #endif /* CONFIG_FAIR_GROUP_SCHED */
7376 #ifdef CONFIG_RT_GROUP_SCHED
7377 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7378 ptr += nr_cpu_ids * sizeof(void **);
7380 root_task_group.rt_rq = (struct rt_rq **)ptr;
7381 ptr += nr_cpu_ids * sizeof(void **);
7383 #endif /* CONFIG_RT_GROUP_SCHED */
7385 #ifdef CONFIG_CPUMASK_OFFSTACK
7386 for_each_possible_cpu(i) {
7387 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7388 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7390 #endif /* CONFIG_CPUMASK_OFFSTACK */
7392 init_rt_bandwidth(&def_rt_bandwidth,
7393 global_rt_period(), global_rt_runtime());
7394 init_dl_bandwidth(&def_dl_bandwidth,
7395 global_rt_period(), global_rt_runtime());
7398 init_defrootdomain();
7401 #ifdef CONFIG_RT_GROUP_SCHED
7402 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7403 global_rt_period(), global_rt_runtime());
7404 #endif /* CONFIG_RT_GROUP_SCHED */
7406 #ifdef CONFIG_CGROUP_SCHED
7407 task_group_cache = KMEM_CACHE(task_group, 0);
7409 list_add(&root_task_group.list, &task_groups);
7410 INIT_LIST_HEAD(&root_task_group.children);
7411 INIT_LIST_HEAD(&root_task_group.siblings);
7412 autogroup_init(&init_task);
7413 #endif /* CONFIG_CGROUP_SCHED */
7415 for_each_possible_cpu(i) {
7419 raw_spin_lock_init(&rq->lock);
7421 rq->calc_load_active = 0;
7422 rq->calc_load_update = jiffies + LOAD_FREQ;
7423 init_cfs_rq(&rq->cfs);
7424 init_rt_rq(&rq->rt);
7425 init_dl_rq(&rq->dl);
7426 #ifdef CONFIG_FAIR_GROUP_SCHED
7427 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7428 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7430 * How much cpu bandwidth does root_task_group get?
7432 * In case of task-groups formed thr' the cgroup filesystem, it
7433 * gets 100% of the cpu resources in the system. This overall
7434 * system cpu resource is divided among the tasks of
7435 * root_task_group and its child task-groups in a fair manner,
7436 * based on each entity's (task or task-group's) weight
7437 * (se->load.weight).
7439 * In other words, if root_task_group has 10 tasks of weight
7440 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7441 * then A0's share of the cpu resource is:
7443 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7445 * We achieve this by letting root_task_group's tasks sit
7446 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7448 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7449 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7450 #endif /* CONFIG_FAIR_GROUP_SCHED */
7452 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7453 #ifdef CONFIG_RT_GROUP_SCHED
7454 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7457 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7458 rq->cpu_load[j] = 0;
7463 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7464 rq->balance_callback = NULL;
7465 rq->active_balance = 0;
7466 rq->next_balance = jiffies;
7471 rq->avg_idle = 2*sysctl_sched_migration_cost;
7472 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7474 INIT_LIST_HEAD(&rq->cfs_tasks);
7476 rq_attach_root(rq, &def_root_domain);
7477 #ifdef CONFIG_NO_HZ_COMMON
7478 rq->last_load_update_tick = jiffies;
7481 #ifdef CONFIG_NO_HZ_FULL
7482 rq->last_sched_tick = 0;
7484 #endif /* CONFIG_SMP */
7486 atomic_set(&rq->nr_iowait, 0);
7489 set_load_weight(&init_task);
7491 #ifdef CONFIG_PREEMPT_NOTIFIERS
7492 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7496 * The boot idle thread does lazy MMU switching as well:
7498 atomic_inc(&init_mm.mm_count);
7499 enter_lazy_tlb(&init_mm, current);
7502 * During early bootup we pretend to be a normal task:
7504 current->sched_class = &fair_sched_class;
7507 * Make us the idle thread. Technically, schedule() should not be
7508 * called from this thread, however somewhere below it might be,
7509 * but because we are the idle thread, we just pick up running again
7510 * when this runqueue becomes "idle".
