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_rt_policy(p))
903 prio = MAX_RT_PRIO-1 - p->rt_priority;
905 prio = __normal_prio(p);
910 * Calculate the current priority, i.e. the priority
911 * taken into account by the scheduler. This value might
912 * be boosted by RT tasks, or might be boosted by
913 * interactivity modifiers. Will be RT if the task got
914 * RT-boosted. If not then it returns p->normal_prio.
916 static int effective_prio(struct task_struct *p)
918 p->normal_prio = normal_prio(p);
920 * If we are RT tasks or we were boosted to RT priority,
921 * keep the priority unchanged. Otherwise, update priority
922 * to the normal priority:
924 if (!rt_prio(p->prio))
925 return p->normal_prio;
930 * task_curr - is this task currently executing on a CPU?
931 * @p: the task in question.
933 * Return: 1 if the task is currently executing. 0 otherwise.
935 inline int task_curr(const struct task_struct *p)
937 return cpu_curr(task_cpu(p)) == p;
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
947 p->sched_class->switched_to(rq, p);
948 } else if (oldprio != p->prio)
949 p->sched_class->prio_changed(rq, p, oldprio);
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
954 const struct sched_class *class;
956 if (p->sched_class == rq->curr->sched_class) {
957 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
959 for_each_class(class) {
960 if (class == rq->curr->sched_class)
962 if (class == p->sched_class) {
963 resched_task(rq->curr);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974 rq->skip_clock_update = 1;
978 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
986 !(task_preempt_count(p) & PREEMPT_ACTIVE));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1000 lockdep_is_held(&task_rq(p)->lock)));
1004 trace_sched_migrate_task(p, new_cpu);
1006 if (task_cpu(p) != new_cpu) {
1007 if (p->sched_class->migrate_task_rq)
1008 p->sched_class->migrate_task_rq(p, new_cpu);
1009 p->se.nr_migrations++;
1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1013 __set_task_cpu(p, new_cpu);
1016 struct migration_arg {
1017 struct task_struct *task;
1021 static int migration_cpu_stop(void *data);
1024 * wait_task_inactive - wait for a thread to unschedule.
1026 * If @match_state is nonzero, it's the @p->state value just checked and
1027 * not expected to change. If it changes, i.e. @p might have woken up,
1028 * then return zero. When we succeed in waiting for @p to be off its CPU,
1029 * we return a positive number (its total switch count). If a second call
1030 * a short while later returns the same number, the caller can be sure that
1031 * @p has remained unscheduled the whole time.
1033 * The caller must ensure that the task *will* unschedule sometime soon,
1034 * else this function might spin for a *long* time. This function can't
1035 * be called with interrupts off, or it may introduce deadlock with
1036 * smp_call_function() if an IPI is sent by the same process we are
1037 * waiting to become inactive.
1039 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1041 unsigned long flags;
1048 * We do the initial early heuristics without holding
1049 * any task-queue locks at all. We'll only try to get
1050 * the runqueue lock when things look like they will
1056 * If the task is actively running on another CPU
1057 * still, just relax and busy-wait without holding
1060 * NOTE! Since we don't hold any locks, it's not
1061 * even sure that "rq" stays as the right runqueue!
1062 * But we don't care, since "task_running()" will
1063 * return false if the runqueue has changed and p
1064 * is actually now running somewhere else!
1066 while (task_running(rq, p)) {
1067 if (match_state && unlikely(p->state != match_state))
1073 * Ok, time to look more closely! We need the rq
1074 * lock now, to be *sure*. If we're wrong, we'll
1075 * just go back and repeat.
1077 rq = task_rq_lock(p, &flags);
1078 trace_sched_wait_task(p);
1079 running = task_running(rq, p);
1082 if (!match_state || p->state == match_state)
1083 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1084 task_rq_unlock(rq, p, &flags);
1087 * If it changed from the expected state, bail out now.
1089 if (unlikely(!ncsw))
1093 * Was it really running after all now that we
1094 * checked with the proper locks actually held?
1096 * Oops. Go back and try again..
1098 if (unlikely(running)) {
1104 * It's not enough that it's not actively running,
1105 * it must be off the runqueue _entirely_, and not
1108 * So if it was still runnable (but just not actively
1109 * running right now), it's preempted, and we should
1110 * yield - it could be a while.
1112 if (unlikely(on_rq)) {
1113 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1115 set_current_state(TASK_UNINTERRUPTIBLE);
1116 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1121 * Ahh, all good. It wasn't running, and it wasn't
1122 * runnable, which means that it will never become
1123 * running in the future either. We're all done!
1132 * kick_process - kick a running thread to enter/exit the kernel
1133 * @p: the to-be-kicked thread
1135 * Cause a process which is running on another CPU to enter
1136 * kernel-mode, without any delay. (to get signals handled.)
1138 * NOTE: this function doesn't have to take the runqueue lock,
1139 * because all it wants to ensure is that the remote task enters
1140 * the kernel. If the IPI races and the task has been migrated
1141 * to another CPU then no harm is done and the purpose has been
1144 void kick_process(struct task_struct *p)
1150 if ((cpu != smp_processor_id()) && task_curr(p))
1151 smp_send_reschedule(cpu);
1154 EXPORT_SYMBOL_GPL(kick_process);
1155 #endif /* CONFIG_SMP */
1159 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1161 static int select_fallback_rq(int cpu, struct task_struct *p)
1163 int nid = cpu_to_node(cpu);
1164 const struct cpumask *nodemask = NULL;
1165 enum { cpuset, possible, fail } state = cpuset;
1169 * If the node that the cpu is on has been offlined, cpu_to_node()
1170 * will return -1. There is no cpu on the node, and we should
1171 * select the cpu on the other node.
1174 nodemask = cpumask_of_node(nid);
1176 /* Look for allowed, online CPU in same node. */
1177 for_each_cpu(dest_cpu, nodemask) {
1178 if (!cpu_online(dest_cpu))
1180 if (!cpu_active(dest_cpu))
1182 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1188 /* Any allowed, online CPU? */
1189 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1190 if (!cpu_online(dest_cpu))
1192 if (!cpu_active(dest_cpu))
1199 /* No more Mr. Nice Guy. */
1200 cpuset_cpus_allowed_fallback(p);
1205 do_set_cpus_allowed(p, cpu_possible_mask);
1216 if (state != cpuset) {
1218 * Don't tell them about moving exiting tasks or
1219 * kernel threads (both mm NULL), since they never
1222 if (p->mm && printk_ratelimit()) {
1223 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1224 task_pid_nr(p), p->comm, cpu);
1232 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1235 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1237 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1240 * In order not to call set_task_cpu() on a blocking task we need
1241 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1244 * Since this is common to all placement strategies, this lives here.
1246 * [ this allows ->select_task() to simply return task_cpu(p) and
1247 * not worry about this generic constraint ]
1249 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1251 cpu = select_fallback_rq(task_cpu(p), p);
1256 static void update_avg(u64 *avg, u64 sample)
1258 s64 diff = sample - *avg;
1264 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1266 #ifdef CONFIG_SCHEDSTATS
1267 struct rq *rq = this_rq();
1270 int this_cpu = smp_processor_id();
1272 if (cpu == this_cpu) {
1273 schedstat_inc(rq, ttwu_local);
1274 schedstat_inc(p, se.statistics.nr_wakeups_local);
1276 struct sched_domain *sd;
1278 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1280 for_each_domain(this_cpu, sd) {
1281 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1282 schedstat_inc(sd, ttwu_wake_remote);
1289 if (wake_flags & WF_MIGRATED)
1290 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1292 #endif /* CONFIG_SMP */
1294 schedstat_inc(rq, ttwu_count);
1295 schedstat_inc(p, se.statistics.nr_wakeups);
1297 if (wake_flags & WF_SYNC)
1298 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1300 #endif /* CONFIG_SCHEDSTATS */
1303 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1305 activate_task(rq, p, en_flags);
1308 /* if a worker is waking up, notify workqueue */
1309 if (p->flags & PF_WQ_WORKER)
1310 wq_worker_waking_up(p, cpu_of(rq));
1314 * Mark the task runnable and perform wakeup-preemption.
1317 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1319 check_preempt_curr(rq, p, wake_flags);
1320 trace_sched_wakeup(p, true);
1322 p->state = TASK_RUNNING;
1324 if (p->sched_class->task_woken)
1325 p->sched_class->task_woken(rq, p);
1327 if (rq->idle_stamp) {
1328 u64 delta = rq_clock(rq) - rq->idle_stamp;
1329 u64 max = 2*rq->max_idle_balance_cost;
1331 update_avg(&rq->avg_idle, delta);
1333 if (rq->avg_idle > max)
1342 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1345 if (p->sched_contributes_to_load)
1346 rq->nr_uninterruptible--;
1349 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1350 ttwu_do_wakeup(rq, p, wake_flags);
1354 * Called in case the task @p isn't fully descheduled from its runqueue,
1355 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1356 * since all we need to do is flip p->state to TASK_RUNNING, since
1357 * the task is still ->on_rq.
1359 static int ttwu_remote(struct task_struct *p, int wake_flags)
1364 rq = __task_rq_lock(p);
1366 /* check_preempt_curr() may use rq clock */
1367 update_rq_clock(rq);
1368 ttwu_do_wakeup(rq, p, wake_flags);
1371 __task_rq_unlock(rq);
1377 static void sched_ttwu_pending(void)
1379 struct rq *rq = this_rq();
1380 struct llist_node *llist = llist_del_all(&rq->wake_list);
1381 struct task_struct *p;
1383 raw_spin_lock(&rq->lock);
1386 p = llist_entry(llist, struct task_struct, wake_entry);
1387 llist = llist_next(llist);
1388 ttwu_do_activate(rq, p, 0);
1391 raw_spin_unlock(&rq->lock);
1394 void scheduler_ipi(void)
1397 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1398 * TIF_NEED_RESCHED remotely (for the first time) will also send
1401 if (tif_need_resched())
1402 set_preempt_need_resched();
1404 if (llist_empty(&this_rq()->wake_list)
1405 && !tick_nohz_full_cpu(smp_processor_id())
1406 && !got_nohz_idle_kick())
1410 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1411 * traditionally all their work was done from the interrupt return
1412 * path. Now that we actually do some work, we need to make sure
1415 * Some archs already do call them, luckily irq_enter/exit nest
1418 * Arguably we should visit all archs and update all handlers,
1419 * however a fair share of IPIs are still resched only so this would
1420 * somewhat pessimize the simple resched case.
1423 tick_nohz_full_check();
1424 sched_ttwu_pending();
1427 * Check if someone kicked us for doing the nohz idle load balance.
1429 if (unlikely(got_nohz_idle_kick())) {
1430 this_rq()->idle_balance = 1;
1431 raise_softirq_irqoff(SCHED_SOFTIRQ);
1436 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1438 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1439 smp_send_reschedule(cpu);
1442 bool cpus_share_cache(int this_cpu, int that_cpu)
1444 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1446 #endif /* CONFIG_SMP */
1448 static void ttwu_queue(struct task_struct *p, int cpu)
1450 struct rq *rq = cpu_rq(cpu);
1452 #if defined(CONFIG_SMP)
1453 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1454 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1455 ttwu_queue_remote(p, cpu);
1460 raw_spin_lock(&rq->lock);
1461 ttwu_do_activate(rq, p, 0);
1462 raw_spin_unlock(&rq->lock);
1466 * try_to_wake_up - wake up a thread
1467 * @p: the thread to be awakened
1468 * @state: the mask of task states that can be woken
1469 * @wake_flags: wake modifier flags (WF_*)
1471 * Put it on the run-queue if it's not already there. The "current"
1472 * thread is always on the run-queue (except when the actual
1473 * re-schedule is in progress), and as such you're allowed to do
1474 * the simpler "current->state = TASK_RUNNING" to mark yourself
1475 * runnable without the overhead of this.
1477 * Return: %true if @p was woken up, %false if it was already running.
1478 * or @state didn't match @p's state.
1481 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1483 unsigned long flags;
1484 int cpu, success = 0;
1487 * If we are going to wake up a thread waiting for CONDITION we
1488 * need to ensure that CONDITION=1 done by the caller can not be
1489 * reordered with p->state check below. This pairs with mb() in
1490 * set_current_state() the waiting thread does.
1492 smp_mb__before_spinlock();
1493 raw_spin_lock_irqsave(&p->pi_lock, flags);
1494 if (!(p->state & state))
1497 success = 1; /* we're going to change ->state */
1500 if (p->on_rq && ttwu_remote(p, wake_flags))
1505 * If the owning (remote) cpu is still in the middle of schedule() with
1506 * this task as prev, wait until its done referencing the task.
1511 * Pairs with the smp_wmb() in finish_lock_switch().
1515 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1516 p->state = TASK_WAKING;
1518 if (p->sched_class->task_waking)
1519 p->sched_class->task_waking(p);
1521 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1522 if (task_cpu(p) != cpu) {
1523 wake_flags |= WF_MIGRATED;
1524 set_task_cpu(p, cpu);
1526 #endif /* CONFIG_SMP */
1530 ttwu_stat(p, cpu, wake_flags);
1532 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1538 * try_to_wake_up_local - try to wake up a local task with rq lock held
1539 * @p: the thread to be awakened
1541 * Put @p on the run-queue if it's not already there. The caller must
1542 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1545 static void try_to_wake_up_local(struct task_struct *p)
1547 struct rq *rq = task_rq(p);
1549 if (WARN_ON_ONCE(rq != this_rq()) ||
1550 WARN_ON_ONCE(p == current))
1553 lockdep_assert_held(&rq->lock);
1555 if (!raw_spin_trylock(&p->pi_lock)) {
1556 raw_spin_unlock(&rq->lock);
1557 raw_spin_lock(&p->pi_lock);
1558 raw_spin_lock(&rq->lock);
1561 if (!(p->state & TASK_NORMAL))
1565 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1567 ttwu_do_wakeup(rq, p, 0);
1568 ttwu_stat(p, smp_processor_id(), 0);
1570 raw_spin_unlock(&p->pi_lock);
1574 * wake_up_process - Wake up a specific process
1575 * @p: The process to be woken up.
1577 * Attempt to wake up the nominated process and move it to the set of runnable
1580 * Return: 1 if the process was woken up, 0 if it was already running.
1582 * It may be assumed that this function implies a write memory barrier before
1583 * changing the task state if and only if any tasks are woken up.
1585 int wake_up_process(struct task_struct *p)
1587 WARN_ON(task_is_stopped_or_traced(p));
1588 return try_to_wake_up(p, TASK_NORMAL, 0);
1590 EXPORT_SYMBOL(wake_up_process);
1592 int wake_up_state(struct task_struct *p, unsigned int state)
1594 return try_to_wake_up(p, state, 0);
1598 * Perform scheduler related setup for a newly forked process p.
1599 * p is forked by current.
1601 * __sched_fork() is basic setup used by init_idle() too:
1603 static void __sched_fork(struct task_struct *p)
1608 p->se.exec_start = 0;
1609 p->se.sum_exec_runtime = 0;
1610 p->se.prev_sum_exec_runtime = 0;
1611 p->se.nr_migrations = 0;
1613 INIT_LIST_HEAD(&p->se.group_node);
1615 #ifdef CONFIG_SCHEDSTATS
1616 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1619 INIT_LIST_HEAD(&p->rt.run_list);
1621 #ifdef CONFIG_PREEMPT_NOTIFIERS
1622 INIT_HLIST_HEAD(&p->preempt_notifiers);
1625 #ifdef CONFIG_NUMA_BALANCING
1626 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1627 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1628 p->mm->numa_next_reset = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1629 p->mm->numa_scan_seq = 0;
1632 p->node_stamp = 0ULL;
1633 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1634 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1635 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1636 p->numa_preferred_nid = -1;
1637 p->numa_work.next = &p->numa_work;
1638 p->numa_faults = NULL;
1639 #endif /* CONFIG_NUMA_BALANCING */
1642 #ifdef CONFIG_NUMA_BALANCING
1643 #ifdef CONFIG_SCHED_DEBUG
1644 void set_numabalancing_state(bool enabled)
1647 sched_feat_set("NUMA");
1649 sched_feat_set("NO_NUMA");
1652 __read_mostly bool numabalancing_enabled;
1654 void set_numabalancing_state(bool enabled)
1656 numabalancing_enabled = enabled;
1658 #endif /* CONFIG_SCHED_DEBUG */
1659 #endif /* CONFIG_NUMA_BALANCING */
1662 * fork()/clone()-time setup:
1664 void sched_fork(struct task_struct *p)
1666 unsigned long flags;
1667 int cpu = get_cpu();
1671 * We mark the process as running here. This guarantees that
1672 * nobody will actually run it, and a signal or other external
1673 * event cannot wake it up and insert it on the runqueue either.
1675 p->state = TASK_RUNNING;
1678 * Make sure we do not leak PI boosting priority to the child.
1680 p->prio = current->normal_prio;
1683 * Revert to default priority/policy on fork if requested.
1685 if (unlikely(p->sched_reset_on_fork)) {
1686 if (task_has_rt_policy(p)) {
1687 p->policy = SCHED_NORMAL;
1688 p->static_prio = NICE_TO_PRIO(0);
1690 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1691 p->static_prio = NICE_TO_PRIO(0);
1693 p->prio = p->normal_prio = __normal_prio(p);
1697 * We don't need the reset flag anymore after the fork. It has
1698 * fulfilled its duty:
1700 p->sched_reset_on_fork = 0;
1703 if (!rt_prio(p->prio))
1704 p->sched_class = &fair_sched_class;
1706 if (p->sched_class->task_fork)
1707 p->sched_class->task_fork(p);
1710 * The child is not yet in the pid-hash so no cgroup attach races,
1711 * and the cgroup is pinned to this child due to cgroup_fork()
1712 * is ran before sched_fork().
