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 || dl_task(p))
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, enqueue_flag = 0;
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);
2811 p->pi_top_task = rt_mutex_get_top_task(p);
2813 prev_class = p->sched_class;
2815 running = task_current(rq, p);
2817 dequeue_task(rq, p, 0);
2819 p->sched_class->put_prev_task(rq, p);
2822 * Boosting condition are:
2823 * 1. -rt task is running and holds mutex A
2824 * --> -dl task blocks on mutex A
2826 * 2. -dl task is running and holds mutex A
2827 * --> -dl task blocks on mutex A and could preempt the
2830 if (dl_prio(prio)) {
2831 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2832 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2833 p->dl.dl_boosted = 1;
2834 p->dl.dl_throttled = 0;
2835 enqueue_flag = ENQUEUE_REPLENISH;
2837 p->dl.dl_boosted = 0;
2838 p->sched_class = &dl_sched_class;
2839 } else if (rt_prio(prio)) {
2840 if (dl_prio(oldprio))
2841 p->dl.dl_boosted = 0;
2843 enqueue_flag = ENQUEUE_HEAD;
2844 p->sched_class = &rt_sched_class;
2846 if (dl_prio(oldprio))
2847 p->dl.dl_boosted = 0;
2848 p->sched_class = &fair_sched_class;
2854 p->sched_class->set_curr_task(rq);
2856 enqueue_task(rq, p, enqueue_flag);
2858 check_class_changed(rq, p, prev_class, oldprio);
2860 __task_rq_unlock(rq);
2864 void set_user_nice(struct task_struct *p, long nice)
2866 int old_prio, delta, on_rq;
2867 unsigned long flags;
2870 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
2873 * We have to be careful, if called from sys_setpriority(),
2874 * the task might be in the middle of scheduling on another CPU.
2876 rq = task_rq_lock(p, &flags);
2878 * The RT priorities are set via sched_setscheduler(), but we still
2879 * allow the 'normal' nice value to be set - but as expected
2880 * it wont have any effect on scheduling until the task is
2881 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2883 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2884 p->static_prio = NICE_TO_PRIO(nice);
2889 dequeue_task(rq, p, 0);
2891 p->static_prio = NICE_TO_PRIO(nice);
2894 p->prio = effective_prio(p);
2895 delta = p->prio - old_prio;
2898 enqueue_task(rq, p, 0);
2900 * If the task increased its priority or is running and
2901 * lowered its priority, then reschedule its CPU:
2903 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2904 resched_task(rq->curr);
2907 task_rq_unlock(rq, p, &flags);
2909 EXPORT_SYMBOL(set_user_nice);
2912 * can_nice - check if a task can reduce its nice value
2916 int can_nice(const struct task_struct *p, const int nice)
2918 /* convert nice value [19,-20] to rlimit style value [1,40] */
2919 int nice_rlim = 20 - nice;
2921 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
2922 capable(CAP_SYS_NICE));
2925 #ifdef __ARCH_WANT_SYS_NICE
2928 * sys_nice - change the priority of the current process.
2929 * @increment: priority increment
2931 * sys_setpriority is a more generic, but much slower function that
2932 * does similar things.
2934 SYSCALL_DEFINE1(nice, int, increment)
2939 * Setpriority might change our priority at the same moment.
2940 * We don't have to worry. Conceptually one call occurs first
2941 * and we have a single winner.
2943 if (increment < -40)
2948 nice = TASK_NICE(current) + increment;
2954 if (increment < 0 && !can_nice(current, nice))
2957 retval = security_task_setnice(current, nice);
2961 set_user_nice(current, nice);
2968 * task_prio - return the priority value of a given task.
2969 * @p: the task in question.
2971 * Return: The priority value as seen by users in /proc.
2972 * RT tasks are offset by -200. Normal tasks are centered
2973 * around 0, value goes from -16 to +15.
2975 int task_prio(const struct task_struct *p)
2977 return p->prio - MAX_RT_PRIO;
2981 * task_nice - return the nice value of a given task.
2982 * @p: the task in question.
2984 * Return: The nice value [ -20 ... 0 ... 19 ].
2986 int task_nice(const struct task_struct *p)
2988 return TASK_NICE(p);
2990 EXPORT_SYMBOL(task_nice);
2993 * idle_cpu - is a given cpu idle currently?
2994 * @cpu: the processor in question.
2996 * Return: 1 if the CPU is currently idle. 0 otherwise.
2998 int idle_cpu(int cpu)
3000 struct rq *rq = cpu_rq(cpu);
3002 if (rq->curr != rq->idle)
3009 if (!llist_empty(&rq->wake_list))
3017 * idle_task - return the idle task for a given cpu.
3018 * @cpu: the processor in question.
3020 * Return: The idle task for the cpu @cpu.
3022 struct task_struct *idle_task(int cpu)
3024 return cpu_rq(cpu)->idle;
3028 * find_process_by_pid - find a process with a matching PID value.
3029 * @pid: the pid in question.
3031 * The task of @pid, if found. %NULL otherwise.
3033 static struct task_struct *find_process_by_pid(pid_t pid)
3035 return pid ? find_task_by_vpid(pid) : current;
3039 * This function initializes the sched_dl_entity of a newly becoming
3040 * SCHED_DEADLINE task.
3042 * Only the static values are considered here, the actual runtime and the
3043 * absolute deadline will be properly calculated when the task is enqueued
3044 * for the first time with its new policy.
3047 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3049 struct sched_dl_entity *dl_se = &p->dl;
3051 init_dl_task_timer(dl_se);
3052 dl_se->dl_runtime = attr->sched_runtime;
3053 dl_se->dl_deadline = attr->sched_deadline;
3054 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3055 dl_se->flags = attr->sched_flags;
3056 dl_se->dl_throttled = 0;
3060 /* Actually do priority change: must hold pi & rq lock. */
3061 static void __setscheduler(struct rq *rq, struct task_struct *p,
3062 const struct sched_attr *attr)
3064 int policy = attr->sched_policy;
3068 if (dl_policy(policy))
3069 __setparam_dl(p, attr);
3070 else if (rt_policy(policy))
3071 p->rt_priority = attr->sched_priority;
3073 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3075 p->normal_prio = normal_prio(p);
3076 p->prio = rt_mutex_getprio(p);
3078 if (dl_prio(p->prio))
3079 p->sched_class = &dl_sched_class;
3080 else if (rt_prio(p->prio))
3081 p->sched_class = &rt_sched_class;
3083 p->sched_class = &fair_sched_class;
3089 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3091 struct sched_dl_entity *dl_se = &p->dl;
3093 attr->sched_priority = p->rt_priority;
3094 attr->sched_runtime = dl_se->dl_runtime;
3095 attr->sched_deadline = dl_se->dl_deadline;
3096 attr->sched_period = dl_se->dl_period;
3097 attr->sched_flags = dl_se->flags;
3101 * This function validates the new parameters of a -deadline task.
3102 * We ask for the deadline not being zero, and greater or equal
3103 * than the runtime, as well as the period of being zero or
3104 * greater than deadline.
3107 __checkparam_dl(const struct sched_attr *attr)
3109 return attr && attr->sched_deadline != 0 &&
3110 (attr->sched_period == 0 ||
3111 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3112 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0;
3116 * check the target process has a UID that matches the current process's
3118 static bool check_same_owner(struct task_struct *p)
3120 const struct cred *cred = current_cred(), *pcred;
3124 pcred = __task_cred(p);
3125 match = (uid_eq(cred->euid, pcred->euid) ||
3126 uid_eq(cred->euid, pcred->uid));
3131 static int __sched_setscheduler(struct task_struct *p,
3132 const struct sched_attr *attr,
3135 int retval, oldprio, oldpolicy = -1, on_rq, running;
3136 int policy = attr->sched_policy;
3137 unsigned long flags;
3138 const struct sched_class *prev_class;
3142 /* may grab non-irq protected spin_locks */
3143 BUG_ON(in_interrupt());
3145 /* double check policy once rq lock held */
3147 reset_on_fork = p->sched_reset_on_fork;
3148 policy = oldpolicy = p->policy;
3150 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3151 policy &= ~SCHED_RESET_ON_FORK;
3153 if (policy != SCHED_DEADLINE &&
3154 policy != SCHED_FIFO && policy != SCHED_RR &&
3155 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3156 policy != SCHED_IDLE)
3161 * Valid priorities for SCHED_FIFO and SCHED_RR are
3162 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3163 * SCHED_BATCH and SCHED_IDLE is 0.
3165 if (attr->sched_priority < 0 ||
3166 (p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3167 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3169 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3170 (rt_policy(policy) != (attr->sched_priority != 0)))
3174 * Allow unprivileged RT tasks to decrease priority:
3176 if (user && !capable(CAP_SYS_NICE)) {
3177 if (fair_policy(policy)) {
3178 if (!can_nice(p, attr->sched_nice))
3182 if (rt_policy(policy)) {
3183 unsigned long rlim_rtprio =
3184 task_rlimit(p, RLIMIT_RTPRIO);
3186 /* can't set/change the rt policy */
3187 if (policy != p->policy && !rlim_rtprio)
3190 /* can't increase priority */
3191 if (attr->sched_priority > p->rt_priority &&
3192 attr->sched_priority > rlim_rtprio)
3197 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3198 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3200 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3201 if (!can_nice(p, TASK_NICE(p)))
3205 /* can't change other user's priorities */
3206 if (!check_same_owner(p))
3209 /* Normal users shall not reset the sched_reset_on_fork flag */
3210 if (p->sched_reset_on_fork && !reset_on_fork)
3215 retval = security_task_setscheduler(p);
3221 * make sure no PI-waiters arrive (or leave) while we are
3222 * changing the priority of the task:
3224 * To be able to change p->policy safely, the appropriate
3225 * runqueue lock must be held.
3227 rq = task_rq_lock(p, &flags);
3230 * Changing the policy of the stop threads its a very bad idea
3232 if (p == rq->stop) {
3233 task_rq_unlock(rq, p, &flags);
3238 * If not changing anything there's no need to proceed further:
3240 if (unlikely(policy == p->policy)) {
3241 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p))
3243 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3245 if (dl_policy(policy))
3248 task_rq_unlock(rq, p, &flags);
3253 #ifdef CONFIG_RT_GROUP_SCHED
3256 * Do not allow realtime tasks into groups that have no runtime
3259 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3260 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3261 !task_group_is_autogroup(task_group(p))) {
3262 task_rq_unlock(rq, p, &flags);
3268 /* recheck policy now with rq lock held */
3269 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3270 policy = oldpolicy = -1;
3271 task_rq_unlock(rq, p, &flags);
3275 running = task_current(rq, p);
3277 dequeue_task(rq, p, 0);
3279 p->sched_class->put_prev_task(rq, p);
3281 p->sched_reset_on_fork = reset_on_fork;
3284 prev_class = p->sched_class;
3285 __setscheduler(rq, p, attr);
3288 p->sched_class->set_curr_task(rq);
3290 enqueue_task(rq, p, 0);
3292 check_class_changed(rq, p, prev_class, oldprio);
3293 task_rq_unlock(rq, p, &flags);
3295 rt_mutex_adjust_pi(p);
3301 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3302 * @p: the task in question.
3303 * @policy: new policy.
3304 * @param: structure containing the new RT priority.
3306 * Return: 0 on success. An error code otherwise.
3308 * NOTE that the task may be already dead.
3310 int sched_setscheduler(struct task_struct *p, int policy,
3311 const struct sched_param *param)
3313 struct sched_attr attr = {
3314 .sched_policy = policy,
3315 .sched_priority = param->sched_priority
3317 return __sched_setscheduler(p, &attr, true);
3319 EXPORT_SYMBOL_GPL(sched_setscheduler);
3321 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3323 return __sched_setscheduler(p, attr, true);
3325 EXPORT_SYMBOL_GPL(sched_setattr);
3328 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3329 * @p: the task in question.
3330 * @policy: new policy.
3331 * @param: structure containing the new RT priority.
3333 * Just like sched_setscheduler, only don't bother checking if the
3334 * current context has permission. For example, this is needed in
3335 * stop_machine(): we create temporary high priority worker threads,
3336 * but our caller might not have that capability.
3338 * Return: 0 on success. An error code otherwise.
3340 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3341 const struct sched_param *param)
3343 struct sched_attr attr = {
3344 .sched_policy = policy,
3345 .sched_priority = param->sched_priority
3347 return __sched_setscheduler(p, &attr, false);
3351 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3353 struct sched_param lparam;
3354 struct task_struct *p;
3357 if (!param || pid < 0)
3359 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3364 p = find_process_by_pid(pid);
3366 retval = sched_setscheduler(p, policy, &lparam);
3373 * Mimics kernel/events/core.c perf_copy_attr().
3375 static int sched_copy_attr(struct sched_attr __user *uattr,
3376 struct sched_attr *attr)
3381 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3385 * zero the full structure, so that a short copy will be nice.
