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
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.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/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
375 static void hrtick_clear(struct rq *rq)
377 if (hrtimer_active(&rq->hrtick_timer))
378 hrtimer_cancel(&rq->hrtick_timer);
382 * High-resolution timer tick.
383 * Runs from hardirq context with interrupts disabled.
385 static enum hrtimer_restart hrtick(struct hrtimer *timer)
387 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
391 raw_spin_lock(&rq->lock);
393 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
394 raw_spin_unlock(&rq->lock);
396 return HRTIMER_NORESTART;
401 static int __hrtick_restart(struct rq *rq)
403 struct hrtimer *timer = &rq->hrtick_timer;
404 ktime_t time = hrtimer_get_softexpires(timer);
406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
410 * called from hardirq (IPI) context
412 static void __hrtick_start(void *arg)
416 raw_spin_lock(&rq->lock);
417 __hrtick_restart(rq);
418 rq->hrtick_csd_pending = 0;
419 raw_spin_unlock(&rq->lock);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq *rq, u64 delay)
429 struct hrtimer *timer = &rq->hrtick_timer;
430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
432 hrtimer_set_expires(timer, time);
434 if (rq == this_rq()) {
435 __hrtick_restart(rq);
436 } else if (!rq->hrtick_csd_pending) {
437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
438 rq->hrtick_csd_pending = 1;
443 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
445 int cpu = (int)(long)hcpu;
448 case CPU_UP_CANCELED:
449 case CPU_UP_CANCELED_FROZEN:
450 case CPU_DOWN_PREPARE:
451 case CPU_DOWN_PREPARE_FROZEN:
453 case CPU_DEAD_FROZEN:
454 hrtick_clear(cpu_rq(cpu));
461 static __init void init_hrtick(void)
463 hotcpu_notifier(hotplug_hrtick, 0);
467 * Called to set the hrtick timer state.
469 * called with rq->lock held and irqs disabled
471 void hrtick_start(struct rq *rq, u64 delay)
473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
474 HRTIMER_MODE_REL_PINNED, 0);
477 static inline void init_hrtick(void)
480 #endif /* CONFIG_SMP */
482 static void init_rq_hrtick(struct rq *rq)
485 rq->hrtick_csd_pending = 0;
487 rq->hrtick_csd.flags = 0;
488 rq->hrtick_csd.func = __hrtick_start;
489 rq->hrtick_csd.info = rq;
492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
493 rq->hrtick_timer.function = hrtick;
495 #else /* CONFIG_SCHED_HRTICK */
496 static inline void hrtick_clear(struct rq *rq)
500 static inline void init_rq_hrtick(struct rq *rq)
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SCHED_HRTICK */
510 * resched_task - mark a task 'to be rescheduled now'.
512 * On UP this means the setting of the need_resched flag, on SMP it
513 * might also involve a cross-CPU call to trigger the scheduler on
516 void resched_task(struct task_struct *p)
520 lockdep_assert_held(&task_rq(p)->lock);
522 if (test_tsk_need_resched(p))
525 set_tsk_need_resched(p);
528 if (cpu == smp_processor_id()) {
529 set_preempt_need_resched();
533 /* NEED_RESCHED must be visible before we test polling */
535 if (!tsk_is_polling(p))
536 smp_send_reschedule(cpu);
539 void resched_cpu(int cpu)
541 struct rq *rq = cpu_rq(cpu);
544 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
546 resched_task(cpu_curr(cpu));
547 raw_spin_unlock_irqrestore(&rq->lock, flags);
551 #ifdef CONFIG_NO_HZ_COMMON
553 * In the semi idle case, use the nearest busy cpu for migrating timers
554 * from an idle cpu. This is good for power-savings.
556 * We don't do similar optimization for completely idle system, as
557 * selecting an idle cpu will add more delays to the timers than intended
558 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 int get_nohz_timer_target(void)
562 int cpu = smp_processor_id();
564 struct sched_domain *sd;
567 for_each_domain(cpu, sd) {
568 for_each_cpu(i, sched_domain_span(sd)) {
580 * When add_timer_on() enqueues a timer into the timer wheel of an
581 * idle CPU then this timer might expire before the next timer event
582 * which is scheduled to wake up that CPU. In case of a completely
583 * idle system the next event might even be infinite time into the
584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
585 * leaves the inner idle loop so the newly added timer is taken into
586 * account when the CPU goes back to idle and evaluates the timer
587 * wheel for the next timer event.
589 static void wake_up_idle_cpu(int cpu)
591 struct rq *rq = cpu_rq(cpu);
593 if (cpu == smp_processor_id())
597 * This is safe, as this function is called with the timer
598 * wheel base lock of (cpu) held. When the CPU is on the way
599 * to idle and has not yet set rq->curr to idle then it will
600 * be serialized on the timer wheel base lock and take the new
601 * timer into account automatically.
603 if (rq->curr != rq->idle)
607 * We can set TIF_RESCHED on the idle task of the other CPU
608 * lockless. The worst case is that the other CPU runs the
609 * idle task through an additional NOOP schedule()
611 set_tsk_need_resched(rq->idle);
613 /* NEED_RESCHED must be visible before we test polling */
615 if (!tsk_is_polling(rq->idle))
616 smp_send_reschedule(cpu);
619 static bool wake_up_full_nohz_cpu(int cpu)
621 if (tick_nohz_full_cpu(cpu)) {
622 if (cpu != smp_processor_id() ||
623 tick_nohz_tick_stopped())
624 smp_send_reschedule(cpu);
631 void wake_up_nohz_cpu(int cpu)
633 if (!wake_up_full_nohz_cpu(cpu))
634 wake_up_idle_cpu(cpu);
637 static inline bool got_nohz_idle_kick(void)
639 int cpu = smp_processor_id();
641 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644 if (idle_cpu(cpu) && !need_resched())
648 * We can't run Idle Load Balance on this CPU for this time so we
649 * cancel it and clear NOHZ_BALANCE_KICK
651 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
655 #else /* CONFIG_NO_HZ_COMMON */
657 static inline bool got_nohz_idle_kick(void)
662 #endif /* CONFIG_NO_HZ_COMMON */
664 #ifdef CONFIG_NO_HZ_FULL
665 bool sched_can_stop_tick(void)
671 /* Make sure rq->nr_running update is visible after the IPI */
674 /* More than one running task need preemption */
675 if (rq->nr_running > 1)
680 #endif /* CONFIG_NO_HZ_FULL */
682 void sched_avg_update(struct rq *rq)
684 s64 period = sched_avg_period();
686 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
688 * Inline assembly required to prevent the compiler
689 * optimising this loop into a divmod call.
690 * See __iter_div_u64_rem() for another example of this.
692 asm("" : "+rm" (rq->age_stamp));
693 rq->age_stamp += period;
698 #endif /* CONFIG_SMP */
700 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
703 * Iterate task_group tree rooted at *from, calling @down when first entering a
704 * node and @up when leaving it for the final time.
706 * Caller must hold rcu_lock or sufficient equivalent.
708 int walk_tg_tree_from(struct task_group *from,
709 tg_visitor down, tg_visitor up, void *data)
711 struct task_group *parent, *child;
717 ret = (*down)(parent, data);
720 list_for_each_entry_rcu(child, &parent->children, siblings) {
727 ret = (*up)(parent, data);
728 if (ret || parent == from)
732 parent = parent->parent;
739 int tg_nop(struct task_group *tg, void *data)
745 static void set_load_weight(struct task_struct *p)
747 int prio = p->static_prio - MAX_RT_PRIO;
748 struct load_weight *load = &p->se.load;
751 * SCHED_IDLE tasks get minimal weight:
753 if (p->policy == SCHED_IDLE) {
754 load->weight = scale_load(WEIGHT_IDLEPRIO);
755 load->inv_weight = WMULT_IDLEPRIO;
759 load->weight = scale_load(prio_to_weight[prio]);
760 load->inv_weight = prio_to_wmult[prio];
763 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766 sched_info_queued(rq, p);
767 p->sched_class->enqueue_task(rq, p, flags);
770 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773 sched_info_dequeued(rq, p);
774 p->sched_class->dequeue_task(rq, p, flags);
777 void activate_task(struct rq *rq, struct task_struct *p, int flags)
779 if (task_contributes_to_load(p))
780 rq->nr_uninterruptible--;
782 enqueue_task(rq, p, flags);
785 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
787 if (task_contributes_to_load(p))
788 rq->nr_uninterruptible++;
790 dequeue_task(rq, p, flags);
793 static void update_rq_clock_task(struct rq *rq, s64 delta)
796 * In theory, the compile should just see 0 here, and optimize out the call
797 * to sched_rt_avg_update. But I don't trust it...
799 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
800 s64 steal = 0, irq_delta = 0;
802 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
803 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806 * Since irq_time is only updated on {soft,}irq_exit, we might run into
807 * this case when a previous update_rq_clock() happened inside a
810 * When this happens, we stop ->clock_task and only update the
811 * prev_irq_time stamp to account for the part that fit, so that a next
812 * update will consume the rest. This ensures ->clock_task is
815 * It does however cause some slight miss-attribution of {soft,}irq
816 * time, a more accurate solution would be to update the irq_time using
817 * the current rq->clock timestamp, except that would require using
820 if (irq_delta > delta)
823 rq->prev_irq_time += irq_delta;
826 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
827 if (static_key_false((¶virt_steal_rq_enabled))) {
830 steal = paravirt_steal_clock(cpu_of(rq));
831 steal -= rq->prev_steal_time_rq;
833 if (unlikely(steal > delta))
836 st = steal_ticks(steal);
837 steal = st * TICK_NSEC;
839 rq->prev_steal_time_rq += steal;
845 rq->clock_task += delta;
847 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
848 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
849 sched_rt_avg_update(rq, irq_delta + steal);
853 void sched_set_stop_task(int cpu, struct task_struct *stop)
855 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
856 struct task_struct *old_stop = cpu_rq(cpu)->stop;
860 * Make it appear like a SCHED_FIFO task, its something
861 * userspace knows about and won't get confused about.
863 * Also, it will make PI more or less work without too
864 * much confusion -- but then, stop work should not
865 * rely on PI working anyway.
867 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
869 stop->sched_class = &stop_sched_class;
872 cpu_rq(cpu)->stop = stop;
876 * Reset it back to a normal scheduling class so that
877 * it can die in pieces.
879 old_stop->sched_class = &rt_sched_class;
884 * __normal_prio - return the priority that is based on the static prio
886 static inline int __normal_prio(struct task_struct *p)
888 return p->static_prio;
892 * Calculate the expected normal priority: i.e. priority
893 * without taking RT-inheritance into account. Might be
894 * boosted by interactivity modifiers. Changes upon fork,
895 * setprio syscalls, and whenever the interactivity
896 * estimator recalculates.
898 static inline int normal_prio(struct task_struct *p)
902 if (task_has_dl_policy(p))
903 prio = MAX_DL_PRIO-1;
904 else if (task_has_rt_policy(p))
905 prio = MAX_RT_PRIO-1 - p->rt_priority;
907 prio = __normal_prio(p);
912 * Calculate the current priority, i.e. the priority
913 * taken into account by the scheduler. This value might
914 * be boosted by RT tasks, or might be boosted by
915 * interactivity modifiers. Will be RT if the task got
916 * RT-boosted. If not then it returns p->normal_prio.
918 static int effective_prio(struct task_struct *p)
920 p->normal_prio = normal_prio(p);
922 * If we are RT tasks or we were boosted to RT priority,
923 * keep the priority unchanged. Otherwise, update priority
924 * to the normal priority:
926 if (!rt_prio(p->prio))
927 return p->normal_prio;
932 * task_curr - is this task currently executing on a CPU?
933 * @p: the task in question.
935 * Return: 1 if the task is currently executing. 0 otherwise.
937 inline int task_curr(const struct task_struct *p)
939 return cpu_curr(task_cpu(p)) == p;
942 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
943 const struct sched_class *prev_class,
946 if (prev_class != p->sched_class) {
947 if (prev_class->switched_from)
948 prev_class->switched_from(rq, p);
949 p->sched_class->switched_to(rq, p);
950 } else if (oldprio != p->prio)
951 p->sched_class->prio_changed(rq, p, oldprio);
954 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
956 const struct sched_class *class;
958 if (p->sched_class == rq->curr->sched_class) {
959 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
961 for_each_class(class) {
962 if (class == rq->curr->sched_class)
964 if (class == p->sched_class) {
965 resched_task(rq->curr);
972 * A queue event has occurred, and we're going to schedule. In
973 * this case, we can save a useless back to back clock update.
975 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
976 rq->skip_clock_update = 1;
980 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
982 #ifdef CONFIG_SCHED_DEBUG
984 * We should never call set_task_cpu() on a blocked task,
985 * ttwu() will sort out the placement.
987 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
988 !(task_preempt_count(p) & PREEMPT_ACTIVE));
990 #ifdef CONFIG_LOCKDEP
992 * The caller should hold either p->pi_lock or rq->lock, when changing
993 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
995 * sched_move_task() holds both and thus holding either pins the cgroup,
998 * Furthermore, all task_rq users should acquire both locks, see
1001 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1002 lockdep_is_held(&task_rq(p)->lock)));
1006 trace_sched_migrate_task(p, new_cpu);
1008 if (task_cpu(p) != new_cpu) {
1009 if (p->sched_class->migrate_task_rq)
1010 p->sched_class->migrate_task_rq(p, new_cpu);
1011 p->se.nr_migrations++;
1012 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1015 __set_task_cpu(p, new_cpu);
1018 static void __migrate_swap_task(struct task_struct *p, int cpu)
1021 struct rq *src_rq, *dst_rq;
1023 src_rq = task_rq(p);
1024 dst_rq = cpu_rq(cpu);
1026 deactivate_task(src_rq, p, 0);
1027 set_task_cpu(p, cpu);
1028 activate_task(dst_rq, p, 0);
1029 check_preempt_curr(dst_rq, p, 0);
1032 * Task isn't running anymore; make it appear like we migrated
1033 * it before it went to sleep. This means on wakeup we make the
1034 * previous cpu our targer instead of where it really is.
1040 struct migration_swap_arg {
1041 struct task_struct *src_task, *dst_task;
1042 int src_cpu, dst_cpu;
1045 static int migrate_swap_stop(void *data)
1047 struct migration_swap_arg *arg = data;
1048 struct rq *src_rq, *dst_rq;
1051 src_rq = cpu_rq(arg->src_cpu);
1052 dst_rq = cpu_rq(arg->dst_cpu);
1054 double_raw_lock(&arg->src_task->pi_lock,
1055 &arg->dst_task->pi_lock);
1056 double_rq_lock(src_rq, dst_rq);
1057 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1060 if (task_cpu(arg->src_task) != arg->src_cpu)
1063 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1066 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1069 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1070 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1075 double_rq_unlock(src_rq, dst_rq);
1076 raw_spin_unlock(&arg->dst_task->pi_lock);
1077 raw_spin_unlock(&arg->src_task->pi_lock);
1083 * Cross migrate two tasks
1085 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1087 struct migration_swap_arg arg;
1090 arg = (struct migration_swap_arg){
1092 .src_cpu = task_cpu(cur),
1094 .dst_cpu = task_cpu(p),
1097 if (arg.src_cpu == arg.dst_cpu)
1101 * These three tests are all lockless; this is OK since all of them
1102 * will be re-checked with proper locks held further down the line.
1104 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1107 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1110 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1113 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1119 struct migration_arg {
1120 struct task_struct *task;
1124 static int migration_cpu_stop(void *data);
1127 * wait_task_inactive - wait for a thread to unschedule.
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1142 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1144 unsigned long flags;
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1190 * If it changed from the expected state, bail out now.
1192 if (unlikely(!ncsw))
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1199 * Oops. Go back and try again..
1201 if (unlikely(running)) {
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1247 void kick_process(struct task_struct *p)
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1257 EXPORT_SYMBOL_GPL(kick_process);
1258 #endif /* CONFIG_SMP */
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1264 static int select_fallback_rq(int cpu, struct task_struct *p)
1266 int nid = cpu_to_node(cpu);
1267 const struct cpumask *nodemask = NULL;
1268 enum { cpuset, possible, fail } state = cpuset;
1272 * If the node that the cpu is on has been offlined, cpu_to_node()
1273 * will return -1. There is no cpu on the node, and we should
1274 * select the cpu on the other node.
1277 nodemask = cpumask_of_node(nid);
1279 /* Look for allowed, online CPU in same node. */
1280 for_each_cpu(dest_cpu, nodemask) {
1281 if (!cpu_online(dest_cpu))
1283 if (!cpu_active(dest_cpu))
1285 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1291 /* Any allowed, online CPU? */
1292 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1293 if (!cpu_online(dest_cpu))
1295 if (!cpu_active(dest_cpu))
1302 /* No more Mr. Nice Guy. */
1303 cpuset_cpus_allowed_fallback(p);
1308 do_set_cpus_allowed(p, cpu_possible_mask);
1319 if (state != cpuset) {
1321 * Don't tell them about moving exiting tasks or
1322 * kernel threads (both mm NULL), since they never
1325 if (p->mm && printk_ratelimit()) {
1326 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1327 task_pid_nr(p), p->comm, cpu);
1335 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1338 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1340 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1343 * In order not to call set_task_cpu() on a blocking task we need
1344 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1347 * Since this is common to all placement strategies, this lives here.
1349 * [ this allows ->select_task() to simply return task_cpu(p) and
1350 * not worry about this generic constraint ]
1352 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1354 cpu = select_fallback_rq(task_cpu(p), p);
1359 static void update_avg(u64 *avg, u64 sample)
1361 s64 diff = sample - *avg;
1367 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1369 #ifdef CONFIG_SCHEDSTATS
1370 struct rq *rq = this_rq();
1373 int this_cpu = smp_processor_id();
1375 if (cpu == this_cpu) {
1376 schedstat_inc(rq, ttwu_local);
1377 schedstat_inc(p, se.statistics.nr_wakeups_local);
1379 struct sched_domain *sd;
1381 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1383 for_each_domain(this_cpu, sd) {
1384 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1385 schedstat_inc(sd, ttwu_wake_remote);
1392 if (wake_flags & WF_MIGRATED)
1393 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1395 #endif /* CONFIG_SMP */
1397 schedstat_inc(rq, ttwu_count);
1398 schedstat_inc(p, se.statistics.nr_wakeups);
1400 if (wake_flags & WF_SYNC)
1401 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1403 #endif /* CONFIG_SCHEDSTATS */
1406 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1408 activate_task(rq, p, en_flags);
1411 /* if a worker is waking up, notify workqueue */
1412 if (p->flags & PF_WQ_WORKER)
1413 wq_worker_waking_up(p, cpu_of(rq));
1417 * Mark the task runnable and perform wakeup-preemption.