1714 * Silence PROVE_RCU.
1716 raw_spin_lock_irqsave(&p->pi_lock, flags);
1717 set_task_cpu(p, cpu);
1718 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1720 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1721 if (likely(sched_info_on()))
1722 memset(&p->sched_info, 0, sizeof(p->sched_info));
1724 #if defined(CONFIG_SMP)
1727 init_task_preempt_count(p);
1729 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1736 * wake_up_new_task - wake up a newly created task for the first time.
1738 * This function will do some initial scheduler statistics housekeeping
1739 * that must be done for every newly created context, then puts the task
1740 * on the runqueue and wakes it.
1742 void wake_up_new_task(struct task_struct *p)
1744 unsigned long flags;
1747 raw_spin_lock_irqsave(&p->pi_lock, flags);
1750 * Fork balancing, do it here and not earlier because:
1751 * - cpus_allowed can change in the fork path
1752 * - any previously selected cpu might disappear through hotplug
1754 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1757 /* Initialize new task's runnable average */
1758 init_task_runnable_average(p);
1759 rq = __task_rq_lock(p);
1760 activate_task(rq, p, 0);
1762 trace_sched_wakeup_new(p, true);
1763 check_preempt_curr(rq, p, WF_FORK);
1765 if (p->sched_class->task_woken)
1766 p->sched_class->task_woken(rq, p);
1768 task_rq_unlock(rq, p, &flags);
1771 #ifdef CONFIG_PREEMPT_NOTIFIERS
1774 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1775 * @notifier: notifier struct to register
1777 void preempt_notifier_register(struct preempt_notifier *notifier)
1779 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1781 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1784 * preempt_notifier_unregister - no longer interested in preemption notifications
1785 * @notifier: notifier struct to unregister
1787 * This is safe to call from within a preemption notifier.
1789 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1791 hlist_del(¬ifier->link);
1793 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1795 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1797 struct preempt_notifier *notifier;
1799 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1800 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1804 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1805 struct task_struct *next)
1807 struct preempt_notifier *notifier;
1809 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1810 notifier->ops->sched_out(notifier, next);
1813 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1815 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1820 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1821 struct task_struct *next)
1825 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1828 * prepare_task_switch - prepare to switch tasks
1829 * @rq: the runqueue preparing to switch
1830 * @prev: the current task that is being switched out
1831 * @next: the task we are going to switch to.
1833 * This is called with the rq lock held and interrupts off. It must
1834 * be paired with a subsequent finish_task_switch after the context
1837 * prepare_task_switch sets up locking and calls architecture specific
1841 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1842 struct task_struct *next)
1844 trace_sched_switch(prev, next);
1845 sched_info_switch(rq, prev, next);
1846 perf_event_task_sched_out(prev, next);
1847 fire_sched_out_preempt_notifiers(prev, next);
1848 prepare_lock_switch(rq, next);
1849 prepare_arch_switch(next);
1853 * finish_task_switch - clean up after a task-switch
1854 * @rq: runqueue associated with task-switch
1855 * @prev: the thread we just switched away from.
1857 * finish_task_switch must be called after the context switch, paired
1858 * with a prepare_task_switch call before the context switch.
1859 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1860 * and do any other architecture-specific cleanup actions.
1862 * Note that we may have delayed dropping an mm in context_switch(). If
1863 * so, we finish that here outside of the runqueue lock. (Doing it
1864 * with the lock held can cause deadlocks; see schedule() for
1867 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1868 __releases(rq->lock)
1870 struct mm_struct *mm = rq->prev_mm;
1876 * A task struct has one reference for the use as "current".
1877 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1878 * schedule one last time. The schedule call will never return, and
1879 * the scheduled task must drop that reference.
1880 * The test for TASK_DEAD must occur while the runqueue locks are
1881 * still held, otherwise prev could be scheduled on another cpu, die
1882 * there before we look at prev->state, and then the reference would
1884 * Manfred Spraul <manfred@colorfullife.com>
1886 prev_state = prev->state;
1887 vtime_task_switch(prev);
1888 finish_arch_switch(prev);
1889 perf_event_task_sched_in(prev, current);
1890 finish_lock_switch(rq, prev);
1891 finish_arch_post_lock_switch();
1893 fire_sched_in_preempt_notifiers(current);
1896 if (unlikely(prev_state == TASK_DEAD)) {
1897 task_numa_free(prev);
1900 * Remove function-return probe instances associated with this
1901 * task and put them back on the free list.
1903 kprobe_flush_task(prev);
1904 put_task_struct(prev);
1907 tick_nohz_task_switch(current);
1912 /* assumes rq->lock is held */
1913 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1915 if (prev->sched_class->pre_schedule)
1916 prev->sched_class->pre_schedule(rq, prev);
1919 /* rq->lock is NOT held, but preemption is disabled */
1920 static inline void post_schedule(struct rq *rq)
1922 if (rq->post_schedule) {
1923 unsigned long flags;
1925 raw_spin_lock_irqsave(&rq->lock, flags);
1926 if (rq->curr->sched_class->post_schedule)
1927 rq->curr->sched_class->post_schedule(rq);
1928 raw_spin_unlock_irqrestore(&rq->lock, flags);
1930 rq->post_schedule = 0;
1936 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1940 static inline void post_schedule(struct rq *rq)
1947 * schedule_tail - first thing a freshly forked thread must call.
1948 * @prev: the thread we just switched away from.
1950 asmlinkage void schedule_tail(struct task_struct *prev)
1951 __releases(rq->lock)
1953 struct rq *rq = this_rq();
1955 finish_task_switch(rq, prev);
1958 * FIXME: do we need to worry about rq being invalidated by the
1963 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1964 /* In this case, finish_task_switch does not reenable preemption */
1967 if (current->set_child_tid)
1968 put_user(task_pid_vnr(current), current->set_child_tid);
1972 * context_switch - switch to the new MM and the new
1973 * thread's register state.
1976 context_switch(struct rq *rq, struct task_struct *prev,
1977 struct task_struct *next)
1979 struct mm_struct *mm, *oldmm;
1981 prepare_task_switch(rq, prev, next);
1984 oldmm = prev->active_mm;
1986 * For paravirt, this is coupled with an exit in switch_to to
1987 * combine the page table reload and the switch backend into
1990 arch_start_context_switch(prev);
1993 next->active_mm = oldmm;
1994 atomic_inc(&oldmm->mm_count);
1995 enter_lazy_tlb(oldmm, next);
1997 switch_mm(oldmm, mm, next);
2000 prev->active_mm = NULL;
2001 rq->prev_mm = oldmm;
2004 * Since the runqueue lock will be released by the next
2005 * task (which is an invalid locking op but in the case
2006 * of the scheduler it's an obvious special-case), so we
2007 * do an early lockdep release here:
2009 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2010 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2013 context_tracking_task_switch(prev, next);
2014 /* Here we just switch the register state and the stack. */
2015 switch_to(prev, next, prev);
2019 * this_rq must be evaluated again because prev may have moved
2020 * CPUs since it called schedule(), thus the 'rq' on its stack
2021 * frame will be invalid.
2023 finish_task_switch(this_rq(), prev);
2027 * nr_running and nr_context_switches:
2029 * externally visible scheduler statistics: current number of runnable
2030 * threads, total number of context switches performed since bootup.
2032 unsigned long nr_running(void)
2034 unsigned long i, sum = 0;
2036 for_each_online_cpu(i)
2037 sum += cpu_rq(i)->nr_running;
2042 unsigned long long nr_context_switches(void)
2045 unsigned long long sum = 0;
2047 for_each_possible_cpu(i)
2048 sum += cpu_rq(i)->nr_switches;
2053 unsigned long nr_iowait(void)
2055 unsigned long i, sum = 0;
2057 for_each_possible_cpu(i)
2058 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2063 unsigned long nr_iowait_cpu(int cpu)
2065 struct rq *this = cpu_rq(cpu);
2066 return atomic_read(&this->nr_iowait);
2072 * sched_exec - execve() is a valuable balancing opportunity, because at
2073 * this point the task has the smallest effective memory and cache footprint.
2075 void sched_exec(void)
2077 struct task_struct *p = current;
2078 unsigned long flags;
2081 raw_spin_lock_irqsave(&p->pi_lock, flags);
2082 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2083 if (dest_cpu == smp_processor_id())
2086 if (likely(cpu_active(dest_cpu))) {
2087 struct migration_arg arg = { p, dest_cpu };
2089 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2090 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2094 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2099 DEFINE_PER_CPU(struct kernel_stat, kstat);
2100 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2102 EXPORT_PER_CPU_SYMBOL(kstat);
2103 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2106 * Return any ns on the sched_clock that have not yet been accounted in
2107 * @p in case that task is currently running.
2109 * Called with task_rq_lock() held on @rq.
2111 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2115 if (task_current(rq, p)) {
2116 update_rq_clock(rq);
2117 ns = rq_clock_task(rq) - p->se.exec_start;
2125 unsigned long long task_delta_exec(struct task_struct *p)
2127 unsigned long flags;
2131 rq = task_rq_lock(p, &flags);
2132 ns = do_task_delta_exec(p, rq);
2133 task_rq_unlock(rq, p, &flags);
2139 * Return accounted runtime for the task.
2140 * In case the task is currently running, return the runtime plus current's
2141 * pending runtime that have not been accounted yet.
2143 unsigned long long task_sched_runtime(struct task_struct *p)
2145 unsigned long flags;
2149 rq = task_rq_lock(p, &flags);
2150 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2151 task_rq_unlock(rq, p, &flags);
2157 * This function gets called by the timer code, with HZ frequency.
2158 * We call it with interrupts disabled.
2160 void scheduler_tick(void)
2162 int cpu = smp_processor_id();
2163 struct rq *rq = cpu_rq(cpu);
2164 struct task_struct *curr = rq->curr;
2168 raw_spin_lock(&rq->lock);
2169 update_rq_clock(rq);
2170 curr->sched_class->task_tick(rq, curr, 0);
2171 update_cpu_load_active(rq);
2172 raw_spin_unlock(&rq->lock);
2174 perf_event_task_tick();
2177 rq->idle_balance = idle_cpu(cpu);
2178 trigger_load_balance(rq, cpu);
2180 rq_last_tick_reset(rq);
2183 #ifdef CONFIG_NO_HZ_FULL
2185 * scheduler_tick_max_deferment
2187 * Keep at least one tick per second when a single
2188 * active task is running because the scheduler doesn't
2189 * yet completely support full dynticks environment.
2191 * This makes sure that uptime, CFS vruntime, load
2192 * balancing, etc... continue to move forward, even
2193 * with a very low granularity.
2195 * Return: Maximum deferment in nanoseconds.
2197 u64 scheduler_tick_max_deferment(void)
2199 struct rq *rq = this_rq();
2200 unsigned long next, now = ACCESS_ONCE(jiffies);
2202 next = rq->last_sched_tick + HZ;
2204 if (time_before_eq(next, now))
2207 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2211 notrace unsigned long get_parent_ip(unsigned long addr)
2213 if (in_lock_functions(addr)) {
2214 addr = CALLER_ADDR2;
2215 if (in_lock_functions(addr))
2216 addr = CALLER_ADDR3;
2221 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2222 defined(CONFIG_PREEMPT_TRACER))
2224 void __kprobes preempt_count_add(int val)
2226 #ifdef CONFIG_DEBUG_PREEMPT
2230 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2233 __preempt_count_add(val);
2234 #ifdef CONFIG_DEBUG_PREEMPT
2236 * Spinlock count overflowing soon?
2238 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2241 if (preempt_count() == val)
2242 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2244 EXPORT_SYMBOL(preempt_count_add);
2246 void __kprobes preempt_count_sub(int val)
2248 #ifdef CONFIG_DEBUG_PREEMPT
2252 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2255 * Is the spinlock portion underflowing?
2257 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2258 !(preempt_count() & PREEMPT_MASK)))
2262 if (preempt_count() == val)
2263 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2264 __preempt_count_sub(val);
2266 EXPORT_SYMBOL(preempt_count_sub);
2271 * Print scheduling while atomic bug:
2273 static noinline void __schedule_bug(struct task_struct *prev)
2275 if (oops_in_progress)
2278 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2279 prev->comm, prev->pid, preempt_count());
2281 debug_show_held_locks(prev);
2283 if (irqs_disabled())
2284 print_irqtrace_events(prev);
2286 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2290 * Various schedule()-time debugging checks and statistics:
2292 static inline void schedule_debug(struct task_struct *prev)
2295 * Test if we are atomic. Since do_exit() needs to call into
2296 * schedule() atomically, we ignore that path for now.
2297 * Otherwise, whine if we are scheduling when we should not be.
2299 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2300 __schedule_bug(prev);
2303 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2305 schedstat_inc(this_rq(), sched_count);
2308 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2310 if (prev->on_rq || rq->skip_clock_update < 0)
2311 update_rq_clock(rq);
2312 prev->sched_class->put_prev_task(rq, prev);
2316 * Pick up the highest-prio task:
2318 static inline struct task_struct *
2319 pick_next_task(struct rq *rq)
2321 const struct sched_class *class;
2322 struct task_struct *p;
2325 * Optimization: we know that if all tasks are in
2326 * the fair class we can call that function directly:
2328 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2329 p = fair_sched_class.pick_next_task(rq);
2334 for_each_class(class) {
2335 p = class->pick_next_task(rq);
2340 BUG(); /* the idle class will always have a runnable task */
2344 * __schedule() is the main scheduler function.
2346 * The main means of driving the scheduler and thus entering this function are:
2348 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2350 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2351 * paths. For example, see arch/x86/entry_64.S.
2353 * To drive preemption between tasks, the scheduler sets the flag in timer
2354 * interrupt handler scheduler_tick().
2356 * 3. Wakeups don't really cause entry into schedule(). They add a
2357 * task to the run-queue and that's it.
2359 * Now, if the new task added to the run-queue preempts the current
2360 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2361 * called on the nearest possible occasion:
2363 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2365 * - in syscall or exception context, at the next outmost
2366 * preempt_enable(). (this might be as soon as the wake_up()'s
2369 * - in IRQ context, return from interrupt-handler to
2370 * preemptible context
2372 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2375 * - cond_resched() call
2376 * - explicit schedule() call
2377 * - return from syscall or exception to user-space
2378 * - return from interrupt-handler to user-space
2380 static void __sched __schedule(void)
2382 struct task_struct *prev, *next;
2383 unsigned long *switch_count;
2389 cpu = smp_processor_id();
2391 rcu_note_context_switch(cpu);
2394 schedule_debug(prev);
2396 if (sched_feat(HRTICK))
2400 * Make sure that signal_pending_state()->signal_pending() below
2401 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2402 * done by the caller to avoid the race with signal_wake_up().
2404 smp_mb__before_spinlock();
2405 raw_spin_lock_irq(&rq->lock);
2407 switch_count = &prev->nivcsw;
2408 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2409 if (unlikely(signal_pending_state(prev->state, prev))) {
2410 prev->state = TASK_RUNNING;
2412 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2416 * If a worker went to sleep, notify and ask workqueue
2417 * whether it wants to wake up a task to maintain
2420 if (prev->flags & PF_WQ_WORKER) {
2421 struct task_struct *to_wakeup;
2423 to_wakeup = wq_worker_sleeping(prev, cpu);
2425 try_to_wake_up_local(to_wakeup);
2428 switch_count = &prev->nvcsw;
2431 pre_schedule(rq, prev);
2433 if (unlikely(!rq->nr_running))
2434 idle_balance(cpu, rq);
2436 put_prev_task(rq, prev);
2437 next = pick_next_task(rq);
2438 clear_tsk_need_resched(prev);
2439 clear_preempt_need_resched();
2440 rq->skip_clock_update = 0;
2442 if (likely(prev != next)) {
2447 context_switch(rq, prev, next); /* unlocks the rq */
2449 * The context switch have flipped the stack from under us
2450 * and restored the local variables which were saved when
2451 * this task called schedule() in the past. prev == current
2452 * is still correct, but it can be moved to another cpu/rq.
2454 cpu = smp_processor_id();
2457 raw_spin_unlock_irq(&rq->lock);
2461 sched_preempt_enable_no_resched();
2466 static inline void sched_submit_work(struct task_struct *tsk)
2468 if (!tsk->state || tsk_is_pi_blocked(tsk))
2471 * If we are going to sleep and we have plugged IO queued,
2472 * make sure to submit it to avoid deadlocks.
2474 if (blk_needs_flush_plug(tsk))
2475 blk_schedule_flush_plug(tsk);
2478 asmlinkage void __sched schedule(void)
2480 struct task_struct *tsk = current;
2482 sched_submit_work(tsk);
2485 EXPORT_SYMBOL(schedule);
2487 #ifdef CONFIG_CONTEXT_TRACKING
2488 asmlinkage void __sched schedule_user(void)
2491 * If we come here after a random call to set_need_resched(),
2492 * or we have been woken up remotely but the IPI has not yet arrived,
2493 * we haven't yet exited the RCU idle mode. Do it here manually until
2494 * we find a better solution.
2503 * schedule_preempt_disabled - called with preemption disabled
2505 * Returns with preemption disabled. Note: preempt_count must be 1
2507 void __sched schedule_preempt_disabled(void)
2509 sched_preempt_enable_no_resched();
2514 #ifdef CONFIG_PREEMPT
2516 * this is the entry point to schedule() from in-kernel preemption
2517 * off of preempt_enable. Kernel preemptions off return from interrupt
2518 * occur there and call schedule directly.
2520 asmlinkage void __sched notrace preempt_schedule(void)
2523 * If there is a non-zero preempt_count or interrupts are disabled,
2524 * we do not want to preempt the current task. Just return..