3387 memset(attr, 0, sizeof(*attr));
3389 ret = get_user(size, &uattr->size);
3393 if (size > PAGE_SIZE) /* silly large */
3396 if (!size) /* abi compat */
3397 size = SCHED_ATTR_SIZE_VER0;
3399 if (size < SCHED_ATTR_SIZE_VER0)
3403 * If we're handed a bigger struct than we know of,
3404 * ensure all the unknown bits are 0 - i.e. new
3405 * user-space does not rely on any kernel feature
3406 * extensions we dont know about yet.
3408 if (size > sizeof(*attr)) {
3409 unsigned char __user *addr;
3410 unsigned char __user *end;
3413 addr = (void __user *)uattr + sizeof(*attr);
3414 end = (void __user *)uattr + size;
3416 for (; addr < end; addr++) {
3417 ret = get_user(val, addr);
3423 size = sizeof(*attr);
3426 ret = copy_from_user(attr, uattr, size);
3431 * XXX: do we want to be lenient like existing syscalls; or do we want
3432 * to be strict and return an error on out-of-bounds values?
3434 attr->sched_nice = clamp(attr->sched_nice, -20, 19);
3440 put_user(sizeof(*attr), &uattr->size);
3446 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3447 * @pid: the pid in question.
3448 * @policy: new policy.
3449 * @param: structure containing the new RT priority.
3451 * Return: 0 on success. An error code otherwise.
3453 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3454 struct sched_param __user *, param)
3456 /* negative values for policy are not valid */
3460 return do_sched_setscheduler(pid, policy, param);
3464 * sys_sched_setparam - set/change the RT priority of a thread
3465 * @pid: the pid in question.
3466 * @param: structure containing the new RT priority.
3468 * Return: 0 on success. An error code otherwise.
3470 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3472 return do_sched_setscheduler(pid, -1, param);
3476 * sys_sched_setattr - same as above, but with extended sched_attr
3477 * @pid: the pid in question.
3478 * @attr: structure containing the extended parameters.
3480 SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr)
3482 struct sched_attr attr;
3483 struct task_struct *p;
3486 if (!uattr || pid < 0)
3489 if (sched_copy_attr(uattr, &attr))
3494 p = find_process_by_pid(pid);
3496 retval = sched_setattr(p, &attr);
3503 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3504 * @pid: the pid in question.
3506 * Return: On success, the policy of the thread. Otherwise, a negative error
3509 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3511 struct task_struct *p;
3519 p = find_process_by_pid(pid);
3521 retval = security_task_getscheduler(p);
3524 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3531 * sys_sched_getparam - get the RT priority of a thread
3532 * @pid: the pid in question.
3533 * @param: structure containing the RT priority.
3535 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3538 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3540 struct sched_param lp;
3541 struct task_struct *p;
3544 if (!param || pid < 0)
3548 p = find_process_by_pid(pid);
3553 retval = security_task_getscheduler(p);
3557 if (task_has_dl_policy(p)) {
3561 lp.sched_priority = p->rt_priority;
3565 * This one might sleep, we cannot do it with a spinlock held ...
3567 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3576 static int sched_read_attr(struct sched_attr __user *uattr,
3577 struct sched_attr *attr,
3582 if (!access_ok(VERIFY_WRITE, uattr, usize))
3586 * If we're handed a smaller struct than we know of,
3587 * ensure all the unknown bits are 0 - i.e. old
3588 * user-space does not get uncomplete information.
3590 if (usize < sizeof(*attr)) {
3591 unsigned char *addr;
3594 addr = (void *)attr + usize;
3595 end = (void *)attr + sizeof(*attr);
3597 for (; addr < end; addr++) {
3605 ret = copy_to_user(uattr, attr, usize);
3618 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3619 * @pid: the pid in question.
3620 * @attr: structure containing the extended parameters.
3621 * @size: sizeof(attr) for fwd/bwd comp.
3623 SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3626 struct sched_attr attr = {
3627 .size = sizeof(struct sched_attr),
3629 struct task_struct *p;
3632 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3633 size < SCHED_ATTR_SIZE_VER0)
3637 p = find_process_by_pid(pid);
3642 retval = security_task_getscheduler(p);
3646 attr.sched_policy = p->policy;
3647 if (task_has_dl_policy(p))
3648 __getparam_dl(p, &attr);
3649 else if (task_has_rt_policy(p))
3650 attr.sched_priority = p->rt_priority;
3652 attr.sched_nice = TASK_NICE(p);
3656 retval = sched_read_attr(uattr, &attr, size);
3664 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3666 cpumask_var_t cpus_allowed, new_mask;
3667 struct task_struct *p;
3672 p = find_process_by_pid(pid);
3678 /* Prevent p going away */
3682 if (p->flags & PF_NO_SETAFFINITY) {
3686 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3690 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3692 goto out_free_cpus_allowed;
3695 if (!check_same_owner(p)) {
3697 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3704 retval = security_task_setscheduler(p);
3708 cpuset_cpus_allowed(p, cpus_allowed);
3709 cpumask_and(new_mask, in_mask, cpus_allowed);
3711 retval = set_cpus_allowed_ptr(p, new_mask);
3714 cpuset_cpus_allowed(p, cpus_allowed);
3715 if (!cpumask_subset(new_mask, cpus_allowed)) {
3717 * We must have raced with a concurrent cpuset
3718 * update. Just reset the cpus_allowed to the
3719 * cpuset's cpus_allowed
3721 cpumask_copy(new_mask, cpus_allowed);
3726 free_cpumask_var(new_mask);
3727 out_free_cpus_allowed:
3728 free_cpumask_var(cpus_allowed);
3734 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3735 struct cpumask *new_mask)
3737 if (len < cpumask_size())
3738 cpumask_clear(new_mask);
3739 else if (len > cpumask_size())
3740 len = cpumask_size();
3742 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3746 * sys_sched_setaffinity - set the cpu affinity of a process
3747 * @pid: pid of the process
3748 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3749 * @user_mask_ptr: user-space pointer to the new cpu mask
3751 * Return: 0 on success. An error code otherwise.
3753 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3754 unsigned long __user *, user_mask_ptr)
3756 cpumask_var_t new_mask;
3759 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3762 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3764 retval = sched_setaffinity(pid, new_mask);
3765 free_cpumask_var(new_mask);
3769 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3771 struct task_struct *p;
3772 unsigned long flags;
3778 p = find_process_by_pid(pid);
3782 retval = security_task_getscheduler(p);
3786 raw_spin_lock_irqsave(&p->pi_lock, flags);
3787 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3788 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3797 * sys_sched_getaffinity - get the cpu affinity of a process
3798 * @pid: pid of the process
3799 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3800 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3802 * Return: 0 on success. An error code otherwise.
3804 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3805 unsigned long __user *, user_mask_ptr)
3810 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3812 if (len & (sizeof(unsigned long)-1))
3815 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3818 ret = sched_getaffinity(pid, mask);
3820 size_t retlen = min_t(size_t, len, cpumask_size());
3822 if (copy_to_user(user_mask_ptr, mask, retlen))
3827 free_cpumask_var(mask);
3833 * sys_sched_yield - yield the current processor to other threads.
3835 * This function yields the current CPU to other tasks. If there are no
3836 * other threads running on this CPU then this function will return.
3840 SYSCALL_DEFINE0(sched_yield)
3842 struct rq *rq = this_rq_lock();
3844 schedstat_inc(rq, yld_count);
3845 current->sched_class->yield_task(rq);
3848 * Since we are going to call schedule() anyway, there's
3849 * no need to preempt or enable interrupts:
3851 __release(rq->lock);
3852 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3853 do_raw_spin_unlock(&rq->lock);
3854 sched_preempt_enable_no_resched();
3861 static void __cond_resched(void)
3863 __preempt_count_add(PREEMPT_ACTIVE);
3865 __preempt_count_sub(PREEMPT_ACTIVE);
3868 int __sched _cond_resched(void)
3870 if (should_resched()) {
3876 EXPORT_SYMBOL(_cond_resched);
3879 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3880 * call schedule, and on return reacquire the lock.
3882 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3883 * operations here to prevent schedule() from being called twice (once via
3884 * spin_unlock(), once by hand).
3886 int __cond_resched_lock(spinlock_t *lock)
3888 int resched = should_resched();
3891 lockdep_assert_held(lock);
3893 if (spin_needbreak(lock) || resched) {
3904 EXPORT_SYMBOL(__cond_resched_lock);
3906 int __sched __cond_resched_softirq(void)
3908 BUG_ON(!in_softirq());
3910 if (should_resched()) {
3918 EXPORT_SYMBOL(__cond_resched_softirq);
3921 * yield - yield the current processor to other threads.
3923 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3925 * The scheduler is at all times free to pick the calling task as the most
3926 * eligible task to run, if removing the yield() call from your code breaks
3927 * it, its already broken.
3929 * Typical broken usage is:
3934 * where one assumes that yield() will let 'the other' process run that will
3935 * make event true. If the current task is a SCHED_FIFO task that will never
3936 * happen. Never use yield() as a progress guarantee!!
3938 * If you want to use yield() to wait for something, use wait_event().
3939 * If you want to use yield() to be 'nice' for others, use cond_resched().
3940 * If you still want to use yield(), do not!
3942 void __sched yield(void)
3944 set_current_state(TASK_RUNNING);
3947 EXPORT_SYMBOL(yield);
3950 * yield_to - yield the current processor to another thread in
3951 * your thread group, or accelerate that thread toward the
3952 * processor it's on.
3954 * @preempt: whether task preemption is allowed or not
3956 * It's the caller's job to ensure that the target task struct
3957 * can't go away on us before we can do any checks.
3960 * true (>0) if we indeed boosted the target task.
3961 * false (0) if we failed to boost the target.
3962 * -ESRCH if there's no task to yield to.
3964 bool __sched yield_to(struct task_struct *p, bool preempt)
3966 struct task_struct *curr = current;
3967 struct rq *rq, *p_rq;
3968 unsigned long flags;
3971 local_irq_save(flags);
3977 * If we're the only runnable task on the rq and target rq also
3978 * has only one task, there's absolutely no point in yielding.
3980 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3985 double_rq_lock(rq, p_rq);
3986 if (task_rq(p) != p_rq) {
3987 double_rq_unlock(rq, p_rq);
3991 if (!curr->sched_class->yield_to_task)
3994 if (curr->sched_class != p->sched_class)
3997 if (task_running(p_rq, p) || p->state)
4000 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4002 schedstat_inc(rq, yld_count);
4004 * Make p's CPU reschedule; pick_next_entity takes care of
4007 if (preempt && rq != p_rq)
4008 resched_task(p_rq->curr);
4012 double_rq_unlock(rq, p_rq);
4014 local_irq_restore(flags);
4021 EXPORT_SYMBOL_GPL(yield_to);
4024 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4025 * that process accounting knows that this is a task in IO wait state.
4027 void __sched io_schedule(void)
4029 struct rq *rq = raw_rq();
4031 delayacct_blkio_start();
4032 atomic_inc(&rq->nr_iowait);
4033 blk_flush_plug(current);
4034 current->in_iowait = 1;
4036 current->in_iowait = 0;
4037 atomic_dec(&rq->nr_iowait);
4038 delayacct_blkio_end();
4040 EXPORT_SYMBOL(io_schedule);
4042 long __sched io_schedule_timeout(long timeout)
4044 struct rq *rq = raw_rq();
4047 delayacct_blkio_start();
4048 atomic_inc(&rq->nr_iowait);
4049 blk_flush_plug(current);
4050 current->in_iowait = 1;
4051 ret = schedule_timeout(timeout);
4052 current->in_iowait = 0;
4053 atomic_dec(&rq->nr_iowait);
4054 delayacct_blkio_end();
4059 * sys_sched_get_priority_max - return maximum RT priority.
4060 * @policy: scheduling class.
4062 * Return: On success, this syscall returns the maximum
4063 * rt_priority that can be used by a given scheduling class.
4064 * On failure, a negative error code is returned.
4066 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4073 ret = MAX_USER_RT_PRIO-1;
4075 case SCHED_DEADLINE:
4086 * sys_sched_get_priority_min - return minimum RT priority.
4087 * @policy: scheduling class.
4089 * Return: On success, this syscall returns the minimum
4090 * rt_priority that can be used by a given scheduling class.
4091 * On failure, a negative error code is returned.
4093 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4102 case SCHED_DEADLINE:
4112 * sys_sched_rr_get_interval - return the default timeslice of a process.
4113 * @pid: pid of the process.
4114 * @interval: userspace pointer to the timeslice value.
4116 * this syscall writes the default timeslice value of a given process
4117 * into the user-space timespec buffer. A value of '0' means infinity.