7512 init_idle(current, smp_processor_id());
7514 calc_load_update = jiffies + LOAD_FREQ;
7517 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7518 /* May be allocated at isolcpus cmdline parse time */
7519 if (cpu_isolated_map == NULL)
7520 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7521 idle_thread_set_boot_cpu();
7522 set_cpu_rq_start_time();
7524 init_sched_fair_class();
7526 scheduler_running = 1;
7529 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7530 static inline int preempt_count_equals(int preempt_offset)
7532 int nested = preempt_count() + rcu_preempt_depth();
7534 return (nested == preempt_offset);
7537 void __might_sleep(const char *file, int line, int preempt_offset)
7540 * Blocking primitives will set (and therefore destroy) current->state,
7541 * since we will exit with TASK_RUNNING make sure we enter with it,
7542 * otherwise we will destroy state.
7544 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7545 "do not call blocking ops when !TASK_RUNNING; "
7546 "state=%lx set at [<%p>] %pS\n",
7548 (void *)current->task_state_change,
7549 (void *)current->task_state_change);
7551 ___might_sleep(file, line, preempt_offset);
7553 EXPORT_SYMBOL(__might_sleep);
7555 void ___might_sleep(const char *file, int line, int preempt_offset)
7557 static unsigned long prev_jiffy; /* ratelimiting */
7559 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7560 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7561 !is_idle_task(current)) ||
7562 system_state != SYSTEM_RUNNING || oops_in_progress)
7564 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7566 prev_jiffy = jiffies;
7569 "BUG: sleeping function called from invalid context at %s:%d\n",
7572 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7573 in_atomic(), irqs_disabled(),
7574 current->pid, current->comm);
7576 if (task_stack_end_corrupted(current))
7577 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7579 debug_show_held_locks(current);
7580 if (irqs_disabled())
7581 print_irqtrace_events(current);
7582 #ifdef CONFIG_DEBUG_PREEMPT
7583 if (!preempt_count_equals(preempt_offset)) {
7584 pr_err("Preemption disabled at:");
7585 print_ip_sym(current->preempt_disable_ip);
7591 EXPORT_SYMBOL(___might_sleep);
7594 #ifdef CONFIG_MAGIC_SYSRQ
7595 void normalize_rt_tasks(void)
7597 struct task_struct *g, *p;
7598 struct sched_attr attr = {
7599 .sched_policy = SCHED_NORMAL,
7602 read_lock(&tasklist_lock);
7603 for_each_process_thread(g, p) {
7605 * Only normalize user tasks:
7607 if (p->flags & PF_KTHREAD)
7610 p->se.exec_start = 0;
7611 #ifdef CONFIG_SCHEDSTATS
7612 p->se.statistics.wait_start = 0;
7613 p->se.statistics.sleep_start = 0;
7614 p->se.statistics.block_start = 0;
7617 if (!dl_task(p) && !rt_task(p)) {
7619 * Renice negative nice level userspace
7622 if (task_nice(p) < 0)
7623 set_user_nice(p, 0);
7627 __sched_setscheduler(p, &attr, false, false);
7629 read_unlock(&tasklist_lock);
7632 #endif /* CONFIG_MAGIC_SYSRQ */
7634 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7636 * These functions are only useful for the IA64 MCA handling, or kdb.
7638 * They can only be called when the whole system has been
7639 * stopped - every CPU needs to be quiescent, and no scheduling
7640 * activity can take place. Using them for anything else would
7641 * be a serious bug, and as a result, they aren't even visible
7642 * under any other configuration.
7646 * curr_task - return the current task for a given cpu.
7647 * @cpu: the processor in question.
7649 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7651 * Return: The current task for @cpu.
7653 struct task_struct *curr_task(int cpu)
7655 return cpu_curr(cpu);
7658 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7662 * set_curr_task - set the current task for a given cpu.
7663 * @cpu: the processor in question.
7664 * @p: the task pointer to set.
7666 * Description: This function must only be used when non-maskable interrupts
7667 * are serviced on a separate stack. It allows the architecture to switch the
7668 * notion of the current task on a cpu in a non-blocking manner. This function
7669 * must be called with all CPU's synchronized, and interrupts disabled, the
7670 * and caller must save the original value of the current task (see
7671 * curr_task() above) and restore that value before reenabling interrupts and
7672 * re-starting the system.