2526 if (likely(!preemptible()))
2530 __preempt_count_add(PREEMPT_ACTIVE);
2532 __preempt_count_sub(PREEMPT_ACTIVE);
2535 * Check again in case we missed a preemption opportunity
2536 * between schedule and now.
2539 } while (need_resched());
2541 EXPORT_SYMBOL(preempt_schedule);
2544 * this is the entry point to schedule() from kernel preemption
2545 * off of irq context.
2546 * Note, that this is called and return with irqs disabled. This will
2547 * protect us against recursive calling from irq.
2549 asmlinkage void __sched preempt_schedule_irq(void)
2551 enum ctx_state prev_state;
2553 /* Catch callers which need to be fixed */
2554 BUG_ON(preempt_count() || !irqs_disabled());
2556 prev_state = exception_enter();
2559 __preempt_count_add(PREEMPT_ACTIVE);
2562 local_irq_disable();
2563 __preempt_count_sub(PREEMPT_ACTIVE);
2566 * Check again in case we missed a preemption opportunity
2567 * between schedule and now.
2570 } while (need_resched());
2572 exception_exit(prev_state);
2575 #endif /* CONFIG_PREEMPT */
2577 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2580 return try_to_wake_up(curr->private, mode, wake_flags);
2582 EXPORT_SYMBOL(default_wake_function);
2585 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2586 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2587 * number) then we wake all the non-exclusive tasks and one exclusive task.
2589 * There are circumstances in which we can try to wake a task which has already
2590 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2591 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2593 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2594 int nr_exclusive, int wake_flags, void *key)
2596 wait_queue_t *curr, *next;
2598 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2599 unsigned flags = curr->flags;
2601 if (curr->func(curr, mode, wake_flags, key) &&
2602 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2608 * __wake_up - wake up threads blocked on a waitqueue.
2610 * @mode: which threads
2611 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2612 * @key: is directly passed to the wakeup function
2614 * It may be assumed that this function implies a write memory barrier before
2615 * changing the task state if and only if any tasks are woken up.
2617 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2618 int nr_exclusive, void *key)
2620 unsigned long flags;
2622 spin_lock_irqsave(&q->lock, flags);
2623 __wake_up_common(q, mode, nr_exclusive, 0, key);
2624 spin_unlock_irqrestore(&q->lock, flags);
2626 EXPORT_SYMBOL(__wake_up);
2629 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2631 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2633 __wake_up_common(q, mode, nr, 0, NULL);
2635 EXPORT_SYMBOL_GPL(__wake_up_locked);
2637 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2639 __wake_up_common(q, mode, 1, 0, key);
2641 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2644 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2646 * @mode: which threads
2647 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2648 * @key: opaque value to be passed to wakeup targets
2650 * The sync wakeup differs that the waker knows that it will schedule
2651 * away soon, so while the target thread will be woken up, it will not
2652 * be migrated to another CPU - ie. the two threads are 'synchronized'
2653 * with each other. This can prevent needless bouncing between CPUs.
2655 * On UP it can prevent extra preemption.
2657 * It may be assumed that this function implies a write memory barrier before
2658 * changing the task state if and only if any tasks are woken up.
2660 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2661 int nr_exclusive, void *key)
2663 unsigned long flags;
2664 int wake_flags = WF_SYNC;
2669 if (unlikely(nr_exclusive != 1))
2672 spin_lock_irqsave(&q->lock, flags);
2673 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2674 spin_unlock_irqrestore(&q->lock, flags);
2676 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2679 * __wake_up_sync - see __wake_up_sync_key()
2681 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2683 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2685 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2688 * complete: - signals a single thread waiting on this completion
2689 * @x: holds the state of this particular completion
2691 * This will wake up a single thread waiting on this completion. Threads will be
2692 * awakened in the same order in which they were queued.
2694 * See also complete_all(), wait_for_completion() and related routines.
2696 * It may be assumed that this function implies a write memory barrier before
2697 * changing the task state if and only if any tasks are woken up.
2699 void complete(struct completion *x)
2701 unsigned long flags;
2703 spin_lock_irqsave(&x->wait.lock, flags);
2705 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2706 spin_unlock_irqrestore(&x->wait.lock, flags);
2708 EXPORT_SYMBOL(complete);
2711 * complete_all: - signals all threads waiting on this completion
2712 * @x: holds the state of this particular completion
2714 * This will wake up all threads waiting on this particular completion event.
2716 * It may be assumed that this function implies a write memory barrier before
2717 * changing the task state if and only if any tasks are woken up.
2719 void complete_all(struct completion *x)
2721 unsigned long flags;
2723 spin_lock_irqsave(&x->wait.lock, flags);
2724 x->done += UINT_MAX/2;
2725 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2726 spin_unlock_irqrestore(&x->wait.lock, flags);
2728 EXPORT_SYMBOL(complete_all);
2730 static inline long __sched
2731 do_wait_for_common(struct completion *x,
2732 long (*action)(long), long timeout, int state)
2735 DECLARE_WAITQUEUE(wait, current);
2737 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2739 if (signal_pending_state(state, current)) {
2740 timeout = -ERESTARTSYS;
2743 __set_current_state(state);
2744 spin_unlock_irq(&x->wait.lock);
2745 timeout = action(timeout);
2746 spin_lock_irq(&x->wait.lock);
2747 } while (!x->done && timeout);
2748 __remove_wait_queue(&x->wait, &wait);
2753 return timeout ?: 1;
2756 static inline long __sched
2757 __wait_for_common(struct completion *x,
2758 long (*action)(long), long timeout, int state)
2762 spin_lock_irq(&x->wait.lock);
2763 timeout = do_wait_for_common(x, action, timeout, state);
2764 spin_unlock_irq(&x->wait.lock);
2769 wait_for_common(struct completion *x, long timeout, int state)
2771 return __wait_for_common(x, schedule_timeout, timeout, state);
2775 wait_for_common_io(struct completion *x, long timeout, int state)
2777 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2781 * wait_for_completion: - waits for completion of a task
2782 * @x: holds the state of this particular completion
2784 * This waits to be signaled for completion of a specific task. It is NOT
2785 * interruptible and there is no timeout.
2787 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2788 * and interrupt capability. Also see complete().
2790 void __sched wait_for_completion(struct completion *x)
2792 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2794 EXPORT_SYMBOL(wait_for_completion);
2797 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2798 * @x: holds the state of this particular completion
2799 * @timeout: timeout value in jiffies
2801 * This waits for either a completion of a specific task to be signaled or for a
2802 * specified timeout to expire. The timeout is in jiffies. It is not
2805 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2806 * till timeout) if completed.
2808 unsigned long __sched
2809 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2811 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2813 EXPORT_SYMBOL(wait_for_completion_timeout);
2816 * wait_for_completion_io: - waits for completion of a task
2817 * @x: holds the state of this particular completion
2819 * This waits to be signaled for completion of a specific task. It is NOT
2820 * interruptible and there is no timeout. The caller is accounted as waiting
2823 void __sched wait_for_completion_io(struct completion *x)
2825 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2827 EXPORT_SYMBOL(wait_for_completion_io);
2830 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2831 * @x: holds the state of this particular completion
2832 * @timeout: timeout value in jiffies
2834 * This waits for either a completion of a specific task to be signaled or for a
2835 * specified timeout to expire. The timeout is in jiffies. It is not
2836 * interruptible. The caller is accounted as waiting for IO.
2838 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2839 * till timeout) if completed.
2841 unsigned long __sched
2842 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2844 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2846 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2849 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2850 * @x: holds the state of this particular completion
2852 * This waits for completion of a specific task to be signaled. It is
2855 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2857 int __sched wait_for_completion_interruptible(struct completion *x)
2859 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2860 if (t == -ERESTARTSYS)
2864 EXPORT_SYMBOL(wait_for_completion_interruptible);
2867 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2868 * @x: holds the state of this particular completion
2869 * @timeout: timeout value in jiffies
2871 * This waits for either a completion of a specific task to be signaled or for a
2872 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2874 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2875 * or number of jiffies left till timeout) if completed.
2878 wait_for_completion_interruptible_timeout(struct completion *x,
2879 unsigned long timeout)
2881 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2883 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2886 * wait_for_completion_killable: - waits for completion of a task (killable)
2887 * @x: holds the state of this particular completion
2889 * This waits to be signaled for completion of a specific task. It can be
2890 * interrupted by a kill signal.
2892 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2894 int __sched wait_for_completion_killable(struct completion *x)
2896 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2897 if (t == -ERESTARTSYS)
2901 EXPORT_SYMBOL(wait_for_completion_killable);
2904 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2905 * @x: holds the state of this particular completion
2906 * @timeout: timeout value in jiffies
2908 * This waits for either a completion of a specific task to be
2909 * signaled or for a specified timeout to expire. It can be
2910 * interrupted by a kill signal. The timeout is in jiffies.
2912 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2913 * or number of jiffies left till timeout) if completed.
2916 wait_for_completion_killable_timeout(struct completion *x,
2917 unsigned long timeout)
2919 return wait_for_common(x, timeout, TASK_KILLABLE);
2921 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2924 * try_wait_for_completion - try to decrement a completion without blocking
2925 * @x: completion structure
2927 * Return: 0 if a decrement cannot be done without blocking
2928 * 1 if a decrement succeeded.
2930 * If a completion is being used as a counting completion,
2931 * attempt to decrement the counter without blocking. This
2932 * enables us to avoid waiting if the resource the completion
2933 * is protecting is not available.
2935 bool try_wait_for_completion(struct completion *x)
2937 unsigned long flags;
2940 spin_lock_irqsave(&x->wait.lock, flags);
2945 spin_unlock_irqrestore(&x->wait.lock, flags);
2948 EXPORT_SYMBOL(try_wait_for_completion);
2951 * completion_done - Test to see if a completion has any waiters
2952 * @x: completion structure
2954 * Return: 0 if there are waiters (wait_for_completion() in progress)
2955 * 1 if there are no waiters.
2958 bool completion_done(struct completion *x)
2960 unsigned long flags;
2963 spin_lock_irqsave(&x->wait.lock, flags);
2966 spin_unlock_irqrestore(&x->wait.lock, flags);
2969 EXPORT_SYMBOL(completion_done);
2972 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2974 unsigned long flags;
2977 init_waitqueue_entry(&wait, current);
2979 __set_current_state(state);
2981 spin_lock_irqsave(&q->lock, flags);
2982 __add_wait_queue(q, &wait);
2983 spin_unlock(&q->lock);
2984 timeout = schedule_timeout(timeout);
2985 spin_lock_irq(&q->lock);
2986 __remove_wait_queue(q, &wait);
2987 spin_unlock_irqrestore(&q->lock, flags);
2992 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2994 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2996 EXPORT_SYMBOL(interruptible_sleep_on);
2999 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3001 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3003 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3005 void __sched sleep_on(wait_queue_head_t *q)
3007 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3009 EXPORT_SYMBOL(sleep_on);
3011 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3013 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3015 EXPORT_SYMBOL(sleep_on_timeout);
3017 #ifdef CONFIG_RT_MUTEXES
3020 * rt_mutex_setprio - set the current priority of a task
3022 * @prio: prio value (kernel-internal form)
3024 * This function changes the 'effective' priority of a task. It does
3025 * not touch ->normal_prio like __setscheduler().
3027 * Used by the rt_mutex code to implement priority inheritance logic.
3029 void rt_mutex_setprio(struct task_struct *p, int prio)
3031 int oldprio, on_rq, running;
3033 const struct sched_class *prev_class;
3035 BUG_ON(prio < 0 || prio > MAX_PRIO);
3037 rq = __task_rq_lock(p);
3040 * Idle task boosting is a nono in general. There is one
3041 * exception, when PREEMPT_RT and NOHZ is active:
3043 * The idle task calls get_next_timer_interrupt() and holds
3044 * the timer wheel base->lock on the CPU and another CPU wants
3045 * to access the timer (probably to cancel it). We can safely
3046 * ignore the boosting request, as the idle CPU runs this code
3047 * with interrupts disabled and will complete the lock
3048 * protected section without being interrupted. So there is no
3049 * real need to boost.
3051 if (unlikely(p == rq->idle)) {
3052 WARN_ON(p != rq->curr);
3053 WARN_ON(p->pi_blocked_on);
3057 trace_sched_pi_setprio(p, prio);
3059 prev_class = p->sched_class;
3061 running = task_current(rq, p);
3063 dequeue_task(rq, p, 0);
3065 p->sched_class->put_prev_task(rq, p);
3068 p->sched_class = &rt_sched_class;
3070 p->sched_class = &fair_sched_class;
3075 p->sched_class->set_curr_task(rq);
3077 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3079 check_class_changed(rq, p, prev_class, oldprio);
3081 __task_rq_unlock(rq);
3084 void set_user_nice(struct task_struct *p, long nice)
3086 int old_prio, delta, on_rq;
3087 unsigned long flags;
3090 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3093 * We have to be careful, if called from sys_setpriority(),
3094 * the task might be in the middle of scheduling on another CPU.
3096 rq = task_rq_lock(p, &flags);
3098 * The RT priorities are set via sched_setscheduler(), but we still
3099 * allow the 'normal' nice value to be set - but as expected
3100 * it wont have any effect on scheduling until the task is
3101 * SCHED_FIFO/SCHED_RR:
3103 if (task_has_rt_policy(p)) {
3104 p->static_prio = NICE_TO_PRIO(nice);
3109 dequeue_task(rq, p, 0);
3111 p->static_prio = NICE_TO_PRIO(nice);
3114 p->prio = effective_prio(p);
3115 delta = p->prio - old_prio;
3118 enqueue_task(rq, p, 0);
3120 * If the task increased its priority or is running and
3121 * lowered its priority, then reschedule its CPU:
3123 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3124 resched_task(rq->curr);
3127 task_rq_unlock(rq, p, &flags);
3129 EXPORT_SYMBOL(set_user_nice);
3132 * can_nice - check if a task can reduce its nice value
3136 int can_nice(const struct task_struct *p, const int nice)
3138 /* convert nice value [19,-20] to rlimit style value [1,40] */
3139 int nice_rlim = 20 - nice;
3141 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3142 capable(CAP_SYS_NICE));
3145 #ifdef __ARCH_WANT_SYS_NICE
3148 * sys_nice - change the priority of the current process.
3149 * @increment: priority increment
3151 * sys_setpriority is a more generic, but much slower function that
3152 * does similar things.
3154 SYSCALL_DEFINE1(nice, int, increment)
3159 * Setpriority might change our priority at the same moment.
3160 * We don't have to worry. Conceptually one call occurs first
3161 * and we have a single winner.
3163 if (increment < -40)
3168 nice = TASK_NICE(current) + increment;
3174 if (increment < 0 && !can_nice(current, nice))
3177 retval = security_task_setnice(current, nice);
3181 set_user_nice(current, nice);
3188 * task_prio - return the priority value of a given task.
3189 * @p: the task in question.
3191 * Return: The priority value as seen by users in /proc.
3192 * RT tasks are offset by -200. Normal tasks are centered
3193 * around 0, value goes from -16 to +15.
3195 int task_prio(const struct task_struct *p)
3197 return p->prio - MAX_RT_PRIO;
3201 * task_nice - return the nice value of a given task.
3202 * @p: the task in question.
3204 * Return: The nice value [ -20 ... 0 ... 19 ].
3206 int task_nice(const struct task_struct *p)
3208 return TASK_NICE(p);
3210 EXPORT_SYMBOL(task_nice);
3213 * idle_cpu - is a given cpu idle currently?
3214 * @cpu: the processor in question.
3216 * Return: 1 if the CPU is currently idle. 0 otherwise.
3218 int idle_cpu(int cpu)
3220 struct rq *rq = cpu_rq(cpu);
3222 if (rq->curr != rq->idle)
3229 if (!llist_empty(&rq->wake_list))
3237 * idle_task - return the idle task for a given cpu.
3238 * @cpu: the processor in question.
3240 * Return: The idle task for the cpu @cpu.
3242 struct task_struct *idle_task(int cpu)
3244 return cpu_rq(cpu)->idle;
3248 * find_process_by_pid - find a process with a matching PID value.
3249 * @pid: the pid in question.
3251 * The task of @pid, if found. %NULL otherwise.
3253 static struct task_struct *find_process_by_pid(pid_t pid)
3255 return pid ? find_task_by_vpid(pid) : current;
3258 /* Actually do priority change: must hold rq lock. */
3260 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3263 p->rt_priority = prio;
3264 p->normal_prio = normal_prio(p);
3265 /* we are holding p->pi_lock already */
3266 p->prio = rt_mutex_getprio(p);
3267 if (rt_prio(p->prio))
3268 p->sched_class = &rt_sched_class;
3270 p->sched_class = &fair_sched_class;
3275 * check the target process has a UID that matches the current process's
3277 static bool check_same_owner(struct task_struct *p)
3279 const struct cred *cred = current_cred(), *pcred;
3283 pcred = __task_cred(p);
3284 match = (uid_eq(cred->euid, pcred->euid) ||
3285 uid_eq(cred->euid, pcred->uid));
3290 static int __sched_setscheduler(struct task_struct *p, int policy,
3291 const struct sched_param *param, bool user)
3293 int retval, oldprio, oldpolicy = -1, on_rq, running;
3294 unsigned long flags;
3295 const struct sched_class *prev_class;
3299 /* may grab non-irq protected spin_locks */
3300 BUG_ON(in_interrupt());
3302 /* double check policy once rq lock held */
3304 reset_on_fork = p->sched_reset_on_fork;
3305 policy = oldpolicy = p->policy;
3307 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3308 policy &= ~SCHED_RESET_ON_FORK;
3310 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3311 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3312 policy != SCHED_IDLE)
3317 * Valid priorities for SCHED_FIFO and SCHED_RR are
3318 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3319 * SCHED_BATCH and SCHED_IDLE is 0.