4119 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4122 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4123 struct timespec __user *, interval)
4125 struct task_struct *p;
4126 unsigned int time_slice;
4127 unsigned long flags;
4137 p = find_process_by_pid(pid);
4141 retval = security_task_getscheduler(p);
4145 rq = task_rq_lock(p, &flags);
4146 time_slice = p->sched_class->get_rr_interval(rq, p);
4147 task_rq_unlock(rq, p, &flags);
4150 jiffies_to_timespec(time_slice, &t);
4151 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4159 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4161 void sched_show_task(struct task_struct *p)
4163 unsigned long free = 0;
4167 state = p->state ? __ffs(p->state) + 1 : 0;
4168 printk(KERN_INFO "%-15.15s %c", p->comm,
4169 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4170 #if BITS_PER_LONG == 32
4171 if (state == TASK_RUNNING)
4172 printk(KERN_CONT " running ");
4174 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4176 if (state == TASK_RUNNING)
4177 printk(KERN_CONT " running task ");
4179 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4181 #ifdef CONFIG_DEBUG_STACK_USAGE
4182 free = stack_not_used(p);
4185 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4187 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4188 task_pid_nr(p), ppid,
4189 (unsigned long)task_thread_info(p)->flags);
4191 print_worker_info(KERN_INFO, p);
4192 show_stack(p, NULL);
4195 void show_state_filter(unsigned long state_filter)
4197 struct task_struct *g, *p;
4199 #if BITS_PER_LONG == 32
4201 " task PC stack pid father\n");
4204 " task PC stack pid father\n");
4207 do_each_thread(g, p) {
4209 * reset the NMI-timeout, listing all files on a slow
4210 * console might take a lot of time:
4212 touch_nmi_watchdog();
4213 if (!state_filter || (p->state & state_filter))
4215 } while_each_thread(g, p);
4217 touch_all_softlockup_watchdogs();
4219 #ifdef CONFIG_SCHED_DEBUG
4220 sysrq_sched_debug_show();
4224 * Only show locks if all tasks are dumped:
4227 debug_show_all_locks();
4230 void init_idle_bootup_task(struct task_struct *idle)
4232 idle->sched_class = &idle_sched_class;
4236 * init_idle - set up an idle thread for a given CPU
4237 * @idle: task in question
4238 * @cpu: cpu the idle task belongs to
4240 * NOTE: this function does not set the idle thread's NEED_RESCHED
4241 * flag, to make booting more robust.
4243 void init_idle(struct task_struct *idle, int cpu)
4245 struct rq *rq = cpu_rq(cpu);
4246 unsigned long flags;
4248 raw_spin_lock_irqsave(&rq->lock, flags);
4250 __sched_fork(0, idle);
4251 idle->state = TASK_RUNNING;
4252 idle->se.exec_start = sched_clock();
4254 do_set_cpus_allowed(idle, cpumask_of(cpu));
4256 * We're having a chicken and egg problem, even though we are
4257 * holding rq->lock, the cpu isn't yet set to this cpu so the
4258 * lockdep check in task_group() will fail.
4260 * Similar case to sched_fork(). / Alternatively we could
4261 * use task_rq_lock() here and obtain the other rq->lock.
4266 __set_task_cpu(idle, cpu);
4269 rq->curr = rq->idle = idle;
4270 #if defined(CONFIG_SMP)
4273 raw_spin_unlock_irqrestore(&rq->lock, flags);
4275 /* Set the preempt count _outside_ the spinlocks! */
4276 init_idle_preempt_count(idle, cpu);
4279 * The idle tasks have their own, simple scheduling class:
4281 idle->sched_class = &idle_sched_class;
4282 ftrace_graph_init_idle_task(idle, cpu);
4283 vtime_init_idle(idle, cpu);
4284 #if defined(CONFIG_SMP)
4285 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4290 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4292 if (p->sched_class && p->sched_class->set_cpus_allowed)
4293 p->sched_class->set_cpus_allowed(p, new_mask);
4295 cpumask_copy(&p->cpus_allowed, new_mask);
4296 p->nr_cpus_allowed = cpumask_weight(new_mask);
4300 * This is how migration works:
4302 * 1) we invoke migration_cpu_stop() on the target CPU using
4304 * 2) stopper starts to run (implicitly forcing the migrated thread
4306 * 3) it checks whether the migrated task is still in the wrong runqueue.
4307 * 4) if it's in the wrong runqueue then the migration thread removes
4308 * it and puts it into the right queue.
4309 * 5) stopper completes and stop_one_cpu() returns and the migration
4314 * Change a given task's CPU affinity. Migrate the thread to a
4315 * proper CPU and schedule it away if the CPU it's executing on
4316 * is removed from the allowed bitmask.
4318 * NOTE: the caller must have a valid reference to the task, the
4319 * task must not exit() & deallocate itself prematurely. The
4320 * call is not atomic; no spinlocks may be held.
4322 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4324 unsigned long flags;
4326 unsigned int dest_cpu;
4329 rq = task_rq_lock(p, &flags);
4331 if (cpumask_equal(&p->cpus_allowed, new_mask))
4334 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4339 do_set_cpus_allowed(p, new_mask);
4341 /* Can the task run on the task's current CPU? If so, we're done */
4342 if (cpumask_test_cpu(task_cpu(p), new_mask))
4345 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4347 struct migration_arg arg = { p, dest_cpu };
4348 /* Need help from migration thread: drop lock and wait. */
4349 task_rq_unlock(rq, p, &flags);
4350 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4351 tlb_migrate_finish(p->mm);
4355 task_rq_unlock(rq, p, &flags);
4359 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4362 * Move (not current) task off this cpu, onto dest cpu. We're doing
4363 * this because either it can't run here any more (set_cpus_allowed()
4364 * away from this CPU, or CPU going down), or because we're
4365 * attempting to rebalance this task on exec (sched_exec).
4367 * So we race with normal scheduler movements, but that's OK, as long
4368 * as the task is no longer on this CPU.
4370 * Returns non-zero if task was successfully migrated.
4372 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4374 struct rq *rq_dest, *rq_src;
4377 if (unlikely(!cpu_active(dest_cpu)))
4380 rq_src = cpu_rq(src_cpu);
4381 rq_dest = cpu_rq(dest_cpu);
4383 raw_spin_lock(&p->pi_lock);
4384 double_rq_lock(rq_src, rq_dest);
4385 /* Already moved. */
4386 if (task_cpu(p) != src_cpu)
4388 /* Affinity changed (again). */
4389 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4393 * If we're not on a rq, the next wake-up will ensure we're
4397 dequeue_task(rq_src, p, 0);
4398 set_task_cpu(p, dest_cpu);
4399 enqueue_task(rq_dest, p, 0);
4400 check_preempt_curr(rq_dest, p, 0);
4405 double_rq_unlock(rq_src, rq_dest);
4406 raw_spin_unlock(&p->pi_lock);
4410 #ifdef CONFIG_NUMA_BALANCING
4411 /* Migrate current task p to target_cpu */
4412 int migrate_task_to(struct task_struct *p, int target_cpu)
4414 struct migration_arg arg = { p, target_cpu };
4415 int curr_cpu = task_cpu(p);
4417 if (curr_cpu == target_cpu)
4420 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4423 /* TODO: This is not properly updating schedstats */
4425 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4429 * Requeue a task on a given node and accurately track the number of NUMA
4430 * tasks on the runqueues
4432 void sched_setnuma(struct task_struct *p, int nid)
4435 unsigned long flags;
4436 bool on_rq, running;
4438 rq = task_rq_lock(p, &flags);
4440 running = task_current(rq, p);
4443 dequeue_task(rq, p, 0);
4445 p->sched_class->put_prev_task(rq, p);
4447 p->numa_preferred_nid = nid;
4450 p->sched_class->set_curr_task(rq);
4452 enqueue_task(rq, p, 0);
4453 task_rq_unlock(rq, p, &flags);
4458 * migration_cpu_stop - this will be executed by a highprio stopper thread
4459 * and performs thread migration by bumping thread off CPU then
4460 * 'pushing' onto another runqueue.
4462 static int migration_cpu_stop(void *data)
4464 struct migration_arg *arg = data;
4467 * The original target cpu might have gone down and we might
4468 * be on another cpu but it doesn't matter.
4470 local_irq_disable();
4471 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4476 #ifdef CONFIG_HOTPLUG_CPU
4479 * Ensures that the idle task is using init_mm right before its cpu goes
4482 void idle_task_exit(void)
4484 struct mm_struct *mm = current->active_mm;
4486 BUG_ON(cpu_online(smp_processor_id()));
4489 switch_mm(mm, &init_mm, current);
4494 * Since this CPU is going 'away' for a while, fold any nr_active delta
4495 * we might have. Assumes we're called after migrate_tasks() so that the
4496 * nr_active count is stable.
4498 * Also see the comment "Global load-average calculations".
4500 static void calc_load_migrate(struct rq *rq)
4502 long delta = calc_load_fold_active(rq);
4504 atomic_long_add(delta, &calc_load_tasks);
4508 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4509 * try_to_wake_up()->select_task_rq().
4511 * Called with rq->lock held even though we'er in stop_machine() and
4512 * there's no concurrency possible, we hold the required locks anyway
4513 * because of lock validation efforts.
4515 static void migrate_tasks(unsigned int dead_cpu)
4517 struct rq *rq = cpu_rq(dead_cpu);
4518 struct task_struct *next, *stop = rq->stop;
4522 * Fudge the rq selection such that the below task selection loop
4523 * doesn't get stuck on the currently eligible stop task.
4525 * We're currently inside stop_machine() and the rq is either stuck
4526 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4527 * either way we should never end up calling schedule() until we're
4533 * put_prev_task() and pick_next_task() sched
4534 * class method both need to have an up-to-date
4535 * value of rq->clock[_task]
4537 update_rq_clock(rq);
4541 * There's this thread running, bail when that's the only
4544 if (rq->nr_running == 1)
4547 next = pick_next_task(rq);
4549 next->sched_class->put_prev_task(rq, next);
4551 /* Find suitable destination for @next, with force if needed. */
4552 dest_cpu = select_fallback_rq(dead_cpu, next);
4553 raw_spin_unlock(&rq->lock);
4555 __migrate_task(next, dead_cpu, dest_cpu);
4557 raw_spin_lock(&rq->lock);
4563 #endif /* CONFIG_HOTPLUG_CPU */
4565 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4567 static struct ctl_table sd_ctl_dir[] = {
4569 .procname = "sched_domain",
4575 static struct ctl_table sd_ctl_root[] = {
4577 .procname = "kernel",
4579 .child = sd_ctl_dir,
4584 static struct ctl_table *sd_alloc_ctl_entry(int n)
4586 struct ctl_table *entry =
4587 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4592 static void sd_free_ctl_entry(struct ctl_table **tablep)
4594 struct ctl_table *entry;
4597 * In the intermediate directories, both the child directory and
4598 * procname are dynamically allocated and could fail but the mode
4599 * will always be set. In the lowest directory the names are
4600 * static strings and all have proc handlers.
4602 for (entry = *tablep; entry->mode; entry++) {
4604 sd_free_ctl_entry(&entry->child);
4605 if (entry->proc_handler == NULL)
4606 kfree(entry->procname);
4613 static int min_load_idx = 0;
4614 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4617 set_table_entry(struct ctl_table *entry,
4618 const char *procname, void *data, int maxlen,
4619 umode_t mode, proc_handler *proc_handler,
4622 entry->procname = procname;
4624 entry->maxlen = maxlen;
4626 entry->proc_handler = proc_handler;
4629 entry->extra1 = &min_load_idx;
4630 entry->extra2 = &max_load_idx;
4634 static struct ctl_table *
4635 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4637 struct ctl_table *table = sd_alloc_ctl_entry(13);
4642 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4643 sizeof(long), 0644, proc_doulongvec_minmax, false);
4644 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4645 sizeof(long), 0644, proc_doulongvec_minmax, false);
4646 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4647 sizeof(int), 0644, proc_dointvec_minmax, true);
4648 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4649 sizeof(int), 0644, proc_dointvec_minmax, true);
4650 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4651 sizeof(int), 0644, proc_dointvec_minmax, true);
4652 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4653 sizeof(int), 0644, proc_dointvec_minmax, true);
4654 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4655 sizeof(int), 0644, proc_dointvec_minmax, true);
4656 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4657 sizeof(int), 0644, proc_dointvec_minmax, false);
4658 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4659 sizeof(int), 0644, proc_dointvec_minmax, false);
4660 set_table_entry(&table[9], "cache_nice_tries",
4661 &sd->cache_nice_tries,
4662 sizeof(int), 0644, proc_dointvec_minmax, false);
4663 set_table_entry(&table[10], "flags", &sd->flags,
4664 sizeof(int), 0644, proc_dointvec_minmax, false);
4665 set_table_entry(&table[11], "name", sd->name,
4666 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4667 /* &table[12] is terminator */
4672 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4674 struct ctl_table *entry, *table;
4675 struct sched_domain *sd;
4676 int domain_num = 0, i;
4679 for_each_domain(cpu, sd)
4681 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4686 for_each_domain(cpu, sd) {
4687 snprintf(buf, 32, "domain%d", i);
4688 entry->procname = kstrdup(buf, GFP_KERNEL);
4690 entry->child = sd_alloc_ctl_domain_table(sd);
4697 static struct ctl_table_header *sd_sysctl_header;
4698 static void register_sched_domain_sysctl(void)
4700 int i, cpu_num = num_possible_cpus();
4701 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4704 WARN_ON(sd_ctl_dir[0].child);
4705 sd_ctl_dir[0].child = entry;
4710 for_each_possible_cpu(i) {
4711 snprintf(buf, 32, "cpu%d", i);
4712 entry->procname = kstrdup(buf, GFP_KERNEL);
4714 entry->child = sd_alloc_ctl_cpu_table(i);
4718 WARN_ON(sd_sysctl_header);
4719 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4722 /* may be called multiple times per register */
4723 static void unregister_sched_domain_sysctl(void)
4725 if (sd_sysctl_header)
4726 unregister_sysctl_table(sd_sysctl_header);
4727 sd_sysctl_header = NULL;
4728 if (sd_ctl_dir[0].child)
4729 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4732 static void register_sched_domain_sysctl(void)
4735 static void unregister_sched_domain_sysctl(void)
4740 static void set_rq_online(struct rq *rq)
4743 const struct sched_class *class;
4745 cpumask_set_cpu(rq->cpu, rq->rd->online);
4748 for_each_class(class) {
4749 if (class->rq_online)
4750 class->rq_online(rq);
4755 static void set_rq_offline(struct rq *rq)
4758 const struct sched_class *class;
4760 for_each_class(class) {
4761 if (class->rq_offline)
4762 class->rq_offline(rq);
4765 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4771 * migration_call - callback that gets triggered when a CPU is added.
