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;
1628 p->mm->numa_next_reset = jiffies;
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_work.next = &p->numa_work;
1637 #endif /* CONFIG_NUMA_BALANCING */
1640 #ifdef CONFIG_NUMA_BALANCING
1641 #ifdef CONFIG_SCHED_DEBUG
1642 void set_numabalancing_state(bool enabled)
1645 sched_feat_set("NUMA");
1647 sched_feat_set("NO_NUMA");
1650 __read_mostly bool numabalancing_enabled;
1652 void set_numabalancing_state(bool enabled)
1654 numabalancing_enabled = enabled;
1656 #endif /* CONFIG_SCHED_DEBUG */
1657 #endif /* CONFIG_NUMA_BALANCING */
1660 * fork()/clone()-time setup:
1662 void sched_fork(struct task_struct *p)
1664 unsigned long flags;
1665 int cpu = get_cpu();
1669 * We mark the process as running here. This guarantees that
1670 * nobody will actually run it, and a signal or other external
1671 * event cannot wake it up and insert it on the runqueue either.
1673 p->state = TASK_RUNNING;
1676 * Make sure we do not leak PI boosting priority to the child.
1678 p->prio = current->normal_prio;
1681 * Revert to default priority/policy on fork if requested.
1683 if (unlikely(p->sched_reset_on_fork)) {
1684 if (task_has_rt_policy(p)) {
1685 p->policy = SCHED_NORMAL;
1686 p->static_prio = NICE_TO_PRIO(0);
1688 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1689 p->static_prio = NICE_TO_PRIO(0);
1691 p->prio = p->normal_prio = __normal_prio(p);
1695 * We don't need the reset flag anymore after the fork. It has
1696 * fulfilled its duty:
1698 p->sched_reset_on_fork = 0;
1701 if (!rt_prio(p->prio))
1702 p->sched_class = &fair_sched_class;
1704 if (p->sched_class->task_fork)
1705 p->sched_class->task_fork(p);
1708 * The child is not yet in the pid-hash so no cgroup attach races,
1709 * and the cgroup is pinned to this child due to cgroup_fork()
1710 * is ran before sched_fork().
1712 * Silence PROVE_RCU.
1714 raw_spin_lock_irqsave(&p->pi_lock, flags);
1715 set_task_cpu(p, cpu);
1716 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1718 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1719 if (likely(sched_info_on()))
1720 memset(&p->sched_info, 0, sizeof(p->sched_info));
1722 #if defined(CONFIG_SMP)
1725 #ifdef CONFIG_PREEMPT_COUNT
1726 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)) {
1898 * Remove function-return probe instances associated with this
1899 * task and put them back on the free list.
1901 kprobe_flush_task(prev);
1902 put_task_struct(prev);
1905 tick_nohz_task_switch(current);
1910 /* assumes rq->lock is held */
1911 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1913 if (prev->sched_class->pre_schedule)
1914 prev->sched_class->pre_schedule(rq, prev);
1917 /* rq->lock is NOT held, but preemption is disabled */
1918 static inline void post_schedule(struct rq *rq)
1920 if (rq->post_schedule) {
1921 unsigned long flags;
1923 raw_spin_lock_irqsave(&rq->lock, flags);
1924 if (rq->curr->sched_class->post_schedule)
1925 rq->curr->sched_class->post_schedule(rq);
1926 raw_spin_unlock_irqrestore(&rq->lock, flags);
1928 rq->post_schedule = 0;
1934 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1938 static inline void post_schedule(struct rq *rq)
1945 * schedule_tail - first thing a freshly forked thread must call.
1946 * @prev: the thread we just switched away from.
1948 asmlinkage void schedule_tail(struct task_struct *prev)
1949 __releases(rq->lock)
1951 struct rq *rq = this_rq();
1953 finish_task_switch(rq, prev);
1956 * FIXME: do we need to worry about rq being invalidated by the
1961 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1962 /* In this case, finish_task_switch does not reenable preemption */
1965 if (current->set_child_tid)
1966 put_user(task_pid_vnr(current), current->set_child_tid);
1970 * context_switch - switch to the new MM and the new
1971 * thread's register state.
1974 context_switch(struct rq *rq, struct task_struct *prev,
1975 struct task_struct *next)
1977 struct mm_struct *mm, *oldmm;
1979 prepare_task_switch(rq, prev, next);
1982 oldmm = prev->active_mm;
1984 * For paravirt, this is coupled with an exit in switch_to to
1985 * combine the page table reload and the switch backend into
1988 arch_start_context_switch(prev);
1991 next->active_mm = oldmm;
1992 atomic_inc(&oldmm->mm_count);
1993 enter_lazy_tlb(oldmm, next);
1995 switch_mm(oldmm, mm, next);
1998 prev->active_mm = NULL;
1999 rq->prev_mm = oldmm;
2002 * Since the runqueue lock will be released by the next
2003 * task (which is an invalid locking op but in the case
2004 * of the scheduler it's an obvious special-case), so we
2005 * do an early lockdep release here:
2007 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2008 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2011 context_tracking_task_switch(prev, next);
2012 /* Here we just switch the register state and the stack. */
2013 switch_to(prev, next, prev);
2017 * this_rq must be evaluated again because prev may have moved
2018 * CPUs since it called schedule(), thus the 'rq' on its stack
2019 * frame will be invalid.
2021 finish_task_switch(this_rq(), prev);
2025 * nr_running and nr_context_switches:
2027 * externally visible scheduler statistics: current number of runnable
2028 * threads, total number of context switches performed since bootup.
2030 unsigned long nr_running(void)
2032 unsigned long i, sum = 0;
2034 for_each_online_cpu(i)
2035 sum += cpu_rq(i)->nr_running;
2040 unsigned long long nr_context_switches(void)
2043 unsigned long long sum = 0;
2045 for_each_possible_cpu(i)
2046 sum += cpu_rq(i)->nr_switches;
2051 unsigned long nr_iowait(void)
2053 unsigned long i, sum = 0;
2055 for_each_possible_cpu(i)
2056 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2061 unsigned long nr_iowait_cpu(int cpu)
2063 struct rq *this = cpu_rq(cpu);
2064 return atomic_read(&this->nr_iowait);
2070 * sched_exec - execve() is a valuable balancing opportunity, because at
2071 * this point the task has the smallest effective memory and cache footprint.
2073 void sched_exec(void)
2075 struct task_struct *p = current;
2076 unsigned long flags;
2079 raw_spin_lock_irqsave(&p->pi_lock, flags);
2080 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2081 if (dest_cpu == smp_processor_id())
2084 if (likely(cpu_active(dest_cpu))) {
2085 struct migration_arg arg = { p, dest_cpu };
2087 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2088 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2092 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2097 DEFINE_PER_CPU(struct kernel_stat, kstat);
2098 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2100 EXPORT_PER_CPU_SYMBOL(kstat);
2101 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2104 * Return any ns on the sched_clock that have not yet been accounted in
2105 * @p in case that task is currently running.
2107 * Called with task_rq_lock() held on @rq.
2109 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2113 if (task_current(rq, p)) {
2114 update_rq_clock(rq);
2115 ns = rq_clock_task(rq) - p->se.exec_start;
2123 unsigned long long task_delta_exec(struct task_struct *p)
2125 unsigned long flags;
2129 rq = task_rq_lock(p, &flags);
2130 ns = do_task_delta_exec(p, rq);
2131 task_rq_unlock(rq, p, &flags);
2137 * Return accounted runtime for the task.
2138 * In case the task is currently running, return the runtime plus current's
2139 * pending runtime that have not been accounted yet.
2141 unsigned long long task_sched_runtime(struct task_struct *p)
2143 unsigned long flags;
2147 rq = task_rq_lock(p, &flags);
2148 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2149 task_rq_unlock(rq, p, &flags);
2155 * This function gets called by the timer code, with HZ frequency.
2156 * We call it with interrupts disabled.
2158 void scheduler_tick(void)
2160 int cpu = smp_processor_id();
2161 struct rq *rq = cpu_rq(cpu);
2162 struct task_struct *curr = rq->curr;
2166 raw_spin_lock(&rq->lock);
2167 update_rq_clock(rq);
2168 curr->sched_class->task_tick(rq, curr, 0);
2169 update_cpu_load_active(rq);
2170 raw_spin_unlock(&rq->lock);
2172 perf_event_task_tick();
2175 rq->idle_balance = idle_cpu(cpu);
2176 trigger_load_balance(rq, cpu);
2178 rq_last_tick_reset(rq);
2181 #ifdef CONFIG_NO_HZ_FULL
2183 * scheduler_tick_max_deferment
2185 * Keep at least one tick per second when a single
2186 * active task is running because the scheduler doesn't
2187 * yet completely support full dynticks environment.
2189 * This makes sure that uptime, CFS vruntime, load
2190 * balancing, etc... continue to move forward, even
2191 * with a very low granularity.
2193 * Return: Maximum deferment in nanoseconds.
2195 u64 scheduler_tick_max_deferment(void)
2197 struct rq *rq = this_rq();
2198 unsigned long next, now = ACCESS_ONCE(jiffies);
2200 next = rq->last_sched_tick + HZ;
2202 if (time_before_eq(next, now))
2205 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2209 notrace unsigned long get_parent_ip(unsigned long addr)
2211 if (in_lock_functions(addr)) {
2212 addr = CALLER_ADDR2;
2213 if (in_lock_functions(addr))
2214 addr = CALLER_ADDR3;
2219 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2220 defined(CONFIG_PREEMPT_TRACER))
2222 void __kprobes add_preempt_count(int val)
2224 #ifdef CONFIG_DEBUG_PREEMPT
2228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2231 add_preempt_count_notrace(val);
2232 #ifdef CONFIG_DEBUG_PREEMPT
2234 * Spinlock count overflowing soon?
2236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2239 if (preempt_count() == val)
2240 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2242 EXPORT_SYMBOL(add_preempt_count);
2244 void __kprobes sub_preempt_count(int val)
2246 #ifdef CONFIG_DEBUG_PREEMPT
2250 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2253 * Is the spinlock portion underflowing?
2255 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2256 !(preempt_count() & PREEMPT_MASK)))
2260 if (preempt_count() == val)
2261 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2262 sub_preempt_count_notrace(val);
2264 EXPORT_SYMBOL(sub_preempt_count);
2269 * Print scheduling while atomic bug:
2271 static noinline void __schedule_bug(struct task_struct *prev)
2273 if (oops_in_progress)
2276 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2277 prev->comm, prev->pid, preempt_count());
2279 debug_show_held_locks(prev);
2281 if (irqs_disabled())
2282 print_irqtrace_events(prev);
2284 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2288 * Various schedule()-time debugging checks and statistics:
2290 static inline void schedule_debug(struct task_struct *prev)
2293 * Test if we are atomic. Since do_exit() needs to call into
2294 * schedule() atomically, we ignore that path for now.
2295 * Otherwise, whine if we are scheduling when we should not be.
2297 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2298 __schedule_bug(prev);
2301 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2303 schedstat_inc(this_rq(), sched_count);
2306 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2308 if (prev->on_rq || rq->skip_clock_update < 0)
2309 update_rq_clock(rq);
2310 prev->sched_class->put_prev_task(rq, prev);
2314 * Pick up the highest-prio task:
2316 static inline struct task_struct *
2317 pick_next_task(struct rq *rq)
2319 const struct sched_class *class;
2320 struct task_struct *p;
2323 * Optimization: we know that if all tasks are in
2324 * the fair class we can call that function directly:
2326 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2327 p = fair_sched_class.pick_next_task(rq);
2332 for_each_class(class) {
2333 p = class->pick_next_task(rq);
2338 BUG(); /* the idle class will always have a runnable task */
2342 * __schedule() is the main scheduler function.
2344 * The main means of driving the scheduler and thus entering this function are:
2346 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2348 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2349 * paths. For example, see arch/x86/entry_64.S.
2351 * To drive preemption between tasks, the scheduler sets the flag in timer
2352 * interrupt handler scheduler_tick().
2354 * 3. Wakeups don't really cause entry into schedule(). They add a
2355 * task to the run-queue and that's it.
2357 * Now, if the new task added to the run-queue preempts the current
2358 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2359 * called on the nearest possible occasion:
2361 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2363 * - in syscall or exception context, at the next outmost
2364 * preempt_enable(). (this might be as soon as the wake_up()'s
2367 * - in IRQ context, return from interrupt-handler to
2368 * preemptible context
2370 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2373 * - cond_resched() call
2374 * - explicit schedule() call
2375 * - return from syscall or exception to user-space
2376 * - return from interrupt-handler to user-space
2378 static void __sched __schedule(void)
2380 struct task_struct *prev, *next;
2381 unsigned long *switch_count;
2387 cpu = smp_processor_id();
2389 rcu_note_context_switch(cpu);
2392 schedule_debug(prev);
2394 if (sched_feat(HRTICK))
2398 * Make sure that signal_pending_state()->signal_pending() below
2399 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2400 * done by the caller to avoid the race with signal_wake_up().
2402 smp_mb__before_spinlock();
2403 raw_spin_lock_irq(&rq->lock);
2405 switch_count = &prev->nivcsw;
2406 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2407 if (unlikely(signal_pending_state(prev->state, prev))) {
2408 prev->state = TASK_RUNNING;
2410 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2414 * If a worker went to sleep, notify and ask workqueue
2415 * whether it wants to wake up a task to maintain
2418 if (prev->flags & PF_WQ_WORKER) {
2419 struct task_struct *to_wakeup;
2421 to_wakeup = wq_worker_sleeping(prev, cpu);
2423 try_to_wake_up_local(to_wakeup);
2426 switch_count = &prev->nvcsw;
2429 pre_schedule(rq, prev);
2431 if (unlikely(!rq->nr_running))
2432 idle_balance(cpu, rq);
2434 put_prev_task(rq, prev);
2435 next = pick_next_task(rq);
2436 clear_tsk_need_resched(prev);
2437 clear_preempt_need_resched();
2438 rq->skip_clock_update = 0;
2440 if (likely(prev != next)) {
2445 context_switch(rq, prev, next); /* unlocks the rq */
2447 * The context switch have flipped the stack from under us
2448 * and restored the local variables which were saved when
2449 * this task called schedule() in the past. prev == current
2450 * is still correct, but it can be moved to another cpu/rq.
2452 cpu = smp_processor_id();
2455 raw_spin_unlock_irq(&rq->lock);
2459 sched_preempt_enable_no_resched();
2464 static inline void sched_submit_work(struct task_struct *tsk)
2466 if (!tsk->state || tsk_is_pi_blocked(tsk))
2469 * If we are going to sleep and we have plugged IO queued,
2470 * make sure to submit it to avoid deadlocks.
2472 if (blk_needs_flush_plug(tsk))
2473 blk_schedule_flush_plug(tsk);
2476 asmlinkage void __sched schedule(void)
2478 struct task_struct *tsk = current;
2480 sched_submit_work(tsk);
2483 EXPORT_SYMBOL(schedule);
2485 #ifdef CONFIG_CONTEXT_TRACKING
2486 asmlinkage void __sched schedule_user(void)
2489 * If we come here after a random call to set_need_resched(),
2490 * or we have been woken up remotely but the IPI has not yet arrived,
2491 * we haven't yet exited the RCU idle mode. Do it here manually until
2492 * we find a better solution.