1420 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1422 check_preempt_curr(rq, p, wake_flags);
1423 trace_sched_wakeup(p, true);
1425 p->state = TASK_RUNNING;
1427 if (p->sched_class->task_woken)
1428 p->sched_class->task_woken(rq, p);
1430 if (rq->idle_stamp) {
1431 u64 delta = rq_clock(rq) - rq->idle_stamp;
1432 u64 max = 2*rq->max_idle_balance_cost;
1434 update_avg(&rq->avg_idle, delta);
1436 if (rq->avg_idle > max)
1445 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1448 if (p->sched_contributes_to_load)
1449 rq->nr_uninterruptible--;
1452 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1453 ttwu_do_wakeup(rq, p, wake_flags);
1457 * Called in case the task @p isn't fully descheduled from its runqueue,
1458 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1459 * since all we need to do is flip p->state to TASK_RUNNING, since
1460 * the task is still ->on_rq.
1462 static int ttwu_remote(struct task_struct *p, int wake_flags)
1467 rq = __task_rq_lock(p);
1469 /* check_preempt_curr() may use rq clock */
1470 update_rq_clock(rq);
1471 ttwu_do_wakeup(rq, p, wake_flags);
1474 __task_rq_unlock(rq);
1480 static void sched_ttwu_pending(void)
1482 struct rq *rq = this_rq();
1483 struct llist_node *llist = llist_del_all(&rq->wake_list);
1484 struct task_struct *p;
1486 raw_spin_lock(&rq->lock);
1489 p = llist_entry(llist, struct task_struct, wake_entry);
1490 llist = llist_next(llist);
1491 ttwu_do_activate(rq, p, 0);
1494 raw_spin_unlock(&rq->lock);
1497 void scheduler_ipi(void)
1500 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1501 * TIF_NEED_RESCHED remotely (for the first time) will also send
1504 if (tif_need_resched())
1505 set_preempt_need_resched();
1507 if (llist_empty(&this_rq()->wake_list)
1508 && !tick_nohz_full_cpu(smp_processor_id())
1509 && !got_nohz_idle_kick())
1513 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1514 * traditionally all their work was done from the interrupt return
1515 * path. Now that we actually do some work, we need to make sure
1518 * Some archs already do call them, luckily irq_enter/exit nest
1521 * Arguably we should visit all archs and update all handlers,
1522 * however a fair share of IPIs are still resched only so this would
1523 * somewhat pessimize the simple resched case.
1526 tick_nohz_full_check();
1527 sched_ttwu_pending();
1530 * Check if someone kicked us for doing the nohz idle load balance.
1532 if (unlikely(got_nohz_idle_kick())) {
1533 this_rq()->idle_balance = 1;
1534 raise_softirq_irqoff(SCHED_SOFTIRQ);
1539 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1541 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1542 smp_send_reschedule(cpu);
1545 bool cpus_share_cache(int this_cpu, int that_cpu)
1547 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1549 #endif /* CONFIG_SMP */
1551 static void ttwu_queue(struct task_struct *p, int cpu)
1553 struct rq *rq = cpu_rq(cpu);
1555 #if defined(CONFIG_SMP)
1556 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1557 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1558 ttwu_queue_remote(p, cpu);
1563 raw_spin_lock(&rq->lock);
1564 ttwu_do_activate(rq, p, 0);
1565 raw_spin_unlock(&rq->lock);
1569 * try_to_wake_up - wake up a thread
1570 * @p: the thread to be awakened
1571 * @state: the mask of task states that can be woken
1572 * @wake_flags: wake modifier flags (WF_*)
1574 * Put it on the run-queue if it's not already there. The "current"
1575 * thread is always on the run-queue (except when the actual
1576 * re-schedule is in progress), and as such you're allowed to do
1577 * the simpler "current->state = TASK_RUNNING" to mark yourself
1578 * runnable without the overhead of this.
1580 * Return: %true if @p was woken up, %false if it was already running.
1581 * or @state didn't match @p's state.
1584 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1586 unsigned long flags;
1587 int cpu, success = 0;
1590 * If we are going to wake up a thread waiting for CONDITION we
1591 * need to ensure that CONDITION=1 done by the caller can not be
1592 * reordered with p->state check below. This pairs with mb() in
1593 * set_current_state() the waiting thread does.
1595 smp_mb__before_spinlock();
1596 raw_spin_lock_irqsave(&p->pi_lock, flags);
1597 if (!(p->state & state))
1600 success = 1; /* we're going to change ->state */
1603 if (p->on_rq && ttwu_remote(p, wake_flags))
1608 * If the owning (remote) cpu is still in the middle of schedule() with
1609 * this task as prev, wait until its done referencing the task.
1614 * Pairs with the smp_wmb() in finish_lock_switch().
1618 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1619 p->state = TASK_WAKING;
1621 if (p->sched_class->task_waking)
1622 p->sched_class->task_waking(p);
1624 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1625 if (task_cpu(p) != cpu) {
1626 wake_flags |= WF_MIGRATED;
1627 set_task_cpu(p, cpu);
1629 #endif /* CONFIG_SMP */
1633 ttwu_stat(p, cpu, wake_flags);
1635 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1641 * try_to_wake_up_local - try to wake up a local task with rq lock held
1642 * @p: the thread to be awakened
1644 * Put @p on the run-queue if it's not already there. The caller must
1645 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1648 static void try_to_wake_up_local(struct task_struct *p)
1650 struct rq *rq = task_rq(p);
1652 if (WARN_ON_ONCE(rq != this_rq()) ||
1653 WARN_ON_ONCE(p == current))
1656 lockdep_assert_held(&rq->lock);
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1664 if (!(p->state & TASK_NORMAL))
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1673 raw_spin_unlock(&p->pi_lock);
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1683 * Return: 1 if the process was woken up, 0 if it was already running.
1685 * It may be assumed that this function implies a write memory barrier before
1686 * changing the task state if and only if any tasks are woken up.
1688 int wake_up_process(struct task_struct *p)
1690 WARN_ON(task_is_stopped_or_traced(p));
1691 return try_to_wake_up(p, TASK_NORMAL, 0);
1693 EXPORT_SYMBOL(wake_up_process);
1695 int wake_up_state(struct task_struct *p, unsigned int state)
1697 return try_to_wake_up(p, state, 0);
1701 * Perform scheduler related setup for a newly forked process p.
1702 * p is forked by current.
1704 * __sched_fork() is basic setup used by init_idle() too:
1706 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1711 p->se.exec_start = 0;
1712 p->se.sum_exec_runtime = 0;
1713 p->se.prev_sum_exec_runtime = 0;
1714 p->se.nr_migrations = 0;
1716 INIT_LIST_HEAD(&p->se.group_node);
1718 #ifdef CONFIG_SCHEDSTATS
1719 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1722 RB_CLEAR_NODE(&p->dl.rb_node);
1723 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1724 p->dl.dl_runtime = p->dl.runtime = 0;
1725 p->dl.dl_deadline = p->dl.deadline = 0;
1726 p->dl.dl_period = 0;
1729 INIT_LIST_HEAD(&p->rt.run_list);
1731 #ifdef CONFIG_PREEMPT_NOTIFIERS
1732 INIT_HLIST_HEAD(&p->preempt_notifiers);
1735 #ifdef CONFIG_NUMA_BALANCING
1736 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1737 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1738 p->mm->numa_scan_seq = 0;
1741 if (clone_flags & CLONE_VM)
1742 p->numa_preferred_nid = current->numa_preferred_nid;
1744 p->numa_preferred_nid = -1;
1746 p->node_stamp = 0ULL;
1747 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1748 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1749 p->numa_work.next = &p->numa_work;
1750 p->numa_faults = NULL;
1751 p->numa_faults_buffer = NULL;
1753 INIT_LIST_HEAD(&p->numa_entry);
1754 p->numa_group = NULL;
1755 #endif /* CONFIG_NUMA_BALANCING */
1758 #ifdef CONFIG_NUMA_BALANCING
1759 #ifdef CONFIG_SCHED_DEBUG
1760 void set_numabalancing_state(bool enabled)
1763 sched_feat_set("NUMA");
1765 sched_feat_set("NO_NUMA");
1768 __read_mostly bool numabalancing_enabled;
1770 void set_numabalancing_state(bool enabled)
1772 numabalancing_enabled = enabled;
1774 #endif /* CONFIG_SCHED_DEBUG */
1775 #endif /* CONFIG_NUMA_BALANCING */
1778 * fork()/clone()-time setup:
1780 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1782 unsigned long flags;
1783 int cpu = get_cpu();
1785 __sched_fork(clone_flags, p);
1787 * We mark the process as running here. This guarantees that
1788 * nobody will actually run it, and a signal or other external
1789 * event cannot wake it up and insert it on the runqueue either.
1791 p->state = TASK_RUNNING;
1794 * Make sure we do not leak PI boosting priority to the child.
1796 p->prio = current->normal_prio;
1799 * Revert to default priority/policy on fork if requested.
1801 if (unlikely(p->sched_reset_on_fork)) {
1802 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1803 p->policy = SCHED_NORMAL;
1804 p->static_prio = NICE_TO_PRIO(0);
1806 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1807 p->static_prio = NICE_TO_PRIO(0);
1809 p->prio = p->normal_prio = __normal_prio(p);
1813 * We don't need the reset flag anymore after the fork. It has
1814 * fulfilled its duty:
1816 p->sched_reset_on_fork = 0;
1819 if (dl_prio(p->prio)) {
1822 } else if (rt_prio(p->prio)) {
1823 p->sched_class = &rt_sched_class;
1825 p->sched_class = &fair_sched_class;
1828 if (p->sched_class->task_fork)
1829 p->sched_class->task_fork(p);
1832 * The child is not yet in the pid-hash so no cgroup attach races,
1833 * and the cgroup is pinned to this child due to cgroup_fork()
1834 * is ran before sched_fork().
1836 * Silence PROVE_RCU.
1838 raw_spin_lock_irqsave(&p->pi_lock, flags);
1839 set_task_cpu(p, cpu);
1840 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1842 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1843 if (likely(sched_info_on()))
1844 memset(&p->sched_info, 0, sizeof(p->sched_info));
1846 #if defined(CONFIG_SMP)
1849 init_task_preempt_count(p);
1851 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1852 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1860 * wake_up_new_task - wake up a newly created task for the first time.
1862 * This function will do some initial scheduler statistics housekeeping
1863 * that must be done for every newly created context, then puts the task
1864 * on the runqueue and wakes it.
1866 void wake_up_new_task(struct task_struct *p)
1868 unsigned long flags;
1871 raw_spin_lock_irqsave(&p->pi_lock, flags);
1874 * Fork balancing, do it here and not earlier because:
1875 * - cpus_allowed can change in the fork path
1876 * - any previously selected cpu might disappear through hotplug
1878 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1881 /* Initialize new task's runnable average */
1882 init_task_runnable_average(p);
1883 rq = __task_rq_lock(p);
1884 activate_task(rq, p, 0);
1886 trace_sched_wakeup_new(p, true);
1887 check_preempt_curr(rq, p, WF_FORK);
1889 if (p->sched_class->task_woken)
1890 p->sched_class->task_woken(rq, p);
1892 task_rq_unlock(rq, p, &flags);
1895 #ifdef CONFIG_PREEMPT_NOTIFIERS
1898 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1899 * @notifier: notifier struct to register
1901 void preempt_notifier_register(struct preempt_notifier *notifier)
1903 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1905 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1908 * preempt_notifier_unregister - no longer interested in preemption notifications
1909 * @notifier: notifier struct to unregister
1911 * This is safe to call from within a preemption notifier.
1913 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1915 hlist_del(¬ifier->link);
1917 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1919 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1921 struct preempt_notifier *notifier;
1923 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1924 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1928 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1929 struct task_struct *next)
1931 struct preempt_notifier *notifier;
1933 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1934 notifier->ops->sched_out(notifier, next);
1937 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1939 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1944 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1945 struct task_struct *next)
1949 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1952 * prepare_task_switch - prepare to switch tasks
1953 * @rq: the runqueue preparing to switch
1954 * @prev: the current task that is being switched out
1955 * @next: the task we are going to switch to.
1957 * This is called with the rq lock held and interrupts off. It must
1958 * be paired with a subsequent finish_task_switch after the context
1961 * prepare_task_switch sets up locking and calls architecture specific
1965 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1966 struct task_struct *next)
1968 trace_sched_switch(prev, next);
1969 sched_info_switch(rq, prev, next);
1970 perf_event_task_sched_out(prev, next);
1971 fire_sched_out_preempt_notifiers(prev, next);
1972 prepare_lock_switch(rq, next);
1973 prepare_arch_switch(next);
1977 * finish_task_switch - clean up after a task-switch
1978 * @rq: runqueue associated with task-switch
1979 * @prev: the thread we just switched away from.
1981 * finish_task_switch must be called after the context switch, paired
1982 * with a prepare_task_switch call before the context switch.
1983 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1984 * and do any other architecture-specific cleanup actions.
1986 * Note that we may have delayed dropping an mm in context_switch(). If
1987 * so, we finish that here outside of the runqueue lock. (Doing it
1988 * with the lock held can cause deadlocks; see schedule() for
1991 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1992 __releases(rq->lock)
1994 struct mm_struct *mm = rq->prev_mm;
2000 * A task struct has one reference for the use as "current".
2001 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2002 * schedule one last time. The schedule call will never return, and
2003 * the scheduled task must drop that reference.
2004 * The test for TASK_DEAD must occur while the runqueue locks are
2005 * still held, otherwise prev could be scheduled on another cpu, die
2006 * there before we look at prev->state, and then the reference would
2008 * Manfred Spraul <manfred@colorfullife.com>
2010 prev_state = prev->state;
2011 vtime_task_switch(prev);
2012 finish_arch_switch(prev);
2013 perf_event_task_sched_in(prev, current);
2014 finish_lock_switch(rq, prev);
2015 finish_arch_post_lock_switch();
2017 fire_sched_in_preempt_notifiers(current);
2020 if (unlikely(prev_state == TASK_DEAD)) {
2021 task_numa_free(prev);
2023 if (prev->sched_class->task_dead)
2024 prev->sched_class->task_dead(prev);
2027 * Remove function-return probe instances associated with this
2028 * task and put them back on the free list.
2030 kprobe_flush_task(prev);
2031 put_task_struct(prev);
2034 tick_nohz_task_switch(current);
2039 /* assumes rq->lock is held */
2040 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2042 if (prev->sched_class->pre_schedule)
2043 prev->sched_class->pre_schedule(rq, prev);
2046 /* rq->lock is NOT held, but preemption is disabled */
2047 static inline void post_schedule(struct rq *rq)
2049 if (rq->post_schedule) {
2050 unsigned long flags;
2052 raw_spin_lock_irqsave(&rq->lock, flags);
2053 if (rq->curr->sched_class->post_schedule)
2054 rq->curr->sched_class->post_schedule(rq);
2055 raw_spin_unlock_irqrestore(&rq->lock, flags);
2057 rq->post_schedule = 0;
2063 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2067 static inline void post_schedule(struct rq *rq)
2074 * schedule_tail - first thing a freshly forked thread must call.
2075 * @prev: the thread we just switched away from.
2077 asmlinkage void schedule_tail(struct task_struct *prev)
2078 __releases(rq->lock)
2080 struct rq *rq = this_rq();
2082 finish_task_switch(rq, prev);
2085 * FIXME: do we need to worry about rq being invalidated by the
2090 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2091 /* In this case, finish_task_switch does not reenable preemption */
2094 if (current->set_child_tid)
2095 put_user(task_pid_vnr(current), current->set_child_tid);
2099 * context_switch - switch to the new MM and the new
2100 * thread's register state.
2103 context_switch(struct rq *rq, struct task_struct *prev,
2104 struct task_struct *next)
2106 struct mm_struct *mm, *oldmm;
2108 prepare_task_switch(rq, prev, next);
2111 oldmm = prev->active_mm;
2113 * For paravirt, this is coupled with an exit in switch_to to
2114 * combine the page table reload and the switch backend into
2117 arch_start_context_switch(prev);
2120 next->active_mm = oldmm;
2121 atomic_inc(&oldmm->mm_count);
2122 enter_lazy_tlb(oldmm, next);
2124 switch_mm(oldmm, mm, next);
2127 prev->active_mm = NULL;
2128 rq->prev_mm = oldmm;
2131 * Since the runqueue lock will be released by the next
2132 * task (which is an invalid locking op but in the case
2133 * of the scheduler it's an obvious special-case), so we
2134 * do an early lockdep release here:
2136 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2137 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2140 context_tracking_task_switch(prev, next);
2141 /* Here we just switch the register state and the stack. */
2142 switch_to(prev, next, prev);
2146 * this_rq must be evaluated again because prev may have moved
2147 * CPUs since it called schedule(), thus the 'rq' on its stack
2148 * frame will be invalid.
2150 finish_task_switch(this_rq(), prev);
2154 * nr_running and nr_context_switches:
2156 * externally visible scheduler statistics: current number of runnable
2157 * threads, total number of context switches performed since bootup.
2159 unsigned long nr_running(void)
2161 unsigned long i, sum = 0;
2163 for_each_online_cpu(i)
2164 sum += cpu_rq(i)->nr_running;
2169 unsigned long long nr_context_switches(void)
2172 unsigned long long sum = 0;
2174 for_each_possible_cpu(i)
2175 sum += cpu_rq(i)->nr_switches;
2180 unsigned long nr_iowait(void)
2182 unsigned long i, sum = 0;
2184 for_each_possible_cpu(i)
2185 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2190 unsigned long nr_iowait_cpu(int cpu)
2192 struct rq *this = cpu_rq(cpu);
2193 return atomic_read(&this->nr_iowait);
2199 * sched_exec - execve() is a valuable balancing opportunity, because at
2200 * this point the task has the smallest effective memory and cache footprint.
2202 void sched_exec(void)
2204 struct task_struct *p = current;
2205 unsigned long flags;
2208 raw_spin_lock_irqsave(&p->pi_lock, flags);
2209 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2210 if (dest_cpu == smp_processor_id())
2213 if (likely(cpu_active(dest_cpu))) {
2214 struct migration_arg arg = { p, dest_cpu };
2216 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2217 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2221 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2226 DEFINE_PER_CPU(struct kernel_stat, kstat);
2227 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2229 EXPORT_PER_CPU_SYMBOL(kstat);
2230 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2233 * Return any ns on the sched_clock that have not yet been accounted in
2234 * @p in case that task is currently running.
2236 * Called with task_rq_lock() held on @rq.
2238 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2242 if (task_current(rq, p)) {
2243 update_rq_clock(rq);
2244 ns = rq_clock_task(rq) - p->se.exec_start;
2252 unsigned long long task_delta_exec(struct task_struct *p)
2254 unsigned long flags;
2258 rq = task_rq_lock(p, &flags);
2259 ns = do_task_delta_exec(p, rq);
2260 task_rq_unlock(rq, p, &flags);
2266 * Return accounted runtime for the task.
2267 * In case the task is currently running, return the runtime plus current's
2268 * pending runtime that have not been accounted yet.
2270 unsigned long long task_sched_runtime(struct task_struct *p)
2272 unsigned long flags;
2276 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2278 * 64-bit doesn't need locks to atomically read a 64bit value.