4772 * Here we can start up the necessary migration thread for the new CPU.
4775 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4777 int cpu = (long)hcpu;
4778 unsigned long flags;
4779 struct rq *rq = cpu_rq(cpu);
4781 switch (action & ~CPU_TASKS_FROZEN) {
4783 case CPU_UP_PREPARE:
4784 rq->calc_load_update = calc_load_update;
4788 /* Update our root-domain */
4789 raw_spin_lock_irqsave(&rq->lock, flags);
4791 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4795 raw_spin_unlock_irqrestore(&rq->lock, flags);
4798 #ifdef CONFIG_HOTPLUG_CPU
4800 sched_ttwu_pending();
4801 /* Update our root-domain */
4802 raw_spin_lock_irqsave(&rq->lock, flags);
4804 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4808 BUG_ON(rq->nr_running != 1); /* the migration thread */
4809 raw_spin_unlock_irqrestore(&rq->lock, flags);
4813 calc_load_migrate(rq);
4818 update_max_interval();
4824 * Register at high priority so that task migration (migrate_all_tasks)
4825 * happens before everything else. This has to be lower priority than
4826 * the notifier in the perf_event subsystem, though.
4828 static struct notifier_block migration_notifier = {
4829 .notifier_call = migration_call,
4830 .priority = CPU_PRI_MIGRATION,
4833 static int sched_cpu_active(struct notifier_block *nfb,
4834 unsigned long action, void *hcpu)
4836 switch (action & ~CPU_TASKS_FROZEN) {
4838 case CPU_DOWN_FAILED:
4839 set_cpu_active((long)hcpu, true);
4846 static int sched_cpu_inactive(struct notifier_block *nfb,
4847 unsigned long action, void *hcpu)
4849 switch (action & ~CPU_TASKS_FROZEN) {
4850 case CPU_DOWN_PREPARE:
4851 set_cpu_active((long)hcpu, false);
4858 static int __init migration_init(void)
4860 void *cpu = (void *)(long)smp_processor_id();
4863 /* Initialize migration for the boot CPU */
4864 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4865 BUG_ON(err == NOTIFY_BAD);
4866 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4867 register_cpu_notifier(&migration_notifier);
4869 /* Register cpu active notifiers */
4870 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4871 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4875 early_initcall(migration_init);
4880 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4882 #ifdef CONFIG_SCHED_DEBUG
4884 static __read_mostly int sched_debug_enabled;
4886 static int __init sched_debug_setup(char *str)
4888 sched_debug_enabled = 1;
4892 early_param("sched_debug", sched_debug_setup);
4894 static inline bool sched_debug(void)
4896 return sched_debug_enabled;
4899 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4900 struct cpumask *groupmask)
4902 struct sched_group *group = sd->groups;
4905 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4906 cpumask_clear(groupmask);
4908 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4910 if (!(sd->flags & SD_LOAD_BALANCE)) {
4911 printk("does not load-balance\n");
4913 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4918 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4920 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4921 printk(KERN_ERR "ERROR: domain->span does not contain "
4924 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4925 printk(KERN_ERR "ERROR: domain->groups does not contain"
4929 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4933 printk(KERN_ERR "ERROR: group is NULL\n");
4938 * Even though we initialize ->power to something semi-sane,
4939 * we leave power_orig unset. This allows us to detect if
4940 * domain iteration is still funny without causing /0 traps.
4942 if (!group->sgp->power_orig) {
4943 printk(KERN_CONT "\n");
4944 printk(KERN_ERR "ERROR: domain->cpu_power not "
4949 if (!cpumask_weight(sched_group_cpus(group))) {
4950 printk(KERN_CONT "\n");
4951 printk(KERN_ERR "ERROR: empty group\n");
4955 if (!(sd->flags & SD_OVERLAP) &&
4956 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4957 printk(KERN_CONT "\n");
4958 printk(KERN_ERR "ERROR: repeated CPUs\n");
4962 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4964 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4966 printk(KERN_CONT " %s", str);
4967 if (group->sgp->power != SCHED_POWER_SCALE) {
4968 printk(KERN_CONT " (cpu_power = %d)",
4972 group = group->next;
4973 } while (group != sd->groups);
4974 printk(KERN_CONT "\n");
4976 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4977 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4980 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4981 printk(KERN_ERR "ERROR: parent span is not a superset "
4982 "of domain->span\n");
4986 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4990 if (!sched_debug_enabled)
4994 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4998 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5001 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5009 #else /* !CONFIG_SCHED_DEBUG */
5010 # define sched_domain_debug(sd, cpu) do { } while (0)
5011 static inline bool sched_debug(void)
5015 #endif /* CONFIG_SCHED_DEBUG */
5017 static int sd_degenerate(struct sched_domain *sd)
5019 if (cpumask_weight(sched_domain_span(sd)) == 1)
5022 /* Following flags need at least 2 groups */
5023 if (sd->flags & (SD_LOAD_BALANCE |
5024 SD_BALANCE_NEWIDLE |
5028 SD_SHARE_PKG_RESOURCES)) {
5029 if (sd->groups != sd->groups->next)
5033 /* Following flags don't use groups */
5034 if (sd->flags & (SD_WAKE_AFFINE))
5041 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5043 unsigned long cflags = sd->flags, pflags = parent->flags;
5045 if (sd_degenerate(parent))
5048 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5051 /* Flags needing groups don't count if only 1 group in parent */
5052 if (parent->groups == parent->groups->next) {
5053 pflags &= ~(SD_LOAD_BALANCE |
5054 SD_BALANCE_NEWIDLE |
5058 SD_SHARE_PKG_RESOURCES |
5060 if (nr_node_ids == 1)
5061 pflags &= ~SD_SERIALIZE;
5063 if (~cflags & pflags)
5069 static void free_rootdomain(struct rcu_head *rcu)
5071 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5073 cpupri_cleanup(&rd->cpupri);
5074 free_cpumask_var(rd->dlo_mask);
5075 free_cpumask_var(rd->rto_mask);
5076 free_cpumask_var(rd->online);
5077 free_cpumask_var(rd->span);
5081 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5083 struct root_domain *old_rd = NULL;
5084 unsigned long flags;
5086 raw_spin_lock_irqsave(&rq->lock, flags);
5091 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5094 cpumask_clear_cpu(rq->cpu, old_rd->span);
5097 * If we dont want to free the old_rd yet then
5098 * set old_rd to NULL to skip the freeing later
5101 if (!atomic_dec_and_test(&old_rd->refcount))
5105 atomic_inc(&rd->refcount);
5108 cpumask_set_cpu(rq->cpu, rd->span);
5109 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5112 raw_spin_unlock_irqrestore(&rq->lock, flags);
5115 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5118 static int init_rootdomain(struct root_domain *rd)
5120 memset(rd, 0, sizeof(*rd));
5122 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5124 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5126 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5128 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5131 if (cpupri_init(&rd->cpupri) != 0)
5136 free_cpumask_var(rd->rto_mask);
5138 free_cpumask_var(rd->dlo_mask);
5140 free_cpumask_var(rd->online);
5142 free_cpumask_var(rd->span);
5148 * By default the system creates a single root-domain with all cpus as
5149 * members (mimicking the global state we have today).
5151 struct root_domain def_root_domain;
5153 static void init_defrootdomain(void)
5155 init_rootdomain(&def_root_domain);
5157 atomic_set(&def_root_domain.refcount, 1);
5160 static struct root_domain *alloc_rootdomain(void)
5162 struct root_domain *rd;
5164 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5168 if (init_rootdomain(rd) != 0) {
5176 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5178 struct sched_group *tmp, *first;
5187 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5192 } while (sg != first);
5195 static void free_sched_domain(struct rcu_head *rcu)
5197 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5200 * If its an overlapping domain it has private groups, iterate and
5203 if (sd->flags & SD_OVERLAP) {
5204 free_sched_groups(sd->groups, 1);
5205 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5206 kfree(sd->groups->sgp);
5212 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5214 call_rcu(&sd->rcu, free_sched_domain);
5217 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5219 for (; sd; sd = sd->parent)
5220 destroy_sched_domain(sd, cpu);
5224 * Keep a special pointer to the highest sched_domain that has
5225 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5226 * allows us to avoid some pointer chasing select_idle_sibling().
5228 * Also keep a unique ID per domain (we use the first cpu number in
5229 * the cpumask of the domain), this allows us to quickly tell if
5230 * two cpus are in the same cache domain, see cpus_share_cache().
5232 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5233 DEFINE_PER_CPU(int, sd_llc_size);
5234 DEFINE_PER_CPU(int, sd_llc_id);
5235 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5236 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5237 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5239 static void update_top_cache_domain(int cpu)
5241 struct sched_domain *sd;
5242 struct sched_domain *busy_sd = NULL;
5246 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5248 id = cpumask_first(sched_domain_span(sd));
5249 size = cpumask_weight(sched_domain_span(sd));
5250 busy_sd = sd->parent; /* sd_busy */
5252 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5254 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5255 per_cpu(sd_llc_size, cpu) = size;
5256 per_cpu(sd_llc_id, cpu) = id;
5258 sd = lowest_flag_domain(cpu, SD_NUMA);
5259 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5261 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5262 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5266 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5267 * hold the hotplug lock.
5270 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5272 struct rq *rq = cpu_rq(cpu);
5273 struct sched_domain *tmp;
5275 /* Remove the sched domains which do not contribute to scheduling. */
5276 for (tmp = sd; tmp; ) {
5277 struct sched_domain *parent = tmp->parent;
5281 if (sd_parent_degenerate(tmp, parent)) {
5282 tmp->parent = parent->parent;
5284 parent->parent->child = tmp;
5286 * Transfer SD_PREFER_SIBLING down in case of a
5287 * degenerate parent; the spans match for this
5288 * so the property transfers.
5290 if (parent->flags & SD_PREFER_SIBLING)
5291 tmp->flags |= SD_PREFER_SIBLING;
5292 destroy_sched_domain(parent, cpu);
5297 if (sd && sd_degenerate(sd)) {
5300 destroy_sched_domain(tmp, cpu);
5305 sched_domain_debug(sd, cpu);
5307 rq_attach_root(rq, rd);
5309 rcu_assign_pointer(rq->sd, sd);
5310 destroy_sched_domains(tmp, cpu);
5312 update_top_cache_domain(cpu);
5315 /* cpus with isolated domains */
5316 static cpumask_var_t cpu_isolated_map;
5318 /* Setup the mask of cpus configured for isolated domains */
5319 static int __init isolated_cpu_setup(char *str)
5321 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5322 cpulist_parse(str, cpu_isolated_map);
5326 __setup("isolcpus=", isolated_cpu_setup);
5328 static const struct cpumask *cpu_cpu_mask(int cpu)
5330 return cpumask_of_node(cpu_to_node(cpu));
5334 struct sched_domain **__percpu sd;
5335 struct sched_group **__percpu sg;
5336 struct sched_group_power **__percpu sgp;
5340 struct sched_domain ** __percpu sd;
5341 struct root_domain *rd;
5351 struct sched_domain_topology_level;
5353 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5354 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5356 #define SDTL_OVERLAP 0x01
5358 struct sched_domain_topology_level {
5359 sched_domain_init_f init;
5360 sched_domain_mask_f mask;
5363 struct sd_data data;
5367 * Build an iteration mask that can exclude certain CPUs from the upwards
5370 * Asymmetric node setups can result in situations where the domain tree is of
5371 * unequal depth, make sure to skip domains that already cover the entire
5374 * In that case build_sched_domains() will have terminated the iteration early
5375 * and our sibling sd spans will be empty. Domains should always include the
5376 * cpu they're built on, so check that.