3321 if (param->sched_priority < 0 ||
3322 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3323 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3325 if (rt_policy(policy) != (param->sched_priority != 0))
3329 * Allow unprivileged RT tasks to decrease priority:
3331 if (user && !capable(CAP_SYS_NICE)) {
3332 if (rt_policy(policy)) {
3333 unsigned long rlim_rtprio =
3334 task_rlimit(p, RLIMIT_RTPRIO);
3336 /* can't set/change the rt policy */
3337 if (policy != p->policy && !rlim_rtprio)
3340 /* can't increase priority */
3341 if (param->sched_priority > p->rt_priority &&
3342 param->sched_priority > rlim_rtprio)
3347 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3348 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3350 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3351 if (!can_nice(p, TASK_NICE(p)))
3355 /* can't change other user's priorities */
3356 if (!check_same_owner(p))
3359 /* Normal users shall not reset the sched_reset_on_fork flag */
3360 if (p->sched_reset_on_fork && !reset_on_fork)
3365 retval = security_task_setscheduler(p);
3371 * make sure no PI-waiters arrive (or leave) while we are
3372 * changing the priority of the task:
3374 * To be able to change p->policy safely, the appropriate
3375 * runqueue lock must be held.
3377 rq = task_rq_lock(p, &flags);
3380 * Changing the policy of the stop threads its a very bad idea
3382 if (p == rq->stop) {
3383 task_rq_unlock(rq, p, &flags);
3388 * If not changing anything there's no need to proceed further:
3390 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3391 param->sched_priority == p->rt_priority))) {
3392 task_rq_unlock(rq, p, &flags);
3396 #ifdef CONFIG_RT_GROUP_SCHED
3399 * Do not allow realtime tasks into groups that have no runtime
3402 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3403 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3404 !task_group_is_autogroup(task_group(p))) {
3405 task_rq_unlock(rq, p, &flags);
3411 /* recheck policy now with rq lock held */
3412 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3413 policy = oldpolicy = -1;
3414 task_rq_unlock(rq, p, &flags);
3418 running = task_current(rq, p);
3420 dequeue_task(rq, p, 0);
3422 p->sched_class->put_prev_task(rq, p);
3424 p->sched_reset_on_fork = reset_on_fork;
3427 prev_class = p->sched_class;
3428 __setscheduler(rq, p, policy, param->sched_priority);
3431 p->sched_class->set_curr_task(rq);
3433 enqueue_task(rq, p, 0);
3435 check_class_changed(rq, p, prev_class, oldprio);
3436 task_rq_unlock(rq, p, &flags);
3438 rt_mutex_adjust_pi(p);
3444 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3445 * @p: the task in question.
3446 * @policy: new policy.
3447 * @param: structure containing the new RT priority.
3449 * Return: 0 on success. An error code otherwise.
3451 * NOTE that the task may be already dead.
3453 int sched_setscheduler(struct task_struct *p, int policy,
3454 const struct sched_param *param)
3456 return __sched_setscheduler(p, policy, param, true);
3458 EXPORT_SYMBOL_GPL(sched_setscheduler);
3461 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3462 * @p: the task in question.
3463 * @policy: new policy.
3464 * @param: structure containing the new RT priority.
3466 * Just like sched_setscheduler, only don't bother checking if the
3467 * current context has permission. For example, this is needed in
3468 * stop_machine(): we create temporary high priority worker threads,
3469 * but our caller might not have that capability.
3471 * Return: 0 on success. An error code otherwise.
3473 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3474 const struct sched_param *param)
3476 return __sched_setscheduler(p, policy, param, false);
3480 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3482 struct sched_param lparam;
3483 struct task_struct *p;
3486 if (!param || pid < 0)
3488 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3493 p = find_process_by_pid(pid);
3495 retval = sched_setscheduler(p, policy, &lparam);
3502 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3503 * @pid: the pid in question.
3504 * @policy: new policy.
3505 * @param: structure containing the new RT priority.
3507 * Return: 0 on success. An error code otherwise.
3509 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3510 struct sched_param __user *, param)
3512 /* negative values for policy are not valid */
3516 return do_sched_setscheduler(pid, policy, param);
3520 * sys_sched_setparam - set/change the RT priority of a thread
3521 * @pid: the pid in question.
3522 * @param: structure containing the new RT priority.
3524 * Return: 0 on success. An error code otherwise.
3526 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3528 return do_sched_setscheduler(pid, -1, param);
3532 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3533 * @pid: the pid in question.
3535 * Return: On success, the policy of the thread. Otherwise, a negative error
3538 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3540 struct task_struct *p;
3548 p = find_process_by_pid(pid);
3550 retval = security_task_getscheduler(p);
3553 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3560 * sys_sched_getparam - get the RT priority of a thread
3561 * @pid: the pid in question.
3562 * @param: structure containing the RT priority.
3564 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3567 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3569 struct sched_param lp;
3570 struct task_struct *p;
3573 if (!param || pid < 0)
3577 p = find_process_by_pid(pid);
3582 retval = security_task_getscheduler(p);
3586 lp.sched_priority = p->rt_priority;
3590 * This one might sleep, we cannot do it with a spinlock held ...
3592 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3601 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3603 cpumask_var_t cpus_allowed, new_mask;
3604 struct task_struct *p;
3610 p = find_process_by_pid(pid);
3617 /* Prevent p going away */
3621 if (p->flags & PF_NO_SETAFFINITY) {
3625 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3629 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3631 goto out_free_cpus_allowed;
3634 if (!check_same_owner(p)) {
3636 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3643 retval = security_task_setscheduler(p);
3647 cpuset_cpus_allowed(p, cpus_allowed);
3648 cpumask_and(new_mask, in_mask, cpus_allowed);
3650 retval = set_cpus_allowed_ptr(p, new_mask);
3653 cpuset_cpus_allowed(p, cpus_allowed);
3654 if (!cpumask_subset(new_mask, cpus_allowed)) {
3656 * We must have raced with a concurrent cpuset
3657 * update. Just reset the cpus_allowed to the
3658 * cpuset's cpus_allowed
3660 cpumask_copy(new_mask, cpus_allowed);
3665 free_cpumask_var(new_mask);
3666 out_free_cpus_allowed:
3667 free_cpumask_var(cpus_allowed);
3674 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3675 struct cpumask *new_mask)
3677 if (len < cpumask_size())
3678 cpumask_clear(new_mask);
3679 else if (len > cpumask_size())
3680 len = cpumask_size();
3682 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3686 * sys_sched_setaffinity - set the cpu affinity of a process
3687 * @pid: pid of the process
3688 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3689 * @user_mask_ptr: user-space pointer to the new cpu mask
3691 * Return: 0 on success. An error code otherwise.
3693 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3694 unsigned long __user *, user_mask_ptr)
3696 cpumask_var_t new_mask;
3699 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3702 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3704 retval = sched_setaffinity(pid, new_mask);
3705 free_cpumask_var(new_mask);
3709 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3711 struct task_struct *p;
3712 unsigned long flags;
3719 p = find_process_by_pid(pid);
3723 retval = security_task_getscheduler(p);
3727 raw_spin_lock_irqsave(&p->pi_lock, flags);
3728 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3729 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3739 * sys_sched_getaffinity - get the cpu affinity of a process
3740 * @pid: pid of the process
3741 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3742 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3744 * Return: 0 on success. An error code otherwise.
3746 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3747 unsigned long __user *, user_mask_ptr)
3752 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3754 if (len & (sizeof(unsigned long)-1))
3757 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3760 ret = sched_getaffinity(pid, mask);
3762 size_t retlen = min_t(size_t, len, cpumask_size());
3764 if (copy_to_user(user_mask_ptr, mask, retlen))
3769 free_cpumask_var(mask);
3775 * sys_sched_yield - yield the current processor to other threads.
3777 * This function yields the current CPU to other tasks. If there are no
3778 * other threads running on this CPU then this function will return.
3782 SYSCALL_DEFINE0(sched_yield)
3784 struct rq *rq = this_rq_lock();
3786 schedstat_inc(rq, yld_count);
3787 current->sched_class->yield_task(rq);
3790 * Since we are going to call schedule() anyway, there's
3791 * no need to preempt or enable interrupts:
3793 __release(rq->lock);
3794 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3795 do_raw_spin_unlock(&rq->lock);
3796 sched_preempt_enable_no_resched();
3803 static void __cond_resched(void)
3805 __preempt_count_add(PREEMPT_ACTIVE);
3807 __preempt_count_sub(PREEMPT_ACTIVE);
3810 int __sched _cond_resched(void)
3812 if (should_resched()) {
3818 EXPORT_SYMBOL(_cond_resched);
3821 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3822 * call schedule, and on return reacquire the lock.
3824 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3825 * operations here to prevent schedule() from being called twice (once via
3826 * spin_unlock(), once by hand).
3828 int __cond_resched_lock(spinlock_t *lock)
3830 int resched = should_resched();
3833 lockdep_assert_held(lock);
3835 if (spin_needbreak(lock) || resched) {
3846 EXPORT_SYMBOL(__cond_resched_lock);
3848 int __sched __cond_resched_softirq(void)
3850 BUG_ON(!in_softirq());
3852 if (should_resched()) {
3860 EXPORT_SYMBOL(__cond_resched_softirq);
3863 * yield - yield the current processor to other threads.
3865 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3867 * The scheduler is at all times free to pick the calling task as the most
3868 * eligible task to run, if removing the yield() call from your code breaks
3869 * it, its already broken.
3871 * Typical broken usage is:
3876 * where one assumes that yield() will let 'the other' process run that will
3877 * make event true. If the current task is a SCHED_FIFO task that will never
3878 * happen. Never use yield() as a progress guarantee!!
3880 * If you want to use yield() to wait for something, use wait_event().
3881 * If you want to use yield() to be 'nice' for others, use cond_resched().
3882 * If you still want to use yield(), do not!
3884 void __sched yield(void)
3886 set_current_state(TASK_RUNNING);
3889 EXPORT_SYMBOL(yield);
3892 * yield_to - yield the current processor to another thread in
3893 * your thread group, or accelerate that thread toward the
3894 * processor it's on.
3896 * @preempt: whether task preemption is allowed or not
3898 * It's the caller's job to ensure that the target task struct
3899 * can't go away on us before we can do any checks.
3902 * true (>0) if we indeed boosted the target task.
3903 * false (0) if we failed to boost the target.
3904 * -ESRCH if there's no task to yield to.
3906 bool __sched yield_to(struct task_struct *p, bool preempt)
3908 struct task_struct *curr = current;
3909 struct rq *rq, *p_rq;
3910 unsigned long flags;
3913 local_irq_save(flags);
3919 * If we're the only runnable task on the rq and target rq also
3920 * has only one task, there's absolutely no point in yielding.
3922 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3927 double_rq_lock(rq, p_rq);
3928 while (task_rq(p) != p_rq) {
3929 double_rq_unlock(rq, p_rq);
3933 if (!curr->sched_class->yield_to_task)
3936 if (curr->sched_class != p->sched_class)
3939 if (task_running(p_rq, p) || p->state)
3942 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3944 schedstat_inc(rq, yld_count);
3946 * Make p's CPU reschedule; pick_next_entity takes care of
3949 if (preempt && rq != p_rq)
3950 resched_task(p_rq->curr);
3954 double_rq_unlock(rq, p_rq);
3956 local_irq_restore(flags);
3963 EXPORT_SYMBOL_GPL(yield_to);
3966 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3967 * that process accounting knows that this is a task in IO wait state.
3969 void __sched io_schedule(void)
3971 struct rq *rq = raw_rq();
3973 delayacct_blkio_start();
3974 atomic_inc(&rq->nr_iowait);
3975 blk_flush_plug(current);
3976 current->in_iowait = 1;
3978 current->in_iowait = 0;
3979 atomic_dec(&rq->nr_iowait);
3980 delayacct_blkio_end();
3982 EXPORT_SYMBOL(io_schedule);
3984 long __sched io_schedule_timeout(long timeout)
3986 struct rq *rq = raw_rq();
3989 delayacct_blkio_start();
3990 atomic_inc(&rq->nr_iowait);
3991 blk_flush_plug(current);
3992 current->in_iowait = 1;
3993 ret = schedule_timeout(timeout);
3994 current->in_iowait = 0;
3995 atomic_dec(&rq->nr_iowait);
3996 delayacct_blkio_end();
4001 * sys_sched_get_priority_max - return maximum RT priority.
4002 * @policy: scheduling class.
4004 * Return: On success, this syscall returns the maximum
4005 * rt_priority that can be used by a given scheduling class.
4006 * On failure, a negative error code is returned.
4008 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4015 ret = MAX_USER_RT_PRIO-1;
4027 * sys_sched_get_priority_min - return minimum RT priority.
4028 * @policy: scheduling class.
4030 * Return: On success, this syscall returns the minimum
4031 * rt_priority that can be used by a given scheduling class.
4032 * On failure, a negative error code is returned.
4034 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4052 * sys_sched_rr_get_interval - return the default timeslice of a process.
4053 * @pid: pid of the process.
4054 * @interval: userspace pointer to the timeslice value.
4056 * this syscall writes the default timeslice value of a given process
4057 * into the user-space timespec buffer. A value of '0' means infinity.
4059 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4062 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4063 struct timespec __user *, interval)
4065 struct task_struct *p;
4066 unsigned int time_slice;
4067 unsigned long flags;
4077 p = find_process_by_pid(pid);
4081 retval = security_task_getscheduler(p);
4085 rq = task_rq_lock(p, &flags);
4086 time_slice = p->sched_class->get_rr_interval(rq, p);
4087 task_rq_unlock(rq, p, &flags);
4090 jiffies_to_timespec(time_slice, &t);
4091 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4099 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4101 void sched_show_task(struct task_struct *p)
4103 unsigned long free = 0;
4107 state = p->state ? __ffs(p->state) + 1 : 0;
4108 printk(KERN_INFO "%-15.15s %c", p->comm,
4109 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4110 #if BITS_PER_LONG == 32
4111 if (state == TASK_RUNNING)
4112 printk(KERN_CONT " running ");
4114 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4116 if (state == TASK_RUNNING)
4117 printk(KERN_CONT " running task ");
4119 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4121 #ifdef CONFIG_DEBUG_STACK_USAGE
4122 free = stack_not_used(p);
4125 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4127 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4128 task_pid_nr(p), ppid,
4129 (unsigned long)task_thread_info(p)->flags);
4131 print_worker_info(KERN_INFO, p);
4132 show_stack(p, NULL);
4135 void show_state_filter(unsigned long state_filter)
4137 struct task_struct *g, *p;
4139 #if BITS_PER_LONG == 32
4141 " task PC stack pid father\n");
4144 " task PC stack pid father\n");
4147 do_each_thread(g, p) {
4149 * reset the NMI-timeout, listing all files on a slow
4150 * console might take a lot of time:
4152 touch_nmi_watchdog();
4153 if (!state_filter || (p->state & state_filter))
4155 } while_each_thread(g, p);
4157 touch_all_softlockup_watchdogs();
4159 #ifdef CONFIG_SCHED_DEBUG
4160 sysrq_sched_debug_show();
4164 * Only show locks if all tasks are dumped:
4167 debug_show_all_locks();
4170 void init_idle_bootup_task(struct task_struct *idle)
4172 idle->sched_class = &idle_sched_class;
4176 * init_idle - set up an idle thread for a given CPU
4177 * @idle: task in question
4178 * @cpu: cpu the idle task belongs to
4180 * NOTE: this function does not set the idle thread's NEED_RESCHED
4181 * flag, to make booting more robust.
4183 void init_idle(struct task_struct *idle, int cpu)
4185 struct rq *rq = cpu_rq(cpu);
4186 unsigned long flags;
4188 raw_spin_lock_irqsave(&rq->lock, flags);
4191 idle->state = TASK_RUNNING;
4192 idle->se.exec_start = sched_clock();
4194 do_set_cpus_allowed(idle, cpumask_of(cpu));
4196 * We're having a chicken and egg problem, even though we are
4197 * holding rq->lock, the cpu isn't yet set to this cpu so the
4198 * lockdep check in task_group() will fail.
4200 * Similar case to sched_fork(). / Alternatively we could
4201 * use task_rq_lock() here and obtain the other rq->lock.
4206 __set_task_cpu(idle, cpu);
4209 rq->curr = rq->idle = idle;
4210 #if defined(CONFIG_SMP)
4213 raw_spin_unlock_irqrestore(&rq->lock, flags);
4215 /* Set the preempt count _outside_ the spinlocks! */
4216 init_idle_preempt_count(idle, cpu);
4219 * The idle tasks have their own, simple scheduling class:
4221 idle->sched_class = &idle_sched_class;
4222 ftrace_graph_init_idle_task(idle, cpu);
4223 vtime_init_idle(idle, cpu);
4224 #if defined(CONFIG_SMP)
4225 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4230 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4232 if (p->sched_class && p->sched_class->set_cpus_allowed)
4233 p->sched_class->set_cpus_allowed(p, new_mask);
4235 cpumask_copy(&p->cpus_allowed, new_mask);
4236 p->nr_cpus_allowed = cpumask_weight(new_mask);
4240 * This is how migration works:
4242 * 1) we invoke migration_cpu_stop() on the target CPU using
4244 * 2) stopper starts to run (implicitly forcing the migrated thread
4246 * 3) it checks whether the migrated task is still in the wrong runqueue.
4247 * 4) if it's in the wrong runqueue then the migration thread removes
4248 * it and puts it into the right queue.
4249 * 5) stopper completes and stop_one_cpu() returns and the migration
4254 * Change a given task's CPU affinity. Migrate the thread to a
4255 * proper CPU and schedule it away if the CPU it's executing on
4256 * is removed from the allowed bitmask.