2501 * schedule_preempt_disabled - called with preemption disabled
2503 * Returns with preemption disabled. Note: preempt_count must be 1
2505 void __sched schedule_preempt_disabled(void)
2507 sched_preempt_enable_no_resched();
2512 #ifdef CONFIG_PREEMPT
2514 * this is the entry point to schedule() from in-kernel preemption
2515 * off of preempt_enable. Kernel preemptions off return from interrupt
2516 * occur there and call schedule directly.
2518 asmlinkage void __sched notrace preempt_schedule(void)
2521 * If there is a non-zero preempt_count or interrupts are disabled,
2522 * we do not want to preempt the current task. Just return..
2524 if (likely(!preemptible()))
2528 add_preempt_count_notrace(PREEMPT_ACTIVE);
2530 sub_preempt_count_notrace(PREEMPT_ACTIVE);
2533 * Check again in case we missed a preemption opportunity
2534 * between schedule and now.
2537 } while (need_resched());
2539 EXPORT_SYMBOL(preempt_schedule);
2542 * this is the entry point to schedule() from kernel preemption
2543 * off of irq context.
2544 * Note, that this is called and return with irqs disabled. This will
2545 * protect us against recursive calling from irq.
2547 asmlinkage void __sched preempt_schedule_irq(void)
2549 enum ctx_state prev_state;
2551 /* Catch callers which need to be fixed */
2552 BUG_ON(preempt_count() || !irqs_disabled());
2554 prev_state = exception_enter();
2557 add_preempt_count(PREEMPT_ACTIVE);
2560 local_irq_disable();
2561 sub_preempt_count(PREEMPT_ACTIVE);
2564 * Check again in case we missed a preemption opportunity
2565 * between schedule and now.
2568 } while (need_resched());
2570 exception_exit(prev_state);
2573 #endif /* CONFIG_PREEMPT */
2575 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2578 return try_to_wake_up(curr->private, mode, wake_flags);
2580 EXPORT_SYMBOL(default_wake_function);
2583 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2584 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2585 * number) then we wake all the non-exclusive tasks and one exclusive task.
2587 * There are circumstances in which we can try to wake a task which has already
2588 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2589 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2591 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2592 int nr_exclusive, int wake_flags, void *key)
2594 wait_queue_t *curr, *next;
2596 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2597 unsigned flags = curr->flags;
2599 if (curr->func(curr, mode, wake_flags, key) &&
2600 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2606 * __wake_up - wake up threads blocked on a waitqueue.
2608 * @mode: which threads
2609 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2610 * @key: is directly passed to the wakeup function
2612 * It may be assumed that this function implies a write memory barrier before
2613 * changing the task state if and only if any tasks are woken up.
2615 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2616 int nr_exclusive, void *key)
2618 unsigned long flags;
2620 spin_lock_irqsave(&q->lock, flags);
2621 __wake_up_common(q, mode, nr_exclusive, 0, key);
2622 spin_unlock_irqrestore(&q->lock, flags);
2624 EXPORT_SYMBOL(__wake_up);
2627 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2629 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2631 __wake_up_common(q, mode, nr, 0, NULL);
2633 EXPORT_SYMBOL_GPL(__wake_up_locked);
2635 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2637 __wake_up_common(q, mode, 1, 0, key);
2639 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2642 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2644 * @mode: which threads
2645 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2646 * @key: opaque value to be passed to wakeup targets
2648 * The sync wakeup differs that the waker knows that it will schedule
2649 * away soon, so while the target thread will be woken up, it will not
2650 * be migrated to another CPU - ie. the two threads are 'synchronized'
2651 * with each other. This can prevent needless bouncing between CPUs.
2653 * On UP it can prevent extra preemption.
2655 * It may be assumed that this function implies a write memory barrier before
2656 * changing the task state if and only if any tasks are woken up.
2658 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2659 int nr_exclusive, void *key)
2661 unsigned long flags;
2662 int wake_flags = WF_SYNC;
2667 if (unlikely(nr_exclusive != 1))
2670 spin_lock_irqsave(&q->lock, flags);
2671 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2672 spin_unlock_irqrestore(&q->lock, flags);
2674 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2677 * __wake_up_sync - see __wake_up_sync_key()
2679 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2681 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2683 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2686 * complete: - signals a single thread waiting on this completion
2687 * @x: holds the state of this particular completion
2689 * This will wake up a single thread waiting on this completion. Threads will be
2690 * awakened in the same order in which they were queued.
2692 * See also complete_all(), wait_for_completion() and related routines.
2694 * It may be assumed that this function implies a write memory barrier before
2695 * changing the task state if and only if any tasks are woken up.
2697 void complete(struct completion *x)
2699 unsigned long flags;
2701 spin_lock_irqsave(&x->wait.lock, flags);
2703 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2704 spin_unlock_irqrestore(&x->wait.lock, flags);
2706 EXPORT_SYMBOL(complete);
2709 * complete_all: - signals all threads waiting on this completion
2710 * @x: holds the state of this particular completion
2712 * This will wake up all threads waiting on this particular completion event.
2714 * It may be assumed that this function implies a write memory barrier before
2715 * changing the task state if and only if any tasks are woken up.
2717 void complete_all(struct completion *x)
2719 unsigned long flags;
2721 spin_lock_irqsave(&x->wait.lock, flags);
2722 x->done += UINT_MAX/2;
2723 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2724 spin_unlock_irqrestore(&x->wait.lock, flags);
2726 EXPORT_SYMBOL(complete_all);
2728 static inline long __sched
2729 do_wait_for_common(struct completion *x,
2730 long (*action)(long), long timeout, int state)
2733 DECLARE_WAITQUEUE(wait, current);
2735 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2737 if (signal_pending_state(state, current)) {
2738 timeout = -ERESTARTSYS;
2741 __set_current_state(state);
2742 spin_unlock_irq(&x->wait.lock);
2743 timeout = action(timeout);
2744 spin_lock_irq(&x->wait.lock);
2745 } while (!x->done && timeout);
2746 __remove_wait_queue(&x->wait, &wait);
2751 return timeout ?: 1;
2754 static inline long __sched
2755 __wait_for_common(struct completion *x,
2756 long (*action)(long), long timeout, int state)
2760 spin_lock_irq(&x->wait.lock);
2761 timeout = do_wait_for_common(x, action, timeout, state);
2762 spin_unlock_irq(&x->wait.lock);
2767 wait_for_common(struct completion *x, long timeout, int state)
2769 return __wait_for_common(x, schedule_timeout, timeout, state);
2773 wait_for_common_io(struct completion *x, long timeout, int state)
2775 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2779 * wait_for_completion: - waits for completion of a task
2780 * @x: holds the state of this particular completion
2782 * This waits to be signaled for completion of a specific task. It is NOT
2783 * interruptible and there is no timeout.
2785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2786 * and interrupt capability. Also see complete().
2788 void __sched wait_for_completion(struct completion *x)
2790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2792 EXPORT_SYMBOL(wait_for_completion);
2795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2796 * @x: holds the state of this particular completion
2797 * @timeout: timeout value in jiffies
2799 * This waits for either a completion of a specific task to be signaled or for a
2800 * specified timeout to expire. The timeout is in jiffies. It is not
2803 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2804 * till timeout) if completed.
2806 unsigned long __sched
2807 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2811 EXPORT_SYMBOL(wait_for_completion_timeout);
2814 * wait_for_completion_io: - waits for completion of a task
2815 * @x: holds the state of this particular completion
2817 * This waits to be signaled for completion of a specific task. It is NOT
2818 * interruptible and there is no timeout. The caller is accounted as waiting
2821 void __sched wait_for_completion_io(struct completion *x)
2823 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2825 EXPORT_SYMBOL(wait_for_completion_io);
2828 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2829 * @x: holds the state of this particular completion
2830 * @timeout: timeout value in jiffies
2832 * This waits for either a completion of a specific task to be signaled or for a
2833 * specified timeout to expire. The timeout is in jiffies. It is not
2834 * interruptible. The caller is accounted as waiting for IO.
2836 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2837 * till timeout) if completed.
2839 unsigned long __sched
2840 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2842 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2844 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2847 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2848 * @x: holds the state of this particular completion
2850 * This waits for completion of a specific task to be signaled. It is
2853 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2855 int __sched wait_for_completion_interruptible(struct completion *x)
2857 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2858 if (t == -ERESTARTSYS)
2862 EXPORT_SYMBOL(wait_for_completion_interruptible);
2865 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2866 * @x: holds the state of this particular completion
2867 * @timeout: timeout value in jiffies
2869 * This waits for either a completion of a specific task to be signaled or for a
2870 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2872 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2873 * or number of jiffies left till timeout) if completed.
2876 wait_for_completion_interruptible_timeout(struct completion *x,
2877 unsigned long timeout)
2879 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2881 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2884 * wait_for_completion_killable: - waits for completion of a task (killable)
2885 * @x: holds the state of this particular completion
2887 * This waits to be signaled for completion of a specific task. It can be
2888 * interrupted by a kill signal.
2890 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2892 int __sched wait_for_completion_killable(struct completion *x)
2894 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2895 if (t == -ERESTARTSYS)
2899 EXPORT_SYMBOL(wait_for_completion_killable);
2902 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2903 * @x: holds the state of this particular completion
2904 * @timeout: timeout value in jiffies
2906 * This waits for either a completion of a specific task to be
2907 * signaled or for a specified timeout to expire. It can be
2908 * interrupted by a kill signal. The timeout is in jiffies.
2910 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2911 * or number of jiffies left till timeout) if completed.
2914 wait_for_completion_killable_timeout(struct completion *x,
2915 unsigned long timeout)
2917 return wait_for_common(x, timeout, TASK_KILLABLE);
2919 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2922 * try_wait_for_completion - try to decrement a completion without blocking
2923 * @x: completion structure
2925 * Return: 0 if a decrement cannot be done without blocking
2926 * 1 if a decrement succeeded.
2928 * If a completion is being used as a counting completion,
2929 * attempt to decrement the counter without blocking. This
2930 * enables us to avoid waiting if the resource the completion
2931 * is protecting is not available.
2933 bool try_wait_for_completion(struct completion *x)
2935 unsigned long flags;
2938 spin_lock_irqsave(&x->wait.lock, flags);
2943 spin_unlock_irqrestore(&x->wait.lock, flags);
2946 EXPORT_SYMBOL(try_wait_for_completion);
2949 * completion_done - Test to see if a completion has any waiters
2950 * @x: completion structure
2952 * Return: 0 if there are waiters (wait_for_completion() in progress)
2953 * 1 if there are no waiters.
2956 bool completion_done(struct completion *x)
2958 unsigned long flags;
2961 spin_lock_irqsave(&x->wait.lock, flags);
2964 spin_unlock_irqrestore(&x->wait.lock, flags);
2967 EXPORT_SYMBOL(completion_done);
2970 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2972 unsigned long flags;
2975 init_waitqueue_entry(&wait, current);
2977 __set_current_state(state);
2979 spin_lock_irqsave(&q->lock, flags);
2980 __add_wait_queue(q, &wait);
2981 spin_unlock(&q->lock);
2982 timeout = schedule_timeout(timeout);
2983 spin_lock_irq(&q->lock);
2984 __remove_wait_queue(q, &wait);
2985 spin_unlock_irqrestore(&q->lock, flags);
2990 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2992 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2994 EXPORT_SYMBOL(interruptible_sleep_on);
2997 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2999 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3001 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3003 void __sched sleep_on(wait_queue_head_t *q)
3005 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3007 EXPORT_SYMBOL(sleep_on);
3009 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3011 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3013 EXPORT_SYMBOL(sleep_on_timeout);
3015 #ifdef CONFIG_RT_MUTEXES
3018 * rt_mutex_setprio - set the current priority of a task
3020 * @prio: prio value (kernel-internal form)
3022 * This function changes the 'effective' priority of a task. It does
3023 * not touch ->normal_prio like __setscheduler().
3025 * Used by the rt_mutex code to implement priority inheritance logic.
3027 void rt_mutex_setprio(struct task_struct *p, int prio)
3029 int oldprio, on_rq, running;
3031 const struct sched_class *prev_class;
3033 BUG_ON(prio < 0 || prio > MAX_PRIO);
3035 rq = __task_rq_lock(p);
3038 * Idle task boosting is a nono in general. There is one
3039 * exception, when PREEMPT_RT and NOHZ is active:
3041 * The idle task calls get_next_timer_interrupt() and holds
3042 * the timer wheel base->lock on the CPU and another CPU wants
3043 * to access the timer (probably to cancel it). We can safely
3044 * ignore the boosting request, as the idle CPU runs this code
3045 * with interrupts disabled and will complete the lock
3046 * protected section without being interrupted. So there is no
3047 * real need to boost.
3049 if (unlikely(p == rq->idle)) {
3050 WARN_ON(p != rq->curr);
3051 WARN_ON(p->pi_blocked_on);
3055 trace_sched_pi_setprio(p, prio);
3057 prev_class = p->sched_class;
3059 running = task_current(rq, p);
3061 dequeue_task(rq, p, 0);
3063 p->sched_class->put_prev_task(rq, p);
3066 p->sched_class = &rt_sched_class;
3068 p->sched_class = &fair_sched_class;
3073 p->sched_class->set_curr_task(rq);
3075 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3077 check_class_changed(rq, p, prev_class, oldprio);
3079 __task_rq_unlock(rq);
3082 void set_user_nice(struct task_struct *p, long nice)
3084 int old_prio, delta, on_rq;
3085 unsigned long flags;
3088 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3091 * We have to be careful, if called from sys_setpriority(),
3092 * the task might be in the middle of scheduling on another CPU.
3094 rq = task_rq_lock(p, &flags);
3096 * The RT priorities are set via sched_setscheduler(), but we still
3097 * allow the 'normal' nice value to be set - but as expected
3098 * it wont have any effect on scheduling until the task is
3099 * SCHED_FIFO/SCHED_RR:
3101 if (task_has_rt_policy(p)) {
3102 p->static_prio = NICE_TO_PRIO(nice);
3107 dequeue_task(rq, p, 0);
3109 p->static_prio = NICE_TO_PRIO(nice);
3112 p->prio = effective_prio(p);
3113 delta = p->prio - old_prio;
3116 enqueue_task(rq, p, 0);
3118 * If the task increased its priority or is running and
3119 * lowered its priority, then reschedule its CPU:
3121 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3122 resched_task(rq->curr);
3125 task_rq_unlock(rq, p, &flags);
3127 EXPORT_SYMBOL(set_user_nice);
3130 * can_nice - check if a task can reduce its nice value
3134 int can_nice(const struct task_struct *p, const int nice)
3136 /* convert nice value [19,-20] to rlimit style value [1,40] */
3137 int nice_rlim = 20 - nice;
3139 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3140 capable(CAP_SYS_NICE));
3143 #ifdef __ARCH_WANT_SYS_NICE
3146 * sys_nice - change the priority of the current process.
3147 * @increment: priority increment
3149 * sys_setpriority is a more generic, but much slower function that
3150 * does similar things.
3152 SYSCALL_DEFINE1(nice, int, increment)
3157 * Setpriority might change our priority at the same moment.
3158 * We don't have to worry. Conceptually one call occurs first
3159 * and we have a single winner.