2279 * So we have a optimization chance when the task's delta_exec is 0.
2280 * Reading ->on_cpu is racy, but this is ok.
2282 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2283 * If we race with it entering cpu, unaccounted time is 0. This is
2284 * indistinguishable from the read occurring a few cycles earlier.
2287 return p->se.sum_exec_runtime;
2290 rq = task_rq_lock(p, &flags);
2291 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2292 task_rq_unlock(rq, p, &flags);
2298 * This function gets called by the timer code, with HZ frequency.
2299 * We call it with interrupts disabled.
2301 void scheduler_tick(void)
2303 int cpu = smp_processor_id();
2304 struct rq *rq = cpu_rq(cpu);
2305 struct task_struct *curr = rq->curr;
2309 raw_spin_lock(&rq->lock);
2310 update_rq_clock(rq);
2311 curr->sched_class->task_tick(rq, curr, 0);
2312 update_cpu_load_active(rq);
2313 raw_spin_unlock(&rq->lock);
2315 perf_event_task_tick();
2318 rq->idle_balance = idle_cpu(cpu);
2319 trigger_load_balance(rq, cpu);
2321 rq_last_tick_reset(rq);
2324 #ifdef CONFIG_NO_HZ_FULL
2326 * scheduler_tick_max_deferment
2328 * Keep at least one tick per second when a single
2329 * active task is running because the scheduler doesn't
2330 * yet completely support full dynticks environment.
2332 * This makes sure that uptime, CFS vruntime, load
2333 * balancing, etc... continue to move forward, even
2334 * with a very low granularity.
2336 * Return: Maximum deferment in nanoseconds.
2338 u64 scheduler_tick_max_deferment(void)
2340 struct rq *rq = this_rq();
2341 unsigned long next, now = ACCESS_ONCE(jiffies);
2343 next = rq->last_sched_tick + HZ;
2345 if (time_before_eq(next, now))
2348 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2352 notrace unsigned long get_parent_ip(unsigned long addr)
2354 if (in_lock_functions(addr)) {
2355 addr = CALLER_ADDR2;
2356 if (in_lock_functions(addr))
2357 addr = CALLER_ADDR3;
2362 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2363 defined(CONFIG_PREEMPT_TRACER))
2365 void __kprobes preempt_count_add(int val)
2367 #ifdef CONFIG_DEBUG_PREEMPT
2371 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2374 __preempt_count_add(val);
2375 #ifdef CONFIG_DEBUG_PREEMPT
2377 * Spinlock count overflowing soon?
2379 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2382 if (preempt_count() == val)
2383 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2385 EXPORT_SYMBOL(preempt_count_add);
2387 void __kprobes preempt_count_sub(int val)
2389 #ifdef CONFIG_DEBUG_PREEMPT
2393 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2396 * Is the spinlock portion underflowing?
2398 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2399 !(preempt_count() & PREEMPT_MASK)))
2403 if (preempt_count() == val)
2404 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2405 __preempt_count_sub(val);
2407 EXPORT_SYMBOL(preempt_count_sub);
2412 * Print scheduling while atomic bug:
2414 static noinline void __schedule_bug(struct task_struct *prev)
2416 if (oops_in_progress)
2419 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2420 prev->comm, prev->pid, preempt_count());
2422 debug_show_held_locks(prev);
2424 if (irqs_disabled())
2425 print_irqtrace_events(prev);
2427 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2431 * Various schedule()-time debugging checks and statistics:
2433 static inline void schedule_debug(struct task_struct *prev)
2436 * Test if we are atomic. Since do_exit() needs to call into
2437 * schedule() atomically, we ignore that path. Otherwise whine
2438 * if we are scheduling when we should not.
2440 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2441 __schedule_bug(prev);
2444 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2446 schedstat_inc(this_rq(), sched_count);
2449 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2451 if (prev->on_rq || rq->skip_clock_update < 0)
2452 update_rq_clock(rq);
2453 prev->sched_class->put_prev_task(rq, prev);
2457 * Pick up the highest-prio task:
2459 static inline struct task_struct *
2460 pick_next_task(struct rq *rq)
2462 const struct sched_class *class;
2463 struct task_struct *p;
2466 * Optimization: we know that if all tasks are in
2467 * the fair class we can call that function directly:
2469 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2470 p = fair_sched_class.pick_next_task(rq);
2475 for_each_class(class) {
2476 p = class->pick_next_task(rq);
2481 BUG(); /* the idle class will always have a runnable task */
2485 * __schedule() is the main scheduler function.
2487 * The main means of driving the scheduler and thus entering this function are:
2489 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2491 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2492 * paths. For example, see arch/x86/entry_64.S.
2494 * To drive preemption between tasks, the scheduler sets the flag in timer
2495 * interrupt handler scheduler_tick().
2497 * 3. Wakeups don't really cause entry into schedule(). They add a
2498 * task to the run-queue and that's it.
2500 * Now, if the new task added to the run-queue preempts the current
2501 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2502 * called on the nearest possible occasion:
2504 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2506 * - in syscall or exception context, at the next outmost
2507 * preempt_enable(). (this might be as soon as the wake_up()'s
2510 * - in IRQ context, return from interrupt-handler to
2511 * preemptible context
2513 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2516 * - cond_resched() call
2517 * - explicit schedule() call
2518 * - return from syscall or exception to user-space
2519 * - return from interrupt-handler to user-space
2521 static void __sched __schedule(void)
2523 struct task_struct *prev, *next;
2524 unsigned long *switch_count;
2530 cpu = smp_processor_id();
2532 rcu_note_context_switch(cpu);
2535 schedule_debug(prev);
2537 if (sched_feat(HRTICK))
2541 * Make sure that signal_pending_state()->signal_pending() below
2542 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2543 * done by the caller to avoid the race with signal_wake_up().
2545 smp_mb__before_spinlock();
2546 raw_spin_lock_irq(&rq->lock);
2548 switch_count = &prev->nivcsw;
2549 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2550 if (unlikely(signal_pending_state(prev->state, prev))) {
2551 prev->state = TASK_RUNNING;
2553 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2557 * If a worker went to sleep, notify and ask workqueue
2558 * whether it wants to wake up a task to maintain
2561 if (prev->flags & PF_WQ_WORKER) {
2562 struct task_struct *to_wakeup;
2564 to_wakeup = wq_worker_sleeping(prev, cpu);
2566 try_to_wake_up_local(to_wakeup);
2569 switch_count = &prev->nvcsw;
2572 pre_schedule(rq, prev);
2574 if (unlikely(!rq->nr_running))
2575 idle_balance(cpu, rq);
2577 put_prev_task(rq, prev);
2578 next = pick_next_task(rq);
2579 clear_tsk_need_resched(prev);
2580 clear_preempt_need_resched();
2581 rq->skip_clock_update = 0;
2583 if (likely(prev != next)) {
2588 context_switch(rq, prev, next); /* unlocks the rq */
2590 * The context switch have flipped the stack from under us
2591 * and restored the local variables which were saved when
2592 * this task called schedule() in the past. prev == current
2593 * is still correct, but it can be moved to another cpu/rq.
2595 cpu = smp_processor_id();
2598 raw_spin_unlock_irq(&rq->lock);
2602 sched_preempt_enable_no_resched();
2607 static inline void sched_submit_work(struct task_struct *tsk)
2609 if (!tsk->state || tsk_is_pi_blocked(tsk))
2612 * If we are going to sleep and we have plugged IO queued,
2613 * make sure to submit it to avoid deadlocks.
2615 if (blk_needs_flush_plug(tsk))
2616 blk_schedule_flush_plug(tsk);
2619 asmlinkage void __sched schedule(void)
2621 struct task_struct *tsk = current;
2623 sched_submit_work(tsk);
2626 EXPORT_SYMBOL(schedule);
2628 #ifdef CONFIG_CONTEXT_TRACKING
2629 asmlinkage void __sched schedule_user(void)
2632 * If we come here after a random call to set_need_resched(),
2633 * or we have been woken up remotely but the IPI has not yet arrived,
2634 * we haven't yet exited the RCU idle mode. Do it here manually until
2635 * we find a better solution.
2644 * schedule_preempt_disabled - called with preemption disabled
2646 * Returns with preemption disabled. Note: preempt_count must be 1
2648 void __sched schedule_preempt_disabled(void)
2650 sched_preempt_enable_no_resched();
2655 #ifdef CONFIG_PREEMPT
2657 * this is the entry point to schedule() from in-kernel preemption
2658 * off of preempt_enable. Kernel preemptions off return from interrupt
2659 * occur there and call schedule directly.
2661 asmlinkage void __sched notrace preempt_schedule(void)
2664 * If there is a non-zero preempt_count or interrupts are disabled,
2665 * we do not want to preempt the current task. Just return..
2667 if (likely(!preemptible()))
2671 __preempt_count_add(PREEMPT_ACTIVE);
2673 __preempt_count_sub(PREEMPT_ACTIVE);
2676 * Check again in case we missed a preemption opportunity
2677 * between schedule and now.
2680 } while (need_resched());
2682 EXPORT_SYMBOL(preempt_schedule);
2683 #endif /* CONFIG_PREEMPT */
2686 * this is the entry point to schedule() from kernel preemption
2687 * off of irq context.
2688 * Note, that this is called and return with irqs disabled. This will
2689 * protect us against recursive calling from irq.
2691 asmlinkage void __sched preempt_schedule_irq(void)
2693 enum ctx_state prev_state;
2695 /* Catch callers which need to be fixed */
2696 BUG_ON(preempt_count() || !irqs_disabled());
2698 prev_state = exception_enter();
2701 __preempt_count_add(PREEMPT_ACTIVE);
2704 local_irq_disable();
2705 __preempt_count_sub(PREEMPT_ACTIVE);
2708 * Check again in case we missed a preemption opportunity
2709 * between schedule and now.
2712 } while (need_resched());
2714 exception_exit(prev_state);
2717 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2720 return try_to_wake_up(curr->private, mode, wake_flags);
2722 EXPORT_SYMBOL(default_wake_function);
2725 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2727 unsigned long flags;
2730 init_waitqueue_entry(&wait, current);
2732 __set_current_state(state);
2734 spin_lock_irqsave(&q->lock, flags);
2735 __add_wait_queue(q, &wait);
2736 spin_unlock(&q->lock);
2737 timeout = schedule_timeout(timeout);
2738 spin_lock_irq(&q->lock);
2739 __remove_wait_queue(q, &wait);
2740 spin_unlock_irqrestore(&q->lock, flags);
2745 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2747 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2749 EXPORT_SYMBOL(interruptible_sleep_on);
2752 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2754 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2756 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2758 void __sched sleep_on(wait_queue_head_t *q)
2760 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2762 EXPORT_SYMBOL(sleep_on);
2764 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2766 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2768 EXPORT_SYMBOL(sleep_on_timeout);
2770 #ifdef CONFIG_RT_MUTEXES
2773 * rt_mutex_setprio - set the current priority of a task
2775 * @prio: prio value (kernel-internal form)
2777 * This function changes the 'effective' priority of a task. It does
2778 * not touch ->normal_prio like __setscheduler().
2780 * Used by the rt_mutex code to implement priority inheritance logic.
2782 void rt_mutex_setprio(struct task_struct *p, int prio)
2784 int oldprio, on_rq, running;
2786 const struct sched_class *prev_class;
2788 BUG_ON(prio > MAX_PRIO);
2790 rq = __task_rq_lock(p);
2793 * Idle task boosting is a nono in general. There is one
2794 * exception, when PREEMPT_RT and NOHZ is active:
2796 * The idle task calls get_next_timer_interrupt() and holds
2797 * the timer wheel base->lock on the CPU and another CPU wants
2798 * to access the timer (probably to cancel it). We can safely
2799 * ignore the boosting request, as the idle CPU runs this code
2800 * with interrupts disabled and will complete the lock
2801 * protected section without being interrupted. So there is no
2802 * real need to boost.
2804 if (unlikely(p == rq->idle)) {
2805 WARN_ON(p != rq->curr);
2806 WARN_ON(p->pi_blocked_on);
2810 trace_sched_pi_setprio(p, prio);
2812 prev_class = p->sched_class;
2814 running = task_current(rq, p);
2816 dequeue_task(rq, p, 0);
2818 p->sched_class->put_prev_task(rq, p);
2821 p->sched_class = &dl_sched_class;
2822 else if (rt_prio(prio))
2823 p->sched_class = &rt_sched_class;
2825 p->sched_class = &fair_sched_class;
2830 p->sched_class->set_curr_task(rq);
2832 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
2834 check_class_changed(rq, p, prev_class, oldprio);
2836 __task_rq_unlock(rq);
2840 void set_user_nice(struct task_struct *p, long nice)
2842 int old_prio, delta, on_rq;
2843 unsigned long flags;
2846 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
2849 * We have to be careful, if called from sys_setpriority(),
2850 * the task might be in the middle of scheduling on another CPU.
2852 rq = task_rq_lock(p, &flags);
2854 * The RT priorities are set via sched_setscheduler(), but we still
2855 * allow the 'normal' nice value to be set - but as expected
2856 * it wont have any effect on scheduling until the task is
2857 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2859 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2860 p->static_prio = NICE_TO_PRIO(nice);
2865 dequeue_task(rq, p, 0);
2867 p->static_prio = NICE_TO_PRIO(nice);
2870 p->prio = effective_prio(p);
2871 delta = p->prio - old_prio;
2874 enqueue_task(rq, p, 0);
2876 * If the task increased its priority or is running and
2877 * lowered its priority, then reschedule its CPU:
2879 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2880 resched_task(rq->curr);
2883 task_rq_unlock(rq, p, &flags);
2885 EXPORT_SYMBOL(set_user_nice);
2888 * can_nice - check if a task can reduce its nice value
2892 int can_nice(const struct task_struct *p, const int nice)
2894 /* convert nice value [19,-20] to rlimit style value [1,40] */
2895 int nice_rlim = 20 - nice;
2897 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
2898 capable(CAP_SYS_NICE));
2901 #ifdef __ARCH_WANT_SYS_NICE
2904 * sys_nice - change the priority of the current process.
2905 * @increment: priority increment
2907 * sys_setpriority is a more generic, but much slower function that
2908 * does similar things.
2910 SYSCALL_DEFINE1(nice, int, increment)
2915 * Setpriority might change our priority at the same moment.
2916 * We don't have to worry. Conceptually one call occurs first
2917 * and we have a single winner.
2919 if (increment < -40)
2924 nice = TASK_NICE(current) + increment;
2930 if (increment < 0 && !can_nice(current, nice))
2933 retval = security_task_setnice(current, nice);
2937 set_user_nice(current, nice);
2944 * task_prio - return the priority value of a given task.
2945 * @p: the task in question.
2947 * Return: The priority value as seen by users in /proc.
2948 * RT tasks are offset by -200. Normal tasks are centered
2949 * around 0, value goes from -16 to +15.
2951 int task_prio(const struct task_struct *p)
2953 return p->prio - MAX_RT_PRIO;
2957 * task_nice - return the nice value of a given task.
2958 * @p: the task in question.
2960 * Return: The nice value [ -20 ... 0 ... 19 ].
2962 int task_nice(const struct task_struct *p)
2964 return TASK_NICE(p);
2966 EXPORT_SYMBOL(task_nice);
2969 * idle_cpu - is a given cpu idle currently?
2970 * @cpu: the processor in question.
2972 * Return: 1 if the CPU is currently idle. 0 otherwise.
2974 int idle_cpu(int cpu)
2976 struct rq *rq = cpu_rq(cpu);
2978 if (rq->curr != rq->idle)
2985 if (!llist_empty(&rq->wake_list))
2993 * idle_task - return the idle task for a given cpu.
2994 * @cpu: the processor in question.
2996 * Return: The idle task for the cpu @cpu.
2998 struct task_struct *idle_task(int cpu)
3000 return cpu_rq(cpu)->idle;
3004 * find_process_by_pid - find a process with a matching PID value.
3005 * @pid: the pid in question.
3007 * The task of @pid, if found. %NULL otherwise.
3009 static struct task_struct *find_process_by_pid(pid_t pid)
3011 return pid ? find_task_by_vpid(pid) : current;
3015 * This function initializes the sched_dl_entity of a newly becoming
3016 * SCHED_DEADLINE task.
3018 * Only the static values are considered here, the actual runtime and the
3019 * absolute deadline will be properly calculated when the task is enqueued
3020 * for the first time with its new policy.
3023 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3025 struct sched_dl_entity *dl_se = &p->dl;
3027 init_dl_task_timer(dl_se);
3028 dl_se->dl_runtime = attr->sched_runtime;
3029 dl_se->dl_deadline = attr->sched_deadline;
3030 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3031 dl_se->flags = attr->sched_flags;
3032 dl_se->dl_throttled = 0;
3036 /* Actually do priority change: must hold pi & rq lock. */
3037 static void __setscheduler(struct rq *rq, struct task_struct *p,
3038 const struct sched_attr *attr)
3040 int policy = attr->sched_policy;
3044 if (dl_policy(policy))
3045 __setparam_dl(p, attr);
3046 else if (rt_policy(policy))
3047 p->rt_priority = attr->sched_priority;
3049 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3051 p->normal_prio = normal_prio(p);
3052 p->prio = rt_mutex_getprio(p);
3054 if (dl_prio(p->prio))
3055 p->sched_class = &dl_sched_class;
3056 else if (rt_prio(p->prio))
3057 p->sched_class = &rt_sched_class;
3059 p->sched_class = &fair_sched_class;
3065 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3067 struct sched_dl_entity *dl_se = &p->dl;
3069 attr->sched_priority = p->rt_priority;
3070 attr->sched_runtime = dl_se->dl_runtime;
3071 attr->sched_deadline = dl_se->dl_deadline;
3072 attr->sched_period = dl_se->dl_period;
3073 attr->sched_flags = dl_se->flags;
3077 * This function validates the new parameters of a -deadline task.
3078 * We ask for the deadline not being zero, and greater or equal
3079 * than the runtime, as well as the period of being zero or
3080 * greater than deadline.
3083 __checkparam_dl(const struct sched_attr *attr)
3085 return attr && attr->sched_deadline != 0 &&
3086 (attr->sched_period == 0 ||
3087 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3088 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0;
3092 * check the target process has a UID that matches the current process's
3094 static bool check_same_owner(struct task_struct *p)
3096 const struct cred *cred = current_cred(), *pcred;
3100 pcred = __task_cred(p);
3101 match = (uid_eq(cred->euid, pcred->euid) ||
3102 uid_eq(cred->euid, pcred->uid));
3107 static int __sched_setscheduler(struct task_struct *p,
3108 const struct sched_attr *attr,
3111 int retval, oldprio, oldpolicy = -1, on_rq, running;
3112 int policy = attr->sched_policy;
3113 unsigned long flags;
3114 const struct sched_class *prev_class;
3118 /* may grab non-irq protected spin_locks */
3119 BUG_ON(in_interrupt());
3121 /* double check policy once rq lock held */
3123 reset_on_fork = p->sched_reset_on_fork;
3124 policy = oldpolicy = p->policy;
3126 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3127 policy &= ~SCHED_RESET_ON_FORK;
3129 if (policy != SCHED_DEADLINE &&
3130 policy != SCHED_FIFO && policy != SCHED_RR &&
3131 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3132 policy != SCHED_IDLE)
3137 * Valid priorities for SCHED_FIFO and SCHED_RR are
3138 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3139 * SCHED_BATCH and SCHED_IDLE is 0.