4258 * NOTE: the caller must have a valid reference to the task, the
4259 * task must not exit() & deallocate itself prematurely. The
4260 * call is not atomic; no spinlocks may be held.
4262 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4264 unsigned long flags;
4266 unsigned int dest_cpu;
4269 rq = task_rq_lock(p, &flags);
4271 if (cpumask_equal(&p->cpus_allowed, new_mask))
4274 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4279 do_set_cpus_allowed(p, new_mask);
4281 /* Can the task run on the task's current CPU? If so, we're done */
4282 if (cpumask_test_cpu(task_cpu(p), new_mask))
4285 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4287 struct migration_arg arg = { p, dest_cpu };
4288 /* Need help from migration thread: drop lock and wait. */
4289 task_rq_unlock(rq, p, &flags);
4290 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4291 tlb_migrate_finish(p->mm);
4295 task_rq_unlock(rq, p, &flags);
4299 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4302 * Move (not current) task off this cpu, onto dest cpu. We're doing
4303 * this because either it can't run here any more (set_cpus_allowed()
4304 * away from this CPU, or CPU going down), or because we're
4305 * attempting to rebalance this task on exec (sched_exec).
4307 * So we race with normal scheduler movements, but that's OK, as long
4308 * as the task is no longer on this CPU.
4310 * Returns non-zero if task was successfully migrated.
4312 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4314 struct rq *rq_dest, *rq_src;
4317 if (unlikely(!cpu_active(dest_cpu)))
4320 rq_src = cpu_rq(src_cpu);
4321 rq_dest = cpu_rq(dest_cpu);
4323 raw_spin_lock(&p->pi_lock);
4324 double_rq_lock(rq_src, rq_dest);
4325 /* Already moved. */
4326 if (task_cpu(p) != src_cpu)
4328 /* Affinity changed (again). */
4329 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4333 * If we're not on a rq, the next wake-up will ensure we're
4337 dequeue_task(rq_src, p, 0);
4338 set_task_cpu(p, dest_cpu);
4339 enqueue_task(rq_dest, p, 0);
4340 check_preempt_curr(rq_dest, p, 0);
4345 double_rq_unlock(rq_src, rq_dest);
4346 raw_spin_unlock(&p->pi_lock);
4351 * migration_cpu_stop - this will be executed by a highprio stopper thread
4352 * and performs thread migration by bumping thread off CPU then
4353 * 'pushing' onto another runqueue.
4355 static int migration_cpu_stop(void *data)
4357 struct migration_arg *arg = data;
4360 * The original target cpu might have gone down and we might
4361 * be on another cpu but it doesn't matter.
4363 local_irq_disable();
4364 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4369 #ifdef CONFIG_HOTPLUG_CPU
4372 * Ensures that the idle task is using init_mm right before its cpu goes
4375 void idle_task_exit(void)
4377 struct mm_struct *mm = current->active_mm;
4379 BUG_ON(cpu_online(smp_processor_id()));
4382 switch_mm(mm, &init_mm, current);
4387 * Since this CPU is going 'away' for a while, fold any nr_active delta
4388 * we might have. Assumes we're called after migrate_tasks() so that the
4389 * nr_active count is stable.
4391 * Also see the comment "Global load-average calculations".
4393 static void calc_load_migrate(struct rq *rq)
4395 long delta = calc_load_fold_active(rq);
4397 atomic_long_add(delta, &calc_load_tasks);
4401 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4402 * try_to_wake_up()->select_task_rq().
4404 * Called with rq->lock held even though we'er in stop_machine() and
4405 * there's no concurrency possible, we hold the required locks anyway
4406 * because of lock validation efforts.
4408 static void migrate_tasks(unsigned int dead_cpu)
4410 struct rq *rq = cpu_rq(dead_cpu);
4411 struct task_struct *next, *stop = rq->stop;
4415 * Fudge the rq selection such that the below task selection loop
4416 * doesn't get stuck on the currently eligible stop task.
4418 * We're currently inside stop_machine() and the rq is either stuck
4419 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4420 * either way we should never end up calling schedule() until we're
4426 * put_prev_task() and pick_next_task() sched
4427 * class method both need to have an up-to-date
4428 * value of rq->clock[_task]
4430 update_rq_clock(rq);
4434 * There's this thread running, bail when that's the only
4437 if (rq->nr_running == 1)
4440 next = pick_next_task(rq);
4442 next->sched_class->put_prev_task(rq, next);
4444 /* Find suitable destination for @next, with force if needed. */
4445 dest_cpu = select_fallback_rq(dead_cpu, next);
4446 raw_spin_unlock(&rq->lock);
4448 __migrate_task(next, dead_cpu, dest_cpu);
4450 raw_spin_lock(&rq->lock);
4456 #endif /* CONFIG_HOTPLUG_CPU */
4458 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4460 static struct ctl_table sd_ctl_dir[] = {
4462 .procname = "sched_domain",
4468 static struct ctl_table sd_ctl_root[] = {
4470 .procname = "kernel",
4472 .child = sd_ctl_dir,
4477 static struct ctl_table *sd_alloc_ctl_entry(int n)
4479 struct ctl_table *entry =
4480 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4485 static void sd_free_ctl_entry(struct ctl_table **tablep)
4487 struct ctl_table *entry;
4490 * In the intermediate directories, both the child directory and
4491 * procname are dynamically allocated and could fail but the mode
4492 * will always be set. In the lowest directory the names are
4493 * static strings and all have proc handlers.
4495 for (entry = *tablep; entry->mode; entry++) {
4497 sd_free_ctl_entry(&entry->child);
4498 if (entry->proc_handler == NULL)
4499 kfree(entry->procname);
4506 static int min_load_idx = 0;
4507 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4510 set_table_entry(struct ctl_table *entry,
4511 const char *procname, void *data, int maxlen,
4512 umode_t mode, proc_handler *proc_handler,
4515 entry->procname = procname;
4517 entry->maxlen = maxlen;
4519 entry->proc_handler = proc_handler;
4522 entry->extra1 = &min_load_idx;
4523 entry->extra2 = &max_load_idx;
4527 static struct ctl_table *
4528 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4530 struct ctl_table *table = sd_alloc_ctl_entry(13);
4535 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4536 sizeof(long), 0644, proc_doulongvec_minmax, false);
4537 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4538 sizeof(long), 0644, proc_doulongvec_minmax, false);
4539 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4540 sizeof(int), 0644, proc_dointvec_minmax, true);
4541 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4542 sizeof(int), 0644, proc_dointvec_minmax, true);
4543 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4544 sizeof(int), 0644, proc_dointvec_minmax, true);
4545 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4546 sizeof(int), 0644, proc_dointvec_minmax, true);
4547 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4548 sizeof(int), 0644, proc_dointvec_minmax, true);
4549 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4550 sizeof(int), 0644, proc_dointvec_minmax, false);
4551 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4552 sizeof(int), 0644, proc_dointvec_minmax, false);
4553 set_table_entry(&table[9], "cache_nice_tries",
4554 &sd->cache_nice_tries,
4555 sizeof(int), 0644, proc_dointvec_minmax, false);
4556 set_table_entry(&table[10], "flags", &sd->flags,
4557 sizeof(int), 0644, proc_dointvec_minmax, false);
4558 set_table_entry(&table[11], "name", sd->name,
4559 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4560 /* &table[12] is terminator */
4565 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4567 struct ctl_table *entry, *table;
4568 struct sched_domain *sd;
4569 int domain_num = 0, i;
4572 for_each_domain(cpu, sd)
4574 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4579 for_each_domain(cpu, sd) {
4580 snprintf(buf, 32, "domain%d", i);
4581 entry->procname = kstrdup(buf, GFP_KERNEL);
4583 entry->child = sd_alloc_ctl_domain_table(sd);
4590 static struct ctl_table_header *sd_sysctl_header;
4591 static void register_sched_domain_sysctl(void)
4593 int i, cpu_num = num_possible_cpus();
4594 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4597 WARN_ON(sd_ctl_dir[0].child);
4598 sd_ctl_dir[0].child = entry;
4603 for_each_possible_cpu(i) {
4604 snprintf(buf, 32, "cpu%d", i);
4605 entry->procname = kstrdup(buf, GFP_KERNEL);
4607 entry->child = sd_alloc_ctl_cpu_table(i);
4611 WARN_ON(sd_sysctl_header);
4612 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4615 /* may be called multiple times per register */
4616 static void unregister_sched_domain_sysctl(void)
4618 if (sd_sysctl_header)
4619 unregister_sysctl_table(sd_sysctl_header);
4620 sd_sysctl_header = NULL;
4621 if (sd_ctl_dir[0].child)
4622 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4625 static void register_sched_domain_sysctl(void)
4628 static void unregister_sched_domain_sysctl(void)
4633 static void set_rq_online(struct rq *rq)
4636 const struct sched_class *class;
4638 cpumask_set_cpu(rq->cpu, rq->rd->online);
4641 for_each_class(class) {
4642 if (class->rq_online)
4643 class->rq_online(rq);
4648 static void set_rq_offline(struct rq *rq)
4651 const struct sched_class *class;
4653 for_each_class(class) {
4654 if (class->rq_offline)
4655 class->rq_offline(rq);
4658 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4664 * migration_call - callback that gets triggered when a CPU is added.
4665 * Here we can start up the necessary migration thread for the new CPU.
4668 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4670 int cpu = (long)hcpu;
4671 unsigned long flags;
4672 struct rq *rq = cpu_rq(cpu);
4674 switch (action & ~CPU_TASKS_FROZEN) {
4676 case CPU_UP_PREPARE:
4677 rq->calc_load_update = calc_load_update;
4681 /* Update our root-domain */
4682 raw_spin_lock_irqsave(&rq->lock, flags);
4684 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4688 raw_spin_unlock_irqrestore(&rq->lock, flags);
4691 #ifdef CONFIG_HOTPLUG_CPU
4693 sched_ttwu_pending();
4694 /* Update our root-domain */
4695 raw_spin_lock_irqsave(&rq->lock, flags);
4697 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4701 BUG_ON(rq->nr_running != 1); /* the migration thread */
4702 raw_spin_unlock_irqrestore(&rq->lock, flags);
4706 calc_load_migrate(rq);
4711 update_max_interval();
4717 * Register at high priority so that task migration (migrate_all_tasks)
4718 * happens before everything else. This has to be lower priority than
4719 * the notifier in the perf_event subsystem, though.
4721 static struct notifier_block migration_notifier = {
4722 .notifier_call = migration_call,
4723 .priority = CPU_PRI_MIGRATION,
4726 static int sched_cpu_active(struct notifier_block *nfb,
4727 unsigned long action, void *hcpu)
4729 switch (action & ~CPU_TASKS_FROZEN) {
4731 case CPU_DOWN_FAILED:
4732 set_cpu_active((long)hcpu, true);
4739 static int sched_cpu_inactive(struct notifier_block *nfb,
4740 unsigned long action, void *hcpu)
4742 switch (action & ~CPU_TASKS_FROZEN) {
4743 case CPU_DOWN_PREPARE:
4744 set_cpu_active((long)hcpu, false);
4751 static int __init migration_init(void)
4753 void *cpu = (void *)(long)smp_processor_id();
4756 /* Initialize migration for the boot CPU */
4757 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4758 BUG_ON(err == NOTIFY_BAD);
4759 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4760 register_cpu_notifier(&migration_notifier);
4762 /* Register cpu active notifiers */
4763 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4764 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4768 early_initcall(migration_init);
4773 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4775 #ifdef CONFIG_SCHED_DEBUG
4777 static __read_mostly int sched_debug_enabled;
4779 static int __init sched_debug_setup(char *str)
4781 sched_debug_enabled = 1;
4785 early_param("sched_debug", sched_debug_setup);
4787 static inline bool sched_debug(void)
4789 return sched_debug_enabled;
4792 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4793 struct cpumask *groupmask)
4795 struct sched_group *group = sd->groups;
4798 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4799 cpumask_clear(groupmask);
4801 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4803 if (!(sd->flags & SD_LOAD_BALANCE)) {
4804 printk("does not load-balance\n");
4806 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4811 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4813 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4814 printk(KERN_ERR "ERROR: domain->span does not contain "
4817 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4818 printk(KERN_ERR "ERROR: domain->groups does not contain"
4822 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4826 printk(KERN_ERR "ERROR: group is NULL\n");
4831 * Even though we initialize ->power to something semi-sane,
4832 * we leave power_orig unset. This allows us to detect if
4833 * domain iteration is still funny without causing /0 traps.
4835 if (!group->sgp->power_orig) {
4836 printk(KERN_CONT "\n");
4837 printk(KERN_ERR "ERROR: domain->cpu_power not "
4842 if (!cpumask_weight(sched_group_cpus(group))) {
4843 printk(KERN_CONT "\n");
4844 printk(KERN_ERR "ERROR: empty group\n");
4848 if (!(sd->flags & SD_OVERLAP) &&
4849 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4850 printk(KERN_CONT "\n");
4851 printk(KERN_ERR "ERROR: repeated CPUs\n");
4855 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4857 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4859 printk(KERN_CONT " %s", str);
4860 if (group->sgp->power != SCHED_POWER_SCALE) {
4861 printk(KERN_CONT " (cpu_power = %d)",
4865 group = group->next;
4866 } while (group != sd->groups);
4867 printk(KERN_CONT "\n");
4869 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4870 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4873 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4874 printk(KERN_ERR "ERROR: parent span is not a superset "
4875 "of domain->span\n");
4879 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4883 if (!sched_debug_enabled)
4887 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4891 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4894 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4902 #else /* !CONFIG_SCHED_DEBUG */
4903 # define sched_domain_debug(sd, cpu) do { } while (0)
4904 static inline bool sched_debug(void)
4908 #endif /* CONFIG_SCHED_DEBUG */
4910 static int sd_degenerate(struct sched_domain *sd)
4912 if (cpumask_weight(sched_domain_span(sd)) == 1)
4915 /* Following flags need at least 2 groups */
4916 if (sd->flags & (SD_LOAD_BALANCE |
4917 SD_BALANCE_NEWIDLE |
4921 SD_SHARE_PKG_RESOURCES)) {
4922 if (sd->groups != sd->groups->next)
4926 /* Following flags don't use groups */
4927 if (sd->flags & (SD_WAKE_AFFINE))
4934 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4936 unsigned long cflags = sd->flags, pflags = parent->flags;
4938 if (sd_degenerate(parent))
4941 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4944 /* Flags needing groups don't count if only 1 group in parent */
4945 if (parent->groups == parent->groups->next) {
4946 pflags &= ~(SD_LOAD_BALANCE |
4947 SD_BALANCE_NEWIDLE |
4951 SD_SHARE_PKG_RESOURCES |
4953 if (nr_node_ids == 1)
4954 pflags &= ~SD_SERIALIZE;
4956 if (~cflags & pflags)
4962 static void free_rootdomain(struct rcu_head *rcu)
4964 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4966 cpupri_cleanup(&rd->cpupri);
4967 free_cpumask_var(rd->rto_mask);
4968 free_cpumask_var(rd->online);
4969 free_cpumask_var(rd->span);
4973 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4975 struct root_domain *old_rd = NULL;
4976 unsigned long flags;
4978 raw_spin_lock_irqsave(&rq->lock, flags);
4983 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4986 cpumask_clear_cpu(rq->cpu, old_rd->span);
4989 * If we dont want to free the old_rt yet then
4990 * set old_rd to NULL to skip the freeing later
4993 if (!atomic_dec_and_test(&old_rd->refcount))
4997 atomic_inc(&rd->refcount);
5000 cpumask_set_cpu(rq->cpu, rd->span);
5001 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5004 raw_spin_unlock_irqrestore(&rq->lock, flags);
5007 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5010 static int init_rootdomain(struct root_domain *rd)
5012 memset(rd, 0, sizeof(*rd));
5014 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5016 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5018 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5021 if (cpupri_init(&rd->cpupri) != 0)
5026 free_cpumask_var(rd->rto_mask);
5028 free_cpumask_var(rd->online);
5030 free_cpumask_var(rd->span);
5036 * By default the system creates a single root-domain with all cpus as
5037 * members (mimicking the global state we have today).
5039 struct root_domain def_root_domain;
5041 static void init_defrootdomain(void)
5043 init_rootdomain(&def_root_domain);
5045 atomic_set(&def_root_domain.refcount, 1);
5048 static struct root_domain *alloc_rootdomain(void)
5050 struct root_domain *rd;
5052 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5056 if (init_rootdomain(rd) != 0) {
5064 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5066 struct sched_group *tmp, *first;
5075 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5080 } while (sg != first);
5083 static void free_sched_domain(struct rcu_head *rcu)
5085 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5088 * If its an overlapping domain it has private groups, iterate and
5091 if (sd->flags & SD_OVERLAP) {
5092 free_sched_groups(sd->groups, 1);
5093 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5094 kfree(sd->groups->sgp);
5100 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5102 call_rcu(&sd->rcu, free_sched_domain);
5105 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5107 for (; sd; sd = sd->parent)
5108 destroy_sched_domain(sd, cpu);
5112 * Keep a special pointer to the highest sched_domain that has
5113 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5114 * allows us to avoid some pointer chasing select_idle_sibling().
5116 * Also keep a unique ID per domain (we use the first cpu number in
5117 * the cpumask of the domain), this allows us to quickly tell if
5118 * two cpus are in the same cache domain, see cpus_share_cache().