5379 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5381 const struct cpumask *span = sched_domain_span(sd);
5382 struct sd_data *sdd = sd->private;
5383 struct sched_domain *sibling;
5386 for_each_cpu(i, span) {
5387 sibling = *per_cpu_ptr(sdd->sd, i);
5388 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5391 cpumask_set_cpu(i, sched_group_mask(sg));
5396 * Return the canonical balance cpu for this group, this is the first cpu
5397 * of this group that's also in the iteration mask.
5399 int group_balance_cpu(struct sched_group *sg)
5401 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5405 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5407 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5408 const struct cpumask *span = sched_domain_span(sd);
5409 struct cpumask *covered = sched_domains_tmpmask;
5410 struct sd_data *sdd = sd->private;
5411 struct sched_domain *child;
5414 cpumask_clear(covered);
5416 for_each_cpu(i, span) {
5417 struct cpumask *sg_span;
5419 if (cpumask_test_cpu(i, covered))
5422 child = *per_cpu_ptr(sdd->sd, i);
5424 /* See the comment near build_group_mask(). */
5425 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5428 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5429 GFP_KERNEL, cpu_to_node(cpu));
5434 sg_span = sched_group_cpus(sg);
5436 child = child->child;
5437 cpumask_copy(sg_span, sched_domain_span(child));
5439 cpumask_set_cpu(i, sg_span);
5441 cpumask_or(covered, covered, sg_span);
5443 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5444 if (atomic_inc_return(&sg->sgp->ref) == 1)
5445 build_group_mask(sd, sg);
5448 * Initialize sgp->power such that even if we mess up the
5449 * domains and no possible iteration will get us here, we won't
5452 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5453 sg->sgp->power_orig = sg->sgp->power;
5456 * Make sure the first group of this domain contains the
5457 * canonical balance cpu. Otherwise the sched_domain iteration
5458 * breaks. See update_sg_lb_stats().
5460 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5461 group_balance_cpu(sg) == cpu)
5471 sd->groups = groups;
5476 free_sched_groups(first, 0);
5481 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5483 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5484 struct sched_domain *child = sd->child;
5487 cpu = cpumask_first(sched_domain_span(child));
5490 *sg = *per_cpu_ptr(sdd->sg, cpu);
5491 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5492 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5499 * build_sched_groups will build a circular linked list of the groups
5500 * covered by the given span, and will set each group's ->cpumask correctly,
5501 * and ->cpu_power to 0.
5503 * Assumes the sched_domain tree is fully constructed
5506 build_sched_groups(struct sched_domain *sd, int cpu)
5508 struct sched_group *first = NULL, *last = NULL;
5509 struct sd_data *sdd = sd->private;
5510 const struct cpumask *span = sched_domain_span(sd);
5511 struct cpumask *covered;
5514 get_group(cpu, sdd, &sd->groups);
5515 atomic_inc(&sd->groups->ref);
5517 if (cpu != cpumask_first(span))
5520 lockdep_assert_held(&sched_domains_mutex);
5521 covered = sched_domains_tmpmask;
5523 cpumask_clear(covered);
5525 for_each_cpu(i, span) {
5526 struct sched_group *sg;
5529 if (cpumask_test_cpu(i, covered))
5532 group = get_group(i, sdd, &sg);
5533 cpumask_clear(sched_group_cpus(sg));
5535 cpumask_setall(sched_group_mask(sg));
5537 for_each_cpu(j, span) {
5538 if (get_group(j, sdd, NULL) != group)
5541 cpumask_set_cpu(j, covered);
5542 cpumask_set_cpu(j, sched_group_cpus(sg));
5557 * Initialize sched groups cpu_power.
5559 * cpu_power indicates the capacity of sched group, which is used while
5560 * distributing the load between different sched groups in a sched domain.
5561 * Typically cpu_power for all the groups in a sched domain will be same unless
5562 * there are asymmetries in the topology. If there are asymmetries, group
5563 * having more cpu_power will pickup more load compared to the group having
5566 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5568 struct sched_group *sg = sd->groups;
5573 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5575 } while (sg != sd->groups);
5577 if (cpu != group_balance_cpu(sg))
5580 update_group_power(sd, cpu);
5581 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5584 int __weak arch_sd_sibling_asym_packing(void)
5586 return 0*SD_ASYM_PACKING;
5590 * Initializers for schedule domains
5591 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5594 #ifdef CONFIG_SCHED_DEBUG
5595 # define SD_INIT_NAME(sd, type) sd->name = #type
5597 # define SD_INIT_NAME(sd, type) do { } while (0)
5600 #define SD_INIT_FUNC(type) \
5601 static noinline struct sched_domain * \
5602 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5604 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5605 *sd = SD_##type##_INIT; \
5606 SD_INIT_NAME(sd, type); \
5607 sd->private = &tl->data; \
5612 #ifdef CONFIG_SCHED_SMT
5613 SD_INIT_FUNC(SIBLING)
5615 #ifdef CONFIG_SCHED_MC
5618 #ifdef CONFIG_SCHED_BOOK
5622 static int default_relax_domain_level = -1;
5623 int sched_domain_level_max;
5625 static int __init setup_relax_domain_level(char *str)
5627 if (kstrtoint(str, 0, &default_relax_domain_level))
5628 pr_warn("Unable to set relax_domain_level\n");
5632 __setup("relax_domain_level=", setup_relax_domain_level);
5634 static void set_domain_attribute(struct sched_domain *sd,
5635 struct sched_domain_attr *attr)
5639 if (!attr || attr->relax_domain_level < 0) {
5640 if (default_relax_domain_level < 0)
5643 request = default_relax_domain_level;
5645 request = attr->relax_domain_level;
5646 if (request < sd->level) {
5647 /* turn off idle balance on this domain */
5648 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5650 /* turn on idle balance on this domain */
5651 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5655 static void __sdt_free(const struct cpumask *cpu_map);
5656 static int __sdt_alloc(const struct cpumask *cpu_map);
5658 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5659 const struct cpumask *cpu_map)
5663 if (!atomic_read(&d->rd->refcount))
5664 free_rootdomain(&d->rd->rcu); /* fall through */
5666 free_percpu(d->sd); /* fall through */
5668 __sdt_free(cpu_map); /* fall through */
5674 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5675 const struct cpumask *cpu_map)
5677 memset(d, 0, sizeof(*d));
5679 if (__sdt_alloc(cpu_map))
5680 return sa_sd_storage;
5681 d->sd = alloc_percpu(struct sched_domain *);
5683 return sa_sd_storage;
5684 d->rd = alloc_rootdomain();
5687 return sa_rootdomain;
5691 * NULL the sd_data elements we've used to build the sched_domain and
5692 * sched_group structure so that the subsequent __free_domain_allocs()
5693 * will not free the data we're using.
5695 static void claim_allocations(int cpu, struct sched_domain *sd)
5697 struct sd_data *sdd = sd->private;
5699 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5700 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5702 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5703 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5705 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5706 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5709 #ifdef CONFIG_SCHED_SMT
5710 static const struct cpumask *cpu_smt_mask(int cpu)
5712 return topology_thread_cpumask(cpu);
5717 * Topology list, bottom-up.
5719 static struct sched_domain_topology_level default_topology[] = {
5720 #ifdef CONFIG_SCHED_SMT
5721 { sd_init_SIBLING, cpu_smt_mask, },
5723 #ifdef CONFIG_SCHED_MC
5724 { sd_init_MC, cpu_coregroup_mask, },
5726 #ifdef CONFIG_SCHED_BOOK
5727 { sd_init_BOOK, cpu_book_mask, },
5729 { sd_init_CPU, cpu_cpu_mask, },
5733 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5735 #define for_each_sd_topology(tl) \
5736 for (tl = sched_domain_topology; tl->init; tl++)
5740 static int sched_domains_numa_levels;
5741 static int *sched_domains_numa_distance;
5742 static struct cpumask ***sched_domains_numa_masks;
5743 static int sched_domains_curr_level;
5745 static inline int sd_local_flags(int level)
5747 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5750 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5753 static struct sched_domain *
5754 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5756 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5757 int level = tl->numa_level;
5758 int sd_weight = cpumask_weight(
5759 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5761 *sd = (struct sched_domain){
5762 .min_interval = sd_weight,
5763 .max_interval = 2*sd_weight,
5765 .imbalance_pct = 125,
5766 .cache_nice_tries = 2,
5773 .flags = 1*SD_LOAD_BALANCE
5774 | 1*SD_BALANCE_NEWIDLE
5779 | 0*SD_SHARE_CPUPOWER
5780 | 0*SD_SHARE_PKG_RESOURCES
5782 | 0*SD_PREFER_SIBLING
5784 | sd_local_flags(level)
5786 .last_balance = jiffies,
5787 .balance_interval = sd_weight,
5789 SD_INIT_NAME(sd, NUMA);
5790 sd->private = &tl->data;
5793 * Ugly hack to pass state to sd_numa_mask()...
5795 sched_domains_curr_level = tl->numa_level;
5800 static const struct cpumask *sd_numa_mask(int cpu)
5802 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5805 static void sched_numa_warn(const char *str)
5807 static int done = false;
5815 printk(KERN_WARNING "ERROR: %s\n\n", str);
5817 for (i = 0; i < nr_node_ids; i++) {
5818 printk(KERN_WARNING " ");
5819 for (j = 0; j < nr_node_ids; j++)
5820 printk(KERN_CONT "%02d ", node_distance(i,j));
5821 printk(KERN_CONT "\n");
5823 printk(KERN_WARNING "\n");
5826 static bool find_numa_distance(int distance)
5830 if (distance == node_distance(0, 0))
5833 for (i = 0; i < sched_domains_numa_levels; i++) {
5834 if (sched_domains_numa_distance[i] == distance)
5841 static void sched_init_numa(void)
5843 int next_distance, curr_distance = node_distance(0, 0);
5844 struct sched_domain_topology_level *tl;
5848 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5849 if (!sched_domains_numa_distance)
5853 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5854 * unique distances in the node_distance() table.
5856 * Assumes node_distance(0,j) includes all distances in
5857 * node_distance(i,j) in order to avoid cubic time.
5859 next_distance = curr_distance;
5860 for (i = 0; i < nr_node_ids; i++) {
5861 for (j = 0; j < nr_node_ids; j++) {
5862 for (k = 0; k < nr_node_ids; k++) {
5863 int distance = node_distance(i, k);
5865 if (distance > curr_distance &&
5866 (distance < next_distance ||
5867 next_distance == curr_distance))
5868 next_distance = distance;
5871 * While not a strong assumption it would be nice to know
5872 * about cases where if node A is connected to B, B is not
5873 * equally connected to A.
5875 if (sched_debug() && node_distance(k, i) != distance)
5876 sched_numa_warn("Node-distance not symmetric");
5878 if (sched_debug() && i && !find_numa_distance(distance))
5879 sched_numa_warn("Node-0 not representative");
5881 if (next_distance != curr_distance) {
5882 sched_domains_numa_distance[level++] = next_distance;
5883 sched_domains_numa_levels = level;
5884 curr_distance = next_distance;
5889 * In case of sched_debug() we verify the above assumption.
5895 * 'level' contains the number of unique distances, excluding the
5896 * identity distance node_distance(i,i).
5898 * The sched_domains_numa_distance[] array includes the actual distance
5903 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5904 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5905 * the array will contain less then 'level' members. This could be
5906 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5907 * in other functions.
5909 * We reset it to 'level' at the end of this function.
5911 sched_domains_numa_levels = 0;
5913 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5914 if (!sched_domains_numa_masks)
5918 * Now for each level, construct a mask per node which contains all
5919 * cpus of nodes that are that many hops away from us.
5921 for (i = 0; i < level; i++) {
5922 sched_domains_numa_masks[i] =
5923 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5924 if (!sched_domains_numa_masks[i])
5927 for (j = 0; j < nr_node_ids; j++) {
5928 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5932 sched_domains_numa_masks[i][j] = mask;
5934 for (k = 0; k < nr_node_ids; k++) {
5935 if (node_distance(j, k) > sched_domains_numa_distance[i])
5938 cpumask_or(mask, mask, cpumask_of_node(k));
5943 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5944 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5949 * Copy the default topology bits..
5951 for (i = 0; default_topology[i].init; i++)
5952 tl[i] = default_topology[i];
5955 * .. and append 'j' levels of NUMA goodness.