3161 if (increment < -40)
3166 nice = TASK_NICE(current) + increment;
3172 if (increment < 0 && !can_nice(current, nice))
3175 retval = security_task_setnice(current, nice);
3179 set_user_nice(current, nice);
3186 * task_prio - return the priority value of a given task.
3187 * @p: the task in question.
3189 * Return: The priority value as seen by users in /proc.
3190 * RT tasks are offset by -200. Normal tasks are centered
3191 * around 0, value goes from -16 to +15.
3193 int task_prio(const struct task_struct *p)
3195 return p->prio - MAX_RT_PRIO;
3199 * task_nice - return the nice value of a given task.
3200 * @p: the task in question.
3202 * Return: The nice value [ -20 ... 0 ... 19 ].
3204 int task_nice(const struct task_struct *p)
3206 return TASK_NICE(p);
3208 EXPORT_SYMBOL(task_nice);
3211 * idle_cpu - is a given cpu idle currently?
3212 * @cpu: the processor in question.
3214 * Return: 1 if the CPU is currently idle. 0 otherwise.
3216 int idle_cpu(int cpu)
3218 struct rq *rq = cpu_rq(cpu);
3220 if (rq->curr != rq->idle)
3227 if (!llist_empty(&rq->wake_list))
3235 * idle_task - return the idle task for a given cpu.
3236 * @cpu: the processor in question.
3238 * Return: The idle task for the cpu @cpu.
3240 struct task_struct *idle_task(int cpu)
3242 return cpu_rq(cpu)->idle;
3246 * find_process_by_pid - find a process with a matching PID value.
3247 * @pid: the pid in question.
3249 * The task of @pid, if found. %NULL otherwise.
3251 static struct task_struct *find_process_by_pid(pid_t pid)
3253 return pid ? find_task_by_vpid(pid) : current;
3256 /* Actually do priority change: must hold rq lock. */
3258 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3261 p->rt_priority = prio;
3262 p->normal_prio = normal_prio(p);
3263 /* we are holding p->pi_lock already */
3264 p->prio = rt_mutex_getprio(p);
3265 if (rt_prio(p->prio))
3266 p->sched_class = &rt_sched_class;
3268 p->sched_class = &fair_sched_class;
3273 * check the target process has a UID that matches the current process's
3275 static bool check_same_owner(struct task_struct *p)
3277 const struct cred *cred = current_cred(), *pcred;
3281 pcred = __task_cred(p);
3282 match = (uid_eq(cred->euid, pcred->euid) ||
3283 uid_eq(cred->euid, pcred->uid));
3288 static int __sched_setscheduler(struct task_struct *p, int policy,
3289 const struct sched_param *param, bool user)
3291 int retval, oldprio, oldpolicy = -1, on_rq, running;
3292 unsigned long flags;
3293 const struct sched_class *prev_class;
3297 /* may grab non-irq protected spin_locks */
3298 BUG_ON(in_interrupt());
3300 /* double check policy once rq lock held */
3302 reset_on_fork = p->sched_reset_on_fork;
3303 policy = oldpolicy = p->policy;
3305 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3306 policy &= ~SCHED_RESET_ON_FORK;
3308 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3309 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3310 policy != SCHED_IDLE)
3315 * Valid priorities for SCHED_FIFO and SCHED_RR are
3316 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3317 * SCHED_BATCH and SCHED_IDLE is 0.
3319 if (param->sched_priority < 0 ||
3320 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3321 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3323 if (rt_policy(policy) != (param->sched_priority != 0))
3327 * Allow unprivileged RT tasks to decrease priority:
3329 if (user && !capable(CAP_SYS_NICE)) {
3330 if (rt_policy(policy)) {
3331 unsigned long rlim_rtprio =
3332 task_rlimit(p, RLIMIT_RTPRIO);
3334 /* can't set/change the rt policy */
3335 if (policy != p->policy && !rlim_rtprio)
3338 /* can't increase priority */
3339 if (param->sched_priority > p->rt_priority &&
3340 param->sched_priority > rlim_rtprio)
3345 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3346 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3348 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3349 if (!can_nice(p, TASK_NICE(p)))
3353 /* can't change other user's priorities */
3354 if (!check_same_owner(p))
3357 /* Normal users shall not reset the sched_reset_on_fork flag */
3358 if (p->sched_reset_on_fork && !reset_on_fork)
3363 retval = security_task_setscheduler(p);
3369 * make sure no PI-waiters arrive (or leave) while we are
3370 * changing the priority of the task:
3372 * To be able to change p->policy safely, the appropriate
3373 * runqueue lock must be held.
3375 rq = task_rq_lock(p, &flags);
3378 * Changing the policy of the stop threads its a very bad idea
3380 if (p == rq->stop) {
3381 task_rq_unlock(rq, p, &flags);
3386 * If not changing anything there's no need to proceed further:
3388 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3389 param->sched_priority == p->rt_priority))) {
3390 task_rq_unlock(rq, p, &flags);
3394 #ifdef CONFIG_RT_GROUP_SCHED
3397 * Do not allow realtime tasks into groups that have no runtime
3400 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3401 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3402 !task_group_is_autogroup(task_group(p))) {
3403 task_rq_unlock(rq, p, &flags);
3409 /* recheck policy now with rq lock held */
3410 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3411 policy = oldpolicy = -1;
3412 task_rq_unlock(rq, p, &flags);
3416 running = task_current(rq, p);
3418 dequeue_task(rq, p, 0);
3420 p->sched_class->put_prev_task(rq, p);
3422 p->sched_reset_on_fork = reset_on_fork;
3425 prev_class = p->sched_class;
3426 __setscheduler(rq, p, policy, param->sched_priority);
3429 p->sched_class->set_curr_task(rq);
3431 enqueue_task(rq, p, 0);
3433 check_class_changed(rq, p, prev_class, oldprio);
3434 task_rq_unlock(rq, p, &flags);
3436 rt_mutex_adjust_pi(p);
3442 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3443 * @p: the task in question.
3444 * @policy: new policy.
3445 * @param: structure containing the new RT priority.
3447 * Return: 0 on success. An error code otherwise.
3449 * NOTE that the task may be already dead.
3451 int sched_setscheduler(struct task_struct *p, int policy,
3452 const struct sched_param *param)
3454 return __sched_setscheduler(p, policy, param, true);
3456 EXPORT_SYMBOL_GPL(sched_setscheduler);
3459 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3460 * @p: the task in question.
3461 * @policy: new policy.
3462 * @param: structure containing the new RT priority.
3464 * Just like sched_setscheduler, only don't bother checking if the
3465 * current context has permission. For example, this is needed in
3466 * stop_machine(): we create temporary high priority worker threads,
3467 * but our caller might not have that capability.
3469 * Return: 0 on success. An error code otherwise.
3471 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3472 const struct sched_param *param)
3474 return __sched_setscheduler(p, policy, param, false);
3478 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3480 struct sched_param lparam;
3481 struct task_struct *p;
3484 if (!param || pid < 0)
3486 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3491 p = find_process_by_pid(pid);
3493 retval = sched_setscheduler(p, policy, &lparam);
3500 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3501 * @pid: the pid in question.
3502 * @policy: new policy.
3503 * @param: structure containing the new RT priority.
3505 * Return: 0 on success. An error code otherwise.
3507 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3508 struct sched_param __user *, param)
3510 /* negative values for policy are not valid */
3514 return do_sched_setscheduler(pid, policy, param);
3518 * sys_sched_setparam - set/change the RT priority of a thread
3519 * @pid: the pid in question.
3520 * @param: structure containing the new RT priority.
3522 * Return: 0 on success. An error code otherwise.
3524 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3526 return do_sched_setscheduler(pid, -1, param);
3530 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3531 * @pid: the pid in question.
3533 * Return: On success, the policy of the thread. Otherwise, a negative error
3536 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3538 struct task_struct *p;
3546 p = find_process_by_pid(pid);
3548 retval = security_task_getscheduler(p);
3551 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3558 * sys_sched_getparam - get the RT priority of a thread
3559 * @pid: the pid in question.
3560 * @param: structure containing the RT priority.
3562 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3565 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3567 struct sched_param lp;
3568 struct task_struct *p;
3571 if (!param || pid < 0)
3575 p = find_process_by_pid(pid);
3580 retval = security_task_getscheduler(p);
3584 lp.sched_priority = p->rt_priority;
3588 * This one might sleep, we cannot do it with a spinlock held ...
3590 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3599 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3601 cpumask_var_t cpus_allowed, new_mask;
3602 struct task_struct *p;
3608 p = find_process_by_pid(pid);
3615 /* Prevent p going away */
3619 if (p->flags & PF_NO_SETAFFINITY) {
3623 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3627 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3629 goto out_free_cpus_allowed;
3632 if (!check_same_owner(p)) {
3634 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3641 retval = security_task_setscheduler(p);
3645 cpuset_cpus_allowed(p, cpus_allowed);
3646 cpumask_and(new_mask, in_mask, cpus_allowed);
3648 retval = set_cpus_allowed_ptr(p, new_mask);
3651 cpuset_cpus_allowed(p, cpus_allowed);
3652 if (!cpumask_subset(new_mask, cpus_allowed)) {
3654 * We must have raced with a concurrent cpuset
3655 * update. Just reset the cpus_allowed to the
3656 * cpuset's cpus_allowed
3658 cpumask_copy(new_mask, cpus_allowed);
3663 free_cpumask_var(new_mask);
3664 out_free_cpus_allowed:
3665 free_cpumask_var(cpus_allowed);
3672 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3673 struct cpumask *new_mask)
3675 if (len < cpumask_size())
3676 cpumask_clear(new_mask);
3677 else if (len > cpumask_size())
3678 len = cpumask_size();
3680 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3684 * sys_sched_setaffinity - set the cpu affinity of a process
3685 * @pid: pid of the process
3686 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3687 * @user_mask_ptr: user-space pointer to the new cpu mask
3689 * Return: 0 on success. An error code otherwise.
3691 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3692 unsigned long __user *, user_mask_ptr)
3694 cpumask_var_t new_mask;
3697 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3700 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3702 retval = sched_setaffinity(pid, new_mask);
3703 free_cpumask_var(new_mask);
3707 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3709 struct task_struct *p;
3710 unsigned long flags;
3717 p = find_process_by_pid(pid);
3721 retval = security_task_getscheduler(p);
3725 raw_spin_lock_irqsave(&p->pi_lock, flags);
3726 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3727 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3737 * sys_sched_getaffinity - get the cpu affinity of a process
3738 * @pid: pid of the process
3739 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3740 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3742 * Return: 0 on success. An error code otherwise.
3744 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3745 unsigned long __user *, user_mask_ptr)
3750 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3752 if (len & (sizeof(unsigned long)-1))
3755 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3758 ret = sched_getaffinity(pid, mask);
3760 size_t retlen = min_t(size_t, len, cpumask_size());
3762 if (copy_to_user(user_mask_ptr, mask, retlen))
3767 free_cpumask_var(mask);
3773 * sys_sched_yield - yield the current processor to other threads.
3775 * This function yields the current CPU to other tasks. If there are no
3776 * other threads running on this CPU then this function will return.
3780 SYSCALL_DEFINE0(sched_yield)
3782 struct rq *rq = this_rq_lock();
3784 schedstat_inc(rq, yld_count);
3785 current->sched_class->yield_task(rq);
3788 * Since we are going to call schedule() anyway, there's
3789 * no need to preempt or enable interrupts:
3791 __release(rq->lock);
3792 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3793 do_raw_spin_unlock(&rq->lock);
3794 sched_preempt_enable_no_resched();
3801 static inline int should_resched(void)
3803 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
3806 static void __cond_resched(void)
3808 add_preempt_count(PREEMPT_ACTIVE);
3810 sub_preempt_count(PREEMPT_ACTIVE);
3813 int __sched _cond_resched(void)
3815 if (should_resched()) {
3821 EXPORT_SYMBOL(_cond_resched);
3824 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3825 * call schedule, and on return reacquire the lock.
3827 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3828 * operations here to prevent schedule() from being called twice (once via
3829 * spin_unlock(), once by hand).
3831 int __cond_resched_lock(spinlock_t *lock)
3833 int resched = should_resched();
3836 lockdep_assert_held(lock);
3838 if (spin_needbreak(lock) || resched) {
3849 EXPORT_SYMBOL(__cond_resched_lock);
3851 int __sched __cond_resched_softirq(void)
3853 BUG_ON(!in_softirq());
3855 if (should_resched()) {
3863 EXPORT_SYMBOL(__cond_resched_softirq);
3866 * yield - yield the current processor to other threads.
3868 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3870 * The scheduler is at all times free to pick the calling task as the most
3871 * eligible task to run, if removing the yield() call from your code breaks
3872 * it, its already broken.
3874 * Typical broken usage is:
3879 * where one assumes that yield() will let 'the other' process run that will
3880 * make event true. If the current task is a SCHED_FIFO task that will never
3881 * happen. Never use yield() as a progress guarantee!!
3883 * If you want to use yield() to wait for something, use wait_event().
3884 * If you want to use yield() to be 'nice' for others, use cond_resched().
3885 * If you still want to use yield(), do not!
3887 void __sched yield(void)
3889 set_current_state(TASK_RUNNING);
3892 EXPORT_SYMBOL(yield);
3895 * yield_to - yield the current processor to another thread in
3896 * your thread group, or accelerate that thread toward the
3897 * processor it's on.
3899 * @preempt: whether task preemption is allowed or not
3901 * It's the caller's job to ensure that the target task struct
3902 * can't go away on us before we can do any checks.
3905 * true (>0) if we indeed boosted the target task.
3906 * false (0) if we failed to boost the target.
3907 * -ESRCH if there's no task to yield to.
3909 bool __sched yield_to(struct task_struct *p, bool preempt)
3911 struct task_struct *curr = current;
3912 struct rq *rq, *p_rq;
3913 unsigned long flags;
3916 local_irq_save(flags);
3922 * If we're the only runnable task on the rq and target rq also
3923 * has only one task, there's absolutely no point in yielding.
3925 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3930 double_rq_lock(rq, p_rq);
3931 while (task_rq(p) != p_rq) {
3932 double_rq_unlock(rq, p_rq);
3936 if (!curr->sched_class->yield_to_task)
3939 if (curr->sched_class != p->sched_class)
3942 if (task_running(p_rq, p) || p->state)
3945 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3947 schedstat_inc(rq, yld_count);
3949 * Make p's CPU reschedule; pick_next_entity takes care of
3952 if (preempt && rq != p_rq)
3953 resched_task(p_rq->curr);
3957 double_rq_unlock(rq, p_rq);
3959 local_irq_restore(flags);
3966 EXPORT_SYMBOL_GPL(yield_to);
3969 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3970 * that process accounting knows that this is a task in IO wait state.
3972 void __sched io_schedule(void)
3974 struct rq *rq = raw_rq();
3976 delayacct_blkio_start();
3977 atomic_inc(&rq->nr_iowait);
3978 blk_flush_plug(current);
3979 current->in_iowait = 1;
3981 current->in_iowait = 0;
3982 atomic_dec(&rq->nr_iowait);
3983 delayacct_blkio_end();
3985 EXPORT_SYMBOL(io_schedule);
3987 long __sched io_schedule_timeout(long timeout)
3989 struct rq *rq = raw_rq();
3992 delayacct_blkio_start();
3993 atomic_inc(&rq->nr_iowait);
3994 blk_flush_plug(current);
3995 current->in_iowait = 1;
3996 ret = schedule_timeout(timeout);
3997 current->in_iowait = 0;
3998 atomic_dec(&rq->nr_iowait);
3999 delayacct_blkio_end();
4004 * sys_sched_get_priority_max - return maximum RT priority.