5120 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5121 DEFINE_PER_CPU(int, sd_llc_size);
5122 DEFINE_PER_CPU(int, sd_llc_id);
5124 static void update_top_cache_domain(int cpu)
5126 struct sched_domain *sd;
5130 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5132 id = cpumask_first(sched_domain_span(sd));
5133 size = cpumask_weight(sched_domain_span(sd));
5136 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5137 per_cpu(sd_llc_size, cpu) = size;
5138 per_cpu(sd_llc_id, cpu) = id;
5142 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5143 * hold the hotplug lock.
5146 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5148 struct rq *rq = cpu_rq(cpu);
5149 struct sched_domain *tmp;
5151 /* Remove the sched domains which do not contribute to scheduling. */
5152 for (tmp = sd; tmp; ) {
5153 struct sched_domain *parent = tmp->parent;
5157 if (sd_parent_degenerate(tmp, parent)) {
5158 tmp->parent = parent->parent;
5160 parent->parent->child = tmp;
5162 * Transfer SD_PREFER_SIBLING down in case of a
5163 * degenerate parent; the spans match for this
5164 * so the property transfers.
5166 if (parent->flags & SD_PREFER_SIBLING)
5167 tmp->flags |= SD_PREFER_SIBLING;
5168 destroy_sched_domain(parent, cpu);
5173 if (sd && sd_degenerate(sd)) {
5176 destroy_sched_domain(tmp, cpu);
5181 sched_domain_debug(sd, cpu);
5183 rq_attach_root(rq, rd);
5185 rcu_assign_pointer(rq->sd, sd);
5186 destroy_sched_domains(tmp, cpu);
5188 update_top_cache_domain(cpu);
5191 /* cpus with isolated domains */
5192 static cpumask_var_t cpu_isolated_map;
5194 /* Setup the mask of cpus configured for isolated domains */
5195 static int __init isolated_cpu_setup(char *str)
5197 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5198 cpulist_parse(str, cpu_isolated_map);
5202 __setup("isolcpus=", isolated_cpu_setup);
5204 static const struct cpumask *cpu_cpu_mask(int cpu)
5206 return cpumask_of_node(cpu_to_node(cpu));
5210 struct sched_domain **__percpu sd;
5211 struct sched_group **__percpu sg;
5212 struct sched_group_power **__percpu sgp;
5216 struct sched_domain ** __percpu sd;
5217 struct root_domain *rd;
5227 struct sched_domain_topology_level;
5229 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5230 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5232 #define SDTL_OVERLAP 0x01
5234 struct sched_domain_topology_level {
5235 sched_domain_init_f init;
5236 sched_domain_mask_f mask;
5239 struct sd_data data;
5243 * Build an iteration mask that can exclude certain CPUs from the upwards
5246 * Asymmetric node setups can result in situations where the domain tree is of
5247 * unequal depth, make sure to skip domains that already cover the entire
5250 * In that case build_sched_domains() will have terminated the iteration early
5251 * and our sibling sd spans will be empty. Domains should always include the
5252 * cpu they're built on, so check that.
5255 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5257 const struct cpumask *span = sched_domain_span(sd);
5258 struct sd_data *sdd = sd->private;
5259 struct sched_domain *sibling;
5262 for_each_cpu(i, span) {
5263 sibling = *per_cpu_ptr(sdd->sd, i);
5264 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5267 cpumask_set_cpu(i, sched_group_mask(sg));
5272 * Return the canonical balance cpu for this group, this is the first cpu
5273 * of this group that's also in the iteration mask.
5275 int group_balance_cpu(struct sched_group *sg)
5277 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5281 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5283 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5284 const struct cpumask *span = sched_domain_span(sd);
5285 struct cpumask *covered = sched_domains_tmpmask;
5286 struct sd_data *sdd = sd->private;
5287 struct sched_domain *child;
5290 cpumask_clear(covered);
5292 for_each_cpu(i, span) {
5293 struct cpumask *sg_span;
5295 if (cpumask_test_cpu(i, covered))
5298 child = *per_cpu_ptr(sdd->sd, i);
5300 /* See the comment near build_group_mask(). */
5301 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5304 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5305 GFP_KERNEL, cpu_to_node(cpu));
5310 sg_span = sched_group_cpus(sg);
5312 child = child->child;
5313 cpumask_copy(sg_span, sched_domain_span(child));
5315 cpumask_set_cpu(i, sg_span);
5317 cpumask_or(covered, covered, sg_span);
5319 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5320 if (atomic_inc_return(&sg->sgp->ref) == 1)
5321 build_group_mask(sd, sg);
5324 * Initialize sgp->power such that even if we mess up the
5325 * domains and no possible iteration will get us here, we won't
5328 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5331 * Make sure the first group of this domain contains the
5332 * canonical balance cpu. Otherwise the sched_domain iteration
5333 * breaks. See update_sg_lb_stats().
5335 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5336 group_balance_cpu(sg) == cpu)
5346 sd->groups = groups;
5351 free_sched_groups(first, 0);
5356 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5358 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5359 struct sched_domain *child = sd->child;
5362 cpu = cpumask_first(sched_domain_span(child));
5365 *sg = *per_cpu_ptr(sdd->sg, cpu);
5366 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5367 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5374 * build_sched_groups will build a circular linked list of the groups
5375 * covered by the given span, and will set each group's ->cpumask correctly,
5376 * and ->cpu_power to 0.
5378 * Assumes the sched_domain tree is fully constructed
5381 build_sched_groups(struct sched_domain *sd, int cpu)
5383 struct sched_group *first = NULL, *last = NULL;
5384 struct sd_data *sdd = sd->private;
5385 const struct cpumask *span = sched_domain_span(sd);
5386 struct cpumask *covered;
5389 get_group(cpu, sdd, &sd->groups);
5390 atomic_inc(&sd->groups->ref);
5392 if (cpu != cpumask_first(span))
5395 lockdep_assert_held(&sched_domains_mutex);
5396 covered = sched_domains_tmpmask;
5398 cpumask_clear(covered);
5400 for_each_cpu(i, span) {
5401 struct sched_group *sg;
5404 if (cpumask_test_cpu(i, covered))
5407 group = get_group(i, sdd, &sg);
5408 cpumask_clear(sched_group_cpus(sg));
5410 cpumask_setall(sched_group_mask(sg));
5412 for_each_cpu(j, span) {
5413 if (get_group(j, sdd, NULL) != group)
5416 cpumask_set_cpu(j, covered);
5417 cpumask_set_cpu(j, sched_group_cpus(sg));
5432 * Initialize sched groups cpu_power.
5434 * cpu_power indicates the capacity of sched group, which is used while
5435 * distributing the load between different sched groups in a sched domain.
5436 * Typically cpu_power for all the groups in a sched domain will be same unless
5437 * there are asymmetries in the topology. If there are asymmetries, group
5438 * having more cpu_power will pickup more load compared to the group having
5441 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5443 struct sched_group *sg = sd->groups;
5448 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5450 } while (sg != sd->groups);
5452 if (cpu != group_balance_cpu(sg))
5455 update_group_power(sd, cpu);
5456 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5459 int __weak arch_sd_sibling_asym_packing(void)
5461 return 0*SD_ASYM_PACKING;
5465 * Initializers for schedule domains
5466 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5469 #ifdef CONFIG_SCHED_DEBUG
5470 # define SD_INIT_NAME(sd, type) sd->name = #type
5472 # define SD_INIT_NAME(sd, type) do { } while (0)
5475 #define SD_INIT_FUNC(type) \
5476 static noinline struct sched_domain * \
5477 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5479 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5480 *sd = SD_##type##_INIT; \
5481 SD_INIT_NAME(sd, type); \
5482 sd->private = &tl->data; \
5487 #ifdef CONFIG_SCHED_SMT
5488 SD_INIT_FUNC(SIBLING)
5490 #ifdef CONFIG_SCHED_MC
5493 #ifdef CONFIG_SCHED_BOOK
5497 static int default_relax_domain_level = -1;
5498 int sched_domain_level_max;
5500 static int __init setup_relax_domain_level(char *str)
5502 if (kstrtoint(str, 0, &default_relax_domain_level))
5503 pr_warn("Unable to set relax_domain_level\n");
5507 __setup("relax_domain_level=", setup_relax_domain_level);
5509 static void set_domain_attribute(struct sched_domain *sd,
5510 struct sched_domain_attr *attr)
5514 if (!attr || attr->relax_domain_level < 0) {
5515 if (default_relax_domain_level < 0)
5518 request = default_relax_domain_level;
5520 request = attr->relax_domain_level;
5521 if (request < sd->level) {
5522 /* turn off idle balance on this domain */
5523 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5525 /* turn on idle balance on this domain */
5526 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5530 static void __sdt_free(const struct cpumask *cpu_map);
5531 static int __sdt_alloc(const struct cpumask *cpu_map);
5533 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5534 const struct cpumask *cpu_map)
5538 if (!atomic_read(&d->rd->refcount))
5539 free_rootdomain(&d->rd->rcu); /* fall through */
5541 free_percpu(d->sd); /* fall through */
5543 __sdt_free(cpu_map); /* fall through */
5549 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5550 const struct cpumask *cpu_map)
5552 memset(d, 0, sizeof(*d));
5554 if (__sdt_alloc(cpu_map))
5555 return sa_sd_storage;
5556 d->sd = alloc_percpu(struct sched_domain *);
5558 return sa_sd_storage;
5559 d->rd = alloc_rootdomain();
5562 return sa_rootdomain;
5566 * NULL the sd_data elements we've used to build the sched_domain and
5567 * sched_group structure so that the subsequent __free_domain_allocs()
5568 * will not free the data we're using.
5570 static void claim_allocations(int cpu, struct sched_domain *sd)
5572 struct sd_data *sdd = sd->private;
5574 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5575 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5577 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5578 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5580 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5581 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5584 #ifdef CONFIG_SCHED_SMT
5585 static const struct cpumask *cpu_smt_mask(int cpu)
5587 return topology_thread_cpumask(cpu);
5592 * Topology list, bottom-up.
5594 static struct sched_domain_topology_level default_topology[] = {
5595 #ifdef CONFIG_SCHED_SMT
5596 { sd_init_SIBLING, cpu_smt_mask, },
5598 #ifdef CONFIG_SCHED_MC
5599 { sd_init_MC, cpu_coregroup_mask, },
5601 #ifdef CONFIG_SCHED_BOOK
5602 { sd_init_BOOK, cpu_book_mask, },
5604 { sd_init_CPU, cpu_cpu_mask, },
5608 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5610 #define for_each_sd_topology(tl) \
5611 for (tl = sched_domain_topology; tl->init; tl++)
5615 static int sched_domains_numa_levels;
5616 static int *sched_domains_numa_distance;
5617 static struct cpumask ***sched_domains_numa_masks;
5618 static int sched_domains_curr_level;
5620 static inline int sd_local_flags(int level)
5622 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5625 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5628 static struct sched_domain *
5629 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5631 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5632 int level = tl->numa_level;
5633 int sd_weight = cpumask_weight(
5634 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5636 *sd = (struct sched_domain){
5637 .min_interval = sd_weight,
5638 .max_interval = 2*sd_weight,
5640 .imbalance_pct = 125,
5641 .cache_nice_tries = 2,
5648 .flags = 1*SD_LOAD_BALANCE
5649 | 1*SD_BALANCE_NEWIDLE
5654 | 0*SD_SHARE_CPUPOWER
5655 | 0*SD_SHARE_PKG_RESOURCES
5657 | 0*SD_PREFER_SIBLING
5658 | sd_local_flags(level)
5660 .last_balance = jiffies,
5661 .balance_interval = sd_weight,
5663 SD_INIT_NAME(sd, NUMA);
5664 sd->private = &tl->data;
5667 * Ugly hack to pass state to sd_numa_mask()...
5669 sched_domains_curr_level = tl->numa_level;
5674 static const struct cpumask *sd_numa_mask(int cpu)
5676 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5679 static void sched_numa_warn(const char *str)
5681 static int done = false;
5689 printk(KERN_WARNING "ERROR: %s\n\n", str);
5691 for (i = 0; i < nr_node_ids; i++) {
5692 printk(KERN_WARNING " ");
5693 for (j = 0; j < nr_node_ids; j++)
5694 printk(KERN_CONT "%02d ", node_distance(i,j));
5695 printk(KERN_CONT "\n");
5697 printk(KERN_WARNING "\n");
5700 static bool find_numa_distance(int distance)
5704 if (distance == node_distance(0, 0))
5707 for (i = 0; i < sched_domains_numa_levels; i++) {
5708 if (sched_domains_numa_distance[i] == distance)
5715 static void sched_init_numa(void)
5717 int next_distance, curr_distance = node_distance(0, 0);
5718 struct sched_domain_topology_level *tl;
5722 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5723 if (!sched_domains_numa_distance)
5727 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5728 * unique distances in the node_distance() table.
5730 * Assumes node_distance(0,j) includes all distances in
5731 * node_distance(i,j) in order to avoid cubic time.
5733 next_distance = curr_distance;
5734 for (i = 0; i < nr_node_ids; i++) {
5735 for (j = 0; j < nr_node_ids; j++) {
5736 for (k = 0; k < nr_node_ids; k++) {
5737 int distance = node_distance(i, k);
5739 if (distance > curr_distance &&
5740 (distance < next_distance ||
5741 next_distance == curr_distance))
5742 next_distance = distance;
5745 * While not a strong assumption it would be nice to know
5746 * about cases where if node A is connected to B, B is not
5747 * equally connected to A.
5749 if (sched_debug() && node_distance(k, i) != distance)
5750 sched_numa_warn("Node-distance not symmetric");
5752 if (sched_debug() && i && !find_numa_distance(distance))
5753 sched_numa_warn("Node-0 not representative");
5755 if (next_distance != curr_distance) {
5756 sched_domains_numa_distance[level++] = next_distance;
5757 sched_domains_numa_levels = level;
5758 curr_distance = next_distance;
5763 * In case of sched_debug() we verify the above assumption.
5769 * 'level' contains the number of unique distances, excluding the
5770 * identity distance node_distance(i,i).
5772 * The sched_domains_numa_distance[] array includes the actual distance
5777 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5778 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5779 * the array will contain less then 'level' members. This could be
5780 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5781 * in other functions.
5783 * We reset it to 'level' at the end of this function.
5785 sched_domains_numa_levels = 0;
5787 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5788 if (!sched_domains_numa_masks)
5792 * Now for each level, construct a mask per node which contains all
5793 * cpus of nodes that are that many hops away from us.
5795 for (i = 0; i < level; i++) {
5796 sched_domains_numa_masks[i] =
5797 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5798 if (!sched_domains_numa_masks[i])
5801 for (j = 0; j < nr_node_ids; j++) {
5802 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5806 sched_domains_numa_masks[i][j] = mask;
5808 for (k = 0; k < nr_node_ids; k++) {
5809 if (node_distance(j, k) > sched_domains_numa_distance[i])
5812 cpumask_or(mask, mask, cpumask_of_node(k));
5817 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5818 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5823 * Copy the default topology bits..
5825 for (i = 0; default_topology[i].init; i++)
5826 tl[i] = default_topology[i];
5829 * .. and append 'j' levels of NUMA goodness.
5831 for (j = 0; j < level; i++, j++) {
5832 tl[i] = (struct sched_domain_topology_level){
5833 .init = sd_numa_init,
5834 .mask = sd_numa_mask,
5835 .flags = SDTL_OVERLAP,
5840 sched_domain_topology = tl;
5842 sched_domains_numa_levels = level;
5845 static void sched_domains_numa_masks_set(int cpu)
5848 int node = cpu_to_node(cpu);
5850 for (i = 0; i < sched_domains_numa_levels; i++) {
5851 for (j = 0; j < nr_node_ids; j++) {
5852 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5853 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5858 static void sched_domains_numa_masks_clear(int cpu)
5861 for (i = 0; i < sched_domains_numa_levels; i++) {
5862 for (j = 0; j < nr_node_ids; j++)
5863 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5868 * Update sched_domains_numa_masks[level][node] array when new cpus
5871 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5872 unsigned long action,
5875 int cpu = (long)hcpu;
5877 switch (action & ~CPU_TASKS_FROZEN) {
5879 sched_domains_numa_masks_set(cpu);
5883 sched_domains_numa_masks_clear(cpu);
5893 static inline void sched_init_numa(void)
5897 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5898 unsigned long action,
5903 #endif /* CONFIG_NUMA */
5905 static int __sdt_alloc(const struct cpumask *cpu_map)
5907 struct sched_domain_topology_level *tl;
5910 for_each_sd_topology(tl) {
5911 struct sd_data *sdd = &tl->data;
5913 sdd->sd = alloc_percpu(struct sched_domain *);
5917 sdd->sg = alloc_percpu(struct sched_group *);
5921 sdd->sgp = alloc_percpu(struct sched_group_power *);
5925 for_each_cpu(j, cpu_map) {
5926 struct sched_domain *sd;
5927 struct sched_group *sg;
5928 struct sched_group_power *sgp;
5930 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5931 GFP_KERNEL, cpu_to_node(j));
5935 *per_cpu_ptr(sdd->sd, j) = sd;
5937 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5938 GFP_KERNEL, cpu_to_node(j));
5944 *per_cpu_ptr(sdd->sg, j) = sg;
5946 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5947 GFP_KERNEL, cpu_to_node(j));
5951 *per_cpu_ptr(sdd->sgp, j) = sgp;
5958 static void __sdt_free(const struct cpumask *cpu_map)
5960 struct sched_domain_topology_level *tl;
5963 for_each_sd_topology(tl) {
5964 struct sd_data *sdd = &tl->data;
5966 for_each_cpu(j, cpu_map) {
5967 struct sched_domain *sd;
5970 sd = *per_cpu_ptr(sdd->sd, j);
5971 if (sd && (sd->flags & SD_OVERLAP))
5972 free_sched_groups(sd->groups, 0);
5973 kfree(*per_cpu_ptr(sdd->sd, j));
5977 kfree(*per_cpu_ptr(sdd->sg, j));
5979 kfree(*per_cpu_ptr(sdd->sgp, j));
5981 free_percpu(sdd->sd);
5983 free_percpu(sdd->sg);
5985 free_percpu(sdd->sgp);
5990 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5991 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5992 struct sched_domain *child, int cpu)
5994 struct sched_domain *sd = tl->init(tl, cpu);
5998 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6000 sd->level = child->level + 1;
6001 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6005 set_domain_attribute(sd, attr);
6011 * Build sched domains for a given set of cpus and attach the sched domains
6012 * to the individual cpus
6014 static int build_sched_domains(const struct cpumask *cpu_map,
6015 struct sched_domain_attr *attr)
6017 enum s_alloc alloc_state;
6018 struct sched_domain *sd;
6020 int i, ret = -ENOMEM;
6022 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6023 if (alloc_state != sa_rootdomain)
6026 /* Set up domains for cpus specified by the cpu_map. */
6027 for_each_cpu(i, cpu_map) {
6028 struct sched_domain_topology_level *tl;
6031 for_each_sd_topology(tl) {
6032 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6033 if (tl == sched_domain_topology)
6034 *per_cpu_ptr(d.sd, i) = sd;
6035 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6036 sd->flags |= SD_OVERLAP;
6037 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6042 /* Build the groups for the domains */
6043 for_each_cpu(i, cpu_map) {
6044 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6045 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6046 if (sd->flags & SD_OVERLAP) {
6047 if (build_overlap_sched_groups(sd, i))
6050 if (build_sched_groups(sd, i))
6056 /* Calculate CPU power for physical packages and nodes */
6057 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6058 if (!cpumask_test_cpu(i, cpu_map))
6061 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6062 claim_allocations(i, sd);
6063 init_sched_groups_power(i, sd);
6067 /* Attach the domains */
6069 for_each_cpu(i, cpu_map) {
6070 sd = *per_cpu_ptr(d.sd, i);
6071 cpu_attach_domain(sd, d.rd, i);
6077 __free_domain_allocs(&d, alloc_state, cpu_map);
6081 static cpumask_var_t *doms_cur; /* current sched domains */
6082 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6083 static struct sched_domain_attr *dattr_cur;
6084 /* attribues of custom domains in 'doms_cur' */
6087 * Special case: If a kmalloc of a doms_cur partition (array of
6088 * cpumask) fails, then fallback to a single sched domain,
6089 * as determined by the single cpumask fallback_doms.