5957 for (j = 0; j < level; i++, j++) {
5958 tl[i] = (struct sched_domain_topology_level){
5959 .init = sd_numa_init,
5960 .mask = sd_numa_mask,
5961 .flags = SDTL_OVERLAP,
5966 sched_domain_topology = tl;
5968 sched_domains_numa_levels = level;
5971 static void sched_domains_numa_masks_set(int cpu)
5974 int node = cpu_to_node(cpu);
5976 for (i = 0; i < sched_domains_numa_levels; i++) {
5977 for (j = 0; j < nr_node_ids; j++) {
5978 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5979 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5984 static void sched_domains_numa_masks_clear(int cpu)
5987 for (i = 0; i < sched_domains_numa_levels; i++) {
5988 for (j = 0; j < nr_node_ids; j++)
5989 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5994 * Update sched_domains_numa_masks[level][node] array when new cpus
5997 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5998 unsigned long action,
6001 int cpu = (long)hcpu;
6003 switch (action & ~CPU_TASKS_FROZEN) {
6005 sched_domains_numa_masks_set(cpu);
6009 sched_domains_numa_masks_clear(cpu);
6019 static inline void sched_init_numa(void)
6023 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6024 unsigned long action,
6029 #endif /* CONFIG_NUMA */
6031 static int __sdt_alloc(const struct cpumask *cpu_map)
6033 struct sched_domain_topology_level *tl;
6036 for_each_sd_topology(tl) {
6037 struct sd_data *sdd = &tl->data;
6039 sdd->sd = alloc_percpu(struct sched_domain *);
6043 sdd->sg = alloc_percpu(struct sched_group *);
6047 sdd->sgp = alloc_percpu(struct sched_group_power *);
6051 for_each_cpu(j, cpu_map) {
6052 struct sched_domain *sd;
6053 struct sched_group *sg;
6054 struct sched_group_power *sgp;
6056 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6057 GFP_KERNEL, cpu_to_node(j));
6061 *per_cpu_ptr(sdd->sd, j) = sd;
6063 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6064 GFP_KERNEL, cpu_to_node(j));
6070 *per_cpu_ptr(sdd->sg, j) = sg;
6072 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6073 GFP_KERNEL, cpu_to_node(j));
6077 *per_cpu_ptr(sdd->sgp, j) = sgp;
6084 static void __sdt_free(const struct cpumask *cpu_map)
6086 struct sched_domain_topology_level *tl;
6089 for_each_sd_topology(tl) {
6090 struct sd_data *sdd = &tl->data;
6092 for_each_cpu(j, cpu_map) {
6093 struct sched_domain *sd;
6096 sd = *per_cpu_ptr(sdd->sd, j);
6097 if (sd && (sd->flags & SD_OVERLAP))
6098 free_sched_groups(sd->groups, 0);
6099 kfree(*per_cpu_ptr(sdd->sd, j));
6103 kfree(*per_cpu_ptr(sdd->sg, j));
6105 kfree(*per_cpu_ptr(sdd->sgp, j));
6107 free_percpu(sdd->sd);
6109 free_percpu(sdd->sg);
6111 free_percpu(sdd->sgp);
6116 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6117 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6118 struct sched_domain *child, int cpu)
6120 struct sched_domain *sd = tl->init(tl, cpu);
6124 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6126 sd->level = child->level + 1;
6127 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6131 set_domain_attribute(sd, attr);
6137 * Build sched domains for a given set of cpus and attach the sched domains
6138 * to the individual cpus
6140 static int build_sched_domains(const struct cpumask *cpu_map,
6141 struct sched_domain_attr *attr)
6143 enum s_alloc alloc_state;
6144 struct sched_domain *sd;
6146 int i, ret = -ENOMEM;
6148 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6149 if (alloc_state != sa_rootdomain)
6152 /* Set up domains for cpus specified by the cpu_map. */
6153 for_each_cpu(i, cpu_map) {
6154 struct sched_domain_topology_level *tl;
6157 for_each_sd_topology(tl) {
6158 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6159 if (tl == sched_domain_topology)
6160 *per_cpu_ptr(d.sd, i) = sd;
6161 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6162 sd->flags |= SD_OVERLAP;
6163 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6168 /* Build the groups for the domains */
6169 for_each_cpu(i, cpu_map) {
6170 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6171 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6172 if (sd->flags & SD_OVERLAP) {
6173 if (build_overlap_sched_groups(sd, i))
6176 if (build_sched_groups(sd, i))
6182 /* Calculate CPU power for physical packages and nodes */
6183 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6184 if (!cpumask_test_cpu(i, cpu_map))
6187 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6188 claim_allocations(i, sd);
6189 init_sched_groups_power(i, sd);
6193 /* Attach the domains */
6195 for_each_cpu(i, cpu_map) {
6196 sd = *per_cpu_ptr(d.sd, i);
6197 cpu_attach_domain(sd, d.rd, i);
6203 __free_domain_allocs(&d, alloc_state, cpu_map);
6207 static cpumask_var_t *doms_cur; /* current sched domains */
6208 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6209 static struct sched_domain_attr *dattr_cur;
6210 /* attribues of custom domains in 'doms_cur' */
6213 * Special case: If a kmalloc of a doms_cur partition (array of
6214 * cpumask) fails, then fallback to a single sched domain,
6215 * as determined by the single cpumask fallback_doms.
6217 static cpumask_var_t fallback_doms;
6220 * arch_update_cpu_topology lets virtualized architectures update the
6221 * cpu core maps. It is supposed to return 1 if the topology changed
6222 * or 0 if it stayed the same.
6224 int __attribute__((weak)) arch_update_cpu_topology(void)
6229 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6232 cpumask_var_t *doms;
6234 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6237 for (i = 0; i < ndoms; i++) {
6238 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6239 free_sched_domains(doms, i);
6246 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6249 for (i = 0; i < ndoms; i++)
6250 free_cpumask_var(doms[i]);
6255 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6256 * For now this just excludes isolated cpus, but could be used to
6257 * exclude other special cases in the future.
6259 static int init_sched_domains(const struct cpumask *cpu_map)
6263 arch_update_cpu_topology();
6265 doms_cur = alloc_sched_domains(ndoms_cur);
6267 doms_cur = &fallback_doms;
6268 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6269 err = build_sched_domains(doms_cur[0], NULL);
6270 register_sched_domain_sysctl();
6276 * Detach sched domains from a group of cpus specified in cpu_map
6277 * These cpus will now be attached to the NULL domain
6279 static void detach_destroy_domains(const struct cpumask *cpu_map)
6284 for_each_cpu(i, cpu_map)
6285 cpu_attach_domain(NULL, &def_root_domain, i);
6289 /* handle null as "default" */
6290 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6291 struct sched_domain_attr *new, int idx_new)
6293 struct sched_domain_attr tmp;
6300 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6301 new ? (new + idx_new) : &tmp,
6302 sizeof(struct sched_domain_attr));
6306 * Partition sched domains as specified by the 'ndoms_new'
6307 * cpumasks in the array doms_new[] of cpumasks. This compares
6308 * doms_new[] to the current sched domain partitioning, doms_cur[].
6309 * It destroys each deleted domain and builds each new domain.
6311 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6312 * The masks don't intersect (don't overlap.) We should setup one
6313 * sched domain for each mask. CPUs not in any of the cpumasks will
6314 * not be load balanced. If the same cpumask appears both in the
6315 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6318 * The passed in 'doms_new' should be allocated using
6319 * alloc_sched_domains. This routine takes ownership of it and will
6320 * free_sched_domains it when done with it. If the caller failed the
6321 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6322 * and partition_sched_domains() will fallback to the single partition
6323 * 'fallback_doms', it also forces the domains to be rebuilt.
6325 * If doms_new == NULL it will be replaced with cpu_online_mask.
6326 * ndoms_new == 0 is a special case for destroying existing domains,
6327 * and it will not create the default domain.
6329 * Call with hotplug lock held
6331 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6332 struct sched_domain_attr *dattr_new)
6337 mutex_lock(&sched_domains_mutex);
6339 /* always unregister in case we don't destroy any domains */
6340 unregister_sched_domain_sysctl();
6342 /* Let architecture update cpu core mappings. */
6343 new_topology = arch_update_cpu_topology();
6345 n = doms_new ? ndoms_new : 0;
6347 /* Destroy deleted domains */
6348 for (i = 0; i < ndoms_cur; i++) {
6349 for (j = 0; j < n && !new_topology; j++) {
6350 if (cpumask_equal(doms_cur[i], doms_new[j])
6351 && dattrs_equal(dattr_cur, i, dattr_new, j))
6354 /* no match - a current sched domain not in new doms_new[] */
6355 detach_destroy_domains(doms_cur[i]);
6361 if (doms_new == NULL) {
6363 doms_new = &fallback_doms;
6364 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6365 WARN_ON_ONCE(dattr_new);
6368 /* Build new domains */
6369 for (i = 0; i < ndoms_new; i++) {
6370 for (j = 0; j < n && !new_topology; j++) {
6371 if (cpumask_equal(doms_new[i], doms_cur[j])
6372 && dattrs_equal(dattr_new, i, dattr_cur, j))
6375 /* no match - add a new doms_new */
6376 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6381 /* Remember the new sched domains */
6382 if (doms_cur != &fallback_doms)
6383 free_sched_domains(doms_cur, ndoms_cur);
6384 kfree(dattr_cur); /* kfree(NULL) is safe */
6385 doms_cur = doms_new;
6386 dattr_cur = dattr_new;
6387 ndoms_cur = ndoms_new;
6389 register_sched_domain_sysctl();
6391 mutex_unlock(&sched_domains_mutex);
6394 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6397 * Update cpusets according to cpu_active mask. If cpusets are
6398 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6399 * around partition_sched_domains().
6401 * If we come here as part of a suspend/resume, don't touch cpusets because we
6402 * want to restore it back to its original state upon resume anyway.
6404 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6408 case CPU_ONLINE_FROZEN:
6409 case CPU_DOWN_FAILED_FROZEN:
6412 * num_cpus_frozen tracks how many CPUs are involved in suspend
6413 * resume sequence. As long as this is not the last online
6414 * operation in the resume sequence, just build a single sched
6415 * domain, ignoring cpusets.
6418 if (likely(num_cpus_frozen)) {
6419 partition_sched_domains(1, NULL, NULL);
6424 * This is the last CPU online operation. So fall through and
6425 * restore the original sched domains by considering the
6426 * cpuset configurations.
6430 case CPU_DOWN_FAILED:
6431 cpuset_update_active_cpus(true);
6439 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6443 case CPU_DOWN_PREPARE:
6444 cpuset_update_active_cpus(false);
6446 case CPU_DOWN_PREPARE_FROZEN:
6448 partition_sched_domains(1, NULL, NULL);
6456 void __init sched_init_smp(void)
6458 cpumask_var_t non_isolated_cpus;
6460 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6461 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6466 * There's no userspace yet to cause hotplug operations; hence all the
6467 * cpu masks are stable and all blatant races in the below code cannot
6470 mutex_lock(&sched_domains_mutex);
6471 init_sched_domains(cpu_active_mask);
6472 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6473 if (cpumask_empty(non_isolated_cpus))
6474 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6475 mutex_unlock(&sched_domains_mutex);
6477 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6478 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6479 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6483 /* Move init over to a non-isolated CPU */
6484 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6486 sched_init_granularity();
6487 free_cpumask_var(non_isolated_cpus);
6489 init_sched_rt_class();
6490 init_sched_dl_class();
6493 void __init sched_init_smp(void)
6495 sched_init_granularity();
6497 #endif /* CONFIG_SMP */
6499 const_debug unsigned int sysctl_timer_migration = 1;
6501 int in_sched_functions(unsigned long addr)
6503 return in_lock_functions(addr) ||
6504 (addr >= (unsigned long)__sched_text_start
6505 && addr < (unsigned long)__sched_text_end);
6508 #ifdef CONFIG_CGROUP_SCHED
6510 * Default task group.
6511 * Every task in system belongs to this group at bootup.