3141 if (attr->sched_priority < 0 ||
3142 (p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3143 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3145 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3146 (rt_policy(policy) != (attr->sched_priority != 0)))
3150 * Allow unprivileged RT tasks to decrease priority:
3152 if (user && !capable(CAP_SYS_NICE)) {
3153 if (fair_policy(policy)) {
3154 if (!can_nice(p, attr->sched_nice))
3158 if (rt_policy(policy)) {
3159 unsigned long rlim_rtprio =
3160 task_rlimit(p, RLIMIT_RTPRIO);
3162 /* can't set/change the rt policy */
3163 if (policy != p->policy && !rlim_rtprio)
3166 /* can't increase priority */
3167 if (attr->sched_priority > p->rt_priority &&
3168 attr->sched_priority > rlim_rtprio)
3173 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3174 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3176 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3177 if (!can_nice(p, TASK_NICE(p)))
3181 /* can't change other user's priorities */
3182 if (!check_same_owner(p))
3185 /* Normal users shall not reset the sched_reset_on_fork flag */
3186 if (p->sched_reset_on_fork && !reset_on_fork)
3191 retval = security_task_setscheduler(p);
3197 * make sure no PI-waiters arrive (or leave) while we are
3198 * changing the priority of the task:
3200 * To be able to change p->policy safely, the appropriate
3201 * runqueue lock must be held.
3203 rq = task_rq_lock(p, &flags);
3206 * Changing the policy of the stop threads its a very bad idea
3208 if (p == rq->stop) {
3209 task_rq_unlock(rq, p, &flags);
3214 * If not changing anything there's no need to proceed further:
3216 if (unlikely(policy == p->policy)) {
3217 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p))
3219 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3221 if (dl_policy(policy))
3224 task_rq_unlock(rq, p, &flags);
3229 #ifdef CONFIG_RT_GROUP_SCHED
3232 * Do not allow realtime tasks into groups that have no runtime
3235 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3236 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3237 !task_group_is_autogroup(task_group(p))) {
3238 task_rq_unlock(rq, p, &flags);
3244 /* recheck policy now with rq lock held */
3245 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3246 policy = oldpolicy = -1;
3247 task_rq_unlock(rq, p, &flags);
3251 running = task_current(rq, p);
3253 dequeue_task(rq, p, 0);
3255 p->sched_class->put_prev_task(rq, p);
3257 p->sched_reset_on_fork = reset_on_fork;
3260 prev_class = p->sched_class;
3261 __setscheduler(rq, p, attr);
3264 p->sched_class->set_curr_task(rq);
3266 enqueue_task(rq, p, 0);
3268 check_class_changed(rq, p, prev_class, oldprio);
3269 task_rq_unlock(rq, p, &flags);
3271 rt_mutex_adjust_pi(p);
3277 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3278 * @p: the task in question.
3279 * @policy: new policy.
3280 * @param: structure containing the new RT priority.
3282 * Return: 0 on success. An error code otherwise.
3284 * NOTE that the task may be already dead.
3286 int sched_setscheduler(struct task_struct *p, int policy,
3287 const struct sched_param *param)
3289 struct sched_attr attr = {
3290 .sched_policy = policy,
3291 .sched_priority = param->sched_priority
3293 return __sched_setscheduler(p, &attr, true);
3295 EXPORT_SYMBOL_GPL(sched_setscheduler);
3297 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3299 return __sched_setscheduler(p, attr, true);
3301 EXPORT_SYMBOL_GPL(sched_setattr);
3304 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3305 * @p: the task in question.
3306 * @policy: new policy.
3307 * @param: structure containing the new RT priority.
3309 * Just like sched_setscheduler, only don't bother checking if the
3310 * current context has permission. For example, this is needed in
3311 * stop_machine(): we create temporary high priority worker threads,
3312 * but our caller might not have that capability.
3314 * Return: 0 on success. An error code otherwise.
3316 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3317 const struct sched_param *param)
3319 struct sched_attr attr = {
3320 .sched_policy = policy,
3321 .sched_priority = param->sched_priority
3323 return __sched_setscheduler(p, &attr, false);
3327 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3329 struct sched_param lparam;
3330 struct task_struct *p;
3333 if (!param || pid < 0)
3335 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3340 p = find_process_by_pid(pid);
3342 retval = sched_setscheduler(p, policy, &lparam);
3349 * Mimics kernel/events/core.c perf_copy_attr().
3351 static int sched_copy_attr(struct sched_attr __user *uattr,
3352 struct sched_attr *attr)
3357 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3361 * zero the full structure, so that a short copy will be nice.
3363 memset(attr, 0, sizeof(*attr));
3365 ret = get_user(size, &uattr->size);
3369 if (size > PAGE_SIZE) /* silly large */
3372 if (!size) /* abi compat */
3373 size = SCHED_ATTR_SIZE_VER0;
3375 if (size < SCHED_ATTR_SIZE_VER0)
3379 * If we're handed a bigger struct than we know of,
3380 * ensure all the unknown bits are 0 - i.e. new
3381 * user-space does not rely on any kernel feature
3382 * extensions we dont know about yet.
3384 if (size > sizeof(*attr)) {
3385 unsigned char __user *addr;
3386 unsigned char __user *end;
3389 addr = (void __user *)uattr + sizeof(*attr);
3390 end = (void __user *)uattr + size;
3392 for (; addr < end; addr++) {
3393 ret = get_user(val, addr);
3399 size = sizeof(*attr);
3402 ret = copy_from_user(attr, uattr, size);
3407 * XXX: do we want to be lenient like existing syscalls; or do we want
3408 * to be strict and return an error on out-of-bounds values?
3410 attr->sched_nice = clamp(attr->sched_nice, -20, 19);
3416 put_user(sizeof(*attr), &uattr->size);
3422 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3423 * @pid: the pid in question.
3424 * @policy: new policy.
3425 * @param: structure containing the new RT priority.
3427 * Return: 0 on success. An error code otherwise.
3429 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3430 struct sched_param __user *, param)
3432 /* negative values for policy are not valid */
3436 return do_sched_setscheduler(pid, policy, param);
3440 * sys_sched_setparam - set/change the RT priority of a thread
3441 * @pid: the pid in question.
3442 * @param: structure containing the new RT priority.
3444 * Return: 0 on success. An error code otherwise.
3446 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3448 return do_sched_setscheduler(pid, -1, param);
3452 * sys_sched_setattr - same as above, but with extended sched_attr
3453 * @pid: the pid in question.
3454 * @attr: structure containing the extended parameters.
3456 SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr)
3458 struct sched_attr attr;
3459 struct task_struct *p;
3462 if (!uattr || pid < 0)
3465 if (sched_copy_attr(uattr, &attr))
3470 p = find_process_by_pid(pid);
3472 retval = sched_setattr(p, &attr);
3479 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3480 * @pid: the pid in question.
3482 * Return: On success, the policy of the thread. Otherwise, a negative error
3485 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3487 struct task_struct *p;
3495 p = find_process_by_pid(pid);
3497 retval = security_task_getscheduler(p);
3500 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3507 * sys_sched_getparam - get the RT priority of a thread
3508 * @pid: the pid in question.
3509 * @param: structure containing the RT priority.
3511 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3514 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3516 struct sched_param lp;
3517 struct task_struct *p;
3520 if (!param || pid < 0)
3524 p = find_process_by_pid(pid);
3529 retval = security_task_getscheduler(p);
3533 if (task_has_dl_policy(p)) {
3537 lp.sched_priority = p->rt_priority;
3541 * This one might sleep, we cannot do it with a spinlock held ...
3543 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3552 static int sched_read_attr(struct sched_attr __user *uattr,
3553 struct sched_attr *attr,
3558 if (!access_ok(VERIFY_WRITE, uattr, usize))
3562 * If we're handed a smaller struct than we know of,
3563 * ensure all the unknown bits are 0 - i.e. old
3564 * user-space does not get uncomplete information.
3566 if (usize < sizeof(*attr)) {
3567 unsigned char *addr;
3570 addr = (void *)attr + usize;
3571 end = (void *)attr + sizeof(*attr);
3573 for (; addr < end; addr++) {
3581 ret = copy_to_user(uattr, attr, usize);
3594 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3595 * @pid: the pid in question.
3596 * @attr: structure containing the extended parameters.
3597 * @size: sizeof(attr) for fwd/bwd comp.
3599 SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3602 struct sched_attr attr = {
3603 .size = sizeof(struct sched_attr),
3605 struct task_struct *p;
3608 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3609 size < SCHED_ATTR_SIZE_VER0)
3613 p = find_process_by_pid(pid);
3618 retval = security_task_getscheduler(p);
3622 attr.sched_policy = p->policy;
3623 if (task_has_dl_policy(p))
3624 __getparam_dl(p, &attr);
3625 else if (task_has_rt_policy(p))
3626 attr.sched_priority = p->rt_priority;
3628 attr.sched_nice = TASK_NICE(p);
3632 retval = sched_read_attr(uattr, &attr, size);
3640 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3642 cpumask_var_t cpus_allowed, new_mask;
3643 struct task_struct *p;
3648 p = find_process_by_pid(pid);
3654 /* Prevent p going away */
3658 if (p->flags & PF_NO_SETAFFINITY) {
3662 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3666 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3668 goto out_free_cpus_allowed;
3671 if (!check_same_owner(p)) {
3673 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3680 retval = security_task_setscheduler(p);
3684 cpuset_cpus_allowed(p, cpus_allowed);
3685 cpumask_and(new_mask, in_mask, cpus_allowed);
3687 retval = set_cpus_allowed_ptr(p, new_mask);
3690 cpuset_cpus_allowed(p, cpus_allowed);
3691 if (!cpumask_subset(new_mask, cpus_allowed)) {
3693 * We must have raced with a concurrent cpuset
3694 * update. Just reset the cpus_allowed to the
3695 * cpuset's cpus_allowed
3697 cpumask_copy(new_mask, cpus_allowed);
3702 free_cpumask_var(new_mask);
3703 out_free_cpus_allowed:
3704 free_cpumask_var(cpus_allowed);
3710 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3711 struct cpumask *new_mask)
3713 if (len < cpumask_size())
3714 cpumask_clear(new_mask);
3715 else if (len > cpumask_size())
3716 len = cpumask_size();
3718 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3722 * sys_sched_setaffinity - set the cpu affinity of a process
3723 * @pid: pid of the process
3724 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3725 * @user_mask_ptr: user-space pointer to the new cpu mask
3727 * Return: 0 on success. An error code otherwise.
3729 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3730 unsigned long __user *, user_mask_ptr)
3732 cpumask_var_t new_mask;
3735 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3738 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3740 retval = sched_setaffinity(pid, new_mask);
3741 free_cpumask_var(new_mask);
3745 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3747 struct task_struct *p;
3748 unsigned long flags;
3754 p = find_process_by_pid(pid);
3758 retval = security_task_getscheduler(p);
3762 raw_spin_lock_irqsave(&p->pi_lock, flags);
3763 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3764 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3773 * sys_sched_getaffinity - get the cpu affinity of a process
3774 * @pid: pid of the process
3775 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3776 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3778 * Return: 0 on success. An error code otherwise.
3780 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3781 unsigned long __user *, user_mask_ptr)
3786 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3788 if (len & (sizeof(unsigned long)-1))
3791 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3794 ret = sched_getaffinity(pid, mask);
3796 size_t retlen = min_t(size_t, len, cpumask_size());
3798 if (copy_to_user(user_mask_ptr, mask, retlen))
3803 free_cpumask_var(mask);
3809 * sys_sched_yield - yield the current processor to other threads.
3811 * This function yields the current CPU to other tasks. If there are no
3812 * other threads running on this CPU then this function will return.
3816 SYSCALL_DEFINE0(sched_yield)
3818 struct rq *rq = this_rq_lock();
3820 schedstat_inc(rq, yld_count);
3821 current->sched_class->yield_task(rq);
3824 * Since we are going to call schedule() anyway, there's
3825 * no need to preempt or enable interrupts:
3827 __release(rq->lock);
3828 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3829 do_raw_spin_unlock(&rq->lock);
3830 sched_preempt_enable_no_resched();
3837 static void __cond_resched(void)
3839 __preempt_count_add(PREEMPT_ACTIVE);
3841 __preempt_count_sub(PREEMPT_ACTIVE);
3844 int __sched _cond_resched(void)
3846 if (should_resched()) {
3852 EXPORT_SYMBOL(_cond_resched);
3855 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3856 * call schedule, and on return reacquire the lock.
3858 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3859 * operations here to prevent schedule() from being called twice (once via
3860 * spin_unlock(), once by hand).
3862 int __cond_resched_lock(spinlock_t *lock)
3864 int resched = should_resched();
3867 lockdep_assert_held(lock);
3869 if (spin_needbreak(lock) || resched) {
3880 EXPORT_SYMBOL(__cond_resched_lock);
3882 int __sched __cond_resched_softirq(void)
3884 BUG_ON(!in_softirq());
3886 if (should_resched()) {
3894 EXPORT_SYMBOL(__cond_resched_softirq);
3897 * yield - yield the current processor to other threads.
3899 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3901 * The scheduler is at all times free to pick the calling task as the most
3902 * eligible task to run, if removing the yield() call from your code breaks
3903 * it, its already broken.
3905 * Typical broken usage is:
3910 * where one assumes that yield() will let 'the other' process run that will
3911 * make event true. If the current task is a SCHED_FIFO task that will never
3912 * happen. Never use yield() as a progress guarantee!!
3914 * If you want to use yield() to wait for something, use wait_event().
3915 * If you want to use yield() to be 'nice' for others, use cond_resched().
3916 * If you still want to use yield(), do not!
3918 void __sched yield(void)
3920 set_current_state(TASK_RUNNING);
3923 EXPORT_SYMBOL(yield);
3926 * yield_to - yield the current processor to another thread in
3927 * your thread group, or accelerate that thread toward the
3928 * processor it's on.
3930 * @preempt: whether task preemption is allowed or not
3932 * It's the caller's job to ensure that the target task struct
3933 * can't go away on us before we can do any checks.
3936 * true (>0) if we indeed boosted the target task.
3937 * false (0) if we failed to boost the target.
3938 * -ESRCH if there's no task to yield to.
3940 bool __sched yield_to(struct task_struct *p, bool preempt)
3942 struct task_struct *curr = current;
3943 struct rq *rq, *p_rq;
3944 unsigned long flags;
3947 local_irq_save(flags);
3953 * If we're the only runnable task on the rq and target rq also
3954 * has only one task, there's absolutely no point in yielding.
3956 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3961 double_rq_lock(rq, p_rq);
3962 if (task_rq(p) != p_rq) {
3963 double_rq_unlock(rq, p_rq);
3967 if (!curr->sched_class->yield_to_task)
3970 if (curr->sched_class != p->sched_class)
3973 if (task_running(p_rq, p) || p->state)
3976 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3978 schedstat_inc(rq, yld_count);
3980 * Make p's CPU reschedule; pick_next_entity takes care of
3983 if (preempt && rq != p_rq)
3984 resched_task(p_rq->curr);
3988 double_rq_unlock(rq, p_rq);
3990 local_irq_restore(flags);
3997 EXPORT_SYMBOL_GPL(yield_to);
4000 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4001 * that process accounting knows that this is a task in IO wait state.
4003 void __sched io_schedule(void)
4005 struct rq *rq = raw_rq();
4007 delayacct_blkio_start();
4008 atomic_inc(&rq->nr_iowait);
4009 blk_flush_plug(current);
4010 current->in_iowait = 1;
4012 current->in_iowait = 0;
4013 atomic_dec(&rq->nr_iowait);
4014 delayacct_blkio_end();
4016 EXPORT_SYMBOL(io_schedule);
4018 long __sched io_schedule_timeout(long timeout)
4020 struct rq *rq = raw_rq();
4023 delayacct_blkio_start();
4024 atomic_inc(&rq->nr_iowait);
4025 blk_flush_plug(current);
4026 current->in_iowait = 1;
4027 ret = schedule_timeout(timeout);
4028 current->in_iowait = 0;
4029 atomic_dec(&rq->nr_iowait);
4030 delayacct_blkio_end();
4035 * sys_sched_get_priority_max - return maximum RT priority.
4036 * @policy: scheduling class.
4038 * Return: On success, this syscall returns the maximum
4039 * rt_priority that can be used by a given scheduling class.
4040 * On failure, a negative error code is returned.
4042 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4049 ret = MAX_USER_RT_PRIO-1;
4051 case SCHED_DEADLINE:
4062 * sys_sched_get_priority_min - return minimum RT priority.
4063 * @policy: scheduling class.
4065 * Return: On success, this syscall returns the minimum
4066 * rt_priority that can be used by a given scheduling class.
4067 * On failure, a negative error code is returned.
4069 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4078 case SCHED_DEADLINE:
4088 * sys_sched_rr_get_interval - return the default timeslice of a process.
4089 * @pid: pid of the process.
4090 * @interval: userspace pointer to the timeslice value.
4092 * this syscall writes the default timeslice value of a given process
4093 * into the user-space timespec buffer. A value of '0' means infinity.
4095 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4098 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4099 struct timespec __user *, interval)
4101 struct task_struct *p;
4102 unsigned int time_slice;
4103 unsigned long flags;
4113 p = find_process_by_pid(pid);
4117 retval = security_task_getscheduler(p);
4121 rq = task_rq_lock(p, &flags);
4122 time_slice = p->sched_class->get_rr_interval(rq, p);
4123 task_rq_unlock(rq, p, &flags);
4126 jiffies_to_timespec(time_slice, &t);
4127 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4135 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4137 void sched_show_task(struct task_struct *p)
4139 unsigned long free = 0;
4143 state = p->state ? __ffs(p->state) + 1 : 0;
4144 printk(KERN_INFO "%-15.15s %c", p->comm,
4145 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4146 #if BITS_PER_LONG == 32
4147 if (state == TASK_RUNNING)
4148 printk(KERN_CONT " running ");
4150 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4152 if (state == TASK_RUNNING)
4153 printk(KERN_CONT " running task ");
4155 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4157 #ifdef CONFIG_DEBUG_STACK_USAGE
4158 free = stack_not_used(p);
4161 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4163 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4164 task_pid_nr(p), ppid,
4165 (unsigned long)task_thread_info(p)->flags);
4167 print_worker_info(KERN_INFO, p);
4168 show_stack(p, NULL);
4171 void show_state_filter(unsigned long state_filter)
4173 struct task_struct *g, *p;
4175 #if BITS_PER_LONG == 32
4177 " task PC stack pid father\n");
4180 " task PC stack pid father\n");
4183 do_each_thread(g, p) {
4185 * reset the NMI-timeout, listing all files on a slow
4186 * console might take a lot of time:
4188 touch_nmi_watchdog();
4189 if (!state_filter || (p->state & state_filter))
4191 } while_each_thread(g, p);
4193 touch_all_softlockup_watchdogs();
4195 #ifdef CONFIG_SCHED_DEBUG
4196 sysrq_sched_debug_show();
4200 * Only show locks if all tasks are dumped:
4203 debug_show_all_locks();
4206 void init_idle_bootup_task(struct task_struct *idle)
4208 idle->sched_class = &idle_sched_class;
4212 * init_idle - set up an idle thread for a given CPU
4213 * @idle: task in question
4214 * @cpu: cpu the idle task belongs to
4216 * NOTE: this function does not set the idle thread's NEED_RESCHED
4217 * flag, to make booting more robust.