4005 * @policy: scheduling class.
4007 * Return: On success, this syscall returns the maximum
4008 * rt_priority that can be used by a given scheduling class.
4009 * On failure, a negative error code is returned.
4011 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4018 ret = MAX_USER_RT_PRIO-1;
4030 * sys_sched_get_priority_min - return minimum RT priority.
4031 * @policy: scheduling class.
4033 * Return: On success, this syscall returns the minimum
4034 * rt_priority that can be used by a given scheduling class.
4035 * On failure, a negative error code is returned.
4037 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4055 * sys_sched_rr_get_interval - return the default timeslice of a process.
4056 * @pid: pid of the process.
4057 * @interval: userspace pointer to the timeslice value.
4059 * this syscall writes the default timeslice value of a given process
4060 * into the user-space timespec buffer. A value of '0' means infinity.
4062 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4065 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4066 struct timespec __user *, interval)
4068 struct task_struct *p;
4069 unsigned int time_slice;
4070 unsigned long flags;
4080 p = find_process_by_pid(pid);
4084 retval = security_task_getscheduler(p);
4088 rq = task_rq_lock(p, &flags);
4089 time_slice = p->sched_class->get_rr_interval(rq, p);
4090 task_rq_unlock(rq, p, &flags);
4093 jiffies_to_timespec(time_slice, &t);
4094 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4102 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4104 void sched_show_task(struct task_struct *p)
4106 unsigned long free = 0;
4110 state = p->state ? __ffs(p->state) + 1 : 0;
4111 printk(KERN_INFO "%-15.15s %c", p->comm,
4112 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4113 #if BITS_PER_LONG == 32
4114 if (state == TASK_RUNNING)
4115 printk(KERN_CONT " running ");
4117 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4119 if (state == TASK_RUNNING)
4120 printk(KERN_CONT " running task ");
4122 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4124 #ifdef CONFIG_DEBUG_STACK_USAGE
4125 free = stack_not_used(p);
4128 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4130 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4131 task_pid_nr(p), ppid,
4132 (unsigned long)task_thread_info(p)->flags);
4134 print_worker_info(KERN_INFO, p);
4135 show_stack(p, NULL);
4138 void show_state_filter(unsigned long state_filter)
4140 struct task_struct *g, *p;
4142 #if BITS_PER_LONG == 32
4144 " task PC stack pid father\n");
4147 " task PC stack pid father\n");
4150 do_each_thread(g, p) {
4152 * reset the NMI-timeout, listing all files on a slow
4153 * console might take a lot of time:
4155 touch_nmi_watchdog();
4156 if (!state_filter || (p->state & state_filter))
4158 } while_each_thread(g, p);
4160 touch_all_softlockup_watchdogs();
4162 #ifdef CONFIG_SCHED_DEBUG
4163 sysrq_sched_debug_show();
4167 * Only show locks if all tasks are dumped:
4170 debug_show_all_locks();
4173 void init_idle_bootup_task(struct task_struct *idle)
4175 idle->sched_class = &idle_sched_class;
4179 * init_idle - set up an idle thread for a given CPU
4180 * @idle: task in question
4181 * @cpu: cpu the idle task belongs to
4183 * NOTE: this function does not set the idle thread's NEED_RESCHED
4184 * flag, to make booting more robust.
4186 void init_idle(struct task_struct *idle, int cpu)
4188 struct rq *rq = cpu_rq(cpu);
4189 unsigned long flags;
4191 raw_spin_lock_irqsave(&rq->lock, flags);
4194 idle->state = TASK_RUNNING;
4195 idle->se.exec_start = sched_clock();
4197 do_set_cpus_allowed(idle, cpumask_of(cpu));
4199 * We're having a chicken and egg problem, even though we are
4200 * holding rq->lock, the cpu isn't yet set to this cpu so the
4201 * lockdep check in task_group() will fail.
4203 * Similar case to sched_fork(). / Alternatively we could
4204 * use task_rq_lock() here and obtain the other rq->lock.
4209 __set_task_cpu(idle, cpu);
4212 rq->curr = rq->idle = idle;
4213 #if defined(CONFIG_SMP)
4216 raw_spin_unlock_irqrestore(&rq->lock, flags);
4218 /* Set the preempt count _outside_ the spinlocks! */
4219 init_idle_preempt_count(idle, cpu);
4222 * The idle tasks have their own, simple scheduling class:
4224 idle->sched_class = &idle_sched_class;
4225 ftrace_graph_init_idle_task(idle, cpu);
4226 vtime_init_idle(idle, cpu);
4227 #if defined(CONFIG_SMP)
4228 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4233 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4235 if (p->sched_class && p->sched_class->set_cpus_allowed)
4236 p->sched_class->set_cpus_allowed(p, new_mask);
4238 cpumask_copy(&p->cpus_allowed, new_mask);
4239 p->nr_cpus_allowed = cpumask_weight(new_mask);
4243 * This is how migration works:
4245 * 1) we invoke migration_cpu_stop() on the target CPU using
4247 * 2) stopper starts to run (implicitly forcing the migrated thread
4249 * 3) it checks whether the migrated task is still in the wrong runqueue.
4250 * 4) if it's in the wrong runqueue then the migration thread removes
4251 * it and puts it into the right queue.
4252 * 5) stopper completes and stop_one_cpu() returns and the migration
4257 * Change a given task's CPU affinity. Migrate the thread to a
4258 * proper CPU and schedule it away if the CPU it's executing on
4259 * is removed from the allowed bitmask.
4261 * NOTE: the caller must have a valid reference to the task, the
4262 * task must not exit() & deallocate itself prematurely. The
4263 * call is not atomic; no spinlocks may be held.
4265 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4267 unsigned long flags;
4269 unsigned int dest_cpu;
4272 rq = task_rq_lock(p, &flags);
4274 if (cpumask_equal(&p->cpus_allowed, new_mask))
4277 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4282 do_set_cpus_allowed(p, new_mask);
4284 /* Can the task run on the task's current CPU? If so, we're done */
4285 if (cpumask_test_cpu(task_cpu(p), new_mask))
4288 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4290 struct migration_arg arg = { p, dest_cpu };
4291 /* Need help from migration thread: drop lock and wait. */
4292 task_rq_unlock(rq, p, &flags);
4293 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4294 tlb_migrate_finish(p->mm);
4298 task_rq_unlock(rq, p, &flags);
4302 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4305 * Move (not current) task off this cpu, onto dest cpu. We're doing
4306 * this because either it can't run here any more (set_cpus_allowed()
4307 * away from this CPU, or CPU going down), or because we're
4308 * attempting to rebalance this task on exec (sched_exec).
4310 * So we race with normal scheduler movements, but that's OK, as long
4311 * as the task is no longer on this CPU.
4313 * Returns non-zero if task was successfully migrated.
4315 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4317 struct rq *rq_dest, *rq_src;
4320 if (unlikely(!cpu_active(dest_cpu)))
4323 rq_src = cpu_rq(src_cpu);
4324 rq_dest = cpu_rq(dest_cpu);
4326 raw_spin_lock(&p->pi_lock);
4327 double_rq_lock(rq_src, rq_dest);
4328 /* Already moved. */
4329 if (task_cpu(p) != src_cpu)
4331 /* Affinity changed (again). */
4332 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4336 * If we're not on a rq, the next wake-up will ensure we're
4340 dequeue_task(rq_src, p, 0);
4341 set_task_cpu(p, dest_cpu);
4342 enqueue_task(rq_dest, p, 0);
4343 check_preempt_curr(rq_dest, p, 0);
4348 double_rq_unlock(rq_src, rq_dest);
4349 raw_spin_unlock(&p->pi_lock);
4354 * migration_cpu_stop - this will be executed by a highprio stopper thread
4355 * and performs thread migration by bumping thread off CPU then
4356 * 'pushing' onto another runqueue.
4358 static int migration_cpu_stop(void *data)
4360 struct migration_arg *arg = data;
4363 * The original target cpu might have gone down and we might
4364 * be on another cpu but it doesn't matter.
4366 local_irq_disable();
4367 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4372 #ifdef CONFIG_HOTPLUG_CPU
4375 * Ensures that the idle task is using init_mm right before its cpu goes
4378 void idle_task_exit(void)
4380 struct mm_struct *mm = current->active_mm;
4382 BUG_ON(cpu_online(smp_processor_id()));
4385 switch_mm(mm, &init_mm, current);
4390 * Since this CPU is going 'away' for a while, fold any nr_active delta
4391 * we might have. Assumes we're called after migrate_tasks() so that the
4392 * nr_active count is stable.
4394 * Also see the comment "Global load-average calculations".
4396 static void calc_load_migrate(struct rq *rq)
4398 long delta = calc_load_fold_active(rq);
4400 atomic_long_add(delta, &calc_load_tasks);
4404 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4405 * try_to_wake_up()->select_task_rq().
4407 * Called with rq->lock held even though we'er in stop_machine() and
4408 * there's no concurrency possible, we hold the required locks anyway
4409 * because of lock validation efforts.
4411 static void migrate_tasks(unsigned int dead_cpu)
4413 struct rq *rq = cpu_rq(dead_cpu);
4414 struct task_struct *next, *stop = rq->stop;
4418 * Fudge the rq selection such that the below task selection loop
4419 * doesn't get stuck on the currently eligible stop task.
4421 * We're currently inside stop_machine() and the rq is either stuck
4422 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4423 * either way we should never end up calling schedule() until we're
4429 * put_prev_task() and pick_next_task() sched
4430 * class method both need to have an up-to-date
4431 * value of rq->clock[_task]
4433 update_rq_clock(rq);
4437 * There's this thread running, bail when that's the only
4440 if (rq->nr_running == 1)
4443 next = pick_next_task(rq);
4445 next->sched_class->put_prev_task(rq, next);
4447 /* Find suitable destination for @next, with force if needed. */
4448 dest_cpu = select_fallback_rq(dead_cpu, next);
4449 raw_spin_unlock(&rq->lock);
4451 __migrate_task(next, dead_cpu, dest_cpu);
4453 raw_spin_lock(&rq->lock);
4459 #endif /* CONFIG_HOTPLUG_CPU */
4461 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4463 static struct ctl_table sd_ctl_dir[] = {
4465 .procname = "sched_domain",
4471 static struct ctl_table sd_ctl_root[] = {
4473 .procname = "kernel",
4475 .child = sd_ctl_dir,
4480 static struct ctl_table *sd_alloc_ctl_entry(int n)
4482 struct ctl_table *entry =
4483 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4488 static void sd_free_ctl_entry(struct ctl_table **tablep)
4490 struct ctl_table *entry;
4493 * In the intermediate directories, both the child directory and
4494 * procname are dynamically allocated and could fail but the mode
4495 * will always be set. In the lowest directory the names are
4496 * static strings and all have proc handlers.
4498 for (entry = *tablep; entry->mode; entry++) {
4500 sd_free_ctl_entry(&entry->child);
4501 if (entry->proc_handler == NULL)
4502 kfree(entry->procname);
4509 static int min_load_idx = 0;
4510 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4513 set_table_entry(struct ctl_table *entry,
4514 const char *procname, void *data, int maxlen,
4515 umode_t mode, proc_handler *proc_handler,
4518 entry->procname = procname;
4520 entry->maxlen = maxlen;
4522 entry->proc_handler = proc_handler;
4525 entry->extra1 = &min_load_idx;
4526 entry->extra2 = &max_load_idx;
4530 static struct ctl_table *
4531 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4533 struct ctl_table *table = sd_alloc_ctl_entry(13);
4538 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4539 sizeof(long), 0644, proc_doulongvec_minmax, false);
4540 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4541 sizeof(long), 0644, proc_doulongvec_minmax, false);
4542 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4543 sizeof(int), 0644, proc_dointvec_minmax, true);
4544 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4545 sizeof(int), 0644, proc_dointvec_minmax, true);
4546 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4547 sizeof(int), 0644, proc_dointvec_minmax, true);
4548 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4549 sizeof(int), 0644, proc_dointvec_minmax, true);
4550 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4551 sizeof(int), 0644, proc_dointvec_minmax, true);
4552 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4553 sizeof(int), 0644, proc_dointvec_minmax, false);
4554 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4555 sizeof(int), 0644, proc_dointvec_minmax, false);
4556 set_table_entry(&table[9], "cache_nice_tries",
4557 &sd->cache_nice_tries,
4558 sizeof(int), 0644, proc_dointvec_minmax, false);
4559 set_table_entry(&table[10], "flags", &sd->flags,
4560 sizeof(int), 0644, proc_dointvec_minmax, false);
4561 set_table_entry(&table[11], "name", sd->name,
4562 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4563 /* &table[12] is terminator */
4568 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4570 struct ctl_table *entry, *table;
4571 struct sched_domain *sd;
4572 int domain_num = 0, i;
4575 for_each_domain(cpu, sd)
4577 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4582 for_each_domain(cpu, sd) {
4583 snprintf(buf, 32, "domain%d", i);
4584 entry->procname = kstrdup(buf, GFP_KERNEL);
4586 entry->child = sd_alloc_ctl_domain_table(sd);
4593 static struct ctl_table_header *sd_sysctl_header;
4594 static void register_sched_domain_sysctl(void)
4596 int i, cpu_num = num_possible_cpus();
4597 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4600 WARN_ON(sd_ctl_dir[0].child);
4601 sd_ctl_dir[0].child = entry;
4606 for_each_possible_cpu(i) {
4607 snprintf(buf, 32, "cpu%d", i);
4608 entry->procname = kstrdup(buf, GFP_KERNEL);
4610 entry->child = sd_alloc_ctl_cpu_table(i);
4614 WARN_ON(sd_sysctl_header);
4615 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4618 /* may be called multiple times per register */
4619 static void unregister_sched_domain_sysctl(void)
4621 if (sd_sysctl_header)
4622 unregister_sysctl_table(sd_sysctl_header);
4623 sd_sysctl_header = NULL;
4624 if (sd_ctl_dir[0].child)
4625 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4628 static void register_sched_domain_sysctl(void)
4631 static void unregister_sched_domain_sysctl(void)
4636 static void set_rq_online(struct rq *rq)
4639 const struct sched_class *class;
4641 cpumask_set_cpu(rq->cpu, rq->rd->online);
4644 for_each_class(class) {
4645 if (class->rq_online)
4646 class->rq_online(rq);
4651 static void set_rq_offline(struct rq *rq)
4654 const struct sched_class *class;
4656 for_each_class(class) {
4657 if (class->rq_offline)
4658 class->rq_offline(rq);
4661 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4667 * migration_call - callback that gets triggered when a CPU is added.
4668 * Here we can start up the necessary migration thread for the new CPU.