7674 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7676 void set_curr_task(int cpu, struct task_struct *p)
7683 #ifdef CONFIG_CGROUP_SCHED
7684 /* task_group_lock serializes the addition/removal of task groups */
7685 static DEFINE_SPINLOCK(task_group_lock);
7687 static void sched_free_group(struct task_group *tg)
7689 free_fair_sched_group(tg);
7690 free_rt_sched_group(tg);
7692 kmem_cache_free(task_group_cache, tg);
7695 /* allocate runqueue etc for a new task group */
7696 struct task_group *sched_create_group(struct task_group *parent)
7698 struct task_group *tg;
7700 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7702 return ERR_PTR(-ENOMEM);
7704 if (!alloc_fair_sched_group(tg, parent))
7707 if (!alloc_rt_sched_group(tg, parent))
7713 sched_free_group(tg);
7714 return ERR_PTR(-ENOMEM);
7717 void sched_online_group(struct task_group *tg, struct task_group *parent)
7719 unsigned long flags;
7721 spin_lock_irqsave(&task_group_lock, flags);
7722 list_add_rcu(&tg->list, &task_groups);
7724 WARN_ON(!parent); /* root should already exist */
7726 tg->parent = parent;
7727 INIT_LIST_HEAD(&tg->children);
7728 list_add_rcu(&tg->siblings, &parent->children);
7729 spin_unlock_irqrestore(&task_group_lock, flags);
7732 /* rcu callback to free various structures associated with a task group */
7733 static void sched_free_group_rcu(struct rcu_head *rhp)
7735 /* now it should be safe to free those cfs_rqs */
7736 sched_free_group(container_of(rhp, struct task_group, rcu));
7739 void sched_destroy_group(struct task_group *tg)
7741 /* wait for possible concurrent references to cfs_rqs complete */
7742 call_rcu(&tg->rcu, sched_free_group_rcu);
7745 void sched_offline_group(struct task_group *tg)
7747 unsigned long flags;
7749 /* end participation in shares distribution */
7750 unregister_fair_sched_group(tg);
7752 spin_lock_irqsave(&task_group_lock, flags);
7753 list_del_rcu(&tg->list);
7754 list_del_rcu(&tg->siblings);
7755 spin_unlock_irqrestore(&task_group_lock, flags);
7758 /* change task's runqueue when it moves between groups.
7759 * The caller of this function should have put the task in its new group
7760 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7761 * reflect its new group.
7763 void sched_move_task(struct task_struct *tsk)
7765 struct task_group *tg;
7766 int queued, running;
7770 rq = task_rq_lock(tsk, &rf);
7772 running = task_current(rq, tsk);
7773 queued = task_on_rq_queued(tsk);
7776 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7777 if (unlikely(running))
7778 put_prev_task(rq, tsk);
7781 * All callers are synchronized by task_rq_lock(); we do not use RCU
7782 * which is pointless here. Thus, we pass "true" to task_css_check()
7783 * to prevent lockdep warnings.
7785 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7786 struct task_group, css);
7787 tg = autogroup_task_group(tsk, tg);
7788 tsk->sched_task_group = tg;
7790 #ifdef CONFIG_FAIR_GROUP_SCHED
7791 if (tsk->sched_class->task_move_group)
7792 tsk->sched_class->task_move_group(tsk);
7795 set_task_rq(tsk, task_cpu(tsk));
7797 if (unlikely(running))
7798 tsk->sched_class->set_curr_task(rq);
7800 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7802 task_rq_unlock(rq, tsk, &rf);
7804 #endif /* CONFIG_CGROUP_SCHED */
7806 #ifdef CONFIG_RT_GROUP_SCHED
7808 * Ensure that the real time constraints are schedulable.
7810 static DEFINE_MUTEX(rt_constraints_mutex);
7812 /* Must be called with tasklist_lock held */
7813 static inline int tg_has_rt_tasks(struct task_group *tg)
7815 struct task_struct *g, *p;
7818 * Autogroups do not have RT tasks; see autogroup_create().