4219 void init_idle(struct task_struct *idle, int cpu)
4221 struct rq *rq = cpu_rq(cpu);
4222 unsigned long flags;
4224 raw_spin_lock_irqsave(&rq->lock, flags);
4226 __sched_fork(0, idle);
4227 idle->state = TASK_RUNNING;
4228 idle->se.exec_start = sched_clock();
4230 do_set_cpus_allowed(idle, cpumask_of(cpu));
4232 * We're having a chicken and egg problem, even though we are
4233 * holding rq->lock, the cpu isn't yet set to this cpu so the
4234 * lockdep check in task_group() will fail.
4236 * Similar case to sched_fork(). / Alternatively we could
4237 * use task_rq_lock() here and obtain the other rq->lock.
4242 __set_task_cpu(idle, cpu);
4245 rq->curr = rq->idle = idle;
4246 #if defined(CONFIG_SMP)
4249 raw_spin_unlock_irqrestore(&rq->lock, flags);
4251 /* Set the preempt count _outside_ the spinlocks! */
4252 init_idle_preempt_count(idle, cpu);
4255 * The idle tasks have their own, simple scheduling class:
4257 idle->sched_class = &idle_sched_class;
4258 ftrace_graph_init_idle_task(idle, cpu);
4259 vtime_init_idle(idle, cpu);
4260 #if defined(CONFIG_SMP)
4261 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4266 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4268 if (p->sched_class && p->sched_class->set_cpus_allowed)
4269 p->sched_class->set_cpus_allowed(p, new_mask);
4271 cpumask_copy(&p->cpus_allowed, new_mask);
4272 p->nr_cpus_allowed = cpumask_weight(new_mask);
4276 * This is how migration works:
4278 * 1) we invoke migration_cpu_stop() on the target CPU using
4280 * 2) stopper starts to run (implicitly forcing the migrated thread
4282 * 3) it checks whether the migrated task is still in the wrong runqueue.
4283 * 4) if it's in the wrong runqueue then the migration thread removes
4284 * it and puts it into the right queue.
4285 * 5) stopper completes and stop_one_cpu() returns and the migration
4290 * Change a given task's CPU affinity. Migrate the thread to a
4291 * proper CPU and schedule it away if the CPU it's executing on
4292 * is removed from the allowed bitmask.
4294 * NOTE: the caller must have a valid reference to the task, the
4295 * task must not exit() & deallocate itself prematurely. The
4296 * call is not atomic; no spinlocks may be held.
4298 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4300 unsigned long flags;
4302 unsigned int dest_cpu;
4305 rq = task_rq_lock(p, &flags);
4307 if (cpumask_equal(&p->cpus_allowed, new_mask))
4310 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4315 do_set_cpus_allowed(p, new_mask);
4317 /* Can the task run on the task's current CPU? If so, we're done */
4318 if (cpumask_test_cpu(task_cpu(p), new_mask))
4321 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4323 struct migration_arg arg = { p, dest_cpu };
4324 /* Need help from migration thread: drop lock and wait. */
4325 task_rq_unlock(rq, p, &flags);
4326 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4327 tlb_migrate_finish(p->mm);
4331 task_rq_unlock(rq, p, &flags);
4335 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4338 * Move (not current) task off this cpu, onto dest cpu. We're doing
4339 * this because either it can't run here any more (set_cpus_allowed()
4340 * away from this CPU, or CPU going down), or because we're
4341 * attempting to rebalance this task on exec (sched_exec).
4343 * So we race with normal scheduler movements, but that's OK, as long
4344 * as the task is no longer on this CPU.
4346 * Returns non-zero if task was successfully migrated.
4348 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4350 struct rq *rq_dest, *rq_src;
4353 if (unlikely(!cpu_active(dest_cpu)))
4356 rq_src = cpu_rq(src_cpu);
4357 rq_dest = cpu_rq(dest_cpu);
4359 raw_spin_lock(&p->pi_lock);
4360 double_rq_lock(rq_src, rq_dest);
4361 /* Already moved. */
4362 if (task_cpu(p) != src_cpu)
4364 /* Affinity changed (again). */
4365 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4369 * If we're not on a rq, the next wake-up will ensure we're
4373 dequeue_task(rq_src, p, 0);
4374 set_task_cpu(p, dest_cpu);
4375 enqueue_task(rq_dest, p, 0);
4376 check_preempt_curr(rq_dest, p, 0);
4381 double_rq_unlock(rq_src, rq_dest);
4382 raw_spin_unlock(&p->pi_lock);
4386 #ifdef CONFIG_NUMA_BALANCING
4387 /* Migrate current task p to target_cpu */
4388 int migrate_task_to(struct task_struct *p, int target_cpu)
4390 struct migration_arg arg = { p, target_cpu };
4391 int curr_cpu = task_cpu(p);
4393 if (curr_cpu == target_cpu)
4396 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4399 /* TODO: This is not properly updating schedstats */
4401 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4405 * Requeue a task on a given node and accurately track the number of NUMA
4406 * tasks on the runqueues
4408 void sched_setnuma(struct task_struct *p, int nid)
4411 unsigned long flags;
4412 bool on_rq, running;
4414 rq = task_rq_lock(p, &flags);
4416 running = task_current(rq, p);
4419 dequeue_task(rq, p, 0);
4421 p->sched_class->put_prev_task(rq, p);
4423 p->numa_preferred_nid = nid;
4426 p->sched_class->set_curr_task(rq);
4428 enqueue_task(rq, p, 0);
4429 task_rq_unlock(rq, p, &flags);
4434 * migration_cpu_stop - this will be executed by a highprio stopper thread
4435 * and performs thread migration by bumping thread off CPU then
4436 * 'pushing' onto another runqueue.
4438 static int migration_cpu_stop(void *data)
4440 struct migration_arg *arg = data;
4443 * The original target cpu might have gone down and we might
4444 * be on another cpu but it doesn't matter.
4446 local_irq_disable();
4447 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4452 #ifdef CONFIG_HOTPLUG_CPU
4455 * Ensures that the idle task is using init_mm right before its cpu goes
4458 void idle_task_exit(void)
4460 struct mm_struct *mm = current->active_mm;
4462 BUG_ON(cpu_online(smp_processor_id()));
4465 switch_mm(mm, &init_mm, current);
4470 * Since this CPU is going 'away' for a while, fold any nr_active delta
4471 * we might have. Assumes we're called after migrate_tasks() so that the
4472 * nr_active count is stable.
4474 * Also see the comment "Global load-average calculations".
4476 static void calc_load_migrate(struct rq *rq)
4478 long delta = calc_load_fold_active(rq);
4480 atomic_long_add(delta, &calc_load_tasks);
4484 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4485 * try_to_wake_up()->select_task_rq().
4487 * Called with rq->lock held even though we'er in stop_machine() and
4488 * there's no concurrency possible, we hold the required locks anyway
4489 * because of lock validation efforts.
4491 static void migrate_tasks(unsigned int dead_cpu)
4493 struct rq *rq = cpu_rq(dead_cpu);
4494 struct task_struct *next, *stop = rq->stop;
4498 * Fudge the rq selection such that the below task selection loop
4499 * doesn't get stuck on the currently eligible stop task.
4501 * We're currently inside stop_machine() and the rq is either stuck
4502 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4503 * either way we should never end up calling schedule() until we're
4509 * put_prev_task() and pick_next_task() sched
4510 * class method both need to have an up-to-date
4511 * value of rq->clock[_task]
4513 update_rq_clock(rq);
4517 * There's this thread running, bail when that's the only
4520 if (rq->nr_running == 1)
4523 next = pick_next_task(rq);
4525 next->sched_class->put_prev_task(rq, next);
4527 /* Find suitable destination for @next, with force if needed. */
4528 dest_cpu = select_fallback_rq(dead_cpu, next);
4529 raw_spin_unlock(&rq->lock);
4531 __migrate_task(next, dead_cpu, dest_cpu);
4533 raw_spin_lock(&rq->lock);
4539 #endif /* CONFIG_HOTPLUG_CPU */
4541 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4543 static struct ctl_table sd_ctl_dir[] = {
4545 .procname = "sched_domain",
4551 static struct ctl_table sd_ctl_root[] = {
4553 .procname = "kernel",
4555 .child = sd_ctl_dir,
4560 static struct ctl_table *sd_alloc_ctl_entry(int n)
4562 struct ctl_table *entry =
4563 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4568 static void sd_free_ctl_entry(struct ctl_table **tablep)
4570 struct ctl_table *entry;
4573 * In the intermediate directories, both the child directory and
4574 * procname are dynamically allocated and could fail but the mode
4575 * will always be set. In the lowest directory the names are
4576 * static strings and all have proc handlers.
4578 for (entry = *tablep; entry->mode; entry++) {
4580 sd_free_ctl_entry(&entry->child);
4581 if (entry->proc_handler == NULL)
4582 kfree(entry->procname);
4589 static int min_load_idx = 0;
4590 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4593 set_table_entry(struct ctl_table *entry,
4594 const char *procname, void *data, int maxlen,
4595 umode_t mode, proc_handler *proc_handler,
4598 entry->procname = procname;
4600 entry->maxlen = maxlen;
4602 entry->proc_handler = proc_handler;
4605 entry->extra1 = &min_load_idx;
4606 entry->extra2 = &max_load_idx;
4610 static struct ctl_table *
4611 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4613 struct ctl_table *table = sd_alloc_ctl_entry(13);
4618 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4619 sizeof(long), 0644, proc_doulongvec_minmax, false);
4620 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4621 sizeof(long), 0644, proc_doulongvec_minmax, false);
4622 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4623 sizeof(int), 0644, proc_dointvec_minmax, true);
4624 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4625 sizeof(int), 0644, proc_dointvec_minmax, true);
4626 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4627 sizeof(int), 0644, proc_dointvec_minmax, true);
4628 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4629 sizeof(int), 0644, proc_dointvec_minmax, true);
4630 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4631 sizeof(int), 0644, proc_dointvec_minmax, true);
4632 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4633 sizeof(int), 0644, proc_dointvec_minmax, false);
4634 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4635 sizeof(int), 0644, proc_dointvec_minmax, false);
4636 set_table_entry(&table[9], "cache_nice_tries",
4637 &sd->cache_nice_tries,
4638 sizeof(int), 0644, proc_dointvec_minmax, false);
4639 set_table_entry(&table[10], "flags", &sd->flags,
4640 sizeof(int), 0644, proc_dointvec_minmax, false);
4641 set_table_entry(&table[11], "name", sd->name,
4642 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4643 /* &table[12] is terminator */
4648 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4650 struct ctl_table *entry, *table;
4651 struct sched_domain *sd;
4652 int domain_num = 0, i;
4655 for_each_domain(cpu, sd)
4657 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4662 for_each_domain(cpu, sd) {
4663 snprintf(buf, 32, "domain%d", i);
4664 entry->procname = kstrdup(buf, GFP_KERNEL);
4666 entry->child = sd_alloc_ctl_domain_table(sd);
4673 static struct ctl_table_header *sd_sysctl_header;
4674 static void register_sched_domain_sysctl(void)
4676 int i, cpu_num = num_possible_cpus();
4677 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4680 WARN_ON(sd_ctl_dir[0].child);
4681 sd_ctl_dir[0].child = entry;
4686 for_each_possible_cpu(i) {
4687 snprintf(buf, 32, "cpu%d", i);
4688 entry->procname = kstrdup(buf, GFP_KERNEL);
4690 entry->child = sd_alloc_ctl_cpu_table(i);
4694 WARN_ON(sd_sysctl_header);
4695 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4698 /* may be called multiple times per register */
4699 static void unregister_sched_domain_sysctl(void)
4701 if (sd_sysctl_header)
4702 unregister_sysctl_table(sd_sysctl_header);
4703 sd_sysctl_header = NULL;
4704 if (sd_ctl_dir[0].child)
4705 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4708 static void register_sched_domain_sysctl(void)
4711 static void unregister_sched_domain_sysctl(void)
4716 static void set_rq_online(struct rq *rq)
4719 const struct sched_class *class;
4721 cpumask_set_cpu(rq->cpu, rq->rd->online);
4724 for_each_class(class) {
4725 if (class->rq_online)
4726 class->rq_online(rq);
4731 static void set_rq_offline(struct rq *rq)
4734 const struct sched_class *class;
4736 for_each_class(class) {
4737 if (class->rq_offline)
4738 class->rq_offline(rq);
4741 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4747 * migration_call - callback that gets triggered when a CPU is added.
4748 * Here we can start up the necessary migration thread for the new CPU.
4751 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4753 int cpu = (long)hcpu;
4754 unsigned long flags;
4755 struct rq *rq = cpu_rq(cpu);
4757 switch (action & ~CPU_TASKS_FROZEN) {
4759 case CPU_UP_PREPARE:
4760 rq->calc_load_update = calc_load_update;
4764 /* Update our root-domain */
4765 raw_spin_lock_irqsave(&rq->lock, flags);
4767 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4771 raw_spin_unlock_irqrestore(&rq->lock, flags);
4774 #ifdef CONFIG_HOTPLUG_CPU
4776 sched_ttwu_pending();
4777 /* Update our root-domain */
4778 raw_spin_lock_irqsave(&rq->lock, flags);
4780 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4784 BUG_ON(rq->nr_running != 1); /* the migration thread */
4785 raw_spin_unlock_irqrestore(&rq->lock, flags);
4789 calc_load_migrate(rq);
4794 update_max_interval();
4800 * Register at high priority so that task migration (migrate_all_tasks)
4801 * happens before everything else. This has to be lower priority than
4802 * the notifier in the perf_event subsystem, though.
4804 static struct notifier_block migration_notifier = {
4805 .notifier_call = migration_call,
4806 .priority = CPU_PRI_MIGRATION,
4809 static int sched_cpu_active(struct notifier_block *nfb,
4810 unsigned long action, void *hcpu)
4812 switch (action & ~CPU_TASKS_FROZEN) {
4814 case CPU_DOWN_FAILED:
4815 set_cpu_active((long)hcpu, true);
4822 static int sched_cpu_inactive(struct notifier_block *nfb,
4823 unsigned long action, void *hcpu)
4825 switch (action & ~CPU_TASKS_FROZEN) {
4826 case CPU_DOWN_PREPARE:
4827 set_cpu_active((long)hcpu, false);
4834 static int __init migration_init(void)
4836 void *cpu = (void *)(long)smp_processor_id();
4839 /* Initialize migration for the boot CPU */
4840 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4841 BUG_ON(err == NOTIFY_BAD);
4842 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4843 register_cpu_notifier(&migration_notifier);
4845 /* Register cpu active notifiers */
4846 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4847 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4851 early_initcall(migration_init);
4856 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4858 #ifdef CONFIG_SCHED_DEBUG
4860 static __read_mostly int sched_debug_enabled;
4862 static int __init sched_debug_setup(char *str)
4864 sched_debug_enabled = 1;
4868 early_param("sched_debug", sched_debug_setup);
4870 static inline bool sched_debug(void)
4872 return sched_debug_enabled;
4875 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4876 struct cpumask *groupmask)
4878 struct sched_group *group = sd->groups;
4881 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4882 cpumask_clear(groupmask);
4884 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4886 if (!(sd->flags & SD_LOAD_BALANCE)) {
4887 printk("does not load-balance\n");
4889 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4894 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4896 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4897 printk(KERN_ERR "ERROR: domain->span does not contain "
4900 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4901 printk(KERN_ERR "ERROR: domain->groups does not contain"
4905 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4909 printk(KERN_ERR "ERROR: group is NULL\n");
4914 * Even though we initialize ->power to something semi-sane,
4915 * we leave power_orig unset. This allows us to detect if
4916 * domain iteration is still funny without causing /0 traps.
4918 if (!group->sgp->power_orig) {
4919 printk(KERN_CONT "\n");
4920 printk(KERN_ERR "ERROR: domain->cpu_power not "
4925 if (!cpumask_weight(sched_group_cpus(group))) {
4926 printk(KERN_CONT "\n");
4927 printk(KERN_ERR "ERROR: empty group\n");
4931 if (!(sd->flags & SD_OVERLAP) &&
4932 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4933 printk(KERN_CONT "\n");
4934 printk(KERN_ERR "ERROR: repeated CPUs\n");
4938 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4940 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4942 printk(KERN_CONT " %s", str);
4943 if (group->sgp->power != SCHED_POWER_SCALE) {
4944 printk(KERN_CONT " (cpu_power = %d)",
4948 group = group->next;
4949 } while (group != sd->groups);
4950 printk(KERN_CONT "\n");
4952 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4953 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4956 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4957 printk(KERN_ERR "ERROR: parent span is not a superset "
4958 "of domain->span\n");
4962 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4966 if (!sched_debug_enabled)
4970 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4974 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4977 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4985 #else /* !CONFIG_SCHED_DEBUG */
4986 # define sched_domain_debug(sd, cpu) do { } while (0)
4987 static inline bool sched_debug(void)
4991 #endif /* CONFIG_SCHED_DEBUG */
4993 static int sd_degenerate(struct sched_domain *sd)
4995 if (cpumask_weight(sched_domain_span(sd)) == 1)
4998 /* Following flags need at least 2 groups */
4999 if (sd->flags & (SD_LOAD_BALANCE |
5000 SD_BALANCE_NEWIDLE |
5004 SD_SHARE_PKG_RESOURCES)) {
5005 if (sd->groups != sd->groups->next)
5009 /* Following flags don't use groups */
5010 if (sd->flags & (SD_WAKE_AFFINE))
5017 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5019 unsigned long cflags = sd->flags, pflags = parent->flags;
5021 if (sd_degenerate(parent))
5024 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5027 /* Flags needing groups don't count if only 1 group in parent */
5028 if (parent->groups == parent->groups->next) {
5029 pflags &= ~(SD_LOAD_BALANCE |
5030 SD_BALANCE_NEWIDLE |
5034 SD_SHARE_PKG_RESOURCES |
5036 if (nr_node_ids == 1)
5037 pflags &= ~SD_SERIALIZE;
5039 if (~cflags & pflags)
5045 static void free_rootdomain(struct rcu_head *rcu)
5047 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5049 cpupri_cleanup(&rd->cpupri);
5050 free_cpumask_var(rd->dlo_mask);
5051 free_cpumask_var(rd->rto_mask);
5052 free_cpumask_var(rd->online);
5053 free_cpumask_var(rd->span);
5057 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5059 struct root_domain *old_rd = NULL;
5060 unsigned long flags;
5062 raw_spin_lock_irqsave(&rq->lock, flags);
5067 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5070 cpumask_clear_cpu(rq->cpu, old_rd->span);
5073 * If we dont want to free the old_rd yet then
5074 * set old_rd to NULL to skip the freeing later
5077 if (!atomic_dec_and_test(&old_rd->refcount))
5081 atomic_inc(&rd->refcount);
5084 cpumask_set_cpu(rq->cpu, rd->span);
5085 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5088 raw_spin_unlock_irqrestore(&rq->lock, flags);
5091 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5094 static int init_rootdomain(struct root_domain *rd)
5096 memset(rd, 0, sizeof(*rd));
5098 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5100 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5102 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5104 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5107 if (cpupri_init(&rd->cpupri) != 0)
5112 free_cpumask_var(rd->rto_mask);
5114 free_cpumask_var(rd->dlo_mask);
5116 free_cpumask_var(rd->online);
5118 free_cpumask_var(rd->span);
5124 * By default the system creates a single root-domain with all cpus as
5125 * members (mimicking the global state we have today).