4671 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4673 int cpu = (long)hcpu;
4674 unsigned long flags;
4675 struct rq *rq = cpu_rq(cpu);
4677 switch (action & ~CPU_TASKS_FROZEN) {
4679 case CPU_UP_PREPARE:
4680 rq->calc_load_update = calc_load_update;
4684 /* Update our root-domain */
4685 raw_spin_lock_irqsave(&rq->lock, flags);
4687 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4691 raw_spin_unlock_irqrestore(&rq->lock, flags);
4694 #ifdef CONFIG_HOTPLUG_CPU
4696 sched_ttwu_pending();
4697 /* Update our root-domain */
4698 raw_spin_lock_irqsave(&rq->lock, flags);
4700 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4704 BUG_ON(rq->nr_running != 1); /* the migration thread */
4705 raw_spin_unlock_irqrestore(&rq->lock, flags);
4709 calc_load_migrate(rq);
4714 update_max_interval();
4720 * Register at high priority so that task migration (migrate_all_tasks)
4721 * happens before everything else. This has to be lower priority than
4722 * the notifier in the perf_event subsystem, though.
4724 static struct notifier_block migration_notifier = {
4725 .notifier_call = migration_call,
4726 .priority = CPU_PRI_MIGRATION,
4729 static int sched_cpu_active(struct notifier_block *nfb,
4730 unsigned long action, void *hcpu)
4732 switch (action & ~CPU_TASKS_FROZEN) {
4734 case CPU_DOWN_FAILED:
4735 set_cpu_active((long)hcpu, true);
4742 static int sched_cpu_inactive(struct notifier_block *nfb,
4743 unsigned long action, void *hcpu)
4745 switch (action & ~CPU_TASKS_FROZEN) {
4746 case CPU_DOWN_PREPARE:
4747 set_cpu_active((long)hcpu, false);
4754 static int __init migration_init(void)
4756 void *cpu = (void *)(long)smp_processor_id();
4759 /* Initialize migration for the boot CPU */
4760 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4761 BUG_ON(err == NOTIFY_BAD);
4762 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4763 register_cpu_notifier(&migration_notifier);
4765 /* Register cpu active notifiers */
4766 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4767 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4771 early_initcall(migration_init);
4776 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4778 #ifdef CONFIG_SCHED_DEBUG
4780 static __read_mostly int sched_debug_enabled;
4782 static int __init sched_debug_setup(char *str)
4784 sched_debug_enabled = 1;
4788 early_param("sched_debug", sched_debug_setup);
4790 static inline bool sched_debug(void)
4792 return sched_debug_enabled;
4795 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4796 struct cpumask *groupmask)
4798 struct sched_group *group = sd->groups;
4801 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4802 cpumask_clear(groupmask);
4804 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4806 if (!(sd->flags & SD_LOAD_BALANCE)) {
4807 printk("does not load-balance\n");
4809 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4814 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4816 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4817 printk(KERN_ERR "ERROR: domain->span does not contain "
4820 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4821 printk(KERN_ERR "ERROR: domain->groups does not contain"
4825 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4829 printk(KERN_ERR "ERROR: group is NULL\n");
4834 * Even though we initialize ->power to something semi-sane,
4835 * we leave power_orig unset. This allows us to detect if
4836 * domain iteration is still funny without causing /0 traps.
4838 if (!group->sgp->power_orig) {
4839 printk(KERN_CONT "\n");
4840 printk(KERN_ERR "ERROR: domain->cpu_power not "
4845 if (!cpumask_weight(sched_group_cpus(group))) {
4846 printk(KERN_CONT "\n");
4847 printk(KERN_ERR "ERROR: empty group\n");
4851 if (!(sd->flags & SD_OVERLAP) &&
4852 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4853 printk(KERN_CONT "\n");
4854 printk(KERN_ERR "ERROR: repeated CPUs\n");
4858 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4860 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4862 printk(KERN_CONT " %s", str);
4863 if (group->sgp->power != SCHED_POWER_SCALE) {
4864 printk(KERN_CONT " (cpu_power = %d)",
4868 group = group->next;
4869 } while (group != sd->groups);
4870 printk(KERN_CONT "\n");
4872 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4873 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4876 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4877 printk(KERN_ERR "ERROR: parent span is not a superset "
4878 "of domain->span\n");
4882 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4886 if (!sched_debug_enabled)
4890 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4894 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4897 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4905 #else /* !CONFIG_SCHED_DEBUG */
4906 # define sched_domain_debug(sd, cpu) do { } while (0)
4907 static inline bool sched_debug(void)
4911 #endif /* CONFIG_SCHED_DEBUG */
4913 static int sd_degenerate(struct sched_domain *sd)
4915 if (cpumask_weight(sched_domain_span(sd)) == 1)
4918 /* Following flags need at least 2 groups */
4919 if (sd->flags & (SD_LOAD_BALANCE |
4920 SD_BALANCE_NEWIDLE |
4924 SD_SHARE_PKG_RESOURCES)) {
4925 if (sd->groups != sd->groups->next)
4929 /* Following flags don't use groups */
4930 if (sd->flags & (SD_WAKE_AFFINE))
4937 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4939 unsigned long cflags = sd->flags, pflags = parent->flags;
4941 if (sd_degenerate(parent))
4944 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4947 /* Flags needing groups don't count if only 1 group in parent */
4948 if (parent->groups == parent->groups->next) {
4949 pflags &= ~(SD_LOAD_BALANCE |
4950 SD_BALANCE_NEWIDLE |
4954 SD_SHARE_PKG_RESOURCES |
4956 if (nr_node_ids == 1)
4957 pflags &= ~SD_SERIALIZE;
4959 if (~cflags & pflags)
4965 static void free_rootdomain(struct rcu_head *rcu)
4967 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4969 cpupri_cleanup(&rd->cpupri);
4970 free_cpumask_var(rd->rto_mask);
4971 free_cpumask_var(rd->online);
4972 free_cpumask_var(rd->span);
4976 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4978 struct root_domain *old_rd = NULL;
4979 unsigned long flags;
4981 raw_spin_lock_irqsave(&rq->lock, flags);
4986 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4989 cpumask_clear_cpu(rq->cpu, old_rd->span);
4992 * If we dont want to free the old_rt yet then
4993 * set old_rd to NULL to skip the freeing later
4996 if (!atomic_dec_and_test(&old_rd->refcount))
5000 atomic_inc(&rd->refcount);
5003 cpumask_set_cpu(rq->cpu, rd->span);
5004 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5007 raw_spin_unlock_irqrestore(&rq->lock, flags);
5010 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5013 static int init_rootdomain(struct root_domain *rd)
5015 memset(rd, 0, sizeof(*rd));
5017 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5019 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5021 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5024 if (cpupri_init(&rd->cpupri) != 0)
5029 free_cpumask_var(rd->rto_mask);
5031 free_cpumask_var(rd->online);
5033 free_cpumask_var(rd->span);
5039 * By default the system creates a single root-domain with all cpus as
5040 * members (mimicking the global state we have today).
5042 struct root_domain def_root_domain;
5044 static void init_defrootdomain(void)
5046 init_rootdomain(&def_root_domain);
5048 atomic_set(&def_root_domain.refcount, 1);
5051 static struct root_domain *alloc_rootdomain(void)
5053 struct root_domain *rd;
5055 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5059 if (init_rootdomain(rd) != 0) {
5067 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5069 struct sched_group *tmp, *first;
5078 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5083 } while (sg != first);
5086 static void free_sched_domain(struct rcu_head *rcu)
5088 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5091 * If its an overlapping domain it has private groups, iterate and
5094 if (sd->flags & SD_OVERLAP) {
5095 free_sched_groups(sd->groups, 1);
5096 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5097 kfree(sd->groups->sgp);
5103 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5105 call_rcu(&sd->rcu, free_sched_domain);
5108 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5110 for (; sd; sd = sd->parent)
5111 destroy_sched_domain(sd, cpu);
5115 * Keep a special pointer to the highest sched_domain that has
5116 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5117 * allows us to avoid some pointer chasing select_idle_sibling().
5119 * Also keep a unique ID per domain (we use the first cpu number in
5120 * the cpumask of the domain), this allows us to quickly tell if
5121 * two cpus are in the same cache domain, see cpus_share_cache().
5123 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5124 DEFINE_PER_CPU(int, sd_llc_size);
5125 DEFINE_PER_CPU(int, sd_llc_id);
5127 static void update_top_cache_domain(int cpu)
5129 struct sched_domain *sd;
5133 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5135 id = cpumask_first(sched_domain_span(sd));
5136 size = cpumask_weight(sched_domain_span(sd));
5139 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5140 per_cpu(sd_llc_size, cpu) = size;
5141 per_cpu(sd_llc_id, cpu) = id;
5145 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5146 * hold the hotplug lock.
5149 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5151 struct rq *rq = cpu_rq(cpu);
5152 struct sched_domain *tmp;
5154 /* Remove the sched domains which do not contribute to scheduling. */
5155 for (tmp = sd; tmp; ) {
5156 struct sched_domain *parent = tmp->parent;
5160 if (sd_parent_degenerate(tmp, parent)) {
5161 tmp->parent = parent->parent;
5163 parent->parent->child = tmp;
5165 * Transfer SD_PREFER_SIBLING down in case of a
5166 * degenerate parent; the spans match for this
5167 * so the property transfers.
5169 if (parent->flags & SD_PREFER_SIBLING)
5170 tmp->flags |= SD_PREFER_SIBLING;
5171 destroy_sched_domain(parent, cpu);
5176 if (sd && sd_degenerate(sd)) {
5179 destroy_sched_domain(tmp, cpu);
5184 sched_domain_debug(sd, cpu);
5186 rq_attach_root(rq, rd);
5188 rcu_assign_pointer(rq->sd, sd);
5189 destroy_sched_domains(tmp, cpu);
5191 update_top_cache_domain(cpu);
5194 /* cpus with isolated domains */
5195 static cpumask_var_t cpu_isolated_map;
5197 /* Setup the mask of cpus configured for isolated domains */
5198 static int __init isolated_cpu_setup(char *str)
5200 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5201 cpulist_parse(str, cpu_isolated_map);
5205 __setup("isolcpus=", isolated_cpu_setup);
5207 static const struct cpumask *cpu_cpu_mask(int cpu)
5209 return cpumask_of_node(cpu_to_node(cpu));
5213 struct sched_domain **__percpu sd;
5214 struct sched_group **__percpu sg;
5215 struct sched_group_power **__percpu sgp;
5219 struct sched_domain ** __percpu sd;
5220 struct root_domain *rd;
5230 struct sched_domain_topology_level;
5232 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5233 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5235 #define SDTL_OVERLAP 0x01
5237 struct sched_domain_topology_level {
5238 sched_domain_init_f init;
5239 sched_domain_mask_f mask;
5242 struct sd_data data;
5246 * Build an iteration mask that can exclude certain CPUs from the upwards
5249 * Asymmetric node setups can result in situations where the domain tree is of
5250 * unequal depth, make sure to skip domains that already cover the entire
5253 * In that case build_sched_domains() will have terminated the iteration early
5254 * and our sibling sd spans will be empty. Domains should always include the
5255 * cpu they're built on, so check that.
5258 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5260 const struct cpumask *span = sched_domain_span(sd);
5261 struct sd_data *sdd = sd->private;
5262 struct sched_domain *sibling;
5265 for_each_cpu(i, span) {
5266 sibling = *per_cpu_ptr(sdd->sd, i);
5267 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5270 cpumask_set_cpu(i, sched_group_mask(sg));
5275 * Return the canonical balance cpu for this group, this is the first cpu
5276 * of this group that's also in the iteration mask.
5278 int group_balance_cpu(struct sched_group *sg)
5280 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5284 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5286 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5287 const struct cpumask *span = sched_domain_span(sd);
5288 struct cpumask *covered = sched_domains_tmpmask;
5289 struct sd_data *sdd = sd->private;
5290 struct sched_domain *child;
5293 cpumask_clear(covered);
5295 for_each_cpu(i, span) {
5296 struct cpumask *sg_span;
5298 if (cpumask_test_cpu(i, covered))
5301 child = *per_cpu_ptr(sdd->sd, i);
5303 /* See the comment near build_group_mask(). */
5304 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5307 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5308 GFP_KERNEL, cpu_to_node(cpu));
5313 sg_span = sched_group_cpus(sg);
5315 child = child->child;
5316 cpumask_copy(sg_span, sched_domain_span(child));
5318 cpumask_set_cpu(i, sg_span);
5320 cpumask_or(covered, covered, sg_span);
5322 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5323 if (atomic_inc_return(&sg->sgp->ref) == 1)
5324 build_group_mask(sd, sg);
5327 * Initialize sgp->power such that even if we mess up the
5328 * domains and no possible iteration will get us here, we won't
5331 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5334 * Make sure the first group of this domain contains the
5335 * canonical balance cpu. Otherwise the sched_domain iteration
5336 * breaks. See update_sg_lb_stats().
5338 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5339 group_balance_cpu(sg) == cpu)
5349 sd->groups = groups;
5354 free_sched_groups(first, 0);
5359 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5361 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5362 struct sched_domain *child = sd->child;
5365 cpu = cpumask_first(sched_domain_span(child));
5368 *sg = *per_cpu_ptr(sdd->sg, cpu);
5369 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5370 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5377 * build_sched_groups will build a circular linked list of the groups
5378 * covered by the given span, and will set each group's ->cpumask correctly,
5379 * and ->cpu_power to 0.
5381 * Assumes the sched_domain tree is fully constructed
5384 build_sched_groups(struct sched_domain *sd, int cpu)
5386 struct sched_group *first = NULL, *last = NULL;
5387 struct sd_data *sdd = sd->private;
5388 const struct cpumask *span = sched_domain_span(sd);
5389 struct cpumask *covered;
5392 get_group(cpu, sdd, &sd->groups);
5393 atomic_inc(&sd->groups->ref);
5395 if (cpu != cpumask_first(span))
5398 lockdep_assert_held(&sched_domains_mutex);
5399 covered = sched_domains_tmpmask;
5401 cpumask_clear(covered);
5403 for_each_cpu(i, span) {
5404 struct sched_group *sg;
5407 if (cpumask_test_cpu(i, covered))
5410 group = get_group(i, sdd, &sg);
5411 cpumask_clear(sched_group_cpus(sg));
5413 cpumask_setall(sched_group_mask(sg));
5415 for_each_cpu(j, span) {
5416 if (get_group(j, sdd, NULL) != group)
5419 cpumask_set_cpu(j, covered);
5420 cpumask_set_cpu(j, sched_group_cpus(sg));
5435 * Initialize sched groups cpu_power.
5437 * cpu_power indicates the capacity of sched group, which is used while
5438 * distributing the load between different sched groups in a sched domain.