7820 if (task_group_is_autogroup(tg))
7823 for_each_process_thread(g, p) {
7824 if (rt_task(p) && task_group(p) == tg)
7831 struct rt_schedulable_data {
7832 struct task_group *tg;
7837 static int tg_rt_schedulable(struct task_group *tg, void *data)
7839 struct rt_schedulable_data *d = data;
7840 struct task_group *child;
7841 unsigned long total, sum = 0;
7842 u64 period, runtime;
7844 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7845 runtime = tg->rt_bandwidth.rt_runtime;
7848 period = d->rt_period;
7849 runtime = d->rt_runtime;
7853 * Cannot have more runtime than the period.
7855 if (runtime > period && runtime != RUNTIME_INF)
7859 * Ensure we don't starve existing RT tasks.
7861 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7864 total = to_ratio(period, runtime);
7867 * Nobody can have more than the global setting allows.
7869 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7873 * The sum of our children's runtime should not exceed our own.
7875 list_for_each_entry_rcu(child, &tg->children, siblings) {
7876 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7877 runtime = child->rt_bandwidth.rt_runtime;
7879 if (child == d->tg) {
7880 period = d->rt_period;
7881 runtime = d->rt_runtime;
7884 sum += to_ratio(period, runtime);
7893 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7897 struct rt_schedulable_data data = {
7899 .rt_period = period,
7900 .rt_runtime = runtime,
7904 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7910 static int tg_set_rt_bandwidth(struct task_group *tg,
7911 u64 rt_period, u64 rt_runtime)
7916 * Disallowing the root group RT runtime is BAD, it would disallow the
7917 * kernel creating (and or operating) RT threads.
7919 if (tg == &root_task_group && rt_runtime == 0)
7922 /* No period doesn't make any sense. */
7926 mutex_lock(&rt_constraints_mutex);
7927 read_lock(&tasklist_lock);
7928 err = __rt_schedulable(tg, rt_period, rt_runtime);
7932 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7933 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7934 tg->rt_bandwidth.rt_runtime = rt_runtime;
7936 for_each_possible_cpu(i) {
7937 struct rt_rq *rt_rq = tg->rt_rq[i];
7939 raw_spin_lock(&rt_rq->rt_runtime_lock);
7940 rt_rq->rt_runtime = rt_runtime;
7941 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7943 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7945 read_unlock(&tasklist_lock);
7946 mutex_unlock(&rt_constraints_mutex);
7951 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7953 u64 rt_runtime, rt_period;
7955 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7956 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7957 if (rt_runtime_us < 0)
7958 rt_runtime = RUNTIME_INF;
7960 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7963 static long sched_group_rt_runtime(struct task_group *tg)
7967 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7970 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7971 do_div(rt_runtime_us, NSEC_PER_USEC);
7972 return rt_runtime_us;
7975 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7977 u64 rt_runtime, rt_period;
7979 rt_period = rt_period_us * NSEC_PER_USEC;
7980 rt_runtime = tg->rt_bandwidth.rt_runtime;
7982 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7985 static long sched_group_rt_period(struct task_group *tg)
7989 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7990 do_div(rt_period_us, NSEC_PER_USEC);
7991 return rt_period_us;
7993 #endif /* CONFIG_RT_GROUP_SCHED */
7995 #ifdef CONFIG_RT_GROUP_SCHED
7996 static int sched_rt_global_constraints(void)
8000 mutex_lock(&rt_constraints_mutex);
8001 read_lock(&tasklist_lock);
8002 ret = __rt_schedulable(NULL, 0, 0);
8003 read_unlock(&tasklist_lock);
8004 mutex_unlock(&rt_constraints_mutex);
8009 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8011 /* Don't accept realtime tasks when there is no way for them to run */
8012 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8018 #else /* !CONFIG_RT_GROUP_SCHED */
8019 static int sched_rt_global_constraints(void)
8021 unsigned long flags;
8024 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8025 for_each_possible_cpu(i) {
8026 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8028 raw_spin_lock(&rt_rq->rt_runtime_lock);
8029 rt_rq->rt_runtime = global_rt_runtime();
8030 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8032 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8036 #endif /* CONFIG_RT_GROUP_SCHED */
8038 static int sched_dl_global_validate(void)
8040 u64 runtime = global_rt_runtime();
8041 u64 period = global_rt_period();
8042 u64 new_bw = to_ratio(period, runtime);
8045 unsigned long flags;
8048 * Here we want to check the bandwidth not being set to some
8049 * value smaller than the currently allocated bandwidth in
8050 * any of the root_domains.