5127 struct root_domain def_root_domain;
5129 static void init_defrootdomain(void)
5131 init_rootdomain(&def_root_domain);
5133 atomic_set(&def_root_domain.refcount, 1);
5136 static struct root_domain *alloc_rootdomain(void)
5138 struct root_domain *rd;
5140 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5144 if (init_rootdomain(rd) != 0) {
5152 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5154 struct sched_group *tmp, *first;
5163 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5168 } while (sg != first);
5171 static void free_sched_domain(struct rcu_head *rcu)
5173 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5176 * If its an overlapping domain it has private groups, iterate and
5179 if (sd->flags & SD_OVERLAP) {
5180 free_sched_groups(sd->groups, 1);
5181 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5182 kfree(sd->groups->sgp);
5188 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5190 call_rcu(&sd->rcu, free_sched_domain);
5193 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5195 for (; sd; sd = sd->parent)
5196 destroy_sched_domain(sd, cpu);
5200 * Keep a special pointer to the highest sched_domain that has
5201 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5202 * allows us to avoid some pointer chasing select_idle_sibling().
5204 * Also keep a unique ID per domain (we use the first cpu number in
5205 * the cpumask of the domain), this allows us to quickly tell if
5206 * two cpus are in the same cache domain, see cpus_share_cache().
5208 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5209 DEFINE_PER_CPU(int, sd_llc_size);
5210 DEFINE_PER_CPU(int, sd_llc_id);
5211 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5212 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5213 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5215 static void update_top_cache_domain(int cpu)
5217 struct sched_domain *sd;
5218 struct sched_domain *busy_sd = NULL;
5222 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5224 id = cpumask_first(sched_domain_span(sd));
5225 size = cpumask_weight(sched_domain_span(sd));
5226 busy_sd = sd->parent; /* sd_busy */
5228 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5230 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5231 per_cpu(sd_llc_size, cpu) = size;
5232 per_cpu(sd_llc_id, cpu) = id;
5234 sd = lowest_flag_domain(cpu, SD_NUMA);
5235 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5237 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5238 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5242 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5243 * hold the hotplug lock.
5246 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5248 struct rq *rq = cpu_rq(cpu);
5249 struct sched_domain *tmp;
5251 /* Remove the sched domains which do not contribute to scheduling. */
5252 for (tmp = sd; tmp; ) {
5253 struct sched_domain *parent = tmp->parent;
5257 if (sd_parent_degenerate(tmp, parent)) {
5258 tmp->parent = parent->parent;
5260 parent->parent->child = tmp;
5262 * Transfer SD_PREFER_SIBLING down in case of a
5263 * degenerate parent; the spans match for this
5264 * so the property transfers.
5266 if (parent->flags & SD_PREFER_SIBLING)
5267 tmp->flags |= SD_PREFER_SIBLING;
5268 destroy_sched_domain(parent, cpu);
5273 if (sd && sd_degenerate(sd)) {
5276 destroy_sched_domain(tmp, cpu);
5281 sched_domain_debug(sd, cpu);
5283 rq_attach_root(rq, rd);
5285 rcu_assign_pointer(rq->sd, sd);
5286 destroy_sched_domains(tmp, cpu);
5288 update_top_cache_domain(cpu);
5291 /* cpus with isolated domains */
5292 static cpumask_var_t cpu_isolated_map;
5294 /* Setup the mask of cpus configured for isolated domains */
5295 static int __init isolated_cpu_setup(char *str)
5297 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5298 cpulist_parse(str, cpu_isolated_map);
5302 __setup("isolcpus=", isolated_cpu_setup);
5304 static const struct cpumask *cpu_cpu_mask(int cpu)
5306 return cpumask_of_node(cpu_to_node(cpu));
5310 struct sched_domain **__percpu sd;
5311 struct sched_group **__percpu sg;
5312 struct sched_group_power **__percpu sgp;
5316 struct sched_domain ** __percpu sd;
5317 struct root_domain *rd;
5327 struct sched_domain_topology_level;
5329 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5330 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5332 #define SDTL_OVERLAP 0x01
5334 struct sched_domain_topology_level {
5335 sched_domain_init_f init;
5336 sched_domain_mask_f mask;
5339 struct sd_data data;
5343 * Build an iteration mask that can exclude certain CPUs from the upwards
5346 * Asymmetric node setups can result in situations where the domain tree is of
5347 * unequal depth, make sure to skip domains that already cover the entire
5350 * In that case build_sched_domains() will have terminated the iteration early
5351 * and our sibling sd spans will be empty. Domains should always include the
5352 * cpu they're built on, so check that.
5355 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5357 const struct cpumask *span = sched_domain_span(sd);
5358 struct sd_data *sdd = sd->private;
5359 struct sched_domain *sibling;
5362 for_each_cpu(i, span) {
5363 sibling = *per_cpu_ptr(sdd->sd, i);
5364 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5367 cpumask_set_cpu(i, sched_group_mask(sg));
5372 * Return the canonical balance cpu for this group, this is the first cpu
5373 * of this group that's also in the iteration mask.
5375 int group_balance_cpu(struct sched_group *sg)
5377 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5381 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5383 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5384 const struct cpumask *span = sched_domain_span(sd);
5385 struct cpumask *covered = sched_domains_tmpmask;
5386 struct sd_data *sdd = sd->private;
5387 struct sched_domain *child;
5390 cpumask_clear(covered);
5392 for_each_cpu(i, span) {
5393 struct cpumask *sg_span;
5395 if (cpumask_test_cpu(i, covered))
5398 child = *per_cpu_ptr(sdd->sd, i);
5400 /* See the comment near build_group_mask(). */
5401 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5404 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5405 GFP_KERNEL, cpu_to_node(cpu));
5410 sg_span = sched_group_cpus(sg);
5412 child = child->child;
5413 cpumask_copy(sg_span, sched_domain_span(child));
5415 cpumask_set_cpu(i, sg_span);
5417 cpumask_or(covered, covered, sg_span);
5419 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5420 if (atomic_inc_return(&sg->sgp->ref) == 1)
5421 build_group_mask(sd, sg);
5424 * Initialize sgp->power such that even if we mess up the
5425 * domains and no possible iteration will get us here, we won't
5428 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5429 sg->sgp->power_orig = sg->sgp->power;
5432 * Make sure the first group of this domain contains the
5433 * canonical balance cpu. Otherwise the sched_domain iteration
5434 * breaks. See update_sg_lb_stats().
5436 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5437 group_balance_cpu(sg) == cpu)
5447 sd->groups = groups;
5452 free_sched_groups(first, 0);
5457 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5459 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5460 struct sched_domain *child = sd->child;
5463 cpu = cpumask_first(sched_domain_span(child));
5466 *sg = *per_cpu_ptr(sdd->sg, cpu);
5467 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5468 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5475 * build_sched_groups will build a circular linked list of the groups
5476 * covered by the given span, and will set each group's ->cpumask correctly,
5477 * and ->cpu_power to 0.
5479 * Assumes the sched_domain tree is fully constructed
5482 build_sched_groups(struct sched_domain *sd, int cpu)
5484 struct sched_group *first = NULL, *last = NULL;
5485 struct sd_data *sdd = sd->private;
5486 const struct cpumask *span = sched_domain_span(sd);
5487 struct cpumask *covered;
5490 get_group(cpu, sdd, &sd->groups);
5491 atomic_inc(&sd->groups->ref);
5493 if (cpu != cpumask_first(span))
5496 lockdep_assert_held(&sched_domains_mutex);
5497 covered = sched_domains_tmpmask;
5499 cpumask_clear(covered);
5501 for_each_cpu(i, span) {
5502 struct sched_group *sg;
5505 if (cpumask_test_cpu(i, covered))
5508 group = get_group(i, sdd, &sg);
5509 cpumask_clear(sched_group_cpus(sg));
5511 cpumask_setall(sched_group_mask(sg));
5513 for_each_cpu(j, span) {
5514 if (get_group(j, sdd, NULL) != group)
5517 cpumask_set_cpu(j, covered);
5518 cpumask_set_cpu(j, sched_group_cpus(sg));
5533 * Initialize sched groups cpu_power.
5535 * cpu_power indicates the capacity of sched group, which is used while
5536 * distributing the load between different sched groups in a sched domain.
5537 * Typically cpu_power for all the groups in a sched domain will be same unless
5538 * there are asymmetries in the topology. If there are asymmetries, group
5539 * having more cpu_power will pickup more load compared to the group having
5542 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5544 struct sched_group *sg = sd->groups;
5549 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5551 } while (sg != sd->groups);
5553 if (cpu != group_balance_cpu(sg))
5556 update_group_power(sd, cpu);
5557 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5560 int __weak arch_sd_sibling_asym_packing(void)
5562 return 0*SD_ASYM_PACKING;
5566 * Initializers for schedule domains
5567 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5570 #ifdef CONFIG_SCHED_DEBUG
5571 # define SD_INIT_NAME(sd, type) sd->name = #type
5573 # define SD_INIT_NAME(sd, type) do { } while (0)
5576 #define SD_INIT_FUNC(type) \
5577 static noinline struct sched_domain * \
5578 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5580 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5581 *sd = SD_##type##_INIT; \
5582 SD_INIT_NAME(sd, type); \
5583 sd->private = &tl->data; \
5588 #ifdef CONFIG_SCHED_SMT
5589 SD_INIT_FUNC(SIBLING)
5591 #ifdef CONFIG_SCHED_MC
5594 #ifdef CONFIG_SCHED_BOOK
5598 static int default_relax_domain_level = -1;
5599 int sched_domain_level_max;
5601 static int __init setup_relax_domain_level(char *str)
5603 if (kstrtoint(str, 0, &default_relax_domain_level))
5604 pr_warn("Unable to set relax_domain_level\n");
5608 __setup("relax_domain_level=", setup_relax_domain_level);
5610 static void set_domain_attribute(struct sched_domain *sd,
5611 struct sched_domain_attr *attr)
5615 if (!attr || attr->relax_domain_level < 0) {
5616 if (default_relax_domain_level < 0)
5619 request = default_relax_domain_level;
5621 request = attr->relax_domain_level;
5622 if (request < sd->level) {
5623 /* turn off idle balance on this domain */
5624 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5626 /* turn on idle balance on this domain */
5627 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5631 static void __sdt_free(const struct cpumask *cpu_map);
5632 static int __sdt_alloc(const struct cpumask *cpu_map);
5634 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5635 const struct cpumask *cpu_map)
5639 if (!atomic_read(&d->rd->refcount))
5640 free_rootdomain(&d->rd->rcu); /* fall through */
5642 free_percpu(d->sd); /* fall through */
5644 __sdt_free(cpu_map); /* fall through */
5650 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5651 const struct cpumask *cpu_map)
5653 memset(d, 0, sizeof(*d));
5655 if (__sdt_alloc(cpu_map))
5656 return sa_sd_storage;
5657 d->sd = alloc_percpu(struct sched_domain *);
5659 return sa_sd_storage;
5660 d->rd = alloc_rootdomain();
5663 return sa_rootdomain;
5667 * NULL the sd_data elements we've used to build the sched_domain and
5668 * sched_group structure so that the subsequent __free_domain_allocs()
5669 * will not free the data we're using.
5671 static void claim_allocations(int cpu, struct sched_domain *sd)
5673 struct sd_data *sdd = sd->private;
5675 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5676 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5678 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5679 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5681 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5682 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5685 #ifdef CONFIG_SCHED_SMT
5686 static const struct cpumask *cpu_smt_mask(int cpu)
5688 return topology_thread_cpumask(cpu);
5693 * Topology list, bottom-up.
5695 static struct sched_domain_topology_level default_topology[] = {
5696 #ifdef CONFIG_SCHED_SMT
5697 { sd_init_SIBLING, cpu_smt_mask, },
5699 #ifdef CONFIG_SCHED_MC
5700 { sd_init_MC, cpu_coregroup_mask, },
5702 #ifdef CONFIG_SCHED_BOOK
5703 { sd_init_BOOK, cpu_book_mask, },
5705 { sd_init_CPU, cpu_cpu_mask, },
5709 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5711 #define for_each_sd_topology(tl) \
5712 for (tl = sched_domain_topology; tl->init; tl++)
5716 static int sched_domains_numa_levels;
5717 static int *sched_domains_numa_distance;
5718 static struct cpumask ***sched_domains_numa_masks;
5719 static int sched_domains_curr_level;
5721 static inline int sd_local_flags(int level)
5723 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5726 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5729 static struct sched_domain *
5730 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5732 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5733 int level = tl->numa_level;
5734 int sd_weight = cpumask_weight(
5735 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5737 *sd = (struct sched_domain){
5738 .min_interval = sd_weight,
5739 .max_interval = 2*sd_weight,
5741 .imbalance_pct = 125,
5742 .cache_nice_tries = 2,
5749 .flags = 1*SD_LOAD_BALANCE
5750 | 1*SD_BALANCE_NEWIDLE
5755 | 0*SD_SHARE_CPUPOWER
5756 | 0*SD_SHARE_PKG_RESOURCES
5758 | 0*SD_PREFER_SIBLING
5760 | sd_local_flags(level)
5762 .last_balance = jiffies,
5763 .balance_interval = sd_weight,
5765 SD_INIT_NAME(sd, NUMA);
5766 sd->private = &tl->data;
5769 * Ugly hack to pass state to sd_numa_mask()...
5771 sched_domains_curr_level = tl->numa_level;
5776 static const struct cpumask *sd_numa_mask(int cpu)
5778 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5781 static void sched_numa_warn(const char *str)
5783 static int done = false;
5791 printk(KERN_WARNING "ERROR: %s\n\n", str);
5793 for (i = 0; i < nr_node_ids; i++) {
5794 printk(KERN_WARNING " ");
5795 for (j = 0; j < nr_node_ids; j++)
5796 printk(KERN_CONT "%02d ", node_distance(i,j));
5797 printk(KERN_CONT "\n");
5799 printk(KERN_WARNING "\n");
5802 static bool find_numa_distance(int distance)
5806 if (distance == node_distance(0, 0))
5809 for (i = 0; i < sched_domains_numa_levels; i++) {
5810 if (sched_domains_numa_distance[i] == distance)
5817 static void sched_init_numa(void)
5819 int next_distance, curr_distance = node_distance(0, 0);
5820 struct sched_domain_topology_level *tl;
5824 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5825 if (!sched_domains_numa_distance)
5829 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5830 * unique distances in the node_distance() table.
5832 * Assumes node_distance(0,j) includes all distances in
5833 * node_distance(i,j) in order to avoid cubic time.
5835 next_distance = curr_distance;
5836 for (i = 0; i < nr_node_ids; i++) {
5837 for (j = 0; j < nr_node_ids; j++) {
5838 for (k = 0; k < nr_node_ids; k++) {
5839 int distance = node_distance(i, k);
5841 if (distance > curr_distance &&
5842 (distance < next_distance ||
5843 next_distance == curr_distance))
5844 next_distance = distance;
5847 * While not a strong assumption it would be nice to know
5848 * about cases where if node A is connected to B, B is not
5849 * equally connected to A.
5851 if (sched_debug() && node_distance(k, i) != distance)
5852 sched_numa_warn("Node-distance not symmetric");
5854 if (sched_debug() && i && !find_numa_distance(distance))
5855 sched_numa_warn("Node-0 not representative");
5857 if (next_distance != curr_distance) {
5858 sched_domains_numa_distance[level++] = next_distance;
5859 sched_domains_numa_levels = level;
5860 curr_distance = next_distance;
5865 * In case of sched_debug() we verify the above assumption.
5871 * 'level' contains the number of unique distances, excluding the
5872 * identity distance node_distance(i,i).
5874 * The sched_domains_numa_distance[] array includes the actual distance
5879 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5880 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5881 * the array will contain less then 'level' members. This could be
5882 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5883 * in other functions.
5885 * We reset it to 'level' at the end of this function.
5887 sched_domains_numa_levels = 0;
5889 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5890 if (!sched_domains_numa_masks)
5894 * Now for each level, construct a mask per node which contains all
5895 * cpus of nodes that are that many hops away from us.
5897 for (i = 0; i < level; i++) {
5898 sched_domains_numa_masks[i] =
5899 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5900 if (!sched_domains_numa_masks[i])
5903 for (j = 0; j < nr_node_ids; j++) {
5904 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5908 sched_domains_numa_masks[i][j] = mask;
5910 for (k = 0; k < nr_node_ids; k++) {
5911 if (node_distance(j, k) > sched_domains_numa_distance[i])
5914 cpumask_or(mask, mask, cpumask_of_node(k));
5919 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5920 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5925 * Copy the default topology bits..
5927 for (i = 0; default_topology[i].init; i++)
5928 tl[i] = default_topology[i];
5931 * .. and append 'j' levels of NUMA goodness.