5439 * Typically cpu_power for all the groups in a sched domain will be same unless
5440 * there are asymmetries in the topology. If there are asymmetries, group
5441 * having more cpu_power will pickup more load compared to the group having
5444 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5446 struct sched_group *sg = sd->groups;
5451 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5453 } while (sg != sd->groups);
5455 if (cpu != group_balance_cpu(sg))
5458 update_group_power(sd, cpu);
5459 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5462 int __weak arch_sd_sibling_asym_packing(void)
5464 return 0*SD_ASYM_PACKING;
5468 * Initializers for schedule domains
5469 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5472 #ifdef CONFIG_SCHED_DEBUG
5473 # define SD_INIT_NAME(sd, type) sd->name = #type
5475 # define SD_INIT_NAME(sd, type) do { } while (0)
5478 #define SD_INIT_FUNC(type) \
5479 static noinline struct sched_domain * \
5480 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5482 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5483 *sd = SD_##type##_INIT; \
5484 SD_INIT_NAME(sd, type); \
5485 sd->private = &tl->data; \
5490 #ifdef CONFIG_SCHED_SMT
5491 SD_INIT_FUNC(SIBLING)
5493 #ifdef CONFIG_SCHED_MC
5496 #ifdef CONFIG_SCHED_BOOK
5500 static int default_relax_domain_level = -1;
5501 int sched_domain_level_max;
5503 static int __init setup_relax_domain_level(char *str)
5505 if (kstrtoint(str, 0, &default_relax_domain_level))
5506 pr_warn("Unable to set relax_domain_level\n");
5510 __setup("relax_domain_level=", setup_relax_domain_level);
5512 static void set_domain_attribute(struct sched_domain *sd,
5513 struct sched_domain_attr *attr)
5517 if (!attr || attr->relax_domain_level < 0) {
5518 if (default_relax_domain_level < 0)
5521 request = default_relax_domain_level;
5523 request = attr->relax_domain_level;
5524 if (request < sd->level) {
5525 /* turn off idle balance on this domain */
5526 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5528 /* turn on idle balance on this domain */
5529 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5533 static void __sdt_free(const struct cpumask *cpu_map);
5534 static int __sdt_alloc(const struct cpumask *cpu_map);
5536 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5537 const struct cpumask *cpu_map)
5541 if (!atomic_read(&d->rd->refcount))
5542 free_rootdomain(&d->rd->rcu); /* fall through */
5544 free_percpu(d->sd); /* fall through */
5546 __sdt_free(cpu_map); /* fall through */
5552 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5553 const struct cpumask *cpu_map)
5555 memset(d, 0, sizeof(*d));
5557 if (__sdt_alloc(cpu_map))
5558 return sa_sd_storage;
5559 d->sd = alloc_percpu(struct sched_domain *);
5561 return sa_sd_storage;
5562 d->rd = alloc_rootdomain();
5565 return sa_rootdomain;
5569 * NULL the sd_data elements we've used to build the sched_domain and
5570 * sched_group structure so that the subsequent __free_domain_allocs()
5571 * will not free the data we're using.
5573 static void claim_allocations(int cpu, struct sched_domain *sd)
5575 struct sd_data *sdd = sd->private;
5577 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5578 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5580 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5581 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5583 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5584 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5587 #ifdef CONFIG_SCHED_SMT
5588 static const struct cpumask *cpu_smt_mask(int cpu)
5590 return topology_thread_cpumask(cpu);
5595 * Topology list, bottom-up.
5597 static struct sched_domain_topology_level default_topology[] = {
5598 #ifdef CONFIG_SCHED_SMT
5599 { sd_init_SIBLING, cpu_smt_mask, },
5601 #ifdef CONFIG_SCHED_MC
5602 { sd_init_MC, cpu_coregroup_mask, },
5604 #ifdef CONFIG_SCHED_BOOK
5605 { sd_init_BOOK, cpu_book_mask, },
5607 { sd_init_CPU, cpu_cpu_mask, },
5611 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5613 #define for_each_sd_topology(tl) \
5614 for (tl = sched_domain_topology; tl->init; tl++)
5618 static int sched_domains_numa_levels;
5619 static int *sched_domains_numa_distance;
5620 static struct cpumask ***sched_domains_numa_masks;
5621 static int sched_domains_curr_level;
5623 static inline int sd_local_flags(int level)
5625 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5628 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5631 static struct sched_domain *
5632 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5634 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5635 int level = tl->numa_level;
5636 int sd_weight = cpumask_weight(
5637 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5639 *sd = (struct sched_domain){
5640 .min_interval = sd_weight,
5641 .max_interval = 2*sd_weight,
5643 .imbalance_pct = 125,
5644 .cache_nice_tries = 2,
5651 .flags = 1*SD_LOAD_BALANCE
5652 | 1*SD_BALANCE_NEWIDLE
5657 | 0*SD_SHARE_CPUPOWER
5658 | 0*SD_SHARE_PKG_RESOURCES
5660 | 0*SD_PREFER_SIBLING
5661 | sd_local_flags(level)
5663 .last_balance = jiffies,
5664 .balance_interval = sd_weight,
5666 SD_INIT_NAME(sd, NUMA);
5667 sd->private = &tl->data;
5670 * Ugly hack to pass state to sd_numa_mask()...
5672 sched_domains_curr_level = tl->numa_level;
5677 static const struct cpumask *sd_numa_mask(int cpu)
5679 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5682 static void sched_numa_warn(const char *str)
5684 static int done = false;
5692 printk(KERN_WARNING "ERROR: %s\n\n", str);
5694 for (i = 0; i < nr_node_ids; i++) {
5695 printk(KERN_WARNING " ");
5696 for (j = 0; j < nr_node_ids; j++)
5697 printk(KERN_CONT "%02d ", node_distance(i,j));
5698 printk(KERN_CONT "\n");
5700 printk(KERN_WARNING "\n");
5703 static bool find_numa_distance(int distance)
5707 if (distance == node_distance(0, 0))
5710 for (i = 0; i < sched_domains_numa_levels; i++) {
5711 if (sched_domains_numa_distance[i] == distance)
5718 static void sched_init_numa(void)
5720 int next_distance, curr_distance = node_distance(0, 0);
5721 struct sched_domain_topology_level *tl;
5725 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5726 if (!sched_domains_numa_distance)
5730 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5731 * unique distances in the node_distance() table.
5733 * Assumes node_distance(0,j) includes all distances in
5734 * node_distance(i,j) in order to avoid cubic time.
5736 next_distance = curr_distance;
5737 for (i = 0; i < nr_node_ids; i++) {
5738 for (j = 0; j < nr_node_ids; j++) {
5739 for (k = 0; k < nr_node_ids; k++) {
5740 int distance = node_distance(i, k);
5742 if (distance > curr_distance &&
5743 (distance < next_distance ||
5744 next_distance == curr_distance))
5745 next_distance = distance;
5748 * While not a strong assumption it would be nice to know
5749 * about cases where if node A is connected to B, B is not
5750 * equally connected to A.
5752 if (sched_debug() && node_distance(k, i) != distance)
5753 sched_numa_warn("Node-distance not symmetric");
5755 if (sched_debug() && i && !find_numa_distance(distance))
5756 sched_numa_warn("Node-0 not representative");
5758 if (next_distance != curr_distance) {
5759 sched_domains_numa_distance[level++] = next_distance;
5760 sched_domains_numa_levels = level;
5761 curr_distance = next_distance;
5766 * In case of sched_debug() we verify the above assumption.
5772 * 'level' contains the number of unique distances, excluding the
5773 * identity distance node_distance(i,i).
5775 * The sched_domains_numa_distance[] array includes the actual distance
5780 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5781 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5782 * the array will contain less then 'level' members. This could be
5783 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5784 * in other functions.
5786 * We reset it to 'level' at the end of this function.
5788 sched_domains_numa_levels = 0;
5790 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5791 if (!sched_domains_numa_masks)
5795 * Now for each level, construct a mask per node which contains all
5796 * cpus of nodes that are that many hops away from us.
5798 for (i = 0; i < level; i++) {
5799 sched_domains_numa_masks[i] =
5800 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5801 if (!sched_domains_numa_masks[i])
5804 for (j = 0; j < nr_node_ids; j++) {
5805 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5809 sched_domains_numa_masks[i][j] = mask;
5811 for (k = 0; k < nr_node_ids; k++) {
5812 if (node_distance(j, k) > sched_domains_numa_distance[i])
5815 cpumask_or(mask, mask, cpumask_of_node(k));
5820 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5821 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5826 * Copy the default topology bits..
5828 for (i = 0; default_topology[i].init; i++)
5829 tl[i] = default_topology[i];
5832 * .. and append 'j' levels of NUMA goodness.
5834 for (j = 0; j < level; i++, j++) {
5835 tl[i] = (struct sched_domain_topology_level){
5836 .init = sd_numa_init,
5837 .mask = sd_numa_mask,
5838 .flags = SDTL_OVERLAP,
5843 sched_domain_topology = tl;
5845 sched_domains_numa_levels = level;
5848 static void sched_domains_numa_masks_set(int cpu)
5851 int node = cpu_to_node(cpu);
5853 for (i = 0; i < sched_domains_numa_levels; i++) {
5854 for (j = 0; j < nr_node_ids; j++) {
5855 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5856 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5861 static void sched_domains_numa_masks_clear(int cpu)
5864 for (i = 0; i < sched_domains_numa_levels; i++) {
5865 for (j = 0; j < nr_node_ids; j++)
5866 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5871 * Update sched_domains_numa_masks[level][node] array when new cpus
5874 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5875 unsigned long action,
5878 int cpu = (long)hcpu;
5880 switch (action & ~CPU_TASKS_FROZEN) {
5882 sched_domains_numa_masks_set(cpu);
5886 sched_domains_numa_masks_clear(cpu);
5896 static inline void sched_init_numa(void)
5900 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5901 unsigned long action,
5906 #endif /* CONFIG_NUMA */
5908 static int __sdt_alloc(const struct cpumask *cpu_map)
5910 struct sched_domain_topology_level *tl;
5913 for_each_sd_topology(tl) {
5914 struct sd_data *sdd = &tl->data;
5916 sdd->sd = alloc_percpu(struct sched_domain *);
5920 sdd->sg = alloc_percpu(struct sched_group *);
5924 sdd->sgp = alloc_percpu(struct sched_group_power *);
5928 for_each_cpu(j, cpu_map) {
5929 struct sched_domain *sd;
5930 struct sched_group *sg;
5931 struct sched_group_power *sgp;
5933 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5934 GFP_KERNEL, cpu_to_node(j));
5938 *per_cpu_ptr(sdd->sd, j) = sd;
5940 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5941 GFP_KERNEL, cpu_to_node(j));
5947 *per_cpu_ptr(sdd->sg, j) = sg;
5949 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5950 GFP_KERNEL, cpu_to_node(j));
5954 *per_cpu_ptr(sdd->sgp, j) = sgp;
5961 static void __sdt_free(const struct cpumask *cpu_map)
5963 struct sched_domain_topology_level *tl;
5966 for_each_sd_topology(tl) {
5967 struct sd_data *sdd = &tl->data;
5969 for_each_cpu(j, cpu_map) {
5970 struct sched_domain *sd;
5973 sd = *per_cpu_ptr(sdd->sd, j);
5974 if (sd && (sd->flags & SD_OVERLAP))
5975 free_sched_groups(sd->groups, 0);
5976 kfree(*per_cpu_ptr(sdd->sd, j));
5980 kfree(*per_cpu_ptr(sdd->sg, j));
5982 kfree(*per_cpu_ptr(sdd->sgp, j));
5984 free_percpu(sdd->sd);
5986 free_percpu(sdd->sg);
5988 free_percpu(sdd->sgp);
5993 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5994 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5995 struct sched_domain *child, int cpu)
5997 struct sched_domain *sd = tl->init(tl, cpu);
6001 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6003 sd->level = child->level + 1;
6004 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6008 set_domain_attribute(sd, attr);
6014 * Build sched domains for a given set of cpus and attach the sched domains
6015 * to the individual cpus
6017 static int build_sched_domains(const struct cpumask *cpu_map,
6018 struct sched_domain_attr *attr)
6020 enum s_alloc alloc_state;
6021 struct sched_domain *sd;
6023 int i, ret = -ENOMEM;
6025 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6026 if (alloc_state != sa_rootdomain)
6029 /* Set up domains for cpus specified by the cpu_map. */
6030 for_each_cpu(i, cpu_map) {
6031 struct sched_domain_topology_level *tl;
6034 for_each_sd_topology(tl) {
6035 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6036 if (tl == sched_domain_topology)
6037 *per_cpu_ptr(d.sd, i) = sd;
6038 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6039 sd->flags |= SD_OVERLAP;
6040 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6045 /* Build the groups for the domains */
6046 for_each_cpu(i, cpu_map) {
6047 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6048 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6049 if (sd->flags & SD_OVERLAP) {
6050 if (build_overlap_sched_groups(sd, i))
6053 if (build_sched_groups(sd, i))
6059 /* Calculate CPU power for physical packages and nodes */
6060 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6061 if (!cpumask_test_cpu(i, cpu_map))
6064 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6065 claim_allocations(i, sd);
6066 init_sched_groups_power(i, sd);
6070 /* Attach the domains */
6072 for_each_cpu(i, cpu_map) {
6073 sd = *per_cpu_ptr(d.sd, i);
6074 cpu_attach_domain(sd, d.rd, i);
6080 __free_domain_allocs(&d, alloc_state, cpu_map);
6084 static cpumask_var_t *doms_cur; /* current sched domains */
6085 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6086 static struct sched_domain_attr *dattr_cur;
6087 /* attribues of custom domains in 'doms_cur' */
6090 * Special case: If a kmalloc of a doms_cur partition (array of
6091 * cpumask) fails, then fallback to a single sched domain,
6092 * as determined by the single cpumask fallback_doms.
6094 static cpumask_var_t fallback_doms;
6097 * arch_update_cpu_topology lets virtualized architectures update the
6098 * cpu core maps. It is supposed to return 1 if the topology changed
6099 * or 0 if it stayed the same.
6101 int __attribute__((weak)) arch_update_cpu_topology(void)
6106 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6109 cpumask_var_t *doms;
6111 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6114 for (i = 0; i < ndoms; i++) {
6115 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6116 free_sched_domains(doms, i);
6123 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6126 for (i = 0; i < ndoms; i++)
6127 free_cpumask_var(doms[i]);
6132 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6133 * For now this just excludes isolated cpus, but could be used to
6134 * exclude other special cases in the future.
6136 static int init_sched_domains(const struct cpumask *cpu_map)
6140 arch_update_cpu_topology();
6142 doms_cur = alloc_sched_domains(ndoms_cur);
6144 doms_cur = &fallback_doms;
6145 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6146 err = build_sched_domains(doms_cur[0], NULL);
6147 register_sched_domain_sysctl();
6153 * Detach sched domains from a group of cpus specified in cpu_map
6154 * These cpus will now be attached to the NULL domain
6156 static void detach_destroy_domains(const struct cpumask *cpu_map)
6161 for_each_cpu(i, cpu_map)
6162 cpu_attach_domain(NULL, &def_root_domain, i);
6166 /* handle null as "default" */
6167 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6168 struct sched_domain_attr *new, int idx_new)
6170 struct sched_domain_attr tmp;
6177 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6178 new ? (new + idx_new) : &tmp,
6179 sizeof(struct sched_domain_attr));
6183 * Partition sched domains as specified by the 'ndoms_new'
6184 * cpumasks in the array doms_new[] of cpumasks. This compares
6185 * doms_new[] to the current sched domain partitioning, doms_cur[].
6186 * It destroys each deleted domain and builds each new domain.
6188 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6189 * The masks don't intersect (don't overlap.) We should setup one
6190 * sched domain for each mask. CPUs not in any of the cpumasks will
6191 * not be load balanced. If the same cpumask appears both in the
6192 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6195 * The passed in 'doms_new' should be allocated using
6196 * alloc_sched_domains. This routine takes ownership of it and will
6197 * free_sched_domains it when done with it. If the caller failed the
6198 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6199 * and partition_sched_domains() will fallback to the single partition
6200 * 'fallback_doms', it also forces the domains to be rebuilt.
6202 * If doms_new == NULL it will be replaced with cpu_online_mask.