6091 static cpumask_var_t fallback_doms;
6094 * arch_update_cpu_topology lets virtualized architectures update the
6095 * cpu core maps. It is supposed to return 1 if the topology changed
6096 * or 0 if it stayed the same.
6098 int __attribute__((weak)) arch_update_cpu_topology(void)
6103 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6106 cpumask_var_t *doms;
6108 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6111 for (i = 0; i < ndoms; i++) {
6112 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6113 free_sched_domains(doms, i);
6120 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6123 for (i = 0; i < ndoms; i++)
6124 free_cpumask_var(doms[i]);
6129 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6130 * For now this just excludes isolated cpus, but could be used to
6131 * exclude other special cases in the future.
6133 static int init_sched_domains(const struct cpumask *cpu_map)
6137 arch_update_cpu_topology();
6139 doms_cur = alloc_sched_domains(ndoms_cur);
6141 doms_cur = &fallback_doms;
6142 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6143 err = build_sched_domains(doms_cur[0], NULL);
6144 register_sched_domain_sysctl();
6150 * Detach sched domains from a group of cpus specified in cpu_map
6151 * These cpus will now be attached to the NULL domain
6153 static void detach_destroy_domains(const struct cpumask *cpu_map)
6158 for_each_cpu(i, cpu_map)
6159 cpu_attach_domain(NULL, &def_root_domain, i);
6163 /* handle null as "default" */
6164 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6165 struct sched_domain_attr *new, int idx_new)
6167 struct sched_domain_attr tmp;
6174 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6175 new ? (new + idx_new) : &tmp,
6176 sizeof(struct sched_domain_attr));
6180 * Partition sched domains as specified by the 'ndoms_new'
6181 * cpumasks in the array doms_new[] of cpumasks. This compares
6182 * doms_new[] to the current sched domain partitioning, doms_cur[].
6183 * It destroys each deleted domain and builds each new domain.
6185 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6186 * The masks don't intersect (don't overlap.) We should setup one
6187 * sched domain for each mask. CPUs not in any of the cpumasks will
6188 * not be load balanced. If the same cpumask appears both in the
6189 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6192 * The passed in 'doms_new' should be allocated using
6193 * alloc_sched_domains. This routine takes ownership of it and will
6194 * free_sched_domains it when done with it. If the caller failed the
6195 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6196 * and partition_sched_domains() will fallback to the single partition
6197 * 'fallback_doms', it also forces the domains to be rebuilt.
6199 * If doms_new == NULL it will be replaced with cpu_online_mask.
6200 * ndoms_new == 0 is a special case for destroying existing domains,
6201 * and it will not create the default domain.
6203 * Call with hotplug lock held
6205 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6206 struct sched_domain_attr *dattr_new)
6211 mutex_lock(&sched_domains_mutex);
6213 /* always unregister in case we don't destroy any domains */
6214 unregister_sched_domain_sysctl();
6216 /* Let architecture update cpu core mappings. */
6217 new_topology = arch_update_cpu_topology();
6219 n = doms_new ? ndoms_new : 0;
6221 /* Destroy deleted domains */
6222 for (i = 0; i < ndoms_cur; i++) {
6223 for (j = 0; j < n && !new_topology; j++) {
6224 if (cpumask_equal(doms_cur[i], doms_new[j])
6225 && dattrs_equal(dattr_cur, i, dattr_new, j))
6228 /* no match - a current sched domain not in new doms_new[] */
6229 detach_destroy_domains(doms_cur[i]);
6235 if (doms_new == NULL) {
6237 doms_new = &fallback_doms;
6238 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6239 WARN_ON_ONCE(dattr_new);
6242 /* Build new domains */
6243 for (i = 0; i < ndoms_new; i++) {
6244 for (j = 0; j < n && !new_topology; j++) {
6245 if (cpumask_equal(doms_new[i], doms_cur[j])
6246 && dattrs_equal(dattr_new, i, dattr_cur, j))
6249 /* no match - add a new doms_new */
6250 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6255 /* Remember the new sched domains */
6256 if (doms_cur != &fallback_doms)
6257 free_sched_domains(doms_cur, ndoms_cur);
6258 kfree(dattr_cur); /* kfree(NULL) is safe */
6259 doms_cur = doms_new;
6260 dattr_cur = dattr_new;
6261 ndoms_cur = ndoms_new;
6263 register_sched_domain_sysctl();
6265 mutex_unlock(&sched_domains_mutex);
6268 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6271 * Update cpusets according to cpu_active mask. If cpusets are
6272 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6273 * around partition_sched_domains().
6275 * If we come here as part of a suspend/resume, don't touch cpusets because we
6276 * want to restore it back to its original state upon resume anyway.
6278 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6282 case CPU_ONLINE_FROZEN:
6283 case CPU_DOWN_FAILED_FROZEN:
6286 * num_cpus_frozen tracks how many CPUs are involved in suspend
6287 * resume sequence. As long as this is not the last online
6288 * operation in the resume sequence, just build a single sched
6289 * domain, ignoring cpusets.
6292 if (likely(num_cpus_frozen)) {
6293 partition_sched_domains(1, NULL, NULL);
6298 * This is the last CPU online operation. So fall through and
6299 * restore the original sched domains by considering the
6300 * cpuset configurations.
6304 case CPU_DOWN_FAILED:
6305 cpuset_update_active_cpus(true);
6313 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6317 case CPU_DOWN_PREPARE:
6318 cpuset_update_active_cpus(false);
6320 case CPU_DOWN_PREPARE_FROZEN:
6322 partition_sched_domains(1, NULL, NULL);
6330 void __init sched_init_smp(void)
6332 cpumask_var_t non_isolated_cpus;
6334 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6335 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6340 mutex_lock(&sched_domains_mutex);
6341 init_sched_domains(cpu_active_mask);
6342 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6343 if (cpumask_empty(non_isolated_cpus))
6344 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6345 mutex_unlock(&sched_domains_mutex);
6348 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6349 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6350 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6354 /* Move init over to a non-isolated CPU */
6355 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6357 sched_init_granularity();
6358 free_cpumask_var(non_isolated_cpus);
6360 init_sched_rt_class();
6363 void __init sched_init_smp(void)
6365 sched_init_granularity();
6367 #endif /* CONFIG_SMP */
6369 const_debug unsigned int sysctl_timer_migration = 1;
6371 int in_sched_functions(unsigned long addr)
6373 return in_lock_functions(addr) ||
6374 (addr >= (unsigned long)__sched_text_start
6375 && addr < (unsigned long)__sched_text_end);
6378 #ifdef CONFIG_CGROUP_SCHED
6380 * Default task group.
6381 * Every task in system belongs to this group at bootup.
6383 struct task_group root_task_group;
6384 LIST_HEAD(task_groups);
6387 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6389 void __init sched_init(void)
6392 unsigned long alloc_size = 0, ptr;
6394 #ifdef CONFIG_FAIR_GROUP_SCHED
6395 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6397 #ifdef CONFIG_RT_GROUP_SCHED
6398 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6400 #ifdef CONFIG_CPUMASK_OFFSTACK
6401 alloc_size += num_possible_cpus() * cpumask_size();
6404 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6406 #ifdef CONFIG_FAIR_GROUP_SCHED
6407 root_task_group.se = (struct sched_entity **)ptr;
6408 ptr += nr_cpu_ids * sizeof(void **);
6410 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6411 ptr += nr_cpu_ids * sizeof(void **);
6413 #endif /* CONFIG_FAIR_GROUP_SCHED */
6414 #ifdef CONFIG_RT_GROUP_SCHED
6415 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6416 ptr += nr_cpu_ids * sizeof(void **);
6418 root_task_group.rt_rq = (struct rt_rq **)ptr;
6419 ptr += nr_cpu_ids * sizeof(void **);
6421 #endif /* CONFIG_RT_GROUP_SCHED */
6422 #ifdef CONFIG_CPUMASK_OFFSTACK
6423 for_each_possible_cpu(i) {
6424 per_cpu(load_balance_mask, i) = (void *)ptr;
6425 ptr += cpumask_size();
6427 #endif /* CONFIG_CPUMASK_OFFSTACK */
6431 init_defrootdomain();
6434 init_rt_bandwidth(&def_rt_bandwidth,
6435 global_rt_period(), global_rt_runtime());
6437 #ifdef CONFIG_RT_GROUP_SCHED
6438 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6439 global_rt_period(), global_rt_runtime());
6440 #endif /* CONFIG_RT_GROUP_SCHED */
6442 #ifdef CONFIG_CGROUP_SCHED
6443 list_add(&root_task_group.list, &task_groups);
6444 INIT_LIST_HEAD(&root_task_group.children);
6445 INIT_LIST_HEAD(&root_task_group.siblings);
6446 autogroup_init(&init_task);
6448 #endif /* CONFIG_CGROUP_SCHED */
6450 for_each_possible_cpu(i) {
6454 raw_spin_lock_init(&rq->lock);
6456 rq->calc_load_active = 0;
6457 rq->calc_load_update = jiffies + LOAD_FREQ;
6458 init_cfs_rq(&rq->cfs);
6459 init_rt_rq(&rq->rt, rq);
6460 #ifdef CONFIG_FAIR_GROUP_SCHED
6461 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6462 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6464 * How much cpu bandwidth does root_task_group get?
6466 * In case of task-groups formed thr' the cgroup filesystem, it
6467 * gets 100% of the cpu resources in the system. This overall
6468 * system cpu resource is divided among the tasks of
6469 * root_task_group and its child task-groups in a fair manner,
6470 * based on each entity's (task or task-group's) weight
6471 * (se->load.weight).
6473 * In other words, if root_task_group has 10 tasks of weight
6474 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6475 * then A0's share of the cpu resource is:
6477 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6479 * We achieve this by letting root_task_group's tasks sit
6480 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6482 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6483 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6484 #endif /* CONFIG_FAIR_GROUP_SCHED */
6486 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6487 #ifdef CONFIG_RT_GROUP_SCHED
6488 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6489 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6492 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6493 rq->cpu_load[j] = 0;
6495 rq->last_load_update_tick = jiffies;
6500 rq->cpu_power = SCHED_POWER_SCALE;
6501 rq->post_schedule = 0;
6502 rq->active_balance = 0;
6503 rq->next_balance = jiffies;
6508 rq->avg_idle = 2*sysctl_sched_migration_cost;
6509 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6511 INIT_LIST_HEAD(&rq->cfs_tasks);
6513 rq_attach_root(rq, &def_root_domain);
6514 #ifdef CONFIG_NO_HZ_COMMON
6517 #ifdef CONFIG_NO_HZ_FULL
6518 rq->last_sched_tick = 0;
6522 atomic_set(&rq->nr_iowait, 0);
6525 set_load_weight(&init_task);
6527 #ifdef CONFIG_PREEMPT_NOTIFIERS
6528 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6531 #ifdef CONFIG_RT_MUTEXES
6532 plist_head_init(&init_task.pi_waiters);
6536 * The boot idle thread does lazy MMU switching as well:
6538 atomic_inc(&init_mm.mm_count);
6539 enter_lazy_tlb(&init_mm, current);
6542 * Make us the idle thread. Technically, schedule() should not be
6543 * called from this thread, however somewhere below it might be,
6544 * but because we are the idle thread, we just pick up running again
6545 * when this runqueue becomes "idle".
6547 init_idle(current, smp_processor_id());
6549 calc_load_update = jiffies + LOAD_FREQ;
6552 * During early bootup we pretend to be a normal task:
6554 current->sched_class = &fair_sched_class;
6557 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6558 /* May be allocated at isolcpus cmdline parse time */
6559 if (cpu_isolated_map == NULL)
6560 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6561 idle_thread_set_boot_cpu();
6563 init_sched_fair_class();
6565 scheduler_running = 1;
6568 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6569 static inline int preempt_count_equals(int preempt_offset)
6571 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6573 return (nested == preempt_offset);
6576 void __might_sleep(const char *file, int line, int preempt_offset)
6578 static unsigned long prev_jiffy; /* ratelimiting */
6580 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6581 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6582 system_state != SYSTEM_RUNNING || oops_in_progress)
6584 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6586 prev_jiffy = jiffies;
6589 "BUG: sleeping function called from invalid context at %s:%d\n",
6592 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6593 in_atomic(), irqs_disabled(),
6594 current->pid, current->comm);
6596 debug_show_held_locks(current);
6597 if (irqs_disabled())
6598 print_irqtrace_events(current);
6601 EXPORT_SYMBOL(__might_sleep);
6604 #ifdef CONFIG_MAGIC_SYSRQ
6605 static void normalize_task(struct rq *rq, struct task_struct *p)
6607 const struct sched_class *prev_class = p->sched_class;
6608 int old_prio = p->prio;
6613 dequeue_task(rq, p, 0);
6614 __setscheduler(rq, p, SCHED_NORMAL, 0);
6616 enqueue_task(rq, p, 0);
6617 resched_task(rq->curr);
6620 check_class_changed(rq, p, prev_class, old_prio);
6623 void normalize_rt_tasks(void)
6625 struct task_struct *g, *p;
6626 unsigned long flags;
6629 read_lock_irqsave(&tasklist_lock, flags);
6630 do_each_thread(g, p) {
6632 * Only normalize user tasks:
6637 p->se.exec_start = 0;
6638 #ifdef CONFIG_SCHEDSTATS
6639 p->se.statistics.wait_start = 0;
6640 p->se.statistics.sleep_start = 0;
6641 p->se.statistics.block_start = 0;
6646 * Renice negative nice level userspace
6649 if (TASK_NICE(p) < 0 && p->mm)
6650 set_user_nice(p, 0);
6654 raw_spin_lock(&p->pi_lock);
6655 rq = __task_rq_lock(p);
6657 normalize_task(rq, p);
6659 __task_rq_unlock(rq);
6660 raw_spin_unlock(&p->pi_lock);
6661 } while_each_thread(g, p);
6663 read_unlock_irqrestore(&tasklist_lock, flags);
6666 #endif /* CONFIG_MAGIC_SYSRQ */
6668 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6670 * These functions are only useful for the IA64 MCA handling, or kdb.
6672 * They can only be called when the whole system has been
6673 * stopped - every CPU needs to be quiescent, and no scheduling
6674 * activity can take place. Using them for anything else would
6675 * be a serious bug, and as a result, they aren't even visible
6676 * under any other configuration.
6680 * curr_task - return the current task for a given cpu.
6681 * @cpu: the processor in question.
6683 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6685 * Return: The current task for @cpu.
6687 struct task_struct *curr_task(int cpu)
6689 return cpu_curr(cpu);
6692 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6696 * set_curr_task - set the current task for a given cpu.
6697 * @cpu: the processor in question.
6698 * @p: the task pointer to set.
6700 * Description: This function must only be used when non-maskable interrupts
6701 * are serviced on a separate stack. It allows the architecture to switch the
6702 * notion of the current task on a cpu in a non-blocking manner. This function
6703 * must be called with all CPU's synchronized, and interrupts disabled, the
6704 * and caller must save the original value of the current task (see
6705 * curr_task() above) and restore that value before reenabling interrupts and
6706 * re-starting the system.