8052 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8053 * cycling on root_domains... Discussion on different/better
8054 * solutions is welcome!
8056 for_each_possible_cpu(cpu) {
8057 rcu_read_lock_sched();
8058 dl_b = dl_bw_of(cpu);
8060 raw_spin_lock_irqsave(&dl_b->lock, flags);
8061 if (new_bw < dl_b->total_bw)
8063 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8065 rcu_read_unlock_sched();
8074 static void sched_dl_do_global(void)
8079 unsigned long flags;
8081 def_dl_bandwidth.dl_period = global_rt_period();
8082 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8084 if (global_rt_runtime() != RUNTIME_INF)
8085 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8088 * FIXME: As above...
8090 for_each_possible_cpu(cpu) {
8091 rcu_read_lock_sched();
8092 dl_b = dl_bw_of(cpu);
8094 raw_spin_lock_irqsave(&dl_b->lock, flags);
8096 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8098 rcu_read_unlock_sched();
8102 static int sched_rt_global_validate(void)
8104 if (sysctl_sched_rt_period <= 0)
8107 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8108 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8114 static void sched_rt_do_global(void)
8116 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8117 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8120 int sched_rt_handler(struct ctl_table *table, int write,
8121 void __user *buffer, size_t *lenp,
8124 int old_period, old_runtime;
8125 static DEFINE_MUTEX(mutex);
8129 old_period = sysctl_sched_rt_period;
8130 old_runtime = sysctl_sched_rt_runtime;
8132 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8134 if (!ret && write) {
8135 ret = sched_rt_global_validate();
8139 ret = sched_dl_global_validate();
8143 ret = sched_rt_global_constraints();
8147 sched_rt_do_global();
8148 sched_dl_do_global();
8152 sysctl_sched_rt_period = old_period;
8153 sysctl_sched_rt_runtime = old_runtime;
8155 mutex_unlock(&mutex);
8160 int sched_rr_handler(struct ctl_table *table, int write,
8161 void __user *buffer, size_t *lenp,
8165 static DEFINE_MUTEX(mutex);
8168 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8169 /* make sure that internally we keep jiffies */
8170 /* also, writing zero resets timeslice to default */
8171 if (!ret && write) {
8172 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8173 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8175 mutex_unlock(&mutex);
8179 #ifdef CONFIG_CGROUP_SCHED
8181 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8183 return css ? container_of(css, struct task_group, css) : NULL;
8186 static struct cgroup_subsys_state *
8187 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8189 struct task_group *parent = css_tg(parent_css);
8190 struct task_group *tg;
8193 /* This is early initialization for the top cgroup */
8194 return &root_task_group.css;
8197 tg = sched_create_group(parent);
8199 return ERR_PTR(-ENOMEM);
8201 sched_online_group(tg, parent);
8206 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8208 struct task_group *tg = css_tg(css);
8210 sched_offline_group(tg);
8213 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8215 struct task_group *tg = css_tg(css);
8218 * Relies on the RCU grace period between css_released() and this.
8220 sched_free_group(tg);
8223 static void cpu_cgroup_fork(struct task_struct *task)
8225 sched_move_task(task);
8228 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8230 struct task_struct *task;
8231 struct cgroup_subsys_state *css;
8233 cgroup_taskset_for_each(task, css, tset) {
8234 #ifdef CONFIG_RT_GROUP_SCHED
8235 if (!sched_rt_can_attach(css_tg(css), task))
8238 /* We don't support RT-tasks being in separate groups */
8239 if (task->sched_class != &fair_sched_class)
8246 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8248 struct task_struct *task;
8249 struct cgroup_subsys_state *css;
8251 cgroup_taskset_for_each(task, css, tset)
8252 sched_move_task(task);
8255 #ifdef CONFIG_FAIR_GROUP_SCHED
8256 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8257 struct cftype *cftype, u64 shareval)
8259 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8262 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8265 struct task_group *tg = css_tg(css);
8267 return (u64) scale_load_down(tg->shares);
8270 #ifdef CONFIG_CFS_BANDWIDTH
8271 static DEFINE_MUTEX(cfs_constraints_mutex);
8273 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8274 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8276 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8278 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8280 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8281 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8283 if (tg == &root_task_group)
8287 * Ensure we have at some amount of bandwidth every period. This is
8288 * to prevent reaching a state of large arrears when throttled via
8289 * entity_tick() resulting in prolonged exit starvation.