6203 * ndoms_new == 0 is a special case for destroying existing domains,
6204 * and it will not create the default domain.
6206 * Call with hotplug lock held
6208 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6209 struct sched_domain_attr *dattr_new)
6214 mutex_lock(&sched_domains_mutex);
6216 /* always unregister in case we don't destroy any domains */
6217 unregister_sched_domain_sysctl();
6219 /* Let architecture update cpu core mappings. */
6220 new_topology = arch_update_cpu_topology();
6222 n = doms_new ? ndoms_new : 0;
6224 /* Destroy deleted domains */
6225 for (i = 0; i < ndoms_cur; i++) {
6226 for (j = 0; j < n && !new_topology; j++) {
6227 if (cpumask_equal(doms_cur[i], doms_new[j])
6228 && dattrs_equal(dattr_cur, i, dattr_new, j))
6231 /* no match - a current sched domain not in new doms_new[] */
6232 detach_destroy_domains(doms_cur[i]);
6238 if (doms_new == NULL) {
6240 doms_new = &fallback_doms;
6241 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6242 WARN_ON_ONCE(dattr_new);
6245 /* Build new domains */
6246 for (i = 0; i < ndoms_new; i++) {
6247 for (j = 0; j < n && !new_topology; j++) {
6248 if (cpumask_equal(doms_new[i], doms_cur[j])
6249 && dattrs_equal(dattr_new, i, dattr_cur, j))
6252 /* no match - add a new doms_new */
6253 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6258 /* Remember the new sched domains */
6259 if (doms_cur != &fallback_doms)
6260 free_sched_domains(doms_cur, ndoms_cur);
6261 kfree(dattr_cur); /* kfree(NULL) is safe */
6262 doms_cur = doms_new;
6263 dattr_cur = dattr_new;
6264 ndoms_cur = ndoms_new;
6266 register_sched_domain_sysctl();
6268 mutex_unlock(&sched_domains_mutex);
6271 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6274 * Update cpusets according to cpu_active mask. If cpusets are
6275 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6276 * around partition_sched_domains().
6278 * If we come here as part of a suspend/resume, don't touch cpusets because we
6279 * want to restore it back to its original state upon resume anyway.
6281 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6285 case CPU_ONLINE_FROZEN:
6286 case CPU_DOWN_FAILED_FROZEN:
6289 * num_cpus_frozen tracks how many CPUs are involved in suspend
6290 * resume sequence. As long as this is not the last online
6291 * operation in the resume sequence, just build a single sched
6292 * domain, ignoring cpusets.
6295 if (likely(num_cpus_frozen)) {
6296 partition_sched_domains(1, NULL, NULL);
6301 * This is the last CPU online operation. So fall through and
6302 * restore the original sched domains by considering the
6303 * cpuset configurations.
6307 case CPU_DOWN_FAILED:
6308 cpuset_update_active_cpus(true);
6316 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6320 case CPU_DOWN_PREPARE:
6321 cpuset_update_active_cpus(false);
6323 case CPU_DOWN_PREPARE_FROZEN:
6325 partition_sched_domains(1, NULL, NULL);
6333 void __init sched_init_smp(void)
6335 cpumask_var_t non_isolated_cpus;
6337 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6338 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6343 mutex_lock(&sched_domains_mutex);
6344 init_sched_domains(cpu_active_mask);
6345 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6346 if (cpumask_empty(non_isolated_cpus))
6347 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6348 mutex_unlock(&sched_domains_mutex);
6351 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6352 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6353 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6357 /* Move init over to a non-isolated CPU */
6358 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6360 sched_init_granularity();
6361 free_cpumask_var(non_isolated_cpus);
6363 init_sched_rt_class();
6366 void __init sched_init_smp(void)
6368 sched_init_granularity();
6370 #endif /* CONFIG_SMP */
6372 const_debug unsigned int sysctl_timer_migration = 1;
6374 int in_sched_functions(unsigned long addr)
6376 return in_lock_functions(addr) ||
6377 (addr >= (unsigned long)__sched_text_start
6378 && addr < (unsigned long)__sched_text_end);
6381 #ifdef CONFIG_CGROUP_SCHED
6383 * Default task group.
6384 * Every task in system belongs to this group at bootup.
6386 struct task_group root_task_group;
6387 LIST_HEAD(task_groups);
6390 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6392 void __init sched_init(void)
6395 unsigned long alloc_size = 0, ptr;
6397 #ifdef CONFIG_FAIR_GROUP_SCHED
6398 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6400 #ifdef CONFIG_RT_GROUP_SCHED
6401 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6403 #ifdef CONFIG_CPUMASK_OFFSTACK
6404 alloc_size += num_possible_cpus() * cpumask_size();
6407 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6409 #ifdef CONFIG_FAIR_GROUP_SCHED
6410 root_task_group.se = (struct sched_entity **)ptr;
6411 ptr += nr_cpu_ids * sizeof(void **);
6413 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6414 ptr += nr_cpu_ids * sizeof(void **);
6416 #endif /* CONFIG_FAIR_GROUP_SCHED */
6417 #ifdef CONFIG_RT_GROUP_SCHED
6418 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6419 ptr += nr_cpu_ids * sizeof(void **);
6421 root_task_group.rt_rq = (struct rt_rq **)ptr;
6422 ptr += nr_cpu_ids * sizeof(void **);
6424 #endif /* CONFIG_RT_GROUP_SCHED */
6425 #ifdef CONFIG_CPUMASK_OFFSTACK
6426 for_each_possible_cpu(i) {
6427 per_cpu(load_balance_mask, i) = (void *)ptr;
6428 ptr += cpumask_size();
6430 #endif /* CONFIG_CPUMASK_OFFSTACK */
6434 init_defrootdomain();
6437 init_rt_bandwidth(&def_rt_bandwidth,
6438 global_rt_period(), global_rt_runtime());
6440 #ifdef CONFIG_RT_GROUP_SCHED
6441 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6442 global_rt_period(), global_rt_runtime());
6443 #endif /* CONFIG_RT_GROUP_SCHED */
6445 #ifdef CONFIG_CGROUP_SCHED
6446 list_add(&root_task_group.list, &task_groups);
6447 INIT_LIST_HEAD(&root_task_group.children);
6448 INIT_LIST_HEAD(&root_task_group.siblings);
6449 autogroup_init(&init_task);
6451 #endif /* CONFIG_CGROUP_SCHED */
6453 for_each_possible_cpu(i) {
6457 raw_spin_lock_init(&rq->lock);
6459 rq->calc_load_active = 0;
6460 rq->calc_load_update = jiffies + LOAD_FREQ;
6461 init_cfs_rq(&rq->cfs);
6462 init_rt_rq(&rq->rt, rq);
6463 #ifdef CONFIG_FAIR_GROUP_SCHED
6464 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6465 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6467 * How much cpu bandwidth does root_task_group get?
6469 * In case of task-groups formed thr' the cgroup filesystem, it
6470 * gets 100% of the cpu resources in the system. This overall
6471 * system cpu resource is divided among the tasks of
6472 * root_task_group and its child task-groups in a fair manner,
6473 * based on each entity's (task or task-group's) weight
6474 * (se->load.weight).
6476 * In other words, if root_task_group has 10 tasks of weight
6477 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6478 * then A0's share of the cpu resource is:
6480 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6482 * We achieve this by letting root_task_group's tasks sit
6483 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6485 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6486 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6487 #endif /* CONFIG_FAIR_GROUP_SCHED */
6489 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6490 #ifdef CONFIG_RT_GROUP_SCHED
6491 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6492 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6495 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6496 rq->cpu_load[j] = 0;
6498 rq->last_load_update_tick = jiffies;
6503 rq->cpu_power = SCHED_POWER_SCALE;
6504 rq->post_schedule = 0;
6505 rq->active_balance = 0;
6506 rq->next_balance = jiffies;
6511 rq->avg_idle = 2*sysctl_sched_migration_cost;
6512 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6514 INIT_LIST_HEAD(&rq->cfs_tasks);
6516 rq_attach_root(rq, &def_root_domain);
6517 #ifdef CONFIG_NO_HZ_COMMON
6520 #ifdef CONFIG_NO_HZ_FULL
6521 rq->last_sched_tick = 0;
6525 atomic_set(&rq->nr_iowait, 0);
6528 set_load_weight(&init_task);
6530 #ifdef CONFIG_PREEMPT_NOTIFIERS
6531 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6534 #ifdef CONFIG_RT_MUTEXES
6535 plist_head_init(&init_task.pi_waiters);
6539 * The boot idle thread does lazy MMU switching as well:
6541 atomic_inc(&init_mm.mm_count);
6542 enter_lazy_tlb(&init_mm, current);
6545 * Make us the idle thread. Technically, schedule() should not be
6546 * called from this thread, however somewhere below it might be,
6547 * but because we are the idle thread, we just pick up running again
6548 * when this runqueue becomes "idle".
6550 init_idle(current, smp_processor_id());
6552 calc_load_update = jiffies + LOAD_FREQ;
6555 * During early bootup we pretend to be a normal task:
6557 current->sched_class = &fair_sched_class;
6560 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6561 /* May be allocated at isolcpus cmdline parse time */
6562 if (cpu_isolated_map == NULL)
6563 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6564 idle_thread_set_boot_cpu();
6566 init_sched_fair_class();
6568 scheduler_running = 1;
6571 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6572 static inline int preempt_count_equals(int preempt_offset)
6574 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6576 return (nested == preempt_offset);
6579 void __might_sleep(const char *file, int line, int preempt_offset)
6581 static unsigned long prev_jiffy; /* ratelimiting */
6583 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6584 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6585 system_state != SYSTEM_RUNNING || oops_in_progress)
6587 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6589 prev_jiffy = jiffies;
6592 "BUG: sleeping function called from invalid context at %s:%d\n",
6595 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6596 in_atomic(), irqs_disabled(),
6597 current->pid, current->comm);
6599 debug_show_held_locks(current);
6600 if (irqs_disabled())
6601 print_irqtrace_events(current);
6604 EXPORT_SYMBOL(__might_sleep);
6607 #ifdef CONFIG_MAGIC_SYSRQ
6608 static void normalize_task(struct rq *rq, struct task_struct *p)
6610 const struct sched_class *prev_class = p->sched_class;
6611 int old_prio = p->prio;
6616 dequeue_task(rq, p, 0);
6617 __setscheduler(rq, p, SCHED_NORMAL, 0);
6619 enqueue_task(rq, p, 0);
6620 resched_task(rq->curr);
6623 check_class_changed(rq, p, prev_class, old_prio);
6626 void normalize_rt_tasks(void)
6628 struct task_struct *g, *p;
6629 unsigned long flags;
6632 read_lock_irqsave(&tasklist_lock, flags);
6633 do_each_thread(g, p) {
6635 * Only normalize user tasks:
6640 p->se.exec_start = 0;
6641 #ifdef CONFIG_SCHEDSTATS
6642 p->se.statistics.wait_start = 0;
6643 p->se.statistics.sleep_start = 0;
6644 p->se.statistics.block_start = 0;
6649 * Renice negative nice level userspace
6652 if (TASK_NICE(p) < 0 && p->mm)
6653 set_user_nice(p, 0);
6657 raw_spin_lock(&p->pi_lock);
6658 rq = __task_rq_lock(p);
6660 normalize_task(rq, p);
6662 __task_rq_unlock(rq);
6663 raw_spin_unlock(&p->pi_lock);
6664 } while_each_thread(g, p);
6666 read_unlock_irqrestore(&tasklist_lock, flags);
6669 #endif /* CONFIG_MAGIC_SYSRQ */
6671 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6673 * These functions are only useful for the IA64 MCA handling, or kdb.
6675 * They can only be called when the whole system has been
6676 * stopped - every CPU needs to be quiescent, and no scheduling
6677 * activity can take place. Using them for anything else would
6678 * be a serious bug, and as a result, they aren't even visible
6679 * under any other configuration.
6683 * curr_task - return the current task for a given cpu.
6684 * @cpu: the processor in question.
6686 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6688 * Return: The current task for @cpu.
6690 struct task_struct *curr_task(int cpu)
6692 return cpu_curr(cpu);
6695 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6699 * set_curr_task - set the current task for a given cpu.
6700 * @cpu: the processor in question.
6701 * @p: the task pointer to set.
6703 * Description: This function must only be used when non-maskable interrupts
6704 * are serviced on a separate stack. It allows the architecture to switch the
6705 * notion of the current task on a cpu in a non-blocking manner. This function
6706 * must be called with all CPU's synchronized, and interrupts disabled, the
6707 * and caller must save the original value of the current task (see
6708 * curr_task() above) and restore that value before reenabling interrupts and
6709 * re-starting the system.
6711 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6713 void set_curr_task(int cpu, struct task_struct *p)
6720 #ifdef CONFIG_CGROUP_SCHED
6721 /* task_group_lock serializes the addition/removal of task groups */
6722 static DEFINE_SPINLOCK(task_group_lock);
6724 static void free_sched_group(struct task_group *tg)
6726 free_fair_sched_group(tg);
6727 free_rt_sched_group(tg);
6732 /* allocate runqueue etc for a new task group */
6733 struct task_group *sched_create_group(struct task_group *parent)
6735 struct task_group *tg;
6737 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6739 return ERR_PTR(-ENOMEM);
6741 if (!alloc_fair_sched_group(tg, parent))
6744 if (!alloc_rt_sched_group(tg, parent))
6750 free_sched_group(tg);
6751 return ERR_PTR(-ENOMEM);
6754 void sched_online_group(struct task_group *tg, struct task_group *parent)
6756 unsigned long flags;
6758 spin_lock_irqsave(&task_group_lock, flags);
6759 list_add_rcu(&tg->list, &task_groups);
6761 WARN_ON(!parent); /* root should already exist */
6763 tg->parent = parent;
6764 INIT_LIST_HEAD(&tg->children);
6765 list_add_rcu(&tg->siblings, &parent->children);
6766 spin_unlock_irqrestore(&task_group_lock, flags);
6769 /* rcu callback to free various structures associated with a task group */
6770 static void free_sched_group_rcu(struct rcu_head *rhp)
6772 /* now it should be safe to free those cfs_rqs */
6773 free_sched_group(container_of(rhp, struct task_group, rcu));
6776 /* Destroy runqueue etc associated with a task group */
6777 void sched_destroy_group(struct task_group *tg)
6779 /* wait for possible concurrent references to cfs_rqs complete */
6780 call_rcu(&tg->rcu, free_sched_group_rcu);
6783 void sched_offline_group(struct task_group *tg)
6785 unsigned long flags;
6788 /* end participation in shares distribution */
6789 for_each_possible_cpu(i)
6790 unregister_fair_sched_group(tg, i);
6792 spin_lock_irqsave(&task_group_lock, flags);
6793 list_del_rcu(&tg->list);
6794 list_del_rcu(&tg->siblings);
6795 spin_unlock_irqrestore(&task_group_lock, flags);
6798 /* change task's runqueue when it moves between groups.
6799 * The caller of this function should have put the task in its new group
6800 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6801 * reflect its new group.