6708 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6710 void set_curr_task(int cpu, struct task_struct *p)
6717 #ifdef CONFIG_CGROUP_SCHED
6718 /* task_group_lock serializes the addition/removal of task groups */
6719 static DEFINE_SPINLOCK(task_group_lock);
6721 static void free_sched_group(struct task_group *tg)
6723 free_fair_sched_group(tg);
6724 free_rt_sched_group(tg);
6729 /* allocate runqueue etc for a new task group */
6730 struct task_group *sched_create_group(struct task_group *parent)
6732 struct task_group *tg;
6734 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6736 return ERR_PTR(-ENOMEM);
6738 if (!alloc_fair_sched_group(tg, parent))
6741 if (!alloc_rt_sched_group(tg, parent))
6747 free_sched_group(tg);
6748 return ERR_PTR(-ENOMEM);
6751 void sched_online_group(struct task_group *tg, struct task_group *parent)
6753 unsigned long flags;
6755 spin_lock_irqsave(&task_group_lock, flags);
6756 list_add_rcu(&tg->list, &task_groups);
6758 WARN_ON(!parent); /* root should already exist */
6760 tg->parent = parent;
6761 INIT_LIST_HEAD(&tg->children);
6762 list_add_rcu(&tg->siblings, &parent->children);
6763 spin_unlock_irqrestore(&task_group_lock, flags);
6766 /* rcu callback to free various structures associated with a task group */
6767 static void free_sched_group_rcu(struct rcu_head *rhp)
6769 /* now it should be safe to free those cfs_rqs */
6770 free_sched_group(container_of(rhp, struct task_group, rcu));
6773 /* Destroy runqueue etc associated with a task group */
6774 void sched_destroy_group(struct task_group *tg)
6776 /* wait for possible concurrent references to cfs_rqs complete */
6777 call_rcu(&tg->rcu, free_sched_group_rcu);
6780 void sched_offline_group(struct task_group *tg)
6782 unsigned long flags;
6785 /* end participation in shares distribution */
6786 for_each_possible_cpu(i)
6787 unregister_fair_sched_group(tg, i);
6789 spin_lock_irqsave(&task_group_lock, flags);
6790 list_del_rcu(&tg->list);
6791 list_del_rcu(&tg->siblings);
6792 spin_unlock_irqrestore(&task_group_lock, flags);
6795 /* change task's runqueue when it moves between groups.
6796 * The caller of this function should have put the task in its new group
6797 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6798 * reflect its new group.
6800 void sched_move_task(struct task_struct *tsk)
6802 struct task_group *tg;
6804 unsigned long flags;
6807 rq = task_rq_lock(tsk, &flags);
6809 running = task_current(rq, tsk);
6813 dequeue_task(rq, tsk, 0);
6814 if (unlikely(running))
6815 tsk->sched_class->put_prev_task(rq, tsk);
6817 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6818 lockdep_is_held(&tsk->sighand->siglock)),
6819 struct task_group, css);
6820 tg = autogroup_task_group(tsk, tg);
6821 tsk->sched_task_group = tg;
6823 #ifdef CONFIG_FAIR_GROUP_SCHED
6824 if (tsk->sched_class->task_move_group)
6825 tsk->sched_class->task_move_group(tsk, on_rq);
6828 set_task_rq(tsk, task_cpu(tsk));
6830 if (unlikely(running))
6831 tsk->sched_class->set_curr_task(rq);
6833 enqueue_task(rq, tsk, 0);
6835 task_rq_unlock(rq, tsk, &flags);
6837 #endif /* CONFIG_CGROUP_SCHED */
6839 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6840 static unsigned long to_ratio(u64 period, u64 runtime)
6842 if (runtime == RUNTIME_INF)
6845 return div64_u64(runtime << 20, period);
6849 #ifdef CONFIG_RT_GROUP_SCHED
6851 * Ensure that the real time constraints are schedulable.
6853 static DEFINE_MUTEX(rt_constraints_mutex);
6855 /* Must be called with tasklist_lock held */
6856 static inline int tg_has_rt_tasks(struct task_group *tg)
6858 struct task_struct *g, *p;
6860 do_each_thread(g, p) {
6861 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6863 } while_each_thread(g, p);
6868 struct rt_schedulable_data {
6869 struct task_group *tg;
6874 static int tg_rt_schedulable(struct task_group *tg, void *data)
6876 struct rt_schedulable_data *d = data;
6877 struct task_group *child;
6878 unsigned long total, sum = 0;
6879 u64 period, runtime;
6881 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6882 runtime = tg->rt_bandwidth.rt_runtime;
6885 period = d->rt_period;
6886 runtime = d->rt_runtime;
6890 * Cannot have more runtime than the period.
6892 if (runtime > period && runtime != RUNTIME_INF)
6896 * Ensure we don't starve existing RT tasks.
6898 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6901 total = to_ratio(period, runtime);
6904 * Nobody can have more than the global setting allows.
6906 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6910 * The sum of our children's runtime should not exceed our own.
6912 list_for_each_entry_rcu(child, &tg->children, siblings) {
6913 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6914 runtime = child->rt_bandwidth.rt_runtime;
6916 if (child == d->tg) {
6917 period = d->rt_period;
6918 runtime = d->rt_runtime;
6921 sum += to_ratio(period, runtime);
6930 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6934 struct rt_schedulable_data data = {
6936 .rt_period = period,
6937 .rt_runtime = runtime,
6941 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6947 static int tg_set_rt_bandwidth(struct task_group *tg,
6948 u64 rt_period, u64 rt_runtime)
6952 mutex_lock(&rt_constraints_mutex);
6953 read_lock(&tasklist_lock);
6954 err = __rt_schedulable(tg, rt_period, rt_runtime);
6958 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6959 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6960 tg->rt_bandwidth.rt_runtime = rt_runtime;
6962 for_each_possible_cpu(i) {
6963 struct rt_rq *rt_rq = tg->rt_rq[i];
6965 raw_spin_lock(&rt_rq->rt_runtime_lock);
6966 rt_rq->rt_runtime = rt_runtime;
6967 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6969 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6971 read_unlock(&tasklist_lock);
6972 mutex_unlock(&rt_constraints_mutex);
6977 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6979 u64 rt_runtime, rt_period;
6981 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6982 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6983 if (rt_runtime_us < 0)
6984 rt_runtime = RUNTIME_INF;
6986 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6989 static long sched_group_rt_runtime(struct task_group *tg)
6993 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6996 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6997 do_div(rt_runtime_us, NSEC_PER_USEC);
6998 return rt_runtime_us;
7001 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7003 u64 rt_runtime, rt_period;
7005 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7006 rt_runtime = tg->rt_bandwidth.rt_runtime;
7011 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7014 static long sched_group_rt_period(struct task_group *tg)
7018 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7019 do_div(rt_period_us, NSEC_PER_USEC);
7020 return rt_period_us;
7023 static int sched_rt_global_constraints(void)
7025 u64 runtime, period;
7028 if (sysctl_sched_rt_period <= 0)
7031 runtime = global_rt_runtime();
7032 period = global_rt_period();
7035 * Sanity check on the sysctl variables.
7037 if (runtime > period && runtime != RUNTIME_INF)
7040 mutex_lock(&rt_constraints_mutex);
7041 read_lock(&tasklist_lock);
7042 ret = __rt_schedulable(NULL, 0, 0);
7043 read_unlock(&tasklist_lock);
7044 mutex_unlock(&rt_constraints_mutex);
7049 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7051 /* Don't accept realtime tasks when there is no way for them to run */
7052 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7058 #else /* !CONFIG_RT_GROUP_SCHED */
7059 static int sched_rt_global_constraints(void)
7061 unsigned long flags;
7064 if (sysctl_sched_rt_period <= 0)
7068 * There's always some RT tasks in the root group
7069 * -- migration, kstopmachine etc..
7071 if (sysctl_sched_rt_runtime == 0)
7074 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7075 for_each_possible_cpu(i) {
7076 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7078 raw_spin_lock(&rt_rq->rt_runtime_lock);
7079 rt_rq->rt_runtime = global_rt_runtime();
7080 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7082 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7086 #endif /* CONFIG_RT_GROUP_SCHED */
7088 int sched_rr_handler(struct ctl_table *table, int write,
7089 void __user *buffer, size_t *lenp,
7093 static DEFINE_MUTEX(mutex);
7096 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7097 /* make sure that internally we keep jiffies */
7098 /* also, writing zero resets timeslice to default */
7099 if (!ret && write) {
7100 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7101 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7103 mutex_unlock(&mutex);
7107 int sched_rt_handler(struct ctl_table *table, int write,
7108 void __user *buffer, size_t *lenp,
7112 int old_period, old_runtime;
7113 static DEFINE_MUTEX(mutex);
7116 old_period = sysctl_sched_rt_period;
7117 old_runtime = sysctl_sched_rt_runtime;
7119 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7121 if (!ret && write) {
7122 ret = sched_rt_global_constraints();
7124 sysctl_sched_rt_period = old_period;
7125 sysctl_sched_rt_runtime = old_runtime;
7127 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7128 def_rt_bandwidth.rt_period =
7129 ns_to_ktime(global_rt_period());
7132 mutex_unlock(&mutex);
7137 #ifdef CONFIG_CGROUP_SCHED
7139 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7141 return css ? container_of(css, struct task_group, css) : NULL;
7144 static struct cgroup_subsys_state *
7145 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7147 struct task_group *parent = css_tg(parent_css);
7148 struct task_group *tg;
7151 /* This is early initialization for the top cgroup */
7152 return &root_task_group.css;
7155 tg = sched_create_group(parent);
7157 return ERR_PTR(-ENOMEM);
7162 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7164 struct task_group *tg = css_tg(css);
7165 struct task_group *parent = css_tg(css_parent(css));
7168 sched_online_group(tg, parent);
7172 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7174 struct task_group *tg = css_tg(css);
7176 sched_destroy_group(tg);
7179 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7181 struct task_group *tg = css_tg(css);
7183 sched_offline_group(tg);
7186 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7187 struct cgroup_taskset *tset)
7189 struct task_struct *task;
7191 cgroup_taskset_for_each(task, css, tset) {
7192 #ifdef CONFIG_RT_GROUP_SCHED
7193 if (!sched_rt_can_attach(css_tg(css), task))
7196 /* We don't support RT-tasks being in separate groups */
7197 if (task->sched_class != &fair_sched_class)
7204 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7205 struct cgroup_taskset *tset)
7207 struct task_struct *task;
7209 cgroup_taskset_for_each(task, css, tset)
7210 sched_move_task(task);
7213 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7214 struct cgroup_subsys_state *old_css,
7215 struct task_struct *task)
7218 * cgroup_exit() is called in the copy_process() failure path.
7219 * Ignore this case since the task hasn't ran yet, this avoids
7220 * trying to poke a half freed task state from generic code.
7222 if (!(task->flags & PF_EXITING))
7225 sched_move_task(task);
7228 #ifdef CONFIG_FAIR_GROUP_SCHED
7229 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7230 struct cftype *cftype, u64 shareval)
7232 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7235 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7238 struct task_group *tg = css_tg(css);
7240 return (u64) scale_load_down(tg->shares);
7243 #ifdef CONFIG_CFS_BANDWIDTH
7244 static DEFINE_MUTEX(cfs_constraints_mutex);
7246 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7247 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7249 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7251 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7253 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7254 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7256 if (tg == &root_task_group)
7260 * Ensure we have at some amount of bandwidth every period. This is
7261 * to prevent reaching a state of large arrears when throttled via
7262 * entity_tick() resulting in prolonged exit starvation.
7264 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7268 * Likewise, bound things on the otherside by preventing insane quota
7269 * periods. This also allows us to normalize in computing quota
7272 if (period > max_cfs_quota_period)
7275 mutex_lock(&cfs_constraints_mutex);
7276 ret = __cfs_schedulable(tg, period, quota);
7280 runtime_enabled = quota != RUNTIME_INF;
7281 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7282 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7283 raw_spin_lock_irq(&cfs_b->lock);
7284 cfs_b->period = ns_to_ktime(period);
7285 cfs_b->quota = quota;
7287 __refill_cfs_bandwidth_runtime(cfs_b);
7288 /* restart the period timer (if active) to handle new period expiry */
7289 if (runtime_enabled && cfs_b->timer_active) {
7290 /* force a reprogram */
7291 cfs_b->timer_active = 0;
7292 __start_cfs_bandwidth(cfs_b);
7294 raw_spin_unlock_irq(&cfs_b->lock);
7296 for_each_possible_cpu(i) {
7297 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7298 struct rq *rq = cfs_rq->rq;
7300 raw_spin_lock_irq(&rq->lock);
7301 cfs_rq->runtime_enabled = runtime_enabled;
7302 cfs_rq->runtime_remaining = 0;
7304 if (cfs_rq->throttled)
7305 unthrottle_cfs_rq(cfs_rq);
7306 raw_spin_unlock_irq(&rq->lock);
7309 mutex_unlock(&cfs_constraints_mutex);
7314 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7318 period = ktime_to_ns(tg->cfs_bandwidth.period);
7319 if (cfs_quota_us < 0)
7320 quota = RUNTIME_INF;
7322 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7324 return tg_set_cfs_bandwidth(tg, period, quota);
7327 long tg_get_cfs_quota(struct task_group *tg)
7331 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7334 quota_us = tg->cfs_bandwidth.quota;
7335 do_div(quota_us, NSEC_PER_USEC);
7340 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7344 period = (u64)cfs_period_us * NSEC_PER_USEC;
7345 quota = tg->cfs_bandwidth.quota;
7347 return tg_set_cfs_bandwidth(tg, period, quota);
7350 long tg_get_cfs_period(struct task_group *tg)
7354 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7355 do_div(cfs_period_us, NSEC_PER_USEC);
7357 return cfs_period_us;
7360 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7363 return tg_get_cfs_quota(css_tg(css));
7366 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7367 struct cftype *cftype, s64 cfs_quota_us)
7369 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7372 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7375 return tg_get_cfs_period(css_tg(css));
7378 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7379 struct cftype *cftype, u64 cfs_period_us)
7381 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7384 struct cfs_schedulable_data {
7385 struct task_group *tg;
7390 * normalize group quota/period to be quota/max_period
7391 * note: units are usecs
7393 static u64 normalize_cfs_quota(struct task_group *tg,
7394 struct cfs_schedulable_data *d)
7402 period = tg_get_cfs_period(tg);
7403 quota = tg_get_cfs_quota(tg);
7406 /* note: these should typically be equivalent */
7407 if (quota == RUNTIME_INF || quota == -1)
7410 return to_ratio(period, quota);
7413 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7415 struct cfs_schedulable_data *d = data;
7416 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7417 s64 quota = 0, parent_quota = -1;
7420 quota = RUNTIME_INF;
7422 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7424 quota = normalize_cfs_quota(tg, d);
7425 parent_quota = parent_b->hierarchal_quota;
7428 * ensure max(child_quota) <= parent_quota, inherit when no
7431 if (quota == RUNTIME_INF)
7432 quota = parent_quota;
7433 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7436 cfs_b->hierarchal_quota = quota;
7441 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7444 struct cfs_schedulable_data data = {
7450 if (quota != RUNTIME_INF) {
7451 do_div(data.period, NSEC_PER_USEC);
7452 do_div(data.quota, NSEC_PER_USEC);
7456 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7462 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7463 struct cgroup_map_cb *cb)
7465 struct task_group *tg = css_tg(css);
7466 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7468 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7469 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7470 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7474 #endif /* CONFIG_CFS_BANDWIDTH */
7475 #endif /* CONFIG_FAIR_GROUP_SCHED */
7477 #ifdef CONFIG_RT_GROUP_SCHED
7478 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7479 struct cftype *cft, s64 val)
7481 return sched_group_set_rt_runtime(css_tg(css), val);
7484 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7487 return sched_group_rt_runtime(css_tg(css));
7490 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7491 struct cftype *cftype, u64 rt_period_us)
7493 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7496 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7499 return sched_group_rt_period(css_tg(css));
7501 #endif /* CONFIG_RT_GROUP_SCHED */
7503 static struct cftype cpu_files[] = {
7504 #ifdef CONFIG_FAIR_GROUP_SCHED
7507 .read_u64 = cpu_shares_read_u64,
7508 .write_u64 = cpu_shares_write_u64,
7511 #ifdef CONFIG_CFS_BANDWIDTH
7513 .name = "cfs_quota_us",
7514 .read_s64 = cpu_cfs_quota_read_s64,
7515 .write_s64 = cpu_cfs_quota_write_s64,
7518 .name = "cfs_period_us",
7519 .read_u64 = cpu_cfs_period_read_u64,
7520 .write_u64 = cpu_cfs_period_write_u64,
7524 .read_map = cpu_stats_show,
7527 #ifdef CONFIG_RT_GROUP_SCHED
7529 .name = "rt_runtime_us",
7530 .read_s64 = cpu_rt_runtime_read,
7531 .write_s64 = cpu_rt_runtime_write,
7534 .name = "rt_period_us",
7535 .read_u64 = cpu_rt_period_read_uint,
7536 .write_u64 = cpu_rt_period_write_uint,
7542 struct cgroup_subsys cpu_cgroup_subsys = {
7544 .css_alloc = cpu_cgroup_css_alloc,
7545 .css_free = cpu_cgroup_css_free,
7546 .css_online = cpu_cgroup_css_online,
7547 .css_offline = cpu_cgroup_css_offline,
7548 .can_attach = cpu_cgroup_can_attach,
7549 .attach = cpu_cgroup_attach,
7550 .exit = cpu_cgroup_exit,
7551 .subsys_id = cpu_cgroup_subsys_id,
7552 .base_cftypes = cpu_files,
7556 #endif /* CONFIG_CGROUP_SCHED */
7558 void dump_cpu_task(int cpu)
7560 pr_info("Task dump for CPU %d:\n", cpu);
7561 sched_show_task(cpu_curr(cpu));