6513 struct task_group root_task_group;
6514 LIST_HEAD(task_groups);
6517 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6519 void __init sched_init(void)
6522 unsigned long alloc_size = 0, ptr;
6524 #ifdef CONFIG_FAIR_GROUP_SCHED
6525 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6527 #ifdef CONFIG_RT_GROUP_SCHED
6528 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6530 #ifdef CONFIG_CPUMASK_OFFSTACK
6531 alloc_size += num_possible_cpus() * cpumask_size();
6534 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6536 #ifdef CONFIG_FAIR_GROUP_SCHED
6537 root_task_group.se = (struct sched_entity **)ptr;
6538 ptr += nr_cpu_ids * sizeof(void **);
6540 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6541 ptr += nr_cpu_ids * sizeof(void **);
6543 #endif /* CONFIG_FAIR_GROUP_SCHED */
6544 #ifdef CONFIG_RT_GROUP_SCHED
6545 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6546 ptr += nr_cpu_ids * sizeof(void **);
6548 root_task_group.rt_rq = (struct rt_rq **)ptr;
6549 ptr += nr_cpu_ids * sizeof(void **);
6551 #endif /* CONFIG_RT_GROUP_SCHED */
6552 #ifdef CONFIG_CPUMASK_OFFSTACK
6553 for_each_possible_cpu(i) {
6554 per_cpu(load_balance_mask, i) = (void *)ptr;
6555 ptr += cpumask_size();
6557 #endif /* CONFIG_CPUMASK_OFFSTACK */
6561 init_defrootdomain();
6564 init_rt_bandwidth(&def_rt_bandwidth,
6565 global_rt_period(), global_rt_runtime());
6567 #ifdef CONFIG_RT_GROUP_SCHED
6568 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6569 global_rt_period(), global_rt_runtime());
6570 #endif /* CONFIG_RT_GROUP_SCHED */
6572 #ifdef CONFIG_CGROUP_SCHED
6573 list_add(&root_task_group.list, &task_groups);
6574 INIT_LIST_HEAD(&root_task_group.children);
6575 INIT_LIST_HEAD(&root_task_group.siblings);
6576 autogroup_init(&init_task);
6578 #endif /* CONFIG_CGROUP_SCHED */
6580 for_each_possible_cpu(i) {
6584 raw_spin_lock_init(&rq->lock);
6586 rq->calc_load_active = 0;
6587 rq->calc_load_update = jiffies + LOAD_FREQ;
6588 init_cfs_rq(&rq->cfs);
6589 init_rt_rq(&rq->rt, rq);
6590 init_dl_rq(&rq->dl, rq);
6591 #ifdef CONFIG_FAIR_GROUP_SCHED
6592 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6593 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6595 * How much cpu bandwidth does root_task_group get?
6597 * In case of task-groups formed thr' the cgroup filesystem, it
6598 * gets 100% of the cpu resources in the system. This overall
6599 * system cpu resource is divided among the tasks of
6600 * root_task_group and its child task-groups in a fair manner,
6601 * based on each entity's (task or task-group's) weight
6602 * (se->load.weight).
6604 * In other words, if root_task_group has 10 tasks of weight
6605 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6606 * then A0's share of the cpu resource is:
6608 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6610 * We achieve this by letting root_task_group's tasks sit
6611 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6613 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6614 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6615 #endif /* CONFIG_FAIR_GROUP_SCHED */
6617 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6618 #ifdef CONFIG_RT_GROUP_SCHED
6619 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6620 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6623 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6624 rq->cpu_load[j] = 0;
6626 rq->last_load_update_tick = jiffies;
6631 rq->cpu_power = SCHED_POWER_SCALE;
6632 rq->post_schedule = 0;
6633 rq->active_balance = 0;
6634 rq->next_balance = jiffies;
6639 rq->avg_idle = 2*sysctl_sched_migration_cost;
6640 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6642 INIT_LIST_HEAD(&rq->cfs_tasks);
6644 rq_attach_root(rq, &def_root_domain);
6645 #ifdef CONFIG_NO_HZ_COMMON
6648 #ifdef CONFIG_NO_HZ_FULL
6649 rq->last_sched_tick = 0;
6653 atomic_set(&rq->nr_iowait, 0);
6656 set_load_weight(&init_task);
6658 #ifdef CONFIG_PREEMPT_NOTIFIERS
6659 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6663 * The boot idle thread does lazy MMU switching as well:
6665 atomic_inc(&init_mm.mm_count);
6666 enter_lazy_tlb(&init_mm, current);
6669 * Make us the idle thread. Technically, schedule() should not be
6670 * called from this thread, however somewhere below it might be,
6671 * but because we are the idle thread, we just pick up running again
6672 * when this runqueue becomes "idle".
6674 init_idle(current, smp_processor_id());
6676 calc_load_update = jiffies + LOAD_FREQ;
6679 * During early bootup we pretend to be a normal task:
6681 current->sched_class = &fair_sched_class;
6684 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6685 /* May be allocated at isolcpus cmdline parse time */
6686 if (cpu_isolated_map == NULL)
6687 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6688 idle_thread_set_boot_cpu();
6690 init_sched_fair_class();
6692 scheduler_running = 1;
6695 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6696 static inline int preempt_count_equals(int preempt_offset)
6698 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6700 return (nested == preempt_offset);
6703 void __might_sleep(const char *file, int line, int preempt_offset)
6705 static unsigned long prev_jiffy; /* ratelimiting */
6707 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6708 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6709 system_state != SYSTEM_RUNNING || oops_in_progress)
6711 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6713 prev_jiffy = jiffies;
6716 "BUG: sleeping function called from invalid context at %s:%d\n",
6719 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6720 in_atomic(), irqs_disabled(),
6721 current->pid, current->comm);
6723 debug_show_held_locks(current);
6724 if (irqs_disabled())
6725 print_irqtrace_events(current);
6728 EXPORT_SYMBOL(__might_sleep);
6731 #ifdef CONFIG_MAGIC_SYSRQ
6732 static void normalize_task(struct rq *rq, struct task_struct *p)
6734 const struct sched_class *prev_class = p->sched_class;
6735 struct sched_attr attr = {
6736 .sched_policy = SCHED_NORMAL,
6738 int old_prio = p->prio;
6743 dequeue_task(rq, p, 0);
6744 __setscheduler(rq, p, &attr);
6746 enqueue_task(rq, p, 0);
6747 resched_task(rq->curr);
6750 check_class_changed(rq, p, prev_class, old_prio);
6753 void normalize_rt_tasks(void)
6755 struct task_struct *g, *p;
6756 unsigned long flags;
6759 read_lock_irqsave(&tasklist_lock, flags);
6760 do_each_thread(g, p) {
6762 * Only normalize user tasks:
6767 p->se.exec_start = 0;
6768 #ifdef CONFIG_SCHEDSTATS
6769 p->se.statistics.wait_start = 0;
6770 p->se.statistics.sleep_start = 0;
6771 p->se.statistics.block_start = 0;
6774 if (!dl_task(p) && !rt_task(p)) {
6776 * Renice negative nice level userspace
6779 if (TASK_NICE(p) < 0 && p->mm)
6780 set_user_nice(p, 0);
6784 raw_spin_lock(&p->pi_lock);
6785 rq = __task_rq_lock(p);
6787 normalize_task(rq, p);
6789 __task_rq_unlock(rq);
6790 raw_spin_unlock(&p->pi_lock);
6791 } while_each_thread(g, p);
6793 read_unlock_irqrestore(&tasklist_lock, flags);
6796 #endif /* CONFIG_MAGIC_SYSRQ */
6798 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6800 * These functions are only useful for the IA64 MCA handling, or kdb.
6802 * They can only be called when the whole system has been
6803 * stopped - every CPU needs to be quiescent, and no scheduling
6804 * activity can take place. Using them for anything else would
6805 * be a serious bug, and as a result, they aren't even visible
6806 * under any other configuration.
6810 * curr_task - return the current task for a given cpu.
6811 * @cpu: the processor in question.
6813 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6815 * Return: The current task for @cpu.
6817 struct task_struct *curr_task(int cpu)
6819 return cpu_curr(cpu);
6822 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6826 * set_curr_task - set the current task for a given cpu.
6827 * @cpu: the processor in question.
6828 * @p: the task pointer to set.
6830 * Description: This function must only be used when non-maskable interrupts
6831 * are serviced on a separate stack. It allows the architecture to switch the
6832 * notion of the current task on a cpu in a non-blocking manner. This function
6833 * must be called with all CPU's synchronized, and interrupts disabled, the
6834 * and caller must save the original value of the current task (see
6835 * curr_task() above) and restore that value before reenabling interrupts and
6836 * re-starting the system.
6838 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6840 void set_curr_task(int cpu, struct task_struct *p)
6847 #ifdef CONFIG_CGROUP_SCHED
6848 /* task_group_lock serializes the addition/removal of task groups */
6849 static DEFINE_SPINLOCK(task_group_lock);
6851 static void free_sched_group(struct task_group *tg)
6853 free_fair_sched_group(tg);
6854 free_rt_sched_group(tg);
6859 /* allocate runqueue etc for a new task group */
6860 struct task_group *sched_create_group(struct task_group *parent)
6862 struct task_group *tg;
6864 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6866 return ERR_PTR(-ENOMEM);
6868 if (!alloc_fair_sched_group(tg, parent))
6871 if (!alloc_rt_sched_group(tg, parent))
6877 free_sched_group(tg);
6878 return ERR_PTR(-ENOMEM);
6881 void sched_online_group(struct task_group *tg, struct task_group *parent)
6883 unsigned long flags;
6885 spin_lock_irqsave(&task_group_lock, flags);
6886 list_add_rcu(&tg->list, &task_groups);
6888 WARN_ON(!parent); /* root should already exist */
6890 tg->parent = parent;
6891 INIT_LIST_HEAD(&tg->children);
6892 list_add_rcu(&tg->siblings, &parent->children);
6893 spin_unlock_irqrestore(&task_group_lock, flags);
6896 /* rcu callback to free various structures associated with a task group */
6897 static void free_sched_group_rcu(struct rcu_head *rhp)
6899 /* now it should be safe to free those cfs_rqs */
6900 free_sched_group(container_of(rhp, struct task_group, rcu));
6903 /* Destroy runqueue etc associated with a task group */
6904 void sched_destroy_group(struct task_group *tg)
6906 /* wait for possible concurrent references to cfs_rqs complete */
6907 call_rcu(&tg->rcu, free_sched_group_rcu);
6910 void sched_offline_group(struct task_group *tg)
6912 unsigned long flags;
6915 /* end participation in shares distribution */
6916 for_each_possible_cpu(i)
6917 unregister_fair_sched_group(tg, i);
6919 spin_lock_irqsave(&task_group_lock, flags);
6920 list_del_rcu(&tg->list);
6921 list_del_rcu(&tg->siblings);
6922 spin_unlock_irqrestore(&task_group_lock, flags);
6925 /* change task's runqueue when it moves between groups.
6926 * The caller of this function should have put the task in its new group
6927 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6928 * reflect its new group.
6930 void sched_move_task(struct task_struct *tsk)
6932 struct task_group *tg;
6934 unsigned long flags;
6937 rq = task_rq_lock(tsk, &flags);
6939 running = task_current(rq, tsk);
6943 dequeue_task(rq, tsk, 0);
6944 if (unlikely(running))
6945 tsk->sched_class->put_prev_task(rq, tsk);
6947 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6948 lockdep_is_held(&tsk->sighand->siglock)),
6949 struct task_group, css);
6950 tg = autogroup_task_group(tsk, tg);
6951 tsk->sched_task_group = tg;
6953 #ifdef CONFIG_FAIR_GROUP_SCHED
6954 if (tsk->sched_class->task_move_group)
6955 tsk->sched_class->task_move_group(tsk, on_rq);
6958 set_task_rq(tsk, task_cpu(tsk));
6960 if (unlikely(running))
6961 tsk->sched_class->set_curr_task(rq);
6963 enqueue_task(rq, tsk, 0);
6965 task_rq_unlock(rq, tsk, &flags);
6967 #endif /* CONFIG_CGROUP_SCHED */
6969 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6970 static unsigned long to_ratio(u64 period, u64 runtime)
6972 if (runtime == RUNTIME_INF)
6975 return div64_u64(runtime << 20, period);
6979 #ifdef CONFIG_RT_GROUP_SCHED
6981 * Ensure that the real time constraints are schedulable.
6983 static DEFINE_MUTEX(rt_constraints_mutex);
6985 /* Must be called with tasklist_lock held */
6986 static inline int tg_has_rt_tasks(struct task_group *tg)
6988 struct task_struct *g, *p;
6990 do_each_thread(g, p) {
6991 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6993 } while_each_thread(g, p);
6998 struct rt_schedulable_data {
6999 struct task_group *tg;
7004 static int tg_rt_schedulable(struct task_group *tg, void *data)
7006 struct rt_schedulable_data *d = data;
7007 struct task_group *child;
7008 unsigned long total, sum = 0;
7009 u64 period, runtime;
7011 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7012 runtime = tg->rt_bandwidth.rt_runtime;
7015 period = d->rt_period;
7016 runtime = d->rt_runtime;
7020 * Cannot have more runtime than the period.
7022 if (runtime > period && runtime != RUNTIME_INF)
7026 * Ensure we don't starve existing RT tasks.
7028 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7031 total = to_ratio(period, runtime);
7034 * Nobody can have more than the global setting allows.
7036 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7040 * The sum of our children's runtime should not exceed our own.