5933 for (j = 0; j < level; i++, j++) {
5934 tl[i] = (struct sched_domain_topology_level){
5935 .init = sd_numa_init,
5936 .mask = sd_numa_mask,
5937 .flags = SDTL_OVERLAP,
5942 sched_domain_topology = tl;
5944 sched_domains_numa_levels = level;
5947 static void sched_domains_numa_masks_set(int cpu)
5950 int node = cpu_to_node(cpu);
5952 for (i = 0; i < sched_domains_numa_levels; i++) {
5953 for (j = 0; j < nr_node_ids; j++) {
5954 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5955 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5960 static void sched_domains_numa_masks_clear(int cpu)
5963 for (i = 0; i < sched_domains_numa_levels; i++) {
5964 for (j = 0; j < nr_node_ids; j++)
5965 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5970 * Update sched_domains_numa_masks[level][node] array when new cpus
5973 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5974 unsigned long action,
5977 int cpu = (long)hcpu;
5979 switch (action & ~CPU_TASKS_FROZEN) {
5981 sched_domains_numa_masks_set(cpu);
5985 sched_domains_numa_masks_clear(cpu);
5995 static inline void sched_init_numa(void)
5999 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6000 unsigned long action,
6005 #endif /* CONFIG_NUMA */
6007 static int __sdt_alloc(const struct cpumask *cpu_map)
6009 struct sched_domain_topology_level *tl;
6012 for_each_sd_topology(tl) {
6013 struct sd_data *sdd = &tl->data;
6015 sdd->sd = alloc_percpu(struct sched_domain *);
6019 sdd->sg = alloc_percpu(struct sched_group *);
6023 sdd->sgp = alloc_percpu(struct sched_group_power *);
6027 for_each_cpu(j, cpu_map) {
6028 struct sched_domain *sd;
6029 struct sched_group *sg;
6030 struct sched_group_power *sgp;
6032 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6033 GFP_KERNEL, cpu_to_node(j));
6037 *per_cpu_ptr(sdd->sd, j) = sd;
6039 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6040 GFP_KERNEL, cpu_to_node(j));
6046 *per_cpu_ptr(sdd->sg, j) = sg;
6048 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6049 GFP_KERNEL, cpu_to_node(j));
6053 *per_cpu_ptr(sdd->sgp, j) = sgp;
6060 static void __sdt_free(const struct cpumask *cpu_map)
6062 struct sched_domain_topology_level *tl;
6065 for_each_sd_topology(tl) {
6066 struct sd_data *sdd = &tl->data;
6068 for_each_cpu(j, cpu_map) {
6069 struct sched_domain *sd;
6072 sd = *per_cpu_ptr(sdd->sd, j);
6073 if (sd && (sd->flags & SD_OVERLAP))
6074 free_sched_groups(sd->groups, 0);
6075 kfree(*per_cpu_ptr(sdd->sd, j));
6079 kfree(*per_cpu_ptr(sdd->sg, j));
6081 kfree(*per_cpu_ptr(sdd->sgp, j));
6083 free_percpu(sdd->sd);
6085 free_percpu(sdd->sg);
6087 free_percpu(sdd->sgp);
6092 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6093 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6094 struct sched_domain *child, int cpu)
6096 struct sched_domain *sd = tl->init(tl, cpu);
6100 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6102 sd->level = child->level + 1;
6103 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6107 set_domain_attribute(sd, attr);
6113 * Build sched domains for a given set of cpus and attach the sched domains
6114 * to the individual cpus
6116 static int build_sched_domains(const struct cpumask *cpu_map,
6117 struct sched_domain_attr *attr)
6119 enum s_alloc alloc_state;
6120 struct sched_domain *sd;
6122 int i, ret = -ENOMEM;
6124 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6125 if (alloc_state != sa_rootdomain)
6128 /* Set up domains for cpus specified by the cpu_map. */
6129 for_each_cpu(i, cpu_map) {
6130 struct sched_domain_topology_level *tl;
6133 for_each_sd_topology(tl) {
6134 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6135 if (tl == sched_domain_topology)
6136 *per_cpu_ptr(d.sd, i) = sd;
6137 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6138 sd->flags |= SD_OVERLAP;
6139 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6144 /* Build the groups for the domains */
6145 for_each_cpu(i, cpu_map) {
6146 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6147 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6148 if (sd->flags & SD_OVERLAP) {
6149 if (build_overlap_sched_groups(sd, i))
6152 if (build_sched_groups(sd, i))
6158 /* Calculate CPU power for physical packages and nodes */
6159 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6160 if (!cpumask_test_cpu(i, cpu_map))
6163 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6164 claim_allocations(i, sd);
6165 init_sched_groups_power(i, sd);
6169 /* Attach the domains */
6171 for_each_cpu(i, cpu_map) {
6172 sd = *per_cpu_ptr(d.sd, i);
6173 cpu_attach_domain(sd, d.rd, i);
6179 __free_domain_allocs(&d, alloc_state, cpu_map);
6183 static cpumask_var_t *doms_cur; /* current sched domains */
6184 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6185 static struct sched_domain_attr *dattr_cur;
6186 /* attribues of custom domains in 'doms_cur' */
6189 * Special case: If a kmalloc of a doms_cur partition (array of
6190 * cpumask) fails, then fallback to a single sched domain,
6191 * as determined by the single cpumask fallback_doms.
6193 static cpumask_var_t fallback_doms;
6196 * arch_update_cpu_topology lets virtualized architectures update the
6197 * cpu core maps. It is supposed to return 1 if the topology changed
6198 * or 0 if it stayed the same.
6200 int __attribute__((weak)) arch_update_cpu_topology(void)
6205 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6208 cpumask_var_t *doms;
6210 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6213 for (i = 0; i < ndoms; i++) {
6214 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6215 free_sched_domains(doms, i);
6222 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6225 for (i = 0; i < ndoms; i++)
6226 free_cpumask_var(doms[i]);
6231 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6232 * For now this just excludes isolated cpus, but could be used to
6233 * exclude other special cases in the future.
6235 static int init_sched_domains(const struct cpumask *cpu_map)
6239 arch_update_cpu_topology();
6241 doms_cur = alloc_sched_domains(ndoms_cur);
6243 doms_cur = &fallback_doms;
6244 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6245 err = build_sched_domains(doms_cur[0], NULL);
6246 register_sched_domain_sysctl();
6252 * Detach sched domains from a group of cpus specified in cpu_map
6253 * These cpus will now be attached to the NULL domain
6255 static void detach_destroy_domains(const struct cpumask *cpu_map)
6260 for_each_cpu(i, cpu_map)
6261 cpu_attach_domain(NULL, &def_root_domain, i);
6265 /* handle null as "default" */
6266 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6267 struct sched_domain_attr *new, int idx_new)
6269 struct sched_domain_attr tmp;
6276 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6277 new ? (new + idx_new) : &tmp,
6278 sizeof(struct sched_domain_attr));
6282 * Partition sched domains as specified by the 'ndoms_new'
6283 * cpumasks in the array doms_new[] of cpumasks. This compares
6284 * doms_new[] to the current sched domain partitioning, doms_cur[].
6285 * It destroys each deleted domain and builds each new domain.
6287 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6288 * The masks don't intersect (don't overlap.) We should setup one
6289 * sched domain for each mask. CPUs not in any of the cpumasks will
6290 * not be load balanced. If the same cpumask appears both in the
6291 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6294 * The passed in 'doms_new' should be allocated using
6295 * alloc_sched_domains. This routine takes ownership of it and will
6296 * free_sched_domains it when done with it. If the caller failed the
6297 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6298 * and partition_sched_domains() will fallback to the single partition
6299 * 'fallback_doms', it also forces the domains to be rebuilt.
6301 * If doms_new == NULL it will be replaced with cpu_online_mask.
6302 * ndoms_new == 0 is a special case for destroying existing domains,
6303 * and it will not create the default domain.
6305 * Call with hotplug lock held
6307 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6308 struct sched_domain_attr *dattr_new)
6313 mutex_lock(&sched_domains_mutex);
6315 /* always unregister in case we don't destroy any domains */
6316 unregister_sched_domain_sysctl();
6318 /* Let architecture update cpu core mappings. */
6319 new_topology = arch_update_cpu_topology();
6321 n = doms_new ? ndoms_new : 0;
6323 /* Destroy deleted domains */
6324 for (i = 0; i < ndoms_cur; i++) {
6325 for (j = 0; j < n && !new_topology; j++) {
6326 if (cpumask_equal(doms_cur[i], doms_new[j])
6327 && dattrs_equal(dattr_cur, i, dattr_new, j))
6330 /* no match - a current sched domain not in new doms_new[] */
6331 detach_destroy_domains(doms_cur[i]);
6337 if (doms_new == NULL) {
6339 doms_new = &fallback_doms;
6340 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6341 WARN_ON_ONCE(dattr_new);
6344 /* Build new domains */
6345 for (i = 0; i < ndoms_new; i++) {
6346 for (j = 0; j < n && !new_topology; j++) {
6347 if (cpumask_equal(doms_new[i], doms_cur[j])
6348 && dattrs_equal(dattr_new, i, dattr_cur, j))
6351 /* no match - add a new doms_new */
6352 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6357 /* Remember the new sched domains */
6358 if (doms_cur != &fallback_doms)
6359 free_sched_domains(doms_cur, ndoms_cur);
6360 kfree(dattr_cur); /* kfree(NULL) is safe */
6361 doms_cur = doms_new;
6362 dattr_cur = dattr_new;
6363 ndoms_cur = ndoms_new;
6365 register_sched_domain_sysctl();
6367 mutex_unlock(&sched_domains_mutex);
6370 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6373 * Update cpusets according to cpu_active mask. If cpusets are
6374 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6375 * around partition_sched_domains().
6377 * If we come here as part of a suspend/resume, don't touch cpusets because we
6378 * want to restore it back to its original state upon resume anyway.
6380 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6384 case CPU_ONLINE_FROZEN:
6385 case CPU_DOWN_FAILED_FROZEN:
6388 * num_cpus_frozen tracks how many CPUs are involved in suspend
6389 * resume sequence. As long as this is not the last online
6390 * operation in the resume sequence, just build a single sched
6391 * domain, ignoring cpusets.
6394 if (likely(num_cpus_frozen)) {
6395 partition_sched_domains(1, NULL, NULL);
6400 * This is the last CPU online operation. So fall through and
6401 * restore the original sched domains by considering the
6402 * cpuset configurations.
6406 case CPU_DOWN_FAILED:
6407 cpuset_update_active_cpus(true);
6415 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6419 case CPU_DOWN_PREPARE:
6420 cpuset_update_active_cpus(false);
6422 case CPU_DOWN_PREPARE_FROZEN:
6424 partition_sched_domains(1, NULL, NULL);
6432 void __init sched_init_smp(void)
6434 cpumask_var_t non_isolated_cpus;
6436 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6437 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6442 * There's no userspace yet to cause hotplug operations; hence all the
6443 * cpu masks are stable and all blatant races in the below code cannot
6446 mutex_lock(&sched_domains_mutex);
6447 init_sched_domains(cpu_active_mask);
6448 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6449 if (cpumask_empty(non_isolated_cpus))
6450 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6451 mutex_unlock(&sched_domains_mutex);
6453 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6454 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6455 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6459 /* Move init over to a non-isolated CPU */
6460 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6462 sched_init_granularity();
6463 free_cpumask_var(non_isolated_cpus);
6465 init_sched_rt_class();
6466 init_sched_dl_class();
6469 void __init sched_init_smp(void)
6471 sched_init_granularity();
6473 #endif /* CONFIG_SMP */
6475 const_debug unsigned int sysctl_timer_migration = 1;
6477 int in_sched_functions(unsigned long addr)
6479 return in_lock_functions(addr) ||
6480 (addr >= (unsigned long)__sched_text_start
6481 && addr < (unsigned long)__sched_text_end);
6484 #ifdef CONFIG_CGROUP_SCHED
6486 * Default task group.
6487 * Every task in system belongs to this group at bootup.
6489 struct task_group root_task_group;
6490 LIST_HEAD(task_groups);
6493 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6495 void __init sched_init(void)
6498 unsigned long alloc_size = 0, ptr;
6500 #ifdef CONFIG_FAIR_GROUP_SCHED
6501 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6503 #ifdef CONFIG_RT_GROUP_SCHED
6504 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6506 #ifdef CONFIG_CPUMASK_OFFSTACK
6507 alloc_size += num_possible_cpus() * cpumask_size();
6510 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6512 #ifdef CONFIG_FAIR_GROUP_SCHED
6513 root_task_group.se = (struct sched_entity **)ptr;
6514 ptr += nr_cpu_ids * sizeof(void **);
6516 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6517 ptr += nr_cpu_ids * sizeof(void **);
6519 #endif /* CONFIG_FAIR_GROUP_SCHED */
6520 #ifdef CONFIG_RT_GROUP_SCHED
6521 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6522 ptr += nr_cpu_ids * sizeof(void **);
6524 root_task_group.rt_rq = (struct rt_rq **)ptr;
6525 ptr += nr_cpu_ids * sizeof(void **);
6527 #endif /* CONFIG_RT_GROUP_SCHED */
6528 #ifdef CONFIG_CPUMASK_OFFSTACK
6529 for_each_possible_cpu(i) {
6530 per_cpu(load_balance_mask, i) = (void *)ptr;
6531 ptr += cpumask_size();
6533 #endif /* CONFIG_CPUMASK_OFFSTACK */
6537 init_defrootdomain();
6540 init_rt_bandwidth(&def_rt_bandwidth,
6541 global_rt_period(), global_rt_runtime());
6543 #ifdef CONFIG_RT_GROUP_SCHED
6544 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6545 global_rt_period(), global_rt_runtime());
6546 #endif /* CONFIG_RT_GROUP_SCHED */
6548 #ifdef CONFIG_CGROUP_SCHED
6549 list_add(&root_task_group.list, &task_groups);
6550 INIT_LIST_HEAD(&root_task_group.children);
6551 INIT_LIST_HEAD(&root_task_group.siblings);
6552 autogroup_init(&init_task);
6554 #endif /* CONFIG_CGROUP_SCHED */
6556 for_each_possible_cpu(i) {
6560 raw_spin_lock_init(&rq->lock);
6562 rq->calc_load_active = 0;
6563 rq->calc_load_update = jiffies + LOAD_FREQ;
6564 init_cfs_rq(&rq->cfs);
6565 init_rt_rq(&rq->rt, rq);
6566 init_dl_rq(&rq->dl, rq);
6567 #ifdef CONFIG_FAIR_GROUP_SCHED
6568 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6569 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6571 * How much cpu bandwidth does root_task_group get?
6573 * In case of task-groups formed thr' the cgroup filesystem, it
6574 * gets 100% of the cpu resources in the system. This overall
6575 * system cpu resource is divided among the tasks of
6576 * root_task_group and its child task-groups in a fair manner,
6577 * based on each entity's (task or task-group's) weight
6578 * (se->load.weight).
6580 * In other words, if root_task_group has 10 tasks of weight
6581 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6582 * then A0's share of the cpu resource is:
6584 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6586 * We achieve this by letting root_task_group's tasks sit
6587 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6589 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6590 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6591 #endif /* CONFIG_FAIR_GROUP_SCHED */
6593 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6594 #ifdef CONFIG_RT_GROUP_SCHED
6595 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6596 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6599 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6600 rq->cpu_load[j] = 0;
6602 rq->last_load_update_tick = jiffies;
6607 rq->cpu_power = SCHED_POWER_SCALE;
6608 rq->post_schedule = 0;
6609 rq->active_balance = 0;
6610 rq->next_balance = jiffies;
6615 rq->avg_idle = 2*sysctl_sched_migration_cost;
6616 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6618 INIT_LIST_HEAD(&rq->cfs_tasks);
6620 rq_attach_root(rq, &def_root_domain);
6621 #ifdef CONFIG_NO_HZ_COMMON
6624 #ifdef CONFIG_NO_HZ_FULL
6625 rq->last_sched_tick = 0;
6629 atomic_set(&rq->nr_iowait, 0);
6632 set_load_weight(&init_task);
6634 #ifdef CONFIG_PREEMPT_NOTIFIERS
6635 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6639 * The boot idle thread does lazy MMU switching as well:
6641 atomic_inc(&init_mm.mm_count);
6642 enter_lazy_tlb(&init_mm, current);
6645 * Make us the idle thread. Technically, schedule() should not be
6646 * called from this thread, however somewhere below it might be,
6647 * but because we are the idle thread, we just pick up running again
6648 * when this runqueue becomes "idle".
6650 init_idle(current, smp_processor_id());
6652 calc_load_update = jiffies + LOAD_FREQ;
6655 * During early bootup we pretend to be a normal task:
6657 current->sched_class = &fair_sched_class;
6660 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6661 /* May be allocated at isolcpus cmdline parse time */
6662 if (cpu_isolated_map == NULL)
6663 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6664 idle_thread_set_boot_cpu();
6666 init_sched_fair_class();
6668 scheduler_running = 1;
6671 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6672 static inline int preempt_count_equals(int preempt_offset)
6674 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6676 return (nested == preempt_offset);
6679 void __might_sleep(const char *file, int line, int preempt_offset)
6681 static unsigned long prev_jiffy; /* ratelimiting */
6683 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6684 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6685 system_state != SYSTEM_RUNNING || oops_in_progress)
6687 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6689 prev_jiffy = jiffies;
6692 "BUG: sleeping function called from invalid context at %s:%d\n",
6695 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6696 in_atomic(), irqs_disabled(),
6697 current->pid, current->comm);
6699 debug_show_held_locks(current);
6700 if (irqs_disabled())
6701 print_irqtrace_events(current);
6704 EXPORT_SYMBOL(__might_sleep);
6707 #ifdef CONFIG_MAGIC_SYSRQ
6708 static void normalize_task(struct rq *rq, struct task_struct *p)
6710 const struct sched_class *prev_class = p->sched_class;
6711 struct sched_attr attr = {
6712 .sched_policy = SCHED_NORMAL,
6714 int old_prio = p->prio;
6719 dequeue_task(rq, p, 0);
6720 __setscheduler(rq, p, &attr);
6722 enqueue_task(rq, p, 0);
6723 resched_task(rq->curr);
6726 check_class_changed(rq, p, prev_class, old_prio);
6729 void normalize_rt_tasks(void)
6731 struct task_struct *g, *p;
6732 unsigned long flags;
6735 read_lock_irqsave(&tasklist_lock, flags);
6736 do_each_thread(g, p) {
6738 * Only normalize user tasks:
6743 p->se.exec_start = 0;
6744 #ifdef CONFIG_SCHEDSTATS
6745 p->se.statistics.wait_start = 0;
6746 p->se.statistics.sleep_start = 0;
6747 p->se.statistics.block_start = 0;
6750 if (!dl_task(p) && !rt_task(p)) {
6752 * Renice negative nice level userspace
6755 if (TASK_NICE(p) < 0 && p->mm)
6756 set_user_nice(p, 0);
6760 raw_spin_lock(&p->pi_lock);
6761 rq = __task_rq_lock(p);
6763 normalize_task(rq, p);
6765 __task_rq_unlock(rq);
6766 raw_spin_unlock(&p->pi_lock);
6767 } while_each_thread(g, p);
6769 read_unlock_irqrestore(&tasklist_lock, flags);
6772 #endif /* CONFIG_MAGIC_SYSRQ */
6774 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6776 * These functions are only useful for the IA64 MCA handling, or kdb.
6778 * They can only be called when the whole system has been
6779 * stopped - every CPU needs to be quiescent, and no scheduling
6780 * activity can take place. Using them for anything else would
6781 * be a serious bug, and as a result, they aren't even visible
6782 * under any other configuration.
6786 * curr_task - return the current task for a given cpu.