8291 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8295 * Likewise, bound things on the otherside by preventing insane quota
8296 * periods. This also allows us to normalize in computing quota
8299 if (period > max_cfs_quota_period)
8303 * Prevent race between setting of cfs_rq->runtime_enabled and
8304 * unthrottle_offline_cfs_rqs().
8307 mutex_lock(&cfs_constraints_mutex);
8308 ret = __cfs_schedulable(tg, period, quota);
8312 runtime_enabled = quota != RUNTIME_INF;
8313 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8315 * If we need to toggle cfs_bandwidth_used, off->on must occur
8316 * before making related changes, and on->off must occur afterwards
8318 if (runtime_enabled && !runtime_was_enabled)
8319 cfs_bandwidth_usage_inc();
8320 raw_spin_lock_irq(&cfs_b->lock);
8321 cfs_b->period = ns_to_ktime(period);
8322 cfs_b->quota = quota;
8324 __refill_cfs_bandwidth_runtime(cfs_b);
8325 /* restart the period timer (if active) to handle new period expiry */
8326 if (runtime_enabled)
8327 start_cfs_bandwidth(cfs_b);
8328 raw_spin_unlock_irq(&cfs_b->lock);
8330 for_each_online_cpu(i) {
8331 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8332 struct rq *rq = cfs_rq->rq;
8334 raw_spin_lock_irq(&rq->lock);
8335 cfs_rq->runtime_enabled = runtime_enabled;
8336 cfs_rq->runtime_remaining = 0;
8338 if (cfs_rq->throttled)
8339 unthrottle_cfs_rq(cfs_rq);
8340 raw_spin_unlock_irq(&rq->lock);
8342 if (runtime_was_enabled && !runtime_enabled)
8343 cfs_bandwidth_usage_dec();
8345 mutex_unlock(&cfs_constraints_mutex);
8351 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8355 period = ktime_to_ns(tg->cfs_bandwidth.period);
8356 if (cfs_quota_us < 0)
8357 quota = RUNTIME_INF;
8359 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8361 return tg_set_cfs_bandwidth(tg, period, quota);
8364 long tg_get_cfs_quota(struct task_group *tg)
8368 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8371 quota_us = tg->cfs_bandwidth.quota;
8372 do_div(quota_us, NSEC_PER_USEC);
8377 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8381 period = (u64)cfs_period_us * NSEC_PER_USEC;
8382 quota = tg->cfs_bandwidth.quota;
8384 return tg_set_cfs_bandwidth(tg, period, quota);
8387 long tg_get_cfs_period(struct task_group *tg)
8391 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8392 do_div(cfs_period_us, NSEC_PER_USEC);
8394 return cfs_period_us;
8397 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8400 return tg_get_cfs_quota(css_tg(css));
8403 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8404 struct cftype *cftype, s64 cfs_quota_us)
8406 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8409 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8412 return tg_get_cfs_period(css_tg(css));
8415 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8416 struct cftype *cftype, u64 cfs_period_us)
8418 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8421 struct cfs_schedulable_data {
8422 struct task_group *tg;
8427 * normalize group quota/period to be quota/max_period
8428 * note: units are usecs
8430 static u64 normalize_cfs_quota(struct task_group *tg,
8431 struct cfs_schedulable_data *d)
8439 period = tg_get_cfs_period(tg);
8440 quota = tg_get_cfs_quota(tg);
8443 /* note: these should typically be equivalent */
8444 if (quota == RUNTIME_INF || quota == -1)
8447 return to_ratio(period, quota);
8450 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8452 struct cfs_schedulable_data *d = data;
8453 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8454 s64 quota = 0, parent_quota = -1;
8457 quota = RUNTIME_INF;
8459 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8461 quota = normalize_cfs_quota(tg, d);
8462 parent_quota = parent_b->hierarchical_quota;
8465 * ensure max(child_quota) <= parent_quota, inherit when no
8468 if (quota == RUNTIME_INF)
8469 quota = parent_quota;
8470 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8473 cfs_b->hierarchical_quota = quota;
8478 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8481 struct cfs_schedulable_data data = {