7042 list_for_each_entry_rcu(child, &tg->children, siblings) {
7043 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7044 runtime = child->rt_bandwidth.rt_runtime;
7046 if (child == d->tg) {
7047 period = d->rt_period;
7048 runtime = d->rt_runtime;
7051 sum += to_ratio(period, runtime);
7060 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7064 struct rt_schedulable_data data = {
7066 .rt_period = period,
7067 .rt_runtime = runtime,
7071 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7077 static int tg_set_rt_bandwidth(struct task_group *tg,
7078 u64 rt_period, u64 rt_runtime)
7082 mutex_lock(&rt_constraints_mutex);
7083 read_lock(&tasklist_lock);
7084 err = __rt_schedulable(tg, rt_period, rt_runtime);
7088 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7089 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7090 tg->rt_bandwidth.rt_runtime = rt_runtime;
7092 for_each_possible_cpu(i) {
7093 struct rt_rq *rt_rq = tg->rt_rq[i];
7095 raw_spin_lock(&rt_rq->rt_runtime_lock);
7096 rt_rq->rt_runtime = rt_runtime;
7097 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7099 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7101 read_unlock(&tasklist_lock);
7102 mutex_unlock(&rt_constraints_mutex);
7107 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7109 u64 rt_runtime, rt_period;
7111 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7112 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7113 if (rt_runtime_us < 0)
7114 rt_runtime = RUNTIME_INF;
7116 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7119 static long sched_group_rt_runtime(struct task_group *tg)
7123 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7126 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7127 do_div(rt_runtime_us, NSEC_PER_USEC);
7128 return rt_runtime_us;
7131 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7133 u64 rt_runtime, rt_period;
7135 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7136 rt_runtime = tg->rt_bandwidth.rt_runtime;
7141 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7144 static long sched_group_rt_period(struct task_group *tg)
7148 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7149 do_div(rt_period_us, NSEC_PER_USEC);
7150 return rt_period_us;
7153 static int sched_rt_global_constraints(void)
7155 u64 runtime, period;
7158 if (sysctl_sched_rt_period <= 0)
7161 runtime = global_rt_runtime();
7162 period = global_rt_period();
7165 * Sanity check on the sysctl variables.
7167 if (runtime > period && runtime != RUNTIME_INF)
7170 mutex_lock(&rt_constraints_mutex);
7171 read_lock(&tasklist_lock);
7172 ret = __rt_schedulable(NULL, 0, 0);
7173 read_unlock(&tasklist_lock);
7174 mutex_unlock(&rt_constraints_mutex);
7179 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7181 /* Don't accept realtime tasks when there is no way for them to run */
7182 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7188 #else /* !CONFIG_RT_GROUP_SCHED */
7189 static int sched_rt_global_constraints(void)
7191 unsigned long flags;
7194 if (sysctl_sched_rt_period <= 0)
7198 * There's always some RT tasks in the root group
7199 * -- migration, kstopmachine etc..
7201 if (sysctl_sched_rt_runtime == 0)
7204 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7205 for_each_possible_cpu(i) {
7206 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7208 raw_spin_lock(&rt_rq->rt_runtime_lock);
7209 rt_rq->rt_runtime = global_rt_runtime();
7210 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7212 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7216 #endif /* CONFIG_RT_GROUP_SCHED */
7218 int sched_rr_handler(struct ctl_table *table, int write,
7219 void __user *buffer, size_t *lenp,
7223 static DEFINE_MUTEX(mutex);
7226 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7227 /* make sure that internally we keep jiffies */
7228 /* also, writing zero resets timeslice to default */
7229 if (!ret && write) {
7230 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7231 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7233 mutex_unlock(&mutex);
7237 int sched_rt_handler(struct ctl_table *table, int write,
7238 void __user *buffer, size_t *lenp,
7242 int old_period, old_runtime;
7243 static DEFINE_MUTEX(mutex);
7246 old_period = sysctl_sched_rt_period;
7247 old_runtime = sysctl_sched_rt_runtime;
7249 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7251 if (!ret && write) {
7252 ret = sched_rt_global_constraints();
7254 sysctl_sched_rt_period = old_period;
7255 sysctl_sched_rt_runtime = old_runtime;
7257 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7258 def_rt_bandwidth.rt_period =
7259 ns_to_ktime(global_rt_period());
7262 mutex_unlock(&mutex);
7267 #ifdef CONFIG_CGROUP_SCHED
7269 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7271 return css ? container_of(css, struct task_group, css) : NULL;
7274 static struct cgroup_subsys_state *
7275 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7277 struct task_group *parent = css_tg(parent_css);
7278 struct task_group *tg;
7281 /* This is early initialization for the top cgroup */
7282 return &root_task_group.css;
7285 tg = sched_create_group(parent);
7287 return ERR_PTR(-ENOMEM);
7292 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7294 struct task_group *tg = css_tg(css);
7295 struct task_group *parent = css_tg(css_parent(css));
7298 sched_online_group(tg, parent);
7302 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7304 struct task_group *tg = css_tg(css);
7306 sched_destroy_group(tg);
7309 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7311 struct task_group *tg = css_tg(css);
7313 sched_offline_group(tg);
7316 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7317 struct cgroup_taskset *tset)
7319 struct task_struct *task;
7321 cgroup_taskset_for_each(task, css, tset) {
7322 #ifdef CONFIG_RT_GROUP_SCHED
7323 if (!sched_rt_can_attach(css_tg(css), task))
7326 /* We don't support RT-tasks being in separate groups */
7327 if (task->sched_class != &fair_sched_class)
7334 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7335 struct cgroup_taskset *tset)
7337 struct task_struct *task;
7339 cgroup_taskset_for_each(task, css, tset)
7340 sched_move_task(task);
7343 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7344 struct cgroup_subsys_state *old_css,
7345 struct task_struct *task)
7348 * cgroup_exit() is called in the copy_process() failure path.
7349 * Ignore this case since the task hasn't ran yet, this avoids
7350 * trying to poke a half freed task state from generic code.
7352 if (!(task->flags & PF_EXITING))
7355 sched_move_task(task);
7358 #ifdef CONFIG_FAIR_GROUP_SCHED
7359 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7360 struct cftype *cftype, u64 shareval)
7362 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7365 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7368 struct task_group *tg = css_tg(css);
7370 return (u64) scale_load_down(tg->shares);
7373 #ifdef CONFIG_CFS_BANDWIDTH
7374 static DEFINE_MUTEX(cfs_constraints_mutex);
7376 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7377 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7379 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7381 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7383 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7384 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7386 if (tg == &root_task_group)
7390 * Ensure we have at some amount of bandwidth every period. This is
7391 * to prevent reaching a state of large arrears when throttled via
7392 * entity_tick() resulting in prolonged exit starvation.
7394 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7398 * Likewise, bound things on the otherside by preventing insane quota
7399 * periods. This also allows us to normalize in computing quota
7402 if (period > max_cfs_quota_period)
7405 mutex_lock(&cfs_constraints_mutex);
7406 ret = __cfs_schedulable(tg, period, quota);
7410 runtime_enabled = quota != RUNTIME_INF;
7411 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7413 * If we need to toggle cfs_bandwidth_used, off->on must occur
7414 * before making related changes, and on->off must occur afterwards
7416 if (runtime_enabled && !runtime_was_enabled)
7417 cfs_bandwidth_usage_inc();
7418 raw_spin_lock_irq(&cfs_b->lock);
7419 cfs_b->period = ns_to_ktime(period);
7420 cfs_b->quota = quota;
7422 __refill_cfs_bandwidth_runtime(cfs_b);
7423 /* restart the period timer (if active) to handle new period expiry */
7424 if (runtime_enabled && cfs_b->timer_active) {
7425 /* force a reprogram */
7426 cfs_b->timer_active = 0;
7427 __start_cfs_bandwidth(cfs_b);
7429 raw_spin_unlock_irq(&cfs_b->lock);
7431 for_each_possible_cpu(i) {
7432 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7433 struct rq *rq = cfs_rq->rq;
7435 raw_spin_lock_irq(&rq->lock);
7436 cfs_rq->runtime_enabled = runtime_enabled;
7437 cfs_rq->runtime_remaining = 0;
7439 if (cfs_rq->throttled)
7440 unthrottle_cfs_rq(cfs_rq);
7441 raw_spin_unlock_irq(&rq->lock);
7443 if (runtime_was_enabled && !runtime_enabled)
7444 cfs_bandwidth_usage_dec();
7446 mutex_unlock(&cfs_constraints_mutex);
7451 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7455 period = ktime_to_ns(tg->cfs_bandwidth.period);
7456 if (cfs_quota_us < 0)
7457 quota = RUNTIME_INF;
7459 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7461 return tg_set_cfs_bandwidth(tg, period, quota);
7464 long tg_get_cfs_quota(struct task_group *tg)
7468 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7471 quota_us = tg->cfs_bandwidth.quota;
7472 do_div(quota_us, NSEC_PER_USEC);
7477 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7481 period = (u64)cfs_period_us * NSEC_PER_USEC;
7482 quota = tg->cfs_bandwidth.quota;
7484 return tg_set_cfs_bandwidth(tg, period, quota);
7487 long tg_get_cfs_period(struct task_group *tg)
7491 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7492 do_div(cfs_period_us, NSEC_PER_USEC);
7494 return cfs_period_us;
7497 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7500 return tg_get_cfs_quota(css_tg(css));
7503 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7504 struct cftype *cftype, s64 cfs_quota_us)
7506 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7509 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7512 return tg_get_cfs_period(css_tg(css));
7515 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7516 struct cftype *cftype, u64 cfs_period_us)
7518 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7521 struct cfs_schedulable_data {
7522 struct task_group *tg;
7527 * normalize group quota/period to be quota/max_period
7528 * note: units are usecs
7530 static u64 normalize_cfs_quota(struct task_group *tg,
7531 struct cfs_schedulable_data *d)
7539 period = tg_get_cfs_period(tg);
7540 quota = tg_get_cfs_quota(tg);
7543 /* note: these should typically be equivalent */
7544 if (quota == RUNTIME_INF || quota == -1)
7547 return to_ratio(period, quota);
7550 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7552 struct cfs_schedulable_data *d = data;
7553 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7554 s64 quota = 0, parent_quota = -1;
7557 quota = RUNTIME_INF;
7559 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7561 quota = normalize_cfs_quota(tg, d);
7562 parent_quota = parent_b->hierarchal_quota;
7565 * ensure max(child_quota) <= parent_quota, inherit when no
7568 if (quota == RUNTIME_INF)
7569 quota = parent_quota;
7570 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7573 cfs_b->hierarchal_quota = quota;
7578 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7581 struct cfs_schedulable_data data = {
7587 if (quota != RUNTIME_INF) {
7588 do_div(data.period, NSEC_PER_USEC);
7589 do_div(data.quota, NSEC_PER_USEC);
7593 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7599 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7600 struct cgroup_map_cb *cb)
7602 struct task_group *tg = css_tg(css);
7603 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7605 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7606 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7607 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7611 #endif /* CONFIG_CFS_BANDWIDTH */
7612 #endif /* CONFIG_FAIR_GROUP_SCHED */
7614 #ifdef CONFIG_RT_GROUP_SCHED
7615 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7616 struct cftype *cft, s64 val)
7618 return sched_group_set_rt_runtime(css_tg(css), val);
7621 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7624 return sched_group_rt_runtime(css_tg(css));
7627 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7628 struct cftype *cftype, u64 rt_period_us)
7630 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7633 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7636 return sched_group_rt_period(css_tg(css));
7638 #endif /* CONFIG_RT_GROUP_SCHED */
7640 static struct cftype cpu_files[] = {
7641 #ifdef CONFIG_FAIR_GROUP_SCHED
7644 .read_u64 = cpu_shares_read_u64,
7645 .write_u64 = cpu_shares_write_u64,
7648 #ifdef CONFIG_CFS_BANDWIDTH
7650 .name = "cfs_quota_us",
7651 .read_s64 = cpu_cfs_quota_read_s64,
7652 .write_s64 = cpu_cfs_quota_write_s64,
7655 .name = "cfs_period_us",
7656 .read_u64 = cpu_cfs_period_read_u64,
7657 .write_u64 = cpu_cfs_period_write_u64,
7661 .read_map = cpu_stats_show,
7664 #ifdef CONFIG_RT_GROUP_SCHED
7666 .name = "rt_runtime_us",
7667 .read_s64 = cpu_rt_runtime_read,
7668 .write_s64 = cpu_rt_runtime_write,
7671 .name = "rt_period_us",
7672 .read_u64 = cpu_rt_period_read_uint,
7673 .write_u64 = cpu_rt_period_write_uint,
7679 struct cgroup_subsys cpu_cgroup_subsys = {
7681 .css_alloc = cpu_cgroup_css_alloc,
7682 .css_free = cpu_cgroup_css_free,
7683 .css_online = cpu_cgroup_css_online,
7684 .css_offline = cpu_cgroup_css_offline,
7685 .can_attach = cpu_cgroup_can_attach,
7686 .attach = cpu_cgroup_attach,
7687 .exit = cpu_cgroup_exit,
7688 .subsys_id = cpu_cgroup_subsys_id,
7689 .base_cftypes = cpu_files,
7693 #endif /* CONFIG_CGROUP_SCHED */
7695 void dump_cpu_task(int cpu)
7697 pr_info("Task dump for CPU %d:\n", cpu);
7698 sched_show_task(cpu_curr(cpu));