6803 void sched_move_task(struct task_struct *tsk)
6805 struct task_group *tg;
6807 unsigned long flags;
6810 rq = task_rq_lock(tsk, &flags);
6812 running = task_current(rq, tsk);
6816 dequeue_task(rq, tsk, 0);
6817 if (unlikely(running))
6818 tsk->sched_class->put_prev_task(rq, tsk);
6820 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6821 lockdep_is_held(&tsk->sighand->siglock)),
6822 struct task_group, css);
6823 tg = autogroup_task_group(tsk, tg);
6824 tsk->sched_task_group = tg;
6826 #ifdef CONFIG_FAIR_GROUP_SCHED
6827 if (tsk->sched_class->task_move_group)
6828 tsk->sched_class->task_move_group(tsk, on_rq);
6831 set_task_rq(tsk, task_cpu(tsk));
6833 if (unlikely(running))
6834 tsk->sched_class->set_curr_task(rq);
6836 enqueue_task(rq, tsk, 0);
6838 task_rq_unlock(rq, tsk, &flags);
6840 #endif /* CONFIG_CGROUP_SCHED */
6842 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6843 static unsigned long to_ratio(u64 period, u64 runtime)
6845 if (runtime == RUNTIME_INF)
6848 return div64_u64(runtime << 20, period);
6852 #ifdef CONFIG_RT_GROUP_SCHED
6854 * Ensure that the real time constraints are schedulable.
6856 static DEFINE_MUTEX(rt_constraints_mutex);
6858 /* Must be called with tasklist_lock held */
6859 static inline int tg_has_rt_tasks(struct task_group *tg)
6861 struct task_struct *g, *p;
6863 do_each_thread(g, p) {
6864 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6866 } while_each_thread(g, p);
6871 struct rt_schedulable_data {
6872 struct task_group *tg;
6877 static int tg_rt_schedulable(struct task_group *tg, void *data)
6879 struct rt_schedulable_data *d = data;
6880 struct task_group *child;
6881 unsigned long total, sum = 0;
6882 u64 period, runtime;
6884 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6885 runtime = tg->rt_bandwidth.rt_runtime;
6888 period = d->rt_period;
6889 runtime = d->rt_runtime;
6893 * Cannot have more runtime than the period.
6895 if (runtime > period && runtime != RUNTIME_INF)
6899 * Ensure we don't starve existing RT tasks.
6901 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6904 total = to_ratio(period, runtime);
6907 * Nobody can have more than the global setting allows.
6909 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6913 * The sum of our children's runtime should not exceed our own.
6915 list_for_each_entry_rcu(child, &tg->children, siblings) {
6916 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6917 runtime = child->rt_bandwidth.rt_runtime;
6919 if (child == d->tg) {
6920 period = d->rt_period;
6921 runtime = d->rt_runtime;
6924 sum += to_ratio(period, runtime);
6933 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6937 struct rt_schedulable_data data = {
6939 .rt_period = period,
6940 .rt_runtime = runtime,
6944 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6950 static int tg_set_rt_bandwidth(struct task_group *tg,
6951 u64 rt_period, u64 rt_runtime)
6955 mutex_lock(&rt_constraints_mutex);
6956 read_lock(&tasklist_lock);
6957 err = __rt_schedulable(tg, rt_period, rt_runtime);
6961 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6962 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6963 tg->rt_bandwidth.rt_runtime = rt_runtime;
6965 for_each_possible_cpu(i) {
6966 struct rt_rq *rt_rq = tg->rt_rq[i];
6968 raw_spin_lock(&rt_rq->rt_runtime_lock);
6969 rt_rq->rt_runtime = rt_runtime;
6970 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6972 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6974 read_unlock(&tasklist_lock);
6975 mutex_unlock(&rt_constraints_mutex);
6980 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6982 u64 rt_runtime, rt_period;
6984 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6985 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6986 if (rt_runtime_us < 0)
6987 rt_runtime = RUNTIME_INF;
6989 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6992 static long sched_group_rt_runtime(struct task_group *tg)
6996 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6999 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7000 do_div(rt_runtime_us, NSEC_PER_USEC);
7001 return rt_runtime_us;
7004 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7006 u64 rt_runtime, rt_period;
7008 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7009 rt_runtime = tg->rt_bandwidth.rt_runtime;
7014 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7017 static long sched_group_rt_period(struct task_group *tg)
7021 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7022 do_div(rt_period_us, NSEC_PER_USEC);
7023 return rt_period_us;
7026 static int sched_rt_global_constraints(void)
7028 u64 runtime, period;
7031 if (sysctl_sched_rt_period <= 0)
7034 runtime = global_rt_runtime();
7035 period = global_rt_period();
7038 * Sanity check on the sysctl variables.
7040 if (runtime > period && runtime != RUNTIME_INF)
7043 mutex_lock(&rt_constraints_mutex);
7044 read_lock(&tasklist_lock);
7045 ret = __rt_schedulable(NULL, 0, 0);
7046 read_unlock(&tasklist_lock);
7047 mutex_unlock(&rt_constraints_mutex);
7052 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7054 /* Don't accept realtime tasks when there is no way for them to run */
7055 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7061 #else /* !CONFIG_RT_GROUP_SCHED */
7062 static int sched_rt_global_constraints(void)
7064 unsigned long flags;
7067 if (sysctl_sched_rt_period <= 0)
7071 * There's always some RT tasks in the root group
7072 * -- migration, kstopmachine etc..
7074 if (sysctl_sched_rt_runtime == 0)
7077 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7078 for_each_possible_cpu(i) {
7079 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7081 raw_spin_lock(&rt_rq->rt_runtime_lock);
7082 rt_rq->rt_runtime = global_rt_runtime();
7083 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7085 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7089 #endif /* CONFIG_RT_GROUP_SCHED */
7091 int sched_rr_handler(struct ctl_table *table, int write,
7092 void __user *buffer, size_t *lenp,
7096 static DEFINE_MUTEX(mutex);
7099 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7100 /* make sure that internally we keep jiffies */
7101 /* also, writing zero resets timeslice to default */
7102 if (!ret && write) {
7103 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7104 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7106 mutex_unlock(&mutex);
7110 int sched_rt_handler(struct ctl_table *table, int write,
7111 void __user *buffer, size_t *lenp,
7115 int old_period, old_runtime;
7116 static DEFINE_MUTEX(mutex);
7119 old_period = sysctl_sched_rt_period;
7120 old_runtime = sysctl_sched_rt_runtime;
7122 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7124 if (!ret && write) {
7125 ret = sched_rt_global_constraints();
7127 sysctl_sched_rt_period = old_period;
7128 sysctl_sched_rt_runtime = old_runtime;
7130 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7131 def_rt_bandwidth.rt_period =
7132 ns_to_ktime(global_rt_period());
7135 mutex_unlock(&mutex);
7140 #ifdef CONFIG_CGROUP_SCHED
7142 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7144 return css ? container_of(css, struct task_group, css) : NULL;
7147 static struct cgroup_subsys_state *
7148 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7150 struct task_group *parent = css_tg(parent_css);
7151 struct task_group *tg;
7154 /* This is early initialization for the top cgroup */
7155 return &root_task_group.css;
7158 tg = sched_create_group(parent);
7160 return ERR_PTR(-ENOMEM);
7165 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7167 struct task_group *tg = css_tg(css);
7168 struct task_group *parent = css_tg(css_parent(css));
7171 sched_online_group(tg, parent);
7175 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7177 struct task_group *tg = css_tg(css);
7179 sched_destroy_group(tg);
7182 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7184 struct task_group *tg = css_tg(css);
7186 sched_offline_group(tg);
7189 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7190 struct cgroup_taskset *tset)
7192 struct task_struct *task;
7194 cgroup_taskset_for_each(task, css, tset) {
7195 #ifdef CONFIG_RT_GROUP_SCHED
7196 if (!sched_rt_can_attach(css_tg(css), task))
7199 /* We don't support RT-tasks being in separate groups */
7200 if (task->sched_class != &fair_sched_class)
7207 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7208 struct cgroup_taskset *tset)
7210 struct task_struct *task;
7212 cgroup_taskset_for_each(task, css, tset)
7213 sched_move_task(task);
7216 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7217 struct cgroup_subsys_state *old_css,
7218 struct task_struct *task)
7221 * cgroup_exit() is called in the copy_process() failure path.
7222 * Ignore this case since the task hasn't ran yet, this avoids
7223 * trying to poke a half freed task state from generic code.
7225 if (!(task->flags & PF_EXITING))
7228 sched_move_task(task);
7231 #ifdef CONFIG_FAIR_GROUP_SCHED
7232 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7233 struct cftype *cftype, u64 shareval)
7235 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7238 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7241 struct task_group *tg = css_tg(css);
7243 return (u64) scale_load_down(tg->shares);
7246 #ifdef CONFIG_CFS_BANDWIDTH
7247 static DEFINE_MUTEX(cfs_constraints_mutex);
7249 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7250 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7252 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7254 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7256 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7257 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7259 if (tg == &root_task_group)
7263 * Ensure we have at some amount of bandwidth every period. This is
7264 * to prevent reaching a state of large arrears when throttled via
7265 * entity_tick() resulting in prolonged exit starvation.
7267 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7271 * Likewise, bound things on the otherside by preventing insane quota
7272 * periods. This also allows us to normalize in computing quota
7275 if (period > max_cfs_quota_period)
7278 mutex_lock(&cfs_constraints_mutex);
7279 ret = __cfs_schedulable(tg, period, quota);
7283 runtime_enabled = quota != RUNTIME_INF;
7284 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7285 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7286 raw_spin_lock_irq(&cfs_b->lock);
7287 cfs_b->period = ns_to_ktime(period);
7288 cfs_b->quota = quota;
7290 __refill_cfs_bandwidth_runtime(cfs_b);
7291 /* restart the period timer (if active) to handle new period expiry */
7292 if (runtime_enabled && cfs_b->timer_active) {
7293 /* force a reprogram */
7294 cfs_b->timer_active = 0;
7295 __start_cfs_bandwidth(cfs_b);
7297 raw_spin_unlock_irq(&cfs_b->lock);
7299 for_each_possible_cpu(i) {
7300 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7301 struct rq *rq = cfs_rq->rq;
7303 raw_spin_lock_irq(&rq->lock);
7304 cfs_rq->runtime_enabled = runtime_enabled;
7305 cfs_rq->runtime_remaining = 0;
7307 if (cfs_rq->throttled)
7308 unthrottle_cfs_rq(cfs_rq);
7309 raw_spin_unlock_irq(&rq->lock);
7312 mutex_unlock(&cfs_constraints_mutex);
7317 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7321 period = ktime_to_ns(tg->cfs_bandwidth.period);
7322 if (cfs_quota_us < 0)
7323 quota = RUNTIME_INF;
7325 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7327 return tg_set_cfs_bandwidth(tg, period, quota);
7330 long tg_get_cfs_quota(struct task_group *tg)
7334 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7337 quota_us = tg->cfs_bandwidth.quota;
7338 do_div(quota_us, NSEC_PER_USEC);
7343 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7347 period = (u64)cfs_period_us * NSEC_PER_USEC;
7348 quota = tg->cfs_bandwidth.quota;
7350 return tg_set_cfs_bandwidth(tg, period, quota);
7353 long tg_get_cfs_period(struct task_group *tg)
7357 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7358 do_div(cfs_period_us, NSEC_PER_USEC);
7360 return cfs_period_us;
7363 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7366 return tg_get_cfs_quota(css_tg(css));
7369 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7370 struct cftype *cftype, s64 cfs_quota_us)
7372 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7375 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7378 return tg_get_cfs_period(css_tg(css));
7381 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7382 struct cftype *cftype, u64 cfs_period_us)
7384 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7387 struct cfs_schedulable_data {
7388 struct task_group *tg;
7393 * normalize group quota/period to be quota/max_period
7394 * note: units are usecs
7396 static u64 normalize_cfs_quota(struct task_group *tg,
7397 struct cfs_schedulable_data *d)
7405 period = tg_get_cfs_period(tg);
7406 quota = tg_get_cfs_quota(tg);
7409 /* note: these should typically be equivalent */
7410 if (quota == RUNTIME_INF || quota == -1)
7413 return to_ratio(period, quota);
7416 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7418 struct cfs_schedulable_data *d = data;
7419 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7420 s64 quota = 0, parent_quota = -1;
7423 quota = RUNTIME_INF;
7425 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7427 quota = normalize_cfs_quota(tg, d);
7428 parent_quota = parent_b->hierarchal_quota;
7431 * ensure max(child_quota) <= parent_quota, inherit when no
7434 if (quota == RUNTIME_INF)
7435 quota = parent_quota;
7436 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7439 cfs_b->hierarchal_quota = quota;
7444 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7447 struct cfs_schedulable_data data = {
7453 if (quota != RUNTIME_INF) {
7454 do_div(data.period, NSEC_PER_USEC);
7455 do_div(data.quota, NSEC_PER_USEC);
7459 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7465 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7466 struct cgroup_map_cb *cb)
7468 struct task_group *tg = css_tg(css);
7469 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7471 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7472 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7473 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7477 #endif /* CONFIG_CFS_BANDWIDTH */
7478 #endif /* CONFIG_FAIR_GROUP_SCHED */
7480 #ifdef CONFIG_RT_GROUP_SCHED
7481 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7482 struct cftype *cft, s64 val)
7484 return sched_group_set_rt_runtime(css_tg(css), val);
7487 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7490 return sched_group_rt_runtime(css_tg(css));
7493 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7494 struct cftype *cftype, u64 rt_period_us)
7496 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7499 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7502 return sched_group_rt_period(css_tg(css));
7504 #endif /* CONFIG_RT_GROUP_SCHED */
7506 static struct cftype cpu_files[] = {
7507 #ifdef CONFIG_FAIR_GROUP_SCHED
7510 .read_u64 = cpu_shares_read_u64,
7511 .write_u64 = cpu_shares_write_u64,
7514 #ifdef CONFIG_CFS_BANDWIDTH
7516 .name = "cfs_quota_us",
7517 .read_s64 = cpu_cfs_quota_read_s64,
7518 .write_s64 = cpu_cfs_quota_write_s64,
7521 .name = "cfs_period_us",
7522 .read_u64 = cpu_cfs_period_read_u64,
7523 .write_u64 = cpu_cfs_period_write_u64,
7527 .read_map = cpu_stats_show,
7530 #ifdef CONFIG_RT_GROUP_SCHED
7532 .name = "rt_runtime_us",
7533 .read_s64 = cpu_rt_runtime_read,
7534 .write_s64 = cpu_rt_runtime_write,
7537 .name = "rt_period_us",
7538 .read_u64 = cpu_rt_period_read_uint,
7539 .write_u64 = cpu_rt_period_write_uint,
7545 struct cgroup_subsys cpu_cgroup_subsys = {
7547 .css_alloc = cpu_cgroup_css_alloc,
7548 .css_free = cpu_cgroup_css_free,
7549 .css_online = cpu_cgroup_css_online,
7550 .css_offline = cpu_cgroup_css_offline,
7551 .can_attach = cpu_cgroup_can_attach,
7552 .attach = cpu_cgroup_attach,
7553 .exit = cpu_cgroup_exit,
7554 .subsys_id = cpu_cgroup_subsys_id,
7555 .base_cftypes = cpu_files,
7559 #endif /* CONFIG_CGROUP_SCHED */
7561 void dump_cpu_task(int cpu)
7563 pr_info("Task dump for CPU %d:\n", cpu);
7564 sched_show_task(cpu_curr(cpu));