8487 if (quota != RUNTIME_INF) {
8488 do_div(data.period, NSEC_PER_USEC);
8489 do_div(data.quota, NSEC_PER_USEC);
8493 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8499 static int cpu_stats_show(struct seq_file *sf, void *v)
8501 struct task_group *tg = css_tg(seq_css(sf));
8502 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8504 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8505 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8506 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8510 #endif /* CONFIG_CFS_BANDWIDTH */
8511 #endif /* CONFIG_FAIR_GROUP_SCHED */
8513 #ifdef CONFIG_RT_GROUP_SCHED
8514 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8515 struct cftype *cft, s64 val)
8517 return sched_group_set_rt_runtime(css_tg(css), val);
8520 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8523 return sched_group_rt_runtime(css_tg(css));
8526 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8527 struct cftype *cftype, u64 rt_period_us)
8529 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8532 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8535 return sched_group_rt_period(css_tg(css));
8537 #endif /* CONFIG_RT_GROUP_SCHED */
8539 static struct cftype cpu_files[] = {
8540 #ifdef CONFIG_FAIR_GROUP_SCHED
8543 .read_u64 = cpu_shares_read_u64,
8544 .write_u64 = cpu_shares_write_u64,
8547 #ifdef CONFIG_CFS_BANDWIDTH
8549 .name = "cfs_quota_us",
8550 .read_s64 = cpu_cfs_quota_read_s64,
8551 .write_s64 = cpu_cfs_quota_write_s64,
8554 .name = "cfs_period_us",
8555 .read_u64 = cpu_cfs_period_read_u64,
8556 .write_u64 = cpu_cfs_period_write_u64,
8560 .seq_show = cpu_stats_show,
8563 #ifdef CONFIG_RT_GROUP_SCHED
8565 .name = "rt_runtime_us",
8566 .read_s64 = cpu_rt_runtime_read,
8567 .write_s64 = cpu_rt_runtime_write,
8570 .name = "rt_period_us",
8571 .read_u64 = cpu_rt_period_read_uint,
8572 .write_u64 = cpu_rt_period_write_uint,
8578 struct cgroup_subsys cpu_cgrp_subsys = {
8579 .css_alloc = cpu_cgroup_css_alloc,
8580 .css_released = cpu_cgroup_css_released,
8581 .css_free = cpu_cgroup_css_free,
8582 .fork = cpu_cgroup_fork,
8583 .can_attach = cpu_cgroup_can_attach,
8584 .attach = cpu_cgroup_attach,
8585 .legacy_cftypes = cpu_files,
8589 #endif /* CONFIG_CGROUP_SCHED */
8591 void dump_cpu_task(int cpu)
8593 pr_info("Task dump for CPU %d:\n", cpu);
8594 sched_show_task(cpu_curr(cpu));
8598 * Nice levels are multiplicative, with a gentle 10% change for every
8599 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8600 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8601 * that remained on nice 0.
8603 * The "10% effect" is relative and cumulative: from _any_ nice level,
8604 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8605 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8606 * If a task goes up by ~10% and another task goes down by ~10% then
8607 * the relative distance between them is ~25%.)
8609 const int sched_prio_to_weight[40] = {
8610 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8611 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8612 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8613 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8614 /* 0 */ 1024, 820, 655, 526, 423,
8615 /* 5 */ 335, 272, 215, 172, 137,
8616 /* 10 */ 110, 87, 70, 56, 45,
8617 /* 15 */ 36, 29, 23, 18, 15,
8621 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8623 * In cases where the weight does not change often, we can use the
8624 * precalculated inverse to speed up arithmetics by turning divisions
8625 * into multiplications:
8627 const u32 sched_prio_to_wmult[40] = {
8628 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8629 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8630 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8631 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8632 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8633 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8634 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8635 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,