6787 * @cpu: the processor in question.
6789 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6791 * Return: The current task for @cpu.
6793 struct task_struct *curr_task(int cpu)
6795 return cpu_curr(cpu);
6798 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6802 * set_curr_task - set the current task for a given cpu.
6803 * @cpu: the processor in question.
6804 * @p: the task pointer to set.
6806 * Description: This function must only be used when non-maskable interrupts
6807 * are serviced on a separate stack. It allows the architecture to switch the
6808 * notion of the current task on a cpu in a non-blocking manner. This function
6809 * must be called with all CPU's synchronized, and interrupts disabled, the
6810 * and caller must save the original value of the current task (see
6811 * curr_task() above) and restore that value before reenabling interrupts and
6812 * re-starting the system.
6814 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6816 void set_curr_task(int cpu, struct task_struct *p)
6823 #ifdef CONFIG_CGROUP_SCHED
6824 /* task_group_lock serializes the addition/removal of task groups */
6825 static DEFINE_SPINLOCK(task_group_lock);
6827 static void free_sched_group(struct task_group *tg)
6829 free_fair_sched_group(tg);
6830 free_rt_sched_group(tg);
6835 /* allocate runqueue etc for a new task group */
6836 struct task_group *sched_create_group(struct task_group *parent)
6838 struct task_group *tg;
6840 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6842 return ERR_PTR(-ENOMEM);
6844 if (!alloc_fair_sched_group(tg, parent))
6847 if (!alloc_rt_sched_group(tg, parent))
6853 free_sched_group(tg);
6854 return ERR_PTR(-ENOMEM);
6857 void sched_online_group(struct task_group *tg, struct task_group *parent)
6859 unsigned long flags;
6861 spin_lock_irqsave(&task_group_lock, flags);
6862 list_add_rcu(&tg->list, &task_groups);
6864 WARN_ON(!parent); /* root should already exist */
6866 tg->parent = parent;
6867 INIT_LIST_HEAD(&tg->children);
6868 list_add_rcu(&tg->siblings, &parent->children);
6869 spin_unlock_irqrestore(&task_group_lock, flags);
6872 /* rcu callback to free various structures associated with a task group */
6873 static void free_sched_group_rcu(struct rcu_head *rhp)
6875 /* now it should be safe to free those cfs_rqs */
6876 free_sched_group(container_of(rhp, struct task_group, rcu));
6879 /* Destroy runqueue etc associated with a task group */
6880 void sched_destroy_group(struct task_group *tg)
6882 /* wait for possible concurrent references to cfs_rqs complete */
6883 call_rcu(&tg->rcu, free_sched_group_rcu);
6886 void sched_offline_group(struct task_group *tg)
6888 unsigned long flags;
6891 /* end participation in shares distribution */
6892 for_each_possible_cpu(i)
6893 unregister_fair_sched_group(tg, i);
6895 spin_lock_irqsave(&task_group_lock, flags);
6896 list_del_rcu(&tg->list);
6897 list_del_rcu(&tg->siblings);
6898 spin_unlock_irqrestore(&task_group_lock, flags);
6901 /* change task's runqueue when it moves between groups.
6902 * The caller of this function should have put the task in its new group
6903 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6904 * reflect its new group.
6906 void sched_move_task(struct task_struct *tsk)
6908 struct task_group *tg;
6910 unsigned long flags;
6913 rq = task_rq_lock(tsk, &flags);
6915 running = task_current(rq, tsk);
6919 dequeue_task(rq, tsk, 0);
6920 if (unlikely(running))
6921 tsk->sched_class->put_prev_task(rq, tsk);
6923 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6924 lockdep_is_held(&tsk->sighand->siglock)),
6925 struct task_group, css);
6926 tg = autogroup_task_group(tsk, tg);
6927 tsk->sched_task_group = tg;
6929 #ifdef CONFIG_FAIR_GROUP_SCHED
6930 if (tsk->sched_class->task_move_group)
6931 tsk->sched_class->task_move_group(tsk, on_rq);
6934 set_task_rq(tsk, task_cpu(tsk));
6936 if (unlikely(running))
6937 tsk->sched_class->set_curr_task(rq);
6939 enqueue_task(rq, tsk, 0);
6941 task_rq_unlock(rq, tsk, &flags);
6943 #endif /* CONFIG_CGROUP_SCHED */
6945 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6946 static unsigned long to_ratio(u64 period, u64 runtime)
6948 if (runtime == RUNTIME_INF)
6951 return div64_u64(runtime << 20, period);
6955 #ifdef CONFIG_RT_GROUP_SCHED
6957 * Ensure that the real time constraints are schedulable.
6959 static DEFINE_MUTEX(rt_constraints_mutex);
6961 /* Must be called with tasklist_lock held */
6962 static inline int tg_has_rt_tasks(struct task_group *tg)
6964 struct task_struct *g, *p;
6966 do_each_thread(g, p) {
6967 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6969 } while_each_thread(g, p);
6974 struct rt_schedulable_data {
6975 struct task_group *tg;
6980 static int tg_rt_schedulable(struct task_group *tg, void *data)
6982 struct rt_schedulable_data *d = data;
6983 struct task_group *child;
6984 unsigned long total, sum = 0;
6985 u64 period, runtime;
6987 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6988 runtime = tg->rt_bandwidth.rt_runtime;
6991 period = d->rt_period;
6992 runtime = d->rt_runtime;
6996 * Cannot have more runtime than the period.
6998 if (runtime > period && runtime != RUNTIME_INF)
7002 * Ensure we don't starve existing RT tasks.
7004 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7007 total = to_ratio(period, runtime);
7010 * Nobody can have more than the global setting allows.
7012 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7016 * The sum of our children's runtime should not exceed our own.
7018 list_for_each_entry_rcu(child, &tg->children, siblings) {
7019 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7020 runtime = child->rt_bandwidth.rt_runtime;
7022 if (child == d->tg) {
7023 period = d->rt_period;
7024 runtime = d->rt_runtime;
7027 sum += to_ratio(period, runtime);
7036 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7040 struct rt_schedulable_data data = {
7042 .rt_period = period,
7043 .rt_runtime = runtime,
7047 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7053 static int tg_set_rt_bandwidth(struct task_group *tg,
7054 u64 rt_period, u64 rt_runtime)
7058 mutex_lock(&rt_constraints_mutex);
7059 read_lock(&tasklist_lock);
7060 err = __rt_schedulable(tg, rt_period, rt_runtime);
7064 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7065 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7066 tg->rt_bandwidth.rt_runtime = rt_runtime;
7068 for_each_possible_cpu(i) {
7069 struct rt_rq *rt_rq = tg->rt_rq[i];
7071 raw_spin_lock(&rt_rq->rt_runtime_lock);
7072 rt_rq->rt_runtime = rt_runtime;
7073 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7075 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7077 read_unlock(&tasklist_lock);
7078 mutex_unlock(&rt_constraints_mutex);
7083 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7085 u64 rt_runtime, rt_period;
7087 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7088 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7089 if (rt_runtime_us < 0)
7090 rt_runtime = RUNTIME_INF;
7092 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7095 static long sched_group_rt_runtime(struct task_group *tg)
7099 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7102 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7103 do_div(rt_runtime_us, NSEC_PER_USEC);
7104 return rt_runtime_us;
7107 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7109 u64 rt_runtime, rt_period;
7111 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7112 rt_runtime = tg->rt_bandwidth.rt_runtime;
7117 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7120 static long sched_group_rt_period(struct task_group *tg)
7124 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7125 do_div(rt_period_us, NSEC_PER_USEC);
7126 return rt_period_us;
7129 static int sched_rt_global_constraints(void)
7131 u64 runtime, period;
7134 if (sysctl_sched_rt_period <= 0)
7137 runtime = global_rt_runtime();
7138 period = global_rt_period();
7141 * Sanity check on the sysctl variables.
7143 if (runtime > period && runtime != RUNTIME_INF)
7146 mutex_lock(&rt_constraints_mutex);
7147 read_lock(&tasklist_lock);
7148 ret = __rt_schedulable(NULL, 0, 0);
7149 read_unlock(&tasklist_lock);
7150 mutex_unlock(&rt_constraints_mutex);
7155 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7157 /* Don't accept realtime tasks when there is no way for them to run */
7158 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7164 #else /* !CONFIG_RT_GROUP_SCHED */
7165 static int sched_rt_global_constraints(void)
7167 unsigned long flags;
7170 if (sysctl_sched_rt_period <= 0)
7174 * There's always some RT tasks in the root group
7175 * -- migration, kstopmachine etc..
7177 if (sysctl_sched_rt_runtime == 0)
7180 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7181 for_each_possible_cpu(i) {
7182 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7184 raw_spin_lock(&rt_rq->rt_runtime_lock);
7185 rt_rq->rt_runtime = global_rt_runtime();
7186 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7188 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7192 #endif /* CONFIG_RT_GROUP_SCHED */
7194 int sched_rr_handler(struct ctl_table *table, int write,
7195 void __user *buffer, size_t *lenp,
7199 static DEFINE_MUTEX(mutex);
7202 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7203 /* make sure that internally we keep jiffies */
7204 /* also, writing zero resets timeslice to default */
7205 if (!ret && write) {
7206 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7207 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7209 mutex_unlock(&mutex);
7213 int sched_rt_handler(struct ctl_table *table, int write,
7214 void __user *buffer, size_t *lenp,
7218 int old_period, old_runtime;
7219 static DEFINE_MUTEX(mutex);
7222 old_period = sysctl_sched_rt_period;
7223 old_runtime = sysctl_sched_rt_runtime;
7225 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7227 if (!ret && write) {
7228 ret = sched_rt_global_constraints();
7230 sysctl_sched_rt_period = old_period;
7231 sysctl_sched_rt_runtime = old_runtime;
7233 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7234 def_rt_bandwidth.rt_period =
7235 ns_to_ktime(global_rt_period());
7238 mutex_unlock(&mutex);
7243 #ifdef CONFIG_CGROUP_SCHED
7245 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7247 return css ? container_of(css, struct task_group, css) : NULL;
7250 static struct cgroup_subsys_state *
7251 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7253 struct task_group *parent = css_tg(parent_css);
7254 struct task_group *tg;
7257 /* This is early initialization for the top cgroup */
7258 return &root_task_group.css;
7261 tg = sched_create_group(parent);
7263 return ERR_PTR(-ENOMEM);
7268 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7270 struct task_group *tg = css_tg(css);
7271 struct task_group *parent = css_tg(css_parent(css));
7274 sched_online_group(tg, parent);
7278 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7280 struct task_group *tg = css_tg(css);
7282 sched_destroy_group(tg);
7285 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7287 struct task_group *tg = css_tg(css);
7289 sched_offline_group(tg);
7292 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7293 struct cgroup_taskset *tset)
7295 struct task_struct *task;
7297 cgroup_taskset_for_each(task, css, tset) {
7298 #ifdef CONFIG_RT_GROUP_SCHED
7299 if (!sched_rt_can_attach(css_tg(css), task))
7302 /* We don't support RT-tasks being in separate groups */
7303 if (task->sched_class != &fair_sched_class)
7310 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7311 struct cgroup_taskset *tset)
7313 struct task_struct *task;
7315 cgroup_taskset_for_each(task, css, tset)
7316 sched_move_task(task);
7319 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7320 struct cgroup_subsys_state *old_css,
7321 struct task_struct *task)
7324 * cgroup_exit() is called in the copy_process() failure path.
7325 * Ignore this case since the task hasn't ran yet, this avoids
7326 * trying to poke a half freed task state from generic code.
7328 if (!(task->flags & PF_EXITING))
7331 sched_move_task(task);
7334 #ifdef CONFIG_FAIR_GROUP_SCHED
7335 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7336 struct cftype *cftype, u64 shareval)
7338 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7341 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7344 struct task_group *tg = css_tg(css);
7346 return (u64) scale_load_down(tg->shares);
7349 #ifdef CONFIG_CFS_BANDWIDTH
7350 static DEFINE_MUTEX(cfs_constraints_mutex);
7352 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7353 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7355 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7357 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7359 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7360 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7362 if (tg == &root_task_group)
7366 * Ensure we have at some amount of bandwidth every period. This is
7367 * to prevent reaching a state of large arrears when throttled via
7368 * entity_tick() resulting in prolonged exit starvation.
7370 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7374 * Likewise, bound things on the otherside by preventing insane quota
7375 * periods. This also allows us to normalize in computing quota
7378 if (period > max_cfs_quota_period)
7381 mutex_lock(&cfs_constraints_mutex);
7382 ret = __cfs_schedulable(tg, period, quota);
7386 runtime_enabled = quota != RUNTIME_INF;
7387 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7389 * If we need to toggle cfs_bandwidth_used, off->on must occur
7390 * before making related changes, and on->off must occur afterwards
7392 if (runtime_enabled && !runtime_was_enabled)
7393 cfs_bandwidth_usage_inc();
7394 raw_spin_lock_irq(&cfs_b->lock);
7395 cfs_b->period = ns_to_ktime(period);
7396 cfs_b->quota = quota;
7398 __refill_cfs_bandwidth_runtime(cfs_b);
7399 /* restart the period timer (if active) to handle new period expiry */
7400 if (runtime_enabled && cfs_b->timer_active) {
7401 /* force a reprogram */
7402 cfs_b->timer_active = 0;
7403 __start_cfs_bandwidth(cfs_b);
7405 raw_spin_unlock_irq(&cfs_b->lock);
7407 for_each_possible_cpu(i) {
7408 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7409 struct rq *rq = cfs_rq->rq;
7411 raw_spin_lock_irq(&rq->lock);
7412 cfs_rq->runtime_enabled = runtime_enabled;
7413 cfs_rq->runtime_remaining = 0;
7415 if (cfs_rq->throttled)
7416 unthrottle_cfs_rq(cfs_rq);
7417 raw_spin_unlock_irq(&rq->lock);
7419 if (runtime_was_enabled && !runtime_enabled)
7420 cfs_bandwidth_usage_dec();
7422 mutex_unlock(&cfs_constraints_mutex);
7427 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7431 period = ktime_to_ns(tg->cfs_bandwidth.period);
7432 if (cfs_quota_us < 0)
7433 quota = RUNTIME_INF;
7435 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7437 return tg_set_cfs_bandwidth(tg, period, quota);
7440 long tg_get_cfs_quota(struct task_group *tg)
7444 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7447 quota_us = tg->cfs_bandwidth.quota;
7448 do_div(quota_us, NSEC_PER_USEC);
7453 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7457 period = (u64)cfs_period_us * NSEC_PER_USEC;
7458 quota = tg->cfs_bandwidth.quota;
7460 return tg_set_cfs_bandwidth(tg, period, quota);
7463 long tg_get_cfs_period(struct task_group *tg)
7467 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7468 do_div(cfs_period_us, NSEC_PER_USEC);
7470 return cfs_period_us;
7473 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7476 return tg_get_cfs_quota(css_tg(css));
7479 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7480 struct cftype *cftype, s64 cfs_quota_us)
7482 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7485 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7488 return tg_get_cfs_period(css_tg(css));
7491 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7492 struct cftype *cftype, u64 cfs_period_us)
7494 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7497 struct cfs_schedulable_data {
7498 struct task_group *tg;
7503 * normalize group quota/period to be quota/max_period
7504 * note: units are usecs
7506 static u64 normalize_cfs_quota(struct task_group *tg,
7507 struct cfs_schedulable_data *d)
7515 period = tg_get_cfs_period(tg);
7516 quota = tg_get_cfs_quota(tg);
7519 /* note: these should typically be equivalent */
7520 if (quota == RUNTIME_INF || quota == -1)
7523 return to_ratio(period, quota);
7526 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7528 struct cfs_schedulable_data *d = data;
7529 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7530 s64 quota = 0, parent_quota = -1;
7533 quota = RUNTIME_INF;
7535 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7537 quota = normalize_cfs_quota(tg, d);
7538 parent_quota = parent_b->hierarchal_quota;
7541 * ensure max(child_quota) <= parent_quota, inherit when no
7544 if (quota == RUNTIME_INF)
7545 quota = parent_quota;
7546 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7549 cfs_b->hierarchal_quota = quota;
7554 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7557 struct cfs_schedulable_data data = {
7563 if (quota != RUNTIME_INF) {
7564 do_div(data.period, NSEC_PER_USEC);
7565 do_div(data.quota, NSEC_PER_USEC);
7569 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7575 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7576 struct cgroup_map_cb *cb)
7578 struct task_group *tg = css_tg(css);
7579 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7581 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7582 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7583 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7587 #endif /* CONFIG_CFS_BANDWIDTH */
7588 #endif /* CONFIG_FAIR_GROUP_SCHED */
7590 #ifdef CONFIG_RT_GROUP_SCHED
7591 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7592 struct cftype *cft, s64 val)
7594 return sched_group_set_rt_runtime(css_tg(css), val);
7597 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7600 return sched_group_rt_runtime(css_tg(css));
7603 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7604 struct cftype *cftype, u64 rt_period_us)
7606 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7609 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7612 return sched_group_rt_period(css_tg(css));
7614 #endif /* CONFIG_RT_GROUP_SCHED */
7616 static struct cftype cpu_files[] = {
7617 #ifdef CONFIG_FAIR_GROUP_SCHED
7620 .read_u64 = cpu_shares_read_u64,
7621 .write_u64 = cpu_shares_write_u64,
7624 #ifdef CONFIG_CFS_BANDWIDTH
7626 .name = "cfs_quota_us",
7627 .read_s64 = cpu_cfs_quota_read_s64,
7628 .write_s64 = cpu_cfs_quota_write_s64,
7631 .name = "cfs_period_us",
7632 .read_u64 = cpu_cfs_period_read_u64,
7633 .write_u64 = cpu_cfs_period_write_u64,
7637 .read_map = cpu_stats_show,
7640 #ifdef CONFIG_RT_GROUP_SCHED
7642 .name = "rt_runtime_us",
7643 .read_s64 = cpu_rt_runtime_read,
7644 .write_s64 = cpu_rt_runtime_write,
7647 .name = "rt_period_us",
7648 .read_u64 = cpu_rt_period_read_uint,
7649 .write_u64 = cpu_rt_period_write_uint,
7655 struct cgroup_subsys cpu_cgroup_subsys = {
7657 .css_alloc = cpu_cgroup_css_alloc,
7658 .css_free = cpu_cgroup_css_free,
7659 .css_online = cpu_cgroup_css_online,
7660 .css_offline = cpu_cgroup_css_offline,
7661 .can_attach = cpu_cgroup_can_attach,
7662 .attach = cpu_cgroup_attach,
7663 .exit = cpu_cgroup_exit,
7664 .subsys_id = cpu_cgroup_subsys_id,
7665 .base_cftypes = cpu_files,
7669 #endif /* CONFIG_CGROUP_SCHED */
7671 void dump_cpu_task(int cpu)
7673 pr_info("Task dump for CPU %d:\n", cpu);
7674 sched_show_task(cpu_curr(cpu));